content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
New methods and tools are needed to meet the challenges in the development of complex socio-technical systems, such as sustainable mobility solutions in metropolitan regions \cite{Ncube2018}.
Systems of connected electrified vehicles can be characterised as a \textit{system of systems} (SoS), where the vehicle can be seen as a \textit{constituent system} (CS) that interacts with changing other CSs to provide an SoS functionality \cite{Hoehne2018}.
An interdisciplinary approach for the realization of these systems is \textit{system of systems engineering} (SoSE).
The definition of stakeholder needs and required functionalities are key elements of SoSE \cite{INCOSE2015}; the precise specification of requirements is a basis for the system decomposition and implementation, or the selection of suitable CSs that form an SoS \cite{Odusd2008}. However, in SoSE, there are different requirements engineering (RE) challenges compared to RE in established systems engineering (SE) processes \cite{Nielsen2015}
According to Maier et al. \cite{Maier1996}, the operational, managerial, and evolutionary independence are the essential characteristics of an SoS. These characteristics have a significant influence on the applicability of existing RE techniques
\cite{Ncube2011,Ncube2018,Nielsen2015}.
In contrast to monolithic systems, SoS consist of individual systems that can operate independently and perform a meaningful task, even when not part of an SoS. The development and operation of the CSs is managed independently, in different organizations with different development- and product life cycles. Also, requirements on the CS- and SoS level change frequently and independently, leading to an evolutionary development \cite{Nielsen2015,Maier1996}.
Based on these SoS characteristics, Ncube and Lim \cite{Ncube2011} describe challenges for the SoS RE process: Due to the different systems in an SoS, requirements cover many different disciplines, can be contradictory, unknown or possibly not fully defined. These difficulties overlap with the fundamental problems in RE \cite{Fernandez2013}, but, according to Ncube and Lim \cite{Ncube2011}, requirements in an SoS additionally must be considered as requirements for the SoS, which describe the properties of the overall system, or requirements for a CS that describe capabilities of a single system. Since requirements on both levels can change continuously and independently, traditional, linear and top-down requirements specification and decomposition techniques can not be used \cite{Ncube2011,Ncube2018,Nielsen2015}.
To address this problem we propose an iterative and scenario-based requirements specification technique.
Based on previous work \cite{Wiecher2020a,Wiecher2019,Wiecher2020} we integrate the Scenario Modeling Language for Kotlin (SMLK) with agile development techniques to support the requirements engineer in the continuous and iterative specification, formalization, and validation of requirements on different levels of abstraction.
This paper makes the following two contributions:
First (1), we extend SMLK to enable requirements engineers to intuitively, but formally model the requirements on the SoS-level as well as the interaction between the CSs (CS-level). With these extensions, requirements can be specified and validated independently, which addresses the managerial and operational independence of systems. Nevertheless, both levels of abstraction are connected to allow for the joint execution and testing of the specified behavior on the SoS- and CS-level, in order to detect and resolve contradictions in the requirements on both these levels.
Second (2), we propose a specification method where we combine behavior-driven development (BDD) and test-driven development (TDD) with the scenario-based modeling technique. This enables the iterative specification of system features and usage scenarios to document stakeholder expectations and generate tests steps, which subsequently drive the scenario-based modeling of the system specification.
While numerous approaches exist that suggest using formal scenario models to bridge the gap from informal requirements to the implementation of software-intensive systems \cite{Damas2006,Whittle2000,Harel2002d,Sutcliffe2003}, the particular contribution of this paper is the extension of sce\-nario-based modeling and programming techniques based on LSC Play-Out~\cite{Harel2002d} and behavioral programming (BP)~\cite{Harel2012} with BDD and TDD. Enabling this combination of agile development techniques with scenario-based requirements modeling addresses the \textit{coverage and sampling concerns} in scenario-based requirements engineering \cite{Sutcliffe2003}: by connecting features with tests (BDD), and tests with the scenario-based requirements model (TDD), we can ensure that every feature is modeled by an appropriate set of scenarios, and that these scenarios are validated by an appropriate set of tests.
We asses the applicability with a proof-of-concept e-mobility application and provide a demonstration tool\footnote{https://bitbucket.org/crstnwchr/besos (includes the proof-of-concept example)}\footnote{https://bitbucket.org/jgreenyer/smlk/ (required to build the example project)} to enable others to use, evolve, and evaluate our approach.
Structure: We describe background in Sect.\,\ref{sec:background}, the scenario-based requirements specification method in Sect.\,\ref{sec:SosRequirementsSpecification}, and the proof-of-concept application in Sect.\,\ref{sec:proofOfConcept}. We report related work in Sect.\,\ref{sec:relatedWork} and conclude in Sect.\,\ref{sec:outlook}.
\section{Background}
\label{sec:background}
\subsection{System of Systems Engineering (SoSE)}
For the description of System of Systems (SoS) no generally valid definition yet exists \cite{Albers2018,Nielsen2015}. Hence, a distinction between complex monolithic systems and SoS is often made by the system characteristics. Therefore Maier describes five key characteristics of SoS \cite{Maier1996}:
(1) \textit{Operational Independence}: Each system that is part of the SoS is independent and can perform a meaningful task, even if it is not integrated into the SoS.
(2) \textit{Managerial Independence}: The individual systems are self-administered and individually managed. Consequently they collaborate with the other systems of the SoS, but they operate independently.
(3) \textit{Geographic Distribution}: The individual systems of the SoS are distributed over large spatial distances, which means that the exchange of information between the individual systems is of primary importance for collaboration.
(4) \textit{Evolutionary Development}: The objectives and functionality of an SoS can change constantly, as they can be added, modified or removed based on experience. Therefore an SoS never appears to be fully completed.
(5) \textit{Emergent Behavior}: By the collaboration of the individual systems, a synergism is achieved in which the SoS fulfils a purpose that cannot be achieved by or attributed to any of the individual systems.
These characteristics have a strong influence on the SoS development. To support a structured SoS development Dahmann et al. \cite{Dahmann2008} describe the differences between systems engineering (SE) and SoS engineering (SoSE).
Accordingly, SE and SoSE both start with identifying and understanding user capability objectives in order to derive technical requirements for the system to be developed. In SE we subsequently continue with a top-down requirements decomposition and system design, with clear responsibilities in the management and engineering of the system \cite{Gausemeier2002}.
In SoSE, by contrast, the identified objectives and requirements serve as a basis for the development of new systems \emph{or} the integration of existing systems to build the SoS. Particularly the operational and managerial independence of individual systems is challenging: the existing systems may also fulfill other purposes that may conflict with the SoS objectives and those of its CS. Therefore it is important to understand how the individual systems behave and how this behavior contributes to the overall SoS behavior.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\linewidth]{SoSTypes}
\caption{Different types of SoS \cite{Maier1996,Dahmann2008}}
\label{fig:SoSTypes}
\end{figure}
When starting the SoS development it is important to categorize the SoS to be developed at an early stage because this has a significant influence on the RE approach that can be applied \cite{Nielsen2015,Dahmann2008,Ncube2018}. Fig.\,\ref{fig:SoSTypes} shows four different SoS types, initially introduced by Maier \cite{Maier1996} and extended by Dahmann and Baldwin \cite{Dahmann2008}:
A \emph{directed} SoS is designed for specific purposes. The individual systems have the ability to operate independently but are managed by a SoSE Team in a way that they fulfill a specific purpose. In an
\emph{acknowledged} SoS the SoSE Team recognises and defines a common purpose and goal, but the CSs retain independent control and goals. The continuous and evolutionary development of the common purpose is based on collaboration between the SoS and the CSs.
In a \emph{collaborative} SoS the individual systems are not bound to follow a central management, but voluntarily participate in a collaboration in order to achieve the SoS goal.
A \emph{virtual} SoS has neither a leading control nor a common goal. This leads to a high degree of emergent behavior where the exact means and structures that produce the functionality of the system are difficult to recognize and distinguish \cite{Nielsen2015,Odusd2008}.
This paper focuses on acknowledged SoS and we introduce an example next.
\subsection{Example of Application}
\label{sec:ExampleOfApplication}
To illustrate our approach, we introduce an e-mobility system of systems. In \cite{Kirpes2019} Kirpes et al. introduce an architecture model that provides an integrative view on former separated areas of electricity, individual mobility, and information and communication technologies to realize future e-mobility SoS.
Based on the example defined in \cite{Kirpes2019}, Fig.\,\ref{fig:SoSExample} shows an SoS user who is interacting with an e-mobility SoS. The main interest of the user is to improve the e-mobility experience and to reduce its costs.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{example}
\caption{Smart charging as an acknowledged SoS. Based on \cite{Ncube2018} and \cite{Kirpes2019}.}
\label{fig:SoSExample}
\end{figure}
These user interests are targeted by a high-level use case that describes how to create an optimized travel plan. At the beginning the user enters travel preferences like start and destination into the smartphone app (APP). The APP then requests further data from other systems that are necessary for the calculation of an optimized route.
For example, GPS data of possible routes are requested from a route-planning service (RPS). Usage data of available charging points along the routes are provided by a charging station operation service (CSOS). Recommendations for a battery-saving charging process are provided by a battery-health service (BHS). And information on current electricity prices in the region is provided by the energy-information service (EIS).
These different systems, which are required to provide information to calculate an optimal route based on user preferences, are developed and managed by four different system owners (OEM, Map Service Provider, Charging Operator, and Energy Provider). We also see one SoSE Team which defines the overall SoS functionality, directs the operations, and has a contractual relationship with the owners of the CSs. According to \cite{Maier1996} and \cite{Dahmann2008} this example has the characteristics of an acknowledged SoS: we have recognised requirements, objectives and responsibilities on the SoS level and a contractual relationship between the SoSE Team and the individual constituent systems owner. However, the constituent systems keep their own management, funding and development approaches (cf. \cite{Ncube2018}).
\subsection{Scenario Modeling Language for Kotlin (SMLK)}
\label{sec:SMLK}
SMLK is a Kotlin-based implementation of the Behavioral Programming (BP) paradigm \cite{Harel2012}. In BP, a program consists of a number of \textit{behavioral threads}, which we also call \textit{scenarios}. Scenarios are loosely coupled via shared events and can model individual behavioral aspects or functional requirements of a system. Scenarios can \textit{request} events that shall happen, be \textit{triggered by} or \textit{wait for} events requested by other scenarios, or (temporarily) \textit{forbid}/\textit{block} events. During execution, the scenarios are interwoven to yield a coherent system behavior that satisfies the requirements of all scenarios.
Listing~\ref{list:interSystemObjectEvent} shows two SMLK scenarios that can be represented graphically as shown in Fig.\,\ref{fig:example-scenario-SD}.
Both scenarios are triggered by the event of a user entering the travel preferences in the app. This event is modeled as an interaction event of the object \lstinlineKotlin{user} sending the object \lstinlineKotlin{app} a message
\lstinlineKotlin{addTravelPreferences}. In the first scenario, the parameters
\lstinlineKotlin{fromLoc} and
\lstinlineKotlin{toLoc} are variables bound to the parameter values carried by the triggering event when the scenario is triggered and initialized.
The SMLK code in the listing shows this binding of the parameter values explicitly (lines 2 and 3). The second scenario is triggered by the same event, but does not use the parameter values; the sequence diagram expresses this by using asterisks.
After the trigger event, the first scenario requests that the \lstinlineKotlin{app} sends the Route Planning Service (\lstinlineKotlin{rps})
a message to calculate the route between
\lstinlineKotlin{fromLoc} and
\lstinlineKotlin{toLoc}, and then requests that the \lstinlineKotlin{rps} shall respond with a route. Then the \lstinlineKotlin{app} shall optimize the route and show it to the user.
The second scenario describes the interaction of the \lstinlineKotlin{app} and the Charging Station Operating System (\lstinlineKotlin{csos}). After the triggering event, the scenario requests that the \lstinlineKotlin{app} sends the \lstinlineKotlin{csos} a request to send GPS position data of available charging stations. The scenario then requests that the \lstinlineKotlin{csos} shall respond with such a list. This interaction must happen before the app optimizes the route, i.e., the event \lstinlineKotlin{app.optimizeRoute()} is blocked until the second scenario terminates; only then can be first scenario proceed.
In these example scenarios, the route details and charging location list contents are not relevant, so mock instances are created by helper functions. When at a later point the behavior is refined, these parameter values may be replaced by other values, e.g., a detailed and correct route may be calculated elsewhere. The scenario method \lstinlineKotlin{requestParamValuesMightVary} allows us to request events with supplied default parameter values, but it will accept also events sent between the same objects, and with the same signature, but with different parameter values.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{example-scenario-SD}
\caption{Graphical representation of the SMLK scenario in Listing~\ref{list:interSystemObjectEvent}}
\label{fig:example-scenario-SD}
\end{figure}
\vspace{-0.1cm}
\begin{lstlisting}[caption=Example scenario from the e-mobility system specification,
label=list:interSystemObjectEvent,
style=KotlinStyle
]
scenario(user sends (app receives App::addTravelPreferences)){
val fromLoc = it.parameters[0] as String
val toLoc = it.parameters[1] as String
request(app sends rps.calculateRoute(fromLoc, toLoc))
val route = createMockRoute()
requestParamValuesMightVary(rps sends app.calculateRouteResponse(route)
request(app.optimizeRoute())
request(app sends user.showMapWithOptimizedRoute())
},
scenario(user sends (app receives App::addTravelPreferences)){
scenario {
request(app sends csos.chargingStationGpsDataRequest())
val chargingStationsList = createMockChargingStationsList()
requestParamValuesMightVary(csos sends app.considerChargingStationLocations(chargingStationsList))
}.before(app.optimizeRoute())
}
\end{lstlisting}
\section{Scenario-based Requirements Specification in a System of Systems Context}
\label{sec:SosRequirementsSpecification}
To develop an SoS, usually existing systems are integrated by new systems to comprise a new SoS. While the new systems may be under a direct managerial and operational control, existing systems may be under the managerial and operational control of another organization.
Over time, systems that are under external control may change, which leads to the necessity to continuously (1) analyze how the changes in one system impact the SoS functionality, and (2) how other systems may have to be adapted to ensure that the SoS functionality can still be provided. This requires the SoSE team to continuously analyze, specify, and align requirements across different hierarchy levels.
Our scenario-based requirements specification approach supports an iterative and integrated behavior modeling and analysis on the SoS and CS level.
Based on the definitions in \cite{Harel2020} we introduce the term \emph{inter-system scenarios} to model the behavior on the SoS level and \emph{intra-system scenarios} to model the CS behavior. Also we show how both views can be integrated to allow for the joint execution and testing of the integrated SoS and CS behavior.
\subsection{Inter-System Scenarios}
The goal of modeling inter-system scenarios is to conceive an validate how SoS use cases can be realized by the interaction of users, existing systems, and new systems to be developed.
The inter-system scenario modeling process starts by defining the use cases, the structural SoS architecture, and then detailing and validating the use cases using scenarios and repeated simulation.
When modeling this behavior, certain assumptions are made about the behavior of the existing systems, possibly based on available documentation or communication with experts from the respective organizations.
Two exemplary inter-system scenarios are already introduced in Listing \ref{list:interSystemObjectEvent}, where we first modeled the interaction between the app, rps and the SoS user, and in the second scenario, between the user, app and csos.
In this example, we see that we are able to model the interaction between selected systems, where new requirements can be considered by iteratively adding new scenarios to the \emph{SoS scenario specification}. By adding these inter-system scenarios the introduced modeling concepts allow to focus on a high level system interaction;
Although we are able to partly ignore specification details (e.g. exact route information in Listing \ref{list:interSystemObjectEvent} line 14), we are able to execute and validate the interaction between the CS. This supports on the SoSE team to get a better understanding of the overall system behavior.
\subsection{Intra-System Scenarios}
\label{sec:intraSystem}
Once a satisfactory concept of the inter-system behavior is established, the inter-system specification must be supplemented and refined in two ways:
First (1), it is necessary to specify the behavior of the existing systems in more detail in order to validate whether the inter-system interaction behavior is indeed aligned with the behavior of the existing systems.
Second (2), the behavior of the new systems to be developed must be detailed, possibly detailing their component structure and internal interactions, in order to provide a thorough basis for their development.
Our approach supports modeling the behavior on this more detailed hierarchy level with scenarios as well, and even to integrate their execution in order to simulate and validate behavioral requirements consistency across the different hierarchy levels.
To better distinguish between these two hierarchy levels, we distinguish the \textit{inter-system} level and \textit{intra-system} level as outlined in Fig.\,\ref{fig:interactionSosCsLevel}.
The SoS scenario specification is located in the inter-system view, and individual CS scenario specifications are located in the intra-system view. When defining the internal behavior of a selected CS, we switch the perspective from the SoSE Team to a systems owner who is responsible for the development of a system. This can be e.g. the map service provider who is responsible for the development of the rps (see Fig.\,\ref{fig:SoSExample}).
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{intraSystem.pdf}
\caption{Inter- and intra-system view to continuously concretise requirements on CS level, while also considering the overall SoS behavior.}
\label{fig:interactionSosCsLevel}
\end{figure}
The intra-system scenarios are added to an individual CS scenario specification, with the goal to model requirements which are needed to build the CS and its subsystems. One example intra-system scenario is shown in Listing \ref{list:intraScenarioSpecification}.
\begin{lstlisting}[caption=CS scenario specification of the RPS,
label=list:intraScenarioSpecification,
style=KotlinStyle
]
scenario(routeRequester sends(rps receives Rps::calculateRoute)){
val fromLocString = it.parameters[0] as String
val toLocString = it.parameters[1] as String
request(rpsController sends gpsService.getLocations(fromLocString, toLocString))
val fromLoc = getLocation(fromLocString)
val toLoc = getLocation(toLocString)
request(gpsService sends rpsController.locations(fromLoc, toLoc))
request(rpsController sends routePlaner.calculateRoute(fromLoc, toLoc))
val route = calculateRoute(fromLoc, toLoc)
request(routePlaner sends rpsController.calculatedRoute(route))
request(rpsController sends routeRequester.calculateRouteResponse(route))
}
\end{lstlisting}
The scenario specifies how the internal components of the rps (rpsController, gspService, and routePlanner) interact when receiving a request to calculate a route. Eventually (line 11), the calculated route will be returned to the requesting object.
When looking at the scenario in more detail, we see that the scenario is triggered when a \lstinlineKotlin{routeRequester} sends the \lstinlineKotlin{rps} the message \lstinlineKotlin{calculateRoute}.
This event is requested on the inter-system level, see the first scenario in Listing~\ref{list:interSystemObjectEvent} (line 4).
One difference is, however, that in the intra-system scenario, we abstract from the app as being the source of the \lstinlineKotlin{calculateRoute} request (and the recipient of the route as a reposonse, see line 11). Instead, we assume that there is an abstract external route-requesting entity that requests a route to be calculated by the rps. We do this to separate the intra-system specification of a system from the particular SoS context defined on the inter-system level, as the system may also be used in other contexts.
The inter-system and intra-system level scenario execution can nevertheless be integrated, because the type of \lstinlineKotlin{routeRequester} is an interface that is also implemented by \lstinlineKotlin{app} (without showing the code in more detail for brevity). Hence it is possible that the event of the app requesting to calculate a route triggers the scenario shown here, and indeed the app would then receive the calculated route as a response.
The event parameters on the intra-system level may vary or be more detailed than the values assumed on the inter-system level where, for example, we used simple mock values (see Listing~\ref{list:interSystemObjectEvent}, lines 5 and 13). It is possible for intra-system scenarios to provide more detailed parameter values where the inter-system level scenarios request events by using the \lstinlineKotlin{requestParamValuesMightVary} command. (see Listing~\ref{list:interSystemObjectEvent} line 14).
\subsection{Specification Method}
To support the requirements engineer in modeling system requirements with SMLK, we propose an iterative method based on agile techniques. Fig.\,\ref{fig:iterativeSpecification} shows an overview of the single steps. We start with the specification of the inter-system behavior by applying the BDD approach. Here we first define the expected system behavior from the SoS user perspective.
Therefor we create a \emph{SoS feature specification} where each feature is defined by one or more \emph{usage scenarios} written in the gherkin syntax\footnote{https://cucumber.io/docs/gherkin/}.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{iterativeSpecification.pdf}
\caption{Continuous and iterative scenario specification}
\label{fig:iterativeSpecification}
\end{figure}
Listing~\ref{list:gherkinFeature1} shows a first feature specification that describes a user interaction with the app. On this hierarchy level, the SoS feature specification allows the SoSE team to define what is expected from the SoS and to document this expectations in a comprehensible form.
\begin{lstlisting}[caption=Initial feature specification including a usage scenario to describe the user interaction with the SoS.,
label=list:gherkinFeature1,
style=GherkinStyle
]
Feature: Retrieve travel preferences and display optimized route
Scenario: Add travel preferences to the app
When the SoS user adds travel preferences to the app
Then the app displays a set of optimized routes
\end{lstlisting}
Based on this SoS feature specification we generate test skeletons as shown in Listing~\ref{list:testSkeletons}.
These test skeletons are then used to drive the modeling of the inter-system behavior. To support a structured and iterative modeling of system requirements, we embed the Test-Driven Scenario Specification (TDSS) \cite{Wiecher2019} into the BDD approach. In this way, we combine the comprehensible specification of expected system behavior with the formal and scenario-based modeling of system requirements.
\begin{lstlisting}[caption=Generated test steps.,
label=list:testSkeletons,
style=GherkinStyle
]
When("^the EV user adds travel preferences to the App$") {
//implement here
}
Then("^the App displays a set of optimized routes$") {
//implement here
}
\end{lstlisting}
The TDSS approach includes the steps outlined in Fig.\,\ref{fig:tdss}. In the first step we extend the generated test skeletons (1). Here, we e.g. model that the user adds travel preferences to the app (Listing~\ref{list:testSteps} line 2) and eventually receives a map with optimized routes (line 5).
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{tdss.pdf}
\caption{Test-Driven Scenario Specification (TDSS) \cite{Wiecher2019}}
\label{fig:tdss}
\end{figure}
After we added these functions we execute the SoS feature specification (2) whereupon the single test steps and finally the events within the test steps are executed. At this point in time we did not model the inter-system behavior and consequently the test fails, because the app will not send the optimized route to the SoS user as expected in line 5.
\begin{lstlisting}[caption=Generated test steps.,
label=list:testSteps,
style=GherkinStyle
]
When("^the EV user adds travel preferences to the App$") {
trigger(user sends app.addTravelPreferences("Dortmund", "Paderborn"))
}
Then("^the App displays a set of optimized routes$") {
eventually(app sends user.mapWithOptimizedRoutes())
}
\end{lstlisting}
Therefore we extend our SoS scenario specification with the inter-system scenarios (3) which we already introduced in Listing \ref{list:interSystemObjectEvent}. We then run the test again to ensure that the modelled system requirements meet the expectations (4). If we have modeled additional tests in previous iterations, we now run them as well to ensure that there are no unexpected interactions between the individual tests and system requirements. If there are more requirements that need to be modeled, we perform further iterations. When all requirements on the SoS level known at this time have been modeled and tested, the SoS feature specification can be cleaned up. Afterwards the detailed specification of selected systems under development follows.
This iterative approach supports the modeling of the interaction of all CSs within the SoS.
In this way we are able to iteratively document the expectations from an SoS user perspective and model and test the interaction between the CSs. Thereby new systems and behavior can be added as needed to realize the expected behavior. When we have gone through several iterations, the SoSE team gets a better understanding of which systems are needed and what information these systems have to exchange with each other. Subsequently we can switch to the intra-system level and focus on the requirements specification for a selected CS within the SoS. Based on our example outlined in Fig.\,\ref{fig:SoSExample} we now switch from the SoSE team perspective to e.g. the perspective of the map service provider, who is responsible for the development of the rps. As shown in Fig.\,\ref{fig:iterativeSpecification} we execute the same specification method, but we create an independent \emph{CS feature specification}, generate independent test steps and create an CS scenario specification.
This allows the independent specification and modeling of the requirements for the CS, which addresses the managerial, operational and evolutionary independence of systems in an SoS. In this way, system requirements can be specified without seeing the system in an SoS context. But, at the same time, both views can be integrated (as described in Sec. \ref{sec:intraSystem}), which allows the joint execution of the SoS behavior and the internal behavior of single already specified systems.
In this way it's possible to detect contradictions between requirements on both levels. For example, if requirements have been specified at CS level that appear to have nothing to do with the SoS behavior but still influence the expected SoS behavior, the joint execution of the scenario specifications can be used to detect and resolve these dependencies.
\section{Proof of Concept}
\label{sec:proofOfConcept}
To assess the applicability of our approach we integrated SMLK with the BDD tool Cucumber and executed the previously described specification method based on the example introduced in Section \ref{sec:ExampleOfApplication}.
On SoS level we started with the feature specification as already shown in Listing~\ref{list:gherkinFeature1}. Subsequently we generated the test skeletons and added the SMLK events as shown in Listing \ref{list:testSteps}.
Following the TDSS approach we executed the SoS feature specification (Step 1 in Fig.\,\ref{fig:tdss}) and got a failed test result as shown in Fig.\,\ref{fig:firstSosTest}.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{firstSosTest.png}
\caption{First TDSS run on SoS level}
\label{fig:firstSosTest}
\end{figure}
Subsequently we extended the SoS scenario specification as shown in Listing \ref{list:interSystemObjectEvent} to specify the SoS behavior. After we added these scenarios we executed the test again and finally received the expected event, resulting in a positive test result as shown in Fig.\,\ref{fig:secondSosTest}.
\begin{figure}[ht]
\centering
\includegraphics[width=1.0\linewidth]{secondSosTest.png}
\caption{Execute tests after adapting the SoS scenario specification}
\label{fig:secondSosTest}
\end{figure}
After we successfully defined a first interaction on inter-system level, we switched to the intra-system level and added a CS scenario specification to model the internal behavior of the rps as shown in Listing \ref{list:intraScenarioSpecification}.
Now we executed the same SoS feature again resulting in a negative result, because the rps internal behavior was not yet specified and hence the CS scenario program didn't send the \lstinlineKotlin{calculateRouteResponse(route)} message to the app.
To fix this we executed the TDSS process within the intra-system view, based on the CS feature specification shown in Listing \ref{list:CSfeature}.
\begin{lstlisting}[caption=Feature on CS level,
label=list:CSfeature,
style=GherkinStyle
]
Feature: Calculate route - RPS
@RpsSystem
Scenario: Calculate route based on user travel preferences
When the app sends travel preferences to the rps
Then the rps responds route information including gps data
\end{lstlisting}
Finally we got passed test results again, but now we also considered the rps internal behavior specification. And, by using \emph{tags} within the different feature specifications (e.g. \lstinlineKotlin{@RpsSystem}) and by applying the concepts described in Sect. \ref{sec:intraSystem}, we were not only able to validate the integrated SoS and CS behavior, but we also could independently test the requirements of single CS.
To allow others to use, validate and evolve our approach, we describe the architecture and functional principles of the developed tool in \cite{Wiecher2021a} as a companion to this paper. Here, we also describe the method we outline in Fig.~\,\ref{fig:iterativeSpecification} in more detail. And we provide information about the necessary resources\footnote{https://bitbucket.org/crstnwchr/besos/}\footnote{https://bitbucket.org/jgreenyer/smlk/}\footnote{https://cucumber.io}\footnote{https://www.jetbrains.com/idea/} to build and execute the example we use in this paper.
\section{Related Work}
\label{sec:relatedWork}
In this paper we use SMLK, which was extended to support an iterative and continuous modeling of system behavior in an SoS context. This modeling language is based on Live Sequence Charts (LSCs) \cite{Damm2001}. A recent LSC variant are Modal Sequence Diagrams (MSDs) \cite{Harel2006}. By modeling behavioral requirements with the help of MSDs, different works argue that this formal requirements modeling can increase the requirements quality (e.g. \cite{Holtmann2014}, \cite{Fockel2016}),
but these approaches are based on traditional SE and do not consider the SoS characteristics and their impact on the requirements specification.
Harel et al. describe an extension to behavioral programming that allows the integration of behavioral programs that operate on different hierarchy levels and time scales~\cite{Harel2011}. Indeed, we also use this approach to integrate different SMLK scenario programs that execute the behavior on the inter- and intra-system level.
Simulation-based analysis and design is commonplace in cyber-physical systems of systems, e.g. using actor-oriented frameworks or co-simulation \cite{Fitzgerald2014,Lee2015}. We aim to provide similar means for the thorough specification and analysis of \textit{requirements} of systems of systems. To the best of our knowledge, this is a new approach.
Other works address model-based RE in the SoS context. Holt et al. describes an ontology for model-based SoS requirements engineering \cite{Holt2012}. Albers et al. show how SoS requirements can be specified based on use-cases and sequence diagrams within SysML \cite{Albers2015}. However, an early, iterative and formal specification of requirements, with the goal to execute and test these requirements specifications is not considered in these approaches.
\section{Summary and Outlook}
\label{sec:outlook}
In this paper, we propose a technology to continuously model behavior requirements in an SoS context. Our approach supports requirements engineers in the iterative specification, modeling and testing of requirements. With the use of SMLK, the system behavior can be modeled textually through scenarios. This scenario-based modeling is close to how engineers communicate system behavior and hence enables a feasible formalization of requirements. To further support and structure the formalization process, we integrated SMLK with agile techniques and appropriate tooling. This fosters the iterative formalization, and by testing the formalized requirements specifications, we get early feedback about the expected system behavior and possible contradictions in requirements. Due to the proposed coupling of inter- and intra- system scenarios, we are also able to execute and test the system behavior on different hierarchy levels. And by integrating the BDD tool cucumber, we are able to specify the expected system behavior with the help of features and usage scenarios written in natural language, which supports the communication of expected system behavior in a multi-disciplinary development team.
For future work, we plan to integrate our previous work \cite{Wiecher2019} and the modeling concepts shown in this paper with an automated test case creation proposed in \cite{Fischbach2020} to further reduce the modeling effort. Also, as already started in previous work \cite{Wiecher2020}, we plan to integrate the results of this paper in an automotive development process and validate the applicability within an ongoing research project. As shown in \cite{Wiecher2019}, we are able to find contradictions in automotive requirements specifications, but the open questions are if the approach is scalable and whether the effort for the requirements modeling is justified.
Another possible direction for future work is focusing on stakeholder needs in a SoS context. In this paper we already integrated the BDD approach to validate requirements and align stakeholder expectations. This could be done more systematically by integrating goal modeling approaches \cite{Aydemir2018}.
\bibliographystyle{splncs04}
|
\section{Introduction}
\subsection{Statement of the problem}
\label{sec:statement}
We consider solving the Helmholtz obstacle-scattering problem, where the obstacle traps geometric-optic rays, by the boundary-element method arising from the Galerkin method applied to boundary-integral-equation formulations of the PDE problem, and then solving the resulting linear systems with the generalised minimum residual method (GMRES). We now give details of each of these aspects.
\subsubsection{The scattering problem.}
Let $\Omega_-\subset \mathbb{R}^d$, $d=2,3$, be a bounded Lipschitz open set such that its open complement $\Omega_+:=\mathbb{R}^d \setminus \overline{\Omega_-}$ is connected; let $\Gamma:= \partial \Omega_-$, and let $n$ be the outward-pointing unit normal vector to $\Omega_-$. We consider the exterior Dirichlet and Neumann scattering problems. For simplicity, we consider the case when the boundary data comes from an incoming plane wave $u^I(x}%\mathbf{x}):= \exp({\rm i} k x}%\mathbf{x}\cdot \hat{a}}%\mathbf{a})$ for $\hat{a}}%\mathbf{a}\in\mathbb{R}^d$ with $\|\hat{a}}%\mathbf{a}\|_2=1$; i.e. we consider
\emph{either} the sound-soft \emph{or} the sound-hard plane-wave scattering problem defined by: given $k>0$ and the incident field $u^I$, find the total field $u$ satisfying
\begin{align}\label{eq:Helmholtz}
\left\{
\begin{aligned}
&\Delta u + k^2 u =0 \hspace{2.5cm}\text{ in }\Omega^+,\\
&\text{either } u \text{ or } \partial_n u = 0 \hspace{1.65cm}\text{ on }\Gamma, \quad\text{ and }\\
&\dfrac{\partial u^S }{\partial r} -{\rm i} ku^S = o \left(\frac{1}{r^{(d-1)/2}}\right) \quad\text{as }r:=|x|\rightarrow \infty, \text{ uniformly in $x/r$},
\end{aligned}
\right.
\end{align}
where \(u^S := u - u^I\) is the scattered field. We are particularly interested in the case when the frequency $k$ is large.
\subsubsection{Trapping and quasimodes}
We consider domains $\Omega_-$ such that there exist stable trapped geometric-optic rays in the exterior $\Omega_+$.
In this situation, the solution operator for the problem \eqref{eq:Helmholtz} can grow exponentially through an increasing sequence of frequencies. This phenomenon can be expressed via the notion of \emph{quasimodes}.
\begin{definition}[Quasimodes]\label{def:quasimodes}
A family of Dirichlet quasimodes of quality $\epsilon(k)$
is a sequence $\{(u_j,k_j)\}_{j=1}^\infty\subset H^1_{\rm loc}(\Omega_+)\times \mathbb{R}$
with $u_j=0$ on $\Gamma$
such that the frequencies $k_j\rightarrow \infty$ as $j \rightarrow \infty$ and there exists a compact subset $\mathcal{K}\subset \Omega_+$ such that, for all $j$, ${\rm supp}\, u_j \subset \mathcal{K}$,
\begin{equation*}
\mathbb{N}{(\Delta +k_j^2) u_j}_{L^2(\Omega_+)} \leq \epsilon(k_j), \quad\text{ and }\quad\mathbb{N}{u_j}_{L^2(\Omega_+)}=1.
\end{equation*}
The definition of Neumann quasimodes is analogous, with $u_j=0$ on $\Gamma$ replaced by $\partial_n u =0$ on $\Gamma$.
\end{definition}
By the results of \cite[Theorem 2]{Bu:98} (see also \cite{Vo:00}), if a family of Dirichlet or Neumann quasimodes exists, the quality $\epsilon(k)$ can be at most exponentially-small in $k$.
For simplicity, in our numerical experiments we focus on the case when $\Omega_-$ is \emph{either} one of the two ``horseshoe-shaped'' 2-d domains shown in Figure \ref{fig:geometries} (and defined precisely below) \emph{or} certain 3-d analogues (defined in \S\ref{sec:3-dexpts} below); in these cases
there exist quasimodes with exponentially-small quality, leading to exponential growth of the solution operator -- see Theorem \ref{thm:ellipse} below.
We emphasise however that
there exist quasimodes with superalgebraically small quality for a much larger class of obstacles (see \cite[Theorem 1]{CaPo:02}, \cite[Theorem 1]{St:00} and the discussion in \S\ref{sec:Weyl_new})
and our bound on the $k$-dependence of the number of GMRES iterations (in Theorem \ref{thm:main2})
hold in these more-general situations.
The existence of quasimodes is linked to the existence of \emph{resonances} (poles of the meromorphic continuation of the solution operator of \eqref{eq:Helmholtz} from $\Im k\geq 0$ to $\Im k<0$); the relationship between trapping, quasimodes, and resonances is a classic topic in scattering theory; see \cite{StVo:95, StVo:96, TaZw:98, St:99, St:00} and \cite[Chapter 7]{DyZw:19}.
\subsubsection{A particular class of $\Omega_-$ for which quasimodes exist.}\label{sec:statement:ellipse}
The following theorem is proved by combining \cite[Equation A.16]{BeChGrLaLi:11} and \cite[Theorem 3.1]{NgGr:13} (see \S\ref{sec:Mathieu}).
\begin{theorem}[Quasimodes when $\Omega_+$ contains part of an ellipse]\label{thm:ellipse}
Let $d=2$. Given $a_1>a_2>0$, let
\begin{equation}\label{eq:ellipse}
E:= \left\{(x_1,x_2) \, : \, \left(\frac{x_1}{a_1}\right)^2+\left(\frac{x_2}{a_2}\right)^2<1\right\}.
\end{equation}
Assume that $\Gamma$ coincides with the boundary of $E$ in the neighborhoods of the points \((0,\pm a_2)\),
and that $\overline{\Omega_+}$ contains the convex hull of the union of these neighbourhoods.
Then there exist families of Dirichlet and Neumann quasimodes with
\begin{equation}\label{eq:quality}
\epsilon(k)=C_1 \exp( - C_2 k) \quad\text{ for all } k>0.
\end{equation}
where $C_1, C_2>0$ are both independent of $k$.
\end{theorem}
For $\Omega_-$ satisfying the assumptions of Theorem \ref{thm:ellipse}, we can compute the frequencies $k_j$ in the quasimodes. Indeed, the functions $u_j$ in the quasimode construction in \cite{BeChGrLaLi:11}/\cite{NgGr:13} are based on the family of eigenfunctions of the ellipse localising around the periodic orbit $\{(0,x_2) : |x_2|\leq a_2\}$ (i.e.~the minor axis of the ellipse);
when the eigenfunctions are sufficiently localised, the eigenfunctions multiplied by a suitable cut-off function form a quasimode, with frequencies $k_j$ equal to the square roots of the respective eigenvalues of the ellipse. By separation of variables, $k_j$ can be expressed as the
solution of a multiparametric eigenvalue problem involving Mathieu functions; see Appendix \ref{sec:Mathieu}.
We use the method introduced in~\cite{Wi:06} and the associated MATLAB toolbox to solve these eigenvalue problems for $k_j$.
When giving values of these $k_j$s we give all the digits computed in double precision.
Note that we are not claiming that all these digits are accurate (see \cite{Wi:06} for some discussion on accuracy), but
some of the quantities we compute below are very sensitive to the precise values of $k$, and so we give the exact values of $k$ used in our computations.
When giving specific values of these $k_j$, we use the notation from \cite[Appendix A]{BeChGrLaLi:11}, recapped in Appendix \ref{sec:Mathieu}, that $k_{m,n}^e$ and $k_{m,n}^o$ are the frequencies associated with the eigenfunctions of the ellipse that are even/odd, respectively, in the angular variable, with $m$ zeros in the radial direction (other than at the centre or the boundary) and $n$ zeros in the angular variable in the interval $[0,\pi)$. Note that the values of $k_{m,n}^e$ and $k_{m,n}^o$ are different for Dirichlet and Neumann boundary conditions, but we do not indicate this difference in our notation.
The eigenfunctions associated to $k_{m,n}^e$, $k_{m,n}^o$ localise about the minor axis as $m\rightarrow \infty$ for fixed $n$ (see the proof of Theorem \ref{thm:ellipse} in Appendix \ref{sec:Mathieu}); therefore quasimodes exist for the families of frequencies $\{k^{e/o}_{m,n}\}_{m=1}^\infty$ for fixed $n$.
\subsubsection{Definitions of the ``small cavity'' and ``large cavity'' obstacles $\Omega_-$.}\label{sec:smalllargegeometries}
Our numerical experiments focus on two specific $\Omega_-$ satisfying the assumptions of Theorem \ref{thm:ellipse}
with $a_1=1$ and $a_2=1/2$. We define the \emph{small cavity} as the region between the two elliptic arcs
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{geometries.pdf}
\caption{The ``small'' and ``large'' cavities $\Omega_-$, with the small cavity shaded in grey, and the large cavity equal to the union of the small cavity and the dashed region. We write the incident-plane-wave direction $\widehat{a}= (\cos\theta,\sin\theta)$, with the angle $\theta$ measured in the positive direction from the horizontal, as pictured. }\label{fig:geometries}
\end{figure}
\begin{figure}[h!]
\centering\includegraphics[width=0.8\textwidth]{elliptic_cavity_2D_output_20_combined_dir_direct_100.pdf}
\caption{Absolute value of the total field \(u\) defined by \eqref{eq:Helmholtz} with Dirichlet boundary conditions, $\Omega_-$ the small cavity, $k=100$, \(\hat{a}=(\cos(\theta),\sin(\theta))\), and \(\theta=4\pi/10\).
}\label{fig:example}
\end{figure}
\begin{align*}
&(\cos (t), 0.5 \sin (t)), \quad t\in [-\phi_0,\phi_0]
\quad\text{ and } \quad
(1.3\cos (t), 0.6 \sin (t)),\quad t\in [-\phi_1,\phi_1] \\
&\qquad\qquad\text{ with } \phi_0=7\pi/10\quad\text{ and } \quad \phi_1= \arccos \left(\frac{1}{1.3}\cos (\phi_0)\right);
\end{align*}
this corresponds to the shaded interior of the solid lines in Figure \ref{fig:geometries}.
We define the \emph{large cavity} as the region between the two arcs now with \(\phi_0=9\pi/10\).
We also consider 3-d analogues of the these cavities, created by rotating them around the $x_1$ axis.
Figure~\ref{fig:example} plots the absolute value of the total field \(u\) satisfying \eqref{eq:Helmholtz} with Dirichlet boundary conditions with $\Omega_-$ the small cavity, $k=100$ and \(\hat{a}=(\cos(\theta),\sin(\theta))\) with \(\theta=4\pi/10\);
this figure was produced by computing the unknown Neumann trace using BEM, and then evaluating the solution given
in terms of layer potentials by Green's integral representation \eqref{eq:Green}.
\subsubsection{Boundary-integral-equation (BIE) formulations of \eqref{eq:Helmholtz}}\label{sec:BIE}
We are primarily interested in solving \eqref{eq:Helmholtz} by reformulating it as an integral equation on $\Gamma$; recall that this procedure has the advantage of converting a problem posed in an unbounded $d$-dimensional domain (i.e. $\Omega_+$) to a problem posed on a bounded $(d-1)$-dimensional domain (i.e. $\Gamma$).
However, the ideas behind our main results are applicable to other methods of solving the Helmholtz equation, and we discuss in \S\ref{sec:FEM} below the standard variational formulation, which is the basis of the finite-element method.
We consider \emph{direct} BIE formulations of \eqref{eq:Helmholtz}, i.e., ones in which the unknown is either the Neumann data (for the Dirichlet problem) or the Dirichlet data (for the Neumann problem); however, our results below also apply to \emph{indirect} formulations (where the unknown has less-immediate physical relevance; see \cite[Page 132]{ChGrLaSp:12}), since there is a close relationship between the integral operators of the direct and indirect formulations; see, e.g., \cite[Remark 2.24, \S2.6]{ChGrLaSp:12}. Once both $u$ and $\partial_n u$ are known on $\Gamma$, the solution in $\Omega_+$ can be obtained from Green's integral representation (\eqref{eq:Green} below).
For the Dirichlet problem we find $\partial_n u$ using the standard ``combined-field'' or ``combined-potential'' BIE
\begin{align}\label{eq:direct_combined_dir}
A_{k,\eta}' \partial_{n} u = \partial_{n} u^I - {\rm i} \eta u^I
\quad\text{ on } \Gamma,
\quad\text{ where } \quad
A_{k,\eta}' := \dfrac{1}{2}I+D'_k-{\rm i} \eta S_k,
\end{align}
where $\eta$ is the (arbitrary) ``coupling parameter'' and $S_k$ and $D'_k$ are the single-layer and adjoint-double-layer operators defined by \eqref{eq:SD'} below.
If $k>0$ and \(\Re(\eta)\neq 0\), then \(A'_{k,\eta}:L^2(\Gamma)\rightarrow L^2(\Gamma)\) is bounded and invertible (see, e.g., \cite[Theorem 2.27]{ChGrLaSp:12}). There has been much research on the question of how to best choose $\eta$,
starting from the works \cite{KrSp:83, Kr:85, Am:90} for the case when $\Omega$ is a ball; see the overviews in \cite[Chapter 5]{ChGrLaSp:12}, \cite[\S7]{BaSpWu:16}, \cite[\S6.5]{ChSpGiSm:20}. Roughly speaking, the best choice for large $k$ is $\eta=k$; therefore, in the rest of the paper we take \(\eta=k\), and let \(A'_{k}:=A'_{k,k}\).
For the Neumann problem, the standard ``combined-field'' or ``combined-potential'' BIE is
\begin{align}\label{eq:direct_combined_neu}
B_{k,\eta} u = {\rm i} \eta u^I - \partial_n u^I \quad\text{ on } \Gamma,
\quad\text{ where }\quad
B_{k,\eta} := {\rm i}\eta \left(\dfrac{1}{2}I-D_k\right) + H_k,
\end{align}
where $D_k$ and $H_k$ are the double-layer and hypersingular operators defined by \eqref{eq:DH} below. In contrast to $A'_k$, $B_{k,\eta}$ is not a bounded operator on ${L^2(\Gamma}%{\partial {\cO}_+})}$ (even when $\Gamma$ is smooth) because of the hypersingular operator $H_k$. If $k>0$, \(\Re(\eta)\neq 0\), and $\Omega_-$ is Lipschitz, then $B_{k,\eta}:H^{s+1/2}(\Gamma)\rightarrow H^{s-1/2}(\Gamma)$ is bounded and invertible for $|s|\leq 1/2$ (see, e.g., \cite[Theorem 2.27]{ChGrLaSp:12}). The standard choice of $\eta$ here is also $\eta=k$, and we let \(B_k:=B_{k,k}\).
The fact that $B_k$ is not bounded from ${L^2(\Gamma}%{\partial {\cO}_+})}\rightarrow {L^2(\Gamma}%{\partial {\cO}_+})}$ means that the condition numbers of $h$-version Galerkin discretisation of \eqref{eq:direct_combined_neu} blow up as $h\rightarrow 0$ for fixed $k$.
There has therefore been much research interest in designing alternative Neumann BIE formulations; see, e.g., \cite{StWe:98, AnDa:05, AnDa:07, BrElTu:12, DaDaLa:13}. We use the following BIE, introduced in \cite{BrElTu:12} (which focused specifically on high-frequency problems)
based on the idea of Calder\'on preconditioning,
\begin{align}\label{eq:direct_combined_neu2}
B_{k,\eta,{\rm reg}} \gamma_+ u = {\rm i} \eta \gamma^+ u^I - S_{{\rm i} k} \partial_{n}^+ u^I \quad\text{ on } \Gamma,
\quad\text{ where } \quad
B_{k,\eta,{\rm reg}} := {\rm i}\eta \left(\dfrac{1}{2}I-D_k\right) + S_{{\rm i} k} H_k.
\end{align}
At least when $\Gamma$ is $C^1$, if $k>0$ and \(\Re(\eta)\neq 0\), then $B_{k,\eta,{\rm reg}}:{L^2(\Gamma}%{\partial {\cO}_+})}\rightarrow {L^2(\Gamma}%{\partial {\cO}_+})}$ is bounded and invertible \cite[Theorem 2.1]{BrElTu:12}. In what follows, we make the same choice for \(\eta\) as in~\cite[Equation 24]{BrElTu:12}, i.e. \(\eta=1/2\), and we let \(B_{k,{\rm reg}}:=B_{k,1/2,{\rm reg}}\).
We highlight that the idea of combatting the ``bad'' behaviour of the hypersingular operator by composing it with a regularising operator (in this case
$S_{{\rm i} k}$) is often called ``operator preconditioning'' (see \cite{Hi:06}).
Because the normality (or not) of an operator is relevant for the analysis of GMRES, we highlight that when $\Omega_-$ is not a ball, $A_{k}'$ and $B_{k, {\rm reg}}$ are non-normal operators on ${L^2(\Gamma}%{\partial {\cO}_+})}$; this is shown by the plots of the numerical range for $A_k'$ in \cite{BeSp:11} and for $B_{k, {\rm reg}}$ in \cite[\S5]{BoTu:13} (see also \cite{BePhSp:13}).
\subsubsection{The boundary-element method (BEM).}\label{sec:statement:BEM}
We solve the BIEs
\eqref{eq:direct_combined_dir} and \eqref{eq:direct_combined_neu2} with the Galerkin method in ${L^2(\Gamma}%{\partial {\cO}_+})}$.
That is, to solve the BIE \eqref{eq:direct_combined_dir} given a finite-dimensional subspace $V_n\subset {L^2(\Gamma}%{\partial {\cO}_+})}$, we
\begin{equation}\label{eq:Galerkin}
\text{ find } v_n \in V_n \text{ such that } \big(A_k' v_n, w_n\big)_{L^2(\Gamma}%{\partial {\cO}_+})} = \big(f, w_n\big)_{L^2(\Gamma}%{\partial {\cO}_+})} \quad\text{ for all } w_n \in V_n,
\end{equation}
where $f$ denotes the right-hand side of the BIE in \eqref{eq:direct_combined_dir}; the Galerkin solution $v_n$ is then an approximation to $\partial_n u$.
We solve the BIE \eqref{eq:direct_combined_neu} via the Galerkin method in $H^{1/2}(\Gamma)$.
That is, given a finite-dimensional subspace $V_n\subset H^{1/2}(\Gamma)$, we
\begin{equation*}
\text{ find } v_n \in V_n \text{ such that } \big\langle B_k v_n, w_n\rangle = \big\langle f, w_n\big\rangle \quad\text{ for all } w_n \in V_n,
\end{equation*}
where $f$ now denotes the right-hand side of the BIE in \eqref{eq:direct_combined_neu}, and $\langle\cdot,\cdot\rangle$ denotes the duality pairing between $H^{-1/2}(\Gamma)$ and $H^{1/2}(\Gamma)$.
Given a basis $\{\phi_j\}_{j=1}^n$ of $V_n$, the Galerkin equations \eqref{eq:Galerkin} are equivalent to the linear system
\begin{equation}\label{eq:1}
\bfA_k' \mathbf{v} = \mathbf{f}
\end{equation}
where
\begin{align}\label{eq:Galerkinmatrix}
(\bfA_k')_{i,j}:= \int_{\Gamma} \big(A_k'\phi_j\big) (x}%\mathbf{x})\phi_i (x}%\mathbf{x}) \dif \sigma(x}%\mathbf{x})
\quad \text{ and }\quad
(\mathbf{f})_i:= \int_{\Gamma} f(x}%\mathbf{x}) \,\phi_i (x}%\mathbf{x}) \dif \sigma(x}%\mathbf{x}).
\end{align}
Regarding notation: we put in bold font the $n\times n$ matrices and $n\times 1$ vectors arising from the Galerkin method -- such as $\bfA_k'$, $\mathbf{v}$, and $\mathbf{f}$ in \eqref{eq:1} -- but do not put in bold font the position vectors in $\mathbb{R}^d$ -- such as $x}%\mathbf{x}$ in \eqref{eq:Galerkinmatrix}.
We consider the $h$-version of the boundary-element method, and choose $V_n$ to be P1 Lagrange elements (i.e.~continuous piecewise-linear polynomials on the reference elements).
To maintain accuracy as $k\rightarrow \infty$, $h$ must be tied to $k$. In applications, one usually chooses $h$ to be proportional to $1/k$, i.e. a fixed number of points per wavelength (see, e.g, \cite{Ma:02}), and we do the same for the numerical experiments in this paper.
At least when $\Omega_-$ is nontrapping, empirically one sees uniform accuracy as $k\rightarrow \infty$ with this choice, although this has not yet been proved.
The current best results proving accuracy of the Galerkin solutions for large $k$ for the Dirichlet problem are in \cite{GaMuSp:19} (following \cite{GrLoMeSp:15}), with these results proving quasioptimality of the Galerkin solution (with quasioptimality constant independent of $k$)
(i) for smooth and strictly convex $\Omega_-$ when $hk^{4/3}$ is sufficiently small, and (ii)
for general nontrapping $\Omega_-$ when
$hk^{3/2}\log k$ is sufficiently small.
There is almost no analogous theory for the Neumann problem for large $k$; the exception is \cite{BoTu:13} whose results about coercivity
of the BIE \eqref{eq:direct_combined_neu2} when $\Omega_-$ is a ball imply a quasioptimality result without any restriction on $h$, albeit with quasioptimality constant growing like $k^{1/3}$.
\subsubsection{Iterative solution of the BEM linear systems via GMRES.}
\label{sec:GMRES}
A popular way of solving the dense linear systems that arise from the BEM is via iterative methods \cite[Chapter 13]{St:08}, \cite[Chapter 6]{SaSc:11}, \cite[\S4]{RjSt:07}. Since the systems arising from the Helmholtz equation are, in general, non-normal (as highlighted in \S\ref{sec:BIE}), a natural choice of iterative method is the generalised minimum residual method (GMRES) \cite{SaSc:86}.
Given \(\bfB \in \bbC^{n\times n}\), \(\bfb\in \bbC^n\), the generalised minimum residual method (GMRES) to find the solution $\bfx$ of $\bfB \bfx = \bfb$ is the following.
Given \(\bfx_0 \in \bbC^n\), let \(\bfr_0(\bfB,\bfb,\bfx_0):=\bfb-\bfB \bfx_0\).
Let the Krylov space \(\calK_m(\bfB,\bfr_0)\) be defined by
\begin{align*}
\calK_m(\bfB,\bfr_0):=\Span \left\{\bfr_0, \bfB\bfr_0, \ldots, \bfB^{m-1}\bfr_0\right\}.
\end{align*}
The $m$th iterate of GMRES, \(\bfx_m\), is defined as the unique vector
in $\bfx_0 + \calK_m(\bfB,\bfb)$ that minimises the residual
$\bfr_m:=\bfb-\bfB \bfx_m$ with respect to the \(\lVert \cdot \rVert_2\) norm (see, e.g., \cite[\S6.5.1]{Sa:03}).
Observe that, since \(\bfx_m \in \bfx_0 + \calK_m(\bfB,\bfb)\), the residual satisfies
\begin{equation}\label{eq:residual}
\bfr_m:=\bfb-\bfB \bfx_m= p_m(\bfB)\bfr_0
\quad\text{ for } p_m \in \bbP_m \text{ with } p_m(0)=1,
\end{equation}
where $\mathbb{P}_m$ denotes the set of polynomials of degree $m$. The definition of GMRES therefore implies that
\begin{align}\label{def:gmres}
\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2= \min_{\substack{p_m\in \bbP_m, \\ p_m(0)=1}}\lVert p_m (\bfB)\bfr_0(\bfB,\bfb,\bfx_0)\rVert_2.
\end{align}
We apply GMRES to the linear system \eqref{eq:1} preconditioned by the mass matrix
\begin{align}\label{eq:massmatrix}
(\bfM)_{i,j}:= \int_{\Gamma} \phi_i(x}%\mathbf{x}) \phi_j(x}%\mathbf{x}) \dif \sigma(x}%\mathbf{x});
\end{align}
i.e.~we solve
\begin{equation}\label{eq:2}
\bfM^{-1}\bfA_k' \mathbf{v} = \bfM^{-1}\mathbf{f}.
\end{equation}
We solve \eqref{eq:2} instead of \eqref{eq:1} because it is easier to translate information about $A_k'$ to information about $\bfM^{-1}\bfA_k' $ rather than information about $\bfA_k'$. This is for the following two reasons.
(a) The eigenvalues of $\bfM^{-1}\bfA_k' $ approximate the eigenvalues of $A_k'$. Indeed, the eigenvalue problem $A_k' v=\lambda_k v$ is equivalent to the variational problem: find $v\in {L^2(\Gamma}%{\partial {\cO}_+})}$ such that $(A_k' v,w)_{L^2(\Gamma}%{\partial {\cO}_+})}= \lambda_k (v,w)_{{L^2(\Gamma}%{\partial {\cO}_+})}}$ for all $w\in {L^2(\Gamma}%{\partial {\cO}_+})}$, and the Galerkin approximation of this is $\bfA_k' \mathbf{v} = \lambda_k \bfM \mathbf{v}$.
(b) If \(\Gamma\) is \(C^1\), then, given $k$, there exists \(h_0\) and \({C_{{\rm approx},1}}\) such that, if \(h\leq h_0\) then
\begin{equation}\label{eq:discon2}
({C_{{\rm approx},1}}^{-1})\mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}\leq
\mathbb{N}{\bfM^{-1}\bfA_k'}_2 \leq {C_{{\rm approx},1}} \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}},
\end{equation}
and analogous bounds hold for $(A_k')^{-1}$ and $(\bfM^{-1}\bfA_k')^{-1}$
(furthermore, if the basis is orthonormal, then $\|\bfM^{-1}\bfA_k'\|_2\rightarrow \|A_k'\|_{{L^2(\Gamma}%{\partial {\cO}_+})}}$
and $\|(\bfM^{-1}\bfA_k')^{-1}\|_2\rightarrow \|(A_k')^{-1}\|_{{L^2(\Gamma}%{\partial {\cO}_+})}}$ as $h\to 0$ for fixed $k$); see Lemma \ref{lem:discon} and Remark \ref{rem:compact} below.
In contrast, in the analogue of \eqref{eq:discon2} with $\bfM^{-1}\bfA_k'$ replaced by $\bfA_k'$, the constant $C$ depends on $h$; see \eqref{eq:Galerkinmatrixbounds2}.
We only consider solving the system \eqref{eq:2} with standard GMRES because our goal is to prove rigorous bounds on the number of iterations and the theory of GMRES convergence is most well-developed for standard GMRES.
We note that GMRES is often used with either restarts or restarts with subspace augmentation (see, e.g., \cite{Mo:95, Mo:02, GiGrPiVa:10}) -- this has the advantage of reducing storage and orthogonalisation costs, but with the number of iterations required to obtain a given relative residual necessarily higher than for standard GMRES (although it is difficult to study this increase theoretically).
\subsection{Four features (F1-F4) observed in numerical experiments on the set-up in \S\ref{sec:statement}, and statement of the main goals of this paper}\label{sec:features}
We now highlight four different features one observes from computing approximations to the scattering problem via the set-up in \S\ref{sec:statement} (i.e.~reformulating as a BIE, creating a linear system via the BEM, and solving the linear system using GMRES). We present numerical experiments illustrating each of the features later in the paper.
These features are about, respectively, (1) the accuracy of the Galerkin solution, (2) the condition number of the Galerkin matrix, (3) the number of GMRES iterations, and (4) the accuracy of the GMRES solution.
\begin{itemize}
\item[F1] When the incoming plane wave enters the cavity, one needs a larger number of points per wavelength for accuracy of the Galerkin solutions than when the wave doesn't enter the cavity.
\item[F2] The norm of $(\bfM^{-1}\bfA_k')^{-1}$ (i) is very sensitive to whether or not $k=k_j$ for $k_j$ a quasimode frequency, and (ii)
grows exponentially through $k_j$,
up to some point, and then grows more slowly.
\item[F3] The number of GMRES iterations required to make the residual $\bfr_m$ arbitrarily small
\begin{itemize}
\item[(a)] grows algebraically with $k$, with no worse growth through $k=k_j$ than $k\neq k_j$,
\item[(b)] depends on whether $\Omega_-$ is the small or large cavity, and
\item[(c)] depends on the direction of the incoming plane wave.
\end{itemize}
\item[F4] The GMRES residual being small does not necessarily mean that the error is small, and the relative sizes of the residual and error depend on both $k$ and the direction of the plane wave.
\end{itemize}
{\bf The main goals of this paper are to explain F3(a) and, to a certain extent,
F3(b).}
The following is an outline of the rest of the introduction. In \S\ref{sec:intro_num} we present numerical results about F3(a), F3(b), and F3(c) for the Dirichlet problem.
In \S\ref{sec:observations} we give plots of the eigenvalues of $\bfM^{-1}\bfA_k'$.
In \S\ref{sec:theorem} we give a general bound on the number of GMRES iterations when the matrix has a ``cluster plus outliers'' structure. In \S\ref{sec:mainresultH} we apply the bound from \S\ref{sec:theorem} to $\bfM^{-1}\bfA_k'$, under assumptions based on the eigenvalue plots in \S\ref{sec:observations}; the result is a $k$-explicit bound on the number of GMRES iterations which explains F3(a).
In the last section of the introduction, \S\ref{sec:FEM}, we discuss how the ideas in this paper can be applied to the Helmholtz FEM.
The partial explanation of F3(b) is contained in \S\ref{sec:smalllarge}.
Although our focus is on F3, we still need to be aware of other features; e.g.~the solution obtained by GMRES is useless if the Galerkin solution itself isn't accurate (F1), or if the GMRES solution isn't close to the Galerkin solution (F4).
We therefore make some brief comments here about to what extent the features F1, F2, F3(c), and F4 are rigorously understood; the summary is that F2 is rigorously understood, whereas F1, F3(c), and F4 are not.
\paragraph{Regarding F1:}
The fact that the accuracy of the Galerkin solution depends on whether or not the wave enters the cavity makes physical sense, but there is currently no rigorous theory on the subject. Indeed, as discussed in \S\ref{sec:statement:BEM}, the current best analysis of how $h$ must depend on $k$ for the $h$-BEM to be uniformly accurate as $k\rightarrow \infty$, \cite{GaMuSp:19}, is not sharp in the nontrapping case, and
is therefore very far from proving rigorous sharp results about the trapping case.
Numerical experiments illustrating F1 are given in Appendix \ref{app:F1F4}.
\paragraph{Regarding F2:}
The exponential growth of $(\bfM^{-1}\bfA_k')^{-1}$ through the sequence of $k_j$ is explained by the following.
The inverses of the boundary integral operators $A'_k$, $B_k$, and $B_{k, {\rm reg}}$ inherit the behaviour of the Helmholtz solution operator, and thus grow when $k=k_j$ for $k_j$ quasimode frequencies. More precisely, if $k_j$ and $\epsilon(k)$ are as in Definition \ref{def:quasimodes}, then there exists $C>0$ (independent of $j$) such that
\begin{equation}\label{eq:Ainvblowup}
\mathbb{N}{(A'_{k_j})^{-1}}_{{\LtG\rightarrow \LtG}} \geq C \left(\frac{1}{\epsilon(k_j)} - \frac{1}{k_j}\right)k_j^{1/4} \quad\text{ for all } j
\end{equation}
\cite[Equation 5.39]{ChGrLaSp:12} \footnote{More precisely, \cite[\S5.6.2, Equation 5.39]{ChGrLaSp:12} proves \eqref{eq:Ainvblowup} with a different power of $k$ on the right-hand side.
The bound \eqref{eq:Ainvblowup} can be proved by following the same steps as in
\cite[\S5.6.2]{ChGrLaSp:12}, but using the sharp bound on the single-layer potential from~\cite[Theorem 1.1, Part (i)]{HaTa:15}.}. Therefore, when $\Omega_-$ is either the small or the large cavity, by \eqref{eq:quality},
$\|(A'_{k_j})^{-1}\|_{{\LtG\rightarrow \LtG}}\geq C_1 \exp (C_2 k_j)$ for some $C_1, C_2>0$ independent of $j$
\cite[Theorem 2.8]{BeChGrLaLi:11}.
We note that, since $\|A'_{k}\|_{{\LtG\rightarrow \LtG}}$ grows algebraically in $k$ for general Lipschitz domains (see
Part (i) of Lemma \ref{lem:normbounds} below), the condition number of $A'_{k_j}$ also grows exponentially as $j\rightarrow \infty$ when $\Omega_-$ is either the small or the large cavity.
An indication (but not a rigorous proof) of why the growth of $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ through $k=k_j$ stagnates,
and why $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ is very sensitive to whether or not $k=k_j$,
is given
by the recent result of \cite[Theorem 1.1]{LaSpWu:20}. This result shows that, for most frequencies, the Helmholtz solution operator (and hence also $\|(A'_{k_j})^{-1}\|_{{\LtG\rightarrow \LtG}}$) is bounded polynomially in $k$. More precisely,
given
$\nu>0$,
there exists $C_3=C_3(\nu)>0$ and a set $J\subset [1,\infty)$ with $|J|\leq \nu$ such,
\begin{equation}\label{eq:Ainv}
\mathbb{N}{(A'_{k})^{-1}}_{{\LtG\rightarrow \LtG}} \leq C_3 k^{5d/2 +1} \quad\text{ for all } k \in [1,\infty)\setminus J
\end{equation}
(for simplicity we have assumed $k\geq 1$, but an analogous bound holds for any $k\geq k_0>0$).
The bounds \eqref{eq:Ainvblowup} and \eqref{eq:Ainv} then imply that the graph of $\|(A'_{k_j})^{-1}\|_{{\LtG\rightarrow \LtG}}$ against $k$ consists of a number of ``spikes'' at $k=k_j$, with the heights of the spikes growing exponentially with $k$, but the widths decreasing with $k$
\footnote{For an illustration of this in a simple 1-d model of resonance behaviour, see \cite[\S23.2]{FeLeSa:64}.}.
Therefore, while
$\|(A'_{k_j})^{-1}\|_{{\LtG\rightarrow \LtG}}$ grows exponentially through $k_j$, the growth is very sensitive to the precise value of $k$ (with this sensitivity increasing as $k$ increases).
This result indicates that the growth of $(\bfM^{-1}\bfA_k')^{-1}$ through $k=k_j$ stagnates since discretisation error collapses the delicate exponential growth.
\paragraph{Regarding F3(c):} This feature arises because the GMRES residual $\bfr_m$ \eqref{eq:residual} depends on the right-hand side vector, which depends on the direction of the plane wave (via the right-hand side of the BIE in \eqref{eq:direct_combined_dir}). There are few rigorous results in the literature describing
the dependence of $\bfr_m$ on the right-hand side vector, but
in Appendix \ref{app:GMRES_rhs} we describe how the results of \cite{Titley2014} give a heuristic explanation
of this feature
for problems with similar eigenvalue distributions to the Helmholtz problems we consider.
\paragraph{Regarding F4:}
Numerical experiments illustrating this feature on our problem are given in Appendix \ref{app:F1F4}.
This feature is poorly understood for non-normal, complex linear systems in general, and thus also for the systems arising from the Helmholtz problems considered here.
There have been many papers that discuss the convergence of GMRES in the sense of residual reduction; in contrast, there is remarkably little known in the literature about the error $\bfx-\bfx_m$. The most recent (and most relevant) results in this area are given in \cite{Meurant2011} and \cite[\S 5.8]{Meurant2020} and even then the results are stated for real systems. In particular, \cite[Theorem 5.35 and Corollary 5.6]{Meurant2020} gives bounds on $\|\bfx-\bfx_m\|_2$ in terms of $\|\mathbf{r}_m\|_2$ multiplied by a computable expression that requires the existence of $\bfH_m^{-1}$ (where $\bfH_m$ is the square $m \times m$ upper Hessenberg matrix arising at the $m$-th step in the Arnoldi process in the GMRES algorithm) and also depends on other entries of $\bfH_m$.
\subsection{Numerical experiments about F3}\label{sec:intro_num}
For ease of exposition, we only present here experiments for the Dirichlet problem \eqref{eq:direct_combined_dir}, i.e., involving the operator $A_k'$, in 2-d.
\S\ref{sec:experiments} contains experiments for the two BIEs for the Neumann problem, \eqref{eq:direct_combined_neu} and \eqref{eq:direct_combined_neu2}, and experiments for the Dirichlet problem in 3-d.
All the experiments in this section use ten points per wavelength.
Furthermore, we plot quantities of interest (such as the number of iterations, the condition number) through \emph{either} integer values of $k$ \emph{or} values of $k$ in a quasimode. As described in \S\ref{sec:statement:ellipse},
for $\Omega_-$ the small or large cavities,
there exists a quasimode with frequencies equal
$\{k^{e/o}_{m,n}\}_{m=1}^\infty$ for fixed $n$.
Our experiments consider $\{k_{m,0}^e\}_{m=1}^\infty$, but we observe very similar behaviour through
$\{k_{m,n}^{e/o}\}_{m=1}^\infty$ for $n\neq 0$ fixed.
\paragraph{Experiments illustrating F3(a) (growth of iterations with $k$).}
Figure~\ref{fig:comparison_spectrum_it}
plots the condition number of $\bfM^{-1}\bfA_k'$ and the number of GMRES iterations against $k$ for the small cavity.
The direction $\hat{a}}%\mathbf{a}$ of the incident plane wave $\exp({\rm i} k x}%\mathbf{x}\cdot\hat{a}}%\mathbf{a})$ is chosen as $a=(\cos\theta, \sin\theta)$ with
$\theta =4\pi/10$; from Figure \ref{fig:geometries} we see that the plane wave is almost vertical and enters the cavity.
The key point from Figure~\ref{fig:comparison_spectrum_it} is that, while the condition number $\bfM^{-1}\bfA_k'$ is very sensitive to whether or not $k$
is near a frequency in the quasimode, the number of iterations is not. This demonstrates the well-known fact that the condition number gives little insight into the behaviour of GMRES for non-normal problems.
In more detail, the left-hand plot in Figure~\ref{fig:comparison_spectrum_it} shows, via the condition number, the sensitivity of $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ to whether or not $k=k_j$ (i.e., the first point in F2).
The green circle outlier at $k=120$ is there because, by chance, the integer frequency 120 lies very close to the quasimode frequency
$k^o_{17,7}=119.997615771724$
(note that to 5 significant figures this approximation of the quasimode frequency is equal to $120$).
This plot also shows the growth of $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ through $k=k_j$ stagnating as $k$ increases (i.e., the second point in F2); this was also seen in
the experiments in \cite[Section IV.H]{BeChGrLaLi:11} on the small cavity, where, even using 20 points per wavelength,
the exponential growth of $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ through $k=k_j$ levelled off after $k=60$.
This sensitivity of $\|(\bfM^{-1}\bfA_k')^{-1}\|_2$ to whether or not $k=k_j$ was also shown in \cite[Figure 4.7]{LoMe:11}
for a cavity similar to both our large and small cavities.
The plot of the number of iterations in Figure~\ref{fig:comparison_spectrum_it} is also included in Figure \ref{fig:iterations_main} below, where we see that the number of iterations grows like $k^{0.66}$ (in the range considered).
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{comparison_spectrum_it.pdf}
\caption{Condition number and number of GMRES iterations for small cavity with incident plane wave at angle $\theta =4\pi/10$ to the horizontal (illustrating F3(a)).
}\label{fig:comparison_spectrum_it}
\end{figure}
\begin{figure}[h!!]
\centering
\includegraphics[width=0.9\textwidth]{comparison_size_cavity.pdf}
\caption{Comparison of the condition number and the number of GMRES iterations for the small and large cavities
with incident plane wave at angle $\theta =4\pi/10$ to the horizontal (illustrating F3(b)).}\label{fig:comparison_size_cavit}
\end{figure}
\begin{figure}[h!!]
\centering
\includegraphics[width=0.9\textwidth]{comparison_rhs.pdf}
\caption{The number of GMRES iterations for the small cavity and three different right-hand sides, corresponding to three different angles $\theta$ of the incident plane wave (illustrating F3(c)).}\label{fig:comparison_rhs}
\end{figure}
\paragraph{Experiments illustrating F3(b) (dependence of iterations on cavity size).}
Figure~\ref{fig:comparison_size_cavit}
plots the condition number of $\bfM^{-1}\bfA_k'$ and the number of GMRES iterations against $k_{m,0}^e$ for both the small and large cavities.
As in the previous figure, $\hat{a}}%\mathbf{a}= (\cos\theta, \sin\theta)$ with
$\theta =4\pi/10$; i.e. the plane wave is almost vertical and enters the cavity.
While the condition numbers behave very similarly,
the growth in the number of iterations is different, again illustrating the fact that the condition number is not relevant for understanding the convergence of GMRES for non-normal matrices. For the small cavity the number of iterations grows approximately like $k^{0.66}$, and for the large cavity like $k^{0.82}$; see Figure \ref{fig:iterations_main} below.
\paragraph{Experiments illustrating F3(c) (dependence of iterations on plane-wave direction).}
Figure~\ref{fig:comparison_rhs}
plots the number of GMRES iterations against $k$ for the small cavity and varying $\theta$, with the incident plane wave $\hat{a}}%\mathbf{a}= (\cos\theta, \sin\theta)$. From Figure \ref{fig:geometries}, we see that when $\theta =4\pi/10$ the plane wave is almost vertical and enters the cavity, when
$\theta =0$ the plane wave is horizontal and enters the cavity, and when $\theta=\pi$ the plane wave is horizontal and doesn't enter the cavity.
Physically, we expect the worst behaviour to occur when $\theta=4\pi/10$, because of the multiple reflections in the cavity, and the best behaviour to occur when $\theta=\pi$, and this is indeed what we see in Figure \ref{fig:comparison_rhs}.
\paragraph{Links with other experiments/results in the literature.}
Both the iterative solution of BEM linear systems
and solving scattering problem involving cavities
have received a lot of interest in the literature; see, e.g., the books \cite[Chapter 13]{St:08}, \cite[Chapter 6]{SaSc:11}, \cite[\S4]{RjSt:07} for the former, and, e.g., \cite{BaSu:05, WaDuSu:09, GmPh:07, GiGrPiVa:10, Du:11, DaDaLa:13, LaAmGr:14, LaGrON:17} for the latter.
Nevertheless, the features F1-F4 do not appear to have been systemically identified and studied before now.
We highlight here one previous study where the features F2 and F3(a) are visible in numerical experiments. Indeed,
\cite{DaDaLa:13} considers solving the Neumann problem with a BIE similar to \eqref{eq:direct_combined_neu2}, but
with $S_{{\rm i} k}$ replaced by a different regularising operator.
The figures \cite[Figures 11(a), 18, and 19]{DaDaLa:13} plot the condition number against $k$ when $\Omega_-$
are cavity domains similar to those in Figure \ref{fig:geometries} (although supporting weaker trapping),
and display spikes; i.e., F2.
The figure \cite[Figure 28(a)]{DaDaLa:13} plots the number of GMRES iterations against $k$ and sees growth with no spikes; i.e., F3(a).
\subsection{Plots of the eigenvalues of $\bfM^{-1}\bfA_k'$.}\label{sec:observations}
\paragraph{Summary of the figures.} Figure \ref{fig:example_spectrum} plots the eigenvalues of $\bfM^{-1}\bfA_k'$ for the small and large cavities at $k=100$ and $k=290$.
Figure ~\ref{fig:outliers} plots the eigenvalues for the small and large cavities at both $k=100$ and $k=290$, as well as the eigenvalues at $k=290$ for two other $\Omega_-$ for which Theorem \ref{thm:ellipse} applies; these two other $\Omega_-$ are plotted in Figure \ref{fig:circles_minus_ellipse}.
Figures \ref{fig:spectra_illustrations_elliptic_cavity_2D_dir} and \ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir}
plot the eigenvalues and singular values of $\bfM^{-1}\bfA_k'$ for several frequencies $k^{e/o}_{m,n}$ and $\Omega_-$ the small and large cavities, respectively.
Figure~\ref{fig:flow_eig} plots the paths of the near-zero eigenvalues as functions of $k$ for $k\in (5,15)$; the spectra are computed
every \(0.025\), and the arrows placed at these points.
\begin{figure}[h!]
\centering
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{example_spectrum_small_0.pdf}
\caption{Small cavity at \(k=100\)}
\end{subfigure}\hfill%
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{example_spectrum_large_0.pdf}
\caption{Large cavity at \(k=100\)}
\end{subfigure}%
\par\bigskip
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{example_spectrum_small_1.pdf}
\caption{Small cavity at \(k=290\)}
\end{subfigure}\hfill%
\hfill
\begin{subfigure}[t]{0.45\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{example_spectrum_large_1.pdf}
\caption{Large cavity at \(k=290\)}
\end{subfigure}%
\caption{The eigenvalues of $\bfM^{-1}\bfA_k'$ for the small and large cavities at $k=100$ and $k=290$.}\label{fig:example_spectrum}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{spectrum_outliers.pdf}
\caption{Plots of the near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ for a variety of different domains and frequencies. The black rectangle $[-0.1,0.1]\times[-0.6,0.6]$ is a choice of the set ${\cal N}$ in Observation O1.}
\label{fig:outliers}
\end{figure}
\begin{figure}[h!]
\centering
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{two_circles_m_ellipse.pdf}
\end{minipage}%
\hfill
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=0.7\textwidth]{two_circles_m_ellipse_close.pdf}
\end{minipage}
\caption{The ``two circles minus ellipse'' and ``two close circles minus ellipse'' $\Omega_-$ considered in Figure \ref{fig:outliers}
with $a_1=1$ and $a_2=1/2$.
}
\label{fig:circles_minus_ellipse}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{spectra_illustrations_elliptic_cavity_2D_dir.pdf}
\caption{The eigenvalues and singular values of $\bfM^{-1}\bfA_k'$ for several frequencies $k^{e/o}_{m,n}$ and $\Omega_-$ the small cavity}\label{fig:spectra_illustrations_elliptic_cavity_2D_dir}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{spectra_illustrations_elliptic_cavity_bigger_2D_dir.pdf}
\caption{The eigenvalues and singular values of $\bfM^{-1}\bfA_k'$ for several frequencies $k^{e/o}_{m,n}$ and $\Omega_-$ the large cavity}
\label{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir}
\end{figure}
\begin{figure}[h!]
\centering
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{flow_illustrations_elliptic_cavity_2D_dir.pdf}
\end{minipage}%
\hfill
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{flow_illustrations_elliptic_cavity_bigger_2D_dir.pdf}
\end{minipage}
\caption{Paths of the eigenvalues for \(k\in (5,15)\) for the small cavity (left) and the large cavity (right)
The eigenvalues that enter the rectangle
are coloured green and have shaded arrowheads.}
\label{fig:flow_eig}
\end{figure}
\paragraph{Observations from these figures.} We make five observations from these figures.
We number them O1, O2(a)-(d), corresponding, respectively, to Assumptions A1 and A2 below, under which we prove a bound on the $k$-dependence of the number of GMRES iterations (Theorem \ref{thm:main2}).
\begin{enumerate}
\item[O1] There exists a bounded open set ${\cal N}$ (${\cal N}$ for ``near-zero'') containing zero and a closed half-plane ${\cal H}$ not containing zero such that
(i) all the eigenvalues of $\bfM^{-1}\bfA_k'$ are contained in ${\cal N}\cup {\cal H}$, and
(ii) ${\cal N}$ and ${\cal H}$ can both be chosen to be independent of $k$.
\end{enumerate}
Point (i) is clear from Figure \ref{fig:example_spectrum} that
plots the eigenvalues of $\bfM^{-1}\bfA_k'$ for the small and large cavities at $k=100$ and $k=290$.
Point (ii)
is shown in Figure \ref{fig:outliers};
indeed, all the near-zero eigenvalues for these two different values of $k$ lie on the same curve, and thus ${\cal N}$ can be taken as the black rectangle in Figure \ref{fig:outliers}. In addition, the curve is the same for the four different $\Omega_-$ considered; this is perhaps expected since the four different $\Omega_+$ all contain a neighbourhood of the minor axis of the same ellipse (namely \eqref{eq:ellipse} with $a_1=1$ and $a_2=1/2$) and the near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ are generated by the trapped ray in this neighbourhood.\footnote{However, with $a_1=1$ and $a_2=1/4$ (i.e., a different ellipse)
we see the eigenvalues lying on, by eye, the same curve as Figure \ref{fig:outliers} for analogous small and large cavities at $k=100$ and $k=290$.}
\begin{enumerate}
\item[O2(a)] (Family of quasimodes.) There exists a sequence $0<k_1<k_2<\ldots$, with $k_j\rightarrow \infty$ as $j\rightarrow \infty$, such that $\bfM^{-1}\bfA_{k_j}'$ has a near-zero singular value for $j$ sufficiently large.
\end{enumerate}
Recall that O2(a) is guaranteed on the continuous level by the lower bound \eqref{eq:Ainvblowup}, and we see small singular values (orange crosses) in three of the four plots in Figure \ref{fig:spectra_illustrations_elliptic_cavity_2D_dir} and all four plots in Figure \ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir}.
\begin{enumerate}
\item[O2(b)] (Quasimode $\implies$ near-zero eigenvalue.) If $j$ is sufficiently large, $\bfM^{-1}\bfA_{k_j}'$ has a near-zero eigenvalue.
\end{enumerate}
This can be seen from the fact that near-zero singular values are accompanied by near-zero eigenvalues
in both Figures \ref{fig:spectra_illustrations_elliptic_cavity_2D_dir} and \ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir}.
\begin{enumerate}
\item[O2(c)] (Near-zero eigenvalues.)
All the eigenvalues of $\bfM^{-1}\bfA_k'$ in the set ${\cal N}$ in O1 move at a speed that can be bounded above and below by constants independent of $k$.
\end{enumerate}
In fact, Figure \ref{fig:flow_eig} indicates that the speed of the eigenvalues is independent of $k$ because
the arrows in the box in Figure \ref{fig:flow_eig} are all evenly spaced; furthermore,
we observe numerically that the speed is approximately one (at least for that range of $k$).
\begin{enumerate}
\item[O2(d)] The large cavity has more near-zero eigenvalues than the small cavity.
\end{enumerate}
This can be seen from Figure \ref{fig:example_spectrum}. In addition, comparing the top-left plots of Figures~\ref{fig:spectra_illustrations_elliptic_cavity_2D_dir} and~\ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir} we see that
when $k=k^o_{0,3}$ there is no near-zero eigenvalue for the small cavity, but there is for the large cavity.
This observation is the reason for the feature F3(b) (the number of iterations is larger for the large cavity than the small cavity).
Observation O2(d) can be partially explained from the fact that a larger number of the Laplace eigenfunctions of the ellipse $E$ \eqref{eq:ellipse} (from which the quasimodes in Theorem \ref{thm:ellipse} are constructed) are localised in the large cavity than in the small cavity. In the FEM case there is a close connection between the functions in the quasimodes and the eigenfunctions of the Galerkin matrix (see \cite[Remark 1.7]{GaMaSp:21}), and thus these localisation considerations immediately explain why the large cavity has more near-zero eigenvalues than the small cavity. However, in the BEM case it is less clear how the eigenvalues of $\bfM^{-1}\bfA_k'$ (which are discretisations of functions living on $\Gamma$) are connected to the functions in the quasimodes (which live in $\Omega_+$).
We return to Observation O2(d) in \S\ref{sec:Weyl_new}
where we use heuristics from Weyl asymptotics to estimate how many more Laplace eigenfunctions of the ellipse are
localised in $\Omega_+$
for the large cavity than for the small cavity. We then compare these heuristics to the number of eigenvalues of $\bfM^{-1}\bfA_k'$ observed computationally (see \S\ref{sec:numA2}).
\paragraph{Link with other experiments/results in the literature.}
Similar eigenvalue plots for BIEs
when $\Omega_-$ is nontrapping or weakly trapping can be found for $A_k'$ in \cite[Figure 9]{Ba:06} and \cite[Figure 3.1 and \S5]{BeSp:11}, and for the indirect analogue of $B_{k, {\rm reg}}$ in \cite[Figure 1]{BrElTu:12}, \cite[Figures 3-5]{BoTu:13}, and \cite[Figure 2]{ViGrGi:14}.
Furthermore, the analogue of O2(c) (the eigenvalues move at $O(1)$ speed)
was used to compute large eigenvalues of the Laplacian in \cite{TuSc:07} for BIEs related to $A_k'$
and \cite{BaHa:14} for a related boundary-based method; see \cite[Figure 5]{TuSc:07} and \cite[Theorem 4.1]{BaHa:14}, respectively.
\subsection{First main result: general bound on number of GMRES iterations for matrix with ``cluster plus outlier'' structure.}
\label{sec:theorem}
For simplicity we consider matrices with simple eigenvalues; the modifications to our assumptions and arguments for matrices with repeated eigenvalues are outlined in Remark \ref{rem:BF1}.
For $\lambda$ a simple eigenvalue, let $\kappa (\lambda)$
be the condition number of $\lambda$ defined by
\begin{align}\label{eq:eigenvaluecondition}
\kappa (\lambda):= \dfrac{\lVert \bfu \rVert \lVert \bfv \rVert}{\lvert \bfu \cdot \bfv \rvert },
\end{align}
where \(\bfu,\bfv\in\bbC^n\) are the right and left eigenvectors, respectively, corresponding to $\lambda$.
Recall that \(\kappa (\lambda)\geq 1\), and equality holds when \(\bfu\) and \(\bfv\) are collinear (which is guaranteed if the matrix
is normal).
\begin{theorem}[Bound on the GMRES relative residual]\label{thm:main1}
Let \(\bfB\in \bbC^{n\times n}\)
be diagonalisable with simple eigenvalues.
Assume that all the eigenvalues of $\bfB$ are contained in ${\cal N} \cup {\cal H}$, where ${\cal N}$ is a bounded open set containing zero and ${\cal H}$ a closed half plane not containing zero. Without loss of generality, let ${\cal H}:=\{\Re z \geq S\}$ for some $S>0$ and assume that ${\cal H}$ contains at least one eigenvalue of $\bfB$, so that $S\leq \|\bfB\|_2$.
Let $\lambda_1,\ldots \lambda_\ell$ be the eigenvalues in ${\cal N}$, and let
$\kappa^*$ be the maximum eigenvalue condition number of $\bfB$.
Given $L_0, L_1$ with $0<L_0<L_1\leq S$, let
\begin{equation*}
N_{\rm eig}:= \big|\{j \, :\, \lambda_j \in \{ L_0 < \Re z < L_1 \} \cap{\cal N}\} \big| + 1.
\end{equation*}
Let
\begin{equation}\label{eq:delta}
\delta:= \frac{L_1-L_0}{4 n\,\kappa^* \, N_{\rm eig}}.
\end{equation}
Let $\beta\in (0,\pi/2)$ be defined by
\begin{equation}
\cos \beta =\frac{L_0}{\lVert \bfB\rVert_2+\delta}, \quad\text{ and let }\quad
\label{eq:gammabeta}
\gamma_\beta := 2\sin \left( \dfrac{\beta}{4-2\beta/\pi}\right)<1.
\end{equation}
Then, when GMRES is applied to the equation $\bfB \bfx=\bfb$, the $m$th GMRES residual \eqref{eq:residual} satisfies
\begin{align}\label{eq:Tues1new}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2}
\leq \left(\prod_{j=1}^\ell \dfrac{1}{\lvert \lambda_j \rvert}\right) \,(\|\bfB\|_2 + \delta)^{\ell+1}\,3^{\ell+1} \,\delta^{-1} \gamma_{\beta}^{m-\ell}.
\end{align}
\end{theorem}
\begin{corollary}[Sufficient condition on the number of GMRES iterations for convergence]\label{cor:SCW}
Under the assumptions of Theorem \ref{thm:main1}, given $\varepsilon>0$, if
\begin{align}
m\geq \ell+ \left(\log(\gamma_{\beta}^{-1})\right)^{-1} \left(
\sum_{j=1}^\ell\log\dfrac{1}{\lvert \lambda_j \rvert}+
\log (\varepsilon^{-1})+\log \left(\delta^{-1}\right)+
(\ell+1)\Big(\log\big(\lVert \bfB\rVert_2+\delta\big)+ \log 3\Big)\right),
\label{eq:mlowerbound}
\end{align}
then, when GMRES is applied to the equation $\bfB \bfx=\bfb$, the $m$th GMRES residual \eqref{eq:residual} satisfies
\begin{equation*}
\frac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2 }{ \lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2}\leq \varepsilon.
\end{equation*}
\end{corollary}
Figure \ref{fig:A3A4} shows the half-plane ${\cal H}$ in Theorem \ref{thm:main1}, and an example of a possible ${\cal N}$ in Theorem \ref{thm:main1}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.35\textwidth]{spectrum_decomposition_figure.pdf}
\caption{The sets ${\cal N}$ and ${\cal H}$ in the assumptions of Theorem \ref{thm:main1}.}\label{fig:A3A4}
\end{figure}
\begin{remark}[The dependence of $\gamma_\beta$ on $\|\bfB\|_2$.]
How $\gamma_\beta$ depends on $\|\bfB\|_2$ is not immediately clear from the definitions in \eqref{eq:gammabeta}.
However, if $\|\bfB\|_2 \gg \delta$ and $\|\bfB\|_2 \gg L_0$, then
\begin{equation}\label{eq:Wed1}
\left(\log(\gamma_{\beta}^{-1})\right)^{-1} = \frac{3\sqrt{3}}{4} \left(\frac{\|\bfB\|_2}{L_0} \right)\left( 1 + O\left(\frac{\delta}{\|\bfB\|_2}\right)\right)
\left( 1 + O\left(\frac{L_0}{\|\bfB\|_2}\right)\right).
\end{equation}
Indeed, let $\alpha := \pi/2 -\beta$ so that $\cos \beta = \sin \alpha$
and, from the definition of $\beta$,
\begin{equation}\label{eq:Wed2}
\alpha = \frac{L_0}{\|\bfB\|_2} \left( 1 + O\left(\frac{\delta}{\|\bfB\|_2}\right)\right).
\end{equation}
The definition of $\gamma_\beta$ then implies that, as $\alpha\rightarrow 0$,
\begin{equation}\label{eq:Wed3}
\gamma_\beta = 1 - \frac{4\alpha}{3\sqrt{3}} + O(\alpha^2) \quad\text{ so that } \quad -\log \gamma_\beta = \frac{4\alpha}{3\sqrt{3}} + O(\alpha^2).
\end{equation}
The asymptotics \eqref{eq:Wed1} then follow from combining \eqref{eq:Wed2} and \eqref{eq:Wed3}.
\end{remark}
\paragraph{Interpreting the bound \eqref{eq:mlowerbound}.}
The bound \eqref{eq:mlowerbound} shows that each outlier $\lambda_j$ contributes $C_1 + C_2 \log(1/|\lambda_j|)$ to the number of iterations needed to guarantee a prescribed residual reduction, where $C_1$ and $C_2$ are independent of $\lambda_j$ but depend on $\|\bfB\|_2$.
Therefore, if each $|\lambda_j|$ is large, only the number of outliers contributes to the required number of iterations.
If $|\lambda_j|$ is small, its value can have more of an effect on the required number of iterations, but this effect is mitigated by the fact that $|\lambda_j|$ appears in a logarithm.
\paragraph{The ideas behind the proof of Theorem \ref{thm:main1}.}
A convergence theory for GMRES based on modelling the eigenvalues as a ``cluster plus outliers'' was famously used in \cite{CaIpKeMe:96}, with the idea arising in the context of the conjugate-gradient method \cite{Je:77} and used subsequently, e.g., in \cite{ElSiWa:02}.
This theory in \cite{CaIpKeMe:96} forms the starting point for proving Theorem \ref{thm:main1}; see Lemma \ref{lem:Campbell} below.
However, a crucial difference is that we are interested in matrices depending on a parameter, namely $k$.
We therefore augment the theory in \cite{CaIpKeMe:96}, with
(i) the results in \cite{BeGoTy:05} about polynomial min-max problems, and
(ii) results about pseudospectra appearing in, e.g., \cite{TrEm:05}.
The result is that when the bound on the number of GMRES iterations \eqref{eq:mlowerbound} is applied with $\bfB = \bfM^{-1}\bfA_k'$, the $k$-dependence of the quantities in the bound (i.e.~$S, L_0, L_1, N_{\rm eig}, \delta, \beta$) is given from \emph{either}
existing $k$-explicit bounds on the norm and the norm of the inverse of $A_k'$ \emph{or} assumptions about the $k$-dependence of both the number and the condition numbers of the eigenvalues (see Assumptions A2 and A3 below).
We highlight that, in our use of the pseudospectrum, we choose $\delta$ as a function of $k$ to compensate for the growth of the non-normality with $k$.
This flexibility in choosing $\delta$ is mentioned in~\cite[Page 6]{Em:99}
when analysing different stages of the GMRES iteration for a single linear system; in contrast, here we use this flexibility applied to a family of linear systems parametrised by $k$.
\subsection{Second main result: $k$-explicit bound on the number of GMRES iterations
for Helmholtz BIEs under strong trapping}
\label{sec:mainresultH}
\subsubsection{Statement of assumptions.}
We write $A\lesssim B$ if there exists $C>0$, independent of all parameters of interest (including $h$ and $k$), such that $A\leq C B$, and $A\sim B$ if both $A\lesssim B$ and $A\gtrsim B$.
\begin{enumerate}
\item[A0] The meshwidth $h$ is chosen as a function of $k$ so that, given \({C_{{\rm approx},1}}>1\), \({C_{{\rm approx},2}}>1\), for all $k$,
(i) the Galerkin solution exists, is unique, and the relative $L^2(\Gamma)$-error of the Galerkin solution is bounded uniformly in $k$,
(ii) the second inequality in \eqref{eq:discon2} holds, i.e., $\mathbb{N}{\bfM^{-1}\bfA_k'}_{2} \leq {C_{{\rm approx},1}}\mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}$, and
(iii) the eigenvalues of both $A_k'$ and $\bfM^{-1}\bfA_k'$ are simple and
the eigenvalues of $A_k'$ are approximated by the eigenvalues
of $\bfM^{-1}\bfA_k'$ in the following sense:
at a given $k$, let the eigenvalues of $\bfM^{-1}\bfA_k'$ be $\lambda_1,\ldots,\lambda_n$ (where $n=n(k)$); there exists an injective function $f_k : \{1,\ldots,n\}\to \{\mu : \mu \text{ an eigenvalue of } A_k'\}$ such that
\begin{equation}\label{eq:A0iii}
|\lambda_j|\geq {C_{{\rm approx},2}} |f_k(j)|
\end{equation}
\item[A1]
There exists a bounded open set ${\cal N}$ containing zero and a closed half-plane ${\cal H}$ not containing zero such that
(i) all the eigenvalues of $\bfM^{-1}\bfA_k'$ are contained in ${\cal N}\cup {\cal H}$, and
(ii) ${\cal N}$ and ${\cal H}$ are both independent of $k$.
\item[A2] The number of eigenvalues of $\bfM^{-1}\bfA_k'$ in the set ${\cal N}$ in A1 is $\leq {C_{{\rm Weyl}}} k^{d-1}$, where ${C_{{\rm Weyl}}}>0$ is independent of $k$.
\item[A3] With $\kappa^*(k)$ the maximum eigenvalue condition number of $\bfM^{-1}\bfA_k'$, there exists ${C_{{\rm cond}}}, M>0$ and independent of $k$ such that $\kappa^*(k)\leq {C_{{\rm cond}}} k^M$.
\end{enumerate}
\paragraph{Why do we expect Assumptions A0-A3 to hold?}
As recapped in \S\ref{sec:statement:BEM}, there exist results on which functions $h=h(k)$ ensure A0(i), although they do not appear to be sharp.
A0(ii) and A0(iii) are ensured at least as $h\rightarrow 0$ for fixed $k$ when $\Gamma$ is $C^1$, with this regularity of $\Gamma$ ensuring that $A_k'$ (and also $B_{k, {\rm reg}}$) is a multiple of the identity plus a compact operator on ${L^2(\Gamma}%{\partial {\cO}_+})}$ (see Remark \ref{rem:compact}).
Indeed, in this case A0(ii) holds by Lemma \ref{lem:discon} below and stronger results than A0(iii) (showing that the eigenvalues of $\bfM^{-1}\bfA_k'$ converge to those of $A_k'$, with multiplicity) hold by \cite[Theorems 2, 3]{At:67}, \cite[Theorem Page 214]{At:75} (see also \cite[Theorem 7]{Sp:75}, \cite[Theorem 4.1]{SpTh:83}).
Note that \eqref{eq:A0iii} specifies that, with $\mu(k)$ an eigenvalue of $A_k'$ and $\lambda(k)$ an eigenvalue of $\bfM^{-1}\bfA_k'$, $|\lambda(k)|\gtrsim |\mu(k)|$. We do not require that $|\lambda(k)|\sim |\mu(k)|$ because we expect at least one $|\mu(k_j)|$ to be exponentially small when $k_j$ is a quasimode frequency (this is proved for the standard variational formulation, i.e., the basis of FEM, in \cite[Theorem 1.5]{GaMaSp:21}), but we expect that $|\lambda(k_j)|$ will be only algebraically small because of the sensitivity in F2. Note also that A0(iii) assumes that this sensitivity does not cause $|\lambda(k)|\ll |\mu(k)|$ for $k$ near $k_j$.
Regarding A1: first note that this corresponds to Observation O1 in \S\ref{sec:observations}.
When $\Omega_+$ is nontrapping $\|(A_k')^{-1}\|_{{\LtG\rightarrow \LtG}}\lesssim 1$ \cite[Theorem 1.13]{BaSpWu:16} and thus the smallest singular value of $\bfM^{-1}\bfA_k'\sim 1$ in this case.
This implies that when $\Omega_-$ is nontrapping the eigenvalues of $\bfM^{-1}\bfA_k'$ are $\sim 1$ away from zero.
Furthermore, at least for some nontrapping $\Omega_-$, the eigenvalues are contained in a $k$-independent half-plane away from zero since $A_k'$ is coercive (with constant independent of $k$) \cite{SpKaSm:15}, \cite{BeSp:11}. These facts suggest that the second part of A1 holds (i.e.~the half-plane ${\cal H}$ is independent of $k$), but are far from a proof.
Regarding Assumption A2: in \S\ref{sec:Weyl_new} we give heuristic arguments backing up this assumption, one based on Weyl-type asymptotics for
eigenvalues of the Laplacian on bounded domains, and the other based on the Observations O2(a)-(c) in \S\ref{sec:observations} and results about the number of resonances of the exterior Helmholtz problem.
Regarding Assumption A3: we did not give any experiments about this assumption in \S\ref{sec:intro_num}, but these are contained in \S\ref{sec:experiments}, and indicate that A3 holds for BEM discretisations of each of the BIEs \eqref{eq:direct_combined_dir},
\eqref{eq:direct_combined_neu}, and \eqref{eq:direct_combined_neu2}.
\subsubsection{$k$-explicit bounds on the number of GMRES iterations via Theorem \ref{thm:main1}}
\label{sec:Helmholtz}
Before applying Theorem \ref{thm:main1} to $\bfM^{-1}\bfA_k'$, we recall that GMRES applied to an $n\times n$ matrix converges in at most $n$ iterations (in exact arithmetic). This bound is well known to have ``little practical content'' \cite[Page 270]{TrBa:97} since one never reaches this number of iterations; nevertheless, it does give a theoretical upper bound on the $k$-dependence of the number of iterations. For example, when the $h$-BEM uses a fixed number of points per wavelength, $n\sim k^{d-1}$, and thus there exists $C>0$ such that if $m\geq C k^{d-1}$ then GMRES converges. However, the constant $C$ is
both large and dependent on the number of points per wavelength, and this is not what one sees in practice. For example, for the large cavity in 2-d with $k=100$, \(\theta = 4\pi/10\), and ten points per wavelength, $n= 1766$ and GMRES converges to tolerance \(10^{-6}\) in 165 iterations. For twenty points per wavelength, $n= 3528$ and GMRES converges to the same tolerance in 167 iterations.
\begin{theorem}[Bound for Helmholtz BIEs]
\label{thm:main2}
Let $\Omega_-$ be piecewise smooth. Assume that $n\leq C_{\rm{dis}}k^{M'}$ for some $M'>0$ and $C_{\rm{dis}}>0$.
Consider GMRES applied to the linear system
\begin{equation*}
\bfM^{-1}\bfA_k' \mathbf{u} = \bfM^{-1}\mathbf{f}.
\end{equation*}
where $\bfA_k'$ \eqref{eq:Galerkinmatrix} is the Galerkin matrix from the BEM discretisation of
the Dirichlet BIE \eqref{eq:direct_combined_dir} and $\bfM$ \eqref{eq:massmatrix} is the mass matrix.
If Assumptions A0-A3 hold, and $\lambda_1,\ldots \lambda_\ell$ are the eigenvalues of $\bfM^{-1}\bfA_k'$ in ${\cal N}$,
then there exists $C_j>0$, $j=1,2$ (independent of $k$) such that given $\varepsilon>0$, for all $k\geq 1$,
if
\begin{equation}\label{eq:mlowerbound2}
m\geq {C_{{\rm Weyl}}} k^{d-1} + C_1 \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}\Bigg(
\sum_{j=1}^\ell\log\dfrac{1}{\lvert \lambda_j \rvert}
+ C_2 k^{d-1}\log k
+ \log (\varepsilon^{-1})
\Bigg),
\end{equation}
then the $m$th GMRES residual satisfies $\|\bfr_m \|_2/\|\bfr_0 \|_2 \leq \varepsilon$.
Furthermore, $C_1$ only depends on ${C_{{\rm approx},1}}$, $\Omega_-$, and $d$, and $C_2$ only depends on ${C_{{\rm approx},1}}, {C_{{\rm Weyl}}}, {C_{{\rm cond}}}, C_{\rm{dis}},d, M,$ $M'$, and $\Omega_-$ with these constants as defined in Assumptions A0-A3.
\end{theorem}
An analogous result holds with $A_k'$ replaced by $B_{k,{\rm reg}}$, if $B_{k,{\rm reg}}$ satisfies appropriate analogues of Assumptions A0-A3; see Remark \ref{rem:Neumann} below.
The bound~\eqref{eq:mlowerbound2} gives insight into how the \(k\)-dependence of the number of iterations arises from the eigenvalue distribution of $\bfM^{-1}\bfA_k'$. Moreover, with $M'$ and $C_{\rm dis}$ fixed, the constants in \eqref{eq:mlowerbound2} are independent of the choice of $n\leq C_{\rm{dis}} k^{M'}$ (provided that Assumptions A0-A3 hold). Therefore, choosing $M'>d-1$ and $C_{\rm dis}>0$, discretizations satisfying $n\leq C_{\rm{dis}} k^{M'}$ include those with an arbitrary number of points per wavelength, and the bound \eqref{eq:mlowerbound2} holds, at least for sufficiently large $k$, uniformly across all of them (which appears consistent with the specific examples of numbers of iterations stated above the theorem). However, the right-hand side of \eqref{eq:mlowerbound2} contains terms that grow faster than $k^{d-1}$ and so, if $n\sim k^{d-1}$, then the $k$-dependence of \eqref{eq:mlowerbound2} is worse than that of the crude bound that GMRES converges in at most $n$ iterations (in exact arithmetic).
\paragraph{Informal explanation of how \eqref{eq:mlowerbound2} arises from \eqref{eq:mlowerbound}.}
When $\bfB =\bfM^{-1}\bfA_k'$, Assumption A1 implies that $S\sim 1$, and thus we can choose $L_0, L_1\sim 1$. Then $\delta$ defined by \eqref{eq:delta} is $\lesssim 1$. By Assumption A0, $\mathbb{N}{\bfM^{-1}\bfA_k'}_2 \lesssim \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}$, which grows at most algebraically with $k$ (see Lemma \ref{lem:normbounds} below). Using this, along with \eqref{eq:Wed1}, we find that the bound on $m$ \eqref{eq:mlowerbound} holds if
\begin{align}
m -\ell \gtrsim \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}\left(
\sum_{j=1}^\ell\log\dfrac{1}{\lvert \lambda_j \rvert}+
\log (\varepsilon^{-1})+\log \left(\delta^{-1}\right)+ \ell \,C \log k
\right),
\label{eq:mlowerbound_explain}
\end{align}
for some $C>0$ independent of $k$.
Assumption A2 is that $\ell \leq{C_{{\rm Weyl}}} k^{d-1}$, and, by definition, $N_{\rm eig}\leq \ell +1$. Assumption A3 and the bound $N_{\rm eig}\leq {C_{{\rm Weyl}}} k^{d-1} +1$ then imply that $\delta^{-1}$ grows at most polynomially in $k$; the bound \eqref{eq:mlowerbound2} then follows from using these bounds in \eqref{eq:mlowerbound_explain}.
\subsubsection{Discussion of the $k$-dependence of the bound in Theorem \ref{thm:main2}, how this bound explains F3(a), and how this bound could be improved.}\label{sec:discussion}
To investigate the sharpness of the bound \eqref{eq:mlowerbound2} in Theorem \ref{thm:main2}, we summarise the results of the numerical experiments from \S\ref{sec:numA2} and \S\ref{sec:experiments} in Table \ref{tab:kdependence}. This table
plots the $k$-dependence for $k\in (50,290)$, for both the small and large cavities, of (i) the number of iterations, (ii) the number of outlier eigenvalues $\ell$ when ${\cal N}:= [-0.1,0.1]\times[-0.6,0.6]$ (i.e., the black rectangle in Figure \ref{fig:outliers}), (iii) $\|\bfM^{-1}\bfA_k'\|_2$ (and its analogue for the two Neumann BIEs), and (iv) the quantity
\begin{equation}\label{eq:cL}
{\cal L}:= \sum_{j=1}^\ell\log\dfrac{1}{\lvert \lambda_j \rvert}.
\end{equation}
Each of $\|\bfM^{-1}\bfA_k'\|_2$ and its Neumann analogues has the same $k$-dependence for both the small and large cavities -- see the top-left plots in Figures \ref{fig:Dirichlet}, \ref{fig:Neumann}, and \ref{fig:reg_Neumann} below -- and so the norm only appears in one column in Table \ref{tab:kdependence}.
The exponents in Table \ref{tab:kdependence} are obtained using
the nonlinear least-squares Marquardt-Levenberg algorithm (the basis of the `fit' command in gnuplot).
\begin{table}[h!]
\begin{tabular}{c|| c| c| c| c|c|c|c|c|c|}
&
$\#$ it.~small & $\#$ it.~large
& $\ell$ small & $\ell$ large & norm & ${\cal L}$ small
& ${\cal L}$ large \\
\hline\hline
Dirichlet $A_k'$ & 0.66 & 0.82 & 0.95 & 1.00 & 0.31 & 0.77 & 0.90 \\
Neumann $B_k$ & 0.55 & 0.77 & 0.93 & 0.97 & 0 & 0.78 &0.91 \\
reg.~Neumann $B_{k, {\rm reg}}$ & 0.60 & 0.80 & 0.95 & 0.95 & 0 &0.79 & 0.89
\end{tabular}
\caption{The exponents in how the quantities in the columns vary with $k$ in 2-d (i.e.~if a quantity $\sim k^a$, then $a$ is displayed), as determined by numerical experiments through $k=k^e_{m,0}$ for $k\in (50,290)$.}\label{tab:kdependence}
\end{table}
We structure our discussion around the following points.
\paragraph{The number of iterations grows slightly less than $k^{d-1}$ for the large cavity.} Table \ref{tab:kdependence} shows that, in 2-d, the number of iterations roughly $\sim k^{0.6}$ for the small cavity and $\sim k^{0.8}$ for the large cavity, for each of the three BIEs. Figure \ref{fig:3d} below shows that for the Dirichlet problem in 3-d the number of iterations
grows roughly like $k^2$ for both the small and large cavities
over the range $k\in (20,110.5)$ -- note that this is a smaller range than we consider in 2-d.
Similarly, Figure \ref{fig:3d_neu} below shows that for the Neumann problem in 3-d the number of iterations for both $B_k$ and $B_{k, {\rm reg}}$ grows roughly like $k^2$ for the small cavity over the range $k\in (20,50)$.
In both 2- and 3-d, the number of iterations therefore grows with $k$ at roughly the same rate as the number of degrees of freedom, illustrating how difficult a problem this is.
\paragraph{The bound \eqref{eq:mlowerbound2} will always give $m\gtrsim k^{d-1}$ because $m\gtrsim \ell$ and $\ell \sim k^{d-1}$.}
Assumption A2 is that $\ell \lesssim k^{d-1}$; in \S\ref{sec:Weyl_new} we present heuristic arguments based on Weyl asymptotics why A2 holds, and the numerical experiments for $d=2$ indicate that $\ell \sim k$. Therefore, since the right-hand side of \eqref{eq:mlowerbound} contains $\ell$, the bound \eqref{eq:mlowerbound2} gives that $m\gtrsim k^{d-1}$ no matter what bounds we obtain on the other quantities in \eqref{eq:mlowerbound}.
\paragraph{The near-zero eigenvalues are, at worst, exponentially small.}
By Assumption A3(iii), $|\lambda_j|^{-1}\lesssim |\mu|^{-1}$, where $\mu:=f_k(\lambda_j)$ is an eigenvalue of $A_k'$.
Since
$|\mu|^{-1}\leq \|(A_k')^{-1}\|_{{\LtG\rightarrow \LtG}}$
and, at least for smooth $\Omega_-$, $\|(A_k')^{-1}\|_{{\LtG\rightarrow \LtG}}\lesssim \exp(\alpha k)$ by \cite[Equation 1.35 and Lemma 6.2]{ChSpGiSm:20},
$|\mu|^{-1} \lesssim \exp(\alpha k)$.
\paragraph{The bound \eqref{eq:mlowerbound2} explains F3(a)
because it shows that the actual position of each near-zero eigenvalue is less important than the total number of near-zero eigenvalues.}
Indeed Figures \ref{fig:example_spectrum}, \ref{fig:spectra_illustrations_elliptic_cavity_2D_dir}, and \ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir} indicate that the near-zero eigenvalues are not all simultaneously close to zero.
Since each $\lambda_j$ enters the bound on $m$ \eqref{eq:mlowerbound2} via the $\log(1/|\lambda_j|)$ in ${\cal L}$ \eqref{eq:cL}, even if one of the $\lambda_j$ is exponentially small (which we expect to be very unlikely by F2), we expect the growth of ${\cal L}$ to still be dominated by the overall number of near-zero eigenvalues, i.e., $\ell \sim k^{d-1}$. This is consistent with the fact that $\ell \sim k$ and ${\cal L}\sim k^{0.9}$ in the experiments for the large cavity.
Making this argument rigorous and proving that ${\cal L}\lesssim k^{d-1}$ would involve first bounding
\begin{equation}\label{eq:boundcL}
{\cal L}:=\sum_{j=1}^\ell \log\frac{1}{|\lambda_j|}\lesssim
\sum_{\substack{\mu \in \sigma(A_k'),\, \mu \in \widetilde{{\cal N}}}}
\log\frac{1}{|\mu|},
\end{equation}
where $\widetilde{{\cal N}}$ is a neighbourhood of ${\cal N}$ (depending on ${C_{{\rm approx},2}}$) containing the images of each $\lambda_j\in {\cal N}$ under $f_k$, and then controlling the number of eigenvalues that can simultaneously be exponentially-close to zero. To our knowledge, the question of whether there exists strong trapping with high multiplicity of quasimodes/resonances is still open.
Even if we knew that all near-zero eigenvalues correspond to localised eigenfunctions of the ellipse, controlling this number
involves understanding the number of eigenvalues exponentially-close together. This could be obtained from proving a Weyl law with $k^{d-2}$ remainder, but this has only been established so far for the torus for $d\geq 5$ \cite{Fr:82}, and is known to be typically false (e.g., on the torus for $d=2$ \cite{Ha:15}).
\paragraph{The norms of the operators grow with $k$ in the limit $k\rightarrow \infty$, and so, for $k$ sufficiently large and assuming ${\cal L}\lesssim k^{d-1}$, the bound \eqref{eq:mlowerbound2} will be dominated by $\|A_k'\|_{{\LtG\rightarrow \LtG}} k^{d-1}\log k$.}
How $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ depends on both $k$ and the geometry of $\Omega_-$ is now well-understood thanks to \cite{ChGrLaLi:09}, \cite[Appendix A]{HaTa:15}, \cite{GaSm:15}, \cite[Chapter 4]{Ga:19}, and \cite{GaSp:19}. These results show that, as $k\rightarrow \infty$, $k^{1/2}\lesssim \|A_k'\|_{{\LtG\rightarrow \LtG}}\lesssim k^{1/2}\log k$ for both the small and large cavities, and we expect that the ideas behind these results can be used to show that
$\|B_{k, {\rm reg}}\|_{{\LtG\rightarrow \LtG}}\gtrsim k^{1/6}$ for these domains;
see the discussion in \S\ref{sec:numbounds} about the top-left plots.
Table \ref{tab:kdependence} shows, however, that in the range $k \in(50,290)$ the growth of $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ is slower than $k^{1/2}$, and $\|B_{k, {\rm reg}}\|_{{\LtG\rightarrow \LtG}}$ does not grow. For $\|A_k'\|_{{\LtG\rightarrow \LtG}}$, this discrepancy is explained in \S\ref{sec:numbounds}.
\paragraph{Limitations of the ``cluster plus outliers'' model applied to $\bfM^{-1}\bfA_k'$ and how it could be improved.}
The limitations of the ``cluster plus outliers'' model where the ``outliers'' are the near-zero eigenvalues are shown in
Figures \ref{fig:example_spectrum}, \ref{fig:spectra_illustrations_elliptic_cavity_2D_dir}, and \ref{fig:spectra_illustrations_elliptic_cavity_bigger_2D_dir}. Indeed, these plots show that the ``cluster'' of eigenvalues away from zero is itself a cluster with outliers. Furthermore, this ``cluster within the cluster'' appears to be contained in a $k$-independent set (see Figure \ref{fig:example_spectrum}).
We therefore expect that a bound with improved $k$-dependence could be obtained by
taking this additional structure into account. Indeed, $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ currently enters the bound \eqref{eq:mlowerbound2} and as a bound on the modulus of the cluster eigenvalues.
If one could prove that the number of the ``outliers of the cluster'' is $\lesssim k^{d-1}$ and that the ``cluster within the cluster'' is contained in a $k$-independent set, then this would replace the bound \eqref{eq:mlowerbound2} with a bound of the form $m \gtrsim k^{d-1} + {\cal L}$. If one could, in addition, prove that ${\cal L}\lesssim k^{d-1}$ (as discussed above), then this would prove the sharp bound $m\gtrsim k^{d-1}$ (with the omitted constant independent of $k$ and properties of the discretisation).
\subsection{Applicability of the ideas in this paper to Helmholtz FEM.}\label{sec:FEM}
Until now we have focused on solving the scattering problem \eqref{eq:Helmholtz} using BIEs and BEM, however
the general result of Theorem \ref{thm:main1} can be applied to other Helmholtz discretisations satisfying Assumptions A0-A3
(or suitably modified versions of these).
\paragraph{Location and number of the near-zero eigenvalues for FEM.} As mentioned in \S\ref{sec:observations}, the connection between quasimodes and near-zero eigenvalues of the standard domain-based variational formulation (i.e.~the basis of FEM) is much clearer than for BEM, and this is subject of the companion paper \cite{GaMaSp:21}. Indeed, \cite[Theorem 1.5]{GaMaSp:21} proves that if $k=k_j$, then there exists a near-zero eigenvalue
of the standard domain-based variational formulation, with the distance of this eigenvalue from zero given in terms of the quality $\epsilon(k_j)$ of the quasimode. Furthermore,
\cite[Theorem 1.8]{GaMaSp:21} shows that the eigenvalues inherit the multiplicities of the quasimodes. These results are proved using arguments from microlocal and complex analysis, inspired by the celebrated ``quasimodes to resonances'' results of
\cite{TaZw:98}, \cite{St:99} (following \cite{StVo:95, StVo:96}); see also \cite[Theorem 7.6]{DyZw:19}.
We highlight that while the number of near-zero eigenvalues in the FEM case is the same as for BEM (namely $\sim k^{d-1}$) the number of degrees of freedom for FEM is much larger than that for BEM. Indeed, while BEM is commonly used with a fixed number of points per wavelength, leading to systems of size $\sim k^{d-1}$, the pollution effect means that FEM is used with systems of size $\gg k^d$.
Therefore, the issue we encountered in \S\ref{sec:Helmholtz} for BEM that the number of near-zero eigenvalues is the same order as total number of degrees of freedom does not occur for FEM.
\paragraph{Location of the other eigenvalues for FEM.}
While the near-zero eigenvalues for FEM are easier to understand rigorously than those for BEM, one subtlety in the FEM case is that the eigenvalues away from zero need not be in a half-plane (as in A1). If \emph{either} the exact Dirichlet-to-Neumann map \emph{or} an impedance boundary condition is used on the truncation boundary, then the numerical range (and hence the eigenvalues) is contained in the lower-half plane. Furthermore, if the problem is nontrapping, then the eigenvalues are contained in a half-plane an $O(1)$ distance below the real axis \cite[Theorem 1.12]{BaSpWu:16}, suggesting that A1 holds with ${\cal N}$ and ${\cal H}$ independent of $k$ (see also \cite[Theorem 5.1]{LiXiSaDe:20} for stronger results on the eigenvalue distribution of a simple nontrapping problem).
However, if a perfectly-matched layer (PML) is used, then the numerical range of the operator contains elements in the upper-half plane, as can be seen from \cite[Equation after (2.12)]{LiWu:19}.
Nevertheless, we expect that a similar result to Theorem \ref{thm:main1} can be proved under a modified version of A1 by
replacing the domain $K_\beta\subset \mathbb{C}$ in Lemma \ref{lem:bound_polynomial_on_lens} below by a non-convex domain, such as one of the class introduced in \cite{KoLi:00};
see, e.g., the discussion in \cite[\S3.1.2]{LiTi:04}.
\paragraph{Preconditioning FEM discretisations.} The reason
we have focused on BEM (and not FEM) in this paper is that
one usually seeks to precondition GMRES applied to the FEM discretisation of the standard variational formulation of the Helmholtz equation. This is because the number of GMRES iterations without preconditioning grows rapidly with $k$ even in non-trapping scenarios. This is in contrast to BEM, where the number of GMRES iterations for discretisations of BIEs \eqref{eq:direct_combined_dir} and \eqref{eq:direct_combined_neu2} enjoy mild growth with $k$ in nontrapping situations; see \cite[Theorem 1.16 and Figure 1]{GaMuSp:19} for \eqref{eq:direct_combined_dir} and \cite[Tables 1 and 2]{BoTu:13} for \eqref{eq:direct_combined_neu2}.
The design of good preconditioners for GMRES applied to the Helmholtz FEM in nontrapping scenarios is a very active area of research; see the literature reviews in \cite{Er:08, ErGa:12, GaZh:19}, and \cite[\S1.3]{GrSpZo:20}.
Since our theory below only proves bounds for GMRES applied to unpreconditioned matrices,
our results are less interesting for FEM than for BEM.
Nevertheless, our results still provide insight into the design of preconditioners for trapping problems -- this is discussed in the conclusions \S\ref{sec:conclusions}.
\section{Proofs of Theorems \ref{thm:main1} and \ref{thm:main2}}
\subsection{Definition of spectral projectors}\label{sec:spectral_proj}
Given $\bfB \in \mathbb{C}^{n\times n}$,
let $\lambda_1, \ldots, \lambda_\ell$ be a subset of the eigenvalues of $\bfB$ (we later choose this subset to be the eigenvalues in ${\cal N}$ for a matrix satisfying the assumptions of Theorem \ref{thm:main1}, but the results in this subsection hold more generally).
Let $\Gamma_j$, $j=1,\ldots,\ell$, be a circle enclosing $\lambda_j$ but no other eigenvalue of $\bfB$, and let $\Gamma = \cup_{j=1}^\ell \Gamma_j$. Let $\widetilde{\Gamma}$ be a positively-oriented curve enclosing the rest of the spectrum.
Let \(\bfR_k(z):=\left(z \bfI - \bfB\right)^{-1}\), i.e.~$\bfR_k(z)$ is the resolvent of $\bfB$.
As in, e.g.,~\cite{CaIpKeMe:96,Em:99,Sa:11}, we define the spectral projectors of $\bfB$ on \(\Gamma_j\) and \(\widetilde{\Gamma}\) by
\begin{align*}
\projJ:= \dfrac{1}{2\pi i}\int_{\Gamma_j} \bfR_k(z)
\dif z,\quad 1\leq j \leq \ell,\quad\text{ and }\quad \projcl:= \dfrac{1}{2\pi i}\int_{\widetilde{\Gamma}}
\bfR_k(z) \dif z.
\end{align*}
Let
\begin{equation}\label{eq:projout}
\projout=\sum_{j=0}^\ell \projJ.
\end{equation}
The residue theorem implies that
\begin{align}\label{eq:projector_relation}
\projout+\projcl=\bfI
\end{align}
(see, e.g., \cite[Theorem 3.3., Page 53]{Sa:11})
and properties of holomorphic functional calculus imply that
\begin{align*}
\projout \projout =\ \projout,
\quad \projcl \projcl = \projcl,
\quad \projcl\projout=0,
\quad \projcl\bfB = \bfB \projcl,
\quad \projout\bfB = \bfB \projout.
\end{align*}
Let $r_j$ be the index of $\lambda_j$, i.e., the dimension of the largest Jordan block associated with $\lambda_j$. Then
\(\range (\projJ)= \ker ((\lambda_j \bfI - \bfB)^{r_j})\) where \(r_j\) is the index of \(\lambda_j\), see~\cite[Lemma 3.1]{Sa:11}.
This last property implies that
\begin{align}\label{eq:Z_w_minimal_poly}
\begin{aligned}
\projJ (\bfI - \lambda_j^{-1}\bfB)^{r_j}&= (\bfI - \lambda_j^{-1}\bfB)^{r_j}\projJ = 0,
\end{aligned}
\end{align}
for $1\leq j\leq \ell$. Finally, let \(r:= \sum_{j=1}^{\ell}r_j\).
\subsection{The ideas of the proofs}\label{sec:idea}
\paragraph{Idea 1: use the ``cluster plus outliers'' model from \cite{CaIpKeMe:96}.}
The starting point is the bound in the following lemma (proved in \S\ref{sec:proofCampbell} below), which is essentially that in \cite[Proposition 4.1]{CaIpKeMe:96}.
\begin{lemma}\label{lem:Campbell}
Let \(\bfB\in \bbC^{n\times n}\) and $\widetilde{\Gamma}$ be as in \S\ref{sec:spectral_proj}.
If $m\geq r$, then
\begin{align}\label{eq:bound_campbell}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2}
\leq \dfrac{1}{2\pi} \lvert \widetilde{\Gamma} \rvert \min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}} \max_{z \in \widetilde{\Gamma}} \left(\prod_{j=1}^\ell \dfrac{\lvert \lambda_j - z \rvert }{\lvert \lambda_j \rvert} \lVert \bfR_k(z) \rVert_2 \lvert p_{m-r}(z) \rvert \right).
\end{align}
\end{lemma}
We now need to use the freedom we have in choosing
\(\widetilde{\Gamma}\) to bound on this curve the three terms in brackets on the right-hand side of \eqref{eq:bound_campbell}, namely,
the distance of the outliers to \(\widetilde{\Gamma}\) (i.e.~$|\lambda_j-z|$), the norm of the resolvent (i.e. \(\lVert \bfR_k(z) \rVert_2\)), and the polynomial \(p_{m-r}\).
\paragraph{Idea 2: choose the shape of $\widetilde{\Gamma}$ so that one can use the min-max result of \cite{BeGoTy:05}.}
To bound the polynomial on the right-hand side of \eqref{eq:bound_campbell}, we use the following result of \cite{BeGoTy:05}. Given a compact set \(K \subset \mathbb{C}\), let
\begin{align}\label{eq:Em}
E_m(K):=\min_{\substack{p_{m}\in \bbP_{m}, \\ p_{m}(0)=1}} \max_{z \in K} \big| p_{m}(z) \big|
\end{align}
\begin{lemma}[{\cite[Lemma 2.2]{BeGoTy:05}}]\label{lem:bound_polynomial_on_lens}
Given \(\beta\in (0,\pi/2)\),
let \(K_\beta \subset \bbC\) be defined by
\begin{equation*}
K_\beta := \big\{z\,:\, |z|\leq 1\big\} \cap \big\{z\,:\, \Re(z) \geq \cos (\beta) \big\}.
\end{equation*}
Let $\gamma_\beta$ be defined by \eqref{eq:gammabeta}.
Then, for any $m\in \mathbb{Z}^+$,
\begin{align*}
E_m(K_\beta) \leq \min \left\{2+\gamma_\beta,\dfrac{2}{1-\gamma_\beta^{m+1}}\right\}\gamma_\beta^m.
\end{align*}
\end{lemma}
Observe that, for $a\in \mathbb{C}\setminus\{0\}$, $E_m(aK)= E_m(K)$, since
\begin{equation*}
E_m(aK) = \min_{\substack{p_{m}\in \bbP_{m}, \\ p_{m}(0)=1}} \max_{z \in aK} \big| p_{m}(z) \big|
=\min_{\substack{p_{m}\in \bbP_{m}, \\ p_{m}(0)=1}} \max_{\widetilde{z} \in K} \big| p_{m}(a\widetilde{z}) \big|
=\min_{\substack{\widetilde{p}_{m}\in \bbP_{m}, \\ \widetilde{p}_{m}(0)=1}} \max_{\widetilde{z} \in K} \big| \widetilde{p}_{m}(\widetilde{z}) \big|
\end{equation*}
where $\widetilde{p}_m(\widetilde{z})= p_m(a\widetilde{z})$.
We therefore choose $\widetilde{\Gamma}$ to be a scaling of $\partial K_\beta$, i.e. $\widetilde{\Gamma} = R(\partial K_\beta)$ for some $R>0$ -- see Figure \ref{fig:lens} -- and use Lemma \ref{lem:bound_polynomial_on_lens} to bound the term involving $p_{m-r}$ in \eqref{eq:bound_campbell}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{spectrum_decomposition.pdf}
\caption{The contour $\widetilde{\Gamma}:= \widetilde{\Gamma}_1 + \widetilde{\Gamma}_2$ in the complex $z$ plane.}\label{fig:lens}
\end{figure}
\paragraph{Idea 3: choose the parameters defining $\widetilde{\Gamma}$ to control the resolvent.}
We choose $\widetilde{\Gamma}$ to be the boundary of $R K_\beta$ with $R:= \|\bfB\|_2+\delta$ for some $\delta>0$. This choice of $R$ ensures that $B(0,R)$ encloses all the eigenvalues of $\bfB$.
We now use the freedom we have in choosing $\delta>0$ and $\beta \in (0,\pi/2)$ to control $\|\bfR(z)\|_2$ on $\widetilde{\Gamma}$.
Since
\begin{equation*}
\big\|(z \bfI - \bfB)^{-1}\big\|_2 \leq \frac{1}{|z| - \|\bfB\|_2} \quad\text{ for } |z| \geq \|\bfB\|_2,
\end{equation*}
we have
\begin{align}\label{eq:resolvent_bound}
\lVert \bfR_k(z) \rVert_2\leq \delta^{-1} \quad \text{ for } z \in \partial B(0,R).
\end{align}
To bound $\lVert \bfR_k(z)\rVert_2$ on the straight part of \((\lVert \bfB\rVert_2+\delta)K_\beta\) we
choose $\delta$ and $\beta$ so that this straight parts avoids
the $\delta$-pseudospectrum of $\bfB$, $\Lambda_{\delta}(\bfB)$, defined by
\begin{align}\label{eq:pseudospectrum}
\Lambda_{\delta}(\bfB):= \big\{z \in \bbC \,\, :\,\, \lVert (z \bfI - \bfB)^{-1}\rVert_2\geq \delta^{-1} \big\}.
\end{align}
Avoiding $\Lambda_{\delta}(\bfB)$ is possible with $\delta$ sufficiently small because of the following result.
\begin{theorem}[{Bauer-Fike-type theorem \cite[Theorem 52.2]{TrEm:05}}]\label{thm:BF1}
If $\bfB \in \mathbb{C}^{n\times n}$ has $n$ simple eigenvalues, then, for all $\delta>0$,
\begin{equation*}
\Lambda_{\delta}(\bfB) \subseteq \bigcup_{j=1}^n \Big(\lambda_j+
B\big(0, \delta \,n\, \kappa(\lambda_j)\big)\Big).
\end{equation*}
\end{theorem}
The price we pay for using this general result is that $\delta$ can potentially be small. However, when the resulting bound is applied to Helmholtz BEM, the smallness of $\delta$ is not the limiting factor in the $k$-dependence of the bound.
\subsection{Proof of Lemma \ref{lem:Campbell}}\label{sec:proofCampbell}
As in~\cite{CaIpKeMe:96}, we define the minimal polynomial associated with \((\lambda_j)_{j=1}^\ell \). Let
\begin{align}\label{eq:qr}
q_r(z):=\prod_{j=1}^\ell (1 -\lambda_j^{-1} z)^{r_j};
\end{align}
observe that \(q_r\in \bbP_r\) (since \(r:= \sum_{j=1}^{\ell}r_j\)) and \(q_r(0)=1\).
The significance of $q_r(z)$ is shown by the following lemma.
\begin{lemma}\label{lem:mainevent}
\begin{equation*}
\projout q_r(\bfB) = 0.
\end{equation*}
\end{lemma}
\begin{proof}
By the definitions of $q_r$ \eqref{eq:qr} and $\projout$ \eqref{eq:projout}, and then \eqref{eq:Z_w_minimal_poly},
\begin{equation*}
\projout q_r(\bfB)
= \left(\sum_{m=1}^\ell \bfP_{\Gamma_m}\right)\left( \prod_{j=1}^\ell \big(\bfI - \lambda_j^{-1}\bfB\big)^{r_j} \right)
= \sum_{m=1}^\ell \left( \prod_{j=1}^\ell \underbrace{
\bfP_{\Gamma_m}\big(\bfI - \lambda_j^{-1}\bfB\big)^{r_j}
}_{=0 \text{ when } j=m}
\right).
\end{equation*}
\end{proof}
Let \(p_{m-r}\in \bbP_{m-r}\) be any polynomial of order \(m-r\) such that \(p_{m-r}(0)=1\). Let
\begin{equation*}
\overline{p}_m (\bfB):= q_r(\bfB) p_{m-r}(\bfB),
\end{equation*}
so that \(\overline{p}_m\in \bbP_m\) with \(\overline{p}_m(0)=1\).
Then, by
\eqref{eq:projector_relation} and Lemma \ref{lem:mainevent},
\begin{align*}
\overline{p}_m (\bfB)&=\projout q_r(\bfB) p_{m-r}(\bfB) + \projcl q_r(\bfB) p_{m-r}(\bfB)
= \projcl q_r(\bfB) p_{m-r}(\bfB).
\end{align*}
Using this in the characterisation \eqref{def:gmres} of the GMRES residual, we find that
\begin{align}\label{eq:local_bound_on_outliers}
\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2 &\leq \min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}}\lVert \projcl q_r(\bfB)p_{m-r}(\bfB) \bfr_0(\bfB,\bfb,\bfx_0)\rVert_2.
\end{align}
By the definition of $\projcl$,
\begin{align*}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2} &\leq \min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}} \dfrac{1}{2\pi} \int_{\widetilde{\Gamma}}\lVert \bfR_k(z) \rVert_2 \lvert q_r(z)p_{m-r}(z) \rvert \dif z,
\end{align*}
and then \eqref{eq:bound_campbell} follows from the definition of $q_r$.
\subsection{Proof of Theorem \ref{thm:main1}}
We first observe that it is sufficient to prove that
\begin{align}\label{eq:Tues1}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2}
\leq \left(\prod_{j=1}^\ell \dfrac{1}{\lvert \lambda_j \rvert}\right) \,(\|\bfB\|_2 + \delta)^{\ell+1}\,2^{\ell} \,\delta^{-1} (\gamma_{\beta}+2) \gamma_{\beta}^{m-r}.
\end{align}
Indeed, \eqref{eq:mlowerbound} then follows by using the inequalities $2\leq 3$ and $\gamma_\beta + 2\leq 3$, and noting that the assumption that $\bfB$ has simple eigenvalues implies that $r_j=1$ for $j=1,\ldots,\ell$, so $r=\ell$.
To prove \eqref{eq:Tues1}, we start from the bound \eqref{eq:bound_campbell}, and then follow Ideas 2 and 3 in \S\ref{sec:idea}.
Indeed, we set $\widetilde{\Gamma}:= R K_{\widetilde{\beta}}$, where $R:= \|\bfB\|_2 + \delta$, with $\widetilde{\beta}\in (0,\pi/2)$ and $\delta>0$ free parameters to be fixed later.
Let $\widetilde{\Gamma}_1$ and $\widetilde{\Gamma}_2$ be as in Figure \ref{fig:lens}.
Since the spectrum of $\bfB$ is discrete, for \(\delta\) small enough there exists $L$ with $L_0<L<L_1$ such that that line $\Re z = L$
does not intersect the $\delta$-pseudospectrum \(\Lambda_\delta(\bfB)\).
Indeed, combining Theorem \ref{thm:BF1} and the definitions of $N_{\rm eig}$ and $\kappa^*$, we see that this is possible if
\begin{equation*}
2 \delta \, n \, \kappa^*\, N_{\rm eig} < L_1-L_0,
\end{equation*}
and thus certainly if $\delta$ is given by \eqref{eq:delta}.
With this choice of $\delta$ and the associated $L$,
let $\widetilde{\beta}$ and $\beta$ be defined so that
\begin{equation}\label{eq:widetildebeta}
\cos\widetilde{\beta} = \frac{L}{R} \quad\text{ and }\quad \cos\beta = \frac{L_0}{R},
\end{equation}
and observe that $K_{\widetilde{\beta}}\subset K_{\beta}$ since $L>L_0$.
In summary, $\widetilde{\Gamma}:= R K_{\widetilde{\beta}}$ with $\delta$ defined by \eqref{eq:delta},
$R:= \|\bfB\|_2 + \delta$,
$L$ is defined so that
the line $\Re z = L$
does not intersect the $\delta$-pseudospectrum \(\Lambda_\delta(\bfB)\), and
$\widetilde{\beta}$ is defined by \eqref{eq:widetildebeta}.
Having defined $\widetilde{\Gamma}$, we now bound the quantities appearing on the right-hand side of \eqref{eq:bound_campbell}.
Since $\widetilde{\Gamma}\subset B(0,R)$,
\begin{equation}\label{eq:Tues2}
\max_{z \in \widetilde{\Gamma}}\left(\prod_{j=1}^\ell \lvert \lambda_j - z \rvert \right)\leq ( 2R)^\ell
\quad\text{ on }\quad\widetilde{\Gamma}.
\end{equation}
Furthermore, the bound \eqref{eq:resolvent_bound} implies that $\lVert \bfR(z) \rVert_2 \leq \delta^{-1}$ on $\widetilde{\Gamma}_2$ and the
choice of $L$ and the definition of $\Lambda_\delta(\bfB)$ \eqref{eq:pseudospectrum} implies that $\lVert \bfR(z) \rVert_2 \leq \delta^{-1}$ on $\widetilde{\Gamma}_2$; therefore
\begin{equation}\label{eq:Tues3}
\lVert \bfR(z) \rVert_2 \leq \delta^{-1} \quad\text{ on } \widetilde{\Gamma}.
\end{equation}
Using \eqref{eq:Tues2} and \eqref{eq:Tues3} in \eqref{eq:bound_campbell}, we find that
\begin{align*}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_2}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_2}
\leq \left(\prod_{j=1}^\ell \dfrac{1}{\lvert \lambda_j \rvert}\right) \, R\, (2R)^\ell \, \delta^{-1} \min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}} \max_{z \in \widetilde{\Gamma}} \lvert p_{m-r}(z) \rvert.
\end{align*}
Using the fact that $\widetilde{\Gamma}= \partial( R \,K_\beta)$,
the definition of $E_{m-r}(K_\beta)$, and the fact that $K_{\widetilde{\beta}}\subset K_{\beta}$, we have
\begin{align*}
\min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}} \max_{z \in \widetilde{\Gamma}} \lvert p_{m-r}(z) \rvert &\leq \min_{\substack{p_{m-r}\in \bbP_{m-r}, \\ p_{m-r}(0)=1}} \max_{z \in R K_{\beta}} \lvert p_{m-r}(z) \rvert = E_{m-r}(R K_{\widetilde\beta})
\leq E_{m-r}(R K_{\beta}).
\end{align*}
The result \eqref{eq:Tues1} then follows from using
Lemma \ref{lem:bound_polynomial_on_lens}.
\begin{remark}[Removing the assumption that the eigenvalues are simple]
\label{rem:BF1}
We assumed that the eigenvalues of $\bfB$ were simple to use Theorem \ref{thm:BF1}.
To remove this assumption,
one can use the Bauer-Fike theorem (see, e.g, ~\cite[Theorem 2.3]{TrEm:05}) that
\begin{equation*}
\Lambda_{\delta}(\bfB)\subset \Lambda(\bfB)+B(0,\delta \kappa(\bfV)),
\end{equation*}
for \(\bfB=\bfV {\bf \Lambda} \bfV^{-1}\) with ${\bf \Lambda}$ a diagonal matrix with eigenvalues on the diagonal and \(\bfV\) the corresponding matrix of eigenvectors.
Assumption A3 would then be replaced with an assumption that $\kappa(\bfV)$ grows at most polynomially with increasing $k$.
\end{remark}
\subsection{Proof of Theorem \ref{thm:main2}}
\begin{lemma}[Bound on $\|A_k'\|_{{\LtG\rightarrow \LtG}}$]
\label{lem:normbounds}
If $\Omega_-$ is piecewise smooth (in the sense of, e.g., \cite[Definition 1.3]{GaSp:19}, then, given $k_0>0$, there exists $C_1$ (depending on $k_0$, $d$, and $\Omega$) such that
\begin{equation}\label{eq:upper}
\mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}} \leq C_1 k^{1/2}\log (k+2)
\quad\text{ for all } k\geq k_0.
\end{equation}
\end{lemma}
\begin{proof}[References for the proof]
This follows from the bounds in
\cite[Appendix A]{HaTa:15}, \cite{GaSm:15}, \cite[Chapter 4]{Ga:19} (see also \cite{GaSp:19}).
\end{proof}
We now prove Theorem \ref{thm:main2}.
First observe that, by the definitions of $N_{\rm eig}$ and $\ell$ and Assumption A2,
\begin{equation}\label{eq:proof1}
1\leq N_{\rm eig}\leq \ell +1 \leq {C_{{\rm Weyl}}} k^{d-1} +1.
\end{equation}
By Assumption A1, $S$ is independent of $k$; we then choose $L_0$ and $L_1$ to be also independent of $k$; i.e. $L_0, L_1\sim 1$.
By Assumption A3 and the definition \eqref{eq:eigenvaluecondition}, $1\leq \kappa^* \leq {C_{{\rm cond}}} k^M$.
Finally, by assumption, $1\leq n\leq C_{\rm dis} k^{M'}$.
Using in \eqref{eq:delta} all these inequalities, we find that
\begin{equation}\label{eq:proof2}
\frac{1}{{C_{{\rm Weyl}}} C_{\rm dis} {C_{{\rm cond}}}} k^{-M-M'-(d-1)} \lesssim \delta \lesssim 1.
\end{equation}
By the bound \eqref{eq:discon2} from Assumption A0 and then \eqref{eq:upper},
\begin{equation}\label{eq:proof3}
\mathbb{N}{\bfM^{-1}\bfA_k'}_{2} \leq {C_{{\rm approx},1}} \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}\lesssim {C_{{\rm approx},1}} k^{1/2}\log (k+2),
\end{equation}
where the omitted constant depends only on $\Omega_-$.
Using \eqref{eq:proof2} and \eqref{eq:proof3} in \eqref{eq:mlowerbound} (and recalling the asymptotics \eqref{eq:Wed1}), we obtain that
if $m$ satisfies
\begin{align}
m\geq {C_{{\rm Weyl}}} k^{d-1}+ C_1 \mathbb{N}{A_k'}_{{\LtG\rightarrow \LtG}}
\left(
\sum_{j=1}^\ell
\log\dfrac{1}{\lvert \lambda_j \rvert}
+
\log (\varepsilon^{-1})+ C\log k +
C_2 k^{d-1}\log k
\right)
\label{eq:proof4}
\end{align}
then the $m$th GMRES residual satisfies $\|\bfr_m \|_2/\|\bfr_0 \|_2 \leq \varepsilon$; in \eqref{eq:proof4}, $C_1$ depends on ${C_{{\rm approx},1}}$, $\Omega_-$ and $d$, $C_2$ depends on ${C_{{\rm Weyl}}}, {C_{{\rm approx},1}}$, $\Omega_-$, and $d$, and $C$ depends on ${C_{{\rm Weyl}}}$, ${C_{{\rm cond}}}$, $C_{\rm dis}$, $M$, $M'$, and $d$.
The bound on $m$ \eqref{eq:mlowerbound2} now follow from absorbing the $\log k$ term into the $k^{d-1}\log k$ term, and modifying the definition of $C_2$ appropriately, so that it now depends also on ${C_{{\rm cond}}}$, $C_{\rm dis}$, $M$, $M'$, and $d$.
\begin{remark}[The analogue of Theorem \ref{thm:main2} for $B_{k, {\rm reg}}$]\label{rem:Neumann}
If $B_{k, {\rm reg}}$ satisfies appropriate analogues of Assumptions A0-A3, then Theorem \ref{thm:main2} holds with $A_k'$ replaced by $B_{k, {\rm reg}}$ (and the Galerkin matrices modified accordingly). Indeed, the bound
\begin{equation}\label{eq:upperN}
\mathbb{N}{B_{k, {\rm reg}}}_{{\LtG\rightarrow \LtG}} \leq C_1 k^{1/4}\log (k+2)
\quad\text{ for all } k\geq k_0
\end{equation}
for piecewise-smooth $\Omega_-$ is proved in \cite{GaMaSp:21N} using results from \cite[Appendix A]{HaTa:15}, \cite{GaSm:15}, \cite[Chapter 4]{Ga:19} and results about semiclassical pseudodifferential operators (to bound $S_{{\rm i} k}$.
The proof of Theorem \ref{thm:main2} therefore goes through for $B_{k,{\rm reg}}$ with \eqref{eq:upperN} replacing \eqref{eq:upper}.
In \S\ref{sec:discussion}, we recalled that $\|(A_k')^{-1}\|_{{\LtG\rightarrow \LtG}}\lesssim \exp(\alpha k)$ when $\Omega_-$ is smooth, with this proved in
\cite[Equation 1.35 and Lemma 6.2]{ChSpGiSm:20} by combining bounds on the exterior Dirichlet problem from \cite[Theorem 1.2]{Bu:98} and bounds on the interior impedance problem in \cite[Theorem 1.8 and Corollary 1.9]{BaSpWu:16}.
The analogous bound for $B_{k, {\rm reg}}$ is proved in \cite{GaMaSp:21N} by expressing $(B_{k, {\rm reg}})^{-1}$ in terms of appropriate exterior and interior solution operators using \cite[Lemma 6.1, Equation 83]{BaSpWu:16},
and then using the bounds on the exterior Neumann problem from \cite{Bu:98}/\cite{Vo:00} and the relevant interior problem (an interior impedance-like problem involving $S_{{\rm i} k}$ in the boundary condition) from the combination of \cite[Theorem 4.6]{GaLaSp:21} and \cite[Lemma 3.2]{GaMaSp:21}.
\end{remark}
\section{Weyl asymptotics, Assumption A2, and why the large cavity has more near-zero eigenvalues than the small cavity}
\label{sec:Weyl_new}
Recall that Assumption A2 is that the number of near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ is $\lesssim k^{d-1}$, where the near-zero eigenvalues are defined as those in the set $k$-independent open set ${\cal N}$ in O1/A1.
In this section we use Weyl asymptotics to give heuristic arguments about why this assumption holds, and also why the large cavity has more near-zero eigenvalues than the small cavity. We begin by giving numerical evidence that Assumption A2 holds.
\subsection{Numerical evidence for Assumption A2}\label{sec:numA2}
Figure \ref{fig:counting_function} plots the the number of eigenvalues of $\bfM^{-1}\bfA_k'$ in the rectangle $[-0.1, 0.1]\times [-0.6,0.6]$ as a function of $k$ (plotted at $k=k_{m,0}^e$ and $k=k^o_{m,n}$) for both the small and large cavities in 2-d \footnote{Since the number of eigenvalues is discrete, the rectangle needs to be sufficiently large so that it includes enough eigenvalues for the Weyl asymptotics to hold.}.
The dashed lines are best-fit lines fitted using the nonlinear least-squares Marquardt-Levenberg algorithm;
we see the number of eigenvalues growing very close to linearly with $k$.
The same experiments for the Galerkin matrices (preconditioned with the mass matrix) of the operators $B_k$ and $B_{k, {\rm reg}}$ result in very similar plots; we do not show them here, but the exponents in the best fit lines are displayed in Table \ref{tab:kdependence}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{counting_function_dir.pdf}
\caption{The number of eigenvalues of $\bfM^{-1}\bfA_k'$ in the rectangle $[-0.1, 0.1]\times [-0.6,0.6]$ plotted against $k=k_{m,0}^e$ and $k=k^o_{m,n}$ for both the small and large cavities.}\label{fig:counting_function}
\end{figure}
The solid lines in Figure \ref{fig:counting_function} show how many Laplace eigenfunctions of the ellipse $E$ \eqref{eq:ellipse} are localised in the respective cavity; we explain in \S\ref{sec:smalllarge} below how we calculate this.
These localised eigenfunctions produce quasimodes (in the sense of Definition \ref{def:quasimodes}) with small quality.
Figure \ref{fig:counting_function} therefore gives strong evidence for
near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ in ${\cal N}$ correspond to localised eigenfunctions of the Laplacian in the ellipse.
\subsection{Recap of Weyl asymptotics for the number of eigenvalues of the Laplacian}\label{sec:Weyl_recap}
On a bounded domain $\Omega\subset \mathbb{R}^d$, the standard Weyl law states that, for either the Dirichlet or Neumann problem in $\Omega$,
\begin{equation}
\label{eq:WeylStandard}
N_{\Omega}(\Lambda)= \frac{\Lambda^{d/2}\omega_d\operatorname{Vol}(\Omega)}{(2\pi )^d}+O\big(\Lambda^{(d-1)/2}\big)
\quad\text{ as }\Lambda\rightarrow \infty;
\end{equation}
see e.g.~\cite[Theorem 1.2.1]{SaVa:92}, \cite{Se:78,Ph:81}.
In fact, if the periodic billiard trajectories on $\Omega$ with speed one
form a set of zero measure,
\cite{Iv:80} (see also~\cite[Corollary 29.1.6]{Ho:85a}~\cite[Theorem 1.6.1]{SaVa:92}) proved that
\begin{equation}\label{eq:Weylimproved}
N_{\Omega}(\Lambda)= \frac{\Lambda^{d/2}\omega_d\operatorname{Vol}(\Omega)}{(2\pi )^d}\pm \frac{\omega_{d-1}}{4(2\pi )^{d-1}}\Lambda^{(d-1)/2}\operatorname{Vol}(\partial\Omega)+o\big(\Lambda^{(d-1)/2}\big) \quad\text{ as }\Lambda\rightarrow \infty,
\end{equation}
where $\omega_d$ is the volume of the unit ball in $\mathbb{R}^d$, the plus sign is taken for the Dirichlet problem, and the minus sign for the Neumann problem.
\subsection{Two non-rigorous arguments about why we expect A2 to hold}
The first uses Observations O2(a)-(c) from \S\ref{sec:observations} along with results from \cite{St:99} linking the number of quasimodes to the number of resonances, and Weyl-type bounds on the number of resonances from \cite{PeZw:99}.
The second uses the Weyl asymptotics for the eigenfunctions of the Laplacian on $\Gamma$ (i.e.~the surface Laplacian) along with results from \cite[Chapter 4]{Ga:19} about the properties of the boundary-integral operators $D_k'$ and $S_k$ as semiclassical pseudodifferential operators.
\subsubsection{Why we expect A2 to hold when there exist quasimodes with $\epsilon(k)=O(k^{-\infty})$.}\label{sec:heuristic1}
We use the notation that
$a = O(k^{-\infty})$ as $k\rightarrow \infty$ if, given $N>0$, there exists $C_N$ and $k_0$ such that $|a|\leq C_N k^{-N}$ for all $k\geq k_0$, i.e.~$a$ decreases superalgebraically in $k$.
The steps in this argument are as follows:
\begin{enumerate}
\item We assume that there exists a one-to-one mapping between eigenvalues of $\bfM^{-1}\bfA_k'$ in ${\cal N}$ and quasimode frequencies $k_j$, such that an eigenvalue corresponding to frequency $k_{j^*}$ is closest to the origin when $k=k_{j^*}$.
\item By Point 1, O1/A1, and O2(c) (the eigenvalues move at $O(1)$ speed), the number of eigenvalues in ${\cal N}$ at $k$ equals the number of quasimode frequencies $k_j$ in
an interval $[k+p,k-p]$ with $p$ independent of $k$.
\item If the quality of the quasimode is $O(k^{-\infty})$, then the number of quasimode frequencies $k_j$ in
an interval $[k+p,k-p]$, with $p$ independent of $k$, is $\lesssim k^{d-1}$.
\item By Points 2 and 3, the number of eigenvalues in ${\cal N}$ at $k$ is $\lesssim k^{d-1}$.
\end{enumerate}
Regarding Point 2: Recall that ${\cal N}$ is independent of $k$ by O1/A1. By O2(c), the eigenvalues in ${\cal N}$ move at $k$-independent speed, and we assume that the paths of the eigenvalues are similar to those in Figure \ref{fig:outliers} and \ref{fig:flow_eig} (i.e.,~the eigenvalues don't move, e.g., in circles in ${\cal N}$).
Therefore, the number of eigenvalues in ${\cal N}$ at $k$ is equal to the number of eigenvalues that pass close to zero in an interval $[k+p,k-p]$ with $p$ independent of $k$. By Point 1, this number is equal to the number of $k_j$ in $[k+p,k-p]$.
Regarding Point 3:
If the quality of the quasimode is $O(k^{-\infty})$, then, by \cite[Theorem 2 and Corollary 2]{St:99}, the number of quasimodes is bounded by the number of resonances
in an $O(k^{-\infty})$ neighbourhood below the real axis.
Using the notation
\begin{equation}\label{eq:countingN}
N(K):= \big|\big\{ k_j \leq K\, :\, k_j \text{ is a quasimode frequency}\big\}\big|,
\end{equation}
where $|\cdot|$ denotes cardinality of a set, the bound
\begin{equation}\label{eq:counting2}
N(k+p) - N(k-p) \lesssim k^{d-1}
\end{equation}
follows since, by
\cite[Proposition 2]{PeZw:99}, the counting function of the number of resonances in an $O(1)$ neighbourhood of the real axis satisfies \eqref{eq:counting2}.
(Note that the assumption~\cite[Equation 1.9]{PeZw:99} about the Weyl asymptotics for the reference operator in the black-box framework holds by the results recapped in \S\ref{sec:Weyl_recap}).
\begin{remark}[When do there exist quasimodes with $\epsilon(k)=\mathcal{O}(k^{-\infty})$?]
There exist quasimodes with $\epsilon(k)=\mathcal{O}(k^{-\infty})$ if $\Omega_-$ satisfies the assumptions of Theorem \ref{thm:ellipse}, and also in the following two situations
by \cite[Theorem 1]{CaPo:02} and \cite[Theorem 1]{St:00} respectively.
(i) $\Gamma$ has zero Dirichlet boundary conditions and $\Omega_+$ contains an elliptic-trapped ray such that (a) $\Gamma$ is analytic in a neighbourhood of the ray and (b) the ray satisfies the stability condition \cite[(H1)]{CaPo:02}. In this situation, if $q >11/2$ when $d=2$ and $q>2d+1$ when $d\geq 3$, then there exists a family of quasimodes (in the sense of Definition \ref{def:quasimodes}) with
\begin{equation*}
\epsilon(k)= C_1 \exp( - C_2 k^{1/q})
\end{equation*}
for some $C_1, C_2>0$ and independent of $k$.
(ii) There exists a sequence of resonances $\{\lambda_\ell\}_{\ell=1}^\infty$ of the exterior Dirichlet/Neumann problem with
\begin{equation*}
0\leq -\Im \lambda_\ell = \mathcal{O}\big(|\lambda_\ell|^{-\infty}\big) \quad\text{ and } \quad \Re \lambda_\ell \rightarrow \infty \quad\text{ as }\quad \ell \rightarrow \infty
\end{equation*}
(recall that the resonances of the exterior Dirichlet/Neumann problem are the poles of the meromorphic continuation of the solution operator
from $\Im k\geq 0$ to $\Im k<0$; see, e.g., \cite[Theorem 4.4. and Definition 4.6]{DyZw:19}).
\end{remark}
\subsubsection{A second argument why we expect Assumption A2 to hold.}
We consider the case when $\Gamma\in C^\infty$ since a great deal of information is then available about the structure of $D_k'$ and $S_k$ (see~\cite[Chapter 4]{Ga:19}). In particular, these operators have the following two important features.
\begin{itemize}
\item[(i)] For any $\varepsilon>0$, there is $M>0$ such that if $v$ is a function with frequency $\geq Mk$, then
$$
\|D_k'v\|_{L^2} +k\|S_k v\|_{L^2}\leq \varepsilon \|v\|_{L^2}.
$$
\item[(ii)] $S_k$ and $D_k'$ almost map the Hilbert space of functions with frequency $\geq Mk$ to itself.
\end{itemize}
(When we say ``a function $v$ with frequency $\geq Mk$'' we mean that $v= \sum_{\lambda_j \geq M k} a_{\lambda_j}\phi_{\lambda_j}$ for some $a_{\lambda_j}\in\mathbb{C}$, where $(-\Delta_g -\lambda_j^2 )\phi_{\lambda_j} =0$ are the eigenfunctions of $-\Delta_g$ on $\Gamma$.)
If (ii) were exactly true (i.e., $S_k$ and $D_k'$ exactly preserve the space of functions with frequency larger than $Mk$) then we could decompose $L^2=F_{\leq Mk}\oplus F_{>Mk}$ where $F_{\leq MK}$ denotes the Hilbert space of functions with frequency $\leq Mk$, and $F_{> Mk}$ its orthogonal complement. In particular, since $F_{\leq Mk}$ and $F_{>Mk}$ would be invariant under the action of $A_k'$, the eigenvalues of $A_k'$ would then be the union of the eigenvalues of
\begin{equation*}
A'_{<}:=\Pi_{F_{\leq Mk}}A'_k\Pi_{F_{\leq Mk}}:F_{\leq Mk}\to F_{\leq Mk}
\end{equation*}
and
\begin{equation*}
A'_>:=\Pi_{F>Mk}A'_k\Pi_{F>Mk}:F_{>Mk}\to F_{>Mk}.
\end{equation*}
Then, by (i) all of the eigenvalues of $A'_{>}$ would be contained in $B(\frac{1}{2},\varepsilon)$ and hence eigenvalues outside this ball would correspond to eigenvalues of $A'_{<}$. By Weyl asymptotics, $\dim_{F_{\leq Mk}}\leq Ck^{d-1}$; therefore the number of eigenvalues of $A'_k$ outside the ball of radius $\varepsilon$ around $\frac{1}{2}$ would be $\lesssim k^{d-1}$.
We now make this heuristic argument slightly more precise. Let $-\Delta_g$ be the Laplacian on $\Gamma$, and $\chi\in C_c^\infty((-2,2);[0,1])$ with $\chi \equiv 1$ on $[-1,1]$. Then, writing
$$
R_k:=(D_k'-{\rm i} k S_k)(I-\chi(-k^{-2}\Delta_g)),
$$
we have that $R_k$ is semiclassical pseudodifferential operator of order $-1$. Furthermore, inspecting the semiclassical principle symbols of $D_k'$ and $S_k$, we see that there is $f$ such that $|\langle t\rangle f(t)|\leq C$ and
\begin{align*}
&\big\|\langle -k^{-2}\Delta_g\rangle^{1/2}(R_k-f(-k^{-2}\Delta_g)))\big\|_{{\LtG\rightarrow \LtG}}\\
&\qquad\qquad+\big\|(R_k-f(-k^{-2}\Delta_g)))\langle -k^{-2}\Delta_g\rangle^{1/2}\big\|_{{\LtG\rightarrow \LtG}}\leq Ck^{-1}.
\end{align*}
Furthermore, for any $\varepsilon>0$ there exists $M$ large enough such that
\begin{equation}
\label{e:oneAndTwo}
\begin{gathered}
\big\|(D_k'-{\rm i} k S_k)(I-1_{[-M^2,M^2]}(-k^{-2}\Delta_g))\big\|_{{\LtG\rightarrow \LtG}}< \varepsilon,\\
k\big\|\big[(D_k'-{\rm i} k S_k),1_{[-M^2,M^2]}(-k^{-2}\Delta_g)\big]\big\|_{{\LtG\rightarrow \LtG}}<\varepsilon,
\end{gathered}
\end{equation}
and then both (i) and (ii) above follow from \eqref{e:oneAndTwo}.
Define $F_{>Mk}$ to be the cokernel of $1_{[-M^2,M^2]}(-k^{-2}\Delta_g)$ in $L^2(\Gamma)$ and then $F_{\leq Mk}$ its orthogonal complement. Then let $\Pi=1_{[-M^2,M^2]}(-k^{-2}\Delta_g)$ be the orthogonal projector onto $F_{\leq Mk}$.
By \eqref{e:oneAndTwo},
$$
D_k'-{\rm i} k S_k=\widetilde{D}'_k-ik\widetilde{S}_k+O(k^{-1}\varepsilon)_{L^2\to L^2},\qquad \widetilde{D}_k'-ik\widetilde{S}_k:= (I-\Pi)(D_k'-{\rm i} k S_k)(I-\Pi)+\Pi (D_k'-{\rm i} k S_k)\Pi.
$$
We now argue with $D_k'-{\rm i} k S_k$ replaced by $\widetilde{D}_k-ik\widetilde{S}_k$ and choose $\varepsilon<\frac{1}{4}$ in~\eqref{e:oneAndTwo}. In this case, we can orthogonally decompose $L^2(\Gamma)$ into the subspaces $F_{\leq Mk}$ and $ F_{>Mk}$ which are invariant under application of $\tilde{A}_k':=\frac{1}{2}I+\widetilde{D}'_k-ik\widetilde{S}_k$. Then, since
$$
(\tfrac{1}{2}-z)I+(I-\Pi)(D_k-{\rm i} k S_k)(I-\Pi)
$$
is invertible for $|z-\frac{1}{2}|> \frac{1}{4}$, $\tilde{A}_k'$ has at most $\dim F_{\leq Mk}$ eigenvalues in $|z-\frac{1}{2}|>\frac{1}{4}$.
By the Weyl law on $\Gamma$ (which follows from \cite{Ho:68,Le:52,Av:56} since $\Gamma$ has no boundary),
$$
\dim F_{\leq MK}=\big|\{ \lambda_j\leq Mk\,:\, \lambda_j^2\text{ is an eigenvalue of }-\Delta_g\}\big|= \frac{\operatorname{Vol}_g(\Gamma) \omega_{d-1}}{(2\pi )^{d-1}}(Mk)^{d-1}+O(k^{d-2}),
$$
and, in particular, $\tilde{A}_k'$ has at most $Ck^{d-1}$ eigenvalues in $|z-\frac{1}{2}|>\frac{1}{4}$.
\begin{remark}
Although the difference is small, replacing $D_k'-{\rm i} k S_k$ by $\widetilde{D}_k'-{\rm i} k\widetilde{S}_k$ as we did in the arguments above is a \emph{serious} simplification and a more sophisticated argument would be needed to obtain a genuine bound on the number of eigenvalues away from $z=1/2$.
\end{remark}
\subsection{Why $\bfM^{-1}\bfA_k'$ has more near-zero eigenvalues for the large cavity than the small cavity}\label{sec:smalllarge}
\paragraph{How the pink lines in Figure \ref{fig:counting_function} were determined.}
Figure \ref{fig:counting_function} shows that the number of eigenvalues of $\bfM^{-1}\bfA_k'$ in ${\cal N}= [-0.1,0.1]\times[-0.6,0.6]$ grows with $k$ for both the small and large cavities, but the rate of growth is higher for the large cavity than the small cavity.
Recall that the pink lines in Figure \ref{fig:counting_function} show how many eigenfunctions of the Laplacian in the ellipse are localised in the respective cavity.
In \S\ref{sec:heuristic1} we assumed that all the near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ correspond to quasimode frequencies $k_j$, and we argued that
\begin{equation}\label{eq:Sunday1}
\big| \big\{ \lambda \text{ eigenvalue of } \bfM^{-1}\bfA_k' \, :\, \lambda \in {\cal N} \big\}\big|
= N(k+p)- N(k-p),
\end{equation}
for an appropriate $p$, where $N(k)$ is given by \eqref{eq:countingN}; i.e., $N(k)$ is the counting function of the quasimode frequencies.
We assume further that all the quasimode frequencies correspond to eigenvalues of Laplacian in the ellipse
$E$ \eqref{eq:ellipse} whose eigenfunctions localised about the minor axis, so that
\begin{equation}\label{eq:Sunday2}
N(k+p)- N(k-p)= N_{\rm loc}(k+p)- N_{\rm loc}(k-p),
\end{equation}
where $N_{\rm loc}$ is the counting function of these eigenvalues of the Laplacian.
We now use a microlocal version of the Weyl asymptotics \eqref{eq:Weylimproved} to determine the asymptotics of $N_{\rm loc}(k)$.
Assume that the ellipse is cut at $(x_{\rm cut}, y_{\rm cut})$, so that the small cavity corresponds to
$x_{\rm cut}=-\cos(3\pi/10)=\cos(7\pi/10)$, and the large cavity corresponds to $x_{\rm cut}= -\cos(\pi/10)=\cos(9\pi/10)$; see Figure \ref{fig:geometries}.
Let
\begin{equation}\label{eq:alphacut}
\alpha_{\rm cut}:= - \left( 1- \frac{x_{\rm cut}^2}{a_1^2}\right) = -\frac{y_{\rm cut}^2}{a_2^2}
\end{equation}
and let $a:= \sqrt{a_1^2-a_2^2}$ (as in Appendix \ref{sec:Mathieu}).
We show below that the asymptotics of $N_{\rm loc}$ for eigenfunctions localised in the cut ellipse is given by
\begin{equation}\label{eq:Nloc}
N_{\rm loc}(K) = \frac{V_{\rm loc}(\alpha_{\rm cut})}{(2\pi)^2} K^d + c_1 K^{d-1} + o\big(K^{d-1}\big) \quad\text{ as } K\rightarrow \infty,
\end{equation}
where
\begin{align}
V_{\rm loc}(\alpha):=8\int^{\pi/2}_{\arcsin(\sqrt{-\alpha})}\int_0^{\cosh^{-1}(\frac{a_1}{a})}
\Phi(\alpha,\omega,\theta)\, a^2(\sinh^2(\omega)+\sin^2(\theta)) {\rm d}\omega {\rm d}\theta,
\label{eq:Vloc}
\end{align}
where
\begin{equation}\label{eq:Phi}
\Phi(\alpha,\omega,\theta):=
\begin{cases}\arcsin\left(\sqrt{ \frac{1}{\sinh^2(\omega)+\sin^2(\theta)}\Big( \alpha+\sin^2\theta\Big)}\right)& \text{ if } \sin^2(\theta)\geq -\alpha\\0&\text{ otherwise.}
\end{cases}
\end{equation}
Calculating the integral in \eqref{eq:Vloc}, we find $V_{\rm loc}= 0.9895$ for the small cavity and $V_{\rm loc} = 3.0710$ for the large cavity.
Then, combining \eqref{eq:Sunday1} and \eqref{eq:Sunday2}, we find that
\begin{align}\nonumber
\big| \big\{ \lambda \text{ eigenvalue of } \bfM^{-1}\bfA_k' \, :\, \lambda \in {\cal N} \big\}\big|
&= N(k+p)- N(k-p)\\
&= N_{\rm loc}(k+p)- N_{\rm loc}(k-p) = 2 p d \frac{V_{\rm loc}}{(2\pi)^2} K^{d-1} + o\big(K^{d-1}\big).
\label{eq:blueline}
\end{align}
We now determine an appropriate value of $p$ when ${\cal N}= [-0.1,0.1]\times[-0.6,0.6]$ (since this is the ${\cal N}$ we chose in Figure \ref{fig:counting_function}).
Figure \ref{fig:outliers} indicates that the eigenvalues all move on roughly the same trajectory through ${\cal N}$.
Since the eigenvalues move with speed observed numerically to be approximately one, the appropriate $p$ is half the length of the portion of the curve that intersects ${\cal N}$.
With the imaginary part the $x$ variable and the real part the $y$ variable, we fit a polynomial of degree two in $x$ to
this portion of the curve, and find its length to be $0.90$, i.e., we take $p=0.45$.
The pink lines in Figure \ref{fig:counting_function} are then the linear function of $k$ on the right-hand side \eqref{eq:blueline} with $p=0.45$ and $d=2$.
As mentioned above, the fact the these pink lines match so well the number of eigenvalue of $\bfM^{-1}\bfA_k'$ in ${\cal N}$ give
strong evidence for the assumptions that (i) all near-zero eigenvalues of $\bfM^{-1}\bfA_k'$ correspond to quasimodes, and
(ii) the majority of quasimodes correspond to localised eigenfunctions of the Laplacian in the ellipse.
\paragraph{How we obtained \eqref{eq:Nloc} and \eqref{eq:Vloc}.}
In \S\ref{sec:Weyl_recap} we recapped the standard \eqref{eq:WeylStandard} and improved \eqref{eq:Weylimproved} Weyl asymptotics for the number of eigenvalues of the Laplacian on a bounded domain. Furthermore, there are microlocal versions of~\eqref{eq:Weylimproved} (see e.g., \cite[Theorems 1.8.5, 1.8.7]{SaVa:92}) that can be integrated to state, roughly, that the total $L^2$ mass of the normalised eigenfunctions with eigenvalue $\leq \Lambda$ in any subset, $U\Subset \{x\in \Omega,|\xi|\leq 1\}$ is given by
\begin{equation}\label{eq:WeylimprovedMicrolocal}
M_{\rm eig}(U,\Lambda) = \frac{\int 1_{U}(x,\xi){\rm d} x{\rm d}\xi}{(2\pi)^d} \Lambda^{d/2} +c_1(U)\Lambda^{(d-1)/2}+o\big(\Lambda^{(d-1)/2}\big) \quad\text{ as } \Lambda\rightarrow \infty,
\end{equation}
where $x$ is the position variable, and $\xi$ the momentum variable.
(For domains without boundary, these estimates can be recovered from
\cite{DuGu:75} and the full statement together with more quantitative versions can be found in~\cite[Theorem 6]{CaGa:20}.)
We now apply \eqref{eq:WeylimprovedMicrolocal} to the ellipse.
We could not find a proof that the periodic billiard trajectories with speed one on an ellipse
form a set of zero measure, under which the improved Weyl asymptotics \eqref{eq:Weylimproved}/\eqref{eq:WeylimprovedMicrolocal} hold. However, the results of \cite[\S4]{Co:10} indicate that the counting function of eigenvalues of the ellipse does indeed satisfy the improved Weyl asymptotics \eqref{eq:Weylimproved}/\eqref{eq:WeylimprovedMicrolocal}.
When $\Omega\subset \mathbb{R}^2$ is the ellipse, the Laplacian is quantum
completely integrable (see, e.g., \cite{GaTo:20} for the definition of quantum complete integrability); one consequence of this is the separation of variables used in Section~\ref{sec:Mathieu}. Moreover, this complete integrability implies the existence of a basis of eigenfunctions that concentrate along integrable tori. In particular, one expects that for $U$ a union of integrable tori,
\begin{equation}
\label{e:localisedWeyl}
M_{\rm eig}(U,\Lambda)\sim \big|\big\{\lambda_j^2\leq \Lambda\,:\, u_{\lambda_j}\text{ is localised inside }U\big\}\big|
\end{equation}
(given $N>0$, we say that $u_{\lambda_j}$ is localised inside $U$ if, for all symbols $a\in C_0^\infty(\overset{\circ}{U^c})$, $\|{\rm Op}_{\lambda_j^{-1}}(a) u_{\lambda_j} \|_{L^2} \leq \lambda_j^{-N}\|u\|_{L^2}$, where ${\rm Op}_{\lambda_j^{-1}}$ is semiclassical quantisation; see, e.g., \cite[Chapter 4]{Zw:12}, \cite[Page 543]{DyZw:19}).
In addition to the the fact that we could not find a reference for the improved Weyl law on the ellipse, this last step is also non-rigorous; a sophisticated analysis of the quantum completely integrable system would be required to justify~\eqref{e:localisedWeyl}.
The integrable tori on the ellipse correspond to billiard trajectories that remain tangent to some confocal conic; see, e.g., \cite[Page 8]{Ly:19}. The eigenfunctions that localise inside the elliptic cavity correspond precisely to eigenfunctions localised on integrable tori generated by confocal hyperbole that intersect the boundary of the ellipse where it has not been truncated; see Figure~\ref{fig:integrableTori}. In Appendix~\ref{sec:Weyl}, we compute the volume, $V_{\rm loc}$, in phase space of the integrable tori contained entirely inside the
small and large cavities and show that \eqref{eq:Vloc} holds.
The fact that these are \emph{all} the eigenfunctions localised inside the cavity (by \eqref{e:localisedWeyl}) then implies that \eqref{eq:Nloc} holds.
(We note that a similar phase-space volume calculation in an integrable setting occurs in \cite[Appendix B]{BaBe:07}.)
\begin{figure}
\centering
\begin{tikzpicture}
\def \a{1.3};
\def \b{.6};
\def \t{45};
\def 1{1}
\def 65{65}
\begin{scope}[scale=2 ]
\draw[thick,black] (0,0) ellipse (1 and .5);
\draw[thick,black] (0,0) ellipse ({\a} and {\b});
\fill[white] ({cos(126)},-.7)-- ({cos(126)},.7)-- (-\a-.1,.7)--(-\a-.1,-.7)--cycle;
\draw[thick,black] ({cos(126)},{.5*sqrt(1-cos(126)^2)})--({cos(126)},{\b*sqrt(1-cos(126)^2/(\a)^2)});
\draw[thick,black] ({cos(126)},-{.5*sqrt(1-cos(126)^2)})--({cos(126)},-{\b*sqrt(1-cos(126)^2/(\a)^2)});
\draw[scale=1, domain=-1:1, smooth, variable=\x, blue, dashed,thick] plot ({sqrt(.75)*cos(\t)*(pow(2,\x)+pow(2,-\x))/2)},{sqrt(.75)*sin(\t)*(pow(2,\x)-pow(2,-\x))/2} );
\draw[scale=1, domain=-1:1, smooth, variable=\x, blue, dashed,thick] plot (-{sqrt(.75)*cos(\t)*(pow(2,\x)+pow(2,-\x))/2)},{sqrt(.75)*sin(\t)*(pow(2,\x)-pow(2,-\x))/2} );
\draw[scale=1, domain=-1:1, smooth, variable=\x, red, dotted,thick] plot ({sqrt(.75)*cos(65)*(pow(2,\x)+pow(2,-\x))/2)},{sqrt(.75)*sin(65)*(pow(2,\x)-pow(2,-\x))/2} );
\draw[scale=1, domain=-1:1, smooth, variable=\x, red,dotted, thick] plot (-{sqrt(.75)*cos(65)*(pow(2,\x)+pow(2,-\x))/2)},{sqrt(.75)*sin(65)*(pow(2,\x)-pow(2,-\x))/2} );
\end{scope}
\end{tikzpicture}
\caption{The small cavity along with two confocal hyperbole. Eigenfunctions of the ellipse associated to trajectories tangent to the red (dotted) hyperbola localise inside the cavity and hence produce good quasimodes. Those associated to trajectories tangent to the blue (dashed) hyperbola do not localise in the cavity and hence do not produce good quasimodes.}\label{fig:integrableTori}
\end{figure}
\section{Numerical experiments.}
\label{sec:experiments}
\subsection{Description of the set-up used for the experiments}
The BIEs \eqref{eq:direct_combined_dir}, \eqref{eq:direct_combined_neu}, and \eqref{eq:direct_combined_neu2}, involving the operators
$A_k'$, $B_k$, and $B_{k, {\rm reg}}$ respectively, were discretised using the BEM with ten points per wavelength (as described in \S\ref{sec:statement:BEM}).
The eigenvalues and singular values of the resulting Galerkin matrices preconditioned with the mass matrix in 2-d were computed
using the library BemTool\footnote{\url{https://github.com/xclaeys/BemTool}} and LAPACK~\cite{Anderson1999}.
The largest matrices were around \(4,500\times 4,500\), and no distributed memory parallelisation was used.
The results about the Galerkin error, GMRES residual, and GMRES error were obtained using the libraries PETSc~\cite{petsc-efficient,petsc-user-ref}, BemTool, Htool\footnote{\url{https://github.com/htool-ddm/htool}}, and SuperLU\_DIST~\cite{lidemmel03} via the software FreeFEM~\cite{Hecht2012}. Note that no compression was used and the largest matrices were around \(4,500\times 4,500\) in 2-d, and \(450,000\times 450,000\) in 3-d.
In some of the figures we plot best-fit lines; these are computed with the nonlinear least-squares Marquardt-Levenberg algorithm (the basis of the `fit' command in gnuplot).
\subsection{Experiments about F3(a) and F3(b) in 2-d}
Figure \ref{fig:iterations_main} shows experiments about F3(a) ($k$-dependence of number of iterations) and F3(b) (dependence of number of iterations on the cavity size) for the Dirichlet BIE \eqref{eq:direct_combined_dir}, the Neumann BIE \eqref{eq:direct_combined_neu}, and the regularised Neumann BIE \eqref{eq:direct_combined_neu2}, through both integer frequencies and $k=k^e_{m,0}$. We saw a subset of these for the Dirichlet BIE in Figures \ref{fig:comparison_spectrum_it} and \ref{fig:comparison_size_cavit}.
The key point is that the growth of the number of iterations is the same through both sets of frequencies and for all three BIEs -- around $O(k^{0.8})$ for the large cavity and around $O(k^{0.6})$ for the small cavity -- even though
(i) $B_{k}$ has a very different distribution of eigenvalues to $A'_k$ and $B_{k, {\rm reg}}$ as shown in \S\ref{sec:Nevalue} below, and
(ii) the $k$-dependence of the norms of $B_k$ and $B_{k, {\rm reg}}$ is very different from that of $A_k'$ -- see \S\ref{sec:numbounds} below.
\begin{figure}[h!]
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{nb_it_dir.pdf}
\caption{Number of iterations for the Dirichlet BIE involving $A_k'$}
\label{fig:iterations_main_dir}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{nb_it_neu.pdf}
\caption{Number of iterations for the Neumann BIE involving $B_k$}
\label{fig:iterations_main_neu}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{nb_it_neu_reg.pdf}
\caption{Number of iterations for the regularised Neumann BIE involving $B_{k, {\rm reg}}$}
\label{fig:iterations_main_neu2}
\end{subfigure}
\caption{Number of iterations for the BIEs in 2-d with \(\theta=4\pi/10\) and $\Omega_-$ the small or large cavity}\label{fig:iterations_main}
\end{figure}
\subsection{Experiments about F3(a) in 3-d}\label{sec:3-dexpts}
Figure \ref{fig:3d} shows the number of GMRES iterations when the Dirichlet BIE \eqref{eq:direct_combined_dir} is solved with $\Omega_-$ the 3-d analogues of the small and large cavities (as described in \S\ref{sec:smalllargegeometries})
Similarly, Figure \ref{fig:3d_neu}
(plotted on the same scale as Figure \ref{fig:3d} for ease of comparison) shows the number of GMRES iterations when the Neumann BIEs \eqref{eq:direct_combined_neu} and
\eqref{eq:direct_combined_neu2}
are solved with $\Omega_-$ the 3-d analogue of the small cavity. For the Neumann BIEs we were unable to go up to as high a frequency as for the Dirichlet BIE because of memory issues.
The fact that in all cases the number of iterations grows roughly like $O(k^2)$, in contrast to roughly like $O(k)$ in 2-d, is consistent with the factors of $k^{d-1}$ appearing in Theorem \ref{thm:main2}.
These 3-d experiments only consider the number of GMRES iterations at integer frequencies. Our definitions of the 3-d small and large cavities are such that the 3-d analogue of the ellipse $E$ \eqref{eq:ellipse} is now a prolate spheroid.
Since the Laplacian is separable in prolate spheroidal coordinates (see, e.g., \cite[\S30.13]{Di:21}),
the eigenvalues of the Laplacian, and hence the corresponding quasimode frequencies (analogous to $k_{m,n}^{e/o}$), can be computed in a similar way to those in 2-d (as described in Appendix \ref{sec:Mathieu}), but we have not pursued this here.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{nb_it_3d_b.pdf}
\caption{Number of iterations for the Dirichlet BIE involving $A_k'$ with \(\theta=4\pi/10\) and $\Omega_-$ the small and large cavities in 3-d}\label{fig:3d}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{nb_it_3d_neu.pdf}
\caption{Number of iterations for the Neumann BIEs involving $B_k$ and $B_{k,{\rm reg}}$ with \(\theta=4\pi/10\) and $\Omega_-$ the small cavity in 3-d}\label{fig:3d_neu}
\end{figure}
\subsection{Plots of the eigenvalues of $B_{k}$ and $B_{k, {\rm reg}}$ in 2-d.}\label{sec:Nevalue}
Figure \ref{fig:eigenvalues_Neumann} plots the eigenvalues and singular values of the discretisations of $(2/{\rm i}) B_k$ (``Neumann'') and $(2/{\rm i})B_{k, {\rm reg}}$ (``regularised Neumann'') for $\Omega_-$ the small cavity at $k=k^e_{0,0}$;
the division by ${\rm i}$ rotates the eigenvalues so that they are in the same half-plane as the eigenvalues of $A_k'$. The accumulation point in the right-hand plot of Figure \ref{fig:eigenvalues_Neumann} is at $(1+{\rm i})/2$, which is consistent with $B_{k, {\rm reg}}$ being a compact perturbation of ${\rm i} (1+{\rm i})/4$ when $\Gamma$ is $C^1$ by \eqref{eq:Breg_compact_pert}.
These plots show the effect of regularising the hypersingular operator $H_k$.
Indeed, the discretisation of $B_k$ contains a vertical ``tail'' of eigenvalues that $\sim 1/h$ as $h\to 0$ for fixed $k$
\cite[Exercise 4.5.2]{SaSc:11}, \cite[Lemma 12.9]{St:08}. Since $H_k: H^1(\Gamma)\to {L^2(\Gamma}%{\partial {\cO}_+})}$ and ${L^2(\Gamma}%{\partial {\cO}_+})}\to H^{-1}(\Gamma)$, $H_k$ restricted to low frequencies maps to ${L^2(\Gamma}%{\partial {\cO}_+})}$,
and the eigenvalues in the tail are associated with high frequencies. Furthermore, the direction of the tail can be explained by the symbol of $H_k$ as a pseudodifferential operator on high frequencies; see \cite[\S4.3]{Ga:19}, \cite[\S4.1]{GaMaSp:21N}.
This tail of eigenvalues is not present for $B_{k, {\rm reg}}$ because $S_{{\rm i} k}H_k:{L^2(\Gamma}%{\partial {\cO}_+})}\to{L^2(\Gamma}%{\partial {\cO}_+})}$.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{spectra_illustrations_elliptic_cavity_2D_neu.pdf}
\caption{The eigenvalues and singular values for the discretisations of the Neumann BIEs involving $B_k$ (``Neumann'') and $B_{k, {\rm reg}}$ (``Regularised Neumann'') for the small cavity at \(k=k^e_{0,0}\).}\label{fig:eigenvalues_Neumann}
\end{figure}
\subsection{Experiments about quantities in the bound of Theorem \ref{thm:main1}.}\label{sec:numbounds}
Figures \ref{fig:Dirichlet}, \ref{fig:Neumann}, and \ref{fig:reg_Neumann} plot the following quantities for discretisations of each of the operators $A_k', B_k,$ and $B_{k, {\rm reg}}$ through \(k=k^e_{m,0}\).
\begin{itemize}
\item Top-left plot: the maximum singular value of $\bfM^{-1}\bfA$, i.e., $\|\bfM^{-1}\bfA\|_2$, where $\bfA$ is the respective Galerkin matrix,
\item Top-right plot: the minimum singular value of $\bfM^{-1}\bfA$, i.e., $\|(\bfM^{-1}\bfA)^{-1}\|_2$, and the eigenvalue of $\bfM^{-1}\bfA$ with the smallest modulus.
\item Bottom-left plot: the quantity
\begin{equation}\label{eq:keyquant}
\sum_{\lambda \in {\cal N}}\log \left(\frac{1}{|\lambda|}\right)
\end{equation}
where ${\cal N} = [-0.1,0.1]\times [-0.6,0.6]$ (as in Figures \ref{fig:outliers} and \ref{fig:counting_function}).
\item Bottom-right plot: $\log \big(\max_{\lambda}(\kappa(\lambda))\big)$, where $\kappa(\lambda)$ is the eigenvalue condition number defined by \eqref{eq:eigenvaluecondition}.
\end{itemize}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{sing_eig_values_elliptic_cavity_2D_dir_b.pdf}
\caption{Properties of the discretisation of the Dirichlet BIE involving $A_k'$ for \(k=k^e_{m,0}\)
and ${\cal N}= [-0.1,0.1]\times[-0.6,0.6]$.
}\label{fig:Dirichlet}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{sing_eig_values_elliptic_cavity_2D_neu_b.pdf}
\caption{Properties of the discretisation of the Neumann BIE involving $B_k$ for \(k=k^e_{m,0}\)
and ${\cal N}= [-0.1,0.1]\times[-0.6,0.6]$.
}\label{fig:Neumann}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{sing_eig_values_elliptic_cavity_2D_neu_reg_b.pdf}
\caption{Properties of the discretisation of the regularised Neumann BIE involving $B_{k, {\rm reg}}$ for \(k=k^e_{m,0}\)
and ${\cal N}= [-0.1,0.1]\times[-0.6,0.6]$.
}\label{fig:reg_Neumann}
\end{figure}
\paragraph{Regarding the top-left plots:} recalling that $\|\bfM^{-1}\bfA_k'\|_2$ approximates $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ (see Lemma \ref{lem:discon} and Assumption A0), we
used this information about $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ and $\|B_{k, {\rm reg}}\|_{{\LtG\rightarrow \LtG}}$ in our discussion in \S\ref{sec:discussion} about the sharpness of Theorem \ref{thm:main2}.
Figure \ref{fig:Dirichlet} shows $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ growing approximately like $k^{1/3}$.
How the geometry of $\Omega_-$ affects the $k$-dependence of $\|A_k'\|_{{\LtG\rightarrow \LtG}}$
is now well-understood thanks to the results of \cite{ChGrLaLi:09}, \cite[Appendix A]{HaTa:15}, \cite{GaSm:15}, \cite[Chapter 4]{Ga:19}, and \cite{GaSp:19}.
In fact, these results show that the $k$-dependence of $\|A_k'\|_{{\LtG\rightarrow \LtG}}$ as $k\rightarrow \infty$ is dominated by the $k$-dependence of $k\|S_k\|_{{\LtG\rightarrow \LtG}}$, and this $\sim k^{1/3}$ on curved parts of $\Gamma$ and $\sim k^{1/2}$ on flat parts, with the omitted constants dependent on the surface measure of these parts of the boundary. For both the small and large cavities, the surface measure of the flat parts of $\Omega_-$ is much smaller than the surface measure of the curved parts of $\Omega_-$ (see Figure \ref{fig:geometries}), and this is the reason why we only see the $k^{1/3}$ growth for the range $k\in (50,290)$ in Figure \ref{fig:Dirichlet}.
Similarly, Figure \ref{fig:reg_Neumann} shows $\|B_{k, {\rm reg}}\|_{{\LtG\rightarrow \LtG}}$ being essentially constant for the range of $k$ considered, although,
at least in 2-d, $\|B_{k, {\rm reg}}\|_{{\LtG\rightarrow \LtG}} \gtrsim k^{1/4}$ for large enough $k$; indeed, \cite[Theorem 4.6]{ChGrLaLi:09} shows that $\|D_k'\|_{{\LtG\rightarrow \LtG}}\gtrsim k^{1/4}$ for a certain class of 2-d domains (to see that the elliptic cavity falls in this class, take the points $x_1$ and $x_2$ in the statement of \cite[Theorem 4.6]{ChGrLaLi:09} to lie on one of the flat ends of the cavity, with $x_2$ in the middle of this end, and $x_1$ at one of the corners)
and \cite[Theorems 4.6 and 4.8]{GaMaSp:21N} show that $\|S_{{\rm i} k }H_k\|_{{\LtG\rightarrow \LtG}}\lesssim (\log k)^{3/2}$.
\paragraph{Regarding the top-right plots:} these show both
(i) the feature F2, i.e.~that while the norms of the inverses of the boundary-integral operators grow exponentially through $k^e_{m,0}$,
and thus the smallest singular values should decrease exponentially, this growth/decay stagnates, and
(ii) that the smallest eigenvalue modulus is very close the smallest singular value,
giving indirect evidence for Assumption A2, i.e., that at $k^e_{m,0}$ (for large enough $m$), the matrix has both a small singular value and a near-zero eigenvalue.
\paragraph{Regarding the bottom-left plots:}
these plots show the quantity \eqref{eq:keyquant} growing differently for the small and large cavities, and we used this information in our discussion in \S\ref{sec:discussion} about the sharpness of Theorem \ref{thm:main2}.
\paragraph{Regarding the bottom-right plots:} these verify Assumption A3, i.e., that the maximum eigenvalue condition number does not grow exponentially with $k$ (at least for the range of $k$ considered, i.e., $k\in (50, 290)$).
\section{Conclusions}\label{sec:conclusions}
In \S\ref{sec:features}, we stated that the main goals of this paper were to explain the feature F3(a) (i.e., why the number of GMRES iterations grows algebraically with $k$, with no worse growth through quasimode frequencies)
and, to a certain extent,
F3(b) (i.e., why the number of iterations depends on whether $\Omega_-$ is the small or large cavity).
Theorem \ref{thm:main2} addresses the $k$-dependence of the number of iterations (i.e., F3(a)). Although the bound \eqref{eq:mlowerbound2} in Theorem \ref{thm:main2}
does not directly distinguish between the small and large cavities, and hence does not explain F3(b), the coefficient of the highest-order terms in the bound \eqref{eq:mlowerbound2} depends on the number of the near-zero eigenvalues (via the constant ${C_{{\rm Weyl}}}$), and the arguments in \S\ref{sec:smalllarge} then explain heuristically the difference between this number for the small and large cavities.
\
For future investigations of GMRES applied to Helmholtz trapping scenarios, we have the following conclusions/messages.
\paragraph{The difference between $k= k_j$ and $k\neq k_j$ (where $k_j$ is a frequency in a quasimode) is not important.}
We saw from Figures \ref{fig:comparison_spectrum_it}
and \ref{fig:iterations_main} that the growth in the number of GMRES iterations did not depend on whether the frequency was in a quasimode (in contrast to the condition number, which does depend strongly on this).
\paragraph{The important quantities are (i) the rate of growth of the cluster, and (ii) the number of near-zero eigenvalues (governed by the number of quasimode frequencies).}
\
Regarding (i): the norm is a proxy for this, but comparing the experiments for $A_k'$ (the BIE \eqref{eq:direct_combined_dir} for the Dirichlet problem) in Figure \ref{fig:Dirichlet} and $B_{k, {\rm reg}}$ (the regularised BIE \eqref{eq:direct_combined_neu2} for the Neumann problem)
in Figure \ref{fig:reg_Neumann} we see one norm growing with $k$ (i.e.~$\|A_k'\|_{\LtG\rightarrow \LtG}$), the other norm remaining constant (i.e.~$\|B_{k, {\rm reg}}\|_{\LtG\rightarrow \LtG}$), but the
number of GMRES iterations for both growing at the same rate -- see Figure \ref{fig:iterations_main}.
Regarding (ii): the arguments in \S\ref{sec:Weyl_new} show that this number is governed by the Weyl law, and hence depends on dimension.
We highlight that, once the frequency is high enough, the density of these near-zero eigenvalues becomes too high for them to be considered as true ``outliers'' -- see Figure \ref{fig:example_spectrum} -- but the bounds of Theorems \ref{thm:main1} and \ref{thm:main2} still hold.
We advocate that these two quantities (i) and (ii) should play the role for Helmholtz trapping problems that the condition number plays in \emph{both} understanding the behaviour of the conjugate gradient method (CG) \emph{and} designing preconditioners for symmetric, positive-definite matrices. Indeed, if \(\bfB\) is symmetric positive-definite, it is well-known that
\begin{align}\label{eq:CG}
\dfrac{\lVert \bfr_m (\bfB,\bfb,\bfx_0) \rVert_{\bfB}}{\lVert \bfr_0 (\bfB,\bfb,\bfx_0) \rVert_{\bfB}} \leq2 \left(\sqrt{\dfrac{\kappa(\bfB)-1}{\kappa(\bfB)+1}}\right)^m,
\end{align}
where \(\lVert \cdot \rVert_{\bfB}\) denotes the norm induced by \(\bfB\).
In a similar way to the bounds on GMRES, the bound \eqref{eq:CG} can overestimate the number of iterations because it does not take into account either the right-hand side, or the fact that CG (like GMRES) can have a superlinear convergence, which depends on the particular distribution of eigenvalues.
Therefore, CG does not converge with the same speed for all matrices with the same maximum/minimum eigenvalues (and therefore for which the right-hand side of \eqref{eq:CG} is the same).
Despite these drawbacks, the bound \eqref{eq:CG} is useful in at least the two following ways:
(a) It indicates that preconditioners should be designed with the goal of decreasing the condition number and guarantees a reduction in the number of iterations if the resulting condition number is sufficiently small.
(b) If one can show that the condition number of $\bfB$ is independent of a certain parameter, it shows that the number of GMRES iterations to achieve a prescribed tolerance can be bounded independently of this parameter; this fact is used in, e.g., domain-decompositions methods where the number of iterations must be independent of the number of subdomains for the method to scale.
We advocate the use of the quantities (i) and (ii) above in a similar way. In particular the bounds of Theorems \ref{thm:main1} and \ref{thm:main2} show that a sufficient condition for a preconditioner to be robust at high frequency is for it to mitigate against the quantities (i) and (ii) growing with frequency.
|
\section{Introduction}
Quantum teleportation makes possible the deterministic transmission of unknown quantum states from one location to another~\cite{Bennett93}. It has been acknowledged as a fundamental scheme of state transfer. A shared quantum channel between a sender and a receiver is one of the essential ingredients for quantum teleportation, and quantum entanglement in the channel is necessary to ensure that the fidelity is superior to that of classical communication protocols. Moreover, quantum teleportation provides a useful framework to study quantum nonlocality~\cite{Popescu94,Gisin96,Barrett01,Cavalcanti13} and is one of the basic steps in constructing element gates, for example, single-qubit and CNOT gates, which are used in continuous-variable (CV) quantum computation~\cite{Jeong02,Ralph03}.
Fidelity $f$ is used to measure the closeness between the input and the teleported states. One of the methods to quantify the performance of teleportation based on $f$ without any dependence on the input states involves averaging $f$ over all the possible inputs. The average fidelity $F$, defined as a uniform average of $f$, has been widely used as a relevant input-independent measure of teleportation performance. It is well-known that $F$ can reach $1$ in quantum teleportation, whereas the maximally attainable $F$ is limited in any classical state-transfer scheme without quantum entanglement~\cite{Bennett93}. However, $F$ does not consider universality, which indicates whether teleportation is performed equally for all of the input states. For example, a non universal teleportation protocol would be successful only for a specific set of inputs. This limitation can hinder the use of a teleportation protocol to implement element gates in CV quantum computation~\cite{Jeong02,Ralph03}. To quantify the universality condition, a fidelity deviation $D$ defined in terms of the standard deviation of $f$ was introduced~\cite{unot:Bang12,Bang2018}.
In this study, we analyze the two measures $F$ and $D$ in the context of noisy qudit---a $d$-level quantum system---teleportation.
We show that perfect universality, $D=0$, is attainable without any dependence on the quantum channel condition, while the maximum average fidelity $F_\text{max}$ is a function of the degree of entanglement in the quantum channel. We prove that the condition of $F_\text{max}$ is adequate to guarantee perfect universality.
For the case of a qubit, that is, $d=2$, we demonstrate a more general and tighter relationship between $F$ and $D$ that addresses the aforementioned behaviors in a clearer manner. Then, we consider a realistic situation in which operational noises can deteriorate the teleportation performance, as represented by a decrease in $F$ and an increase in $D$. We propose a significant machine-learning-based method to alleviate such deterioration. We numerically demonstrate that the proposed machine-learning-based method is effective within a certain rate of noise occurrence. Note that the recovery of $F$ and $D$ can be implemented solely by using the process of $F \to F_\text{max}$. This is because $F_\text{max}$ itself suffices the condition of zero deviation. This feature allows us to reduce the control and/or learning time, which increases the recovery rate.
\section{Average fidelity and fidelity deviation}
We first review the two measures, namely, average fidelity $F$ and fidelity deviation $D$. Consider a map (or a general quantum operation) ${\cal T}: \ket{\phi} \rightarrow \hat{\rho}_\text{out}$, where $\ket{\phi}$ is an input and $\hat{\rho}_\text{out}$ an output. The fidelity $f$ is defined as~\cite{Jozsa94}
\begin{eqnarray}
f = \tr{\left[ \left(\sqrt{\hat{\rho}_\text{t}}\hat{\rho}_\text{out}\sqrt{\hat{\rho}_\text{t}}\right)^{1/2}\right]},
\label{eq:fidelity}
\end{eqnarray}
where $\hat{\rho}_\text{t}$ is a target state that is assumed to be a pure state $\ket{\phi_t}$. Then, formula~(\ref{eq:fidelity}) can be rewritten as
\begin{eqnarray}
f = \bra{\phi_t}\hat{\rho}_\text{out}\ket{\phi_t}.
\end{eqnarray}
In a well-designed ${\cal T}$, almost all of the input states $\ket{\phi}$ are correctly transferred to their corresponding targets $\ket{\phi_t}$. The average fidelity represents {\em how well} a state transfer is performed. It is defined by averaging $f$ over all possible inputs $\ket{\phi}$ as
\begin{eqnarray}
F=\int{\text{d}\phi}\,f,
\label{eq:avgf}
\end{eqnarray}
where $\text{d} \phi$ is the Haar measure that satisfies $\int \text{d} \phi=1$.
In general, $f$ varies with respect to the input states for a given channel ${\cal T}$. If $f$ is uniform for a task, the task is said to be universal. To devise a measure of universality, we employ the fidelity deviation $D$ in terms of the fluctuation of $f$~\cite{unot:Bang12,Bang2018}:
\begin{eqnarray}
D = \left(\int{\text{d}\phi}\, f^2 - F^2\right)^\frac{1}{2}.
\label{eq:stdd}
\end{eqnarray}
One can show perfect universality, that is, $D=0$ when $f=F$. Otherwise, it is strictly positive. Additionally,
\begin{eqnarray}
D^2 \leq\int{\text{d}\phi}\,{f} - F^2=F(1-F)\leq\frac{1}{4},
\label{eq:limit_ug}
\end{eqnarray}
where the last equality holds when $F=1/2$. Thus, $D$ is bounded as $0 \le D \le 1/2$.
Before going further, we note that $F=1$ holds iff $f=1$ for all possible inputs, and therefore, $F=1$ implies perfect universality. However, when the attainable maximum of $F$ is less than one ($F_\text{max} < 1$), perfect universality is not guaranteed at $F_\text{max}$ in general. For example, if $F_\text{max} < 1$, then $0 \leq D \leq \sqrt{F(1-F)}$ [see Eq.~(\ref{eq:limit_ug})]. This is true for almost all probabilistic tasks, for instance, universal-NOT \cite{Buzek99,Buzek00-1} and quantum cloning \cite{Buzek96}. Therefore, it is natural to consider the minimization of $D$ independently of the maximization of $F$.
\section{Noisy quantum teleportation}
In this section, we describe noisy quantum teleportation~\cite{Braunstein00}.
First, a sender, say Alice, has a $d$-dimensional pure state $\hat{\rho}_\phi = \ketbra{\phi}{\phi}$ that is to be transferred to a receiver, say Bob (in general, Alice has no information about the input state $\hat{\rho}_\phi$). A maximally entangled state $\ket{\Psi_0}=\frac{1}{\sqrt{d}}\sum_{j=0}^{d-1}\ket{j}\otimes\ket{j}$ is shared by Alice and Bob. Second, Alice performs a joint measurement by using a set of maximally entangled bases $\{\ket{\Psi_\alpha}\}$ ($\alpha=0,1,\ldots,d^2-1$) on the composite system of the unknown state and one of the entangled pairs. The entangled basis is obtained as
\begin{eqnarray}
\ket{\Psi_\alpha} = (\hat{U}_\alpha \otimes \hat{\openone}_d) \ket{\Psi_0},
\label{eq:A_m}
\end{eqnarray}
where $\hat{\openone}_d$ is the identity of the $d$-dimensional Hilbert space and $\hat{U}_\alpha$ is a unitary conditioned by completeness, that is, $\sum_{\alpha=0}^{d^2-1}\ket{\Psi_\alpha}\bra{\Psi_\alpha}=\hat{\openone}_{d^2}$~\cite{Braunstein00, Son01}. The outcome of Alice's joint measurement is communicated to Bob through a classical channel. Lastly, Bob applies a local operation $\hat{V}_\alpha$ based on the measurement outcomes received from Alice. Then, the state $\ket{\phi}$ is reproduced on Bob's side (see Fig.~\ref{fig:tele_scheme}).
Sharing the entanglement between Alice and Bob is the crucial step in ensuring that the better performance is superior to that of any classical protocol~\cite{Horodecki99}. However, under realistic conditions, quantum entanglement often becomes noisy~\cite{Carvalho04,Buric08}, and thus, we need to resolve the noisy case. Herein, we consider the following ($d \times d$)-dimensional noisy entanglement as the quantum channel~\cite{Werner89}:
\begin{eqnarray}
\hat{\rho}_{\Psi_0} = \gamma \ket{\Psi_0}\bra{\Psi_0} + \frac{1-\gamma}{d^2}\hat{\openone}_{d^2},
\label{eq:ws}
\end{eqnarray}
which is a statistical mixture of the maximally entangled state $\ket{\Psi_0}$ and white (symmetric) noise, and $\gamma$ is a fraction of $\ket{\Psi_0}$.
\begin{figure}[t]
\includegraphics[width=0.8\textwidth]{tele_scheme.eps}
\caption{Schematic of $d$-dimensional quantum state teleportation. The quantum channel, i.e., shared entangled state $\hat{\rho}_{\Psi_0}$, can be noisy [as in Eq.~(\ref{eq:ws})]. The errors can also arise in $\hat{U}_\alpha$ and $\hat{V}_\alpha$ due to the imperfections of control. It can be cured in our scheme (see Sect.~V for details).}.
\label{fig:tele_scheme}
\end{figure}
\section{Analysis of $F$ and $D$ in noisy quantum teleportation}
In this section, we discuss the relationship between the two measures, namely, the average fidelity $F$ and fidelity deviation $D$, according to the noise parameter $\gamma$ and dimension $d$.
\subsection{Qudit-teleportation}
By definition, the value of $F$ ranges between 0 and 1. However, for a given set of $\gamma$ and $d$, one can restrict the range by using the following relationship:
\begin{result}\label{res:fF}
For a given set of noise parameter $\gamma$ and dimension $d$, $F$ has a relation such that
\begin{eqnarray}
\frac{1}{d} - \frac{\gamma}{d(d+1)} \le F \le \gamma + \frac{(1-\gamma)}{d}.
\end{eqnarray}
\end{result}
This result is derived from the following extensive analysis of $F$. First, we write the transmitted state $\hat{\rho}_\text{out}$ as
\begin{eqnarray}
\hat{\rho}_\text{out} = \sum_{\alpha=0}^{d^2-1} \left(\hat{\openone}_{d^2} \otimes \hat{V}_\alpha\right) \bra{\Psi^\text{A}_\alpha} \hat{\rho}_\text{tot} \ket{\Psi^\text{A}_\alpha} \left(\hat{\openone}_{d^2} \otimes \hat{V}_\alpha^\dagger\right),
\label{eq:output_init}
\end{eqnarray}
where $\ket{\Psi_\alpha^\text{A}} = \ket{\Psi_\alpha}\otimes\hat{\openone}_d$ denotes Alice's joint measurement, $\hat{\openone}_{d^2} \otimes \hat{V}_\alpha$ represents Bob's local operation, and $\hat{\rho}_\text{tot} = \hat{\rho}_\phi \otimes \hat{\rho}_{\Psi_0}$ is the total initial state of $\hat{\rho}_\phi$ that is to be transferred, where $\hat{\rho}_{\Psi_0}$ is the shared entangled state. We can rewrite Eq.~(\ref{eq:output_init}) as
\begin{eqnarray}
\hat{\rho}_\text{out} = \frac{\gamma}{d^2}\sum_{\alpha=0}^{d^2-1}\hat{X}_\alpha \hat{\rho}_\phi \hat{X}_\alpha^\dagger + \frac{1-\gamma}{d}\hat{\openone}_d,
\end{eqnarray}
where $\hat{X}_\alpha = \hat{V}_\alpha \hat{U}_\alpha^\dagger$. The fidelity $f$ for a given $\hat{\rho}_\phi=\ket{\phi}\bra{\phi}$ is
\begin{eqnarray}
f = \frac{\gamma}{d^2}\sum_{\alpha=0}^{d^2-1}\xi_\alpha + \frac{1-\gamma}{d},
\label{eq:f_d}
\end{eqnarray}
where $\xi_\alpha$ is the fidelity between $\hat{\rho}_\phi$ and $\hat{X}_\alpha\hat{\rho}_\phi\hat{X}_\alpha^\dagger$, and it is expressed as
\begin{eqnarray}
\xi_\alpha = \tr{(\hat{X}_\alpha\hat{\rho}_\phi\hat{X}_\alpha^\dagger \hat{\rho}_\phi}) = \abs{\bra{\phi}\hat{X}_\alpha\ket{\phi}}^2.
\label{eq:df_xi}
\end{eqnarray}
By using the above descriptions, we evaluate $F$ and $D$. First, we consider the average fidelity $F$. From Eqs.~(\ref{eq:avgf}) and (\ref{eq:f_d}), we have
\begin{eqnarray}
F =\int \text{d}\phi \, f = \frac{\gamma}{d^2}\sum_{\alpha=0}^{d^2-1}\int \text{d} \phi \, \xi_\alpha + \frac{1-\gamma}{d}.
\label{eq:F_d}
\end{eqnarray}
The integral $\int \text{d} \phi \, \xi_\alpha$ can be calculated by using a lemma of identity, called Schur's lemma~\cite{Albeverio02,Braunstein00}:
\begin{eqnarray}
\int_G \text{d}g \, \left( \hat{U}_g^\dagger \otimes \hat{U}_g^\dagger \right) \hat{X} \left( \hat{U}_g \otimes \hat{U}_g \right) = a \hat{\openone}_{d^2} + b \hat{P},
\label{eq:sch1}
\end{eqnarray}
where
\begin{eqnarray}
a = \frac{d^2 \tr{(\hat{X})} - d \tr{(\hat{X}\hat{P})}}{d^2 (d^2 -1)}, \nonumber \\
b = \frac{d^2 \tr{(\hat{X}\hat{P})} - d \tr{(\hat{X})}}{d^2 (d^2 -1)},
\end{eqnarray}
for any operator $\hat{X}$ in the $d \times d$-dimensional Hilbert space. Here, $\text{d}g$ denotes the (normalized) Haar measure on the unitary group $G=U(d)$, satisfying $\int_G \text{d}g = 1$; $\hat{U}_g$ is an irreducible representation of $g \in G$; and $\hat{P}$ denotes the swap operator defined by $\hat{P}\ket{i j} = \ket{j i}$. By applying Schur's lemma, we can compute the integral $\int \text{d}\phi \, \xi_\alpha$ such that
\begin{eqnarray}
\int \text{d}\phi \, \xi_\alpha &=& \bra{\mathbf{0}\mathbf{0}} \int \text{d} \phi \, \left( \hat{U}_\phi^\dagger \otimes \hat{U}_\phi^\dagger \right) \left(\hat{X}_\alpha \otimes \hat{X}_\alpha^\dagger \right) \left( \hat{U}_\phi \otimes \hat{U}_\phi \right) \ket{\mathbf{0}\mathbf{0}} \nonumber \\
&=& \frac{1}{d(d+1)} \left(\abs{\tr{(\hat{X}_\alpha)}}^2 + d\right),
\label{eq:re_int1}
\end{eqnarray}
where $\ket{\phi}=\hat{U}_\phi\ket{\mathbf{0}}$. Then, we arrive at the final form of $F$ as
\begin{eqnarray}
F = F_\text{max} - \frac{\gamma}{d+1} \left( d - \frac{1}{d^3} \sum_{\alpha=0}^{d^2-1} \abs{\tr{(\hat{X}_\alpha)}}^2 \right),
\label{eq:fF_d}
\end{eqnarray}
where $F_\text{max} = \gamma + \frac{(1-\gamma)}{d}$, which is the maximum value of $F$ for a given $\gamma$.
We rewrite Eq.~(\ref{eq:fF_d}) in a more useful form as follows:
\begin{eqnarray}
F = \frac{d E + 1}{d+1},
\label{eq:FE}
\end{eqnarray}
where $E$ is defined as
\begin{eqnarray}
E = \frac{1}{d^2}\sum_{\alpha=0}^{d^2-1}\bra{\Psi_\alpha}\hat{\rho}_{\Psi_0}\ket{\Psi_\alpha} = \frac{\gamma}{d^4}\sum_{\alpha=0}^{d^2-1}\abs{\tr{(\hat{X}_\alpha)}}^2 + \frac{1-\gamma}{d^2}.
\label{eq:ef}
\end{eqnarray}
We find that $E$ is maximized when $\tr{(\hat{X}_\alpha)}=d$ (or equivalently, $\hat{X}_\alpha=\hat{\openone}_d$) for all $\alpha$ and minimized when $\tr{(\hat{X}_\alpha)}=0$ for all $\alpha$. Then, the value of $E$ is bounded by $\gamma$ such that
\begin{eqnarray}
\frac{(1-\gamma)}{d^2} \le E \le \gamma + \frac{(1-\gamma)}{d^2},
\end{eqnarray}
where the upper bound $\gamma + \frac{(1-\gamma)}{d^2}$ is called the (fullest) entanglement fraction of the channel~\cite{Albeverio02}. In this manner, by using Eq.~(\ref{eq:FE}) we can finally prove Result~\ref{res:fF}. Notably, this result is consistent with the results obtained in previous studies~\cite{Horodecki99,Albeverio02}.
Next, we consider the fidelity deviation $D$. Moreover, because the range of $F$ is limited, we derive a condition of $D$ for a given $\gamma$ and $d$ as follows.
\begin{result}
For a given set of noise parameter $\gamma$ and dimension $d$, the fidelity deviation $D$ has a condition such that
\begin{eqnarray}
0 \le D \le \gamma \,\overline{\Delta},
\end{eqnarray}
where $\overline{\Delta} = \frac{1}{d^2}\sum_{\alpha=0}^{d^2-1}\Delta_\alpha$, and
\begin{eqnarray}
\Delta_\alpha^2 = \int \normalfont \text{d} \phi \, \xi_\alpha^2 - \left(\int \text{d} \phi \, \xi_\alpha\right)^2,
\label{eq:delta_xi}
\end{eqnarray}
which can be regarded as the fidelity deviation of $\xi_\alpha$.
\end{result}
To prove this result, we express $D$ as follows by using Eqs.~(\ref{eq:f_d}) and (\ref{eq:F_d})
\begin{eqnarray}
D = \left( \int \text{d}\phi \, f^2 - F^2 \right)^\frac{1}{2} = \frac{\gamma}{d^2}\left(\sum_{\alpha,\beta=0}^{d^2-1} C_{\alpha\beta}\right)^{\frac{1}{2}},
\label{eq:D_d}
\end{eqnarray}
where $C_{\alpha\beta}$ are elements of covariance matrix $\mathbf{C}$ given by
\begin{eqnarray}
C_{\alpha\beta} = \int \text{d}\phi \, \xi_\alpha \xi_\beta - \int \text{d}\phi \, \xi_\alpha \int \text{d}\phi \, \xi_\beta.
\end{eqnarray}
Note that $\mathbf{C}$ is symmetric, that is, $C_{\alpha\beta} = C_{\beta\alpha}$, and its diagonal elements $C_{\alpha\alpha}$ are equal to $\Delta_\alpha^2$ in Eq. \eqref{eq:delta_xi}. Furthermore, each element of $\mathbf{C}$ is bounded as
\begin{eqnarray}
\abs{C_{\alpha\beta}} \le \sqrt{\Delta_\alpha^2 \Delta_\beta^2},
\label{eq:ineq_C}
\end{eqnarray}
which is known as the variance-covariance inequality~\cite{Bhatia00}. Then, by using Eq.~(\ref{eq:ineq_C}), we obtain Result 2. The perfect universality $D=0$ can be achieved when $\xi_\alpha$ is constant for all $\alpha$.
Based on the above results, we discuss the conditions of the sets $\{ \hat{U}_\alpha \}$ and $\{ \hat{V}_\alpha \}$ from the viewpoint of achieving the maximum average fidelity and perfect universality. First, we consider the maximization of $F$. According to Eq.~(\ref{eq:fF_d}), $F_\text{max}$ is achieved when
\begin{eqnarray}
\hat{X}_\alpha = \hat{\openone}_d,~\text{or equivalently,}~\hat{U}_\alpha = \hat{V}_\alpha,~(\forall\alpha).
\label{eq:condi_opt}
\end{eqnarray}
That implies that the maximization of $E$ straightforwardly leads to $F_\text{max}$. Therefore, our main result is as follows:
\begin{result}
The condition of Eq.~(\ref{eq:condi_opt}) naturally suffices perfect universality, that is,
\label{result:U}
\begin{eqnarray}
\hat{X}_\alpha = \hat{\openone}_d ~\rightarrow ~\text{the perfect universality}~D=0.
\end{eqnarray}
\end{result}
The proof is simple. By using Eq.~(\ref{eq:df_xi}), we rewrite condition~(\ref{eq:condi_opt}) as
\begin{eqnarray}
\xi_\alpha = 1~(\forall\alpha).
\end{eqnarray}
Then, $\mathbf{C}$ becomes a zero (or null) matrix because $C_{\alpha\beta}=0$ for $\xi_\alpha=const$ ($\forall\alpha$), thus leading to $D=0$, see Eq.~(\ref{eq:D_d}).
However, the opposite is not always true, that is, perfect universality does not guarantee the maximum fidelity. We emphasize that Result~\ref{result:U} is a non-common trait in limited maximum fidelity tasks, for instance, universal-NOT or cloning~\cite{unot:Bang12}.
\subsection{Qubit-teleportation}
In the case of a qubit, that is, $d=2$, we can derive a tighter relationship between $F$ and $D$. To this end, we first write the input state $\hat{\rho}_\phi$ in terms of the Bloch representation as
\begin{eqnarray}
\hat{\rho}_\phi = \frac{1}{2}\left(\hat{\openone}_2 + \boldsymbol{\phi}^\text{T} \boldsymbol{\sigma}\right),
\label{eq:Bphi_st}
\end{eqnarray}
where $\boldsymbol{\phi} = (\phi_x, \phi_y, \phi_z)^\text{T}$ is a Bloch vector of unit norm (i.e., $|\boldsymbol{\phi}|=1$) in three-dimensional real vector space $\mathbb{R}^3$, and $\boldsymbol{\sigma} = (\hat{\sigma}_x, \hat{\sigma}_y, \hat{\sigma}_z)^\text{T}$ is a vector operator whose components $\hat{\sigma}_j$ ($j=x,y,z$) are Pauli operators. Then, we can express $\xi_\alpha$ in Eq.~(\ref{eq:df_xi}) as
\begin{eqnarray}
\xi_\alpha = \frac{1}{2} \left( 1 + \boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi} \right),
\label{eq:B_xi}
\end{eqnarray}
where $\mathbf{R}_\alpha$ is a $3 \times 3$ rotation matrix in $\mathbb{R}^3$, whose elements $[\boldsymbol{R}_\alpha]_{jk}$ are given as
\begin{eqnarray}
[\boldsymbol{R}_\alpha]_{jk} =\frac{1}{2}\tr{(\hat{X}_\alpha \hat{\sigma}_j \hat{X}_\alpha^\dagger \hat{\sigma}_k)}~\text{for}~j,k = x,y,z.
\end{eqnarray}
The rotation angles $\vartheta_\alpha$ and axes $\mathbf{n}_\alpha$ of $\mathbf{R}_\alpha$ are found in the general expression of a single-qubit unitary as
\begin{eqnarray}
\hat{X}_\alpha = e^{-i\frac{\vartheta_\alpha}{2}\mathbf{n}_\alpha^\text{T} \boldsymbol{\sigma}} = \cos{\frac{\vartheta_\alpha}{2}}\,\hat{\openone}_2 - i \sin{\frac{\vartheta_\alpha}{2}}\mathbf{n}_\alpha^\text{T} \boldsymbol{\sigma}.
\end{eqnarray}
Then, by using Eqs.~(\ref{eq:F_d}) and (\ref{eq:B_xi}), the $F$ corresponding to $d=2$ can be expressed as
\begin{eqnarray}
F = \frac{\gamma}{4}\sum_{\alpha=0}^3 \frac{1}{2}\left[1 + \int d\boldsymbol{\phi} \left( \boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi} \right) \right] + \frac{1-\gamma}{2},
\label{eq:rwF_1q}
\end{eqnarray}
where $\text{d}\boldsymbol{\phi}$ is the Haar measure over the surface of the Bloch sphere, and it is normalized as $\int \text{d}\boldsymbol{\phi}=1$. Here, we again employ Schur's lemma to calculate the integral $\int \text{d}\boldsymbol{\phi}(\boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi})$:
\begin{eqnarray}
\int_G \text{d}g \, \mathbf{O}_g \mathbf{X} \mathbf{O}_g^\text{T} = \frac{1}{r}\tr{(\mathbf{X})}\,\mathbf{I}_{r},
\label{eq:Bschur}
\end{eqnarray}
where $\mathbf{I}_r$ is the identity matrix in $\mathbb{R}^r$, $\mathbf{O}_g$ is an irreducible orthogonal representation of an element $g \in G$, and $\text{d}g$ is the measure, normalized as $\int_g \text{d}g =1$. By applying this lemma to $O(3)$ of three-dimensional rotations, we can calculate the integral in Eq.~(\ref{eq:rwF_1q}) as
\begin{eqnarray}
\int \text{d}\boldsymbol{\phi} (\boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi}) = \frac{1}{3}\tr{(\mathbf{R}_\alpha)},
\label{eq:intB_xi}
\end{eqnarray}
and we immediately obtain the following form of $F$:
\begin{eqnarray}
F = \frac{1}{2} + \frac{\gamma}{24}\sum_{\alpha=0}^{3} \tr{(\mathbf{R}_\alpha)}.
\label{eq:fF_d2}
\end{eqnarray}
Next, we consider $D$. First, we express $D$ as
\begin{eqnarray}
D = \frac{\gamma}{4}\left( \sum_{\alpha=0} ^{3} \Delta_\alpha^2 + \sum_{\alpha \neq \beta} C_{\alpha\beta} \right)^\frac{1}{2}.
\label{eq:rwD_1q}
\end{eqnarray}
Subsequently, we (re)calculate the inequality of Eq.~(\ref{eq:ineq_C}) for $d=2$, and a tighter lower bound can be found as follows (for more details, see appendix~B in Ref.~\cite{unot:Bang12}):
\begin{eqnarray}
-\frac{1}{2}\Delta_\alpha \Delta_\beta \le C_{\alpha\beta}\le \Delta_\alpha \Delta_\beta ~(\alpha \neq \beta),
\label{eq:ineq_Cd2}
\end{eqnarray}
where we obtain the lower bound when the two rotation axes $\mathbf{n}_\alpha$ and $\mathbf{n}_\beta$ are orthogonal to each other, that is, $\mathbf{n}_\alpha^\text{T}\mathbf{n}_\beta = 0$, and the upper bound when $\mathbf{n}_\alpha$ and $\mathbf{n}_\beta$ are parallel or antiparallel, that is, $\mathbf{n}_\alpha^\text{T}\mathbf{n}_\beta = \pm 1$. Then, the upper bound of $D$ is [from Eq.~(\ref{eq:ineq_Cd2})]
\begin{eqnarray}
D \le \frac{\gamma}{4} \sqrt{\sum_{\alpha=0} ^{3} \Delta_\alpha^2 + \sum_{\alpha \neq \beta} \Delta_\alpha \Delta_\beta} = \gamma\overline{\Delta}_{d=2},
\label{eq:rw2D_1q}
\end{eqnarray}
where $\overline{\Delta}_{d=2} = \frac{1}{4}\sum_{\alpha=0}^{3}\Delta_\alpha$. Here, $\Delta_\alpha$ are given in the Bloch form, such that
\begin{eqnarray}
\Delta_\alpha = \frac{1}{2} \left[ \int \text{d} \boldsymbol{\phi}\, \left(\boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi}\right)^2 - \frac{1}{9}\tr{(\mathbf{R}_\alpha)}^2 \right]^\frac{1}{2}.
\label{eq:wrB_delta_xi}
\end{eqnarray}
Now, we calculate $\int \text{d} \boldsymbol{\phi}\, (\boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi})^2$ by using the product form of Schur's lemma:
\begin{eqnarray}
\int \text{d}g \left(\mathbf{O}_g^\text{T}\otimes\mathbf{O}_g^\text{T}\right)\mathbf{X}\left(\mathbf{O}_g\otimes\mathbf{O}_g\right) = a \mathbf{I}_{d^2} + b \mathbf{D} + c \mathbf{P},
\label{eq:Bschur2}
\end{eqnarray}
where
\begin{eqnarray}
a &=& \frac{(r+1)\tr{(\mathbf{X})} - \tr{(\mathbf{X}\mathbf{D})} - \tr{(\mathbf{X}\mathbf{P})}}{r(r-1)(r+2)}, \nonumber \\
b &=& \frac{-\tr{(\mathbf{X})} + (r+1)\tr{(\mathbf{X}\mathbf{D})} - \tr{(\mathbf{X}\mathbf{P})}}{r(r-1)(r+2)}, \nonumber \\
c &=& \frac{-\tr{(\mathbf{X})} - \tr{(\mathbf{X}\mathbf{D})} + (r+1)\tr{(\mathbf{X}\mathbf{P})}}{r(r-1)(r+2)}. \nonumber
\end{eqnarray}
Here, $\mathbf{P}$ is a swap matrix $\mathbf{P}\,(\mathbf{x}_i \otimes \mathbf{x}_j) = \mathbf{x}_j \otimes \mathbf{x}_i$, or equivalently, $\mathbf{P}=\sum_{i,j=0}^{r-1}\left(\mathbf{x}_j\otimes\mathbf{x}_i \right) \left(\mathbf{x}_i\otimes\mathbf{x}_j \right)^\text{T}$, and $\mathbf{D}=\left(\sum_{i=0}^{r-1} \mathbf{x}_i\otimes\mathbf{x}_i \right) \left( \sum_{j=0}^{r-1}\mathbf{x}_j\otimes\mathbf{x}_j \right)^\text{T}$
where $\{\mathbf{x}_i\}$ is an orthonormal basis set in $\mathbb{R}^r$. Then, by using this lemma, we can rewrite $\int \text{d} \boldsymbol{\phi}\, (\boldsymbol{\phi}^\text{T}\mathbf{R}_\alpha\boldsymbol{\phi})^2$ as
\begin{eqnarray}
\mathbf{x}_{00}^\text{T} \int \text{d} \boldsymbol{\phi} \Big(\mathbf{O}_{\boldsymbol{\phi}}^\text{T} \otimes \mathbf{O}_{\boldsymbol{\phi}}^\text{T} \Big) \left(\mathbf{R}_\alpha \otimes \mathbf{R}_\alpha\right) \Big(\mathbf{O}_{\boldsymbol{\phi}} \otimes \mathbf{O}_{\boldsymbol{\phi}} \Big) \mathbf{x}_{00},
\label{eq:stdD_1q-int}
\end{eqnarray}
where $\mathbf{x}_{00}=\mathbf{x}_0 \otimes \mathbf{x}_0$ and $\boldsymbol{\phi} = \mathbf{O}_{\boldsymbol{\phi}}\mathbf{x}_0$. Eq.~(\ref{eq:wrB_delta_xi}) is then calculated and we have
\begin{eqnarray}
\Delta_\alpha = \frac{1}{2\sqrt{5}}\left(1-\frac{1}{3}\tr{(\mathbf{R}_\alpha)}\right),
\end{eqnarray}
where we used the following properties:
\begin{eqnarray}
\tr{(\mathbf{R}_\alpha\otimes\mathbf{R}_\alpha)} &=& \tr{(\mathbf{R}_\alpha)}^2, \nonumber \\
\tr{(\mathbf{R}_\alpha\otimes\mathbf{R}_\alpha\,\mathbf{D})} &=& \tr{(\mathbf{R}_\alpha\mathbf{R}_\alpha^\text{T})}=\tr{(\mathbf{I}_3)}=3, \nonumber \\
\tr{(\mathbf{R}_\alpha\otimes\mathbf{R}_\alpha\,\mathbf{P})} &=& \tr{(\mathbf{R}_\alpha^2)}.
\end{eqnarray}
Consequently, we can derive a tighter relationship between $F$ and $D$ as follows:
\begin{eqnarray}
D \le \frac{1}{\sqrt{5}} \left( F_\text{max} - F \right),
\label{eq:upb_d2}
\end{eqnarray}
where $F_\text{max} = \frac{1 + \gamma}{2}$. This confirms Result~\ref{result:U}, that is, $D=0$ iff $F=F_\text{max}$. Here, we can prove that the maximum value of $D$, that is, the worst case of universality, implies the minimum average fidelity, $F_\text{min}=\frac{1}{2} - \frac{\gamma}{6}$.
\section{Machine-learning-based stabilization of control in teleportation}
Result~\ref{result:U} implies the following: [{\bf T.1}] {\em The maximization of $F$ naturally includes the minimization of $D$ in noisy teleportation.} This is indeed a structural trait of teleportation. We conjecture that quantum teleportation degraded by operational noises can be effectively cured and further provide a benefit of utilizing the aforementioned trait [{\bf T.1}] can be achieved. To investigate this, we assume the operational noises in the control of Alice's joint measurements (i.e., $\{\hat{U}_\alpha\}$) and Bob's local operations (i.e., $\{\hat{V}_\alpha\}$), which deteriorate $F$ and $D$. We propose a machine-learning-based algorithm to stabilize the teleportation system against noise. For qubit teleportation (i.e., $d=2$), we demonstrate by means of numerical simulations that our machine-learning-based algorithm can cure the deterioration of both $F$ and $D$ solely by maximizing $F$ (see Fig.~\ref{fig:tele_scheme}).
\subsection{Effects of operational noise}
In general, a unitary operation in the $d$-dimensional Hilbert space can be parameterized as follows:
\begin{eqnarray}
\label{eq:unit_op}
\hat{U}(\mathbf{p}) = \exp{(-i\,\mathbf{p}^\text{T}\mathbf{G})},
\end{eqnarray}
where $\mathbf{p}=(p_1,p_2,\ldots,p_{d^2-1})^\text{T}$ is a $(d^2-1)$-dimensional real vector and $\mathbf{G}=(\hat{g}_1, \hat{g}_2, \ldots,\hat{g}_{d^2-1})^\text{T}$ is a vector operator whose components are SU($d$) group generators $\hat{g}_j$ ($j=1,2,\ldots d^2-1$)~\cite{Hioe81,Son04,Bang08}. Here, the components of $\mathbf{p}$ correspond to a set of control parameters in a real experiment, such as multiport beam splitters and phase shifters in linear optical systems~\cite{Reck94} or the pulse sequences of solid system qubits~\cite{Lee00}. The operational noises of $\{\hat{U}_\alpha\}$ and $\{\hat{V}_\alpha\}$ can be formulated as follows: the control parameters of $\hat{U}_\alpha$ and $\hat{V}_\alpha$ fluctuates such that
\begin{eqnarray}
\label{eq:err}
\mathbf{p} \rightarrow \mathbf{p} +\eta\,\boldsymbol\epsilon,
\end{eqnarray}
where $\boldsymbol\epsilon = (\epsilon_1, \epsilon_2, \ldots, \epsilon_{d^2-1})^\text{T}$ is a vector of the stochastic errors $\epsilon_j \in [-\pi, \pi]$. The factor $\eta \in [0,1]$ is casted to represent the degree of immaturity in control.
We investigate the effects of operational noises on qubit teleportation. First, we assume that the protocol is already set with the optimal condition as in Eq.~(\ref{eq:condi_opt}), and in this case, operational noise occurs continuously in both $\{\hat{U}_\alpha\}$ and $\{\hat{V}_\alpha\}$. To understand the effects of operational noise, we perform Monte-Carlo simulations with increasing $\eta$. The results are shown in Fig.~\ref{fig:FD_operr}, where each data point is created by averaging $10^4$ simulations. $F$ decreases from $1$ to $F_\text{err}$, and $D$ increases from $0$ to $D_\text{err}$. The fullest deterioration of $F$, that is, $F_\text{err} \simeq 0.5102$ is approximately $98\%$ to $F_\text{rand}=\frac{1}{2}$, whereas $D_\text{err}$ increases to $\simeq 0.0431$, or by approximately $39\%$, to $\frac{1}{2}D_\text{err}^\text{max} \simeq \frac{1}{4\sqrt{5}}$ even in the worst case of $F_\text{err} \simeq \frac{1}{2}$. This implies that the deterioration of $F$ is more conspicuous than that of $D$. Here, $F_\text{rand}$ is the average fidelity of the purely random protocol and $\frac{1}{2}D_\text{err}^\text{max}$ represents half (mean) of the maximum fidelity deviation at $F_\text{err}$, that is, $\frac{1}{2}D_\text{err}^\text{max} = \frac{1}{2\sqrt{5}}(F_\text{max}-F_\text{err})$ (see the dashed line in Fig.~\ref{fig:FD_operr}).
\begin{figure}[t]
\includegraphics[angle=270,width=0.37\textwidth]{avgF_operr.eps}
\includegraphics[angle=270,width=0.37\textwidth]{stdD_operr.eps}
\caption{Deteriorations of $F$ (left) and $D$ (right) with respect to $\eta$. Each data point is created by averaging $10^4$ simulations. The error bar denotes the standard deviation. The teleportation system breaks down due to operational noise, as indicated by $F=1 \rightarrow F_\text{err} < 1$ and $D=0 \rightarrow D_\text{err} > 0$. The dashed lines are created based on simulations of the random protocol.}
\label{fig:FD_operr}
\end{figure}
\subsection{Machine-learning-based algorithm for stabilization of control}
To cure the unstable control, we propose a machine-learning-based algorithm built on the so-called differential evolution concept~\cite{unot:Bang12}, where the control parameters of $\{\hat{U}_\alpha\}$ and $\{\hat{V}_\alpha\}$ are allowed to evolve during the process. The algorithm runs as follows: First, $N_\text{pop}$ sets of the control parameter vectors are prepared as the candidate solutions $\{\mathbf{p}^{(u)}_{\alpha,i}, \mathbf{p}^{(v)}_{\alpha,i}\}$, where $\mathbf{p}^{(u)}_{\alpha,i}$ and $\mathbf{p}^{(v)}_{\alpha,i}$ are respectively the control parameter vectors of the candidate operations $\hat{U}_{\alpha,i}$ and $\hat{V}_{\alpha,i}$ ($i=1,2,\ldots,N_\text{pop}$). Thus, we have $2 d^2 N_\text{pop}$ parameter vectors. Then, the prepared sets of the candidate solutions are allowed to evolve through the following steps: (1) We generate $2N_\text{pop}$ mutant vectors $\boldsymbol\nu^{(u,v)}_{\alpha,i}$ for $\hat{U}_{\alpha,i}$ and $\hat{V}_{\alpha,i}$ according to
\begin{eqnarray}
\boldsymbol\nu^{(u,v)}_{\alpha,i} = \mathbf{p}^{(u,v)}_{\alpha,a} + W \left(\mathbf{p}^{(u,v)}_{\alpha,b} - \mathbf{p}^{(u,v)}_{\alpha,c}\right),
\end{eqnarray}
where $\mathbf{p}^{(u,v)}_{\alpha,a}$, $\mathbf{p}^{(u,v)}_{\alpha,b}$, and $\mathbf{p}^{(u,v)}_{\alpha,c}$ are randomly selected for $a,b,c \in \{1,2,\ldots,N_\text{pop}\}$. These vectors are selected to be different from each other. The free parameter $W$, also called a differential weight, is a real and constant number. (2) Thereafter, all $2 d^2 N_\text{pop}$ parameter vectors,
\begin{eqnarray}
\mathbf{p}^{(u)}_{\alpha,i} =
\begin{pmatrix}
p^{(u)}_{1} \\
p^{(u)}_{2} \\
\vdots \\
p^{(u)}_{d^2-1},
\end{pmatrix}_{\alpha,i},~
\mathbf{p}^{(v)}_{\alpha,i} =
\begin{pmatrix}
p^{(v)}_{1} \\
p^{(v)}_{2} \\
\vdots \\
p^{(v)}_{d^2-1},
\end{pmatrix}_{\alpha,i}
\end{eqnarray}
are reformed to trial vectors,
\begin{eqnarray}
\boldsymbol{\tau}^{(u)}_{\alpha,i} =
\begin{pmatrix}
\tau^{(u)}_{1} \\
\tau^{(u)}_{2} \\
\vdots \\
\tau^{(u)}_{d^2-1},
\end{pmatrix}_{\alpha,i},~
\boldsymbol{\tau}^{(v)}_{\alpha,i} =
\begin{pmatrix}
\tau^{(v)}_{1} \\
\tau^{(v)}_{2} \\
\vdots \\
\tau^{(v)}_{d^2-1},
\end{pmatrix}_{\alpha,i}
\end{eqnarray}
by the rule: For each $j = 1,2,\ldots,d^2-1$,
\begin{eqnarray}
\label{eq:crossover}
\left\{
\begin{array}{ll}
\tau^{(u,v)}_{j} \leftarrow p^{(u,v)}_{j} & ~~\text{if}~R_j > C_r,\\
\tau^{(u,v)}_{j} \leftarrow \nu^{(u,v)}_{j} & ~~\text{otherwise}, \\
\end{array}
\right.
\end{eqnarray}
where $R_j \in [0, 1]$ is a randomly generated number, and the crossover rate $C_r$ is another free parameter ranging between $0$ and $1$. Note that these free parameters $W$ and $C_r$ are set to achieve the best learning efficiency. (3) Finally, the control parameter vectors are evaluated by using the fitness criteria, that is, how well do the given parameters fit to the protocol. More specifically, $\{\boldsymbol\tau^{(u)}_{\alpha,i}, \boldsymbol\tau^{(v)}_{\alpha,i}\}$ are taken if they yield a higher level of fitness; if not, $\{\mathbf{p}^{(u)}_{\alpha,i}, \mathbf{p}^{(v)}_{\alpha,i}\}$ are retained. In our algorithm, we extract the best fitness among $N_\text{pop}$ and retain the corresponding parameters $\{\mathbf{p}_{\alpha,\text{best}}, \mathbf{p}_{\alpha,\text{best}}\}$. Steps (1)-(3) are then repeated.
\begin{figure}[t]
\includegraphics[angle=270,width=0.37\textwidth]{avgF_find.eps}
\includegraphics[angle=270,width=0.37\textwidth]{stdD_find.eps}
\caption{Remediation of fully broken teleportation. The graphs of $F$ (left) and $D$ (right) are plotted based on the results of the numerical simulations. Here, we consider three cases: $\gamma=\gamma_\text{C}=\frac{1}{3}$, $\gamma=\gamma_\text{BV}=\frac{1}{\sqrt{2}}$, and $\gamma=1$. Each data point is created based on $10^4$ repeating simulations.}
\label{fig:de_find}
\end{figure}
We investigate numerically whether even a fully broken teleportation system can be cured using the proposed machine-learning-based algorithm. The numerical simulation is performed for $d=2$. Here, we take $N_\text{pop}=100$, and the free parameters of our algorithm are selected such that: $W=0.5$ and $C_r = 0.1$. To use the structural trait [{\bf T.1}] of teleportation, we define fitness in terms of $F$; in other words, there is no minimization of $D$. Note that in general, fitness should be defined as a function of $F$ and $D$ (for example, see Ref.~\cite{unot:Bang12}). Such a setting is indeed beneficial, as described later. In Fig.~\ref{fig:de_find}, we present the results in the form of graphs of $F$ and $D$ for three cases: $\gamma=\gamma_\text{C}=\frac{1}{3}$, $\gamma=\gamma_\text{BV}=\frac{1}{\sqrt{2}}$, and $\gamma=1$. Here, $\gamma_\text{C}$ denotes the condition of the separability of the channel and $\gamma_\text{BV}$ is the critical value that the entanglement of the channel allows the violation of CHSH inequality. Each data point is created based on $10^4$ repeated simulations. The results indicates that the broken teleportation system can be recovered; $F$ approaches $F_\text{max}$, and $D$ decreases to zero. Specifically, we obtain ($F \simeq 0.6640$, $D \simeq 0.0002$) for $\gamma=\gamma_\text{C}$, ($F \simeq 0.8478$, $D \simeq 0.0005$) for $\gamma=\gamma_\text{BV}$, and ($F \simeq 0.9915$, $D \simeq 0.0009$) for $\gamma=1$.
\begin{figure}[t]
\includegraphics[angle=270,width=0.37\textwidth]{avgF_recover.eps}
\includegraphics[angle=270,width=0.37\textwidth]{stdD_recover.eps}
\caption{Real-time remediation of control fluctuations. The results of the numerical simulations are given as graphs of $F$ (left) and $D$ (right). Each data point is created based on $10^4$ simulations. It is assumed that fluctuation occurs after every $10$ or $50$ iterations of our algorithm. For more convincing analysis, we consider the worst case, namely, $\eta =1$. It is observed that the teleportation can be cured continuously.}
\label{fig:comp_err}
\end{figure}
We further investigate through numerical simulations whether the system can be stabilized, by assuming that the fluctuation occurs abruptly after intervals of some iterations of our algorithm. Note that such a model is realistic~\cite{Viola99, Khaneja01}. Here, we consider the scenarios in which the controls fluctuate after every $10$ and $50$ iterations. For more faithful and confident analysis, we consider the scenario with the worst deterioration, that is, $\eta=1$. Fitness is defined solely by $F$. In Fig.~\ref{fig:comp_err}, we present our simulation results. $F$ and $D$ deteriorate to their fullest extents, but the system is cured continuously.
\section{Summary and remarks}
We analyzed the average fidelity and fidelity deviation for noisy teleportation. We proved that teleportation can be zero fidelity deviations (or equivalently, the perfect universality) independently of the quantum channel condition, while the achievable maximum average fidelity is limited by the fraction of entanglement in the channel. Based on these analyses, we derived Result~\ref{result:U}: the maximum average fidelity ensures perfect universality in quantum teleportation. For the case of $d=2$, we derived a tighter relationship between the two measures. Taking into account other realistic noises, namely, the fluctuations in system control, we proposed a machine-learning-based algorithm to stabilize teleportation. We demonstrated by means of numerical simulations that even the fullest deteriorations can be cured.
It is remarkable that the process of fidelity maximization guarantees the minimization of fidelity deviation without additional processes.
The aforementioned trait (coming from Result~\ref{result:U}) is indeed beneficial to reduce the algorithm time and realize faster system remediation; in fact, if the minimization of fidelity deviation was considered in the algorithm, we may not have obtained a cure cycle (i.e., sufficient time for iterations to cure abrupt fluctuations). Such a gain is expected to be more conspicuous in large-$d$ teleportation.
\section*{Acknowledgments}
JB thanks to M. Wie\'{s}niak and T. V\'{e}rtesi for discussions. WS and JB thank to the financial support of the National Research Foundation (NRF) of Korea Grants (no.~2019R1A2C2005504 and no.~NRF-2019M3E4A1079666), funded by the MSIP (Ministry of Science, ICT and Future Planning), Korea government. JR acknowledges the National Research Foundation of Korea (NRF) Grants no. NRF-2020M3E4A1079792). JB and KB was supported by KIAS Individual Grants (no.~CG061003 and no.~CG074701), respectively. JB and KB was supported by the NRF Grant funded by the Korea government(MSIT) (no.~2020M3E4A1079939).
|
\section{Introduction}
This article is a continuation of the previous work, carried out in (\cite{A1}, \cite{akhmedov}, \cite{AP1}, \cite{ABBKP}, \cite{AP2}, \cite{AP3}, \cite{AP4}, \cite{AP5}, \cite{AP6}, \cite{AHP}, \cite{AGP}, \cite{AS}, \cite{APS}), on the geography of symplectic $4$-manifolds. For some background and concise history on symplectic geography problem, we refer the reader to the introduction found in \cite{AHP}, \cite{AP4}, and \cite{AGP}.
Our work here is greatly motivated and influenced by the recent work of Donald Cartwright, Vincent Koziarz, and third author in \cite{CKY} and the earlier work of Gopal Prasad and the third author in \cite{PY, PY1}. The main purpose of our article is to construct new minimal symplectic $4$-manifolds that are interesting with respect to the symplectic geography problem. Starting from Cartwright-Steger surfaces, and their normal covers on Bogomolov-Miyaoka-Yau line $c_1^2 = 9\chi_h$, the Hirzebruch's line-arrangement surfaces and their quotients, by forming their symplectic connected sum with the exotic symplectic $4$-manifolds constructed in \cite{AP2, AS}, or the product $4$-manifolds $\Sigma_{g} \times \Sigma_{h}$, and applying the sequence of Luttinger surgeries along the lagrangian tori, we obtain a family of new symplectic $4$-manifolds with non-negative signatures. As a consequence of our work, we produce (i) an irreducible symplectic and infinitely many non-symplectic $4$-manifolds that are homeomorphic but not diffeomorphic to $(2n-1){\mathbb{CP}}{}^{2}\#(2n-1)\overline{\mathbb{CP}}{}^{2}$ for each integer $n \geq 9$, (ii) the families of simply connected irreducible nonspin symplectic $4$-manifolds that have the smallest Euler characteristics among the all known simply connected $4$-manifolds with positive signature and with more than one smooth structure. We also construct a complex surface on Bogomolov-Miyaoka-Yau line $c_1^2 = 9\chi_h$ using Hirzebruch's certain line-arrangement surface.
Before stating our main results, let us fix some notations that will be used throughout this paper. Given two $4$-manifolds, $X$\/ and $Y$, we will denote their connected sum by $X\# Y$. For a positive integer $k\geq 2$, the connected sum of $k$\/ copies of $X$\/ will be denoted by $kX$\/. Let ${\mathbb{CP}}{}^{2}$ denote the complex projective plane, with its standard orientation, and let $\overline{\mathbb{CP}}{}^{2}$ denote the underlying smooth $4$-manifold ${\mathbb{CP}}{}^{2}$ equipped with the opposite orientation. Our main results are the following theorems.
\begin{theo}\label{thm:main1} Let $M$ be $(2n-1){\mathbb{CP}}{}^{2}\#(2n-1)\overline{\mathbb{CP}}{}^{2}$ for any integer $n \geq 9$. Then there exist an infinite family of irreducible symplectic and an infinite family of irreducible non-symplectic $4$-manifolds that all are homeomorphic but not diffeomorphic to $M$. \end{theo}
The theorem above improves one of the main results of \cite{AP3, AS} where exotic irreducible smooth structures on $(2n-1){\mathbb{CP}}{}^{2}\#(2n-1)\overline{\mathbb{CP}}{}^{2}$ for $n\geq 25$ and for $n\geq 12$ were constructed, respectively. The next theorem improves the main results of \cite{AP3, AHP, AS} for the positive signature cases.
\begin{theo}\label{thm:main2}
Let $M$ be one of the following\/ $4$-manifolds.
\begin{itemize}
\item[(i)] $(2n-1){\mathbb{CP}}{}^{2}\#(2n-2)\overline{\mathbb{CP}}{}^{2}$ for any integer $n \geq 9$.
\item[(ii)] $(2n-1){\mathbb{CP}}{}^{2}\#(2n-3)\overline{\mathbb{CP}}{}^{2}$ for any integer $n \geq 10$.
\end{itemize}
Then there exist an infinite family of irreducible symplectic\/ $4$-manifolds and an infinite family of irreducible non-symplectic\/ $4$-manifolds that are homeomorphic but not diffeomorphic to\/ $M$.
\end{theo}
The second theorem above, which deals with the cases of signature equal $1$ and $2$, can be extended to the signature grater than equal $3$ cases as well.
Let us recall that exotic irreducible smooth structures on $(2n-1){\mathbb{CP}}{}^{2}\#(2n-1)\overline{\mathbb{CP}}{}^{2}$ for $n\geq 12$, on $(2n-1){\mathbb{CP}}{}^{2}\#(2n-2)\overline{\mathbb{CP}}{}^{2}$ for $n\geq 14$, on $(2n-1){\mathbb{CP}}{}^{2}\#(2n-3)\overline{\mathbb{CP}}{}^{2}$ for $n\geq 13$, and on $(2n-1){\mathbb{CP}}{}^{2}\#(2n-4)\overline{\mathbb{CP}}{}^{2}$ for $n\geq 15$ were constructed in \cite{AS} (see also earlier work in \cite{AP3} and \cite{AHP}).
Our paper is organized as follows. In Sections~\ref{sec: model} and \ref{sec:Luttinger}, we discuss some background material and collect some building blocks that are needed in our constructions of symplectic $4$-manifolds. In Sections~\ref{sec:ball quotients}, \ref{sec:positive}, \ref{sec:positive CS}, we present the proofs of our main results. A preliminary report on this work has been presented by the first author at Purdue University and by the second author at MPIM and in various research seminars and workshops since November 2018.
\section{Complex surfaces on Bogomolov-Miyaoka-Yau line}\label{sec: model}
\subsection{Fake projective planes}\label{sec: model1}
A fake projective plane is a smooth complex surface which is not the complex projective plane, but has the same Betti numbers as the complex projective plane. The first fake projective plane was constructed by David Mumford in 1979 using p-adic uniformization \cite{Mu}. He also showed that there could only be a finite number of such surfaces. Two more examples were found by Ishida and Kato \cite{IsKa} in 1998, and another by Keum \cite{Ke} in 2006. In 2007 \cite{PY} (see also Addendum \cite{PY1}), the third author and Gopal Prasad almost completely classified fake projective planes by proving that they fall into “28 classes”. Using the arithmeticity of the fundamental group of fake projective planes, and the formula for the covolume of principal arithmetic subgroups, they found twenty eight distinct classes of fake projective planes. For a very small number of classes, they left open the question of existence of fake projective planes in that class, but conjectured that there are none. Finally, Donald Cartwright and Tim Steger verified their conjecture and found all the fake projective planes, up to isomorphism, in each of the 28 classes \cite{CS}.
\begin{exam} In this example, we recall some properties of fake projective plane $M$. We refer the reader to \cite{Y} and also \cite{Ye}, where a complete classification of all smooth surfaces of general type with Euler number $3$ is given. There are $50$ pairs of fake projectives planes as classified in \cite{PY, PY1, CKY} and one Cartwright-Steger surface to be explained in {\bf 2.2}.
For fake projective planes, the Euler characteristic and the Betti numbers of $M$ are $e(M)=3$, $b_{1}(M)=0$ and $b_{2}(M)=1$. $M$ is a minimal complex surface of general type with $\sigma(M) =1$, ${c_{1}}^2(M)=3e(M) = 9$ and $\chi_{h}(M) = 1$. The intersection form of $M$ is odd, and has rank $1$. The fundamental group $\Pi$ of $M$ is a torsion-free cocompact arithmetic subgroup of $PU(2, 1)$, thus $M$ is a ball quotient $B_{\mathbb{C}}^{2}/\Pi$. For $46$ pairs of fake projecitve planes, the canonical line bundle $K_M$ is divisible by $3$, i.e., there is a line bundle $L$ such that $K_M=3L$. For the remaining four pairs of fake projective planes, we know that $K=3H+\tau$ for some torsion line bundle $\tau$. It was mentioned that the class of $H$ can be represented by a symplectic surface $H$ of self-intersection $1$ (see discussion in \cite{Ko}, pages 212-213), but notice that Taubes result concerning existence of pseudoholomorphic curves does not apply to the classes $H$ or $2H$ in this case (\cite{L-T}). By considering the classes $pH$ for any positive integer $p \geq 3$, we can produce symplectic surface $H(p)$ of self-intersection $p^2$ and the genus $g(H(p)) = 1 + 1/2(pH \cdot pH + 3H \cdot pH) = 1 + p(p+3)/2$. These symplectic surfaces in $M$ are quite useful. Using the symplectic connected sum operation \cite{gompf}, the pair $(M, H(p))$ (or the covers of $M$ on BMY line) and the knot surged homotopy elliptic surfaces $E(n)_{K}$ along with the symplectic submanifold $S_{K}$ can be used to construct exotic symplectic $4$-manifolds with positive signature and near the Bogomolov-Miyaoka-Yau line $c_1^2 = 9\chi_h$. The symplectic surface $S_{K}$ above in $E(n)_{K}$ is a higher genus section of self-intersection $-n$ resulting from $-n$ sphere section of $E(n)$ under the the knot surgery along a fibered knot $K$ of genus $g$. To make this construction work, one needs to set $n=p^2$ and $g=1+p(p+3)/2$. Since $\pi_1(E(n)_{K} \setminus S_{K})$ is trivial and $\pi_{1}(H(p))$ surjects into $\pi_{1}(M)$, the resulting symplectic $4$-manifold is simply connected.
\end{exam}
\subsection{Complex surfaces of Cartwright and Steger}
\label{sec: model2}
The study of enumerating the set of all fake projective planes in the so-called class $\mathcal{C}_{11}$ in the notation of \cite{PY} led Donald Cartwright and Tim Steger to discover a complex surface with irregularity $q=1$ and Euler characteristic $e=3$, named as Cartwright-Steger surface. Cartwright and Steger showed that a certain maximal arithmetic subgroup $\bar{\Gamma}$ of $PU(2; 1)$ contains a torsion-free subgroup $\Pi$ of index $864$ which has abelianization $\mathbb{Z}^2$. Such subgroup $\Pi$ is unique up to conjugation. Furthermore, for each ineger $n \geq 1$, $\Pi$ has a normal subgroup $\Pi_{n}$ of index $n$. Let $M_{n} = B^{2}(\mathbb{C})/\Pi_{n}$ denote the quotient of a complex hyperbolic space by a torsion free lattice $\Pi_{n}$ of $PU(2; 1)$. The Euler characteristic of $M_{n}$\/ is $e(M_{n})=ne(M_{1})=3n$. $M_{n}$\/ is a minimal complex surface of general type with $\sigma(M_{n}) = n$, ${c_{1}}^2(M_{n})=3e(M_{n}) = 9n$ and $\chi_{h}(M_{n}) = n$. The intersection form of $M_{1}$ is odd, indefinite and modulo torsion is isomorphic to $3\langle 1 \rangle \oplus 2\langle -1 \rangle$. The Betti numbers of $M_{1}$ are: $1, 2, 5, 2, 1$. It is known that the Albanese map of $M_{1}$ gives rise to an Albanese fibration with generic fiber of genus $19$ \cite{CKY}.
\subsection{The covers of Cartwright-Steger surface}
\label{forM1}
Let us recall the following from \cite{CKY}. In one of our constructions we will be using the curves $b(M_c)$ or $ b^{-1}(M_c)$ in Proposition 2.4 of \cite{CKY}. For simplicity, let us consider $D = b(M_c)$.
Recall that in the notation of \cite{CS} and \cite{PY}, the maximal arithmetic lattice considered in this case is denoted by $\overline \Gamma$ summarized in Theorem 1 of \cite{CKY}. The lattice of the Cartwright-Steger surface is denoted by $\Pi$ with generators given by $a_1$, $a_2$, $a_3$ explained in Theorem 2 of \cite{CKY}.
The map $\pi:M=B_{{\mathbb C}}^2/\Pi\rightarrow B_{{\mathbb C}}^2/\overline \Gamma$ is a covering map of order $864$.
The quotient $B_{{\mathbb C}}^2/\overline \Gamma$ is represented by the right hand side of Figure 1 of \cite{CKY}.
$D$ is a component of $\pi^{-1}(D_A)$ in the picture and $\pi^{-1}(D_A)$ is an immersed totally geodesic curve. The singularities of $D$ could only be found in $\pi^{-1}(P_1)$ and $\pi^{-1}(P_2)$. $D$ is a component of genus $4$ in $\pi^{-1}(D_A)$. According to Proposition 2.4 of \cite{CKY}, the only singular points of the curve $D$ is given by a point of normal crossing given by $n_{-1}(D)=2$.
By Proposition 2.4(d) and its proof in \cite{CKY}, $\Pi_M\backslash M$ has genus~4 by the Riemann-Hurwitz formula, and we can find explicit generators $u_i$, $v_i$ of~$\Pi_M$ such that $[u_1,v_1][u_2,v_2][u_3,v_3][u_4,v_4]=1$. When $M=b(M_c)$, the following eight elements generate $\Pi_M$:
\begin{displaymath}
\begin{aligned}
p_1&=a_2^3a_1^{-1}a_3^{-1}j^8a_2^{-2}a_1^{-1}j^4,\\
p_2&=a_3^3a_1a_3^2a_2a_1j^4a_3^{-1}j^8a_3^{-2}a_1^{-1}a_3^{-3},\\
p_3&=j^8a_1^{-1}a_3^{-3}a_2^2j^4a_3^{-2}a_1^{-1}a_3^{-3},\\
p_4&=j^8a_2a_1a_2^{-2}a_1^{-1}j^4a_3^3a_1^2a_2^{-1},\\
\end{aligned}
\quad
\begin{aligned}
p_5&=a_3^3a_1a_3^2j^4a_1^{-1}j^8a_3^2a_1a_2^{-3},\\
p_6&=a_3^3a_1a_2a_1a_3a_2^{-3},\\
p_7&=a_3^3a_1j^8a_1a_2^{-2}a_1^{-1}a_3^2j^4,\\
p_8&=j^4a_3^{-2}j^8a_2a_1a_2a_1a_2^{-2},\\
\end{aligned}
\end{displaymath}
and satisfy the single relation
\begin{displaymath}
p_5^{-1}p_2^{-1}p_5p_1p_3p_8^{-1}p_4p_1^{-1}p_7^{-1}p_6^{-1}p_7p_2p_3^{-1}p_8p_4^{-1}p_6=1.
\end{displaymath}
Here the group elements such as $j$ is given by {\bf 1.1} of \cite{CKY}.
The above gives a set of generators for $\pi_1(\hat{D})$, where $\hat{D}$ is the normalization of $D$.
To compute $i_*(\pi_1(D))\subset \pi_1(M)$, where $i:D\rightarrow M$ is the inclusion, note that
$i_*(\pi_1(D))$ is generated by $i_*p_j, j=1,\dots,8$, together with loops around the nodal point.
The following two elements $\pi_1$ and $\pi_2$ of $\Pi$ satisfy $\pi(b^{-1}(O))\in b(M_c)$, are taken from the third table on page 41 of the arXiv version of the paper \cite{CKY}:\\
$$\pi_1:=a_3^3a_1a_2^{-1}, \ \
\pi_2:=a_2a_1^{-2}a_3^{-1}a_1a_3^{-1}a_1^{-1}a_2^{-2}$$
Consider the subgroup $G_1$ of $G$ given by $G_1 = \, < G \, | \, p_1,p_2,p_3,p_4,p_5,p_6,p_7,p_8,\pi_1\pi_2^{-1}>.$
By using the Magma program, one can verify that $G_1$ is a normal subgroup of $\Pi$ of order $4$. Moreover, it can be verified that the quotient group is ${\mathbb Z}_2\times{\mathbb Z}_2$. No larger subgroup of $G$ containing $G_1$ could
be found and hence $G_1$ is our candidate $i_*(\pi_1(D))$. The authors are grateful to Donald Cartwright for his help with Magma computations.
In conclusion, we have the following.
\begin{prop}
$D$ is an immersed totally geodesic curve satisfying the following properties.
\begin{enumerate}
\item The normalization $\hat{D}$ of $D$ is a Riemann surface of genus $4$.\\
\item $D\cdot D=-1$.\\
\item $i_*\pi_1(D)$ is a normal subgroup of $\pi_1(M)$ of index $4$,
and $\pi_1(M)/i_*\pi_1(D)={\mathbb Z}_2\times{\mathbb Z}_2$.
\end{enumerate}
\end{prop}
Denote by $H$ the covering group $\pi_1(M)/i_*\pi_1(D)={\mathbb Z}_2\times{\mathbb Z}_2$ in Proposition 1.
We have
$$1\rightarrow i_*\pi_1(D)\rightarrow \pi_1(M)\rightarrow H.$$
Consider now a normal unramified covering $\tilde{M}$ of $M$ with covering group given by $H$. Let $p:\tilde{M}\rightarrow M$ be the covering map.
From construction, $p^{-1}(D)$ consists of four connected components. Let $E$ be one such connected component. Then from construction,
inclusion $i_*\pi_1(E)\rightarrow \pi_1(\tilde{M})$ is an isomorphism. Hence we have
\begin{lemm}\label{lemm1}
$E$ is a curve of self-intersection $-1$ on $\tilde{M}$. The normalization of $E$ is a Riemann surface of genus $4$. Moreover,
$i_*\pi_1(E)\rightarrow \pi_1(\tilde{M})$ is an isomorphism.
\end{lemm}
This follows from construction. Note that a neighborhood of $D$ in $M$ is isomorphic to a neighborhood of $E$ in $\tilde{M}$, as
the covering is a normal covering with $\pi_1(\tilde{M})$ a normal subgroup of $\Pi$.
\begin{lemm}\label{lemm2}
The Chern numbers of $\tilde{M}$ are given by
$c_1^2(\tilde{M})=36$, $c_2(\tilde{M})=12$.
\end{lemm}
This follows from the fact that the Chern numbers involved are multiplicative.
\begin{lemm}\label{lemm3}
$\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ contains a symplectic genus $5$ curve $\Sigma_5$ of self intersection $-2$ that carries the fundamental group of $\tilde{M}\#\overline{\mathbb{CP}}{}^{2}$
\end{lemm}
\begin{proof}
It was shown in Lemma \ref{lemm1} that $\tilde{M}$ contains a curve $E$ of self intersection $-1$, whose normalization is a Riemann surface of genus $4$. Since genus is a birational invariant, the genus of $E$ is $4$ as well. We symplectically blow up $E$ at its self intersection, so that it becomes square $-5$ curve and the exceptional sphere $e_1$ intersects it twice. We symplectically smooth two intersections points of the proper transform of $E$ with $e_1$, which gives us genus $5$ symplectic curve $\Sigma_5$ of self intersection $-2$ inside $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$. Since $i_*\pi_1(E)\rightarrow \pi_1(\tilde{M})$ is an isomorphism, we see that $\Sigma_5$
carries the fundamental group of $\tilde{M}\#\overline{\mathbb{CP}}{}^{2}$.
\end{proof}
\section{Luttinger surgery and symplectic cohomology $(2n-3)(\mathbb{S}^2\times \mathbb{S}^2)$}\label{sec:Luttinger}
We briefly review the Luttinger surgery, and collect some symplectic building blocks that will be used later in our constructions. For the details on Luttinger surgery, the reader is referred to the papers \cite{lu} and \cite{ADK}.
\begin{defi} Let $X$\/ be a symplectic $4$-manifold with a symplectic form $\omega$, and the torus $\Lambda$ be a Lagrangian submanifold of $X$. Given a simple loop $\lambda$ on $\Lambda$, let $\lambda'$ be a simple loop on $\partial(\nu\Lambda)$ that is parallel to $\lambda$ under the Lagrangian framing. For any integer $n$, the $(\Lambda,\lambda,1/n)$ Luttinger surgery\/ on $X$\/ defined to be the $X_{\Lambda,\lambda}(1/n) = ( X - \nu(\Lambda) ) \cup_{\phi} (\mathbb{S}^1 \times \mathbb{S}^1 \times \mathbb{D}^2)$, the $1/n$\/ surgery on $\Lambda$ with respect to $\lambda$ under the Lagrangian framing. Here $\phi : \mathbb{S}^1 \times \mathbb{S}^1 \times \partial \mathbb{D}^2 \to \partial(X - \nu(\Lambda))$ denotes a gluing map satisfying $\phi([\partial \mathbb{D}^2]) = n[{\lambda'}] + [\mu_{\Lambda}]$ in $H_{1}(\partial(X - \nu(\Lambda))$, where $\mu_{\Lambda}$ is a meridian of $\Lambda$. \end{defi}
It is shown in \cite{ADK} that $X_{\Lambda,\lambda}(1/n)$ possesses a symplectic form that restricts to the original symplectic form $\omega$ on $X\setminus\nu\Lambda$. The proof of the following lemma is easy to verify and is left to the reader as an exercise.
\begin{lemm}\label{invariants}
\begin{enumerate}
\noindent \item $\pi_1(X_{\Lambda,\lambda}(1/n)) = \pi_1(X- \Lambda)/N(\mu_{\Lambda} \lambda'^n)$, where
$N(\mu_{\Lambda} \lambda'^n)$ denote the smallest normal subgroup of $\pi_1(X- \Lambda)$ that contains $\mu_{\Lambda} \lambda'^n$
\noindent \item $\sigma(X)=\sigma(X_{\Lambda,\lambda}(1/n))$ and $e(X)=e(X_{\Lambda,\lambda}(1/n))$.
\end{enumerate}
\end{lemm}
\subsection{Luttinger surgeries on product manifolds $\Sigma_{n}\times \Sigma_{2}$ and $\Sigma_{n}\times \mathbb{T}^2$}\label{L}
Recall from \cite{FPS, AP1} that for each integer $n\geq2$, there is a family of irreducible pairwise non-diffeomorphic 4-manifolds $\{Y_n(m)\mid m=1,2,3,\dots\}$ that have the same integer cohomology ring as $(2n-3)(\mathbb{S}^2\times \mathbb{S}^2)$. $Y_n(m)$ are obtained by performing $2n+3$ Luttinger surgeries (cf.\ \cite{ADK, lu}) and a single $m$\/ torus surgery on $\Sigma_2\times \Sigma_n$. These $2n+4$ torus surgeries are performed as follows
\begin{eqnarray}\label{first 8 Luttinger surgeries}
&&(a_1' \times c_1', a_1', -1), \ \ (b_1' \times c_1'', b_1', -1), \ \
(a_2' \times c_2', a_2', -1), \ \ (b_2' \times c_2'', b_2', -1),\\ \nonumber
&&(a_2' \times c_1', c_1', +1), \ \ (a_2'' \times d_1', d_1', +1),\ \
(a_1' \times c_2', c_2', +1), \ \ (a_1'' \times d_2', d_2', +m),
\end{eqnarray}
together with the following $2(n-2)$ additional Luttinger surgeries
\begin{gather*}
(b_1'\times c_3', c_3', -1), \ \
(b_2'\times d_3', d_3', -1), \ \dots ,\
(b_1'\times c_n', c_n', -1), \ \
(b_2'\times d_n', d_n', -1).
\end{gather*}
Here, $a_i,b_i$ ($i=1,2$) and $c_j,d_j$ ($j=1,\dots,n$) denote the standard loops that generate $\pi_1(\Sigma_2)$ and $\pi_1(\Sigma_n)$, respectively. See Figure~\ref{fig:lagrangian-pair} for a typical Lagrangian tori along which the surgeries are performed.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.89]{fig9.pdf}
\caption{Lagrangian tori $a_i'\times c_j'$ and $a_i''\times d_j'$}
\label{fig:lagrangian-pair}
\end{center}
\end{figure}
Since $m$-torus surgery is non-symplectic for $m\geq 2$, the manifold $Y_n(m)$ is symplectic only when $m=1$. Using the Lemma~\ref{invariants}, we see that the Euler characteristic of $Y_n(m)$ is $4n-4$ and its signature is $0$. $\pi_1(Y_n(m))$ is generated by $a_i,b_i,c_j,d_j$ ($i=1,2$ and $j=1,\dots,n$) and the following relations hold in $\pi_1(Y_n(m))$:
\begin{gather}\label{Luttinger relations}
[b_1^{-1},d_1^{-1}]=a_1,\ \ [a_1^{-1},d_1]=b_1,\ \ [b_2^{-1},d_2^{-1}]=a_2,\ \ [a_2^{-1},d_2]=b_2,\\ \nonumber
[d_1^{-1},b_2^{-1}]=c_1,\ \ [c_1^{-1},b_2]=d_1,\ \ [d^{-1}_2,b^{-1}_1]=c_2,\ \ [c_2^{-1},b_1]^m=d_2,\\ \nonumber
[a_1,c_1]=1, \ \ [a_1,c_2]=1,\ \ [a_1,d_2]=1,\ \ [b_1,c_1]=1,\\ \nonumber
[a_2,c_1]=1, \ \ [a_2,c_2]=1,\ \ [a_2,d_1]=1,\ \ [b_2,c_2]=1,\\ \nonumber
[a_1,b_1][a_2,b_2]=1,\ \ \prod_{j=1}^n[c_j,d_j]=1,\\ \nonumber
[a_1^{-1},d_3^{-1}]=c_3, \ \ [a_2^{-1} ,c_3^{-1}] =d_3, \ \dots, \
[a_1^{-1},d_n^{-1}]=c_n, \ \ [a_2^{-1} ,c_n^{-1}] =d_n,\\ \nonumber
[b_1,c_3]=1,\ \ [b_2,d_3]=1,\ \dots, \
[b_1,c_n]=1,\ \ [b_2,d_n]=1.
\end{gather}
The surfaces $\Sigma_2\times\{{\rm pt}\}$ and $\{{\rm pt}\}\times \Sigma_n$ in $\Sigma_2\times\Sigma_n$ are not affected by the above Luttinger surgeries,
and descend to surfaces in $Y_n(m)$. They are symplectic submanifolds in $Y_n(1)$. Let us denote these symplectic submanifolds n $Y_n(1)$ by $\Sigma_2$ and $\Sigma_n$.
Note that $[\Sigma_2]^2=[\Sigma_n]^2=0$ and $[\Sigma_2]\cdot[\Sigma_n]=1$. Let $\mu(\Sigma_2)$ and $\mu(\Sigma_n)$ denote the meridians of these surfaces in $Y_n(m)$. The above construction easily generalizes to $\Sigma_3\times\Sigma_n$. We will denote the resulting smooth manifold in this case as $Z_n(m)$.
Next, we consider a slightly different construction. Let us fix integers $n \geq 2$, and $m \geq 1$. Let $Y_{n}(1,m)$ denote smooth $4$-manifold obtained by performing the following $2n$ torus surgeries on $\Sigma_n\times \mathbb{T}^2$:
\begin{eqnarray}\label{eq: Luttinger surgeries for Y_1(m)}
&&(a_1' \times c', a_1', -1), \ \ (b_1' \times c'', b_1', -1),\\ \nonumber
&&(a_2' \times c', a_2', -1), \ \ (b_2' \times c'', b_2', -1),\\ \nonumber
&& \cdots, \ \ \cdots \\ \nonumber
&&(a_{n-1}' \times c', a_{n-1}', -1), \ \ (b_{n-1}' \times c'', b_{n-1}', -1),\\ \nonumber
&&(a_{n}' \times c', c', +1), \ \ (a_{n}'' \times d', d', +m).
\end{eqnarray}
Let $a_i,b_i$ ($i=1,2, \cdots, n$) and $c,d$\/ denote the standard generators of $\pi_1(\Sigma_{n})$ and $\pi_1(\mathbb{T}^2)$, respectively. Since all the torus surgeries listed above are Luttinger surgeries when $m = 1$ and the Luttinger surgery preserves minimality, $Y_{n}(1/p,1/q)$ is a minimal symplectic $4$-manifold. The fundamental group of $Y_{n}(1/p,m/q)$ is generated by $a_i,b_i$ ($i=1,2,3 \cdots, n$) and $c,d$, and the Lemma~\ref{invariants} implies that the following relations hold in $\pi_1(Y_{n}(1,m))$:
\begin{gather}\label{Luttinger relations for Y_1(m)}
[b_1^{-1},d^{-1}]=a_1,\ \ [a_1^{-1},d]=b_1,\ \
[b_2^{-1},d^{-1}]=a_2,\ \ [a_2^{-1},d]=b_2,\\ \nonumber
\cdots, \ \ \cdots, \\ \nonumber
[b_{n-1}^{-1},d^{-1}]=a_{n-1},\ \ [a_{n-1}^{-1},d]=b_{n-1},\ \
[d^{-1},b_{n}^{-1}]=c,\ \ {[c^{-1},b_{n}]}^{-m}=d,\\ \nonumber
[a_1,c]=1,\ \ [b_1,c]=1,\ \ [a_2,c]=1,\ \ [b_2,c]=1,\\ \nonumber
[a_3,c]=1,\ \ [b_3,c]=1,\\ \nonumber
\cdots, \ \ \cdots, \\ \nonumber
[a_{n-1},c]=1,\ \ [b_{n-1},c]=1,\\ \nonumber
[a_{n},c]=1,\ \ [a_{n},d]=1,\\ \nonumber
[a_1,b_1][a_2,b_2] \cdots [a_n,b_n]=1,\ \ [c,d]=1.
\end{gather}
Let us denote by $\Sigma'_n \subset Y_{n}(1,m)$ a genus $n$ surface that desend from the surface $\Sigma_{n}\times\{{\rm pt}\}$ in $\Sigma_{n}\times \mathbb{T}^2$.
\section{Construction of a smooth complex algebraic surface with $K^2 = 144$ and $\chi_h = 16$}\label{sec:ball quotients}
In this section, we construct a smooth complex algebraic surface with invariants $K^2 = 144$ and $\chi_h = 16$. This complex surface of general type is on the BMY line $c_1^2 = 9\chi_h$, and thus is a ball quotient. It is obtained as an abelian covering of the complex projective plane branched over an arrangement of $12$ lines shown as in Figure~\ref{fig:H}, known in the literature as the Hesse configuration. Such complex surfaces with bigger invariants, $K^2$ and $\chi_h$, was initially studied by by Friedrich Hirzebruch (for example, see \cite{Hirze}, page 134). Our construction is motivated and similar in spirit to that of Bauer-Catanese in \cite{Main}, where the complex ball quotients obtained from a complete quadrangle arrangement in $\mathbb{CP}^2$.
In $\mathbb{CP}^2$, let us consider the Hesse arrangement $H$, which is a configuration of $9$ points $p_i$ ($1\leq i \leq 9$) and $12$ lines $l_j$ ($1 \leq j \leq 12$), such that each line passes through $3$ of the points $p_i$ and each point lies at the intersection of $4$ of the lines $l_j$ (see Figure~\ref{fig:H}). We blow up $\mathbb{CP}^2$ at the points $p_1, \cdots, p_9$, and denote the blow up map by $\pi: T:=\widehat{\mathbb{CP}^2} \rightarrow \mathbb{CP}^2$. Let $E_i$ be the exceptional divisor corresponding to the blow up at the point $p_i$ for $i=1,\cdots, 9$. In the sequel, we will slightly abuse our notation and denote the proper transform of a line $l_j$ using the same symbol, or $\tilde l_j$ when distinction needed.
Let us now take the formal sum of the proper transforms $l_j$ of the $12$ lines of the arrangement and the $9$ exceptional divisors $E_i$'s, and denote it by $D$. The divisor $D$ in $T$ has only simple normal crossings. The homology classes of simple closed loops around the $l_j$'s and the $E_i$'s generate $H_1(T-D, \mathbb{Z})$. Let us denote a loop encircling a line $E_i$ or $l_j$ by using the same letter. Then for each $i=1,\cdots,9$, the class of $E_i$ can be written as a sum of the homology classes of $4$ loops around the $4$ lines intersecting $E_i$. To illustrate this, notice that we have $E_1 = l_1 + l_4 + l_7 + l_{10}$ and similar relations hold for the other $E_i$'s. Moreover, the sum of the homology classes of $12$ loops $l_j$'s are $0$, which shows that $H_1(T-D, \mathbb{Z})$ is a free group of rank $11$.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=.65]{HH3.pdf}
\caption{Hesse arrangement in $\mathbb{CP}^2$}
\label{fig:H}
\end{center}
\end{figure}
It is known that a surjective homomorphism $\varphi: \mathbb{Z}^{11} \simeq H_1(T-D,\mathbb{Z}) \rightarrow (\mathbb{Z}/3\mathbb{Z})^2$ determines an abelian $(\mathbb{Z}/3\mathbb{Z})^2$-cover $p:W \rightarrow T=\widehat{\mathbb{CP}^2}$. We need that $p$ is branched exactly in $D$. Notice that there are various epimorphisms satisfying this. To illustrate one, let us define $\varphi$ as follows:
\begin{align*}
&\varphi(l_1) = \varphi(l_2) = \varphi(l_3) = (1,0),\\
&\varphi(l_4) = (2,0), \varphi(l_5) = (0,1), \varphi(l_6) = (2,2),\\
&\varphi(l_7) = (2,1), \varphi(l_8) = (1,2), \varphi(l_9) = (2,0),\\
&\varphi(l_{10}) = \varphi(l_{11}) = (1,1), \varphi(l_{12}) = (0,2).
\end{align*}
Then
\begin{align*}
&\varphi(E_1) = \varphi(l_1+l_4+l_7+l_{10}) = (0,2),
&\varphi(E_2) = \varphi(l_2+l_4+l_9+l_{11}) = (0,1),\\
&\varphi(E_3) = \varphi(l_3+l_4+l_{12}+l_8) = (1,1),
&\varphi(E_4) = \varphi(l_1+l_5+l_{11}+l_{12}) = (2,1),\\
&\varphi(E_5) = \varphi(l_2+l_5+l_7+l_8) = (1,1),
&\varphi(E_6) = \varphi(l_3+l_5+l_9+l_{10}) = (1,2),\\
&\varphi(E_7) = \varphi(l_1+l_6+l_8+l_9) = (0,1),
&\varphi(E_8) = \varphi(l_2+l_6+l_{12}+l_{10}) = (1,2),\\
&\varphi(E_9) = \varphi(l_3+l_6+l_7+l_{11}) = (0,1)
\end{align*}are all nonzero.
Also, $\varphi(l_1+l_2+l_3+l_{7} + l_9+l_{10}) \neq (0,0)$. These conditions ensure that $\varphi$ gives a $(\mathbb{Z}/3\mathbb{Z})^2$ Galois cover branched exactly in $D$ (see Lemma 2.3, part 1 in \cite{Main}, also \cite{Ku}).
We also note that the following are linearly independent:
\begin{center}
$\varphi(E_1)$ and $\varphi(l_i)$, $i=1,4,7,10$;
$\varphi(E_2)$ and $\varphi(l_i)$, $i=2,4,9,11$; \\
$\varphi(E_3)$ and $\varphi(l_i)$, $i=3,4,12,8$;
$\varphi(E_4)$ and $\varphi(l_i)$, $i=1,5,11,12$;\\
$\varphi(E_5)$ and $\varphi(l_i)$, $i=2,5,7,8$;
$\varphi(E_6)$ and $\varphi(l_i)$, $i=3,5,9,10$;\\
$\varphi(E_7)$ and $\varphi(l_i)$, $i=1,6,8,9$;
$\varphi(E_8)$ and $\varphi(l_i)$, $i=2,6,12,10$;\\
$\varphi(E_9)$ and $\varphi(l_i)$, $i=3,6,7,11$.
\end{center}
Moreover, $D$ has simple normal crossings, we deduce that the total space $W$ is smooth (see Lemma 1.4 in \cite{Ku}).
Let us compute some invariants of the surface $W$, and verify that $c_1^2(W)=K_W^2=144$ and $\chi_h(W) = 16$. Let $H$ be the divisor class corresponding to the invertible sheaf $\mathcal{O}(1)$ on $\mathbb{CP}^2$. The canonical sheaf $w_{\mathbb{CP}^2}$ of $\mathbb{CP}^2$ is $\mathcal{O}(-2-1)=\mathcal{O}(-3)$ which corresponds to the canonical divisor $-3H$. Then, the canonical divisor $K_Y$ of $Y$ is $-3H + \sum\limits_{i=1}^{9}E_i$ where we denoted the pullback of $H$ by itself. By using the canonical divisor formula for abelian covers (Proposition 4.2 in \cite{Par}), we compute
\begin{eqnarray*}
K_W &=& \pi^* \Big( (-3H + \sum\limits_{i=1}^{9}E_i) + \frac{2}{3} \sum\limits_{i=1}^{9}E_i + \frac{2}{3}(12H- 4\sum\limits_{i=1}^{9}(E_i) \Big)\\
&=& \pi^* \Big(5H - \sum\limits_{i=1}^{9}E_i \Big).
\end{eqnarray*}
Since $H \cdot E_i = 0,\, H^2=1$ and $E_i^2 = -1$, the above equality gives $K_W^2 = 9 (25-9) = 144$.
The Euler number $e(W)$ of $W$ can be found as follows.
\begin{equation*}
e(W) = 9 e(\widehat{\mathbb{CP}^2} = \mathbb{CP}^2 \# 9\overline{\mathbb{CP}^2}) - 6\cdot 21 e(\mathbb{CP}^1)+4 \cdot 48 = 48.
\end{equation*}
Thus $c_1^2(W) = 3 c_2(W)$, and $W$ is a ball quotient. Since $12 \chi_h(W) - c_1^2(W) = e(W)$, we have $\chi_h(W) = 16$. In summary, we proved the following theorem.
\begin{theo}
There exists a smooth complex algebraic surface $W$ with invariants $c_1^2(W)=144$ and $\chi_h(W) = 16$ constructed as $(\mathbb{Z}/3\mathbb{Z})^2$-cover of ${\mathbb{CP}}{}^{2}$ branched over the Hesse configuration.
\end{theo}
Now we consider the map $p \circ \pi : W \rightarrow \mathbb{CP}^2$, where $\pi$ is the blow up map, $p$ is the abelian cover. Let us take $p_1$, one of the blown up points in $\mathbb{CP}^2$ which is the intersection point of $l_1,l_4,l_7,l_{10}$ (see Figure~\ref{fig:H}). The pencil of lines in $\mathbb{CP}^2$ passing through $p_1$ lifts to a fibration on $W$. To determine the genus of the generic fiber of this fibration, we take a line $K$ passing through $p_1$ such that its only intersection with the lines $l_1,l_4,l_7,l_{10}$ is $p_1$. In addition, $K$ intersects the remaining $8$ lines of the arrangement. These $8$ intersection points and the point $p_1$ are $9$ branch points on $K$. The preimage of $K-E_{1}$ in $W$, which is the generic fiber of the given fibration, is a degree $3$ cover of $K-E_{1}$ (cf. \cite{BHPV}, p.241), branched at $9$ points. For the determination of the genus $g$ of the surface above $K-E_{1}$, we apply the Riemann-Hurwitz ramification formula
\begin{equation}
2g-2 = 3 (-2) + 9 \cdot 2 \Rightarrow g=7.
\end{equation}
Therefore, generic fibers are of genus $7$ surfaces. Moreover, there are at least $9$ distinct fibrations in $W$ coming from the points $p_i$'s.
Let us consider the $12$ lines $l_j$ of the Hesse arrangement and determine their inverse images in $W$ under $p \circ \pi$. We observe that on each $l_j$, $j=1, \cdots, 12$, there are $5$ branch points. By the Riemann-Hurwitz formula, we have
\begin{equation}
2g-2 = 3 (-2) + 5 \cdot 2 \Rightarrow g=3.
\end{equation}
Therefore, they lift to genus $3$ curves. To find their self-intersections, we apply the adjunction formula. Firstly, we note that each $l_j$ is blown up at three points, say $p_k,p_l,p_m$. For its proper transform $\tilde l_j$ in $\widehat{\mathbb{CP}^2}$, we have
\begin{equation}
[\tilde l_j] = H-E_k-E_l-E_m.
\end{equation}
Thus,
\begin{eqnarray*}
K_W \cdot [\Sigma_3] &=& \pi^* \Big((5H - \sum\limits_{i=1}^{9}E_i) \cdot (H-E_k-E_l-E_m) \Big)\\
&=&3(5-1-1-1) = 6.
\end{eqnarray*}
Using the adjunction formula $2g-2 = 4 = K_W \cdot [\Sigma_3] + [\Sigma_3]^2$, we have $[\Sigma_3]^2 = -2$. On the other hand, on each exceptional sphere $E_i$, there are $4$ branch points. Thus, their preimages are genus $2$ curves in $W$:
\begin{equation}
2g-2 = 3 (-2) + 4 \cdot 2 \Rightarrow g=2.
\end{equation}
Similarly as above,
\begin{equation*}
K_W \cdot [\Sigma_2] = \pi^* \Big((5H - \sum\limits_{i=1}^{9}E_i) \cdot (E_i) \Big)=3
\end{equation*}
and by the adjunction formula we have $2g-2 = 2 = K_W \cdot [\Sigma_2] + [\Sigma_2]^2$; which shows that $[\Sigma_2]^2 = -1$.
Let us reconsider the pencil of lines in $\mathbb{CP}^2$ passing through $p_1$ and take the line $l_1$. The preimage of its proper transform $\tilde l_1$ is a genus three surface $\Sigma_3$ with self-intersection $-2$ in $W$. The exceptional divisors $E_1$, $E_4$ and $E_7$ intersecting $\tilde l_1$ lift to genus $2$ curves with self-intersections $-1$, each of which intersects $\Sigma_3$ transversally once. Notice that the lift of $E_1$ gives rise to a section, and the union of lifts of the exceptional divisors $E_4$, $E_7$, and the proper transform of intersecting $\tilde l_1$ corresponds to a singular fiber of the given fibration. We symplectically resolve their three transversal intersection points and obtain genus $9$ symplectic submanifold of $W$ with self intersection $+1$. As in Section 2.3 of \cite{AS}, we have the following proposition.
\begin{prop}
\label{genus9}
$W \# \overline{\mathbb{CP}}{}^{2}$ contains an embedded symplectic genus $9$ curve $\Sigma_9$ with self intersection $0$. Furthermore, there is a surjection $f_{*}: \pi_{1}(\Sigma_9) \rightarrow \pi_{1}(W \# \overline{\mathbb{CP}}{}^{2})$.
\end{prop}
\section{Constructions of Symplectic 4-Manifolds with Positive Signatures from Hirzebruch's line-arrangement surface}\label{sec:positive}
In what follows, we will construct families of simply connected, minimal, symplectic and smooth 4-manifolds with positive signatures, by making use of the complex surface $W \# \overline{\mathbb{CP}}{}^{2}$, which we constructed above using the Hesse configuration. By Proposition \ref{genus9}, we know that $W \# \overline{\mathbb{CP}}{}^{2}$ contains an embedded symplectic genus $9$ curve of self-intersection zero. We endow $W\#\overline{\mathbb{CP}}{}^{2}$ with the symplectic structure induced from the K\"{a}hler structure. It will be the first building block in our construction, which has the following invariants: $e(W\#\overline{\mathbb{CP}}{}^{2})= 49$, $\sigma(W\#\overline{\mathbb{CP}}{}^{2}) = 15$, $c_1^2(W\#\overline{\mathbb{CP}}{}^{2})= 143$ and $\chi_h(W\#\overline{\mathbb{CP}}{}^{2})= 16$. Our second piece will be a minimal, simply connected and symplectic $4$-manifold $X_{g,g+2}$ (\cite{AS}, Theorem 3.12 and Section 5):
\begin{theo}\label{thm2}
For any integer $g \geq 1$, there exist a minimal symplectic 4-manifold $X_{g,g+2}$ obtained via Luttinger surgery such that
\begin{itemize}
\item[(i)] $X_{g,g+2}$ is simply connected
\item[(ii)] $e(X_{g,g+2})= 4g+2$, $\sigma (X_{g,g+2}) = - 2$, $c_1^{2}(X_{g,g+2}) = 8g-2$, and\/ $\chi(X_{g,g+2}) = g$.
\item[(iii)] $X_{g,g+2}$ contains the symplectic surface $\Sigma$ of genus $2$ with self-intersection $0$ and two genus $g$ surfaces with self-intersection $-1$ intersecting $\Sigma$ positively and transversally.
\end{itemize}
\end{theo}
\subsection{Symplectic and smooth 4-manifolds with signatures equal to 12}
We now present our first construction of exotic symplectic $4$-manifolds with $\sigma = 12$. Our first building block is the complex surface $W\#\overline{\mathbb{CP}}{}^{2}$ containing genus $9$ symplectic surface with self-intersection $0$. The second building block is obtained from the symplectic $4$-manifold $X_{7,9}$, in the notation of Theorem~\ref{thm2}. We will use the fact that $X_{7,9}$ contains a symplectic genus two surface $\Sigma_2$ with self-intersection $0$ and two genus $7$ symplectic surfaces with self intersections $-1$ intersecting $\Sigma_2$ positively and transversally.
Let us review the construction of $X_{7,9}$ (see \cite{AS} for the details). We take a copy of $\mathbb{T}^2 \times \{pt\}$ and $\{pt\} \times \mathbb{T}^2$ in $\mathbb{T}^2 \times \mathbb{T}^2$ equipped with the product symplectic form, and symplectically resolve the intersection point of these dual symplectic tori. The resolution produces symplectic genus two surface of self intersection $+2$ in $\mathbb{T}^2 \times \mathbb{T}^2$. By symplectically blowing up this surface twice, in $\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}$, we obtain a symplectic genus 2 surface $\Sigma_2$ with self-intersection $0$, with two $-1$ spheres (i.e. the exceptional spheres resulting from the blow-ups) intersecting it positively and transversally. Next, we form the symplectic connected sum of $\mathbb{T}^4\#2\overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_7$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. By performing the sequence of appropriate $\pm 1$ Luttinger surgeries on $(\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_7)$, we obtain the symplectic $4$-manifold $X_{7,9}$ (\cite{AS}). It can be seen from the construction that, $X_{7,9}$ contains a symplectic surface $\Sigma_2$ with self intersection $0$ and two genus $7$ surfaces $S_{1}$ and $S_{2}$ with self intersections $-1$ which have positive and transverse intersections with $\Sigma_2$. Notice that the surfaces $S_{1}$ and $S_{2}$ result from the internal sum of the punctured exceptional spheres in $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and the punctured genus 7 surfaces in $\Sigma_2 \times \Sigma_7 \setminus \nu (\Sigma_2 \times \{pt\})$. Moreover, $X_{7,9}$ contains a pair of disjoint Lagrangian tori $T_{1}$ and $T_{2}$ of self-intersections $0$ such that $\pi_{1}(X\setminus(T_{1}\cup T_{2})) = 1$. Note that these Lagrangian tori descend from $\Sigma_2 \times \Sigma_7$, and survive in $X_{7,9}$ after symplectic connected sum and the Luttinger surgeries. This is because there are at least two pairs of Lagrangian tori in $\Sigma_2 \times \Sigma_7$ that were away from the standard symplectic surfaces $\Sigma_2 \times \{pt\}$ and $ \{pt\} \times \Sigma_7$, and the Lagrangian tori that were used for Luttinger surgeries. Also, the fact that $\pi_{1}(X_{7,9}\setminus(T_{1}\cup T_{2})) = 1$ is explained in details in \cite{AP3} (see proof of Theorem 8, page 272). Next, we symplectically resolve the intersection of $\Sigma_2$ and one of the genus $7$ surfaces, say $S_1$, in $X_{7,9}$. This produces the genus $9$ surface $\Sigma_9$ of square $+1$ intersecting the other genus $7$ surface $S_{2}$ with self-intersection $-1$. We blow up $\Sigma_9$ at one point (away from its intersection point with $S_{2}$). Thus we obtained a genus $9$ surface $\Sigma_9'$ of square 0 inside $X_{7,9} \# \overline{\mathbb{CP}}{}^{2}$.
Since the two symplectic building blocks $W \# \overline{\mathbb{CP}}{}^{2}$ and $X_{7,9} \# \overline{\mathbb{CP}}{}^{2}$ contain symplectic genus $9$ surfaces of self intersections zero, we can form their symplectic connected sum along these surfaces. Let
\begin{equation*}
Y = (W \# \overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_9=\Sigma_9'} (X_{7,9} \# \overline{\mathbb{CP}}{}^{2}).
\end{equation*}
\begin{lemm}
The symplectic manifold $Y$ has $e(Y) = 112$, $\sigma(Y) = 12$.
\end{lemm}
\begin{proof} We have stated the topological invariants of $W \# \overline{\mathbb{CP}}{}^{2}$ above and by Theorem~\ref{thm2}, we have $e(X_{7,9}) = 30$, $\sigma(X_{7,9}) = -2$. Applying the symplectic connected sum formula, we compute the topological invariants of $Y$ as above.
\end{proof}
Next, we proceed by following the same lines in \cite{AS}, Section 5 and the references therein. We show that $Y$ is symplectic and simply connected, using Gompf's Symplectic Connected Sum Theorem and Van Kampen's Theorem, respectively. Using Freedman's classification theorem for simply-connected $4$-manifolds, the lemma above and the fact that $W\#\overline{\mathbb{CP}}{}^{2}$\/ contains genus two surface of self-intersection $-1$ disjoint from $\Sigma_9$, we conclude that $Y$\/ is homeomorphic to $61{\mathbb{CP}}{}^{2}\#49\overline{\mathbb{CP}}{}^{2}$. Since $Y$ is symplectic, by Taubes's theorem, $Y$ has non-trivial Seiberg-Witten invariant. Next, using the connected sum theorem, we deduce that the Seiberg-Witten invariant of $61{\mathbb{CP}}{}^{2}\#49\overline{\mathbb{CP}}{}^{2}$ is trivial. Therefore, $Y$ is not diffeomorphic to $61{\mathbb{CP}}{}^{2}\#49\overline{\mathbb{CP}}{}^{2}$. Furthermore, $Y$ is a minimal symplectic $4$-manifold by Usher's Minimality Theorem \cite{usher}. Since symplectic minimality implies smooth minimality, $Y$ is also smoothly minimal, and thus is smoothly irreducible.
Moreover, as explained above, $Y$ contains a pair of disjoint Lagrangian tori $T_{1}$ and $T_{2}$ of self-intersection $0$ such that $\pi_{1}(Y\setminus(T_{1}\cup T_{2})) = 1$. We can perturb the symplectic form on $Y$ in such a way that one of the tori, say $T_{1}$, becomes symplectically embedded. We perform a knot surgery, (using a knot $K$ with non-trivial Alexander polynomial) on $Y$ along $T_{1}$ to obtain irreducible 4-manifold $(Y)_K$ that is homeomorphic but not diffeomorphic to $Y$. By varying our choice of the knot $K$, we can realize infinitely many pairwise non-diffeomorphic, irreducible 4-manifolds, either symplectic or nonsymplectic. (see Theorem 3.7 in \cite{AS})
\subsection{Symplectic and smooth 4-manifolds with signatures equal to 11}
In this section, we will construct simply connected, minimal, symplectic and smooth $4$-manifolds with signature is equal to $11$ in two different ways. The first construction gives the exotic $59\mathbb{CP}^{2} \# 48 \overline{\mathbb{CP}}^{2}$ and the second gives the exotic $63\mathbb{CP}^{2} \# 52\overline{\mathbb{CP}}^{2}$.
In first construction, one of our building block is again $W\#\overline{\mathbb{CP}}{}^{2}$, containing a symplectic genus $9$ surface $\Sigma_9$ of square $0$. To obtain the second symplectic building block, we form the symplectic connected sum of $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_6$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. Let
\begin{equation*}
X_{6,8} = (\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_6).
\end{equation*}
Similar to the discussion above, we see that $X_{6,8}$ contains a symplectic surface $\Sigma_2$ with self intersection $0$ and two genus $6$ surfaces with self intersections $-1$ which have positive and transverse intersections with $\Sigma_2$. Furthermore, $X_{6,8}$ contains a symplectic genus $7$ surface $\Sigma_7$ of square $0$ resulting from the internal sum of a punctured genus one surface in $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and a punctured genus 6 surface $\Sigma_6$ in $ \Sigma_2 \times \Sigma_6 \setminus \nu (\Sigma_2 \times \{pt\})$. In addition, $\Sigma_7$ intersects $\Sigma_2$ positively and transversely once. We symplectically resolve this intersection and get symplectic genus 9 surface of self intersection $+2$. We blow it up at two points and hence we obtain symplectic genus 9 surface $\Sigma_9''$ of square 0 inside $X_{6,8} \#2 \overline{\mathbb{CP}}{}^{2}$.
Next, we form the symplectic connected sum of $W \#\overline{\mathbb{CP}}{}^{2}$ and $X_{6,8} \# 2 \overline{\mathbb{CP}}{}^{2}$ along the symplectic genus 9 surfaces $\Sigma_9$ and $\Sigma_9''$ of squares zero. Let
\begin{equation*}
V= (W \#\overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_9 = \Sigma_9''} (X_{6,8} \# 2\overline{\mathbb{CP}}{}^{2}).
\end{equation*}
The invariants of $X_{6,8} \# 2 \overline{\mathbb{CP}}{}^{2}$ are as follows. $e(X_{6,8} \#2 \overline{\mathbb{CP}}{}^{2}) = 28$, $\sigma(X_{6,8} \#2 \overline{\mathbb{CP}}{}^{2}) = -4$, thus we have
\begin{lemm}
The symplectic manifold $V$ has $e(V) = 109$, $\sigma(V) = 11$.
\end{lemm}
We conclude as above that $V$ is symplectic, simply connected and an exotic copy of $59\mathbb{CP}^{2} \# 48 \overline{\mathbb{CP}}^{2}$ which is also smoothly minimal. As in the previous case, by performing knot surgery, we obtain infinitely many pairwise non-diffeomorphic, irreducible, symplectic and non-symplectic 4-manifolds.
\vspace{0.1in}
Let us now build another simply connected, minimal, symplectic and smooth 4-manifolds with signature $11$, but with different $\chi$. Our first building block is $Y= (W \# \overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_9=\Sigma_9'} (X_{7,9} \# \overline{\mathbb{CP}}{}^{2})$ constructed above. We note that $Y$ contains a genus $2$ surface of self intersection $-1$ coming from $W$, which is not affected by the symplectic connected sum operation in the construction of $Y$. For instance, one can consider $E_6$ in $\widehat{\mathbb{CP}^2}$, which is the exceptional sphere coming from the blow-up of the point $p_6$ in the Hesse arrangement. We have shown that the exceptional spheres lift to genus $2$ surfaces of self intersections $-1$ in $W$. Take one of the preimages of $E_6$ inside $W$. It is a symplectic genus $2$ surface of square $-1$ embedded in $W$, and descends to $Y$ after the symplectic connected sum. Let us denote it by $\Sigma_2'$. On the other hand, we take copies of $\mathbb{T}^2 \times \{pt\}$ and $\{pt\} \times \mathbb{T}^2$ in $\mathbb{T}^2 \times \mathbb{T}^2$ equipped with the product symplectic form, and symplectically resolve the intersection point of these dual symplectic tori. The resolution produces symplectic genus two surface of self intersection $+2$ in $\mathbb{T}^2 \times \mathbb{T}^2$. By symplectically blowing up this surface, we obtain a symplectic genus $2$ surface $\Sigma_2 \subset \mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$ of self-intersection $+1$, with the $-1$ exceptional sphere intersecting it positively and transversally. Next, we form the symplectic connected sum of $Y$ with $\mathbb{T}^4\#\overline{\mathbb{CP}}^{2}$ along the genus two surfaces $\Sigma_2'$ and $\Sigma_2$. Let
\begin{equation*}
L= Y \#_{\Sigma_2' = \Sigma_2'} (\mathbb{T}^4\#\overline{\mathbb{CP}}{}^{2}).
\end{equation*}
The invariants of $\mathbb{T}^4\#\overline{\mathbb{CP}}{}^{2}$ are as follows. $e(\mathbb{T}^4\#\overline{\mathbb{CP}}{}^{2}) = 1$, $\sigma(\mathbb{T}^4\#\overline{\mathbb{CP}}{}^{2}) = -1$, thus we have
\begin{lemm}
$e(L) = 117$, $\sigma(L) = 11$.
\end{lemm}
We can conclude that $L$ is symplectic, simply connected, an exotic copy of $63\mathbb{CP}^{2} \# 52\overline{\mathbb{CP}}^{2}$, which is also smoothly minimal. As in the previous case, by performing knot surgery we realize infinitely many pairwise non-diffeomorphic, irreducible, symplectic and nonsymplectic 4-manifolds.
\section{Constructions of exotic $4$-manifolds with nonnegative signatures from Cartwright-Steger surfaces}\label{sec:positive CS}
In this section, we will construct families of simply connected non-spin symplectic and smooth 4-manifolds with nonnegative signatures and small $\chi$. We consider the complex surface $\tilde{M}$ that we constructed from Cartwright-Steger surfaces in Section \ref{forM1}), with $c_1^2(\tilde{M})= 36$ and $e(\tilde{M})=12$. Using the formulas $\sigma = (c_1^2 - 2e)/3$ and $\chi = (e + \sigma)/4$, we have $\sigma(\tilde{M})=\chi(\tilde{M})=4$. Recall that by Lemma \ref{lemm3}, $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ contains a genus $5$ symplectic curve $\Sigma_5$ of self intersection $-2$ and $i_*\pi_1(\Sigma_5)\rightarrow \pi_1(\tilde{M}\#\overline{\mathbb{CP}}{}^{2})$ being a surjective homomorphism. In our construction of symplectic $4$-manifolds with nonnegative signatures, $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ along with $\Sigma_5$ will serve as our first building block. For our second building block, we will use the minimal, simply connected and symplectic 4-manifolds $X_{g,g+2}$ and $X_{g,g+1}$ \cite{AP2} (see also \cite{AS}) for which the following theorems hold:
\begin{theo}\label{theorem}
For any integer $g \geq 1$, there exist a minimal symplectic 4-manifold $X_{g,g+2}$ obtained via Luttinger surgery such that
\begin{itemize}
\item[(i)] $X_{g,g+2}$ is simply connected
\item[(ii)] $e(X_{g,g+2})= 4g+2$, $\sigma (X_{g,g+2}) = - 2$, $c_1^{2}(X_{g,g+2}) = 8g-2$, and\/ $\chi(X_{g,g+2}) = g$.
\item[(iii)] $X_{g,g+2}$ contains the symplectic surface $\Sigma$ of genus $2$ with self-intersection $0$ and $2$ genus $g$ surfaces with self-intersection $-1$ intersecting $\Sigma$ positively and transversally.
\end{itemize}
\end{theo}
\begin{theo}\label{thm3}
There exist a minimal symplectic 4-manifold $X_{g,g+1}$ obtained via Luttinger surgery such that
\begin{itemize}
\item[(i)] $X_{g,g+1}$ is simply connected
\item[(ii)] $e(X_{g,g+1})= 4g+1$, $\sigma (X_{g,g+1}) = - 1$, $c_1^{2}(X_{g,g+2}) = 8g-1$, and\/ $\chi(X_{g,g+1}) = g$.
\item[(iii)] $X_{g,g+1}$ contains the symplectic surface $\Sigma$ of genus $2$ with self-intersection $0$, genus $\Sigma_{g+1}$ symplectic surface with self-intersection $0$ intersecting $\Sigma$ positively and transversally.
\end{itemize}
\end{theo}
\begin{proof}
For the details of the constructions of $X_{g,g+2}$ and $X_{g,g+1}$, we refer the reader to \cite{AP2} and Section 5 of \cite{AS}.
\end{proof}
\subsection{Symplectic and smooth manifolds with $(\sigma,\chi)=(1,10)$}
To construct simply connected, symplectic and smooth $4$-manifolds with $(\sigma,\chi)=(1,10)$, we use $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ containing genus $5$ curve $\Sigma_5$ of self intersection $-2$ and $X_{2,4}$ in the notation of Theorem \ref{theorem}.
For the convenience of the reader, we briefly review the construction of $X_{2,4}$. Take a copy of $\mathbb{T}^2 \times \{pt\}$ and $\{pt\} \times \mathbb{T}^2$ in $\mathbb{T}^2 \times \mathbb{T}^2$ equipped with the product symplectic form, and symplectically resolve the intersection point of these dual symplectic tori. The resolution produces symplectic genus two surface of self intersection $+2$ in $\mathbb{T}^2 \times \mathbb{T}^2$. By symplectically blowing up this surface twice, in $\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}$, we obtain a symplectic genus 2 surface $\Sigma_2$ with self-intersection $0$, with two $-1$ spheres (i.e. the exceptional spheres resulting from the blow-ups) intersecting it positively and transversally. We also note that $\Sigma_2$ has a dual symplectic torus $\mathbb{T}^2$ of self intersection zero intersecting $\Sigma_2$ positively and transversally at one point. Next, we form the symplectic connected sum of $\mathbb{T}^4\#2\overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_2$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. By performing the sequence of $8$ appropriate $\pm 1$ Luttinger surgeries on $(\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_2)$, we obtain the symplectic $4$-manifold $X_{2,4}$.
It can be seen from the construction that there are genus $3$ curves of self intersections $0$ inside $X_{2,4}$. Each of them comes from the internal sum of the one of the punctured tori in $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and one of the punctured genus two surfaces in $\Sigma_2 \times \Sigma_2 \setminus \nu (\Sigma_2 \times \{pt\})$. Such a genus 3 curve of square zero intersects $\Sigma_2$ positively and transversally at one point. We symplectically resolve this intersection and obtain a genus $5$ surface $\Sigma_5'$ of square $+2$ in $X_{2,4}$.
Since the two symplectic building blocks $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ and $X_{2,4}$ contain symplectic genus $5$ surfaces of self intersections $-2$ and $+2$ respectively, we can form their symplectic connected sum along these surfaces $\Sigma_5$ and $\Sigma_5'$. Let
\begin{equation*}
M_{1,10} = (\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_5 = \Sigma_5'} X_{2,4}.
\end{equation*}
\begin{lemm}
$\sigma(M_{1,10}) = 1$, $\chi_h(M_{1,10}) =10$, $e(M_{1,10}) = 39$ and $c_1^2 (M_{1,10}) = 81$.
\end{lemm}
\begin{proof}
We have $\sigma(M_{1,10}) = \sigma(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \sigma(X_{2,4}) = 3+(-2) = 1$ and $\chi_h(M_{1,10}) = \chi(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \chi(X_{2,4}) + (5-1) = 4+2+4= 10$. Using the
formulas $c_1^2 = 3\sigma + 2e$ and $e = 4\chi - \sigma$, we compute $e(M_{1,10})$ and $c_1^2 (M_{1,10})$ as given.
\end{proof}
Let us now show that $M_{1,10}$ is an exotic copy of $19\mathbb{CP}^{2} \# 18\overline{\mathbb{CP}}^{2}$. Notice that $M_{1,10}$ is symplectic and simply connected, which follows from Gompf's Symplectic Connected Sum Theorem \cite{gompf} and Seifert-Van Kampen's Theorem respectively. Using Freedman's classification theorem for simply-connected $4$-manifolds and the lemma above, $M_{1,10}$ is homeomorphic to $19\mathbb{CP}^{2} \# 18\overline{\mathbb{CP}}^{2}$. Since $M_{1,10}$ is symplectic, by Taubes's theorem it has a non-trivial Seiberg-Witten invariant. Next, by appealing to the connected sum theorem for the Seiberg-Witten invariants, we deduce that the Seiberg-Witten invariant of $19{\mathbb{CP}}{}^{2}\#18\overline{\mathbb{CP}}{}^{2}$ is trivial. Thus, $M_{1,10}$ is not diffeomorphic to $19{\mathbb{CP}}{}^{2}\#18\overline{\mathbb{CP}}{}^{2}$. Furthermore, $M_{1,10}$ is a minimal symplectic $4$-manifold by Usher's Minimality Theorem \cite{usher}. Since symplectic minimality implies smooth minimality, $M_{1,10}$ is also smoothly minimal, and thus is smoothly irreducible \cite{HK}. As in Section~\ref{sec:positive}, by performing appropriate generalized torus surgeries, we realize infinitely many pairwise non-diffeomorphic, irreducible, symplectic and nonsymplectic $4$-manifolds (see \cite{AP2} for the further details of such construction).
\subsection{Symplectic and smooth manifolds with $(\sigma,\chi)=(0,9)$}
In this construction, our first building block is again $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ containing genus $5$ symplectic curve $\Sigma_5$ of self intersection $-2$. For our second building block, we use $X_{1,3}$ in the notation of Theorem \ref{theorem}.
Let us recall the construction of $X_{1,3}$. In constructing $X_{1,3}$, we first obtain a symplectic genus 2 surface $\Sigma_2$ with self-intersection $0$, with two $-1$ spheres intersecting it positively and transversally in $\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}$. In addition, there are symplectic tori $\mathbb{T}^2$ of self intersections zero each of which intersects $\Sigma_2$ positively and transversally once. Next, we form the symplectic connected sum of $\mathbb{T}^4\#2\overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_1$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. By performing the sequence of $6$ appropriate $\pm 1$ Luttinger surgeries on $(\mathbb{T}^4 \# 2 \overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_1)$, we obtain the symplectic $4$-manifold $X_{1,3}$. Therefore, we see that $X_{1,3}$ contains a symplectic surface $\Sigma_2$ with self intersection $0$ and two tori $T_{1}$ and $T_{2}$ with self intersections $-1$ which have positive and transverse intersections with $\Sigma_2$. Note that $T_{1}$ and $T_{2}$ result from the internal sum of the punctured exceptional spheres in $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and the punctured tori in $\Sigma_2 \times \Sigma_1 \setminus \nu (\Sigma_2 \times \{pt\})$. Moreover, there are genus 2 curves of self intersections 0 inside $X_{1,3}$. Each of them comes from the internal sum of the one of the punctured tori in $\mathbb{T}^4 \# 2\overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and one of the punctured tori in $\Sigma_2 \times \Sigma_1 \setminus \nu (\Sigma_2 \times \{pt\})$. Such a genus 2 curve $\Sigma_2'$ of square zero intersects $\Sigma_2$ positively and transversally at one point. We symplectically resolve the intersections of $\Sigma_2$ with $T_{1}$ and $\Sigma_2$ with $\Sigma_2'$. Thus we obtain a genus 5 surface $\Sigma_5$ of square $+3$ in $X_{1,3}$. By blowing up $\Sigma_5$ at one point, we obtain a genus $5$ surface $\Sigma_5'$ of square $+2$ in $X_{1,3}\# \overline{\mathbb{CP}}{}^{2}$.
Since the two symplectic building blocks $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ and $X_{1,3}\#\overline{\mathbb{CP}}{}^{2}$ contain symplectic genus $5$ surfaces of self intersections $-2$ and $+2$ respectively, we can form their symplectic connected sum along these surfaces $\Sigma_5$ and $\Sigma_5'$. Let
\begin{equation*}
M_{0,9} = (\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_5 = \Sigma_5'} (X_{1,3}\#\overline{\mathbb{CP}}{}^{2}).
\end{equation*}
\begin{lemm}
$\sigma(M_{0,9}) = 0$, $\chi(M_{0,9}) =9$, $e(M_{0,9}) = 36$ and $c_1^2 (M_{0,9}) = 72$.
\end{lemm}
\begin{proof}
We have $\sigma(M_{0,9}) = \sigma(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \sigma(X_{1,3}\#\overline{\mathbb{CP}}{}^{2}) = 3+(-3) = 0$ and $\chi(M_{0,9}) = \chi(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \chi(X_{1,3}\#\overline{\mathbb{CP}}{}^{2}) + (5-1) = 4+1+4= 9$. Consequently, we compute $e(M_{0,9})$ and $c_1^2 (M_{0,9})$ as given in the statement.
\end{proof}
As above, we show that $M_{0,9}$ is an exotic copy of $17\mathbb{CP}^{2} \# 17\overline{\mathbb{CP}}^{2}$ and $M_{0,9}$ is also smoothly irreducible. As in Section~\ref{sec:positive}, by performing generalized torus surgeries, we realize infinitely many pairwise non-diffeomorphic, irreducible, symplectic and nonsymplectic $4$-manifolds (see \cite{AP2} for the further details of such construction).
\subsection{Symplectic and smooth manifolds with $(\sigma,\chi)=(2,10)$}\label{2,10}
In this case, the first symplectic building blocks is $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ along the genus $5$ curve $\Sigma_5$ of self intersection $-2$. Our the second symplectic building block is $X_{2,3}$ in the notation of Theorem \ref{thm3}, which was constructed in \cite{AP2}.
Let us recall the construction of $X_{2,3}$. We take a copy of $\mathbb{T}^2 \times \{pt\}$ and the braided torus $T_{\beta}$ representing the homology class $2[\{pt\} \times \mathbb{T}^2]$ in $\mathbb{T}^2 \times \mathbb{T}^2$ (see ~\cite{AP2}, page 4 for the construction of $T_{\beta}$). The tori $\mathbb{T}^2 \times \{pt\}$ and $T_{\beta}$ intersect at two points. We symplectically blow up one of these intersection points, and symplectically resolve the other intersection point to obtain the symplectic genus two surface of self intersection $0$ in $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$ (see ~\cite{AP2}, pages 3-4). The symplectic genus $2$ surface $\Sigma_2$ has a dual symplectic torus $\mathbb{T}^2$ of self intersections zero intersecting $\Sigma_2$ positively and transversally at one point. We form the symplectic connected sum of $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_2$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. By performing the sequence of $4$ appropriate $\pm 1$ Luttinger surgeries on $(\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_2)$, we obtain the symplectic $4$-manifold $X_{2,3}$ constructed in \cite{AP2}. It can be seen from the construction that, $X_{2,3}$ contains a symplectic surface $\Sigma_3$ with self intersection $0$, resulting from the internal sum of the punctured torus
in $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and one of the punctured genus two surfaces in $\Sigma_2 \times \Sigma_2 \setminus \nu (\Sigma_2 \times \{pt\})$. $\Sigma_3$ intersects $\Sigma_2$ positively and transversally at one point. (The reader may see Section 5.3 and Figure 7 in \cite{AS} showing the construction steps for a similar case.) We now symplectically resolve their intersection which gives genus five surface $\Sigma_5'$ of self intersection $+2$ in $X_{2,3}$.
Let
\begin{equation*}
M_{2,10} = (\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_5 = \Sigma_5'} (X_{2,3}).
\end{equation*}
\begin{lemm}\label{l2}
$\sigma(M_{2,10}) = 2$, $\chi_h(M_{2,10}) =10$, $e(M_{2,10}) = 38$ and $c_1^2 (M_{2,10}) = 82$.
\end{lemm}
\begin{proof}
We have $\sigma(M_{2,10}) = \sigma(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \sigma(X_{2,3}) = 3+(-1) = 2$ and $\chi_h(M_{2,10}) = \chi(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \chi(X_{2,3}) + (5-1) = 4+2+4= 10$. Consequently, we compute $e(M_{2,10})$ and $c_1^2 (M_{2,10})$.
\end{proof}
Similarly, using the Lemma \ref{l2} and the above mentioned theorems, we see that $M_{2,10}$ is an exotic copy of $19\mathbb{CP}^{2} \# 17\overline{\mathbb{CP}}^{2}$, which is smoothly irreducible.
\subsection{Symplectic and smooth manifolds with $(\sigma,\chi)=(1,9)$}
Similar to the previous cases, we use $\tilde{M} \#\overline{\mathbb{CP}}{}^{2}$ containing genus 5 curve $\Sigma_5$ of self intersection $-2$, and $X_{1,2}\#\overline{\mathbb{CP}}{}^{2}$ in the notation of Theorem \ref{thm3}, constructed in \cite{AP2}.
To construct $X_{1,2}$, we first obtain a symplectic genus two surface of self intersection $0$ in $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$ as follows. Let us take a copy of $\mathbb{T}^2 \times \{pt\}$ and the braided torus $T_{\beta}$ representing the homology class $2[\{pt\} \times \mathbb{T}^2]$ in $\mathbb{T}^2 \times \mathbb{T}^2$. The tori $\mathbb{T}^2 \times \{pt\}$ and $T_{\beta}$ intersect at two points. We symplectically blow up one of these two intersection points, and symplectically resolve the other intersection point to obtain the symplectic genus two surface $\Sigma_2$ of self intersection $0$ in $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$. Note that the exceptional sphere $S^2$ intersects $\Sigma_2$ positively and transversally twice. Next, we form the symplectic connected sum of $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}$ with $\Sigma_2 \times \Sigma_1$ along the genus two surfaces $\Sigma_2$ and $\Sigma_2 \times \{pt\}$. By performing the sequence of $6$ appropriate $\pm 1$ Luttinger surgeries on $(\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2}) \#_{\Sigma_2 = \Sigma_2 \times \{pt\}} (\Sigma_2 \times \Sigma_1)$, we obtain the symplectic $4$-manifold $X_{1,2}$. It was shown in \cite{AP2}, $X_{1,2}$ is an exotic copy of $\mathbb{CP}^{2} \# 2\overline{\mathbb{CP}}^{2}$. Observe that as a result of the internal sum of the twice punctured sphere $S^2$ in $\mathbb{T}^4 \# \overline{\mathbb{CP}}^{2} \setminus \nu (\Sigma_2)$ and the twice punctured tori in $\Sigma_2 \times \Sigma_1 \setminus \nu (\Sigma_2 \times \{pt\})$, we acquire a symplectic genus $2$ surface of self intersection $-1$ in $X_{1,2}$ intersecting $\Sigma_2$ positively and transversally twice. We symplectically resolve the two intersections and get symplectic genus $5$ surface of square $+3$ in $X_{1,2}$. We blow up this surface at one point and obtain symplectic genus $5$ surface $\Sigma_5'$ of self intersection $+2$ in $X_{1,2}\#\overline{\mathbb{CP}}{}^{2}$.
Let us define
\begin{equation*}
M_{1,9} = (\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) \#_{\Sigma_5 = \Sigma_5'} (X_{1,2}\#\overline{\mathbb{CP}}{}^{2}).
\end{equation*}
\begin{lemm}\label{l3}
$\sigma(M_{1,9}) = 1$, $\chi_h(M_{1,9}) =9$, $e(M_{1,9}) = 35$ and $c_1^2 (M_{1,9}) = 73$.
\end{lemm}
\begin{proof}
We have $\sigma(M_{1,9}) = \sigma(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \sigma(X_{1,2}\#\overline{\mathbb{CP}}{}^{2}) = 3+(-2) = 1$ and $\chi (M_{1,9}) = \chi(\tilde{M} \#\overline{\mathbb{CP}}{}^{2}) + \chi(X_{1,2}\#\overline{\mathbb{CP}}{}^{2}) + (5-1) = 4+1+4= 9$. Consequently, we compute $e(M_{1,9})$ and $c_1^2 (M_{1,9})$ as given.
\end{proof}
Similarly, using the Lemma \ref{l3} and the above mentioned theorems, we show that the minimal symplectic $4$-manifold $M_{1,9}$ is an exotic copy of $17\mathbb{CP}^{2} \# 16\overline{\mathbb{CP}}^{2}$.
\begin{rem} In this remark, we discuss how to obtain a minimal symplectic $4$-manifold with the fundamental group $\mathbb{Z}_{2}$ and the invariants $(\sigma,\chi)=(0,8)$. Since $e = 4\chi - \sigma = 32$, such a symplectic $4$-manifold yields to a homology $15\mathbb{CP}^{2} \# 15\overline{\mathbb{CP}}^{2}$ with $\pi_{1} \cong \mathbb{Z}_{2}$. Since the covering group of the complex surface $M$ (see Proposition 1) is $\mathbb{Z}_2\times \mathbb{Z}_2$, it has a degree two unramified covering. Let us consider the normal unramified covering $M_2$ of $M$ with covering group given by index two subgroup $H'$ of $\pi_1(M)$. Let $p: M_2\rightarrow M$ be the covering map. Notice that in this case the pull-back of $D$ under this $\mathbb{Z}_2$ covering is not isomorphic to the fundamental group of the ambient manifold, but rather a normal subgroup of index $2$. Using the symplectic pair $(M_{2}\#\overline{\mathbb{CP}}{}^{2}, \Sigma_5)$ instead of $(\tilde{M}\#\overline{\mathbb{CP}}{}^{2}, \Sigma_5)$, and $(X_{2,3}, \Sigma_5')$ in our above constructions (see \ref{2,10}) leads to a minimal symplectic $4$-manifold with $(\sigma,\chi)=(0,8)$ and $\pi_{1} \cong \mathbb{Z}_{2}$. Previously no such examples were known.
\end{rem}
\section*{Acknowledgments}
The first author was partially supported by a Simons Research Fellowship and Collaboration Grant for Mathematicians from the Simons Foundation. He would like to thank the Departments of Mathematics at Purdue and at Harvard Universities for their hospitality, where part of this work was completed. The second author would like to thank Max Planck Institute for Mathematics in Bonn for its support and hospitality. The third author is partially supported by NSF grant DMS-$1501282$. All authors would like to thank Donald Cartwright for his help related to Magma computations.
|
\section{Introduction and preliminaries}
Informally, the \emph{cutoff phenomenon} is an abrupt transition of a Markov chain to its equilibrium when the system under consideration is sufficiently large (see Section \ref{subsection 1.2} for a rigorous definition).
To the author's knowledge, the first rapid mixing result appeared in \cite{diaconis} on the symmetric group while considering random transpositions.
Shortly afterward, \textcite{aldous} showed that the top-in-at-random card-shuffle precisely exhibits a cutoff phenomenon, initiating the whole industry of the cutoff phenomenon.
As pointed out in \cite{Lubetzky_2012}, only a few examples of cutoff were known regarding the Glauber dynamics of the Ising model (see Section \ref{subsection1.1} for formal definitions), such as that of \cites{ding}{levin} on complete graphs and of \cites{Lubetzky_2012}{lubetzky2012cutoff}{lubetzky2014} on lattices.
Recent researches have mainly focused on lattices.
A breakthrough paper by \textcite{Lubetzky_2012} showed cutoff with a continuous-time window $O(\ln\ln n)$ for this longstanding problem.
An improvement on the window size to optimal $O(1)$ was made by the same authors in \cite{lubetzky2014} with the information percolation framework.
By the same technique, the authors illustrated the existence of cutoff in high enough temperatures for the Ising model of any sequence of graphs with a bounded degree in \cite{lubetzky2017}.
Mean-field Potts model on complete graphs was comprehensively explored in \cite{potts}, again verifying the cutoff phenomenon in high temperatures.
For the bipartite Potts model, \textcite{Yevgeniy} proved the cutoff phenomena in the high temperatures using their aggregate path coupling method.
The purpose of this paper is to investigate the Glauber dynamics for the Ising model on complete multipartite graphs.
(Exact definitions are given in the rest of the introduction.)
Indeed, we identify the critical temperature and establish cutoff in the high temperature regime.
On the other hand, exponentially slow mixing is established in the low temperature regime.
The significance of our setting is that complete multipartite graphs have an intermediate geometry between the complete graphs which have no geometry at all (e.g. \cite{levin}), and lattices which have a strong geometry (e.g. \cite{Lubetzky_2012}).
Thus, our result serves as a midway example between those two extreme cases.
The method of proof hinges on generalizations of the tools in \cite{levin}, notably the two-coordinate chain thereof.
Due to the nature of complete multipartite graphs, our model can be considered as a block spin Ising model with no interaction inside each block.
Such mean-field block models naturally occur in statistical physics when modelling metamagnets (see \cite{KINCAID197557}) and in studies on social interactions (see, e.g., \cite{Gallo}).
A recent paper by \textcite{Knöpfel2020} contains an excellent introduction to this line of work.
When it comes to cutoff phenomenon on finite graphs, it is easy to convert the discrete-time results to that of the continuous-time and vice versa.
Hence, we only consider discrete-time chains.
\subsection{Notations}
Boldface letters are used to denote vectors or matrices.
Inequalities between vectors and matrices are defined element-wise.
The dependence of any quantities on the number of vertices $n$ is understood throughout the paper.
Some important quantities not depending on $n$ will be explicitly mentioned.
We will write $\mathbf e_j$ to be the $j$th vector in the standard basis of $\mathbb R^m$.
The lower case $t$ will always denote time.
Let $\circ$ denote the Hadamard product between matrices.
More precisely, $B\circ C = (B_{ij} C_{ij})$ whenever $B=(B_{ij})$ and $C=(C_{ij})$ are matrices with the same dimensions.
\subsection{Ising model and Glauber dynamics}\label{subsection1.1}
Let $G=(V,E)$ be a finite graph with the vertex set $V$ and the edge set $E$.
Elements of $\Omega \colonequals \{\pm 1\}^V$ are called \emph{configurations}.
In the absence of external fields, the \emph{Ising model} on $G$ is a distribution $\mu$ called the \emph{Gibbs distribution} on $\Omega$ given by \[\mu (\sigma) \colonequals \frac{e^{-\beta H(\sigma)}}{Z(\beta) }\] where $\sigma \in \Omega$, $\beta \geq 0$, $H(\sigma)=- \sum_{ij \in E}h_{ij}\sigma(i) \sigma(j)$, and $Z(\beta)$ is a normalizing factor.
Assuming an isotropic interaction strength between the vertices, we set $h_{ij}=1/|V|$.
The physical interpretation of $H(\sigma)$ is the energy of the whole spin system with the configuration $\sigma$.
We call each $\sigma(v)$ the \emph{spin} at \emph{site} $v$.
The \emph{Glauber dynamics} for the Ising model is a reversible Markov chain with respect to the Gibbs distribution satisfying the following rule.
At each time, choose a site uniformly at random in $V$ and update the spin at the chosen site according to $\mu$ conditioned on the set of configurations having the same spins at all the sites except the chosen one.
The Glauber dynamics for the Gibbs distribution $\mu$ is irreducible, aperiodic, and reversible with $\mu$ as its unique stationary distribution.
For the Ising model, it is easy to see that the probability of updating to $\pm 1$ at the chosen site $v$ is $r_{\pm}(S)$ where \begin{gather}
r_{\pm}(x) \colonequals \frac{e^{\pm \beta x}}{ e^{\beta x}+e^{-\beta x}} = \frac{1 \pm \tanh (\beta x)}{2}; \quad x\in\mathbb R \label{r+}
\end{gather} and $S=\sum_{vv' \in E} \sigma(v')/|V|$ is the mean-field at $v$.
\subsection{Markov chain mixing and cutoff phenomenon} \label{subsection 1.2}
The \emph{total variation distance} between two probability measures $\nu_1$ and $\nu_2$ on $\Omega$ is defined by \[\|\nu_1 -\nu_2\|_{TV} \colonequals \sup_{A\subseteq \Omega}|\nu_1(A)-\nu_2(A)|= \frac{1}{2}\sum_{x\in \Omega}|\nu_1(x)-\nu_2(x)|.\]
The total variation distance is half of the $L^1$-distance between the probability measures.
Let $({\sigma_{t}})$ be the Markov chain of the Glauber dynamics for the Ising model.
Define the worst-case total variation distance of the chains to the stationary distribution $\mu$ at time $t$ by \[d(t) \colonequals \max_{\sigma \in \Omega} \|\mathbb P_{\sigma }({\sigma_{t}} \in \cdot) -\mu\|_{TV}\] where here and thereafter $\mathbb P_{\sigma}$ denotes the probability given $\sigma_0 =\sigma$.
The \emph{mixing time} is defined by \[t_{\mathrm{mix}}(\varepsilon) \colonequals \min\{t : d(t) \leq \varepsilon\}; \quad \varepsilon \in (0,1).\]
We say a sequence of Markov chains with corresponding mixing times $t_{\mathrm{mix}}^{(n)}(\varepsilon)$ exhibit a \emph{cutoff phenomenon} if for every $0<\varepsilon<1/2$, \[\lim_{n \to \infty} \frac{t_{\mathrm {mix}}^{(n)}(\varepsilon)}{t_{\mathrm mix}^{(n)}(1-\varepsilon)}=1.\]
Furthermore, we say that the cutoff occurs at $t_{\mathrm {mix}}^{(n)}$ with \emph{window size} $O(w_n)$ if $w_n= o(t_{\mathrm {mix}}^{(n)})$ and \begin{gather*}
\lim_{\gamma \to \infty }\liminf_{n \to \infty }d_n(t_{\mathrm {mix}}^{(n)}-\gamma w_n)=1,\quad \lim_{\gamma \to \infty }\limsup_{n \to \infty }d_n(t_{\mathrm {mix}}^{(n)}+\gamma w_n)=0.
\end{gather*}
\subsection{Magnetization chain on complete multipartite graphs}
Now, we are in a place to consider a complete $m$-partite graph, a graph whose vertices are partitioned into $m$ different independent sets, and every pair of vertices from different independent sets is connected by an edge.
Each edge represents an interaction between the vertices.
Denote this graph by $K_{np_1,np_2,\dots, np_m}$ which has $n$ vertices and $m$ partitions where $\sum_{i=1}^m p_i =1$ and $p_i>0$ for $i=1,2,\dots, m$.
We fix the parameters $m$ and $p_i$'s hereafter.
Without loss of generality, we assume $p_1 \leq p_2 \leq \dots \leq p_m$.
We may also assume that $np_i \in \mathbb N$ for every $i$ so that $K_{np_1,np_2,\dots, np_m}$ is well defined whenever such considerations are required.
Let $V = \bigcup_{i=1}^m J_i$ be the set of all vertices where $J_i$ denotes the set of the $i$th partition of the vertices.
Note $np_i = |J_i|$.
We define $\Omega_i \colonequals \{\pm 1\}^{J_i}$ for $i=1,\dots,m$ so that $\Omega= \prod_{i=1}^m \Omega_i$ is our configuration space.
Each configuration $\sigma \in \Omega$ has a unique representation $(\sigma^{(1)},\dots,\sigma^{(m)}) \in \prod_{i=1}^m \Omega_i$ and both representations are understood throughout this paper.
For each $\sigma \in \Omega$, define the \textit{magnetization} on $J_i$ by $S^{(i)}(\sigma) \colonequals \sum_{v \in J_i} \sigma(v) /n$, $i=1,\dots,m$.
For the Markov chain $(\sigma_t)_{t \geq 0}=(\sigma_t^{(1)},\dots,\sigma_{t}^{(m)})_{t\geq 0}$ starting at $\sigma=(\sigma^{(1)},\dots,\sigma^{(m)}) \in \prod_{i=1}^m \Omega_i$, we define the corresponding magnetization on $J_i$ by \[S_{t}^{(i)} \colonequals \frac{1}{n}\sum_{v\in V_i}\sigma_{t}^{(i)}(v) \ \text{for}\ i\in \{1,\dots,m\}, \ t\geq 0.\]
We sometimes use the vector notation $\mathbf S_t \colonequals (S^{(1)}_t, \dots, S^{(m)}_{t})$ for $t\geq 0$.
We call the process $(\mathbf S_t)_{t\geq 0}$ a \emph{magnetization chain}.
\thref{magmarkov} shows that $(\mathbf S_t)_{t\geq 0}$ is in fact a Markov chain.
Note that it is a projection of the whole Markov chain $({\sigma_{t}})_{t\geq 0}$, so mixing of the whole chain $({\sigma_{t}})_{t\geq 0}$ implies the mixing of the chain $(\mathbf S_t)_{t\geq 0}$.
Our aim is to show the converse in a certain sense.
\subsection{Main results}
Given the above definitions and notations, our main result establishes the cutoff phenomenon on complete multipartite graphs.
\begin{thm}[Main result]
For $m\in\mathbb N$ and $p_i>0$ such that $\sum_{i=1}^m p_i =1$, the Glauber dynamics for the Ising model on the complete multipartite graph $K_{np_1,\dots, np_m}$ exhibits a cutoff at $\frac{1}{2(1-\beta/\beta_{cr})} n\ln n$ with window size $O(n)$ in the high temperature regime $\beta<\beta_{cr}$ where $\beta_{cr}=\beta_{cr}(p_1,\dots,p_m)$ is a constant defined in equation \eqref{betacrdef}.
\end{thm}
\begin{thm}\thlabel{lowtempresult}
In the low temperature regime $\beta > \beta_{cr}$, the dynamics is exponentially slow mixing, i.e., $t_{\mathrm{mix}}\geq C_1 \exp(C_2 n)$ for some constants $C_1,\ C_2 >0$ not depending on $n$.
\end{thm}
A few remarks are in order.
Our main result is obtained as a consequence of \thref{upperbound} and \thref{lowerbound}.
In the low temperature regime $\beta>\beta_{cr}$, the mixing time is exponentially slow, therefore identifying the critical temperature $\beta_{cr}$.
In the $m=1$ case, there are no spin interactions so the chain is equivalent to the lazy random walk on an $n$-dimensional hypercube, which has a cutoff at $(n\ln n) /2 $ with window size $O(n)$ (see \parencite[]{hypercube} or \parencite[Chapter 18]{Peres}).
This result can be seen as a consequence of our main result since $m=1$ implies $\beta_{cr}=\infty$ (see equation \eqref{betacrdef}).
\subsection{Organization of the article}
As mentioned earlier, our proof is based on the ideas of \textcite{levin}.
We assume high temperatures until Section \ref{section6}.
We first observe that the magnetization chain is a Markov chain in its own right (\thref{magmarkov}).
A suitable scaling of the magnetization chain leads to a contraction property (\thref{normcontraction}).
This in turn gives a uniform variance bound of magnetizations in time (Sections \ref{section2} and \ref{section3}).
In Section \ref{section4}, we construct a coupling of the magnetization chain so that it couples in $\frac{1}{2(1-\beta/\beta_{cr})} n\ln n+O(n)$ steps with high probability.
After the magnetization coupling phase, by considering the "$2m$-coordinate chain" inspired by \cite{levin}, we can construct a post magnetization coupling to reach the full-mixing in another $O(n)$ steps.
This proves the upper bound (\thref{upperbound}).
We construct a suitable distinguishing-statistic of the magnetization chain \cite[see][Chapter 7.3]{Peres} to obtain the lower bound (\thref{lowerbound}).
These upper and lower bound results establish the cutoff in the high temperature regime.
Exponentially slow mixing in the low temperature regime is shown in Section \ref{section6}.
\section{Contraction of the magnetization chain in high temperatures}\label{section2}
We describe the \emph{monotone coupling}.
Let $I$ and $U$ be independent uniform random variables over $V$ and $[0,1]$, respectively.
We consider the collection of Markov chains with starting configurations $\sigma \in \Omega$.
Simultaneously define the next configurations at time $t=1$ by \[\sigma_{1} (i)= \begin{cases} \sigma (i) & \text{if $I \neq i$} \\ \mathbbm 1_{U < r_+(\sum_{j\neq k}S^{(j)}(\sigma))} -\mathbbm 1_{U \geq r_+(\sum_{j\neq k}S^{(j)}(\sigma))} & \text{if $I=i \in J_k$} \end{cases} \quad \]
where $r_+$ is defined in equation \eqref{r+}.
Repeat this procedure independently for each time.
It is clear that each Markov chain $(\sigma_t)_{t \geq 0}$ above is a version of the Glauber dynamics on the complete multipartite graph with starting state $\sigma$'s, defined on a common probability space.
The above coupling is called a \emph{monotone coupling} in the sense that if $\sigma \leq \tilde \sigma$ are starting states for $(\sigma_t)_{t\geq 0}$ and $(\tilde \sigma _t)_{t\geq 0}$, respectively, then $S^{(i)}(\sigma) \leq S^{(i)} (\tilde \sigma )$ for $i=1,\dots,m$ so that $\sigma_{1} \leq \tilde \sigma_{1}$, and $\sigma_t \leq \tilde \sigma_t$ for any $t\geq 0$ accordingly.
Define \[\mathcal S \colonequals \prod_{i=1}^m\{-p_i, -p_i+2/n,\dots, p_i\}.\]
\begin{prop}[Magnetization chain]\thlabel{magmarkov}
The process $(S^{(1)}_t, \dots, {S^{(m)}_{t}})_{t \geq 0}$ is a \\Markov chain on the magnetization state space $\mathcal S$.
\end{prop}
\begin{proof}
Note that \begin{align*}
\mathbb P\bigl(({S^{(1)}_{t+1}},\dots, {S^{(m)}_{t+1}} )=(S^{(1)}_{t} -\frac{2}{n},\dots, {S^{(m)}_{t}})\bigr)&= p_1 \frac{n{S^{(1)}_{t}} +|J_1|}{2|J_1|} r_- \Bigl(\sum_{j\neq 1}{S^{(j)}_{t}}\Bigr)
\\&=\frac{p_1+{S^{(1)}_{t}}}{2} r_-\Bigl(\sum_{j\neq 1}{S^{(j)}_{t}}\Bigr)\end{align*} is measurable with respect to the $\sigma$-algebra generated by $({S^{(1)}_{t}},\dots, {S^{(m)}_{t}})$.
Other cases can be dealt with similarly.
\end{proof}
\begin{rmk}
By symmetry, $({S^{(1)}_{t}},\dots, {S^{(m)}_{t}})$ starting from $\sigma$ and $(-{S^{(1)}_{t}},\dots, -{S^{(m)}_{t}})$ starting from $-\sigma$ have the same distributions.
This can also be seen by the physical fact that the map $\sigma \mapsto -\sigma$ just corresponds to flipping the reference axis to which we are measuring the spins of each site.
This does not change the dynamics of the spin system.
\end{rmk}
\begin{defi}[Hamming distance]
For two configurations $\sigma$ and $\sigma '$, denote the Hamming distance by $\mathrm{dist}(\sigma, \sigma') \colonequals \frac{1}{2}\sum_{k \in V} |\sigma(k)-\sigma'(k)|$.
\end{defi}
\begin{rmk}
This is a metric on $\Omega$, which is equal to the number of sites with different spins for two configurations.
Similarly, we can define $\mathrm{dist}_i$ on $\Omega_i$, respectively, but $\mathrm{dist}_i$'s merely satisfy the triangle inequality.
\end{rmk}
\begin{lem}[Contraction in mean for monotone coupling]\thlabel{monotonecontraction}
For a monotone coupling $({\sigma_{t}}, {\sigma'_{t}} )_{t \geq 0}$ starting at $(\sigma,\sigma') = ((\sigma^{(1)},\dots,\sigma^{(2)}), (\sigma'^{(1)},\dots,\sigma'^{(2)} ))$, we have \[\begin{pmatrix}\mathbb E \mathrm{dist}_1({\sigma^{(1)}_t},{\sigma'^{(1)}_t})\\ \vdots \\ \mathbb E\mathrm{dist}_m({\sigma^{(m)}_{t}},{\sigma'^{(m)}_t}) \end{pmatrix} \leq \mathbf A^t \begin{pmatrix} \mathrm{dist}_1(\sigma^{(1)}, \sigma_0') \\ \vdots \\ \mathrm{dist}_m(\sigma^{(m)},\sigma'^{(m)}) \end{pmatrix}\] where \[\mathbf A=\mathbf A_n \colonequals \begin{pmatrix}a&b_1&b_1&\dots&b_1
\\ b_2 & a&b_2 &\dots&b_2
\\b_3&b_3&a&\dots&b_3
\\\vdots &\dots&&&\vdots
\\b_m &\dots&&\dots&a
\end{pmatrix}\] with
$a\colonequals 1-1/n$, $b_k \colonequals p_k {\beta }/{n}$.
\end{lem}
\begin{proof}
Assume $d(\sigma,\sigma')=1$ with $-1=\sigma(v) = -\sigma'(v)$ for some vertex $v$.
Note $\sigma \leq \sigma'$.
Since we are considering a monotone coupling, it holds that for each $i=1,\dots,m$, \begin{align*}
\mathrm{dist}_i({\sigma^{(i)}_{1}},{\sigma'^{(i)}_{1}})= \mathbbm 1_{v\in J_i}(1-\mathbbm 1_{I=v}) + \mathbbm 1_{v \notin J_i}(\mathbbm 1_{I\in J_i} \mathbbm 1_{B_i})
\end{align*} where \[B_i=\Biggl\{r_+\biggl(\sum_{l\neq i}S_l(\sigma)\biggr)\leq U < r_+\biggl(\sum_{l\neq i}S_l(\sigma')\biggr)\Biggl\}.\]
Note that \begin{align*}
\mathbb P(B_i) &=\frac{1}{2}\Biggl( \tanh\biggl(\beta \sum_{l\neq i}S_l(\sigma')\biggr) - \tanh\biggl(\beta \sum_{l\neq i}S_l(\sigma)\biggr)\Biggr)
\\&=\frac{1}{2}\Biggl( \tanh\biggl(\beta \biggl(\sum_{l\neq i}S_l(\sigma)+\frac{2}{n}\biggr)\biggr) - \tanh\biggl(\beta \sum_{l\neq i}S_l(\sigma)\biggr)\Biggr) \mathbbm1_{v \notin J_i}
\\ &\leq \tanh \frac{\beta}{n} \mathbbm1_{v \notin J_i}.
\end{align*}
Since $I$ and $U$ are independent, for $i=1,\dots,m$, \begin{align*} \mathbb E\mathrm{dist}_i({\sigma^{(i)}_{1}},{\sigma'^{(i)}_{1}}) \leq \mathbbm 1_{v\in J_i}(1-\frac{1}{n})+\mathbbm 1_{v \notin J_i} p_i \tanh \frac{\beta}{n}.
\end{align*}
Suppose $\mathrm{dist}(\sigma,\sigma')= k >1$.
There exists $\sigma^0 \colonequals \sigma$, $\sigma^{1}$, $\dots$, $\sigma^k \colonequals \sigma'$ such that $\mathrm{dist}(\sigma^i, \sigma^{i+1})=1$.
By the triangular inequality for $\mathrm{dist}_i$ and the fact $\tanh(\beta /n) \leq \beta/n$, \begin{align*}
\mathbb E\mathrm{dist}_i({\sigma^{(i)}_{1}},{\sigma'^{(i)}_{1}}) \leq (1-\frac{1}{n})\mathrm{dist}_i(\sigma^{(i)},\sigma'^{(i)})+ p_i \frac{\beta}{n} \sum_{ l\neq i }\mathrm{dist}_l(\sigma^{(l)},\sigma'^{(l)}).
\end{align*}
Furthermore, by the Markov property,
\begin{multline*}
\mathbb E[\mathrm{dist}_i ({\sigma^{(i)}_{t+1}}, \sigma'^{(i)}_{t+1})| {\sigma_{t}}, {\sigma'_{t}} ] \leq (1-\frac{1}{n})\mathrm{dist}_i({\sigma^{(i)}_{t}}, {\sigma'^{(i)}_{t}}) + \frac{p_i \beta}{n} \sum_{l\neq i }\mathrm{dist}_l({\sigma^{(l)}_{t}},{\sigma'^{(l)}_{t}}).
\end{multline*}
By taking expectation and putting $x_{i,t} \colonequals \mathbb E\mathrm{dist}_i({\sigma^{(i)}_{t}}, {\sigma'^{(i)}_{t}})$, we have
\begin{align*}
\begin{pmatrix} x_{1,t} \\ \vdots \\ x_{m,t} \end{pmatrix} \leq \mathbf A \begin{pmatrix} x_{1,t-1} \\ \vdots \\ x_{m,t-1} \end{pmatrix}.
\end{align*}
Iterating gives \[ \begin{pmatrix} x_{1,t} \\ \vdots \\ x_{m,t} \end{pmatrix} \leq \mathbf A^t \begin{pmatrix} \mathrm{dist}_1(\sigma^{(1)},\sigma'^{(1)}) \\ \vdots \\ \mathrm{dist}_m(\sigma^{(m)},\sigma'^{(m)}) \end{pmatrix}.\]
\end{proof}
From now on, $\mathbf A$ (which depends on the number of vertices $n$) always denotes the matrix defined in \thref{monotonecontraction}.
Note that $\mathbf A$ is a positive matrix, so by the Perron-Frobenius theorem, there exists the largest eigenvalue $g=g_n>0$ with the left eigenvector $\mathbf a^T\colonequals (a_1, \dots, a_m) >\mathbf 0$ normalized in $l^1$ norm. Note that $g$ has algebraic multiplicity $1$ (see \parencite[Section 8.2]{Meyer} for a proof), so $\mathbf a^T$ is unique.
We fix the following notations \begin{gather}
\upsilon \colonequals n(1-g) \text{ and }\label{upsilondef}
\\ \beta_{cr} \colonequals \frac{1}{(m-1)\sum_{i=1}^m a_ip_i} \label{betacrdef}
\end{gather}
where $g$ and $(a_1,\dots,a_m)$ are defined in the previous paragraph.
Another characterization of $\beta_{cr}$ is given in \thref{slowmixinglemma}.
Insuk Seo commented\footnote{personal communication} that it can also be characterized as the threshold value of $\beta$ that makes $\mathbf K$ positive definite where $\mathbf K$ is defined through the equation $\mathbf A=\mathbf I-\frac{1}{n}\mathbf K$, $\mathbf I$ being the $m$-by-$m$ identity matrix.
\thref{upsilon} connects the quantities $\upsilon$ and $\beta_{cr}$.
\begin{prop}\thlabel{upsilon}
The left eigenvector $\mathbf a^T$ only depends on $p_1,\dots,p_m$.
Moreover, $\upsilon$ only depends on $p_1,\dots,p_m,$ and $\beta$ through the following equation: \[\upsilon = 1-\beta(m-1)\sum_{i=1}^ma_ip_i.\]
Therefore, $\beta_{cr}$ only depends on $p_1,\dots,p_m$, and we have $\upsilon = 1-\beta/\beta_{cr}$.
\end{prop}
\begin{proof}
Since $g$ satisfies \begin{align*}
0&=(n/\beta)^m\det(\mathbf A-gI)=\det (n \mathbf A/\beta -n gI/\beta) \\&= \det \begin{pmatrix}\frac{\upsilon-1}{\beta}&p_1&p_1&\dots&p_1
\\ p_2 & \frac{\upsilon-1}{\beta}&p_2 &\dots&p_2
\\p_3&p_3&\frac{\upsilon-1}{\beta}&\dots&p_3
\\\vdots &\dots&&&\vdots
\\p_m &\dots&&\dots&\frac{\upsilon-1}{\beta}
\end{pmatrix},
\end{align*}
it holds that $(\upsilon-1)/\beta$ is a root of a polynomial with coefficients only depending on $p_1,\dots,p_m$.
Since $\mathbf a$ is in the kernel of the transpose of the above matrix, it only depends on $p_1,\dots,p_m$.
Finally, $g= \|\mathbf A^T \mathbf a\|_1= 1-1/n+\frac{\beta}{n}(m-1) \sum_{k}a_ip_i$ implies $\upsilon =1 -\beta (m-1)\sum_i a_ip_i$.
\end{proof}
We collect further properties of the matrix $\mathbf A$ and its left eigenvector $\mathbf a^T$ in the next two lemmas.
\begin{lem}\thlabel{finalfinallemma}
We have \[a_1\geq \dots\geq a_m \ \text{and}\ \sum_{i=1}^ma_ip_i \leq \frac{1}{m}.\]
The equality in the latter holds if and only if $p_1=\dots=p_m$.
\end{lem}
\begin{proof}
Recall that we assumed $p_1\leq \dots \leq p_m$.
We claim that $a_1\geq \dots\geq a_m$.
To that end, fix $i<j$.
From $\mathbf a^T\mathbf A=g\mathbf a^T$, we have $(1-\frac{1}{n})a_i +\frac{\beta}{n}\sum_{k\neq i}a_kp_k-ga_i=0=(1-\frac{1}{n})a_j +\frac{\beta}{n}\sum_{k\neq j}a_kp_k-ga_j$.
Then $(1-\frac{1}{n}-g-\frac{\beta p_i}{n})a_i=(1-\frac{1}{n}-g-\frac{\beta p_j}{n})a_j$, i.e., $(\beta p_i +1-\upsilon )a_i=(\beta p_j +1-\upsilon )a_j$.
Thus, $p_i\leq p_j$ implies $a_i\geq a_j$, proving the claim.
By Chebyshev's sum inequality, since $a_i\geq a_j$ and $p_i\leq p_j$ whenever $i<j$, \begin{align*}
\sum_{i=1}^ma_ip_i \leq \frac{1}{m} \left(\sum_{i=1}^m a_i\right)\left(\sum_{i=1}^m p_i\right)=\frac{1}{m}.
\end{align*}
The equality holds if and only if $a_1= \dots =a_m$ or $p_1= \dots =p_m$.
The proof is now complete by noticing the fact that $(\beta p_i +1-\upsilon )a_i=(\beta p_j +1-\upsilon )a_j$ and $a_1= \dots =a_m=1/m$ imply $p_1= \dots =p_m$.
\end{proof}
\begin{rmk}
As a consequence, we obtain a lower bound $\beta_{cr}\geq m/(m-1)$.
\end{rmk}
\begin{lem}\thlabel{newlemma}
For $\mathbf 0 \leq \mathbf s \in \mathcal S$ and $\mathbf p \colonequals (p_1,\dots, p_m)^T$, we have \[\|\mathbf A^t\mathbf s\|_1 \leq g^t \left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2},\quad \mathbf e_j^T \mathbf A^t \mathbf s \leq \sqrt{p_j} g^t \left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2} . \]
In particular, it holds that \[\|\mathbf A^t \mathbf{p}\|_1 \leq g^t, \quad \mathbf e_j^T \mathbf A^t \mathbf p \leq \sqrt{p_j}g^t. \]
\end{lem}
\begin{proof}
We want to find a symmetric matrix $\mathbf{C}$ which is similar to $\mathbf A$.
To that end, suppose that there exists an invertible diagonal matrix $\mathbf{D}=\mathrm{diag}(d_1, \dots, d_m)$ and a symmetric matrix $\mathbf{C}$ such that $\mathbf{C}=\mathbf{D}^{-1}\mathbf A\mathbf{D}$.
Then $\mathbf{D}\mathbf A^T\mathbf{D}^{-1}=\mathbf{C}^T=\mathbf{C}=\mathbf{D}^{-1}\mathbf A\mathbf{D}$, so $\mathbf{D}^2\mathbf A^T=\mathbf A\mathbf{D}^2$, which leads to $d_i^2 p_j = p_i d_j^2$ for $i,j \in \{1,2,\dots, m\}$.
With the above in mind, let $\mathbf{D}\colonequals \mathrm{diag}(\sqrt{p_1},\dots ,\sqrt{p_m})$ and $\mathbf{C}\colonequals (c_{ij}) $ where $c_{ii}=1-1/n$ and $c_{ij}=\sqrt{p_ip_j}{\beta}/{n}$ for $i\neq j$.
Note that $\mathbf{C}$ is real-symmetric and $\mathbf{C}=\mathbf{D}^{-1}\mathbf A\mathbf{D}$.
Then, by the spectral theorem for real symmetric matrices, $\|\mathbf{C}\|_2=g$.
Note that $\mathbf{C}$ and $\mathbf A$ have the same real eigenvalues since they are similar.
Observe that for $\mathbf x, \mathbf y \in \mathbb R^m$, $\|\mathbf x\mathbf y^T\|_2=\|\mathbf x\|_2\|\mathbf y\|_2$.
This can be easily checked by the equalities \[\|\mathbf x \mathbf y^T\|_2=\sup_{\|\mathbf z\|_2=1} \|\mathbf x \mathbf y^T \mathbf z\|_2 =\sup_{\|\mathbf z\|_2=1} |\mathbf y^T \mathbf z |\|\mathbf x \|_2 = \|\mathbf y\|_2 \|\mathbf x\|_2 .\]
Let $\mathbbm 1 \colonequals (1,\dots,1)^T$.
The case $\mathbf s=\mathbf 0$ is trivial, so assume $\mathbf s>\mathbf 0$.
Since $\mathbf s \mathbbm 1^T$ has rank 1, $\mathbf{C}^t\mathbf{D}^{-1}\mathbf s\mathbbm 1^T \mathbf{D}$ has rank $1$ .
Also, its elements are positive, so it has a positive eigenvalue by the Perron-Frobenius theorem.
Thus, $\mathrm{Tr}(\mathbf{C}^t\mathbf{D}^{-1}\mathbf s\mathbbm 1^T \mathbf{D})$ is equal to its spectral radius, from which the following inequality follows:
\begin{align*}
\|\mathbf A^t \mathbf s\|_1 &= \mathbbm 1^T \mathbf A^t \mathbf s = \mathbbm 1^T \mathbf{D}\cdot \mathbf{C}^t\mathbf{D}^{-1} \mathbf s =\mathrm{Tr}(\mathbf{C}^t\mathbf{D}^{-1}\mathbf s\mathbbm 1^T \mathbf{D}) \leq \|\mathbf{C}^t\mathbf{D}^{-1}\mathbf s\mathbbm 1^T \mathbf{D} \|_2
\\&\leq \|\mathbf{C}\|_2^t \|\mathbf{D}^{-1}\mathbf s \mathbbm 1^T \mathbf{D}\|_2 = g^t \|\mathbf{D}^{-1}\mathbf s\|_2 \|\mathbf{D} \mathbbm 1 \|_2 = g^t \left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2} .
\end{align*}
Similarly, \[\mathbf e_j^T \mathbf A^t \mathbf s \leq \|\mathbf{C}\|_2^t \|\mathbf{D}^{-1}\mathbf s \mathbf e_j^T \mathbf{D}\|_2 =g^t\|\mathbf{D}^{-1}\mathbf s\|_2\|\mathbf{D}\mathbf e_j\|_2= \sqrt{p_j}g^t\Biggl(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\Biggr)^{1/2}.\]
\end{proof}
\begin{rmk}
Another relatively simple proof of $\|\mathbf A^t\mathbf s\|_1 \leq g^t \left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2}$ can be given as follows.
By the Cauchy-Schwartz inequality, we have $\sqrt{\sum_i s_i^2 /p_i}\geq \sum_i s_i$.
Then $\|\mathbf A^t \mathbf s\|_1\leq \|\mathbf D^{-1}\mathbf A^t \mathbf s\|_2=\|\mathbf D^{-1}\mathbf A^t \mathbf D \mathbf D^{-1} \mathbf s\|_2=\|\mathbf C^t \mathbf D^{-1}\mathbf s\|_2\leq g^t \|\mathbf D^{-1} \mathbf s\|_2$.
\end{rmk}
From now on, for brevity, we use the notation \[\mathbf p \colonequals (p_1,\dots,p_m)^T.\]
\begin{lem}\thlabel{lemma}
For a monotone coupling $({\sigma_{t}},{\sigma'_{t}})_{t\geq 0}$ starting at $(\sigma,\sigma')$, we have
\[\mathbb E \sum_{i=1}^m a_i \mathrm{dist}_i (\sigma_t,\sigma'_t) \leq g^t \sum_{i=1}^m a_i \mathrm{dist}_i (\sigma,\sigma'). \]
Moreover, for $i=1,\dots,m$, \[\mathbb E \mathrm{dist}_i ({\sigma^{(i)}_t}, {\sigma'^{(i)}_t}) \leq n\sqrt{p_i}g^t.\]
\end{lem}
\begin{proof}
From \thref{monotonecontraction}, \begin{align*}
\mathbb E \sum_{i=1}^m a_i \mathrm{dist}_i (\sigma_t,\sigma'_t) &= \mathbf a^T \begin{pmatrix}\mathbb E\mathrm{dist}_1({\sigma^{(1)}_t},{\sigma'^{(1)}_t})\\ \vdots \\ \mathbb E\mathrm{dist}_m({\sigma^{(m)}_{t}},{\sigma'^{(m)}_{t}}) \end{pmatrix} \leq \mathbf a^T \mathbf A^t \begin{pmatrix} \mathrm{dist}_1(\sigma^{(1)}, \sigma'^{(1)}) \\ \vdots \\ \mathrm{dist}_m(\sigma^{(m)},\sigma'^{(m)}) \end{pmatrix}
\\ &\leq g^t \mathbf a^T \begin{pmatrix} \mathrm{dist}_1(\sigma^{(1)}, \sigma'^{(1)}) \\ \vdots \\ \mathrm{dist}_m(\sigma^{(m)},\sigma'^{(m)}) \end{pmatrix} \leq g^t \sum_{i=1}^m a_i \mathrm{dist}_i (\sigma,\sigma').
\end{align*}
Notice that $\mathrm{dist}_k({\sigma^{(k)}_{t}},{\sigma'^{(k)}_{t}}) \leq np_k$ for each $k$, so \thref{newlemma} implies \[
\mathbb E \mathrm{dist}_i ({\sigma^{(i)}_t}, {\sigma'^{(i)}_t}) \leq n\mathbf e_i^T \mathbf A^t \mathbf p \leq n\sqrt{p_i}g^t.
\]
\end{proof}
We would like to translate \thref{lemma} to the case of magnetization chains, which is done in \thref{normcontraction}.
\begin{lem}\thlabel{generalcontraction}
For starting magnetizations $\mathbf s = (s^{(1)},\dots,s^{(m)})\geq(s'^{(1)},\dots, s'^{(m)})= \mathbf s'$, the magnetization chains satisfy
\[\mathbf 0 \leq \begin{pmatrix}\mathbb E_{\mathbf s}{S^{(1)}_{t}} -\mathbb E_{\mathbf s'}{S'^{(1)}_{t}}\\ \vdots \\ \mathbb E_{\mathbf s}{S^{(m)}_{t}} -\mathbb E_{\mathbf s'}{S'^{(m)}_{t}} \end{pmatrix} \leq \mathbf A^t \begin{pmatrix} s^{(1)}-s'^{(1)} \\ \vdots \\ s^{(m)}-s'^{(m)} \end{pmatrix}.\]
\begin{rmk}
We say such pairs of starting magnetizations are \emph{monotone pairs}.
\end{rmk}
\end{lem}
\begin{proof}
Let $({\sigma_{t}},{\sigma'_{t}})$ be a monotone coupling starting from $(\sigma,\sigma')$ where $\sigma \geq \sigma'$ and $S^{(i)}(\sigma)=s_i$, $S'^{(i)}(\sigma')=s_i'$ for $i=1,\dots,m$.
Such a monotone coupling exists because of the given condition $s_i \geq s_i'$ for each $i$.
Since $\sigma_i \geq \sigma_i'$, we have $s_i-s_i' =\frac{2}{n}\mathrm{dist}_i(\sigma_i,\sigma_i')$ for each $i$.
By monotonicity, ${\sigma^{(i)}_t} \geq {\sigma'^{(i)}_t}$ for each $i$.
Thus, ${S^{(i)}_{t}}-{S'^{(i)}_{t}}=|{S^{(i)}_{t}}-{S'^{(i)}_{t}}|=\frac{2}{n}\mathrm{dist}_i({\sigma^{(i)}_t},{\sigma'^{(i)}_t})\geq0$ for each $i$.
Then, by \thref{monotonecontraction}, \[\mathbf 0 \leq \begin{pmatrix}\mathbb E_{\sigma}{S^{(1)}_{t}} -\mathbb E_{\sigma'}{S'^{(1)}_{t}}\\ \vdots \\ \mathbb E_{\sigma}{S^{(m)}_{t}} -\mathbb E_{\sigma'}{S'^{(m)}_{t}} \end{pmatrix} =
\begin{pmatrix}\mathbb E_{\sigma,\sigma'}|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|\\ \vdots \\ \mathbb E_{\sigma,\sigma'}|{S^{(m)}_{t}}-{S'^{(m)}_{t}}| \end{pmatrix} \leq \mathbf A^t \begin{pmatrix} s^{(1)}-s'^{(1)} \\ \vdots \\ s^{(m)}-s'^{(m)} \end{pmatrix}.\]
Now, we can complete the proof since we have $\mathbb E_{\sigma}{S^{(i)}_{t}}-\mathbb E_{\sigma'}{S'^{(i)}_{t}}= \mathbb E_{\mathbf s}{S^{(i)}_{t}} -\mathbb E_{\mathbf s'}{S'^{(i)}_{t}}$ for each $i$ by \thref{magmarkov}.
\end{proof}
Recall that $\circ$ denotes a Hadamard product.
\begin{prop}\thlabel{normcontraction}
For a monotone coupling $(\sigma_t,\sigma_t')_{t\geq 0}$ starting at $(\sigma,\sigma')$ with magnetizations $(\mathbf s, \mathbf s')$, we have \begin{align*}
\mathbb E_{\sigma, \sigma'}\|\mathbf a \circ \mathbf S_t - \mathbf a \circ \mathbf S_t'\|_1
\leq g ^t \|\mathbf a \circ \mathbf s - \mathbf a \circ \mathbf s'\|_1
.\end{align*}
Moreover, not depending on the coupling, we have
\begin{align*}
\|\mathbb E_{\mathbf s}\mathbf a \circ \mathbf S_t - \mathbb E_{\mathbf s'}\mathbf a \circ \mathbf S_t'\|_1
\leq g ^t \|\mathbf a \circ \mathbf s - \mathbf a \circ \mathbf s'\|_1.
\end{align*}
\end{prop}
\begin{proof}
For any magnetizations $\mathbf s \equiv \mathbf s_{(0)}$ and $\mathbf s' \equiv \mathbf s_{(m)}$, there exists $\mathbf s _{(1)} ,\dots, \mathbf s_{(m-1)} \in \mathcal S \subset \mathbb R^m$ such that $\mathbf s_{(i-1)} - \mathbf s_{(i)} = \mathbf e_i (s^{(i)}-s'^{(i)})$ for $i=1,\dots, m$.
In particular, $\mathbf s_{(i-1)}$ and $\mathbf s_{(i)}$ are a monotone pair for each $i$.
Then we can consider a monotone coupling $(\sigma_{(0),t},\dots,\sigma_{(m),t})_{t\geq 0}$ with starting states $(\sigma_{(0)},\dots ,\sigma_{(m)})$ such that $\sigma_t=\sigma_{(0),t}$, $\sigma_t'=\sigma_{(m),t}$ for $t\geq 0$, and the magnetization of the starting configuration $\sigma_{(i)}$ is $\mathbf s_{(i)}$ for $i=0,\dots,m$.
Let $\mathbf S_{(j),t}$ be the magnetization chain corresponding to $\sigma_{(j),t}$ for $j=0,\dots,m$.
By telescoping, \thref{generalcontraction} gives \begin{align*}
&\mathbb E_{\sigma,\sigma'} \|\mathbf a \circ \mathbf S_{t} - \mathbf a \circ \mathbf S_{t}'\|_1 \leq \sum_{j=1}^m \mathbb E_{\sigma_{(j-1)},\sigma_{(j)}} \| \mathbf a \circ \mathbf S_{(j-1),t} - \mathbf a \circ \mathbf S_{(j),t}\|_1
\\&\leq \sum_{j=1}^m \mathbf a^T \mathbf A^t \mathbf {e}_j |s^{(j)}-s'^{(j)} | = g^t \sum_{j=1}^m a_j |s^{(j)}-s'^{(j)} | = g ^t \|\mathbf a \circ \mathbf s - \mathbf a \circ \mathbf s'\|_1.
\end{align*}
Then, the triangle inequality and \thref{magmarkov} imply \[\|\mathbb E_{\mathbf s}\mathbf a \circ \mathbf S_t - \mathbb E_{\mathbf s'}\mathbf a \circ \mathbf S_t'\|_1
\leq g ^t \|\mathbf a \circ \mathbf s - \mathbf a \circ \mathbf s'\|_1.\]
\end{proof}
\section{Variance bound of the magnetization in high temperatures} \label{section3}
The next lemma is a generalization of Lemma 2.6 in \cite{levin} to Markov chains with a finite state space in $\mathbb R^m$.
Observe that for square-integrable $\mathbb R^m$-valued i.i.d. random vectors $X,Y$, we have $\mathbb{V}\mathrm{ar}X=\frac{1}{2}\mathbb E\|X-Y\|_2^2$.
\begin{lem}\thlabel{variancebound}
Let $(\mathbf Z_t)_{t\geq 0}$ be a Markov chain in a finite state space $\tilde {\mathcal S} \subseteq \mathbb R^m$.
Suppose that there exists $0<r<1$ such that for any $\theta, \theta' \in \tilde {\mathcal S}$, \[\|\mathbb E_\theta \mathbf Z_t - \mathbb E_{\theta'}\mathbf Z_t'\|_1 \leq r^t \|\theta-\theta'\|_1.\]
Then, for the $l^2$ norm variance, \[\sup_{\theta \in \mathcal S} \mathbb{V}\mathrm{ar}_{\theta} \mathbf Z_t \leq m\sup_{\theta \in \mathcal S} \mathbb{V}\mathrm{ar}_{\theta}\mathbf Z_1 \: \min \{t, (1-r ^2)^{-1}\}.\]
\end{lem}
\begin{proof}
Put $v_t \colonequals \sup_{\theta \in \mathcal S} {\mathbb{V}\mathrm{ar}}_{\theta}\mathbf Z_t$.
Let $(\mathbf Z_t)$ and $(\mathbf Z_t')$ be independent copies of the chain starting from $\theta \in \tilde{\mathcal S}$.
The idea is to condition on the first step.
Note that $\|\mathbf x\|_2 \leq \|\mathbf x\|_1 \leq \sqrt{m}\|\mathbf x\|_2$ for $\mathbf x \in \mathbb R^m$.
Then by the observation right before the statement of this lemma, \[ \frac{1}{2}\mathbb E_{\theta}\|\mathbf Z_1-\mathbf Z_1'\|_1^2 \leq m\frac{1}{2}\mathbb E_{\theta}\|\mathbf Z_1-\mathbf Z_1'\|_2^2 \leq mv_1.\]
By the assumption and Markov property, we have \[\|\mathbb E_\theta[\mathbf Z_t | \mathbf Z_1] - \mathbb E_\theta[\mathbf Z_t' | \mathbf Z_1']\|_1=\|\mathbb E_{\mathbf Z_1}[\mathbf Z_{t-1}]-\mathbb E_{\mathbf Z_1'}[\mathbf Z_{t-1}']\|_1 \leq r^{t-1}\|\mathbf Z_1-\mathbf Z_1'\|_1.\]
Thus, for $\theta \in \tilde{\mathcal S}$, \begin{align*}
\mathbb{V}\mathrm{ar}_\theta [\mathbb E_\theta (\mathbf Z_t|\mathbf Z_1)]&= \frac{1}{2}\mathbb E_\theta \|\mathbb E_{\mathbf Z_1} \mathbf Z_{t-1}-\mathbb E_{\mathbf Z_1'} \mathbf Z_{t-1}'\|_2^2 \leq \frac{1}{2}\mathbb E_\theta \|\mathbb E_{\mathbf Z_1} \mathbf Z_{t-1}-\mathbb E_{\mathbf Z_1'} \mathbf Z_{t-1}'\|_1^2
\\ &\leq \frac{1}{2}\mathbb E_\theta \Big[r^{2(t-1)}\|\mathbf Z_1-\mathbf Z_1'\|_1^2\Big] \leq mv_1 r^{2(t-1)}.
\end{align*}
By the Markov property, for every $\theta \in \tilde{\mathcal S}$, $\mathbb{V}\mathrm{ar}_\theta [\mathbf Z_t | \mathbf Z_1 ] \leq v_{t-1}$, so \[\sup_{\theta \in \mathcal S}\mathbb E_\theta [\mathbb{V}\mathrm{ar}_\theta [\mathbf Z_t|\mathbf Z_1]] \leq v_{t-1}.\]
The total variance formula holds since we are using the $l^2$ norm.
Thus, taking supremum over $\theta \in \tilde{\mathcal S}$ in the total variance formula $\mathbb{V}\mathrm{ar}_\theta \mathbf Z_t= \mathbb E_\theta \big[\mathbb{V}\mathrm{ar}_\theta [\mathbf Z_t|\mathbf Z_1]\big] + \mathbb{V}\mathrm{ar}_\theta \big[\mathbb E_\theta [\mathbf Z_t|\mathbf Z_1]\big]$, we have $v_t \leq v_{t-1}+mv_1r^{2(t-1)}$.
Upon iterating, \[v_t \leq mv_1 \sum_{t=1}^{t} r^{2(t-1)} \leq mv_1 \min\big\{t, (1-r^2)^{-1}\big\}.\]
\end{proof}
The following proposition is an important result bounding the variance of magnetization chains uniformly in time.
\begin{prop}\thlabel{magvariancebound}
Let $\beta<\beta_{cr}$.
For an arbitrary starting configuration $\mathbf s$ and $t \geq 0$, we have \[\sum_{i=1}^m\mathbb{V}\mathrm{ar}_{\mathbf s}({S^{(i)}_{t}})=C/n\] where $C>0$ only depends on $p_1,\dots,p_m$, and $\beta$.
\end{prop}
\begin{proof}
Observe that $\sum_{i=1}^m \mathbb{V}\mathrm{ar}_{\mathbf s}(a_i S^{(i)}_{t}) = \mathbb{V}\mathrm{ar}_{\mathbf s}(\mathbf a \circ \mathbf S_t)$.
Note that increments of $\mathbf S_t$ are bounded by $2/n$ in absolute value.
Then, from \thref{finalfinallemma}, we have \[\sum_{i=1}^m \mathbb{V}\mathrm{ar}_{\mathbf s}{a_i S^{(i)}_{1}} \leq a_1^2 (2/n)^2.\]
By \thref{finalfinallemma}, \thref{normcontraction}, and \thref{variancebound}, we have
\[a_m^2\sum_{i=1}^m\mathbb{V}\mathrm{ar}_{\mathbf s}({S^{(i)}_{t}}) \leq \sum_{i=1}^m\mathbb{V}\mathrm{ar}_{\mathbf s}(a_i{S^{(i)}_{t}}) \leq m\frac{4a_1^2}{n^2}\frac{1}{1-g^2} = \frac{4ma_1^2}{\upsilon n(1+g)}\leq \frac{4ma_1^2}{\upsilon n}. \]
Note that \thref{upsilon} assures $\upsilon>0$.
\end{proof}
We also establish a bound for the expected magnetization on subsets of partitions.
To that end, we need the following observation.
\begin{lem}\thlabel{zerospin}
For each $i \in V$, $\mathbb E_{\mu}(\sigma(i))=0$ where $\mu$ is the Gibbs distribution.
In particular, we have $\mathbb E_\mu (S^{(i)})=0$.
\end{lem}
\begin{proof}
Since $\mu(\sigma)=\mu(-\sigma)$ for each configuration $\sigma$ and $\sigma \mapsto -\sigma$ is a bijection from $\Omega$ into itself, we have $\mathbb E_{\mu}(\sigma(i))= \sum_\sigma \sigma(i) \mu(\sigma) = \sum _{\sigma : \sigma(i)=1} \mu(\sigma)-\sum _{\sigma : \sigma(i)=-1} \mu(\sigma)=0$.
\end{proof}
\begin{prop}[Expected magnetization bound]\thlabel{expectedmagbound}
Let $\beta< \beta_{cr}$ and $1\leq i \leq m$. For any $B \subseteq J_i$ and a chain $(\sigma_t)_{t \geq 0}$ starting at $\sigma \in \Omega$, define $M_t (B) \colonequals \frac{1}{2}\sum_{k \in B}{\sigma_{t}} (k)$.
Then \[|\mathbb E_\sigma M_t(B)| \leq |B| g^t /\sqrt{p_i}.\]
Furthermore, for $t \geq \frac{1}{2(1-\beta/\beta_{cr})} n\ln n$, we have \[\mathbb{V}\mathrm{ar}_\sigma(M_t (B)) =O(n)\ , \quad \mathbb E_\sigma|M_t (B)| = O(\sqrt{n}).\]
\end{prop}
\begin{proof}
Let "+" denote the configuration such that all spins are $1$ and "-" denote the configuration with all spins $-1$.
Let $(\sigma_{t}^+, \sigma_{t}^\mu ,\sigma_{t} ^-)$ be a monotone coupling with starting configuration $(+,\mu,-)$ where $\mu$ is the stationary distribution.
Let $i \in \{1,\dots,m\}$.
By \thref{lemma} and \thref{zerospin}, \begin{align*}
\mathbb E_+[M_t(J_i)^+] \leq \mathbb E_{+,\mu}|M_t(J_i)^+ -M_t(J_i)^\mu| +\mathbb E_\mu [M_t(J_i)^\mu] \leq n\sqrt{p_i} g^t.
\end{align*}
Then, by symmetry, for $v \in J_i$, $\mathbb E_+[M_t(v)] \leq n\sqrt{p_i}g^t/|J_i| =g^t/\sqrt{p_i}$.
Thus, by summing over sites in $B$, $\mathbb E_+[M_t(B)^+] \leq |B|g^t/\sqrt{p_i}$.
However, for any configuration $\sigma$, by monotonicity, $\mathbb E_+[M_t(B)^+] \geq \mathbb E_\sigma [M_t(B)] \geq \mathbb E_-[M_t(B)^-]$.
Considering the remark after \thref{magmarkov}, $\mathbb E_- [M_t(B)^-]= -\mathbb E_+ [M_t(B)^+]$.
Thus, $|\mathbb E_\sigma [M_t(B)]| \leq |\mathbb E_+[M_t(B)^+]| \leq |B|g^t/\sqrt{p_i}$ for any $\sigma$.
Now, by \thref{magvariancebound}, $ O(1/n)=\mathbb{V}\mathrm{ar}{S^{(i)}_{t}} =\mathbb{V}\mathrm{ar}(M_t(J_i)2/n) $, so \[\mathbb{V}\mathrm{ar}_+( M_t(J_i) ) = O(n). \]
Thus, for $t \geq \frac{1}{2(1-\beta/\beta_{cr})} n\ln n$, \[\mathbb E_+(M_t(J_i)^2)=\mathbb{V}\mathrm{ar}_+(M_t(J_i))+(\mathbb E_+M_t(J_i))^2=O(n)\]
However, by symmetry, for any fixed $v_1,v_2 \in J_i$, \[\mathbb E_+(M_t(J_i)^{2})=np_i+\binom{np_i}{2}\mathbb E_+(\sigma_{t}^+(v_1)\sigma_{t}^+(v_2)). \]
Thus, \[|\mathbb E_+\sigma_{t}^+(v_1)\sigma_{t}^+(v_2)|=O(1/n).\]
Likewise, for $B\subseteq J_i$, \[\mathbb E_+(M_t(B)^2) =|B|+ \binom{|B|}{2}\mathbb E_+(\sigma_{t}^+(v_1)\sigma_{t}^+(v_2)) \leq O(n). \]
Similarly, $\mathbb E_-M_t(B)^2 \leq O(n)$, so from $(M_t(B))^2 \leq (M_t(B)^+)^2+(M_t(B)^-)^2$, \[\mathbb E(M_t(B)^2) =O(n) \] whenever $t \geq \frac{1}{2(1-\beta/\beta_{cr})} n\ln n$.
Thus, for $t \geq \frac{1}{2(1-\beta/\beta_{cr})} n\ln n$, \[\mathbb{V}\mathrm{ar}_\sigma(M_t(B))=O(n).\]
Lastly, for $t \geq \frac{1}{2(1-\beta/\beta_{cr})} n\ln n$, from Jensen's inequality, \begin{align*}
\mathbb E_\sigma|M_t (B)| &\leq \sqrt{\mathbb E_\sigma|M_t (B)| ^2} =\sqrt{(\mathbb E_\sigma[M_t (B)])^2 +\mathbb{V}\mathrm{ar}_\sigma(M_t (B))}
\\ &\leq |\mathbb E_\sigma[M_t (B)]| + \sqrt{\mathbb{V}\mathrm{ar}_\sigma(M_t (B))}=O(\sqrt{n}).
\end{align*}
\end{proof}
\section{Couplings}\label{section4}
Fix the notation \[t_n \colonequals \frac{1}{2(1-\beta/\beta_{cr})} n \ln n.\]
\begin{defi}[Modified matching]
Let $\sigma\in \Omega$ and $\sigma'\in \Omega$ have magnetizations $\mathbf s \in \mathcal S$ and $\mathbf s'\in \mathcal S$, respectively.
Consider two copies of the graph, $V=\bigcup_i J_i$ and $V'=\bigcup_i J_i'$.
Let $i \in \{1,\dots,m\}$.
If $s^{(i)} \geq s'^{(i)}$, then it is possible to match each site in $J_i'$ with $+1$ spin to a site in $J_i$ with $+1$ spin.
Any leftover sites in $J_i'$ are arbitrarily matched to the leftover sites in $J_i$.
We match the sites in a similar way whenever $s^{(i)}\leq s'^{(i)}$.
This defines a bijection $f_{\sigma,\sigma'}\colon V \to V'$.
We call this bijection a \emph{modified matching of $\sigma$ and $\sigma'$}.
\end{defi}
\begin{defi}[Modified monotone update and coupling]
Let $f_{\sigma,\sigma'}\colon V \to V'$ be a modified matching of $\sigma,\sigma' \in\Omega$.
Let $I$ and $U$ be uniformly distributed over $V=\bigcup_{i=1}^m J_i$ and $[0,1]\subseteq \mathbb R$, respectively, and be independent.
Suppose $I \in J_\eta$ for some $\eta \in \{1,\dots, m\}$ is the chosen site in $V$.
Consider the case $\sum_{v \notin J_\eta }\sigma(v) \leq \sum_{v \notin J_\eta }\sigma'(v)$.
If \[U< \frac{1+\tanh\left(\beta \sum_{v \notin J_\eta }\sigma(v) \right)}{2},\] then update the chosen site $I$ of $V$ by +1 and $f_{\sigma,\sigma'}(I)$ of $V'$ by +1.
If \[ U\geq \frac{1+\tanh\left(\beta \sum_{v \notin J_\eta }\sigma'(v) \right)}{2},\] then update the chosen site $I$ of $V$ by -1 and $f_{\sigma,\sigma'}(I)$ of $V'$ by -1.
Otherwise, if \[\frac{1+\tanh\left(\beta \sum_{v \notin J_\eta }\sigma(v) \right)}{2} \leq U < \frac{1+\tanh\left(\beta \sum_{v \notin J_\eta }\sigma'(v) \right)}{2},\] then update the chosen site $I$ of $V$ by -1 and $f_{\sigma,\sigma'}(I)$ of $V'$ by +1.
The other case $\sum_{v \notin J_\eta }\sigma(v) > \sum_{v \notin J_\eta }\sigma'(v)$ can similarly be updated.
Given the chosen site $I$, we call the above procedure of deciding the updating spin in the two chains a \emph{modified monotone update} with respect to the given modified matching.
Now, fix a modified matching $f_{\sigma,\sigma'}$ of $\sigma$ and $\sigma'$.
Let ${\sigma_{t}}$ and ${\sigma'_{t}}$ be chains starting at $\sigma$ and $\sigma'$, respectively.
Repeating the above procedure independently for each step with respect to $f_{\sigma,\sigma'}$ gives a coupling of the Glauber dynamics.
We call this coupling a \emph{modified monotone coupling} with respect to the given modified matching.
\end{defi}
\begin{rmk}
\thref{monotonecontraction} and its consequences hold with a suitable distance function for a modified coupling with respect to a given modified matching.
\end{rmk}
We first construct a coupling such that the magnetizations agree after $t_n +O(n)$ steps in the next two lemmas.
\begin{lem}[Lemma 2.4, \cite{levin}]\thlabel{supermartingale}
Let $(W_t)_{t \geq 0 }$ be a non-negative supermartingale with a stopping time $\tau$ satisfying
\\\emph{(i)} \ $W_{0} =k$
\\\emph{(ii)} \ $W_{t+1}-W_{t} \leq B <\infty$
\\\emph{(iii)} \ $\mathbb{V}\mathrm{ar}(W_{t+1}|\mathcal F_t) >\sigma ^2 >0 \ \text{on the event} \ \{\tau >t\}$.
Then for $u> \frac{4B^2}{3\sigma^2}$, \[\mathbb P_k(\tau >u) \leq \frac{4k}{\sigma\sqrt{u}}.\]
\end{lem}
\begin{lem}[Magnetization coupling]\thlabel{magcoupling}
Let $\beta< \beta_{cr}$.
For any configurations $\sigma$ and $\sigma'$, there exists a coupling $({\sigma_{t}} ,{\sigma'_{t}})$ with starting states $(\sigma, \sigma')$ satisfying the following condition.
If $\tau_{mag} \colonequals \min \{t \geq 0: \mathbf S_t = \mathbf S_t'\}$, then for large $\gamma n$, \[\mathbb P_{\sigma,\sigma'}(\tau_{mag} >t_n +\gamma n) \leq \frac{c}{\sqrt{\gamma}}\] where $c>0$ is a constant not depending on $\sigma$, $\sigma'$, or $n$.
\end{lem}
\begin{proof}
Let $({\sigma_{t}}, {\sigma'_{t}})$ be a monotone coupling with starting states $(\sigma,\sigma')$.
Put $Y_{i,t} \colonequals \frac{n}{2}a_i|{S^{(i)}_{t}}-{S'^{(i)}_{t}}|$ for $i=1,\dots,m$ and $Y_{tot,t}\colonequals \sum_{i=1}^m Y_{i,t}$.
Define \[\tau \colonequals \min \{t\geq t_n : \max_{1\leq i\leq m}Y_{i,t}/a_i\leq 1\}.\]
By \thref{normcontraction}, \[\mathbb E_{\sigma, \sigma '} [Y_{tot,t_n}] \leq c \sqrt{n}\] for some $c>0$.
We construct a coupling such that $(Y_{tot,t})_{t_n\leq t< \tau}$ is a positive supermartingale with bounded increments and the conditional probability of not being lazy is bounded away from zero uniformly in time and $n$.
To that end, consider a time $t_n\leq t< \tau$.
Define $K_t\colonequals \bigcup_{i: Y_{i,t}/a_i \leq 1} J_i$, $L_t\colonequals \bigcup_{i: Y_{i,t}/a_i > 1} J_i$, and $L_t'\colonequals \bigcup_{i: Y_{i,t}/a_i > 1} J_i'$.
Note that $L_t \neq \emptyset$ since $t < \tau$.
Choose a site equiprobably over $V=K_t \dot\cup L_t$.
Let $f_t$ be the modified matching of ${\sigma_{t}}$ and ${\sigma'_{t}}$.
If a site in $K_t$ is chosen, then use the modified monotone update with respect to $f_t$ to update $({\sigma_{t}},{\sigma'_{t}})$.
If a site in $L_t$ is chosen, then independently choose another site equiprobably over $L_t'$ (which can be the same site) to update ${\sigma'_{t}}$ independent of ${\sigma_{t}}$.
It is easy to check that the above is a coupling of the Glauber dynamics.
Clearly, $Y_{tot,t}$ has bounded increment with the above coupling.
Let $I$ be a random variable uniformly distributed over $V$ which is independent of $\mathcal F_t$.
Let $E= \{I\in L_t,{\sigma_{t}}(I)=+1, \sigma_{t+1}(I)=-1, \sigma_{t+1}'(f_t(I))=1\}$ and $F= \{I\in L_t,{\sigma_{t}}(I)=-1, \sigma_{t+1}(I)=+1, \sigma_{t+1}'(f_t(I))=-1\}$.
Since $L_t \neq \emptyset$ implies $|L_t|/n\geq p_1$, we obtain that $\mathbb P(Y_{tot,t+1}\neq Y_{tot,t}|\mathcal F_t)$ is bounded below by \begin{align*}
&\geq \mathbb P(Y_{tot,t+1}\neq Y_{tot,t}, I \in L_t|\mathcal F_t) \geq \mathbb P(E\dot\cup F | \mathcal F_t)
\\&\geq \frac{|L_t| +\sum_{i\in L_t}{\sigma_{t}}(i)}{2n}\biggl(\frac{1-\tanh(\beta(1-p_1))}{2}\biggr)^2
\\&\enspace+ \frac{|L_t| -\sum_{i\in L_t}{\sigma_{t}}(i)}{2n}\biggl(\frac{1-\tanh(\beta(1-p_1))}{2}\biggr)^2
\\&\geq p_1\biggl(\frac{1-\tanh(\beta(1-p_1))}{2}\biggr)^2>0.
\end{align*}
Finally, we need to show the supermartingale property.
Consider $Y_{1,t+1}/a_1-Y_{1,t}/a_1$.
Suppose $J_1 \subseteq K_t$.
Then by a direct calculation, on the event $\{J_1 \subseteq K_t\}$, it holds that $\mathbb E(Y_{1,t+1}/a_1-Y_{1,t}/a_1|\mathcal F_t)$ is bounded above by \begin{align*}
&\leq \biggl(p_1 -\frac{|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|}{2}\biggr)\frac{|\tanh(\beta \sum_{j\neq1}{S^{(j)}_{t}})-\tanh(\beta \sum_{j\neq1}{S'^{(j)}_{t}})|}{2}
\\&\enspace -\frac{|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|}{2}\biggl(1-\frac{|\tanh(\beta \sum_{j\neq1}{S^{(j)}_{t}})-\tanh(\beta \sum_{j\neq1}{S'^{(j)}_{t}})|}{2}\biggr)
\\&\leq \frac{1}{2}\biggl(-|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|+p_1\tanh\biggl(\beta \Bigl|\sum_{j\neq1}{S^{(j)}_{t}}-\sum_{j\neq1}{S'^{(j)}_{t}}\Bigl|\biggr)\biggr).
\end{align*}
Suppose $J_1 \subseteq L_t$.
Note that $Y_{1,t} >1$ implies $({S^{(1)}_{t+1}}-{S'^{(1)}_{t+1}})({S^{(1)}_{t}}-{S'^{(1)}_{t}})\geq 0 $ and $|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|>0$.
Let $\xi= ({S^{(1)}_{t}}-{S'^{(1)}_{t}})/|{S^{(1)}_{t}}-{S'^{(1)}_{t}}| \in \{\pm 1\}$.
Then by equation \eqref{dynamics} in Section \ref{subsection4.2}, on the event $\{J_1 \subseteq L_t\}$, $\mathbb E(Y_{1,t+1}/a_1-Y_{1,t}/a_1|\mathcal F_t)$ is equal to \begin{align*}
&= \xi \frac{n}{2}\biggl(\mathbb E({S^{(1)}_{t+1}}-{S^{(1)}_{t}}| {\sigma_{t}})-\mathbb E({S'^{(1)}_{t+1}}-{S'^{(1)}_{t}}| {\sigma'_{t}})\biggr)
\\&=\xi \frac{n}{2}\frac{1}{n}\biggl(-{S^{(1)}_{t}}+p_1\tanh(\beta \sum_{j\neq1}{S^{(j)}_{t}}) \biggr)
\\&\quad - \xi \frac{n}{2}\frac{1}{n}\biggl(-{S'^{(1)}_{t}}+p_1\tanh(\beta \sum_{j\neq1}{S'^{(j)}_{t}}) \biggr)
\\&= \frac{\xi}{2}\biggl(-({S^{(1)}_{t}}-{S'^{(1)}_{t}})+p_1\biggl(\tanh(\beta \sum_{j\neq1}{S^{(j)}_{t}})-\tanh(\beta\sum_{j\neq1}{S'^{(j)}_{t}})\biggr) \biggr)
\\&\leq \frac{1}{2}\biggl(-|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|+p_1\tanh\biggl(\beta \Bigl|\sum_{j\neq1}{S^{(j)}_{t}}-\sum_{j\neq1}{S'^{(j)}_{t}}\Bigl|\biggr) \biggr).
\end{align*}
Since either $J_1 \subseteq L_t$ or $J_1 \subseteq K_t$ must hold, $\mathbb E(Y_{1,t+1}/a_1-Y_{1,t}/a_1|\mathcal F_t)$ is equal to \begin{align*}
&= \mathbbm 1_{J_1 \subseteq K_t}\mathbb E(Y_{1,t+1}-Y_{1,t}|\mathcal F_t) +\mathbbm 1_{J_1 \subseteq L_t}\mathbb E(Y_{1,t+1}-Y_{1,t}|\mathcal F_t)
\\&\leq \frac{1}{2}\biggl(-|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|+p_1\tanh\biggl(\beta \Bigl|\sum_{j\neq1}{S^{(j)}_{t}}-\sum_{j\neq1}{S'^{(j)}_{t}}\Bigr|\biggr) \biggr)
\\&\leq\frac{1}{2}\biggl(-|{S^{(1)}_{t}}-{S'^{(1)}_{t}}|+p_1\beta \sum_{j\neq1}\Bigl|{S^{(j)}_{t}}-{S'^{(j)}_{t}}\Bigl| \biggr).
\end{align*}
Thus, \[ \mathbb E(Y_{1,t+1}/a_1|\mathcal F_t) \leq (1-\frac{1}{n})Y_{1,t}/a_1 +\frac{\beta p_1}{n}\sum_{j\neq 1} Y_{j,t}/a_j.\]
Putting in the matrix form with $ \tilde {\mathbf Y}_t \colonequals (Y_{1,t}/a_1,\dots, Y_{m,t}/a_m)^T$, we have \begin{align*}
\mathbb E(Y_{tot,t+1}|\mathcal F_t)= \mathbf a^T \mathbb E(\tilde {\mathbf Y}_{t+1}|\mathcal F_t) \leq \mathbf a^T \mathbf A \tilde {\mathbf Y}_t = g \mathbf a^T \tilde {\mathbf Y}_t = g Y_{tot,t}.
\end{align*}
Since $\beta<\beta_{cr}$ implies $g<1$ by \thref{upsilon}, the supermartingale property is established.
With the above coupling, by \thref{supermartingale}, for large $\gamma n$, \[\mathbb P_{\sigma,\sigma'}(\tau >t_n +\gamma n|\sigma_{t_n},\sigma'_{t_n}) \leq c' \frac{n\|(S^{(1)}_{t_n},\dots, S^{(m)}_{t_n})-(S'^{(1)}_{t_n},\dots,S'^{(m)}_{t_n})\|_1}{\sqrt{\gamma n}}\] for some $c' >0$ not depending on $n$.
Taking expectation, \[\mathbb P_{\sigma,\sigma'}(\tau >t_n +\gamma n) \leq O(\gamma ^{-1/2}).\]
Note $\sigma_{\tau}$ has at most $m$ more +1 spin sites than $\sigma_{\tau}'$, so $0\leq Y_{tot,\tau} \leq a_1 m$ by \thref{finalfinallemma}.
At $\tau$, construct a modified matching of $\sigma_{\tau}$ and $\sigma_{\tau}'$, and use the modified monotone coupling with respect to this modified matching from then on.
At $\tau_{mag}$, we construct another modified matching of the sites to do a new modified monotone coupling so that $({S^{(1)}_{t}},\dots, {S^{(m)}_{t}})=({S'^{(1)}_{t}},\dots,{S'^{(m)}_{t}})$ forever after $\tau_{mag}$.
By \thref{finalfinallemma}, a modified version of \thref{normcontraction}, and the strong Markov property, we have \begin{align*}
\mathbb P_{\sigma,\sigma'}(\tau_{mag}>\tau +\gamma' n|\sigma_{\tau}, \sigma_{\tau}') &\leq \mathbb P_{\sigma,\sigma'}(Y_{tot,\tau +\gamma' n}\geq a_m |\sigma_{\tau}, \sigma_{\tau}')
\\& \leq \mathbb E_{\sigma,\sigma'}[Y_{tot,\tau +\gamma' n}|\sigma_{\tau}, \sigma_{\tau}'] /a_m
\\& \leq g^{\gamma'n} Y_{tot,\tau}/a_m \leq g^{\gamma'n} a_1m/a_m \leq e^{-\upsilon \gamma'}a_1m/a_m.
\end{align*}
Thus, \begin{align*}
\mathbb P_{\sigma,\sigma'}(\tau_{mag} > t_n + (\gamma+\gamma')n) &\leq O(\gamma^{-1/2})+ e^{-\upsilon \gamma'}a_1m/a_m,
\end{align*} and putting $\gamma=\gamma'$ yields \begin{align*}
\mathbb P_{\sigma,\sigma'}(\tau_{mag} > t_n + \gamma n) \leq O(\gamma^{-1/2}).
\end{align*}
\end{proof}
\begin{defi}[Good configurations]
Define the set of "good" configurations by \[\tilde \Omega \colonequals \{\sigma \in \Omega : |S^{(i)}(\sigma )|\leq p_i/2, \ i=1,\dots,m\}.\]
For $\sigma=(\sigma^{(1)}, \dots,\sigma^{(m)}) \in \tilde \Omega$ and each $i$, define \begin{align*}
&u_{i }^{\sigma} \colonequals |\{v\in J_i: \sigma^{(i)}(v)=1\}|, \enspace v_{i }^{\sigma} \colonequals |\{v\in J_i: \sigma^{(i)}(v)=-1\}|.
\end{align*}
Define \[\tilde \Lambda \colonequals \{(u_1,v_1,u_2,v_2,\dots,u_m,v_m) \in \mathbb N^{2m} : {|J_i|}/{4} \leq u_i \wedge v_i,\ i=1,\dots,m\}.\]
\end{defi}
\begin{rmk}
Note that $\sigma \in \tilde\Omega \iff (u_1^{\sigma},v_1^{\sigma},\dots,u_m^{\sigma},v_m^{\sigma} ) \in \tilde \Lambda$.
In other words, $\tilde \Lambda$ is another representation of good configurations $\tilde \Omega$.
We omit the starting state and write $u_i$ instead of $u_i^\sigma$ for convenience.
\end{rmk}
\begin{lem}[Lemma 3.3, \cite{levin}]\thlabel{goodstartingstatesdistance}
For any subset $A\subseteq \Omega$ and stationary distribution $\pi$,
\begin{align*}
d_n(t_0+t)&=\max_{\sigma \in \Omega} \|\mathbb P_{\sigma }(\sigma_{t_0+t} \in \cdot) -\pi\|_{TV}
\\&\leq \max_{\sigma \in A} \|\mathbb P_{\sigma }({\sigma_{t}} \in \cdot) -\pi\|_{TV} +\max_{\sigma \in \Omega} \mathbb P_{\sigma}(\sigma_{t_0} \notin A).
\end{align*}
\end{lem}
Recall that we are assuming the high temperature regime.
By \thref{expectedmagbound}, there exists $\delta >0$ such that $\max_{\sigma \in \Omega,1\leq i\leq m}|\mathbb E_{\sigma} S^{(i)}_{\delta n}| \leq p_1 /4$.
Hence, by \thref{magvariancebound}, for large $n$,
\begin{align*} \mathbb P_{\sigma}(\sigma_{\delta n}\notin \tilde \Omega ) &\leq \sum_{ i=1}^m \mathbb P_{\sigma}(|S^{(i)}_{\delta n} | >p_i/2) \leq \sum_{i =1}^m \mathbb P_{\sigma}(|S^{(i)}_{\delta n}-\mathbb E_{\sigma}S^{(i)}_{\delta n} | >p_i/4)
\\&\leq \frac{16}{p_1 ^2} \sum_{i=1}^m \mathbb{V}\mathrm{ar}_{\sigma}S^{(i)}_{\delta n}= O(1/n).\end{align*}
Combining with \thref{goodstartingstatesdistance},
\begin{equation}
d_n(\delta n+t) \leq \max_{\sigma \in \tilde \Omega} \|\mathbb P_{\sigma }({\sigma_{t}} \in \cdot) -\mu\|_{TV} +O(1/n). \label{distance}
\end{equation}
\begin{defi}[$2m$-coordinate chain]
Let $\tilde \sigma \in \Omega$ be a reference configuration.
For $\sigma \in \Omega$ and each $i$, define \begin{align*}
&U_i(\sigma) \colonequals |\{v\in J_i: \sigma^{(i)}(v)=\tilde \sigma^{(i)}(v)=1\}|,
\\& V_i(\sigma) \colonequals |\{v\in J_i: \sigma^{(i)}(v)=\tilde \sigma^{(i)}(v)=-1\}|.
\end{align*}
For a chain $({\sigma_{t}})$ with the starting configuration $ \sigma_0 \in \Omega$, define the \emph{$2m$-coordinate chain with respect to $\tilde \sigma$} by \begin{align*}
\mathbf U_t \colonequals ({U^{(1)}_{t}},{V^{(1)}_{t}},\dots,{U^{(m)}_{t}},{V^{(m)}_{t}})\colonequals (U_{1}({\sigma_{t}}),V_{1}({\sigma_{t}}),\dots,U_{m}({\sigma_{t}}),V_{m}({\sigma_{t}})).
\end{align*}
It is easy to see that the $2m$-coordinate chain is again a Markov chain in its state space $\mathcal U \subseteq \mathbb N^{2m}$ and determines the magnetization chain $({S^{(1)}_{t}},\dots,{S^{(m)}_{t}})$ through the relation ${S^{(i)}_{t}}=2({U^{(i)}_{t}}-V^{(i)}_t)/n-( \tilde u_{i}- \tilde v_{i})/{n}$ for $i=1,\dots,m$.
\end{defi}
Symmetry gives us the following lemma which is an adaptation of Lemma 3.4 in \cite{levin}.
\begin{lem}\thlabel{totalvariationdistance}
Let $({\sigma_{t}})$ be a chain starting at $\sigma \in \Omega$.
Consider the corresponding $2m$-coordinate chain starting at $\mathbf u \in \mathcal U$.
Then \[\|\mathbb P_{\sigma}({\sigma_{t}} \in \cdot) - \mu\|_{TV}= \|\mathbb P_{\mathbf u }(({U^{(1)}_{t}},{V^{(1)}_{t}},\dots, {U^{(m)}_{t}},{V^{(m)}_{t}})\in \cdot )-\nu\|_{TV}\]
where $\nu$ is the stationary distribution of the $2m$-coordinate chain.
\end{lem}
\begin{proof}
Since $\mu(\sigma)=e^{\beta n\sum_{i\neq j}S^{(i)}(\sigma) S^{(j)}(\sigma)}/Z(\beta)$, given the $2m$-coordinate $\mathbf u' \in \mathcal U$, the conditional $\mu$-probability of the configurations is equiprobable.
In other words, $\mu(\cdot | \Omega(\mathbf u'))$ is uniform where $\Omega(\mathbf u')$ is the set of configurations having the $2m$-coordinate $\mathbf u'$.
Also, by symmetry, \[\mathbb P_{\sigma}({\sigma_{t}} \in \cdot \ |\mathbf U_t = \mathbf u')\] is uniform over $\Omega(\mathbf u')$.
Thus, \begin{align*}
&\mathbb P_\sigma ({\sigma_{t}} = \eta) -\mu(\eta)
=\sum_{\mathbf u' \in \mathcal U}\frac{\mathbbm 1\{\eta \in \Omega(\mathbf u')\}}{|\Omega(\mathbf u')|}\left(\mathbb P_{\mathbf u'}\left(\mathbf U_t = \mathbf u'\right)-\mu (\Omega(\mathbf u'))\right).
\end{align*}
Taking absolute values, applying the triangular inequality, summing over $\eta$, and changing the order of summation shows \[\|\mathbb P_{\sigma}({\sigma_{t}} \in \cdot) - \mu\|_{TV}\leq \|\mathbb P_{\mathbf u }(({U^{(1)}_{t}},{V^{(1)}_{t}},\dots, {U^{(m)}_{t}},{V^{(m)}_{t}})\in \cdot )-\nu\|_{TV}.\]
The reverse inequality holds since the $2m$-coordinate chain is a function of the original chain $({\sigma_{t}})$.
\end{proof}
\begin{rmk}
This lemma lets us look at the $2m$-coordinate chain instead of the original chain when considering the total variation distance.
\end{rmk}
Fix a good configuration $\tilde \sigma \in \tilde \Omega$.
Recall $\tau_{mag}$ defined in \thref{magcoupling}.
We use the following coupling after $\tau_{mag}$, which is a generalization of Lemma 3.5 of \cite{levin}.
\begin{lem}[Post magnetization coupling]\thlabel{postmagcoupling}
Let $\tilde \sigma \in \tilde \Omega$ be a good configuration.
Suppose that two configurations $\sigma_0, \sigma_0'$ satisfy $S^{(i)}(\sigma_0)=S^{(i)}(\sigma_0')$ for $i=1,\dots,m$.
With respect to the good configuration $\tilde \sigma$, define \begin{align*}
\Theta_i \colonequals \Big\{\sigma \in \Omega : \min \{U _i(\sigma), \tilde u_i-U _i(\sigma), V_i(\sigma), \tilde v_i-V_i(\sigma)\} \geq \frac{|J_i|}{16}\Big\},\enspace \Theta \colonequals \bigcap _{ i=1}^m\Theta_i \end{align*} for each $i$.
Then there exists a coupling $({\sigma_{t}}, {\sigma'_{t}})$ of the Glauber dynamics with starting states $(\sigma_0,\sigma_0')$ satisfying: \begin{align*}
&\mathrm {(i)} \; \mathbf S_{t}=\mathbf S_{t}'\ \text{for all}\ t\geq0
\\&\mathrm {(ii)} \; \text{If ${R^{(i)}_{t}}\colonequals {U'^{(i)}_{t}}-{U^{(i)}_{t}}$, then }
\mathbb E_{\sigma_0,\sigma_0'}\left({R^{(i)}_{t+1}}-{R^{(i)}_{t}}| {\sigma_{t}}, {\sigma'_{t}} \right) = \frac{-{R^{(i)}_{t}}}{n}, \\&\qquad i=1,\dots,m
\\&\mathrm {(iii)} \; \text{There exists $c>0$ not depending on $n$ such that on the event}\ \{{\sigma_{t}},{\sigma'_{t}} \in \Theta\},\\& \qquad \mathbb P_{\sigma_0,\sigma_0'}\left({R^{(i)}_{t+1}}-{R^{(i)}_{t}} \neq 0| {\sigma_{t}}, {\sigma'_{t}} \right) \geq c>0 \; \text{for all} \ i=1,\dots,m.
\end{align*}
\end{lem}
\begin{proof}
We inductively define the coupling.
The random spin $S$ determined by the randomness $I$ and $U$ is \begin{align*}
S=\sum_{i=1}^m (\mathbbm 1_{I\in J_i,\; U\leq r_+(\sum_{j\neq i}{S^{(j)}_{t}})}-\mathbbm1_{I\in J_i,\; U > r_+(\sum_{j\neq i}{S^{(j)}_{t}})}).
\end{align*}
Suppose that $({\sigma_{t}}, {\sigma'_{t}})$ is given such that the statements hold for some $t\geq0$.
Let $\sigma_{t+1}$ be determined $I$ and $U$.
If $I\in J_i$ for some $i$, then choose $I'$ randomly from $\{v \in J_i' : {\sigma'_{t}}(v) = {\sigma_{t}}(I)\}$.
Update the primed chain by
\[\sigma_{t+1}'(v)= \begin{cases} {\sigma'_{t}}(v) & \text{if $v \neq I'$} \\ S & \text{if $v = I'$} \end{cases} \quad . \]
By the induction hypothesis $\mathbf S_t=\mathbf S_t'$, we have $\{v \in J_i' : {\sigma'_{t}}(v) = {\sigma_{t}}(I)\} \neq \emptyset$ and $({\sigma'_{t}})$ satisfies the Glauber dynamics.
Also, $\mathbf S_{t+1}=\mathbf S_{t+1}'$ with this coupling.
For $i=1,\dots,m$, put \begin{align*}
&A_i(\sigma) \colonequals \{v \in J_i : \sigma(v)=\tilde \sigma(v)=1\},
\\&B_i(\sigma) \colonequals \{v \in J_i : \sigma(v)=-1, \ \tilde\sigma(v)=1\},
\\&C_i(\sigma) \colonequals \{v \in J_i : \sigma(v)=1, \ \tilde\sigma(v)=-1\},
\\&D_i(\sigma) \colonequals \{v \in J_i : \sigma(v)=\tilde\sigma(v)=-1\},
\end{align*}
so $|A_i(\sigma )|=U_i(\sigma)$, $|B_i(\sigma )|=\tilde u_i-U_i(\sigma)$, $|C_i(\sigma )|=\tilde v_i- V_i(\sigma)$, and $|D_i(\sigma )|=V_i(\sigma)$.
Now we calculate ${R^{(1)}_{t+1}}-{R^{(i)}_{t}}$ with the above coupling.
The following table shows the one-step dynamics of ${R^{(1)}_{t}}$.
\begin{center}
\begin{tabular}{|c| c| c|c|}
\hline
$I$ & $I'$ & $S$ & ${R^{(1)}_{t+1}}-{R^{(1)}_{t}}$ \\
\hline\hline
$B_{1}({\sigma_{t}})$ & $D_{1}({\sigma'_{t}})$ & 1 & -1 \\
\hline
$C_{1}({\sigma_{t}})$ & $A_{1}({\sigma'_{t}})$ & -1 & -1 \\
\hline
$A_{1}({\sigma_{t}})$ & $C_{1}({\sigma'_{t}})$ & -1 & 1 \\
\hline
$D_{1}({\sigma_{t}})$ & $B_{1}({\sigma'_{t}})$ & 1 & 1 \\
\hline
otherwise &otherwise&otherwise&0\\
\hline
\end{tabular}
\end{center}
Since ${S^{(1)}_{t}}={S'^{(1)}_{t}}$ implies ${R^{(1)}_{t}}\equiv {U'^{(1)}_{t}}-{U^{(1)}_{t}}={V'^{(1)}_{t}}-{V^{(1)}_{t}}$, \begin{align*}
&\mathbb P_{\sigma_0,\sigma_0'}({R^{(1)}_{t+1}}-{R^{(1)}_{t}}=-1|{\sigma_{t}}, {\sigma'_{t}})\equalscolon a({U^{(1)}_{t}},{V^{(1)}_{t}},U_{2,t},V_{2,t})
\\&= \frac{\tilde u_1 -{U^{(1)}_{t}}}{n}\frac{{V'^{(1)}_{t}}}{\tilde u_1-{U'^{(1)}_{t}}+{V'^{(1)}_{t}}}r_+(\sum_{j\neq 1}{S^{(j)}_{t}})
+\frac{\tilde v_1 -{V^{(1)}_{t}}}{n}\frac{{U'^{(1)}_{t}}}{\tilde v_1-{V'^{(1)}_{t}}+{U'^{(1)}_{t}}}r_-(\sum_{j\neq 1}{S^{(j)}_{t}})
\\&=\frac{\tilde u_1-{U^{(1)}_{t}}}{n}\frac{{V^{(1)}_{t}}+{R^{(1)}_{t}}}{\tilde u_1-{U^{(1)}_{t}}+{V^{(1)}_{t}}}r_+(\sum_{j\neq 1}{S^{(j)}_{t}})
+\frac{\tilde v_1-{V^{(1)}_{t}}}{n}\frac{{U^{(1)}_{t}}+{R^{(1)}_{t}}}{\tilde v_1-{V^{(1)}_{t}}+{U^{(1)}_{t}}}r_-(\sum_{j\neq 1}{S^{(j)}_{t}}).
\end{align*}
Likewise, \begin{align*}
&\mathbb P_{\sigma_0,\sigma_0'}({R^{(1)}_{t+1}}-{R^{(1)}_{t}}=1|{\sigma_{t}}, {\sigma'_{t}})\equalscolon b({U^{(1)}_{t}},{V^{(1)}_{t}},U_{2,t},V_{2,t})
\\&= \frac{{U^{(1)}_{t}}}{n}\frac{\tilde v_1 - {V'^{(1)}_{t}}}{{U'^{(1)}_{t}}+\tilde v_1-{V'^{(1)}_{t}}}r_-(\sum_{j\neq 1}{S^{(j)}_{t}})
+\frac{{V^{(1)}_{t}}}{n}\frac{\tilde u_1 - {U'^{(1)}_{t}}}{\tilde u_1-{U'^{(1)}_{t}}+{V'^{(1)}_{t}}}r_+(\sum_{j\neq 1}{S^{(j)}_{t}})
\\&= \frac{{U^{(1)}_{t}}}{n}\frac{\tilde v_1 - ({V^{(1)}_{t}}+{R^{(1)}_{t}})}{{U^{(1)}_{t}}+\tilde v_1-{V^{(1)}_{t}}}r_-(\sum_{j\neq 1}{S^{(j)}_{t}})
+\frac{{V^{(1)}_{t}}}{n}\frac{\tilde u_1 - ({U^{(1)}_{t}}+{R^{(1)}_{t}})}{\tilde u_1-{U^{(1)}_{t}}+{V^{(1)}_{t}}}r_+(\sum_{j\neq 1}{S^{(j)}_{t}}).
\end{align*}
Thus, by a direct calculation, \begin{align*}
&\mathbb E_{\sigma_0,\sigma_0'}({R^{(1)}_{t+1}}-{R^{(1)}_{t}}|{\sigma_{t}},{\sigma'_{t}})=b-a
\\&=\frac{-{R^{(1)}_{t}}}{n}\biggl(r_+(\sum_{j\neq 1}{S^{(j)}_{t}})+r_-(\sum_{j\neq 1}{S^{(j)}_{t}})\biggr)= \frac{-{R^{(1)}_{t}}}{n}.
\end{align*}
Moreover, on the event $\{{\sigma_{t}},{\sigma'_{t}} \in \Theta\}$, $(\tilde u_1,\tilde v_1,\dots,\tilde u_m,\tilde v_m) \in \tilde \Lambda$ implies ${U^{(1)}_{t}}\leq \tilde u_1 -|J_1|/16 \leq 3|J_1|/4 -|J_1|/16 = 11|J_1|/16$, and $\tilde u_1 -{U^{(1)}_{t}} \leq 3|J_1|/4 -|J_1|/16 = 11|J_1|/16$.
The same upper bound holds for $\tilde v_1-{V^{(1)}_{t}}$ and ${V^{(1)}_{t}}$.
Thus, on the event $\{{\sigma_{t}},{\sigma'_{t}} \in \Theta\}$, \begin{align*}
\mathbb P_{\sigma_0,\sigma_0'}({R^{(1)}_{t+1}}-{R^{(1)}_{t}} \neq 0|{\sigma_{t}},{\sigma'_{t}}) &\geq b\geq \frac{p_1}{16}\frac{\frac{1}{16}r_-(\sum_{j\neq 1}{S^{(j)}_{t}})}{\frac{11}{16}+\frac{11}{16}}+\frac{p_1}{16}\frac{\frac{1}{16}r_+(\sum_{j\neq 1}{S^{(j)}_{t}})}{\frac{11}{16}+\frac{11}{16}}
\\ &=\frac{p_1}{352}. \end{align*}
Similarly, for $i>1$, $\mathbb P_{\sigma_0,\sigma_0'}({R^{(i)}_{t+1}}-{R^{(i)}_{t}} \neq 0|{\sigma_{t}},{\sigma'_{t}}) \geq {p_i}/{352} \geq {p_1}/{352}>0$, which concludes the induction.
\end{proof}
\section{Upper and Lower Bounds in the high temperature regime}\label{section5}
\subsection{Upper Bound}
\begin{thm}\thlabel{upperbound}
For $\beta<\beta_{cr}$, we have \[\lim_{\gamma \to \infty }\limsup_{n \to \infty}d_n(t_n +\gamma n)=0.\]
\end{thm}
\begin{proof}
Let $\nu$ be the stationary measure for the $2m$-coordinate chain.
For any $A\subseteq \mathcal U$, \begin{align*}
|\mathbb P_{\mathbf u}(\mathbf U_t \in A)-\nu(A)| &= \Big|\sum_{\mathbf u' \in \mathcal U} \nu(\mathbf u') \left(\mathbb P_{\mathbf u}(\mathbf U_t \in A)-\mathbb P_{\mathbf u'}(\mathbf U_t' \in A)\right) \Big|
\\ &\leq \sum_{\mathbf u' \in \mathcal U} \nu(\mathbf u') \|\mathbb P_{\mathbf u}(\mathbf U_t \in \cdot)-\mathbb P_{\mathbf u'}(\mathbf U_t' \in \cdot)\|_{TV}
\\ &\leq \max_{\mathbf u' \in \mathcal U}\|\mathbb P_{\mathbf u}(\mathbf U_t \in \cdot)-\mathbb P_{\mathbf u'}(\mathbf U_t' \in \cdot)\|_{TV}.
\end{align*}
Thus, taking supremum over $A\subseteq \mathcal U$ and $\mathbf u\in \tilde \Lambda$,
\begin{align*}
\max_{\mathbf{u}\in \tilde \Lambda}\|\mathbb P_{\mathbf u}\left(\mathbf U_t\in \cdot \right)-\nu\|_{TV} \leq \max_{\substack{\mathbf{u}\in \tilde \Lambda,\\ \mathbf{u}' \in \mathcal U}}\| \mathbb P_{\mathbf{u}}\left(\mathbf U_t\in \cdot \right)-\mathbb P_{\mathbf{u}'}\left(\mathbf U_t'\in \cdot \right)\|_{TV}.
\end{align*}
Also, from inequality \eqref{distance} and \thref{totalvariationdistance}, \begin{align*}
d_n(\delta n+t) &\leq \max_{\sigma \in \tilde \Omega} \|\mathbb P_{\sigma }({\sigma_{t}} \in \cdot) -\mu\| +O(1/n)
\\&= \max_{\mathbf{u} \in \tilde \Lambda }\|\mathbb P_{\mathbf{u}}(\mathbf U_t\in \cdot )-\nu\|_{TV} +O(1/n).
\end{align*}
For $2m$-coordinate chains $\mathbf U_t$ and $\mathbf U_t'$ with respect to a fixed $\tilde \sigma\in \tilde \Omega$ starting at $\mathbf u \in \mathcal U$ and $\mathbf u'\in \mathcal U$, respectively, put \[\tau_{tot,c} \colonequals \min\{t \geq 0 :\mathbf U_t= \mathbf U_t'\}.\]
It is a standard fact \parencite[Section 5.2]{Peres} that \begin{align*}
\|\mathbb P_{\mathbf u}(\mathbf U_t\in \cdot )-\mathbb P_{\mathbf u'}(\mathbf U_t'\in \cdot )\|_{TV}
\leq \mathbb P_{\mathbf u,\mathbf u'}(\tau_{tot,c} >t).
\end{align*}
Combining all the above results, it suffices to bound \[
\max_{\substack{\mathbf{u}\in \tilde \Lambda,\\ \mathbf{u}' \in \mathcal U}}\mathbb P_{\mathbf u,\mathbf u'}(\tau_{tot,c} >t).\]
With the above considerations, fix a good starting configuration $\tilde \sigma \in \tilde \Omega$ with the associated $2m$-coordinates $ \tilde {\mathbf u}=(\tilde u_1,\tilde v_1,\dots, \tilde u_m,\tilde v_m) \in \tilde \Lambda$ and an arbitrary starting configuration $\sigma' \in \Omega$.
Put \[t_n(\gamma) \colonequals t_n +\gamma n , \enspace H_M \colonequals \{\tau_{mag}\leq t_n(\gamma)\}.\]
The first step is the magnetization coupling phase.
By \thref{magcoupling}, there exists a coupling $({\sigma_{t}}, {\sigma'_{t}})$ for $t\leq t_n(\gamma)$ with starting configurations $(\tilde \sigma, \sigma')$ such that \[\mathbb P_{\tilde \sigma,\sigma'}(H_M^c) \leq O(1/ \sqrt{\gamma}).\]
The next step is the $2m$-coordinate chain coupling phase.
For $i=1,\dots,m$, define \begin{align*}
&\tau_{i,c} \colonequals \min\{t \geq 0 : ({U^{(i)}_{t}},V^{(i)}_t)=({U'^{(i)}_{t}},V'^{(i)}_t)\},
\\ &\Theta_i \colonequals \Big\{\sigma \in \Omega : \min \{U _i(\sigma), \tilde u_i-U _i(\sigma), V_i(\sigma), \tilde v_i-V_i(\sigma)\} \geq \frac{|J_i|}{16}\Big\},
\\ &H_{i}(t) \colonequals \{{\sigma^{(i)}_t}, {\sigma'^{(i)}_t} \in \Theta_i \},
\quad H_i \colonequals \bigcap_{t\in [t_n(\gamma), t_n(2\gamma)]}H_i (t), \quad H_{tot} \colonequals \bigcap_{i=1}^mH_i.
\end{align*}
We have defined the two coordinate chains with respect to $\tilde \sigma$.
On the event $H_M$, for $t \geq t_n(\gamma)$, we use the coupling in \thref{postmagcoupling}, while on the event $H_M ^c$, we let the chains run independently for $t \geq t_n(\gamma)$ since we do not care about this un-probable event.
Our first claim is that
\begin{align*}
\mathbb P_{\tilde\sigma,\sigma'}(H_i^c) \leq \gamma O(1/n), \enspace i=1,\dots,m.
\end{align*}
To that end, observe that
\begin{equation*}
\begin{split}
\{{\sigma^{(i)}_t}\notin \Theta_i\} \subseteq \{{U^{(i)}_{t}}<|J_i|/16\} &\cup \{\tilde u_i-{U^{(i)}_{t}}<|J_i|/16\}
\\& \cup \{V^{(i)}_t<|J_i|/16\}\cup \{\tilde v_i-V^{(i)}_t<|J_i|/16\}.
\end{split}
\end{equation*}
Notice $\tilde u_i \geq |J_i|/4$ implies
\begin{align*}
\{{U^{(i)}_{t}}<|J_i|/16\}&\subseteq \{\tilde u_i-{U^{(i)}_{t}}> 3|J_i|/16\},
\\ \{\tilde u_i-{U^{(i)}_{t}}<|J_i|/16\}&\subseteq\{{U^{(i)}_{t}}> 3|J_i|/16\}.
\end{align*}
Similarly, $\tilde v_i \geq |J_i|/4$ implies
\begin{equation*}
\begin{split}
\{V^{(i)}_t<|J_i|/16\}&\subseteq \{\tilde v_i-V^{(i)}_t> 3|J_i|/16\},
\\ \{\tilde v_i-V^{(i)}_t<|J_i|/16\}&\subseteq \{V^{(i)}_t> 3|J_i|/16\}.
\end{split}
\end{equation*}
Put \[\tilde A_{i}\colonequals \{k\in J_i : \tilde \sigma (k)=1\},\enspace i=1,\dots,m.\]
Then, following the notation in \thref{expectedmagbound}, $|M_t(\tilde A_i)|=|{U^{(i)}_{t}}-(\tilde u_i - {U^{(i)}_{t}})|$ implies \[\{{U^{(i)}_{t}}<|J_i|/16\}\cup \{\tilde u_i-{U^{(i)}_{t}}<|J_i|/16\} \subseteq \{|M_t(\tilde A_i)|\geq |J_i|/8\}.\]
Similarly, $|M_t(J_i\setminus \tilde A_i)|=|V^{(i)}_t-(\tilde v_i - V^{(i)}_t)|$ implies \[ \{V^{(i)}_t<|J_i|/16\}\cup\{\tilde v_i-V^{(i)}_t<|J_i|/16\}\subseteq \{|M_t(J_i\setminus \tilde A_i)|\geq |J_i|/8\}.\]
Combining all the above results, we obtain \[\{{\sigma^{(i)}_t} \notin \Theta_i\} \subseteq \{|M_t(\tilde A_i)|\geq |J_i|/8 \} \cup \{|M_t(J_i\setminus \tilde A_i)|\geq |J_i|/8\}.\]
A parallel argument for the primed chain shows \[\{{\sigma'^{(i)}_t} \notin \Theta_i\} \subseteq \{|M_t'(\tilde A_i)|\geq |J_i|/8 \} \cup \{|M_t'(J_i\setminus \tilde A_i)|\geq |J_i|/8\}.\]
In conclusion,
\begin{align*}
H_{i}(t) ^c&=\{{\sigma^{(i)}_t} \notin \Theta_i\} \cup \{{\sigma'^{(i)}_t} \notin \Theta_i\}
\\ & \subseteq \{|M_t(\tilde A_i)|\geq |J_i|/8 \} \cup \{|M_t(J_i\setminus \tilde A_i)|\geq |J_i|/8\}
\\ & \qquad \qquad \qquad \qquad \quad \enspace \; \cup \{|M_t'(\tilde A_i)|\geq |J_i|/8 \} \cup \{|M_t'(J_i\setminus \tilde A_i)|\geq |J_i|/8\}.
\end{align*}
Define \begin{equation*}
B\colonequals \bigcup_{t\in [t_n(\gamma), t_n(2\gamma)]} \{|M_t(\tilde A_i)|\geq |J_i|/8\} , \quad Y \colonequals \sum_{t\in [t_n(\gamma), t_n(2\gamma)]} \mathbbm 1_{\{ | M_t(\tilde A_i)| \geq |J_i|/16 \}}.
\end{equation*}
Since $M_t(\tilde A_i)$ has increments in $\{-1,0,1\}$, we have $B\subseteq \{Y\geq |J_i|/16\}$.
By Chebyshev's inequality, $\mathbb P_{\tilde\sigma,\sigma'}(B)\leq c\mathbb E_{\tilde\sigma,\sigma'}(Y)/n$ for some constant $c>0$.
From \thref{expectedmagbound}, for $t\geq t_n$, $\mathbb P_{\tilde\sigma,\sigma'}( | M_t(\tilde A_i)| \geq |J_i|/16 )=O(1/n)$, so $\mathbb E_{\tilde\sigma,\sigma'}(Y)=\gamma O(1)$.
Thus, $\mathbb P_{\tilde\sigma,\sigma'}(B)=\gamma O(1/n)$.
Similar results hold for $\bigcup_{t\in [t_n(\gamma), t_n(2\gamma)]} \{|M_t(J_i\setminus \tilde A_i)|\geq |J_i|/8\}$, $\bigcup_{t\in [t_n(\gamma), t_n(2\gamma)]} \{|M_t'( \tilde A_i)|\geq |J_i|/8\}$, and $\bigcup_{t\in [t_n(\gamma), t_n(2\gamma)]} \{|M_t'(J_i\setminus \tilde A_i)|\geq |J_i|/8\}$.
In conclusion, \begin{align*}
\mathbb P_{\tilde\sigma,\sigma'}(H_i^c)=\mathbb P_{\tilde\sigma,\sigma'}\Biggl(\bigcup_{t\in [t_n(\gamma), t_n(2\gamma)]} H_i(t)^c\Biggr) \leq 4\gamma O(1/n),
\end{align*}
which proves our first claim.
From the first claim, \[\mathbb P_{\tilde \sigma,\sigma'}(H_{tot}^c) \leq \sum_{i=1}^m \mathbb P_{\sigma,\sigma'}(H_{i}^c)= \gamma O(1/n).\]
Now, condition on the event $H_M$.
Recalling the fact that \thref{postmagcoupling} assures $\mathbf S_t= \mathbf S_t'$ for $t \geq t_n(\gamma)$ on the event $H_M$, we can make ${R^{(i)}_{t}}$ stay zero after $\tau_{i,c}$, using the modified monotone update on $J_i$ whenever a site in $J_i$ is chosen to be updated.
Thus, on $H_M$, \[\tau_{tot,c}=\max_{1\leq i\leq m} \tau_{i,c}.\]
Our second claim is that \begin{align*}
\mathbb P_{\tilde \sigma,\sigma'}(\tau_{i,c} > t_n(2\gamma) , H_i , H_M)= O(1/\sqrt{\gamma}), \enspace i=1,\dots,m.
\end{align*}
From \thref{supermartingale} and \thref{postmagcoupling}, $\mathbb P_{\tilde \sigma,\sigma'}(\tau_{i,c} > t_n(2\gamma) , H_i , H_M|\sigma_{t_n(\gamma)},\sigma_{t_n(\gamma)}')\leq {c|R^{(i)}_{t_n (\gamma)}|}/{\sqrt{n\gamma}}$ for some $c>0$.
Taking expectation yields, \[\mathbb P_{\tilde \sigma,\sigma'}(\tau_{i,c} > t_n(2\gamma) , H_i , H_M)\leq \frac{c\mathbb E_{\tilde \sigma,\sigma'}|R^{(i)}_{t_n (\gamma)}|}{\sqrt{n\gamma}}\]
However, for any $t>0$, $|{R^{(i)}_{t}}|=|U_t'-U_t|=|M_t'(\tilde A_i)-M_t(\tilde A_i)|$, so from \thref{expectedmagbound}, $\mathbb E_{\tilde \sigma,\sigma'}|R^{(i)}_{t_n (\gamma)}| \leq \mathbb E_{\sigma'}|M_{t_n(\gamma)}'(\tilde A_i)|+\mathbb E_{\tilde \sigma}|M_{t_n(\gamma)}(\tilde A_i)| =O(\sqrt{n})$, which proves our second claim.
From the second claim, \begin{align*}
&\mathbb P_{\tilde \sigma,\sigma'}(\tau_{tot,c}>t_n( 2\gamma), H_{tot}, H_M)
\leq \sum_{i=1}^m \mathbb P_{\tilde \sigma,\sigma'}(\tau_{i,c}>t_n( 2\gamma), H_{tot}, H_M)
\\&\leq \sum_{i=1}^m \mathbb P_{\tilde \sigma,\sigma'}(\tau_{i,c}>t_n( 2\gamma), H_{i}, H_M)
= O(1/\sqrt{\gamma}).
\end{align*}
Combining all the above results, \begin{align*}
&\mathbb P_{\tilde \sigma,\sigma'}(\tau_{tot,c}>t_n( 2\gamma))
\\&\leq \mathbb P_{\tilde \sigma,\sigma'}(\tau_{tot,c}>t_n( 2\gamma), H_{tot},H_M)+\mathbb P_{\tilde \sigma,\sigma'}(H_{tot}^c) +\mathbb P_{\tilde \sigma,\sigma'}(H_M^c)
\\&= O(1/\sqrt{\gamma})+\gamma O(1/n) +O(1/\sqrt{\gamma}).
\end{align*}
Finally, \begin{align*}
d_n(t_n+(2\gamma+\delta)n) \leq O(1/\sqrt{\gamma}) + \gamma O(1/n)+O(1/n),
\end{align*} which gives us the result upon taking limits.
\end{proof}
\subsection{Lower Bound} \label{subsection4.2}
We first analyze the drift of magnetization chains.
Let $1\leq i\leq m$ and $\mathcal F_t$ be the $\sigma$-algebra generated by ${S^{(1)}_{t}},\dots, {S^{(m)}_{t}}$.
By a direct calculation,
\begin{align}
\mathbb E[{S^{(i)}_{t+1}}-{S^{(i)}_{t}}|\mathcal F_t] &= \frac{2}{n}p_i\frac{|J_i|-n{S^{(i)}_{t}}}{2|J_i|}r_+(\sum_{j\neq i}{S^{(j)}_{t}}) -\frac{2}{n}p_i\frac{|J_i|+n{S^{(i)}_{t}}}{2|J_i|}r_-(\sum_{j\neq i}{S^{(j)}_{t}}) \nonumber
\\ &=\frac{2}{n}\frac{p_i-{S^{(i)}_{t}}}{2}r_+(\sum_{j\neq i}{S^{(j)}_{t}}) -\frac{2}{n}\frac{p_i+{S^{(i)}_{t}}}{2}r_-(\sum_{j\neq i}{S^{(j)}_{t}}) \nonumber
\\ &= \frac{1}{n}\biggl(-{S^{(i)}_{t}}+p_i\tanh ({\beta} \sum_{j\neq i}{S^{(j)}_{t}})\biggr). \label{dynamics}
\end{align}
The following simple lemma is the main tool to get the lower bound in \thref{lowerbound}.
\begin{lem}[Proposition 7.9, \cite{Peres}]\thlabel{statistics}
Let $f\colon\mathcal{S} \to \mathbb R$ be a measurable function and $\nu_1$, $\nu_2$ be two probability measures on $\mathcal S$.
Let $\sigma_*^2 \colonequals \max\{\mathbb{V}\mathrm{ar}_{\nu_1}f, \ \mathbb{V}\mathrm{ar}_{\nu_2}f\}$.
If $|\mathbb E_{\nu_1}f -\mathbb E_{\nu_2}f|\geq r \sigma_*$, then \[\|\nu_1 -\nu_2\|_{TV}\geq 1-\frac{8}{r^2}\]
\end{lem}
Positive starting configurations give us the following result.
\begin{lem}\thlabel{finallemma}
Let $\mathbf s\geq \mathbf 0$ be the starting magentization.
Then, for $t\geq0$, \[ \mathbb E_{\mathbf s}\|\mathbf S_t\|_1 \leq g^t\left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2}+O(1/\sqrt{n}).\]
\end{lem}
\begin{proof}
Consider the case that $|J_i|$ is odd for each $i=1,\dots,m$.
Let $\nu $ be the starting distribution such that $\mathbf s_+'=(\frac{1}{n},\dots,\frac{1}{n})$ with probability $\frac{1}{2}$ and $\mathbf s_-'=(-\frac{1}{n},\dots,-\frac{1}{n})$ with probability $\frac{1}{2}$.
By \thref{generalcontraction}, since $\mathbf s \geq\mathbf s_+'$ in this case, \begin{align*}
\mathbf 0 &\leq \mathbb E_{\mathbf s,\nu} (\mathbf S_t-\mathbf S_t')
\leq \frac{1}{2}\mathbf A^t(\mathbf s - \mathbf s_+')+\frac{1}{2}\mathbf A^t(\mathbf s - \mathbf s_-') =\mathbf A^t\mathbf s.
\end{align*}
However, $\mathbb E_{\nu} {S'^{(i)}_{t}}=0$ for $i=1,\dots,m$ by the remark after \thref{magmarkov}.
Thus, $\mathbf 0 \leq \mathbb E_{\mathbf s} \mathbf S_t \leq \mathbf A^t \mathbf s$, so by \thref{newlemma}, \begin{align*}
0 \leq \sum_{i=1}^m \mathbb E_{\mathbf s} {S^{(i)}_{t}} \leq \|\mathbf A^t \mathbf s\|_1 \leq g^t \left(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\right)^{1/2}.
\end{align*}
From \thref{magvariancebound} and Cauchy-Schwartz inequality, since $0\leq \mathbb E_{\mathbf s}{S^{(i)}_{t}}$ for $i=1,\dots,m$, \begin{align*}
&\mathbb E_{\mathbf s}\|\mathbf S_t\|_1=\sum_{i=1}^m \mathbb E_{\mathbf s}|{S^{(i)}_{t}}| \leq \sum_{i=1}^m\left(|\mathbb E_{\mathbf s}{S^{(i)}_{t}}|+\sqrt{\mathbb{V}\mathrm{ar}_{\mathbf s}{S^{(i)}_{t}}}\right)
=\sum_{i=1}^m\mathbb E_{\mathbf s}{S^{(i)}_{t}}+\sum_{i=1}^m\sqrt{\mathbb{V}\mathrm{ar}_{\mathbf s}{S^{(i)}_{t}}}
\\&\leq g^t \Biggl(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\Biggr)^{1/2}+ \Biggl(m \sum_{i=1}^m\mathbb{V}\mathrm{ar} {S^{(i)}_{t}}\Biggr)^{1/2}
= g^t \Biggl(\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}\Biggr)^{1/2} + O(\frac{1}{\sqrt{n}}).
\end{align*}
Other cases of $|J_i|$ can similarly be shown by considering $0$ instead of $\frac{1}{n}$ whenever the partition has even number of sites.
\end{proof}
Finally, we prove the lower bound.
\begin{thm}\thlabel{lowerbound}
For $\beta< \beta_{cr}$, we have \[\lim_{\gamma \to \infty }\liminf_{n \to \infty}d_n(t_n -\gamma n)=1.\]
\end{thm}
\begin{proof}
Since the magnetization chain is a projection of the original chain, it suffices to provide a lower bound on the total variation norm of the magnetization chain.
Using $\tanh x \geq x-x^2/3$ for $x\in\mathbb R$, from equations \eqref{dynamics}, we have \begin{align*}
\mathbb E({S^{(i)}_{t+1}}|\mathcal F_t) &\geq (1-\frac{1}{n}){S^{(i)}_{t}}+ \frac{p_i}{n}\Biggl(\beta \sum_{j\neq i}{S^{(j)}_{t}} - \frac{\beta^2 (\sum_{j\neq i}{S^{(j)}_{t}})^2}{3}\Biggr)
\end{align*}
for each $i=1,\dots,m$.
In the matrix form, \begin{align*}
\mathbb E(\mathbf S_{t+1}|\mathcal F_t) &\geq \mathbf A \mathbf S_t - \mathbf x
\end{align*}
where $\mathbf x = \frac{\beta ^2}{3n} (p_1(\sum_{j\neq 1}{S^{(j)}_{t}})^2,\dots, p_m(\sum_{j\neq m}{S^{(j)}_{t}})^2)^T$.
Recall the definition of $\mathbf a^T \colonequals (a_1, \dots, a_m)>\mathbf 0$ with $\|\mathbf a\|_1=1$ being the left eigenvector of $\mathbf A$ with eigenvalue $g$.
Then $\mathbb E(\mathbf a^T \mathbf S_{t+1}|\mathcal F_t )\geq \mathbf a^T \mathbf A\mathbf S_t - \mathbf a^T \mathbf x= g\mathbf a^T \mathbf S_t - \mathbf a^T \mathbf x$, i.e.,
\begin{equation}\label{compute}
\mathbb E\Bigl(\sum_{i=1}^m a_i{S^{(i)}_{t+1}}|\mathcal F_t\Bigr) \geq g \sum_{i=1}^m a_i {S^{(i)}_{t}}- \frac{\beta^2}{3n} \sum_{i=1}^m a_ip_i\biggl(\sum_{j\neq i}{S^{(j)}_{t}}\biggr)^2 .
\end{equation}
Observe that \begin{align*}
\sum_{i=1}^m a_ip_i\biggl(\sum_{j\neq i}{S^{(j)}_{t}}\biggr)^2 \leq \sum_{k=1}^m a_kp_k\bigg( \sum_{j=1}^m |S^{(j)}_{t}| \bigg)^2= \biggl(\sum_{k=1}^m a_k p_k\biggr)\|\mathbf S_t\|_1^2.
\end{align*}
Thus, upon taking expectation in equation \eqref{compute}, \begin{align*}
\mathbb E\left(\sum_{i=1}^m a_i{S^{(i)}_{t+1}}\right) \geq g \mathbb E\left(\sum_{i=1}^m a_i {S^{(i)}_{t}}\right)- \frac{\beta^2}{3n}\left(\sum_{i=1}^m a_ip_i\right)\mathbb E \|\mathbf S_t\|_1^2 .
\end{align*}
We claim that, \begin{align*}
\mathbb E\|\mathbf S_t\|_1^2 \leq (\mathbb E\|\mathbf S_t\|_1)^2 + O(1/n).
\end{align*}
Since $\mathbb E\|\mathbf S_t\|_1^2 = (\mathbb E\|\mathbf S_t\|_1)^2 + \mathbb{V}\mathrm{ar}\|\mathbf S_t\|_1$, it suffices to show $\mathbb{V}\mathrm{ar}\|\mathbf S_t\|_1 \leq O(1/n)$.
However, from \thref{magvariancebound}, \begin{align*}
\mathbb{V}\mathrm{ar}\|\mathbf S_t\|_1 &=\sum_{i=1}^m \mathbb{V}\mathrm{ar}|{S^{(i)}_{t}}| + 2\sum_{i> j}\mathrm{Cov}(|{S^{(i)}_{t}}|,|{S^{(j)}_{t}}|)
\\&\leq \sum_{i=1}^m \mathbb{V}\mathrm{ar}{S^{(i)}_{t}} + 2\sum_{i>j}\sqrt{\mathbb{V}\mathrm{ar}{S^{(i)}_{t}}}\sqrt{\mathbb{V}\mathrm{ar}{S^{(j)}_{t}}}
\\&\leq \sum_{i=1}^m \mathbb{V}\mathrm{ar}{S^{(i)}_{t}} + \sum_{i>j}(\mathbb{V}\mathrm{ar}{S^{(i)}_{t}}+\mathbb{V}\mathrm{ar}{S^{(j)}_{t}})=m\sum_{i=1}^m\mathbb{V}\mathrm{ar}{S^{(i)}_{t}}=O(1/n), \end{align*}
which proves the claim.
Put $Z_t \colonequals \sum_{i=1}^m a_i {S^{(i)}_{t}}/g^t$.
Then, from the claim above, \begin{align*}
\mathbb EZ_{t+1}-\mathbb EZ_t\geq -\frac{\beta^2 \sum_{i} a_ip_i}{3ng^{t+1}}\left((\mathbb E\|\mathbf S_t\|_1)^2 + O(1/n)\right).
\end{align*}
Assume that $\mathbf s \geq \mathbf 0$ is a non-negative starting magnetization.
Recalling the definition $\upsilon \colonequals n(1-g)$, from \thref{finallemma} and the fact $\sum_{i} {(s^{(i)})^2}/{p_i} \leq 1$, \begin{align*}
\mathbb E_{\mathbf s}Z_{t+1}-\mathbb E_{\mathbf s}Z_t &\geq -\frac{\beta^2\sum_{i} a_ip_i}{3ng^{t+1}}\Biggl(\biggl(g^t\biggl(\sum_{i} {(s^{(i)})^2}/{p_i}\biggr)^{1/2}+O(1/\sqrt{n})\biggr)^2+O(1/n)\Biggr)
\\&\geq -\frac{\beta^2\sum_{i} a_ip_i }{3(n-\upsilon)}\Biggl(g^t\sum_{i} {(s^{(i)})^2}/{p_i}+O(1/\sqrt{n})+\frac{1}{g^t}O(1/n)\Biggr).
\end{align*}
Iterating from $0$ to $t-1$, \begin{align*}
\mathbb E_{\mathbf s}Z_{t}-Z_0 &\geq -\frac{\beta^2\sum_{i}a_ip_i}{3(n-\upsilon)}\left(\frac{1-g^t}{\upsilon/n}\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}+tO(1/\sqrt{n})+\frac{n-\upsilon}{\upsilon}(\frac{1}{g^t}-1)O(1/n)\right)
\\&=-\frac{\beta^2\sum_{i}a_ip_i}{3\upsilon(1-\upsilon/n)}(1-g^t)\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}-\frac{\beta^2\sum_{i}a_ip_i}{3(n-\upsilon)}tO(1/\sqrt{n})
\\&\quad\;-\frac{\beta^2\sum_{i}a_ip_i}{3\upsilon}(\frac{1}{g^t}-1)O(1/n).
\end{align*}
For brevity, let us prefer to use $\upsilon$ rather than use $\beta_{cr}$ in view of \thref{upsilon}.
Consider the step $t_*\colonequals t_n-\gamma n/\upsilon=\frac{1}{2\upsilon} n\ln n- \frac{\gamma n}{\upsilon}$.
Observe that $1-1/x \geq e^{-1/(x-1)}$ for $x>1$ implies \[g^{t_*} \geq \frac{e^{\gamma}}{n^{n/(2(n-\upsilon))}}.\]
Then \begin{align*}
\mathbb E_{\mathbf s}Z_{t_*}-\sum_{i=1}^ma_is_i \geq &-\frac{\beta^2\sum_{i}a_ip_i}{3\upsilon(1-{\upsilon}/{n})}\left(1-\frac{e^{\gamma }}{n^{{n}/{(2(n-\upsilon))}}}\right)\sum_{i=1}^m \frac{(s^{(i)})^2}{p_i}
\\&-\frac{\beta^2\sum_{i}a_ip_i}{3(n-\upsilon)}\left(\frac{1}{2\upsilon} n\ln n- \frac{\gamma n}{\upsilon}\right)O({1}/{\sqrt{n}})
\\&-\frac{\beta^2\sum_{i}a_ip_i}{3\upsilon}\left(\frac{n^{n/(2(n-\upsilon))}}{e^{\gamma }}-1\right)O({1}/{n}).
\end{align*}
The right-hand side of the above inequality converges to $-\frac{\beta^2\sum_{i}a_ip_i\sum_{i} {(s^{(i)})^2}/{p_i}}{3\upsilon}$ as $n\to \infty$ for every $\gamma>0$.
We claim that if $n$ is large enough, then there exists $ \mathbf s > \mathbf 0 $ such that \[\sum_{i=1}^ma_is_i-\frac{\beta^2\sum_{i}a_ip_i\sum_{i} {(s^{(i)})^2}/{p_i}}{3\upsilon}>0.\]
Consider $\mathbf s =\zeta \mathbf p$ where $0<\zeta <1$ is a constant to be determined.
We want to find $\zeta$ such that \[\sum_{i=1}^ma_ip_i\zeta-\frac{\beta^2\sum_{i}a_ip_i\sum_{i} {(p_i\zeta)^2}/{p_i}}{3\upsilon}>0,\] which is equivalent to \[3\upsilon>\beta^2 \zeta.\]
From \thref{upsilon}, $\upsilon >0$, so $\frac{3\upsilon}{\beta^2} \mathbf p>\mathbf s>\mathbf 0 $ assures that the inequality in the claim holds, and such a positive magnetization $\mathbf s \in\mathcal S$ exists since $n$ is large and $0 \leq \beta<\beta_{cr}$ (if $\beta=0$, choose $\mathbf s= \mathbf p$).
By the last claim, for large $n$, there exists $\mathbf s \in \mathcal S $ and $\varepsilon>0$ such that \[\mathbb E_{\mathbf s}(\sum_{i=1}^m a_iS^{(i)}_{t_*}) \geq 2\varepsilon g^{t_*}\geq 2\varepsilon \frac{e^{\gamma}}{n^{n/(2(n-\upsilon))}}\geq \varepsilon\frac{e^{\gamma}}{\sqrt{n}}. \]
\thref{magvariancebound} and the Cauchy-Schwartz inequality imply $\mathbb{V}\mathrm{ar}(\sum_{i=1}^m a_iS^{(i)}_{t_*}) = O(\frac{1}{n})$ as $n\to \infty$.
Thus, by \thref{zerospin} and \thref{statistics}, for some $c>0$, \begin{align*}
\lim _{\gamma \to \infty }\liminf_{n\to \infty}d_n(t_n- \frac{\gamma n}{\upsilon} ) \geq \lim_{\gamma \to \infty}1-\frac{c}{\varepsilon^2 e^{2\gamma}}=1.
\end{align*}
\end{proof}
\section{Exponentially slow mixing in the low temperature regime}\label{section6}
Using a standard bottleneck ratio argument, we can show that the mixing time for the Glauber dynamics is exponential in the low temperature regime.
The bottleneck ratio is defined as \[\Phi \colonequals \min_{A: \mu(A) \leq 1/2} \frac{\sum_{x\in A,y\notin A} \mu(x)P(x,y)}{\mu(A)}\] where $P$ is the transition matrix of the Glauber dynamics.
The bottleneck ratio gives a lower bound of the mixing time (see \parencite[Theorem 7.4]{Peres}): \[t_{\mathrm{mix}}\geq \frac{1}{4\Phi}.\]
We need another characterization of the critical temperature $\beta_{cr}$.
\begin{lem}\thlabel{slowmixinglemma}
We have that \[\beta_{cr}= \frac{\sum_i a_i^2 p_i}{(\sum_i a_i p_i)^2 - \sum_i a_i ^2 p_i ^2}\]
\end{lem}
\begin{proof}
From $\mathbf a^T \mathbf A= g \mathbf a^T$, equation \eqref{betacrdef}, and \thref{upsilon}, we have \[\sum_i a_i p_i=\Big(p_k+\frac{1}{\beta_{cr}} \Big)a_k\] for each $k=1,\dots,m$.
Multiplying $a_k p_k$ to both sides and summing over $k$ yields the result.
\end{proof}
\begin{proof}[Proof of \thref{lowtempresult}]
It suffices to show that $\Phi \leq c_1 \exp(-c_2 n)$ for some positive constants $c_1, c_2 >0$.
By symmetry of the Hamiltonian, we have that $\mu (A) \leq 1/2$ where $A \colonequals \{\sigma : \sum_i S^{(i)}(\sigma) > 0 \}$.
Since the only way to go from $A$ to $A^c$ is to go through $B\colonequals \{\sigma : | \sum_i S^{(i)}(\sigma) | \leq 1/n \}$, it holds that \[\sum_{x\in A,y\notin A} \mu(x)P(x,y) \leq \mu(B).\]
Note that for any $\sigma \in \Omega$, \[\mu(\sigma)=\frac{\exp \bigg(\frac{\beta n}{2}\Big(\big(\sum_i S^{(i)}(\sigma)\big)^2-\sum_i \big(S^{(i)}(\sigma)\big)^2\Big)\bigg)}{Z(\beta)}.\]
By the Cauchy-Schwartz inequality, \[\mu(B)\leq \binom{n}{\lceil n/2 \rceil} \frac{\exp \Big(\frac{\beta n}{2}\big(1-\frac{1}{m}\big) \big(\frac{1}{n}\big)^2\Big)}{Z(\beta)}\lesssim \binom{n}{\lceil n/2 \rceil}/Z(\beta)\] where $\lesssim$ denotes that the inequality holds for sufficiently large $n$ up to a constant not depending on $n$.
Using Stirling's formula, \[\Phi \lesssim \frac{\exp(n\ln 2)}{Z(\beta)\mu(A)}.\]
Now, consider the configurations with exactly $k_i n p_i$ many "$+$" spins in $J_i$ where $1/2\leq k_i\leq 1$ for each $i=1,\dots,m$ and there exists at least one $i$ such that $1/2 <k_i$.
These configurations are members of $A$ and there are at least $\prod_{i=1} ^m\binom{np_i}{k_inp_i}$ many such configurations.
Using Stirling's formula again, we obtain \[
Z(\beta)\mu(A) \gtrsim \Bigg(\frac{1}{\prod_{i=1}^m (1-k_i)^{p_i(1-k_i)}k_i^{p_i k_i}}\Bigg)^n e^{ \frac{\beta n}{2}\big((\sum_i (2k_i -1)p_i)^2-\sum_i (2k_i -1)^2p_i^2\big)}
.\]
Define a function $f$ through the equation \[e^{nf(k_1,\dots,k_m)}\colonequals \Bigg(\frac{1}{\prod_{i=1}^m (1-k_i)^{p_i(1-k_i)}k_i^{p_i k_i}}\Bigg)^n e^{ \frac{\beta n}{2}\big((\sum_i (2k_i -1)p_i)^2-\sum_i (2k_i -1)^2p_i^2\big)}. \]
Put $(k_1,\dots,k_m)= (1/2,\dots,1/2)+\gamma (v_1,\dots,v_m)$ where $v_i \geq 0$ for each $i=1,\dots,m$, $\gamma \in \mathbb R$, and $\sum_i v_i ^2 \neq 0$.
Fixing $v_i$'s, we can regard $f$ as a one-variable function of $\gamma$, say $f=f(\gamma)$, and this is equivalent to fixing a direction in $\mathbb R^m$.
A little calculation shows that \begin{align*}
f(\gamma)&=2\beta \gamma^2 \bigg(\Big(\sum_i v_i p_i\Big)^2-\sum_i v_i^2 p_i^2 \bigg)
\\& \quad -\sum_i p_i \big((1/2-\gamma v_i)\ln(1/2-\gamma v_i)+ (1/2+\gamma v_i)\ln(1/2 +\gamma v_i)\big)
\\f'(\gamma)&= 4\beta \gamma \bigg(\Big(\sum_i v_i p_i\Big)^2-\sum_i v_i^2 p_i^2 \bigg)-\sum_i p_i v_i\big(-\ln(1/2-\gamma v_i)+ \ln(1/2 +\gamma v_i)\big)
\\ f''(\gamma)&=4\beta \bigg(\Big(\sum_i v_i p_i\Big)^2-\sum_i v_i^2 p_i^2 \bigg)-\sum_ip_iv_i^2 \bigg(\frac{1}{1/2-\gamma v_i}+\frac{1}{1/2 +\gamma v_i}\bigg)
\end{align*}
where $'$ denotes a differentiation in $\gamma$.
Note that $f(0)= \ln 2$ and $f'(0)=0$.
Thus, it suffices to show that there is a direction $(v_1,\dots,v_m)$ such that $f''(0)>0$.
\thref{slowmixinglemma} shows that the direction $(v_1,\dots,v_m)=(a_1,\dots,a_m)$ satisfies $f''(0)>0$ whenever $\beta>\beta_{cr}$, which completes the proof.
\end{proof}
\begin{rmk}
Combined with the non-exponential mixing time of $O(n\ln n)$ whenever $\beta <\beta_{cr}$, the above proof shows that $\inf_{\mathbf v\geq \mathbf 0, \mathbf v\neq \mathbf 0}\frac{\sum_i v_i^2 p_i}{(\sum_i v_i p_i)^2 - \sum_i v_i ^2 p_i ^2}$ is achieved with the direction $(v_1,\dots,v_m)=\mathbf a^T$.
\end{rmk}
\begin{ack}
The author would like to thank Professor Insuk Seo for introducing the problem and sharing his limitless insight through numerous discussions. The author also acknowledges an anonymous user at \textbf{math.stackexchange.com} \footnote{https://math.stackexchange.com/q/3553425} for the main idea of the proof in \thref{newlemma}.
Finally, the author acknowledges the anonymous reviewers for their helpful comments and careful reading of the paper.
\end{ack}
\printbibliography
\end{document}
|
\section{Conclusion}\label{sec:conclsuion}
In this paper we presented a decentralized multi-robot MPC-based motion planning approach that accounts for the robot's interactions with obstacles and other robots through the use of a RNN-based trajectory prediction model. We showed that our proposed interaction-aware RNN model generalizes well with different number of robots and obstacles, and is able to provide more accurate trajectory predictions than the constant velocity model in a variety of scenarios. In simulation with six quadrotors, we showed that our decentralized planner outperforms the planner using a constant velocity model for trajectory prediction and can achieve a comparable level of performance to the centralized sequential planner while being communication-free. We also validated our approach in real-world experiments with three quadrotors flying in a shared space with walking humans. Future work shall take into account sensing uncertainties and consider more complex unstructured environments with static obstacles.
\section{Introduction}\label{sec:intro}
Autonomous navigation of a team of robots in dynamic environments is important when deploying them in various applications such as coverage and inspection \cite{breitenmoser2016combining}, search and rescue \cite{baxter2007multi}, formation flying \cite{Zhu2019ICRA} and multi-view videography \cite{Nageli2017}. In these scenarios, the robots navigate in a shared space that may also have moving obstacles. To achieve predictive collision avoidance and ensure safety, each robot needs to know the future motion predictions of other robots in the environment. These motion predictions can be obtained among robots by sharing their future planned trajectories with each other via communication \cite{Zhu2019RAL}. However, such communication may not be available nor reliable in practice. Alternatively, some approaches \cite{Kamel2017} employ a constant velocity model to predict other robots' trajectories. Even though communication among robots is not required, the planned robot motions may not be safe, particularly in crowded dynamic environments \cite{Zhu2019RAL}.
In this paper, we propose an interaction- and obstacle-aware trajectory prediction model and combine it with the model predictive control (MPC) framework to achieve decentralized multi-robot motion planning in dynamic environments. Fig. \ref{fig:overview} gives an overview of the proposed method. In particular, we first generate a demonstration dataset consisting of robot trajectories using a multi-robot collision avoidance simulator \cite{Zhu2019RAL}. It utilizes a centralized sequential MPC for local motion planning in which inter-robot communication is employed. Next, we formulate the robot trajectory prediction problem as a sequence modeling task and hence design a model based on recurrent neural networks (RNN).
By training the model using the generated dataset, it learns to imitate the centralized sequential MPC and thus can predict the planning behaviors of the robots.
Finally by combining the trajectory prediction model with the MPC framework, multi-robot local motion planning is achieved in a decentralized manner.
\begin{figure}[t]
\centering
\setlength{\textfloatsep}{0.7\baselineskip plus 0.2\baselineskip minus 0.5\baselineskip}
\includegraphics[width=0.86\linewidth]{method_overview.pdf}
\caption{The proposed decentralized communication-free motion planner that relies on a RNN model for interaction-aware trajectory prediction and a MPC for local motion planning.}
\label{fig:overview}
\end{figure}
The main contributions of this work are:
\begin{itemize}
\item A RNN-based interaction- and obstacle-aware model that is able to provide robot trajectory predictions in a multi-robot scenario.
\item Incorporation of the trajectory prediction model with MPC to achieve decentralized multi-robot local motion planning in dynamic environments.
\end{itemize}
We show that our designed model can make accurate trajectory predictions, thanks to which the proposed decentralized multi-robot motion planner can achieve a comparable level of performance to the centralized planner while being communication-free. We also validate our approach with a team of quadrotors in real-world experiments.
\section{Related Work}\label{sec:relatedWork}
\subsection{Multi-Robot Collision Avoidance}
We focus our work on online local motion planning for multi-robot systems (also referred as multi-robot collision avoidance), which has been actively studied over the past years.
Traditional reactive controller-level approaches include the optimal reciprocal collision avoidance (ORCA) method \cite{vandenBerg2011} that builds on the concept of reciprocal velocity obstacles (RVO) \cite{vandenBerg2008}, artificial potential field (APF) based method \cite{yongjie2009collision}, buffered Voronoi cell (BVC) approach \cite{Zhou2017,Zhu2019MRS}, and control barrier functions (CBF) \cite{Wang2017}. While these \rebuttal{reactive} methods are computationally efficient, the robot dynamics are not fully modeled and the robot motion is typically limited by only planning one time step ahead. Recently, there have emerged new learning-based methods for multi-robot collision avoidance, such as deep imitation learning \cite{Shi2020,Riviere2020} and those that are reinforcement learning (RL) based \cite{Chen2017a,semnani2020multi}.
\rebuttal{
RL-based methods can learn policies that have a long-term cumulative reward for the robots and thus are considered to be non-myopic \cite{Chen2017a}. However, they are generally not able to handle hard state constraints, such as collision avoidance constraints.}
These issues can be overcome by using the model predictive control (MPC) framework for collision-free trajectory generation in which an optimization problem is solved for each robot in a receding horizon manner. In this paper, we study the multi-robot MPC-based collision avoidance problem.
For each robot to solve a local trajectory optimization problem, it needs to know the future trajectories of other robots. One approach is to let each robot communicate its planned trajectory with every other robot in the team. Hence, robots can then update their own trajectories to be collision free with other robots' trajectory plans, as in these distributed MPC works \cite{Zhu2019RAL,Luis2020}. While these methods can achieve safe collision avoidance, the communication burden across the team is large and may not be available nor reliable in practice \cite{Serra2020}. Another approach is to let each robot predict other robots' future motions based on its own observations. For instance, \cite{Kamel2017} employs a constant velocity model when predicting other robots' future trajectories. In that case, communication among robots is not required. However, such a prediction can be inaccurate and may lead to unsafe trajectory planning \cite{Zhu2019RAL}. In this paper, we will develop an interaction- and obstacle-aware model for the trajectory prediction taking into account surrounding information of the robot to model the interaction and environment constraints. By incorporating the model with the MPC framework, we can achieve safe and communication-free decentralized collision avoidance for multiple robots in dynamic environments.
\subsection{Motion Prediction}
\rebuttal{
Our proposed approach decouples motion prediction and trajectory planning to achieve decentralized and communica- tion-free collision avoidance. Such a decoupling is also seen in \cite{Schmerling2018,Fridovich-Keil2020}, where the motion prediction of humans are used to plan a safe trajectory for the ego robot.}
Motion prediction for decision-making agents has drawn significant research efforts over the past years, with most works focusing on human trajectory prediction \cite{Rudenko2020}. Early works on motion prediction are typically model-based such as the renowned social force-based method \cite{Helbing1995} which models pedestrian behaviors through the use of attractive and repulsive potentials. The model is later generalized and adapted to modeling traffic car behaviors \cite{helbing1998generalized}. While these methods are computationally efficient, the prediction accuracy is quite low. There have also been several notable attempts to utilize game theory to model interacting decision-making agents and predict their future trajectories \cite{Turnwald2016, Oyler2016}, in which the agents are assumed to play a non-cooperative game and their trajectory predictions can be obtained from computing the Nash equilibria of the game.
While interaction-aware trajectory predictions can be obtained, these methods are limited to specific road scenarios and cannot be directly applied to general multi-robot systems.
The class of approaches that have achieved state-of-the-art performance in trajectory prediction problems are the learning-based methods. Some of these include inverse reinforcement learning (IRL) \cite{Kuderer2015}, recurrent neural networks (RNN) \cite{Alahi2016,Brito2020}, variational autoencoders \cite{Lee2017}, generative adversarial networks (GAN) \cite{Gupta2018} \rebuttal{that provide predicted human trajectories in two-dimensional (2D) environments, Gaussian mixture regression (GMR) \cite{mainprice2013human} and Gaussian process regression (GPR) \cite{park2019planner} that can predict human actions in 3D workspaces.}
\rebuttal{
Our approach of predicting the trajectories of other robots is based on previous works on human motion prediction since both can be formulated as a sequence modelling problem. In particular, our prediction model is based on RNN, inspired by the works in \cite{Pfeiffer2018} for interaction-aware pedestrian motion prediction in which static obstacles are considered and represented using a grid map. We adapt the model to predict robots trajectories in multi-robot scenarios with moving obstacles described by their positions and velocities, and further apply the model to decentralized multi-robot motion planning by incorporating it within MPC.
}
\section{Preliminaries}\label{sec:preliminary}
Throughout this paper vectors are denoted in bold lowercase letters, $\mathbf{x}$, matrices in plain uppercase, $M$, and sets in calligraphic, $\mathcal{S}$. $\norm{\mathbf{x}}$ denotes the Euclidean norm of $\mathbf{x}$ and $\norm{\mathbf{x}}_{Q}^{\rebuttal{2}} = \mathbf{x}^TQ\mathbf{x}$ denotes the weighted squared norm.
\subsection{Robot and Obstacle Model}\label{subsec:robot_obs_model}
Following \cite{Zhu2019RAL}, we consider a team of $n$ robots moving in a shared workspace $\mathcal{W} \subseteq \mathbb{R}^3$, where each robot $i \in \mathcal{I} = \{ 1,2,\dots, n \} \subset \mathbb{N} $ is modeled as an enclosing sphere with radius $r$. The robots follow the same dynamical model that is described by a discrete-time equation as follows,
\begin{equation}\label{eq:nonDyn}
\begin{aligned}
\mathbf{x}_i^{k+1} = \mathbf{f}(\mathbf{x}_i^k, \mathbf{u}_i^k), \quad \mathbf{x}_i^0 = \mathbf{x}_i(0),
\end{aligned}
\end{equation}
where $\mathbf{x}_i^k \in \mathbb{X} \subset \mathbb{R}^{n_x}$ denotes the state of the robot, typically including its position $\mathbf{p}_i^k$ and velocity $\mathbf{v}_i^k$, and $\mathbf{u}_i^k \in \mathbb{U} \subset \mathbb{R}^{n_u}$ the control inputs at time $k$. Without loss of generality, $k=0$ indicates the current time. $\mathbb{X}$ and $\mathbb{U}$ are the admissible state space and control space, respectively. $\mathbf{x}_i(0)$ is the current state of robot $i$.
In addition, moving obstacles for example pedestrians in the environment are considered. For each obstacle $o \in \mathcal{I}_o = \{1,2,\dots,n_o \} \subset \mathbb{N} $ at position $\mathbf{p}_o \in \mathbb{R}^3$, we model it as an upright non-rotating enclosing \emph{ellipsoid} centered at $\mathbf{p}_o$ with semi-principal axes $(a, b, c)$.
\rebuttal{
In this paper, we assume that each robot can observe the states (positions and velocities) of all other robots and moving obstacles and keep their history information.}
\subsection{Multi-Robot Collision Avoidance}\label{subsec:mrca}
Multi-robot local motion planning is considered in this paper, in which the goal is to achieve real-time collision-free navigation for multiple robots. Each robot has a given goal location $\mathbf{g}_i$, which generally comes from some high-level path planner \cite{Honig2018} or is specified by the user. Any pair of robots $i$ and $j$ from the group are mutually collision-free if $\norm{\mathbf{p}_i^k - \mathbf{p}_j^k} \geq 2r, \forall i\neq j \in \mathcal{I}, k = 0, 1, \dots $. Regarding robot-obstacle collision avoidance, we approximate the obstacle with an enlarged ellipsoid and check if the robot's position is inside it. Hence, the robot $i$ is collision-free with the obstacle $o$ at time step $k$ if $\norm{\mathbf{p}_i^k-\mathbf{p}_o^k}_{\Omega} \geq 1$, where $\Omega = \text{diag}(1/(a+r)^2, 1/(b+r)^2, 1/(c+r)^2)$.
The objective is to compute a local motion $\mathbf{u}_i^k$ for each robot in the group, that respects its dynamic constraints, makes progress towards its goal location $\mathbf{g}_i$ and is collision-free with other robots in the group as well as moving obstacles within a planning time horizon.
\subsection{Model Predictive Control}\label{subsec:mpc}
The multi-robot collision avoidance problem can be solved using model predictive control by formulating a receding horizon constrained optimization problem. For each robot $i \in \mathcal{I}$, a discrete-time constrained optimization formulation with $N$ time steps and planning horizon $N\Delta t$, where $\Delta t$ is the sampling time, is derived as follows,
\begin{subequations}\label{eq:dmpc}
\begin{alignat}{2}
\min_{\substack{\mathbf{x}_i^{0:N}, \mathbf{u}_i^{0:N-1}, \\ s^{0:N}}} ~~
& \sum_{k=0}^{N-1} J_i^k(\mathbf{x}_i^k, \mathbf{u}_i^k, s^k) + J_i^N(\mathbf{x}_i^N, \mathbf{g}_i, s^N) \\
\text{s.t.} ~~ & \mathbf{x}_i^0 = \mathbf{x}_i(0), \\
& \mathbf{x}_i^{k} = \mathbf{f}(\mathbf{x}_i^{k-1}, \mathbf{u}_i^{k-1}), \\
& \norm{\mathbf{p}_i^k - \mathbf{p}_j^k} \geq 2r - s^{k}, \label{subeq:inter_robot_ca}\\
& \norm{\mathbf{p}_i^k - \mathbf{p}_o^k}_{\Omega} \geq 1 - s^{k},\label{subeq:robot_obs_ca}\\
& s^{k} \geq 0, \mathbf{u}_i^{k-1} \in \mathbb{U}, \mathbf{x}_i^k \in \mathbb{X},\\
&\forall j \neq i\in\mathcal{I}; \, \forall o\in\mathcal{I}_o; \, \forall k\in \{1,\dots,N\},
\end{alignat}
\end{subequations}
where $J_i^k(\mathbf{x}_i^k, \mathbf{u}_i^k, s^k)$ and $J_i^N(\mathbf{x}_i^N, \mathbf{g}_i, s^N)$ are the stage and terminal costs, and $s$ is the slack variable. At each time step, each robot in the team solves online the constrained optimization problem (\ref{eq:dmpc}) and then executes the first step control inputs, in a receding horizon fashion.
Note that for each robot to solve the optimization problem (\ref{eq:dmpc}), it has to know the future trajectories of other robots and moving obstacles, as shown in Eq. (\ref{subeq:inter_robot_ca}) and Eq. (\ref{subeq:robot_obs_ca}). For moving obstacles (pedestrians), we assume their motions follow a constant velocity model (CVM) in the short planning horizon and predict their future trajectories accordingly.
\rebuttal{
This assumption is reasonable since CVM can achieve state-of-the-art performance when used for pedestrian motion prediction \cite{Scholler2020}.}
For robots' future trajectories, denote by $\mathcal{T}_{i}^0 = \{ \mathbf{p}_i^{1:N} \}$ the robot $i$'s current planned trajectory. Further denote by $\hat{\mathcal{T}}_{i,j}^0 = \{\mathbf{p}_j^{1:N}\}$ the trajectory of robot $j\in\mathcal{I}, j\neq i$ that robot $i$ assumes and uses in solving the problem (\ref{eq:dmpc}), where the hat $\hat{\cdot}$ indicates that it is robot $i$'s estimation of the other robot's trajectory.
Typically, there are two ways for robot $i$ to obtain the future trajectory of the other robot $j$. The first way is via communication: all robots in the team communicate their planned trajectories to each other at each time step. It can be implemented using a centralized sequential planning framework as in \cite{Zhu2019RAL}, that is, $\hat{\mathcal{T}}_{i,j}^0 = \mathcal{T}_j^0$. Although this method guarantees collision avoidance by construction, it does not scale well with a large number of robots. Moreover, communication is not always available and reliable in practice.
The other way is without communication. Hence, robot $i$ has to predict another robot $j$'s future trajectory based on its observation of the environment:
\begin{equation}\label{eq:prediction_general}
\hat{\mathcal{T}}_{i,j}^0 = \tn{\tb{prediction}}(\mathcal{H}_i^0),
\end{equation}
where $\mathcal{H}_i^0$ is the information that robot $i$ can acquire until current time from its observation. Previous works \cite{Zhu2019RAL,Kamel2017} use a constant velocity model to perform the prediction only based on the other robot's current state, that is, $\mathcal{H}_i^0 = \mathbf{x}_j^0$. However, such a prediction can be inaccurate and may lead to unsafe trajectory planning \cite{Zhu2019RAL}. In this paper, we will develop an interaction- and obstacle-aware model for the trajectory prediction taking into account surrounding environment information of the robot to model the interaction and environment constraints.
\section{Approach}\label{sec:method}
In this section, we present our interaction and obstacle-aware trajectory prediction method and incorporate it with the MPC framework to achieve decentralized multi-robot collision avoidance in dynamic environments.
\subsection{Trajectory Prediction Problem Formulation}
As shown in Eq. (\ref{eq:prediction_general}), robot $i\in\mathcal{I}$ needs to predict the future trajectories of other robots $j \neq i \in \mathcal{I}$ to plan its safe motion. Hereafter, we refer to the robot $i$ as the ego robot and the robot $j$ as the \emph{query robot} that is indicated by the sub-script $\cdot_q$. In addition, we use the sub-script $\cdot_{-q}$ to indicate the collection of all the other robots except for the query robot.
We aim to address the problem of finding a trajectory prediction model for the query robot $q$ that gives a sequence of its future positions $\mathbf{p}_q^{1:T_H}$ in a multi-robot scenario. Here $T_H \geq N$ is the prediction horizon that should not be smaller than the local motion planning horizon. As has been shown in previous trajectory prediction works \cite{Pfeiffer2018,Brito2020}, we will instead work with sequences of velocities $\mathbf{v}_q^{1:T_H}$ for prediction to avoid overfitting when based on position sequences, and numerically integrate them afterwards starting from the query robot's current position $\mathbf{p}_q^0$.
Denote by $\mathbf{v}_q^{-T_O:0}$ the past sequence of velocities of the query robot within an observation time $T_O \geq 1$.
Denote by $\mathbf{p}_{-q,r}^{-T_O:0}$ and $\mathbf{v}_{-q,r}^{-T_O:0}$ the past relative positions and velocities of other robots with respect to the query robot.
Further denote by $\mathbf{p}_{\mathcal{I}_o,r}^{0}$ and $\mathbf{v}_{\mathcal{I}_o,r}^{0}$ the current relative positions and velocities of the moving obstacles $o\in\mathcal{I}_o$ with respect to the query robot. By observing history states of the query robot and its surrounding other robots as well as moving obstacles, we want to find an interaction- and obstacle-aware model $\mathbf{h}_{\boldsymbol\theta}$ with parameters $\boldsymbol\theta$:
\begin{equation}\label{eq:prediction_formulation}
\mathbf{v}_q^{1:T_H} = \mathbf{h}_{\boldsymbol\theta}(\mathbf{v}_q^{-T_O:0}, \mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0}, \mathbf{p}_{\mathcal{I}_o,r}^{0}, \mathbf{v}_{\mathcal{I}_o,r}^{0}),
\end{equation}
that outputs a prediction of the query robot's future states.
\subsection{Demonstration Data Generation}\label{subsec:dataset}
We use a simulation dataset to train our designed network model. The dataset is generated using demonstrations from a multi-robot collision avoidance simulator \cite{Zhu2019RAL} which employs a centralized sequential planner to solve the problem (\ref{eq:dmpc}). This involves each robot solving a MPC problem sequentially and communicates its planned trajectory to other robots to avoid.
Note that the planner differs from the prioritized planning approach since each robot has to avoid all other robots and hence it shows cooperation among robots.
Specifically, we create a three-dimensional environment in which a team of robots and moving obstacles are simulated.
In the simulation, each robot navigates to a randomly generated goal position, which is changed dynamically to a new location after being reached. The generated robots' goal positions are ensured to be collision-free with each other and the obstacles. Moving obstacles are simulated in the environment by randomly specifying an initial position and velocity \rebuttal{(with speed between 0.5 m/s and 1.2 m/s)} to each of them and then make them move at a constant velocity. Once any obstacle moves out of the environment, a new initial position and velocity will be set to it. Moreover, we add small Gaussian noise to the velocities of the moving obstacles in simulation. We perform the simulation for $N_{\tn{sim}}$ time steps and record the positions and velocities of all robots and obstacles at each time step.
After running the simulation, for each time step $t$ and robot $q$, we retrieve its future sequence of velocities and observation of the past states of the system from the recorded data. Hence, our dataset is as follows
\begin{align}
\mathcal{D} = \{ (\mathcal{H}_q^t, \mathbf{v}_q^{t+1:t+T_H}) | \forall q \in \mathcal{I}, \forall t \in \{1,\dots, N_{\tn{sim}-T_H}\} \},
\end{align}
where the observation information $\mathcal{H}_q^t$ is
\begin{equation}
\mathcal{H}_i^t = \{\mathbf{v}_q^{t-T_O:t}, \mathbf{p}_{-q,r}^{t-T_O:t}, \mathbf{v}_{-q,r}^{t-T_O:t}, \mathbf{p}_{\mathcal{I}_o,r}^{t}, \mathbf{v}_{\mathcal{I}_o,r}^{t}\}.
\end{equation}
\subsection{Interaction- and Obstacle-Aware Model}
We now present our recurrent neural network (RNN) model for interaction- and obstacle-aware trajectory prediction, as shown in Fig. \ref{fig:architecture}. The model first creates a joint representation of three input channels: the query robot's history state, information of other interacting robots and moving obstacles, via a query robot state encoder and an environment encoder module. Then a decoder module is adopted to output a predicted trajectory of the query robot. The recurrent layers in the model are of the LSTM type \cite{Hochreiter1997} that has been shown able to learn time dependencies over a long period of time.
Next, we describe the three main modules of the model in detail.
\begin{figure*}[t]
\centering
\setlength{\textfloatsep}{0.1\baselineskip plus 0.0\baselineskip minus 0.4\baselineskip}
\includegraphics[width=0.82\linewidth]{network_architecture}
\caption{Network architecture of the interaction- and obstacle-aware model. Three channels of information are taken as inputs: the query robot's past velocities $\mathbf{v}_q^{-T_O:0}$, past relative states of other robots $(\mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0})$ and current relative states of obstacles $(\mathbf{p}_{\mathcal{I}_o,r}^0, \mathbf{v}_{\mathcal{I}_o,r}^0)$. A joint representation of the inputs is created through a query robot encoder and an environment encoder. A decoder is adopted to output a sequence of velocities $\mathbf{v}_q^{1:T_H}$ predicted for the query robot's future trajectory.}
\label{fig:architecture}
\end{figure*}
\subsubsection{Query robot state encoder}
It consists of a recurrent layer that produces a flat encoding $\mathbf{z}_q^0$ from the history velocities of the query robot $\mathbf{v}_q^{-T_O:0}$.
This layer learns a dynamical model of the query robot, so that the network can leverage it to obtain better predictions.
\subsubsection{Environment encoder module}
It includes $n-1$ parallel recurrent layers with shared weights to encode the sequences of past relative positions and velocities of other robots with respect to the query robot $(\mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0})$ into a set $\mathcal{Z}_{-q,r}^{0}$, and $n_o$ parallel dense layers with shared weights that encode the current relative positions and velocities of moving obstacles with respect to the query robot $(\mathbf{p}_{\mathcal{I}_o,r}^{0}, \mathbf{v}_{\mathcal{I}_o,r}^{0})$ into a set $\mathcal{Z}_{\mathcal{I}_o,r}^{0}$. The encodings from both of these sets, which are made to have the same length, are then stacked together and followed by a global max pooling operation executed along the new data axis. Thus, this module can capture the interaction of the query robot with a variable number of other robots and obstacles in the environment and encode it into a single flat vector $\mathbf{z}_e^0$.
This framework also makes it possible to account for potentially different types of agents and obstacles by training their own set of encoders and stacking them with the rest of intermediate encodings.
\subsubsection{Decoder module}
It takes in the concatenation of $\mathbf{z}_q^0$ with $\mathbf{z}_e^0$ and passes it through a recurrent decoder followed by a dense decoder and an output layer that finally generates a sequence of predicted future velocities $\mathbf{v}_{q}^{1:T_H}$ for the query robot over the prediction horizon.
\subsection{Model Training}
Using the generated demonstration data in Section \ref{subsec:dataset}, the designed model is trained end-to-end using back-propagation through time (BTTP) \cite{werbos1990backpropagation} with a fixed truncation depth $t_{\tn{trunc}}$. We learn the trajectory prediction model by minimizing the following loss function,
\begin{equation} \label{eq:mse}
\begin{aligned}
L(\mathbf{v}_q^{1:T_H}, \boldsymbol\theta) = \frac{1}{T_H}\sum_{k=1}^{T_H}\norm{\mathbf{v}_q^{k} - \mathbf{v}_{q, \tn{true}}^{k}}^2 + \lambda\cdot l(\boldsymbol\theta),
\end{aligned}
\end{equation}
where $\mathbf{v}_{q,\tn{true}}^k$ is the ground truth velocity from the demonstration dataset, $l(\boldsymbol\theta)$ represents the regularization terms and $\lambda$ is the regularization factor. In our model, the $L2$ regularization method is adopted.
\subsection{Decentralized Multi-Robot Motion Planning}
Having the trained trajectory prediction model, we can incorporate it with the MPC framework and solve the problem (\ref{eq:dmpc}) in a decentralized manner. As shown in Fig. \ref{fig:overview}, in a multi-robot navigation scenario, each robot first performs inference with the trained neutral network to predict the future trajectories of its neighboring robots and then plans a collision-free trajectory accordingly. Hence, decentralized multi-robot motion planning in dynamic environments is achieved. To be able to perform the inference, each robot needs to measure its own state as well as its neighbors', and keep a history memory of the information for a time horizon $T_O$. In addition, the robot also needs to measure the current states of moving obstacles in the environment.
\section{Results}\label{sec:result}
We now present results of simulation comparing the proposed approach with other methods and real-world experiments with quadrotors.
A video accompanying this paper includes additional simulation and experimental results.
\subsection{Implementation Details}
To generate the dataset, we use an existing MATLAB multi-robot collision avoidance simulator\footnote{Code: \href{https://github.com/tud-amr/mrca-mav}{\color{blue}{https://github.com/tud-amr/mrca-mav}}} \cite{Zhu2019RAL} and simulate $N_{\tn{sim}} = 10^5$ time steps in a $10\times 10 \times 3$ m environment with \rebuttal{10 robots and 10 moving obstacles}. The robot we simulate is the Parrot Bebop 2 quadrotor with a radius set as 0.4 m.
Ellipsoids representing the moving obstacles have semi-axes $(0.4, 0.4, 0.9)$ m. The sampling time and MPC planning horizon length are $\Delta t = 0.05$ s and $N = 20$, respectively.
\rebuttal{
We employ the same dynamics model and cost functions in the MPC problem (2) of our previous work \cite{Zhu2019RAL}.}
The Forces Pro solver \cite{domahidi2014forces} is used to solve the MPC problem. We set $T_O = 20$ and $T_H = 20$ the horizon length for robot past states observation and trajectory prediction. We further generate another test dataset by running the simulator in \rebuttal{six} different scenarios for $2\times 10^4$ time steps for each one of them.
\rebuttal{
All computations are performed in a commodity computer with an Intel i7 CPU@2.60GHz and an NVIDIA GTX 1060 GPU.}
The designed learning network is implemented in Python using TensorFlow 2. All layers in the network have 64 neurons except for the recurrent decoder that has 128 neurons and the output layer that has 3 neurons. While the activation function of the output layer is linear, all other layers in the network use a hyperbolic activation function. The regularization factor used during model training is $\lambda = 0.01$.
\begin{figure*}[t]
\centering
\setlength{\textfloatsep}{0.2\baselineskip plus 0.0\baselineskip minus 0.4\baselineskip}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_4_0.pdf}
\caption{4 quads, no obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_10_0.pdf}
\caption{10 quads, no obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_10_10.pdf}
\caption{10 quads, 10 obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_20_0.pdf}
\caption{\rebuttal{20 quads, no obstacles.}}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_20_10.pdf}
\caption{\rebuttal{20 quads, 10 obstacles.}}
\end{subfigure}
\caption{Performance results of our proposed interaction-aware RNN model for trajectory prediction compared to the baselines. The solid lines represent the average errors along the prediction horizon and the filled patches around them are 30\% of the standard deviation. The sampling period is 50 ms and the prediction horizon has 20 timesteps. }
\label{fig:prediction_results}%
\end{figure*}
\subsection{Trajectory Prediction Evaluation}
We first evaluate our trajectory prediction model on a test dataset that has not been used for training nor validation. The dataset includes different test scenarios: an open environment with 4, 10, and \rebuttal{20} quadrotors, and with 10 moving obstacles. We compare our interaction-aware RNN-based model to three alternative methods: a) the constant velocity model (CVM) that is widely used in decentralized multi-robot motion planning; b) a simple RNN model that only considers the query robot's past states for trajectory prediction while ignoring its surrounding environment (this allows us to highlight the interaction awareness of our designed model); and c) an open-loop MPC planner assuming that the goal, robot model and constraints are known.
In Fig. \ref{fig:prediction_results} we present quantitative results of the prediction error with respect to ground truth in the test dataset. Recall that ground truths are the recorded robot traveled trajectories computed with the centralized sequential MPC (closed-loop). As expected, the prediction error of the open-loop MPC has the smallest prediction error among the methods since it was used for data generation and has perfect knowledge about the goal locations of all robots, which are not available for prediction in our proposed RNN-based model. Our proposed model can still achieve accurate trajectory predictions and significantly outperforms the CVM method across all scenarios. Moreover, compared to the simple RNN model, our interaction-aware approach achieves more accurate trajectory predictions, particularly in cluttered scenarios where interactions among robots are more frequent, as shown in Fig. \ref{fig:prediction_results}(b)-(e).
\rebuttal{
Furthermore, to evaluate the generalization capability of the learned network, we perform simulations in the scenarios (d) and (e) with 20 quadrotors which are beyond our training dataset. The results show that the proposed model still performs well on trajectory prediction in the two scenarios.}
\subsection{Decentralized Motion Planning}
We then evaluate performance of the proposed decentralized planner that incorporates the learned trajectory prediction model.
\subsubsection{\rebuttal{Comparisons to other methods}}
We compare our method to the centralized sequential planning method \cite{Zhu2019RAL} with full communication among robots and the decentralized planning method \cite{Kamel2017} that uses the constant velocity model (CVM) for trajectory prediction to analyze whether more accurate trajectory forecasts of our RNN-based model lead to better planning performance. \rebuttal{Besides, another decentralized method, the buffered Voronoi cell (BVC) \cite{Zhou2017}, which guarantees collision avoidance is also implemented for comparison.}
Six quadrotors flying in four types of scenarios that represents different levels of difficulty \cite{Serra2020} are considered. Moreover, in order to avoid potential bias results, each scenario includes 50 instances where the robots have different starting and goal locations. The four scenarios are: \emph{1) symmetric swap}, in which the robots initially located at the vertices of a virtual horizontal regular hexagon are required to exchange their positions; \emph{2) asymmetric swap}, which differs from the previous scenario in that the hexagons are irregular, thus leading to more challenging collision-avoidance problems; \emph{3) pair-wise swap}, in which the robots are placed at random starting positions and assigned to three pairs within which the two robots need to swap their positions; and \emph{4) random moving}, in which each robot moves from a random starting position to a random goal in the environment.
Qualitatively, Fig. \ref{fig:sim_planning} shows the sample trajectory trails of the six quadrotors for one instance from the asymmetric swap scenario. \rebuttal{It can be seen that our RNN-based decentralized planner achieves results that are closer to the centralized sequential planner than the CVM-based planner.}
To quantitatively evaluate the performance of different motion planners, we consider a wide range of metrics: the number of instances that lead to collisions within the entire 50 runs, the average trajectory length and trajectory duration of the team of robots, and the overall robot average speed during the whole simulation. The last three metrics are only computed for those successful runs. Table \ref{tab:planning_comparison} summaries the simulation results. It can be seen that our RNN-based planner significantly outperforms the planner using the CVM for trajectory prediction in terms of safety, in particular in the challenging asymmetric swap scenario. In addition, \rebuttal{our planner also achieves consistently smaller trajectory lengths and durations compared to the CVM-based planner} in all scenarios.
\rebuttal{Compared to the BVC method, our proposed approach achieves significantly shorter trajectory durations, particularly in the (a)symmetric swapping scenarios, which shows superiority of the MPC framework over the reactive BVC method.}
Finally, compared to the centralized sequential planner with full communication, our planner can achieve a comparable level of performance in terms of safety and trajectory efficiency while being decentralized and communication-free. However, three instances out of 50 in the challenging asymmetric swap scenario is still observed with collisions using the RNN-based method, indicating that in few rare cases, highly-accurate trajectory predictions of other robots, for example obtained via communication, are necessary to ensure safety.
\rebuttal{
In the simulation, on average the computation time of the proposed decentralized MPC planner with the learned predictor is 36.3 ms, which is smaller than that of the centralized sequential planner which plans trajectories for all six robots (43.9 ms). Besides, our decentralized approach is communication-free.}
\begin{figure}[t]
\centering
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_cent_xy.pdf}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_CVM_xy.pdf}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_RNN_xy.pdf}
\end{subfigure}
\\
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_cent_xz.pdf}
\caption{Cen. Comm.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_CVM_xz.pdf}
\caption{Decen. CVM}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_RNN_xz.pdf}
\caption{Decen. RNN}
\end{subfigure}
\caption{Simulation results of six quadrotors exchanging positions in the asymmetric swap scenario. Solid lines represent the trajectories. The upper and lower plots show the top view (X-Y) and side view(X-Z), respectively.}
\label{fig:sim_planning}
\end{figure}
\begin{table*}[t]
\centering
\caption{Performance comparison of different multi-robot motion planners (centralized with communication \cite{Zhu2019RAL}, \rebuttal{decentralized buffered Voronoi cell (BVC) method \cite{Zhou2017}}, decentralized with constant velocity model (CVM) \cite{Kamel2017} and decentralized with our RNN-based model) across the four different types of scenarios (symmetric swap, asymmetric swap, pair-wise swap and random moving). Each scenario includes 50 running instances. }
\begin{tabular}{ll||c|ccc|ccc|c}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{\textbf{Scenario}}} & \multicolumn{1}{c||}{\multirow{2}{*}{\textbf{Motion Planner}}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Num. of coll. \\ instances\end{tabular}}} & \multicolumn{3}{c|}{\textbf{Trajectory length (m)}} & \multicolumn{3}{c|}{\textbf{Trajectory duration (s)}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Average speed\\ (m/s)\end{tabular}}} \\
\multicolumn{1}{c}{} & \multicolumn{1}{c||}{} & & Min. & Average & Max. & Min. & Average & Max. & \\ \hline \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Symmetric\\ swap\end{tabular}} & Cen. Comm. & {0} & 5.46 & 7.22$\pm$0.85 & 8.96 & 5.15 & 5.62$\pm$0.24 & 6.05 & 1.27$\pm$0.05 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{5.75} & \rebuttal{7.63$\pm$0.87} & \rebuttal{9.63} & \rebuttal{10.00} & \rebuttal{12.21$\pm$1.25} & \rebuttal{14.80} & \rebuttal{0.63$\pm$0.03} \\
& Decen. (CVM) & 4 & 5.43 & 7.45$\pm$0.88 & 9.55 & 4.95 & 6.02$\pm$0.55 & 7.30 & 1.23$\pm$0.05 \\
& {Decen. (RNN)} & \tb{{0}} & {5.41} & \tb{7.35$\pm$0.91} & {10.60} & {4.75} & \tb{5.59$\pm$0.40} & {9.70} & \tb{1.30$\pm$0.05} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Asymmetric\\ swap\end{tabular}} & Cen. Comm. & {0} & 5.08 & 6.77$\pm$0.80 & 9.02 & 4.65 & 5.17$\pm$0.30 & 5.85 & 1.30$\pm$0.05 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{5.31} & \rebuttal{7.25$\pm$0.87} & \rebuttal{9.29} & \rebuttal{9.70} & \rebuttal{11.34$\pm$1.00} & \rebuttal{13.90} & \rebuttal{0.65$\pm$0.03} \\
& Decen. (CVM) & 15 & 5.32 & 7.76$\pm$1.89 & 18.06 & 5.05 & 5.96$\pm$0.51 & 7.30 & 1.28$\pm$0.04 \\
& {Decen. (RNN)} & {3} & {5.16} & \tb{7.14$\pm$1.10} & {12.60} & {4.80} & \tb{5.48$\pm$0.42} & {6.85} & \tb{1.29$\pm$0.05} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Pair-wise\\ swap\end{tabular}} & Cen. Comm. & {0} & 1.64 & 5.10$\pm$1.87 & 9.92 & 3.50 & 4.76$\pm$0.44 & 5.85 & 1.06$\pm$0.12 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{1.83} & \rebuttal{5.25$\pm$1.93} & \rebuttal{10.13} & \rebuttal{5.20} & \rebuttal{7.58$\pm$1.66} & \rebuttal{13.20} & \rebuttal{0.67$\pm$0.07} \\
& Decen. (CVM) & 3 & 1.74 & 5.54$\pm$2.53 & 17.42 & 3.60 & 5.00$\pm$0.65 & 5.75 & \tb{1.06$\pm$0.10} \\
& {Decen. (RNN)} & \tb{0} & {1.70} & \tb{4.94$\pm$2.02} & {9.94} & {3.40} & \tb{4.83$\pm$0.45} & {5.95} & {1.01$\pm$0.14} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Random\\ moving\end{tabular}} & Cen. Comm. & {0} & 0.39 & 4.66$\pm$1.94 & 8.63 & 3.50 & 7.72$\pm$0.40 & 5.50 & 0.98$\pm$0.13 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{0.39} & \rebuttal{4.81$\pm$2.00} & \rebuttal{8.84} & \rebuttal{4.70} & \rebuttal{7.13$\pm$1.30} & \rebuttal{10.10} & \rebuttal{0.69$\pm$0.09} \\
& Decen. (CVM) & \tb{0} & 0.39 & 4.82$\pm$2.06 & 9.53 & 3.60 & 4.89$\pm$0.57 & 6.35 & \tb{0.98$\pm$0.12} \\
& {Decen. (RNN)} & \tb{0} & {0.39} & \tb{4.36$\pm$2.11} & {9.19} & {3.85} & \tb{4.76$\pm$0.43} & {5.95} & {0.91$\pm$0.11} \\
\hline
\end{tabular}
\label{tab:planning_comparison}
\end{table*}
\rebuttal{
\subsubsection{Effect of non-MPC robots on performance}
Our proposed decentralized approach assumes that all robots interact and adopt the same motion planning strategy, namely MPC-based trajectory optimization with the learned motion prediction model. We now evaluate the performance of our approach in a mixture scenario where some robots employ the BVC method \cite{Zhou2017} for collision avoidance.
We simulate 50 instances with six quadrotors in the symmetric swap scenario of Section V-C-1. Table \ref{table:nonMPC_robot} presents the simulation results. When there is only one BVC robot, no collisions are observed. However, when more BVC robots are in the team, particularly when half of the robots (3) are BVC-based, collisions will happen due to incorrect motion predictions of them by other MPC robots. This indicates that the assumption that the robots interact with the same planning strategy is necessary to ensure safety.
\begin{table}[t]
\centering
\caption{\rebuttal{Simulation results of six quadrotors in the symmetric swap scenario where a varying number of BVC-based robots are in the team. 50 running instances are simulated.}}
\begin{tabular}{lcccccc}
\hline
Num. of BVC robots & 0 & 1 & 2 & 3 & 4 \\ \hline
Num. of coll. instan. & 0 & 0 & 2 & 4 & 1 \\
Ave. traj. time (s) & 5.59 & 6.82 & 8.517 & 9.63 & 10.27 \\ \hline
\end{tabular}
\label{table:nonMPC_robot}
\end{table}
}
\subsection{Experimental Validation}
\subsubsection{Setup}
We validate our proposed approach with a team of Parrot Bebop 2 quadrotors flying in a shared space with walking human obstacles. The pose of each quadrotor and obstacle (human) is obtained using an external motion capture system (OptiTrack) and their velocities obtained via a standard Kalman filter running at a high rate. Control commands are sent to the quadrotor via ROS.
\rebuttal{
During the experiment, the humans walked at a speed with mean 0.8 m/s and the maximum 1.2 m/s. They could change their speeds and make small turns in the workspace.
}
\begin{figure}[t]
\centering
~
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=1.0\textwidth]{exp_gopro.png}
\caption{}
\label{subfig:exp_gopro}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=1.0\textwidth]{exp_matlab.pdf}
\caption{}
\label{subfig:exp_matlab}
\end{subfigure}
\begin{subfigure}{0.21\textwidth}
\includegraphics[width=1.0\textwidth]{3q_dis_time.pdf}
\caption{}
\label{subfig:quad_dis}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=1.0\textwidth]{3q_obs_dis_his.pdf}
\caption{}
\label{subfig:obs_dis}
\end{subfigure}
\caption{Experimental results with three quadrotors flying in a shared space with two walking humans. (a) A snapshot of the experiment. (b) Schematic of quadrotors, humans, and planned trajectories. (c) Distance between the quadrotors over time. The shaded grey area indicates the two walking humans join the space. (d) Histogram of the quadrotor-obstacle distance during the experiments.}
\label{fig:exp_results}%
\end{figure}
\subsubsection{Results}
Experiments in two representative scenarios are conducted: with and without obstacles. In the first scenario, three quadrotors, initially distributed in a virtual horizontal circle, are required to swap their positions multiple times. Then in the second scenario, two moving obstacles (walking humans) join the space while the three quadrotors keeps changing their positions while avoiding the humans at the same time. Fig. \ref{subfig:exp_gopro} presents a snapshot from the experiment.
Fig. \ref{subfig:quad_dis} shows distance between each pair of the three quadrotors over time during the experiment. It can be seen that they maintained a safe distance of 0.8 m over the entire run even after the two walking humans joined the space which makes it more confined. In Fig. \ref{subfig:obs_dis} we cumulate the distance between each quadrotor and human obstacle that is computed as the closest distance from the quadrotor center to the obstacle ellipsoid's surface. The results show that a minimum safe separation of 0.4 m to the obstacles is achieved. In sum, the demonstration shows that our proposed approach works well for multi-robot motion planning in dynamic environments which is decentralized and communication-free.
\section{Conclusion}\label{sec:conclsuion}
In this paper we presented a decentralized multi-robot MPC-based motion planning approach that accounts for the robot's interactions with obstacles and other robots through the use of a RNN-based trajectory prediction model. We showed that our proposed interaction-aware RNN model generalizes well with different number of robots and obstacles, and is able to provide more accurate trajectory predictions than the constant velocity model in a variety of scenarios. In simulation with six quadrotors, we showed that our decentralized planner outperforms the planner using a constant velocity model for trajectory prediction and can achieve a comparable level of performance to the centralized sequential planner while being communication-free. We also validated our approach in real-world experiments with three quadrotors flying in a shared space with walking humans. Future work shall take into account sensing uncertainties and consider more complex unstructured environments with static obstacles.
\section{Introduction}\label{sec:intro}
Autonomous navigation of a team of robots in dynamic environments is important when deploying them in various applications such as coverage and inspection \cite{breitenmoser2016combining}, search and rescue \cite{baxter2007multi}, formation flying \cite{Zhu2019ICRA} and multi-view videography \cite{Nageli2017}. In these scenarios, the robots navigate in a shared space that may also have moving obstacles. To achieve predictive collision avoidance and ensure safety, each robot needs to know the future motion predictions of other robots in the environment. These motion predictions can be obtained among robots by sharing their future planned trajectories with each other via communication \cite{Zhu2019RAL}. However, such communication may not be available nor reliable in practice. Alternatively, some approaches \cite{Kamel2017} employ a constant velocity model to predict other robots' trajectories. Even though communication among robots is not required, the planned robot motions may not be safe, particularly in crowded dynamic environments \cite{Zhu2019RAL}.
In this paper, we propose an interaction- and obstacle-aware trajectory prediction model and combine it with the model predictive control (MPC) framework to achieve decentralized multi-robot motion planning in dynamic environments. Fig. \ref{fig:overview} gives an overview of the proposed method. In particular, we first generate a demonstration dataset consisting of robot trajectories using a multi-robot collision avoidance simulator \cite{Zhu2019RAL}. It utilizes a centralized sequential MPC for local motion planning in which inter-robot communication is employed. Next, we formulate the robot trajectory prediction problem as a sequence modeling task and hence design a model based on recurrent neural networks (RNN).
By training the model using the generated dataset, it learns to imitate the centralized sequential MPC and thus can predict the planning behaviors of the robots.
Finally by combining the trajectory prediction model with the MPC framework, multi-robot local motion planning is achieved in a decentralized manner.
\begin{figure}[t]
\centering
\setlength{\textfloatsep}{0.7\baselineskip plus 0.2\baselineskip minus 0.5\baselineskip}
\includegraphics[width=0.86\linewidth]{method_overview.pdf}
\caption{The proposed decentralized communication-free motion planner that relies on a RNN model for interaction-aware trajectory prediction and a MPC for local motion planning.}
\label{fig:overview}
\end{figure}
The main contributions of this work are:
\begin{itemize}
\item A RNN-based interaction- and obstacle-aware model that is able to provide robot trajectory predictions in a multi-robot scenario.
\item Incorporation of the trajectory prediction model with MPC to achieve decentralized multi-robot local motion planning in dynamic environments.
\end{itemize}
We show that our designed model can make accurate trajectory predictions, thanks to which the proposed decentralized multi-robot motion planner can achieve a comparable level of performance to the centralized planner while being communication-free. We also validate our approach with a team of quadrotors in real-world experiments.
\section{Related Work}\label{sec:relatedWork}
\subsection{Multi-Robot Collision Avoidance}
We focus our work on online local motion planning for multi-robot systems (also referred as multi-robot collision avoidance), which has been actively studied over the past years.
Traditional reactive controller-level approaches include the optimal reciprocal collision avoidance (ORCA) method \cite{vandenBerg2011} that builds on the concept of reciprocal velocity obstacles (RVO) \cite{vandenBerg2008}, artificial potential field (APF) based method \cite{yongjie2009collision}, buffered Voronoi cell (BVC) approach \cite{Zhou2017,Zhu2019MRS}, and control barrier functions (CBF) \cite{Wang2017}. While these \rebuttal{reactive} methods are computationally efficient, the robot dynamics are not fully modeled and the robot motion is typically limited by only planning one time step ahead. Recently, there have emerged new learning-based methods for multi-robot collision avoidance, such as deep imitation learning \cite{Shi2020,Riviere2020} and those that are reinforcement learning (RL) based \cite{Chen2017a,semnani2020multi}.
\rebuttal{
RL-based methods can learn policies that have a long-term cumulative reward for the robots and thus are considered to be non-myopic \cite{Chen2017a}. However, they are generally not able to handle hard state constraints, such as collision avoidance constraints.}
These issues can be overcome by using the model predictive control (MPC) framework for collision-free trajectory generation in which an optimization problem is solved for each robot in a receding horizon manner. In this paper, we study the multi-robot MPC-based collision avoidance problem.
For each robot to solve a local trajectory optimization problem, it needs to know the future trajectories of other robots. One approach is to let each robot communicate its planned trajectory with every other robot in the team. Hence, robots can then update their own trajectories to be collision free with other robots' trajectory plans, as in these distributed MPC works \cite{Zhu2019RAL,Luis2020}. While these methods can achieve safe collision avoidance, the communication burden across the team is large and may not be available nor reliable in practice \cite{Serra2020}. Another approach is to let each robot predict other robots' future motions based on its own observations. For instance, \cite{Kamel2017} employs a constant velocity model when predicting other robots' future trajectories. In that case, communication among robots is not required. However, such a prediction can be inaccurate and may lead to unsafe trajectory planning \cite{Zhu2019RAL}. In this paper, we will develop an interaction- and obstacle-aware model for the trajectory prediction taking into account surrounding information of the robot to model the interaction and environment constraints. By incorporating the model with the MPC framework, we can achieve safe and communication-free decentralized collision avoidance for multiple robots in dynamic environments.
\subsection{Motion Prediction}
\rebuttal{
Our proposed approach decouples motion prediction and trajectory planning to achieve decentralized and communica- tion-free collision avoidance. Such a decoupling is also seen in \cite{Schmerling2018,Fridovich-Keil2020}, where the motion prediction of humans are used to plan a safe trajectory for the ego robot.}
Motion prediction for decision-making agents has drawn significant research efforts over the past years, with most works focusing on human trajectory prediction \cite{Rudenko2020}. Early works on motion prediction are typically model-based such as the renowned social force-based method \cite{Helbing1995} which models pedestrian behaviors through the use of attractive and repulsive potentials. The model is later generalized and adapted to modeling traffic car behaviors \cite{helbing1998generalized}. While these methods are computationally efficient, the prediction accuracy is quite low. There have also been several notable attempts to utilize game theory to model interacting decision-making agents and predict their future trajectories \cite{Turnwald2016, Oyler2016}, in which the agents are assumed to play a non-cooperative game and their trajectory predictions can be obtained from computing the Nash equilibria of the game.
While interaction-aware trajectory predictions can be obtained, these methods are limited to specific road scenarios and cannot be directly applied to general multi-robot systems.
The class of approaches that have achieved state-of-the-art performance in trajectory prediction problems are the learning-based methods. Some of these include inverse reinforcement learning (IRL) \cite{Kuderer2015}, recurrent neural networks (RNN) \cite{Alahi2016,Brito2020}, variational autoencoders \cite{Lee2017}, generative adversarial networks (GAN) \cite{Gupta2018} \rebuttal{that provide predicted human trajectories in two-dimensional (2D) environments, Gaussian mixture regression (GMR) \cite{mainprice2013human} and Gaussian process regression (GPR) \cite{park2019planner} that can predict human actions in 3D workspaces.}
\rebuttal{
Our approach of predicting the trajectories of other robots is based on previous works on human motion prediction since both can be formulated as a sequence modelling problem. In particular, our prediction model is based on RNN, inspired by the works in \cite{Pfeiffer2018} for interaction-aware pedestrian motion prediction in which static obstacles are considered and represented using a grid map. We adapt the model to predict robots trajectories in multi-robot scenarios with moving obstacles described by their positions and velocities, and further apply the model to decentralized multi-robot motion planning by incorporating it within MPC.
}
\section{Preliminaries}\label{sec:preliminary}
Throughout this paper vectors are denoted in bold lowercase letters, $\mathbf{x}$, matrices in plain uppercase, $M$, and sets in calligraphic, $\mathcal{S}$. $\norm{\mathbf{x}}$ denotes the Euclidean norm of $\mathbf{x}$ and $\norm{\mathbf{x}}_{Q}^{\rebuttal{2}} = \mathbf{x}^TQ\mathbf{x}$ denotes the weighted squared norm.
\subsection{Robot and Obstacle Model}\label{subsec:robot_obs_model}
Following \cite{Zhu2019RAL}, we consider a team of $n$ robots moving in a shared workspace $\mathcal{W} \subseteq \mathbb{R}^3$, where each robot $i \in \mathcal{I} = \{ 1,2,\dots, n \} \subset \mathbb{N} $ is modeled as an enclosing sphere with radius $r$. The robots follow the same dynamical model that is described by a discrete-time equation as follows,
\begin{equation}\label{eq:nonDyn}
\begin{aligned}
\mathbf{x}_i^{k+1} = \mathbf{f}(\mathbf{x}_i^k, \mathbf{u}_i^k), \quad \mathbf{x}_i^0 = \mathbf{x}_i(0),
\end{aligned}
\end{equation}
where $\mathbf{x}_i^k \in \mathbb{X} \subset \mathbb{R}^{n_x}$ denotes the state of the robot, typically including its position $\mathbf{p}_i^k$ and velocity $\mathbf{v}_i^k$, and $\mathbf{u}_i^k \in \mathbb{U} \subset \mathbb{R}^{n_u}$ the control inputs at time $k$. Without loss of generality, $k=0$ indicates the current time. $\mathbb{X}$ and $\mathbb{U}$ are the admissible state space and control space, respectively. $\mathbf{x}_i(0)$ is the current state of robot $i$.
In addition, moving obstacles for example pedestrians in the environment are considered. For each obstacle $o \in \mathcal{I}_o = \{1,2,\dots,n_o \} \subset \mathbb{N} $ at position $\mathbf{p}_o \in \mathbb{R}^3$, we model it as an upright non-rotating enclosing \emph{ellipsoid} centered at $\mathbf{p}_o$ with semi-principal axes $(a, b, c)$.
\rebuttal{
In this paper, we assume that each robot can observe the states (positions and velocities) of all other robots and moving obstacles and keep their history information.}
\subsection{Multi-Robot Collision Avoidance}\label{subsec:mrca}
Multi-robot local motion planning is considered in this paper, in which the goal is to achieve real-time collision-free navigation for multiple robots. Each robot has a given goal location $\mathbf{g}_i$, which generally comes from some high-level path planner \cite{Honig2018} or is specified by the user. Any pair of robots $i$ and $j$ from the group are mutually collision-free if $\norm{\mathbf{p}_i^k - \mathbf{p}_j^k} \geq 2r, \forall i\neq j \in \mathcal{I}, k = 0, 1, \dots $. Regarding robot-obstacle collision avoidance, we approximate the obstacle with an enlarged ellipsoid and check if the robot's position is inside it. Hence, the robot $i$ is collision-free with the obstacle $o$ at time step $k$ if $\norm{\mathbf{p}_i^k-\mathbf{p}_o^k}_{\Omega} \geq 1$, where $\Omega = \text{diag}(1/(a+r)^2, 1/(b+r)^2, 1/(c+r)^2)$.
The objective is to compute a local motion $\mathbf{u}_i^k$ for each robot in the group, that respects its dynamic constraints, makes progress towards its goal location $\mathbf{g}_i$ and is collision-free with other robots in the group as well as moving obstacles within a planning time horizon.
\subsection{Model Predictive Control}\label{subsec:mpc}
The multi-robot collision avoidance problem can be solved using model predictive control by formulating a receding horizon constrained optimization problem. For each robot $i \in \mathcal{I}$, a discrete-time constrained optimization formulation with $N$ time steps and planning horizon $N\Delta t$, where $\Delta t$ is the sampling time, is derived as follows,
\begin{subequations}\label{eq:dmpc}
\begin{alignat}{2}
\min_{\substack{\mathbf{x}_i^{0:N}, \mathbf{u}_i^{0:N-1}, \\ s^{0:N}}} ~~
& \sum_{k=0}^{N-1} J_i^k(\mathbf{x}_i^k, \mathbf{u}_i^k, s^k) + J_i^N(\mathbf{x}_i^N, \mathbf{g}_i, s^N) \\
\text{s.t.} ~~ & \mathbf{x}_i^0 = \mathbf{x}_i(0), \\
& \mathbf{x}_i^{k} = \mathbf{f}(\mathbf{x}_i^{k-1}, \mathbf{u}_i^{k-1}), \\
& \norm{\mathbf{p}_i^k - \mathbf{p}_j^k} \geq 2r - s^{k}, \label{subeq:inter_robot_ca}\\
& \norm{\mathbf{p}_i^k - \mathbf{p}_o^k}_{\Omega} \geq 1 - s^{k},\label{subeq:robot_obs_ca}\\
& s^{k} \geq 0, \mathbf{u}_i^{k-1} \in \mathbb{U}, \mathbf{x}_i^k \in \mathbb{X},\\
&\forall j \neq i\in\mathcal{I}; \, \forall o\in\mathcal{I}_o; \, \forall k\in \{1,\dots,N\},
\end{alignat}
\end{subequations}
where $J_i^k(\mathbf{x}_i^k, \mathbf{u}_i^k, s^k)$ and $J_i^N(\mathbf{x}_i^N, \mathbf{g}_i, s^N)$ are the stage and terminal costs, and $s$ is the slack variable. At each time step, each robot in the team solves online the constrained optimization problem (\ref{eq:dmpc}) and then executes the first step control inputs, in a receding horizon fashion.
Note that for each robot to solve the optimization problem (\ref{eq:dmpc}), it has to know the future trajectories of other robots and moving obstacles, as shown in Eq. (\ref{subeq:inter_robot_ca}) and Eq. (\ref{subeq:robot_obs_ca}). For moving obstacles (pedestrians), we assume their motions follow a constant velocity model (CVM) in the short planning horizon and predict their future trajectories accordingly.
\rebuttal{
This assumption is reasonable since CVM can achieve state-of-the-art performance when used for pedestrian motion prediction \cite{Scholler2020}.}
For robots' future trajectories, denote by $\mathcal{T}_{i}^0 = \{ \mathbf{p}_i^{1:N} \}$ the robot $i$'s current planned trajectory. Further denote by $\hat{\mathcal{T}}_{i,j}^0 = \{\mathbf{p}_j^{1:N}\}$ the trajectory of robot $j\in\mathcal{I}, j\neq i$ that robot $i$ assumes and uses in solving the problem (\ref{eq:dmpc}), where the hat $\hat{\cdot}$ indicates that it is robot $i$'s estimation of the other robot's trajectory.
Typically, there are two ways for robot $i$ to obtain the future trajectory of the other robot $j$. The first way is via communication: all robots in the team communicate their planned trajectories to each other at each time step. It can be implemented using a centralized sequential planning framework as in \cite{Zhu2019RAL}, that is, $\hat{\mathcal{T}}_{i,j}^0 = \mathcal{T}_j^0$. Although this method guarantees collision avoidance by construction, it does not scale well with a large number of robots. Moreover, communication is not always available and reliable in practice.
The other way is without communication. Hence, robot $i$ has to predict another robot $j$'s future trajectory based on its observation of the environment:
\begin{equation}\label{eq:prediction_general}
\hat{\mathcal{T}}_{i,j}^0 = \tn{\tb{prediction}}(\mathcal{H}_i^0),
\end{equation}
where $\mathcal{H}_i^0$ is the information that robot $i$ can acquire until current time from its observation. Previous works \cite{Zhu2019RAL,Kamel2017} use a constant velocity model to perform the prediction only based on the other robot's current state, that is, $\mathcal{H}_i^0 = \mathbf{x}_j^0$. However, such a prediction can be inaccurate and may lead to unsafe trajectory planning \cite{Zhu2019RAL}. In this paper, we will develop an interaction- and obstacle-aware model for the trajectory prediction taking into account surrounding environment information of the robot to model the interaction and environment constraints.
\section{Approach}\label{sec:method}
In this section, we present our interaction and obstacle-aware trajectory prediction method and incorporate it with the MPC framework to achieve decentralized multi-robot collision avoidance in dynamic environments.
\subsection{Trajectory Prediction Problem Formulation}
As shown in Eq. (\ref{eq:prediction_general}), robot $i\in\mathcal{I}$ needs to predict the future trajectories of other robots $j \neq i \in \mathcal{I}$ to plan its safe motion. Hereafter, we refer to the robot $i$ as the ego robot and the robot $j$ as the \emph{query robot} that is indicated by the sub-script $\cdot_q$. In addition, we use the sub-script $\cdot_{-q}$ to indicate the collection of all the other robots except for the query robot.
We aim to address the problem of finding a trajectory prediction model for the query robot $q$ that gives a sequence of its future positions $\mathbf{p}_q^{1:T_H}$ in a multi-robot scenario. Here $T_H \geq N$ is the prediction horizon that should not be smaller than the local motion planning horizon. As has been shown in previous trajectory prediction works \cite{Pfeiffer2018,Brito2020}, we will instead work with sequences of velocities $\mathbf{v}_q^{1:T_H}$ for prediction to avoid overfitting when based on position sequences, and numerically integrate them afterwards starting from the query robot's current position $\mathbf{p}_q^0$.
Denote by $\mathbf{v}_q^{-T_O:0}$ the past sequence of velocities of the query robot within an observation time $T_O \geq 1$.
Denote by $\mathbf{p}_{-q,r}^{-T_O:0}$ and $\mathbf{v}_{-q,r}^{-T_O:0}$ the past relative positions and velocities of other robots with respect to the query robot.
Further denote by $\mathbf{p}_{\mathcal{I}_o,r}^{0}$ and $\mathbf{v}_{\mathcal{I}_o,r}^{0}$ the current relative positions and velocities of the moving obstacles $o\in\mathcal{I}_o$ with respect to the query robot. By observing history states of the query robot and its surrounding other robots as well as moving obstacles, we want to find an interaction- and obstacle-aware model $\mathbf{h}_{\boldsymbol\theta}$ with parameters $\boldsymbol\theta$:
\begin{equation}\label{eq:prediction_formulation}
\mathbf{v}_q^{1:T_H} = \mathbf{h}_{\boldsymbol\theta}(\mathbf{v}_q^{-T_O:0}, \mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0}, \mathbf{p}_{\mathcal{I}_o,r}^{0}, \mathbf{v}_{\mathcal{I}_o,r}^{0}),
\end{equation}
that outputs a prediction of the query robot's future states.
\subsection{Demonstration Data Generation}\label{subsec:dataset}
We use a simulation dataset to train our designed network model. The dataset is generated using demonstrations from a multi-robot collision avoidance simulator \cite{Zhu2019RAL} which employs a centralized sequential planner to solve the problem (\ref{eq:dmpc}). This involves each robot solving a MPC problem sequentially and communicates its planned trajectory to other robots to avoid.
Note that the planner differs from the prioritized planning approach since each robot has to avoid all other robots and hence it shows cooperation among robots.
Specifically, we create a three-dimensional environment in which a team of robots and moving obstacles are simulated.
In the simulation, each robot navigates to a randomly generated goal position, which is changed dynamically to a new location after being reached. The generated robots' goal positions are ensured to be collision-free with each other and the obstacles. Moving obstacles are simulated in the environment by randomly specifying an initial position and velocity \rebuttal{(with speed between 0.5 m/s and 1.2 m/s)} to each of them and then make them move at a constant velocity. Once any obstacle moves out of the environment, a new initial position and velocity will be set to it. Moreover, we add small Gaussian noise to the velocities of the moving obstacles in simulation. We perform the simulation for $N_{\tn{sim}}$ time steps and record the positions and velocities of all robots and obstacles at each time step.
After running the simulation, for each time step $t$ and robot $q$, we retrieve its future sequence of velocities and observation of the past states of the system from the recorded data. Hence, our dataset is as follows
\begin{align}
\mathcal{D} = \{ (\mathcal{H}_q^t, \mathbf{v}_q^{t+1:t+T_H}) | \forall q \in \mathcal{I}, \forall t \in \{1,\dots, N_{\tn{sim}-T_H}\} \},
\end{align}
where the observation information $\mathcal{H}_q^t$ is
\begin{equation}
\mathcal{H}_i^t = \{\mathbf{v}_q^{t-T_O:t}, \mathbf{p}_{-q,r}^{t-T_O:t}, \mathbf{v}_{-q,r}^{t-T_O:t}, \mathbf{p}_{\mathcal{I}_o,r}^{t}, \mathbf{v}_{\mathcal{I}_o,r}^{t}\}.
\end{equation}
\subsection{Interaction- and Obstacle-Aware Model}
We now present our recurrent neural network (RNN) model for interaction- and obstacle-aware trajectory prediction, as shown in Fig. \ref{fig:architecture}. The model first creates a joint representation of three input channels: the query robot's history state, information of other interacting robots and moving obstacles, via a query robot state encoder and an environment encoder module. Then a decoder module is adopted to output a predicted trajectory of the query robot. The recurrent layers in the model are of the LSTM type \cite{Hochreiter1997} that has been shown able to learn time dependencies over a long period of time.
Next, we describe the three main modules of the model in detail.
\begin{figure*}[t]
\centering
\setlength{\textfloatsep}{0.1\baselineskip plus 0.0\baselineskip minus 0.4\baselineskip}
\includegraphics[width=0.82\linewidth]{network_architecture}
\caption{Network architecture of the interaction- and obstacle-aware model. Three channels of information are taken as inputs: the query robot's past velocities $\mathbf{v}_q^{-T_O:0}$, past relative states of other robots $(\mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0})$ and current relative states of obstacles $(\mathbf{p}_{\mathcal{I}_o,r}^0, \mathbf{v}_{\mathcal{I}_o,r}^0)$. A joint representation of the inputs is created through a query robot encoder and an environment encoder. A decoder is adopted to output a sequence of velocities $\mathbf{v}_q^{1:T_H}$ predicted for the query robot's future trajectory.}
\label{fig:architecture}
\end{figure*}
\subsubsection{Query robot state encoder}
It consists of a recurrent layer that produces a flat encoding $\mathbf{z}_q^0$ from the history velocities of the query robot $\mathbf{v}_q^{-T_O:0}$.
This layer learns a dynamical model of the query robot, so that the network can leverage it to obtain better predictions.
\subsubsection{Environment encoder module}
It includes $n-1$ parallel recurrent layers with shared weights to encode the sequences of past relative positions and velocities of other robots with respect to the query robot $(\mathbf{p}_{-q,r}^{-T_O:0}, \mathbf{v}_{-q,r}^{-T_O:0})$ into a set $\mathcal{Z}_{-q,r}^{0}$, and $n_o$ parallel dense layers with shared weights that encode the current relative positions and velocities of moving obstacles with respect to the query robot $(\mathbf{p}_{\mathcal{I}_o,r}^{0}, \mathbf{v}_{\mathcal{I}_o,r}^{0})$ into a set $\mathcal{Z}_{\mathcal{I}_o,r}^{0}$. The encodings from both of these sets, which are made to have the same length, are then stacked together and followed by a global max pooling operation executed along the new data axis. Thus, this module can capture the interaction of the query robot with a variable number of other robots and obstacles in the environment and encode it into a single flat vector $\mathbf{z}_e^0$.
This framework also makes it possible to account for potentially different types of agents and obstacles by training their own set of encoders and stacking them with the rest of intermediate encodings.
\subsubsection{Decoder module}
It takes in the concatenation of $\mathbf{z}_q^0$ with $\mathbf{z}_e^0$ and passes it through a recurrent decoder followed by a dense decoder and an output layer that finally generates a sequence of predicted future velocities $\mathbf{v}_{q}^{1:T_H}$ for the query robot over the prediction horizon.
\subsection{Model Training}
Using the generated demonstration data in Section \ref{subsec:dataset}, the designed model is trained end-to-end using back-propagation through time (BTTP) \cite{werbos1990backpropagation} with a fixed truncation depth $t_{\tn{trunc}}$. We learn the trajectory prediction model by minimizing the following loss function,
\begin{equation} \label{eq:mse}
\begin{aligned}
L(\mathbf{v}_q^{1:T_H}, \boldsymbol\theta) = \frac{1}{T_H}\sum_{k=1}^{T_H}\norm{\mathbf{v}_q^{k} - \mathbf{v}_{q, \tn{true}}^{k}}^2 + \lambda\cdot l(\boldsymbol\theta),
\end{aligned}
\end{equation}
where $\mathbf{v}_{q,\tn{true}}^k$ is the ground truth velocity from the demonstration dataset, $l(\boldsymbol\theta)$ represents the regularization terms and $\lambda$ is the regularization factor. In our model, the $L2$ regularization method is adopted.
\subsection{Decentralized Multi-Robot Motion Planning}
Having the trained trajectory prediction model, we can incorporate it with the MPC framework and solve the problem (\ref{eq:dmpc}) in a decentralized manner. As shown in Fig. \ref{fig:overview}, in a multi-robot navigation scenario, each robot first performs inference with the trained neutral network to predict the future trajectories of its neighboring robots and then plans a collision-free trajectory accordingly. Hence, decentralized multi-robot motion planning in dynamic environments is achieved. To be able to perform the inference, each robot needs to measure its own state as well as its neighbors', and keep a history memory of the information for a time horizon $T_O$. In addition, the robot also needs to measure the current states of moving obstacles in the environment.
\section{Results}\label{sec:result}
We now present results of simulation comparing the proposed approach with other methods and real-world experiments with quadrotors.
A video accompanying this paper includes additional simulation and experimental results.
\subsection{Implementation Details}
To generate the dataset, we use an existing MATLAB multi-robot collision avoidance simulator\footnote{Code: \href{https://github.com/tud-amr/mrca-mav}{\color{blue}{https://github.com/tud-amr/mrca-mav}}} \cite{Zhu2019RAL} and simulate $N_{\tn{sim}} = 10^5$ time steps in a $10\times 10 \times 3$ m environment with \rebuttal{10 robots and 10 moving obstacles}. The robot we simulate is the Parrot Bebop 2 quadrotor with a radius set as 0.4 m.
Ellipsoids representing the moving obstacles have semi-axes $(0.4, 0.4, 0.9)$ m. The sampling time and MPC planning horizon length are $\Delta t = 0.05$ s and $N = 20$, respectively.
\rebuttal{
We employ the same dynamics model and cost functions in the MPC problem (2) of our previous work \cite{Zhu2019RAL}.}
The Forces Pro solver \cite{domahidi2014forces} is used to solve the MPC problem. We set $T_O = 20$ and $T_H = 20$ the horizon length for robot past states observation and trajectory prediction. We further generate another test dataset by running the simulator in \rebuttal{six} different scenarios for $2\times 10^4$ time steps for each one of them.
\rebuttal{
All computations are performed in a commodity computer with an Intel i7 CPU@2.60GHz and an NVIDIA GTX 1060 GPU.}
The designed learning network is implemented in Python using TensorFlow 2. All layers in the network have 64 neurons except for the recurrent decoder that has 128 neurons and the output layer that has 3 neurons. While the activation function of the output layer is linear, all other layers in the network use a hyperbolic activation function. The regularization factor used during model training is $\lambda = 0.01$.
\begin{figure*}[t]
\centering
\setlength{\textfloatsep}{0.2\baselineskip plus 0.0\baselineskip minus 0.4\baselineskip}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_4_0.pdf}
\caption{4 quads, no obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_10_0.pdf}
\caption{10 quads, no obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_10_10.pdf}
\caption{10 quads, 10 obstacles.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_20_0.pdf}
\caption{\rebuttal{20 quads, no obstacles.}}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.195\textwidth}
\includegraphics[width=1.0\textwidth]{predict_20_10.pdf}
\caption{\rebuttal{20 quads, 10 obstacles.}}
\end{subfigure}
\caption{Performance results of our proposed interaction-aware RNN model for trajectory prediction compared to the baselines. The solid lines represent the average errors along the prediction horizon and the filled patches around them are 30\% of the standard deviation. The sampling period is 50 ms and the prediction horizon has 20 timesteps. }
\label{fig:prediction_results}%
\end{figure*}
\subsection{Trajectory Prediction Evaluation}
We first evaluate our trajectory prediction model on a test dataset that has not been used for training nor validation. The dataset includes different test scenarios: an open environment with 4, 10, and \rebuttal{20} quadrotors, and with 10 moving obstacles. We compare our interaction-aware RNN-based model to three alternative methods: a) the constant velocity model (CVM) that is widely used in decentralized multi-robot motion planning; b) a simple RNN model that only considers the query robot's past states for trajectory prediction while ignoring its surrounding environment (this allows us to highlight the interaction awareness of our designed model); and c) an open-loop MPC planner assuming that the goal, robot model and constraints are known.
In Fig. \ref{fig:prediction_results} we present quantitative results of the prediction error with respect to ground truth in the test dataset. Recall that ground truths are the recorded robot traveled trajectories computed with the centralized sequential MPC (closed-loop). As expected, the prediction error of the open-loop MPC has the smallest prediction error among the methods since it was used for data generation and has perfect knowledge about the goal locations of all robots, which are not available for prediction in our proposed RNN-based model. Our proposed model can still achieve accurate trajectory predictions and significantly outperforms the CVM method across all scenarios. Moreover, compared to the simple RNN model, our interaction-aware approach achieves more accurate trajectory predictions, particularly in cluttered scenarios where interactions among robots are more frequent, as shown in Fig. \ref{fig:prediction_results}(b)-(e).
\rebuttal{
Furthermore, to evaluate the generalization capability of the learned network, we perform simulations in the scenarios (d) and (e) with 20 quadrotors which are beyond our training dataset. The results show that the proposed model still performs well on trajectory prediction in the two scenarios.}
\subsection{Decentralized Motion Planning}
We then evaluate performance of the proposed decentralized planner that incorporates the learned trajectory prediction model.
\subsubsection{\rebuttal{Comparisons to other methods}}
We compare our method to the centralized sequential planning method \cite{Zhu2019RAL} with full communication among robots and the decentralized planning method \cite{Kamel2017} that uses the constant velocity model (CVM) for trajectory prediction to analyze whether more accurate trajectory forecasts of our RNN-based model lead to better planning performance. \rebuttal{Besides, another decentralized method, the buffered Voronoi cell (BVC) \cite{Zhou2017}, which guarantees collision avoidance is also implemented for comparison.}
Six quadrotors flying in four types of scenarios that represents different levels of difficulty \cite{Serra2020} are considered. Moreover, in order to avoid potential bias results, each scenario includes 50 instances where the robots have different starting and goal locations. The four scenarios are: \emph{1) symmetric swap}, in which the robots initially located at the vertices of a virtual horizontal regular hexagon are required to exchange their positions; \emph{2) asymmetric swap}, which differs from the previous scenario in that the hexagons are irregular, thus leading to more challenging collision-avoidance problems; \emph{3) pair-wise swap}, in which the robots are placed at random starting positions and assigned to three pairs within which the two robots need to swap their positions; and \emph{4) random moving}, in which each robot moves from a random starting position to a random goal in the environment.
Qualitatively, Fig. \ref{fig:sim_planning} shows the sample trajectory trails of the six quadrotors for one instance from the asymmetric swap scenario. \rebuttal{It can be seen that our RNN-based decentralized planner achieves results that are closer to the centralized sequential planner than the CVM-based planner.}
To quantitatively evaluate the performance of different motion planners, we consider a wide range of metrics: the number of instances that lead to collisions within the entire 50 runs, the average trajectory length and trajectory duration of the team of robots, and the overall robot average speed during the whole simulation. The last three metrics are only computed for those successful runs. Table \ref{tab:planning_comparison} summaries the simulation results. It can be seen that our RNN-based planner significantly outperforms the planner using the CVM for trajectory prediction in terms of safety, in particular in the challenging asymmetric swap scenario. In addition, \rebuttal{our planner also achieves consistently smaller trajectory lengths and durations compared to the CVM-based planner} in all scenarios.
\rebuttal{Compared to the BVC method, our proposed approach achieves significantly shorter trajectory durations, particularly in the (a)symmetric swapping scenarios, which shows superiority of the MPC framework over the reactive BVC method.}
Finally, compared to the centralized sequential planner with full communication, our planner can achieve a comparable level of performance in terms of safety and trajectory efficiency while being decentralized and communication-free. However, three instances out of 50 in the challenging asymmetric swap scenario is still observed with collisions using the RNN-based method, indicating that in few rare cases, highly-accurate trajectory predictions of other robots, for example obtained via communication, are necessary to ensure safety.
\rebuttal{
In the simulation, on average the computation time of the proposed decentralized MPC planner with the learned predictor is 36.3 ms, which is smaller than that of the centralized sequential planner which plans trajectories for all six robots (43.9 ms). Besides, our decentralized approach is communication-free.}
\begin{figure}[t]
\centering
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_cent_xy.pdf}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_CVM_xy.pdf}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.151\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_RNN_xy.pdf}
\end{subfigure}
\\
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_cent_xz.pdf}
\caption{Cen. Comm.}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_CVM_xz.pdf}
\caption{Decen. CVM}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=1.0\textwidth]{testAsy_RNN_xz.pdf}
\caption{Decen. RNN}
\end{subfigure}
\caption{Simulation results of six quadrotors exchanging positions in the asymmetric swap scenario. Solid lines represent the trajectories. The upper and lower plots show the top view (X-Y) and side view(X-Z), respectively.}
\label{fig:sim_planning}
\end{figure}
\begin{table*}[t]
\centering
\caption{Performance comparison of different multi-robot motion planners (centralized with communication \cite{Zhu2019RAL}, \rebuttal{decentralized buffered Voronoi cell (BVC) method \cite{Zhou2017}}, decentralized with constant velocity model (CVM) \cite{Kamel2017} and decentralized with our RNN-based model) across the four different types of scenarios (symmetric swap, asymmetric swap, pair-wise swap and random moving). Each scenario includes 50 running instances. }
\begin{tabular}{ll||c|ccc|ccc|c}
\hline
\multicolumn{1}{c}{\multirow{2}{*}{\textbf{Scenario}}} & \multicolumn{1}{c||}{\multirow{2}{*}{\textbf{Motion Planner}}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Num. of coll. \\ instances\end{tabular}}} & \multicolumn{3}{c|}{\textbf{Trajectory length (m)}} & \multicolumn{3}{c|}{\textbf{Trajectory duration (s)}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Average speed\\ (m/s)\end{tabular}}} \\
\multicolumn{1}{c}{} & \multicolumn{1}{c||}{} & & Min. & Average & Max. & Min. & Average & Max. & \\ \hline \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Symmetric\\ swap\end{tabular}} & Cen. Comm. & {0} & 5.46 & 7.22$\pm$0.85 & 8.96 & 5.15 & 5.62$\pm$0.24 & 6.05 & 1.27$\pm$0.05 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{5.75} & \rebuttal{7.63$\pm$0.87} & \rebuttal{9.63} & \rebuttal{10.00} & \rebuttal{12.21$\pm$1.25} & \rebuttal{14.80} & \rebuttal{0.63$\pm$0.03} \\
& Decen. (CVM) & 4 & 5.43 & 7.45$\pm$0.88 & 9.55 & 4.95 & 6.02$\pm$0.55 & 7.30 & 1.23$\pm$0.05 \\
& {Decen. (RNN)} & \tb{{0}} & {5.41} & \tb{7.35$\pm$0.91} & {10.60} & {4.75} & \tb{5.59$\pm$0.40} & {9.70} & \tb{1.30$\pm$0.05} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Asymmetric\\ swap\end{tabular}} & Cen. Comm. & {0} & 5.08 & 6.77$\pm$0.80 & 9.02 & 4.65 & 5.17$\pm$0.30 & 5.85 & 1.30$\pm$0.05 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{5.31} & \rebuttal{7.25$\pm$0.87} & \rebuttal{9.29} & \rebuttal{9.70} & \rebuttal{11.34$\pm$1.00} & \rebuttal{13.90} & \rebuttal{0.65$\pm$0.03} \\
& Decen. (CVM) & 15 & 5.32 & 7.76$\pm$1.89 & 18.06 & 5.05 & 5.96$\pm$0.51 & 7.30 & 1.28$\pm$0.04 \\
& {Decen. (RNN)} & {3} & {5.16} & \tb{7.14$\pm$1.10} & {12.60} & {4.80} & \tb{5.48$\pm$0.42} & {6.85} & \tb{1.29$\pm$0.05} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Pair-wise\\ swap\end{tabular}} & Cen. Comm. & {0} & 1.64 & 5.10$\pm$1.87 & 9.92 & 3.50 & 4.76$\pm$0.44 & 5.85 & 1.06$\pm$0.12 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{1.83} & \rebuttal{5.25$\pm$1.93} & \rebuttal{10.13} & \rebuttal{5.20} & \rebuttal{7.58$\pm$1.66} & \rebuttal{13.20} & \rebuttal{0.67$\pm$0.07} \\
& Decen. (CVM) & 3 & 1.74 & 5.54$\pm$2.53 & 17.42 & 3.60 & 5.00$\pm$0.65 & 5.75 & \tb{1.06$\pm$0.10} \\
& {Decen. (RNN)} & \tb{0} & {1.70} & \tb{4.94$\pm$2.02} & {9.94} & {3.40} & \tb{4.83$\pm$0.45} & {5.95} & {1.01$\pm$0.14} \\ \hline
\multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Random\\ moving\end{tabular}} & Cen. Comm. & {0} & 0.39 & 4.66$\pm$1.94 & 8.63 & 3.50 & 7.72$\pm$0.40 & 5.50 & 0.98$\pm$0.13 \\
& \rebuttal{Decen. (BVC)} & \tb{\rebuttal{0}} & \rebuttal{0.39} & \rebuttal{4.81$\pm$2.00} & \rebuttal{8.84} & \rebuttal{4.70} & \rebuttal{7.13$\pm$1.30} & \rebuttal{10.10} & \rebuttal{0.69$\pm$0.09} \\
& Decen. (CVM) & \tb{0} & 0.39 & 4.82$\pm$2.06 & 9.53 & 3.60 & 4.89$\pm$0.57 & 6.35 & \tb{0.98$\pm$0.12} \\
& {Decen. (RNN)} & \tb{0} & {0.39} & \tb{4.36$\pm$2.11} & {9.19} & {3.85} & \tb{4.76$\pm$0.43} & {5.95} & {0.91$\pm$0.11} \\
\hline
\end{tabular}
\label{tab:planning_comparison}
\end{table*}
\rebuttal{
\subsubsection{Effect of non-MPC robots on performance}
Our proposed decentralized approach assumes that all robots interact and adopt the same motion planning strategy, namely MPC-based trajectory optimization with the learned motion prediction model. We now evaluate the performance of our approach in a mixture scenario where some robots employ the BVC method \cite{Zhou2017} for collision avoidance.
We simulate 50 instances with six quadrotors in the symmetric swap scenario of Section V-C-1. Table \ref{table:nonMPC_robot} presents the simulation results. When there is only one BVC robot, no collisions are observed. However, when more BVC robots are in the team, particularly when half of the robots (3) are BVC-based, collisions will happen due to incorrect motion predictions of them by other MPC robots. This indicates that the assumption that the robots interact with the same planning strategy is necessary to ensure safety.
\begin{table}[t]
\centering
\caption{\rebuttal{Simulation results of six quadrotors in the symmetric swap scenario where a varying number of BVC-based robots are in the team. 50 running instances are simulated.}}
\begin{tabular}{lcccccc}
\hline
Num. of BVC robots & 0 & 1 & 2 & 3 & 4 \\ \hline
Num. of coll. instan. & 0 & 0 & 2 & 4 & 1 \\
Ave. traj. time (s) & 5.59 & 6.82 & 8.517 & 9.63 & 10.27 \\ \hline
\end{tabular}
\label{table:nonMPC_robot}
\end{table}
}
\subsection{Experimental Validation}
\subsubsection{Setup}
We validate our proposed approach with a team of Parrot Bebop 2 quadrotors flying in a shared space with walking human obstacles. The pose of each quadrotor and obstacle (human) is obtained using an external motion capture system (OptiTrack) and their velocities obtained via a standard Kalman filter running at a high rate. Control commands are sent to the quadrotor via ROS.
\rebuttal{
During the experiment, the humans walked at a speed with mean 0.8 m/s and the maximum 1.2 m/s. They could change their speeds and make small turns in the workspace.
}
\begin{figure}[t]
\centering
~
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.19\textwidth}
\includegraphics[width=1.0\textwidth]{exp_gopro.png}
\caption{}
\label{subfig:exp_gopro}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=1.0\textwidth]{exp_matlab.pdf}
\caption{}
\label{subfig:exp_matlab}
\end{subfigure}
\begin{subfigure}{0.21\textwidth}
\includegraphics[width=1.0\textwidth]{3q_dis_time.pdf}
\caption{}
\label{subfig:quad_dis}
\end{subfigure}
\captionsetup[subfigure]{position=b}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=1.0\textwidth]{3q_obs_dis_his.pdf}
\caption{}
\label{subfig:obs_dis}
\end{subfigure}
\caption{Experimental results with three quadrotors flying in a shared space with two walking humans. (a) A snapshot of the experiment. (b) Schematic of quadrotors, humans, and planned trajectories. (c) Distance between the quadrotors over time. The shaded grey area indicates the two walking humans join the space. (d) Histogram of the quadrotor-obstacle distance during the experiments.}
\label{fig:exp_results}%
\end{figure}
\subsubsection{Results}
Experiments in two representative scenarios are conducted: with and without obstacles. In the first scenario, three quadrotors, initially distributed in a virtual horizontal circle, are required to swap their positions multiple times. Then in the second scenario, two moving obstacles (walking humans) join the space while the three quadrotors keeps changing their positions while avoiding the humans at the same time. Fig. \ref{subfig:exp_gopro} presents a snapshot from the experiment.
Fig. \ref{subfig:quad_dis} shows distance between each pair of the three quadrotors over time during the experiment. It can be seen that they maintained a safe distance of 0.8 m over the entire run even after the two walking humans joined the space which makes it more confined. In Fig. \ref{subfig:obs_dis} we cumulate the distance between each quadrotor and human obstacle that is computed as the closest distance from the quadrotor center to the obstacle ellipsoid's surface. The results show that a minimum safe separation of 0.4 m to the obstacles is achieved. In sum, the demonstration shows that our proposed approach works well for multi-robot motion planning in dynamic environments which is decentralized and communication-free.
|
\section{Introduction}
Motivated by the abundance of unlabeled data and the expensiveness of obtaining labels, the paradigm of active learning has been proposed and extensively studied in the literature~\cite[see e.g.][for comprehensive surveys]{settles2009active,hanneke2014theory}.
In active learning, a learner starts with a set of unlabeled examples,
and can adaptively select subsets of them to query for their labels. Thanks to its adaptivity, an active learner can focus on obtaining informative labels, and can thus substantially reduce labeling effort compared to conventional supervised learning.
Halfspaces, also known as linear separators, are arguably one of the most fundamental concept classes studied in machine learning and data analysis.
Significant research efforts on halfspace learning from computational and statistical perspectives have resulted in rich theory~\cite[e.g.][]{vapnik1998statistical,blum1996polynomial,zhang2004statistical,bartlett2006convexity, kalai2008agnostically} and many practical algorithms~\citep[e.g.][]{cortes1995support, cristianini2000introduction}.
Label noise is ubiquitous in machine learning applications due to various factors, such as human error, sensor failure, etc~\citep{balcan2020noise}, and it is therefore important to design learning algorithms that are robust to label noise.
If computational efficiency is not of concern, classical methods such as empirical risk minimization are known to achieve statistical consistency~\citep{vapnik1998statistical}.
However, in many practical applications, it is often necessary for learners to process its training examples in a computationally efficient manner.
Therefore, it is of importance to develop computationally efficient, noise-tolerant learning algorithms with statistical consistency guarantees.
However, it is now well-understood that without additional assumptions on the label noise, agnostically learning halfspaces is computationally hard~\citep{feldman2006new,guruswami2009hardness,daniely2016complexity}, even under well-behaved unlabeled data distributions such as standard Gaussian~\citep{klivans2014embedding,diakonikolas2020near}. This motivates the study of learning halfspaces under more benign label noise conditions. Massart noise~\citep{massart2006risk} and Tsybakov noise~\citep{tsybakov2004optimal} are two noise models widely studied in the literature; specialized to the halfspace learning setting, they are formally defined as:
\begin{definition}[Massart noise condition]
Given $\eta \in [0,\frac12)$,
a distribution $D$ over $\RR^d \times \cbr{-1,+1}$ is said to satisfy the {\em $\eta$-Massart noise condition} with respect to halfspace $w^\star \in \RR^d$, if for all examples $x$, $\eta(x) \leq \eta$, where $\eta(x) = \PP_D(y \neq \sign(\inner{w^\star}{x}) \mid x)$.
\label{def:mnc}
\end{definition}
\begin{definition}[Tsybakov noise condition]
Given $A > 0$ and $\alpha \in (0,1]$, a distribution $D$ over $\RR^d \times \cbr{-1,+1}$ is said to satisfy the {\em $(A, \alpha)$-Tsybakov noise condition} with respect to halfspace $w^\star \in \RR^d$, if for all $t \in [0,\frac12)$, $\PP_D\rbr{ \frac12 - \eta(x) \leq t} \leq A t^{\frac{\alpha}{1-\alpha}}$,
where $\eta(x) = \PP_D(y \neq \sign(\inner{w^\star}{x}) \mid x)$.
\label{def:tnc}
\end{definition}
Although nearly-matching upper and lower bounds on sample and label complexities have been established in both supervised (passive) and active learning settings under these two noise conditions~\citep[e.g.][]{massart2006risk, tsybakov2004optimal, castro2008minimax, hanneke2011rates, balcan2013active, wang2016noise}, most of these results are only {\em statistical}: the algorithms that achieve the sample or label complexity upper bounds are computationally inefficient. Only recently have computationally efficient algorithms been proposed in the literature~\citep[e.g.][]{awasthi2015efficient, diakonikolas2019distribution, diakonikolas2020polynomial} under these noise conditions; see Section~\ref{sec:relwork} for detailed discussions. Still, these works leave out two important open questions:
\begin{enumerate}
\item {\sloppy Are there efficient active halfspace learning algorithms that tolerate Massart noise with near-optimal label complexity, under a broad range of unlabeled data distributions? Specifically, the algorithm of~\cite{yan2017revisiting}
achieves an information-theoretically near-optimal label complexity of $O\rbr{\frac{d}{(1-2\eta)^2} \polylog(\frac1\epsilon)}$, but relies on the strong assumption that the unlabeled data distribution is uniform on the unit sphere; under broader distributional assumptions such as log-concave distributions,
the state-of-the-art algorithm of~\cite{zhang2020efficient} only achieves a suboptimal label complexity of $O\rbr{\frac{d}{(1-2\eta)^4} \polylog(\frac1\epsilon)}$. Can we design efficient algorithms with $\tilde{O}\rbr{\frac{d}{(1-2\eta)^2} \polylog(\frac1\epsilon)}$ label complexity under broader unlabeled data distributions, for example, the family of isotropic log-concave distributions~\citep{lovasz2007geometry}? }
\item Are there efficient active halfspace learning algorithms that tolerate Tsybakov noise with label complexities better than passive learning? The state-of-the-art work of~\cite{diakonikolas2020polynomial} propose efficient algorithms for passive learning halfspaces with Tsybakov noise, with sample complexities $O\rbr{(\frac{d}{\epsilon})^{O(\frac{1}{\alpha})}}$ and $O\rbr{ \poly(d) (\frac 1 \epsilon)^{O(\frac1{\alpha^2})} }$; moreover, computationally inefficient algorithms such as empirical risk minimization achieve a sharper sample complexity of $O\rbr{ d(\frac{1}{\epsilon})^{2-\alpha}}$~\cite[e.g.][Chapter 3]{hanneke2014theory}.
Can we design efficient active learning algorithms with label complexities of strictly lower order than these? A positive answer to this question can serve as a stepping stone towards developing efficient active algorithms with label complexity matching those of computationally inefficient active learning algorithms~\citep[e.g.][]{balcan2013active}, which is $O\rbr{ d(\frac 1 \epsilon)^{2-2\alpha}}$.
\end{enumerate}
\paragraph{Our results.} Our work answers the above two questions in the affirmative. Specifically, under a set of structural assumptions on the unlabeled data distribution~\citep{diakonikolas2020polynomial} (see also Definition~\ref{def:well-behaved} in Section~\ref{sec:setting}), we give an efficient PAC active halfspace learning algorithm, such that with appropriate settings of its parameters:
\begin{enumerate}
\item under the $\eta$-Massart noise condition, it has an information-theoretically near-optimal label complexity of $O\rbr{\frac{d}{(1-2\eta)^2} \polylog(\frac1\epsilon)}$. This substantially weakens the distributional requirements to achieve such near-optimal label complexity results; before our work, such result is only known when the unlabeled data distribution is uniform over the unit sphere~\citep{yan2017revisiting}, or when the noise parameter $\eta$ is $3 \times 10^{-6}$, a tiny constant~\citep{awasthi2015efficient}. Furthermore, when the unlabeled data distribution is isotropic log-concave, our result improves over the recent work of of~\cite{zhang2020efficient}, where an efficient algorithm with suboptimal label complexity $O\rbr{\frac{d}{(1-2\eta)^4} \polylog(\frac1\epsilon)}$ is proposed.
\item under the $(A, \alpha)$-Tsybakov noise condition with $\alpha \in (\frac12, 1]$, it has a label complexity of $\tilde{O}(d (\frac{1}{\epsilon})^{\frac{2-2\alpha}{2\alpha-1}})$.
Specifically, when $\alpha > \frac{7-\sqrt{17}}{4} \approx 0.719$, $\frac{2-2\alpha}{2\alpha-1} < 2-\alpha$, in which case our algorithm has a better label complexity than passive learning.
Furthermore, in the special case of $(B,\alpha)$-geometric Tsybakov noise condition (see Definition~\ref{def:gtnc} in Section~\ref{sec:setting}), our algorithm achieves a lower label complexity of $O\rbr{ d (\frac{1}{\epsilon})^{\frac{2}{\alpha} - 2} }$ for all $\alpha \in (0,1]$; specifically, when $\alpha > 2 - \sqrt{2} \approx 0.585$, $\frac{2-2\alpha}{\alpha} < 2 - \alpha$, in which case our algorithm has a better label complexity than passive learning.
\end{enumerate}
\paragraph{Techniques.} Our algorithm and analysis bear similarities to the recent work of~\cite{zhang2020efficient}, who observe that online mirror descent-style updates, when composed with a margin-based active sample selection rule~\citep{balcan2007margin}, implicitly minimizes a nonstandard proximity measure against the Bayes-optimal halfspace $w^\star$.
Compared to~\citet{zhang2020efficient}, our results are novel in two aspects.
First, the algorithm of~\cite{zhang2020efficient} is specialized to Massart noise, as its update rule crucially relies on the knowledge of the Massart noise level $\eta$. In contrast, we propose a new and simpler update rule that can work under both Massart and Tsybakov noise conditions.
Second, under the $\eta$-Massart noise setting, ~\cite{zhang2020efficient} use an averaging-based initialization procedure, which leads to an algorithm and analysis requiring a suboptimal label complexity of $\tilde{O}\rbr{ \frac{d}{(1-2\eta)^4} \polylog(\frac1\epsilon)}$. In this paper, we design a new initialization procedure with improved label efficiency, leading to an algorithm with an information-theoretically near-optimal label complexity of $\tilde{O}\rbr{ \frac{d}{(1-2\eta)^2} \polylog(\frac1\epsilon) }$ under Massart noise.
\section{Related work}
\label{sec:relwork}
\paragraph{Learning halfspaces under Massart and Tsybakov noise: statistical rates.} For passive learning, it is well-known that empirical risk minimization achieves minimax-optimal sample complexities of $\tilde{O}\rbr{\frac{d}{(1-2\eta) \epsilon}}$ and $\tilde{O}\rbr{d (\frac{1}{\epsilon})^{2-\alpha}}$ under $\eta$-Massart noise and $(A,\alpha)$-Tsybakov noise conditions respectively~\citep[e.g.][Chapter 3]{hanneke2014theory}.
For active learning, many works have provided distribution-specific label complexity upper bounds, including the general analyses of~\cite{hanneke2011rates,beygelzimer2010agnostic,zhang2014beyond} and more specialized analyses of~\cite{balcan2013active,wang2016noise}.
In the setting of $\eta$-Massart noise, under the assumption that the unlabeled data distribution is isotropic log-concave, the state-of-the-art algorithms of~\cite{balcan2013active, zhang2014beyond} have a label complexity of $\tilde{O}\rbr{ \frac{d}{(1-2\eta)^2} \polylog( \frac1\epsilon) }$; in the setting of $(A,\alpha)$-Tsybakov noise under log-concave unlabeled distributions, the state-of-the-art algorithms of~\cite{balcan2013active,zhang2014beyond,wang2016noise} achieve a label complexity of $\tilde{O}\rbr{ d (\frac{1}{\epsilon})^{2-2\alpha} }$.
Although these algorithms provide sharp label complexity guarantees,
they all suffer from computational inefficiency: they need to perform empirical 0-1 loss minimization, which is known to be NP-hard in general~\citep{arora1997hardness}.
\paragraph{Efficient passive learning halfspaces with Massart noise.}
The study of computationally efficient halfspace learning under Massart noise is initiated by the work of~\cite{awasthi2015efficient}, who provide a PAC learning algorithm that works under the assumptions that the unlabeled distribution is isotropic log-concave, and the Massart noise parameter $\eta$ is smaller than a tiny constant ($3 \times 10^{-6}$). Prior to this work, positive results mainly focus on the much weaker random classification noise~\citep[e.g.][]{blum1996polynomial,balcan2013statistical}. Under similar distributional assumptions,~\citet{awasthi2016learning} propose an algorithm with sample complexity $O\rbr{ d^{O\rbr{\frac{1}{(1-2\eta)^4}}} \frac{1}{\epsilon} }$ for any $\eta \in [0,\frac12)$. Recent works of~\cite{zhang2020efficient} and~\cite{diakonikolas2020learning} provide passive learning algorithms with fully-polynomial sample complexities in this setting, achieving sample complexities of $O\rbr{ \frac{d}{\epsilon(1-2\eta)^5} }$ and $O\rbr{ \frac{d^9}{\epsilon^4 (1-2\eta)^{10}} }$ respectively.
In the distribution-free PAC learning setting, that is, when no assumptions are imposed on the unlabeled data distribution, efficient halfspace learning is much more challenging.
Recent breakthrough of~\cite{diakonikolas2019distribution} provides an efficient improper learner that can guarantee to output a halfspace with error $\eta+\epsilon$ with sample complexity $\poly(d,\frac1\epsilon)$. ~\cite{chen2020classification} improves over this result by proposing a proper learner, along with a generic ``distillation'' procedure that converts any improper learner to a proper one. In the same paper, they also show that for any statistical query algorithm, obtaining a classifier that achieves an error rate of $\mathrm{opt} + o(1)$ requires a superpolynomial number of statistical queries, where $\mathrm{opt}$ denotes the error rate of the Bayes-optimal halfspace $w^\star$. This lower bound is recently strengthened by~\cite{diakonikolas2020hardness}, showing that even achieving a weaker $\poly(\mathrm{opt})$ error rate requires a superpolynomial number of statistical queries.
\paragraph{Efficient passive learning halfspaces with Tsybakov noise.} Recently,~\cite{diakonikolas2020learningb} obtains an algorithm with a quasi-polynomial time and sample complexity of $O\rbr{d^{O(\frac{1}{\alpha^2} \ln\frac{1}{\epsilon})}}$ for PAC learning halfspaces under $(A,\alpha)$-Tsybakov noise condition, under distributions with certain structural properties.
This result is further improved by~\cite{diakonikolas2020polynomial}, who obtain two algorithms with time and sample complexities of $O\rbr{\poly(d) \cdot (\frac1\epsilon)^{O(\frac{1}{\alpha^2})}}$ and $O\rbr{(\frac d \epsilon)^{O(\frac 1 \alpha)}}$ respectively.
\paragraph{Efficient active learning halfspaces with Massart and Tsybakov noise.} By combining the agnostic halfspace learning algorithm of~\cite{kalai2008agnostically}, and margin-based sampling~\citep{balcan2007margin, balcan2013active}, \cite{awasthi2016learning} obtains an active halfspace learning algorithm that tolerates $\eta$-Massart noise with a sample complexity of $O\rbr{ d^{O(\frac{1}{(1-2\eta)^4})} \ln\frac1\epsilon }$ under the isotropic log-concavity assumption on the unlabeled data distribution. This result is recently substantially improved by~\cite{zhang2020efficient}, who obtain a label complexity of $O\rbr{ \frac{d}{(1-2\eta)^4} \polylog(\frac1\epsilon) }$ in the same setting; however this label complexity bound still does not match the information-theoretic lower bound of $\Omega\rbr{ \frac{d}{(1-2\eta)^2} \ln\frac1\epsilon }$. The only label-optimal result on PAC active halfspace learning under Massart noise we are aware of is~\cite{yan2017revisiting}, however it relies on the strong assumption that the unlabeled data distribution is uniform over the unit sphere. Under Tsybakov noise condition, to the best of our knowledge,
all prior active learning works require well-specified model assumptions on the conditional distribution of label given feature \citep{cesa2009robust,dekel2012selective,agarwal2013selective,krishnamurthy2017active}, e.g. assuming $\EE[ y \mid x ] = \sigma(w^\star \cdot x)$ for some known function $\sigma: \RR \to \RR$.
This is relatively strong, as it requires all examples $x$ that have the same projection on $w^\star$ to have the same value of $\eta(x)$. In contrast, our work does not require such assumptions.
\section{Preliminaries}
\label{sec:setting}
We consider the standard PAC active learning for binary classfication setup~\citep{valiant1985learning,balcan2009agnostic}. Specifically, the instance space $\Xcal$ is $\RR^d$, the label space $\Ycal$ is $\cbr{-1,+1}$, and there is a data distribution $D$ supported on $\Xcal \times \Ycal$. The hypothesis class of interest is the set of linear classifiers, also known as halfspaces, defined as $\Hcal = \cbr{h_w: w \in \RR^d}$, where for every $w \in \RR^d$, $h_w$ denotes the corresponding linear classifier that maps $x$ to $\sign(\inner{w}{x})$. We use {\em error rate} to measure the performance of a classifier $h: \Xcal \to \Ycal$, defined as $\err(h, D) = \PP_{(x,y) \sim D}(h(x) \neq y)$.
Given classifier $h$, and a set of labeled examples $S$, denote $\err(h, S) = \frac1{|S|} \sum_{(x,y) \in S} \ind\rbr{h(x) \neq y}$ as the empirical error rate of $h$ on $S$.
Throughout this paper, we assume that the Bayes-optimal classifier is a halfspace $h_{w^\star}$, where $w^\star \in \RR^d$ is a unit vector. It can be verified that $h_{w^\star}$ is indeed Bayes-optimal under Massart or Tsybakov noise conditions (recall Definitions \ref{def:mnc} and \ref{def:tnc}).
An active learning algorithm has access to two labeling oracles: first, an unlabeled example oracle $\mathrm{EX}$, which, upon query, returns an unlabeled example $x$ drawn from $D_X$, the marginal distribution of $D$ over $\Xcal$; second, a labeling oracle $\Ocal$, which, upon query with input example $x$, returns a label $y$ drawn from $D_{Y \mid X=x}$, the conditional distribution of $Y$ given $X=x$. A learner is said to achieve $(\epsilon,\delta)$-PAC active learning guarantee, if, by interactively querying the unlabeled example oracle $\mathrm{EX}$ and the labeling oracle $\Ocal$, it outputs a classifier $\tilde{h}$, such that with probability $1-\delta$, $\err(\tilde{h}, D) - \err(h_{w^\star}, D) \leq \epsilon$. Its label complexity is the total number of queries to $\Ocal$ throughout the learning process.
For a natural number $N$, denote by $[N] := \cbr{1,2,\ldots,N}$.
For a vector $w$ in $\RR^d$, denote by its $\ell_2$-normalization $\hat{w} := \frac{w}{\|w\|_2}$ if $w \neq \overrightarrow{0}$, and $\hat{w} = (1,0,\ldots,0)$ if $w = \overrightarrow{0}$. {\em Throughout this paper, we reserve the ``hat'' symbol and notations such as $\hat{w}$, $\hat{v}$ for $\ell_2$-normalization and $\ell_2$-normalized vectors.} Unless explicitly stated, we use $\| \cdot \|$ to denote the vector $\ell_2$ norm.
For two vectors $u$ and $v$, denote by $\theta(u,v) = \arccos \rbr{\frac{\inner{u}{v}}{\|u\|\|v\|}} \in [0,\pi]$ the angle between them; also, denote by $\tilde{\theta}(u,v) = \min(\theta(u,v), \pi - \theta(u,v)) \in [0, \frac \pi 2]$.
In our algorithm and analysis below, we will be frequently using the following definition of the distribution $D$ conditioned on a band: given a unit vector $\hat{w}$ and a threshold $b > 0$, denote by $B_{\hat{w},b} = \cbr{x \in \RR^d: \abr{\inner{\hat{w}}{x}} \leq b}$; in addition, denote by $D_{\hat{w}, b}$ the conditional distribution of $D$ on the set $\cbr{(x,y) \in \RR^d \times \cbr{-1,+1}: x \in B_{\hat{w},b}}$; similarly, denote by $D_{X \mid \hat{w}, b}$ the conditional distribution of $D_X$ on the set $B_{\hat{w},b}$.
Computational hardness results~\citep{chen2020classification,diakonikolas2020hardness} strongly suggest that efficient learning halfspaces with noise may be computationally intractable if no assumptions on the unlabeled data distribution are made, even under benign noise conditions such as Massart noise. Therefore, throughout this paper, following~\cite{diakonikolas2020polynomial}, we assume the unlabeled distribution to lie in a family of structured, or well-behaved distributions, defined as follows:
\begin{definition}[Well-behaved distributions~\citep{diakonikolas2020polynomial}]
\label{def:well-behaved}
Fix $L, R, U, \beta > 0$. A distribution $D_X$ over $\RR^d$ is said to be $(2,L,R, U, \beta)$ well-behaved, if for any $2$-dimensional linear subspace $V$ of $\RR^d$, we have:
given an $x$ randomly drawn from $D_X$,
$x_V$, the projected coordinates of $x$ onto $V$\footnote{Formally, pick $(v_1,v_2)$ as an orthonormal basis of $V$; define $x_V := (\inner{v_1}{x}, \inner{v_2}{x}) \in \RR^2$.}, has a probability density function $p_V$ on $\RR^2$, such that:
\begin{enumerate}
\item $p_V(z) \geq L$, for all $z$ such that $\| z \|_2 \leq R$;
\item $p_V(z) \leq U$, for all $z \in \RR^2$;
\end{enumerate}
in addition, for any unit vector $w$ in $\RR^d$ and any $t > 0$, $\PP_{D_X}(\abr{\inner{w}{x}} \geq t) \leq \exp(1-\frac{t}{\beta})$.
\end{definition}
The well-behavedness assumption captures the well-studied family of isotropic log-concave distributions~\citep{lovasz2007geometry,balcan2013active}, and can potentially be more general.\footnote{However, the well-behavedness assumption here does not capture the family of $s$-concave distributions ($s \geq -\frac{1}{2d+3}$) studied in recent works~\citep{balcan2017sample}, as it requires any 1-d projection of distribution to have sub-exponential tail. Whether our analysis can be extended to $s$-concave distributions is an interesting open question.}
In addition to Massart and Tsybakov noise conditions, we also study a subfamily of Tsybakov noise, namely geometric Tsybakov noise. Such noise assumption was first considered in nonparametric active learning literature~\citep{castro2008minimax}. It generalizes the ``strong Massart noise'' condition considered in~\cite{diakonikolas2020learning,zhang2017hitting}, in that it allows $\frac12 - \eta(x)$ to grow polynomially with $\abr{\inner{w^\star}{x}}$, the distance between $x$ and the Bayes-optimal decision boundary $\cbr{x \in \RR^d: \inner{w^\star}{x} = 0}$.
\begin{definition}[Geometric Tsybakov noise condition]
\label{def:gtnc}
Given $B > 0$ and $\alpha \in (0,1]$, a distribution $D$ over $\RR^d \times \cbr{-1,+1}$ is said to satisfy the {\em $(B,\alpha)$-geometric Tsybakov noise condition} with respect to halfspace $w^\star$, if for all $x$ in $\RR^d$, $\frac12 - \eta(x) \geq \min\rbr{\frac12, B \abr{\inner{w^\star}{x}}^{\frac{1-\alpha}{\alpha}} }$, where $\eta(x) = \PP(y \neq \sign(\inner{w^\star}{x}) \mid x)$.
\end{definition}
It can be shown that, if the unlabeled distribution is well-behaved (Definition~\ref{def:well-behaved}), modulo a logarithmic factor, $(B,\alpha)$-geometric Tsybakov noise condition implies $(A,\alpha)$-Tsybakov noise condition with $A = \tilde{O}\rbr{(\frac 1 B)^{\frac \alpha {1-\alpha}}}$; see Lemma~\ref{lem:geomtnc-tnc} in Appendix~\ref{sec:auxiliary} for a formal statement.
\section{Algorithm}
We now describe our noise-tolerant active halfspace learning algorithm in detail. The main algorithm, Algorithm 1, has a simple structure: it first calls subprocedure $\ensuremath{\textsc{Initialize}}\xspace$ (line~\ref{step:init}) to generate a vector $v_1$ with $\ell_2$ distance at most $\frac{1}{4}$ to $w^\star$ with high probability. After obtaining $v_1$, it repeatedly calls $\ensuremath{\textsc{Optimize}}\xspace$ to refine its iterates $v_j$; as we will see, $\ensuremath{\textsc{Optimize}}\xspace$ guarantees that, with high probability, after iteration $j$, iterate $v_{j+1}$ is such that $\| v_{j+1} - w^\star \| \leq 4^{-(j+1)}$, i.e. after each iteration, an upper bound on $\|v_j - w^\star \|$ shrinks by a constant factor.
The algorithm returns after its final iteration $k_\epsilon$ is finished; after this iteration, we have $\theta(v_{k_\epsilon+1}, w^\star) \leq r_{\epsilon} = \tilde{O}\rbr{ \epsilon }$, which implies that
$\err(h_{\tilde{v}}, D) - \err(h_{w^\star}, D) \leq \PP_{D}(h_{\tilde{v}}(x) \neq h_{w^\star}(x)) \leq \epsilon$ (see Lemma~\ref{lem:prob-angle} in Appendix \ref{sec:auxiliary}).
Depending on different noise conditions on $D$, we use different schedules of sampling region bandwidths $\cbr{b_j}$ and numbers of label queries $\cbr{T_j}$:
{
\sloppy
\begin{enumerate}
\item Under the $\eta$-Massart noise condition, $b_j = b_{\mathsf{MNC}}(\eta, 4^{-(j+1)})$ and $T_j = T_{\mathsf{MNC}}(\eta, 4^{-(j+1)})$ for all $j \in \NN$, where $b_{\mathsf{MNC}}(\eta, r) = \tilde{\Theta}\rbr{ (1-2\eta) r}$ and $T_{\mathsf{MNC}}(\eta, r) = \tilde{O}\rbr{ \frac{d}{(1-2\eta)^2} (\ln\frac 1 {\delta r})^3 }$.
\item Under the $(A,\alpha)$-Tsybakov noise condition for $\alpha \in (\frac12, 1]$, $b_j = b_{\mathsf{TNC}}(A, \alpha, 4^{-(j+1)})$ and $T_j = T_{\mathsf{TNC}}(A, \alpha, 4^{-(j+1)})$ for all $j \in \NN$, where $b_{\mathsf{TNC}}(A, \alpha, r) = \tilde{\Theta}\rbr{ \min(r, (\frac{1}{A})^{\frac{1-\alpha}{2\alpha-1}} r^{\frac \alpha {2\alpha-1}}) }$ and $T_{\mathsf{TNC}}(A, \alpha, r) = \tilde{O}\rbr{ d (\ln\frac 1 {\delta r})^3 \cdot (1 + (\frac{A}{r})^{\frac{2-2\alpha}{2\alpha-1}})}$.
\item Under the $(B, \alpha)$-geometric Tsybakov noise condition, $b_j = b_{\mathsf{GTNC}}(B,\alpha, 4^{-(j+1)})$ and $T_j = T_{\mathsf{GTNC}}(B, \alpha, 4^{-(j+1)})$ for all $j \in \NN$, where $b_{\mathsf{GTNC}}(B, \alpha, r) = \tilde{\Theta}\rbr{ \min(r, B r^{\frac1\alpha}) }$, and
$T_{\mathsf{GTNC}}(B, \alpha, r) = \tilde{O}\left( d (\ln\frac 1 {\delta r})^3 (1 + \frac{1}{B^2 } (\frac{1}{r})^{\frac{2-2\alpha}{\alpha}} ) \right)$.
\end{enumerate}
}
For brevity, in the above definitions, the dependence on the unlabeled distribution parameters $(L, R, U, \beta)$ is ignored; we refer the readers to Appendix~\ref{sec:params} for more precise definitions of these functions.
The choices of $\cbr{b_j}$ and $\cbr{T_j}$ are to ensure that the algorithm's iterates $v_j$ are brought progressively closer to $w^\star$ with increasing $j$; this will be discussed in greater detail in Section~\ref{sec:theory}.
\begin{algorithm}[t]
\caption{Main algorithm}
\label{alg:main}
\begin{algorithmic}[1]
\REQUIRE{Bandwidth schedule $\cbr{b_j}$, iteration schedule $\cbr{T_j}$, target excess error $\epsilon \in (0,1)$, failure probability $\delta \in (0,1)$.}
\ENSURE{A halfspace $\tilde{v}$ such that $\err(h_{\tilde{v}}, D) - \err(h_{w^\star}, D) \leq \epsilon$.}
\STATE Define $r_\epsilon := \frac{\epsilon}{32 U \beta^2 (\ln\frac{12}{\epsilon})^2}$, and set $k_\epsilon := \lceil \log_4 \frac{1}{r_\epsilon} \rceil$ be the total number of iterations.
\STATE $v_1 \gets \ensuremath{\textsc{Initialize}}\xspace(\cbr{b_j}, \cbr{T_j})$.
\label{step:init}
\FOR{$j=1,\ldots,k_\epsilon$}
\STATE $v_{j+1} \gets \ensuremath{\textsc{Optimize}}\xspace(v_{j}, 4^{-(j+1)}, b_j, T_j, \mathsf{average})$.
\label{step:refine}
\ENDFOR
\RETURN $v_{k_\epsilon+1}$.
\end{algorithmic}
\end{algorithm}
We now discuss the two subprocedures employed by the main algorithm, \ensuremath{\textsc{Optimize}}\xspace and \ensuremath{\textsc{Initialize}}\xspace, in detail.
\subsection{Procedure \ensuremath{\textsc{Optimize}}\xspace and its guarantees}
\label{subsec:optimize}
Procedure \ensuremath{\textsc{Optimize}}\xspace (Algorithm~\ref{alg:optimize}) aims at refining its input halfspace $w_1$, so that it outputs a halfspace $\tilde{w}$ whose $\ell_2$ distance to $w^\star$ has an upper bound ($r$) at most a factor of $\frac14$ times the original $\ell_2$ distance upper bound between $w_1$ and $w^\star$ ($4r$), with good probability.
To this end, it maintains an iterate $w_t$; at each iteration, it performs adaptive sampling to obtain a labeled example $(x_t, y_t)$ drawn from $D_{\hat{w}_t, b}$, and updates with this new example using the well-known online gradient descent algorithm~\citep[e.g.][Chapter 11]{cesa2006prediction}.
Following standard active learning sampling strategies~\cite[e.g.][]{balcan2009agnostic}, every draw from distributions $D_{X \mid \hat{w}_t, b}$ is done by rejection sampling, i.e., keep querying $\mathrm{EX}$ until it returns an example in $B_{\hat{w}_t,b}$.
After generating the iterates $\cbr{w_t}_{t=1}^T$, it aggregates them using either a normalize-and-average step (lines~\ref{line:avg-start} to~\ref{line:avg-end}), or draws a vector uniformly at random from this set, multiplied by a random sign (lines~\ref{line:rnd-start} to~\ref{line:rnd-end}), depending on its aggregation mode $\mathrm{agg}$.
\ensuremath{\textsc{Optimize}}\xspace is similar to \ensuremath{\textsc{Refine}}\xspace in~\cite{zhang2020efficient}, but has two key differences. First, the update vector $g_t$ used in our \ensuremath{\textsc{Optimize}}\xspace algorithm is $-y_t x_t$, whereas the update vector $g_t$ in \ensuremath{\textsc{Refine}}\xspace depends on the Massart noise parameter $\eta$; this undesirable dependence on $\eta$ implies that \ensuremath{\textsc{Refine}}\xspace cannot be used for handling broader noise conditions such as Tsybakov noise. Second, it allows two aggregation modes to be used, and when to use which aggregation mode depends on the precision of the input (i.e. the $\ell_2$ closeness of input $w_1$ and $w^\star$): as we will see, in early learning stages, we will call \ensuremath{\textsc{Optimize}}\xspace with mode $\mathsf{random}$; in later stages, we will call \ensuremath{\textsc{Optimize}}\xspace with mode $\mathsf{average}$. \ensuremath{\textsc{Optimize}}\xspace is algorithmically similar to the nonconvex SGD algorithm of~\cite{diakonikolas2020learning} (see also earlier algorithmic insights of~\citet{guillory2009active}), who carefully construct a nonconvex learning objective such that under Massart noise, any stationary point of the objective corresponds to a vector close to $w^\star$; however, as we will see next, our analysis techniques are fairly different from theirs.
As its update rule suggest, \ensuremath{\textsc{Optimize}}\xspace performs online linear optimization with adaptively-chosen linear functions. Generalizing insights from prior work~\citep{zhang2020efficient}, our key observation is that, somewhat intriguingly, \ensuremath{\textsc{Optimize}}\xspace can be alternatively viewed as minimizing the following ``proximity function'' to $w^\star$:
\begin{definition}
Given distribution $D$ with its Bayes optimal classifier being a halfspace $h_{w^\star}$, and a positive number $b > 0$, define $\psi_{D,b}(w) := \EE_{D_{\hat{w},b}} \sbr{ (1-2\eta(x)) \abr{\inner{w^\star}{x}} }$.
\end{definition}
\begin{algorithm}[t]
\caption{$\ensuremath{\textsc{Optimize}}\xspace$}
\label{alg:optimize}
\begin{algorithmic}[1]
\REQUIRE vector $w_1$, target proximity $r$ (such that $\| w_1 - w^\star \| \leq 4 r$), bandwidth $b$, number of iterations $T$, aggregation mode $\mathrm{agg}$.
\ENSURE Optimized halfspace $\tilde{w}$ such that $\| \tilde{w} - w^\star \| \leq r$ with good probability.
\STATE Define $\Kcal = \cbr{w: \| w - w_1 \| \leq 4 r}$, step size $\alpha = \frac{r}{\beta} \sqrt{\frac{1}{dT}} / \rbr{\ln\frac{Td}{\delta r b R L}}$.
\FOR{$t=1, 2, \dots, T$}
\label{line:omd-loop-start}
\STATE Sample $x_t$ from $D_{X \mid \hat{w}_t, b}$ using rejection sampling, and query $\Ocal$ for its label $y_t$.
\STATE Update $w_{t+1} \gets\arg\min_{w \in \Kcal} \| w_t - \alpha g_t \|$,
where $g_t = - y_t x_t$.
\label{line:omd}
\ENDFOR
\label{line:omd-loop-end}
\IF{$\mathrm{agg} = \mathsf{average}$}
\label{line:avg-start}
\RETURN $\tilde{w} \gets \frac{1}{T} \sum_{t=1}^T \hat{w}_t$.
\label{line:avg-end}
\ELSIF{$\mathrm{agg} = \mathsf{random}$}
\label{line:rnd-start}
\RETURN $\tilde{w} \gets \sigma \cdot \hat{w}_\tau$, where $\tau$ is chosen uniformly at random from $[T]$, and $\sigma$ is chosen uniformly at random from $\cbr{-1,+1}$.
\label{line:rnd-end}
\ENDIF
\end{algorithmic}
\end{algorithm}
Recall that the three noise conditions considered in our paper all assume that $h_{w^\star}$ is the Bayes optimal classifier. Therefore, $\eta(x) \leq \frac12$ for all $x$, and consequently $\psi_{D,b}$ always takes nonnegative values.
In addition, $\psi_{D,b}$ is scale-invariant: as $B_{\widehat{\alpha w}, b} = B_{\hat{w}, b}$, $\psi_{D,b}(\alpha w) = \psi_{D,b}(w)$ for any $w \in \RR$ and $\alpha \neq 0$.
Informally, $\psi_{D,b}$ is a distance proxy function that measures the closeness of input $w$ and the optimal $w^\star$, although the closeness here is defined by a nonstandard measure $\tilde{\theta}(\cdot, \cdot)$ (recall its definition in Section~\ref{sec:setting}). This is formalized in the following lemma (proof in Appendix~\ref{sec:pot}):
\begin{lemma}
Suppose $D_X$ is $(2, L, R, U, \beta)$-well-behaved. In addition, $w \in \RR^d$ and $b > 0$ is such that $\tilde{\theta} = \tilde{\theta}(w, w^\star) \geq \Omega(b)$. Then:
\begin{enumerate}
\item if $D$ satisfies $\eta$-Massart noise condition, $\psi_{D,b}(w) \geq \tilde{\Omega}\rbr{ (1-2\eta) \tilde{\theta} }$.
\item If $D$ satisfies $(A, \alpha)$-Tsybakov noise condition, $\psi_{D,b}(w) \geq \Omega\rbr{ (\frac b A)^{\frac{1-\alpha}{\alpha}} \tilde{\theta} }$.
\item If $D$ satisfies $(B, \alpha)$-geometric Tsybakov noise condition, then $\psi_{D,b}(w) \geq \tilde{\Omega}\rbr{ \min( \tilde{\theta}, B\tilde{\theta}^{\frac{1}{\alpha}}) }$.
\end{enumerate}
\label{lem:pot-lb-main-text}
\end{lemma}
The following key lemma formalizes the aforementioned claim that \ensuremath{\textsc{Optimize}}\xspace produces iterates $w_t$'s that approximately minimize $\psi_{D,b}$; specifically, the average value of $\psi_{D,b}(w_t)$'s is well-controlled, if the sampling bandwidth $b$ is small and the number of iterations $T$ is large.
\begin{lemma}
Suppose $D_X$ is $(2, L, R, U, \beta)$-well-behaved. There exists a numerical constant $c > 0$ such that the following holds. \ensuremath{\textsc{Optimize}}\xspace, with input initial vector $w_1$, target proximity $r \in (0,\frac 14]$ such that $\| w_1 - w^\star \| \leq 4r$, bandwidth $b \leq \frac R 2$, number of iterations $T$, produces iterates $\cbr{w_t}_{t=1}^T$, such that
with probability $1-\delta r$,
\[
\frac1T \sum_{t=1}^T \psi_{D,b}(w_t)
\leq
c \rbr{ b
+
(b + \beta r) \cdot \rbr{\ln\frac{Td}{\delta r b R L}} (\sqrt{\frac{d + \ln\frac1{\delta r}}{T}} + \frac{\ln\frac1{\delta r}}{T}) }.
\]
\label{lem:optimize-g}
\end{lemma}
The proof of Lemma~\ref{lem:optimize-g} can be found at Appendix~\ref{sec:optimize-g}.
Its key insight is similar to the ideas in~\cite{zhang2020efficient}: we derive a regret guarantee of the online linear optimization problem induced by the adaptively-chosen gradient vectors $\cbr{g_t}_{t=1}^T$, which implies an upper bound on the {\em negative benchmark} term. Thanks to the adaptive sampling scheme, the negative benchmark concentrates to $\sum_{t=1}^T \psi_{D,b}(w_t)$. Although the proof of Lemma~\ref{lem:optimize-g} uses standard regret results on online linear optimization, it is not a direct consequence of the standard reduction from online convex optimization to online linear optimization -- $\psi_{D,b}(w)$ is not necessarily convex in $w$.
We now discuss the aggregation mode $\mathrm{agg}$ in more detail. As \ensuremath{\textsc{Optimize}}\xspace is called by the main algorithm (Algorithm~\ref{alg:main}) repeatedly, we discuss its different settings in earlier and later stages of the main algorithm respectively.
In later stages of calling $\ensuremath{\textsc{Optimize}}\xspace$ (specifically, when its input target proximity $r \leq \frac{1}{16}$), the constraint set $\Kcal$ ensures that all $w_t$'s have acute angles with $w^\star$; in this case, for all $t$, $\tilde{\theta}(w_t, w^\star) = \theta(w_t, w^\star)$, and therefore upper bounds on $\psi_{D,b}(w_t)$ imply upper bounds on $\theta(w_t, w^\star)$.
In this ``local convergence'' regime, applying the guarantees provided by Lemma~\ref{lem:optimize-g}, a deterministic average over the normalized iterates $\hat{w}_t$'s achieves the target $\ell_2$ proximity to $w^\star$; see item~\ref{item:avg} of Lemma~\ref{lem:optimize-main} or Lemma \ref{lem:optimize-main-avg-only} for a precise statement. This corresponds to the $\mathsf{average}$ mode (lines~\ref{line:avg-start} to~\ref{line:avg-end}).
In contrast, in early stages of calling \ensuremath{\textsc{Optimize}}\xspace (specifically, when $r > \frac1{16}$), its iterates $\cbr{w_t}_{t=1}^T$ may not yet be in acute angle with $w^\star$; in this case, it is hard to guarantee that the average normalized iterate $\frac{1}{T} \sum_{t=1}^T \hat{w}_t$ is close to $w^\star$. This motivates the second mode $\mathsf{random}$ (lines~\ref{line:rnd-start} to~\ref{line:rnd-end}), which returns a vector uniformly at random from $\hat{w}_t$'s, times a random sign $\sigma$. By the guarantees of Lemma~\ref{lem:optimize-g}, it can be shown that with appropriate settings of parameters $\alpha, T, b$, \ensuremath{\textsc{Optimize}}\xspace guarantees to return $\tilde{w}$ with a constant $\ell_2$ distance to $w^\star$ with a constant probability; see item~\ref{item:rnd} of Lemma~\ref{lem:optimize-main} for a precise statement.
\subsection{Procedure \ensuremath{\textsc{Initialize}}\xspace and its guarantees}
Procedure \ensuremath{\textsc{Initialize}}\xspace (Algorithm~\ref{alg:initialize}) aims at label-efficiently learning a halfspace $\hat{u}_0$ amenable to local refinement: it guarantees that with high probability, the output halfspace $\hat{u}_0$ satisfies that $\| \hat{u}_0 - w^\star \| \leq \frac14$.
At a high level, \ensuremath{\textsc{Initialize}}\xspace uses \ensuremath{\textsc{Optimize}}\xspace as a black box in a label efficient manner.
Recall from the previous subsection that, with appropriate settings of input parameters, running \ensuremath{\textsc{Optimize}}\xspace with aggregation mode $\mathsf{random}$ guarantees to output a halfspace $\frac14$-close to $w^\star$ with constant probability. \ensuremath{\textsc{Initialize}}\xspace ``boosts'' the above guarantee, in that it increases the probability of outputting a vector $\frac14$-close to $w^\star$ from a small constant to $1-O(\delta)$.
\ensuremath{\textsc{Initialize}}\xspace consists of two stages. In the first stage (lines~\ref{line:stage-1-start} to~\ref{line:stage-1-end}), it generates $U$, a set of halfspaces of size $N = \lceil 10 \ln\frac 4 \delta \rceil$, such that at least one element in $U$ has a small excess error rate, specifically $O(\epsilon_0)$; here the choice of $\epsilon_0$ depends on the noise condition of $D$, as will be discussed next. To achieve this target excess error rate, it runs \ensuremath{\textsc{Optimize}}\xspace for $k_0 = \tilde{O}(\log\frac{1}{\epsilon_0})$ iterations; Claim~\ref{claim:stage1} in Appendix~\ref{sec:initialize} shows that each trial $i$ generates $v_{i,k_0+1}$ with excess error $O(\epsilon_0)$ with constant probability. Because of the independence of the $N$ trials, with high probability, one of the $v_{i,k_0+1}$'s will have excess error $O(\epsilon_0)$.
\begin{algorithm}[t]
\caption{\ensuremath{\textsc{Initialize}}\xspace: a label efficient acute initialization procedure}
\label{alg:initialize}
\begin{algorithmic}[1]
\REQUIRE{Bandwidth schedule $\cbr{b_j}$, iterations schedule $\cbr{T_j}$.}
\ENSURE{Unit vector $\hat{u}_0$ such that $\| \hat{u}_0 - w^\star \| \leq \frac14$.}
\STATE Define $r_0 = \frac{\epsilon_0}{64 U \beta^2 (\ln\frac{24}{\epsilon_0})^2}$, where $\epsilon_0$ is defined according to the noise condition of $D$; $k_0 = \lceil \log_4 (\frac{1}{r_0}) \rceil$ is the number of iterations per trial; $N = \lceil 10 \ln \frac {4}{\delta} \rceil$ the number of trials in the first stage.
\FOR{$i=1,2,\ldots,N$}
\label{line:stage-1-start}
\STATE Initialize $v_{i,0} \gets \overrightarrow{0}$.
\FOR{$j=0,1,2,\ldots,k_0$}
\STATE Set $\mathrm{agg} \gets \mathsf{random}$ if $j = 0$; $\mathrm{agg} \gets \mathsf{average}$ otherwise.
\STATE $v_{i,j+1} \gets \ensuremath{\textsc{Optimize}}\xspace(v_{i,j}, 4^{-(j+1)}, b_j, T_j, \mathrm{agg})$.
\ENDFOR
\ENDFOR
\STATE $U \gets \cbr{v_{i,k_0+1}: i \in [N]}$.
\label{line:stage-1-end}
\STATE $S \leftarrow$ draw $O(\frac{1}{\epsilon_0^2}\ln \frac {N}{\delta} )$ random unlabeled examples from $D_X$, and query $\Ocal$ for their labels.
\label{line:stage-2-start}
\RETURN $\hat{u}_0 = \frac{u_0}{\| u_0\|}$, where $u_0 = \argmin_{u \in U} \err(h_u, S)$.
\label{line:stage-2-end}
\end{algorithmic}
\end{algorithm}
In the second stage (lines~\ref{line:stage-2-start} to~\ref{line:stage-2-end}), it draws $S$, a set of labeled examples from $D$, and selects the halfspace $\hat{u}$ in $U$ with the smallest empirical error on $S$.
Combining the guarantees of $U$ in the first stage, the choice of $|S| = \tilde{O}(\epsilon_0^{-2})$, and standard guarantees of empirical risk minimization, it is guaranteed that $\hat{u}$ has excess error $\tilde{O}(\epsilon_0)$ with high probability.
Parameter $\epsilon_0$ is set to ensure that the above excess error guarantee can be translated to a geometric $\ell_2$-proximity guarantee, and therefore depends on different noise conditions on $D$. Specifically, we set $\epsilon_0$ as:
\begin{enumerate}
\item $\epsilon_{\mathsf{MNC}}(\eta) = \tilde{O}\rbr{ 1-2\eta }$, under the $\eta$-Massart noise condition;
\item $\epsilon_{\mathsf{TNC}}(A, \alpha) = \tilde{O}\rbr{ (\frac{1}{A})^{\frac{1-\alpha} {\alpha}} }$, under the $(A, \alpha)$-Tsybakov noise condition with $\alpha \in (\frac12, 1]$;
\item $\epsilon_{\mathsf{GTNC}}(B, \alpha) = \tilde{O}\rbr{ B }$, under the $(B,\alpha)$-geometric Tsybakov noise condition.
\end{enumerate}
We again refer the readers to Appendix~\ref{sec:params} for more precise definitions of these functions, with their dependence on $(L, R, U, \beta)$ explicit.
We show the following guarantee of \ensuremath{\textsc{Initialize}}\xspace, under any one of the three noise conditions considered.
\begin{lemma}
Suppose $D$ is $(2, L, R, U, \beta)$-well behaved and satisfies one of the three noise conditions. With the respective settings of $\cbr{b_j}, \cbr{T_j}, \epsilon_0$, \ensuremath{\textsc{Initialize}}\xspace outputs a unit vector $\hat{u}_0$, such that with probability $1-\delta/2$, $\| \hat{u}_0 - w^\star \| \leq \frac{1}{4}$. The total number of label queries by \ensuremath{\textsc{Initialize}}\xspace is at most:
\begin{enumerate}
\item $\tilde{O} \rbr{ \frac{d}{(1-2\eta)^2} }$, if $D$ satisfies $\eta$-Massart noise;
\item $\tilde{O} \rbr{ d \cdot \rbr{ 1 + A^{\frac{2-2\alpha}{\alpha(2\alpha-1)}} } }$, if $D$ satisfies $(A, \alpha)$-Tsybakov noise with $\alpha \in (\frac12, 1]$;
\item $\tilde{O} \rbr{ d \cdot \rbr{ 1 + (\frac{1}{B})^{\frac 2 \alpha} } }$, if $D$ satisfies $(B,\alpha)$-geometric Tsybakov noise.
\end{enumerate}
\label{lem:initialize}
\end{lemma}
Specifically, under the $\eta$-Massart noise condition, this yields a procedure that can output a vector $\hat{u}$ with constant proximity to the optimal halfspace $w^\star$ with high probability, using $\tilde{O}(\frac{d}{(1-2\eta)^2})$ label queries.
This label complexity matches the $\Omega(\frac{d}{(1-2\eta)^2})$ information-theoretic lower bound~\cite[e.g.][Theorem 1]{yan2017revisiting}. When specialized to isotropic log-concave unlabeled distribution settings, this resolves an open problem by~\cite{zhang2020efficient} on whether there is an efficient and label-optimal initialization procedure that reliably computes a vector with small constant angle with $w^\star$.
\section{Performance guarantees}
\label{sec:theory}
We now present Theorem~\ref{thm:main}, the main result of this paper.
\begin{theorem}
Fix $\epsilon \in (0,1)$ and $\delta \in (0, \frac{1}{10})$. Suppose $D$ is $(2, L, R, U, \beta)$-well behaved and satisfies one of the three noise conditions. With the settings of $\cbr{b_j}, \cbr{T_j}$, and $\epsilon_0$ under the respective noise conditions, with probability $1-\delta$, Algorithm~\ref{alg:main} outputs a halfspace $\tilde{v}$, such that $\err(h_{\tilde{v}}, D) - \err(h_{w^\star}, D) \leq \epsilon$. In addition, its total number of label queries is at most:
\begin{enumerate}
\item $\tilde{O}\rbr{ \frac{d}{(1-2\eta)^2} \polylog(\frac1\epsilon)}$, if $D$ satisfies $\eta$-Massart noise;
\item $\tilde{O}\rbr{ d \cdot \rbr{1 + A^{\frac{2-2\alpha}{\alpha(2\alpha-1)}} + (\frac{A}{\epsilon})^{\frac{2-2\alpha}{2\alpha-1}}} }$, if $D$ satisfies $(A, \alpha)$-Tsybakov noise with $\alpha \in (\frac12, 1]$;
\item $\tilde{O}\rbr{ d \cdot \rbr{ 1 + (\frac{1}{B})^{\frac 2 \alpha} + \frac{1}{B^2} (\frac 1 \epsilon)^{\frac{2-2\alpha}{\alpha}}} }$, if $D$ satisfies $(B,\alpha)$-geometric Tsybakov noise.
\end{enumerate}
\label{thm:main}
\end{theorem}
Specialized to the $\eta$-Massart noise condition, our label complexity bound matches information-theoretic lower bounds~\cite[e.g.][Theorem 1]{yan2017revisiting} up to polylogarithmic factors, and improves over the state-of-the-art active halfspace learning algorithm of~\cite{zhang2020efficient} in two aspects. First, our label complexity is a factor of $O\rbr{ \frac{1}{(1-2\eta)^2} }$ lower than~\cite{zhang2020efficient}, thanks to the new initialization procedure; second, our algorithm and analysis allows for dealing with a broader set of unlabeled data distributions beyond isotropic log-concave, matching the assumptions employed in recent works~\citep[e.g.][]{diakonikolas2020polynomial}.
Under the $(A, \alpha)$-Tsybakov noise condition, our theorem provides nontrivial label complexity results when $\alpha \in (\frac12, 1]$. In the extreme case when $\alpha = 1$, our algorithm has a label complexity of $\tilde{O}\rbr{ d \polylog(\frac1\epsilon) }$. The label complexity bound becomes higher when $\alpha$ is further away from $1$. Compared to the recent passive learning algorithm of~\cite{diakonikolas2020polynomial} that can tolerate $(A,\alpha)$-Tsybakov noise for any $\alpha \in (0,1]$, our results cannot allow $\alpha$ to be in $(0,\frac12]$, but our algorithm has better label efficiency when $\alpha$ is close to 1.
Under the $(B, \alpha)$-geometric Tsybakov noise condition, our label complexity bound $\tilde{O}\rbr{ d (\frac{1}{\epsilon})^{\frac{2-2\alpha}{\alpha}} }$ is higher than the computationally inefficient algorithm of~\cite{balcan2013active}, which has a label complexity of $\tilde{O}\rbr{ d (\frac1\epsilon)^{2-2\alpha} }$.
This is due to a limitation of our current proof technique: we reduce the goal of achieving excess error guarantee to achieving geometric proximity.
Our proof in fact yields a stronger result: with $\tilde{O}\rbr{ d (\frac{1}{\epsilon})^{\frac{2-2\alpha}{\alpha}} }$ label queries, our algorithm outputs a halfspace that has angle $O(\epsilon)$ with $w^\star$ with high probability; this result matches the information-theoretic lower bound of~\cite{wang2016noise} in achieving closeness-in-angle guarantees, in the dependence on $\epsilon$. We leave whether it is possible to develop efficient active learning algorithms with label complexity guarantees matching those of computationally inefficient algorithms in this setting as an important open question.
\begin{remark}[Unlabeled sample complexity of Algorithm~\ref{alg:main}]
Our active learning algorithm consumes a total of $O(\poly(d, \frac1\epsilon))$ unlabeled examples with high probability.
To see this, note that our sampling regions' bandwidths all satisfy $b = \Omega(\poly(\epsilon))$ under each of the three noise conditions and therefore have probability masses $\Omega(\poly(\epsilon))$ by Lemma~\ref{lem:1d-prob-ub}. This implies that with high probability, each active sampling invokes at most $O(\poly(\frac1\epsilon))$ calls to the unlabeled example oracle $\mathrm{EX}$; this implies that the total number of calls to $\mathrm{EX}$ is also $O(\poly(d, \frac1\epsilon))$.
\end{remark}
\begin{remark}[Attribute efficiency]
Our algorithm and analysis can be straightforwardly modified to achieve attribute efficiency \citep[e.g.][]{littlestone1987learning,blum1990learning,awasthi2016learning,zhang2018efficient}, i.e. achieving label complexities that exploit the sparsity of the Bayes-optimal linear classifier $w^\star$. Specifically, under the extra assumption that $w^\star$ is $s$-sparse ($s \ll d$), a variant of our algorithm achieves a guarantee similar to Theorem~\ref{thm:main}, with the dimension $d$ in the label complexity bounds replaced with $s \polylog(d)$. We provide the details in Appendix~\ref{sec:sparsity}.
\end{remark}
\subsection{Proof sketch of Theorem~\ref{thm:main}}
We now outline the proof of Theorem~\ref{thm:main}.
Recall that from Lemma~\ref{lem:initialize}, line~\ref{step:init} of the main algorithm calls \ensuremath{\textsc{Initialize}}\xspace to generate vector $v_1$ such that $\| v_1 - w^\star \| \leq \frac14$ with probability $1-\delta/2$. This step uses $\tilde{O}\rbr{\frac{d}{(1-2\eta)^2}}$, $\tilde{O}\rbr{d \cdot (1 + A^{\frac{2-2\alpha}{\alpha(2\alpha-1)}}) }$, and $\tilde{O}\rbr{ d \cdot (1 + (\frac{1}{B})^{\frac 2 \alpha})}$ label queries to $\Ocal$, under the three noise conditions respectively.
For the guarantees in subsequent rounds, we rely on the following lemma, which shows that repeatedly applying \ensuremath{\textsc{Optimize}}\xspace yields local convergence guarantees.
Specifically, this lemma implies that, given an input halfspace $v_j \in \RR^d$ such that $\| v_j - w^\star \| \leq 4^{-j}$ at the beginning of the $j$-th iteration of the main algorithm (Algorithm~\ref{alg:main}), \ensuremath{\textsc{Optimize}}\xspace, with settings of bandwidth parameter $b_j$ and number of iterations $T_j$, outputs a refined halfspace $v_{j+1}$ such that its $\ell_2$ distance with $w^\star$ is at most $4^{-(j+1)}$ with high probability.
\begin{lemma}
Fix $r \in (0,\frac{1}{16}]$, and $\delta \in (0,\frac1{10})$.
Suppose $D$ is $(2, L, R, U, \beta)$-well behaved and satisfies one of the three noise conditions; in addition,
$b$ and $T$ are such that:
\begin{enumerate}
\item $b = b_{\mathsf{MNC}}(\eta, r), T = T_{\mathsf{MNC}}(\eta, r)$, if $D$ satisfies $\eta$-Massart noise;
\item $b = b_{\mathsf{TNC}}(A, \alpha, r), T = T_{\mathsf{TNC}}(A, \alpha, r)$, if $D$ satisfies $(A, \alpha)$-Tsybakov noise with $\alpha \in (\frac12, 1]$;
\item $b = b_{\mathsf{GTNC}}(B, \alpha, r), T = T_{\mathsf{GTNC}}(B, \alpha, r)$, if $D$ satisfies $(B, \alpha)$-geometric Tsybakov noise.
\end{enumerate}
Then \ensuremath{\textsc{Optimize}}\xspace, with input initial $w_1$ satisfying $\| w_1 - w^\star \| \leq 4 r$, target proximity $r$, bandwidth $b$, number of iterations $T$, aggregation method $\mathrm{agg} = \mathsf{average}$, outputs $\tilde{w}$ such that probability $1-\delta r$,
$\| \tilde{w} - w^\star \| \leq r$.
\label{lem:optimize-main-avg-only}
\end{lemma}
The proof of Lemma~\ref{lem:optimize-main-avg-only} can be found at Appendix~\ref{sec:optimize}.
Some intuitions on this lemma have been given in Section~\ref{subsec:optimize}, and we elaborate on its key ideas in greater detail here. Recall that
Lemma~\ref{lem:optimize-g} shows that running \ensuremath{\textsc{Optimize}}\xspace gives an upper bound on $\frac 1 T \sum_{t=1}^T \psi_{D,b}(w_t)$, the average value of $\psi_{D,b}(w_t)$'s, in terms of $b$ and $T$.
We set $b$ and $T$ differently under different noise conditions, so that $\frac 1 T \sum_{t=1}^T \psi_{D,b}(w_t)$ can be controlled at an appropriate level.
Markov's Inequality implies that there is an overwhelming fraction ($\geq \frac{31}{32}$) of $w_t$'s with small $\psi_{D,b}(w_t)$ - denote by $S$ the set of such $t$'s. Now, we conduct a case analysis:
\begin{enumerate}
\item For every $t$ in $S$, $\psi_{D,b}(w_t)$ is small. Recall that Lemma~\ref{lem:pot-lb-main-text} shows that a small value of $\psi_{D,b}(w_t)$ implies a small value of $\tilde{\theta}(w_t, w^\star)$.
In addition, the diameter of the constraint set $\Kcal$ is at most $8 r \leq \frac12$, and both $w_t$ and $w^\star$ are in $\Kcal$, so $\theta(w_t, w^\star)$ is acute (see Lemma~\ref{lem:angle-l2}) and is equal to $\tilde{\theta}(w_t, w^\star)$.
Hence, for all $t$ in $S$, $\theta(w_t, w^\star) \leq \frac{r}{2}$ and consequently, $\| \hat{w}_t - w^\star \| \leq \frac r 2$.
\item On the other hand, for every $t$ in $[T] \setminus S$, we still have $w_t \in \Kcal$, so $\| \hat{w}_t - w^\star \|$ is at most $16 r$.
\end{enumerate}
By averaging over the upper bounds on $\| \hat{w}_t - w^\star \|$, and using the convexity of $\ell_2$ norm, we conclude that $\|\tilde{w} - w^\star \| = \| \frac1T \sum_{t=1}^T \hat{w}_t - w^\star \| \leq \frac1T \sum_{t=1}^T \| \hat{w}_t - w^\star \| \leq r$.
Equipped with the above initialization and local convergence guarantees, the proof of Theorem~\ref{thm:main} is now straightforward; its details can be found at Appendix~\ref{sec:proof-main}.
\section{Conclusions and open problems}
We provide an efficient active halfspace learning algorithm that can achieve new label complexity guarantees under Massart and Tsybakov noise conditions, under certain structural assumptions on the unlabeled data distribution. Specifically, our algorithm achieves a near-optimal label complexity under the Massart noise condition, and achieves new label complexity guarantees under two subfamilies of Tsybakov noise conditions. A key open problem is to develop efficient algorithms with label complexities matching those of computationally inefficient approaches under $(A,\alpha)$-Tsybakov noise, for all $\alpha \in (0,1]$.
Another interesting open question is to design efficient active learning algorithms that can adapt to unknown noise conditions.
\paragraph{Acknowledgments.} We thank Yining Wang for helpful discussions on label complexity lower bounds in~\citep{wang2016noise} for active learning halfspaces under Tsybakov noise under the uniform distribution. We also thank the anonymous reviewers for their constructive feedback.
|
\section{Introduction}
Networks or graphs are frequently used to capture various relationships that exist in the real world, and thus we witness the emergence of social networks~\cite{xuan2019self, kim2018social, 8281007}, traffic networks~\cite{ruan2019empirical, tang2020predictability, xu2020ge}, biological networks~\cite{walter2004visualization, wale2008comparison, zhou2020m}, literature citation networks~\cite{hosseini2018analysis, yasunaga2019scisummnet}, etc. The recently proposed graph representation methods allow us to better understand the structures of these networks and promote the development of various disciplines. Interestingly, the early graph embedding methods were benefited from natural language processing~\cite{MotifBasedAttention}, while now the graph neural networks (GNN) are used to successfully deal with visual semantic segmentation~\cite{2001.00335}. Furthermore, these graph embedding methods have made remarkable achievements in such areas as recommendation systems\cite{cheng2016wide,wang2019knowledge}, QA sites~\cite{zhang2018diffusion, fu2019nes}, and even drug discovery~\cite{jing2018deep, lane2018comparing}. In fact, network science, together with machine learning (especially deep learning), has made an important contribution to the development of cross-disciplines.
Subgraphs or motifs \cite{liu2019link,xuan2015temporal}, as basic building blocks, can be used to describe the mesoscale structure of a network. The networks constructed by different subgraphs may have vastly different topological properties and functions, and thus could be integrated into many graph algorithms to improve their performances.
For instance, after extracting the root subgraph with a modified skip-gram model, Narayanan et al.~\cite{1606.08928} proposed Subgraph2Vec as an unsupervised representation learning method, leading to good performance on graph classification. Ugander et al.~\cite{ugander2013subgraph} treated subgraph frequencies in social networks as local attributes and found that subgraph frequencies do provide unique insights for identifying social and graph structures of large networks. Inspired by neural document embedding models, Nguyen et al.~\cite{doi:10.1137/1.9781611975321.35} proposed the GE-FSG method, which adopts a series of frequent subgraphs as the inputs of the PV-DBOW model to obtain the entire-graph embeddings, achieving good performance in graph classification and clustering.
These studies focus more on the basic statistics, e.g., the number of subgraphs, but lack analysis of the underlying structure among these subgraphs. The recently proposed subgraph network (SGN) model~\cite{xuan2019subgraph} takes the above issue into consideration and connects different subgraphs to construct a new network at a higher level. This process can be iterated to form a series of SGNs of different orders. It has been proven that SGNs can effectively expand the structural space and further improve the performance of network algorithms.
However, SGN model has the following two shortages. First, the rule to establish SGN is deterministic, i.e., users can generate only one SGN of each order for a network. Such lack of diversity will limit the capacity of SGN to expand the latent structure space. Second, when the number of subgraphs exceeds the number of nodes in a network, the generated SGN can be even larger than the original network, which makes it extremely time-consuming to process SGNs of the higher-order, letting alone integrating these SGNs to design algorithms of better performances. On the other hand, it is noted that network sampling can increase the diversity by introducing the randomness, and meanwhile control the scale, providing an effective and inexpensive solution for network analysis. This merit thus is exactly complementary to the SGN model.
In this paper, we introduce network sampling into the SGN model, and proposes \underline{S}ampling \underline{S}ub\underline{G}raph \underline{N}etwork ($\text{S}^2$GN). In particular, we utilize the following four network sampling strategies, including random walk, biased walk, link selection, and spanning tree, to sample a subnetwork containing certain numbers of nodes and links, and then map the subnetwork to SGN based on certain rules. Network sampling and SGN construction can be used iteratively, so as to create a series of $\text{S}^2$GN of different orders, whose structural features can then be fused with those of the original network, so as to enhance a number of network algorithms. Specifically, we have the following contributions:
\begin{itemize}
\item We propose a new network model, sampling subgraph network ($\text{S}^2$GN), by introducing network sampling into SGN. Compared with SGN, our $\text{S}^2$GN can increase the diversity and decrease the complexity to a certain extent, benefiting the subsequent network algorithms.
\item We propose hierarchical fusion to fully utilize the structural information extracted from $\text{S}^2$GNs of different orders, generated by different sampling strategies, to enhance various graph classification algorithms based on manual attributes, Graph2Vec, DeepKernel, and CapsGNN.
\item We apply the new method to eight real-world network datasets, and our experimental results demonstrate the effectiveness and efficiency of $\text{S}^2$GN. The fusion of $\text{S}^2$GNs generated by different sampling strategies can increase the performance of graph classification algorithms in 30 out of 32 cases, with a relative improvement of 10.75\% on average (4.68\% for SGN). This value increases to 14.49\% (2.06\% for SGN) when only CapsGNN is considered, i.e., the combination of $\text{S}^2$GN-Fusion and CapsGNN achieves the $F_1$-$Score$ 80.98\% on average, greatly improving the graph classification performance. More remarkably, compared with SGN, generating S$^2$GNs needs much less time, reduced by almost two orders of magnitude.
\end{itemize}
The rest of the paper is structured as follows. In Sec.~\ref{sec:related}, we briefly describe the related work in network sampling and feature extraction. In Sec.~\ref{sec:SSGN}, we introduced the construction method of $\text{S}^2$GN. In Sec.~\ref{sec:experiments}, we give several feature extraction methods, which together with $\text{S}^2$GN are applied to eight real-world network datasets. Finally, we conclude the paper and highlight some promising directions for future work in Sec.~\ref{sec:Con}.
\section{Related work}\label{sec:related}
In this section, to supply some necessary background information, we give a brief overview of network sampling strategies and graph representation algorithms in graph mining and network science.
\subsection{Network Sampling}\label{sec:sampling}
Our work is closely related to the line of research in the network analysis based on sampling. Sampling methods in graph mining have two main tasks: generating node sequences and limiting the scale of the network. For the former, many studies utilize sampling strategies to extract node sequences to provide materials for subsequent network representation. Random walk~\cite{noh2004random} is one of the most famous node sampling methods, which has a wide influence in the field of graph mining~\cite{andersen2006local, fouss2007random}. For example, DeepWalk~\cite{perozzi2014deepwalk} combined the random walk with the language model in NLP, which was applied to node classification as a graph embedding method. In addition, Grover and Leskovec~\cite{grover2016node2vec} designed a biased walk mechanism based on random walk, which had a further improvement in node classification. Breadth-First Sampling~\cite{kurant2010bias} is a node sampling algorithm, which is biased to the nodes of high degrees and has been successfully applied in the measurement and topological analysis of OSNs. By limiting the scale of a network, Satuluri et al.~\cite{10.1145/1989323.1989399} sparsified graphs and achieved faster graph clustering without sacrificing quality. Moreover, sampling on graphs also has a wide spectrum of applications on network visualization~\cite{devi2019graph}. The sampling method can simplify the network while preserving significant structure information, which
is of ultra importance in graph mining.
\subsection{Graph Representation}\label{sec:graphrep}
The most naive network representation method is to calculate graph attributes according to certain typical topological metrics~\cite{li2011graph}. Early graph embedding methods were considerably affected by NLP. For example, as graph-level embedding algorithms, Narayanan et al. proposed Subgraph2Vec~\cite{1606.08928} and Graph2Vec~\cite{narayanan2017graph2vec}, which achieve good performances on graph classification.
Another popular approach is to use graph kernel methods to capture the similarity between graphs.
Although representing networks well, they generally have relatively high computational complexity~\cite{li2011graph}. It is worth mentioning that the WL kernel~\cite{10.5555/1953048.2078187} was used to make the subgraph isomorphism check more effective. On this basis, Yanardag and Vishwanathan~\cite{yanardag2015deep} proposed an alternative kernel formulation termed as Deep Graph Kernel (DeepKernel) which achieved good performances on several datasets.
With the rise of spectral analysis of graph data in recent years, graph convolutional neural network (GCN) has been developed. It uses the Laplace decomposition of graphs to achieve convolutional operation in the spectral domain. Kipf et al.~\cite{kipf2016semi} used this neural network structure for semi-supervised learning, and achieved excellent results. Later, mathematical analysis on GCN went further and proved that the Laplacian decomposition used by GCN and Laplacian smoothing on images have mathematically equivalent forms~\cite{Li2018DeeperII}. At the same time, GCNs in the spatial domain have also been proposed. Inspired by the idea of convolution kernels in CNN, Mathias et al.~\cite{10.5555/3045390.3045603} proposed the method of PATCHY-SAN, which can determine the direction of the convolutions and the order of the nodes in the convolution window, and this model also achieved good results in graph classification. In this way, GCN treats the obtained information without weighting, i.e. the information of important neighbors and non-important neighbors will be put into the convolution layer in an unbiased manner. GAT overcomes this shortage by supplementing a self-attention coefficient before the convolution layer~\cite{veli2018graph}. Based on the newly proposed capsule network architecture, Zhang et al.~\cite{xinyi2018capsule} designed a CapsGNN to generate multiple embeddings for each graph, thereby capturing the classification-related information and the potential information with respect to the graph properties at the same time, which achieved the good performance.
Although the above graph representation methods have relatively high expressiveness and learning ability, largely improving the performance of graph classification, they do not have good interpretability, and in addition, they only rely on a single network structure, limiting their ability to exploit the latent structural space. Therefore, we generate multiple S$^2$GNs to fully expand the latent structural space, so as to enhance the network algorithms. Our experiments have demonstrated that S$^2$GNs can be naturally integrated with many graph representation methods by our feature fusion framework for the further improvement of their effectiveness.
\section{Methodology}\label{sec:SSGN}
We first briefly review SGN and the four network sampling methods. Then we introduce the framework to establish $\text{S}^2$GN.
\subsection{Subgraph network}\label{sec:sgn}
Subgraph network (SGN)~\cite{xuan2019subgraph} is considered as a mapping function in network space. It provides a scalable model that transforms the original node-level network into a subgraph-level network. As shown in Fig.~\ref{fig:SGN}, the SGN in Fig.~\ref{fig:SGN} (b) can be obtained by SGN mapping from the original network in Fig.~\ref{fig:SGN} (a). One can see that the edges of different colors in (a) are mapped into the corresponding nodes in (b), which are naturally connected depending on whether they share the same node in the original network.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\linewidth]{subgraph-network.pdf}
\caption{Schematic diagram of SGN construction.}
\label{fig:SGN}
\end{figure}
Formally, given an undirected network $G=(V,E)$ as an original network, where $V$ and $E$ are the node and edge sets, respectively. Let $V_{i}\subseteq V$ and $E_{i}\subseteq E$. Then, $g_{i}=(V_{i},E_{i})$ is a subgraph of $G$. The SGN, denoted by $G_{s} = \mathscr{L}(G)$, is a mapping from $G$ to $G_{s}=(V_{s},E_{s})$, where the node and edge sets are denoted by $V_{s}$=$\left\{g_{i}|i=0,1,2,...,n\right\}$ and $E_{s}\subseteq(V_{s}\times V_{s})$. If $g_{a}\cap g_{b} \neq \emptyset$, i.e., $g_{a}\cap g_{b} \in V$, in the original network, then they are connected in the SGN, i.e.,
$(g_{a}, g_{b})\in E_{s}$. It can be seen that the construction of SGN has three steps: (i) detect subgraphs $\left\{g_i \right\}$ from the original network; (ii) clear and define the connection rules between subgraphs; (iii) build SGN by leveraging the subgraphs.
For simplicity, here for the case of 1st-order SGN, denoted by SGN$^{\textbf{(1)}}$, pairwise linked nodes are chosen as building units, and the adjacent node pairs are connected. In this case, SGN$^{\textbf{(1)}}$ is equivalent to the line graph~\cite{harary1960some}, which reveals the topological interaction between edges of the original network. Fu et al.~\cite{8281007} used this method to map the original network to an SGN, and then used the node centrality in SGN to predict the weights of edges of the original network. As the SGN gradually maps to the higher-order network space, one can observe more abundant feature information. For example, the 2nd-order subgraph network, denoted by SGN$^{\textbf{(2)}}$, is obtained by repeating the mapping process on the SGN$^{\textbf{(1)}}$. The building unit of SGN$^{\textbf{(2)}}$ is a 2-hop structure (open triangle), which maintains the 2nd-order interactive information of the edge structures and can provide more insights about the local structure of a network~\cite{eckmann2002curvature}. To reduce the density of SGN, in the case of SGN$^{\textbf{(2)}}$, two building units are connected when they share the same edge. The latent structural information provided by higher-order SGNs may steadily diminish as the order increases. Therefore, SGN generally works best with the first two orders~\cite{xuan2019subgraph}.
\subsection{Network Sampling Strategies}\label{sec:ssd}
In this paper, we adopt the following four sampling strategies, including random walk, biased walk, link selection, and spanning tree, to design our $\text{S}^2$GN.
\textbf{Random walk.} Random walk ~\cite{pearson1905problem} can be used to obtain the co-occurrence relationship between nodes during network sampling.
A node in a network can be described by the wandering sequence starting from it. The wandering sequence obtained from the node contains both local and higher-order neighbors. When the wandering scope is extended to the graph level, one can peek into the topology of the whole network.
In our model, given a network $G=(V, E)$, the random walk algorithm is described as follows:
\begin{itemize}
\item \emph{Start with an initial node $v^{0}\in V$.}
\item \emph{At step $i$, choose one neighbouring node $u$ $\in$ $\mathcal{N}(v^{i-1})$.}
\item \emph{Let $v^{i}$ $\leftarrow$ $u$ be the next node and get the edge $\widehat{E}$ $\leftarrow$ $\widehat{E}+\left\{(v^{i-1},v^{i})\right\}$.}
\item \emph{Repeat the steps until $|\widehat{E}|= |V|$.}
\end{itemize}
Node $v^{i}$ is generated by the following distribution:
\[P({v^{i}} = x|{v^{i - 1}} = m)=\left\{ \begin{array}{l}
{\textstyle{{{\alpha}} \over N}},if(m,x) \in E\\
0,otherwise
\end{array} \right.\]
where $\alpha$ is the transition probability between nodes $m$ and $x$, and $N$ is the normalizing constant. One can follow the above steps to simulate a random walk and get the final substructure $\widehat{G}=(\widehat{V},\widehat{E})$.
\textbf{Biased walk.} In the field of network science, biased walk ~\cite{azar1992biased} is different from the random walk where the probability of a potential new state is independent of external conditions. When the network is too complex to be analyzed by statistical methods, the biased walk provides an effective method for structural analysis by extracting the symmetry of an undirected network. The concept of the biased walk has attracted considerable attention, especially in the fields of transportation and social networks~\cite{adal2010biased}. Here, we adopt the walking mechanism of Node2Vec~\cite{grover2016node2vec}, where the homogeneity equivalence and structural equivalence of nodes are preserved by integrating the depth-first search and breadth-first search. Specifically, we adopt the 2nd-order random walk with parameters $p$ and $q$, which takes into account the topological distance between the next node and the previous node as well as the connectivity of the current node. Thus, the transition probability $\alpha$ between $v^{i}$ and $v^{i+1}$ is determined by
\[{\alpha _{(v^{i},v^{i+1})}} = {\omega _{pq}}(v^{i-1},v^{i+1}) = \left\{ \begin{array}{l}
\frac{1}{p},{d_{(v^{i-1},v^{i+1})}} = 0\\
1,{d_{(v^{i-1},v^{i+1})}} = 1\\
\frac{1}{q},{d_{(v^{i-1},v^{i+1})}} = 2
\end{array} \right.\]
where $v^{i-1}$, $v^{i}$, and $v^{i+1}$ are the previous, current, and next nodes, respectively, and $d_{(v^{i-1},v^{i+1})} \in (0,1,2)$ indicates the shortest path between $v^{i-1}$ and $v^{i+1}$. Note that $\alpha$ is equal to $\omega_{pq}$ when the network is unweighted. Various substructures of network can be obtained by controlling $p$ and $q$.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{linkselectiongraph-blue.pdf}
\caption{Illustration of the walk procedure in link selection.}
\label{fig:LSpic}
\end{figure}
\textbf{Link selection.} We also propose a new edge-based sampling method, namely link selection. Given a network $G=(V, E)$, we first sample an initial edge $e^0=(v^{0}, v^{1})$, and then randomly select a node of this edge as the source node of the next sampling edge. The nodes of all the sampled edges form the source node pool $V_{pool}$ for the next sampling. The sampling process will not terminate until the stop condition is met. The substructures after this sampling strategy are obtained by a diffuse search from a central edge, which ensures the acquisition of important network structures to a certain extent. As shown in Fig.~\ref{fig:LSpic}, the node pair (1,2) is selected as the initial edge and then we can get the substructure that contains nodes (1,2,3) after one iteration through node "2" and get an expanding substructure that contains nodes (1,2,3,4) after second iteration through another node "1". After several iterations, one can get the final substructure, which contains 7 nodes and 8 edges while the program satisfies the stop condition.
\begin{itemize}
\item \emph{Start with an initial edge $e^{0}=(v^{0}, v^{1})\in E$, and let $V_{pool}=\left\{v^{0}, v^{1}\right\}$, $E_{pool}=\left\{e^{0}\right\}$.}
\item \emph{At step $i$, choose one node $u\in V_{pool}$. }
\item \emph{Let $u^{i}$ $\leftarrow$ $u$ be the next start node and select an edge $(u^{i},u^{i+1}) \notin E_{pool}$.}
\item \emph{Update $V_{pool}$ $\leftarrow$ $V_{pool}+\left\{u^{i+1}\right\}$ and get the edge pool $E_{pool}$ $\leftarrow$ $E_{pool}+\left\{(u^{i},u^{i+1})\right\}$.}
\item \emph{Repeat the above steps until $|E_{pool}|= |V|$.}
\end{itemize}
Note that $(u^{i},u^{i+1})$ has the same transition probability with the random walk, and $V_{pool}$ and $E_{pool}$ are the node and edge sets of the final substructure $\widehat{G}$. This method differs from random walk in that it can search the network on the basis of the current substructure rather than a single node, which can reduce the appearance of a chain structure to a greater extent.
\begin{figure} [htpb]
\centering
\includegraphics[width=1.0\linewidth]{spanningtreegraph-blue.pdf}
\caption{The substructure obtained by spanning tree.}
\label{fig:STpic}
\end{figure}
\textbf{Spanning tree.} A spanning tree~\cite{dey2019new} is a minimally connected substructure that contains all nodes in the graph, as shown in Fig.~\ref{fig:STpic}. Different spanning trees can be obtained by traversing from different nodes. Here we randomly select a node as the initial node. The maximum and minimum spanning trees are unified without considering the edge weights. In this section, we use the typical Kruskal algorithm~\cite{najman2013playing} to generate spanning trees and the weight values of edges are all set to 1.
\subsection{Framework for Constructing $\text{S}^2$GN}\label{sec:ssgn}
Most real-world networks have large scale and complex structure. Typically, SGN could be even larger and denser, making the follow-up network algorithms less efficient. It may also introduce extra noisy structural information, disturbing the network algorithms to a certain extent. In view of this, we focus on optimizing the SGN model and propose a framework for constructing a sampling subgraph network (S$^2$GN) by integrating different network sampling methods.
The pseudocodes of constructing S$^2$GN and sampling substructures are given in Algorithms~\ref{alg:1} and \ref{alg:2}, respectively. In Algorithms~\ref{alg:1}, GetMaxSubstructure(·) is to obtain the maximally connected substructure of original network if it is not connected; NodeRanking(·) is to rank the input nodes; SGNAlgorithms(·) is to construct SGNs. GetNextEdgeWithStrategy(·) in Algorithms~\ref{alg:2} is to get the next edge according to a given sampling strategy.
\begin{figure*}[!t]
\centering
\includegraphics[width=1\linewidth]{ssgn-framework-new1.pdf}
\caption{The overall framework of the S$^2$GN algorithm for network structure feature fusion.}
\label{fig:SSGN}
\end{figure*}
\begin{algorithm}[!t]
\caption{\textbf{Construction of S$^2$GN.}}
\LinesNumbered
\label{alg:1}
\KwIn{A network $G$($V$,$E$) with node set $V$ and link set $E\subseteq(V\times{V})$\;
Sampling strategy $f_s(\cdot)$\;
The order of SGN $h$.}
\KwOut{S$^2$GN, denoted by $G_s$($V_s$,$E_s$).}
Initialize a temporary object $G_s$ = $G$\;
\While {$h$}
{ \If {the $G_s$ is not full-connected}
{
GetMaxSubstructure($G_s$)\;
Initial node $u$ = NodeRanking($V_s$)\;
Get sampling substructure $\widehat{G_s}$ through executing Algorithm \ref{alg:2}\;
$G_{sgn}$ = SGNAlgorithms($\widehat{G_s}$)\;
$G_s$ $\leftarrow$ Relabeled($G_{sgn}$)\;
}
\Else {Repeat 5-8\;}
$h = h - 1$\;
}
\Return $G_s$($V_s$,$E_s$)
\end{algorithm}
\begin{algorithm}[!t]
\caption{\textbf{Sampling substructure.}}
\LinesNumbered
\label{alg:2}
\KwIn{A network $G$($V$,$E$)\;
Source node $u$\;
Sampling walks $l$.}
\KwOut{Sampling substructure, denoted by $\widehat{G_s}$=$g$($\widehat{v}$,$\widehat{e}$).}
Let $v_0$=$u$, initial $walk_v$ to [$v_0$], $walk_e$ to $\emptyset$\;
Select first edge $e_1$ with a given probability of sampling strategy\;
Append the $v_1$ = $dst$($e_1$) to $walk_v$, $e_1$ to $walk_e$\;
\For {$i=2$ to $l-1$}
{ $cur_v$ = $walk_v$[-1], $cur_e$ = $walk_e$[-1]\;
$e_i$ = GetNextEdgeWithStrategy($cur_v$, $cur_e$)\;
Append $e_i$ to $walk_e$, $v_i$=$dst(e_i)$ to $walk_v$\;}
$\widehat{v}$ = $walk_v$, $\widehat{e}$ = $walk_e$\;
\Return $\widehat{G_s}$=$g$($\widehat{v}$,$\widehat{e}$)
\end{algorithm}
In general, S$^2$GN can be constructed in three steps: source node selection, sampling substructure and S$^2$GN construction, which are introduced in the following.
\begin{itemize}
\item \textbf{\textbf{Source node selection}}: There are many ways to choose the initial node: (i) Randomly select a node as the source node; (ii) Select an initial node according to its importance measured by closeness centrality~\cite{okamoto2008ranking}, K-shell~\cite{lu2016vital}, PageRank~\cite{langville2004deeper} or others. In this paper, we use the K-shell method in order to capture the key structure more likely.
\item \textbf{\textbf{Sampling substructure}}: After the initial source node is determined, a substructure can be obtained by conducting a certain sampling strategy to extract the main context of the current network. According to different sampling strategies, diverse sampling substructures can be generated, reflecting the different aspects of the original network and further benefiting the subsequent network algorithms.
\item \textbf{\textbf{S$^2$GN construction}}: Based on the sampling substructure, we use SGN model to construct S$^2$GN. Note that network sampling and SGN are adopted iteratively so as to get the S$^2$GNs of higher orders. This method can control the size of S$^2$GNs and meanwhile increase their diversity. Therefore, compared with SGN, the S$^2$GN could further enhance both efficiency and effectiveness of the subsequent network algorithms.
\end{itemize}
\begin{figure*}[!t]
\centering
\includegraphics[width=1\linewidth]{12_order_new.pdf}
\caption{Visualization of 1st-order and 2nd-order S$^2$GNs using four network sampling strategies on positive and negative samples from the MUTAG dataset.}
\label{fig:NetV}
\end{figure*}
Now, we use various feature extraction methods to get structural features from S$^2$GNs of different orders, which are first fused and then used to establish the graph classification models. The overall framework of S$^2$GN construction for structural feature space expansion is shown in Fig.~\ref{fig:SSGN}. Note that, generally, information fusion tries to integrate information from multiple aspects to improve algorithm performance, which has a wide range of applications in practice. For instance, in speech recognition, the visual features of the lip motion are fused with the speech signal features to predict the words expressed~\cite{zhou2019modality}. In image recognition, Xuan et al.~\cite{8219720} developed a multistream convolutional neural network to automatically merge the features of multi-view pearl images, so as to improve the accuracy of pearl classification. In this paper, we use different sampling strategies to capture the structural features from different aspects. As an example, we visualize different 1st-order and 2nd-order S$^2$GNs generated by the four network sampling strategies on positive and negative samples from the MUTAG dataset, as shown in Fig.~\ref{fig:NetV}. It can be seen that the S$^2$GNs generated by different sampling strategies have quite different structures, and the structural difference between the positive and negative samples may be enlarged in S$^2$GNs. Therefore, it can be expected that the fusion of these diverse S$^2$GNs could improve the performance of graph classification.
\section{Experiments}\label{sec:experiments}
Now, we compare S$^2$GN and SGN models on their abilities to enhance graph classification based on four feature extraction methods. We first introduce the datasets, followed by the feature extraction methods and the parameter setting. After that, we show the experimental results with discussion.
\subsection{Datasets}
We test our S$^2$GN method on eight real-world network datasets, as introduced in the following. IMDB-BINARY is about social networks, while the others are about bio- and chemo-informatics networks. The basic statistics of these datasets are presented in Table~\ref{data}.
\begin{table}[!h]\renewcommand{\arraystretch}{1.2}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\caption{Basic statistics of eight datasets. $N_G$ is the number of graphs, \#$C_{max}$ is the number of graphs belonging to the largest class, $N_C$ is the number of classes, and \#Nodes and \#Edges are the average numbers of nodes and edges, respectively, of the graphs in the dataset.}
\centering
\begin{centering}
\begin{tabular}{l|ccccc}
\hline\hline
Dataset& $N_G$ & \#$C_{max}$ & $N_C$ & \#Nodes & \#Edges \tabularnewline
\hline
MUTAG& \tabincell{c}{188} &125 &2 &18 &20\tabularnewline
PTC& \tabincell{c}{344} &192 &2 &14 &14\tabularnewline
PROTEINS& \tabincell{c}{1113} &663 &2 &39 &73\tabularnewline
ENZYMES& \tabincell{c}{600} & 100 &6 &32 &63\tabularnewline
NCI1& \tabincell{c}{4110} &2057 &2 &30 &32\tabularnewline
NCI109& \tabincell{c}{4127} &2079 &2 &30 &32\tabularnewline
IMDB-BINARY& \tabincell{c}{1000} &500 &2 &20 &193\tabularnewline
D$\&$D & \tabincell{c}{1178} &691 &2 &284 &716\tabularnewline
\hline\hline
\end{tabular}
\end{centering}
\label{data}
\end{table}
\begin{itemize}
\item \textsc{\textit{MUTAG}}~\cite{debnath1991structure} contains 188 mutagenic aromatic and heteroaromatic compounds, with nodes and edges representing atoms and the chemical bonds between them, respectively. They are labeled according to whether there is a mutagenic effect on a special bacteria.
\item \textsc{\textit{PTC}}~\cite{toivonen2003statistical} includes 344 chemical compound graphs, with nodes and edges representing atoms and the chemical bonds between them, respectively. Their labels are determined by their carcinogenicity for rats.
\item \textsc{\textit{PROTEINS}}~\cite{borgwardt2005protein} comprises of 1113 graphs. The nodes are Secondary Structure Elements (SSEs) and the edges are neighbors in the amino-acid sequence or in the 3D space. These graphs represent either enzyme or non-enzyme proteins.
\item \textsc{\textit{ENZYMES}}~\cite{nr} contains 600 protein tertiary structures, and each enzyme belongs to one of the 6 EC top-level classes.
\item \textsc{\textit{NCI1}} \& \textsc{\textit{NCI109}}~\cite{wale2008comparison} comprise of 4110 and 4127 graphs, respectively. The nodes and edges represent atoms and chemical bonds between them, respectively. They are two balanced subsets of the datasets of chemical compounds screened for the activities against non-small cell lung cancer and ovarian cancer cell lines, respectively. The positive and negative samples are distinguished according to whether they are effective against cancer cells.
\item \textsc{\textit{IMDB-BINARY}}~\cite{nguyen2018learning} is about movie collaboration including 1000 graphs, which is collected from IMDB and contains lots of information about different movies. Each graph is an ego-network, where nodes represent actors or actresses and edges indicate whether they appear in the same movie. Each graph is categorized into one of the two genres (Action and Romance).
\item \textsc{\textit{D$\&$D}}~\cite{dobson2003distinguishing} contains 1178 graphs of protein structures. A node represents an amino acid and edges are constructed if the distance between two nodes is less than 6 $\mathring{A}$. A label denotes whether a protein is an enzyme or non-enzyme.
\end{itemize}
\subsection{Feature Extraction Methods}\label{sec:fem}
We adopt four typical methods to generate graph representation, namely manual attributes, Graph2Vec, DeepKernel, and CapsGNN, which are introduced in the following.
\begin{itemize}
\item \textit{Attributes}: Here, we use the same 11 manual attributes as those introduced in \cite{xuan2019subgraph}, including the number of nodes, the number of edges, average degree, network density, average clustering coefficient, the percentage of leaf nodes, the largest eigenvalue of the adjacency matrix, average betweenness centrality, average closeness centrality, and average eigenvector centrality.
\item \textit{Graph2Vec}~\cite{narayanan2017graph2vec}: This is the first unsupervised embedding approach for an entire network, which is based on the extending word-and-document embedding techniques that has shown great advantages in natural language processing (NLP).
\item \textit{DeepKernel}~\cite{yanardag2015deep}: This method provides a unified framework that leverages the dependency information of sub-structures by learning latent representations. The sub-structure similarity matrix, $\mathcal{M}$, is calculated by the matrix $\mathcal{V}$ with each column representing a sub-structure vector. Denote by $\mathcal{P}$ the matrix with each column representing a sub-structure frequency vector. According to the definition of kernel: $\mathcal{K} = \mathcal{P}\mathcal{M}\mathcal{P}^\mathrm{T} = \mathcal{P}\mathcal{V}\mathcal{V}^\mathrm{T}\mathcal{P}^\mathrm{T}=\mathcal{H}\mathcal{H}^\mathrm{T}$, one can use the columns in the matrix $\mathcal{H}=\mathcal{P}\mathcal{V}$ as the inputs to the classifier.
\item \textit{CapsGNN}~\cite{xinyi2018capsule}: This method was inspired by CapsNet~\cite{sabour2018matrix}, which adopts the concept of capsules to overcome the weakness of existing GNN-based graph embedding algorithms. In particular, CapsGNN extracts node features in the form of capsules and utilizes the routing mechanism to capture important information at the graph level. The model generates multiple embeddings for each graph so as to capture graph properties from different aspects.
\end{itemize}
\subsection{Parameter Setting}\label{sec:ps}
For source node selection, we choose the node of the largest K-shell~\cite{lu2016vital} as the source node for random walk (RW) and biased walk (BW), and choose the edge of the largest betweenness centrality as the source edge for link selection (LS). We randomly pick up a node as the source node for the spanning tree (ST) to increase the diversity of S$^2$GN, since the sampled subnetworks will be quite similar if we fix the source node for this method. Moreover, we set the two parameters of BW as $p=4$ and $q=1$.
\begin{table*}[!t]
\caption{Classification results measured by $F1$-$Score$ on eight datasets by using different feature extraction methods.}
\centering
\renewcommand\arraystretch{1.2}
\setlength{\tabcolsep}{0.7mm}{
\begin{tabular}{c|cccccccc|c}
\hline\hline
\textbf{Algorithm} & \multicolumn{ 8}{c}{\textbf{Classification results} ($F1$-$Score$, \%)} \\
\hline
\textbf{Attributes} & MUTAG & PTC & PROTEINS & ENZYMES & NCI1 & NCI109 & IMDB-BINARY & D$\&$D & Avg. \\
\hline
Original & $86.58\pm{3.61}$ &$63.52\pm{4.55}$ &$78.30\pm{2.49}$ &$43.37\pm{2.29}$ & $67.48\pm{0.87}$ & $67.34\pm{1.25}$ &$73.00\pm{3.68}$ & $75.85\pm{1.61}$ & 69.43\\
\hline
SGN & $91.58\pm{4.21}$ & $67.94\pm{6.36}$ & $79.46\pm{2.96}$ & $50.22\pm{2.91}$ & $69.84\pm{1.59}$ & $69.73\pm{1.97}$ & ${\bf77.65\pm{4.50}}$ & $76.65\pm{1.59}$ & 72.88\\
$RIMP$-SGN &5.78\% &6.96\% &1.48\% &15.79\% &3.50\% &3.55\% &6.37\% &1.05\% &4.97\%\\
\hline
S$^2$GN-RW & $90.53\pm{2.11}$ & $66.71\pm{1.62}$ & $77.76\pm{1.52}$ & $52.17\pm{2.36}$& $74.82\pm{0.69}$ & $73.96\pm{0.84}$ & $71.85\pm{2.74}$ & $77.37\pm{2.91}$ & 73.15 \\
S$^2$GN-BW & $93.94\pm{2.37}$ & $69.11\pm{2.94}$ & ${\bf79.83\pm{1.46}}$ & $53.37\pm{2.78}$& $75.47\pm{0.98}$ & $73.99\pm{1.04}$ & $76.05\pm{1.29}$ & ${\bf77.75\pm{1.68}}$ & 74.94 \\
S$^2$GN-LS & $89.21\pm{2.48}$ & $66.18\pm{2.55}$ & $78.57\pm{1.44}$ &$49.50\pm{2.14}$& $75.85\pm{1.03}$ & $74.65\pm{0.62}$ & $71.80\pm{2.53}$ & $76.91\pm{1.94}$ & 72.83 \\
S$^2$GN-ST & $90.79\pm{2.12}$ & $70.44\pm{2.75}$ & $76.28\pm{1.89}$ & $45.33\pm{1.29}$& $72.25\pm{1.08}$ & $73.26\pm{0.76}$ & {\bf $77.60\pm{1.50}$} & $76.90\pm{2.47}$ & 72.81 \\
\hline
S$^2$GN-Fusion & ${\bf94.74\pm{1.84}}$ & ${\bf72.06\pm{3.29}}$ & $79.14\pm{0.84}$ & ${\bf55.25\pm{1.90}}$& ${\bf76.03\pm{1.32}}$ & ${\bf74.89\pm{1.18}}$ & $76.97\pm{1.21}$ & $77.03\pm{2.46}$ & ${\bf75.76}$ \\
$RIMP$-Fusion &9.42\% &13.44\% &1.07\% & 27.39\% & 12.67\% & 11.21\% & 5.44\% & 1.56\% & 9.12\% \\
\hline\hline
\textbf{Graph2Vec} & MUTAG & PTC & PROTEINS & ENZYMES & NCI1 & NCI109 & IMDB-BINARY & D$\&$D & Avg. \\
\hline
Original &$83.15\pm{9.25}$ &$60.17\pm{6.86}$ &$73.30\pm{2.05}$ &$45.17\pm{2.73}$ &$73.22\pm{1.81}$ & $74.26\pm{1.47}$ & $62.47\pm{3.99}$ & $70.25\pm{2.18}$ & 67.75\\
\hline
SGN & $86.84\pm{5.70}$ & $63.24\pm{6.70}$ & $74.44\pm{3.09}$ & $48.73\pm{2.56}$ & $76.64\pm{3.21}$ & $74.86\pm{2.76}$ & $70.65\pm{5.55}$ & $80.42\pm{3.06}$ & 70.73 \\
$RIMP$-SGN &4.44\% & 5.10\% & 1.56\% & 7.88\% & 4.67\% & 0.81\% & 13.09\% & 14.48\% & 4.39\% \\
\hline
S$^2$GN-RW & $80.26\pm{2.69}$ & $61.47\pm{2.06}$ & $76.37\pm{1.12}$ & $48.67\pm{2.53}$& $76.88\pm{1.35}$ & $74.39\pm{1.40}$ & $68.35\pm{1.57}$ & $81.86\pm{1.80}$ & 71.03\\
S$^2$GN-BW & ${\bf86.84\pm{3.07}}$ & ${\bf64.71\pm{2.85}}$ & ${\bf77.13\pm{1.09}}$ & $52.33\pm{2.30}$ & $77.39\pm{1.12}$ & $75.69\pm{1.46}$ & $71.64\pm{2.00}$ & $82.12\pm{2.22}$ & $73.48$\\
S$^2$GN-LS & $81.05\pm{2.57}$ & $62.35\pm{2.88}$ & $76.91\pm{2.21}$ & $47.68\pm{1.73}$ & ${\bf79.18\pm{1.71}}$ & ${\bf77.42\pm{1.13}}$ & $67.25\pm{2.16}$ & $81.77\pm{1.98}$ & 71.70\\
S$^2$GN-ST & $81.84\pm{2.99}$ & $63.97\pm{2.39}$ & $75.20\pm{2.15}$ & $49.87\pm{2.91}$ & $76.30\pm{1.21}$ & $72.95\pm{0.89}$ & $72.49\pm{2.11}$ & $74.92\pm{2.89}$ & 70.94 \\
\hline
S$^2$GN-Fusion & $81.73\pm{3.37}$ & $64.38\pm{2.42}$ & $75.10\pm{0.89}$ & ${\bf54.78\pm{2.29}}$ & $76.91\pm{0.72}$ & $75.72\pm{1.31}$ & ${\bf76.43\pm{2.17}}$ & ${\bf82.75\pm{2.79}}$ & ${\bf73.48}$\\
$RIMP$-Fusion &-1.71\% & 7.00\% & 2.46\% & 21.28\% & 5.04\% & 1.97\% & 22.35\% & 17.79\% & 8.46\% \\
\hline\hline
\textbf{DeepKernel} & MUTAG & PTC & PROTEINS & ENZYMES & NCI1 & NCI109 & IMDB-BINARY & D$\&$D & Avg. \\
\hline
Original &$82.95\pm{2.68}$ &$59.04\pm{1.09}$ &$73.30\pm{0.82}$ &$45.04\pm{3.73}$ &$67.06\pm{1.91}$ &$67.04\pm{1.36}$ &$67.50\pm{2.45}$ & $75.97\pm{1.91}$ & 67.24\\
\hline
SGN & $93.68\pm{5.15}$ & $65.88\pm{5.05}$ & $76.78\pm{2.41}$ & $45.93\pm{3.75}$ & $70.26\pm{1.24}$ & $71.06\pm{1.61}$ & $75.70\pm{1.55}$ & $77.84\pm{2.08}$ & 72.14\\
$RIMP$-SGN &12.94\% & 11.59\% & 4.75\% & 1.98\% & 4.77\% & 6.00\% & 12.15\% & 2.46\% & 7.29\%\\
\hline
S$^2$GN-RW & $93.68\pm{5.66}$ & $61.76\pm{3.77}$ & $75.80\pm{4.21}$ & $43.32\pm{3.64}$ & $69.15\pm{1.63}$ & $69.06\pm{1.70}$ & $72.30\pm{2.68}$ & $83.47\pm{1.00}$ & 71.07\\
S$^2$GN-BW & $94.00\pm{5.43}$ & $67.35\pm{4.48}$ & $76.69\pm{2.97}$ & $47.75\pm{2.68}$ & $71.51\pm{1.38}$ & $69.83\pm{2.05}$ & $74.10\pm{3.33}$ & $81.57\pm{1.11}$ & 72.85\\
S$^2$GN-LS & $93.68\pm{4.59}$ & $66.18\pm{4.21}$ & $76.16\pm{1.92}$ & $50.28\pm{3.04}$ & ${\bf71.55\pm{1.15}}$ & $70.19\pm{2.26}$ & $75.80\pm{3.43}$ & $\bf{83.98\pm{1.77}}$ & 73.48\\
S$^2$GN-ST & $88.95\pm{3.68}$ & $65.29\pm{4.59}$ & $74.73\pm{4.54}$ & $48.02\pm{3.52}$ & $70.77\pm{1.20}$ & ${\bf71.04\pm{1.03}}$ & $75.90\pm{2.07}$ & $78.94\pm{1.38}$ & 71.71\\
\hline
S$^2$GN-Fusion & ${\bf94.73\pm{4.07}}$ & ${\bf70.88\pm{4.25}}$ & ${\bf77.14\pm{2.97}}$ & ${\bf52.21\pm{2.24}}$ & $71.06\pm{1.01}$ & $70.48\pm{1.22}$ & ${\bf76.50\pm{3.75}}$ & $83.77\pm{1.87}$ & ${\bf74.60}$\\
$RIMP$-Fusion &14.20\% & 20.05\% & 5.24\% & 15.92\% & 5.96\% & 5.13\% & 13.33\% & 10.27\% & 10.94\%\\
\hline\hline
\textbf{CapsGNN} & MUTAG & PTC & PROTEINS & ENZYMES & NCI1 & NCI109 & IMDB-BINARY & D$\&$D & Avg. \\
\hline
Original &$86.32\pm{7.52}$& $62.06\pm{4.25}$& $75.89\pm{3.51}$ &$49.78\pm{3.02}$ & $78.30\pm{1.80}$& $72.99\pm{2.15}$& $72.71\pm{4.36}$ & $67.75\pm{2.57}$ & 70.73\\
\hline
SGN & $89.47\pm{7.44}$ & $64.12\pm{3.67}$ & $76.34\pm{4.13}$ & $50.04\pm{2.70}$ & $78.61\pm{1.87}$ & $73.72\pm{2.39}$ & $76.47\pm{5.74}$ & $68.71\pm{1.91}$ & 72.19\\
$RIMP$-SGN &3.65\% & 3.32\% & 0.59\% & 0.52\% & 0.40\% & 1.00\% & 5.17\% & 1.42\% & 2.06\%\\
\hline
S$^2$GN-RW & $88.70\pm{4.59}$ & $77.81\pm{4.96}$ & $84.73\pm{2.09}$ & $51.33\pm{1.14}$ & $74.23\pm{1.40}$ & $75.16\pm{1.39}$ & $92.50\pm{3.15}$ & $78.05\pm{1.91}$ & 77.81\\
S$^2$GN-BW & $92.63\pm{4.82}$ & $81.91\pm{5.45}$ & $84.10\pm{3.72}$ & $52.77\pm{2.11}$ & ${\bf78.83\pm{2.35}}$ & $75.25\pm{1.69}$ & $93.35\pm{1.12}$ & ${\bf78.66\pm{2.32}}$ & 79.69\\
S$^2$GN-LS & $90.53\pm{4.59}$ & $79.11\pm{4.16}$ & $84.28\pm{1.96}$ & $52.17\pm{1.23}$ & $76.57\pm{1.26}$ & $75.43\pm{1.46}$ & $93.92\pm{1.75}$ & $77.03\pm{1.52}$ & 78.63\\
S$^2$GN-ST & $89.21\pm{5.73}$ & $78.67\pm{5.06}$ & $84.03\pm{2.58}$ & $52.56\pm{1.18}$ & $76.52\pm{1.42}$ & $75.16\pm{1.60}$ & $94.20\pm{1.26}$ & $72.31\pm{2.73}$ & 77.83 \\
\hline
S$^2$GN-Fusion & ${\bf93.15\pm{4.11}}$ & ${\bf84.12\pm{6.47}}$ & ${\bf85.18\pm{1.84}}$ & ${\bf56.08\pm{3.15}}$ & $78.13\pm{2.27}$ & ${\bf78.23\pm{1.05}}$ & ${\bf95.10\pm{2.30}}$ & $77.85\pm{1.95}$ & ${\bf80.98}$\\
$RIMP$-Fusion &7.91\% & 35.55\% & 12.24\% & 12.66\% & -0.22\% & 7.18\% & 30.79\% & 14.90\% & 14.49\% \\
\hline\hline
\end{tabular}}
\label{Results}
\end{table*}
\begin{figure*}[!t]
\centering
\includegraphics[width=1\linewidth]{plot-new6-2-9-8-F1Score-2021-02-09-eps-converted-to.pdf}
\caption{Average $F1$-$Score$ as functions of the training set size (represented by the fraction of samples in the training set), for various feature extraction methods on different datasets, based on RW, BW, LS, ST and Fusion, respectively.}
\label{fig:plot}
\end{figure*}
In this study, for \emph{Graph2Vec}, the embedding dimension is adopted according to~\cite{narayanan2017graph2vec}. Since the embedding dimension is predominant for learning performances, a commonly-used value of 1024 is adopted. The other parameters are set to default values: the learning rate is set to 0.5, the batch size is set to 512 and the number of epochs is set to 1000. For \emph{DeepKernel}, according to ~\cite{yanardag2015deep}, the Weisfelier-Lehman subtree kernel is used to build the corpus and its height is set to 2. Furthermore, the embedding dimension is set to 10, the window size is set to 5 and skip-gram is used for the word2vec model. We adopt the default parameters for \emph{CapsGNN} and flatten the multiple embeddings of each graph as the input.
Without loss of generality, the well-known Random Forest is chosen as the classification model. Meanwhile, for each feature extraction method, the feature space is first expanded by using S$^2$GNs, and then the dimension of the feature vectors is reduced to the same value as that of the feature vector obtained from the original network using PCA in the experiments, for a fair comparison. Each dataset is randomly split into 8 folds for training and 2 fold for testing. Here, the $F1$-$Score$ is adopted as the metric to evaluate the classification performance:
\begin{equation}
F_{1}= \frac{2PR}{P + R}\,,
\end{equation}
where $P$ and $R$ are the precision and recall, respectively. In order to diminish the random effect of the fold assignment to some extent, the experiment is repeated 100 times and then the average $F_1$-$Score$ and its standard deviation are reported.
We further define the relative improvement rate (RIMP) of SGN or S$^2$GN model as
\begin{equation}
RIMP= (F1_{model}-F1_{ori})/F1_{ori}\,
\end{equation}
where $F1_{ori}$ and $F1_{model}$ refer to the $F1$-$Score$ of the graph classification algorithm without and with the SGN model (or S$^2$GN-Fusion model), respectively.
\subsection{Experimental Results}\label{sec:er}
We use the four network sampling strategies to generate sampling substructures, and further construct the corresponding 1st-order and 2nd-order S$^2$GNs, denoted by S$^2$GN-RW, S$^2$GN-BW, S$^2$GN-LS, and S$^2$GN-ST, respectively\footnote{It has been proven that the graph classification models can be significantly enhanced by appropriately using the structural information of the SGNs in the first two orders, while such gain will be reduced soon as more SGNs of higher orders are integrated~\cite{xuan2019subgraph}. This is why we only use the S$^2$GNs of the first two orders here.}. After that, we adopt the four feature extraction methods, namely manual attributes, Graph2Vec, DeepKernel, and CapsGNN, to get structural feature vectors. For each feature extraction method, we fuse the vectors generated from the different S$^2$GNs to a single vector. Finally, this vector is fed into the Random Forest model to produce the classification result. Note that we also produce the results for a single sampling strategy for a more comprehensive comparison. Here, a ten-fold cross-validation method is used to calculate $F1$-$Score$ of graph classification. To enrich the sampling structures and reduce the probability of sampling repetition, 10 sampling averaging processes were carried out for each sampling strategy.
\subsubsection{Enhancement on classification performance}
The experimental results are shown in Table~\ref{Results}, where one can see that the four S$^2$GN models based on a single sampling strategy, i.e., S$^2$GN-RW, S$^2$GN-BW, S$^2$GN-LS, and S$^2$GN-ST, are comparable with the SGN model, which all produce similar classification results under different datasets and feature extraction methods. Interestingly, S$^2$GN-BW outperforms SGN in enhancing the classification models based on the four feature extraction methods in most cases, leading to a relative improvement of 4.52\% on average. Such results are consistent with the experience that Node2Vec is a powerful method to capture the structural properties of a network. Moreover, since different S$^2$GNs generated by different sampling strategies can capture the different aspects of a network, as visualized in Fig.~\ref{fig:NetV}, one may expect that the fusion of these S$^2$GNs can produce even better classification results. Indeed, we find that the fusion of $\text{S}^2$GNs increases the performance of the original graph classification algorithms in 30 out of 32 cases, with a relative improvement of 10.75\% on average (much better than 4.68\% by SGN). The value increases to 14.49\% (much better than 2.06\% by SGN) when only CapsGNN is considered. This result is quite impressive, since CapsGNN, together with $\text{S}^2$GN, achieves the state-of-the-art performance on PROTEINS and IMDB-BINARY datasets.
To address the robustness of our S$^2$GN model against the size variation of the training set, the $F_1$-$Score$ is calculated by using various sizes of training sets (from 10 to 90 percent, within a 20 percent interval). For each size, the training and test sets are randomly divided, which is repeated 100 times with the average result recorded. The results are shown in Fig.~\ref{fig:plot} for various feature extraction methods on eight datasets. It can be seen that still the curves of S$^2$GN-Fusion are relatively higher than those of S$^2$GNs generated by a single sampling strategy in most cases, indicating that the superiority of S$^2$GN-Fusion is robust in enhancing graph classification algorithms. In particular, such superiority seems much more significant when enhancing CapsGNN, which is interesting and may indicate that the potential of S$^2$GN-Fusion could be exploited further by connecting a better embedding method or end-to-end graph neural network, and meanwhile there could be much room for further improvement for graph classification.
\subsubsection{Reduction of time complexity}
Note that one important motivation to introduce sampling strategies into SGN is to control the network size so as to improve the efficiency of the network algorithms based upon them. Therefore, here to address the computational complexity of our method, we record the average computational time of SGN and S$^2$GN generated by the four sampling strategies on the eight datasets, namely MUTAG, PTC, PROTEINS, ENZYMES, NCI1, NCI109, IMDB-BINARY, and D$\&$D. The results are presented in Table~\ref{Complexity}, where one can see that, overall, the computational time of S$^2$GN is much less than that of SGN for each sampling strategy on each dataset, decreasing from hundreds of seconds to less than 19 seconds. In fact, the computational time of S$^2$GNs generated by different sampling strategies is comparable to each other. Considering that S$^2$GN-Fusion method needs to generate all the four S$^2$GNs, its computational time is close to the sum of individual ones, which is still less than 25 seconds. Such results suggest that, by comparing with SGN, our S$^2$GN model can indeed largely increase the efficiency of the network algorithms.
In fact, we can estimate the time complexity of our model in theory. For random walk, it is a computationally efficient sampling method, which only requires $\mathcal{O}(|E|)$ space complexity to store the neighbors of each node in the graph. As for the time complexity, by imposing graph connectivity in the sample generation process, random walk provides a convenient mechanism to increase the effective sampling rate by reusing samples across different source nodes. For biased walk, we adopt the 2nd random walk mechanism of Node2Vec, where each step of random walk is based on the transition probability $\alpha$ which can be precomputed, so the time consuming of each step using alias sampling is $\mathcal{O}(1)$. Link selection broadens the scope of the start node at each step in the random walk process, thereby accelerating the time to reach the stop condition. Kruskal algorithm to generate spanning trees is a greedy algorithm, which has $\mathcal{O}(|E|log(|E|))$ time complexity. We know that the computational complexity of SGN$^{(1)}$ is $\mathcal{O}(|E|^2)$ and that of constructing SGN$^{(2)}$ is $\mathcal{O}(|E|^4)$. Our S$^2$GN model constrains the expansion of the network scale and reduces the cost of constructing SGNs to the fixed $\mathcal{O}(|E|^2)$. Thus, the time computational complexity $\mathcal{T}$ of our S$^2$GN model is $\mathcal{O}(|E|+|E|^2)\leq \mathcal{T}\leq \mathcal{O}(|E|log|E|+|E|^2|)$ according to the different sampling strategies, which is much lower than that of SGN.
\begin{table}[!t]
\caption{Average computational time to establish SGN and S$^2$GNs by the four sampling strategies on the eight datasets.}
\centering
\renewcommand\arraystretch{1.1}
\begin{tabular}{lccccc}
\hline\hline
\multirow{2}{*}{Time (Seconds)} & \multirow{2}{*}{SGN} & \multicolumn{4}{c}{S$^2$GN} \\ \cline{3-6}
& & RW & BW & LS & ST \\
\hline
MUTAG & $1.58\times{10^2}$ & 0.677 & 0.252 & 0.600 & 0.090 \\
PTC & $1.93\times{10^3}$ & 1.216 & 0.804 & 1.170 & 0.607 \\
PROTEINS & $3.20\times{10^3}$ & 1.192 & 1.161 & 2.018 & 1.625 \\
ENZYMES & $3.97\times{10^3}$ & 1.284 & 1.230 & 2.106 & 1.598 \\
NCI1 & $1.75\times{10^2}$ & 2.670 & 2.099 & 2.484 & 1.746 \\
NCI109 & $1.75\times{10^2}$ & 2.682 & 2.114 & 2.495 & 1.749 \\
IMDB-BINARY & $1.11\times{10^4}$ & 1.478 & 1.580 & 1.256 & 1.106\\
D$\&$D & $7.90\times{10^2}$ & 2.701 & 3.162 & 18.32 & 0.805\\\hline
\end{tabular}
\label{Complexity}
\end{table}
\subsubsection{Visualization}
\begin{figure}[!t]
\centering
\includegraphics[width=1\linewidth]{figure_iar_sp-cut-eps-converted-to1.pdf}
\caption{The t-SNE visualization of structural features using CapsGNN without (left) and with (right) S$^2$GN-ST. The same color of points represent the same class of graphs in IMDB-BINARY dataset.}
\label{fig:TSNE}
\end{figure}
As a simple case study, we visualize the results of classification on IMDB-BINARY dataset based on CapsGNN method to verify the effectiveness of our S$^2$GN model. Here, we choose S$^2$GN-ST to visualize since this is the best S$^2$GN generated by the single sampling strategy that enhances the classification performance of CapsGNN most. As shown in Fig.~\ref{fig:TSNE}, the structural features are located in different places by utilizing t-SNE. The left shows the original classification result using CapsGNN without S$^2$GN-ST, while the right depicts the optimized distribution of the same dataset using CapsGNN with S$^2$GN-ST. One can see that the graphs in IMDB-BINARY dataset can indeed be distinguished by the original features of CapsGNN, but it appears that the distinction of graphs could become more explicit after hierarchical representation through network sampling and SGN mapping, demonstrating the effectiveness of our S$^2$GN model.
\section{Conclusions}\label{sec:Con}
In this paper, we present a novel sampling subgraph network (S$^2$GN) model as well as a hierarchical feature fusion framework for graph classification by introducing network sampling strategies into the SGN model. Compared with the latter, the S$^2$GNs are of higher diversity and controllable scale, and thus benefit the network feature extraction methods to capture more various aspects of the network structure with higher efficiency.
We use different sampling strategies, namely random walk (RW), biased walk (BW), link selection (LS), and spanning tree (ST), to generate the corresponding sampling subgraph networks S$^2$GN-RW, S$^2$GN-BW, S$^2$GN-LS, and S$^2$GN-ST, respectively. The experimental results show that, compared with SGN, S$^2$GN has much lower time complexity, which was reduced by almost two orders of magnitude, and meanwhile they have comparable effects on graph classification. In fact, the network algorithms based on S$^2$GN-BW behave even better than those based on SGN, although each sampling subnetwork is only a part of the original network. More interestingly, when the features of all the four S$^2$GNs are fused and then fed into graph classification models, the classification performance can be significantly enhanced. In particular, when CapsGNN is used to extract the features of these S$^2$GNs, we can achieve the-state-of-the-art results on the PROTEINS and IMDB-BINARY datasets.
In the future, we will try more sampling strategies and then integrate them with SGN to generate more diverse S$^2$GNs; we will also apply our framework to more tasks beyond graph classification, such as link prediction, node classification, etc.
\section{Acknowledgments}\label{sec:ack}
The authors would like to thank all the members in the IVSN Research Group, Zhejiang University of Technology for the valuable discussions about the ideas and technical details presented in this paper. This work was partially supported by the National Natural Science Foundation of China under Grant 61973273, by the Zhejiang Provincial Natural Science Foundation of China under Grant LR19F030001, and by the Hong Kong Research Grants Council under the GRF Grant CityU11200317.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
This article deals with forecasting and/or clustering of times series. We focus on applications where one knows that times series at hand have some intrinsic structure, such as entry-wise \textit{nonnegativity}. In that case one can exploit Nonnegative Matrix Factorization (NMF) approaches which have been introduced by \cite{paatero1994} for spectral unmixing problems in analytical chemistry and popularized by \cite{lee1999}. For further details we refer the interested reader to the surveys \cite{wang2013,gillis2015,gillis2017} and references therein. NMF has been widely applied in many different contexts: document analysis \cite{xu2003,essid2013}, hidden Markov chain \cite{fu2018}, representation learning in image recognition \cite{lee1999}, community discovery \cite{wang2011}, and clustering methods \cite{turkmen2015}. This paper introduces two NMF-like procedures for forecasting and clustering of time series. Forecasting for temporal time series has been previously done before through a mixed linear regression and matrix factorization in \cite{yu2016temporal}, matrix completion for one temporal time serie in \cite{gillard2018structured}, and tensor factorization \cite{de2017tensorcast,yokota2018missing,tan2016short}.
Sliding Mask Method (SMM) inputs the forecast values and it can be viewed as a ‘‘nonnegative'' matrix completion algorithm under low nonnegative rank assumption. This framework raises two issues. A first challenge is uniqueness of the decomposition, also referred to as {\it identifiability} of the model. In Theorem~\ref{thm:uniqueness}, we introduce a new condition that ensures uniqueness from partial observation of the target matrix $\mathbf M$. An other challenge, as pointed by \cite{vavasis2009} for instance, is that solving \textit{exactly} NMF decomposition problem is $\mathrm{NP}$-hard. Nevertheless NMF-type problems can be solved efficiently using (accelerated) proximal gradient descent method \lm{\cite{parikh2013}} for block-matrix coordinate descent in an \textit{alternating projection scheme}, \textit{e.g.}, \cite{javadi2017} and references therein. We rely on these techniques to introduce algorithms inputting the forecast values based on NMF decomposition, see Section~\ref{sec:algo}. Theorem~\ref{thm:robust} complements proving robustness of solutions of NMF-type algorithms when entries are missing or corrupted by noise.
In practical situation, one may face a large number of time series to forecast, {\it e.g.} supply chain optimization, electricity consumption forecast... In this case, one cannot forecast each time series separately and/or clustering the set of time series without facing large algorithmic complexity. We can address this issue using Latent Clustered Forecast (LCF). This method uses NMF decomposition $\mathbf M={\mathbf W}\H$ as a dimension reduction step. As we will see in Section~\ref{sec:blackbox}, it performs forecast on rows of~$\H$ (using any black-box supervised forecast solver) and/or clustering on the rows of ${\mathbf W}$.
\pagebreak[3]
\textbf{Notation: }
Denote by $A^\top$ the transpose of matrix~$A$. We use $\mathds{R}_+^{n\times p}$ to denote $n\times p$ nonnegative matrices. It would be useful to consider a columns description of matrix
$\mathbf A=[A_1\cdots A_{n_2}]$ and row decomposition $\mathbf A^T=[(A^{(1)})^\top\cdots (A^{(n_1)})^\top]$ for $\mathbf A\in\mathds R^{n_1\times n_2}$ where $A_k$ denotes the columns of $\mathbf A$ and~$A^{(k)}$ denotes the rows of~$\mathbf A$. $A_{i,j}$ indicates the elements of matrix~$\mathbf A$. \lm{$[n]$ represents the set $\{1,2,\dots,n\}$}, \lm{while~$\mathbf 1_d$ is the all-ones vector of size $d$.} $\mathds{1}_{\mathcal A}$ is the indicator function of $\mathcal A$, such that $\mathds{1}_{\mathcal A} = 0$ if condition~$\mathcal A$ is verified, $\infty$ otherwise.
\subsection{The time series forecasting problem}
\begin{table*}[t]
\centering
\begin{tabular}{llll}
\toprule
Acronym & Name & Objective & Constraints: ${\mathbf W}\ge \mathbf 0$\ +\ \\
\midrule
\multirow{2}{*}{NMF} & Nonnegative Matrix Factorization & \multirow{2}{*}{$\mathbf F_1$} & \multirow{2}{*}{$\H\ge \mathbf 0$} \\
& \lm{\cite{cichocki2006}} & & \\
{SNMF} & Semi NMF \lm{\cite{gillis2015a}} & {$\mathbf F_1$} & \\
NNMF & Normalized NMF & $\mathbf F_1$ & $\H\ge \mathbf 0$, ${\mathbf W}\mathbf{1}=\mathbf{1}$ \\
SNNMF & Semi Normalized NMF & $\mathbf F_1$ & ${\mathbf W}\mathbf{1}=\mathbf{1}$ \\
\multirow{2}{*}{AMF} & Archetypal Matrix Factorization & \multirow{2}{*}{$\mathbf F_2$} & \multirow{2}{*}{${\mathbf W}\mathbf{1}=\mathbf{1}$, $\mathbf V\ge \mathbf 0$, $\mathbf V\mathbf{1}=\mathbf{1}$} \\
& \lm{\cite{javadi2017}} & \\
ANMF & Archetypal NMF & $\mathbf F_2$ & $\H\ge \mathbf 0$, $\mathbf V\ge \mathbf 0$, $\mathbf V\mathbf{1}=\mathbf{1}$ \\
ANNMF & Archetypal Normalized NMF & $\mathbf F_2$ & ${\mathbf W}\mathbf{1}=\mathbf{1}$, $\H\ge \mathbf 0$, $\mathbf V\ge \mathbf 0$, $\mathbf V\mathbf{1}=\mathbf{1}$ \\
\bottomrule
\bottomrule
mNMF & Mask NNMF & $\mathbf F_3$ & $\mathbf T \mathbf N=\mathbf X, {\mathbf W}\mathbf{1}=\mathbf{1}, \H\ge \mathbf 0$ \\
mAMF & Mask AMF & $\mathbf F_4$ & $\mathbf T \mathbf N=\mathbf X, {\mathbf W}\mathbf{1}=\mathbf{1}$, $\mathbf V\ge \mathbf 0$, $\mathbf V\mathbf{1}=\mathbf{1}$ \\
\bottomrule
\end{tabular}
\caption{The seven block convex programs achieving matrix factorization of nonnegative matrices. The objectives are $\mathbf F_1:=\|\mathbf M-{\mathbf W}\H\|_F^2$ and $\mathbf F_2:=\|\mathbf M-{\mathbf W}\H\|_F^2+\lambda\|\H-\mathbf V\mathbf M\|_F^2$. The two last lines are SMM procedures with sliding operator ${\boldsymbol \Pi}$ and objectives $\mathbf F_3:=\|\mathbf N-{\mathbf W}\H\|_F^2$ and $\mathbf F_4:=\|\mathbf N-{\mathbf W}\H\|_F^2+\lambda\|\H-\mathbf V\mathbf N\|_F^2$.
}
\label{table:NMF_family}
\end{table*}
This article considers $N\geq1$ times series on the same temporal period with $T\geq1$ timestamps in a setting where $N\geq T$ and possibly $N\gg T$. We would like to forecast the next $F\geq 1$ times. Additionally, one may also aim at clustering these $N$ time series, and/or reduce the ambient dimension $N\times T$ while maintaining a good approximation of these times series. The observed times series can be represented as a matrix~$\mathbf M$ of size $N\times T$. A row $\mathbf M^{(i)}$ of~$\mathbf M$ represents a time series and a column $\mathbf M_j$ of $\mathbf M$ represents a timestamp record. We assume that there exists a {\it target} matrix~$\mathbf M^\star$ with
\[
\mathbf M=\mathbf M^\star_{{T}}+\mathbf E\,,
\]
where $\mathbf E$ is some noise term and $\mathbf M^\star_{{T}}\in\mathds{R}_+^{N\times T}$ is a sub-matrix
\[
\mathbf M^\star:=\Big[\underbrace{\mathbf M^\star_{{T}}}_{\text{past}}\ \underbrace{\mathbf M^\star_{{F}}}_{\text{future}}\Big]
\]
of the {\it target matrix} of size $N\times (T+F)$ that can be split into timestamps up today $\mathbf M^\star_{{T}}\in\mathds{R}_+^{N\times T}$ and future timestamps $\mathbf M^\star_{{F}}\in\mathds{R}_+^{N\times F}$ to be forecast.
The statistical task is the following: given the observation $\mathbf M$ predict the future target values $\mathbf M^\star_F$, and incidentally the ‘‘denoised'' $\mathbf M^\star_{{T}}$.
\subsection{Nonnegative and Archetypal Analysis}
We aim to decompose nonnegative matrix $\mathbf M\in\mathds{R}_+^{N\times T}$ as the product of nonnegative matrix ${\mathbf W}\in\mathds{R}^{N\times K}$ and matrix $\H\in\mathds{R}^{K\times T}$ by minimizing the Frobenius norm of the difference between $\mathbf M$ and the reconstructed matrix $\widehat\mathbf M:={\mathbf W}\H$ \lm{\cite{cichoki2009}}:
\begin{align}
\tag{NMF}
\min_{{\mathbf W}\ge \mathbf 0, \H\ge \mathbf 0}\mbox{ } \|\mathbf M-{\mathbf W}\H\|_F^2 \,. \label{eq:nmf_original}
\end{align}
Another approach consists in the {\it Archetypal Analysis}:
\begin{align}
\tag{AMF}
\min_{\substack{{\mathbf W}\ge \mathbf 0,{\mathbf W}\mathbf{1}=\mathbf{1} \\ \mathbf V\ge \mathbf 0,\mathbf V\mathbf{1}=\mathbf{1}}}\mbox{ }& \|\mathbf M-{\mathbf W}\H\|_F^2 + \lambda\|\H-\mathbf V\mathbf M\|_F^2\,, \label{eq:amf_original}
\end{align}
where $\lambda>0$ is a tuning parameter, see for instance \lm{\cite{javadi2017}}. Different normalisation and constraints can be considered, we exhibit~$7$ variants in Table~\ref{table:NMF_family}. We will be particularly interested in
\begin{align}
\tag{NNMF}
\min_{\substack{{\mathbf W}\ge \mathbf 0,{\mathbf W}\mathbf 1_K=\mathbf 1_N\\ \H\ge \mathbf 0}}\mbox{ } \|\mathbf M-{\mathbf W}\H\|_F^2 \,. \label{eq:nnmf}
\end{align}
\vspace*{-0.3cm}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=5cm]{coding.png}
\vspace*{-0.5cm}
\end{center}
\caption{Nonnegative and Archetypal representations performs dimension reduction using $\H$ as decoding, while coding is based on conic (resp.~barycentric) coordinates (w.r.t.~rows of $\H$) in the non-normalized case (resp. normalized case ${\mathbf W}\mathbf{1}_K=\mathbf{1}_N$).
\vspace*{-0.5cm}}
\label{fig:dim_red}
\end{figure}
An important parameter is the so-called \textit{nonnegative rank} $K\geq1$ \lm{\cite{gillis2012a}} that governs dimension reduction performed by these matrix factorization, see Figure~\ref{fig:dim_red}. Note that the left hand side matrix ${\mathbf W}$ is always nonnegative and may be or may not be normalized so that ${\mathbf W}\mathbf{1}_K=\mathbf{1}_N$. Each row of the observation~$\mathbf M$ is then a weighted sum of rows of~$\H$, these weights being encoded by~${\mathbf W}$. Hence, each time series (row $\mathbf M^{(i)}$) is encoded by ${\mathbf W}^{(i)}$ a row of~${\mathbf W}$. Decoding is performed multiplying ${\mathbf W}^{(i)}$ by $\H$. When~${\mathbf W}$ is normalized, the aforementioned factorisation algorithms search for the best polytope encapsulating the cloud of point given by the rows of~$\mathbf M$. The points generating this polytope are the rows of~$\H$ while each cloud point (here a row~$\mathbf M^{(i)}$ of~$\mathbf M$) is localized thanks to its barycentric coordinates given by~${\mathbf W}^{(i)}$, for further details see \lm{\cite{gillis2014}}. When~${\mathbf W}$ is not normalized, the aforementioned factorisation algorithms search for the best cone encapsulating the cloud of point given by the rows of $\mathbf M$. The rays generating this cone are the rows of $\H$ while each cloud point (here a row $\mathbf M^{(i)}$ of $\mathbf M$) is localized thanks to its conic coordinates given by ${\mathbf W}^{(i)}$, for further details see \lm{\cite{ge2015}}.
The parameter $\lambda$ in the archetypal analysis enforces some kind of reciprocity. In NMF, the cloud of points given by the rows of $\mathbf M$ should be encapsulated by the rows of $\H$. Archetypal analysis penalises the reciprocal: the rows of $\H$ should be encapsulated in the convex/conic volume given by the cloud of points of the rows of $\mathbf M$. Both parameters $K$ and $\lambda$ can be tuned by cross-validation \lm{\cite{arlot2010}}, as done in our experiments, see Section~\ref{sec:experiments}.
\subsection{Contribution: Towards Nonnegative Matrix Completion for Time series}
We are interested in a {\it Matrix Completion} problem using Nonnegative Matrix Factorization. We consider $n\times p$ matrices $\mathbf X_0,\mathbf X^\star,{\mathcal Z}$ which, in the Sliding Mask Method (SMM), are linear transformations of matrices $\mathbf M^\star, \mathbf M^\star_T,\mathbf E$ respectively, see Section~\ref{sec:mask}. Note that $n:=(B-W+1)N>N$ and $p:=WP>F$ where $B,W,F$ will be introduced in Section~\ref{sec:mask}.
Consider a ‘‘{\it mask}'' operator $\mathbf T(\mathbf N)$ that sets to zero $N\times F$ values of $\mathbf N$. Namely, given $\mathbf N\in\mathds R^{n\times p}$, define
\[
\mathbf T(\mathbf N) = \left[\begin{array}{c|c}
\mathbf N_1& \mathbf N_2\\ \hline
\mathbf N_3 & \mathbf 0_{N \times F}
\end{array}\right]\,,
\]
where $(\mathbf N_i)_{i=1}^4$ are blocks of $\mathbf N = \left[\begin{array}{c|c}
\mathbf N_1& \mathbf N_2\\ \hline
\mathbf N_3 & \mathbf N_4
\end{array}\right]$.
Our goal is the following matrix completion problem: Given a noisy and incomplete observation
\begin{equation}
\label{eq:observations}
\mathbf X:=\mathbf T(\mathbf X_0)+{\mathcal F}\,,
\end{equation}
where ${\mathcal F}$ is some noise term, find a good estimate of the target $\mathbf X_0$. One can consider the {\it Nonnegative Matrix Completion} referred to as Mask NNMF:
\begin{align}
\min_{\substack{{\mathbf W}\mathbf 1=\mathbf 1,{\mathbf W}\ge \mathbf 0\\\H\ge \mathbf 0\\\mathbf T \mathbf N = \mathbf X}}\mbox{ }& \|\mathbf N - {\mathbf W}\H\|_F^2\,, \label{eq:nmf3}
\tag{mNMF}
\end{align}
where solutions $\mathbf N\in\mathds R^{n\times p}$ are such that $\mathbf T \mathbf N=\mathbf X$ (observed values) and $\mathbf T^\perp \mathbf N = \mathbf T^\perp ({\mathbf W}\H)$ (forecast values). This latter formulation is an instance of Matrix Completion \lm{\cite{nguyen2019}}. Forecasting problem reduces to Matrix Completion problem, whose aim is finding the nonnegative matrix factorization $\mathbf N\simeq {\mathbf W}\H$ of observed matrix~$\mathbf X$ such that $\mathbf T \mathbf N=\mathbf X$.
\begin{remark}
Problem \eqref{eq:nmf3} is NNMF when $\mathbf T=\mathds{I}$, where~$\mathds{I}$ is the identity operator.
\end{remark}
Dropping $\H\geq0$, another approach is the {\it Archetypal Matrix Completion} referred to as Mask AMF:
\begin{align}
\min_{\substack{{\mathbf W}\ge \mathbf 0,{\mathbf W}\mathbf{1}=\mathbf{1} \\ \mathbf V\ge \mathbf 0,\mathbf V\mathbf{1}=\mathbf{1}\\ \mathbf T \mathbf N = \mathbf X}}& \|\mathbf N-{\mathbf W}\H\|_F^2 + \lambda\|\H-\mathbf V\mathbf N\|_F^2 \label{eq:amf1}
\tag{mAMF}
\end{align}
\begin{remark}
When $\mathbf T=\mathds{I}$, Problem~\eqref{eq:amf1} reduces to standard AMF formulation (AMF).
\end{remark}
\subsubsection{Uniqueness from partial observations}
Define the mask $\mathbf X^\star$ of $\mathbf X_0$ by
\[
\mathbf X^\star:=\mathbf T(\mathbf X_0) =: \left[\begin{array}{c|c}
\mathbf X_1 & \mathbf X_2\\ \hline
\mathbf X_3 & \mathbf 0_{N \times F}
\end{array}\right]\,,
\]
where $\mathbf X_1\in\mathds R^{(n-N)\times (p-F)}$, $\mathbf X_2\in\mathds R^{(n-N)\times F}$, and $\mathbf X_3\in\mathds R^{N\times (p-F)}$ are blocks of $\mathbf X_0$. Let us consider
\[\begin{aligned}
&\mathbf T_{{\mathrm{train}}}(\mathbf X_0) := [\mathbf X_1 \; \mathbf X_2]\,, && \mathbf T_{\mathrm{test}}(\mathbf X_0) := [\mathbf X_3 \; \mathbf 0_{N \times F}]\,, \\
&\mathbf T_{T}(\mathbf X_0) := \Big[\begin{array}{c}\mathbf X_1\\ \mathbf X_3\end{array}\Big]\,, && \mathbf T_{F}(\mathbf X_0) := \Big[\begin{array}{c}\mathbf X_2\\ \mathbf 0_{N \times F}\end{array}\Big]\,.
\end{aligned}\]
\begin{remark}
Let $\mathbf X_0:={\mathbf W}_0\H_0$, $\H_0:=[{\H_0}_T\; {\H_0}_F]$, and ${\mathbf W}_0^\top:=[{{\mathbf W}_0}_{{\mathrm{train}}}^\top\; {{\mathbf W}_0}_{\mathrm{test}}^\top]$, then
\[\begin{aligned}
&\mathbf T_{{\mathrm{train}}}(\mathbf X_0) = {{\mathbf W}_0}_{{\mathrm{train}}} \H_0\,, && \mathbf X_3 = {{\mathbf W}_0}_{\mathrm{test}} {\H_0}_T\,, \\
&\mathbf T_{T}(\mathbf X_0) = {{\mathbf W}_0} {\H_0}_T\,, && \mathbf X_2 = {{\mathbf W}_0}_{{\mathrm{train}}} {\H_0}_F\,.
\end{aligned}\]
\end{remark}
At first glance, the coding-decoding scheme of Figure~\ref{fig:dim_red} can be ill-posed and/or not robust to noise. The first issue is the {\it uniqueness} of the decomposition ${\mathbf W}_0\H_0$ {\it given partial observations}, namely proving that {\it Partial Observation Uniqueness} \eqref{eq:pou} property holds:
\begin{equation}
\label{eq:pou}\tag{\bf POU}
\begin{aligned}
&\text{If }\mathbf T({\mathbf W}\H)=\mathbf T({\mathbf W}_0\H_0)\notag\\&\text{Then }({\mathbf W},\H)\equiv({\mathbf W}_0,\H_0)\,,
\end{aligned}
\end{equation}
where $\equiv$ means up to positive scaling and permutation: if an entry-wise nonnegative pair $({\mathbf W},\H)$ is given then $({\mathbf W}\mathbf P{\mathcal{D}},{\mathcal{D}}^{-1}\mathbf P^\top\H)$ is also a nonnegative decomposition ${\mathbf W}\H={\mathbf W}\mathbf P{\mathcal{D}}\times{\mathcal{D}}^{-1}\mathbf P^\top\H$, where~${\mathcal{D}}$ scales and~$\mathbf P$ permutes the columns (resp.~rows) of~${\mathbf W}$ (resp.~$\H$).
When we observe the full matrix $\mathbf X_0={\mathbf W}_0\H_0$, the uniqueness issue has been addressed under some sufficient conditions on ${\mathbf W},\H$, {\it e.g.}, {\it Strongly boundary closeness} of \cite{laurberg2008}, \lm{{\it Complete factorial sampling} of \cite{donoho2004}, and {\it Separability} of \cite{recht2012}}. A necessary and sufficient condition exists:
\begin{theorem}[\cite{thomas1974}]\label{theo:cns}
The decomposition $\mathbf X_0:={\mathbf W}_0\H_0$ is unique up to permutation and positive scaling of columns (resp.~rows) of ${\mathbf W}_0$ (resp.~$\H_0$) {\bf if and only if} the $K$-dimensional positive orthant is the only $K$-simplicial cone verifying
\[
\mathrm{Cone}({\mathbf W}_0^\top) \subseteq \mathcal C \subseteq \mathrm{Cone}(\H_0)\,,
\]
where $\mathrm{Cone}(\mathbf A)$ is the cone generated by the rows of $\mathbf A$.
\end{theorem}
\medskip
\noindent
Our first main assumption is that:\\
\noindent
$\bullet$ {({\bf A1})} {\it In the set given by the union of sets:}
\begin{align*}
&\{\mathcal C \ :\ \mathrm{Cone}({{{\mathbf W}_0}_{{\mathrm{train}}}}^\top) \subseteq \mathcal C \subseteq \mathrm{Cone}(\H_0)\}\\
&\bigcup \{\mathcal C \ :\ \mathrm{Cone}({\mathbf W}_0^\top) \subseteq \mathcal C \subseteq \mathrm{Cone}({\H_0}_T)\}\,,
\end{align*}
{\it the nonnegative orthant is the only $K$-simplicial cone.}
\medskip
\noindent
It is clear that this property is implied by the following one, {({\bf A'1})} $\Rightarrow$ {({\bf A1})}.
\medskip
\noindent
$\bullet$ {({\bf A'1})} {\it In the set}
\begin{align*}
&\{\mathcal C \ :\ \mathrm{Cone}({{{\mathbf W}_0}_{{\mathrm{train}}}}^\top) \subseteq \mathcal C \subseteq \mathrm{Cone}({\H_0}_T)\}
\end{align*}
{\it the nonnegative orthant is the only $K$-simplicial cone.}
\medskip
\noindent
Given the standard definition:
\begin{definition}[\cite{javadi2017}] The convex hull ${\rm{conv}}(\mathbf X_0)$ has an internal radius $\mu>0$ if it contains an $K-1$ dimensional ball of radius $\mu$.
\end{definition}
\medskip
\noindent
Our second main assumption is that:\\
\noindent
$\bullet$ {({\bf A2})} {\it Assume that}
\begin{align}
{\rm{conv}}(\underbrace{\mathbf T_{{\mathrm{train}}}(\mathbf X_0)}_{={\mathbf W}_{0\mathrm{train}}\H_0}) \text{ has internal radius } \mu>0\,.
\label{hyp:radius}
\tag{\bf A2}
\end{align}
\begin{theorem}
\label{thm:uniqueness}
Condition\! {$(${\bf A1}$)$} is sufficient for \eqref{eq:pou}.\\
If {$(${\bf A1}$)$} and \eqref{hyp:radius} holds, $\mathbf T({\mathbf W}\H)=\mathbf T({\mathbf W}_0\H_0)$ and ${\mathbf W}_0\mathbf{1}={\mathbf W}\mathbf{1}=\mathbf{1}$ then $({\mathbf W},\H)=({\mathbf W}_0,\H_0)$ up to permutation of columns (resp.~rows) of ${\mathbf W}$ (resp.~$\H$), and there is no scaling
\end{theorem}
\vspace*{-0.3cm}
\begin{proof}
Proofs are given in Supplement Material.
\end{proof}
\begin{remark}
By Theorem~\ref{theo:cns}, observe that {({\bf A'1})} is a necessary and sufficient condition for the uniqueness of the decomposition $\mathbf X_1 = {{\mathbf W}_0}_{{\mathrm{train}}} {\H_0}_T$. Then, using {$(${\bf A'1}$)$} $\Rightarrow$ {$(${\bf A1}$)$}, we understand that if decomposition of $\mathbf X_1 = {{\mathbf W}_0}_{{\mathrm{train}}} {\H_0}_T$ is unique then \eqref{eq:pou} holds.
\end{remark}
\subsubsection{Robustness under partial observations}
The second issue is {\it robustness to noise}. To the best of our knowledge, all the results addressing this issue assume that the noise error term is small enough, {\it e.g.}, \cite{laurberg2008}, \cite{recht2012}, or \cite{javadi2017}. In this paper, we extend these stability result to the nonnegative matrix completion framework (partial observations) and we also assume that noise term $\|{\mathcal F}\|_F$ is small enough.
In the normalized case ({\it i.e.}, ${\mathbf W}\mathbf 1=\mathbf 1$), both issues (uniqueness and robustness) can be handle with the notion of $\alpha$-uniqueness, introduced by \cite{javadi2017}. This notion does not handle the matrix completion problem we are addressing. To this end, let us introduce the following notation. Given two matrices $\mathbf A\in\mathds{R}^{n_a\times p}$ and ${\mathcal B}\in\mathds{R}^{n_b\times p}$ with same row dimension, and $\mathcal{C}\in\mathds{R}^{n_a\times n_b}$, define the divergence $\mathcal D(\mathbf A,{\mathcal B})$ as
\begin{align*}
\mathcal D(\mathbf A,{\mathcal B}) &:=\min_{\mathcal{C}\ge \mathbf 0\,,\ \mathcal{C}\mathbf{1}_{n_b}=\mathbf{1}_{n_a}}
\sum_{a=1}^{n_{a}} \Big\|A^{(a)}- \sum_{b=1}^{n_b} C_{ab} B^{(b)}\Big\|_F^2\,, \\
&=\min_{\mathcal{C}\ge \mathbf 0\,,\ \mathcal{C}\mathbf{1}_{n_b}=\mathbf{1}_{n_a}}
\|\mathbf A - \mathcal{C}{\mathcal B}\|_F^2\,.
\end{align*}
which is the squared distance between rows of $\mathbf A$ and ${\rm{conv}}({\mathcal B})$, the convex hull of rows of ${\mathcal B}$.
For ${\mathcal B}\in\mathds R^{n\times p}$ define
\[
\widetilde{\mathcal D}(\mathbf A,{\mathcal B}) :=\min_{\substack{\mathcal{C}\ge \mathbf 0\,,\ \mathcal{C}\mathbf{1}_{n}=\mathbf{1}_{n_a}\\ \mathbf T(\mathbf N-{\mathcal B})=0}}
\|\mathbf A - \mathcal{C}\mathbf N\|_F^2\,.
\]
\begin{definition}[$\mathbf T_\alpha$-unique] Given $\mathbf X_0\in\mathds R^{n\times p}, {\mathbf W}_0\in\mathds R^{n\times K}, \H_0\in\mathds R^{K\times p}$, the factorization $\mathbf X_0={\mathbf W}_0\H_0$ is $\mathbf T_\alpha$-unique with parameter $\alpha>0$ if for all $\H\in\mathds R^{K\times p}$ with ${\rm{conv}}(\mathbf X_0)\subseteq{\rm{conv}}(\H)$:
\begin{align*}
\widetilde{\mathcal D}(\H,\mathbf X_0)^{1\slash 2} & \ge \widetilde{\mathcal D}(\H_0,\mathbf X_0)^{1\slash 2} \\
& + \alpha\left\{\mathcal D(\H,\H_0)^{1\slash 2} + \mathcal D(\H_0,\H)^{1\slash 2}\right\}\,.
\end{align*}
\end{definition}
\medskip
\noindent
Our third main assumption is given by:\\%that there exists a \textit{nonnegative rank} $K\geq1$ such that:
\noindent
$\bullet$ {({\bf A3})} {\it Assume that}
\begin{align}
&\mathbf X_0={\mathbf W}_0\H_0\text{ is }\mathbf T_\alpha\text{-unique
\label{hyp:mask}
\tag{\bf A3}
\end{align}
\medskip
\begin{theorem}
\label{thm:robust}
If \eqref{hyp:radius} and~\eqref{hyp:mask} hold then there exists positive reals $\Delta$ and $\Lambda$ (depending on $\mathbf X_0$) such that, for all ${\mathcal F}$ such that $\|{\mathcal F}\|_F\leq\Delta$ and $0\leq \lambda\leq\Lambda$, any solution $({\widehat{\mathbf W},\widehat\H})$ to \eqref{eq:amf1} (if $\lambda\neq 0$) or \eqref{eq:nmf3} (if $\lambda=0$) with observation \eqref{eq:observations} is such that:
\[\begin{aligned}
\sum_{\ell\le[K]} \min_{\ell'\le[K]} \|{\H}_{0\ell} - {\widehat \H}_{\ell'}\|_2^2
& \le c\,\|{\mathcal F}\|_F^2 \,,
\end{aligned}\]
where $c$ is a constant depending only on $\mathbf X_0$.
\end{theorem}
\begin{proof}
Proofs are given in Supplement Material.
\end{proof}
\subsection{Outline}
The rest of the paper is organized as follows. In Section~\ref{sec:mask} we discuss the {\it Sliding Mask Method} (SMM), while Section~\ref{sec:blackbox} is devoted to describe the {\it Latent Clustered Forecast} (LCF). Numerical experiments and conclusions are presented in Sections~\ref{sec:experiments} and~\ref{sec:conclusions}, respectively. A repository on the numerical expermients can be found at \url{https://github.com/Luca-Mencarelli/Nonnegative-Matrix-Factorization-Time-Series}.
\section{The Sliding Mask Method}\label{sec:mask}
\subsection{Sliding window as forecasting}
One is given $N$ time series $\mathbf M^{(1)},\ldots, \mathbf M^{(N)}\in\mathds R^T$ over a period of $T$ dates. Recall~$\mathbf M\in\mathds R^{N\times T}$ is the matrix of observation such that $\mathbf M^\top=[(\mathbf M^{(1)})^\top\cdots (\mathbf M^{(N)})^\top]$ and assumed entry-wise nonnegative. We assume some {\it periodicity} in our time series, namely that~$\mathbf M^\star$ can be split into $B$ matrix blocks of size $N\times P$ where $P=(T+F)/B$, see Figure~\ref{fig:blocks}.
\vspace*{-0.3cm}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=8cm]{block.png}
\vspace*{-0.5cm}
\end{center}
\caption{Target matrix $\mathbf M^\star$ can be split into $B$ blocks of same time length~$P$.
}
\label{fig:blocks}
\end{figure}
Given $W\geq 1$ and a $T\times N$ matrix $\mathbf M$, we define~$\mathbf \Pi(\mathbf M)$ the linear operator that piles up $W$ consecutive sub-blocks in a row, as depicted in Figure~\ref{fig:mask}. This process looks at $W$ consecutive blocks in a \textit{sliding} manner. Note that~$\mathbf \Pi(\mathbf M)$ is an {\it incomplete} matrix where the missing values are depicted in orange in Figure~\ref{fig:mask}, they correspond to the time-period to be forecasted. Unless otherwise specified, these unobserved values are set to zero. Remark that $\mathbf \Pi(\mathbf M)$ has $W$ columns blocks, namely $WP$ columns and $(B-W+1)N$ rows. By an abuse of notation, we also denote
\[
\mathbf \Pi:\mathds R^{N\times(T+F)}\to \mathds R^{(B-W+1)N\times WP}
\]
the same one-to-one linear matrix operation on matrices of size $N\times (T+F)$. In this case, $\mathbf X_0:=\mathbf \Pi(\mathbf M^\star)$ is a {\it complete} matrix where the orange values have been implemented with the future values of the target~$\mathbf M_F^\star$.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=12cm]{new_fig.png}
\vspace*{-0.3cm}
\end{center}
\caption{The operator ${\boldsymbol \Pi}(\mathbf M)$ outputs an incomplete $(B-W+1)N\times WP$ matrix given by a mask where the $NF$ orange entries are not observed. These entries corresponds to future times that should be forecasted.}
\label{fig:mask}
\end{figure}
The rationale behind is recasting the forecasting problem as a supervised learning problem where one observes, at each line of ${\boldsymbol \Pi}(\mathbf M)$, the $WP-F$ first entries and learn the next $F$ entries. The training set is given by rows $1,3,5,7$ in ${\boldsymbol \Pi}(\mathbf M)$ of Figure~\ref{fig:mask} and the validation set is given by rows $2,4,6,8$ where one aims at predicting the $F$ missing values from the $WP-F$ first values of these rows.
\subsection{Mask NNMF and Mask AMF}
Consider a matrix completion version of NMF with observations
\[
\mathbf X:=\mathbf \Pi(\mathbf M)=\underbrace{\mathbf \Pi(\mathbf M_T^\star)}_{\mathbf X^\star}+\underbrace{\mathbf \Pi(\mathbf E)}_{{\mathcal Z}}\,,
\]
\vspace*{-0.7cm}
\noindent
and
\vspace*{-0.2cm}
\begin{align}
\min_{{\mathbf W}\mathbf 1=\mathbf 1,{\mathbf W}\ge \mathbf 0,\H\ge \mathbf 0}\mbox{ }& \|\mathbf X - \mathbf T ({\mathbf W}\H)\|_F^2\,, \label{eq:nmf2}
\end{align}
where the ‘‘mask'' operator $\mathbf T$ is defined by zeroing the ‘‘future'' values (in orange in Figure~\ref{fig:mask}). Note that
\[
\mathbf T(\underbrace{{\boldsymbol \Pi}(\mathbf M^\star)}_{\mathbf X_0})={\boldsymbol \Pi}(\mathbf M^\star_T)=\mathbf X^\star\,.
\]
Moreover, note that Problem \eqref{eq:nmf2} is equivalent to mask NNMF \eqref{eq:nmf3}. If we drop the nonnegative constraints on $\H$ and consider the archetypal approach, we obtain mask AMF \eqref{eq:amf1}. In particular, Theorem~\ref{thm:robust} applies proving that \eqref{eq:nmf3} and \eqref{eq:amf1} are robust to small noise.
\subsection{Algorithms}
\label{sec:algo}
\subsubsection{Alternating Least Squares for \eqref{eq:nmf3}}
The basic algorithmic framework for matrix factorization problems is {\it Block Coordinate Descent} (BCD) method, which can be straightforwardly adapted to~\eqref{eq:nmf3} (see Supplement Material). BCD for~\eqref{eq:nmf3} reduces to {\it Alternating Least Squares} (ALS) algorithm (see Algorithm~\ref{algo:als}), when an alternative minimization procedure is performed and matrix ${\mathbf W}\H$ is projected onto the linear subspace $\mathbf T\mathbf N=\mathbf X$ by means of operator $\mathcal P_{\mathbf X}$, as follows:
\[
\mathbf N:=\mathcal P_{\mathbf X}({\mathbf W}\H) : \mathbf T\mathbf N=\mathbf X \mbox{ and } \mathbf T^\perp\mathbf N={\mathbf W}\H\,.
\]
\begin{comment}
\begin{algorithm}
\caption{BCD for mNMF} \label{algo:nmf1}
\begin{algorithmic}[1]
\STATE {\bf Initialization}: choose $\H^0\ge \mathbf 0,{\mathbf W}^0\ge \mathbf 0$, and $\mathbf N^0\ge \mathbf 0$, set $i:=0$.
\WHILE{stopping criterion is not satisfied}
\STATE $\H^{i+1}:=\mbox{update}(\H^{i},{\mathbf W}^{i},\mathbf N^{i})$ \label{updateH}
\STATE ${\mathbf W}^{i+1}:=\mbox{update}(\H^{i+1},{\mathbf W}^{i},\mathbf N^{i})$ \label{updateW}
\STATE $\mathbf N^{i+1}:=\mbox{update}(\H^{i+1},{\mathbf W}^{i+1},\mathbf N^{i})$ \label{updateM}
\STATE $i:=i+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\end{comment}
\begin{algorithm}
\caption{ALS for mNMF} \label{algo:als}
\begin{algorithmic}[1]
\STATE {\bf Initialization}: choose $\H^0\ge \mathbf 0,{\mathbf W}^0\ge 0$, set $\mathbf N^0:=\mathcal P_{\mathbf X}(\H^0{\mathbf W}^0)$ and $i:=0$.
\WHILE{stopping criterion is not satisfied}
\STATE $\H^{i+1}:=\min_{\H\ge 0}\|\mathbf N^i-{\mathbf W}^i \H\|_F^2$ \label{updateH_als}
\STATE ${\mathbf W}^{i+1}:=\min_{{\mathbf W}\ge 0, {\mathbf W}\mathbf{1}=\mathbf{1}}\|\mathbf N^i-{\mathbf W} \H^{i+1}\|_F^2$ \label{updateW_als}
\STATE set $\mathbf N^{i+1}:=\mathcal P_{\mathbf X}({\mathbf W}^{i+1}\H^{i+1})$ \label{updateM_als}
\STATE $i:=i+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
{\it Hierarchical Alternating Least Squares} (HALS) is an ALS-like algorithm obtained by applying an exact coordinate descent method \cite{gillis2014}. Moreover, an accelerated version of HALS is proposed in \cite{gillis2012} (see Supplement Material).
\subsubsection{Projected Gradient for \eqref{eq:amf1}}
{\it Proximal Alternative Linear Minimization} (PALM) method, introduced in \cite{bolte2014} and applied to AMF by \cite{javadi2017}, can be also generalized to \eqref{eq:amf1} (see Algorithm~\ref{algo:nmf2}).
\begin{algorithm}[!htbp]
\caption{PALM for mAMF} \label{algo:nmf2}
\begin{algorithmic}[1]
\STATE {\bf Initialization}: chose $\H^0$, ${\mathbf W}^0\ge \mathbf 0$ such that ${\mathbf W}^0\mathbf 1=\mathbf 1$, set $\mathbf N^0:=\mathcal{P}_{\mathbf X}({\mathbf W}^0\H^0)$ and $i:=0$.
\WHILE{stopping criterion is not satisfied}
\STATE $\widetilde{\H}^i := \H^i - \frac{1}{\gamma_1^i} {{\mathbf W}^i}^\top\left(\H^i {\mathbf W}^i- \mathbf N^i\right)$ \label{step:palm1}
\STATE $\H^{i+1} := \widetilde{\H}^i-\frac{\lambda}{\lambda+\gamma_1^i} \left(\widetilde{\H}^i-\mathcal{P}_{{\rm{conv}}{(\mathbf N^i)}}(\tilde{\H}^i)\right) $
\STATE ${\mathbf W}^{i+1} := \mathcal{P}_{\Delta}\left({\mathbf W}^i-\frac{1}{\gamma_2^i}\left(\H^{i+1} {\mathbf W}^i- \mathbf N^i\right){\H^{i+1}}^\top\right)$ \label{step:palm2}
\STATE $\mathbf N^{i+1} := \mathcal{P}_{\mathbf X}\left(\mathbf N^i+\frac{1}{\gamma_3^i}\left(\H^{i+1} {\mathbf W}^{i+1}- \mathbf N^{i}\right)\right)$
\STATE $i:=i+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
$\mathcal{P}_{{\rm{conv}}(\mathbf A)}$ is the projection operator onto ${\rm{conv}}(\mathbf A)$ and $\mathcal{P}_{\Delta}$ is the projection operator onto the $(N-1)$-dimensional standard simplex $\Delta^N$. The two projections can be efficiently computed by means, {\it e.g.}, Wolfe algorithm \cite{wolfe1976}) and active set method \cite{condat2016}, respectively.
\begin{theorem}\label{prop:palm1}
Let $\varepsilon>0$. If $\gamma_1^i > \|{{\mathbf W}^i}^\top {\mathbf W}^i\|_F$, $\gamma_2^i > \max\left\{\|\H^{i+1}{\H^{i+1}}^\top\|_F,\,\varepsilon\right\}$, and $\gamma_3^i > 1$, for each iteration $i$,
then the sequence $\left(\H^i,{\mathbf W}^i,\mathbf N^i\right)$ generated by Algorithm~\ref{algo:nmf2} converges to a stationary point of $\Psi(\H,{\mathbf W},\mathbf N):=f(\H) + g({\mathbf W}) + p(\mathbf N) + h(\H,{\mathbf W},\mathbf N)$, where: \begin{align*}
&\textstyle f(\H) = \lambda\mathcal{D}(\H,\mathbf N)\,, && \textstyle g({\mathbf W}) = \sum_{k=1}^K \mathds{1}_{\{W_k\in\Delta\}}\,, \\
&p(\mathbf N) = \mathds{1}_{\{\mathbf N=\mathcal P_{\mathbf X}({\mathbf W}\H)\}}\,, && h(\H,{\mathbf W},\mathbf N) = \|\mathbf N-{\mathbf W}\H\|_F^2\,.
\end{align*}
\end{theorem}
\begin{proof}
Proof is given in Supplement Material.
\end{proof}
Finally, the inertial PALM (iPALM) method, introduced for NMF in \cite{pock2016}, is generalized to \eqref{eq:amf1} in Algorithm~\ref{algo:nmf3}.
\begin{algorithm}[!htbp]
\caption{iPALM for mAMF} \label{algo:nmf3}
\begin{algorithmic}[1]
\STATE {\bf Initialization}: $\H^0$, ${\mathbf W}^0\ge 0$ such that ${\mathbf W}^0\mathbf 1=\mathbf 1$, set $\mathbf N^0:=\mathcal{P}_{\mathbf X}({\mathbf W}^0\H^0)$, $\H^{-1}:=\H^0$, ${\mathbf W}^{-1}:={\mathbf W}^0$, $\mathbf N^{-1}:=\mathbf N^0$, and $i:=0$.
\WHILE{stopping criterion is not satisfied}
\STATE $\H^i_1 := {\H}^i + \alpha^i_1\left({\H}^i-{\H}^{i-1}\right)$, $\H^i_2 := {\H}^i + \beta^i_1\left({\H}^i-{\H}^{i-1}\right)$
\STATE $\widetilde{\H}^i := \H^i_1 - \frac{1}{\gamma_1^i} {{\mathbf W}^i}^\top\left(\H^i_2 {\mathbf W}^i- \mathbf N^i\right)$
\STATE $\H^{i+1} := \widetilde{\H}^i-\frac{\lambda}{\lambda+\gamma_1^i} \left(\widetilde{\H}^i-\mathcal{P}_{{\rm{conv}}{(\mathbf N^i)}}(\tilde{\H}^i)\right)$
\STATE ${\mathbf W}^i_1 := {{\mathbf W}}^i + \alpha^i_2\left({{\mathbf W}}^i-{{\mathbf W}}^{i-1}\right)$, ${\mathbf W}^i_2 := {{\mathbf W}}^i_1 + \beta^i_2\left({{\mathbf W}}^i-{{\mathbf W}}^{i-1}\right)$
\STATE ${\mathbf W}^{i+1} := \mathcal{P}_{\Delta}\left({\mathbf W}^i_1-\frac{1}{\gamma_2^i}\left(\H^{i+1} {\mathbf W}^i_2- N^i\right){\H^{i+1}}^\top\right)$
\STATE $\mathbf N^i_1 := {\mathbf N}^i_1 + \alpha^i_3\left({\mathbf N}^i-{\mathbf N}^{i-1}\right)$, $\displaystyle \mathbf N^i_2 := {\mathbf N}^i_1 + \beta^i_3\left({\mathbf N}^i-{\mathbf N}^{i-1}\right)$
\STATE $\mathbf N^{i+1} := \mathcal{P}_{\mathbf X}\left(\mathbf N^i_1+\frac{1}{\gamma_3^i}\left(\H^{i+1} {\mathbf W}^{i+1}- \mathbf N^{i}_2\right)\right)$
\STATE $i:=i+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\begin{remark}
If, for all iterations $i$, $\alpha^i_1=\alpha^i_2=0$ and $\beta^i_1=\beta^i_2=0$, iPALM reduces to PALM.
\end{remark}
\subsubsection{Stopping criterion for normalized NMF}
For NNMF, KKT conditions regarding matrix ${\mathbf W}$ are the following (see Supplement Material):
\begin{equation*}
{\mathbf W}\circ\left(({\mathbf W}\H-\mathbf N)\H^\top + t \, \mathbf 1_K^\top\right) = 0\,.
\end{equation*}
By complementary condition, it follows that, $\forall j$, $t_i = (({\mathbf W}\H-\mathbf N)\H^\top)_{i,j}$. Hence, we compute $t_i$ by selecting, for each row $W^{(i)}$, any positive entry $W_{i,j}>0$.
\begin{remark}
Numerically to obtain a robust estimation of $t_i$, we can average the corresponding values calculated per entry $W_{i,j}$.
\end{remark}
Let $\varepsilon_{\mathbf W}$, $\varepsilon_\H$, and $\varepsilon_\mathbf R$ be three positive thresholds. The stopping criterion for the previous algorithms consists in a combination of:
\begin{enumerate}
\item the maximum number of iterations;
\item the Frobenius norm of the difference of ${\mathbf W}$ and $\H$ at two consecutive iterations, {\it i.e.}, the algorithm stops if
\begin{equation*}
\|{\mathbf W}^{i+1}-{\mathbf W}^{i}\|_F \le \varepsilon_{\mathbf W} \; \wedge \; \|\H^{i+1}-\H^{i}\|_F \le \varepsilon_\H\,;
\end{equation*}
\begin{comment}
\item a criterion based on KKT condition and introduced in \cite{mei2017}, {\it i.e.}, the algorithm stops if
\begin{equation*}
\|\mathbf R({\mathbf W}^{i+1})\|_F + \|\mathbf R(\H^{i+1})\|_F \le \varepsilon_\mathbf R\,,
\end{equation*}
where matrices $\mathbf R({\mathbf W})$ and $\mathbf R(\H)$ are defined as
\begin{equation*}\begin{aligned}
\mathbf R({\mathbf W})_{i,j} &:= |({\mathbf W}\H-\mathbf N)\H^\top)_{i,j}|{\mathds{1}_{\{W_{i,j}\ne 0\}}}\,, \\
\mathbf R(\H)_{i,j} &:= |{\mathbf W}^\top({\mathbf W}\H-\mathbf N))_{i,j}|{\mathds{1}_{\{H_{i,j}\ne 0\}}}\,. \\
\end{aligned}\end{equation*}
\end{comment}
\item a novel criterion based on KKT condition, {\it i.e.}, the algorithm stops if
\begin{equation*}
\|\mathbf R({\mathbf W}^{i+1})\|_F + \|\mathbf R(\H^{i+1})\|_F \le \varepsilon_\mathbf R
\end{equation*}
where matrices $\mathbf R({\mathbf W})$ and $\mathbf R(\H)$ are defined as
\begin{equation*}\begin{aligned}
\mathbf R({\mathbf W})_{i,j} &:= |({\mathbf W}\H-\mathbf N)\H^\top)_{i,j}+t_i|\mathds{1}_{\{W_{i,j}\ne 0\}}\,, \\
\mathbf R(\H)_{i,j} &:= |{\mathbf W}^\top({\mathbf W}\H-\mathbf N))_{i,j}|\mathds{1}_{\{H_{i,j}\ne 0\}}\,. \\
\end{aligned}\end{equation*}
\end{enumerate}
\begin{comment}
\item a criterion similar to the one introduced in \cite{gillis2014}) and based on the first-order optimality conditions. Let $c_0=1$ if the non-negativity constraints $\H\ge\mathbf 0$ belongs to the problem, and 0 otherwise; analogous definition is given for constant $c_1$ with respect to constraint ${\mathbf W}\ge\mathbf 0$. Given a matrix $\mathbf A$, let $[\mathbf A]_0=\|\min\{\mathbf A,\mathbf 0\}\|_F$ We set:
\begin{equation*}\begin{aligned}
C_H(\H) := c_0 [\H]_0 & + c_1 ([\nabla_{\H} h(\H,{\mathbf W},\mathbf M)]_0 \\
& + \left\|\H\circ\nabla_{\H} h(\H,{\mathbf W},\mathbf M)\right\|_F) \\
& +(1-c_1) \|\nabla_{\H} h(\H,{\mathbf W},\mathbf M)\|_F\,.
\end{aligned}\end{equation*}
\end{comment}
\subsection{Large-scale data-set}
Assume the observed matrix $\mathbf X={\boldsymbol \Pi}(\mathbf M)$ is large scaled, namely one has to forecast a large number $N$ of times series (e.g. more than $100,000$) and possibly a large number of time stamps $T$. The strategy, described in Section 1.3.1 in \cite{cichoki2009} for NMF, is to learn the $\H\in\mathds R^{K\times T}$ matrix from a sub-matrix $\mathbf N_r\in\mathds R^{r\times T}$ of $K\leq r\ll N$ rows of $\mathbf N\in\mathds R^{n\times T}$, and learn the ${\mathbf W}\in\mathds R^{N\times K}$ matrix from a sub-matrix $\mathbf N_c\in\mathds R^{N\times c}$ of $K\leq c\ll T$ columns of $\mathbf N\in\mathds R^{N\times T}$.
We denote by $\H_c$ the sub-matrix of $\H$ given by the columns appearing in $\mathbf N_c$ and ${\mathbf W}_r$ the sub-matrix of~$\H$ given by the columns appearing in~$\mathbf N_c$.
This strategy can be generalized to \eqref{eq:nmf3} and \eqref{eq:amf1}. For \eqref{eq:nmf3} this generalization is straightforward, and for \eqref{eq:amf1} one need to change Steps~3-5 in Algorithm~\ref{algo:nmf2} as follows:
\begin{align*}
\textstyle \widetilde{\H}^i &:= \H^i - \frac{1}{\gamma_1^i} ({\mathbf W}_r^i)^\top\left(\H^i {\mathbf W}_r^i- \mathbf N_r^i\right) \\
\textstyle \H^{i+1} &:= \widetilde{\H}^i-\frac{\lambda}{\lambda+\gamma_1^i}\left(\widetilde{\H}^i-\mathcal{P}_{{\rm{conv}}{(\mathbf N^i)}}(\tilde{\H}^i)\right) \\
\textstyle{\mathbf W}^{i+1} &:= \mathcal{P}_{\Delta}\left({\mathbf W}^i-\frac{1}{\gamma_2^i}\left(\H_c^{i+1} {\mathbf W}^i- \mathbf N_c^i\right)(\H_c^{i+1})^\top\right)\,.
\end{align*}
\begin{comment}
as follows:
\begin{equation*}
{\mathbf W}^i := \max\{\varepsilon,\, \mathbf N^i_c {\H^i_c}^\dagger\}\,, \quad \H^i := \max\{\varepsilon,\, {{\mathbf W}^{i+1}_r}^\dagger \mathbf N^i_r \}\,,
\end{equation*}
where $\dagger$ indicate the Moore-Penrose inverse.
\end{comment}
Same approach is exploited for Algorithm~\ref{algo:nmf3}.
\section{Latent Clustered Forecast}\label{sec:blackbox}
In this section we describe a second approach to time series forecasting problem, namely {\it Latent Clustered Forecast} (LCF) (see Algorithm~\ref{algo:forecasting}). Let
\[
\mathbf M^\star = {\mathbf W}^\star\H^\star = {\mathbf W}^\star\Big[\underbrace{\H^\star_{{T}}}_{\text{past}}\ \underbrace{\H^\star_{{F}}}_{\text{future}}\Big] = \Big[\underbrace{\mathbf M^\star_{{T}}}_{\text{past}}\ \underbrace{\mathbf M^\star_{{F}}}_{\text{future}}\Big] \,.
\]
Rather than directly forecasting the time series future data $\mathbf M_F^\star$, we firstly forecast the profiles $\H_F$ and we secondly reconstruct $\mathbf M_F^\star$. LCF initially cluster the $N$ time series in $I\ll N$ groups. For each cluster $i\in[I]$, matrix $(\mathbf M_T^\star)_i$ is defined by selecting the time series belonging to cluster and is factorized as ${\mathbf W}_i(\H_T)_i$. Then, $(\H_F)_i$ ($i\in [I]$) are forecasted by means of a black-box procedure, {\it e.g.}, random forest regression. Finally, the forecasted value are retrieved as $\mathbf M_F := {\mathbf W}_i (\H_F)_i$.
\begin{algorithm}[!htbp]
\caption{LCF for time series forecasting} \label{algo:forecasting}
\begin{algorithmic}[1]
\STATE factorize $\mathbf M_T^\star \simeq {\mathbf W}_0\H_0$ \label{step:factorize}
\STATE divide the rows of $\mathbf M^\star_T$ in $I$ clusters \label{step:forecast1}
\FOR{\textbf{each} $i\in[I]$}
\STATE define $(\mathbf M_T^\star)_i$ by selecting the time series in $i$ \label{ref:loop1}
\STATE factorize $(\mathbf M_T^\star)_i \simeq {\mathbf W}_i(\H_T)_i$ \label{step:forecast3}
\STATE forecast $(\H_F)_i$ \label{step:forecast2}
\ENDFOR
\STATE ${\mathbf M}_F := {\mathbf W}_0 {\H}_F$
\end{algorithmic}
\end{algorithm}
\vspace*{-0.2cm}
\begin{remark}
Algorithm~\ref{algo:forecasting} defines a class of algorithms, whose instances are defined by choosing clustering strategy at Steps~\ref{step:forecast1}, matrix factorization approach at Steps~\ref{step:factorize} and~\ref{step:forecast3}, and forecasting algorithm at Step~\ref{step:forecast2}. Algorithm~\ref{algo:forecasting} can be trivially parallelized by implementing multi-threading for loop at Steps~\ref{ref:loop1}-\ref{step:forecast2}.
\end{remark}
\noindent
We implemented two specific procedures for clustering and forecasting steps, detailed in next subsections.
\vspace*{-0.2cm}
\subsection{Clustering (Step~\ref{step:forecast1})}
We performs initially a (nonnegative) factorization of observed matrix $\mathbf M^\star_T\simeq {\mathbf W}_0\H_0$. The clustering of time series is based on weight matrix~${\mathbf W}_0$: the rationale of this strategy is considering time series as \invertedcommas{similar} if the corresponding linear combination of the columns of $(\H_T)_0$ has \invertedcommas{similar weights}. Hence, the full dendrogram of matrix ${\mathbf W}_0$ is computed via hierarchical clustering the rows of ${\mathbf W}_0$ with complete linkage and $\ell_1$ affinity, implemented in the Python {\tt sklearn 0.23} package under the routine {\tt AgglomerativeClustering(compute\_full\_tree=True)} \cite{scikit-learn}.
The dendrogram is explored from top to the bottom by looking at the children of each node (see Algorithm~\ref{algo:dendrogram}). The node is split into two children if and only if at least one of its children has size at least $d$. If it is not the case, the node is the leaf of the output tree. Algorithm~\ref{algo:dendrogram} is executed at the root node of the full dendrogram and returns the set of time series clusters.
Note that other clustering approaches might have been considered here. In practice, one need to pay attention that clusters should have maximal size $d$. The choice of the size $d$ is discussed for the numerical experiments of Section~\ref{sec:experiments}.
\begin{algorithm}[!htbp]
\caption{Exploring dendrogram} \label{algo:dendrogram}
\begin{algorithmic}[1]
\STATE \textbf{Input}: node $s$, cluster maximal dimension $d$
\STATE let child{\tiny \_}1 and child{\tiny \_}2 be the children of $s$
\IF{size of child{\tiny \_}1 $\le$ $d$ and size of child{\tiny \_}2 $\le$ $d$}
\STATE {\bf return} the set of all children of node $s$
\ELSE
\FORALL{children $c$ of $s$}
\IF{size of $c$ $\le$ $d$}
\STATE {\bf return} the set of all children of node $c$
\ELSE
\STATE recursively execute Algorithm~\ref{algo:dendrogram} on node $c$
\ENDIF
\ENDFOR
\ENDIF
\end{algorithmic}
\end{algorithm}
\vspace*{-0.3cm}
\subsection{Forecasting (Step~\ref{step:forecast2})}
In order to forecast $\H_F$ we apply deep learning approach. The samples are defined via {\it sliding window transformation}: the time series are subdivided in overlapping intervals of time window $D\ge 1$ and, in supervised learning fashion, the relationship between the values of each interval ({\it inputs}) and the consecutive~$F$ ({\it outputs}) values is learnt by mean of a regression algorithm, {\it e.g.}, random forest regression.
\section{Numerical Experiments}\label{sec:experiments}
The interested reader may find a github repository on the numerical experiments at \url{https://github.com/Luca-Mencarelli/Nonnegative-Matrix-Factorization-Time-Series}
\subsection{Real-world Data-sets}
The numerical experiments refer to the weekly and daily electricity consumption data-set of $N=370$ Portuguese customers during the period 2011-2014 ($T+F=208$ and $T+F=1456$ for weekly and daily measurements, respectively) \cite{trindade2016}. We chose the the last four weeks of 2014 as forecasting period ($F=4$ and $F=28$ for weekly and daily consumption, respectively).
We tests both SMM and LCF with a random initialization of matrices $\H^0, {\mathbf W}^0$. Each entry in $\H^0$ is randomly selected in $[0,h]$ where $h>0$ is chosen by practitioner. Each row of matrix ${\mathbf W}^0$ is randomly generated in the corresponding standard simplex.
For SMM we implement both HALS for \eqref{eq:nmf3} and PALM for \eqref{eq:amf1}. Moreover, we consider two different strategies to define matrix ${\boldsymbol \Pi}(\mathbf M)$: with {\it non-overlapping} and {\it overlapping} sliding intervals (for the overlapping case, we consider intervals with a common period of a week).
For LCF we consider different regression algorithms to forecast profiles $\H_F$, {\it e.g.}, {\it Long Short-Term Memory} (LSTM) and {\it Gated Recurrent Units} (GRU) deep neural networks with preliminary data standardization \cite{shewalkar2019} and {\it Autoregressive Integrated Moving Average} (ARIMA) models \cite{douc2014}: the best results in terms of forecasting quality and elapsed time are produced by {\it random forest regression}. Maximal dimension $d$ in Algorithm~\ref{algo:dendrogram} is set equal to nonnegative rank. For the matrix factorizations we apply PALM algorithm to Semi Normalized NMF (SNNMF)
Moreover, we have compared our method with other existing time series forecasting methods such as {\it Random Forest Regression} (RFR) and {\it EXponential Smoothing} (EXS).
The quality of the forecasted matrix $\mathbf M_F$ is measured by the relative root-mean-squared error (RRMSE) and the relative mean-percentage error (RMPE):
\begin{equation*}\begin{aligned}
\mbox{RRMSE} = \frac{\|\mathbf M_F-\mathbf M_F^\star\|_F}{\|\mathbf M_F^\star\|_F}\,,\;
\mbox{RMPE} = \frac{\|\mathbf M_F-\mathbf M_F^\star\|_1}{\|\mathbf M_F^\star\|_1}\,.
\end{aligned}\end{equation*}
We run all the test on a MacBook Pro mounting macOS Sierra with 2.6 GHz Intel Core i5 processor and 8 GB 1600 MHz DDR3 memory.
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Algorithm & RRMSE(K,W) & RMPE(K,W) \\
\midrule
SMM on mAMF & 5.24\%(4,5) & {\bf 6.60\%}(4,5) \\
SMM on mNMF & {\bf 5.15\%}(4,5) & 7.63\%(4,4) \\
LCF & 18.40\%(10,--) & 25.84\%(4,--) \\
RFR & 8.10\% & 7.98\% \\
\bottomrule
\end{tabular}
\caption{Indices for weekly consumption data-set.}
\label{tb:week}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{lll}
\toprule
Algorithm & RRMSE(K,W) & RMPE(K,W) \\
\midrule
SMM on mAMF & 14.35\%(4,5) & {\bf 34.50\%}(4,20) \\
SMM on mNMF & {\bf 14.17\%}(4,5) & 41.90\%(4,40) \\
LCF & 21.22\%(10,--) & 52.97\%(10,--) \\
RFR & 14.50\% & 47.72\%\\
\bottomrule
\end{tabular}
\caption{Indices for daily consumption data-set.}
\label{tb:day}
\end{table}
Tables~\ref{tb:week}-\ref{tb:day} report the cross-validated RRMSE and RMPE on observed values obtained during the computational tests for each methods for the weekly and daily consumption data-set, respectively. In parenthesis we detail the corresponding values of the eventual hyper-parameters K and W. SMM outperforms RFR, while performances of LCF are comparable with RFR.
\subsection{Synthetic Data-sets}
Further computational experiments have been realized by consider three synthetic data-sets. Each data-set has been generated by replicating 5000 small time series (with 10 time periods) 10 times and adding white noise multiplied by a constant factor $\sigma$ to each time series entry separately. We choose $\sigma\in\{0.005,0.1,1\}$. We refer to the three data-sets as \invertedcommas{low noise}, \invertedcommas{medium noise}, and \invertedcommas{high noise}.
We tested the SMM both for mAMF and mNMF, and LCF with random forest, exponential smoothing, SARIMAX model, and LSTM or GRU deep learning network as regression algorithm to extend the archetypes. We set a time limit of 2 hours for all the methods: SARIMAX, LSTM and GRU exceed time limit. We compare the proposed methodologies against the same regression methods applied to forecast each time series separately.
In Tables \ref{tb:low-noise}-\ref{tb:high-noise} we report the performance indexes for each method. RFR stands for random forest, while EXS for exponential smoothing. For the LCF methodology the best results is obtained with rank equal to 10 for the first matrix factorization and equal to 3 for the matrix factorization corresponding to the matrix factorization for each cluster. Moreover, we set to 600 the number of clusters, obtaining 1800 archetypes to forecast and significantly reducing the number of forecast time series.
For all the data-sets the SMM outperforms all the other methods in terms of performance indexes. As the noise increases, the LCF methodology produces a forecast with performance indexes comparable with the ones corresponding to the plain forecasting of each single time series.
The CPU time of SMM increases with the noise, but it significantly depend on the rank. For instance, in the medium noise data-set a forecast comparable with the best one (reported in Table \ref{tb:medium-noise}) is obtained for rank equal to 10 in a CPU time of 1386.74 seconds.
EXS is the fastest method but one of the less accurate with respect to reconstruction error. LCF with RFR is faster than the plain RFR by taking advantage from the smaller number of time series to extends. Unfortunately, LCF with EXS is slower that EXS since the small number of extended time series does not compensate the time devoted to performs the matrix factorization for each cluster.
\begin{table}[t]
\centering
\begin{tabular}{llll}
\toprule
Algorithm & RRMSE(K,W) & RMPE(K,W) & CPU (sec.) \\
\midrule
SMM on mAMF & {\bf 0.28\%(20,5)} & {\bf 0.27\%(20,5)} & 210.07 \\
SMM on mNMF & 0.28\%(20,5) & 0.27\%(20,5) & 867.83 \\
LCF with RFR & 36.32\% & 35.78\% & 393.58 \\
LCF with EXP & 49.47\% & 49.73\% & 86.39 \\
RFR & 23.89\% & 22.88\% & 953.67 \\
EXS & 48.44\% & 48.11\% & {\bf 34.39} \\
\bottomrule
\end{tabular}
\caption{Indices for low noise data-set.}
\label{tb:low-noise}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{llll}
\toprule
Algorithm & RRMSE(K,W) & RMPE(K,W) & CPU (sec.) \\
\midrule
SMM on mAMF & 0.79\%(20,5) & 0.83\%(20,5) & 5216.62 \\
SMM on mNMF & {\bf 0.59\%(20,5)} & {\bf 0.57\%(20,5)} & 5372.65 \\
LCF with RFR & 34.55\% & 34.57\% & 332.98 \\
LCF with EXS & 46.53\% & 46.02\% & 90.14 \\
RFR & 23.30\% & 21.75\% & 776.40 \\
EXP & 45.39\% & 44.10\% & {\bf 29.84} \\
\bottomrule
\end{tabular}
\caption{Indices for medium noise data-set.}
\label{tb:medium-noise}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{llll}
\toprule
Algorithm & RRMSE(K,W) & RMPE(K,W) & CPU (sec.) \\
\midrule
SMM on mAMF & {\bf 30.85\%(10,5)} & {\bf 29.86(10,5)} & 2120.79 \\
SMM on mNMF & 38.83\%(10,5) & 30.23\%(10,5) & 4576.86 \\
LCF with RFR & 35.40\% & 32.85\% & 427.90 \\
LCF with EXS & 37.51\% & 33.66\% & 191.94 \\
RFR & 33.84\% & 30.63\% & 772.89 \\
EXP & 37.13\% & 33.33\% & {\bf 28.86} \\
\bottomrule
\end{tabular}
\caption{Indices for high noise data-set.}
\label{tb:high-noise}
\end{table}
\section{Conclusions and Perspectives}\label{sec:conclusions}
In this paper, we have introduced and described two novel approaches for the time series forecasting problem relying on nonnegative matrix factorization. We apply these algorithms to a realistic data-sets, namely the daily and weekly Portuguese electricity consumption data-set, and synthetics data-sets, showing the forecasting capabilities of the proposed methodology.
Moreover, we have shown several uniqueness and robustness theoretical results for the solution of the matrix factorization problems faced by the two algorithms, namely the {\it Sliding Mask Method} and the {\it Latent Clustered Forecast}.
The strength of the proposed methodology consists in its relatively loose assumptions, mainly by supposing that time series matrix can be efficiently described by a low rank nonnegative decomposition, and that the time series are periodic for the {\it Sliding Mask Method}.
Future works consists in embedding side information in the forecasting procedure by extending algorithms in \cite{mei2019} to the {\it Sliding Mask Method} and the {\it Latent Clustered Forecast}.
|
\section{Introduction}
Modern neural networks are usually sensitive to small, adversarially chosen perturbations to the inputs \citep{DBLP:journals/corr/SzegedyZSBEGF13, biggio2013evasion}. Given an image ${\bm{x}}$ that is correctly classified by a neural network, a malicious attacker may find a small adversarial perturbation $\boldsymbol{\delta}$ such that the perturbed image ${\bm{x}}+\boldsymbol{\delta}$, though visually indistinguishable from the original image, is assigned to a wrong class with high confidence by the network. Such vulnerability creates security concerns in many real-world applications.
Developing a model that can resist small $\ell_\infty$ perturbations has been extensively studied in the literature. Adversarial training methods \citep{DBLP:journals/corr/SzegedyZSBEGF13,goodfellow2014, madry2017towards, pmlr-v97-zhang19p,ding2020mma} first generate on-the-fly adversarial examples of the inputs, then update model parameters using these perturbed samples together with the original labels. While such approaches can achieve decent empirical robustness, the evaluation is restricted to a particular (class of) attack method, and there are no formal guarantees whether the resulting model is robust against other attacks \citep{DBLP:journals/corr/abs-1802-00420,tramer2020adaptive,tjeng2018evaluating}.
Another line of algorithms train provably robust models for standard networks by maximizing the certified radius provided by robust certification methods, typically using linear relaxation \citep{wong2018provable,pmlr-v80-weng18a,mirman2018differentiable,zhang2018efficient,wang2018mixtrain,NEURIPS2018_f2f44698}, semidefinite relaxation \citep{raghunathan2018certified,pmlr-v115-dvijotham20a}, interval bound relaxation \citep{mirman2018differentiable,gowal2018effectiveness} or their combinations \citep{zhang2020towards}. However, most of these methods are sophisticated to implement and computationally expensive.
Besides these approaches, \citet{pmlr-v97-cohen19c,salman2019provably, zhai2020macer} study the certified guarantee on $\ell_2$ perturbations for Gaussian smoothed classifiers. However, recent works suggest that such methods are hard to extend to the $\ell_\infty$-perturbation scenario if the input dimension is large.
In this work, we propose a new approach by introducing a novel type of neural network that naturally resists local adversarial attacks, and can be easily certified under $\ell_\infty$ perturbation. In particular, we propose a novel neuron called $\ell_\infty$-dist neuron. Unlike the standard neuron design that uses a linear transformation followed by a non-linear activation, the $\ell_\infty$-dist neuron is purely based on computing the $\ell_\infty$-distance between the inputs and the parameters. It is straightforward to see that such a neuron is 1-Lipschitz with respect to $\ell_\infty$-norm, and the neural networks constructed with $\ell_\infty$-dist neurons (called $\ell_{\infty}$-dist nets) enjoy the same property. Based on such a property, we can efficiently obtain the certified robustness for any $\ell_{\infty}$-dist net using the margin of the prediction outputs.
Theoretically, we investigate the expressive power of $\ell_{\infty}$-dist nets and their robust generalization ability. We first prove a Lipschitz-universal approximation theorem which shows that $\ell_{\infty}$-dist nets can approximate any 1-Lipschitz function (with respect to $\ell_{\infty}$-norm) arbitrarily well. We then give upper bounds of robust test error, which would be small if the $\ell_{\infty}$-dist net learns a large margin classifier on the training data. These results demonstrate the excellent expressivity and generalization ability of the $\ell_{\infty}$-dist net function class.
While $\ell_{\infty}$-dist nets have nice theoretical guarantees, training such a network is still challenging. For example, the gradient of the parameters for $\ell_{\infty}$-norm distance is sparse, which makes the optimization difficult. In addition, we find that commonly used tricks and techniques in conventional network training cannot be taken for granted for this fundamentally different architecture. We address these challenges by proposing a holistic strategy for $\ell_{\infty}$-dist net training. Specifically, we show how to initialize the model parameters, apply proper normalization, design suitable weight decay mechanism, and overcome the sparse gradient problem via smoothed approximated gradients. Using the above methods, training an $\ell_\infty$-dist net is just as easy as training a standard network without any adversarial training, even though the resulting model is already provably robust.
Furthermore, the $\ell_{\infty}$-dist net has wide adaptability by serving as a robust feature extractor and combining itself with conventional networks for practical applications. After building a simple 2-layer perceptron on top of an $\ell_{\infty}$-dist net, we show that the model allows fast training and certification, and consistently achieves state-of-the-art certified robustness on a wide range of classification tasks.
Concretely, we reach \textbf{93.09\%} certified accuracy on MNIST under perturbation $\epsilon=0.3$, \textbf{79.23\%} on FashionMNIST under $\epsilon=0.1$, \textbf{35.42\%} on CIFAR-10 under $\epsilon=8/255$, and \textbf{16.31\%} on TinyImageNet under $\epsilon=1/255$.
As a comparison, these results outperform the previous best-known results \citep{xu2020automatic}, in which they achieve 33.38\% certified accuracy on CIFAR-10 dataset, and achieve 15.86\% certified accuracy on TinyImageNet using a WideResNet model which is \textit{33 times larger} than the $\ell_\infty$-dist net.
Our contributions are summarized as follows:
\begin{itemize}[topsep=0pt]
\setlength{\itemsep}{0pt}
\item We propose a novel neural network using $\ell_{\infty}$-dist neurons, called $\ell_{\infty}$-dist net. We show that any $\ell_{\infty}$-dist net is 1-Lipschitz with respect to $\ell_\infty$-norm, which directly guarantees the certified robustness (Section 3).
\item In the theoretical part, we prove that $\ell_{\infty}$-dist nets can approximate any 1-Lipschitz function with respect to $\ell_\infty$-norm. We also prove that $\ell_{\infty}$-dist nets have a good robust generalization ability (Section 4).
\item In the algorithmic part, we provide a holistic training strategy for $\ell_{\infty}$-dist nets, including parameter initialization, normalization, weight decay and smoothed approximated gradients (Section 5).
\item We show how to combine $\ell_{\infty}$-dist nets with standard networks and obtain robust models more effectively (Section 6). Experimental results show that we can consistently achieve state-of-the-art certified accuracy on MNIST, Fashion-MNIST, CIFAR-10 and TinyImageNet dataset (Section 7).
\item Finally, we provide all the implementation details and codes at \href{https://github.com/zbh2047/L\_inf-dist-net}{https://github.com/zbh2047/L\_inf-dist-net}.
\end{itemize}
\section{Related Work}
\label{sec_related_work}
\paragraph{Robust Training Approaches.}
Adversarial training is the most successful method against adversarial attacks. By adding adversarial examples to the training set on the fly, adversarial training methods
\citep{DBLP:journals/corr/SzegedyZSBEGF13,goodfellow2014,madry2017towards, huang2015learning,pmlr-v97-zhang19p, wong2020fast,ding2020mma} can significantly improve the robustness of conventional neural networks. However, all the methods above are evaluated according to the empirical robust accuracy against pre-defined adversarial attack algorithms, such as projected gradient decent. These methods cannot guarantee whether the resulting model is also robust against other attacks.
\paragraph{Certified Robustness for Conventional Networks.}
Many recent works focus on certifying the robustness of learned neural networks under \emph{any} attack.
These approaches are mainly based on bounding the certified radius layer by layer using some convex relaxation methods \citep{wong2018provable,NIPS2018_8060,pmlr-v80-weng18a,mirman2018differentiable,dvijotham2018dual,zhang2018efficient,wang2018mixtrain,NEURIPS2018_f2f44698,xiao2018training,Balunovic2020Adversarial,raghunathan2018certified,pmlr-v115-dvijotham20a}. However, such approaches are usually complicated, computationally expensive and have difficulties in applying to deep and large models. To overcome these drawbacks, \citet{mirman2018differentiable,gowal2018effectiveness} considered interval bound propagation (IBP), a special form of convex relaxation which is much simpler and computationally cheaper. However, the produced bound is loose which results in unstable
training. \citet{zhang2020towards,xu2020automatic} took a further step to combine IBP with linear relaxation to make the bound tighter, which achieves current state-of-the-art performance. Fundamentally different from all these approaches that target to certify conventional networks, we proposed a novel network that provides robustness guarantee by its nature.
\paragraph{Certified Robustness for Smoothed Classifiers.} Randomized smoothing can provide a (probabilistic) certified robustness guarantee for general models. \citet{lecuyer2018certified, DBLP:journals/corr/abs-1809-03113, pmlr-v97-cohen19c, salman2019provably, zhai2020macer,zhang2020black} showed that if a Gaussian random noise is added to the input, a certified guarantee on small $\ell_2$ perturbation can be computed for Gaussian smoothed classifiers. However, \citet{yang2020randomized, blum2020random, kumar2020curse} showed that randomized smoothing cannot achieve nontrivial certified accuracy against larger than $\Omega\rbr{\min\rbr{1, d^{1/p-1/2}}}$ radius for $\ell_{p}$ perturbations, where $d$ is the input dimension. Therefore it cannot provide meaningful results for a relatively large $\ell_{\infty}$ perturbation due to the curse of dimensionality.
\paragraph{Lipschitz Networks.} Another line of approaches sought to bound the global Lipschitz constant of the neural network. Lipschitz networks can be very useful in certifying adversarial robustness \citep{tsuzuku2018lipschitz}, proving generalization bounds \citep{7934087}, or estimating Wasserstein distance \citep{pmlr-v70-arjovsky17a}. Previous works trained Lipschitz ReLU networks by either directly constraining the spectral norm of each weight matrix to be less than one, or optimizing a loss constructed using the global Lipschitz constant which is upper bounded by these spectral norms \citep{cisse2017parseval,yoshida2017spectral,gouk2018regularisation,tsuzuku2018lipschitz,qian2018lnonexpansive}. However, as pointed out by \citet{huster2018limitations} and \citet{anil2019sorting}, such Lipschitz networks lack expressivity to some simple Lipschitz functions and the global Lipschitz bound is not tight. Recently, \citet{anil2019sorting} proposed a new Lipschitz network that is a Lipschitz-universal approximator. \citep{NEURIPS2019_1ce3e6e3} extended their work to convolutional architectures. However, the robustness performances are still not as good as other certification methods, and none of these methods can provide good certified results for $\ell_\infty$ robustness. In this
work, we show the proposed $\ell_\infty$-dist net is also a Lipschitz-universal approximator (under $\ell_\infty$-norm), and the Lipschitz $\ell_\infty$-dist net can substantially outperform other Lipschitz networks in term of certified accuracy.
\paragraph{$\ell_p$-dist Neurons.} We notice that recently \citet{chen2020addernet,xu2020kernel} proposed a new network called AdderNet, which leverages $\ell_1$-norm operation to build the network for sake of efficient inference. \citet{wang2019kervolutional} also considered replacing dot-product neurons by $\ell_1$-dist or $\ell_2$-dist neurons in order to enhance the model's non-linearity and expressivity. Although $\ell_\infty$-dist net looks similar to these networks on the surface, they are designed for fundamentally different problems, and in fact $\ell_1$-dist neurons (or $\ell_2$-dist neurons) can not give any robust guarantee for norm-bounded perturbations (see Appendix \ref{sec_addernet} for more discussions).
\section{$\ell_\infty$-dist Network and its Robustness Guarantee}
\label{sec_property}
\subsection{Preliminaries}
\label{sec:bound}
Consider a standard classification task. Suppose we have an underlying data distribution $\mathcal{D}$ over pairs of examples ${\bm{x}} \in {\mathcal{X}}$ and corresponding labels $y\in {\mathcal{Y}} =\{1,2,\cdots,M\}$ where $M$ is the number of classes. Usually $\mathcal{D}$ is unknown and we can only access a training set ${\mathcal{T}}=\{({\bm{x}}_1,y_1),\cdots,({\bm{x}}_n,y_n)\}$ in which $({\bm{x}}_i,y_i)$ is \textit{i.i.d.} drawn from $\mathcal D$. Let $f\in\mathcal{F}$ be the classifier of interest that maps any ${\bm{x}}\in{\mathcal{X}}$ to ${\mathcal{Y}}$. We call ${\bm{x}}'={\bm{x}}+\boldsymbol{\delta}$ an \textit{adversarial example} of ${\bm{x}}$ to classifier $f$ if $f$ can correctly classify ${\bm{x}}$ but assigns a different label to ${\bm{x}}'$. In real practice, the most commonly used setting is to consider the attack under $\epsilon$-bounded $\ell_{\infty}$-norm constraint, i.e., $\boldsymbol{\delta}$ satisfies $\|\boldsymbol{\delta}\|_{\infty} \leq {\epsilon}$, which is also called $\ell_\infty$ perturbations.
Our goal is to learn a model from ${\mathcal{T}}$ that can resist attacks at $({\bm{x}},y)$ over $({\bm{x}},y)\sim \mathcal{D}$ for any small $\ell_\infty$ perturbation. It relates to compute the radius of the largest $\ell_{\infty}$ ball centered at ${\bm{x}}$ in which $f$ does not change its prediction. This radius is called the \textit{robust radius}, which is defined as \citep{zhai2020macer,pmlr-v97-zhang19p}:
\begin{equation}
\label{eqn:def_distance}
R(f;{\bm{x}},y)= \left\{
\begin{aligned}
\inf_{f({\bm{x}}') \neq f({\bm{x}})} \left\| {\bm{x}}'-{\bm{x}} \right\|_{\infty}, & \quad f({\bm{x}})=y \\
0 \quad\quad\quad\quad, & \quad f({\bm{x}}) \neq y
\end{aligned}
\right .
\end{equation}
Unfortunately, exactly computing the robust radius of a classifier induced by a standard deep neural network is very difficult. For example, \citet{katz2017reluplex} showed that calculating such radius for a DNN with ReLU activation is NP-hard. Researchers then seek to derive a tight \textit{lower bound} of $R(f;{\bm{x}},y)$ for general $f$. Such lower bound is called \textit{certified radius} and we denote it as $CR(f;{\bm{x}},y)$. It follows that $CR(f;{\bm{x}},y) \leq R(f;{\bm{x}},y)$ for any $f, {\bm{x}}, y$.
\subsection{Networks with $\ell_\infty$-dist Neurons}
In this subsection, we propose a novel neuron called the $\ell_\infty$-dist neuron, which is inherently robust with respect to $\ell_\infty$-norm perturbations. Using these neurons as building blocks, we then show how to obtain robust neural networks dubbed $\ell_\infty$-dist nets.
\begin{figure}
\centering
\vspace{0pt}
\includegraphics[width=0.48\textwidth]{figure.pdf}
\vspace{-20pt}
\caption{Illustration of the conventional neuron (left) and the $\ell_\infty$-dist neuron (right).}
\label{fig:neuron}
\vspace{-10pt}
\end{figure}
Denote ${\bm{x}}$ as the input vector to a neuron. A standard neuron processes the input by first projecting ${\bm{x}}$ to a scalar value using a linear transformation, then applying a non-linear activation function $\sigma$ on it, i.e., $\sigma({\bm{w}}^{\top}{\bm{x}} +b)$ with ${\bm{w}}$ and $b$ as parameters and function $\sigma$ being sigmoid or ReLU activation.
Unlike the previous design paradigm, we introduce a new type of neuron using $\ell_\infty$ distance as the basic operation, called $\ell_\infty$-dist neuron:
\begin{equation}\label{eq:acti}
u({\bm{x}},\theta)=\|{\bm{x}}-{\bm{w}}\|_{\infty}+b,
\end{equation}
where $\theta = \{{\bm{w}},b\}$ is the parameter set (see Figure \ref{fig:neuron} for an illustration). From Eqn.~\ref{eq:acti} we can see that the $\ell_\infty$-dist neuron is non-linear as it calculates the $\ell_\infty$-distance between input ${\bm{x}}$ and parameter ${\bm{w}}$ with a bias term $b$. As a result, there is no need to further apply a non-linear activation function.
\begin{remark}
\label{remark_similar}
Conventional neurons use dot-product to represent the similarity between input ${\bm{x}}$ and weight ${\bm{w}}$. Likewise, $\ell_\infty$-distance is also a similarity measure. Note that $\ell_\infty$-distance is always non-negative, and a smaller $\ell_\infty$-distance indicates a stronger similarity.
\end{remark}
Without loss of generality, we study the properties of multi-layer perceptron (MLP) networks constructed using $\ell_\infty$-dist neurons. All theoretical results can be easily extended to other neural network architectures, such as convolutional networks. We use ${\bm{x}}\in \mathbb{R}^{d_{\mathrm{input}}}$ to denote the input vector of an MLP network. An MLP network using $\ell_\infty$-dist neurons can be formally defined as follows.
\begin{definition} ($\ell_\infty$-dist Net)
\label{def_ell_inf_net}
Define an $L$ layer $\ell_\infty$-dist net as follows. Assume the $l$-th hidden layer contains $d_l$ hidden units. The network takes ${\bm{x}}^{(0)} \triangleq {\bm{x}} \in \mathbb{R}^{d_{\mathrm{input}}}$ as input, and the $k$-th unit in the $l$-th hidden layer $x^{(l)}_k$ is computed by
\begin{equation}
\begin{aligned}
x^{(l)}_k=u({\bm{x}}^{(l-1)}, \theta^{(l,k)})=\|{\bm{x}}^{(l-1)}-{\bm{w}}^{(l,k)}\|_{\infty}+b^{(l,k)},\\
1\le l\le L, 1\le k \le d_l
\end{aligned}
\end{equation}
where ${\bm{x}}^{(l)}=(x^{(l)}_1,x^{(l)}_2,\cdots,x^{(l)}_{d_{l}})$ is the output of the $l$-th layer.
\end{definition}
For classification tasks, the dimension of the final outputs of an $\ell_\infty$-dist net matches the number of categories, i.e., $M$. Based on Remark \ref{remark_similar}, we use the negative of the final layer to be the outputs of the $\ell_\infty$-dist net ${\bm{g}}$, i.e. ${\bm{g}}({\bm{x}})=(-x_1^{(L)}, -x_2^{(L)}, \cdots, -x_M^{(L)})$ and define the predictor $f({\bm{x}}) = \argmax_{i \in [M]} g_i({\bm{x}})$. Similar to conventional networks, we can apply any standard loss function on the $\ell_\infty$-dist net, such as the cross-entropy loss or hinge loss.
\subsection{Lipschitz and Robustness Facts about $\ell_{\infty}$-dist Nets}
In this subsection, we will show that the $\ell_\infty$-dist neurons and the neural networks constructed using them have nice theoretical properties in controlling the robustness of the model. We first show that $\ell_{\infty}$-dist nets are 1-Lipschitz with respect to $\ell_\infty$-norm, then derive the certified robustness of the model based on such property.
\begin{definition} (Lipschitz Function)
A function ${\bm{g}}({\bm{z}}): \mathbb{R}^m \rightarrow \mathbb{R}^n$ is called $\lambda$-Lipschitz with respect to $\ell_p$-norm $\|\cdot \|_p$, if for any ${\bm{z}}_1, {\bm{z}}_2$, the following holds:
$$\|{\bm{g}}({\bm{z}}_1)-{\bm{g}}({\bm{z}}_2)\|_p\le \lambda \|{\bm{z}}_1-{\bm{z}}_2\|_p$$
\end{definition}
\begin{fact}\label{thm:lip}
Any $\ell_\infty$-dist net ${\bm{g}}(\cdot)$ is 1-Lipschitz with respect to $\ell_\infty$-norm, i.e., for any ${\bm{x}}_1, {\bm{x}}_2\in \mathbb{R}^{d_{\mathrm{input}}}$, we have
$\|{\bm{g}}({\bm{x}}_1)-{\bm{g}}({\bm{x}}_2)\|_{\infty}\le \|{\bm{x}}_1-{\bm{x}}_2\|_{\infty}$.
\end{fact}
\begin{proof}
It's easy to check that every basic operation $u({\bm{x}}^{(l-1)}, \theta^{(l,k)})$ is 1-Lipschitz, and therefore the mapping from one layer to the next ${\bm{x}}^{(l)}\to {\bm{x}}^{(l+1)}$ is 1-Lipschitz. Finally by composition we have for any ${\bm{x}}_1, {\bm{x}}_2\in \mathbb{R}^{d_{input}}$, $\|{\bm{g}}({\bm{x}}_1)-{\bm{g}}({\bm{x}}_2)\|_{\infty}\le \|{\bm{x}}_1-{\bm{x}}_2\|_{\infty}$.
\end{proof}
\begin{remark}
\label{remark_lip}
Fact \ref{thm:lip} is only true when the basic neuron uses \emph{infinity-norm} distance. For a network constructed using $\ell_p$-dist neurons where $p<\infty$ (e.g. 2-norm), such network will not be 1-Lipschitz (even with respect to $\ell_p$-norm) because the mapping from one layer to the next ${\bm{x}}^{(l)}\to {\bm{x}}^{(l+1)}$ is not 1-Lipschitz.
\end{remark}
Since ${\bm{g}}$ is 1-Lipschitz with respect to $\ell_\infty$-norm, if the perturbation over ${\bm{x}}$ is rather small, the change of the output can be bounded and the prediction of the perturbed data ${\bm{x}}'$ will not change as long as $\argmax_{i \in [M]} g_i({\bm{x}}) = \argmax_{i \in [M]} g_i({\bm{x}}')$, which directly bounds the certified radius.
\begin{fact}\label{thm:mar}
Given model $f({\bm{x}})=\argmax_{i\in[M]}g_i({\bm{x}})$ defined above, and ${\bm{x}}$ being correctly classified, we define $\mathrm{margin}({\bm{x}};{\bm{g}})$ as the difference between the largest and second-largest elements of ${\bm{g}}({\bm{x}})$. Then for any ${\bm{x}}'$ satisfying $\|{\bm{x}}-{\bm{x}}'\|_{\infty}< \mathrm{margin}({\bm{x}};{\bm{g}})/2$, we have that $f({\bm{x}})=f({\bm{x}}')$. In other words,
\begin{equation}\label{eq:cor1}
CR(f,{\bm{x}},y)\ge \mathrm{margin}({\bm{x}};{\bm{g}})/2
\end{equation}
\end{fact}
\begin{proof}
Since ${\bm{g}}({\bm{x}})$ is 1-Lipschitz, each element of ${\bm{g}}({\bm{x}})$ can move at most $\mathrm{margin}({\bm{x}};{\bm{g}})/2$ when ${\bm{x}}$ changes to ${\bm{x}}'$, therefore the largest element will remain the same.
\end{proof}
Using this bound, we can certify the robustness of an $\ell_{\infty}$-dist net \emph{of any size} under $\ell_{\infty}$-norm perturbations with little computational cost (\emph{only a forward pass}). In contrast, existing certified methods may suffer from either poor scalability (methods based on linear relaxation) or curse of dimensionality (randomized smoothing).
\section{Theoretical Properties of $\ell_{\infty}$-dist Nets}
The expressive power of a model family and its generalization are two central topics in machine learning. Since we have shown that $\ell_{\infty}$-dist nets are 1-Lipschitz with respect to $\ell_\infty$-norm, it's natural to ask whether $\ell_{\infty}$-dist nets can approximate \emph{any} 1-Lipschitz function (with respect to $\ell_\infty$-norm) and whether we can give generalization guarantee on the robust test error based on the Lipschitz property. In this section, we give affirmative answers to both questions. Without loss of generality, we consider \emph{binary classification} problems and assume the output dimension is 1. All the omitted proofs in this section can be found in Appendix \ref{sec_proof}.
\subsection{Lipschitz-Universal Approximation of $\ell_{\infty}$-dist Nets}
It is well-known that the conventional network is a universal approximator, in that it can approximate any continuous function arbitrarily well \citep{cybenko1989approximation}.
Similarly, in this section we will prove a Lipschitz-universal approximation theorem for $\ell_{\infty}$-dist nets, formalized in the following:
\begin{theorem}\label{thm:inflayer}
For any 1-Lipschitz function $\tilde{g}({\bm{x}})$ (with respect to $\ell_{\infty}$-norm) on a bounded domain ${\mathbb{K}}\in \mathbb R^{d_{\text{input}}}$ and any $\epsilon>0$, there exists an $\ell_{\infty}$-dist net $g({\bm{x}})$ with width no more than $d_{input}+2$, such that for all ${\bm{x}}\in {\mathbb{K}}$, we have $\|g({\bm{x}})-\tilde{g}({\bm{x}})\|_{\infty} \le \epsilon$.
\end{theorem}
We briefly present a proof sketch of Theorem $\ref{thm:inflayer}$. The proof has the same structure to \citet{NIPS2017_7203}, who first proved such universal approximation theorem for width-bounded ReLU networks, by constructing a special network that approximates a target function $\tilde{g}({\bm{x}})$ by the sum of indicator functions of grid points, i.e. $\tilde{g}({\bm{x}})\approx\sum_{{\bm{z}}\in \mathbb S} \tilde{g}({\bm{z}})\mathbf 1_{\{{\bm{x}}:\|{\bm{x}}-{\bm{z}}\|_{\infty}\le \epsilon/2\}}({\bm{x}})$ where $\mathbb S=\{{\bm{z}}\in \mathbb K:z_i=C_i\epsilon,C_i\in \mathbb Z\}$ . For $\ell_{\infty}$-dist nets, such an approach cannot be directly applied as the summation will break the Lipschitz property. We employ a novel ``max of pyramids'' construction to overcome the issue. The key idea is to approximate the target function using the maximum of many ``pyramid-like'' basic 1-Lipschitz functions, i.e. $\tilde{g}({\bm{x}})\approx g(x):=\max_{{\bm{z}}\in \mathbb S} (\tilde{g}({\bm{z}})-\|{\bm{x}}-{\bm{z}}\|_{\infty})$. To represent such a function, we first show that the $\ell_\infty$-dist neuron can express the following basic functions: $f({\bm{x}})=\|{\bm{x}}-{\bm{z}}\|_{\infty}$, $f({\bm{x}})=x_i+b$, $f({\bm{x}})=-x_i+b$ and $f({\bm{x}})=\max(x_i,x_j)$; Then $g(x)$ can be constructed using these functions as building blocks. Finally, we carefully design a computation pattern for a width-bounded $\ell_\infty$-dist net to perform such max-reduction layer by layer.
Theorem \ref{thm:inflayer} implies that an $\ell_{\infty}$-dist net can approximate any 1-Lipschitz function with respect to $\ell_{\infty}$-norm on a compact set, using width barely larger than the input dimension. Combining it with Fact \ref{thm:lip}, we conclude that $\ell_{\infty}$-dist nets are a good class of models to approximate 1-Lipschitz functions.
\subsection{Bounding Robust Test Error of $\ell_{\infty}$-dist Nets}
In this subsection, we give a generalization bound for the \emph{robust test error} of $\ell_{\infty}$-dist nets. Let $({\bm{x}}, y)$ be an instance-label couple where ${\bm{x}}\in {\mathbb{K}}$ and $y\in \{1,-1\}$ and denote $\mathcal{D}$ as the distribution of $({\bm{x}}, y)$. For a function $g({\bm{x}}): \mathbb{R}^{d_{\mathrm{input}}} \rightarrow \mathbb{R}$, we use $\sign(g({\bm{x}}))$ as the classifier. The $r$-robust test error $\gamma_r$ of a classifier $g$ is defined as
$$\gamma_r=\mathbb E_{(x,y)\sim\mathcal{ D}}\sbr{\sup_{\|{\bm{x}}'-{\bm{x}}\|_{\infty}\le r}\mathbb I_{y g({\bm{x}}')\le 0}}.$$
Then $\gamma_r$ can be upper bounded by the margin error on training data and the size of the network, as stated in the following theorem:
\begin{theorem}\label{thm:rob}
Let ${\mathbb{F}}$ denote the set of all $g$ represented by an $\ell_{\infty}$-dist net with width $W(W\ge d_{\text{input}})$ and depth $L$. For every $t>0$, with probability at least $1-2e^{-2t^2}$ over the random drawing of $n$ samples, for all $ r>0$ and $g\in {\mathbb{F}}$ we have that
\begin{equation}\label{eq:lipgene}
\begin{aligned}
\gamma_r \le \inf_{\delta\in (0,1]} \left[\frac{1}{n}\sum_{i=1}^n \right.\underbrace{\mathbb I_{y_i g({\bm{x}}_i)\le \delta+r}}_{\text{large training margin}}+\underbrace{\tilde{O}\rbr{\frac{LW^2}{\delta \sqrt{n}}}}_{\text{network size}}\\
\left.+\rbr{\frac{\log\log_2(\frac{2}{\delta})}{n}}^{\frac{1}{2}}\right]+\frac{t}{\sqrt{n}}.
\end{aligned}
\end{equation}
\end{theorem}
Theorem \ref{thm:rob} demonstrates that when a large margin classifier is found on training data, and the size of the $\ell_{\infty}$-dist net is not too large, then with high probability, the model can generalize well in terms of \textit{adversarial robustness}.
Note that our bound does not depend on the input dimension, while previous generalization bounds for the general Lipschitz model class (i.e. \citet{luxburg2004distance}) suffer from the curse of dimensionality \cite{neyshabur2017exploring}.
\section{Training $\ell_{\infty}$-dist Nets}
\label{sec_optimize}
In this section we will focus on how to train an $\ell_{\infty}$-dist net successfully. Motivated by the theoretical analysis in previous sections, it
suffices to find a large margin solution on the training data. Therefore we can simply use the standard multi-class hinge loss to obtain a large training margin, similar to \citet{anil2019sorting}. Since the loss is differentiable almost everywhere, any gradient-based optimization method can be used to train $\ell_{\infty}$-dist nets.
However, we empirically find that the optimization is challenging and directly training the network usually \textit{fails} to obtain a good performance. Moreover, conventional wisdom like batch normalization cannot be taken as a grant in the $\ell_{\infty}$-dist net since it will hurt the model's robustness. In the following we will dig into the optimization difficulties and provide a holistic training strategy to overcome them.
\subsection{Normalization}
\label{sec_bn}
One important difference between $\ell_\infty$-dist nets and conventional networks is that for conventional networks, the output of a linear layer is unbiased (zero mean in expectation) under random initialization over parameters \citep{glorot2010understanding,he2015delving}, while the output of an $\ell_\infty$-dist neuron is biased (always being non-negative, assuming no bias term). Indeed, for a weight vector initialized using a standard Gaussian distribution, assume a zero input ${\bm{x}}^{(0)}=\mathbf 0\in \mathbb R^d$ is fed into an $\ell_\infty$-dist net, the expected output of the first layer can be approximated by $x_j^{(1)}=\mathbb E_{{\bm{w}}^{(1,j)}}\|{\bm{x}}^{(0)}-{\bm{w}}^{(1,j)}\|_\infty\approx \sqrt{2\log d}$. The output vector ${\bm{x}}^{(1)}$ is then fed into subsequent layers, making the outputs in upper layers linearly increase.
Normalization is a useful way to control the scale of the layer's outputs to a standard range. Batch Normalization \citep{ioffe2015batch}, which shifts and scales feature values in each layer, is shown to be one of the most important components in training deep neural networks. However, if we directly apply batch normalization in $\ell_{\infty}$-dist nets, the Lipschitz constant will change due to the scaling operation, and the robustness of the model cannot be guaranteed.
Fortunately, we find using the shift operation alone already helps the optimization. Therefore we apply the shift operation in all intermediate layers after calculating the $\ell_{\infty}$ distance. As a result, we remove the bias terms in the corresponding $\ell_{\infty}$-dist neurons as they are redundant. We do not use normalization in the final layer. Similar to BatchNorm, we use the running mean during inference, which serves as additional bias terms in $\ell_{\infty}$-dist neurons and does not affect the Lipschitz constant of the model. We do not use affine transformation, which is typically used in BatchNorm.
\subsection{Smoothed Approximated Gradients}
\label{sec_app_grad}
We find that training an $\ell_{\infty}$-dist net from scratch is usually inefficient, and the optimization can easily be stuck at some bad solution. One important reason is that the gradients of the $\ell_\infty$-dist operation (i.e., $\nabla_{{\bm{w}}}\|{\bm{z}}-{\bm{w}}\|_{\infty}$ and $\nabla_{{\bm{z}}}\|{\bm{z}}-{\bm{w}}\|_{\infty}$) are very sparse which typically contain only one non-zero element. In practice, we observe that there are less than 1\% parameters updated in an epoch if we directly train the $\ell_{\infty}$-dist net using SGD/Adam from random initialization.
To improve the optimization, we relax the $\ell_{\infty}$-dist neuron by using the $\ell_{p}$-dist neuron for the whole network to get an approximate and non-sparse gradient of the model parameters. During training, we set $p$ to be a small value in the beginning and increase it in each iteration until it approaches infinity. For the last few epochs, we set $p$ to infinity and train the model to the end. Empirically, using smoothed approximated gradients significantly boosts the performance, as will be shown in Section \ref{sec_ablation}.
\subsection{Parameter Initialization}
We find that there are still optimization difficulties in training \emph{deep} models using normalization and the smoothed gradient method. In particular, we find that a deeper model performs worse than its shallow counterpart in term of \emph{training accuracy} (see Appendix \ref{sec_initialization}), a phenomenon similar to \citet{he2016deep}. To fix the problem, \citet{he2016deep} proposed ResNet architecture by modifying the network using identity mapping as skip connections. According to Proposition \ref{prop:3p} (see Appendix \ref{sec_proof}), an $\ell_\infty$-dist layer can also perform identity mapping by assigning proper weights and biases at initialization, and a deeper $\ell_\infty$-dist net then can act as a shallow one by these identity mappings.
Given such findings, we can directly construct identity mappings at initialization.
Concretely, for an $\ell_\infty$-dist layer with the same input-output dimension, we first initialize the weights randomly from a standard Gaussian distribution as common, then modify the diagonal elements (i.e. $w_j^{(l,j)}$ in Definition \ref{def_ell_inf_net}) to be a large negative number $C_0$. Throughout all experiments, we set $C_0=-10$. We do not need the bias in $\ell_\infty$-dist neurons after applying mean shift normalization, and the running mean automatically makes an identity mapping.
\subsection{Weight Decay}
Weight decay is a commonly used trick in training deep neural networks. It is equivalent to adding an $\ell_2$ regularization term to the loss function. However, we empirically found that using weight decay in the $\ell_\infty$-dist net gives inferior performance (see Section \ref{sec_ablation}). The problem might be the incompatibility of weight decay ($\ell_2$ regularization) with $\ell_\infty$-norm used in $\ell_\infty$-dist nets, as we will explain below.
Conventional networks use dot-product as the basic operation, therefore an $\ell_2$-norm constraint on the weight vectors directly controls the scale of the output magnitude. However, it is straightforward to see that the $\ell_2$-norm of the weight vector does not correspond to the output scale of the $\ell_\infty$ distance operation. A more reasonable choice is to use $\ell_\infty$-norm regularization instead of $\ell_2$-norm. In fact, we have $\|{\bm{x}}-{\bm{w}}\|_\infty\le \|{\bm{x}}\|_\infty+\|{\bm{w}}\|_\infty$, analogous to $\langle {\bm{x}},{\bm{w}}\rangle \le \|{\bm{x}}\|_2\|{\bm{w}}\|_2$.
For general $\ell_p$-dist neurons during training, we can use $\ell_p$-norm regularization analogously. By taking derivative with respect to the weight ${\bm{w}}$, we derive the corresponding weight decay formula:
\vspace{-8pt}
\begin{equation}
\label{eq:weight_decay}
\Delta_{w_i}=-\lambda \nabla_{w_i} \|{\bm{w}}\|_p^2=-\lambda \left(\frac {|w_i|}{\|{\bm{w}}\|_p}\right)^{p-2} w_i
\end{equation}
where $\lambda$ is the weight decay coefficient. Note that Eqn.~\ref{eq:weight_decay} reduces to commonly used weight decay if $p=2$. When $p\rightarrow \infty$, the weight decay tends to take effects only on the element $w_i$ with the largest absolute value.
\section{Certified Robustness by Using $\ell_\infty$-dist Nets as Robust Feature Extractors}
\label{sec_combine}
We have shown that the $\ell_\infty$-dist net is globally Lipschitz in the sense that the function is 1-Lipschitz everywhere over the input space. This constraint is pretty strong when we only require function's Lipschitzness on a specific manifold, such as real image data manifold. To make the model more flexible and fit the practical tasks better, we can build a lightweight conventional network on top of an $\ell_\infty$-dist net. The $\ell_\infty$-dist net ${\bm{g}}$ will serve as a robust feature extractor, and the lightweight network ${\bm{h}}$ (a shallow MLP in our experiments) will focus on task-specific goals, such as classification. We denote the composite network as ${\bm{h}}\circ{\bm{g}}$.
We first show how to certify the robustness for the composite network ${\bm{h}}\circ{\bm{g}}$. Given any input ${\bm{x}}$, consider a perturbation set $\mathcal B_{\infty}^{\epsilon}({\bm{x}})=\{{\bm{x}}':\|{\bm{x}}'-{\bm{x}}\|_\infty\le \epsilon\}$, where $\epsilon$ is the pre-defined perturbation level. If $({\bm{h}}\circ{\bm{g}})({\bm{x}}')$ predicts the correct label $y$ for all ${\bm{x}}'\in \mathcal B_{\infty}^{\epsilon}({\bm{x}})$, we can guarantee robustness of the network ${\bm{h}}\circ{\bm{g}}$ for input ${\bm{x}}$. Since ${\bm{g}}$ is 1-Lipschitz with respect to $\ell_\infty$-norm, we have ${\bm{g}}({\bm{x}}')\in \mathcal B_{\infty}^{\epsilon}({\bm{g}}({\bm{x}}))$ by definition. Based on this property, to guarantee robustness of the network ${\bm{h}}\circ{\bm{g}}$ for input ${\bm{x}}$, it suffices to check whether for all ${\bm{z}}'\in \mathcal B_{\infty}^{\epsilon}({\bm{g}}({\bm{x}}))$, ${\bm{h}}({\bm{z}}')$ predicts the correct label $y$. This is equivalent to certifying the robustness for a conventional network ${\bm{h}}$ given input ${\bm{z}}={\bm{g}}({\bm{x}})$, which can be calculated using any previous certification method such as convex relaxation.
We describe one of the simplest convex relaxation method named IBP \citep{gowal2018effectiveness} in Appendix \ref{sec_ibp}, which is used in our experiments. Other more advanced approaches such as CROWN-IBP \citep{zhang2020towards,xu2020automatic} can also be considered.
After obtaining the bound, we can set it as the training objective function to train the neural network parameters, similar to \citet{gowal2018effectiveness,zhang2020towards}. All calculations are differentiable, and gradient-based optimization methods can be applied. Note that since ${\bm{h}}$ and ${\bm{g}}$ have entirely different architectures, we apply the training strategy introduced in Section \ref{sec_optimize} to the $\ell_\infty$-dist net ${\bm{g}}$ only. More details will be presented in Section \ref{sec_experiment_setting}.
\section{Experiments \& Results}
\label{sec_experiment}
In this section, we conduct extensive experiments for the proposed network. We train our models on four popular benchmark datasets: MNIST, Fashion-MNIST, CIFAR-10 and TinyImagenet.
\subsection{Experimental Setting}
\label{sec_experiment_setting}
\paragraph{Model details.} We mainly study two types of models. The first type is denoted as $\ell_\infty$-dist Net, i.e., a network consists of $\ell_\infty$-dist neurons only. The second type is a composition of $\ell_\infty$-dist Net and a shallow MLP (denoted as $\ell_\infty$-dist Net+MLP). That is, using $\ell_\infty$-dist Net as a robust feature extractor as described in Section \ref{sec_combine}. We use a 5-layer $\ell_{\infty}$-dist Net for MNIST and Fashion-MNIST, and a 6-layer $\ell_{\infty}$-dist Net for CIFAR-10 and TinyImageNet.
Each hidden layer has 5120 neurons, and the top layer has 10 neurons (or 200 neurons for TinyImageNet) for classification. Normalization is applied between each intermediate layer. For $\ell_\infty$-dist Net+MLP, we remove the top layer and add a 2-layer fully connected conventional network on top of it. The hidden layer has 512 neurons with tanh activation. See Table \ref{tbl_architecture} for a complete demonstration of the models on each dataset.
\paragraph{Training configurations.} In all experiments, we train $\ell_\infty$-dist Net and $\ell_\infty$-dist Net+MLP using Adam optimizer with hyper-parameters $\beta_1=0.9$, $\beta_2=0.99$ and $\epsilon=10^{-10}$. The batch size is set to 512. For data augmentation, we use random crop (padding=1) for MNIST and Fashion-MNIST, and use random crop (padding=4) and random horizontal flip for CIFAR-10, following the common practice. For TinyImageNet dataset, we use random horizontal flip and crop each image to $56\times 56$ pixels for training, and use a center crop for testing, which is the same as \citet{xu2020automatic}. As for the loss function, we use multi-class hinge loss for $\ell_\infty$-dist Net and the IBP loss \citep{gowal2018effectiveness} for $\ell_\infty$-dist Net+MLP. The training procedure is as follows. First, we relax the $\ell_{\infty}$-dist net to $\ell_{p}$-dist net by setting $p=8$ and train the network for $e_1$ epochs. Then we gradually increase $p$ from 8 to 1000 exponentially in the next $e_2$ epochs. Finally, we set $p=\infty$ and train the last $e_3$ epochs. Here $e_1,e_2$ and $e_3$ are hyper-parameters varying from the dataset. We use $lr=0.02$ in the first $e_1$ epochs and decease the learning rate using cosine annealing for the next $e_2+e_3$ epochs. We use $\ell_p$-norm weight decay for $\ell_{\infty}$-dist nets and $\ell_2$-norm weight decay for the MLP with coefficient $\lambda=0.005$. All these explicitly specified hyper-parameters are kept fixed across different architectures and datasets. For $\ell_\infty$-dist Net+MLP training, we use the same linear warmup strategy for hyper-parameter $\epsilon_{\text{train}}$ in \citet{gowal2018effectiveness,zhang2020towards}. See Appendix \ref{sec_experiment_details} (Table \ref{tbl_hyperparameter}) for details of training configuration and hyper-parameters.
\begin{table*}[ht]
\vspace{-5pt}
\caption{Comparison of our results with existing methods\footnotemark[1].}
\small
\vspace{2pt}
\label{tbl_results}
\begin{center}
\begin{tabular}{c|cr|rrr}
\hline
Dataset & Method & FLOPs & Test & Robust & Certified \\ \hline
\multirow{6}{*}{\begin{tabular}[c]{@{}c@{}}MNIST\\ ($\epsilon=0.3$)\end{tabular}} & Group Sort~\citep{anil2019sorting} &2.9M & 97.0 & 34.0 & 2.0 \\
& COLT~\citep{Balunovic2020Adversarial} &4.9M & 97.3 & - & 85.7 \\
& IBP ~\citep{gowal2018effectiveness} &114M & 97.88 & 93.22 & 91.79 \\
& CROWN-IBP~\citep{zhang2020towards} &114M & 98.18 & 93.95 & 92.98 \\
& $\ell_\infty$-dist Net &82.7M & 98.54 & 94.71 & 92.64 \\
& $\ell_\infty$-dist Net+MLP &85.3M & \textbf{98.56} & \textbf{95.28} & \textbf{93.09} \\ \hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Fashion\\ MNIST\\ ($\epsilon=0.1$)\end{tabular}} & CAP~\citep{wong2018provable} &0.41M & 78.27 & 68.37 & 65.47 \\
& IBP~\citep{gowal2018effectiveness} &114M & 84.12 & 80.58 & 77.67 \\
& CROWN-IBP~\citep{zhang2020towards} &114M & 84.31 & 80.22 & 78.01 \\
& $\ell_\infty$-dist Net &82.7M & \textbf{87.91} & 79.64 & 77.48 \\
& $\ell_\infty$-dist Net+MLP &85.3M & \textbf{87.91} & \textbf{80.89} & \textbf{79.23} \\ \hline
\multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}CIFAR-10\\ ($\epsilon=8/255$)\end{tabular}} & PVT~\citep{dvijotham2018training} &2.4M & 48.64 & 32.72 & 26.67 \\
& DiffAI~\citep{mirman2019provable} &96.3M & 40.2 & - & 23.2 \\
& COLT~\citep{Balunovic2020Adversarial} &6.9M & 51.7 & - & 27.5 \\
& IBP~\citep{gowal2018effectiveness} &151M & 50.99 & 31.27 & 29.19 \\
& CROWN-IBP~\citep{zhang2020towards} &151M & 45.98 & 34.58 & 33.06 \\
& CROWN-IBP (loss fusion)~\citep{xu2020automatic} &151M & {46.29} & {35.69} & 33.38 \\
& $\ell_\infty$-dist Net &121M & \textbf{56.80} & \textbf{37.46} & 33.30 \\
& $\ell_\infty$-dist Net+MLP &123M & 50.80 & 37.06 & \textbf{35.42} \\ \hline
\end{tabular}
\end{center}
\vspace{-12pt}
\end{table*}
\paragraph{Evaluation.} Following common practice, we test the robustness of the trained models under $\ell_\infty$-perturbation $\epsilon=0.3$ on MNIST, $\epsilon=0.1$ on Fashion-MNIST, $\epsilon=8/255$ on CIFAR-10 and $\epsilon=1/255$ on TinyImageNet. We use two evaluation metrics to measure the robustness of the model. We first evaluate the robust test accuracy under the Projected Gradient Descent (PGD) attack \citep{madry2017towards}. Following standard practice, we set the number of steps of the PGD attack to be 20. We also calculate the certified radius for each sample, and check the percentage of test samples that can be certified to be robust within the chosen radius. Note that the second metric is always lower than the first.
\paragraph{Baselines.}
We compare our proposed models with state-of-the-art methods for each dataset, including relaxation methods: CAP~\citep{wong2018provable}, PVT~\citep{dvijotham2018training}, DiffAI~\citep{mirman2019provable}, IBP~\citep{gowal2018effectiveness}, CROWN-IBP~\citep{zhang2020towards}, CROWN-IBP with loss fusion~\citep{xu2020automatic}, COLT~\citep{Balunovic2020Adversarial}, and Lipschitz networks: GroupSort~\citep{anil2019sorting}. We do not compare with randomized smoothing methods~\citep{pmlr-v97-cohen19c,salman2019provably,zhai2020macer} since it cannot obtain good certification under s relative large $\ell_\infty$-norm perturbation as discussed in Section \ref{sec_related_work}. We report the performances picked from the original papers if not specified otherwise.
\subsection{Experimental Results}
We list our results in Table~\ref{tbl_results} and Table~\ref{tbl_results_imagenet}. We use ``Standard'', ``Robust'' and ``Certified'' as abbreviations of standard (clean) test accuracy, robust test accuracy under PGD attack and certified robust test accuracy. All the numbers are reported in percentage. We use ``FLOPs'' to denote the number of basic floating-point operations needed (i.e., multiplication-add in conventional networks or subtraction in $\ell_\infty$-dist nets) in forward propagation. Note that FLOPs in linear relaxation methods are typically small because, in the original papers, they are implemented on small networks due to high costs.
\begin{table*}[ht]
\vspace{-5pt}
\caption{Comparison of our results with \citet{xu2020automatic} on TinyImageNet dataset ($\epsilon=1/255$).}
\vspace{2pt}
\small
\label{tbl_results_imagenet}
\begin{center}
\begin{tabular}{c|cr|rrr}
\hline
Method & Model & \multicolumn{1}{c|}{FLOPs} & \multicolumn{1}{c}{Test} & \multicolumn{1}{c}{Robust} & \multicolumn{1}{c}{Certified} \\ \hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}CROWN-IBP\\ (loss fusion)\\\citep{xu2020automatic}\end{tabular}} & CNN7+BN & 458M & 21.58 & 19.04 & 12.69 \\
& ResNeXt & 64M & 21.42 & 20.20 & 13.05 \\
& DenseNet & 575M & 22.04 & 19.48 & 14.56 \\
& WideResNet & 5.22G & 27.86 & 20.52 & 15.86 \\ \hline
$\ell_\infty$-dist net & $\ell_\infty$-dist Net+MLP & 156M & 21.82 & 18.09 & \textbf{16.31} \\ \hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[ht]
\vspace{-5pt}
\caption{Ablation studies for Section \ref{sec_optimize} on CIFAR-10 dataset.}
\vspace{2pt}
\small
\label{tbl_ablation}
\begin{center}
\begin{tabular}{c|ccc|rrr|rrr}
\hline
\multicolumn{1}{l|}{} & \multicolumn{1}{c}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Smooth\\ gradient\end{tabular}}} & \multicolumn{1}{c}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Identity\\ init\end{tabular}}} & \multicolumn{1}{c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Weight\\ decay\end{tabular}}} & \multicolumn{3}{c|}{$\ell_\infty$-dist Net} & \multicolumn{3}{c}{$\ell_\infty$-dist Net+MLP} \\
\multicolumn{1}{l|}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{Test} & \multicolumn{1}{c}{Robust} & \multicolumn{1}{c|}{Certified} & \multicolumn{1}{c}{Test} & \multicolumn{1}{c}{Robust} & \multicolumn{1}{c}{Certified} \\ \hline
A & \xmark & \xmark & \xmark & 28.21 & 8.42 & 7.02 & 37.68 &28.72 &27.76 \\
B & \cmark & \xmark & \xmark & 55.63 & 35.28 & 32.56 & 37.61 & 30.99 & 29.71 \\
C & \cmark & \cmark & \xmark & 56.15 & 35.96 & 32.71 & 48.97 & 36.21 & 35.02 \\
D & \cmark & \cmark & $\ell_2$-norm & 53.64 & 32.12 & 29.01 & 45.34 & 33.48 & 32.49 \\
E & \cmark & \cmark & \cmark & 56.80 & 37.46 & 33.30 & 50.80 & 37.06 & 35.42 \\ \hline
\end{tabular}
\vspace{-10pt}
\end{center}
\end{table*}
\paragraph{General Performance of $\ell_\infty$-dist Net.} From Table \ref{tbl_results} we can see that using $\ell_\infty$-dist Net alone already achieves decent certified accuracy on all datasets. Notably, $\ell_\infty$-dist Net reaches the start-of-the-art certified accuracy on CIFAR-10 dataset while achieving a significantly higher standard accuracy than all previous methods. Note that we just use a standard loss function to train the $\ell_\infty$-dist Net \textit{without any adversarial training}.
\paragraph{General Performance of $\ell_\infty$-dist Net+MLP.} For all these datasets, $\ell_\infty$-dist Net+MLP achieves better certified accuracy than $\ell_\infty$-dist Net, establishing new state-of-the-art results. For example, on the MNIST dataset, the model can reach 93.09\% certified accuracy and 98.56\% standard accuracy; on CIFAR-10 dataset, the model can reach 35.42\% certified accuracy, which is 6.23\% higher than IBP and 2.04\% higher than the previous best result; on the TinyImageNet dataset (see Table \ref{tbl_results_imagenet}), the simple $\ell_\infty$-dist Net+MLP model (156M FLOPs computational cost) already beats the previous best result in \citet{xu2020automatic} although their model is 33 times larger than ours (5.22G FLOPs). Furthermore, the clean test accuracy of $\ell_\infty$-dist Net+MLP is also better than CROWN-IBP on MNIST, Fashion-MNIST and CIFAR-10 dataset.
\paragraph{Efficiency.} Both the training and the certification of $\ell_\infty$-dist Net are very fast. As stated in previous sections, the computational cost per iteration for training $\ell_\infty$-dist Net is roughly the same as training a conventional network of the same size, and the certification process only requires a forward pass. In Table \ref{tab:speed}, we quantitatively compare the per-epoch training speed of our method with previous methods such as IBP or CROWN-IBP on CIFAR-10 dataset. All these experiments are run on a single NVIDIA-RTX 3090 GPU. As we can see, for both $\ell_\infty$-dist Net and $\ell_\infty$-dist Net+MLP, the per-epoch training time is less than 20 seconds, which is significantly faster than CROWN-IBP and is comparable to IBP.
\begin{table}[ht]
\vspace{-5pt}
\caption{Comparison of per-epoch training speed for different methods on CIFAR-10 dataset.}
\vspace{2pt}
\small
\label{tab:speed}
\centering
\begin{tabular}{cc}
\hline
Method & Per-epoch Time (seconds)\\
\hline
IBP & 17.4 \\
CROWN-IBP & 112.4 \\
CROWN-IBP (loss fusion) & 43.3\\
$\ell_\infty$-dist Net & 19.7\\
$\ell_\infty$-dist Net+MLP & 19.7\\
\hline
\end{tabular}
\vspace{-10pt}
\end{table}
\footnotetext[1]{For GroupSort, the results are obtained from \citet{anil2019sorting}, Fig. 8,9. For COLT, the certified result uses MILP (mixed integer linear programming) solver which is much slower than other methods. For IBP, the results are obtained from \citet{zhang2020towards}. For PVT, the certified accuracy are under perturbation $\epsilon=0.03$ which is smaller than 8/255.}
\paragraph{Discussions with GroupSort Network.} We finally make a special comparison with GroupSort, because both GroupSort network and our proposed $\ell_\infty$-dist Net design 1-Lipschitz networks explicitly and can be directly trained using a standard loss function. However, in the GroupSort network, all weight matrices $\mathbf W$ are constrained to have bounded $\ell_\infty$-norm, i.e., $\|\mathbf W\|_\infty\le 1$, leading to a time-consuming projection operation (see Appendix C in \citet{anil2019sorting}). This operation brings optimization difficulty \citep{cohen2019universal} and further limit the scalability of the network structure. We hypothesis that this is the major reason why $\ell_\infty$-dist Net substantially outperforms GroupSort on the MNIST dataset, as shown in Table \ref{tbl_results}.
\subsection{Ablation Studies}
\label{sec_ablation}
In this section, we conduct ablation experiments to see the effect of smoothed approximated gradients, parameter initialization using identity map construction, and $\ell_p$-norm weight decay. The results are shown in Table \ref{tbl_ablation}. We use the model described in Section \ref{sec_experiment_setting} on CIFAR-10 dataset with hyper-parameters provided in Table \ref{tbl_hyperparameter}. From Table \ref{tbl_ablation} we can clearly see that:
\begin{itemize}[topsep=0pt]
\setlength{\itemsep}{0pt}
\item Smoothed approximated gradient technique is crucial to train a good model for $\ell_\infty$-dist Net. After applying smoothed approximated gradient only, we can already achieve 32.56\% certified accuracy;
\item Both smoothed approximated gradient and identity-map initialization are crucial to train a good model for $\ell_\infty$-dist Net+MLP. Combining the two techniques results in 35.02\% certified accuracy;
\item $\ell_p$-norm weight decay can further boost the results, although the effect may be marginal (0.59\% and 0.4\% improvement of certified accuracy for the two models).
\item Conventional $\ell_2$-norm weight decay does harm to the performance of $\ell_\infty$-dist nets.
\end{itemize}
In summary, these optimization strategies in Section \ref{sec_optimize} all contribute to the final performance of the model.
\section{Conclusion}
In this paper, we design a novel neuron that uses $\ell_\infty$ distance as its basic operation. We show that the neural network constructed with $\ell_\infty$-dist neuron is naturally a 1-Lipschitz function with respect to $\ell_\infty$ norm. This directly provides a theoretical guarantee of the certified robustness based on the margin of the prediction outputs. We further formally analyze the expressive power and the robust generalization ability of the network, and provide a holistic training strategy to handle optimization difficulties encountered in training $\ell_\infty$-dist nets. Experiments show promising results on MNIST, Fashion-MNIST, CIFAR-10 and TinyImageNet datasets. As this structure is entirely new, plenty of aspects are needed to investigate, such as how to further handle the optimization difficulties for this network. For future work we will study these aspects and extend our model to more challenging tasks like ImageNet.
\section*{Acknowledgements}
This work was supported by National Key R\&D Program of China (2018YFB1402600), Key-Area Research and Development Program of Guangdong Province (No. 2019B121204008), BJNSF (L172037) and Beijing Academy of Artificial Intelligence. Project 2020BD006 supported by PKU-Baidu Fund.
|
\section{Introduction}
\label{sec:Intro}
In this paper, we consider the following nonconvex minimization problem with linear constraint
\begin{equation}
\label{model}
\begin{split}
\min_{x,y} & \quad F(x_1,\ldots,x_s) + h(y) \\
\text{ such that} & \quad \sum_{i=1}^s \mathcal A_i x_i + \mathcal By = b,
\end{split}
\end{equation}
where $y \in \mathbb R^{\mathbf q}$, $x_i \in \mathbb R^{\mathbf n_i}$, $x:=[x_1;\ldots; x_s] \in \mathbb R^{\mathbf n}$, $\mathbf n=\sum_{i=1}^s \mathbf n_i$, $\mathcal A_i$ is a linear map from $\mathbb R^{\mathbf n_i}$ to $\mathbb R^{\mathbf m}$, $\mathcal B$ is a linear map from $\mathbb R^{\mathbf q}$ to $\mathbb R^{\mathbf m}$, $b\in \mathbb R^{\mathbf m} $, $h:\mathbb R^{\mathbf q} \to \mathbb R$ is a differentiable function, and $F(x)=f(x)+\sum_{i=1}^s g_i(x_i)$, where $f: \mathbb R^{\mathbf n} \to \mathbb R$ is a nonconvex nonsmooth function and $g_i:\mathbb R^{\mathbf n_i} \to \mathbb R\cup \{+\infty\}$ are proper lower semi-continuous functions for $i=1,2,\dots,s$. We assume that $F$ satisfies\footnote{This condition is satisfied when $f$ is a sum of a continuously differentiable function and a block separable function that
has limiting subdifferential, see~\cite[Proposition 2.1]{Attouch2010}.}
$\partial F(x) = \partial_{x_1} F(x) \times \ldots \times \partial_{x_s} F(x)$, where $\partial F$ denote the limiting subdifferential of $F$ (see the definition in the supplementary document).
\textbf{Notation.} We denote $[s]:=\{1,\ldots,s\}$. For the $\mathbf p$-dimensional Euclidean space $\mathbb R^{\mathbf p}$, we use $\langle\cdot,\cdot\rangle$ to denote the inner product and $\|\cdot\|$ to denote the corresponding induced norm.
For a linear map $\mathcal M$, $\mathcal M^*$ denotes the adjoint linear map with
respect to the inner product and $\|\mathcal M\|$ is the induced operator norm of $\mathcal M$. We use $\mathcal I $ to denote the identity map. For a positive definite self-adjoint operator $\mathcal Q$, we denote $\|x\|^2_{\mathcal Q}:=\langle x, \mathcal Qx\rangle $. We denote the smallest eigenvalue of a symmetric linear self-map (that is, $\mathcal M=\mathcal M^*$) by $\lambda_{\min}(\mathcal M)$. We use $Im(\mathcal B)$ to denote the image of $\mathcal B$
\subsection{Nonconvex low-rank representation problem}
\label{sec:intro-LRR}
One of the important applications of
Problem~\eqref{model} is the following generalized nonconvex low-rank representation problem: given a data matrix $D \in \mathbb{R}^{d \times n}$, solve
\begin{equation}\label{LRR}
\begin{aligned}
\min_{X,Y,Z} & \,\,\sum_{i=1}^{\min(m,n)} r_1(\sigma_i(X)) + r_2(Y) + r_3(Z)\\
\text{subject to} & \qquad D= A_1X + Y A_2 + Z,
\end{aligned}
\end{equation}
where $X\in\mathbb{R}^{m\times n}, Y\in\mathbb{R}^{d\times q}, Z\in \mathbb{R}^{d\times n}$,
$A_1\in\mathbb{R}^{d\times m}$,
$A_2\in\mathbb{R}^{q\times n}$,
$r_1(\cdot)$ is an increasing concave function to promote X to be of low rank,
$r_2(\cdot)$ is regularization function, and $r_3(\cdot)$ is a function that models some noise (for example, if we take $r_3(Z) = \frac{1}{2}\|Z\|_F^2$ then $Z$ represents a Gaussian noise).
Problem \eqref{LRR} generalizes several important problems in machine learning.
Let us mention some examples:
\begin{itemize}
\item[(i)] When $A _1$ and $A_2$ are identity matrices, $r_1(t) = t^\chi$ with $0< \chi \leq 1$, $r_2(Y) = \sum_{i=1}^{q-1}\|Y_{i} - Y_{i+1}\|$, where $Y_{i}$ is the
$i$-th column of $Y$,
Problem~\eqref{LRR} decomposes the data matrix $D$ into three components, $X$, $Y$ and~$Z$.
For example, in video surveillance,
each column of $D$ is a vectorized image of a video frame,
$X$ is a low-rank matrix that plays the role of the background, $Y$ is the foreground that has small variations betweet its columns (such as slowly moving objectives), and $Z$ represents some noise \cite{Wang2019}.
\item[(ii)] When $A _1$ and $A_2$ are identity matrices, $ r_1(t) = t$, $r_2(Y) = \lambda\|Y\|_1$, where $\lambda$ is some constant, Problem \eqref{LRR} recovers the robust principal component analysis model, see, e.g., \cite{Cande2011}, where $X$ is a low-rank matrix, $Y$ represents a sparse noise, and $Z$ represents additional noise. It is also used for foreground-background separation in video surveillance.
\item[(iii)] When $r_1(t) = t$ and $r_2(Y) = \|Y\|_*$, Problem~\eqref{LRR} is the latent low-rank representation problem \cite{Liu2011}.
The authors~\cite{Liu2011} used
$A_1 = DP_1$ and $A_2 = P_2^*D$, where $P_1$ and $P_2$ are computed by orthogonalizing the columns of $D^*$ and $D$, respectively. {We will use this application to illustrate the effectiveness of our proposed framework, iADMM, in Section~\ref{sec:numerical}.} \vspace{-0.1cm}
\end{itemize}
{Other applications of Problem~\eqref{model} include statistical learning, see, e.g., \cite{Bach2011,WangTSP2011},
and minimization on compact manifolds,
see, e.g., \cite{Lai2014,Wen2010}.}
\subsection{Motivation and related works}
Let $$\mathcal A:=[\mathcal A_1 \ldots \mathcal A_s],\qquad\mathcal A x:=\sum_{i=1}^s \mathcal A_i x_i\in \mathbb R^{\mathbf m}.$$ The augmented Lagrangian for Problem~\eqref{model} is given by
\begin{equation}
\label{Lagrangian}
\begin{split}
\mathcal L(x, y, \omega):= F(x) + h(y)
+ \langle \omega,\mathcal A x + \mathcal By-b \rangle + \frac{\beta}{2} \|\mathcal A x + \mathcal By-b\|^2,
\end{split}
\end{equation}
where $\beta>0$ is a penalty parameter. ADMM was first introduced by \cite{Glowinski1975} and \cite{Gabay1976}. It has recently become popular because of its efficacy in solving emerging large-scale problems in machine learning and computer vision \cite{Boyd2011,Scheinberg2010,Yang2009,Yang2017,Yin2008}.
For simplicity, let us describe the iteration scheme of a classical ADMM for solving Problem~\eqref{model} with 2 blocks $x$ and $y$:
\begin{subequations}
\label{ADMM-classic}
\begin{align}
x^{k+1} &\in\argmin_{x} \mathcal L(x,y^{k},\omega^k),\label{subproblem_x}\\
y^{k+1} &\in\argmin_{y} \mathcal L(x^{k+1},y,\omega^k),\label{subproblem_y}\\
\omega^{k+1} &=\omega^k+ \beta (\mathcal A x^{k+1} +\mathcal B y^{k+1} -b). \label{subproblem_w}
\end{align}
\end{subequations}
For a multi-block problem, with $s>1$, the scheme is similar, see for example \cite{Wang2019}.
The update of $x$ in~\eqref{subproblem_x} (a similar discussion is applicable to~\eqref{subproblem_y}) can be rewritten as
\begin{equation*}
x^{k+1} \; \in \; \argmin_{x} F(x) + \varphi^k(x),
\end{equation*}
where
\begin{equation}
\label{eq:varphi}
\varphi^k(x)=\frac{\beta}{2}\|\mathcal A x + \mathcal By^k -b \|^2 + \langle \omega^k,\mathcal A x + \mathcal By^k -b \rangle.
\end{equation}
Solving the subproblem~\eqref{subproblem_x} is usually very expensive especially when $F$ is not smooth.
A remedy is minimizing a suitable surrogate function of $\mathcal L(\cdot, y^k,\omega^k)$ that allows a more efficient update for $x$. For example, since $\varphi^k(x)$ is upper bounded by
\begin{equation}
\label{eq:varphihat}
\hat \varphi(x)= \varphi^k(x^k)+ \langle \nabla \varphi^k(x^k),x-x^k\rangle + \frac{\kappa\beta}{2} \|x-x^k\|^2
\end{equation}
where $\kappa\geq \|\mathcal A^* \mathcal A \|$ (because $\nabla \varphi^k(x)$ is $\beta \|\mathcal A^* \mathcal A \|$-Lipschitz continuous), $x$ can be updated by
\begin{equation}
\label{eq:LADMM}
x^{k+1} \; \in \; \argmin_{x} F(x) + \hat{\varphi}(x),
\end{equation}
which leads to the linearized ADMM method, see \cite{Lin2011,XuWu2011}. The update in~\eqref{eq:LADMM}
has a closed form for some nonsmooth $F$; see \cite{Boyd2014}.
When $F= f + g$ and $f$ is $L_f$-smooth then we can also use the upper bound $ \hat F(x)=f(x^k)+ \langle \nabla f(x^k),x-x^k\rangle + \frac{L_f}{2} \|x-x^k\|^2 + g(x)$ of $ F$ to derive the following update for $x$:
\begin{equation}
\label{eq:PLADMM}
x^{k+1} \; \in \; \argmin_{x} \hat F(x) + \hat{\varphi}(x) .
\end{equation}
This leads to the proximal linearized ADMM method, see \cite{Bot2020,LiuShen2011}.
We note that $\mathcal L(\cdot, y^k,\omega^k)$ is always upper bounded by $\mathcal L(\cdot, y^k,\omega^k) +\mathbf D_\phi(x,x^k)$, where $\mathbf D_\phi$ is the Bregman distance associated with a continuously differentiable convex function $\phi$ on $\mathbb R^n$:
\begin{equation}
\label{def:Bregman}
\mathbf D_\phi(a,b):=\phi(a)-\phi(b)-\langle \nabla \phi(b),a-b\rangle, \forall a,b\in\mathbb R^n.
\end{equation}
For example, if $\phi(x)=\|x\|_{\mathcal Q}^2=\langle x, \mathcal Q x\rangle$ then $\mathbf D_\phi(a,b)=\|a - b\|^2_{\mathcal Q}$.
This upper bound leads to proximal ADMM,
see \cite{DengYin2012,Li2015}.
The above mentioned upper bound functions are specific examples of surrogate functions for $\mathcal L(\cdot, y^k,\omega^k)$ (see Definition~\ref{def:surrogate}) while each method of updating $x$ corresponds to a majorization-minimization (MM) step. In the convex setting (that is, $f(x,y)$ is convex), \cite{Lu2016} and \cite{HongADMM2020} use the MM principle to unify and generalize the convergence analysis of many ADMM for multi-blocks problems (that is, $s>1$). However, ADMM with the MM principle has not been studied for the nonconvex problem~\eqref{model}, to the best of our knowledge.
When the linear coupling constraint is absent, the block coordinate descent (BCD) method is a standard approach to solve~\eqref{model}. \cite{Razaviyayn2013} proposed the block successive upper-bound minimization (BSUM) framework that employs the MM principle in each block update. By employing suitable surrogate functions in each block update, BSUM recovers the typical BCD methods, for example of \cite{GRIPPO20001,Hildreth,Powell1973,Tseng2001,Beck2013,Bolte2014,Tseng2009}. In the non-convex setting, BCD methods with inertial terms have also been studied and they have showed significant improvement in their practical performance, see for example \cite{Ochs2019,Xu2013,Xu2017,Pock2016,Hien_ICML2020}.
Recently, the authors in \cite{Titan2020} propose an inertial block MM framework for solving~\eqref{model} without the linear coupling constraint.
To the best of our knowledge, ADMM with inertial terms for the primal variables have not been studied for the nonconvex setting although they have been analysed for the convex setting; see \cite{LiLin2019,Ouyang2015}.
\subsection{Contribution}
In this paper, we propose iADMM, a framework of inertial alternating direction methods of multipliers, for solving the nonconvex nonsmooth problem~\eqref{model}. When no extrapolation is used, iADMM becomes a general ADMM framework that employs the minimization-majorization principle in each block update. For the first time in the \emph{nonconvex} nonsmooth setting of Problem~\eqref{model}, we study ADMM and its inertial version combined with the MM principle when updating each block of variables.
Moreover, our framework allows to use an over-relaxation parameter $\alpha \in (0,2)$ to set $\alpha \beta$ as the constant stepsize for updating the dual variable $\omega$. Note that $\alpha=1$ is the standard choice in the nonconvex setting, see, e.g., \cite{HongADMM2020,Li2015,Wang2019}. In the convex setting, \cite{Gabay1976} showed that $ \alpha$ can be chosen in $(0,2)$. However, in the nonconvex setting, most of the works assume that $ \alpha\in \big( 0, \frac{1+\sqrt{5}}{2}\big)$, see, e.g.,~\cite{Fazel2013,Yang2017}. Recently, \cite{Bot2020} proposed proximal ADMM that use $\alpha\in (0,2)$ for solving a special case of the nonconvex Problem~\eqref{model} with $s=1$ and $\mathcal A=-\mathcal I$.
Under mild assumptions, we analyse the subsequential convergence guarantee for the generated sequence of iADMM and ADMM in parallel. When $F(x) + h(y)$ satisfies the K{\L} property and $\alpha=1$, we prove the global convergence for the generated sequence. Finally, we apply the proposed framework to solve a class of Problem~\eqref{LRR} and report its numerical results to illustrate the efficacy of iADMM.
\section{An inertial ADMM framework}
\label{sec:algorithms}
In this section, we describe the iADMM framework and prove its subsequential and global convergence.
Throughout the paper, we make the following assumptions that are standard for studying Problem~\eqref{model} and the convergence of ADMMs in the nonconvex setting, see for example \cite{Wang2019,Bot2020,Li2015}.
\begin{assumption}
\label{assume}
(i) $\sigma_{\mathcal B}:=\lambda_{\min}(\mathcal B \mathcal B^*) >0 $.
(ii)
$F(x) + h(y)$ is lower bounded.
(iii) $h$ is a $L_h$-smooth function (that is, $\nabla h$ is $L_h$-Lipschitz continuous with constant $L_h$).
\end{assumption}
\subsection{iADMM description}
\label{sec:iADMM-description}
Let us first formally define a surrogate function. Some examples were given in the introduction.
\begin{definition}[Surrogate function]
\label{def:surrogate}
Let $\mathcal X\subseteq \mathbb R^{\mathbf n}$.
A function $u:\mathcal X \times \mathcal X \to \mathbb R$ is called a surrogate function of a function $f$ on $\mathcal X$
if the following conditions are satisfied:
\begin{itemize}
\item[(a)] $u(z,z) = f(z)$ for all $z\in \mathcal X$,
\item[(b)] $u(x,z) \geq f(x)$ for all $x,z\in\mathcal X$.
\end{itemize}
\end{definition}
As we are considering multi-block problems, we need the following definition of a block surrogate function, which is a generalization of Definition~\ref{def:surrogate}.
\begin{definition}[Block surrogate function]
\label{def:surrogate-block} Let Let $\mathcal X_i\subseteq \mathbb R^{\mathbf n_i}$, $\mathcal X\subseteq \mathbb R^{\mathbf n}$.
A function $u_i:\mathcal X_i \times \mathcal X \to \mathbb R $ is called a block $i$ surrogate function of $f$ on $\mathcal X$
if the following conditions are satisfied:
\begin{itemize}
\item[(a)] $u_i(z_i,z) = f(z)$ for all $z\in \mathcal X$,
\item[(b)] $u_i(x_i,z) \geq f(x_i,z_{\ne i})$ for all $x_i\in\mathcal X_i$ and $z\in \mathcal X$, where
$$(x_i,z_{\ne i}):= (z_1,\ldots,z_{i-1},x_i,z_{i+1},\ldots,z_s).$$
\end{itemize}
{The block approximation error is defined as $$ e_i(x_i,z):=u_i(x_i,z) - f(x_i,z_{\ne i}).$$}
\end{definition}
The inertial alternating direction method of multipliers (iADMM) framework is described in Algorithm~\ref{alg:iADMM}.
\begin{algorithm}[tb]
\caption{iADMM for solving Problem~\eqref{model}}
\label{alg:iADMM}
\begin{algorithmic}
\STATE Choose $x^{0}= x^{-1}$, $y^0=y^{-1}$, $\omega^0$. Let $u_i$, $i\in [s]$, be block $i$ surrogate functions of $f(x)$ on $\mathbb R^{\mathbf n}$.
\FOR{$k=0,\ldots$}
\STATE Set $x^{k,0}=x^k$
\FOR{ $i = 1,...,s$}
\STATE\label{step-update} Compute $\bar x_i^k=x_i^k + \zeta_i^k (x_i^k - x_i^{k-1})$.
\STATE Update block $x_i$ by
\begin{equation}
\label{eq:xi-updatenew}
\begin{split}
& x_i^{k,i}\in\argmin_{x_i}\Big\{ u_i(x_i,x^{k,i-1})+g_i(x_i) \\
& \,+ \langle \mathcal A_i^*\big( \omega^k+\beta(\mathcal A \bar x^{k,i-1} +\mathcal By^k-b) \big),x_i\rangle+\frac{\kappa_i\beta}{2}\|x_i-\bar x_i^k\|^2\Big\},
\end{split}
\end{equation}
where $\kappa_i\geq \|\mathcal A_i^* \mathcal A_i\|$, and
$\bar x^{k,i-1}=(x^{k+1}_1,\ldots,x^{k+1}_{i-1},\bar x^{k}_i, x^{k}_{i+1},\ldots,x^k_s)$.
\STATE Set $x^{k,i}_j = x^{k,i-1}_j$ for all $j\neq i$.
\ENDFOR
\STATE Set $x^{k+1} = x^{k,s}$.
\STATE Compute $\hat y^k= y^k + \delta_k (y^k - y^{k-1})$.
\STATE Update $y$ by
\begin{equation}
\label{eq:y_update}
\begin{split}
y^{k+1}\in\argmin_{y}\Big\{ \langle \mathcal B^*\omega^k +\nabla h(\hat y^k),y \rangle + \frac{\beta}{2} \|\mathcal A x^{k+1} + \mathcal By-b\|^2+\frac{L_h}{2}\|y-\hat y^k\|^2\Big\}.
\end{split}
\end{equation}
\STATE Update $\omega$ by
\begin{equation}
\label{eq:omega_update}
\omega^{k+1}=\omega^k +\alpha\beta (\mathcal A x^{k+1}+ \mathcal B y^{k+1} -b).
\end{equation}
\ENDFOR
\end{algorithmic}
\end{algorithm}
iADMM cyclically update the blocks $x_1,\ldots,x_s$ and $y$. Let us use
$
x^{k,i}=(x^{k+1}_1, \ldots,x^{k+1}_{i},x^{k}_{i+1},\ldots,x^{k}_s)$ and $x^{k+1} = x^{k,s}$, {where $k$ is the outer iteration index, and $i$ the cyclic inner iteration index ($i \in [s]$).}
The update of block $x_i$ in~\eqref{eq:xi-updatenew} (note that $x^{k+1}_{ i}=x^{k,i}_{ i} $) means that iADMM chooses a surrogate function for $x_i\mapsto \mathcal L(x_i,x^{k,i}_{\ne i},y^k,\omega^k)$, which is formed by summing a surrogate function of $x_i\mapsto f(x_i,x^{k,i}_{\ne i}) + g_i(x_i)$ and a surrogate function of $x_i\mapsto \varphi^k(x_i,x^{k,i}_{\ne i})$ where $\varphi^k(x)$ is defined in \eqref{eq:varphi}, then apply extrapolation to the latter
surrogate function\footnote{It is important noting that it is possible to embed the general inertial term $\mathcal G_i^k$ to the surrogate of $x_i\mapsto \mathcal L(x_i,x^{k,i}_{\ne i},y^k,\omega^k)$ as in~\cite{Titan2020}. This inertial term may also lead to the extrapolation for the block surrogate function of $ f(x)$ or for both the two block surrogates. However, to simplify our analysis, we only consider here the effect of the inertial term for the block surrogate of $\varphi^k(x)$.}. To update block $y$, as $h(y)$ is $L_h$-smooth, we apply Nesterov type acceleration on $h$ as in \eqref{eq:y_update}.
Together with Assumption~\ref{assume}, we make the following standard assumption for $u_i$ throughout the paper.
\begin{assumption}
\label{assump:Lipschitz_ui}
(i) The block surrogate function $u_i(x_i,z)$ is continuous.
(ii) Given $z\in \mathbb R^{\mathbf n}$, for $i\in [s]$, there exists a function $x_i\mapsto \bar e_i(x_i,z)$ such that $ \bar e_i(\cdot,z)$ is continuously differentiable at $z_i$, $\bar e_i(z_i,z)=0$, $\nabla_{x_i} \bar e_i(z_i,z)=0$, and the block approximation error $x_i\mapsto e_i(x_i,z)$ satisfies \begin{equation}
\label{lemma:h_property}
e_i(x_i,z) \leq \bar e_i(x_i,z) \; \text{ for all } \; x_i.
\end{equation}
\end{assumption}
The condition in Assumption~\ref{assump:Lipschitz_ui} (ii) is satisfied when we simply choose $u_i(x_i,z) = f(x_i,z_{\ne i})$ (that is, $f(x_i,z_{\ne i})$ is a surrogate function of itself), or when $e_i(\cdot,z)$ is continuously differentiable at $z_i$ and $\nabla_{x_i} e_i(z_i,z)=0$, or when $e_i(x_i,z)\leq c \|x_i-z_i\|^{1+\epsilon} $ for some $\epsilon>0$ and $c>0$; see \cite[Lemma 3]{Titan2020}.
\begin{remark} Before proceeding to the convergence analysis of iADMM, we make the following remark. As we target Nesterov-type acceleration in the update of $y$ (note that $h$ is assumed to be $L_h$-smooth), we analyse the update rule as in \eqref{eq:y_update} for $y$. In case $y$ is updated by
\begin{equation*}
\begin{split}
& y^{k+1} \; \in \; \argmin_{y} \mathcal L(x^{k+1},y,\omega^k),
\end{split}
\end{equation*}
iADMM still works and the convergence analysis would be simplified by using the same rationale to obtain subsequential as well as global convergence. We hence omit this case in our analysis.
\end{remark}
\subsection{Convergence analysis}
Let us start by defining some additional notations and their convention that will be used later. Let $x^{k,i}$, $y^k$ and $\omega^k$ be the iterates generated by iADMM. We denote $\Delta x^k_i= x^{k}_i -x^{k-1}_i$, $\Delta y^k=y^{k} -y^{k-1}$, $\Delta \omega^k =\omega^{k} -\omega^{k-1}$, $\alpha_1=\frac{|1-\alpha|}{\alpha \sigma_{\mathcal B}(1-|1-\alpha|)}$, $ \alpha_2=\frac{3\alpha}{\sigma_{\mathcal B}(1-|1-\alpha|)^2}$ and $\mathcal L^k=\mathcal L(x^{k},y^k,\omega^k )$. We let $\nu_i$, $i\in [s]$, and $\nu_y$ be arbitrary constants in $(0,1)$. We take the following convention in the notation that allows us to analyse iADMM and its non-inertial version in parallel:
\begin{itemize}
\item If $\zeta_i^k = 0$, that is, when we do not apply extrapolation in the update of $x_i^k$, we take $\zeta_i^k/\nu_i=0$ and $\nu_i=0$.
\item If $\delta_k =0$, that is, when we do not apply extrapolation in the update of $y$, we take $\delta_k/\nu_y=0$ and $\nu_y=0$.
\end{itemize}
Now we present our main convergence results. Their proofs can be found in the supplementary material.
As iADMM allows to use extrapolation in the update of $x_i^k$ and $y^k$, the Lagrangian is not guaranteed to satisfy the sufficient descent property; in fact,
it is not guaranteed to decrease at each iteration.
Instead, it has the following nearly sufficiently decreasing property as stated in the following Propositions~\ref{prop:NSDP} and~\ref{prop:NSDP-fory}
\begin{proposition}
\label{prop:NSDP}
(i) Considering the update in~\eqref{eq:xi-updatenew}, in general when $x_i\mapsto u_i(x_i,z) + g_i(x_i)$ is nonconvex,
we choose $\kappa_i> \|A_i^*A_i\|$. Denote $a_i^k= \beta \zeta_i^k(\kappa_i+\| \mathcal A_i^* \mathcal A_i \|)$. Then
\begin{equation}
\label{NSDP}
\begin{array}{l}
\mathcal L(x^{k,i},y^k,\omega^k)+\eta_i \|\Delta x^{k+1}_i\|^2
\leq
\mathcal L(x^{k,i-1},y^k,\omega^k)+ \gamma_i^k \|\Delta x^k_i\|^2,
\end{array}
\end{equation}
where
\begin{equation}
\begin{array}{l}
\eta_i = \frac{(1-\nu_i)(\kappa_i-\|\mathcal A_i^*\mathcal A_i\|)\beta}{2}, \\
\gamma_i^k= \frac{(a^k_i)^2}{2\nu_i(\kappa_i-\|\mathcal A_i^*\mathcal A_i\|)\beta}.
\end{array}
\end{equation}
(ii) When $x_i\mapsto u_i(x_i,z) + g_i(x_i)$ is convex, we choose $\kappa_i =\|\mathcal A_i^* \mathcal A_i\|$
and Inequality~\eqref{NSDP} is satisfied with
\begin{equation}
\label{eq:gammastrong}
\gamma^{k}_i =\frac{\beta \|\mathcal A_i^* \mathcal A_i\| (\zeta_i^k)^2}{2} , \quad \eta_i=\frac{\beta \|\mathcal A_i^* \mathcal A_i\|}{2}.
\end{equation}
\end{proposition}
\begin{proposition}
\label{prop:NSDP-fory}
Considering the update in~\eqref{eq:y_update}, we have
\begin{equation*}
\label{eq:NSDP-y}
\begin{split}
&\mathcal L(x^{k+1},y^{k+1},\omega^k)+ \eta_y \|\Delta y^{k+1}\|^2 \leq \mathcal L(x^{k+1},y^k,\omega^k) + \gamma_y^k\|\Delta y^{k}\|^2,
\end{split}
\end{equation*}
where $\eta_y=\frac{(1-\nu_y)(\beta\|\mathcal B^* \mathcal B\|+L_h) }{2}$ and $\gamma_y^k=\frac{2L_h^2 \delta_k^2}{\nu_y (\beta\|\mathcal B^*\mathcal B\|+L_h)} $ when $h(y)$ is nonconvex, and $\eta_y=\frac{L_h }{2}$ and $\gamma_y^k= \frac{L_h \delta_k^2}{2}$ when $h(y)$ is convex.
\end{proposition}
From Proposition~\ref{prop:NSDP} and Proposition~\ref{prop:NSDP-fory}, we obtain the following recursive inequality for $\{\mathcal L^k\}$ in Proposition~\ref{prop:recursive} that serves as cornerstone to derive the bound for the extrapolation parameters $\zeta_i^k$ and $\delta_k$ in Proposition~\ref{prop:subsequential}.
\begin{proposition}
\label{prop:recursive}
We have
\begin{equation}
\label{eq:recursive2}
\begin{split}
\mathcal L^{k+1}+ \eta_y\|\Delta y^{k+1}\|^2 +\sum_{i=1}^s \eta_i \|\Delta x^{k+1}_i\|^2
\leq \mathcal L^k +\sum_{i=1}^s \gamma_i^k \|\Delta x^{k}_i\|^2
+\gamma_y^k \|\Delta y^{k}\|^2\\
+ \frac{\alpha_1}{\beta } (\|B^* \Delta \omega^{k}\|^2- \|B^* \Delta \omega^{k+1} \|^2)+ \frac{\alpha_2}{\beta}L_h^2 \|\Delta y^{k+1}\|^2\\
+ \frac{\alpha_2}{\beta}\big( \bar \delta_kL_h^2 \|\Delta y^{k}\|^2 + 4L_h^2\delta_{k-1}^2\|\Delta y^{k-1}\|^2\big),
\end{split}
\end{equation}
where $\bar \delta_k = 2$ if $\delta_k=0$ for all $k$ and $4(1+\delta_k)^2$ otherwise
\end{proposition}
Now we characterize the chosen parameters for Algorithm~\ref{alg:iADMM} in the following proposition.
\begin{proposition}
\label{prop:subsequential}
Let $\eta_y$, $\gamma_y^k$, $\eta_i$, $\gamma_i^k$, $i\in [s]$, and $\bar \delta_k$ be defined in Proposition~\ref{prop:NSDP} and Proposition \ref{prop:NSDP-fory}.
Denote
\begin{equation}
\label{muk}
\begin{array}{l}
\mu= \eta_y-\frac{\alpha_2L_h^2}{\beta}.
\end{array}
\end{equation}
For $k\geq 1$, suppose the parameters are chosen such that $\mu>0$, $\eta_i>0$, and the following conditions are satisfied for some constants $0<C_x,C_y<1$:
\begin{equation}
\label{requirement}
\begin{split}
& \gamma_i^k \leq C_x \eta_i, \quad \frac{ 4\alpha_2 L_h^2\delta_{k-1}^2}{\beta}\leq C_2 \mu,
\\
&\quad
\frac{ \alpha_2 L_h^2\bar \delta_k}{\beta} + \gamma_y^k \leq C_1 \mu,
\end{split}
\end{equation}
where
$\begin{cases} C_1=C_y \,\text{and} \,C_2=0 &\mbox{if}\, \delta_k=0 \,\forall\,k,
\\
0<C_1<C_y \, \text{and}\, C_2=C_y-C_1 & \mbox{otherwise}.
\end{cases} $
(i) For $K>1$ we have
\begin{equation}
\label{eq:upperbounded}
\begin{split}
\mathcal L^{K+1} + \mu \|\Delta y^{K+1} \|^2 + \sum_{i=1}^s\eta_i \| \Delta x^{K+1}_i \|^2
+ \frac{\alpha_1}{ \beta } \|\mathcal B^* \Delta w^{K+1}\|^2 \\ + (1-C_1)\mu \|\Delta y^{K}\|^2
+ \sum_{k=1}^{K-1}\big[(1-C_y)\mu \|\Delta y^{k}\|^2 + (1-C_x)\sum_{i=1}^s\eta_i \|\Delta x^{k+1}_i\|^2 \big] \\
\leq \mathcal L^1+ \frac{\alpha_1}{\beta } \|\mathcal B^* \Delta \omega^{1}\|^2 + C_x\sum_{i=1}^s \eta_i \|\Delta x^{1}_i\|^2 + \mu \|\Delta y^{1}\|^2 + C_2 \mu \|\Delta y^{0}\|^2.
\end{split}
\end{equation}
(ii) If we use one of the following methods:
\begin{itemize}
\item we choose $\delta_k=0$ for all $k$, that is, there is no extrapolation in the update of $y$,
\item we use extrapolation in the update of $y$ and choose the parameters such that
\begin{equation}
\label{requirement2}
\begin{split}
\beta &\geq \frac{4L_h \alpha}{\sigma_{\mathcal{ B}} (1-|1-\alpha| )}, \\
\beta &\geq\frac{6\alpha L_h^2 }{\mu \sigma_{\mathcal B} (1-|1-\alpha|)} \max\Big\{1,\frac{12\delta_k^2}{1-C_1}\Big \}
\end{split}
\end{equation}
\end{itemize}
then $\{\Delta y^{k}\}$, $\{\Delta x^{k}_i\}$ and $\{\Delta \omega^{k}\}$ converge to 0.
\end{proposition}
We will assume that Algorithm~\ref{alg:iADMM} generates a bounded sequence in our subsequential and global convergence results. Let us provide a sufficient condition that guarantees this boundedness assumption in the following proposition.
\begin{proposition}
\label{prop:bounded-sequence}
If $b+ Im(\mathcal A) \subseteq Im(\mathcal B)$, $\lambda_{\min}(\mathcal B^* \mathcal B)>0 $ and $F(x) + h(y)$ is coercive over the feasible set $\{(x,y): \mathcal Ax + \mathcal By = b\}$ then the sequences $\{x^k\}$, $\{y^k\}$ and $\{\omega^k\}$ generated by Algorithm~\ref{alg:iADMM} are bounded.
\end{proposition}
It is important noting that the coercive condition of $F(x) + h(y)$ over the feasible set is weaker than the coercive condition of $F(x) + h(y)$ over $x\in\mathbb R^{\mathbf n}, y\in \mathbb R^{\mathbf q}$. Let us now present the subsequential convergence of the generated sequence.
\begin{theorem}[Subsequential convergence]
\label{thrm:subsequential}
Suppose the parameters {of Algorithm~\ref{alg:iADMM}} are chosen such that the conditions in \eqref{requirement} of Proposition~\ref{prop:subsequential} are satisfied. If the generated sequence of Algorithm~\ref{alg:iADMM} is bounded, then every limit point of the generated sequence is a critical point of $\mathcal L$.
\end{theorem}
To obtain a global convergence, we need the following Kurdyka-{\L}ojasiewicz (K{\L}) property for $F(x) + h(y)$.
\begin{definition}
\label{def:KL}
A function $\phi(\cdot)$ is said to have the K{\L} property
at $\bar{\mathbf x}\in{\rm dom}\,\partial\, \phi$ if there exists $\varsigma\in(0,+\infty]$,
a neighborhood $U$ of $\bar{\mathbf x}$ and a concave function $\Upsilon:[0,\varsigma)\to\mathbb{R}_{+}$
that is continuously differentiable on $(0,\varsigma)$, continuous at
$0$, $\Upsilon(0)=0$, and $\Upsilon'(t)>0$ for all $t\in(0,\eta),$ such that for all
$\mathbf x\in U\cap[\phi(\bar{\mathbf x})<\phi(\mathbf x)<\phi(\bar{\mathbf x})+\varsigma],$ we have
\begin{equation}
\label{ieq:KL}
\Upsilon'\left(\phi(\mathbf x)-\phi(\bar{\mathbf x})\right) \, \dist\left(0,\partial\phi(\mathbf x)\right)\geq1,
\end{equation}
where $\dist\left(0,\partial\phi(\mathbf x)\right)=\min\left\{ \|\mathbf z\|:\mathbf z\in\partial\phi(\mathbf x)\right\}$.
If $\phi(\mathbf x)$ has the K{\L}
property at each point of ${\rm dom}\, \partial\phi$ then $\phi$ is a K{\L}
function.
\end{definition}
Many non-convex non-smooth functions in practical applications belong to the class of K{\L} functions, for examples, real analytic functions, semi-algebraic functions, and locally strongly convex functions, see for example~\cite{Bochnak1998,Bolte2014}.
\begin{theorem}[Global convergence]
\label{thrm:global}
Suppose we do not use extrapolation to update $y$ (that is, $\delta_k=0$ for all $k$) and we take $\alpha=1$. Then the conditions in \eqref{requirement} become
\begin{equation}
\label{requirement-global}
\begin{split}
\gamma_i^k \leq C_x \eta_i, \quad
\frac{2 \alpha_2 L_h^2}{\beta} \leq C_y \mu,
\end{split}
\end{equation}
for some constants $0< C_x,C_y <1$.
Furthermore, we assume that (i) for any $x,z\in \mathbb R^{\mathbf n}$, $x_i\in {\rm dom} (g_i)$ we have
\begin{equation}
\label{assume:partial}
\begin{array}{ll}
\partial_{x_i} \big(f(x) + g_i(x_i)\big)=\partial_{x_i} f(x) + \partial_{x_i} g_i(x_i), \\
\partial_{x_i} \big(u_i(x_i,z) + g_i(x_i) \big)= \partial_{x_i} u_i(x_i,z) + \partial_{x_i} g_i(x_i),
\end{array}
\end{equation}
and (ii) for any $x,z$ in a bounded subset of $\mathbb R^{\mathbf n}$, if $\mathbf s_i\in\partial u_i(x_i,z)$, there exists $\xi_i\in \partial_{x_i} f(x)$ such that
\begin{equation}
\label{eq:l1}
\|\xi_i - \mathbf s_i\| \leq L_i\|x-z\| \, \text{for some constant}\, L_i.
\end{equation}
If the generated sequence of Algorithm~\ref{alg:iADMM} is bounded and $F(x) + h(y)$ has the K{\L} property, then the whole generated sequence of Algorithm~\ref{alg:iADMM} converges to a critical point of $\mathcal L$.
\end{theorem}
We refer the readers to \cite[Corollary 10.9]{RockWets98} for a sufficient condition for~\eqref{assume:partial} (see supplementary material for more details). Some specific examples that satisfy \eqref{assume:partial} include: (i) $g_i=0$, (ii) the functions $x_i\mapsto f(x)$ and $x_i\mapsto u_i(x_i,z)$ are strictly differentiable (see \cite[Exercise 10.10]{RockWets98}), (iii) the functions $x_i\mapsto f(x)$ and $x_i\mapsto u_i(x_i,z)$ are convex and the relative interior qualification conditions are satisfied: ${\rm ri} ({\rm dom} (f(\cdot,x_{\ne i})) \cap {\rm ri} ({\rm dom} g_i) \ne \emptyset $ and ${\rm ri} ({\rm dom} (g(\cdot,z)) \cap {\rm ri} ({\rm dom} g_i) \ne \emptyset $. We note that although the condition in \eqref{eq:l1} is necessary for our convergence proof, the Lipschitz constant $L_i$ does not influence how to choose the parameters in our framework. We end this section by noting that a convergence rate for the generated sequence of iADMM can be derived using the same technique as in the proof of \cite[Theorem~2]{Attouch2009}. Some examples of using the technique of \cite[Theorem~2]{Attouch2009} to derive the convergence rate include \cite[Theorem~2.9]{Xu2013} and \cite[Theorem~3]{Hien_ICML2020}. Other than the convergence rate which appears to be the same in different papers using the technique in \cite{Attouch2009}, determining the K{\L} exponent, that is, the coefficient $\mathbf a$ when $\Upsilon(t) = c t^{1-\mathbf a}$, where $c$ is a constant, is an active and challenging topic. The type of the convergence rate depends on the value of $\mathbf a$. Specifically, when $\mathbf a=0$, the algorithm converges after a finite number of steps; when $\mathbf a\in (0,1/2]$ \ngi{it has} linear convergence, and when $\mathbf a \in (1/2, 1]$ \ngi{it has} sublinear convergence. Determining the value of $\mathbf a$ is out of the scope of this paper.
\section{Numerical results}
\label{sec:numerical}
\begin{figure*}[h!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.492\textwidth]{Images/error_hopkins155} &
\includegraphics[width=0.4954\textwidth]{Images/obj_hopkins155} \\
\includegraphics[width=0.492\textwidth]{Images/error_umist10} &
\includegraphics[width=0.4954\textwidth]{Images/obj_umist10} \\
\includegraphics[width=0.492\textwidth]{Images/error_yaleb10} &
\includegraphics[width=0.5\textwidth]{Images/obj_yaleb10}
\end{tabular}
\caption{Evolution of the value of the segmentation error rate and the objective function value with respect to time. For Hopkins155, the results are the average values over 156 sequences
\label{fig:model1}}
\end{center}
\end{figure*}
In this section, we apply iADMM to solve a latent low-rank representation problem of the form of Problem~\eqref{LRR}; see Section~\ref{sec:intro-LRR}. Specifically, we choose $r_1(t) = \lambda_1 t$, $r_3(Z) = \frac{1}{2}\|Z\|^2$ (hence $Z$ represents a Gaussian noise), and consider a \emph{nonconvex} regularization function for $Y$, $r_2(Y) = \lambda\sum_{i=1}^q\phi(\|Y_i\|_2)$, where $Y_i$ is the $i$-th column of $Y$ and $\phi(t) = 1 - \exp(-\theta t)$ \cite{brafea}. In the upcoming experiments, we choose $A_1 = DP_1$ and $A_2 = P_2^*D$ as proposed in \cite{Liu2011}, where $P_1$ and $P_2$ are computed by orthogonalizing the columns of $D^*$ and $D$, respectively.
Problem \eqref{LRR} in this case takes the form of \eqref{model} with $B$ being the identity operator, $b$ being the data set $D$, $x_1$ and $x_2$ being the matrices $X$ and $Y$, $y$ being the matrix $Z$, $f(X,Y) = \lambda_1\|X\|_* + r_2(Y)$, $g_i = 0$ and $h(Z) = \frac{1}{2}\|Z\|^2$.
We choose {the following} block surrogate functions for $f$: $u_1(X,X^k,Y^k) = \lambda_1\|X\|_* + r_2(Y^k)$, $u_2(Y,X^{k+1},Y^k) = r_2(Y^k) + \sum_{i=1}^q\varsigma_i^k\|Y_i\|_2 + \lambda_1\|X^{k+1}\|_*$,
where $\varsigma_i^k \in \lambda\nabla\phi(\|Y_i^k\|_2)$.
Obviously $u_1$ satisfies Assumption \ref{assump:Lipschitz_ui} and $u_2$ satisfies Assumption \ref{assump:Lipschitz_ui} (i). Since $\phi$ is continuously differentiable with Lipschitz gradient on $[0,+\infty)$ and the Euclidean norm is Lipschitz continuous, it follows from Section 4.5 of \cite{Titan2020} that $u_2$ satisfies
Assumption~\ref{assump:Lipschitz_ui}~(ii).
For updating $X$, according to the update \eqref{eq:xi-updatenew}, $X^{k+1}$ is computed by solving the following nuclear norm problem
\begin{equation}\label{subLRR}
\begin{split}
\min_X \lambda_1\|X\|_* + \bigg\langle A_1^* \Big(\beta(A_1\bar X^k+ Y^kA_2 + Z^k - D)
+ W^k\Big), X\bigg\rangle
+ \frac{\kappa_1\beta}{2}\|X - \bar X^k\|^2,
\end{split}
\end{equation}
where $\kappa_1 \geq \|A_1^*A_1\|$ and $\bar X^k = X^k + \zeta_1^k(X^k - X^{k-1})$. The sub-problem \eqref{subLRR} has a closed-form solution given by
\begin{equation*}
X^{k+1} = US_{\lambda_1/\kappa_1\beta} V^T,
\end{equation*}
where $USV^T$ is the SVD of $\bar X^k - A_1^*(A_1\bar X^k+ Y^kA_2 + Z^k - D+ W^k)/\kappa_1\beta$ and $S_{\lambda_1/\kappa_1\beta} = \text{diag}([S_{ii} - \lambda_1/\kappa_1\beta]_+)$, {where $\text{diag}(u)$ is a diagonal matrix whose diagonal elements are the entries of $u$, and $[.]_+$ is the projection onto the nonnegative orthant.}
The update \eqref{eq:y_update} for $Y$ is
\begin{equation*}
\begin{split}
Y^{k+1}\in\arg\min_Y \sum_{i=1}^q\varsigma_i^k\|Y_i\|_2+ \langle (W^k + \beta(A_1X^{k+1} \\+ \bar Y^kA_2 + Z^k - D))A_2^*, Y\rangle
+ \frac{\kappa_2\beta}{2}\|Y - \bar Y^k \|^2,
\end{split}
\end{equation*}
where $\kappa_2 \geq \|A_2A_2^*\|$ and $\bar Y^k = Y^k + \zeta_2^k(Y^k - Y^{k-1})$. The sub-problem above has a closed-form solution
\begin{equation*}
Y^{k+1}_{i} =
\left[
\|P_{i}^k\| - \varsigma_i^k/(\kappa_2\beta) \right]_+\frac{P_{i}^k}{\|P_{i}^k\|},
\end{equation*}
where $P_{i}^k$ is the $i$-th column of $\bar Y^k - (A_1X^{k+1} + \bar Y^kA_2 + Z^k - D)/\kappa_2 - W^k/(\kappa_2\beta)$.
The updates \eqref{eq:y_update} and \eqref{eq:omega_update} for $Z$ and $W$ are respectively given by
\begin{equation*}
\begin{split}
Z^{k+1} & = -(W^k +\beta(A_1X^{k+1} + y^{k+1}A_2 - D))/(1+\beta), \\
W^{k+1}& = W^k + \alpha\beta(A_1X^{k+1} + y^{k+1}A_2 + Z^{k+1}- D).
\end{split}
\end{equation*}
Let us determine the parameters. Note that $L_h = 1$, $\sigma_{\mathcal B} = 1$, and $\delta_k = 0$. Since $h(Z)$ is convex and we do not apply extrapolation for $Z$, hence by Proposition~\ref{prop:NSDP-fory} we have $\eta_y=\frac12$ and $\gamma_y^k=0$. Since $\|X\|_*$ and $\sum_{i=1}^q\varsigma_i^k\|Y_i\|_2$ are convex, we choose $\kappa_1 = \|A_1^*A_1\|$, $\kappa_2 = \|A_2A_2^*\|$ and the conditions in~\eqref{requirement-global} become
$\zeta_i^k \leq \sqrt{C_x}$ (for $i=1,2$) and $\frac{(2+C_y)\alpha_2}{\beta} \leq \frac{C_y}{2}$.
In our experiments, we choose
$C_x= 1 - 10^{-15}$,
$\alpha = 1$,
{$C_y = 1 - 10^{-6}$},
$\beta = 2(2 + C_y)\alpha_2/C_y$, $a_0=1$,
\mbox{$a_k=\frac12(1+\sqrt{1+4a_{k-1}^2})$}, and
$\zeta_i^k =\min\Big\{\frac{a_{k-1}-1}{a_k},\sqrt{C_x }\Big\}$.
We compare iADMM without extrapolation denoted by ADMM-mm, and iADMM with the extrapolation denoted by iADMM-mm, with a linearized ADMM that is only different from ADMM-mm for updating $Y$. In particular, the linearizedADMM method updates $Y$ by solving the following nonconvex sub-problems
\begin{equation*}
\min -\lambda\exp(\|Y_i\|_2) + \frac{\kappa_2\beta}{2}\|Y_i - V_i^k\|^2,
\end{equation*}
where $V^k_i$ is the $i$-the column of $X^k - (W^k + \beta(A_1X^{k+1} \\+ \bar Y^kA_2 + Z^k - D))A_2^*/(\kappa_2\beta)$. Since the sub-problems above do not have closed-form solutions, we employ {an} MM scheme to solve them.
\begin{table}[h!]
\centering
\caption{Comparison of segmentation error rate and final objective function values obtained within the allotted time.
Bold values indicate the best results}\label{results}
\vspace{0.05in}
\begin{tabular}{@{}clll@{}}
\toprule
& \multirow{2}{*}{Method} & Error & Obj. value \\
& & mean $\pm$ std & mean $\pm$ std \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{90}{Hopkins}} & linearizedADMM & 0.1579 $\pm$ 0.1550 & 3.0254 $\pm$ 2.4189 \\
& ADMM-mm & 0.1472 $\pm$ 0.1513 & 1.8081 $\pm$ 1.6674 \\
& iADMM-mm & \textbf{0.0562} $\pm$ 0.1006 & \textbf{0.2023} $\pm$ 0.1062 \\
\hline
\multirow{3}{*}{\rotatebox[origin=c]{90}{Umist}} & linearizedADMM & 0.5170 & 1.0838$\times10^9$ \\
& ADMM-mm & 0.5170 & 1.0167$\times10^9$ \\
& iADMM-mm & \textbf{0.2604} & \textbf{0.1694}$\times10^9$ \\
\hline
\multirow{3}{*}{\rotatebox[origin=c]{90}{Yaleb}} & linearizedADMM & 0.7656 & 5.2317$\times10^3$ \\
& ADMM-mm & 0.7047 & 4.4829$\times10^3$ \\
& iADMM-mm & \textbf{0.1984} & \textbf{0.6951}$\times10^3$ \\
\hline
\end{tabular}
\end{table}
To examine the performance of the comparative algorithms, we consider subspace segmentation tasks. In particular, after obtaining $X^*$, we follow the setting in \cite{Lui2013} to construct the affinity matrix $Q$ by $Q_{ij} = (\tilde{U}\tilde{U}^T)_{ij}$, where $\tilde{U}$ is formed by $U^*(\Sigma^*)^{1/2}$ with normalized rows and $U^*\Sigma^*(V^*)^T$ being the SVD of $X^*$. Finally, we apply the Normalized Cuts \cite{Shi2000} on $W$ to cluster the data into groups.
The experiments are run on three data sets: Hopkins 155, extended Yale B and Umist. Hopskins 155 consists of 156
sequences, each of which has from 39 to 550 vectors drawn from two or three motions (one motion corresponds to one subspace). Each sequence
is a sole segmentation task and thus there are 156
clustering tasks in total. Yale B contains 2414 frontal face
images of 38 classes while Umist contains
564 images of 20 classes. To avoid computational issue when computing the segmentation error rate, we construct clustering tasks by using only the first 10 classes of these two data sets as proposed in \cite{Lu2015}.
All tests are preformed using Matlab
R2019a on a PC 2.3 GHz Intel Core i5
of 8GB RAM. The code is available from \url{https://github.com/nhatpd/iADMM}
In our experiments, we choose $\theta = 5$, $\lambda_1 = \lambda = 0.01$ for Hopkins 155, and $\lambda_1 = \lambda = 1$ for the two other data sets. We note that we do not optimize numerical results by tweaking the parameters as this is beyond the scope of this work.
It is important noting that we evaluate the algorithms on the same models. We set the initial points to zero. We run each algorithm 10, 300, and 500 seconds for each sequence of Hopkins 155, Umist10, and Yaleb10, respectively. We plot the curves of the value of the segmentation error rate and the objective function value versus the training time in Figure~\ref{fig:model1}, and report the final values in Table~\ref{results}. Since there are 156 sequences (data sets) in Hopkins 155, we plot the average values, and report the final average results and standard deviation over these sequences.
We observe that iADMM-mm converges the fastest on all the data sets, providing a significant acceleration of ADMM-mm. iADMM-mm achieves not only the best final objective function values but also the best segmentation error rates.
This illustrates the usefulness of the acceleration technique.
In addition, ADMM-mm outperforms linearizedADMM which illustrates the usefulness of properly choosing a proper surrogate function.
\section{Conclusion}
We have analysed iADMM, a framework of inertial alternating direction methods of multipliers, for solving a class of nonconvex nonsmooth optimization problem with linear constraints. The preliminary computational results in solving a class of nonconvex low-rank representation problems not only show the efficacy of using inertial terms for ADMM but also show the advantage of using suitable block surrogate functions that may lead to closed-form solutions in the block update of ADMM. We conclude the paper by mentioning two important questions that we consider as a future research directions:
\begin{itemize}
\item Can we extend the cyclic update rule of iADMM to randomized/non-cyclic setting?
\item To guarantee the global convergence, iADMM does not allow extrapolation in the update of $y$; see Theorem~\ref{thrm:global}. Can we extend the analysis to allow the extrapolation in the update of $y$?
\end{itemize}
\section{\hskip 0em~~#1}\vspace{-1pt}}
\newcommand{\Subsection}[1]{\vspace{-8pt}\subsection{\hskip -1em~~#1}\vspace{-3pt}}
\newcommand{\overset{\raisebox{-1mm}{\scriptsize$\mathrm{c}$}}{\le}}{\overset{\raisebox{-1mm}{\scriptsize$\mathrm{c}$}}{\le}}
\newcommand{\overset{\mathrm{c}}{\ge}}{\overset{\mathrm{c}}{\ge}}
\newcommand{\mathbbm{1}}{\mathbbm{1}}
\newcommand{\overline{\nabla}}{\overline{\nabla}}
\newcommand{\mathrm{rand}}{\mathrm{rand}}
\newcommand{\mathrm{opt}}{\mathrm{opt}}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\newcommand*{\samethanks}[1][\value{footnote}]{\footnotemark[#1]}
\newcommand{\ep}[1]{\textup{(}#1\textup{)}}
\newcommand{\overline{\bf{a}}}{\overline{\bf{a}}}
\newcommand{\overline{\bf b}}{\overline{\bf b}}
\newcommand{\overline{\bf c}}{\overline{\bf c}}
\newcommand{\overline{\bf d}}{\overline{\bf d}}
\newcommand{\overline{\bf e}}{\overline{\bf e}}
\newcommand{\overline{\bf f}}{\overline{\bf f}}
\newcommand{\overline{\bf g}}{\overline{\bf g}}
\newcommand{\overline{\bf h}}{\overline{\bf h}}
\newcommand{\overline{\bf i}}{\overline{\bf i}}
\newcommand{\overline{\bf j}}{\overline{\bf j}}
\newcommand{\overline{\bf k}}{\overline{\bf k}}
\newcommand{\overline{\bf l}}{\overline{\bf l}}
\newcommand{\overline{\bf m}}{\overline{\bf m}}
\newcommand{\overline{\bf n}}{\overline{\bf n}}
\newcommand{\overline{\bf o}}{\overline{\bf o}}
\newcommand{\overline{\bf p}}{\overline{\bf p}}
\newcommand{\overline{\bf q}}{\overline{\bf q}}
\newcommand{\overline{\bf r}}{\overline{\bf r}}
\newcommand{\overline{\bf s}}{\overline{\bf s}}
\newcommand{\overline{\bf t}}{\overline{\bf t}}
\newcommand{\overline{\bf u}}{\overline{\bf u}}
\newcommand{\overline{\bf v}}{\overline{\bf v}}
\newcommand{\overline{\bf w}}{\overline{\bf w}}
\newcommand{\overline{\bf x}}{\overline{\bf x}}
\newcommand{\overline{\bf y}}{\overline{\bf y}}
\newcommand{\overline{\bf z}}{\overline{\bf z}}
\newcommand{\overline{\bm{\lambda}}}{\overline{\bm{\lambda}}}
\newcommand{\overline{\bf A}}{\overline{\bf A}}
\newcommand{\overline{\bf B}}{\overline{\bf B}}
\newcommand{\overline{\bf C}}{\overline{\bf C}}
\newcommand{\overline{\bf D}}{\overline{\bf D}}
\newcommand{\overline{\bf E}}{\overline{\bf E}}
\newcommand{\overline{\bf F}}{\overline{\bf F}}
\newcommand{\overline{\bf G}}{\overline{\bf G}}
\newcommand{\overline{\bf H}}{\overline{\bf H}}
\newcommand{\overline{\bf I}}{\overline{\bf I}}
\newcommand{\overline{\bf J}}{\overline{\bf J}}
\newcommand{\overline{\bf K}}{\overline{\bf K}}
\newcommand{\overline{\bf L}}{\overline{\bf L}}
\newcommand{\overline{\bf M}}{\overline{\bf M}}
\newcommand{\overline{\bf N}}{\overline{\bf N}}
\newcommand{\overline{\bf O}}{\overline{\bf O}}
\newcommand{\overline{\bf P}}{\overline{\bf P}}
\newcommand{\overline{\bf Q}}{\overline{\bf Q}}
\newcommand{\overline{\bf R}}{\overline{\bf R}}
\newcommand{\overline{\bf S}}{\overline{\bf S}}
\newcommand{\overline{\bf T}}{\overline{\bf T}}
\newcommand{\overline{\bf U}}{\overline{\bf U}}
\newcommand{\overline{\bf V}}{\overline{\bf V}}
\newcommand{\overline{\bf W}}{\overline{\bf W}}
\newcommand{\overline{\bf X}}{\overline{\bf X}}
\newcommand{\overline{\bf Y}}{\overline{\bf Y}}
\newcommand{\overline{\bf Z}}{\overline{\bf Z}}
\newcommand{\overline{\ell}}{\overline{\ell}}
\newcommand{\overline{\kappa}}{\overline{\kappa}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{B}}{\mathcal{B}}
\newcommand{\mathcal{C}}{\mathcal{C}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathcal{E}}{\mathcal{E}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\newcommand{\mathcal{G}}{\mathcal{G}}
\newcommand{\mathcal{H}}{\mathcal{H}}
\newcommand{\mathcal{I}}{\mathcal{I}}
\newcommand{\mathcal{J}}{\mathcal{J}}
\newcommand{\mathcal{K}}{\mathcal{K}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{O}}{\mathcal{O}}
\newcommand{\mathcal{P}}{\mathcal{P}}
\newcommand{\mathcal{Q}}{\mathcal{Q}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{U}}{\mathcal{U}}
\newcommand{\mathcal{V}}{\mathcal{V}}
\newcommand{\mathcal{W}}{\mathcal{W}}
\newcommand{\mathcal{X}}{\mathcal{X}}
\newcommand{\mathcal{Y}}{\mathcal{Y}}
\newcommand{\mathcal{Z}}{\mathcal{Z}}
\newcommand{\hatcalA}{\widehat{\mathcal{A}}}
\newcommand{\hatcalB}{\widehat{\mathcal{B}}}
\newcommand{\hatcalC}{\widehat{\mathcal{C}}}
\newcommand{\hatcalD}{\widehat{\mathcal{D}}}
\newcommand{\hatcalE}{\widehat{\mathcal{E}}}
\newcommand{\hatcalF}{\widehat{\mathcal{F}}}
\newcommand{\hatcalG}{\widehat{\mathcal{G}}}
\newcommand{\hatcalH}{\widehat{\mathcal{H}}}
\newcommand{\hatcalI}{\widehat{\mathcal{I}}}
\newcommand{\hatcalJ}{\widehat{\mathcal{J}}}
\newcommand{\hatcalK}{\widehat{\mathcal{K}}}
\newcommand{\hatcalL}{\widehat{\mathcal{L}}}
\newcommand{\hatcalM}{\widehat{\mathcal{M}}}
\newcommand{\hatcalN}{\widehat{\mathcal{N}}}
\newcommand{\hatcalO}{\widehat{\mathcal{O}}}
\newcommand{\hatcalP}{\widehat{\mathcal{P}}}
\newcommand{\hatcalQ}{\widehat{\mathcal{Q}}}
\newcommand{\hatcalR}{\widehat{\mathcal{R}}}
\newcommand{\hatcalS}{\widehat{\mathcal{S}}}
\newcommand{\hatcalT}{\widehat{\mathcal{T}}}
\newcommand{\hatcalU}{\widehat{\mathcal{U}}}
\newcommand{\hatcalV}{\widehat{\mathcal{V}}}
\newcommand{\hatcalW}{\widehat{\mathcal{W}}}
\newcommand{\hatcalX}{\widehat{\mathcal{X}}}
\newcommand{\hatcalY}{\widehat{\mathcal{Y}}}
\newcommand{\hatcalZ}{\widehat{\mathcal{Z}}}
\newcommand{\tilcalA}{\widetilde{\mathcal{A}}}
\newcommand{\tilcalB}{\widetilde{\mathcal{B}}}
\newcommand{\tilcalC}{\widetilde{\mathcal{C}}}
\newcommand{\tilcalD}{\widetilde{\mathcal{D}}}
\newcommand{\tilcalE}{\widetilde{\mathcal{E}}}
\newcommand{\tilcalF}{\widetilde{\mathcal{F}}}
\newcommand{\tilcalG}{\widetilde{\mathcal{G}}}
\newcommand{\tilcalH}{\widetilde{\mathcal{H}}}
\newcommand{\tilcalI}{\widetilde{\mathcal{I}}}
\newcommand{\tilcalJ}{\widetilde{\mathcal{J}}}
\newcommand{\tilcalK}{\widetilde{\mathcal{K}}}
\newcommand{\tilcalL}{\widetilde{\mathcal{L}}}
\newcommand{\tilcalM}{\widetilde{\mathcal{M}}}
\newcommand{\tilcalN}{\widetilde{\mathcal{N}}}
\newcommand{\tilcalO}{\widetilde{\mathcal{O}}}
\newcommand{\tilcalP}{\widetilde{\mathcal{P}}}
\newcommand{\tilcalQ}{\widetilde{\mathcal{Q}}}
\newcommand{\tilcalR}{\widetilde{\mathcal{R}}}
\newcommand{\tilcalS}{\widetilde{\mathcal{S}}}
\newcommand{\tilcalT}{\widetilde{\mathcal{T}}}
\newcommand{\tilcalU}{\widetilde{\mathcal{U}}}
\newcommand{\tilcalV}{\widetilde{\mathcal{V}}}
\newcommand{\tilcalW}{\widetilde{\mathcal{W}}}
\newcommand{\tilcalX}{\widetilde{\mathcal{X}}}
\newcommand{\tilcalY}{\widetilde{\mathcal{Y}}}
\newcommand{\tilcalZ}{\widetilde{\mathcal{Z}}}
\newcommand{\barcalA}{\overline{\mathcal{A}}}
\newcommand{\barcalB}{\overline{\mathcal{B}}}
\newcommand{\barcalC}{\overline{\mathcal{C}}}
\newcommand{\barcalD}{\overline{\mathcal{D}}}
\newcommand{\barcalE}{\overline{\mathcal{E}}}
\newcommand{\barcalF}{\overline{\mathcal{F}}}
\newcommand{\barcalG}{\overline{\mathcal{G}}}
\newcommand{\barcalH}{\overline{\mathcal{H}}}
\newcommand{\barcalI}{\overline{\mathcal{I}}}
\newcommand{\barcalJ}{\overline{\mathcal{J}}}
\newcommand{\barcalK}{\overline{\mathcal{K}}}
\newcommand{\barcalL}{\overline{\mathcal{L}}}
\newcommand{\barcalM}{\overline{\mathcal{M}}}
\newcommand{\barcalN}{\overline{\mathcal{N}}}
\newcommand{\barcalO}{\overline{\mathcal{O}}}
\newcommand{\barcalP}{\overline{\mathcal{P}}}
\newcommand{\barcalQ}{\overline{\mathcal{Q}}}
\newcommand{\barcalR}{\overline{\mathcal{R}}}
\newcommand{\barcalS}{\overline{\mathcal{S}}}
\newcommand{\barcalT}{\overline{\mathcal{T}}}
\newcommand{\barcalU}{\overline{\mathcal{U}}}
\newcommand{\barcalV}{\overline{\mathcal{V}}}
\newcommand{\barcalW}{\overline{\mathcal{W}}}
\newcommand{\barcalX}{\overline{\mathcal{X}}}
\newcommand{\barcalY}{\overline{\mathcal{Y}}}
\newcommand{\barcalZ}{\overline{\mathcal{Z}}}
\newcommand{\mathbf{a}}{\mathbf{a}}
\newcommand{\mathbf{A}}{\mathbf{A}}
\newcommand{\mathbf{b}}{\mathbf{b}}
\newcommand{\mathbf{B}}{\mathbf{B}}
\newcommand{\mathbf{c}}{\mathbf{c}}
\newcommand{\mathbf{C}}{\mathbf{C}}
\newcommand{\mathbf{d}}{\mathbf{d}}
\newcommand{\mathbf{D}}{\mathbf{D}}
\newcommand{\mathbf{e}}{\mathbf{e}}
\newcommand{\mathbf{E}}{\mathbf{E}}
\newcommand{\mathbf{f}}{\mathbf{f}}
\newcommand{\mathbf{F}}{\mathbf{F}}
\newcommand{\mathbf{g}}{\mathbf{g}}
\newcommand{\mathbf{G}}{\mathbf{G}}
\newcommand{\mathbf{h}}{\mathbf{h}}
\newcommand{\mathbf{H}}{\mathbf{H}}
\newcommand{\mathbf{i}}{\mathbf{i}}
\newcommand{\mathbf{I}}{\mathbf{I}}
\newcommand{\mathbf{j}}{\mathbf{j}}
\newcommand{\mathbf{J}}{\mathbf{J}}
\newcommand{\mathbf{k}}{\mathbf{k}}
\newcommand{\mathbf{K}}{\mathbf{K}}
\newcommand{\mathbf{l}}{\mathbf{l}}
\newcommand{\mathbf{L}}{\mathbf{L}}
\newcommand{\mathbf{m}}{\mathbf{m}}
\newcommand{\mathbf{M}}{\mathbf{M}}
\newcommand{\mathbf{n}}{\mathbf{n}}
\newcommand{\mathbf{N}}{\mathbf{N}}
\newcommand{\mathbf{o}}{\mathbf{o}}
\newcommand{\mathbf{O}}{\mathbf{O}}
\newcommand{\mathbf{p}}{\mathbf{p}}
\newcommand{\mathbf{P}}{\mathbf{P}}
\newcommand{\mathbf{q}}{\mathbf{q}}
\newcommand{\mathbf{Q}}{\mathbf{Q}}
\newcommand{\mathbf{r}}{\mathbf{r}}
\newcommand{\mathbf{R}}{\mathbf{R}}
\newcommand{\mathbf{s}}{\mathbf{s}}
\newcommand{\mathbf{S}}{\mathbf{S}}
\newcommand{\mathbf{t}}{\mathbf{t}}
\newcommand{\mathbf{T}}{\mathbf{T}}
\newcommand{\mathbf{u}}{\mathbf{u}}
\newcommand{\mathbf{U}}{\mathbf{U}}
\newcommand{\mathbf{v}}{\mathbf{v}}
\newcommand{\mathbf{V}}{\mathbf{V}}
\newcommand{\mathbf{w}}{\mathbf{w}}
\newcommand{\mathbf{W}}{\mathbf{W}}
\newcommand{\mathbf{x}}{\mathbf{x}}
\newcommand{\mathbf{X}}{\mathbf{X}}
\newcommand{\mathbf{y}}{\mathbf{y}}
\newcommand{\mathbf{Y}}{\mathbf{Y}}
\newcommand{\mathbf{z}}{\mathbf{z}}
\newcommand{\mathbf{Z}}{\mathbf{Z}}
\newcommand{\mathrm{a}}{\mathrm{a}}
\newcommand{\mathrm{A}}{\mathrm{A}}
\newcommand{\mathrm{b}}{\mathrm{b}}
\newcommand{\mathrm{B}}{\mathrm{B}}
\newcommand{\mathrm{c}}{\mathrm{c}}
\newcommand{\mathrm{C}}{\mathrm{C}}
\newcommand{\mathrm{d}}{\mathrm{d}}
\newcommand{\mathrm{D}}{\mathrm{D}}
\newcommand{\mathrm{e}}{\mathrm{e}}
\newcommand{\mathrm{E}}{\mathrm{E}}
\newcommand{\mathrm{f}}{\mathrm{f}}
\newcommand{\mathrm{F}}{\mathrm{F}}
\newcommand{\mathrm{g}}{\mathrm{g}}
\newcommand{\mathrm{G}}{\mathrm{G}}
\newcommand{\mathrm{h}}{\mathrm{h}}
\newcommand{\mathrm{H}}{\mathrm{H}}
\newcommand{\mathrm{i}}{\mathrm{i}}
\newcommand{\mathrm{I}}{\mathrm{I}}
\newcommand{\mathrm{j}}{\mathrm{j}}
\newcommand{\mathrm{J}}{\mathrm{J}}
\newcommand{\mathrm{k}}{\mathrm{k}}
\newcommand{\mathrm{K}}{\mathrm{K}}
\newcommand{\mathrm{l}}{\mathrm{l}}
\newcommand{\mathrm{L}}{\mathrm{L}}
\newcommand{\mathrm{m}}{\mathrm{m}}
\newcommand{\mathrm{M}}{\mathrm{M}}
\newcommand{\mathrm{n}}{\mathrm{n}}
\newcommand{\mathrm{N}}{\mathrm{N}}
\newcommand{\mathrm{o}}{\mathrm{o}}
\newcommand{\mathrm{O}}{\mathrm{O}}
\newcommand{\mathrm{p}}{\mathrm{p}}
\newcommand{\mathrm{P}}{\mathrm{P}}
\newcommand{\mathrm{q}}{\mathrm{q}}
\newcommand{\mathrm{Q}}{\mathrm{Q}}
\newcommand{\mathrm{r}}{\mathrm{r}}
\newcommand{\mathrm{R}}{\mathrm{R}}
\newcommand{\mathrm{s}}{\mathrm{s}}
\newcommand{\mathrm{S}}{\mathrm{S}}
\newcommand{\mathrm{t}}{\mathrm{t}}
\newcommand{\mathrm{T}}{\mathrm{T}}
\newcommand{\mathrm{u}}{\mathrm{u}}
\newcommand{\mathrm{U}}{\mathrm{U}}
\newcommand{\mathrm{v}}{\mathrm{v}}
\newcommand{\mathrm{V}}{\mathrm{V}}
\newcommand{\mathrm{w}}{\mathrm{w}}
\newcommand{\mathrm{W}}{\mathrm{W}}
\newcommand{\mathrm{x}}{\mathrm{x}}
\newcommand{\mathrm{X}}{\mathrm{X}}
\newcommand{\mathrm{y}}{\mathrm{y}}
\newcommand{\mathrm{Y}}{\mathrm{Y}}
\newcommand{\mathrm{z}}{\mathrm{z}}
\newcommand{\mathrm{Z}}{\mathrm{Z}}
\newcommand{\mathbb{A}}{\mathbb{A}}
\newcommand{\mathbb{B}}{\mathbb{B}}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{\mathbb{D}}{\mathbb{D}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbb{F}}{\mathbb{F}}
\newcommand{\mathbb{G}}{\mathbb{G}}
\newcommand{\mathbb{H}}{\mathbb{H}}
\newcommand{\mathbb{I}}{\mathbb{I}}
\newcommand{\mathbb{J}}{\mathbb{J}}
\newcommand{\mathbb{K}}{\mathbb{K}}
\newcommand{\mathbb{L}}{\mathbb{L}}
\newcommand{\mathbb{M}}{\mathbb{M}}
\newcommand{\mathbb{N}}{\mathbb{N}}
\newcommand{\mathbb{O}}{\mathbb{O}}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathbb{Q}}{\mathbb{Q}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{S}}{\mathbb{S}}
\newcommand{\mathbb{T}}{\mathbb{T}}
\newcommand{\mathbb{U}}{\mathbb{U}}
\newcommand{\mathbb{V}}{\mathbb{V}}
\newcommand{\mathbb{W}}{\mathbb{W}}
\newcommand{\mathbb{X}}{\mathbb{X}}
\newcommand{\mathbb{Y}}{\mathbb{Y}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{\vphantom{\bbR}\smash{\overline \bbR}}{\vphantom{\mathbb{R}}\smash{\overline \mathbb{R}}}
\newcommand{\mathfrak{A}}{\mathfrak{A}}
\newcommand{\mathfrak{B}}{\mathfrak{B}}
\newcommand{\mathfrak{C}}{\mathfrak{C}}
\newcommand{\mathfrak{D}}{\mathfrak{D}}
\newcommand{\mathfrak{E}}{\mathfrak{E}}
\newcommand{\mathfrak{F}}{\mathfrak{F}}
\newcommand{\mathfrak{G}}{\mathfrak{G}}
\newcommand{\mathfrak{H}}{\mathfrak{H}}
\newcommand{\mathfrak{I}}{\mathfrak{I}}
\newcommand{\mathfrak{J}}{\mathfrak{J}}
\newcommand{\mathfrak{K}}{\mathfrak{K}}
\newcommand{\mathfrak{L}}{\mathfrak{L}}
\newcommand{\mathfrak{M}}{\mathfrak{M}}
\newcommand{\mathfrak{N}}{\mathfrak{N}}
\newcommand{\mathfrak{O}}{\mathfrak{O}}
\newcommand{\mathfrak{P}}{\mathfrak{P}}
\newcommand{\mathfrak{Q}}{\mathfrak{Q}}
\newcommand{\mathfrak{R}}{\mathfrak{R}}
\newcommand{\mathfrak{S}}{\mathfrak{S}}
\newcommand{\mathfrak{T}}{\mathfrak{T}}
\newcommand{\mathfrak{U}}{\mathfrak{U}}
\newcommand{\mathfrak{V}}{\mathfrak{V}}
\newcommand{\mathfrak{W}}{\mathfrak{W}}
\newcommand{\mathfrak{X}}{\mathfrak{X}}
\newcommand{\mathfrak{Y}}{\mathfrak{Y}}
\newcommand{\mathfrak{Z}}{\mathfrak{Z}}
\newcommand{\mathscr{A}}{\mathscr{A}}
\newcommand{\mathscr{B}}{\mathscr{B}}
\newcommand{\mathscr{C}}{\mathscr{C}}
\newcommand{\mathscr{D}}{\mathscr{D}}
\newcommand{\mathscr{E}}{\mathscr{E}}
\newcommand{\mathscr{F}}{\mathscr{F}}
\newcommand{\mathscr{G}}{\mathscr{G}}
\newcommand{\mathscr{H}}{\mathscr{H}}
\newcommand{\mathscr{I}}{\mathscr{I}}
\newcommand{\mathscr{J}}{\mathscr{J}}
\newcommand{\mathscr{K}}{\mathscr{K}}
\newcommand{\mathscr{L}}{\mathscr{L}}
\newcommand{\mathscr{M}}{\mathscr{M}}
\newcommand{\mathscr{N}}{\mathscr{N}}
\newcommand{\mathscr{O}}{\mathscr{O}}
\newcommand{\mathscr{P}}{\mathscr{P}}
\newcommand{\mathscr{Q}}{\mathscr{Q}}
\newcommand{\mathscr{R}}{\mathscr{R}}
\newcommand{\mathscr{S}}{\mathscr{S}}
\newcommand{\mathscr{T}}{\mathscr{T}}
\newcommand{\mathscr{U}}{\mathscr{U}}
\newcommand{\mathscr{V}}{\mathscr{V}}
\newcommand{\mathscr{W}}{\mathscr{W}}
\newcommand{\mathscr{X}}{\mathscr{X}}
\newcommand{\mathscr{Y}}{\mathscr{Y}}
\newcommand{\mathscr{Z}}{\mathscr{Z}}
\DeclareMathAlphabet{\mathbsf}{OT1}{cmss}{bx}{n}
\DeclareMathAlphabet{\mathssf}{OT1}{cmss}{m}{sl
\newcommand{\mathsf{a}}{\mathsf{a}}
\newcommand{\mathsf{A}}{\mathsf{A}}
\newcommand{\mathbsf{a}}{\mathbsf{a}}
\newcommand{\mathbsf{A}}{\mathbsf{A}}
\newcommand{\mathsf{b}}{\mathsf{b}}
\newcommand{\mathsf{B}}{\mathsf{B}}
\newcommand{\mathbsf{b}}{\mathbsf{b}}
\newcommand{\mathbsf{B}}{\mathbsf{B}}
\newcommand{\mathsf{c}}{\mathsf{c}}
\newcommand{\mathsf{C}}{\mathsf{C}}
\newcommand{\mathbsf{c}}{\mathbsf{c}}
\newcommand{\mathbsf{C}}{\mathbsf{C}}
\newcommand{\mathsf{d}}{\mathsf{d}}
\newcommand{\mathsf{D}}{\mathsf{D}}
\newcommand{\mathbsf{d}}{\mathbsf{d}}
\newcommand{\mathbsf{D}}{\mathbsf{D}}
\newcommand{\mathsf{e}}{\mathsf{e}}
\newcommand{\mathsf{E}}{\mathsf{E}}
\newcommand{\mathbsf{e}}{\mathbsf{e}}
\newcommand{\mathbsf{E}}{\mathbsf{E}}
\newcommand{\mathsf{f}}{\mathsf{f}}
\newcommand{\mathsf{F}}{\mathsf{F}}
\newcommand{\mathbsf{f}}{\mathbsf{f}}
\newcommand{\mathbsf{F}}{\mathbsf{F}}
\newcommand{\mathsf{g}}{\mathsf{g}}
\newcommand{\mathsf{G}}{\mathsf{G}}
\newcommand{\mathbsf{g}}{\mathbsf{g}}
\newcommand{\mathbsf{G}}{\mathbsf{G}}
\newcommand{\mathsf{h}}{\mathsf{h}}
\newcommand{\mathsf{H}}{\mathsf{H}}
\newcommand{\mathbsf{h}}{\mathbsf{h}}
\newcommand{\mathbsf{H}}{\mathbsf{H}}
\newcommand{\mathsf{i}}{\mathsf{i}}
\newcommand{\mathsf{I}}{\mathsf{I}}
\newcommand{\mathbsf{i}}{\mathbsf{i}}
\newcommand{\mathbsf{I}}{\mathbsf{I}}
\newcommand{\mathsf{j}}{\mathsf{j}}
\newcommand{\mathsf{J}}{\mathsf{J}}
\newcommand{\mathbsf{j}}{\mathbsf{j}}
\newcommand{\mathbsf{J}}{\mathbsf{J}}
\newcommand{\mathsf{k}}{\mathsf{k}}
\newcommand{\mathsf{K}}{\mathsf{K}}
\newcommand{\mathbsf{k}}{\mathbsf{k}}
\newcommand{\mathbsf{K}}{\mathbsf{K}}
\newcommand{\mathsf{l}}{\mathsf{l}}
\newcommand{\mathsf{L}}{\mathsf{L}}
\newcommand{\mathbsf{l}}{\mathbsf{l}}
\newcommand{\mathbsf{L}}{\mathbsf{L}}
\newcommand{\mathsf{m}}{\mathsf{m}}
\newcommand{\mathsf{M}}{\mathsf{M}}
\newcommand{\mathbsf{m}}{\mathbsf{m}}
\newcommand{\mathbsf{M}}{\mathbsf{M}}
\newcommand{\mathsf{n}}{\mathsf{n}}
\newcommand{\mathsf{N}}{\mathsf{N}}
\newcommand{\mathbsf{n}}{\mathbsf{n}}
\newcommand{\mathbsf{N}}{\mathbsf{N}}
\newcommand{\mathsf{o}}{\mathsf{o}}
\newcommand{\mathsf{O}}{\mathsf{O}}
\newcommand{\mathbsf{o}}{\mathbsf{o}}
\newcommand{\mathbsf{O}}{\mathbsf{O}}
\newcommand{\mathsf{p}}{\mathsf{p}}
\newcommand{\mathsf{P}}{\mathsf{P}}
\newcommand{\mathbsf{p}}{\mathbsf{p}}
\newcommand{\mathbsf{P}}{\mathbsf{P}}
\newcommand{\mathsf{q}}{\mathsf{q}}
\newcommand{\mathsf{Q}}{\mathsf{Q}}
\newcommand{\mathbsf{q}}{\mathbsf{q}}
\newcommand{\mathbsf{Q}}{\mathbsf{Q}}
\newcommand{\mathsf{r}}{\mathsf{r}}
\newcommand{\mathsf{R}}{\mathsf{R}}
\newcommand{\mathbsf{r}}{\mathbsf{r}}
\newcommand{\mathbsf{R}}{\mathbsf{R}}
\newcommand{\mathsf{s}}{\mathsf{s}}
\newcommand{\mathsf{S}}{\mathsf{S}}
\newcommand{\mathbsf{s}}{\mathbsf{s}}
\newcommand{\mathbsf{S}}{\mathbsf{S}}
\newcommand{\mathsf{t}}{\mathsf{t}}
\newcommand{\mathsf{T}}{\mathsf{T}}
\newcommand{\mathbsf{t}}{\mathbsf{t}}
\newcommand{\mathbsf{T}}{\mathbsf{T}}
\newcommand{\mathsf{u}}{\mathsf{u}}
\newcommand{\mathsf{U}}{\mathsf{U}}
\newcommand{\mathbsf{u}}{\mathbsf{u}}
\newcommand{\mathbsf{U}}{\mathbsf{U}}
\newcommand{\mathsf{v}}{\mathsf{v}}
\newcommand{\mathsf{V}}{\mathsf{V}}
\newcommand{\mathbsf{v}}{\mathbsf{v}}
\newcommand{\mathbsf{V}}{\mathbsf{V}}
\newcommand{\mathsf{w}}{\mathsf{w}}
\newcommand{\mathsf{W}}{\mathsf{W}}
\newcommand{\mathbsf{w}}{\mathbsf{w}}
\newcommand{\mathbsf{W}}{\mathbsf{W}}
\newcommand{\mathsf{x}}{\mathsf{x}}
\newcommand{\mathsf{X}}{\mathsf{X}}
\newcommand{\mathbsf{x}}{\mathbsf{x}}
\newcommand{\mathbsf{X}}{\mathbsf{X}}
\newcommand{\mathsf{y}}{\mathsf{y}}
\newcommand{\mathsf{Y}}{\mathsf{Y}}
\newcommand{\mathbsf{y}}{\mathbsf{y}}
\newcommand{\mathbsf{Y}}{\mathbsf{Y}}
\newcommand{\mathsf{z}}{\mathsf{z}}
\newcommand{\mathsf{Z}}{\mathsf{Z}}
\newcommand{\mathbsf{z}}{\mathbsf{z}}
\newcommand{\mathbsf{Z}}{\mathbsf{Z}}
\newcommand{\ssfTheta}{\ssfTheta}
\newcommand{\Theta}{\Theta}
\newcommand{\bsfTheta}{\bsfTheta}
\newcommand{\boldsymbol{\Theta}}{\boldsymbol{\Theta}}
\newcommand{\ssfPhi}{\ssfPhi}
\newcommand{\Phi}{\Phi}
\newcommand{\bsfPhi}{\bsfPhi}
\newcommand{\boldsymbol{\Phi}}{\boldsymbol{\Phi}}
\newcommand{\mathsf{\Xi}}{\mathsf{\Xi}}
\DeclareSymbolFont{bsfletters}{OT1}{cmss}{bx}{n}
\DeclareSymbolFont{ssfletters}{OT1}{cmss}{m}{n}
\DeclareMathSymbol{\bsfGamma}{0}{bsfletters}{'000}
\DeclareMathSymbol{\ssfGamma}{0}{ssfletters}{'000}
\DeclareMathSymbol{\bsfDelta}{0}{bsfletters}{'001}
\DeclareMathSymbol{\ssfDelta}{0}{ssfletters}{'001}
\DeclareMathSymbol{\bsfTheta}{0}{bsfletters}{'002}
\DeclareMathSymbol{\ssfTheta}{0}{ssfletters}{'002}
\DeclareMathSymbol{\bsfLambda}{0}{bsfletters}{'003}
\DeclareMathSymbol{\ssfLambda}{0}{ssfletters}{'003}
\DeclareMathSymbol{\bsfXi}{0}{bsfletters}{'004}
\DeclareMathSymbol{\ssfXi}{0}{ssfletters}{'004}
\DeclareMathSymbol{\bsfPi}{0}{bsfletters}{'005}
\DeclareMathSymbol{\ssfPi}{0}{ssfletters}{'005}
\DeclareMathSymbol{\bsfSigma}{0}{bsfletters}{'006}
\DeclareMathSymbol{\ssfSigma}{0}{ssfletters}{'006}
\DeclareMathSymbol{\bsfUpsilon}{0}{bsfletters}{'007}
\DeclareMathSymbol{\ssfUpsilon}{0}{ssfletters}{'007}
\DeclareMathSymbol{\bsfPhi}{0}{bsfletters}{'010}
\DeclareMathSymbol{\ssfPhi}{0}{ssfletters}{'010}
\DeclareMathSymbol{\bsfPsi}{0}{bsfletters}{'011}
\DeclareMathSymbol{\ssfPsi}{0}{ssfletters}{'011}
\DeclareMathSymbol{\bsfOmega}{0}{bsfletters}{'012}
\DeclareMathSymbol{\ssfOmega}{0}{ssfletters}{'012}
\newcommand{\widehat{a}}{\widehat{a}}
\newcommand{\widehat{A}}{\widehat{A}}
\newcommand{\widetilde{a}}{\widetilde{a}}
\newcommand{\widetilde{A}}{\widetilde{A}}
\newcommand{\widehat{\ba}}{\widehat{\mathbf{a}}}
\newcommand{\widehat{\bA}}{\widehat{\mathbf{A}}}
\newcommand{\widetilde{\ba}}{\widetilde{\mathbf{a}}}
\newcommand{\widetilde{\bA}}{\widetilde{\mathbf{A}}}
\newcommand{\widehat{b}}{\widehat{b}}
\newcommand{\widehat{B}}{\widehat{B}}
\newcommand{\widetilde{b}}{\widetilde{b}}
\newcommand{\widetilde{B}}{\widetilde{B}}
\newcommand{\widehat{\bb}}{\widehat{\mathbf{b}}}
\newcommand{\widehat{\bB}}{\widehat{\mathbf{B}}}
\newcommand{\widetilde{\bb}}{\widetilde{\mathbf{b}}}
\newcommand{\widetilde{\bB}}{\widetilde{\mathbf{B}}}
\newcommand{\underline{\tilb}}{\underline{\widetilde{b}}}
\newcommand{\widehat{c}}{\widehat{c}}
\newcommand{\widehat{C}}{\widehat{C}}
\newcommand{\widetilde{c}}{\widetilde{c}}
\newcommand{\widetilde{C}}{\widetilde{C}}
\newcommand{\widehat{\bc}}{\widehat{\mathbf{c}}}
\newcommand{\widehat{\bC}}{\widehat{\mathbf{C}}}
\newcommand{\widetilde{\bc}}{\widetilde{\mathbf{c}}}
\newcommand{\widetilde{\bC}}{\widetilde{\mathbf{C}}}
\newcommand{\widehat{d}}{\widehat{d}}
\newcommand{\widehat{D}}{\widehat{D}}
\newcommand{\widetilde{d}}{\widetilde{d}}
\newcommand{\widetilde{D}}{\widetilde{D}}
\newcommand{\widehat{\bd}}{\widehat{\mathbf{d}}}
\newcommand{\widehat{\bD}}{\widehat{\mathbf{D}}}
\newcommand{\widetilde{\bd}}{\widetilde{\mathbf{d}}}
\newcommand{\widetilde{\bD}}{\widetilde{\mathbf{D}}}
\newcommand{\widehat{e}}{\widehat{e}}
\newcommand{\widehat{E}}{\widehat{E}}
\newcommand{\widetilde{e}}{\widetilde{e}}
\newcommand{\widetilde{E}}{\widetilde{E}}
\newcommand{\widehat{\be}}{\widehat{\mathbf{e}}}
\newcommand{\widehat{\bE}}{\widehat{\mathbf{E}}}
\newcommand{\widetilde{\be}}{\widetilde{\mathbf{e}}}
\newcommand{\widetilde{\bE}}{\widetilde{\mathbf{E}}}
\newcommand{\widehat{f}}{\widehat{f}}
\newcommand{\widehat{F}}{\widehat{F}}
\newcommand{\widetilde{f}}{\widetilde{f}}
\newcommand{\widetilde{F}}{\widetilde{F}}
\newcommand{\widehat{\boldf}}{\widehat{\mathbf{f}}}
\newcommand{\widehat{\bF}}{\widehat{\mathbf{F}}}
\newcommand{\widetilde{\boldf}}{\widetilde{\mathbf{f}}}
\newcommand{\widetilde{\bF}}{\widetilde{\mathbf{F}}}
\newcommand{\widehat{g}}{\widehat{g}}
\newcommand{\widehat{G}}{\widehat{G}}
\newcommand{\widetilde{g}}{\widetilde{g}}
\newcommand{\widetilde{G}}{\widetilde{G}}
\newcommand{\widehat{\bg}}{\widehat{\mathbf{g}}}
\newcommand{\widehat{\bG}}{\widehat{\mathbf{G}}}
\newcommand{\widetilde{\bg}}{\widetilde{\mathbf{g}}}
\newcommand{\widetilde{\bG}}{\widetilde{\mathbf{G}}}
\newcommand{\widehat{h}}{\widehat{h}}
\newcommand{\widehat{H}}{\widehat{H}}
\newcommand{\widetilde{h}}{\widetilde{h}}
\newcommand{\widetilde{H}}{\widetilde{H}}
\newcommand{\widehat{\bh}}{\widehat{\mathbf{h}}}
\newcommand{\widehat{\bH}}{\widehat{\mathbf{H}}}
\newcommand{\widetilde{\bh}}{\widetilde{\mathbf{h}}}
\newcommand{\widetilde{\bH}}{\widetilde{\mathbf{H}}}
\newcommand{\widehat{i}}{\widehat{i}}
\newcommand{\widehat{I}}{\widehat{I}}
\newcommand{\widetilde{i}}{\widetilde{i}}
\newcommand{\widetilde{I}}{\widetilde{I}}
\newcommand{\widehat{\bi}}{\widehat{\mathbf{i}}}
\newcommand{\widehat{\bI}}{\widehat{\mathbf{I}}}
\newcommand{\widetilde{\bi}}{\widetilde{\mathbf{i}}}
\newcommand{\widetilde{\bI}}{\widetilde{\mathbf{I}}}
\newcommand{\widehat{j}}{\widehat{j}}
\newcommand{\widehat{J}}{\widehat{J}}
\newcommand{\widetilde{j}}{\widetilde{j}}
\newcommand{\widetilde{J}}{\widetilde{J}}
\newcommand{\widehat{\bj}}{\widehat{\mathbf{j}}}
\newcommand{\widehat{\bJ}}{\widehat{\mathbf{J}}}
\newcommand{\widetilde{\bj}}{\widetilde{\mathbf{j}}}
\newcommand{\widetilde{\bJ}}{\widetilde{\mathbf{J}}}
\newcommand{\widehat{k}}{\widehat{k}}
\newcommand{\widehat{K}}{\widehat{K}}
\newcommand{\widetilde{k}}{\widetilde{k}}
\newcommand{\widetilde{K}}{\widetilde{K}}
\newcommand{\widehat{\bk}}{\widehat{\mathbf{k}}}
\newcommand{\widehat{\bK}}{\widehat{\mathbf{K}}}
\newcommand{\widetilde{\bk}}{\widetilde{\mathbf{k}}}
\newcommand{\widetilde{\bK}}{\widetilde{\mathbf{K}}}
\newcommand{\widehat{l}}{\widehat{l}}
\newcommand{\widehat{L}}{\widehat{L}}
\newcommand{\widetilde{l}}{\widetilde{l}}
\newcommand{\widetilde{L}}{\widetilde{L}}
\newcommand{\widehat{\bl}}{\widehat{\mathbf{l}}}
\newcommand{\widehat{\bL}}{\widehat{\mathbf{L}}}
\newcommand{\widetilde{\bl}}{\widetilde{\mathbf{l}}}
\newcommand{\widetilde{\bL}}{\widetilde{\mathbf{L}}}
\newcommand{\widehat{m}}{\widehat{m}}
\newcommand{\widehat{M}}{\widehat{M}}
\newcommand{\widetilde{m}}{\widetilde{m}}
\newcommand{\widetilde{M}}{\widetilde{M}}
\newcommand{\widehat{\boldm}}{\widehat{\mathbf{m}}}
\newcommand{\widehat{\bM}}{\widehat{\mathbf{M}}}
\newcommand{\widetilde{\boldm}}{\widetilde{\mathbf{m}}}
\newcommand{\widetilde{\bM}}{\widetilde{\mathbf{M}}}
\newcommand{\widehat{n}}{\widehat{n}}
\newcommand{\widehat{N}}{\widehat{N}}
\newcommand{\widetilde{n}}{\widetilde{n}}
\newcommand{\widetilde{N}}{\widetilde{N}}
\newcommand{\widehat{\bn}}{\widehat{\mathbf{n}}}
\newcommand{\widehat{\bN}}{\widehat{\mathbf{N}}}
\newcommand{\widetilde{\bn}}{\widetilde{\mathbf{n}}}
\newcommand{\widetilde{\bN}}{\widetilde{\mathbf{N}}}
\newcommand{\widehat{o}}{\widehat{o}}
\newcommand{\widehat{O}}{\widehat{O}}
\newcommand{\widetilde{o}}{\widetilde{o}}
\newcommand{\widetilde{O}}{\widetilde{O}}
\newcommand{\widehat{\bo}}{\widehat{\mathbf{o}}}
\newcommand{\widehat{\bO}}{\widehat{\mathbf{O}}}
\newcommand{\widetilde{\bo}}{\widetilde{\mathbf{o}}}
\newcommand{\widetilde{\bO}}{\widetilde{\mathbf{O}}}
\newcommand{\widehat{p}}{\widehat{p}}
\newcommand{\widehat{P}}{\widehat{P}}
\newcommand{\widetilde{p}}{\widetilde{p}}
\newcommand{\widetilde{P}}{\widetilde{P}}
\newcommand{\widehat{\bp}}{\widehat{\mathbf{p}}}
\newcommand{\widehat{\bP}}{\widehat{\mathbf{P}}}
\newcommand{\widetilde{\bp}}{\widetilde{\mathbf{p}}}
\newcommand{\widetilde{\bP}}{\widetilde{\mathbf{P}}}
\newcommand{\widehat{q}}{\widehat{q}}
\newcommand{\widehat{Q}}{\widehat{Q}}
\newcommand{\widetilde{q}}{\widetilde{q}}
\newcommand{\widetilde{Q}}{\widetilde{Q}}
\newcommand{\widehat{\bq}}{\widehat{\mathbf{q}}}
\newcommand{\widehat{\bQ}}{\widehat{\mathbf{Q}}}
\newcommand{\widetilde{\bq}}{\widetilde{\mathbf{q}}}
\newcommand{\widetilde{\bQ}}{\widetilde{\mathbf{Q}}}
\newcommand{\widehat{r}}{\widehat{r}}
\newcommand{\widehat{R}}{\widehat{R}}
\newcommand{\widetilde{r}}{\widetilde{r}}
\newcommand{\widetilde{R}}{\widetilde{R}}
\newcommand{\widehat{\br}}{\widehat{\mathbf{r}}}
\newcommand{\widehat{\bR}}{\widehat{\mathbf{R}}}
\newcommand{\widetilde{\br}}{\widetilde{\mathbf{r}}}
\newcommand{\widetilde{\bR}}{\widetilde{\mathbf{R}}}
\newcommand{\widehat{s}}{\widehat{s}}
\newcommand{\widehat{S}}{\widehat{S}}
\newcommand{\widetilde{s}}{\widetilde{s}}
\newcommand{\widetilde{S}}{\widetilde{S}}
\newcommand{\widehat{\bs}}{\widehat{\mathbf{s}}}
\newcommand{\widehat{\bS}}{\widehat{\mathbf{S}}}
\newcommand{\widetilde{\bs}}{\widetilde{\mathbf{s}}}
\newcommand{\widetilde{\bS}}{\widetilde{\mathbf{S}}}
\newcommand{\widehat{t}}{\widehat{t}}
\newcommand{\widehat{T}}{\widehat{T}}
\newcommand{\widetilde{t}}{\widetilde{t}}
\newcommand{\widetilde{T}}{\widetilde{T}}
\newcommand{\widehat{\bt}}{\widehat{\mathbf{t}}}
\newcommand{\widehat{\bT}}{\widehat{\mathbf{T}}}
\newcommand{\widetilde{\bt}}{\widetilde{\mathbf{t}}}
\newcommand{\widetilde{\bT}}{\widetilde{\mathbf{T}}}
\newcommand{\widehat{u}}{\widehat{u}}
\newcommand{\widehat{U}}{\widehat{U}}
\newcommand{\widetilde{u}}{\widetilde{u}}
\newcommand{\widetilde{U}}{\widetilde{U}}
\newcommand{\widehat{\bu}}{\widehat{\mathbf{u}}}
\newcommand{\widehat{\bU}}{\widehat{\mathbf{U}}}
\newcommand{\widetilde{\bu}}{\widetilde{\mathbf{u}}}
\newcommand{\widetilde{\bU}}{\widetilde{\mathbf{U}}}
\newcommand{\widehat{v}}{\widehat{v}}
\newcommand{\widehat{V}}{\widehat{V}}
\newcommand{\widetilde{v}}{\widetilde{v}}
\newcommand{\widetilde{V}}{\widetilde{V}}
\newcommand{\widehat{\bv}}{\widehat{\mathbf{v}}}
\newcommand{\widehat{\bV}}{\widehat{\mathbf{V}}}
\newcommand{\widetilde{\bv}}{\widetilde{\mathbf{v}}}
\newcommand{\widetilde{\bV}}{\widetilde{\mathbf{V}}}
\newcommand{\widehat{w}}{\widehat{w}}
\newcommand{\widehat{W}}{\widehat{W}}
\newcommand{\widetilde{w}}{\widetilde{w}}
\newcommand{\widetilde{W}}{\widetilde{W}}
\newcommand{\widehat{\bw}}{\widehat{\mathbf{w}}}
\newcommand{\widehat{\bW}}{\widehat{\mathbf{W}}}
\newcommand{\tilde{\bw}}{\tilde{\mathbf{w}}}
\newcommand{\widetilde{\bW}}{\widetilde{\mathbf{W}}}
\newcommand{\widehat{x}}{\widehat{x}}
\newcommand{\widehat{X}}{\widehat{X}}
\newcommand{\tilde{x}}{\tilde{x}}
\newcommand{\widetilde{X}}{\widetilde{X}}
\newcommand{\widehat{\bx}}{\widehat{\mathbf{x}}}
\newcommand{\widehat{\bX}}{\widehat{\mathbf{X}}}
\newcommand{\widetilde{\bx}}{\widetilde{\mathbf{x}}}
\newcommand{\widetilde{\bX}}{\widetilde{\mathbf{X}}}
\newcommand{\widehat{y}}{\widehat{y}}
\newcommand{\widehat{Y}}{\widehat{Y}}
\newcommand{\widetilde{y}}{\widetilde{y}}
\newcommand{\widetilde{Y}}{\widetilde{Y}}
\newcommand{\widehat{\by}}{\widehat{\mathbf{y}}}
\newcommand{\widehat{\bY}}{\widehat{\mathbf{Y}}}
\newcommand{\widetilde{\by}}{\widetilde{\mathbf{y}}}
\newcommand{\widetilde{\bY}}{\widetilde{\mathbf{Y}}}
\newcommand{\widehat{z}}{\widehat{z}}
\newcommand{\widehat{Z}}{\widehat{Z}}
\newcommand{\widetilde{z}}{\widetilde{z}}
\newcommand{\widetilde{Z}}{\widetilde{Z}}
\newcommand{\widehat{\bz}}{\widehat{\mathbf{z}}}
\newcommand{\widehat{\bZ}}{\widehat{\mathbf{Z}}}
\newcommand{\widetilde{\bz}}{\widetilde{\mathbf{z}}}
\newcommand{\widetilde{\bZ}}{\widetilde{\mathbf{Z}}}
\newcommand{\widehat{\bm{\xi}}}{\widehat{\bm{\xi}}}
\newcommand{\overline{\bm{\xi}}}{\overline{\bm{\xi}}}
\newcommand{\overline{a}}{\overline{a}}
\newcommand{\overline{b}}{\overline{b}}
\newcommand{\overline{c}}{\overline{c}}
\newcommand{\overline{d}}{\overline{d}}
\newcommand{\overline{e}}{\overline{e}}
\newcommand{\overline{f}}{\overline{f}}
\newcommand{\overline{g}}{\overline{g}}
\newcommand{\overline{h}}{\overline{h}}
\newcommand{\overline{i}}{\overline{i}}
\newcommand{\overline{j}}{\overline{j}}
\newcommand{\overline{k}}{\overline{k}}
\newcommand{\overline{l}}{\overline{l}}
\newcommand{\overline{m}}{\overline{m}}
\newcommand{\overline{n}}{\overline{n}}
\newcommand{\overline{o}}{\overline{o}}
\newcommand{\overline{p}}{\overline{p}}
\newcommand{\overline{q}}{\overline{q}}
\newcommand{\overline{r}}{\overline{r}}
\newcommand{\overline{s}}{\overline{s}}
\newcommand{\overline{t}}{\overline{t}}
\newcommand{\overline{u}}{\overline{u}}
\newcommand{\overline{v}}{\overline{v}}
\newcommand{\overline{w}}{\overline{w}}
\newcommand{\bar{x}}{\bar{x}}
\newcommand{\overline{y}}{\overline{y}}
\newcommand{\overline{z}}{\overline{z}}
\newcommand{\bar{A}}{\bar{A}}
\newcommand{\overline{B}}{\overline{B}}
\newcommand{\vphantom{C}\smash{\overline C}}{\vphantom{C}\smash{\overline C}}
\newcommand{\overline{D}}{\overline{D}}
\newcommand{\overline{E}}{\overline{E}}
\newcommand{\overline{F}}{\overline{F}}
\newcommand{\overline{G}}{\overline{G}}
\newcommand{\overline{H}}{\overline{H}}
\newcommand{\overline{I}}{\overline{I}}
\newcommand{\overline{J}}{\overline{J}}
\newcommand{\overline{K}}{\overline{K}}
\newcommand{\vphantom{L}\smash{\overline L}}{\vphantom{L}\smash{\overline L}}
\newcommand{\overline{M}}{\overline{M}}
\newcommand{\overline{N}}{\overline{N}}
\newcommand{\overline{O}}{\overline{O}}
\newcommand{\overline{P}}{\overline{P}}
\newcommand{\overline{Q}}{\overline{Q}}
\newcommand{\overline{R}}{\overline{R}}
\newcommand{\overline{S}}{\overline{S}}
\newcommand{\overline{T}}{\overline{T}}
\newcommand{\overline{U}}{\overline{U}}
\newcommand{\overline{V}}{\overline{V}}
\newcommand{\overline{W}}{\overline{W}}
\newcommand{\overline{X}}{\overline{X}}
\newcommand{\overline{Y}}{\overline{Y}}
\newcommand{\overline{Z}}{\overline{Z}}
\newcommand{\overline{\mu}}{\overline{\mu}}
\newcommand{\overline{\rho}}{\overline{\rho}}
\newcommand{\overline{\lambda}}{\overline{\lambda}}
\newcommand{\bm{\alpha}}{\bm{\alpha}}
\newcommand{\bm{\beta}}{\bm{\beta}}
\newcommand{\bm{\gamma}}{\bm{\gamma}}
\newcommand{\bm{\delta}}{\bm{\delta}}
\newcommand{\bm{\theta}}{\bm{\theta}}
\newcommand{\bm{\tau}}{\bm{\tau}}
\newcommand{\bm{\pi}}{\bm{\pi}}
\newcommand{\bm{\epsilon}}{\bm{\epsilon}}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{\bm{\varepsilon}}{\bm{\varepsilon}}
\newcommand{\bm{\sigma}}{\bm{\sigma}}
\newcommand{\bm{\zeta}}{\bm{\zeta}}
\newcommand{\bm{\eta}}{\bm{\eta}}
\newcommand{\bm{\kappa}}{\bm{\kappa}}
\newcommand{\bm{\chi}}{\bm{\chi}}
\newcommand{\bm{\phi}}{\bm{\phi}}
\newcommand{\bm{\psi}}{\bm{\psi}}
\newcommand{\bm{\omega}}{\bm{\omega}}
\newcommand{\bm{\xi}}{\bm{\xi}}
\newcommand{\bm{\lambda}}{\bm{\lambda}}
\newcommand{\bm{\rho}}{\bm{\rho}}
\newcommand{\bm{\Gamma}}{\bm{\Gamma}}
\newcommand{\bm{\Lambda}}{\bm{\Lambda}}
\newcommand{\bSigma }{\bm{\Sigma}}
\newcommand{\bm{\Psi}}{\bm{\Psi}}
\newcommand{\bm{\Delta}}{\bm{\Delta}}
\newcommand{\bm{\Xi}}{\bm{\Xi}}
\newcommand{\bm{\Upsilon}}{\bm{\Upsilon}}
\newcommand{\bm{\Omega}}{\bm{\Omega}}
\newcommand{\bm{\Phi}}{\bm{\Phi}}
\newcommand{\bm{\Pi}}{\bm{\Pi}}
\newcommand{\bm{\Theta}}{\bm{\Theta}}
\newcommand{\widetilde{\blambda}}{\widetilde{\bm{\lambda}}}
\newcommand{\widetilde{\alpha}}{\widetilde{\alpha}}
\newcommand{\widetilde{\beta}}{\widetilde{\beta}}
\newcommand{\widetilde{\gamma}}{\widetilde{\gamma}}
\newcommand{\widetilde{\Gamma}}{\widetilde{\Gamma}}
\newcommand{\widetilde{\delta}}{\widetilde{\delta}}
\newcommand{\widetilde{\theta}}{\widetilde{\theta}}
\newcommand{\widetilde{\tau}}{\widetilde{\tau}}
\newcommand{\widetilde{\pi}}{\widetilde{\pi}}
\newcommand{\widetilde{\epsilon}}{\widetilde{\epsilon}}
\newcommand{\widetilde{\varepsilon}}{\widetilde{\varepsilon}}
\newcommand{\widetilde{\sigma}}{\widetilde{\sigma}}
\newcommand{\widetilde{\zeta}}{\widetilde{\zeta}}
\newcommand{\widetilde{\eta}}{\widetilde{\eta}}
\newcommand{\widetilde{\kappa}}{\widetilde{\kappa}}
\newcommand{\widetilde{\chi}}{\widetilde{\chi}}
\newcommand{\widetilde{\phi}}{\widetilde{\phi}}
\newcommand{\widetilde{\psi}}{\widetilde{\psi}}
\newcommand{\widetilde{\omega}}{\widetilde{\omega}}
\newcommand{\widetilde{\xi}}{\widetilde{\xi}}
\newcommand{\widetilde{\lambda}}{\widetilde{\lambda}}
\newcommand{\widetilde{\rho}}{\widetilde{\rho}}
\newcommand{\widetilde{\nu}}{\widetilde{\nu}}
\newcommand{\widetilde{\iota}}{\widetilde{\iota}}
\newcommand{\widetilde{\bdelta}}{\widetilde{\bm{\delta}}}
\newcommand{\widetilde{\Delta}}{\widetilde{\Delta}}
\newcommand{\widetilde{\mu}}{\widetilde{\mu}}
\newcommand{\widetilde{\bx}}{\widetilde{\mathbf{x}}}
\newcommand{\widehat{\alpha}}{\widehat{\alpha}}
\newcommand{\widehat{\beta}}{\widehat{\beta}}
\newcommand{\widehat{\gamma}}{\widehat{\gamma}}
\newcommand{\widehat{\delta}}{\widehat{\delta}}
\newcommand{\widehat{\theta}}{\widehat{\theta}}
\newcommand{\widehat{\tau}}{\widehat{\tau}}
\newcommand{\widehat{\pi}}{\widehat{\pi}}
\newcommand{\widehat{\epsilon}}{\widehat{\epsilon}}
\newcommand{\widehat{\varepsilon}}{\widehat{\varepsilon}}
\newcommand{\widehat{\sigma}}{\widehat{\sigma}}
\newcommand{\widehat{\zeta}}{\widehat{\zeta}}
\newcommand{\widehat{\eta}}{\widehat{\eta}}
\newcommand{\widehat{\kappa}}{\widehat{\kappa}}
\newcommand{\widehat{\chi}}{\widehat{\chi}}
\newcommand{\widehat{\phi}}{\widehat{\phi}}
\newcommand{\widehat{\psi}}{\widehat{\psi}}
\newcommand{\widehat{\omega}}{\widehat{\omega}}
\newcommand{\widehat{\xi}}{\widehat{\xi}}
\newcommand{\widehat{\lambda}}{\widehat{\lambda}}
\newcommand{\widehat{\rho}}{\widehat{\rho}}
\newcommand{\underline{a}}{\underline{a}}
\newcommand{\underline{A}}{\underline{A}}
\newcommand{\underline{b}}{\underline{b}}
\newcommand{\underline{B}}{\underline{B}}
\newcommand{\underline{c}}{\underline{c}}
\newcommand{\underline{C}}{\underline{C}}
\newcommand{\underline{d}}{\underline{d}}
\newcommand{\underline{D}}{\underline{D}}
\newcommand{\underline{e}}{\underline{e}}
\newcommand{\underline{E}}{\underline{E}}
\newcommand{\underline{f}}{\underline{f}}
\newcommand{\underline{F}}{\underline{F}}
\newcommand{\underline{g}}{\underline{g}}
\newcommand{\underline{G}}{\underline{G}}
\newcommand{\underline{h}}{\underline{h}}
\newcommand{\underline{\bh}}{\underline{\mathbf{h}}}
\newcommand{\underline{H}}{\underline{H}}
\newcommand{\underline{i}}{\underline{i}}
\newcommand{\underline{I}}{\underline{I}}
\newcommand{\underline{j}}{\underline{j}}
\newcommand{\underline{J}}{\underline{J}}
\newcommand{\underline{k}}{\underline{k}}
\newcommand{\underline{K}}{\underline{K}}
\newcommand{\underline{l}}{\underline{l}}
\newcommand{\underline{L}}{\underline{L}}
\newcommand{\underline{m}}{\underline{m}}
\newcommand{\underline{M}}{\underline{M}}
\newcommand{\underline{n}}{\underline{n}}
\newcommand{\underline{N}}{\underline{N}}
\newcommand{\underline{o}}{\underline{o}}
\newcommand{\underline{O}}{\underline{O}}
\newcommand{\underline{p}}{\underline{p}}
\newcommand{\underline{P}}{\underline{P}}
\newcommand{\underline{q}}{\underline{q}}
\newcommand{\underline{Q}}{\underline{Q}}
\newcommand{\underline{r}}{\underline{r}}
\newcommand{\underline{R}}{\underline{R}}
\newcommand{\underline{s}}{\underline{s}}
\newcommand{\underline{S}}{\underline{S}}
\newcommand{\underline{t}}{\underline{t}}
\newcommand{\underline{T}}{\underline{T}}
\newcommand{\underline{u}}{\underline{u}}
\newcommand{\underline{U}}{\underline{U}}
\newcommand{\underline{v}}{\underline{v}}
\newcommand{\underline{V}}{\underline{V}}
\newcommand{\underline{w}}{\underline{w}}
\newcommand{\underline{W}}{\underline{W}}
\newcommand{\underline{x}}{\underline{x}}
\newcommand{\underline{X}}{\underline{X}}
\newcommand{\underline{y}}{\underline{y}}
\newcommand{\underline{Y}}{\underline{Y}}
\newcommand{\underline{z}}{\underline{z}}
\newcommand{\underline{Z}}{\underline{Z}}
\newcommand{\underline{\bE}}{\underline{\mathbf{E}}}
\newcommand{\underline{\bW}}{\underline{\mathbf{W}}}
\newcommand{\underline{\bH}}{\underline{\mathbf{H}}}
\newcommand{\underline{\lambda}}{\underline{\lambda}}
\newcommand{\dot{B}}{\dot{B}}
\newcommand{\dot{c}}{\dot{c}}
\newcommand{\dot{P}}{\dot{P}}
\newcommand{\dot{L}}{\dot{L}}
\newcommand{\dot{\bx}}{\dot{\mathbf{x}}}
\newcommand{\dot{\by}}{\dot{\mathbf{y}}}
\newcommand{\dot{\bz}}{\dot{\mathbf{z}}}
\def\, \cdot \,{\, \cdot \,}
\def\, \diamond \,{\, \diamond \,}
\def\, \star \,{\, \star \,}
\newcommand{\eexp}[1]{e^{#1}}
\newcommand{i.i.d.\ }{i.i.d.\ }
\newcommand{\stackrel{\mathrm{p}}{\longrightarrow}}{\stackrel{\mathrm{p}}{\longrightarrow}}
\newcommand{\stackrel{\mathrm{w.p.1}}{\longrightarrow}}{\stackrel{\mathrm{w.p.1}}{\longrightarrow}}
\newcommand{\xrightarrow{\mathrm{a.s.}}}{\xrightarrow{\mathrm{a.s.}}}
\newcommand{\stackrel{\mathrm{d}}{\longrightarrow}}{\stackrel{\mathrm{d}}{\longrightarrow}}
\newcommand{\stackrel{\mathrm{D}}{\longrightarrow}}{\stackrel{\mathrm{D}}{\longrightarrow}}
\newcommand{\ceil}[1]{\left\lceil{#1}\right\rceil}
\newcommand{\floor}[1]{\lfloor{#1}\rfloor}
\newcommand{\lrangle}[2]{\left\langle{#1},{#2}\right\rangle}
\newcommand{\lranglet}[2]{\langle{#1},{#2}\rangle}
\newcommand{\stackrel{.}{\leq}}{\stackrel{.}{\leq}}
\newcommand{\stackrel{.}{<}}{\stackrel{.}{<}}
\newcommand{\stackrel{.}{\geq}}{\stackrel{.}{\geq}}
\newcommand{\stackrel{.}{>}}{\stackrel{.}{>}}
\newcommand{\stackrel{\,..}{=}}{\stackrel{\,..}{=}}
\newcommand{\stackrel{\rm(a)}{=}}{\stackrel{\rm(a)}{=}}
\newcommand{\stackrel{\rm(b)}{=}}{\stackrel{\rm(b)}{=}}
\newcommand{\stackrel{\rm(c)}{=}}{\stackrel{\rm(c)}{=}}
\newcommand{\stackrel{\rm(d)}{=}}{\stackrel{\rm(d)}{=}}
\newcommand{\stackrel{\rm(e)}{=}}{\stackrel{\rm(e)}{=}}
\newcommand{\stackrel{\rm(f)}{=}}{\stackrel{\rm(f)}{=}}
\newcommand{\stackrel{\rm(g)}{=}}{\stackrel{\rm(g)}{=}}
\newcommand{\stackrel{\rm(h)}{=}}{\stackrel{\rm(h)}{=}}
\newcommand{\stackrel{\rm(a)}{\le}}{\stackrel{\rm(a)}{\le}}
\newcommand{\stackrel{\rm(b)}{\le}}{\stackrel{\rm(b)}{\le}}
\newcommand{\stackrel{\rm(c)}{\le}}{\stackrel{\rm(c)}{\le}}
\newcommand{\stackrel{\rm(d)}{\le}}{\stackrel{\rm(d)}{\le}}
\newcommand{\stackrel{\rm(e)}{\le}}{\stackrel{\rm(e)}{\le}}
\newcommand{\stackrel{\rm(f)}{\le}}{\stackrel{\rm(f)}{\le}}
\newcommand{\stackrel{\rm(g)}{\le}}{\stackrel{\rm(g)}{\le}}
\newcommand{\stackrel{\rm(h)}{\le}}{\stackrel{\rm(h)}{\le}}
\newcommand{\stackrel{\rm(a)}{\ge}}{\stackrel{\rm(a)}{\ge}}
\newcommand{\stackrel{\rm(b)}{\ge}}{\stackrel{\rm(b)}{\ge}}
\newcommand{\stackrel{\rm(c)}{\ge}}{\stackrel{\rm(c)}{\ge}}
\newcommand{\stackrel{\rm(d)}{\ge}}{\stackrel{\rm(d)}{\ge}}
\newcommand{\stackrel{\rm(e)}{\ge}}{\stackrel{\rm(e)}{\ge}}
\newcommand{\stackrel{\rm(f)}{\ge}}{\stackrel{\rm(f)}{\ge}}
\newcommand{\stackrel{\rm(g)}{\ge}}{\stackrel{\rm(g)}{\ge}}
\newcommand{\stackrel{\rm(h)}{\ge}}{\stackrel{\rm(h)}{\ge}}
\newcommand{P_{\mathrm{e}}^{(n)}}{P_{\mathrm{e}}^{(n)}}
\newcommand{P_{\mathrm{e}, 1}^{(n)}}{P_{\mathrm{e}, 1}^{(n)}}
\newcommand{P_{\mathrm{e}, 2}^{(n)}}{P_{\mathrm{e}, 2}^{(n)}}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{argmin}
\DeclareMathOperator*{\dist}{dist}
\DeclareMathOperator*{\argsup}{arg\,sup}
\DeclareMathOperator*{\arginf}{arg\,inf}
\DeclareMathOperator{\minimize}{minimize}
\DeclareMathOperator{\maximize}{maximize}
\DeclareMathOperator{\st}{s.t.\;}
\DeclareMathOperator{\erfc}{erfc}
\DeclareMathOperator{\cum}{cum}
\DeclareMathOperator{\sgn}{sgn}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\spn}{span}
\DeclareMathOperator{\supp}{supp}
\DeclareMathOperator{\adj}{adj}
\DeclareMathOperator{\var}{\mathsf{Var}}
\DeclareMathOperator{\Vol}{Vol}
\DeclareMathOperator{\cov}{\mathsf{Cov}}
\DeclareMathOperator{\sech}{sech}
\DeclareMathOperator{\sinc}{sinc}
\DeclareMathOperator{\rank}{rank}
\DeclareMathOperator{\poly}{poly}
\DeclareMathOperator{\polylog}{polylog}
\DeclareMathOperator{\vect}{vec}
\newcommand{\Hb}{H_{\mathrm{b}}
\newcommand{\mathrm{Bern}}{\mathrm{Bern}}
\DeclareMathOperator*{\lms}{l.i.m.\,}
\newcommand{\varop}[1]{\var\left[{#1}\right]}
\newcommand{\covop}[2]{\cov\left({#1},{#2}\right)}
\newcommand{\mathbf{0}}{\mathbf{0}}
\newcommand{\mathbf{1}}{\mathbf{1}}
\def\independenT#1#2{\mathrel{\rlap{$#1#2$}\mkern5mu{#1#2}}}
\newcommand\indep{\protect\mathpalette{\protect\independenT}{\perp}}
|
\section*{}\label{sect:intro}
\vskip -1.5cm
\section{Introduction}\label{sect:intro}
From a fundamental point of view, presumably the most basic parameters are given at a mass-defining scale which is considered to be the Planck mass $\Mp\simeq 2.43\times 10^{18\,}\gev$. Any other scale involves a small coupling which is that scale divided by $\Mp$. Below $\Mp$, one tries to understand elementary particles in this sense.
Figure \ref{fig:Gross} is a theorist's design to plan his idea, starting with a grand framework, presumably inclusive of as much natural phenomena as possible. Within this framework, he builds a theory. The theory contains models. These models must be working examples, explaining the observed phenomena. Without a model example, some will say that it is a religion even though the design is fantastic. Our job is to find working models in this FRAMEWORK/THEORY/MODEL inclusion cartoon. In this sense, ``symmetry'' as a framework has worked for a century.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.5\textwidth]{figGross.eps}
\end{center}
\caption{An inclusive relation of frameworks due to Gross \cite{Gross16}.
}
\label{fig:Gross}
\end{figure}
Nowadays, flavor symmetries are studied mainly by some discrete symmetries, due to the observed large mixing angles in the leptonic sector. But, flavor symmetry may be a gauge symmetry in which case a true unification is GUTs with the flavor symmetry included there. The first attempt along this line was due to Georgi in SU(11) \cite{Georgi79} on the unification of GUT families (UGUTF). The next try with spinor representation of SO($4n+2$) groups was in SO(14) \cite{KimPRL80}. A more attempt along this line is from string compactification based on ${\bf Z}_{12-I}$ orbifold compactification \cite{KimJHEP15}. But, mostly one tries to obtain an electroweak-scale massive particle by a discrete
symmetry and a light scalar by the Goldstone theorem.
Among Goldstone bosons,
``invisible'' axion is the most interesting one.
\section{CP's}
To discuss violation of a symmetry, first one has to define the symmetry. Parity P is the most well-known example for the definition and violation of a symmetry. Even though kinetic mixings of U(1) gauge bosons have been considered for some time, the definition of a symmetry is usually done such that the kinetic energy terms preserve the symmetry.
If there exists a possibility of a Lagrangian that satisfies $(\rm CP){\cal L}(CP)^{-1}={\cal L}$, then the CP symmetry is preserved. Here,
the first thing to do is to define fields with CP quantum numbers. Next, find out terms breaking CP and search for its physical implications.
CP violation is an interference phenomenon. At this workshop, Domencico \cite{OkunCited} cited Okun's statement, ``Neutral K mesons are a unique physical system which appears to be created by nature to demonstrate, in the most impressive manner, a number of most spectacular phenomenon''. One may replace (spectacular)$\to$(interference).
CP violation in the SM is an interference phenomenon, encompassing all three families. This will become clearer below when we express the Jarlskog determinant $J$.
After the discovery of weak CP violation in 1964, needs for
theories of weak CP violation became very important. Some of these weak CP violation were
\begin{eqnarray}
&1.& \textrm{by light colored scalars \cite{KM73},}\nonumber \\
&2.& \textrm{by right-handed current(s) \cite{Mohapatra72,KM73},}\nonumber \\
&3.& \textrm{by three left-handed families \cite{KM73}} \\
&4.& \textrm{by propagators of light color-singlet scalars \cite{Weinberg76},}\nonumber \\
&5.& \textrm{by an extra-U(1) gauge interaction \cite{Wolfenstein64},} \nonumber
\end{eqnarray}
where some examples interpretable in modern gauge theories are cited. Item 3 is known as the Cabibbo\cite{Cabibbo63}-Kobayashi-Maskwa(CKM) model.
In the standard model(SM), the kinetic energy terms of quarks, leptons, and Higgs doublets are CP conserving. The CP violation in the SM arises in the interaction terms, typically through the Yukawa couplings. If the VEVs of Higgs doublets vanish, then there is no CP violation because all fermions are massless. Below the VEV scale of the Higgs doublets, all the SM fields obtain masses, and one can locate the CP phase in the left-handed currents, coupling to $W^\pm_\mu$. The charged current couplings are defined in this setup for the CKM (for quarks) and Pontecorvo-Maki-Nakagawa-Sakata(PMNS) (for leptons) \cite{PMNS} matrices.
The CKM matrix is unitary, which is the only condition for the CKM matrix. The physical significance of the weak CP violation is given by the Jarlskog determinant $J$ which is obtained from the imaginary part of a product of two elements of $V$ and two elements of $V^*$ of the CKM matrix, e.g. of the type $J=|{\rm Im}(V_{12}V_{23}V_{13}^*V_{22}^*)|$ \cite{Jarlskog85}.
Let us choose the CKM and PMNS matrices such that the 1st row real. Then, the CKM matrix can be chosen as \cite{KimSeo11},
\begin{eqnarray}
V_{\rm KS}= \left(\begin{array}{ccc} c_1,&s_1c_3,&s_1s_3 \\ [0.2em]
-c_2s_1,&e^{-i\delq}s_2s_3 +c_1c_2c_3,&-e^{-i\delq} s_2c_3+c_1c_2s_3\\[0.2em]
-e^{i\delq} s_1s_2,&-c_2s_3 +c_1s_2c_3 e^{i\delq},& c_2c_3 +c_1s_2s_3 e^{i\delq}
\end{array}\right),~ {\rm~ with~Det\,}V_{\rm KS}=1 \label{eq:KSexact}
\end{eqnarray}
where $c_i=\cos\theta_i$ and $s_i=\sin\theta_i$ for $i=1,2,3$. In this form, the invariant quantity for the CP violation, the Jarlskog determinant is directly seen from $V_{\rm KS}$ itself \cite{KimNam15},\footnote{For the PMNS matrix, we use different parameters by replacing $\theta_i\to\Theta_i$ and $\delq\to \dell$.}
\begin{eqnarray}
J=|{\rm Im}\,V_{13} V_{22} V_{31} |\simeq O(\lambda^6),
\end{eqnarray}
where $\lambda\simeq 0.22$ is the Cabibbo parameter $\lambda=\sin\theta_C$ \cite{Cabibbo63}. Note that all three families participate in the evaluartion of $J$, fulfilling the claim that CP violation is an `` interference phenomenon''. For the CP violation to be nonzero, in addition,
all u-type quark masses
must be different and all d-type
quark masses must be different, because one can rotate the phases of identical-mass quarks such that $\delq$ changes, which implies that $\delq$ is unobservable.
\begin{figure}[t!]
{\includegraphics[width=0.5
\textwidth]{figJarlskog}} \hskip 0.5cm {\includegraphics[width=0.45
\textwidth]{figLong}} \\
~ \hskip 3cm (a) \hskip 8.5cm (b)
\caption{The Jarlskog triangles. (a) B-meson decay including a K-meson, and (b) B-meson decay to $\pi$ mesons.}\label{Fig:Jarls}
\end{figure}
There are many different parametrization schemes. Different parametrizations give different CP phases $\delq$ \cite{KimNam15}. This can be understtod by the Jarlskog triangle for B-meson decay, shown in Fig. \ref{Fig:Jarls}\,(a).
For the parametrization of Eq. (\ref{eq:KSexact}), $J$ is given with the red parameters
\begin{equation}
J=c_1 c_2c_3 s_1^2 s_2 s_3 \sin\alpha
\end{equation}
while the PDG parametrization with the blue parameters fives $J=c_{12}c_{31}^2c_{23}s_{12} s_{23}s_{13}\sin\gamma$ \cite{PDG15}. In any parametrization, the area of Fig. \ref{Fig:Jarls}\,(a) gives the same value.
We can use the KS--parametrization, Eq. (\ref{eq:KSexact}), to show the maximality of the weak CP violation for the measured values of real angles $\theta_i$. Let us use the fact that any Jarlskog triangle has the same area. So, consider Fig. \ref{Fig:Jarls}\,(b).
With the $\lambda=\sin\theta_C$ expansion, the area of the Jarlskog triangle is of order $\lambda^6$, which is the product of two sides enclosing the angle $\delta$. Rotate the side of O($\lambda^5$) in the red arrow direction, making a triangle implied by dashed lines. This area of the triangle is maximal when the rotating angle $\delta\simeq 90^{\rm o}$, which is the triangle enclosing the yellow in Fig. \ref{Fig:Jarls}\,(b). Note that $\alpha$ of Fig. \ref{Fig:Jarls}\,(a) is determined by the fitting groups close to $\alpha= \frac\pi{2}$,
\begin{eqnarray}
\alpha=\left(85.4^{+3.9}_{-3.8}
\right)^{\rm o}~[\rm PDG,\cite{PDG15}],~\left(88.6^{+3.3}_{-3.3}
\right)^{\rm o}~[\rm U_{fit}, \cite{HarnewH16}],~\left(90.6^{+3.9}_{-1.1}
\right)^{\rm o}~[\rm CKM_{fit},\cite{Pich16}].
\end{eqnarray}
With the KS parametrization, $\delta\simeq \frac\pi{2}$ is in the allowed region. So, we proved that the weak CP violation is maximal with the pre-fixed real angles in the KS--parametrization. Since physical statements are parametrization independent, this maximality must be the case in the PDG--parametrization also.
In the leptonic sector also, there is a preliminary hint that $\dell\ne 0$, and close to $-\frac{\pi}{2}$ even though the error bar is large \cite{T2KCP}. The quark mixing angles are $\theta_i$ and $\delq$, and lepton mixing angles are $\Theta_i$ and $\dell$.
Even if $\theta_i$ and $\Theta_i$ cannot be related, we can relate $\delq$ and $\dell$ if there is only one CP phase in the whole theory. Indeed, this has been shown in Ref. \cite{KimPLB11} where the weak CP violation is spontaneous and one unremovable phase is located at the weak interaction singlet {\it \`a la} the Froggatt-Nielsen(FN) mechanism \cite{FN79}.\footnote{After the talk, another method for the flavor solution has been proposed \cite{UanomFl}.}
In the supersymmetric model, it was shown that one phase in the ultra-violet completion gives \cite{KimNam15}:
$\dell=\pm\delq.$
Then, the Jarlskog triangles of the quark and lepton sectors will have one common phase.
In addition to the CKM and PMNS phases, there are Majorana and leptogenesis phases also.
If there is only one phase in the ultra-violet completed theory, all of these must be expressed in terms of one phase. So, the Majorana phase determined at the intermediate scale and the leptogenesis phase can be also expressed in terms of this one phase, as shown in Ref. \cite{CoviKim16}.
\section{The strong CP problem}
Because of instanton solutions of QCD, there exists an effective
interaction term containing the gluon anomaly: $\bar{\theta}\{G\tilde{G}\}$. It is the flavor singlet and the source solving the U(1) problem of QCD by 't Hooft \cite{Hooft86}.
This gluon anomaly term is physical, but leads to
\begin{itemize}
\item
The strong CP problem, ``Why is the nEDM so small?'' This leads to the three classes of natural solutions.
\item
The remaining `natural solution' is ``invisible'' axion \cite{KimRMP10}.
\end{itemize}
The three classes of solutions are (1) calculable models, (2) massless up quark, and (3) ``invisible'' axion. Calculable models is not separable from the discussion of the weak CP violation.
\begin{figure}[t!]
\centerline{\includegraphics[width=0.7\textwidth]{figCorfu16}}
\caption{(a) CP violation by the insertion of $\langle \pi^0,\eta'\rangle$. The bullets in (b,c,d) are (a) of the CP violating interaction. (c) and (d) lead to an nEDM.}\label{Fig:nEDM}
\end{figure}
If the CP violating coupling $\overline{g_{\pi NN}}$ is present, the neutron electric dipole moment(nEDM) is calculated as (with the CP conserving ${g_{\pi NN}}$ term),
\begin{equation}
\frac{d_n}{e}=\frac{{g_{\pi NN}}\overline{g_{\pi NN}}}{4\pi^2 m_N} \ln\left(\frac{m_N}{m_\pi}\right).\label{eq:nEDM}
\end{equation}
Figure \ref{Fig:nEDM}\,(a) shows $\overline{g_{\pi NN}}$, and Figure \ref{Fig:nEDM}\,(c,d) give Eq. (\ref{eq:nEDM}) \cite{Crewther79}.
If the $\bar\theta$ term is present in QCD, then $\pi^0$ can obtain a VEV. VEVs of $\pi^0$ and $\eta'$ break CP, and give $|\overline{g_{\pi NN}}|\simeq\bar\theta/3$ \cite{KimRMP10}. The non-observation of nEDM put a limit on $|\bar\theta|$ as less than $10^{-10}$. For the class of calculable solutions, the so-called Nelson-Barr type weak CP violation is close to a solution \cite{Nelson,Barr}, but the limit $10^{-10}$ is difficult to realize.
For the massless up-quark solution, it seems not favored by the measured current quark masses as shown in Fig. \ref{Fig:mu} \cite{Manohar14}.
\begin{figure}[t!]
\centerline{\includegraphics[width=0.4\textwidth]{figmu}}
\caption{$m_d$ vs. $m_u$ \cite{Manohar14}.}\label{Fig:mu}
\end{figure}
This leads us to a brief historical introduction, eventually leading to the ``invisible'' axion \cite{KSVZ1,KSVZ2,DFSZ}. Pre- ``invisible'' axion developments are the following.
At the time when the third quark family was not discovered, Weinberg
tried to introduce the weak CP violation in the Higgs potential. Satisfying the Glashow-Weinberg condition that up-type quarks couple to $H_u$ and down-type quarks couple to $H_d$ \cite{GW77}, he introduced many Higgs doublets. Then, the weak CP violation introduced in the potential, with a discrete symmetry $\phi_I\to-\phi_I$,
\begin{eqnarray}
V_{\rm W}=\frac12\sum_I m_I^2\phi_I^\dagger\phi_I+\frac14 \sum_{IJ} \left\{a_{IJ} \phi_I^\dagger\phi_I\phi_J^\dagger\phi_J+b_{IJ}\phi_I^\dagger\phi_I\phi_J^\dagger\phi_J +(c_{IJ} \phi_I^\dagger\phi_I\phi_J^\dagger\phi_J+{\rm H.c.})
\right\}\label{eq:Weinberg}
\end{eqnarray}
Weinberg's necessary condition for the existence of CP violation is non-zero $c_{IJ}$ terms \cite{Weinberg76}. If one removes the $c_{IJ}$ terms, Peccei and Quinn (PQ) noticed that there emerges a global symmetry which is now called the \UPQ~symmetry \cite{PQ77}. This is an example that keeping only a few terms among the discrete symmetry allowed terms in the potential produces a global symmetry.
With the PQ symmetry by removing the $c_{ij}$ term in Eq. (\ref{eq:Weinberg}) \cite{PQ77}, Weinberg and Wilczek at Ben Lee Memorial Conference noted the existence of a pseudoscalar, the PQWW axion
\cite{PQWW77}, which was soon declared to be non-existent \cite{Peccei78}. This has led to calculable models discussed in
\cite{Beg78,Georgi78,Moha78,
Segre79,Langacker79}.
\section{The ``invisible'' axion}
However, a good symmetry principle of Fig. \ref{fig:Gross} is so an attractive framework, the PQ symmetry is re-introduced by a SM singlet field \cite{KSVZ1}, which has been named as ``invisible'' axion $a$. It was noted that the ``invisible'' axion was harmful in the evolution of the Universe
\cite{Preskill83}, which has been later turned into a bonus after realizing that the CDM contribution was important in the evolution of the Universe \cite{Primack84}.
``Invisible'' can be made ``visible'' if one invents a clever cavity detector
\cite{Sikivie83}, which is used in many axion search labs now \cite{AxionLabs}.
\begin{figure}[t!]
\centerline{\includegraphics[width=0.85\textwidth]{figAxionData}}
\caption{Axion detection bounds on the $c_{a\gamma\gamma}$ vs. $m_a$ plane with regions for several model parameters \cite{KimPLBagg15}. The white square on the upper left corner is the MI-axion point \cite{KimNamPLB16}.}\label{Fig:AxData}
\end{figure}
There have been the cosmic experiment and solar axion search
experiments \cite{ADMX1}. In Fig. \ref{Fig:AxData}, the current bounds on the ``invisible'' QCD axion search are shown.
An SU(2)xU(1) singlet housing the ``invisible'' axion gives the effective Lagrangian of $a$ as
\begin{eqnarray}
{\cal L}=c_1\frac{\partial_\mu a}{f_a} \sum_q\bar{q}\,\gamma^\mu\gamma_5\, q-\sum_q(\bar{q}_L\, m\, q_R\, e^{ic_2 a/f_a} +\textrm{h.c.})+\frac{c_3}{32\pi^2 f_a}a\, G_{\mu\nu}\tilde{G}^{\mu\nu}
\label{eq:invAxion}\\
+\frac{c_{aWW}}{32\pi^2 f_a}a\, W_{\mu\nu}\tilde{W}^{\mu\nu}
+\frac{c_{aYY}}{32\pi^2 f_a}a\, Y_{\mu\nu}\tilde{Y}^{\mu\nu} + {\cal L}_{\rm\, leptons},\nonumber
\end{eqnarray}
where $\tilde{G}^{\mu\nu}, \tilde{W}^{\mu\nu}$, and $\tilde{Y}^{\mu\nu}$ are dual field strengths of gluon, $W$, and hypercharge fields, respectively.
It is a key question how the PQ symmetry is defined. These couplings arise from the following renormalizable couplings,
\begin{eqnarray}
{\cal L}_{\rm KSVZ}& =&- f\overline{Q}_R Q_L+\textrm{h.c.} ,\label{eq:KSVZ}\\[0.5em]
{V}_{\rm DFSZ} &=& -\mu_1^2 H_u^*H_u -\mu_2^2 H_d^*H_d+\lambda_1( H_u^*H_u)^2+\lambda_2( H_d^*H_d)^2+{\sigma\rm~terms} \label{eq:DFSZ}\\
&& (+M H_uH_d\sigma) +\lambda' H_uH_d\sigma^2+\textrm{h.c.}\nonumber
\end{eqnarray}
In the KSVZ model, the heavy quark $Q$ is introduced and the $f$-term Yukawa coupling is the definition of the PQ symmetry. In the DFSZ model, the $\lambda'$-term is the definition of the PQ symmetry.
Note, however, that there must be a fine-tuning in the coefficient $\lambda'$ such that $\vew\ll f_a$ \cite{Dreiner14}. The axion-photon-photon couplings are listed in Table \ref{tab:Inv}.
\begin{table}
\begin{tabular}{|r|c|}
\hline &\\[-1em]
KSVZ: $Q_{\rm em}$ ~&$c_{a\gamma\gamma}$\\\hline
$0$ ~~ &$-2$~ \\[0.3em]
$\pm \frac13$~~ &$-\frac{4}3$~ \\[0.3em]
$\pm \frac23$~~ &$ \frac{2}3$ \\[0.3em]
$\pm 1 $~~ &$4$ \\[0.3em]
$(m,m)$~&$-\frac{1}3$~ \\[0.2em]
\hline
\end{tabular}
\hskip 1cm
\begin{tabular}{|r c|c|}
\hline &&\\[-1em]
DFSZ: $(q^c$-$e_L)$ pair ~& Higgs &$c_{a\gamma\gamma}$\\\hline
\\[-1.35em]
&& \\[-0.85em]
non-SUSY $(d^c,e)$ ~~ & $H_d$&$\frac23$~ \\[0.3em]
non-SUSY $(u^c,e)$ ~~ & $H_u^*$&$-\frac43$~
\\[0.2em]
GUTs ~~ & &$\frac23$~ \\[0.3em]
SUSY ~~ & &$\frac23$~
\\[0.2em]
\hline
\end{tabular}
\caption{$c_{a\gamma\gamma}$ in the KSVZ and DFSZ models. For the $u$ and $d$ quark masses, $m_u=0.5\, m_d$ is assumed for simplicity. $(m,m)$ in the last row the KSVZ means $m$ quarks of $Q_{\rm em}=\frac23\,e$ and $m$ quarks of $Q_{\rm em}=-\frac13\,e$. SUSY in the DFSZ includes contributions of color partners of Higgsinos. If we do not include the color partners, \ie in the MSSM without heavy colored particles, $c_{a\gamma\gamma}\simeq 0$ \cite{BaeComm}.
}
\label{tab:Inv}
\end{table}
\begin{figure}[!t]
\centerline{\includegraphics[width=0.45\textwidth]{figWormFlux}}
\caption{A wormhole connection to a shadow world.}\label{Fig:WormFlux}
\end{figure}
The fine-tuning problem in the DFSZ model is resolved in the supersymmetric(SUSY) extension of the model. There is no renormalizable term of the singlet superfield $\sigma$ with the SM fields. The leading term is the so-called Kim-Nilles term \cite{KN84},
\begin{eqnarray}
W_{\rm KN}= \frac{1}{M}H_uH_d\sigma^2,
\end{eqnarray}
where $M$ is determined from a theory. It is shown in Table \ref{tab:Inv} as the SUSY $c_{a\gamma\gamma}$.
In Table \ref{tab:Inv}, $H_d$ and $H_u^*$ imply that they give mass to $e$. GUTs and SUSY choose appropriate Higgs doublets and always give $c_{a\gamma\gamma}=\frac23$.
In the discussion of ``invisible'' axion, gravity effects was considered to be crucial. It started with the wormhole effects in the Euclidian quantum gravity \cite{Coleman88}. In fact, gravity equation with the antisymmetric tensor field $B_{\mu\nu}$ gives wormhole solutions \cite{GS88}. This triggered the discrete symmetries allowable as subgroups of gauge groups \cite{Krauss89}, and \UPQ~global symmetry needed for ``invisible'' axion
was considered to be problematic \cite{GravSpoil92}. It is euristically presented in Fig. \ref{Fig:WormFlux} for a flow of gauge charges from an observer O to a shadow world S through a wormhole. If he disconnects $S$, he recovers all gauge charges and concludes that no gauge charge is lost. This is because gauge charges carry hairs of flux lines. Thus, O confirms that gauge symmetries are not broken by wormholes. But, global symmetries do not carry flux lines and quantum gravity does not allow global symmetries. Thus, ``invisible'' axion has the gravity spoil problem \cite{GravSpoil92}.
\begin{figure}[!b]
\hskip 1cm {\includegraphics[width=0.47\textwidth]{figDiscV}}\hskip 1cm {\includegraphics[width=0.33\textwidth]{figOnlyA}}\vskip 0.3cm
\centerline{\hskip 1cm(a)\hskip 6.7cm (b)}
\caption{Terms allowed in interactions. (a) Terms allowed by discrete symmetries (the most left column) and non-Abelian anomalies, and (b) terms allowed only by non-Abelian anomalies. Considering only terms in the lavender color, one finds a global symmetry.}\label{fig:Terms}
\end{figure}
One may consider all terms allowed with some discrete symmetries, which is symbolized in the most left column of Fig. \ref{fig:Terms}\,(a). Discrete symmetries appear in most string compactification and the wormhole effects do not break these discrete symmetries. If only a few terms in all possible terms in $V$ are considered as symbolized in the lavender color in Fig. \ref{fig:Terms}, there may appear global symmetries. For example, we can keep terms except the $c_{ij}$ terms in Eqs. (\ref{eq:Weinberg}), which led to the \UPQ~symmetry as discussed before.
These global symmetries are approximate and broken by terms in the red parts in Fig. \ref{fig:Terms}. The red part above the lavender symbolizes the terms in $V$.
One large violation of a global symmetry was suggested for heavy axions or axizillas at the TeV scale \cite{KimAxizilla}. The important one for ``invisible'' axion is the case of no $\Delta V$ as shown in Fig. \ref{fig:Terms}\,(b). Then, the minimum of the potential is at $\bar{\theta}=0$ \cite{VW84}. If a term in $\Delta V$ is present, it must be sufficiently small such that the ``invisible'' axion solution of the strong CP problem is intact.
In any case, a PQ global symmetry can be obtained at least approximately from an ultra-violet completed theory.
The well-known global symmetric operators are the effective neutrino mass term in the Lagrangian (instead of $V$) for U(1)$_L$ \cite{Weinberg79} and the Kim-Nilles term in SUSY for \UPQ,
\begin{equation}
{-\cal L}_{\nu\rm mass}=\frac{f_{ij}}{M}\ell_i\ell_j H_uH_u,~~~W_{\rm KN}=\frac{1}{M}\sigma_1\sigma_2 H_uH_d,\label{eq:def}
\end{equation}
where $\ell$'s are lepton doublets in the SM. Equations in (\ref{eq:def}) define the lepton number $L$ and the PQ quantum number $Q$,
\begin{eqnarray}
&&L(\ell\ell)=2,~\quad L(H_uH_u)=-2,\label{eq:Lnumber0}\\
&&Q(\sigma_1\sigma_2)=2,
~\quad Q(H_uH_d)=-2.\nonumber
\end{eqnarray}
For masses of the SM neutrinos, $\langle H_uH_u\rangle\ne 0$ is enough. The renormalizable interactions with right-handed neutrinos $N_R$ leading to Eq. (\ref{eq:Lnumber0}), $\ell_L H_u N_R$, are not needed for phenomenological neutrino masses at low energy.
For the KN term also, renormalizable couplings, $\sigma_1 H_u X_{\rm doublet}$, $\sigma_2 H_d X'_{\rm doublet}, \overline{Q}_LQ_R\sigma_1,\cdots$, can give the term.
In this road toward detecting an ``invisible'' QCD axion, there has been a few theoretical development starting from an ultra-violet completed theory. The scale must be intermediate. The model-indepent(MI) superstring axion \cite{MIaxion} is not suitable for this because the decay constant is about $10^{16\,}\gev$ \cite{ChoiKim85} which is the white square on the upper left corner in Fig. \ref{Fig:AxData}.
The question is, ``is it possible to obtain exact global symmetries?''
From string compactification, there is one way to make the ``invisible'' QCD axion to be located at the intermediate scale starting with an exact global symmetry,
\begin{equation}
10^9\,\gev\le f_a\le 10^{11.5\,}\gev.
\end{equation}
It starts from the appearance of
an anomalous U(1) gauge symmetry in string compactification. In compactifying the E$_8\times$E$_8'$ heterotic string, there appears an anomalous U(1)$_a$ gauge symmetry in many cases \cite{Anom86},
\begin{equation}
{\rm E}_8\times {\rm E}_8\to {\rm U(1)}_a\times \cdots
\end{equation}
Thus, the anomalous U(1)$_a$ is belonging to a gauge symmetry of E$_8\times$E$_8'$. In the original E$_8\times$E$_8'$ heterotic string, there is also the MI-axion degree $B_{\mu\nu}$. The gauge boson corresponding to this anomalous \Uanom~ obtains mass by absorbing the MI-axion degree as its longitudinal degree. Therefore, the harmful MI-axion disappears, but not quite completely. Below the compactification scale of $10^{17\,}\gev$, there appears a global symmetry which works as the PQ symmetry. This PQ symmetry can be broken by a SM singlet Higgs scalar(s), producing the ``invisible'' axion. So, this ``invisible'' axion arises from an exact global symmetry \Uanom, and is free from the gravity spoil problem because its origin is gauge symmetry.
Within this string compactification scheme, the axion-photon-photon coupling has been calculated \cite{Kim88,KimPLBagg15,KimNamPLB16} and presented in Table \ref{tab:StAxion}.
\begin{table}
\begin{tabular}{|l| r|c|}
\hline &&\\[-1em]
String: &$c_{a\gamma\gamma}$ &Comments\\\hline\\[-1.35em]
&& \\[-0.9em]
Ref. \cite{ChoiIWKim} & $-\frac13$~ &Approximate\\[0.3em]
Ref. \cite{KimPLBagg15,KimNamPLB16} & $\frac23$~ &Anom. U(1)\\[0.3em]
\hline
\end{tabular}
\caption{String model prediction of $c_{a\gamma\gamma}$. In the last line, $c_{a\gamma\gamma}= (1-2\sin^2\theta_W)/\sin^2\theta_W$ with $m_u=0.5\, m_d$.
}
\label{tab:StAxion}
\end{table}
\section{CP and cosmology}
Consideration of CP in cosmological
dark energy (DE), CDM, and $\Delta B$ can be realated CP symmetry and its violation. Quintessential axion \cite{QuintAxion}, QCD axion, and Sakharov's conditions are related to CP. $\Delta B$ needs
CP violation. In the Type-II leptogenesis, which will be given shortly, involves the weak CP violation.
The axion solution of the strong CP problem is a cosmological solution. QCD
axions oscillate with the CP violating
vacuum angle $\bar\theta$, but the average value is 0. If the axion vacuum starts from $a/f_a=\theta_1\ne 0$, then the vacuum oscillates and this collective motion behaves like cold dark matter(CDM) as commented around Fig. \ref{Fig:AxData} \cite{BCM14}, for which a recent calculation energy density of coherent oscillation, $\rho_a$, is given in \cite{Bae08}.
\begin{figure}[!b]
\hskip 1cm {\includegraphics[width=0.67\textwidth]{figDW3}}
\caption{The axion vacuum with $\NDW=3$.}\label{fig:DWN3}
\end{figure}
The axion vacuum is identified by the shift of axion field by $2\pi\NDW f_a$,
\begin{equation}
a\to a+2\pi \,\NDW\,f_a.
\end{equation}
It is because the $\bar\theta$ term has the periodicity $2\pi$,
\begin{equation}
{\cal L}_{\bar\theta}=-\frac{a}{32\pi^2\,f_a}\int d^4x \,G^a_{\mu\nu}\tilde{G}^{a\,\mu\nu},~~a=a+2\pi f_a,
\end{equation}
while matter fields $\Phi$ may not have the periodicity of $2\pi f_a$, but only after $2\pi\NDW f_a$,
\begin{equation}
\Phi\to e^{i\bar\theta/\NDW}\Phi,~~
\bar\theta=\frac{a}{f_a}.
\end{equation}
The axion vacuum with $\NDW=3$ is shown in Fig. \ref{fig:DWN3}, where the red triangle vacuum is returning to itsef after going over three maxima. Between different vacua, there are domain walls.
\begin{figure}[!t]
\hskip 1cm {\includegraphics[width=0.47\textwidth]{f52}}\hskip 1cm {\includegraphics[width=0.33\textwidth]{f5Hor2}}\vskip 0.3cm
\centerline{\hskip 1cm(a)\hskip 6.7cm (b)}
\caption{Domain walls in $\Z_2$. (a) Domain wall ball, and (b) a cosmological scale domain wall.}\label{fig:DW2}
\end{figure}
Topological defects of global \UG~produce an additional axion energy density by the decay of string-wall system, $\rho_{st}$. Contribution of axionic string to energy density was known for a long time \cite{Davis86}. In addition, axionic domain walls carry huge energy density \cite{Vilenkin82, Sikivie82}. Because of the difficulty of removing comological scale domain walls for $\NDW\ge 2$ as sketched in Fig. \ref{fig:DW2}, it was suggested that
the axionic domain wall number should be 1 \cite{Sikivie82}. If $\NDW=1$, the horizon scale walls can be annihilated as in Fig. \ref{fig:DW1} \cite{Vilenkin82,Barr87}. A small wall bounded by string collides with a horizon scale wall (a), and eats up the wall with the light velocity (b), eventually annihilating the horizon scale string-wall system (c).
Computer simulations use axion models with $\NDW=1$.
Three groups have calculated these which vary from O(1) to O(100),
\begin{eqnarray}
&&\textrm{Florida group:}~\textrm{O(1)} \cite{Florida01},\nonumber\\
&&\textrm{Cambridge group:}~\textrm{O(100)} \cite{Cambridge94}, \label{eq:StContri}\\
&&\textrm{Tokyo group:}~\textrm{O(10)} \cite{Kawasaki12}.\nonumber
\end{eqnarray}
A recent calculation for $\NDW=1$ models has been given $\rho_{st}\sim {\rm O(10)}\,\rho_a $ \cite{SekiguchiT}.
\begin{figure}[!t]
\hskip 1cm {\includegraphics[width=0.67\textwidth]{f51}}
\caption{Domain walls in $\NDW=1$. (a) A DW ball is approaching to a horizon scale DW, (b) the DW ball is punching the huge wall and expands with the light velocity, and (c) it eats up the huge wall \cite{Barr87}.}\label{fig:DW1}
\end{figure}
Therefore, it is important to realize axion models with $\NDW=1$. The KSVZ axion model with one heavy quark achieves $\NDW=1$. There are two other methods. One is identifying different vacua modulo the center number of the GUT gauge group \cite{LS82}. Another important one is obtaining $\NDW=1$ by the Goldstone boson direction \cite{Choi85,KimPLB16},
\begin{figure}[!t]
\hskip 1cm {\includegraphics[width=0.67\textwidth]{f5Gold}}
\caption{Vacua identification by Goldstone boson direction.}\label{fig:Goldstone}
\end{figure}
which is shown in Fig. \ref{fig:Goldstone}.
There are two degrees for the shifts, $N_1$ and $N_2$ directions. For the torus of $N_1=3$ and $N_2=2$ models, seemingly there are 6 vacua represended by red bullets in Fig. \ref{fig:Goldstone}. The Goldstone boson directions are shown as arrow lines and torus identifications are shown as dashed arrows. So, all six vacua are connected by one way or the other, and the $N_1=3$ and $N_2=2$ model gives $\NDW=1$. One always obtain $\NDW=1$ if $N_1$ and $N_2$ are relatively prime. The reason that the ``invisible'' axion from \Uanom~ has $\NDW=1$ is because $N_{\rm from~E_8\times E_8'}=\rm large~integer$ but $N_{\rm MI~axion}=1$ \cite{Witten85,KimPLB16}, and $N_{\rm from~E_8\times E_8'}$ and $N_{\rm MI~axion}$ are relatively prime.
There is another cosmological solution that the axionic string-wall system created at $f_a$ is allevated by another confining force at high temperature \cite{BarrKim14}.
The instantons of this force, acting below $f_a$, can generate an axion potential that erases the axion strings long before QCD effects become important, preventing QCD-generated axion walls from ever appearing.
Axionic string contribution is important if strings are created after PQ
symmetry breaking. On the other hand, with a high scale inflation this string contribution to energy density is important, as shown in Eq. (\ref{eq:StContri}). There can be a more important constraint if a large $r(=\,$tensor/scalar ratio) is observed. Two groups reported this constraint \cite{Gondolo14} after the BICEP2 report \cite{BICEP2}. Probably, this is the most significant impact of BICEP2 result on
$\NDW=1$ axion physics. The region is marked around $m_a\sim 71\,\mu$eV in Fig. \ref{Fig:AxData}.
\section{Type-II leptogenesis}
The mere 5\,\% of the energy pie, mainly atoms composed of baryons of the Universe, belongs to the problem on the chiral representations of quarks in the GUT scheme \cite{Georgi79,KimICHEP1,KimJHEP15}. In cosmology, it belongs to Sakharov's three conditions \cite{Sakharov67},
\begin{itemize}
\item $B$ number (or global quantum number) violation,
\item CP and C violation,
\item Out of thermal equilibrium.
\end{itemize}
For the last condition, we just make sure that the process proceeds
in non-equilibrium conditions. Usually, heavy particle decays proceed out of thermal equilibrium.
The first one is the existence of baryon number violating interaction implied in GUTs, and the second condition is on the CP and C violation.
Sphaleron processes at the electroweak scale changes baryon ($B$) and lepton($L$) numbers but do not change the combination $(B-L)$. If the sphaleron processes are 100\,\% effective in converting these global quantum numbers, a net baryon number is $\Delta B|_{\rm after~sphaleron}\propto (B-L)$.
Thus, SU(5) GUT is problematic because it conserves $B-L$ and cannot generate $(B-L)$ by the SU(5) processes. So, if one tries to use GUTs, choose one which violates $B-L$ such as in the SO(10) GUT.
But, a simple way is to work in the SM,
SU(3)$\times$SU(2)$\times$U(1), and introduce SM singlet fermions $N$ which are called neutrinos. These $N$ particles are considered to be heavy. This issue is the first condition on the global quantum number. If the SM singlets are present, there always appear neutrino masses, which implies that $L$ is broken. In defining the lepton number $L$, let us remember the Weinberg operator for neutrino masses \cite{Weinberg79,Minkowski77},
\begin{equation}
{\cal L}_{\nu\rm mass}=\frac{f_{ij}}{M}\ell_i\ell_j H_uH_u \label{eq:Wein}
\end{equation}
where $\ell$ are left-handed lepton doublets, and gauge invariant indices are implied. $M$ is an effective mass in the non-renormalizable operator. The lepton number is defined with the left-handed SM doublets $\ell_L(\ni\nu_{e,\mu,\tau})$ carrying $L=+1$.
Non-zero neutrino masses break $L$. So, as far as $\langle H_u\rangle=0$, neutrinos do not obtain mass by the above operator. Therefore, $L$ can be properly defined below the scale of $\langle H_u\rangle\ne 0$. One may argue that there is also $H_d$, which is exactly the reason that $H_u$ can carry a global quantum number, because both $H_u$ and $H_d$ can together define a gauge charge $Y$ and a global charge $L$. Two $\ell$'s carry $+2$. However, who cares about renormalizable terms very importantly at low energy?
In cosmology, however, it is important.
In cosmology, lepton number of $H_u$ is defined by
\begin{itemize}
\item[]{\bf Type-I}:~~~ $L(H_uH_u)=0$,
\item[]{\bf Type-II}:~~ $L(H_uH_u)=-2$.
\end{itemize}
The quantum numbers of Type-I \cite{FY86} and Type-II \cite{CoviKim16} leptogeneses are suggested by the following renormalizable term, and the quantum numbers of $L$,
\begin{eqnarray}
\ell_L~~~H_u~~~~~N_L&&\\
{\rm Type-I}:~ +1,~~~0,~~~-1&&\\
{\rm Type-II}:~ +1,~ -1,~~~~~0&&
\end{eqnarray}
The early idea was to define a right(left)-handed $N$ with lepton number $L=+1(-1)$ \cite{FY86}. In fact, the definition of lepton number, related to neutrino masses, is a combination of defining the lepton number of the up-type Higgs doublets together with that of $N$.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.55\textwidth]{figIntab.eps}\\ ~\vskip 0.3cm
\includegraphics[width=0.65\textwidth]{figIntcd.eps}\hskip 1.2cm \includegraphics[width=0.2\textwidth]{figInttree.eps}
\end{center}
\caption{The Feynman diagrams interfering in the $N$ decay: (a) the lowest order diagram in Type-II, (b) the $W$ exchange diagram, (c) the wave function renormalization diagram, and (d) the heavy neutral lepton exchange diagram. There exist similar ${\cal N}$-decay diagrams. In all figures, the final leptons can be both charged leptons and neutrinos.
The lepton number violations are inserted with blue bullets and phases are inserted at red bullets. (a) and (b) interfere. (c) and (d) give a vanishing contribution in $N_0$ domination with one complex VEV.} \label{fig:IntCP}
\end{figure}
`Type-II leptogenesis' is proposed under a different definition on $L$ from that of Type-I and the electroweak symmetry breaking at high temperature. There exists a finite region of parameter space in the multi-Higgs model that the electroweak symmetry is broken at high temperature \cite{Senj79}. In SUSY, the electroweak symmetry breaking at high temperature is more probable since the temperature dependent terms in $V$ are SUSY breaking terms.
In Fig. \ref{fig:IntCP}, we present diagrams appearing in Type-I and Type-II leptogeneses. The SM Higgs doublet is $h$, and an inert Higgs doublet is $H$. (e), (c), and (d) give Type-I where both the $L$ violation and CP violation appear in blue bullets. In Type-I, CP violation is introduced at the mass term of $N$, needing at least two $N$'s for the interference of Figs. \ref{tab:DefQno}\,(e), (c), and (d). So, if there is a hierarchy of masses such as $m_{N_0}/m_{N_1}\ll 1$, then the lepton asymmetry is suppressed by that factor.
\begin{table}
\centerline{\begin{tabular}{|l| ccc|}
\hline &&&\\[-1em]
&$\ell_L$ &$h_u\,(H_u)$ & $N$\\ \hline\\[-1.4em]
&&& \\[-1em]
Type-I leptogenesis & ~$+1$~ &~$0$~&~$-1$~\\[0.3em]
Type-II leptogenesis & ~$+1$~ &~$-1\,(-2)$~ &$0$~\\[0.3em]
\hline
\end{tabular}}
\caption{Definition of lepton numbers, in Type-I and Type-II leptogeneses.}
\label{tab:DefQno}
\end{table}
\begin{table}
\centerline{\begin{tabular}{|c| c ccccc|}
\hline &&&&&&\\[-1em]
&~$\ell_L$~ &$H_u$&$H_d$ & $h_{u}$&$N$ & ${\cal N}$\\ \hline\\[-1.4em]
&&&&&& \\[-1em]
Type-II leptogen. & ~$+1$~ &~$-2$~&~$+2$~ &~$-1$~&~$0$~&~$+1$~\\[0.3em]
VEV & ~$\times$~ &~\textrm{inert}~ & $ \textrm{inert}$ & ~$v_{\rm ew}\,s_\beta$ &~$\times$~&~$\times$~\\[0.3em]
\hline
\end{tabular}}
\caption{Definition of lepton numbers in Type-II leptogeneses. We introduced an inert Higgs $H_u$ carrying $L=-2$ with zero VEV and singlet leptons ${\cal N}$ carrying $L=+1$.}
\label{tab:Type2}
\end{table}
In Fig. \ref{fig:IntCP}, (a) and (b) give Type-II, and CP violation is introduced by the PMNS matrix. Thus, we must have $W$ boson is not massless, \ie SU(2)$\times$U(1) symmetry is broken at the high temperature. Indeed, such possibility has been suggested long time ago \cite{Senj79}. One can define the {\it lepton number} $L$ as shown in Table \ref{tab:DefQno}. In Type-II, we detail the lepton numbers in Table \ref{tab:Type2}. Here, we need an inert Higgs doublet $H_u$ carrying $L=-2$ and singlet leptons ${\cal N}$ carrying $L=+1$. Here, the lepton number violation appears in blue bullets and the CP violation appears in red bullets. Anyway, the fields $H_{u,d}$ and ${\cal N}$ introduced at high energy scale are not visible at low energy scale. We introduce interactions
\begin{eqnarray}
&N_0\,\ell_L h_u,~{\cal N}_0\,\ell_L H_u,~N_0\,N_0,~H_uH_d,\cdots\\
&h_u^*H_u, ~N_0\,{\cal N}_0,~{\cal N}_0{\cal N}_0, \cdots
\end{eqnarray}
where the first line conserves $L$ and the second line violates $L$.
Within this framework, we calculated the lepton asymmetry which turned out to be
consistent with the fact that in these diagrams
the lepton violation is on the left side of the cut diagram as discussed in \cite{Adhikari:2001yr}.
The lepton asymmetry we obtain is a form,
\begin{eqnarray}
\epsilon_L^{N_0}(W)&
\approx \frac{\alpha_{\rm em}}{2\sqrt{2}\sin^2\theta_W} \frac{\Delta m_h^2 }{m_0^2}\\
&\cdot\sum_{i,j}{\cal A}_{ij}\sin[(\pm n_P+n'-n_i+n_j)\delta_X],
\label{eq:asymmExp4}
\end{eqnarray}
where ${\cal A}_{ij}$ are given by Yukawa couplings, $n_P,n',n_i,n_j$ are integers, and we assumed that only one phase $\delta_X$ appears in the full theory. Two independent $n'$ are Majorana phases multiplied to the PMNS matrix.
Using the sphaleron calculation of \cite{D'Onofrio:2014kta}, we obtain
\begin{eqnarray}
\frac{\Gamma_{\rm sph}^{\rm broken}}{T^3 H(T)} =
\kappa \alpha_W^4 \left(\frac{ 4\pi k}{g_W} \right)^7 e^{- 1.52 k \frac{4\pi}{g_W}} \sqrt{\frac{90}{\pi^2 g_*}}
\frac{M_P}{T} \geq 1,
\end{eqnarray}
leading to the constraint
\begin{eqnarray}
C_2 C_3 \sin\delta_c + C_1 S_2 S_3 \sin (\delta_c +
\dell) \simeq 2.4 \times 10^{-2},
\end{eqnarray}
where $\delta_c$ is a Majorana neutrino phase. So, there is an enough parameter space to allow an acceptable $\Delta L$.
\section{Conclusion}
My talk is centered around CP symmetry/violation and its cosmological effects. A few emphases were:
(1) the Jarlskog determinant $J$ is $|{\rm Im\,} V_{31} V_{22} V_{13}|$ in the KS form,
(2) it is shown that the Jarlskog determinant $J$ is almost maximum with the current determination of quark (real) mixing angles, (3) why we need ``invisible'' axions for a solution of strong CP problem, (4) we commented some cosmological problems, (5) the ``invisible'' axion from an exact \UG~with $\NDW=1$ is possible if it arises from \Uanom~ in string compactification, and (6) we discussed also the recent idea of Type-II leptogenesis.
\paragraph{Note added.} After the talk, there appeared some relevant papers \cite{UanomFl,Dvali16,DiscBUp}.
|
\section{Introduction}
\label{sec:intro}
The cost sharing problem asks for an equitable way to split the cost of a service among all of the participants. Formally, there is a cost function defined over all subsets of a ground set of elements (or players) and the objective is to fairly divide the cost of the full set among the participants. Cost sharing is central to cooperative game theory, and there is a rich literature developing the key concepts and principles to reason about this topic. Two popular cost sharing concepts are the {\em core}~\cite{gillies1959solutions}, where no group of players has an incentive to deviate, and the {\em Shapley value}~\cite{shapley2016value}, which is the unique vector of cost shares satisfying four natural axioms.
While both the core and the Shapley value are easy to define, computing them poses additional challenges. One obstacle is that the computation of the cost shares requires knowledge of costs in myriad different scenarios. For example, computing the exact Shapley value requires one to look at the marginal contribution of a player over {\em all possible subsets}. Recent work ~\cite{liben2012computing} shows that one can find approximate Shapley values for a restricted subset of cost functions by looking at the costs for polynomially many specifically chosen examples. In practice, however, another roadblock emerges: one cannot simply query for the cost of a hypothetical scenario. Rather, the costs for scenarios that have not occurred are simply unknown. We share the opinion of \citet{BPZ-15} that the main difficulty with using cost sharing methods in concrete applications is the information needed to compute them.
Concretely, consider the following cost sharing applications.
\paragraph{Attributing Battery Consumption on Mobile Devices.}
A modern mobile phone or tablet is typically running a number of distinct apps concurrently. In addition to foreground processes, a lot of activity may be happening in the background: email clients may be fetching new mail, GPS may be active for geo-fencing applications, messaging apps are polling for new notifications, and so on. All of these activities consume power; the question is how much of the total battery consumption should be attributed to each app? This problem is non-trivial because the operating system induces cooperation between apps to save battery power. For example there is no need to activate the GPS sensor twice if two different apps request the current location almost simultaneously.
\paragraph{Moneyball and Player Ratings}
The impact of an individual player on the overall performance of the team typically depends on the other players currently playing. One can infer the total benefit from the players on the field (or on the court) from metrics like number of points scored, time of possession, etc., the question here is how to allocate this impact to the individuals. Recently many such metrics have been proposed (for example plus/minus ratio in hockey, wins above replacement in baseball.), our goal here is to find scores compatible with cooperative game theory concepts.
\paragraph{Understanding Black Box Learning}
Deep neural networks are prototypical examples of black box learning, and it is almost impossible to tease out the contribution of a particular feature to the final output. Particularly, in situations where the features are binary, cooperative game theory gives a formal way to analyze and derive these contributions. While one can evaluate the objective function on any subset of features, deep networks are notorious for performing poorly on certain out of sample examples~\cite{AdversarialNN, AdversarialNN2}, which may lead to misleading conclusions when using traditional cost sharing methods.
We model these cost sharing questions as follows. Let $N$ be the set of possible players (apps or features), and for a subset $S \subseteq N$, let $C(S)$ denote the cost of $S$. This cost represents the total power consumed over a standard period of time, or the number of points scored by the team, and so on. We are given ordered pairs $(S_1, C(S_1)), (S_2, C(S_2)), \ldots,$ $(S_m, C(S_m))$, where each $S_i \subseteq N$ is drawn independently from some distribution $\mathcal{D}$. The problem of \textsc{Statistical Cost Sharing}\ asks to look for reasonable cost sharing strategies in this setting.
\subsection{Our results}
We build on the approach from \citet{BPZ-15}, which studied \textsc{Statistical Cost Sharing}\ in the context of the core, and assume that only partial data about the cost function is observed. The authors showed that cost shares that are likely to respect the core property can be obtained for certain restricted classes of functions. Our main result is an algorithm that generalizes these results for {\em all} games where the core is non-empty and we derive sample complexity bounds showing exactly the number of samples required to compute cost shares (Theorems~\ref{t:vccore} and \ref{thm:approx_core_learn}).\footnote{Concurrently and independently of our work, \citet{BPZ-16} proved a polynomial sample complexity bound for this problem of computing cost shares that are likely to respect the core property, for all functions with a non-empty core.} While the main approach of \citet{BPZ-15} relied on first learning the cost function and then computing cost shares, we show how to proceed directly, computing cost shares without explicitly learning a good estimate of the cost function. We also show that approximately satisfying the core with probability one is impossible in general (Theorem~\ref{t:core}).
We then focus on the Shapley value, which has never been studied in the \textsc{Statistical Cost Sharing} \ context. We introduce a new cost sharing method called \emph{data-dependent Shapley value} which is the unique solution (Theorem~\ref{t:ddshapley}) satisfying four natural axioms resembling the Shapley axioms (Definition~\ref{d:ddshapley}), and which can be approximated arbitrarily well from samples for any bounded function and any distribution (Theorem~\ref{t:apxddshapley}). Regarding the traditional Shapley value, we obtain a tight $\sqrt{1 - \kappa}$ multiplicative approximation for submodular functions with bounded curvature $\kappa$ over the uniform distribution (Theorems~\ref{t:curv} and \ref{t:lbcurv}), but show that they cannot be approximated by a bounded factor in general, even for the restricted class of coverage functions, which are learnable, over the uniform distribution (Theorem~\ref{thm:lower}).
\subsection{Related work}
There are two avenues of work which we build upon. The first is the notion of cost sharing in cooperative games, first introduced by \citet{von2007theory}. We consider the Shapley value and the core, two popular solution concepts for cost-sharing in cooperative games. The Shapley value \cite{shapley2016value} is studied in algorithmic mechanism design \cite{anshelevich2008price,balkanski2015mechanisms,feigenbaum2000sharing,moulin1999incremental}. For applications of the Shapley value, see the surveys by \citet{roth1988shapley} and \citet{winter2002shapley}. A naive computation of the Shapley value of a cooperative game would take exponential time; recently, methods for efficiently approximating the Shapley value have been suggested \cite{bachrach2010approximating,fatima2008linear,liben2012computing,mann1960values} for some restricted settings.
The core, introduced by \citet{gillies1959solutions}, is another well-studied solution concept for cooperative games. \citet{bondareva1963some} and \citet{shapley1967balanced} characterized when the core is non-empty. The core has been studied in the context of multiple combinatorial games, such as facility location \cite{goemans2004cooperative} and maximum flow \cite{deng1999algorithmic}. In cases with no solutions in the core or when it is computationally hard to find one, the balance property has been relaxed to hold approximately \cite{devanur2005strategyproof, immorlica2008limitations}. In applications where players submit bids, cross-monotone cost sharing, a concept stronger than the core that satisfies the group strategy proofness property, has attracted a lot of attention~\cite{immorlica2008limitations,jain2002equitable,moulin2001strategyproof,pal2003group}. We note that these applications are sufficiently different from the ones we are studying in this work.
The second is the recent work in econometrics and computational economics that aims to estimate critical concepts directly from a limited data set, and reason about the sample complexity of the computational problems. Specifically, in all of the above papers, the algorithm must be able to query or compute $C(S)$ for an arbitrary set $S \subseteq N$. In our work, we are instead given a collection of samples from some distribution; importantly the algorithm does not know $C(S)$ for sets $S$ that were not sampled. This approach was first introduced by \citet{BPZ-15}, who showed how to compute an approximate core for some families of games. Their main technique is to first learn the cost function $C$ from samples and then to use the learned function to compute cost shares. The authors also showed that there exist games that are not PAC-learnable but that have an approximate core that can be computed.
Outside of cooperative game theory, this data-driven approach has attracted renewed focus. In auction design, a line of work \cite{chawla2014mechanism,cole2014sample, dughmi2014sampling, morgenstern2015pseudo} has studied revenue maximization from samples instead of being given a Bayesian prior. In the inductive clustering setting, the algorithm is only given a small random subset of the data set it wishes to cluster \cite{balcan2009finding,balcan2009agnostic}. More closely related to our work is the problem of optimization from samples \cite{ curvature, BRS17} where, the goal is to approximate $\max_{S \in M} C(S)$ for some constraint $M \subseteq 2^N$ and $C : 2^N \rightarrow \mathbb{R}$ from samples for some class of combinatorial functions.
\section{Preliminaries}
A \emph{cooperative game} is defined by an ordered pair $(N, C)$, where $N$ is the ground set of \emph{elements}, also called \emph{players}, and $C : 2^N \rightarrow \mathbb{R}_{\geq 0}$ is the \emph{cost function} mapping each \emph{coalition} $S \subseteq N$ to its cost, $C(S)$. The ground set of size $n = |N|$ is called the \emph{grand coalition} and we denote the elements by $N = \{1, \ldots, n\} = [n]$. We assume that $C(\emptyset) = 0$, $C(S) \ge 0$ for all $S \subseteq N$, and that $\max_S C(S)$ is bounded by a polynomial in $n$, which are standard assumptions. We will slightly abuse notation and use $C(i)$ instead of $C(\{i\})$ for $i \in N$ when it is clear from the context.
We recall three specific classes of functions. \emph{Submodular} functions exhibit the property of diminishing returns: $C_S(i) \geq C_T(i)$ for all $S \subseteq T \subseteq N$ and $i \in N$ where $C_S(i)$ is the marginal contribution of element $i$ to set $S$, i.e., $C_S(i) = C(S \cup \{i\}) - C(S)$.
\emph{Coverage} functions are the canonical example of submodular functions. A function is coverage if it can be written as $C(S) = | \cup_{i \in S} T_i|$ where $T_i \subseteq U$ for some universe $U$.
Finally, we also consider the simple class of additive functions that are such that $C(S) = \sum_{i\in S}C(i)$.
A \emph{cost allocation} is a vector $\psi \in \mathbb{R}^n$ where $\psi_i$ is the \emph{share} of element $i$.
We call a cost allocation $\psi$ \emph{balanced} if $\sum_{i \in N} \psi_i = C(N)$. Given a cooperative game $(N, C)$ the goal in the cost sharing literature is to find ``desirable" balanced cost allocations. Most proposals take an axiomatic approach, defining a set of axioms that a cost allocation should satisfy. These lead to the concepts of Shapley value and the core, which we define next. A useful tool to describe and compute these cost sharing concepts is permutations. We denote by $\sigma$ a uniformly random permutation of $N$ and by $S_{\sigma < i}$ the players before $i$ in permutation $\sigma$.
\subsection{The core}
The core is a balanced cost allocation where no player has an incentive to deviate from the grand coalition---for any subset of players the sum of their shares does not cover their collective cost.
\begin{definition} A cost allocation $\psi$ is in the \emph{core} of function $C$ if the following properties are satisfied:
\begin{itemize}
\item \textbf{Balance:} $\sum_{i \in N} \psi_i = C(N)$,
\item \textbf{Core property:} for all $S \subseteq N$, $\sum_{i \in S} \psi_i \leq C(S)$.
\end{itemize}
\end{definition}
The core is a natural cost sharing concept. For example, in the battery blame scenario it translates to the following assurance: No matter what other apps are running concurrently, an app is never blamed for more battery consumption than if it were running alone. Given that app developers are typically business competitors, and that a mobile device's battery is a very scarce resource, such a guarantee can rather neatly avoid a great deal of finger-pointing. Unfortunately, for a given cost function $C$ the core may not exist (we say the core is empty), or there may be multiple (or even infinitely many) cost allocations in the core. For submodular functions $C$, the core is guaranteed to exist and one allocation in the core can be computed in polynomial time. Specifically, for any permutation $\sigma$, the cost allocation $\psi$ such that $\psi_i = C(S_{\sigma < i} \cup \{i\}) - C(S_{\sigma < i})$ is in the core.
\subsection{The Shapley value}
The Shapley value provides an alternative cost sharing method. For a game $(N, C)$ we denote it by $\phi^C$, dropping the superscript when it is clear from the context.
While the Shapley value may not satisfy the core property, they satisfy the following four equally natural axioms:
\begin{itemize}
\item \textbf{Balance:} $\sum_{i \in N} \phi_i = C(N)$.
\item \textbf{Symmetry:} For all $i, j \in N$, if $C(S \cup \{i\}) = C(S \cup \{j\})$ for all $S \subseteq N \setminus \{i, j\}$ then $\phi_i = \phi_j$.
\item \textbf{Zero element:} For all $i \in N$, if $C(S \cup \{i\}) = C(S)$ for all $S \subseteq N$ then $\phi_i = 0$.
\item \textbf{Additivity:} For two games $(N, C_1)$ and $(N, C_2)$ with the same players, but different cost functions $C_1$ and $C_2$, let $\phi^1$ and $\phi^2$ be the respective cost allocations. Consider a new game $(N, C_1 + C_2)$, and let $\phi'$ be the cost allocation for this game. Then for all elements, $i \in N$, $\phi'_i = \phi^1_i + \phi^2_i$.
\end{itemize}
Each of these axioms is natural: balance ensures that the cost of the grand coalition is distributed among all of the players. Symmetry states that two identical players should have equal shares. Zero element verifies that a player that adds zero cost to any coalition should have zero share. Finally, additivity just confirms that costs combine in a linear manner. It is surprising that the set of cost allocations that satisfies all four axioms is unique. Moreover, the Shapley value $\phi$ can be written as the following summation:
$$ \phi_i = \EU[\sigma][C(S_{\sigma < i} \cup\{i\}) - C(S_{\sigma < i})] = \sum_{S \subseteq N \setminus\{i\}}\frac{|S|!(n - |S| - 1)!}{n!} (C(S \cup\{i\}) - C(S)).$$
This expression is the expected marginal contribution $C(S \cup\{i\}) - C(S)$ of $i$ over a set of players $S$ who arrived before $i$ in a random permutation of $N$. As the summation is over exponentially many terms, the Shapley value generally cannot be computed exactly in polynomial time. However, several sampling approaches have been suggested to approximate the Shapley value for specific classes of functions~\cite{bachrach2010approximating,fatima2008linear,liben2012computing,mann1960values}.
\subsection{Statistical cost sharing}
With the sole exception of ~\cite{BPZ-15}, previous work in cost-sharing critically assumes that the algorithm is given oracle access to $C$, i.e., it can query, or determine, the cost $C(S)$ for any $S \subseteq N$. In this paper, we aim to (approximately) compute the Shapley value and other cost allocations from \emph{samples}, without oracle access to $C$, and with a number of samples that is polynomial in $n$.
\begin{definition}\label{def:scs}
Consider a cooperative game with players $N$ and cost function $C$. In the \textsc{Statistical Cost Sharing}\ problem we are given pairs $(S_1, C(S_1)), (S_2, C(S_2)), \ldots, (S_m, C(S_m))$ where each $S_i$ is drawn i.i.d. from a distribution $\mathcal{D}$ over $2^N$. The goal is to find a cost allocation $\psi \in \mathbb{R}^n$.
\end{definition}
In what follows we will often refer to an individual $(S, C(S))$ pair as a {\em sample}. It is tempting to reduce \textsc{Statistical Cost Sharing}\ to classical cost sharing by simply collecting enough samples to use known algorithms. For example, \citet{liben2012computing} showed how to approximate the Shapley value with polynomially many queries $C(S)$. However, if the distribution $\mathcal{D}$ is not aligned with these specific queries, which is the case for the uniform distribution, emulating these algorithms in our setting requires exponentially many samples. \citet{BPZ-15} showed how to instead first learn an approximation to $C$ from the given samples and then compute cost shares for the learned function, but their results hold only for a limited number of games and cost functions $C$. We show that a more powerful approach is to compute cost shares directly from the data, without explicitly learning the cost function first.
\subsection{Warm up: linear functions and product distributions}
As a simple example, we consider the special case of additive functions with $C(i) \geq 1 / \poly(n)$ and on bounded product distributions $\mathcal{D}$.\footnote{A bounded product distribution has marginal probabilities bounded below and above by $1/\poly(n)$ and $1 - 1/\poly(n)$.} In this setting, the core property and the Shapley value can be approximated arbitrarily well. It is easy to verify that the cost allocation $\psi$ such that $\psi_i = C(i)$ is in the core and that it is the Shapley value.
To compute these cost shares, we
estimate the expected marginal contribution of an element $i$ to a random set, i.e., $v_i^{\mathcal{D}} := \E_{S \sim \mathcal{D}|i \not \in S}[C(S \cup \{i\}) - C(S)]$.
Note that in the case of additive functions, $v_i^{\mathcal{D}} = C(i)$. In addition,
$$ v_i^{\mathcal{D}} = \EU[S \sim \mathcal{D}|i \not \in S][C(S \cup \{i\}) - C(S)] = \EU[S \sim \mathcal{D}|i \in S][C(S)] - \EU[S \sim \mathcal{D}|i \not \in S][C(S)],$$
when $\mathcal{D}$ is a product distribution. Thus, this value can be estimated arbitrarily well by looking at the difference in cost between the average value of the samples containing $i$ and the average of those not containing $i$. The analysis is a simple concentration bound and is deferred to the appendix.
\begin{restatable}{rLem}{lconcentration}
\label{l:addconcentration}
Let $C$ be an additive function with $C(i) \geq 1 / \poly(n)$ for all $i$. Then, given $\poly(n, \nicefrac{1}{\delta}, \nicefrac{1}{\epsilon})$ samples, we can compute an estimate $\tilde{v}_i$ such that with probability $(1 - \delta)$:
$$ (1 - \epsilon) C(i) < \tilde{v}_i < (1+\epsilon) C(i).$$
\end{restatable}
Thus an algorithm which computes $\psi_i =\tilde{v}_i$ approximates the Shapley value and the core property arbitrarily well (the formal definitions of approximating the core and the Shapley value are deferred to the respective sections devoted to those concepts).
\section{Approximating the Core from Samples}
\label{s:corefromsamples}
In this section, we consider the problem of finding cost allocations from samples that satisfy relaxations of the core. A natural approach to this problem is to first learn the underlying model, $C$, from the data and to then compute a cost allocation for the learned function. As shown in \cite{BPZ-15}, this approach works if $C$ is PAC-learnable, but there exist functions $C$ that are not PAC-learnable and for which a cost allocation that approximately satisfies the core can still be computed. The main result of this section shows that a cost allocation that approximates the core property can be computed for \emph{any} function with a non-empty core. Moreover, we show that the number of samples from $\mathcal{D}$ needed to accurately learn the core is low.
The approach is to directly compute a cost allocation, which empirically satisfies the core property, i.e., it satisfies the core property on all of the samples. We then argue that the same cost shares will likely satisfy the core property on newly drawn samples as well. This difference between the empirical performance of a function and its expected performance is known as the generalization error of a function and its analysis is central to theoretical machine learning. Intuitively, the generalization error is small when the number of samples, $m$, used to train the function is large, and the function itself is relatively simple. Two of the most common tools that formally capture these notions are the VC-dimension and the $m$-sample Rademacher complexity of a function class. We will use both of these to highlight different trade-offs in computing statistical cost shares.
We begin by defining three notions of approximate core: the probably stable \cite{BPZ-15}, approximately stable, and probably approximately stable cores.
\begin{definition} Given $\delta, \epsilon > 0$, a cost allocation $\mathbf{\psi}$ such that $\sum_{i \in N} \psi_i = C(N)$ is in
\begin{itemize}
\item the \textbf{probably stable} core \cite{BPZ-15} if, for all $\mathcal{D}$, $$\Pr_{S \sim \mathcal{D}} \left[ \sum_{i \in S} \psi_i \leq C(S) \right] \geq 1 - \delta,$$
\item the \textbf{approximately stable} core over $\mathcal{D}$ if for all $S \subseteq N$, $$(1- \epsilon) \cdot \sum_{i \in S} \psi_i \leq C(S), $$
\item the \textbf{probably approximately stable} core if, for all $\mathcal{D}$, $$\Pr_{S \sim \mathcal{D}} \left[ (1- \epsilon)\cdot \sum_{i \in S} \psi_i \leq C(S) \right] \geq 1 - \delta.$$
\end{itemize}
\end{definition}
The algorithms we consider compute cost shares in polynomial time. The hardness results are information theoretic and are not due to running time limitations.
\begin{definition}
Cost shares $\mathbf{\psi}$ are \emph{computable} for the class of functions $\mathcal{C}$ over distribution $\mathcal{D}$, if for all $C \in \mathcal{C}$ and any $\Delta, \delta, \epsilon > 0$, given $C(N)$ and $m = \poly(n, \nicefrac{1}{\Delta}, \nicefrac{1}{\delta}, \nicefrac{1}{\epsilon})$ samples $(S_j, C(S_j))$ with each $S_j$ drawn i.i.d. from distribution $\mathcal{D}$, there exists an algorithm that computes $\mathbf{\psi}$ with probability at least $1-\Delta$ over both the samples and the choices of the algorithm. If the algorithm has $\poly(m)$ running time, then the cost shares $\mathbf{\psi}$ are \emph{efficiently computable}.
\end{definition}
Finally, we will refer to the number of samples $m$ required to compute approximate cores as the {\em sample complexity} of the algorithm.
\paragraph{Our Results.} We give algorithms that efficiently compute cost shares in the probably stable core for functions with a non-empty (traditional) core with a simple approach using the VC-dimension (Section~\ref{s:probstable}), the algorithm has sample complexity linear in $n$. With a more complex analysis and using the Rademacher complexity, we obtain efficiently computable cost shares in the probably approximately stable core with an improved sample complexity dependence of $\log n$ but with an additional dependence on the spread of the function $C$ (Section~\ref{s:probapxstable}). Finally, we show that cost shares in the approximately stable core are not computable even for the uniform distribution and the well-behaved class of monotone submodular functions (Section~\ref{s:apxstable}).
\subsection{Cost shares in the probably stable core are efficiently computable}
\label{s:probstable}
\citet{BPZ-15} showed that several families of functions $\mathcal{D}$ have a core that is probably stable. These families include network flow, threshold task, and induced subgraph games which are all well-known classes of cooperative games; and the class of monotone simple games which are games that take values in $\{0,1\}$. We generalize their result so that it is not constrained to specific classes of functions, and show how to compute a probably stable core for any game with a non-empty core.
Technically, we use the VC-dimension of the class of halfspaces to show that the performance on the samples generalizes well to the performance on the distribution $\mathcal{D}$. We review the definition of the VC-dimension in Appendix~\ref{s:appcore} and only state VC-dimension results needed for our purposes. We first state the generalization error obtained for a class of functions with VC-dimension $d$.
\begin{theorem}[\cite{ML}, Theorem 6.8]
\label{t:vcmain}
Let $\mathcal{H}$ be a hypothesis class of functions from a domain $\mathcal{X}$ to $\{-1, 1\}$ and $f : \mathcal{X} \mapsto \{-1, 1\}$ be some ``correct" function. Assume that $\mathcal{H}$ has VC-dimension $d$. Then, there is an absolute constant $c$ such that with $m \geq c (d + \log(1/ \Delta))/\delta^2$ i.i.d. samples $\mathbf{x}^1, \ldots, \mathbf{x}^m \sim \mathcal{D}$,
$$\left|\Pr_{\mathbf{x} \sim \mathcal{D}}\left[h(\mathbf{x}) \neq f(\mathbf{x})\right] - \frac{1}{m}\sum_{i=1}^m \mathds{1}_{h(\mathbf{x}^i) \neq f(\mathbf{x}^i)}\right| \leq \delta$$
for all $h \in \mathcal{H}$, with probability $1- \Delta$ over the samples.
\end{theorem}
We use a special case of the class of halfspaces, for which we know the VC-dimension.
\begin{theorem}[\cite{ML}, Theorem 9.2]
\label{t:vchalfspace}
The class of functions $\{\mathbf{x} \mapsto \sign(\mathbf{w}^\intercal \mathbf{x}) : \mathbf{w} \in \mathbb{R}^n\}$ has VC-dimension $n$.
\end{theorem}
We first define a class of functions that contains the core, and prove that it has low VC-dimension. Given a sample $S$, define $\mathbf{x}^S$ such that $x^S_i = \mathds{1}_{i \in S}$ for $i \in [n]$ and $x^S_{n+1} = C(S)$. Note that if the core property is satisfied for sample $S$, then $\sign\left(\sum_{i = 1}^n \psi_i x^S_i - x^S_{n+1}\right) = -1$ . We now bound the VC-dimension of this hypothesis class of functions induced by cost allocations $\psi$.
\begin{corollary}
The class of functions $\mathcal{H}^{core} = \{\mathbf{x} \mapsto sign(\sum_{i = 1}^n \psi_i x_i - x_{n+1}) \ : \ \psi \in \mathbb{R}^n, \sum_i \psi_i = C(N)\}$ has VC-dimension at most $n+1$.
\end{corollary}
\begin{proof}
We combine the observation that $\{\mathbf{x} \mapsto \sign(\sum_{i = 1}^n w_i x_i - x_{n+1}) \ : \ \mathbf{w} \in \mathbb{R}^n, \sum_i w_i = C(N)\} \subseteq \{\mathbf{x} \mapsto \sign(\mathbf{w}^\intercal \mathbf{x}) \ : \ \mathbf{w} \in \mathbb{R}^{n+1}\}$ with the well-known fact that the VC-dimension of $\mathcal{H'}$ is at most the VC-dimension of $\mathcal{H}$ for $\mathcal{H'} \subseteq \mathcal{H}$.
\end{proof}
It remains to show how to optimize over functions in this class.
\begin{theorem}
\label{t:vccore}
The class of functions with a non-empty core has cost shares in the probably stable core that are efficiently computable.
The sample complexity is $$O\left(\frac{n + \log(1/ \Delta)}{\delta^2}\right).$$
\end{theorem}
\begin{proof}
Let $\psi$ be a cost allocation which satisfies both the core property on all the samples and the balance property, i.e., $\sum_{i \in S} \psi_i \leq C(S)$ for all samples $S$ and $\sum_{i \in N} \psi_i = C(N)$. Note that such a cost allocation exists since we assume that $C$ has a non-empty core. Given the set of samples, it can be computed with a simple linear program. We argue that $\psi$ is probably stable.
Define $h(x) = \sign\left(\sum_{i = 1}^n \psi_i x^S_i - x^S_{n+1}\right)$ and $f(\mathbf{x}) = - 1$ for all $\mathbf{x}$. Since the core property is satisfied on all the samples, $\frac{1}{m}\sum_{i=1}^m \mathds{1}_{h(x) \neq f(x)} = 0$. Thus, by Theorem~\ref{t:vcmain},
\begin{align*}
\Pr_{S \sim \mathcal{D}}\left[ \sum_{i \in S} \psi_i \leq C(S) \right] & = 1 - \Pr_{\mathbf{x}^S : S \sim \mathcal{D}}\left[\sign\left(\sum_{i = 1}^n \psi_i x^S_i - x^S_{n+1}\right) \neq -1\right] \\
& = 1 - \Pr_{\mathbf{x}^S : S \sim \mathcal{D}}\left[h\left(\mathbf{x}^S\right) \neq f\left(\mathbf{x}^S\right)\right] \\
& = 1 - \left|\Pr_{\mathbf{x}^S : S \sim \mathcal{D}}\left[h\left(\mathbf{x}^S\right) \neq f\left(\mathbf{x}^S\right)\right] - \frac{1}{m}\sum_{i=1}^m \mathds{1}_{h\left(\mathbf{x}^S\right) \neq f\left(\mathbf{x}^S\right)}\right| \\
& \geq 1 - \delta
\end{align*}
with $O((n + \log(1/ \Delta))/\delta^2)$ samples.
\end{proof}
\subsection{Logarithmic sample complexity for probably approximately stable cores
\label{s:probapxstable}
We show that the sample complexity dependence on $n$ can be improved from linear to logarithmic. However, this improvement comes at a cost. We now find a probably approximately stable core instead of probably stable core, and the sample complexity depends on the {\em spread} of the function $C$, defined as $\frac{\max_S C(S)}{\min_{S \not= \emptyset} C(S)}$. This approach assumes that $\min_{S \not= \emptyset} C(S) > 0$.
We start with an overview.
\begin{enumerate}
\item As previously, we find a cost allocation which satisfies the core property on all samples. However, we restrict this search to cost allocations with bounded $\ell_1$-norm. Such a cost allocation can be found efficiently since the space of such cost allocations is convex.
\item The analysis begins by bounding the $\ell_1$-norm of any vector in the core (Lemma~\ref{l:boundedcore}). Combined with the assumption that the core is non-empty, this implies that a cost allocation $\psi$ satisfying the previous conditions exists.
\item Let $[x]_+$ denote the function $x \mapsto \max(x, 0)$. Consider the following ``loss" function:
$$ \left[\frac{\sum_{i \in S} \psi_i}{C(S)} - 1\right]_+$$
This loss function is convenient since it is equal to $0$ if and only if the core property is satisfied for $S$ and it is $1$-Lipschitz, which is used in the next step.
\item Next, we bound the difference between the empirical loss and the expected loss for all $\psi$ with a known result using the Rademacher complexity of linear predictors with low $\ell_1$ norm over $\rho$-Lipschitz loss functions (Theorem~\ref{thm:rad_main}).
\item Finally, given $\psi$ which approximately satisfies the core property in expectation, we show that $\psi$ is in the probably approximately stable core by Markov's inequality (Lemma~\ref{lem:markov}).
\end{enumerate}
We review the definition of the Rademacher complexity in Appendix~\ref{s:appcore}. For our purposes, the following result which follows from the Rademacher complexity of linear classes is sufficient.
\begin{theorem}[\cite{ML}, Theorem 26.15]
\label{thm:rad_main}
Suppose that $\mathcal{D}$ is a distribution over $\mathcal{X} \times \mathbb{R}$ such that with probability $1$ we have that $\|\mathbf{x}\|_{\infty} \leq R.$ Let $\mathcal{H} = \{\mathbf{w} \in \mathbb{R}^d : \|\mathbf{w}\|_1 \leq B\}$ and let $\ell : \mathcal{H} \times (\mathcal{X} \times \mathbb{R}) \rightarrow \mathbb{R}$ be a loss function of the form
$$\ell(\mathbf{w}, (\mathbf{x}, y)) = \phi(\mathbf{w}^{\intercal} \mathbf{x}, y)$$
such that for all $y \in \mathbb{R}$, $a \mapsto \phi(a, y)$ is an $\rho$-Lipschitz function and such that $\max_{a \in [-BR, BR]}|\phi(a,y)| \leq c$. Then, for all $\mathbf{w} \in \mathcal{H}$ and any $\Delta \in (0,1)$, with probability of at least $1 - \Delta$ over $m$ i.i.d. samples $(\mathbf{x}_1, y_1), \ldots, (\mathbf{x}_m, y_m)$ from $\mathcal{D}$,
$$\EU[(\mathbf{x}, y) \sim \mathcal{D}][\ell(\mathbf{w}, (\mathbf{x}, y))] \leq \frac1m \sum_{i=1}^m \ell(\mathbf{w}, (\mathbf{x}_i, y_i))+ 2\rho B R \sqrt{\frac{2 \log(2d)}{m} } + c \sqrt{\frac{2 \log(2/\Delta)}{m}}.$$
\end{theorem}
We first bound the $\ell_1$ norm of vectors in the core to bound the space of linear functions that we search over.
\begin{lemma}
\label{l:boundedcore}
Assume that $\psi$ is a vector in the core, then $\|\psi\|_1 \leq 2 \max_S |C(S)|$.
\end{lemma}
\begin{proof}
Fix some vector $\psi$ in the core. Let $A$ be the set of elements $i$ such that $\psi_i \geq 0$ and $B$ be the remaining elements. First note that $$\sum_{i \in A} \psi_i \leq C(A) \leq \max_S |C(S)|$$ where the first inequality is by the core property. Next, note that $$0 \leq C(N) = \sum_{i \in A} \psi_i + \sum_{i \in B} \psi_i \leq \max_S |C(S)| + \sum_{i \in B} \psi_i$$ where the equality is by the balance property, so $\sum_{i \in B} \psi_i \geq -\max_S |C(S)|.$ Thus, $$\|\psi\|_1 = \sum_{i \in A} \psi_i - \sum_{i \in B} \psi_i \leq \max_S |C(S)| + \max_S |C(S)|.$$
\end{proof}
We can thus focus on bounded cost allocations $\psi \in \mathcal{H}$ where
$$\mathcal{H} := \left\{\psi \ : \ \psi \in \mathbb{R}^n, \|\psi\|_1 \leq 2 \max_S |C(S)|\right\}.$$
The next lemma shows that if the core property approximately holds in expectation, then it is likely to approximately hold.
\begin{lemma} \label{lem:markov} For any $0 < \epsilon, \delta < 1$ and cost allocation $\psi$, $$\EU[S \sim \mathcal{D}]\left[\left[\frac{\sum_{i \in S} \psi_i}{C(S)} - 1\right]_+\right] \le \frac{\epsilon\delta}{1 - \epsilon} \Rightarrow \Pr_{S \sim \mathcal{D}}\left[(1 - \epsilon) \sum_{i \in S} \psi_i \le C(S)\right] \ge 1 - \delta.$$\end{lemma}
\begin{proof} For any $a > 0$ and nonnegative random variable $X$, by Markov's inequality we have $\Pr[X \le a] \ge 1 - E[X]/ a$. By letting $a = \epsilon /(1 - \epsilon)$, $X = \left[(\sum_{i \in S} \psi_i)/ C(S) - 1\right]_+$, and observing that
$$
\left[\frac{\sum_{i \in S} \psi_i}{C(S)} - 1\right]_+ \le \frac{\epsilon}{1 - \epsilon} \Rightarrow \frac{\sum_{i \in S} \psi_i}{C(S)} - 1 \le \frac{\epsilon}{1 - \epsilon} \Rightarrow (1 - \epsilon) \sum_{i \in S} \psi_i \le C(S),
$$
we obtain $\Pr_{S \sim \mathcal{D}}\left[(1 - \epsilon) \sum_{i \in S} \psi_i \le C(S)\right] \ge 1 - \delta$.
\end{proof}
Combining Theorem~\ref{thm:rad_main}, Lemma~\ref{l:boundedcore}, and Lemma~\ref{lem:markov}, the dependence of the sample complexity on $n$ is improved.
\begin{theorem} \label{thm:approx_core_learn} The class of functions with a non-empty core has cost shares in the probably approximately stable core that are efficiently computable with sample complexity
$$
\left(\frac{1 - \epsilon}{\epsilon\delta}\right)^2 \left(128 \tau(C)^2 \log(2 n) + 8 \tau(C)^2 \log (2/\Delta) \right) = O\left(\left(\frac{\tau(C)}{\epsilon \delta}\right)^2 \left(\log n +\log(1/\Delta)\right)\right).
$$
where $\tau(C) = \frac{ \max_S C(S)}{\min_{S \neq \emptyset} C(S)}$ is the spread of $C$.
\end{theorem}
\begin{proof}Fix $C \in \mathcal{C}$. Suppose we are given $m$ samples from $\mathcal{D}$.
We pick $\psi^{\star} \in \mathcal{H}$ such that core property holds on all the samples and such that the balance property holds ($\sum_{i \in N} \psi_i = C(N)$). This cost allocation $\psi^{\star}$ can be found efficiently since the collection of such $\psi$ is a convex set. By the assumption that $C$ has at least one vector in the core and by Lemma~\ref{l:boundedcore}, such a $\psi^\star$ exists. Given $S \sim \mathcal{D}$, define $\mathbf{x}^{S}$ such that $x^S_i = \mathds{1}_{i \in S} / C(S)$. Fix $y = 1$. Define the loss function $\ell$ as follows,
$$\ell\left(\psi, \left(\mathbf{x}^S, y\right)\right) := \left[\psi^{\intercal} \mathbf{x}^S - y\right]_+ = \left[\frac{\sum_{i \in S} \psi_i}{C(S)} - 1\right]_+$$
We wish to use Theorem~\ref{thm:rad_main} with $R = 1 / \min_{S \neq \emptyset} |C(S)|$, $B = 2 \max_S |C(S)|$, $\phi(a,y) = \left[a - y\right]_+$, $\rho = 1$, and $c = \tau(C)$. We verify that all the conditions hold. First note that without loss of generality, samples where $S = \emptyset$ can be ignored, so $\norm{\mathbf{x}^S}_\infty \le 1 / \min_{S \neq \emptyset} |C(S)|$ for all $S$. Next, $\|\psi\|_1 \leq 2 \max_S |C(S)|$ for $\psi \in \mathcal{H}$ by definition of $\mathcal{H}$. The loss function $\ell$ is of the form $\ell(\psi, (\mathbf{x}, y)) = \phi(\psi^{\intercal} \mathbf{x}, y) = \left[\psi^{\intercal} \mathbf{x} - y\right]_+$ such that
$a \mapsto \phi(a,y) = \left[ a - y\right]_+ $ is an $1$-Lipschitz function and such that $\max_{a \in [-BR, BR]} |\phi(a,y) | \leq 2 \max_S |C(S)|/ \min_{S \neq \emptyset} |C(S)| = 2 \tau(C)$. In addition, note that
$$\frac1m \sum_{i=1}^m \ell\left(\psi^{\star}, \left(\mathbf{x}^{S_i}, 1\right)\right) = 0$$
since the core property holds on all the samples. Thus, by Theorem~\ref{thm:rad_main},
$$\EU[S \sim \mathcal{D}]\left[\frac{\sum_{i \in S} \psi^{\star}_i}{C(S)} - 1\right]_+ = \EU[\mathbf{x}^S : S \sim \mathcal{D}]\left[\ell\left(\psi^{\star}, \left(\mathbf{x}^S, 1\right)\right)\right] \leq 4 \tau(C) \sqrt{\frac{2 \log(2n)}{m} } + \tau(C) \sqrt{\frac{2 \log(2/\Delta)}{m}}.$$
Choose any $\epsilon, \delta > 0$. If the number of samples $m$ is chosen as in the statement of the theorem, then the righthand side of the above inequality will be less than $\frac{\epsilon\delta}{1 - \epsilon}$. Thus by Lemma \ref{lem:markov},
$$
\Pr_{S \sim \mathcal{D}}\left[(1 - \epsilon) \sum_{i \in S} \psi_i \le C(S)\right] \ge 1 - \delta,
$$
which completes the proof.\end{proof}
\subsection{The approximately stable core is not computable}
\label{s:apxstable}
Since we obtained a probably approximately stable core, a natural question is if it is possible to compute cost allocations that are approximately stable over natural distributions. The answer is negative in general: even for the restricted class of monotone submodular functions, which always have a solution in the core, the core cannot be approximated from samples, even over the uniform distribution.
\begin{restatable}{rThm}{tcore}
\label{t:core}
Cost shares $\psi$ in the $(1/2 + \epsilon)$-approximately stable core, i.e., such that for all $S$,
$$\left(\frac{1}{2} + \epsilon\right) \cdot \sum_{i \in S} \psi_i \leq C(S), $$
cannot be computed for monotone submodular functions over the uniform distribution, for any constant $\epsilon > 0$.
\end{restatable}
\begin{proof}
The ground set of elements is partitioned in $\epsilon^{-1}$ sets $A_1, \ldots A_{\epsilon^{-1}}$ of size $\epsilon n$ for some small constant $\epsilon > 0$. Let $\mathcal{C} = \{C^{A_i} \ : \ i \in [\epsilon^{-1}]\}$ where
$$C^A(S) = |(N \setminus A) \cap S| + \min \left(|A \cap S|, (1 + \epsilon) \frac{\epsilon n}{2}\right).$$
The expected number of elements of $A_i$ in a sample $S$ from the uniform distribution is $|A_i| / 2 = \epsilon n /2$, so by the Chernoff bound
$$\Pr \left[|A_i \cap S| \geq (1 + \epsilon)\frac{n\epsilon}{2}\right] \leq e^{-\frac{\epsilon^2 n }{6}},$$
Thus, by a union bound, $C^{A_i}(S) = |S|$ over all $i$ and all samples $S$ with probability $1 - O(e^{- n})$ and we henceforth assume this is the case. It is therefore impossible to learn any information about the partition $A_1, \ldots A_{\epsilon^{-1}}$ from samples. Any cost allocation $\psi$ computed by an algorithm given samples from $C^{A_i}$ is thus \emph{independent} of $i$.
Next, consider such a cost allocation $\mathbf{\psi}$ independent of $i$ satisfying the balance property. There exists $A_i$ such that $\sum_{j \in A_i} \psi_j > (1 - \epsilon) n \epsilon$ since $\sum_{j \in N} \psi_j = C(N) > (1 - \epsilon)n$ by the balance property. In addition, $C^{A_i}(A_i) = (1 + \epsilon)\epsilon n / 2$. We obtain
$$\sum_{j \in A_i} \psi_j > (1 - \epsilon) n \epsilon = \frac{(1 - \epsilon) }{(1 + \epsilon)} 2 C^{A_i}(A_i). $$
Thus, the core property is violated by a $1/2 + \epsilon'$ factor for set $A_i$ and function $C^{A_i}$, and for any constant $\epsilon' > 0$ by picking $\epsilon$ sufficiently small.
\end{proof}
\section{Approximating the Shapley Value from Samples}
\label{s:shapleyFunction}
We turn our attention to the \textsc{Statistical Cost Sharing} \ problem in the context of the Shapley value. Since the properties (axioms) of the Shapley value are over elements and not sets, there is no simple relaxation of the Shapley value where the properties hold ``probably" over $\mathcal{D}$ as we had for the core. However, since the Shapley value exists and is unique for all functions, a natural relaxation is to simply approximate this value from samples.
We begin by observing that there exists a distribution such that Shapley value can be approximated arbitrarily well from samples. However, in this paper, we are motivated by applications where the algorithm does not control the distribution over the samples, but where the samples are drawn from some ``natural" distribution. Thus, the distributions we consider in this section are the uniform distribution, and more generally product distributions, which are the standard distributions studied in the learning literature for combinatorial functions \cite{balcan2011learning,balcan2012learning,feldman2014learning,
feldman2014optimal}. It is easy to see that we need some restrictions on the distribution $\mathcal{D}$ (for example, if the empty set if drawn with probability one, the Shapley value cannot be approximated).
In the case of submodular functions with bounded curvature, we prove a tight approximation bound in terms of the curvature when samples are drawn from bounded product distributions. However, we show that the Shapley value cannot be approximated from samples even for coverage functions (which are a special case of submodular functions) and the uniform distribution. Since coverage functions are learnable from samples, this implies the counter-intuitive observation that learnability does not imply that the Shapley value is approximable from samples. We begin by formally defining $\alpha$-approximability of the Shapley value in the statistical setting.
\begin{definition} An algorithm $\alpha$-approximates, $\alpha \in (0,1]$, the Shapley value of a family of cost functions $\mathcal{C}$ over distribution $\mathcal{D}$, if, for all $C \in \mathcal{C}$ and all $\delta > 0$, given $\poly(n, \nicefrac{1}{\delta}, \nicefrac{1}{1-\alpha})$ samples from $\mathcal{D}$, it computes Shapley value estimates $\tilde{\phi}_C$ such that for
\begin{itemize}
\item positive bounded Shapley value, if $\phi_i \geq 1 / \poly(n)$, then $\alpha \phi_i \leq \tilde{\phi}_i \leq \frac{1}{\alpha}\phi_i$;
\item negative bounded Shapley value, if $\phi_i \leq - 1 / \poly(n)$, then $\frac{1}{\alpha} \phi_i \leq \tilde{\phi}_i \leq \alpha\phi_i$;
\item small Shapley value, if $ |\phi_i| < 1 / \poly(n) $, then $|\phi_i - \tilde{\phi}_i| = o(1) $ .
\end{itemize}
for all $i \in N$ with probability at least $1 - \delta$ over both the samples and the choices made by the algorithm.
\end{definition}
\paragraph{Controlling $\mathcal{D}$.} We begin by noting that there exists a distribution $\mathcal{D}$ such that the Shapley value can be approximated arbitrarily well for bounded functions. Other sampling methods have previously been suggested (\cite{bachrach2010approximating,fatima2008linear,liben2012computing,
mann1960values}), but the samples $(S, C(S))$ used in these methods are not i.i.d. and the value query model is assumed.
\begin{definition}
The \emph{Shapley distribution} $\mathcal{D}^{sh}$ is the distribution which first picks a size $j \in \{0, \ldots, n\}$ uniformly at random and then draws a uniformly random set of size $j$.
\end{definition}
Let $\mathcal{S}^j_{i}$ and $\mathcal{S}^j_{-i}$ be the collections of samples of size $j$ containing element $i$ and not containing it respectively. Define $\texttt{avg}(\mathcal{S}) := (\sum_{S \in \mathcal{S}} C(S)) / |\mathcal{S}| $ to be the average value of the samples in $\mathcal{S}$. Consider the following cost allocation:
$$\tilde{\phi}_i = \sum_{j=1}^n \texttt{avg}(\mathcal{S}^j_{i}) - \sum_{j=0}^{n-1} \texttt{avg}(\mathcal{S}^j_{-i})$$
When the distribution is the Shapley distribution $\mathcal{D}^{sh}$, the expected value of this cost allocation is the Shapley value and concentration bounds kick in.
\begin{restatable}{rPro}{pcontrolledD}
\label{p:controlledD} The Shapley value is $(1 - \epsilon)$-approximable, for any constant $\epsilon > 0$, over the Shapley distribution $\mathcal{D}^{sh}$.
\end{restatable}
\begin{proof}
Recall the definition of the Shapley value $\phi$ and observe that
\begin{align*}
\phi_i & = \EU[\sigma][C(S_{\sigma < i} \cup\{i\}) - C(S_{\sigma < i})]\\
& = \EU[S:|S| \sim \mathcal{U}(\{1,\ldots,n\}), i \in S][C(S)] - \EU[S:|S| \sim \mathcal{U}(\{0,\ldots,n-1\}), i \not \in S][C(S)] \\
& = \sum_{j=1}^n \E[\texttt{avg}(\mathcal{S}^j_{i})] - \sum_{j=0}^{n-1} \E[\texttt{avg}(\mathcal{S}^j_{-i})] \\
& = \E[\tilde{\phi}_i]
\end{align*}
Next, observe that by standard concentration bounds and a sufficiently large polynomial number of samples drawn from the Shapley distribution $\mathcal{D}^{sh}$, $m/\poly(n)$ samples are in $\mathcal{S}^j_{i}$ and $\mathcal{S}^j_{-i}$ for all $j$ and $i$. Then, by Hoeffding's inequality,
$$\Pr\left[\left|\texttt{avg}(\mathcal{S}^j_{i}) - \E[\texttt{avg}(\mathcal{S}^j_{i})]\right| \geq \frac{\epsilon |\phi_i| }{ 2n}\right] \leq 2e^{-\frac{m(\epsilon |\phi_i|)^2}{\poly(n)}}$$
By a union bound, we have
$$\Pr\left[\tilde{\phi}_i - \left(\left|\sum_{j=1}^n \E[\texttt{avg}(\mathcal{S}^j_{i})] - \sum_{j=0}^{n-1} \E[\texttt{avg}(\mathcal{S}^j_{-i})] \right) \right| \geq \epsilon|\phi_i|\right] \leq 2e^{-\frac{m (\epsilon|\phi_i|)^2}{\poly(n)}}.$$
We get that either $(1 - \epsilon) \phi_i \leq \tilde{\phi}_i \leq (1+\epsilon) \phi_i$ if $\phi_i > 0$ or $(1 + \epsilon) \phi_i \leq \tilde{\phi}_i \leq (1-\epsilon) \phi_i$ if $\phi_i < 0$ with probability $1 - 2e^{-\frac{m \epsilon^2}{\poly(n)}}$, if $|\phi_i| \geq 1/ \poly(n)$.
If $|\phi_i| = o(1/\poly(n))$, we bound the first inequality by $\epsilon/ 2n$ instead of $\epsilon |\phi_i| / 2n$ and obtain
$|\phi_i - \tilde{\phi}_i | = o(1) $ with probability $1 - 2e^{-\frac{m\epsilon^2}{\poly(n)}}$.
\end{proof}
A known method to estimate an expected value according to some distribution while given samples from another distribution is called \emph{importance sampling}. Importance sampling reweighs samples according to their probabilities of being sampled. Although the above method achieves accurate estimates with a sufficiently large number of samples, the number of samples required may be exponential, see Theorem~\ref{thm:lower}.
\subsection{Submodular functions with bounded curvature}
We consider submodular functions with bounded curvature, a common assumption in the submodular maximization literature \cite{iyer2013submodular,iyer2013curvature,sviridenko2015optimal,vondrak2010submodularity}. We show that the Shapley value of these functions is approximable from samples, for which we derive a tight bound. Intuitively, the curvature of a submodular function bounds by how much the marginal contribution of an element can decrease. This property is useful since the Shapley value of an element can be written as a weighted sum of its marginal contributions over all sets.
\begin{definition} A monotone submodular function $C$ has curvature $\kappa \in [0,1]$ if $C_{N \setminus \{i\}}(i) \geq (1-\kappa) C(i)$ for all $i \in N$. This curvature is bounded if $\kappa < 1$.
\end{definition}
An immediate consequence of this definition is that $C_S(i) \geq (1-\kappa) C_T(i)$ for all $S,T$ such that $i \not \in S \cup T$ by monotonicity and submodularity.
The main idea for the approximation is that the expected marginal contribution of an element $i$ to a random set approximates the Shapley value of $i$ by the curvature property. We use the same tool $\tilde{v}_i$ to estimate expected marginal contributions $v_i = \E_{S \sim \mathcal{D}| i \not \in S}[C_S(i)]$ as for additive functions. Recall that $\tilde{v}_i = \texttt{avg}(\mathcal{S}_i) - \texttt{avg}(\mathcal{S}_{-i})$ is the difference between the average value of samples containing $i$ and the average value of samples not containing $i$.
\begin{restatable}{rThm}{tcurv}
\label{t:curv}
Monotone submodular functions with bounded curvature $\kappa$ have Shapley value that is $\sqrt{1-\kappa}-\epsilon$ approximable from samples over the uniform distribution and $1 - \kappa-\epsilon$ approximable over any bounded product distribution for any constant $\epsilon >0$.
\end{restatable}
First, the Shapley value of monotone functions is non-negative since marginal contributions are non-negative by monotonicity. Next, if a Shapley value $\phi_i$ is ``small" ($o(1/\poly(n))$), then $C_S(i)$ is small for all $S$ by the curvature property, implying that $v_i = \E_{S \sim \mathcal{D}| i \not \in S}[C_S(i)]$ is small as well. By Lemma~\ref{l:concentration}, a generalization of Lemma~\ref{l:addconcentration} showing that $\tilde{v}_i$ is a good estimate of $v_i$, and with $\epsilon = 1/n$, $|v_i - \tilde{v}_i| = o(1)$. With $\tilde{\phi}_i = \tilde{v}_i$, we then obtain $|\tilde{\phi}_i - \phi_i| = o(1)$.
The interesting case is positive bounded Shapley value $\phi_i$, which we assume for the rest of the analysis. We first show a $1 - \kappa$ approximation for product distributions, which is a straightforward application of the curvature property combined with Lemma~\ref{l:concentration}. Consider the algorithm which computes $\tilde{\phi}_i = \tilde{v}_i$. Note that
$$\phi_i =\EU[\sigma][C_{A_{\sigma<i}}(i)] \geq (1-\kappa) v_i > \frac{1-\kappa}{1 +\epsilon} \tilde{v}_i > (1- \kappa-\epsilon)\tilde{v}_i$$
where the first inequality is by curvature and the second by Lemma~\ref{l:concentration}. Similarly, for the other direction, $\phi_i\leq v_i / (1-\kappa) < \tilde{v}_i/(1-\kappa -\epsilon)$. The $\sqrt{1-\kappa}$ result is the main technical component of this proof. We begin with a technical overview.
\begin{enumerate}
\item Denote the uniform distribution over all sets of size $j$ by $\mathcal{U}_{j}$. Lemma~\ref{l:decreasing} shows that the expected marginal contribution $\E_{S \sim \mathcal{U}_{j}| i \not \in S}[C_S(i)]$ of $i$ to a uniformly random set $S$ of size $j$ is decreasing in $j$, which is by submodularity.
\item Consider $L := (1-\epsilon)n/2$, $H := (1+\epsilon)n/2$. Lemma~\ref{l:vi} shows that $(1+ \epsilon) \cdot \E_{S \sim \mathcal{U}_{L}| i \not \in S}[C_S(i)] \geq v_i$ and $(1- \epsilon)\cdot \E_{S \sim \mathcal{U}_{H}| i \not \in S}[C_S(i)]\leq v_i$, which is because a uniformly random set $S$ is likely to have size between $L$ and $H$ and by submodularity.
\item Combining these two lemmas, roughly \emph{half} of the terms (when $j \leq L$) in the summation $\phi_i = ( \sum_{j=0}^{n-1} \E_{S \sim \mathcal{U}_{j}| i \not \in S}[C_S(i)])/n$ are greater than $v_i$ and the other half (when $j \geq H$) of the terms are smaller. This is the main observation for the improvement from $1 - \kappa$.
\item The above and curvature imply that $(1/2 + (1 - \kappa)/2) v_i \leq \phi_i \leq (1/2 + 1/(2(1-\kappa))v_i $.
\item By scaling $v_i$ to obtain the best approximation possible with the previous inequality, we obtain a $\sqrt{1-\kappa}$ approximation.
\end{enumerate}
Let $$\tilde{\phi}_i = \frac{2-\kappa}{2\sqrt{1-\kappa}} \cdot \tilde{v}_i$$ be the estimated Shapley value. Denote by $\mathcal{U}_j$ the uniform distribution over all sets of size $j$, so $$\phi_i = \EU[\sigma][C_{A_{\sigma < i}}(i)] = \frac{1}{n}\sum_{j = 0}^{n-1}\EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] .$$ The main idea to improve the loss from $1-\kappa$ to $\sqrt{1 - \kappa}$ is to observe that $v_i$ can be a factor $1-\kappa$ away from the contribution $\E_{S \sim \mathcal{U}_{j_l}| i \not \in S}[C_S(i)]$ of $j$ to sets of low sizes $j_l \leq L := (1-\epsilon') \cdot n / 2$ or $1-\kappa$ away from its contribution $\E_{S \sim \mathcal{U}_{j_h}| i \not \in S}[C_S(i)]$ to sets of high sizes $j_h \geq H := (1+\epsilon') \cdot n / 2 $, but not both, otherwise the curvature property would be violated as illustrated in Figure~\ref{f:curv}. The following lemma shows that $\E_{S \sim \mathcal{U}_{j}| i \not \in S}[C_S(i)]$ is decreasing in $j$ by submodularity.
\begin{figure}
\centering
\includegraphics[width = 0.6\linewidth]{general}
\caption{The expected marginal contribution $E_{S \sim \mathcal{U}_{j}| i \not \in S}[C_S(i)]$ of an element $i$ to a set of size $j$ as a function of $j$. The curvature property implies that any two points are at most a factor $1-\kappa$ from each other. Lemma~\ref{l:decreasing} shows that it is decreasing. Lemma~\ref{l:vi} shows that the expected marginal contribution $v_i$ of $i$ to a uniformly random set is approximately between $E_{S \sim \mathcal{U}_{L}| i \not \in S}[C_S(i)]$ and $E_{S \sim \mathcal{U}_{H}| i \not \in S}[C_S(i)]$. The Shapley value of $i$ is the average value of this expected marginal contribution over all integers $j \in \{0, \ldots, n-1\}$.}
\label{f:curv}
\end{figure}
\begin{lemma}
\label{l:decreasing}
Let $C$ be a submodular function, then for all $j \in \{0, \ldots, n-1\}$ and all $i \in N$,
$$\EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] \geq \EU[S \sim \mathcal{U}_{j+1}| i\not \in S][C_S(i)] $$
\end{lemma}
\begin{proof}
By submodularity,
$$\sum_{S : |S| = j, i \not \in S} C_{S}(i) \geq \sum_{S : |S| = j, i \not \in S} \frac{1}{n -j -1}\sum_{i' \not \in S \cup \{i\}} C_{S \cup \{i'\}}(i). $$
In addition, observe that by counting in two ways,
$$\sum_{S : |S| = j, i \not \in S} \sum_{i' \not \in S \cup \{i\}} C_{S \cup \{i'\}}(i)= (j+1) \sum_{S: |S| = j+1, i \not \in S}C_{S}(i).$$
By combining the two previous observations,
\begin{align*}
\EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] & = \frac{1}{|\{S : |S| = j, i \not \in S\} |}\sum_{S : |S| = j, i \not \in S} C_{S}(i)\\
&\geq \frac{1}{\binom{n-1}{j}} \frac{j+1}{n-j-1}\sum_{S: |S| = j+1, i \not \in S}C_{S}(i) \\
&= \frac{1}{|\{S : |S| = j+1, i \not \in S\} |}\sum_{S: |S| = j+1, i \not \in S}C_{S}(i) \\
&= \EU[S \sim \mathcal{U}_{j+1}| i\not \in S][C_S(i)]
\end{align*}
\end{proof}
The next lemma shows that for $j$ slightly lower than $n/2$, the expected marginal contribution $v_i$ of element $i$ to a random set cannot be much larger than $\E_{S \sim \mathcal{U}_{j}| i \not \in S}[C_S(i)]$, and similarly for $j$ slightly larger than $n/2$, it cannot be much smaller.
\begin{lemma}
\label{l:vi}
Let $C$ be a submodular function, then for all $i \in N$, $$\left(1 + \frac{e^{-\frac{\epsilon n}{6}}}{1 - \kappa} \right) \cdot \EU[S \sim \mathcal{U}_{L}| i \not \in S][C_S(i)] \geq v_i \geq \left(1 - e^{-\frac{\epsilon n}{6}}\right) \cdot \EU[S \sim \mathcal{U}_{H}| i \not \in S][C_S(i)].$$
\end{lemma}
\begin{proof}
By Chernoff bound, $L \leq |S|$ and $|S| \leq H$ with probability at least $1-e^{-\frac{\epsilon n}{6}}$ each for $S$ drawn from the uniform distribution. Denote the uniform distribution over all sets by $\mathcal{U}$. So,
\begin{align*}
v_i & = \sum_{j=0}^{n-1} \Pr_{S \sim \mathcal{U}| i \not \in S}[|S| = j ] \cdot \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] &\\
& \geq \sum_{j=0}^{H} \Pr_{S \sim \mathcal{U}| i \not \in S}[|S| = j] \cdot \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] &\\
& \geq \Pr_{S \sim \mathcal{U}| i \not \in S}\left[|S| \leq H \right]\cdot \EU[S \sim \mathcal{U}_{H}| i \not \in S][C_S(i)] & \text{Lemma~\ref{l:decreasing}} \\
& \geq (1-e^{-\frac{\epsilon n}{6}}) \cdot \EU[S \sim \mathcal{U}_{H}| i \not \in S][C_S(i)] . &
\end{align*}
Similarly,
\begin{align*}
v_i & = \sum_{j=0}^{n-1} \Pr_{S \sim \mathcal{U}| i \not \in S}[|S| = j ] \cdot \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] & \\
& \leq \Pr_{S \sim \mathcal{U}| i \not \in S}\left[|S| < L \right] \cdot C(i) + \Pr_{S \sim \mathcal{U}|i \not \in S}\left[|S| \geq L \right] \cdot \EU[S \sim \mathcal{U}_{L}| i \not \in S][C_S(i)] & \text{Lemma~\ref{l:decreasing}} \\
& \leq e^{-\frac{\epsilon n}{6}} \cdot \frac{1}{1-\kappa} \cdot \EU[S \sim \mathcal{U}_{L}| i \not \in S][C_S(i)] + \EU[S \sim \mathcal{U}_{L}| i \not \in S][C_S(i)] & \text{curvature} \\
\end{align*}
\end{proof}
We are now ready to prove Theorem~\ref{t:curv}:
\begin{align*}
\phi_i & = \frac{1}{n}\sum_{j=0}^{n-1} \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] &\\
& = \frac{1}{n} \left(\sum_{j=0}^{H -1} \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] + \sum_{i = H}^{n-1} \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] \right) & \\
& \leq \frac{1+\epsilon'}{2} \cdot C(i) + \frac{1}{2} \cdot \EU[S \sim \mathcal{U}_{H}| i \not \in S][C_S(i)] & \text{Lemma~\ref{l:decreasing}} \\
& \leq \frac{1+\epsilon'}{2(1-\kappa)} \cdot v_i + \frac{1}{2(1-e^{-\frac{\epsilon' n}{6}})} \cdot v_i & \text{curvature and Lemma~\ref{l:vi}} \\
& \leq \left(\frac{2- \kappa}{2(1-\kappa)} + c_1 \epsilon' \right) \cdot v_i \\
& \leq \left(\frac{2- \kappa}{2(1-\kappa)} + c_2 \epsilon'\right) \cdot \tilde{v}_i & \text{Lemma~\ref{l:concentration}} \\
& = \left(\frac{1}{\sqrt{1-\kappa} - c_3 \epsilon'} \right) \cdot \tilde{\phi}_i & \text{ definition of } \tilde{\phi}_i \\
\end{align*}
for some constants $c_1, c_2, c_3$ and let $\epsilon' = \epsilon / c_3 $ to obtain the desired result for any $\epsilon$.
Similarly,
\begin{align*}
\phi_i & = \frac{1}{n} \sum_{j=0}^{L} \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] + \frac{1}{n} \sum_{j = L + 1}^{n-1} \EU[S \sim \mathcal{U}_{j}| i \not \in S][C_S(i)] & \\
& \geq \frac{1-\epsilon'}{2} \cdot \EU[S \sim \mathcal{U}_{L}| i \not \in S][C_S(i)] + \frac{1}{2} (C(N) - C(N \setminus\{i\}) & \text{Lemma~\ref{l:decreasing}} \\
& \geq \frac{1-\epsilon'}{2\left(1 + \frac{e^{-\frac{\epsilon' n}{6}}}{1 - \kappa} \right)} \cdot v_i + \frac{1-\kappa}{2} \cdot v_i & \text{Lemma~\ref{l:vi} and curvature} \\
& \geq \left(\frac{2- \kappa}{2} - c_1 \epsilon' \right) v_i \\
& \geq \left(\frac{2- \kappa}{2} - c_2 \epsilon' \right) \tilde{v}_i & \text{Lemma~\ref{l:concentration}} \\
&= \left(\sqrt{1-\kappa} - c_3 \epsilon'\right) \tilde{\phi}_i & \text{ definition of } \tilde{\phi}_i \\
\end{align*}
\qed
We show that this approximation is optimal. We begin with a general lemma to derive information theoretic inapproximability results for the Shapley value. This lemma shows that if there exists two functions in $\mathcal{C}$ that cannot be distinguished from samples with high probability and that have an element with Shapley value which differs by an $\alpha^2$ factor, then $\mathcal{C}$ does not have a Shapley value that is $\alpha$-approximable from samples.
\begin{lemma}
\label{l:imp}
Consider a family of cost functions $\mathcal{C}$, a constant $\alpha \in (0, 1)$, and assume there exist $C^1, C^2 \in \mathcal{C}$ such that:
\begin{itemize}
\item \textbf{Indistinguishable from samples.} With probability $1 - O(e^{- \beta n})$ over $S \sim \mathcal{D}$ for some constant $\beta > 0$, $$C^1(S) = C^2(S).$$
\item \textbf{Gap in Shapley value.} There exists $i \in N$ such that $$\phi^{C^1}_i < \alpha^2 \phi^{C^2}_i.$$
\end{itemize}
Then, $\mathcal{C}$ does not have Shapley value that is $\alpha$-approximable from samples over $\mathcal{D}$.
\end{lemma}
\begin{proof}
By a union bound, given $m$ sets $S_1, \ldots, S_m$ drawn i.i.d. from $\mathcal{D}$ with $m$ polynomial in $n$, $C^1(S_j) = C^2(S_j)$ for all $S_j$ with probability $1 - O(e^{- \beta n})$.
Let $C = C^1$ or $C = C^2$ with probability $1/2$ each. Assume the algorithm is given $m$ samples such that $C^1(S_j) = C^2(S_j) $ and consider its (possibly randomized) choice $\tilde{\phi}_i$. Note that $\tilde{\phi}_i$ is independent of the randomization of $C$ since $C^1$ and $C^2$ are indistinguishable to the algorithm. Since $\phi^{C^1}_i / \phi^{C^2}_i < \alpha^2$, $\tilde{\phi}_i$ is at least a factor $\alpha$ away from $\phi^{C}_i$ with probability at least $1/2$ over the choices of the algorithm and $C$. Label the cost functions so that $\tilde{\phi}_i$ is at least a factor $\alpha$ away from $\phi^{C^1}_i$ with probability at least $1/2$ over the choices of the algorithm. Thus, with $\delta = 1/4$, there exists no algorithm such that for all $C' \in \{C_1, C^2\}$, $\alpha \cdot \phi^{C'}_i \leq \tilde{\phi}^{C'}_i \leq \frac{1}{\alpha} \cdot \phi^{C'}_i $ with probability at least $3/4$ over both the samples and the choices of the algorithm.
\end{proof}
We obtain the inapproximability result by constructing two such functions.
\begin{restatable}{rThm}{tlbcurv}
\label{t:lbcurv}
For every $\kappa < 1$, there exists a hypothesis class of submodular functions with curvature $\kappa$ that have Shapley value that is not $\sqrt{1-\kappa} + \epsilon$-approximable from samples over the uniform distribution, for every constant $\epsilon > 0$.
\end{restatable}
\begin{proof}
\begin{figure}
\captionsetup{width=0.85\textwidth}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{lb1}
\caption{}
\label{f:1}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{lb2}
\label{f:2}
\caption{}
\end{subfigure}
\caption{The marginal contributions $C^1_S(i^{\star})$, (a), and $C^2_S(i^{\star})$, (b), of $i^{\star}$ to a set $S$ of size $j$. }
\label{f:lb}
\end{figure}
We first give a technical overview.
\begin{itemize}
\item We construct two functions $C^1$ and $C^2$ which are indistinguishable from samples but that have an element $i^{\star}$ for which its marginal contribution differs by a factor of $1- \kappa$ for the two functions and then Lemma~\ref{l:imp} concludes the proof.
\item The expected marginal contribution $\E_{S \sim \mathcal{U}_{j}| i^{\star} \not \in S}[C_S(i^{\star})]$ for both of these functions is illustrated in Figure~\ref{f:lb} as a function of $j$. Informally, $\E_{S \sim \mathcal{U}_{j}| i^{\star} \not \in S}[C_S(i^{\star})]$ is equal for both functions between $L$ and $H$ to obtain indistinguishability from samples since samples are of size between $L$ and $H$ with high probability. Combining this constraint with the submodular and curvature constraints, the gap between $\E_{S \sim \mathcal{U}_{j}| i^{\star} \not \in S}[C^1_S(i^{\star})]$ and $\E_{S \sim \mathcal{U}_{j}| i^{\star} \not \in S}[C^2_S(i^{\star})]$ is maximized for all $j < L$ and $j > H$ to maximize the gap in the Shapley value for $i^{\star}$.
\end{itemize}
These two functions have a simpler definition via their marginal contributions, so we start by defining them in terms of these marginal contributions and we later give their formal definition to show that they are well-defined. The marginal contributions of $i^{\star}$ are illustrated in Figure~\ref{f:lb}.
$$C^1_S(i) = \begin{cases} 1 & \text{ if } |S| < L \\ 1 - \kappa & \text{ otherwise} \end{cases}$$
$$C^2_S(i^{\star}) = \begin{cases} 1 - \kappa & \text{ if } |S| \leq H \\ 1- \kappa - (|S| - H) \cdot \frac{1 - \kappa - (1 - \kappa)^2}{\sqrt{n}} & \text{ if } H < |S| \leq H+ \sqrt{n} \\ (1 - \kappa)^2 & \text{ otherwise} \end{cases}$$
For $i \neq i^{\star}$:
$$C^2_S(i) = \begin{cases} \frac{L - (1 - \kappa)}{L-1} & \text{ if } |S| < L-1 \text{ or } (|S| = L - 1 \text{ and } i^{\star} \in S) \\ 1 - \kappa & \text{ if } (|S| = L - 1 \text{ and } i^{\star} \not \in S) \text{ or } L \leq |S| \leq H \text{ or } \\ & \ \ \ (H \leq |S| \leq H + \sqrt{n} \text{ and } i^{\star} \not \in S) \\ 1- \kappa - \frac{1 - \kappa - (1 - \kappa)^2}{\sqrt{n}} & \text{ otherwise } \end{cases}$$
The formal definitions of the functions are
$$C^1(S) = \begin{cases} |S| & \text{ if } |S| < L \\ L + (|S| - L) \cdot (1 - \kappa) & \text{ otherwise} \end{cases}$$
and
$$C^2(S) = \begin{cases} 1_{i^{\star} \in S} \cdot (1 - \kappa) + (|S| - 1_{i^{\star} \in S}) \cdot \frac{L - (1 - \kappa)}{L-1} & \text{if } |S| < L \\
L + (|S| - L) \cdot (1- \kappa) & \text{if } L \leq |S| \leq H \\ & \text{or } (H < |S| \leq H + \sqrt{n} \text{ and } i^{\star} \not \in S) \\
L + (|S| - L) \cdot (1- \kappa) & \\ \hspace{0.35cm} + 1- \kappa - (|S| - H) \cdot \frac{1 - \kappa - (1 - \kappa)^2}{\sqrt{n}} & \text{if } H < |S| \leq H + \sqrt{n} \text{ and } i^{\star} \in S \\
L + (H + \sqrt{n} - L) \cdot (1- \kappa) + 1_{i^{\star} \in S} \cdot (1 - \kappa)^2 &\\ \hspace{0.35cm} + (|S| - 1_{i^{\star} \in S} - (H + \sqrt{n}))(1 - \kappa - \frac{1 - \kappa - (1 - \kappa)^2}{\sqrt{n}}) & \text{otherwise}
\end{cases}$$
The Shapley value of $i^{\star}$ with respect to $C^1$ and $C^2$ is then:
$$\phi^{C^1}_{i^{\star}} = 1 \cdot \frac{1 - \epsilon'}{2} + (1- \kappa) \cdot \frac{1 + \epsilon'}{2} \geq \frac{2 - \kappa}{2} - \epsilon$$
and
$$\phi^{C^2}_{i^{\star}} \leq (1-\kappa) \cdot \frac{1 + \epsilon'}{2} + (1 - \kappa)^2 \cdot \frac{1 - \epsilon'}{2} \leq \frac{(1- \kappa)(2 - \kappa)}{2} + \epsilon$$
for an appropriate choice of $\epsilon'$. Next, by Chernoff bound and a union bound, $L \leq |S| \leq H$ for polynomially many samples $S$ from the uniform distirbution, with probability $1 - e^{-\Omega(n)}$. Thus, $C^1(S) = C^2(S)$ for all samples $S$ with probability $1 - e^{-\Omega(n)}$.
It remains to show that $C^1$ and $C^2$ are submodular with curvature $\kappa$, i.e., for any $S \subseteq T$ and $i \not \in T$, $$C_S(i) \geq C_T(i) \geq (1 - \kappa) C_S(i),$$ which is immediate for $C^1$. Regarding $C^2$, it is also immediate that $C^2_S(i^{\star}) \geq C^2_T(i^{\star}) \geq (1 - \kappa) C^2_S(i^{\star})$. For $i \neq i^{\star}$, observe that $$\frac{L - (1 - \kappa)}{L-1} \leq 1 + \epsilon\hspace{0.5cm} \text{ and } \hspace{0.5cm} 1- \kappa - \frac{1- \kappa - (1 - \kappa)^2}{\sqrt{n}} \geq 1 - \kappa - \epsilon,$$ so $C^2_S(i) \geq C^2_T(i) \geq (1 - \kappa - \epsilon) C^2_S(i)$.
\end{proof}
\subsection{Learnability does not imply approximability of the Shapley value}
Although the Shapley value is approximable for the class of submodular functions with bounded curvature, we show that the Shapley value of coverage (and submodular) functions are not approximable from samples in general. The impossibility result is information theoretic and is not due to computational limitations. Coverage functions are an interesting class of functions because they are learnable from samples over any distribution \cite{badanidiyuru2012sketching}, according to the PMAC learning model \cite{balcan2011learning}, which is a generalization of PAC learnability for real valued functions. In addition, by Theorem~\ref{p:controlledD}, coverage functions have Shapley value that can efficiently be approximated arbitrarily well in the value query model. Thus, this impossibility result implies that learnability and approximability in the value query model are not strong enough conditions for approximability of the Shapley value from samples.
\begin{theorem}
\label{thm:lower}
There exists no constant $\alpha > 0$ such that coverage functions have Shapley value that is $\alpha$-approximable from samples over the uniform distribution.
\end{theorem}
\begin{proof}
Partition $N$ into two parts $A$ and $B$ of equal size. Consider the following two functions:
$$ C^1(S) = \begin{cases} 0 & \text{if } S = \emptyset \\ 1 & \text{if } |S \cap A| > 0 \text{, } |S \cap B| = 0 \\ \frac{1}{\alpha^2} & \text{if } |S \cap A| = 0 \text{, } |S \cap B| > 0 \\ 1 + \frac{1}{\alpha^2} & \text{otherwise}\end{cases} \hspace{0.3cm} C^2(S) = \begin{cases} 0 & \text{if } S = \emptyset \\ 1 & \text{if } |S \cap B| > 0 \text{, } |S \cap A| = 0 \\ \frac{1}{\alpha^2} & \text{if } |S \cap B| = 0 \text{, } |S \cap A| > 0 \\ 1 + \frac{1}{ \alpha^2} & \text{otherwise}\end{cases}$$
These functions are coverage functions with $U = \{a, b_1, \ldots, b_{1/ \alpha^2}\}$ and $T_i = \{a\}$ or $T_i = \{b_1, \ldots, b_{1/ \alpha^2}\}$. By the Chernoff bound (Lemma~\ref{l:chernoff}) with $\delta = 1/2$ and $\mu = n/2$, if $S$ is a sample from the uniform distribution, then $$ \Pr\left[|S \cap A| = 0\right] = \Pr\left[|S \cap B| = 0\right] < \Pr\left[|S \cap B| \leq n / 4\right] \leq e^{-n/16},$$ so $C^1(S) \neq C^2(S)$ with probability at most $2e^{-n/16}$. It is easy to see that for any $i$, its Shapley value is either $2/n$ or $2/(\alpha^2 n)$ depending on which partition it is in. Combining this with Lemma~\ref{l:imp} concludes the proof.
\end{proof}
\section{Data Dependent Shapley Value}
The general impossibility result for computing the Shapley value from samples arises from the fact that the concept was geared towards the query model, where the algorithm can ask for the cost of any set $S \subseteq N$. In this section, we develop an analogue that is data- or distribution-dependent. We denote it by $\phi^{C, \mathcal{D}}$ with respect to both $C$ and $\mathcal{D}$. We define four natural distribution-dependent axioms resembling the Shapley value axioms, and then prove that our proposed value is the unique solution satisfying them. This value can be approximated arbitrarily well in the statistical model for all functions. We start by stating the four axioms.
\begin{definition}
\label{d:ddshapley}
The \emph{data-dependent axioms} for cost sharing functions are:
\begin{itemize}
\item \textbf{Balance:} $\sum_{i \in N} \phi^{\mathcal{D}}_i = \E_{S \sim \mathcal{D}}[C(S)]$,
\item \textbf{Symmetry:} for all $i$ and $j$, if $\Pr_{S \sim \mathcal{D}}\left[|S \cap \{i, j\}| =1\right] = 0 $ then $\phi^{\mathcal{D}}_i = \phi^{\mathcal{D}}_j$,
\item \textbf{Zero element:} for all $i$, if
$\Pr_{S \sim \mathcal{D}}\left[i \in S\right] = 0 $ then $\phi^{\mathcal{D}}_i = 0$,
\item \textbf{Additivity:} for all $i$, if $\mathcal{D}_1$, $\mathcal{D}_2$, $\alpha$, and $\beta$ such that $\alpha + \beta = 1$,
$\phi^{\alpha \mathcal{D}_1 + \beta \mathcal{D}_2}_i = \alpha \phi^{\mathcal{D}_1}_i + \beta \phi^{\mathcal{D}_2}_i$
where $\Pr\left[S \sim \alpha \mathcal{D}_1 + \beta \mathcal{D}_2\right] = \alpha \cdot \Pr\left[S \sim \mathcal{D}_1\right] + \beta \cdot \Pr\left[S \sim \mathcal{D}_2\right]$.
\end{itemize}
\end{definition}
The similarity to the original Shapley value axioms is readily apparent. The main distinction is that we expect these to hold with regard to $\mathcal{D}$, which captures the frequency with which different coalitions $S$ occur. Note that we no longer require that $\mathcal{D}$ has full support over $2^N$. Interpreting the axioms one by one, the balance property ensures that the expected cost is always accounted for. The symmetry axiom states that if two elements always occur together, they should have the same share, since they are indistinguishable. If an element is never observed, then it should have zero share. Finally costs should combine in a linear manner according to the distribution.
These axioms are specifically designed to provide some guarantees on the shares of elements to functions with complex interactions where recovery is hard from samples.
We define the \emph{data-dependent Shapley value}:
$$\phi^{\mathcal{D}}_i := \sum_{S \ : \ i \in S} \Pr\left[S \sim \mathcal{D}\right] \cdot \frac{C(S)}{|S|}.$$
Informally, for all set $S$, the cost $C(S)$ is divided equally between all elements in $S$ and is weighted with the probability that $S$ occurs according to $\mathcal{D}$. The main appeal of this cost allocation is the following theorem.
\begin{theorem}
\label{t:ddshapley}
The data-dependent Shapley value is the unique value satisfying the four data-dependent axioms.
\end{theorem}
We first show that if there exists a value satisfying the axioms, it must be the data-dependent Shapley value. Then, we show that the data-dependent Shapley value satisfies the axioms, which concludes the proof.
\begin{lemma}
If there exists a value satisfying the four data-dependent Shapley axioms, then this value is the data-dependent Shapley value.
\end{lemma}
\begin{proof}
Define $\mathcal{D}_S$ to be the distribution such that $\Pr\left[S \sim \mathcal{D}_S\right] = 1$. Observe that the unique value satisfying the balance, symmetry, and zero element axioms must satisfy
$$\phi^{\mathcal{D}_S}_i = \begin{cases} \frac{C(S)}{|S|} & \text{ if } i\in S \\ 0 & \text{ otherwise.} \end{cases}$$
Since $\mathcal{D} = \sum_{S} \Pr\left[S \sim \mathcal{D}\right] \cdot \mathcal{D}_S$, the unique value satisfying the four axioms must satisfy
$$\phi^{\mathcal{D}}_i = \sum_{S} \Pr\left[S \sim \mathcal{D}\right] \cdot \phi^{\mathcal{D}_S}_i = \sum_{S : i \in S} \Pr\left[S \sim \mathcal{D}\right] \cdot \frac{C(S)}{|S|}$$
where the first equality is by additivity and the second equality by the above observation.
\end{proof}
\begin{lemma}
The data-dependent Shapley value satisfies the four data-dependent Shapley axioms.
\end{lemma}
\begin{proof} We show that each axiom is satisfied.
\begin{itemize}
\item \textbf{Balance:} By definition,
$\sum_{i \in N} \phi^{\mathcal{D}}_i = \sum_{i \in N}\sum_{S:i \in S} \Pr\left[S \sim \mathcal{D}\right] C(S) /|S|$, then by switching the order of the summations, $$\sum_{i \in N}\sum_{S : i \in S} \Pr\left[S \sim \mathcal{D}\right] \frac{C(S)}{|S|} = \sum_{S \subseteq N} \sum_{i \in S} \Pr\left[S \sim \mathcal{D}\right] \frac{C(S)}{|S|} = \sum_{S \subseteq N} \Pr\left[S \sim \mathcal{D}\right] C(S) = \EU[S \sim \mathcal{D}][C(S)].$$
\item \textbf{Symmetry:} Let $\mathcal{S}_i = \{S: i \in S, \Pr\left[S \sim \mathcal{D}\right] > 0\}$. If $\Pr_{S \sim \mathcal{D}}\left[|S \cap \{i, j\}| =1\right] = 0$, then $\mathcal{S}_{i} = \mathcal{S}_{j}$ and
$$\phi^{\mathcal{D}}_i = \sum_{S \in \mathcal{S}_{i}} \Pr\left[S \sim \mathcal{D}\right] \frac{C(S)}{|S|} = \sum_{S \in \mathcal{S}_{j}} \Pr\left[S \sim \mathcal{D}\right] \frac{C(S)}{|S|} = \phi^{\mathcal{D}}_j.$$
\item \textbf{Zero element:} If $\Pr_{S \sim \mathcal{D}}\left[i \in S\right] = 0$, then $\Pr\left[S \sim \mathcal{D}\right] = 0$ if $i \in S$. Thus, $\phi^{\mathcal{D}}_i = 0$.
\item \textbf{Additivity:} By definition of $\phi$ and $\alpha \mathcal{D}_1 + \beta \mathcal{D}_2$,
\begin{align*}
\phi^{\alpha \mathcal{D}_1 + \beta \mathcal{D}_2}_i & = \sum_{S \ : \ i \in S} \Pr\left[S \sim \alpha \mathcal{D}_1 + \beta \mathcal{D}\right] \frac{C(S)}{|S|}
\\
& = \alpha \sum_{S: i \in S} \Pr\left[S \sim \mathcal{D}_1\right] \frac{C(S)}{|S|} + \beta \sum_{S:i \in S} \Pr\left[S \sim \mathcal{D}_2\right] \frac{C(S)}{|S|} \\
& = \alpha \phi^{\mathcal{D}_1}_i + \beta \phi^{\mathcal{D}_2}_i.
\end{align*}
\end{itemize}
\end{proof}
The data-dependent Shapley value can be approximated from samples with the following empirical data-dependent Shapley value:
$$\tilde{\phi}^{\mathcal{D}}_i = \frac{1}{m} \sum_{S_j \ : \ i \in S_j} \frac{C(S_j)}{|S_j|}.$$
These estimates are arbitrarily good with arbitrarily high probability.
\begin{theorem}
\label{t:apxddshapley}
The empirical data-dependent Shapley value approximates the data-dependent Shapley value arbitrarily well, i.e., $|\tilde{\phi}^{\mathcal{D}}_i - \phi^{\mathcal{D}}_i| < \epsilon$ with $\poly(n, 1/\epsilon, 1/\delta)$ samples and with probability at least $1 - \delta$ for any $\delta, \epsilon > 0$.
\end{theorem}
\begin{proof}
Define $X_j = \begin{cases} \frac{C(S_j)}{ |S_j|} & \text{ if } i \in S_j \\ 0 & \text{ otherwise} \end{cases}$
and observe that $(\sum_{j=1}^m X_j)/m = \tilde{\phi}^{\mathcal{D}}_i$ and\\ $\E[(\sum_{j=1}^m X_j)/m] = \phi^{\mathcal{D}}_i$. Clearly, $X_j \in [0, b]$ where $b := \max_S C(S) / |S|$, so by Hoeffding's inequality,
$\Pr\left[|\tilde{\phi}^{\mathcal{D}}_i - \phi^{\mathcal{D}}_i| \geq \epsilon\right] \leq 2e^{-\frac{2m \epsilon^{2}}{\poly(n)}}$ with $0 < \epsilon < 1$.
\end{proof}
\section{Discussion and Future Work}
We follow a recent line of work that studies classical algorithmic problems from a statistical perspective, where the input is restricted to a collection of samples. Our results fall into two categories, we give results for approximating the Shapley value and the core and propose new cost sharing concepts that are tailored for the statistical framework. We use techniques from multiple fields that encompass statistical machine learning, combinatorial optimization, and, of course, cost sharing. The cost sharing literature being very rich, the number of directions for future work are considerable. Obvious avenues include studying other cost sharing methods in this statistical framework, considering other classes of functions to approximate known methods, and improving the sample complexity of previous algorithms. More conceptually, an exciting modeling question arises when designing ``desirable" axioms from data. Traditionally these axioms only depended on the cost function, whereas in this model they can depend on both the cost function and the distribution, providing an interesting interplay.
\newpage
\bibliographystyle{plainnat}
|
\section{Introduction and main result}
For $d\ge 3$, the classical Hardy inequality states that
\begin{equation}
\label{ec:hardy}
\frac{(d-2)^2}{4}\int_{\mathbb{R}^d}\frac{u^2(x)}{|x|^2}\, dx\le
\int_{\mathbb{R}^d}|\nabla u(x)|^2\, dx.
\end{equation}
Due to its applicability, there is an extensive literature about the topic (see the references in \cite{RT}) covering many extensions of this estimate in several and different directions. We are interested in one involving the fractional powers of the Laplacian. We can rewrite \eqref{ec:hardy} as
\begin{equation*}
\frac{(d-2)^2}{4}\int_{\mathbb{R}^d}\frac{u^2(x)}{|x|^2}\, dx\le \int_{\mathbb{R}^d} u(x) (-\Delta u(x)) \, dx
\end{equation*}
and, taking the fractional Laplacian $(-\Delta)^\sigma$ defined by $\widehat{(-\Delta)^\sigma u}=|\cdot|^{2\sigma} \widehat{u}$, a natural extension is the inequality
\begin{equation}
\label{ec:Hardy-frac}
C_{\sigma,d}\int_{\mathbb{R}^d}\frac{u^2(x)}{|x|^{2\sigma}}\, dx\le \int_{\mathbb{R}^d} u(x) (-\Delta)^{\sigma} u(x) \, dx,
\end{equation}
for which the sharp constant $C_{\sigma,d}$ is well known (see \cite{Beckner3,Yafaev}).
From \eqref{ec:Hardy-frac}, we deduce the positivity (in a distributional sense) of the operator
\[
(-\Delta)^\sigma-\frac{C_{\sigma,d}}{|\cdot|^{2\sigma}}.
\]
Our target is to provide a Hardy inequality like \eqref{ec:Hardy-frac} related to ultraspherical expansions and apply it to prove the positivity of certain operator on the sphere with a potential having singularities in both poles of the sphere.
Let $C_n^\lambda(x)$ be the ultraspherical polynomial of degree $n$ and order $\lambda>-1/2$. We consider $c_n^\lambda(x)=d_n^{-1}C_n^\lambda(x)$ with
\[
d_n^2=
\int_{-1}^{1}\left(C_n^\lambda(x)\right)^2\,d\mu_\lambda(x),\qquad
d\mu_\lambda(x)=(1-x^2)^{\lambda-1/2}\, dx.
\]
The sequence of polynomials $\{c_n^\lambda\}_{n\ge 0}$ forms an orthonormal basis of the space $L_\lambda^2:=L^2((-1,1),d\mu_\lambda)$. For each $c_n^\lambda$, it holds that $\mathcal{L}_\lambda c_n^\lambda=-(n+\lambda)^2 c_n^\lambda$, where
\[
\mathcal{L}_\lambda =(1-x^2)\frac{d^2}{dx^2}-(2\lambda+1)x\frac{d}{dx}-\lambda^2.
\]
The ultraspherical expansion of each appropriate function $f$ defined in $(-1,1)$ is given by
\[
f\longmapsto \sum_{n=0}^\infty a_n^\lambda(f)c_n^\lambda,
\]
where $a_n^\lambda(f)$ is the $n$-th Fourier coefficient of $f$ respect to $\{c_n^\lambda\}_{n\ge 0}$, i.e.,
\[
a_n^\lambda(f)=\int_{-1}^{1}f(y)c_n^\lambda(y)\,d\mu_\lambda(y).
\]
The fractional powers of the operator $\mathcal{L}_\lambda$ are defined by
\[
(-\mathcal{L}_\lambda)^{\sigma/2}f=\sum_{n=0}^\infty(n+\lambda)^{\sigma} a_n^\lambda(f)c_n^\lambda, \qquad \sigma >0.
\]
This operator should be the natural candidate to prove a Hardy type inequality for the ultraspherical expansion but, however, it is not the most appropriate in this setting. We have to consider other one with an analogous behaviour to $(-\mathcal{L}_\lambda)^{\sigma/2}$, in order to deduce some results on the sphere. For each $\sigma>0$ we define (spectrally) the operator
\[
A_\sigma^{\lambda} =\frac{\Gamma(\sqrt{-\mathcal{L}_\lambda}+\frac{1+\sigma}{2})}
{\Gamma(\sqrt{-\mathcal{L}_\lambda}+\frac{1-\sigma}{2})}.
\]
Then for $f$ defined on the interval $(-1,1)$
$$
A_\sigma^{\lambda} f(x)=\sum_{n=0}^{\infty}
\frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}a_n^\lambda(f)c_n^\lambda(x).
$$
Note that
\begin{equation}
\label{eq:asym}
\frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}\simeq (n+\lambda)^{\sigma},
\end{equation}
then the behaviour of $(-\mathcal{L}_\lambda)^{\sigma/2}$ and $A_\sigma^\lambda$ is similar. The natural Sobolev space to analyse Hardy type inequalities is
\[
H^{\sigma}_\lambda=\Big\{f\in L^2_\lambda : \|f\|_{H^{\sigma}_\lambda}:=\Big(\sum_{n=0}^\infty (n+\lambda)^{\sigma}(a_n^\lambda(f))^2\Big)^{1/2}<\infty\Big\}.
\]
We have to note that $H^\sigma_\lambda$ is equivalent to the space $\mathcal{L}_{\lambda,\sigma}^2$ introduced in \cite{Betancoretal}.
With the previous notation our Hardy inequality for ultraspherical expansions is given in the following result.
\begin{Thm}\label{theorem01}
Let $\lambda>0$ and $0<\sigma<1$. Then for $u\in H_\lambda^\sigma$
\begin{equation}
\label{ec:Hardy-Ultra}
Q_{\sigma,\lambda}\int_{-1}^{1}\frac{u^2(x)}{(1-x^2)^{\sigma/2}}\,d\mu_\lambda(x)\leq
\int_{-1}^{1}u(x)A_\sigma^{\lambda} u(x)\,d\mu_\lambda(x),
\end{equation}
where
\begin{equation}
\label{ec:cons-Hardy}
Q_{\sigma,\lambda}=2^\sigma
\frac{\Gamma(\frac{\lambda}{2}+\frac{1+\sigma}{4})^2}{\Gamma(\frac{\lambda}{2}+\frac{1-\sigma}{4})^2}.
\end{equation}
\end{Thm}
Inequality \eqref{ec:Hardy-Ultra} can be rewritten in terms of the Fourier coefficients
\begin{equation}
\label{ec:Pitt}
Q_{\sigma,\lambda}\int_{-1}^{1}\frac{u^2(x)}{(1-x^2)^{\sigma/2}}\,d\mu_\lambda(x)\leq
\sum_{n=0}^\infty \frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}(a_n^\lambda(u))^2,
\end{equation}
which is a kind of Pitt inequality for the ultraspherical expansions (for other Pitt inequalities see \cite{Beckner2,GIT}). Note that for the right hand side of \eqref{ec:Hardy-Ultra} we have, by \eqref{eq:asym},
\[
\int_{-1}^{1}u(x)A_\sigma^{\lambda} u(x)\,d\mu_\lambda(x)=\sum_{n=0}^\infty \frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}(a_n^\lambda(u))^2
\simeq \|u\|_{H^{\sigma}_\lambda}^2,
\]
so the space $H_\lambda^\sigma$ is the adequated one.
The proof of Theorem \ref{theorem01} will be a consequence of a proper ground state representation in our setting, analogous to the given one in the Euclidean case in \cite{FLS}. Following the ideas in that paper, we can see that the constant $Q_{\sigma,\lambda}$ is sharp but not achieved. Similar ideas have been recently exploited in \cite{CRT,RT}.
From \eqref{ec:Hardy-Ultra}, by using Cauchy–-Schwarz inequality, we can obtain a Heisenberg type uncertainty principle as it was done for the sublaplacian of the Heisenberg group in \cite{GL}, and for the fractional powers of the same sublaplacian in~\cite{RT}.
\begin{Cor}
Let $\lambda>0$ and $0<\sigma<1$. Then for $u\in H_\lambda^\sigma$
\begin{equation*}
Q_{\sigma,\lambda}\left(\int_{-1}^{1} u^2(x)\,d\mu_\lambda(x)\right)^{2}\leq
\int_{-1}^{1}u^2(x)(1-x^2)^{\sigma/2}\,d\mu_\lambda(x)
\int_{-1}^{1}u(x)A_\sigma^{\lambda} u(x)\,d\mu_\lambda(x),
\end{equation*}
where $Q_{\sigma,\lambda}$ is the constant given in \eqref{ec:cons-Hardy}.
\end{Cor}
Pitt inequality \eqref{ec:Pitt} allows us to prove a logarithmic uncertainty principle for the ultraspherical expansions. The main idea comes from \cite{Beckner3}. By an elementary argument, for a derivable function such that $\phi(0)=0$ and $\phi(\sigma)>0$ for $\sigma\in(0,\varepsilon)$, with $\varepsilon >0$, it is verified that $\phi'(0_+)\ge 0$. Then, taking the function
\[
\phi(\sigma)=
\sum_{n=0}^\infty \frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}(a_n^\lambda(u))^2
-Q_{\sigma,\lambda}\int_{-1}^{1}\frac{u^2(x)}{(1-x^2)^{\sigma/2}}\,d\mu_\lambda(x),
\]
we have $\phi(0)=0$ (this is Parseval identity) and, by \eqref{ec:Pitt}, $\phi(\sigma)>0$ for $\sigma \in (0,1)$, then $\phi'(0_+)\ge 0$ and this inequality gives the logarithmic uncertainty principle, which is written as
\begin{multline*}
\left(\log 2+\psi\left(\frac{\lambda}{2}+\frac14\right)\right)\int_{-1}^1 u^2(x)\, d\mu_\lambda(x)\\\le
\sum_{n=0}^\infty \psi\left(n+\lambda+\frac12\right)(a_n(u))^2+\int_{-1}^1\log(\sqrt{1-x^2})u^2(x)\, d\mu_\lambda(x),
\end{multline*}
where $\psi(a)=\frac{\Gamma'(a)}{\Gamma(a)}$.
In next section we will show an application of Theorem \ref{theorem01} to obtain a Hardy inequality on the sphere. The results in Section \ref{sec:aux} are the main ingredients in the proof of Theorem \ref{theorem01} which is given in last section of the paper.
\section{An application to the sphere}
It is well known that $L^2(\mathbb{S}^d)=\oplus_{n=0}^\infty \mathcal{H}_n(\mathbb{S}^d)$, where $\mathcal{H}_n(\mathbb{S}^d)$ is the set of spherical harmonics of degree $n$ in $d+1$ variables. If we consider the shifted Laplacian on the sphere
\[
-\Delta_{\mathbb{S}^d}=\tilde{-\Delta_{\mathbb{S}^d}}+\left(\frac{d-1}{2}\right)^2,
\]
where $\tilde{-\Delta_{\mathbb{S}^d}}$ is the Laplace-Beltrami operator on $\mathbb{S}^d$, it is verified that
\[
-\Delta_{\mathbb{S}^d} \mathcal{H}_n(\mathbb{S}^d)=\left(n+\frac{d-1}{2}\right)^2\mathcal{H}_n(\mathbb{S}^d).
\]
In this way, the analogous of the operator $A_\sigma^{\lambda}$ on $\mathbb{S}^d$ is defined by
\begin{align*}
\mathbf{A}_\sigma f&=\frac{\Gamma\left(\sqrt{-\Delta_{\mathbb{S}^d}}+\frac{1+\sigma}{2}\right)}
{\Gamma\left(\sqrt{-\Delta_{\mathbb{S}^d}}+\frac{1-\sigma}{2}\right)}f\\
&=\sum_{n=0}^\infty\frac{\Gamma\left(n+\frac{d-1}{2}+\frac{1+\sigma}{2}\right)}
{\Gamma\left(n+\frac{d-1}{2}+\frac{1-\sigma}{2}\right)}\proj_{\mathcal{H}_n(\mathbb{S}^d)}f,
\end{align*}
where $\proj_{\mathcal{H}_n(\mathbb{S}^d)}f$ denotes the projection of $f$ onto the eigenspace $\mathcal{H}_n(\mathbb{S}^d)$.
The operator $\mathbf{A}_\sigma$ becomes the fractional powers of the Laplacian in the Euclidean space through conformal transforms as was observed by T. P. Branson in~\cite{Branson}. So $\mathbf{A}_\sigma$ is the natural operator to prove a Hardy type inequality on the sphere. In our proof, we will write $\mathbf{A}_\sigma$ in terms of $A_\sigma^\lambda$ and this is the main reason to consider $A_\sigma^\lambda$ in the case of the ultraspherical expansions. An analogous of the Hardy-Littlewood-Sobolev inequality for $\mathbf{A}_\sigma$ and some other inequalities for it were given by W. Beckner in \cite{Beckner-Annals}. The operators $\mathbf{A}_\sigma$ also appear in \cite[p. 151]{Samko} and \cite[p. 525]{Rubin}.
Each point $x\in \mathbb{S}^d$ can be written as
\[
x=(t,\sqrt{1-t^2}x'_1,\dots,\sqrt{1-t^2}x'_d),
\]
for $t\in (-1,1)$ and $x':=(x'_1,\dots,x'_d)\in \mathbb{S}^{d-1}$, and so
\[
\int_{\mathbb{S}^d} f(x)\, dx=\int_{-1}^1 \int_{\mathbb{S}^{d-1}}f(t,\sqrt{1-t^2}x')(1-t^2)^{(d-2)/2}\, dx'\, dt.
\]
With these coordinates, see \cite[Section 3]{Sherman}, we have that an orthonormal basis for each $\mathcal{H}_n(\mathbb{S}^d)$ is given by
\[
\phi_{n,j,k}(x)=\psi_{n,j}(t)Y_{j,k}^d (x'), \qquad j=0,\dots,n,
\]
with
\[
\psi_{n,j}(t)=(1-t^2)^{j/2}c_{n-j}^{j+(d-1)/2}(t)
\]
and $\{Y_{j,k}^d\}_{k=1,\dots,d(j)}$ an orthonormal basis of spherical harmonics on $\mathbb{S}^{d-1}$ of degree $j$. The value $d(j)$ indicates the dimension of $\mathcal{H}_j(\mathbb{S}^{d-1})$; i.e.,
\[
d(j)=(2j+d-2)\frac{(j+d-3)!}{j!(d-2)!}.
\]
Then, the orthogonal projection of $f$ onto the eigenspace $\mathcal{H}_n(\mathbb{S}^d)$ can be written as
\[
\proj_{\mathcal{H}_n(\mathbb{S}^d)}f=\sum_{j=0}^n \sum_{k=1}^{d(j)} f_{n,j,k} \phi_{n,j,k},
\]
with
\[
f_{n,j,k}=\int_{-1}^{1} G_{j,k}(t) c_{n-j}^{j+(d-1)/2}(t) (1-t^2)^{j+(d-2)/2}\, dt,
\]
\[
G_{j,k}(t)=(1-t^2)^{-j/2}F_{j,k}(t)\quad\text{ and }\quad
F_{j,k}(t)=\int_{\mathbb{S}^{d-1}}f(t,\sqrt{1-t^2}x')Y_{j,k}^d(x')\, dx'.
\]
It is easy to observe that
\[
f(x)=\sum_{j=0}^\infty \sum_{k=1}^{d(j)} F_{j,k}(t)Y_{j,k}^d(x')= \sum_{j=0}^\infty \sum_{k=1}^{d(j)} (1-t^2)^{j/2}G_{j,k}(t)Y_{j,k}^d(x').
\]
Moreover, from the definition of $\mathbf{A}_\sigma$, we have
\[
\mathbf{A}_\sigma f(x)=\sum_{j=0}^\infty \sum_{k=1}^{d(j)}(1-t^2)^{j/2} A_{\sigma}^{j+(d-1)/2} G_{j,k}(t)Y_{j,k}^d(x').
\]
Now, considering the Sobolev space
\[
\mathbf{H}^{\sigma}=\Big\{f\in L^2(\mathbb{S}^d) : \|f\|_{\mathbf{H}^{\sigma}}:=\Big(\sum_{n=0}^\infty \Big(n+\frac{d-1}{2}\Big)^{\sigma}
\|\proj_{\mathcal{H}_n(\mathbb{S}^d)}f\|^2_{L^2(\mathbb{S}^d)}\Big)^{1/2}<\infty\Big\},
\]
we have the following Hardy inequality on the sphere.
\begin{Thm}\label{theorem02}
Let $d\ge 2$, $0<\sigma<1$, and $e_d$ be the north pole of the sphere $\mathbb{S}^d$. Then for $f\in \mathbf{H}^\sigma$
\begin{equation}
\label{ec:Hardy-sphere}
2^\sigma Q_{\sigma,(d-1)/2}\int_{\mathbb{S}^d}\frac{f^2(x)}{(|x-e_d||x+e_d|)^{\sigma}}\,dx\leq
\int_{\mathbb{S}^d}f(x)\mathbf{A}_\sigma f(x)\,dx,
\end{equation}
where $Q_{\sigma,(d-1)/2}$ is the constant given in \eqref{ec:cons-Hardy}.
\end{Thm}
\begin{proof}
By the orthogonality of the spherical harmonics, it is elementary to show that
\[
\int_{\mathbb{S}^d}f(x)\mathbf{A}_\sigma f(x)\,dx=\sum_{j=0}^\infty \sum_{k=1}^{d(j)}\int_{-1}^1 G_{j,k}(t) A_{\sigma}^{j+(d-1)/2} G_{j,k}(t)\, d\mu_{j+(d-1)/2}(t).
\]
Now, applying Theorem \ref{theorem01}, we deduce that
\[
\int_{\mathbb{S}^d}f(x)\mathbf{A}_\sigma f(x)\,dx\ge
\sum_{j=0}^\infty \sum_{k=1}^{d(j)}Q_{\sigma, j+(d-1)/2} \int_{-1}^1 \frac{F_{j,k}^2(t)}{(1-t^2)^{\sigma/2}} \, d\mu_{(d-1)/2}(t).
\]
It is known (see \cite{Yafaev}) that for $0<x\le y$ and $j\ge 0$ we have that $\frac{\Gamma(j+y)}{\Gamma(j+x)}\ge \frac{\Gamma(y)}{\Gamma(x)}$. So, $Q_{\sigma,j+(d-1)/2}\ge Q_{\sigma,(d-1)/2}$ and
\[
\int_{\mathbb{S}^d}f(x)\mathbf{A}_\sigma f(x)\,dx\ge Q_{\sigma,(d-1)/2}
\sum_{j=0}^\infty \sum_{k=1}^{d(j)}\int_{-1}^1 \frac{F_{j,k}^2(t)}{(1-t^2)^{\sigma/2}} \, d\mu_{(d-1)/2}(t).
\]
The proof of \eqref{ec:Hardy-sphere} is finished by using the identity
\[
\sum_{j=0}^\infty \sum_{k=1}^{d(j)}\int_{-1}^1 \frac{F_{j,k}^2(t)}{(1-t^2)^{\sigma/2}} \, d\mu_{(d-1)/2}(t)=
2^\sigma \int_{\mathbb{S}^d}\frac{f^2(x)}{(|x-e_d||x+e_d|)^{\sigma}}\, dx.
\]
\end{proof}
The analogous role on the sphere of radially symmetric functions is played by functions which are invariant under the action of $SO(d-1)$. By $SO(d-1)$-invariance we mean that $f$ is invariant under the action of the group $SO(d-1)$ on $\mathbb{S}^{d-1}$ whenever $SO(d-1)$ is embedded into $SO(d)$ in a suitable way. Each function $f$ of this kind can be written as $f(x)=g(\langle x,e_d\rangle )$, for a certain function $g$ defined in $(-1,1)$. Then for this kind of functions Theorem \ref{theorem02} reduces to Theorem \ref{theorem01} with $\lambda=(d-1)/2$, in this way we can deduce that the constant $2^\sigma Q_{\sigma,(d-1)/2}$ in \eqref{ec:Hardy-sphere} is sharp.
As in the classic case, from Theorem \ref{theorem02} we deduce that in a distributional sense
\[
\mathbf{A}_\sigma-\frac{2^\sigma Q_{\sigma,(d-1)/2}}{(|x-e_d||x+e_d|)^{\sigma}}\ge 0.
\]
Note that in this case we are perturbing the operator $\mathbf{A}_\sigma$ adding a potential with singularities in both poles of the sphere.
\section{Auxiliary results}
\label{sec:aux}
The following lemmas give the tools to prove Theorem \ref{theorem01}. To be more precise, Lemma \ref{lemma01} provides a nonlocal representation of the operator $A_\sigma^{\lambda}$ with a kernel having nice properties for our target. Lemma \ref{lemma02} shows the action of the operator $A_\sigma^{\lambda}$ on the family of weights $(1-x^2)^{-(\lambda/2+(1-\sigma)/4)}$.
For $f,g\in L_\lambda^2$ we are going to set up the notation
$$
\langle f,g\rangle_\lambda=\int_{-1}^{1}f(x)g(x)\,d\mu_\lambda(x)
$$
to simplify the writing.
\begin{Lem}\label{lemma01}
Let $\lambda>0$ and $0<\sigma<1$. If $f$ is a finite linear combination of ultraspherical polynomials, then
\begin{equation}\label{eq:Lem01_principal}
A_\sigma^{\lambda} f(x)
=
\int_{-1}^1\left(f(x)-f(y)\right)K_\sigma^\lambda(x,y)\,d\mu_\lambda(y)+E_{\sigma,\lambda}f(x), \qquad x\in (-1,1),
\end{equation}
where the kernel is given by
\[
K_\sigma^\lambda(x,y)
=
D_{\sigma,\lambda}\int_{-1}^{1}\frac{d\mu_{\lambda-1/2}(t)}{(1-xy-\sqrt{1-x^2}\sqrt{1-y^2}t)^{\lambda+(1+\sigma)/2}},
\]
with
\[
D_{\sigma,\lambda}=\frac{c_\lambda^2}{2^{\lambda+(1+\sigma)/2}}
\frac{\Gamma(\frac{1-\sigma}{2})\Gamma(\lambda+\frac{1+\sigma}{2})}{|\Gamma(-\sigma)|\Gamma(1+\lambda)}, \qquad c_\lambda=\frac{\Gamma(2\lambda+1)}{2^{2\lambda}(\Gamma(\lambda+1/2))^2},
\]
and
\[
E_{\sigma,\lambda}=\frac{\Gamma(\lambda+\frac{1+\sigma}{2})}{\Gamma(\lambda+\frac{1-\sigma}{2})}.
\]
Moreover, for $f\in H_\lambda^\sigma$ we have
\begin{equation}
\label{eq:esc-prod}
\langle A_\sigma^\lambda f, f\rangle_\lambda=\frac{1}{2}\int_{-1}^1\int_{-1}^1 (f(x)-f(y))^2 K_\sigma^\lambda (x,y) \, d\mu_\lambda(y)\, d \mu_\lambda(x)+E_{\sigma,\lambda}\langle f, f \rangle_\lambda
\end{equation}
\end{Lem}
\begin{proof}
We start with the identity
\begin{equation}\label{eq:Lema0}
\int_{0}^{\infty}\left(e^{-(n+\lambda)t}-e^{-(\sigma-1)t/2}\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt
=
2^{1+\sigma}\Gamma(-\sigma)\frac{\Gamma(n+\lambda+\frac{1+\sigma}{2})}{\Gamma(n+\lambda+\frac{1-\sigma}{2})}
\end{equation}
for $\lambda>0$ (actually it is also true for values $\lambda>-1/2$) and $0<\sigma<1$. To deduce the previous identity it is enough to apply integration by parts with $u=e^{-(n+\lambda+(1-\sigma)/2)t}-1$ and $v=-2e^{-\sigma t/2}(\sinh t/2)^{-\sigma}/\sigma$, and use \cite[eq. 8, p. 367]{PrudnikovI}
$$
\int_{0}^{\infty}e^{-\rho t}\left(\cosh (ct)-1\right)^{\nu}\,dt=\frac{\Gamma(\frac{\rho}{c}-\nu)\Gamma(2\nu+1)}{2^\nu c\Gamma(\frac{\rho}{c}+\nu+1)}
$$
for $c>0$, $2\nu>-1$, and $\rho>c\nu$.
Now, we consider the Poisson operator for ultraspherical expansions. It is given by
\[
e^{-t\sqrt{-\mathcal{L}_\lambda}}f(x)=\sum_{n=0}^{\infty}e^{-(n+\lambda)t}a_n^\lambda(f)c_n^\lambda(x)=\int_{-1}^1 f(y)P_t^\lambda(x,y)\, d\mu_\lambda(y),
\]
with
\[
P_t^\lambda(x,y)=\sum_{n=0}^\infty e^{-(n+\lambda)t}c_n^\lambda(x)c_n^\lambda(y).
\]
By the product formula for ultraspherical polynomials \cite[eq. B.2.9, p. 419]{DaiXu}
\[
\frac{C_n^\lambda(x)C_n^\lambda(y)}{C_n^\lambda(1)}=c_\lambda\int_{-1}^{1}
C_n^\lambda(xy+\sqrt{1-x^2}\sqrt{1-y^2}t)\,d\mu_{\lambda-1/2}(t), \qquad \lambda>0,
\]
the identity \cite[eq. B.2.8. p. 419]{DaiXu}
\[
\sum_{n=0}^{\infty}\frac{n+\lambda}{\lambda}C_n^\lambda(x)r^n=\frac{1-r^2}{(1-2xr+r^2)^{\lambda+1}}, \qquad 0\leq r<1,
\]
and the relation $d_n^2=\frac{\lambda}{c_\lambda(n+\lambda)} C_n^\lambda(1)$, we deduce the expression
\[
P_t^\lambda(x,y)=\frac{c_\lambda^2}{2^\lambda} \int_{-1}^{1}\frac{\sinh t}{(\cosh t -w(s))^{\lambda+1}}\,d\mu_{\lambda-1/2}(s),
\]
with $w(s)=xy+\sqrt{1-x^2}\sqrt{1-y^2}s$. The previous identity for $P_t^\lambda$ is not new, it appears as formula (2.12) in \cite{MS}.
Combining \eqref{eq:Lema0} and the definition of the Poisson operator, it is clear that
\[
A_\sigma^{\lambda} f(x)
=
\frac{1}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}
\left(e^{-t\sqrt{-\mathcal{L}_{\lambda}}}f(x)-f(x)e^{-(\sigma-1)t/2}\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt,
\]
which can be splitted in
\begin{multline}\label{eq:IntSplitLem1}
A_\sigma^{\lambda} f(x)
\\=
\frac{1}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}
\left(e^{-t\sqrt{-\mathcal{L}_\lambda}}f(x)-f(x)e^{-t\sqrt{-\mathcal{L}_\lambda}}1(x)\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt
\\
+\frac{f(x)}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}
\left(e^{-t\sqrt{-\mathcal{L}_\lambda}}1(x)-e^{-(\sigma-1)t/2}\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt.
\end{multline}
From the obvious identity
\[
e^{-t\sqrt{-\mathcal{L}_\lambda}}1(x)=\int_{-1}^{1}P_t^\lambda(x,y)\,d\mu_\lambda(y)=e^{-\lambda t},
\]
for the second term in \eqref{eq:IntSplitLem1} we have
\begin{equation*}
\begin{split}
\frac{f(x)}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}&
\left(e^{-t\sqrt{-\mathcal{L}_\lambda}}1(x)-e^{-(\sigma-1)t/2}\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt
\\ &=
\frac{f(x)}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}\left(e^{-\lambda t}-e^{-(\sigma-1)t/2}\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt
\\ &=
E_{\sigma,\lambda}f(x),
\end{split}
\end{equation*}
where we have used \eqref{eq:Lema0} with $n=0$.
The first integral in \eqref{eq:IntSplitLem1} verifies
\begin{multline*}
\frac{1}{2^{1+\sigma}\Gamma(-\sigma)}\int_{0}^{\infty}
\left(e^{-t\sqrt{-\mathcal{L}_\lambda}}f(x)-f(x)e^{-t\sqrt{-\mathcal{L}_\lambda}}1(x)\right)\left(\sinh t/2\right)^{-\sigma-1}\,dt\\
\begin{aligned}
&= \frac{1}{2^{1+\sigma}|\Gamma(-\sigma)|}\int_{0}^{\infty}\int_{-1}^{1} P_t^\lambda(x,y)(f(x)-f(y))\, d\mu_\lambda(y)\left(\sinh t/2\right)^{-\sigma-1}\,dt
\\& =
\frac{1}{2^{1+\sigma}|\Gamma(-\sigma)|}\int_{-1}^{1}\left(f(x)-f(y)\right)
\int_{0}^{\infty}P_t^\lambda(x,y)\left(\sinh t/2\right)^{-\sigma-1}\,dt\,d\mu_\lambda(y)
\\& =
\int_{-1}^{1}\left(f(x)-f(y)\right)K_\sigma^\lambda(x,y)\,d\mu_\lambda(y),
\end{aligned}
\end{multline*}
with
\[
K_\sigma^\lambda(x,y)= \frac{1}{2^{1+\sigma}|\Gamma(-\sigma)|}\int_{0}^{\infty}P_t^\lambda (x,y)\left(\sinh t/2\right)^{-\sigma-1}\,dt.
\]
In last computation we have used Fubini theorem. This is justified for finite combinations of ultraspherical polynomials by using the estimate
\begin{equation*}
P_t^\lambda(x,y)\le \frac{C \sinh t}{(1-x^2)^{\lambda/2}(1-y^2)^{\lambda/2}(\cosh t-xy-\sqrt{1-x^2}\sqrt{1-y^2})},
\end{equation*}
which follows from the elementary inequality
\[
\int_{-1}^1 \frac{(1-s^2)^{\lambda-1}}{(A-Bs)^{\lambda+1}}\, ds\le \frac{C}{B^\lambda(A-B)}, \qquad A>B>0, \quad \lambda>0,
\]
and the mean value theorem.
Indeed, taking $C_f=\max\{|f'(x)|:x\in [-1,1]\}$ and using the inequality $1-xy-\sqrt{1-x^2}\sqrt{1-y^2}\ge C|x-y|^2$, we have
\begin{multline*}
\int_{0}^{\infty}\int_{-1}^{1} P_t^\lambda(x,y)|f(x)-f(y)|\, d\mu_\lambda(y)\left(\sinh t/2\right)^{-\sigma-1}\,dt\\\le \frac{C_f}{(1-x^2)^{\lambda/2}} \left(C_1\int_0^1\int_{-1}^1 \frac{t^{-\sigma}|x-y|}{t^2+|x-y|^2}(1-y^2)^{\lambda/2-1/2}\, dy \, dt\right.\\\left.+ C_2\int_1^\infty\int_{-1}^1 e^{-(\sigma+1)t/2}|x-y|(1-y^2)^{\lambda/2-1/2}\, dy \, dt\right)=: \frac{C_f}{(1-x^2)^{\lambda/2}} ( I_1+I_2).
\end{multline*}
Obviously, $I_2$ is a finite integral. For $I_1$ the change of variable $t=|x-y|s$ gives
\[
I_1\le C_1 \int_0^\infty\frac{s^{-\sigma}}{s^2+1}\, ds\int_{-1}^1 |x-y|^{-\sigma}(1-y^2)^{\lambda/2-1/2}\, dy<\infty.
\]
To obtain the expression of $K_\sigma^\lambda$ we observe that
\begin{multline*}
K_\sigma^\lambda(x,y)\\
\begin{aligned}
&=\frac{c_\lambda^2}{2^{\lambda+1+\sigma}|\Gamma(-\sigma)|}\int_{0}^{\infty}\int_{-1}^{1}\frac{\sinh t}{(\cosh t -w(s))^{\lambda+1}}\,d\mu_{\lambda-1/2}(s)\left(\sinh t/2\right)^{-\sigma-1}\,dt
\\ &=
\frac{c_\lambda^2}{2^{\lambda+(1+\sigma)/2}}
\frac{\Gamma(\frac{1-\sigma}{2})\Gamma(\lambda+\frac{1+\sigma}{2})}{|\Gamma(-\sigma)|\Gamma(\lambda+1)}
\int_{-1}^{1}\frac{d\mu_{\lambda-1/2}(s)}{(1-w(s))^{\lambda+(1+\sigma)/2}},
\end{aligned}
\end{multline*}
where we have applied Fubini theorem and the change of variable $2(\sinh t/2)^2=z(1-w(s))$ in last equality. With the last identity we have concluded the proof of \eqref{eq:Lem01_principal}.
To prove \eqref{eq:esc-prod} we follow the argument in \cite[Lemma 5.1]{RT}. First, we observe that the kernel $K_\sigma^\lambda(x,y)$ is positive and symmetric in the sense that $K_\sigma^\lambda(x,y)=K_\sigma^\lambda(y,x)$. Then, \eqref{eq:esc-prod} is clear when $f$ is a finite linear combination of ultraspherical polynomials. For $f\in H_\lambda^\sigma$ we consider a sequence of finite linear combinations of ultraspherical polynomials $\{p_k\}_{k\ge 0}$ such that $p_k$ converges to $f$ in $H_\lambda^\sigma$. Then, by using the definition of $A_\sigma^\lambda$, it is clear that $\langle A_\sigma^\lambda p_k, p_k\rangle_{\lambda}$ converges to $\langle A_\sigma^\lambda f, f\rangle_{\lambda}$. Moreover, the result for polynomial functions implies
\begin{multline}
\label{eq:esc-prod-pk}
\langle A_\sigma^\lambda p_k, p_k\rangle_{\lambda}=\frac{1}{2}\int_{-1}^1\int_{-1}^1 (p_k(x)-p_k(y))^2 K_\sigma^\lambda (x,y) \, d\mu_\lambda(y)\, d \mu_\lambda(x)\\+E_{\sigma,\lambda}\langle p_k, p_k \rangle_\lambda<\infty.
\end{multline}
Consequently, the functions $P_k(x,y)=p_k(x)-p_k(y)$ form a Cauchy sequence in $L^2((-1,1)\times (-1,1), d\omega)$ where $d\omega(x,y)=K_\sigma^\lambda(x,y)\, d\mu_\lambda(x)\, d\mu_\lambda(y)$ which converges to $f(x)-f(y)$ in this norm. Hence, passing to the limit in \eqref{eq:esc-prod-pk}, we complete the proof of the lemma.
\end{proof}
\begin{Lem}\label{lemma02}
Let $\lambda>0$ and $2\lambda+1>\sigma>0$. Then
\begin{equation}\label{eq:Lem02_principal}
A_\sigma^{\lambda}\left(\frac{1}{(1-x^2)^{\lambda/2+(1-\sigma)/4}}\right)=
\frac{Q_{\sigma,\lambda}}{(1-x^2)^{\lambda/2+(1+\sigma)/4}},
\end{equation}
where $Q_{\sigma,\lambda}$ is the constant given in \eqref{ec:cons-Hardy}.
\end{Lem}
\begin{proof}
First of all, we have to realize that the ultraspherical polynomial $C_{n}^\lambda(x)$ is odd for $n=2m+1$, $m\in\mathbb{Z}^{+}$; therefore, for $\beta>0$, the function $(1-x^2)^{\beta-1} C_{2m+1}^\lambda(x)$ is an odd function and its integral over the interval $(-1,1)$ is zero. For $n=2m$ we use \cite[eq. 15, p. 519]{PrudnikovII} to obtain
\begin{equation*}
\begin{split}
\int_{-1}^{1}&(1-x^2)^{\beta-1}C_{2m}^\lambda(x)\,dx
\\ &=
\sqrt{\pi}\frac{(2\lambda)_{2m}}{(2m)!}\frac{\Gamma(\beta)}{\Gamma(\beta+1/2)}
{}_3F_2(-2m,2\lambda+2m,\beta;2\beta,\lambda+1/2;1)
\\ &=
\pi\frac{(2\lambda)_{2m}}{(2m)!}
\frac{\Gamma(\beta)\Gamma(\lambda+1/2)\Gamma(\beta-\lambda+1/2)}
{\Gamma(1/2-m)\Gamma(\lambda+m+1/2)\Gamma(\beta+m+1/2)\Gamma(\beta-\lambda-m+1/2)},
\end{split}
\end{equation*}
where in last identity we have evaluated the hypergeometric function with the so-called Watson formula \cite[eq. 16.4.6, p. 406]{Olver}. Therefore, if we denote $\alpha=\lambda/2+(1-\sigma)/4$, we obtain that
\begin{equation}\label{eq:Lem02_key}
\int_{-1}^{1}(1-x^2)^{\alpha-1}C_{2m}^\lambda(x)\,dx=
R_{\sigma,\lambda}\int_{-1}^{1}(1-x^2)^{\alpha+\sigma/2-1}C_{2m}^{\lambda}(x)\,dx,
\end{equation}
with
\begin{multline*}
R_{\sigma,\lambda}=
\frac{\Gamma(\alpha)\Gamma(\alpha-\lambda+1/2)}{\Gamma(\alpha+\sigma/2)\Gamma(\alpha-\lambda+1/2+\sigma/2)}\\\times
\frac{\Gamma(\alpha+m+1/2+\sigma/2)\Gamma(\alpha-\lambda-m+1/2+\sigma/2)}{
\Gamma(\alpha+m+1/2)\Gamma(\alpha-\lambda-m+1/2)}.
\end{multline*}
In this way, if we prove the identity
\begin{equation}\label{eq:Lem02_relation}
R_{\sigma,\lambda}=Q_{\sigma,\lambda}^{-1}\frac{\Gamma(2m+2\alpha+\sigma)}{\Gamma(2m+2\alpha)}
\end{equation}
we will conclude the proof, because \eqref{eq:Lem02_key} implies
\[
a_n^\lambda\left(\frac{1}{(1-x^2)^{\alpha+\sigma/2}}\right)
=Q_{\sigma,\lambda}^{-1}\frac{\Gamma(n+2\alpha+\sigma)}{\Gamma(n+2\alpha)}
a_n^\lambda\left(\frac{1}{(1-x^2)^\alpha}\right),
\]
where we have had in mind that the $n$-th Fourier coefficient is null when $n=2m+1$.
Let us check that \eqref{eq:Lem02_relation} actually holds. Using the reflection formula \cite[eq. 6.1.17, p. 256]{AS} twice we have
\begin{align*}
\frac{\Gamma(\alpha-\lambda-m+1/2+\sigma/2)}{\Gamma(\alpha-\lambda-m+1/2)}
&=
\frac{\Gamma(\alpha+m+\sigma/2)}{\Gamma(\alpha+m)}\frac{\sin(\pi(\alpha-\lambda-m+1/2))}
{\sin(\pi(\alpha-\lambda-m+1/2+\sigma/2))} \\
&=
\frac{\Gamma(\alpha+m+\sigma/2)}{\Gamma(\alpha+m)}
\frac{\Gamma(\alpha)\Gamma(\alpha-\lambda+1/2+\sigma/2)}{\Gamma(\alpha+\sigma/2)\Gamma(\alpha-\lambda+1/2)},
\end{align*}
and then
\begin{align*}
R_{\sigma,\lambda}
&=
\frac{\Gamma(\alpha)^2}{\Gamma(\alpha+\sigma/2)^2}
\frac{\Gamma(\alpha+m+\sigma/2)\Gamma(\alpha+m+\sigma/2+1/2)}{\Gamma(\alpha+m)\Gamma(\alpha+m+1/2)}
\\ &=
Q_{\sigma,\lambda}^{-1}\frac{\Gamma(2m+2\alpha+\sigma)}{\Gamma(2m+2\alpha)},
\end{align*}
by the duplication formula \cite[eq. 6.1.18, p. 256]{AS}.
\end{proof}
\section{Proof of Theorem \ref{theorem01}}
Polarizing the identity \eqref{eq:esc-prod} in Lemma \ref{lemma01} we obtain
\begin{equation}\label{eq:Thm01_key}
\langle g,A_\sigma^{\lambda} f\rangle_\lambda=\frac{1}{2}\int_{-1}^{1}\int_{-1}^{1}F(x,y)
K_\sigma^\lambda(x,y)\,d\mu_\lambda(y)\,d\mu_\lambda(x)+E_{\sigma,\lambda}\langle g,f\rangle_\lambda,
\end{equation}
with $F(x,y)=(g(x)-g(y))(f(x)-f(y))$.
Let us take $g(x)=(1-x^2)^{-\lambda/2-(1-\sigma)/4}$ and $f(x)=u^2(x)/g(x)$ for $u\in H_\lambda^\sigma$. Then
$$
F(x,y)=\left(u(x)-u(y)\right)^2-g(x)g(y)\left(\frac{u(x)}{g(x)}-\frac{u(y)}{g(y)}\right)^2
$$
and \eqref{eq:Thm01_key} becomes
\begin{multline*}
\langle g,A_\sigma^{\lambda} f\rangle_\lambda\\
=
\langle u,A_\sigma^{\lambda} u\rangle_\lambda-\frac{1}{2}\int_{-1}^{1}\int_{-1}^{1}g(x)g(y)\left(\frac{u(x)}{g(x)}-\frac{u(y)}{g(y)}\right)^2 K_\sigma^\lambda(x,y)\,d\mu_\lambda(y)\,d\mu_\lambda(x).
\end{multline*}
Now, by \eqref{eq:Lem02_principal}, we have
\[
\langle g,A_\sigma^{\lambda} f\rangle_\lambda
=
\langle A_\sigma^{\lambda} g,f\rangle_\lambda
=
Q_{\sigma,\lambda}\int_{-1}^{1}\frac{u^2(x)}{(1-x^2)^{\sigma/2}}\,d\mu_\lambda(x)
\]
and then we can deduce the ground state representation
\begin{multline}
\label{ec:ground}
\langle u,A_\sigma^{\lambda} u\rangle_\lambda-Q_{\sigma,\lambda}\int_{-1}^{1}\frac{u^2(x)}{(1-x^2)^{\sigma/2}}\,d\mu_\lambda(x)
\\=\frac{1}{2}\int_{-1}^{1}\int_{-1}^{1}g(x)g(y)\left(\frac{u(x)}{g(x)}-\frac{u(y)}{g(y)}\right)^2 K_\sigma^\lambda(x,y)\,d\mu_\lambda(y)\,d\mu_\lambda(x).
\end{multline}
So, due to the positivity of the kernel $K_{\sigma}^\lambda$, we conclude the proof.
|
\section*{}
\vspace{26pt}
\pagestyle{plain}
\setcounter{page}{1}
\vfill
\vspace{0.5cm}
\newpage
\tableofcontents
\newpage
\section{Introduction}
The topological recursion (TR) is a formalism developed by Eynard, Orantin \cite{EOFg,EORev} and Chekhov \cite{CEO06} which has in recent years found many applications in random matrices \cite{E1MM,BEO}, enumerative geometry \cite{BMconj,EMS,Ebook,ACEH}, intersection theory on the moduli space of curves \cite{Mirza1,EOwp,EInter}, integrable systems \cite{BEInt,MDHitchin,BeEMgenus}, topological strings \cite{BKMP,EOBKMP,MCCLiuBKMP}, quantum field theories \cite{BEknots,BESeifert,ABO1}, see \cite{EynardICM} for a recent overview. In its simplest version, it takes as input a spectral curve $\Sigma$ embedded as a Lagrangian in $(\mathbb{C} \times \mathbb{C},\mathrm{d} x \wedge \mathrm{d} y)$, and returns a collection of meromorphic forms $\omega_{g,n}$ defined on ${\rm Sym}^n \Sigma$, indexed by integers $g \geq 0$ and $n \geq 1$. It also returns scalars $F_{g} = \omega_{g,0}$ for $g \geq 0$, which enjoy a property of symplectic invariance \cite{EO2MM,EOxy}.
The proof of the symplectic invariance assumes $\Sigma$ is compact, the embedding algebraic, and it is a computational tour de force: it does not explain why this property is true and does not allow an easy generalization to weaker conditions. These applications hint at interpreting the topological recursion as a quantization procedure, but a thorough understanding of its underlying (symplectic) geometric nature is still incomplete.
Kontsevich and Soibelman \cite{KSTR} recently proposed a new point of view and setting for TR which generalizes the TR of \cite{EOFg}. We refer to it as KS-TR. Their starting point is the notion of classical Airy structure, \textit{i.e.} a Lagrangian defined by quadratic equations in a symplectic vector space $T^*V$. The initial data for KS-TR is a lift of the former to a sub-Lie algebra of the Weyl algebra of $V$, which they call a \emph{quantum Airy structure} (Definition~\ref{Def:QAiry}). Quantum Airy structures are equivalently determined by their coefficients, collected in four tensors $(A,B,C,D)$, which must satisfy relations \eqref{SymA}-\eqref{Drel} coming from the ``Lie subalgebra'' condition. The outcome of KS-TR is a formal function on $V$ of the form $$Z = \exp\big(\sum_{g \geq 0} \hbar^{g - 1}S_{g}\big)$$ annihilated by the differential operators determining the quantum Airy structure (Proposition~\ref{Propmain}). The $n$-th order Taylor coefficients $F_{g,n}$ of $S_{g}$ are computed as in TR by induction on $2g - 2 + n$ using $(A,B,C,D)$, and encode the same information as the $\omega_{g,n}$ did in TR. More precisely, the data of a spectral curve can be used to produce a quantum Airy structure, such that the $F_{g,n}$ computed by KS-TR are the coefficients of the decomposition of $\omega_{g,n}$ computed by TR in a suitable basis of meromorphic differentials (Section 3.5 in \cite{KSTR}, and Section 6.1 here).
Kontsevich and Soibelman emphasize in \cite{KSTR} the geometry of Lagrangians in $T^*V$ and the relations between KS-TR and deformation quantization. The present work is complementary to \cite{KSTR}. It focuses on the study of the relations defining quantum Airy structures, with the aim to exhibiting initial data for KS-TR.
\subsubsection*{Outline}
Let us summarize the content of the article.
In Section~\ref{S2} we concisely present the KS-TR formalism. We write down explicitly in Section~\ref{S22} the relations satisfied by $(A,B,C,D)$, for which we give a graphical interpretation as three coupled IHX-like relations (Figure~\ref{3RelIHX}). The existence of the partition function (Proposition~\ref{Propmain}) is proved in \cite{KSTR} by general holonomicity arguments. We prove it in Section~\ref{S24} by direct computations. Section~\ref{SS3} shows that the partition function can be explicitly computed when some of the tensors $(A,B,C)$ are zero.
In Section~\ref{SS3bis}, we give an equivalent characterization of classical and quantum Airy structures in terms of ``torsion free'' symplectic representations of Lie algebras $\rho_{1}\,;\,V \rightarrow \mathfrak{sp}(T^*V)$, together with the data of a Lagrangian linear embedding $\mathcal{I}\,;\,V \rightarrow T^*V$. As a result, we describe in Section~\ref{SS4} an action of the group of at most quadratic differential operators on quantum Airy structures (the analog of a gauge group) and their partition function. Therefore, we are especially interested in quantum Airy structure modulo the action of this group. One can define in this way the moduli space of quantum Airy structures (Section~\ref{SModspace}), and deformation theory of quantum Airy structures is governed by twisted Lie algebra cohomology. We also define (Section~\ref{S34}) an action of commuting flows corresponding to translations in $V$. It means that, from a given quantum Airy structure $(L_i)_{i \in I}$, we can obtain a deformed Airy structure $(L_i^{(t)})_{i \in I}$ parametrized by $t$ in a formal neighborhood of $0$ in $V$. The action of translation is non-linear even at the infinitesimal level, thus non-trivial modulo the gauge group action.
The remaining of the paper is devoted to exhibiting examples of quantum Airy structures. In Section~\ref{SS5}, we study general properties of the above symplectic representations, and apply them to prove some results aiming towards a of classification finite-dimensional quantum Airy structures forming semi-simple Lie algebras. In particular, representation theory allows us to construct a quantum Airy structure forming the Lie algebra $\mathfrak{sl}_{2}(\mathbb{C})$. In Section~\ref{S5}, we give a classification of abelian quantum Airy structures in dimension two and three, a partial classification of two-dimensional quantum Airy structures, and an example of a non-trivial quantum Airy structure for a non semi-simple three-dimensional Lie algebra.
We then progress towards more geometric examples. In Section~\ref{S4}, we describe four classes of quantum Airy structures, associated respectively to Frobenius algebras, non-commutative Frobenius algebras, the loop space of Frobenius algebras, and a $\mathbb{Z}_{2}$-invariant version of the latter. Our proposal for the Frobenius algebra class (Section~\ref{S4Frob}) satisfies rather trivially the axioms of a quantum Airy structure, and contains as a special case the enumeration of the trivalent graphs underlying TR and the partition functions of 2d TQFT's (Lemma~\ref{TQFTpart}). For the three other classes, checking that our proposal is a quantum Airy structure is a computation; it is perhaps not enlightening, but we have not found more elegant proofs. For the Frobenius and non-commutative Frobenius algebra class, we are able to give an explicit formula -- in the form of a finite-dimensional path integral, well-defined at the level of formal power series -- for the partition function in full generality (Section~\ref{S4NCFrob}). For the loop space of Frobenius algebras class (Section~\ref{S4Loop}-\ref{S444}), the partition function necessarily has $S_{0} = 0$, but $(S_{g})_{g \geq 1}$ can be non-trivial. For the $\mathbb{Z}_{2}$-invariant version, a priori all $S_{g}$ can be non-trivial.
The class of quantum Airy structures we describe in Proposition~\ref{ThmLoopZ2TQFT} for $\mathbb{Z}_{2}$-invariant loop spaces of Frobenius algebras are in correspondence with local spectral curves, and KS-TR gets identified with TR in this case. Section~\ref{comparTR} explains this correspondence to TR in more detail. In Section~\ref{SJUING}, we explain how the gauge group action on quantum Airy structures relates to Givental group action on Lagrangian cones. This puts our understanding the correspondence between TR and correlation functions of semi-simple cohomological field theories established in \cite{DBOSS} within the original spirit of Givental quantization procedure \cite{GiventalQuad}. Independently, Section~\ref{SYoung} interpretes the recursion for quantum Airy structure on loop spaces as a dynamic on Young diagrams.
We conclude in Section~\ref{SConclu} with a list of open problems raised throughout the article.
\subsubsection*{Comment}
We stress that KS-TR is not only a reformulation of TR. It comes with new non-trivial examples of initial data, \textit{e.g.} having a finite-dimensional $V$ (Sections~\ref{S4Frob}-\ref{S4NCFrob}), and the case where $V$ is infinite-dimensional and attached to a curve without reference to a local involution (Proposition~\ref{ThmLoopTQFT}). The latter may be used to propose a TR for spectral curves without ramification points, see Section~\ref{SWithout}. Although the motivations mainly come from geometry, KS-TR can be presented only resorting to multilinear algebra and combinatorics, without complex analysis. The beginners or non-geometers interested in the theory of topological recursion -- \textit{e.g.} for the enumeration of maps \cite{Ebook} -- may find the simplicity of this new framework (concentrated in Section~\ref{S2}) appealing.
\subsubsection*{Acknowledgments}
We thank M.~Kontsevich and Y.~Soibelman for communicating preliminary versions of their work with us, P.~Biane, M.~Shapiro, P.~Teichner and F.~Wagemann for discussions, M.~Karev and D.~Noshchenko for comments. We also thank the organizers of the AMS Symposium \emph{Topological recursion and its applications}, Charlotte in July 2016 -- where this collaboration was initiated -- and the organizers of the thematic month on \emph{Topological recursion and modularity} at the Matrix, Creswick in December 2016 -- where a preliminary version of this work was presented. J.E.A. is funded in part by the Danish National Research Foundation grant DNRF95 (Centre for Quantum Geometry of Moduli Spaces, QGM). The work of G.B. is supported by the Max Planck Gesellschaft. The work of L.C. was supported by the Russian Foundation for Basic Research (Grant No. 15-01-99504a) and the ERC Advance Grant 291092 ``Exploring the Quantum Universe'' (EQU).
\section{Kontsevich-Soibelman approach to topological recursion}
\label{S2}
\subsection{Setting}
\label{S21}
Let $V$ be a vector space over $\mathbb{C}$. It could be finite or infinite-dimensional. We will mostly work in a basis $(e_i)_{i \in I}$ of $V$, and with its dual basis $(x_i)_{i \in I}$ which form a set of linear coordinates on $V$. In the cases where $\dim V = \infty$, convergence issues will not play a role in this article. For general discussions it will be implicitly assumed that all seemingly infinite sums are actually finite or make sense after introducing if necessary suitable filtrations or completions. For specific examples where $\dim\,V = \infty$ we will justify that the sums contain only finitely many non-zero terms. We equip $T^*V$ with its canonical symplectic structure, and consider its Weyl algebra
$$
\mathcal{W}_{V}^{\hbar} = \mathbb{C}[\hbar]\big\langle (x_i,\partial_{i})_{i \in I})\big\rangle/\langle[\partial_i,x_i] = \hbar \rangle.
$$
Kontsevich and Soibelman \cite{Soibeltalk,KSTR} proposed the following setting, motivated by the problem of quantization of Lagrangians in $T^*V$ defined by quadratic equations.
Let $(x_i)_{i \in I}$ be the linear coordinates on $V$. By convention, $i,j,k,\ldots$ are fixed indices, while indices $a,b,c,\ldots$ should be summed over $I$. For instance, $A^i_{a,b}x_ax_b := \sum_{a,b \in I} A^i_{a,b}x_ax_b$.
\begin{definition}
\label{Def:QAiry} A \emph{quantum Airy structure} on $V$ is a sequence $(L_i)_{i \in I}$ of elements of $\mathcal{W}_{V}^{\hbar}$ of the form
\begin{equation}
\label{Lform} L_i = \hbar \partial_{i} - \tfrac{1}{2}A^i_{a,b}x_ax_b - \hbar B^i_{a,b}x_a\partial_{b} - \tfrac{\hbar^2}{2}C^i_{a,b}\partial_a\partial_b - \hbar D^i,
\end{equation}
where $\hbar$ is a formal parameter and $A^i_{j,k},B^i_{j,k},C^i_{j,k}$ and $D^i$ are scalars, which form a Lie subalgebra of $\mathcal{W}_{V}^{\hbar}$, \textit{i.e.}
\begin{equation}
\label{Liealg} [L_i,L_j] = \hbar\,f_{i,j}^a L_a
\end{equation}
for some scalars $f_{i,j}^k$.
\end{definition}
In this definition, we can always assume that $A^i_{j,k} = A^i_{k,j}$ and $C^i_{j,k} = C^i_{k,j}$. The coefficients defining a quantum Airy structure can be rearranged in a basis-free way
\begin{equation}
\label{tensorini} A \in {\rm Hom}(V^{\otimes 3},\mathbb{C}),\qquad B \in {\rm Hom}(V^{\otimes 2},V),\qquad C \in {\rm Hom}(V,V^{\otimes 2}),\qquad D \in {\rm Hom}(V,\mathbb{C}),
\end{equation}
by the assignments
$$
A(e_i \otimes e_j \otimes e_k) = A^i_{j,k},\qquad B(e_i \otimes e_j) = B^i_{j,a}e_{a},\qquad C(e_i) = C^i_{a,b}e_{a} \otimes e_{b},\qquad D(e_i) = D^i.
$$
Equation \eqref{Liealg} puts strong constraints on $A,B,C,D$. They will be studied in Section~\ref{S22} in a pedestrian way, and in Section~\ref{S3} in a more abstract way. We remark that for any choice of $A$ and $D$, $(A,B = C = 0,D)$ defines a (rather trivial) quantum Airy structure. The justification for the name ``Airy structures'' will appear in the examples provided in Section~\ref{S4Frob}. The notion of classical Airy structure will only be presented in Section~\ref{SSymp}, as it does not play a central role here although it served as motivation in \cite{KSTR}.
The Weyl algebra $\mathcal{W}_{V}^{\hbar}$ naturally acts by differential operators on functions on $V$. Equation \eqref{Liealg} is a sufficient condition for the existence of a function $Z$ on $V$ which is a common solution to $L_i\cdot Z = 0$ for all $i \in I$. More precisely, we have
\begin{proposition}
\label{Propmain}There exists a unique formal series
\begin{equation}
\label{Zpart} Z = \exp\bigg(\sum_{\substack{g \geq 0 \\ n \geq 1}} \frac{\hbar^{g - 1}}{n!} \sum_{i_1,\ldots,i_n \in I} F_{g,n}(i_1,\ldots,i_n)\,x_{i_1}\cdots x_{i_n}\bigg),
\end{equation}
where $F_{g,n}(i_1,\ldots,i_n)$ are scalars, invariant under permutation of the $(i_m)_{m = 1}^n$, such that $F_{0,1}(i) = F_{0,2}(i,j) = 0$ for all $i,j$, and
$$
\forall i\qquad L_i\cdot Z = 0.
$$
More precisely,
\begin{equation}
\label{F03} F_{0,3}(i,j,k) = A^i_{j,k},\qquad F_{1,1}(i) = D^i,
\end{equation}
and for $2g - 2 + n \geq 2$
\begin{eqnarray}
\label{TRForm} F_{g,n}(i_1,J) & = & \sum_{m = 2}^n B^{i_1}_{i_m,a} F_{g,n - 1}(a,J \setminus \{i_m\}) \\
&& + \tfrac{1}{2}C^{i_1}_{a,b}\bigg(F_{g - 1,n + 1}(a,b,J) + \sum_{\substack{J' \sqcup J'' = I \\ h' + h'' = g}} F_{h',1 + |J'|}(a,J') F_{h'',1 + |J''|}(b,J'')\bigg), \nonumber
\end{eqnarray}
where $J = \{i_2,\ldots,i_n\}$ is a $(n-1)$-uple of indices in $I$.
\end{proposition}
\noindent\textbf{Proof.} The uniqueness is obvious. We take $i_1 \in I$, insert \eqref{Zpart} into the equation $L_{i_1}\cdot Z = 0$, and for each $g \geq 0$, $n \geq 1$ and $(i_2,\ldots,i_n) \in I^{n - 1}$ we collect the coefficient of $\hbar^{g - 1}x_{i_2}\cdots x_{i_n}$. The equations for $(g,n) = (0,1),(0,2),(0,3)$ are
\begin{eqnarray}
0 & = & -F_{0,1}(i_1) + C_{ab}^i F_{0,1}(a)F_{0,1}(b), \nonumber \\
0 & = & - F_{0,2}(i_1,i_2) + B_{i_2,a}^{i_1} F_{0,1}(a) + C^{i_1}_{a,b} F_{0,1}(a) F_{0,2}(b,i_2), \nonumber \\
0 & = & - F_{0,3}(i_1,i_2,i_3) + A_{i_2,i_3}^{i_1} + \big(B_{i_2,a}^{i_1} F_{0,2}(a,i_3) + B_{k,a}^{i_1} F_{0,2}(a,i_2)\big) \nonumber \\
&& +\tfrac{1}{2} C^{i_1}_{a,b}\big(F_{0,1}(a) F_{0,3}(b,i_2,k) + F_{0,2}(a,i_2)F_{0,2}(b,i_3)\big).
\end{eqnarray}
As we take $F_{0,1}(i) = F_{0,2}(i,j) = 0$, we indeed solve the two first equations. The third equation yields $F_{0,3}(i_1,i_2,i_3) = A^{i_1}_{i_2,i_3}$. For $(g,n) = (1,1)$ we find $F_{1,1}(i_1) = D^{i_1}$. In general, isolating the term $F_{g,n}(i_1,\ldots,i_n)$ readily gives \eqref{TRForm}. In this equation, the order of the indices in the set $J'$ and $J''$ do not matter as the $F_{g',n'}(j_1,\ldots,j_n)$ were assumed symmetric under permutation of $(j_m)_{m = 1}^n$.
Conversely, the existence is guaranteed by the constraints \eqref{Liealg}. Alternatively, and effectively, we can define $F_{g,n}$ by formula \eqref{TRForm} inductively on $2g - 2 + n$, provided we justify that the result is symmetric when $i_1$ is permuted with the other $i_m$s. We will show this is true later in Proposition~\ref{Explicitsym}, by direct computations involving the relations between $(A,B,C,D)$ following from \eqref{Liealg}. We see for instance that the symmetry of $F_{0,3}$ imposes that $A^{i}_{j,k} = A^j_{i,k}$, hence $A$ must be fully symmetric in its three indices. It will indeed be a consequence (see Section~\ref{S31}) of \eqref{Liealg} for operators of the form \eqref{Lform}. \hfill $\Box$
Formula \eqref{TRForm} has a graphical interpretation (Figure~\ref{TRgraph}), which contains the same kind of terms as the topological recursion introduced in \cite{EOFg}. One can therefore propose, following Kontsevich and Soibelman, an elementary definition of the topological recursion.
\begin{itemize}
\item[$\bullet$] the initial data is a quantum Airy structure, \textit{i.e.} $(A,B,C,D)$ as in \eqref{tensorini} satisfying the relations we will write in Section~\ref{S2}.
\item[$\bullet$] the outcome are symmetric tensors $F_{g,n} \in {\rm Hom}(V^{\otimes n},\mathbb{C})$ indexed by $2g - 2 + n > 0$, which we consider as the Taylor expansion coefficients of a formal series/function on $V$, whose exponential is denoted $Z$ and called the partition function.
\end{itemize}
The topological recursion of \cite{EOFg} rather takes as initial data a spectral curve, \textit{i.e.} a simple branched covering $x\,:\,\Sigma \rightarrow \Sigma_0$ between two Riemann surfaces, together with a meromorphic $1$-form $\omega_{0,1}$ on $\Sigma$, and a fundamental meromorphic bidifferential of the second kind $\omega_{0,2}$ on $\Sigma^2$. We will show in Section~\ref{TRcomp} that this data determines a quantum Airy structure based on $V = H^0(U,K_U)/x^*H^0(U_0,K_{U_0})$ where $U_0$ is a (small enough) neighborhood of the ramification points (zeroes of $\mathrm{d} x$) in $\Sigma_0$, and $U := x^{-1}(U_0)$. The $F_{g,n}$ computed by \eqref{TRForm} are then coefficients of the decompositions of the meromorphic $n$-differentials $\omega_{g,n}$ defined by \cite{EOFg} in a suitable basis of meromorphic forms (Proposition~\ref{FWid}). Therefore, Kontsevich-Soibelman topological recursion can be seen as a generalisation of \cite{EOFg}.
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=\textwidth]{TRSurf.eps}
\caption{\label{TRgraph} $F_{g,n}(i_1,\ldots,i_n)$ is represented as a surface of genus $g$ with $n$ boundaries, carrying the labels $i_1,\ldots,i_n$. In this pictorial language, the terms appearing in the recursion \eqref{TRForm} are all the topologies resulting from removal of a pair of pant $P$ bounding the first boundary. The weight of $P$ is a $B$ or a $C$ depending on whether it has one or two external boundary components.}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{TRGraph.eps}
\caption{\label{TRGraph} Unfolding \eqref{TRForm} gives \cite[Section 3]{EORev} a set $\mathfrak{G}_{g,n}(1)$ of pairs $(G,T)$ where $G$ is a trivalent graph with first Betti number $g$ and $n$ leaves, with cyclic order at each vertex, and $T$ is a spanning tree rooted at the first leaf, having the extra property that edges which are not in $T$ must connect vertices $v$ and $v'$ which are \emph{parent}. This means that the geodesic in $T$ from the root to $v$ contains (or is contained) in the geodesic from the root to $v'$. Vertices incident to a loop are assigned a $D$, vertices incident to one external leg are $B$s, vertices incident to two external legs are $A$s. Internal vertices can be $A,B,C$ as prescribed by the recursive construction of the graph -- which is remembered by the spanning tree rooted at the first leg. We have listed these graphs for low values of $(g,n)$. The $\tfrac{1}{2^p}$ is the symmetry factor which arises from the repetition of factor of $\tfrac{1}{2}$ in the $C$-term of \eqref{TRForm}. A $(k)$ indicates that there are $k$ such graphs, which differ by the labeling $2,\ldots,n$ of the legs. When two such graphs are related by an exchange of the two legs outgoing from a $C$, the two graphs give the same contribution to $F_{g,n}$, and we listed it as a single graph with a factor of $\tfrac{1}{2}$ less.}
\end{figure}
\end{center}
\subsection{The relations between $(A,B,C,D)$}
\label{S22}
Let $L := (L_i)_{i \in I}$ be differential operators of the form \eqref{Lform}. We now describe the necessary and sufficient conditions on $(A,B,C,D)$ for $L$ to be a quantum Airy structure, \textit{i.e.} satisfy \eqref{Liealg}. Evaluating the commutator between the first terms with a pure single derivative and the $B$-terms we again obtain terms with pure single derivatives. Because this commutator is the only source of such terms in the right-hand side, we immediately conclude that the structure constants $f^k_{i,j}$ are determined by the $B$-terms alone
\begin{equation}
\forall i,j,k,\qquad f^k_{i,j}=B^i_{j,k}-B^j_{i,k}.
\label{f2B}
\end{equation}
Evaluating now the commutator between $L_i$ and $L_j$ and comparing with the right-hand side in Equation~\eqref{Liealg} we obtain further constraints on $(A,B,C,D)$. First, the absence of a linear term in $x_i$ immediately implies the full symmetry of the coefficients $A$
\begin{equation}
\forall i,j,k,\qquad A^i_{j,k}=A^j_{i,k},
\label{SymA}
\end{equation}
as anticipated for the symmetry of $F_{0,3}$ in \eqref{F03}. We obtain three more relations matching the coefficients of the terms $x_kx_\ell$, $\partial_{k}\partial_{\ell}$, $x_k\partial_{\ell}$, for any $i,j,k,\ell$
\begin{eqnarray}
\label{BBACeq} B^i_{j,a}B^a_{k,\ell} + B^i_{k,a}B^j_{a,\ell} + C^i_{\ell,a}A^j_{a,k} & = & (i \leftrightarrow j), \\
\label{BCeq} B^i_{j,a}C^a_{k,\ell} + C^i_{k,a}B^j_{a,\ell} + C^i_{\ell,a}B^j_{a,k} & = & (i \leftrightarrow j),\\
\label{BAeq} B^i_{j,a}A^a_{k,\ell} + B^i_{k,a}A^j_{a,\ell} + B^i_{\ell,a}A^j_{a,k} & = & (i \leftrightarrow j).
\end{eqnarray}
And matching the coefficient of $\hbar.1$ we find, for all $i,j$
\begin{equation}
\label{Drel} B^i_{j,a}D^a + \tfrac{1}{2}C^i_{a,b}A^j_{a,b} = (i \leftrightarrow j).
\end{equation}
Consequently we have the lemma
\begin{lemma}
$(L_i)_{i \in I}$ is a quantum Airy structure if and only if $(A,B,C,D)$ satisfy the zero torsion conditions \eqref{f2B}-\eqref{SymA}, the \textbf{BB-AC} relation \eqref{BBACeq}, the \textbf{BC} relation \eqref{BCeq}, the \textbf{BA} relation \eqref{BAeq}, and the \textbf{D} relation \eqref{Drel}. \hfill $\Box$
\end{lemma}
Equation \eqref{f2B} can be taken as a definition of the structure constants $f$, and one can check by direct computation that the above relations imply the Jacobi identity for $f$. The full symmetry of $A$ could be added to the axioms of quantum Airy structures. We call \eqref{f2B}-\eqref{SymA} ``zero torsion condition" for a reason explained in Section~\ref{S32}. The three relations for $(A,B,C)$ are rather non-trivial. If $d = {\rm dim}\,V$, let us count the number of unknowns and \textit{a priori} independent equations determining quantum Airy structures. $A$ has $\tfrac{d(d + 1)(d + 2)}{6}$ independent coefficients, $B$ has $d^3$ coefficients, and $C$ has $\frac{d^2(d + 1)}{2}$ coefficients. The \textbf{BB-AC} relation is antisymmetric in $i,j$, so gives $\tfrac{d(d - 1)}{2}\cdot d^2$ constraints. The \textbf{BC} and the \textbf{AC} relations are antisymmetric in $i,j$, symmetric in $k,\ell$, so gives $\frac{d(d - 1)}{2}\cdot \tfrac{d(d + 1)}{2}$ constraints. So, as far as $(A,B,C)$ are concerned, we have $\tfrac{d}{3}(5d^2 + 3d + 1)$ unknowns, and $\tfrac{d^2(d - 1)(2d + 1)}{2}$ constraints. The first values are
$$\begin{array}{|c||c|c|c|c|c|c|}
\hline
d & 1 & 2 & 3 & 4 & 5 & 6 \\
\hline
\hline
\#{\rm unknowns} & 3 & 18 & 55 & 124 & 235 & 398 \\
\hline
\#{\rm constraints} & 0 & 10 & 63 & 216 & 550 & 1170 \\
\hline
\end{array}
$$
For $d \geq 3$, we find that the three relations form an overdetermined system. Therefore, it is \textit{a priori} not obvious that non-zero solutions for $(A,B,C)$ can be found at all. If $(A,B,C)$ is a solution, the set of allowed $D$ satisfying \eqref{Drel} is an affine space, hence easier to describe. We will however show in Sections~\ref{S4} and \ref{S5} that many non-trivial solutions can be found.
\begin{lemma}\label{trL}
\label{basrem} Assume $(A,B,C)$ solves the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations, as well as \eqref{f2B} and \eqref{SymA}.
\begin{itemize}
\item[$(i)$] If $V$ is an abelian Lie algebra (namely $f_{i,j}^k = 0$), any choice of $D$ gives a quantum Airy structure.
\item[$(ii)$] If ${\rm tr}\,B^i := \sum_{a} B^i_{a,a}$ exists (for instance, it is always the case when $\dim V\,< \infty$), then
$$
D^i_{{\rm ref}} = \tfrac{1}{2}\,{\rm tr}\,B^i
$$
completes $(A,B,C)$ into a quantum Airy structure. Further, $(A,B,C,D)$ completes $(A,B,C)$ into a quantum Airy structure if and only if for any $v_1,v_2 \in V$, $(D - D_{{\rm ref}})([v_1,v_2]) = 0$ where we consider $D \in V^*$.
\end{itemize}
\end{lemma}
In particular, when $V$ is finite-dimensional and semi-simple, $[V,V] = V$ thus, for given $(A,B,C)$ satisfying the relations, there is a unique way to complete into a quantum Airy structure $(A,B,C,D)$.
\noindent \textbf{Proof.} The claim follows from the observation that, summing \eqref{BBACeq} over $k = \ell \in I$, we find
$$
f^i_{j,a} {\rm Tr}\,B^a + C^i_{a,b}A^j_{a,b} - C^j_{a,b}A^i_{a,b} = 0.
$$
\hfill $\Box$
\vspace{0.2cm}
Note that
$$
\hbar B^i_{a,b} x_a\partial_{b} = \tfrac{\hbar}{2} B^i_{a,b}(x_a\partial_b + \partial_b x_a) - \tfrac{\hbar}{2}\,{\rm tr}\,B^i
$$
when it makes sense. So, the solution exhibited in $(ii)$ corresponds to choosing the symmetric ordering of $x_a$ and $\partial_b$, rather than the normal order $x_a\partial_b$. When $V$ is infinite-dimensional, we will in Section~\ref{S4Loop} see cases where ${\rm tr}\,B^i$ is not well-defined, but we can nevertheless find other solutions for $D$.
\subsection{Graphical interpretation of the relations}
The structure of the indices $(i,j,k,l)$ and the summation index $a$ is the same in the three relations \textbf{BB-AC}, \textbf{BC} and \textbf{BA}. They can in fact be presented as a system of three IHX relations (Figure~\ref{3RelIHX}). Another graphical form is given in Figure~\ref{3REL}.
This IHX form is in fact not surprising. The original IHX relation is the graphical interpretation of the Jacobi relation for a Lie algebra $\mathfrak{l}$, and the Jacobi relation itself expresses that the adjoint representation $\mathfrak{l} \rightarrow {\rm GL}(\mathfrak{l})$ is a homomorphism of Lie algebras. It is a special case, for the adjoint representation, of the STU relation expressing that one has a representation $\mathfrak{l} \rightarrow {\rm End}(\mathfrak{m})$, where $\mathfrak{m}$ is a module for $\mathfrak{l}$. We will see in Section~\ref{S32} that the three relations for $(A,B,C)$ are equivalent to requiring that the adjoint action ${\rm ad}_{L}\,:\,V \rightarrow {\rm End}(T^*_{\hbar}V)$ is a representation of the Lie algebra $(V,(f_{i,j}^k)_{i,j,k})$. So, these 3 relations come from specializing the STU relation to this representation which has the special form $[L_i,\cdot]$ with $L_i$ of the form \eqref{Lform}.
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{IHXsym3.eps}
\caption{\label{3RelIHX} A red dot indicates a $B^{i_1}_{i_2,i_3}$ with incident edges labeling as indicated by the numbers. A blue dot represents an $A^{\bullet}_{\bullet\bullet}$. A green dot represents a $C^{i_1}_{\bullet\bullet}$, with first incident edge carrying the upper index. The three relations are that these combinations are symmetric with respect to permutation of the two left legs.}
\end{figure}
\end{center}
\vspace{0.5cm}
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{3Rel.eps}
\caption{\label{3REL} Each edge carries an index. Indices carried by dashed edge are summed over. $\bullet$ means symmetrisation of the indices of the edge incident at that vertex, while $\circ$ means antisymmetrization. The arrow indicates which index is placed first, and thus defines the order of composition of the two operators.}
\end{figure}
\end{center}
\subsection{Proof of symmetry}
\label{S24}
\begin{proposition}
\label{Explicitsym}
If $(L_i)_{i \in I}$ is a quantum Airy structure, $F_{g,n}(i_1,\ldots,i_n)$ defined recursively by \eqref{TRForm} for $2g - 2 + n > 0$ is symmetric under permutation of $(i_m)_{m = 1}^n$.
\end{proposition}
\begin{remark}
The converse of Proposition~\ref{Explicitsym} is not true. Indeed, if $A = 0$ and $D = 0$, one finds that $F_{g,n} = 0$ for all $g,n$, hence it is symmetric, whether or not $B$ and $C$ satisfy the -- still non-trivial -- relations \textbf{BB-AC} and \textbf{BC}.
\end{remark}
\noindent\textbf{Proof.} We already saw in Section~\ref{S21} that $F_{0,3}(i,j,k) = A^i_{j,k}$, which is fully symmetric due to \eqref{SymA}, and for $F_{1,1}$ there is nothing to check. We compute from \eqref{TRForm}
$$
F_{0,4}(i,j,k,\ell) = B_{j,a}^i A_{k,\ell}^{a} + {\rm cycl.}\,\,(j,k,\ell)
$$
which is fully symmetric thanks to \eqref{BAeq}, and
$$
F_{1,2}(i,j) = B_{j,a}^iD^a + \tfrac{1}{2}C^i_{a,b}A_{a,b}^j
$$
which is fully symmetric thanks to \eqref{Drel}. So the result holds for $2g - 2 + n \leq 2$. Take $(g,n)$ such that $2g -2 + n > 2$, and assume we have proved fully symmetry of $F_{g',n'}$ for all $2g' - 2 + n' \leq 2g - 2 + n$. Let $K = \{k_1,\ldots,k_n\}$. Let us define $F_{g,n + 2}(i,j,K)$ by applying \eqref{TRForm} with first index $i$. The resulting terms are in the range of the induction hypothesis, thus completely symmetric under permutation of $(j,k_1,\ldots,k_n)$. So, we only need to prove it is symmetric under permutation of $i$ and $j$. For this purpose, we use again \eqref{TRForm} with first index $j$, except for the term involving $B^i_{j,a}F_{g,n + 1}(a,K)$, for which we use \eqref{TRForm} with first index $a$. Denote $K^{[k]} := K \setminus\{k\}$ and $K^{[k,\ell]} := K\setminus\{k,\ell\}$. We also implicitly use the full symmetry of $A$ and the symmetry of $C$ in its two lower indices. We find that
\begin{eqnarray}
&& F_{g,n + 2}(i,j,K) \nonumber \\
& = & B_{j,a}^iF_{g,n + 1}(a,K) + \sum_{k \in K} B_{k,a}^iF_{g,n + 1}(a,j,K^{[k]}) + \tfrac{1}{2}C_{a,b}^i F_{g - 1,n + 3}(a,b,j,K) \nonumber \\
&& + \sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K}} C_{a,b}^i F_{h',2 + |K'|}(a,j,K') F_{h'',1 + |K''|}(b,K'') + \delta_{g,0}\delta_{n,1}A_{j,k_1}^i \nonumber \\
& = & B_{j,a}^i\bigg\{\sum_{k \in K} B_{k,b}^a F_{g,n}(b,K^{[k]}) + \tfrac{1}{2}C_{b,c}^aF_{g - 1,n + 2}(b,c,K) \nonumber \\
&& + \sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K}} \tfrac{1}{2}C_{b,c}^a F_{h',1+|K'|}(b,K')F_{h'',1 + |K''|}(c,K'') + \delta_{g,0}\delta_{n,2}A_{k_1,k_2}^a + \delta_{g,1}\delta_{n,0}D^a \bigg\} \nonumber \\
&& + \sum_{k \in K} B_{k,a}^i\bigg\{B_{a,b}^j F_{g,n}(b,K^{[k]}) + \sum_{\ell \in K^{[k]}} B_{\ell,b}^j F_{g,n}(b,a,K^{[k,\ell]}) + \tfrac{1}{2} C_{b,c}^j F_{g - 1, n + 2}(b,c,a,K^{[k]}) \nonumber \\
&& + \sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K^{[k]}}} \tfrac{1}{2} F_{h',2 + |K'|}(b,a,K') F_{h'',1 + |K''|}(c,K'') + \delta_{g,0}\delta_{n,2} A_{a,k}^j\bigg\} \nonumber \\
&& + \tfrac{1}{2}C_{a,b}^i\bigg\{B_{a,c}^jF_{g - 1,n + 2}(b,c,K) + B^j_{b,c}F_{g - 1,n + 2}(a,c,K) \nonumber \\
&& + \sum_{k \in K} B_{k,c}^j F_{g - 1,n + 2}(b,c,a,K^{[k]}) + \tfrac{1}{2}C_{c,d}^j F_{g - 2,n + 4}(c,d,a,b,K) \nonumber \\
&& + \sum_{\substack{h' + h'' = g - 1 \\ K' \sqcup K'' = K}} C_{c,d}^jF_{h',3+|K'|}(a,b,c,K')F_{h'',1 + |K''|}(d,K'') + C_{c,d}^j F_{h',2 + |K'|}(a,c,K')F_{h'',2 + |K''|}(b,d,K'') \nonumber \\
&& + \delta_{g,1}\delta_{n,0}A_{a,b}^j\bigg\} \nonumber \\
&& + \sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K}} C_{a,b}^i F_{h'',1 + |K''|}(b,K'')\bigg\{B_{a,c}^j F_{h',1 + |K'|}(c,K') + \sum_{k \in K'} B_{k,c}^j F_{h',1 + |K'|}(c,a,K^{[k]}) \nonumber \\
&& + \tfrac{1}{2}C_{c,d}^j F_{h' - 1,3 + |K'|}(c,d,a,K') + \sum_{\substack{s + s' = h' \\ L \sqcup L' = K'}} C_{c,d}^j F_{s,2 + |L|}(a,c,L)F_{s',1 + |L'|}(d,L') + \delta_{h',0}\delta_{|K'|,1}A_{a,k'}^j\bigg\}. \nonumber
\end{eqnarray}
We now collect the various terms
\begin{eqnarray}
&& F_{g,n + 2}(i,j,K) \nonumber \\
&= & \delta_{g,0}\delta_{n,2}\,F_{0,4}(i,j,k_1,k_2) + \delta_{g,1}\delta_{n,0} F_{1,2}(i,j) \nonumber \\
&& + \sum_{k \in K} \textcolor{red}{F_{g,n}(b,K^{[k]}) \Big(B_{j,a}^iB_{k,b}^a + B_{a,b}^jB_{k,a}^i + A_{a,k}^jC_{a,b}^i\Big)} + \sum_{k \neq \ell \in K} F_{g,n}(a,b,K^{[k,\ell]})\,B_{k,a}^i B_{\ell,b}^j \nonumber \\
&& + \textcolor{red}{\tfrac{1}{2} F_{g - 1,n + 2}(b,c,K)\big(B_{j,a}^iC_{b,c}^a + B_{a,c}^jC_{a,b}^i + B^j_{b,c}C^i_{a,b}\big)} + \sum_{k \in K} \tfrac{1}{2} F_{g - 1,n + 2}(b,c,a,K^{[k]})\big(B^i_{k,a}C^j_{b,c} + B^j_{k,c}C^i_{a,b}\big) \nonumber \\
&& + \tfrac{1}{4} F_{g - 2,n + 4}(a,b,c,d,K)\,C_{a,b}^i C_{c,d}^j + \textcolor{red}{\sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K}} \tfrac{1}{2} F_{h',1+|K'|}(b,K')F_{h'',1+|K''|}(c,K'')\big(C_{b,c}^aB_{j,a}^i + 2C_{a,b}^iB_{a,c}^j\big)} \nonumber \\
&&+ \sum_{k \in K} \sum_{\substack{h' + h'' = g \\ K' \sqcup K'' = K^{[k]}}} \tfrac{1}{2} F_{h',2+|K'|}(b,a,K')F_{h'',1 + |K''|}(c,K'')\big(B_{k,a}^iC_{b,c}^j + B^j_{k,c}C^i_{a,b}\big) \nonumber \\
&& + \sum_{\substack{h' + h'' = g - 1\\ K' \sqcup K'' = K}} \tfrac{1}{2}F_{h',3 + |K'|}(a,b,c,K') F_{h'',1 + |K''|}(d,K'')\big(C_{c,d}^jC_{a,b}^i + C_{a,d}^iC_{c,b}^j\big) \nonumber \\
&& + \sum_{\substack{h' + h'' = g - 1\\ K' \sqcup K'' = K}} \frac{1}{2}F_{h',2 + |K'|}(a,c,K')F_{h'',2 + |K''|}(b,d,K'')\,C_{a,b}^iC_{c,d}^j \nonumber \\
&& + \sum_{\substack{h'' + s + s' = g \\ K'' \sqcup L \sqcup L' = K}} F_{h'',1 + |K''|}(b,K'')F_{s,2 + |L|}(a,c,L)F_{s,1 + |L'|}(d,L')C_{c,d}^j C_{a,b}^i. \nonumber
\end{eqnarray}
We already proved that $F_{0,4}$ and $F_{1,2}$ are symmetric, and all the non-red terms are obviously symmetric by permutation of $i$ and $j$. The three relations \eqref{BBACeq}-\eqref{BCeq}-\eqref{BAeq} allow the conclusion that the red terms are also symmetric. So, $F_{g,n}(i,j,K)$ is fully symmetric, and by induction this entails the claim. \hfill $\Box$
\section{Lie algebraic approach to Airy structures}
\label{SS3bis}
\subsection{Reformulation \textit{via} the adjoint representation}
\label{S32}
In $\mathcal{W}_{V}^{\hbar}$, we have two notions of degree: the $\hbar$-degree, and the variable degree assigning degree $1$ to $x_i$ and $\partial_{x_i}$. We denote $T^*_{\hbar}V = V^* \oplus V.\hbar$ and
consider the subspace $\mathcal{D}_{V} \subset \mathcal{W}_{V}^{\hbar}$
$$
\mathcal{D}_{V} = \bigoplus_{d = 0}^{2} \mathcal{D}_{V,d},\qquad \mathcal{D}_{V,0} = \mathbb{C}.\hbar,\qquad \mathcal{D}_{V,1} = T^*_{\hbar}V,\qquad \mathcal{D}_{V,2} = {\rm Sym}^2 T^*_{\hbar}V.
$$
Note that the copy of $V^*$ in $\mathcal{D}_{V}$ contains the linear functions $x_i$, while the copy of $V$ correspond to differential operators $\partial_{i}$. $\mathcal{D}_{V}$ is naturally a sub-Lie algebra of $\mathcal{W}_{V}^{\hbar}$. Let $\pi_1$ and $\pi_1^*$ be the linear projections to the subspaces $V^*$ and $V.\hbar$ of $\mathcal{D}_{V,1}$. We are interested in linear maps $V \rightarrow \mathcal{D}_{V}$, which we parametrize -- after a choice of basis $(e_i)_{i \in I}$ of $V$ -- in the form
\begin{equation}
\label{Lpreform} L_i = M^i_{a} \hbar \partial_{a} + N^i_{a} x_a - \bigg(\tfrac{1}{2}A^i_{a,b}x_ax_b + \hbar B^i_{a,b}x_a\partial_{b} + \tfrac{\hbar^2}{2} C^i_{a,b}\partial_{a}\partial_{b} + \hbar D^i\bigg).
\end{equation}
\begin{definition}
We say that a linear map $L\,:\,V \rightarrow \mathcal{D}_V$ has normal form -- with respect to a choice of basis $(e_i)_{i \in I}$ of $V$ -- if $M^i_{j} = \delta_{i,j}$ and $N^i_{j} = 0$. If the basis is not specified, we mean that there exists a basis for which $L$ has normal form.
\end{definition}
\begin{definition}
A \emph{quasi-Airy structure} on $V$ is the data of a Lie algebra structure on $V$, together with a homomorphism of Lie algebras $L\,:\,V \rightarrow \mathcal{D}_V$. In this case, we denote $\mathcal{Z}_{L}$ the space of solutions of
$$
\forall v \in V,\qquad L(v)\cdot Z = 0
$$
of the form
$$
Z = \exp\bigg(\sum_{g \geq 0,n \geq 1} \frac{\hbar^{g - 1}}{n!}\,\mathcal{F}_{g,n}\bigg),\qquad \mathcal{F}_{g,n} \in {\rm Sym}^n V^*.
$$
Note that the relations imposed by $[L_i,L_j] = f_{i,j}^a L_{a}$ are \textit{a priori} different than those described in Section~\ref{S22} due to the presence of $M$ and $N \neq 0$.
\end{definition}
Tautologically, quasi-Airy structures having normal form are quantum Airy structures.
Let $L\,:\,V \rightarrow \mathcal{D}_V$ be a quasi-Airy structure. In particular, $V$ has the structure of a Lie algebra, and we can consider the adjoint representation ${\rm ad}_{L}\,:\,V \rightarrow {\rm End}(\mathcal{D}_V)$ defined by ${\rm ad}_{L}(v) = \hbar^{-1}[L(v),\cdot]$. Computation with \eqref{Lpreform} shows that ${\rm ad}_{L}$ has a block decomposition with respect to $\mathcal{D}_{V} = \bigoplus_{d = 0}^{2} \mathcal{D}_{V,d}$
$$
{\rm ad}_{L} = \left(\begin{array}{ccc} 0 & \rho_{0,1} & \rho_{0,2} \\ 0 & \rho_1 & \rho_{1,2} \\ 0 & 0 & \rho_2 \end{array}\right).
$$
If we choose a basis of $V$, since $L$ has the form \eqref{Lpreform} the central block has a further decomposition with respect to $\mathcal{D}_{V,1} = V^* \oplus V.\hbar$ of the form
\begin{equation}
\label{rho1} \rho_1(e_i) = \left(\begin{array}{cc} -B^i & A^i \\ -C^i & (B^i)^{T} \end{array}\right),
\end{equation}
where $X^i = (X^i_{j,k})_{j,k}$ is considered as a matrix for $X \in \{A,B,C\}$. Note that \eqref{rho1} is the general form of a matrix $Y$ such that $YS$ is symmetric, $S$ being the symplectic transformation
$$
S = \left(\begin{array}{cc} 0 & - 1 \\ 1 & 0 \end{array}\right).
$$
We note that $\rho_1$ is a representation of $V$ on $T_{\hbar}^*V$ if and only if $[\rho_1(e^i),\rho_1(e^j)] = f_{i,j}^a \rho_1(e^a)$.
Further, with respect to the same decomposition of $\mathcal{D}_{V,1}$
\begin{equation}
\label{rho01ti} \rho_{0,1} = \big( M^i \quad N^i \big).
\end{equation}
We also define $\tilde{\rho}_{1}\,:\,V \rightarrow {\rm End}(\mathbb{C}.\hbar \oplus T^*_{\hbar}V)$ by
\begin{equation}
\label{rho1ti}\tilde{\rho}_{1} = \left(\begin{array}{cc} 0 & \rho_{0,1} \\ 0 & \rho_{1} \end{array}\right).
\end{equation}
An easy computation shows
\begin{lemma}
\label{LLL1}Let $L\,:\,V \rightarrow \mathcal{D}_V$ be a linear map of normal form \eqref{Lform}. Then $\rho_1$ defined by \eqref{rho1} is a representation of $V$ on $T^*_{\hbar}V$ if and only if the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations are satisfied.
\end{lemma}
\begin{lemma}
\label{LLLfdsf} Let $L\,:\,V \rightarrow \mathcal{D}_V$ be a linear map of normal form, such that $\rho_1$ is a representation of $V$ on $T^*_{\hbar}V$. The three properties below are equivalent.
\begin{itemize}
\item[$(i)$] $\tilde{\rho}_{1}$ is a representation of $V$ on $\mathbb{C}.\hbar \oplus T^*_{\hbar} V$.
\item[$(ii)$] the two extra relations \eqref{f2B} and \eqref{SymA} are satisfied.
\item[$(iii)$] if $\mathcal{I}\,:\,V \rightarrow T^*_{\hbar} V$ is the natural inclusion, the representation $\rho_{1}$ satisfies
\begin{equation}
\label{commuts}\forall v_1,v_2 \in V,\qquad \rho_{1}(v_1).\mathcal{I}(v_2) - \rho_{1}(v_2).\mathcal{I}(v_1) = \mathcal{I}([v_1,v_2]).
\end{equation}
\end{itemize}
\end{lemma}
The only property which may be missing in Lemma~\ref{LLLfdsf} for $L$ to be a quantum Airy structure is the \textbf{D} relation \eqref{Drel} -- compare however with Lemma~\ref{basrem}, $(ii)$. Indeed the $D^i$ contribute to the constant part of the operators, and thus are not seen at the level of the adjoint action. Condition $(iii)$ is very similar to a \emph{zero torsion condition} for connections in vector bundles, which is the reason why we adopted this name to refer to the linear relation \eqref{commuts}.
Remark that, if $(L_i)_{i \in I}$ are operators of normal form with $A = C = 0$, $\rho_1$ is a representation if and only if $\rho_1$ acting on $V$ is a representation -- given by the matrices $(B^i)^T$. In particular, if $(f_{i,j}^k)$ are the structure constants of a Lie algebra structure on $V$, the choice $B^i_{j,k} = \tfrac{1}{2}f_{i,j}^k$ makes $(V,\rho_1^{V})$ (where $\rho_1^{V}$ means $\rho_1$ acting on $V$) the adjoint representation, and $(V,\rho_{1}^{V^*})$ its dual. One then sees that the \textbf{BB-AC} relation is equivalent to the Jacobi relation that $f_{i,j}^k$ indeed satisfy. In this case, $(L_i)_{i \in I}$ is the well-known expression of the adjoint representation by differential operators. It is however not a quantum Airy structure because the first -- and very important -- term $\hbar\partial_i$ is missing. Or, equivalently, this representation $\rho_{1}$ in general violates the zero torsion condition \eqref{commuts}.
\subsection{Classical Airy structures from symplectic representations}
\label{SSymp}
This paragraph is a digression to relate Section~\ref{S32} to the notion of \emph{classical Airy structure} introduced in \cite{KSTR}, which is the Poisson analog of Definition~\ref{Def:QAiry}. Let $\mathcal{P}_{V,d}$ be the space of polynomial functions on $T^*V$ of degree $d$, and
$$
\mathcal{P}_{V} = \bigoplus_{d = 0}^2 \mathcal{P}_{V,d}
$$
equipped with its canonical Poisson bracket. We denote $p_{V^*}\,:\,\mathcal{P}_{V} \rightarrow V^* \subset \mathcal{P}_{V,1}$ the natural projection.
\begin{definition}
\label{clAiry}A \emph{classical Airy structure} is the data of a linear map $\Lambda\,:\,V \rightarrow \mathcal{P}_{V,1} \oplus \mathcal{P}_{V,2}$ such that $I := p_{V^*}\circ \Lambda$ is an isomorphism, and such that ${\rm Im}\,\Lambda$ is closed under Poisson bracket.
\end{definition}
If $(e_i)_{i \in I}$ is a basis of $V$, we denote $(x_i)_{i \in I}$ and $(y_i)_{i \in I}$ the corresponding linear coordinates on $T^*V$. By defining $\Lambda_i:= I^{-1}_{i,a}\Lambda(e_a)$, a classical Airy structure is uniquely determined by a family $(\Lambda_i)_{i \in I}$ of elements of $\mathcal{P}_{V}$ of the form
$$
\Lambda_i = y_i - \tfrac{1}{2}A^i_{a,b}x_ax_b - B^i_{a,b}x_ay_b - \tfrac{1}{2}C^i_{a,b}y_ay_b,
$$
such that
\begin{equation}
\label{Poissoncom} \{\Lambda_i,\Lambda_j\} = f_{i,j}^a\Lambda_{a}
\end{equation}
for some scalars $f_{i,j}^k$. Note that \eqref{Poissoncom} implies that $\{(x,y) \in T^*V\,\,:\,\,\forall i\,\,\Lambda_i (x,y)= 0\}$ is a Lagrangian subvariety in $T^*V$. This Lagrangian for non-zero $(A,B,C)$ is a perturbation of the canonical Lagrangian given by the zero section $\{(x,y) \in T^*V\,\,:\,\,\forall i\quad y_i = 0\} \subset T^*V$.
The data of such $(\Lambda_i)_{i \in I}$ is the definition adopted by \cite{KSTR} for classical Airy structure. Here, we prefer to adopt the basis-independent Definition~\ref{clAiry}. The analysis we did for quantum Airy structures in the realm of the Weyl algebra can be transposed directly for classical Airy structures in the Poisson realm.
\begin{proposition}
\label{clasicProp}The data $(A,B,C)$ defines a classical Airy structure if and only if it satisfies \textbf{BB-AC}, \textbf{BC}, \textbf{BA} relations, as well as the zero torsion conditions \eqref{f2B}-\eqref{SymA}. This is also equivalent (Lemma~\ref{LLLfdsf}) to requiring that $\tilde{\varrho}_{1}$ is a representation of the Lie algebra $V$ defined by the structure constants $f_{i,j}^k$ in the basis $(e_i)_{i \in I}$. \hfill $\Box$
\end{proposition}
In other words, quantum Airy structures are just classical Airy structures together with the data $D$ satisfying the affine relation \eqref{Drel}.
This allows a more geometric perspective on Airy structures, which we are going to explain again but in a slightly different way. The keypoint, which is also put forward and exploited in \cite{KSTR}, is that suitable family of infinitesimal symplectomorphisms in $T^*V$ give rise to classical Airy structures.
Let us fix a Lie algebra structure on $V$. We then consider the vector space $W = V \oplus V^*$ with the canonical symplectic structure $\omega$. Let $ \pi_V : W \rightarrow V$ be the projection along $V^*$. We denote $\mathfrak{sp}(W)$ the Lie algebra of infinitesimal symplectomorphisms of $W$. Combining Proposition~\ref{clasicProp} with Lemma~\ref{LLLfdsf} we find
\begin{lemma}
There is a one to one correspondence between classical Airy structures, and Lie algebra homomorphisms $\mathcal{L}\,:\,V \rightarrow \mathfrak{sp}(T^*V)$ together with a Lagrangian embedding $\mathcal{I}\,:\,V \rightarrow T^*V$ such that
\begin{equation}\label{Con1}
\forall v_1,v_2 \in V,\qquad {\cal L}_{v_1} \mathcal{I}(v_2) - {\cal L}_{v_2}\mathcal{I}(v_1) = \mathcal{I}([v_1,v_2]).
\end{equation}
The coefficients of the classical Airy structure are given in terms of the symplectic form and the action on $W$ of the linear symplectomorphisms determined by $\mathcal{L}$ in \eqref{ABCs} below.
\end{lemma}
The advantage of this formulation is that given a Lie algebra homomorphism of $V$ into $\mathfrak{sp}(T^*V)$, for instance coming from geometry, then one only need to construct $\mathcal{I}$ such that the \emph{linear} condition \eqref{Con1} is satisfied. We also remark that the Lie algebra structure on $V$ is completely specified by the $\mathcal{L}$ \textit{via} \eqref{Con1}.
\noindent {\bf Proof}. First we will assume that we are given a Lie algebra homomorphism $\mathcal{L}\,:\,V \rightarrow \mathfrak{sp}(T^*V)$ as above, and from this we will construct a classical Airy structure. We pick a basis $(e_i)_{i \in I}$ of $V$. We let $e_i^*$ be the dual basis of $e_i$, then $((0,e_i^*),(e_i,0))_{i \in I}$ is a symplectic basis of $T^*V = W$, \textit{i.e.}
$$
\forall i,j,\qquad \omega(e_i,e_j) = \omega(e_i^*,e_j^*) = 0,\qquad \omega(e_i,e_j^*) = \delta_{i,j} .
$$
We let $f_{i,j}^k$ be the structure constants of $V$ in the basis $e_i$ of $V$. One then has the decomposition
$$
\forall w \in W,\qquad w = -x_a e_a + y_a e_a^*,\qquad x_i = \omega(e_i^*,w),\qquad y_i = \omega(e_i,w) .
$$
Denoting ${\cal L}_i := {\cal L}_{e_i}$, we have by definition of a symplectic representation
$$
\forall (w_1,w_2) \in W^2,\qquad \omega({\cal L}_i w_1,w_2) + \omega(w_1, {\cal L}_i w_2) = 0.
$$
Hence ${\cal L}_i$ is represented in the symplectic basis of $T^*V \cong V^* \oplus V$ by the matrix
$$
H_i = \left(\begin{array}{cc}
-B^i & A^i \cr
C^i & (B^i)^{T} \cr
\end{array}
\right)
$$
where $A^i$ and $C^i$ are symmetric matrices with
\begin{equation}
\label{ABCs} A^i_{j,k} = -\omega(\mathcal{L}_ie_j,e_k),\qquad B^i_{j,k} = \omega(\mathcal{L}_{i}e_j,e_k^*),\qquad C^i_{j,k} = -\omega(\mathcal{L}_{i}e_j^*,e_k^*).
\end{equation}
With this choice of $(A,B,C)$, one defines the hamiltonians
$$
h_i(w) := \tfrac{1}{2}\omega(\mathcal{L}_{i}w,w) = -\Big(\tfrac{1}{2} A_{a,b}^i x_a x_b + B_{a,b}^i x_a y_b + \tfrac{1}{2} C_{a,b}^i y_a y_b\Big)
$$
for $w = -x_a e_a + y_a e_a^*$. Then, we can define $\Lambda\,:\,e_i \rightarrow (w \mapsto \omega(e_i,w) + h_i(w))$, and we claim this is a classical Airy structure. Indeed, one can check using the commutation relations for $\mathcal{L}$ that $\{w \in W\,\,:\,\,\forall i\,\,\,\ \omega(e_i,w) + h_i(w) = 0\}$ is a Lagrangian subvariety of $T^*V$.
Conversely, a classical Airy structure gives the desired representation by just using \eqref{ABCs} to define ${\cal L}_i$ in terms of $(A,B,C)$. \hfill $\Box$
\section{Moduli spaces of Airy structures}
\label{SS4}
\subsection{Group action}
\label{S31}
The affine extended symplectic group $\mathcal{G}_{V} := \exp(\mathcal{D}_{V}/\hbar)$ acts by conjugation on its Lie algebra $\mathcal{D}_{V}$, hence inducing an action on the set of quasi-Airy structures, as well as on the partition function $Z$
$$
\mathcal{U} \in \mathcal{G}_{V},\qquad \tilde{L}_i = \mathcal{U}L_i\mathcal{U}^{-1},\qquad \tilde{Z} = \mathcal{U}\cdot Z.
$$
It contains -- and is generated by -- the Heisenberg subgroup $\exp(\mathcal{D}_{V,1}/\hbar)$, the metaplectic group $\exp(\mathcal{D}_{V,2}/\hbar)$, and an extra $\mathbb{C}^* = \exp(\mathcal{D}_{V,0}/\hbar)$ which has the role of a central extension.
This perspective makes it clear that we ought to study (quasi-)Airy structures only up to the action of ${\mathcal{G}}_{V}$.
Computations show that the subgroup of $\mathcal{G}_{V}$ which preserves the normal form \eqref{Lform} of quantum Airy structures only consists of multiplication by scalars (which are central and do not change the $L_i$), renormalization $(A,B,C,D) \rightarrow (\lambda^3 A,\lambda B,\lambda^{-1}C,\lambda D)$ for $\lambda \in \mathbb{C}^*$ (which can be realized via $\mathcal{U} = \lambda^{x_a\partial_{a}}$) and differential operators of order two
\begin{equation}
\label{Udiff2} \mathcal{U} = \exp\big( \tfrac{\hbar}{2} u_{a,b} \partial_{a}\partial_{b}\big),\qquad u_{i,j} = u_{j,i} \in \mathbb{C}.
\end{equation}
If $(A,B,C,D)$ are the coefficients of a quantum Airy structure, the new quantum Airy structure obtained by the action of \eqref{Udiff2} has coefficients $(\tilde{A},\tilde{B},\tilde{C},\tilde{D})$ given by
\begin{eqnarray}
\tilde{A}^i_{j,k} & = & A^i_{j,k}. \nonumber \\
\tilde{B}^i_{j,k} & = & B^i_{j,k} + u_{j,a}A^i_{a,k}. \nonumber \\
\tilde{C}^i_{j,k} & = & C^i_{j,k} + u_{j,a}B^i_{a,k} + u_{k,a}B^i_{j,a} + u_{j,a}u_{k,b}A^i_{a,b}. \nonumber \\
\label{SymU} \tilde{D}^i & = & D^i + \tfrac{1}{2}u_{a,b}A^i_{a,b}.
\end{eqnarray}
We have just proved
\begin{corollary}
If $(u_{i,j})_{i,j \in I}$ is a symmetric matrix, \eqref{SymU} is a symmetry of the relations of Section~\ref{S22}.
\end{corollary}
This can also be checked directly by inserting \eqref{SymU} in the relations. At the level of the partition functions, if ${\rm u} = (u_{i,j})_{i,j}$ is invertible, Wick's theorem shows that the action of \eqref{Udiff2} can be realized by a formal Gaussian convolution
$$
\exp\Big(\tfrac{\hbar}{2} u_{a,b} \partial_{a}\partial_{b} \Big)\cdot Z (x) = \int_{V} \frac{\mathbf{\mathrm{d}} \xi}{\det(2\pi \hbar\,{\rm id})^{1/2}}\,\exp\bigg(-\frac{({\rm u}^{-1})_{a,b}\xi_a\xi_b}{2\hbar}\bigg)\,Z(x + \xi),
$$
where $\mathbb{\mathrm{d}} \xi = \prod_{i \in I} \mathrm{d}\xi_i$ is the Lebesgue measure on $V$.
\begin{remark}
Another easy transformation of the $L_i$ is the rescaling of $\hbar$ acting as $L_i \rightarrow z^{-1}\exp(z\hbar\partial_{\hbar})L_i$. It transforms $(A,B,C,D)$ into $(z^{-1}A,B,zC,D)$. We prefer not to include it in $\mathcal{G}_{V}$.
\end{remark}
\begin{lemma}
\label{pretoAiry} If $L$ is a quasi-Airy structure, its $\mathcal{G}_{V}$-orbit contains a quantum Airy structure if and only if $(\pi_1 \oplus \pi_1^*) \circ L\,:\,V \rightarrow T^*_{\hbar}V$ is a (linear) Lagrangian embedding of $V$.
\end{lemma}
\noindent\textbf{Proof.} If $L = (L_i)_{i \in I}$ is a quantum Airy structure, then $\pi_1^* \circ L$ is the isomorphism between $V$ and $V^*$ induced by the choice of a basis $(e_i)_{i \in I}$. In particular, $(\pi_1 \oplus \pi_1^*) \circ L$ is a Lagrangian embedding of $V$ into $T^*_{\hbar}V$. These properties remain true for $\tilde{L}$ in the $\mathcal{G}_{V}$-orbit of $L$. Conversely, let $L$ be a quasi-Airy structure such that $(\pi_1 \oplus \pi_1^*) \circ L\,:\,V \rightarrow T^*_{\hbar}V$ is a Lagrangian embedding. We can always compose it with a linear symplectomorphism of $T_{\hbar}^*V$ bringing this Lagrangian to $V^*$. It is well-known that $\mathcal{G}_{V}$ contains elements which can realize as automorphisms of the Weyl algebra the linear symplectomorphisms
$$
\tilde{x}_i = \alpha_{i,a}x_a + \beta_{i,a} \hbar\partial_{a},\qquad \hbar\tilde{\partial}_i = \gamma_{i,a}x_{a} + \epsilon_{i,a}\hbar\partial_{a}
$$
for arbitrary matrices $(\alpha,\beta,\gamma,\varepsilon)$ satisfying the symplectic conditions
$$
\left\{\begin{array}{l} \alpha\beta^{T}\,\,{\rm symmetric,} \\ (\gamma\epsilon^{T})\,\,{\rm symmetric,} \\ \alpha\epsilon^{T} - \beta\gamma^{T} = {\rm id}_{V}. \end{array}\right.
$$
Therefore, we can find $\tilde{L}$ in the $\mathcal{G}_{V}$-orbit of $L$ such that $\pi_1 \circ L = 0$ and $\pi_1^* \circ L$ is an isomorphism. In other words, in a given basis, $\tilde{L}_i$ has the form \eqref{Lpreform} with $\tilde{N}^i_{j} = 0$ and ${\rm M} = (\tilde{M}^i_{j})_{i,j}$ invertible. So, putting $\check{L}_i = ({\rm M}^{-1})_{i,a}\tilde{L}_{a}$ gives an operator in normal form, \textit{i.e.} a quantum Airy structure. \hfill $\Box$
\begin{corollary}
A quasi-Airy structure $(L_i)_{i \in I}$ has a quantum Airy structure in its $\mathcal{G}_{V}$-orbit if and only if $\rho_{1}$ is a representation of the Lie algebra $V$ into $T_{\hbar}^*V$, and $\tilde{\rho}_{1}$ is a one-dimensional extension of this representation such that $v \mapsto (v,\rho_{0,1}(v))$ is a Lagrangian embedding of $V$ into $T^*_{\hbar}V$.
\end{corollary}
\textbf{Proof.} Comparing Lemmas~\ref{LLL1}-\ref{LLLfdsf} with the relations found in Section~\ref{S22} shows that $(L_i)_{i \in I}$ of normal form is a quantum Airy structure if and only if $\rho_1$ is a representation of $V$ in $T^*_{\hbar}V$ and $\tilde{\rho}_{1}$ is a one-dimensional extension of this representation such that ${\rho}_{0,1} = (\psi_{V*},0_{V})$ where $\psi_{V*}\,:\,V \rightarrow V^*$ is the isomorphism determined by the choice of basis in which $(L_i)_{i \in I}$ is defined. In this case, $\tilde{\rho}_{1}$ determines an exact sequence of $V$-modules:
$$
0 \longrightarrow \mathbb{C} \longrightarrow \mathbb{C}.\hbar \oplus T^*_{\hbar}V \longrightarrow (\mathbb{C}.\hbar \oplus T^*_{\hbar}V)/\mathbb{C} \longrightarrow 0,
$$
If $L_i \in \mathcal{D}$ does not have normal form, the block $\rho_{0,1}$ nevertheless gives the map $(\pi_1^* \oplus \pi_1) \circ L$, and the general claim is a consequence of Lemma~\ref{pretoAiry}. \hfill $\Box$
\subsection{Definition of moduli spaces}
\label{SModspace}
\label{S3def}
We shall now introduce various moduli spaces associated to Airy structures.
Let $V$ be a (finite-dimensional) Lie algebra. We denote $\mathfrak{A}_{V}^{{\rm cl}}$ (resp. $\mathfrak{A}_{V}$, $\mathfrak{A}_{V}^{{\rm q}}$) the set of classical (resp quasi-, quantum) Airy structures based on the Lie algebra $V$. Of course, each of them is a subset of the set of all Airy structures where we also vary the Lie algebra structure on the vector space $V$. However, we restrict here to study the set of Airy structures based on a fixed Lie algebra.
As a subset of $\mathbb{A}_{V} := ({\rm Sym}^3 V^*) \times (V^* \otimes V^* \otimes V) \times ({\rm Sym}^2 V^* \otimes V)$ cut out by the (finitely many) quadratic \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations and the linear relation \eqref{f2B} (we have included the relation \eqref{SymA} in the definition of $\mathbb{A}_{V}$), $\mathfrak{A}_{V}^{{\rm cl}}$ naturally has the structure of an affine algebraic variety. Likewise, $\mathfrak{A}_{V}$ is an affine algebraic variety. $\mathfrak{A}_{V}^{{\rm q}}$ is obviously a subvariety of $\mathfrak{A}_{V}$. It can also be seen as a subvariety of $V^* \times \mathbb{A}_{V}$ cut out by the extra \textbf{D} relation, where $D^i$ are the coordinates in the first factor. In fact, as the \textbf{D} relation is affine, it arises as the total space of an affine subbundle $\pi\,:\,\mathfrak{A}_{V}^{{\rm q}} \rightarrow \mathfrak{A}_{V}^{{\rm cl}}$ of the trivial vector bundle $V^* \times \mathbb{A}_{V} \rightarrow \mathbb{A}_{V}$ restricted to $\mathfrak{A}_{V}^{{\rm cl}}$. According to Lemma~\ref{basrem}, $\pi$ has a section given by
$$
D_{{\rm ref}}^i := \tfrac{1}{2}\,{\rm Tr}\,B^i.
$$
Mapping $D$ to $(D - D_{{\rm ref}})$ turns $\mathfrak{A}_{V}^{{\rm q}}$ into a trivial vector bundle over $\mathfrak{A}_{V}^{{\rm cl}}$, with fiber
$$
(V')^{\bot} = \big\{\varphi \in V^*\,\,:\,\,\forall x,y \in V,\quad \varphi([x,y]) = 0\big\}.
$$
As we saw in Section~\ref{S31}, $\mathcal{G}_{V}$ acts algebraically on $\mathfrak{A}_{V}^{{\rm cl}}$ and $\mathfrak{A}_{V}$, and its algebraic subgroup $\tilde{\mathcal{G}}_{V}$ preserving Airy structures acts on $\mathfrak{A}_{V}^{{\rm q}}$. Since we are interested in Airy structures up to the action of $\mathcal{G}_{V}$, the appropriate algebraic way to proceed would be to consider the quotient stack or the GIT quotients $\mathfrak{A}_{V}^{{\rm cl}}/\!/\mathcal{G}_{V}$ and $\mathfrak{A}_{V}/\!/\mathcal{G}_{V}$. In this paper we shall just consider the set-theoretic quotients
$\mathfrak{M}_{V}^{{\rm cl}} = \mathfrak{A}_{V}^{{\rm cl}}/\mathcal{G}_{V}$ and $\mathfrak{M}_{V}= \mathfrak{A}_{V}/\mathcal{G}_{V}$ with the induced topology from $\mathfrak{A}_{V}^{{\rm cl}}$ and $\mathfrak{A}_{V}$ respectively, and present some preliminary remarks about this quotient space, which we will call the moduli space of quasi-Airy structures. We caution the reader that this quotient space will in general not even be Hausdorff. The same comment applies to the moduli space of quantum Airy structures $\mathfrak{M}_{V}^{{\rm q}} = \mathfrak{A}_{V}^{{\rm q}}/\tilde{\mathcal{G}}_{V}$. The fibration $\pi\,:\,\mathfrak{A}_{V}^{{\rm q}} \rightarrow \mathfrak{A}_{V}^{{\rm cl}}$ is $\tilde{\mathcal{G}}_{V}$-equivariant, therefore we have a natural fibration $\mathfrak{M}_{V}^{{\rm q}} \rightarrow \mathfrak{M}_{V}$. The description of its fibers can in principle be obtained by looking at the action of $\mathcal{G}_{V}$ on $D$ via \eqref{SymU}.
\subsection{Deformation theory}
As the moduli space of classical Airy structures just consists of Lie algebra homomorphisms modulo inner automorphisms of the target Lie algebra $\mathcal{D}_{V}$, one can use the theory of deformation of algebraic structures to study their moduli space. The deformations of the Lie algebra homomorphism $L\,:\,V \rightarrow \mathcal{D}_{V}$ are governed by the differential graded algebra $E_{L}^{\bullet} := {\rm Hom}(\Lambda^{\bullet}V,\mathcal{D}_{V})$, which is equipped with the Cartan-Eilenberg differential
$$
(\mathrm{d}_{E_{L}^{\bullet}} \phi)(v_1\wedge \cdots v_{n + 1}) := \sum_{\ell = 1}^{n + 1} (-1)^{\ell + 1} \phi(v_1 \wedge \cdots \widehat{v_{\ell}} \cdots \wedge v_{n + 1}) L(v_{\ell}) + \sum_{\ell < m} (-1)^{\ell + m}\phi\big([v_k,v_{\ell}]_{V} \wedge v_1 \wedge \cdots \widehat{v_k} \cdots \widehat{v_{\ell}} \cdots \wedge v_{n + 1} \big)
$$
for $\varphi \in E_{L}^n$ and the bracket
$$
[\phi,\psi]_{E_{L}^{\bullet}}(v_1 \wedge \cdots \wedge v_{n + m}) := \sum_{\substack{\sigma \in \mathfrak{S}_{n + m} \\ \sigma(1) < \cdots < \sigma(m) \\ \sigma(m + 1) < \cdots < \sigma(m + n)}} \varepsilon(\sigma) \big[\phi(v_{\sigma(1)}\wedge \cdots \wedge v_{\sigma(m)}),\psi(v_{\sigma(m + 1)}\wedge \cdots \wedge v_{\sigma(m + n)})\big]_{\mathcal{D}}
$$
for $\phi \in E^m_{L}$ and $\psi \in E^n_{L}$ where $v_1\wedge \cdots \wedge \widehat{v_{\ell}} \wedge \cdots \wedge v_n$ stands for $v_1\wedge \cdots \wedge v_n$ with $v_{\ell}$ removed and $[\cdot,\cdot]_{X}$ stands for the bracket in $X$. We denote $H_{L}^i(V,\mathcal{D}_{V})$ the cohomology of this complex, and $Z_{L}^{n}(V,\mathcal{D}_{V})$ the space of $n$-cocycles. In particular we get a quadratic map
$$\Omega \,:\, H^1_L(V,{\mathcal D}_{V}) \rightarrow H^2_L(V,{\mathcal D}_{V}),$$
given by
$$\Omega(\phi) = [\phi,\phi]_{E^\bullet_L}.$$
We get the following proposition as a direct consequence of the results of \cite{Nijenhuis}.
\begin{proposition}
\label{THDEF1}Let $L\, :\, V \rightarrow \mathcal{D}_{V}$ be a classical Airy structure. If $H_{L}^1(V,\mathcal{D}_{V}) = 0$, then $L$ is rigid, \textit{i.e.} all continuous deformations of $L$ remain in its $\mathcal{G}_{V}$-orbit. In general there exists an open neighbourhood $U$ of zero in $H_{L}^1(V,\mathcal{D}_{V})$ and an open neighbourhood $\mathfrak{U}$ of $[L]\in \mathfrak{M}^{{\rm cl}}_{V}$ such that
$$ \mathfrak{U} \cong \left(U \cap \Omega^{-1}(0)\right)/\mathcal{G}_{V}(L)$$
where $\mathcal{G}_{V}(L)$ is the stabiliser of $L$ in $\mathcal{G}_{V}$.
\end{proposition}
\begin{corollary}
\label{THDEF2}If we have that $H^2_L(V,{\mathcal D}) = 0$ and that the action of $\mathcal{G}_{V}(L)$ on $H_{L}^1(V,\mathcal{D})$ factors through a finite group, then $M$ will have the structure of an orbifold near $[L]\in \mathfrak{M}_{V}^{{\rm cl}}$ and
$$ T_{[L]}\mathfrak{M}^{{\rm cl}}_{V} \cong H_{L}^1(V,\mathcal{D}_{V}).$$
\end{corollary}
We have seen by explicit computations that $\tilde{\varrho}_{1}\,:\,V \rightarrow {\rm End}(\mathbb{C}.\hbar \oplus T^*_{\hbar}V)$ is a Lie algebra homomorphism if and only if $\varrho\,:\,V \rightarrow \mathcal{D}_{V}$ is a Lie algebra homomorphism. So, we can in fact replace in Proposition~\ref{THDEF1} and Corollary~\ref{THDEF2} the module $\mathcal{D}_{V}$ by the module $\mathbb{C}.\hbar \oplus T^*_{\hbar}V$. However, if we only used the module $T^*_{\hbar}V$, we would miss the constraints imposed the zero torsion condition \eqref{commuts}.
\subsection{Translations}
\label{S34}
So far, the symmetries we have described act linearly on the coefficients of quantum Airy structures. Among them, translations $x_i \rightarrow x_i + t_i$ transform a quasi-Airy structure into operators $(\tilde{L}_i)_{i \in I}$ such that $(\tilde{L}^i - P^i)_{i \in I}$ is a quasi-Airy structure, for some constants $P^i$. The solution of $\tilde{L}_i\cdot \tilde{Z} = 0$ for all $i$ is $Z(t + x)$, which is the Taylor expansion of $Z(\xi)$ around the point $\xi = t$. If we write $Z(x) = \exp\big(\sum_{g \geq 0} \hbar^{g - 1}S_{g}(x)\big)$, and assume momentarily that $S_g$ has a non-zero radius of convergence uniformly in $g$,
$$
Z(t + x) = \exp\Big(\sum_{g \geq 0} \sum_{n \geq 1} \sum_{i_1,\ldots,i_n} \frac{\hbar^{g - 1}}{n!}\,F_{g,n}^{(t)}(i_1,\ldots,i_n)x_{i_1}\cdots x_{i_n}\Big),
$$
where now $F_{0,1}^{(t)}(i) = \partial_{i}S_{0}(t)$ and $F_{0,2}^{(t)}(i,j) = \partial_{i}\partial_{j}S_{0}(t)$ are a priori now zero. It is natural to suspect that
$$
\tilde{Z}(x) := \exp\big(-\tfrac{F_{0,1}^{(t)}(a)x_a}{\hbar} - \tfrac{F_{0,2}^{(t)}x_ax_b}{2\hbar}\big)Z(x + t)
$$
is the partition function of a new, $t$-dependent quantum Airy structure. The next theorem will confirm and make sense of this, using formal series in $t$.
We introduce the graded vector space $\hat{K} = \mathbb{K}[[(t_i)_{i \in I}]]$, by assigning degree $1$ to each $t_i$. We get the decomposition into homogeneous pieces
$$
\hat{K} = \prod_{n \geq 0} \hat{K}^{[n]}.
$$
Let $(A,B,C,D)$ be a quantum Airy structure. We first describe a formal replacement for $\partial_{i}S_{0}(t)$ and $\partial_{i}\partial_{j}S_0(t)$. These are elements of $\hat{K}$ whose homogeneous components are inductively defined
$$
G_{0,1}(i) = \sum_{n \geq 2} G_{0,1}^{[n]}(i),\qquad G_{0,2}(i,j) = \sum_{n \geq 1} G_{0,2}^{[n]}(i,j),
$$
with
\begin{eqnarray}
\label{G01ini}G_{0,1}^{[2]}(i) & = & \tfrac{1}{2}A^i_{a,b}t_{a}t_{b}\,, \\
\label{G01rec} \forall n \geq 3,\quad G_{0,1}^{[n]}(i) & = & B^i_{a,b}t_aG_{0,1}^{[n - 1]}(b) + \tfrac{1}{2}C^i_{a,b}\sum_{\substack{n_1 + n_2 = n \\ n_1,n_2 \geq 2}} G_{0,1}^{[n_1]}(a)G_{0,1}^{[n_2]}(b)\,,
\end{eqnarray}
and
\begin{eqnarray}
\label{G02ini} G_{0,2}^{[1]}(i,j) & = & A^i_{j,a}t_{a}\,, \\
\label{G02rec} \forall n \geq 2,\quad G_{0,2}^{[n]}(i,j) & = & B^i_{a,b}t_{a}G_{0,2}^{[n - 1]}(b,j) + B^i_{j,a}G_{0,1}^{[n]}(a) + C^i_{a,b} \sum_{\substack{n_1 + n_2 = n \\ n_1 \geq 2,\,\,n_2 \geq 1}} G_{0,1}^{[n_1]}(a)G_{0,2}^{[n_2]}(b,j)\,.
\end{eqnarray}
Then, we define $(\tilde{A},\tilde{B},\tilde{C},\tilde{D})$ with coefficients in $\hat{K}$, again inductively by their homogeneous components. The initial conditions are
\begin{equation}
\label{iniform} \tilde{X}^{[0]} = X,\qquad X \in \{A,B,C,D\}
\end{equation}
and the recursions read for $n \geq 1$
\begin{eqnarray}
\label{Arecform} (\tilde{A}^{[n]})^{i}_{j,k} & = & B^i_{a,b}(A^{[n - 1]})^{b}_{j,k}t_{a} + B^i_{j,a}G_{0,2}^{[n]}(a,k) + B^i_{k,a}G_{0,2}^{[n]}(a,j) \\
&& + C^i_{a,b}\bigg(\sum_{\substack{n_1 + n_2 = n \\ n_1 \geq 2\,\,n_2 \geq 0}} G_{0,1}^{[n_1]}(a)(A^{[n_2]})^{b}_{j,k} + \sum_{\substack{n_1 + n_2 = n \\ n_1,n_2 \geq 1}} G_{0,2}^{[n_1]}(a,j)G_{0,2}^{[n_2]}(b,k)\bigg). \\
\label{Brecform} (\tilde{B}^{[n]})^i_{j,k} & = & B^i_{a,b}(B^{[n - 1]})^{b}_{j,k}t_{a} + C^i_{k,a}G_{0,2}^{[n]}(a,j) + \sum_{\substack{n_1 + n_2 = n \\ n_1 \geq 2,\,\,n_2 \geq 0}} C^i_{a,b}G_{0,1}^{[n_1]}(a)(B^{[n_2]})_{j,k}^b. \\
\label{Crecform} (\tilde{C}^{[n]})^i_{j,k} & = & B^i_{a,b}(C^{[n - 1]})^b_{j,k}t_a + \sum_{\substack{n_1 + n_2 = n \\ n_1 \geq 2,\,\,n_2 \geq 0}} C^i_{a,b}G_{0,1}^{[n_1]}(a)(C^{[n_2]})^b_{j,k}. \\
\label{Drecform} (\tilde{D}^{[n]})^i & = & B^i_{a,b}(D^{[n - 1]})^bt_a + C^i_{a,b} \sum_{\substack{n_1 + n_2 = n \\ n_1 \geq 2,\,\,n_2 \geq 0}} G_{0,1}^{[n_1]}(a)(D^{[n_2]})^b.
\end{eqnarray}
\begin{proposition}
If $(A,B,C,D)$ is a quantum Airy structure whose partition function has Taylor coefficients $F_{g,n}(i_1,\ldots,i_n)$, then $(\tilde{A},\tilde{B},\tilde{C},\tilde{D})$ is a quantum Airy structure with coefficients in $\hat{K}$ whose partition function has Taylor coefficients
\begin{equation}
\label{Fgnshift} \tilde{F}_{g,n}(i_1,\ldots,i_n) = \sum_{m \geq 0} \frac{1}{m!} \sum_{j_1,\ldots,j_m} F_{g,n}(i_1,\ldots,i_n,j_1,\ldots,j_m)t_{j_1,\ldots,j_m} \in \hat{K}.
\end{equation}
\end{proposition}
The above formulas form a non-linear\footnote{We mean that $\tilde{X}$ does not depend linearly of $X$.}, infinitesimal symmetry of quantum Airy structures. In the convergent case, this symmetry is the one expected.
\begin{lemma}
Assume $V$ finite-dimensional. Let $Z = \exp\big(\sum_{g \geq 0} \hbar^{g - 1}S_{g}(x)\big)$ be the partition function of a quantum Airy structure, where $S_g \in \mathbb{K}[[(x_i)_{i \in I}]]$. If $S_0$ has positive radius of convergence, then for any $g \geq 0$, $S_{g}$ has a radius of convergence bounded from below by a positive constant independent of $g$. We denote $\underline{S}_g$ the analytic function defined by those series at least in a neighborhood of $0$ in $V$. The formal series defining $(\tilde{A},\tilde{B},\tilde{C},\tilde{D},G_{0,1},G_{0,2})$ have positive radius of convergence, and we also use underlined letters to denote the analytic functions of $t$ they define. Then
$$
\underline{G}_{0,1}(t;i) = \partial_{i} \underline{S}_0(t),\qquad \underline{G}_{0,2}(t;i,j) = \partial_{i}\partial_{j} \underline{S}_{0}(t),\qquad \underline{A}^i_{j,k}(t) = \partial_i\partial_j\partial_k\underline{S}_0(t),\qquad \underline{D}^i_{j,k}(t) = \partial_{i} \underline{S}_{1}(t).
$$
Further, for $x$ and $t$ in a neighborhood of $0$,
$$
\underline{\tilde{S}}_{g}(t;x) = \underline{S}_{g}(x + t).
$$
\end{lemma}
\noindent \textbf{Proof.} A computation shows that the translation $x \rightarrow x + t$, followed by conjugation by
$$
\mathcal{U} = \exp\Big(\hbar^{-1}\big(G_{0,1}(a)x_a + \tfrac{1}{2}G_{0,2}(a,b)x_ax_b\big)\Big)
$$
transforms the quantum Airy structure $(A,B,C,D)$ into the quasi-Airy structure $(\hat{L}_i - \hat{P}^i)_{i \in I}$ with coefficients (see \eqref{Lpreform} for notations, we replaced here $M^i_{j}$ by $M_{i,j}$ and $N^i_{j}$ by $N_{i,j}$ for convenience)
\begin{eqnarray}
\hat{M}_{i,j} & = & \delta_{i,j} - B^i_{j,a}t_a - C^i_{j,a}G_{0,1}(a). \nonumber \\
\hat{N}_{i,j} & = & -A^i_{j,a}t_{a} + (\delta_{i,b} - B^i_{a,b}t_a - C^i_{a,b}G_{0,1}(a))G_{0,2}(b,j) - B^i_{j,a}G_{0,1}(a). \nonumber \\
\hat{P}^i & = & -\tfrac{1}{2}A^i_{a,b}t_{a}t_{b} - \tfrac{1}{2}C^i_{a,b}G_{0,1}(a)G_{0,1}(b) + (\delta_{i,b} - B^i_{a,b}t_{a})G_{0,1}(b). \nonumber \\
\hat{A}^i_{j,k} & = & A^i_{j,k} + B^i_{j,a}G_{0,2}(a,k) + B^i_{k,a}G_{0,2}(a,j) + C^i_{a,b}G_{0,2}(a,j)G_{0,2}(b,k). \nonumber \\
\hat{B}^i_{j,k} & = & B^i_{j,k} + C^i_{k,a}G_{0,2}(a,j). \nonumber \\
\hat{C}^i_{j,k} & = & C^i_{j,k}. \nonumber \\
\hat{D}^i & = & D^i. \nonumber
\end{eqnarray}
We indeed remark that the operations of translation and conjugation by the exponential of a quadratic form preserve the Lie commutation relations, so $(\hat{L}_i - \hat{P}^i)_{i \in I}$ is indeed a quasi-Airy structure, with same structure constants. This determines a quantum Airy structure provided $\hat{M}$ is an invertible matrix, and provided one can choose $\hat{N}_{i,j} = \hat{P}^i = 0$ for all $i,j$. In this case, the quantum Airy structure is $\tilde{L}_i = (\hat{M}^{-1})_{i,a}\hat{L}_a$, \textit{i.e.} its coefficients are $\tilde{X}^i = (\hat{M}^{-1})_{i,a}\hat{X}^a$ for $X \in \{A_{j,k},B_{j,k},C_{j,k},D\}$.
We can indeed solve the equation $\hat{P}^i = 0$ by choosing $G_{0,1}(i)$ as in \eqref{G01ini}-\eqref{G01rec}. Then, $G_{0,2}(i,j)$ is obtained by solving perturbatively $\hat{N}_{i,j} = 0$, leading to \eqref{G02ini}-\eqref{G02rec}. Inserting these series in the expression of the coefficients of $\tilde{L}_i$
$$
\tilde{X}^i = \hat{M}^{-1}_{i,a}\hat{X}^a,\qquad X \in \{A_{j,k},B_{j,k},C_{j,k},D\}
$$
leads to formulas \eqref{iniform} and \eqref{Arecform}-\eqref{Drecform}. The partition function for this new quantum Airy structure is
$$
\tilde{Z}(x) = \exp\big(- \tfrac{G_{0,1}(a)x_{a}}{\hbar} - \tfrac{G_{0,2}(a,b)x_ax_b}{2\hbar}\big)Z(x + t)
$$
and by consistency, we deduce that
\begin{equation}
\label{inininini}G_{0,1}(i) = \partial_{i} S_{0}(t),\qquad G_{0,2}(i,j) = \partial_{i}\partial_{j}S_0(t),\qquad \tilde{A}^i_{j,k} = \partial_{i}\partial_{j}\partial_{k} S_0(t),\qquad \tilde{D}^i = \partial_{i} S_1(t)
\end{equation}
and the Taylor coefficients of $\tilde{Z}$ are given by \eqref{Fgnshift}, both in the sense of formal series in $(t_i)_{i \in I}$.
Now assume that $V$ is finite-dimensional and $S_0(t)$ has a non-zero radius of convergence. The equation $L_i\cdot Z = 0$ implies for $g \geq 1$
$$
\hat{M}_{i,a}(t)\partial_{a}S_g(t) = \delta_{g,1}D^i + \tfrac{1}{2}C^i_{a,b}\bigg(\partial_{a}\partial_{b} S_{g - 1}(t) + \sum_{\substack{g_1 + g_2 = g \\ g_1,g_2 \geq 1}} \partial_{a}S_{g_1}(t) \partial_{b}S_{g_2}(t)\bigg).
$$
We recall that $\hat{M}_{i,a}(t) = \delta_{i,a} - B^i_{a,b}t_{b} - C^i_{a,b}\partial_{b}S_0(t)$. As $S_0(t) = O(t)$, we have $\hat{M}(t) = {\rm Id} + O(t)$. Hence, the (finite-dimensional) matrix $\hat{M}(t)$ is invertible for $t$ small enough. So, we can prove by induction on $g \geq 1$ that $S_g(t)$, as a solution of the (compatible) system of linear ODEs with analytic coefficients in a neighborhood of $t = 0$,
$$
\partial_{i}S_g(t) = [\hat{M}^{-1}(t)]_{i,c}\bigg\{\delta_{g,1}D^c + \tfrac{1}{2}C^c_{a,b}\Big(\partial_a\partial_{b}S_{g - 1}(t) + \sum_{\substack{g_1 + g_2 = g \\ g_1,g_2 \geq 1}} \partial_{a}S_{g_1}(t) \partial_{b}S_{g_2}(t)\Big)\bigg\}
$$
is the formal Taylor series at $0$ of an analytic function $\underline{S}_{g}(t)$ in the neighborhood $\Omega$ of $0$ on which $\hat{M}(t)$ is invertible and $S_0(t)$ is analytic. This contains a neighborhood of $0$ independent of $g$.
Independently, as $S_0(t)$ is analytic in a neighborhood of $0$, the equality
\begin{equation}
\label{G01S0} G_{0,1}(i) = \partial_{i} S_0(t)
\end{equation}
provides a definition of $\underline{G}_{0,1}$ as the analytic function $\partial_{i} \underline{S}_0(t)$, whose Taylor series at $0$ is $G_{0,1}(i)$, hence such that \eqref{G01S0} holds at the level of analytic functions. Then, the expression for the formal series $G_{0,2}(i,j)$ obtained by enforcing $N_{i,j} = 0$ above shows that it is the formal Taylor series at $0$ of an analytic function $\underline{G}_{0,2}(i,j)$ for $t$ in $\Omega$. And the expression of $(\tilde{A},\tilde{B},\tilde{C},\tilde{D})$ in terms of $S_0(t)$ and its first and second order derivatives shows they upgrade in the same way to analytic functions of $t \in \Omega$, in such a way that the equality between formal series at $0$ continue to hold at the level of analytic functions of $t \in \Omega$. \hfill $\Box$
\section{Formulas for the partition function in two simple cases}
\label{SS3}
\subsection{The case $C = 0$}
\label{S35}
Quantum Airy structures with $C = 0$ give rise to a compatible system of linear ODEs for the partition function
$$
\forall i\qquad \hbar(\delta_{i,b} - B^i_{a,b}x_a)\partial_{x_b}\operatorname{ln}\,Z = \tfrac{1}{2}A^i_{a,b}x_{a}x_{b} + \hbar D^i.
$$
The partition function in this case can be computed in exact form. Let us introduce matrices $\mathbf{B}_j = (B^a_{j,b})_{a,b}$ and column vectors $\mathbf{A}_{j,k} = (A_{j,k}^a)_{a}$ and $\mathbf{D} = (D^a)_{a}$, as well as the formal power series
\begin{eqnarray}
\psi_0(z) & := & \tfrac{1}{2z^3}\big(-\operatorname{ln}(1 - z) - z - \tfrac{z^2}{2}\big) = \tfrac{1}{6} + O(z), \nonumber \\
\psi_1(z) & := & -\tfrac{1}{z}\,\operatorname{ln}(1 - z) = 1 + O(z), \nonumber
\end{eqnarray}
which we will apply to matrices.
\begin{proposition}\label{Cequal0}
The partition function of a quantum Airy structure having $C = 0$ reads
$$
Z = \exp\big(\hbar^{-1}S_0(x) + S_1(x)\big),
$$
where
$$
S_0(x) =\Big[\psi_0\Big(\sum_{j}x_j \mathbf{B}_j\Big)\cdot \mathbf{A}_{a,b}\Big]_{c}x_{a}x_{b}x_{c},\qquad S_1(x) = \Big[\psi_1\Big(\sum_{j} x_j\mathbf{B}_j\Big)\cdot \mathbf{D}\Big]_{a}x_a\,,
$$
and $\cdot$ is the multiplication between a matrix and a column vector.
\end{proposition}
\noindent \textbf{Proof.} We use the expression of the Taylor coefficients $F_{g,n}(i_1,\ldots,i_n)$ of the partition function as sums over trivalent graphs. Since $C = 0$, all trivalent vertices should be incident to one leaf (a $B$), two leafs (an $A$), or a loop (a $D$). This drastically simplifies the structure of the graphs which can contribute to the sum, in particular their genus is $0$ and $1$. Thus $F_{g,n} = 0$ for $g \geq 2$.
In genus $0$ (Figure~\ref{Fig:genus0graph}), the graphs are characterized by the sequence $\sigma(1),\ldots,\sigma(n - 3)$ of leaves successively attached when moving away from the root, and the pair $\{m,m'\}$ of leaves terminating the graph. Thus:
\begin{eqnarray}
F_{0,n}(i_1,\ldots,i_n) & = & \sum_{\sigma} \sum_{m < m'} B^{i_1}_{i_{\sigma(1)},a_1}B^{a_2}_{i_{\sigma(2)},a_2} \cdots B^{a_{n - 4}}_{i_{\sigma(n - 3)},a_{n - 3}} A^{a_{n - 3}}_{i_m,i_{m'}} \nonumber \\
& = & \sum_{\sigma} \sum_{m < m'} \big[\mathbf{B}_{i_{\sigma(1)}}\mathbf{B}_{i_{\sigma(2)}}\cdots \mathbf{B}_{i_{\sigma(n - 3)}}\cdot \mathbf{A}_{i_m,i_{m'}}\big]_{i_1} \nonumber
\end{eqnarray}
where the first sum ranges over bijections $\sigma\,:\,\{1,\ldots,n - 3\} \rightarrow \{2,\ldots,n\}\setminus \{m,m'\}$ and we recall that the indices $a_1,a_2,\ldots,a,b,c,\ldots $ are implicitly summed over. Then
$$
\sum_{i_1,\ldots,i_n} F_{0,n}(i_1,\ldots,i_n)\,\frac{x_{i_1}\cdots x_{i_n}}{n!} = \frac{1}{n!} \sum_{\sigma} \sum_{m < m'} \Big[\big(x_d\mathbf{B}_{d}\big)^{n - 3}\cdot \mathbf{A}_{a,b}\Big]_{c}x_ax_bx_c.
$$
The set of $(\sigma,\{m,m'\})$ appearing in this sum is in bijection with permutations of $\{2,\ldots,n\}$, thus
$$
\sum_{i_1,\ldots,i_n} F_{0,n}(i_1,\ldots,i_n)\,\frac{x_{i_1}\cdots x_{i_n}}{n!} = \tfrac{1}{n} \Big[\big(x_d\mathbf{B}_{d}\big)^{n - 3}\cdot \mathbf{A}_{a,b}\Big]_{c}x_ax_bx_c,
$$
and summing over $n \geq 3$ gives the announced expression.
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{genus0graph}
\caption{Genus $0$ graphs without inner trivalent vertex. \label{Fig:genus0graph}}
\end{figure}
\end{center}
In genus $1$ (Figure~\ref{Fig:genus1graph}), the same graphs appear except that the terminal vertex is a loop instead of a pair of leaves. Thus for $n \geq 1$
\begin{eqnarray}
F_{1,n}(i_1,\ldots,i_n) & = & \sum_{\sigma} B^{i_1}_{i_{\sigma(1)},a_1}B^{a_2}_{i_{\sigma(2)},a_2} \cdots B^{a_{n - 2}}_{i_{\sigma(n - 1)},a_{n - 1}} D^{a_{n - 1}} \nonumber \\
& = & \sum_{\sigma} \big[\mathbf{B}_{i_{\sigma(1)}}\mathbf{B}_{i\sigma(2)}\cdots \mathbf{B}_{i_{\sigma(n - 1)}}\cdot \mathbf{D}\big]_{i_1}\,, \nonumber
\end{eqnarray}
where $\sigma$ are bijections from $\{1,\ldots,n - 1\}$ to $\{2,\ldots,n\}$. Therefore
$$
\sum_{i_1,\ldots,i_n} F_{1,n}(i_1,\ldots,i_n)\,\frac{x_{i_1}\cdots x_{i_n}}{n!} = \frac{1}{n} \Big[\big(x_b\mathbf{B}_{b}\big)^{n - 1}\cdot \mathbf{D}\big]_{a}x_{a},
$$
and summing over $n \geq 1$ gives the announced expression. \hfill $\Box$
\begin{center}
\begin{figure}[h!]
\centering
\includegraphics[width=0.6\textwidth]{genus1graph}
\caption{Genus $1$ graphs without inner trivalent vertex. \label{Fig:genus1graph}}
\end{figure}
\end{center}
We remark that, if $(L_i)_{i \in I}$ forms an abelian quantum Airy structure, all matrices $\rho_1(e^i)$ in \eqref{rho1} commute, therefore they can simultaneously be brought into an upper-triangular form. So, abelian quantum Airy structures are always, up to the action of $\mathcal{G}_{V}$, equivalent to an abelian quantum Airy structure with $C = 0$, to which Proposition~\ref{Cequal0} can be applied.
\subsection{The case $A = 0$}
\label{S35bis}
Quantum Airy structures with $A = 0$ are related in principle to quantum Airy structures with $C = 0$ by an element of $\mathcal{G}_{V}$. Indeed, the automorphism of the Weyl algebra $(x_i,\partial_{x_i}) \rightarrow (\hbar\partial_{x_i},-x_i)$ exchanges $C$-terms and $A$-terms. This transformation yields operators which are not anymore in normal form, but it can be pre- and post-composed with elements in $\exp(\mathcal{D}_{V,2}/\hbar)$ so as to be brought back to quantum Airy structures. So the partition function of a quantum Airy structures with $A = 0$ is computable in principle from Proposition~\ref{Cequal0}.
Here we take a direct route, by examining which graphs may give non-zero contributions.
\begin{lemma}
\label{Aequal0genus0} For a quantum Airy structure with $A = 0$, we have $F_{0,n} = 0$ for all $n \geq 1$.
\end{lemma}
\noindent \textbf{Proof.} The recursion for $n \geq 4$
$$
F_{0,n}(i_1,\ldots,i_{n}) = \sum_{m = 2}^{n} B^{i_1}_{i_m,a} F_{0,n - 1}(a,i_2,\ldots,\widehat{i_m},\ldots,i_n)
$$
together with the initial data $F_{0,3}(i_1,i_2,i_3) = A^{i_1}_{i_2,i_3}$ has $F_{0,n} = 0$ as the unique solution. \hfill $\Box$
In higher genera, simplification occurs when furthermore $B = 0$.
\begin{proposition}
\label{ABequal0}
Let $\mathfrak{T}_{g}$ be the set of rooted trivalent trees with $g$ leaves. The root edge is denoted $r$. If $v$ is a vertex, we denote $e_0(v)$ the edge closer to the root, and $e_{\pm}(v)$ the two other edges -- in arbitrary order. If $T \in \mathfrak{T}_{g}$, let $E(T)$ be the set of unoriented edges (including the root edge, denoted $r$), $V(T)$ the set of trivalent vertices, and $L(T)$ the set of leaves (univalent vertices distinct from the root). If $\ell$ is a leaf, we denote $e(\ell)$ its incident edge. Note that the cardinality of ${\rm Aut}\,T$ is a power of $2$, as this group consists of the permutations of $\{e_{-}(v),e_{+}(v)\}$ for each $v$, which preserve $T$.
The partition function of a quantum Airy structure with $A = B = 0$ is computed by
$$
F_{g,n}(i_1,\ldots,i_n) = \delta_{n,1}f_g(i_1),
$$
with $f_0 = 0$ and for $g \geq 1$
\begin{equation}
\label{fgab0} f_g(i) := \sum_{T \in \mathfrak{T}_{g}} \frac{1}{|{\rm Aut}\,T|} \sum_{\substack{a\,:\,E(T) \rightarrow I \\ a(r) = i}} \prod_{v \in V(T)} C^{a(e_0(v))}_{a(e_-(v)),a(e_+(v))}\,D^{a(e(\ell))}.
\end{equation}
\end{proposition}
\noindent \textbf{Proof.} Let us write $Z = \exp\big(\sum_{g \geq 0} \hbar^{g - 1}S_{g}\big)$. $L_i \cdot Z = 0$ turns into the differential recursion
\begin{equation}
\label{Sgrel}\forall g \geq 0,\qquad \partial_{x_i} S_{g} = \tfrac{1}{2}C^i_{a,b}\big(\partial_{x_a}\partial_{x_b}S_{g - 1}(x) + \sum_{h_1 + h_2 = g} \partial_{x_a} S_{h_1} \partial_{x_b}S_{h_2}\big) + \delta_{g,1}D^i.
\end{equation}
We know from Lemma~\ref{Aequal0genus0} that $S_{0} = 0$. Therefore, the equation above for $g = 0$ implies $S_{1} = \sum_{i} D^i x_i$, and by induction that $S_{g}$ must be a linear function. As $A = B = 0$, in the graphs in $\mathfrak{G}_{g,n}(1)$ with non-zero contribution the Betti number $g$ can only arise from loops receiving a $D$-weight. They should thus form $g$ leaves of the graph attached to the spanning tree. \hfill $\Box$
\section{Finite-dimensional Airy structures and representation theory}
\label{SS5}
\subsection{General properties}
\begin{lemma}
\label{Vsemi1}Assume $V$ is a finite-dimensional Lie algebra, $L\,:\,V \rightarrow \mathcal{D}_{V}$ a homomorphism of normal form and that there exists a V-submodule $\rho_M : V \rightarrow \End(M)$ of $ \mathbb{C}.\hbar \oplus T^*_{\hbar}V $ and an isomorphism
\begin{equation}\label{isoM}
\Phi\,:\,\mathbb{C}.\hbar \oplus T^*_{\hbar}V \rightarrow \mathbb{C}.\hbar \oplus M,\qquad \Phi \tilde \rho_1 \Phi^{-1} =\rho_{\text{t}} \oplus \rho_M,
\end{equation}
where $\rho_{\text{t}}$ is the one-dimensional trivial representation and $\Phi|_{\mathbb{C}.\hbar} = \mathop{\rm Id}\nolimits$.
Then there exists $\alpha \in M^*$ such that
$$
\forall v \in V,\qquad \varrho_{0,1}(v) \circ \Phi^{-1} = \alpha \circ \rho_M(v)\circ \pi_M,
$$
where $\pi_M: \mathbb{C}.\hbar \oplus M \rightarrow M$ is the projection onto $M$ with kernel $\mathbb{C}.\hbar$.
Further,
$$
V_{\alpha} := \{v \in V\,\,:\,\,{\rm Im}\,\rho_M(v) \subseteq {\rm Ker}\,\alpha\}
$$
must vanish and
$$\pi_M \circ \Phi \circ \mathcal{I}(V) = \{ m\in M\,\,\,:\,\,\,\forall v \in V\quad \alpha(\rho_M(v)m)=0\},$$
$$\pi_M \circ \Phi \circ \mathcal{I}([V,V]) \subset {\rm Ker}\,\alpha.$$
where $\mathcal{I}$ is the Lagrangian embedding from Lemma~\ref{LLLfdsf}.
\end{lemma}
We note that if $V$ is semi-simple then the trivial submodule $\mathbb{C}.\hbar$ in $\mathbb{C}.\hbar \oplus T^*_{\hbar}V $ will have a complementary submodule $M$ as assumed in Lemma \ref{Vsemi1}, thus the conclusion of this Lemma applies in this case. We will then use the notation
$$ \tilde{\mathcal{I}} = \pi_M \circ \Phi \circ \mathcal{I}.$$
\begin{lemma}\label{Vsemi1.1}
Suppose that $V$ is a finite-dimensional semi-simple Lie-algebra and $L\,:\,V \rightarrow \mathcal{D}_{V}$ a Lie algebra homomorphism of normal form. Let $\rho_M : V \rightarrow \End(M)$ be the symplectic V-submodule of $ \mathbb{C}.\hbar \oplus T^*_{\hbar}V $ with the properties stated in Lemma \ref{Vsemi1}. If $\omega_M$ is the $V$-invariant symplectic structure on $M$, then there exists an $m_\alpha \in \tilde{\mathcal{I}}(V) \subset M$ such that
$$\alpha = \omega_M(m_\alpha, \cdot).$$
Moreover $\tilde{\mathcal{I}}(V)$ is a Lagrangian subspace of $M$ on which $\alpha$ vanishes and
$$ \rho_M(V) m_\alpha = \tilde{\mathcal{I}}(V).$$
There exists a unique $v_\alpha \in V$ such that
$$ \rho_M(v_\alpha) m_\alpha = m_\alpha.$$
\end{lemma}
\noindent\textbf{Proof of Lemma~\ref{Vsemi1}.} We consider some $V$-module structure $\tilde{\rho}_{1}\,:\,V \rightarrow {\rm End}(\mathbb{C}.\hbar \oplus T^*_{\hbar}V)$ of the form (\ref{rho1ti}). Then $\mathbb{C}.\hbar$ is a trivial one-dimensional submodule.
By projection from $\mathbb{C}.\hbar \oplus T^*_{\hbar}V $ to $\mathbb{C}.\hbar$ along $T^*_{\hbar}V$, we get a linear functional $\tilde \alpha \in (\mathbb{C}.\hbar \oplus T^*_{\hbar}V)^*$. If we now let $\beta$ be the projection onto $T^*_{\hbar}V$ along $\mathbb{C}.\hbar$, then $\tilde \alpha + \beta = \mathop{\rm Id}\nolimits_{\mathbb{C}.\hbar \oplus T^*_{\hbar}V }$ and
$$\tilde \rho_1 = (\tilde \alpha + \beta) \circ \Phi^{-1}\circ \rho_M \circ \Phi=
\tilde \alpha \circ \Phi^{-1}\circ\rho_M \circ \Phi + \beta \circ \Phi^{-1}\circ \rho_M \circ \Phi,$$
where the first factor on the right hand side gives $\rho_{0,1}$ by definition. We now let $\alpha = \tilde \alpha\circ \Phi^{-1}|_M$ and get the claimed formula for $\rho_{0,1}$.
If $(e_i)_{i \in I}$ is a basis of $V$, the linear forms $\alpha \circ \rho_M(e_i)$ on $V$ must be linearly independent. Therefore, if $v = \sum_{i \in I} v_i e_i$ is solution to $\rho_{0,1}(v) = \sum_{i \in I} v_i\,\alpha \circ \Phi^{-1} \circ \rho_M(e_i) = 0$, we must have $v = 0$. This implies $V_{\alpha} = 0$. From the normal form of $L$, we see that
$$ V = \{ m\in M\,\,:\,\,\,\forall v \in V\quad \rho_{0,1}(v)(\Phi^{-1}(m))=0\},$$
from which it follows that
$$ \pi_M \circ \Phi \circ \mathcal{I}(V) = \{ m\in M\,\,\,:\,\,\,\forall v \in V\quad \alpha( \rho_M(v)m)=0\}.$$
But then (\ref{commuts}) implies that
$$ \pi_M \circ \Phi \circ \mathcal{I}([V,V]) \subset {\rm Ker}\,\alpha.$$
\hfill $\Box$
\noindent\textbf{Proof of Lemma~\ref{Vsemi1.1}.}
There exists a unique $m_\alpha \in M$ such that $\alpha = \omega_{M}( m_\alpha,\cdot )$, where $\omega_{M}$ is the symplectic form, by the non-degenerateness of $\omega_{M}$. Since
$$\alpha \circ \rho_M(v) \pi_M \circ \Phi \circ \mathcal{I} (u) = 0$$
for $u,v\in V$ by the normal form assumption, we see that
$$\rho_M(V) m_\alpha \subset \pi_M \circ \Phi \circ \mathcal{I}(V).$$
As $V_\alpha = 0$, $\alpha\circ \rho_M : V \rightarrow M^*$ is injective. Thus
$$ \rho_M(V) m_\alpha = \tilde{\mathcal{I}}(V).$$
Since $V$ is semi-simple, we get that $[V,V]=V$. Lemma~\ref{Vsemi1} then implies that $\alpha(\tilde{\mathcal{I}}(V)) =0$. Thus $m_\alpha$ must be in the Lagrangian subspace $\tilde{\mathcal{I}}(V)$, and we see that there must exist $v_\alpha\in V$ such that
$$\rho_M(v_\alpha) m_\alpha = m_\alpha.$$
If $\tilde{v}_\alpha\in V$ was another such element, then $\rho_M(v_\alpha-\tilde{v}_\alpha)m_\alpha = 0$. Thus $v_\alpha - \tilde{v}_\alpha \in V_\alpha$. By Lemma \ref{Vsemi1}, we get that $v_\alpha =\tilde{v}_\alpha$, hence the claimed uniqueness.
\hfill $\Box$
We shall now establish a kind of converse to Lemma \ref{Vsemi1.1}.
\begin{theorem}\label{SSC}
Suppose that $V$ is a finite-dimensional Lie algebra and $\rho_M : V \rightarrow \End(M)$ is a symplectic representation of dimension twice the dimension of $V$ and further assume that there exists $m_\alpha\in M$ such that $\rho_M(V)m_\alpha \subset M$ is a Lagrangian subspace.
Then for any choice of a basis $(w_i)_{i \in I}$ of a Lagrangian complement $W$ to $ \rho_M(V)m_\alpha$, one can construct a quantum Airy structure $L : V \rightarrow \mathcal{D}_V$ as follows. Let $v_i$ be the unique basis of $V$ such that $\big(w_i,(\rho_M(v_i)m_\alpha)\big)_{i \in I}$ is a symplectic basis of $M$. Now let $(A^i,B^i,C^i)$ be defined by the matrix of $\rho_M(v_i)$ in this basis as in Equation~\ref{rho1}, and let $D^i = \frac12 {\hbox{tr}}(B^i)$. Then $L_i$ is determined by $(A^i,B^i,C^i, D^i)$ as in Equation~\eqref{Lform}.
\end{theorem}
\noindent\textbf{Proof.}
We define the map $\tilde{\mathcal{I}}\,:\,V \rightarrow \rho_M(V)m_\alpha$ by the formula
$$ \tilde{\mathcal{I}}(v) = \rho_M(v)m_\alpha.$$
We then immediately see that
$$
\forall u,v \in V,\qquad \rho_M(u)\tilde{\mathcal{I}}(v) - \rho_M(v) \tilde{\mathcal{I}}(v) = \tilde{\mathcal{I}}([u,v]).
$$
Further, since $\rho_M(V)m_\alpha$ is Lagrangian, we see that $\tilde{\mathcal{I}} $ is an isomorphism. For any choice of a basis $(w_i)_i$ of a Lagrangian complement $W$ to $ \rho_M(V)m_\alpha$, we pick the basis $v_i$ of $V$ as specified in the Theorem. By using the basis $\big(w_i,(\rho_M(v_i)m_\alpha)\big)_{i \in I}$ of $M$ and the basis $(v_i^*,v_i)_{i \in I}$ of $T^*_{\hbar}V$, we induce a symplectic representation $\rho_1$ of $V$ on $T^*_{\hbar}V$ and by the above property of $\tilde{\mathcal{I}}$, we see that
$$\forall u,v \in V,\qquad \rho_1(u)\mathcal{I}(v)-\rho_1(v) \mathcal{I}(u) = \mathcal{I}([u,v]).$$
But then we can simply define $L_i$ as specified in the Theorem above and because $\rho_{0,1} = \alpha \circ \rho_M$, where $\alpha = \omega_{M}(m_\alpha, \cdot)$, we see by the way the basis of $V$ is chosen that $L_i$ are of normal form. It then follows from Lemmas \ref{LLL1}-\ref{LLLfdsf} that $(A,B,C)$ satisfies the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations and the zero torsion condition \eqref{commuts} with respect to the Lie algebra structure on $V$. Then letting $D$ be given as in the Theorem, we see by Lemma \ref{trL} that $(A,B,C,D)$ induces a quantum Airy structure.
\hfill $\Box$
\subsection{The example of $\mathfrak{sl}_{2}(\mathbb{C})$}
The previous considerations allow the construction a quantum Airy structure based on $\mathfrak{sl}_{2}(\mathbb{C})$.
\begin{theorem}
\label{sl2th}The simple Lie algebra $\mathfrak{sl}_{2}(\mathbb{C})$ supports the following quantum Airy structure
\begin{eqnarray*}
L_1 &=& \hbar \partial_1 - \hbar(3 x_1\partial_1 + 5 x_2 \partial_2 + x_3\partial_3) - \tfrac{9\hbar}{2},\\
L_2 &=& \hbar \partial_2 - \hbar(\tfrac{8}{3} x_3\partial_1 + 3 x_1 \partial_2 ) - \tfrac{3\hbar^2}{80}\partial_3^2,\\
L_3 &=& \hbar \partial_3 - \hbar(\tfrac{5}{3} x_2\partial_1 + 3 x_1 \partial_3 ) +60 x_3^2.
\end{eqnarray*}
Its partition function is $Z(x) = \tfrac{\mathfrak{Z}(x)}{\mathfrak{Z}(0)}$ with
\begin{eqnarray}
\mathfrak{Z}(x) & = & x_2^{-\frac{3}{5}}(1 - 3x_1)^{-\frac{3}{2}}\big((1 - 3x_1)^2 -10x_2x_3\big)^{\frac{1}{2}} \exp\bigg\{-\frac{8\big((1 - 3x_1)^2 - 10x_2x_3\big)^{\frac{5}{2}}}{125\hbar x_2^3}\bigg\} \nonumber \\
&& \times\widetilde{{\rm H}}^{(2)}_{\frac{1}{5}}\bigg(\frac{8\big((1 - 3x_1)^2 - 10x_2x_3\big)^{\frac{5}{2}}}{125{\rm i}\hbar x_2^3}\bigg), \nonumber
\end{eqnarray}
where $\widetilde{{\rm H}}^{(2)}(z) := {\rm H}^{(2)}(-{\rm i}z)$ and ${\rm H}^{(2)}$ is the Hankel function of the second kind. This function satisfies the differential equation
$$
\big(z^2\partial_{z}^2 + z\partial_{z} - (\nu^2 + z^2)\big)\widetilde{{\rm H}}^{(2)}_{\nu}(z) = 0
$$
\end{theorem}
\noindent\textbf{Proof.}
$V=\mathfrak{sl}_{2}(\mathbb{C})$ has three generators $(H,E,F)$ satisfying the commutation relations
$$ [H,E] = 2E, \ [H,F] = -2F \ \ \text{ and } \ \ [E,F] = H.$$
Let $M$ be the $6$-dimensional irreducible representation of $\mathfrak{sl}_{2}(\mathbb{C})$. Recall that it has a basis $(e_0,\ldots, e_5)$ such that
\begin{eqnarray*}
\rho_M(H) e_i &=& (6-2i-1) e_i,\\
\rho_M( E) e_i &=& (6-i) e_{i-1},\\
\rho_M(F) e_i &=& (i+1) e_{i+1},
\end{eqnarray*}
for $i \in \{0,\ldots,5\}$ where by convention $e_{-1}=e_6=0$.
Further $M$ is symplectic with the following symplectic structure
$$\omega_{M}(e_i,e_{5-i}) = \left\{
\begin{array}{ccc}
1 && i=0\\
-5 && i=1\\
10 && i=2
\end{array} \right.
$$
and $\text{span}\{e_0,e_1,e_2\}$, $\text{span}\{e_3,e_4,e_5\}$ are Lagrangian.
We are now going to pick\footnote{By Lemma \ref{Vsemi1.1}, we know that $m_\alpha$ has to be the eigenvector of some $v_\alpha$ with eigenvalue $1$, which we can assume is a multiple of $H$, but then we see that $m_\alpha$ must be propositional to one of the $e_i$'s. However $\rho_M(V)m_\alpha$ must be Lagrangian. This means that $m_\alpha$ has to be proportional to either $e_1$ or $e_4$.} $m_\alpha = e_1$. We get that
$$\mathcal{I}(H) = 3 e_1,\qquad \mathcal{I}(E) = 5 e_0,\qquad \mathcal{I}(F) = 2 e_2.$$
It is then easy to check that
$$ \forall u,v \in V,\qquad \rho_M(u)\mathcal{I}(v) - \rho_M(v)\mathcal{I}(u) = \mathcal{I}([u,v])$$
as claimed in the proof of Theorem \ref{SSC}. One also finds
$$
\alpha\circ \rho_M(H)(e_4) = 15,\qquad \alpha\circ \rho_M(E)(e_5) = -5 ,\qquad \alpha\circ \rho_M(F)(e_3) = -20
$$
and that all other evaluations on the basis $e_0,\ldots, e_5$ are zero.
Let now $(H^*,E^*,F^*)$ be the basis of $V^*$ which is dual to the basis $(H,E,F)$ of $V$.
We get a linear symplectomorphism $\phi : T^*_{\hbar} V \rightarrow M$ by mapping
$$\phi(H^*) = \tfrac{1}{15} e_4,\qquad \phi(E^*) = -\tfrac{1}{5} e_5,\qquad \phi(F^*) = -\tfrac{1}{20} e_3,$$
and letting $\phi|_V = \mathcal{I}.$
The representation $\rho_1 = \phi^{-1}\rho_M \varphi$ reads in the basis $(e_i)_{i = 0}^5$
$$\rho_1(H) = \left(
\begin{array}{rrrrrr}
-3 & 0& 0& 0& 0& 0\\
0 & -5& 0& 0& 0& 0\\
0 & 0& -1& 0& 0& 0\\
0 & 0& 0& 3& 0& 0\\
0 & 0& 0& 0& 5& 0\\
0 & 0& 0& 0& 0& 1
\end{array}\right)
\ \ \ \
\rho_1(E) = \left(
\begin{array}{rrrrrr}
0 & -3& 0& 0& 0& 0\\
0 & 0& 0& 0& 0& 0\\
-\frac83 & 0& 0& 0& 0& 0\\
0 & 0& 0& 0& 0& \frac83\\
0 & 0& 0& 3& 0& 0\\
0 & 0& -\frac3{40}& 0& 0& 0
\end{array}\right)
$$
$$\rho_1(F) = \left(
\begin{array}{rrrrrr}
0 & 0& -3& 0& 0& 0\\
-\frac53 & 0& 0& 0& 0& 0\\
0 & 0& 0& 0& 0& -120\\
0 & 0& 0& 0& \frac53& 0\\
0 & 0& 0& 0& 0& 0\\
0 & 0& 0& 3& 0& 0
\end{array}\right)
$$
We now let $\alpha' = \alpha \circ \phi.$ Then
$$
\alpha' \circ \rho_1(H) = H,\qquad \alpha' \circ \rho_1(E) = E,\qquad \alpha' \circ \rho_1(F) = F,
$$
where we here think of $H,E,F$ as elements of $(T^*_{\hbar} V)^*.$ If we now define $\rho_{0,1} = \alpha' \circ \rho_1$, then it is of normal form. Comparing with \eqref{rho1}, we get the three stated operators. Alternatively, one can check directly that they satisfy the desired $\mathfrak{sl}_{2}(\mathbb{C})$ commutation relations.
The general solution of $L_1 \cdot Z = 0$ takes the form
$$
Z(x) = (1 - 3x_1)^{-\frac{3}{2}}\tilde{Z}\big((1 - 3x_1)^{-\frac{5}{3}}x_2,(1 - 3x_1)^{-\frac{1}{3}}x_3\big)
$$
In terms of the variables $y_2 = (1 - 3x_1)^{-\frac{5}{3}}$ and $y_3 = (1 - 3x_1)^{-\frac{1}{3}}x_3$, the equations $L_i\cdot Z = 0$ for $i \in \{2,3\}$ become
\begin{eqnarray}
\Big\{\hbar \partial_{y_2} - \hbar\big(\tfrac{40}{3}y_2y_3\partial_{y_2} + \tfrac{8}{3}y_3^2\partial_{y_3} + 12y_3\big) - \tfrac{3\hbar^2}{80}\partial_{y_3}^2 \Big\}\tilde{Z}(y_2,y_3) & = & 0 \nonumber \\
\Big\{\hbar \partial_{y_3} + 60y_3^2 - \hbar\big(\tfrac{25}{3}y_2^2\partial_{y_2} + \tfrac{5}{3}y_2y_3\partial_{y_3} + \tfrac{15}{2}y_2\big)\Big\}\tilde{Z}(y_2,y_3) & = & 0 \nonumber
\end{eqnarray}
The two equations can be decoupled by a gauge transformation and a change of variables, and then reduce to a Bessel differential equation. Taking into account the initial conditions for the sought partition function picks up the solution ${\rm H}^{(2)}$ of this differential equation. The final outcome is
$$
\tilde{Z}(y_2,y_3) = c_{\hbar}\widetilde{{\rm H}}^{(2)}_{\frac{1}{5}}\bigg(\tfrac{8}{125\hbar}\,\frac{(1 - 10y_2y_3)^{\frac{5}{2}}}{y_2^3}\bigg)\,\exp\Big\{\hbar^{-1}(-\tfrac{8}{125}y_2^{-3} + \tfrac{8}{5}y_2^{-2}y_2 - 12y_2^{-1}y_3^2)\Big\}
$$
for some $c_{\hbar}$ independent of $(y_2,y_3)$, which is the announced result.
\hfill $\Box$
\subsection{Towards a classification for simple Lie algebras}
If $V$ is a semi-simple Lie algebra and $L\,:\,V \rightarrow \mathcal{D}_{V}$ is a homomorphism of normal form, $\rho_1$ defines a structure of symplectic $V$-module on $W = T^*V$, which must split in a direct sum of irreducible $V$-modules. Let $\mathcal{R}$ be the set of equivalence class of irreducible $V$-modules which occurs in the decomposition of $W$ and $n\,:\,\mathcal{R} \rightarrow \mathbb{Z}_{> 0}$ the multiplicities. The Frobenius-Schur indicator distinguishes the following properties of an irreducible $V$-module $M$
$$
{\rm FS}(M) = \left\{\begin{array}{ccl} -1 & & {\rm if}\,\,M \cong M^*\,\,{\rm is}\,\,{\rm symplectic} \\
0 & & {\rm if}\,\,M \ncong M^* \\
1 & & {\rm if}\,\,M \cong M^*\,\,{\rm is}\,\,{\rm symmetric}
\end{array}\right.
$$
We denote $\mathcal{R}_{{\rm s}} \subset \mathcal{R}$ the subset consisting of the symplectic $V$-modules.
\begin{lemma}
\label{Ldrop} If $[M] \in \mathcal{R} \setminus \mathcal{R}_{{\rm s}}$, then $[M^*] \in \mathcal{R}\setminus \mathcal{R}_{{\rm s}}$. If ${\rm FS}(M) = 0$, then $n([M]) = n([M^*])$. If ${\rm FS}(M) = 1$, then $n([M])$ is even.
\end{lemma}
\noindent \textbf{Proof.} Suppose we have an $[M] \in \mathcal{R}\setminus \mathcal{R}_{{\rm s}}$ such that $M \ncong M^*$. Let $m = n(M)$ and $m^* = n(M^*)$. The symplectic form of $W$ restricted to $M^{m}\oplus (M^*)^{m^*}$ is then non-degenerate. Indeed, if it would be degenerate, it would induce a non-zero map from one of the $M$'s or $M^*$'s in this sum to some other module in the decomposition of $W$, which would then make $M$ or $M^*$ isomorphic to one of the other modules in the decomposition, thus we would have a contradiction.
Consider one copy of $M$ inside $M^{\oplus m}$ and let $$ M^\perp = \big\{ v \in M^{m}\oplus (M^*)^{m^*}\mid \omega(v, M) = 0\big\}.$$
We see that $M^{m}\subset M^\perp$. Let $M'$ be a $V$-submodule of $M^{m}\oplus (M^*)^{m^*}$, which is a complement of $M^\perp$. But then the projection from $M'$ to one of the $M^*$ factors must be an isomorphism. Then we see that $(M \oplus M')^\perp$ must be isomorphic to $M^{m- 1}\oplus (M^*)^{m^* - 1}$ by projection onto the remaining factors of $M$'s and $M^*$'s and we can by induction conclude that $m -1 = m^* -1$.
We now turn to the case of an $[M] \in \mathcal{R}\setminus \mathcal{R}_{{\rm s}}$ which is symmetric. Likewise, consider one copy of $M$ inside $M^{m}$ and restrict the symplectic form of $W$ to $M$. If the restriction was non-zero, it must mean that $M$ is a symplectic module and the restriction of the symplectic form of $W$ is the symplectic invariant form $M$ admits, which is unique up to non-zero rescaling. Since $M$ is assumed to be symmetric, it must then be isotropic. Consider now
$$ M^\perp = \{ v \in M^{m}\mid \omega(v, M) = 0\}.$$
Since $M^\perp$ is a $V$-submodule, we can find a $V$-submodule $M' \subset M^{m}$ such that
$$ M^{m} = M^{\perp} \oplus M'.$$
But then the symplectic form of $W$ will induce an isomorphism between $M^*$ and $M'$ and will make $M\oplus M'$ a symplectic vector subspace of $M^{m}$, but since $M'\subset M^{m}$, it must also be isomorphic to $M$. To prove that $m$ is now even, we proceed inductively by considering $ (M\oplus M')^\perp \subset M^{m}$. We see that as a $V$-module $(M\oplus M')^\perp \cong M^{m-2}$. Thus, just as before, by induction, this module must be isomorphic to $M^{\oplus 2m'}$, thus $m = 2(m'+1).$ \hfill $\Box$
Let us denote $\mathcal{R}_{2}$ the quotient of $\mathcal{R}\setminus \mathcal{R}_{{\rm s}}$ by the relation identifying the objects $M$ and $M^*$, and $\mathcal{R}' = \mathcal{R}_{{\rm s}} \cup \mathcal{R}_{2}$. Lemma~\ref{Ldrop} tells us that the multiplicity drops to a function $n\,:\,\mathcal{R}' \rightarrow \mathbb{Z}_{> 0}$.
\begin{corollary}
\label{Cdim} The dimension of the $V$-module $W$ is
\begin{equation}
\label{dimf} 2\dim V = \sum_{M \in \mathcal{R}_{{\rm s}}} n(M)\,\dim M + \sum_{M \in \mathcal{R}_{2}} 2n(M)\,\dim M.
\end{equation}
In particular: if $M \in \mathcal{R}_{{\rm s}}$, then $\dim M \leq 2\dim V$; if $M \in \mathcal{R}_{2}$, then $\dim M \leq \dim V$. \hfill $\Box$
\end{corollary}
\begin{lemma}
\label{nohadj}$W$ cannot be isomorphic as a $V$-module to the direct sum of the adjoint module and its dual.
\end{lemma}
\noindent\textbf{Proof.} Assume that $W$ is isomorphic as a $V$-module to $V \oplus V^*$. We use the Killing form $\langle \cdot,\cdot \rangle$ to identify $V$ with $V^*$ as a $V$-module, so we obtain an isomorphism of $V$-modules $\Psi\,:\,W \rightarrow V \oplus V$. Let $m_{\alpha} = (m_{\alpha}',m_{\alpha}'')$ be the element of $V\oplus V$ and $\tilde{\mathcal{I}}\,:\,V \rightarrow V \oplus V$ be the Lagrangian embedding provided by Lemma~\ref{Vsemi1.1}. Then, for all $v_1,v_2 \in V$ we have that
\begin{eqnarray}
0 & = & \omega(\tilde{\mathcal{I}}(v_1),\tilde{\mathcal{I}}(v_2)\big) \nonumber \\
& = & \omega\big(\rho(v_1)m_{\alpha},\rho(v_2)m_{\alpha}\big) \nonumber \\
& = & \big\langle \rho(v_1)m_{\alpha}',\rho(v_2)m_{\alpha}'' \big\rangle - \big\langle \rho(v_1)m_{\alpha}'',\rho(v_2)m_{\alpha}' \big\rangle \nonumber \\
& = & -\big\langle \rho(v_2)\rho(v_1)m_{\alpha}',m_{\alpha}'' \big\rangle + \big\langle \rho(v_1)\rho(v_2)m_{\alpha}',m_{\alpha}''\big\rangle \nonumber \\
& = & \big\langle \rho([v_1,v_2])m_{\alpha}',m_{\alpha}'' \rangle. \nonumber
\end{eqnarray}
The third equality exploits the fact that $W$ is isomorphic to the direct sum of adjoint modules. Since $V$ is semi-simple, we must have $[V,V] = V$, therefore $m_{\alpha}'' \in (\rho(V)m_{\alpha}')^{\bot}$. We can always fix the Cartan subalgebra $\mathfrak{h} \subset V$, such that $m_{\alpha}'$ is a generic element of $\mathfrak{h}$. Then, $\rho(V)m_{\alpha}'$ will span all the root spaces, therefore $m_{\alpha}''$ must belong to $\mathfrak{h}$ as well. But then $v_{\alpha}$ defined in Lemma~\ref{Vsemi1.1} could not be unique, thus contradicting the result of Lemma~\ref{Vsemi1.1}.
\hfill $\Box$
\begin{lemma}
\label{nozero}$\mathcal{R}$ cannot contain the trivial $V$-module.
\end{lemma}
\noindent \textbf{Proof.} Assume that the trivial representation is a $V$-submodule of $W$ which we denote $T$. According to Lemma~\ref{Ldrop}, it must have even multiplicity such that $T^{2n(T)}$ forms a symplectic $V$-submodule of $W$, and there exists a symplectic $V$-submodule $M''$ such that $W$ is isomorphic to $T^{2n(T)} \oplus M''$ as a $V$-module. Take $m_{\alpha} \in W$ as in Lemma~\ref{Vsemi1.1}, which we decompose as $(m_{\alpha},m_{\alpha}'')$ in $T^{2n(T)} \oplus M''$. Then $\rho(V)m_{\alpha} = \rho(V)m_{\alpha}'' \subset M''$. As $M''$ is a symplectic vector space, for dimensional reasons $\rho(V)m_{\alpha}$ cannot be Lagrangian, thus contradicting the result of Lemma~\ref{Vsemi1.1}. \hfill $\Box$
This dimension formula together with Lemmas~\ref{nohadj}-\ref{nozero} already puts strong constraints on $V$ and the sets of $V$-modules $\mathcal{R}_{2}$ and $\mathcal{R}_{{\rm s}}$. Since simple Lie algebras and their finite-dimensional irreducible representations are completely classified, we can use this classification to identify which $V$ and $M$ satisfy the naive dimension bounds of Corollary~\ref{Cdim} and which could belong to $\mathcal{R}$'. For each simple $V$ there are finitely many possible $M$'s. Then, one determines what are the possible $\mathcal{R}'$s made out of these $M$'s and the possible multiplicity functions $n\,:\,\mathcal{R}' \rightarrow \mathbb{Z}_{> 0}$, such that the dimension formula \eqref{dimf} holds. The outcome of this process is summarized in the following proposition.
\begin{proposition}
\label{calsssls}The following simple Lie algebras do not admit classical Airy structure (and therefore, do not admit quantum Airy structures either): $A_n$ for $n \notin \{1,5\}$, $C_n$ for $n \geq 6$, $D_n$ for any $n \geq 4$, $E_6$, $E_7$, $E_8$. For the remaining simple Lie algebras, the dimension bound \eqref{dimf} is respected for the following decompositions in irreducible modules\footnote{We indicate an irreducible module by its dimension in bold characters. If there are several non isomorphic irreducible modules of the same dimension $d$, we use the notation $\textbf{d}$, $\textbf{d}'$, $\textbf{d}''$ to distinguish them. We write $(\textbf{s})$ if and only if the module is symplectic.}.
\begin{itemize}
\item[$\boxed{A_1}$] the only candidate is $\textbf{6(s)}$ -- and it was realized in Theorem~\ref{sl2th}.
\item[$\boxed{A_5}$] the only candidate is $\textbf{20(s)}^2 \oplus \textbf{15}^2$. Here, if we denote $F$ the fundamental module, $\textbf{20(s)} = \Lambda^3F$ and $\textbf{15} = S^2F$.
\item[$\boxed{B_n}$] For $n \geq 3$, the only candidate is $\textbf{(2n + 1)}^{2n}$, where $\textbf{(2n + 1)}$ is the fundamental module. For $n = 2$, we have $\textbf{5}^4$ as well as $\textbf{4(s)} \oplus \textbf{16(s)}$. Here, $\textbf{4(s)}$ is the fundamental module for $\mathfrak{sp}(4)$ -- using the isomorphism of this Lie algebra with $\mathfrak{so}(5) = B_2$.
\item[$\boxed{C_3}$] there are three candidates: $\textbf{6(s)}^{7}$, $\textbf{14'(s)}^3$ and $\textbf{14}^2 \oplus \textbf{14'(s)}$. Here, $\textbf{6}(s)$ is the fundamental module.
\item[$\boxed{C_4}$] there are three candidates: $\textbf{8(s)}^{8}$, $\textbf{8(s)}^{3} \oplus \textbf{48}$ and $\textbf{8(s)}^{2} \oplus \textbf{27}^{2}$.
\item[$\boxed{C_5}$] the only candidate is $\textbf{110(s)}$.
\item[$\boxed{F_4}$] the only candidate is $\textbf{26}^{4}$.
\item[$\boxed{G_2}$] the only candidate is $\textbf{7}^{4}$.
\end{itemize}
\end{proposition}
We leave to future work the question of concluding on the existence of quantum Airy structures with those module contents, and of completing the classification of quantum Airy structure based on simple Lie algebras, and its extension to semi-simple Lie algebras. It is nevertheless remarkable that many simple Lie algebras are already excluded.
\section{Low dimensional examples}
\label{S5}
\subsection{Twisted cohomology of Lie algebras in low dimensions}
Let $\Lie{g}$ be a complex Lie algebra and $M$ be a $\Lie{g}$-module. We will consider a rather special case, which however suffices for our purposes to compute the needed Lie algebra cohomologies, namely that $\Lie{g}$ has a codimension one ideal $\Lie{i}$ such that
$$ \Lie{g} = \Lie{i} \oplus \Lie{q}$$
as vector spaces. Pick $y\in \Lie{q}-\{0\}$. If $v$ is an endomorphism of a vector space $\mathcal{V}$, we denote the $v$-invariant subspace $\mathcal{V}^{\langle v \rangle} \subseteq \mathcal{V}$.
\begin{lemma}\label{q1Lie}
We have that
$$ H^q(\Lie{g},M) \cong H^q(\Lie{i}, M)^{\langle y \rangle} \oplus \frac{H^{q-1}(\Lie{i}, M)}{yH^{q-1}(\Lie{i}, M)}.$$
\end{lemma}
\noindent \textbf{Proof.}
We consider the Hochschild-Serre spectral sequence \cite{HochSerre} for the pair $(\Lie{i}, \Lie{g})$ with $E_2$-page
$$ E^{p,q}_2 = H^p(\Lie{q}, H^q(\Lie{i},M)).$$
But since $\mathfrak{q}$ is one-dimensional, we see that this $E_2$-page is concentrated in the columns $p=0,1$, thus the differential $d_2$ is trivial and the spectral sequence collapses at the $E_2$-page. Since in general this spectral sequence converges to $H^q(\Lie{g},M)$, we get that
$$ H^q(\Lie{g},M) \simeq H^0(\Lie{q},H^q(\Lie{i},M)) \oplus H^1(\Lie{q},H^{q-1}(\Lie{i},M)).$$
But
$$
H^0(\Lie{q},H^q(\Lie{i},M)) \simeq H^q(\Lie{i}, M)^{\langle y \rangle},\qquad H^1(\Lie{q},H^{q-1}(\Lie{i},M)) \simeq \frac{H^{q-1}(\Lie{i}, M)}{yH^{q-1}(\Lie{i}, M)}.$$\hfill $\Box$
We recall that $y\in \Lie{q}$ acts on $u \in Z^q(\Lie{i},M)$ by the formula
\begin{equation}\label{yaction}
y(u)(x_1, \ldots x_q) = yu(x_1, \ldots x_q) - \sum_{i=1}^q u(x_1, \ldots, [y,x_i], \ldots, x_q).
\end{equation}
The case where also $\Lie{i}$ is one-dimensional is particularly simple. Let $x\in \Lie{i}-\{0\}$. Then $[y,x]= ax$ for some $a\in {\mathbb C}$. We then have that
$$
H^0(\Lie{i},M) \simeq M^{\langle x \rangle},\qquad H^1(\Lie{i},M) \simeq M/xM.
$$
Here the last isomorphism is induced by mapping $u\in Z^1(\Lie{i},M)$ to $u(x)\in M$. We then see that
$$y(u)(x) = yu(x) - u([y,x]) = (y-a \mathop{\rm Id}\nolimits)u(x).$$
Thus we get that
$$H^1(\Lie{i},M)^{\langle y \rangle} = \left( M/xM\right)^{{\langle y-a \rangle}}.$$
\begin{lemma}\label{cod2}
In the case where $\Lie{g}$ is two-dimensional, with the notation as above, we have that
$$H^0(\Lie{g},M) \simeq M^{{\langle x,y \rangle}},\qquad H^1(\Lie{g}, M) \simeq \left( M/xM\right)^{{\langle y-a \rangle}} \oplus \left(M^{\langle x \rangle}/yM^{\langle x \rangle}\right)\,,$$
and
$$ H^2(\Lie{g}, M) \simeq \frac{M}{xM + (y-a)M}.$$
\end{lemma}
\noindent \textbf{Proof.}
Since we have already computed $H^0$ and $H^1$, we just need to compute $H^2$. From the Lemma \ref{q1Lie}, we get that
$$ H^2(\Lie{g}, M) \simeq H^1(\Lie{q}, H^1(\Lie{i},M)) \simeq \frac{M/xM}{(y-a)M/xM}$$
from which the result follows.
\hfill $\Box$
Let us now consider the case where $\Lie{i}$ has dimension two. Say $\Lie{i} = {\rm span}\{x_1,x_2\}$, where
$[x_2,x_1] = a x_1.$ We then have a description of the cohomology of $\Lie{i}$ from Proposition \ref{cod2}. However, in order to track the action of $y$, it is easier to work with the following model for $H^1$ of $\Lie{i}$
$$ H^1(\Lie{i},M) \simeq K/N,$$
where
\begin{eqnarray}
K & = & \ker(d_2 : M\times M \rightarrow M),\qquad d_2(m_1,m_2) = x_1m_2 - (x_2-a)m_1 \nonumber \\
N & = & \mathop{\fam0 Im}\nolimits(d_1 : M \rightarrow M\times M), \ \ d_1(m) = (x_1m,x_2m). \nonumber
\end{eqnarray}
If we now define $(a_{ij})_{i,j}$ by
\begin{equation}
[y,x_1] = a_{11} x_1 + a_{12} x_2,\qquad [y,x_2] = a_{21} x_1 + a_{22} x_2 \label{comut}
\end{equation}
then the Jacobi identity demands that
$$ aa_{12} = 0, \qquad a a_{22} = 0.$$
It is easy to check that the action of $y$ on $ H^1(\Lie{i},M)$ is induced from the following action of $y$ on $K$
$$ y(m_1,m_2) = \big((y-a_{11})m_1 -a_{12} m_2, (y-a_{22})m_2 - a_{21} m_1\big),$$
and using the above condition on $a$ and $ (a_{ij})_{i,j}$ one can easily verify that this action of $y$ indeed does preserve $K$ and that it maps $N$ to itself.
In order to compute the action of $y$ on $H^2(\Lie{i},M)$, we simply use (\ref{yaction}) to get that the action is induced by the following action of $y$ on $M$
$$y(m) = (y-a_{11}-a_{22})m.$$ Putting the above together, we get the following result.
\begin{lemma}\label{cod3}
In the case where $\Lie{g}$ is three-dimensional and has an ideal $\Lie{i}$ of codimension one, with the notation as above, we have that
$$H^0(\Lie{g},M) \simeq M^{{\langle x_1,x_2,y \rangle}}, \ \ H^1(\Lie{g}, M) \simeq \left( K/N\right)^{{\langle y \rangle}} \oplus \left(M^{{\langle x_1,x_2 \rangle}}/yM^{{\langle x_1,x_2 \rangle}}\right)\,,$$
and
$$ H^2(\Lie{g}, M) \simeq \left( \frac{M}{x_1M +x_2 M+ (y-a)M}\right)^{\langle y-a_{11}-a_{22}\rangle} \oplus \left(\frac{K}{N+ yK}\right)\,,$$
where $(a_{i,j})_{i,j}$ encode the commutation relations \eqref{comut}.
\end{lemma}
We can give an alternative description of the first cohomology group in the case where $\Lie{g}$ has dimension three with generators $x_1,x_2,y$ with structure constants as above, which is adapted to impose the extra linear constraints necessary to compute the tangent space to the moduli space of quantum (rather than quasi-) Airy structures. Let us introduce
\begin{eqnarray}
K^{(3)} & := & \ker\big(d^{(3)}_2 \,:\,M\times M\times M \longrightarrow M\times M\times M\big), \nonumber \\
d^{(3)}_2(m_1,m_2,m_3) & = & \big(- (x_2-a)m_1 + x_1m_2, -(y-a_{11})m_1 + a_{12}m_2 + x_1 m_3,a_{21}m_1 - (y-a_{22})m_2 + x_2m_3\big), \nonumber \\
N^{(3)} &:= & \mathop{\fam0 Im}\nolimits\big(d^{(3)}_1\,:\,M \longrightarrow M\times M \times M\big), \nonumber \\
d^{(3)}_1(m) & = & (x_1m,x_2m, ym). \nonumber
\end{eqnarray}
Then we have that
$$
H^1(\Lie{g},M) \simeq K^{(3)}/N^{(3)}.
$$
In this description a cocycle $u: \Lie{g} \rightarrow M$ is identified with $(m_1,m_2,m_3)\in K^{(3)}$ by
$$
u(x_1)=m_1,\qquad u(x_2) = m_2,\qquad u(y) = m_3.
$$
\subsection{Abelian cases}
From Lemma~\ref{basrem} and Section~\ref{S32} we know that when $V$ is a finite-dimensional abelian Lie algebra, the data of an abelian Lie subalgebra of $\mathfrak{sp}(T^*V)$ spanned by
$$
\varrho_1(e_i) = \left(\begin{array}{cc} -B^i & A^i \\ - C^i & (B^i)^{T}\end{array}\right)
$$
and of an arbitrary $D$ is equivalent to the data of a quantum Airy structure. To describe all abelian quantum Airy structure, we should take our matrices $\rho_1(e_i)$ belonging to a given maximal abelian Lie subalgebra of $\mathfrak{sp}(2n)$, and impose the constraint that $A$ is fully symmetric. If we wish to do so up to $\mathcal{G}_{V}$-equivalence, one should first list all isomorphism classes of maximal abelian Lie subalgebras of $\mathfrak{sp}(2n)$. The reference \cite{Winternitz} provides tools to do so, and achieves a complete classification in dimension two and three. Imposing further full symmetry of $A$, we describe below the five cases. The order in this list respects the order of the list of maximal abelian Lie subalgebras of \cite{Winternitz}. In dimension $n \geq 4$ and higher, there exists continuous families of isomorphism classes of maximal abelian Lie subalgebras of $\mathfrak{sp}(2n)$, and the classification becomes more intricate. Therefore, in general finite dimension, the classification of abelian quantum Airy structures -- and \textit{a fortiori} of quantum Airy structures -- modulo $\mathcal{G}_{V}$-action seems out of reach.
As commuting matrices can always be simultaneously trigonalized, we can always achieve $C^i = 0$ for all $i$ via $\mathcal{G}_{V}$-transformations. It is then implicit in the lists below that $C = 0$. Any two different cases in this list represent two non-equivalent quantum Airy structures for generic values of the parameters. Within each case, we have not indicated the fundamental domain of parameters for the residual $\mathcal{G}_{V}$-action.
\subsubsection{Dimension two}
\begin{itemize}
\item[\boxed{1}] $A = 0$ and $B^i$ diagonal.
\item[\boxed{2}] $A^1 = A^2 = 0$ and $B^i = \left(\begin{array}{cc} \alpha_i & 0 \\ 0 & 0 \end{array}\right)$.
\item[\boxed{3}] $A^1 = \left(\begin{array}{cc} \alpha & \beta \\ \beta & \gamma \end{array}\right)$, $A^2 = \left(\begin{array}{cc} \beta & \gamma \\ \gamma & \epsilon \end{array}\right)$, and $B^1 = B^2 = 0$.
\item[\boxed{4}] $A^1 = \left(\begin{array}{cc} \alpha & 0 \\ 0 & 0 \end{array}\right)$, $A^2 = \left(\begin{array}{cc} 0 & 0 \\ 0 & \beta \end{array}\right)$, $B^1 = \left(\begin{array}{cc} 0 & 0 \\ -\beta & 0 \end{array}\right)$, $B^2 = 0$.
\item[\boxed{5}] $A^1 = A^2 = 0$ and $B^i = \left(\begin{array}{cc} \alpha_i & \beta_i \\ 0 & \alpha_i \end{array}\right)$.
\end{itemize}
\subsubsection{Dimension three}
\begin{itemize}
\item[\boxed{$1$}] $A = 0$ and $B^i$ diagonal.
\item[\boxed{$2$}] $A^1 = A^3 = 0$, $A^2 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \alpha & 0 \\ 0 & 0 & 0 \end{array}\right)$ and $B^i = \left(\begin{array}{ccc} \beta_i & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & \gamma_i \end{array}\right)$.
\item[\boxed{$3$}] $A^1 = 0$, $A^2 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \alpha & \beta \\ 0 & \beta & \gamma \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \beta & \gamma \\ 0 & \gamma & \epsilon \end{array}\right)$, and $B^i = \left(\begin{array}{ccc} \kappa_i & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$.
\item[\boxed{$4$}] $A^1 = 0$, $A^2 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \alpha_2 & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & \gamma \end{array}\right)$, and $B^i = \left(\begin{array}{ccc} - \beta_i & 0 & 0 \\ 0 & 0 & 0 \\ 0 & -\alpha_i \end{array}\right)$ with $\alpha_1 = \alpha_3 = 0$.
\item[\boxed{$5$}] $A^i = 0$ and $B^i = \left(\begin{array}{ccc} \alpha_i & 0 & 0 \\ 0 & \beta_i & \gamma_i \\ 0 & 0 & \beta_i\end{array}\right)$.
\item[\boxed{$6$}] $A^1 = \left(\begin{array}{ccc} \alpha & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^2 = A^3 = 0$ and $B^i = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \beta_i & \gamma_i \\ 0 & 0 & \beta_i \end{array}\right)$.
\item[\boxed{$7$}] $A$ fully symmetric and $B^i = 0$.
\item[\boxed{$8$}] $A^1 = \left(\begin{array}{ccc} \alpha & \beta & 0 \\ \beta & \gamma & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^2 = \left(\begin{array}{ccc} \beta & \gamma & 0 \\ \gamma & \epsilon & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0\\ 0 & 0 & 0 \\ 0 & 0 & \kappa \end{array}\right)$,
\item[\phantom{\boxed{$8$}}] and $B^1 = B^2 = 0$, $B^3 = \left(\begin{array}{ccc} 0 & 0 & -\kappa \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$.
\item[\boxed{$9$}] $A^1 = \left(\begin{array}{ccc} \alpha & \beta & 0 \\ \beta & \gamma & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^2 = \left(\begin{array}{ccc} \beta & \gamma & 0 \\ \gamma & \epsilon & \kappa \\ 0 & \kappa & 0 \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \kappa & 0 \\ 0 & 0 & 0 \end{array}\right)$,
\item[\phantom{$9$}] and $B^1 = B^3 = 0$, $B^2 = \left(\begin{array}{ccc} 0 & 0 & -\kappa \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$.
\item[\boxed{$10$}] $A^1 = 0$, $A^2 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \alpha & \beta \\ 0 & \beta & 0 \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \beta & 0 \\ 0 & 0 & 0 \end{array}\right)$,
\item[\phantom{\boxed{$10$}}] and $B^1 = 0$, $B^2 = \left(\begin{array}{ccc} 0 & -\beta & -\alpha \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$, $B^3 = \left(\begin{array}{ccc} 0 & 0 & -\beta \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$.
\item[\boxed{$11$}] $A^1 = \left(\begin{array}{ccc} \alpha & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$, $A^2 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \beta & \beta + \gamma \\ \beta + \gamma & \gamma \end{array}\right)$, $A^3 = \left(\begin{array}{ccc} 0 & 0 & 0 \\ 0 & \beta + \gamma & \gamma \\ 0 & \gamma & -\beta \end{array}\right)$,
\item[\phantom{\boxed{$11$}}] and $B^1 = 0$, $B^2 = \left(\begin{array}{ccc} 0 & -\gamma & - \beta \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$, $B^3 = \left(\begin{array}{ccc} 0 & \beta & - \beta - \gamma \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{array}\right)$.
\item[\boxed{$12$}] $A = 0$ and $B^i = \left(\begin{array}{ccc} \alpha_i & 0 & \beta_i \\ 0 & \alpha_i & \gamma_i \\ 0 & 0 & \alpha_i \end{array}\right)$.
\item[\boxed{$13$}] $A = 0$ and $B^i = \left(\begin{array}{ccc} \alpha_i & \beta_i & \gamma_i \\ 0 & \alpha_i & 0 \\ 0 & 0 & \alpha_i \end{array}\right)$.
\item[\boxed{$14$}] $A = 0$ and $B^i = \left(\begin{array}{ccc} \alpha_i & \beta_i & \gamma_i \\ 0 & \alpha_i & \beta_i \\ 0 & 0 & \alpha_i \end{array}\right)$.
\end{itemize}
\subsection{Non-abelian two-dimensional Lie algebra}
In dimension two, there is a unique non-abelian Lie algebra up to isomorphism. It is the Lie algebra of affine transformations of $\mathbb{C}$, generated by $L_0$ and $L_1$ with the commutation relation
\begin{equation}
\label{affinLie}[L_0,L_1] = -L_1\,.
\end{equation}
We introduce a $\mathbb{Z}_2$-grading by declaring $t_i$ and $\partial_{t_i}$ to have degree $i$ for $i \in \{0,1\}$. We remark that, using the conjugation by an operator of the form $\exp\big(\tfrac{\hbar}{2} u_{a,b}\partial_{x_a}\partial_{x_b}\big)$, one can generically bring $L_0$ to a form where $C^0 = 0$.
\begin{proposition}
\label{d2Leonid} The complete list of classical Airy structures such that
\begin{itemize}
\item[$(\mathbf{1})$] $L_i$ has degree $i$ for $i \in \{0,1\}$,
\item[$(\mathbf{2})$] $L_0$ has no differential operators of order two, \textit{i.e.} $C^0 = 0$,
\end{itemize}
reads as follows
\begin{equation}
\begin{array}{crclcrcl} \mathbf{Ia} & L_0 & = & \hbar \partial_{x_0} - \alpha x_0^2 + \tfrac{\beta\alpha}{4}x_1^2 + 2\hbar x_0\partial_{x_0} + \hbar x_1\partial_{x_1} && L_1 & = & \hbar \partial_{x_1} + \tfrac{\alpha\beta}{2} x_0x_1 - \hbar \beta x_1\partial_{x_0} \\
\mathbf{Ib} & L_0 & = & \hbar \partial_{x_0} - \alpha x_0^2 - \beta x_1^2 - 2\hbar x_0\partial_{x_0} - \hbar x_1 \partial_{x_1} && L_1 & = & \hbar \partial_{x_1} - 2\beta x_0x_1 - 2\hbar x_0\partial_{x_1} \\
\mathbf{Ic} & L_0 & = & \hbar \partial_{x_0} - \alpha x_0^2 - \hbar(\beta + 1)x_0\partial_{x_0} - \hbar\beta x_1\partial_{x_1} && L_1 & = & \hbar \partial_{x_1} - \hbar(\beta + 1)x_0\partial_{x_1} \\
\mathbf{IIa} & L_0 & = & \hbar \partial_{x_0} + \tfrac{4}{\alpha}x_0^2 - \beta x_1^2 + 2\hbar x_0\partial_{x_0} - \hbar x_1\partial_{x_1} && L_1 & = & \hbar \partial_{x_1} - 2\beta x_0x_1 - \hbar \alpha\beta x_1\partial_{x_0} - 2\hbar x_0\partial_{x_1} - \hbar^2 \alpha \partial_{x_0}\partial_{x_1} \\
\mathbf{IIb} & L_0 & = & \hbar \partial_{x_0} + \frac{\alpha^2}{\beta}x_0^2 + \hbar \alpha x_0\partial_{x_0} + \hbar (1 - \alpha) x_1\partial_{x_1} && L_1 & = & \hbar \partial_{x_1} - \hbar \alpha x_0\partial_{x_1} - \hbar^2\gamma\partial_{x_0}\partial_{x_1} \end{array} \nonumber
\end{equation}
where $\alpha,\beta,\gamma \in \mathbb{C}$ (and non-zero when they appear in the denominator). Quantum Airy structures are obtained from these by adding an arbitrary constant $-\hbar D^0$ to $L_1$, while $D^1 = 0$. \hfill $\Box$
\end{proposition}
This result is obtained by inserting the form of the most general quantum Airy structure with properties (\textbf{1}) and (\textbf{2}) into the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations, and solving the resulting equations. The \textbf{D} relation can be analyzed separately, and we indeed find that quantum Airy structures based on \eqref{affinLie} must have $D^1 = 0$, while $D^0$ is unconstrained. We omit the details of this proof. The groups \textbf{I} and \textbf{II} correspond to $C = 0$ or $C^1 \neq 0$.
Now we want to consider the existence of deformations modulo $\mathcal{G}_{V}$ of these structures. We remark that this may not preserve the homogeneity condition (\textbf{1}). In particular, with the following transformations -- which belong to $\mathcal{G}_{V}$
\begin{equation}
t_0\to e^{\xi_1}t_0,\,\,\,\partial_0\to e^{-\xi_1}\partial_0,\,\,\, t_1\to e^{\xi_2}t_1\,\,\, \partial_1\to e^{-\xi_2}\partial_1,\,\,\,\hbar\to e^{-\xi_1}\hbar,\,\,\,L_1\to e^{\xi_2-\xi_1}L_1, \qquad \xi_i \in \mathbb{C}
\label{scaling}
\end{equation}
and assuming the parameters $\alpha,\beta,\gamma$ non-zero, and in each case one can use them to set the parameters to $1$, except in case \textbf{Ic} where one can only set $\alpha = 1$ while $\beta$ remains a free parameter.
Now let us assume that $\alpha,\beta,\gamma$ are generic. We compute the Lie algebra cohomologies governing the deformations of quantum Airy structures modulo $\mathcal{G}_{V}$ using Lemma~\ref{cod2}, which in our case applies with $x = L_1$, $y = L_0$ and $a = -1$. The result is
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|}
\hline & \textbf{Ia} & \textbf{Ib} & \textbf{Ic} & \textbf{IIa} & \textbf{IIb} \\
\hline
\hline $\dim H^0$ & $1$ & $1$ & $2$ & $1$ & $2$ \\
\hline $\dim H^1$ & $1$ & $1$ & $2$ & $1$ & $2$ \\
\hline $\dim H^2$ & $0$ & $0$ & $0$ & $0$ & $0$ \\
\hline
\end{tabular}
\end{center}
We observe that $H^2$ always vanishes, \textit{i.e.} there are no obstructions to deformations. As the conjugation by scalars acts trivially on $\mathcal{D}$, $H^0$ is at least one-dimensional. In the two cases where $H^0$ has dimension two, the extra generator in fact has a trivial action on $H^1$, so $\dim H^1$ gives the number of independent deformations of the corresponding quantum Airy structure. In all cases, the value of $D^0$ gives a one-parameter deformation. In \textbf{Ic}, $\beta$ gives the second deformation parameter -- one sees for instance that it cannot be gauged out by the scaling transformations \eqref{scaling}. In case \textbf{IIb}, one finds that the second one-parameter deformation (called $t$) is
\begin{equation}
\label{IIbt} \textbf{IIb}\,\,:\,\, L_i^{(t)} = L_0 + tL_i',\qquad L_0' = \hbar (1 - 2\alpha)x_0\partial_{x_1},\qquad L_1' = \hbar^2\beta \partial_{x_1}^2
\end{equation}
Although the cohomology groups we computed only give the number of independent deformations within quasi-Airy structures, we actually find in these five examples that they are always realized within Airy structures.
The partition functions for each of these quantum Airy structures -- with an arbitrary $D^0$ -- can be computed by hand. As the group \textbf{I} has $C = 0$, we can also use Lemma~\ref{Cequal0}.
\begin{eqnarray}
Z_{{\rm Ia}} & = & (1 + 2x_0 + \beta x_1^2)^{\frac{a}{8\hbar} + \frac{D^0}{2}}\exp\big\{\hbar^{-1}\big(\tfrac{\alpha x_0(x_0 - 1)}{4} - \tfrac{\alpha\beta x_1^2}{8}\big)\big\} \nonumber \\
Z_{{\rm Ib}} & = & (1 - 2x_0)^{-\tfrac{\alpha}{8\hbar} -\frac{D^0}{2}}\exp\big\{\hbar^{-1}\big(-\tfrac{\alpha x_0(x_0 + 1)}{4}\tfrac{\beta x_0x_1^2}{1 - 2x_0}\big)\big\} \nonumber \\
Z_{{\rm Ic}} & = & (1 - (\beta + 1)x_0)^{-\frac{\alpha}{\hbar(\beta + 1)^3} - \frac{D^0}{\beta + 1}}\exp\big(-\tfrac{\alpha x_0(2 + (\beta + 1)x_0)}{2\hbar ( \beta + 1)^2}\big) \nonumber \\
Z_{{\rm IIa}} & = & (1 + 2x_0)^{ - \frac{1}{4\hbar\alpha} + \frac{D^0}{4}}x_1^{\frac{1}{2\alpha\hbar} -\frac{D^0}{2}} \exp\big\{\hbar^{-1}\big(\tfrac{x_0(1 - x_0)}{
\alpha} -\tfrac{\beta x_1^2}{2}\big)\big\}\,J_{- \frac{1}{2\alpha\hbar} + \frac{D^0}{2}}\Big[\hbar^{-1}x_1\big(\tfrac{\beta(1 + 2x_0)}{\alpha}\big)^{\frac{1}{2}}\Big] \nonumber \\
Z_{{\rm IIb}} & = & (1 + \alpha x_0)^{-\frac{1}{\hbar\alpha\beta} + \frac{D^0}{\alpha}}\,\exp\big(\tfrac{x_0(2 - \alpha x_0)}{\beta\hbar}\big) \nonumber
\end{eqnarray}
The most interesting case is $Z_{{\rm IIa}}$, where we see an appearance of the Bessel function $J_{\nu}(z)$. It is characterized by
$$
\big(z^2\partial_{z}^2 + z\partial_{z} + (z^2 - \nu^2)\big)J_{\nu}(z) = 0,\qquad J_{\nu}(z) := \sum_{m \geq 0} \frac{\Gamma(\nu + 1)}{\Gamma(m + \nu + 1)}\,\frac{(-1)^{m}}{m!}\,\big(\tfrac{z}{2}\big)^{2\nu + m}\,.
$$
Our normalization by a constant prefactor of $J_{\nu}(z)$ is not the conventional one, but we made it so that $J_{\nu}(z) = z^{2\nu}(1 + O(z))$ when $z \rightarrow 0$.
In the case \textbf{IIb}, the partition function for the one-parameter deformation \eqref{IIbt} is in fact independent of the parameter $t$, and independent of $x_1$.
\subsection{The case $n=3$.}
When considering isomorphism classes of three-dimensional non-trivial Lie algebra over $\mathbb{C}$, one finds \cite{Winternitzbook} four rigid cases and a one-parameter family.
\begin{itemize}
\item[$\bullet$] $\mathfrak{sl}_{2}$, which was treated in Theorem~\ref{sl2th}.
\item[$\bullet$] The direct sum of the non-abelian two-dimensional algebra and the abelian Lie algebra of dimension $1$. We do not study this direct sum case.
\item[$\bullet$] The Heisenberg Lie algebra $[y_1,y_2] = y_3$, $[y_1,y_3] = [y_2,y_3] = 0$, which we do not treat.
\item[$\bullet$] The Lie algebra $[y_1,y_2] = y_2$ and $[y_1,y_3] = y_2 + y_3$, which we also do not treat.
\item[$\bullet$] The Lie algebra $\mathfrak{l}_{q}$ defined by $[y_1,y_2] = y_2$, $[y_1,y_3] = qy_3$, $[y_2,y_3] = 0$, for $q \in \mathbb{C}^*$. We have $\mathfrak{l}_{q} \simeq \mathfrak{l}_{r}$ if and only if $q = r$ or $qr = 1$. For $q = -1$, this is the Lie algebra of the group of isometries of euclidean $\mathbb{R}^2$.
\end{itemize}
We will not attempt here at the classification of Airy structures supported by each of these Lie algebras. We rather look for a non-trivial example of quantum Airy structure based on $\mathfrak{l}_{q}$, keeping as many non-zero elements as possible, which illustrates the non-triviality of the problem of exhibiting finite-dimensional quantum Airy structures. We rename the variables $(x_0,x_0',x_1)$, and assign a $\mathbb{Z}_{2}$-degree $0$ to $x_0,x_0'$, and $1$ to $x_1$. We postulate the commutations relation of $\mathfrak{l}_{q}$ in the following form
$$
[L_0,L_1] = L_1,\qquad [L_0,L_0'] = qL_0'\qquad [L_0',L_1] = 0,\qquad \qquad {\rm that}\,\,{\rm is}\,\,L\,:\,\,(y_1,y_2,y_3) \rightarrow (L_0,L_1,L_0')
$$
and look for a quantum Airy structure such that $L_0,L_0'$ have degree $0$ and $L_1$ has degree $1$. For the same Lie algebra, computations shows that the other choice of grading $(y_1,y_2,y_3) \mapsto (L_0,L_1,L_1')$ does not support non-trivial Airy structures.
\begin{lemma}
\label{Ldim3} Let $\zeta$ be a root of $P(\zeta) := 2\zeta^3 - 2\zeta^2 + 3\zeta - 1$ and $\alpha$ a complex parameter. The matrices, which we write with respect to the basis $(e_0,e_0',e_1)$, $C^0 = 0$ and
\begin{eqnarray}
A^0 = \left(\begin{array}{ccc} \tfrac{2}{3}(-5\zeta + 3)2 & \tfrac{4}{3}(-2\zeta^2 + \zeta - 1) & 0 \\ \tfrac{4}{3}(-2\zeta^2 + \zeta - 1) & \tfrac{2}{3}(-\zeta + 1) & 0 \\ 0 & 0 & 2\alpha \end{array}\right) && B^0 = \left(\begin{array}{ccc} 1 -3\zeta & 3(-2\zeta^2 + \zeta - 1) & 0 \\ 3(-2\zeta^2 + \zeta - 1) & 3\zeta + 1 & 0 \\ 0 & 0 & 1 \end{array}\right) \nonumber \\
A'^0 = \left(\begin{array}{ccc} \tfrac{4}{3}(-2\zeta^2 + \zeta - 1) & \tfrac{2}{3}(-\zeta + 1) & 0 \\ \tfrac{2}{3}(-\zeta + 1) & 0 & 0 \\ 0 & 0 & -\alpha(2\zeta^2 + 1) \end{array}\right) &&
B'^0 = \left(\begin{array}{ccc} 3(-2\zeta^2 + \zeta - 1) & 3\zeta - 1 & 0 \\ -\zeta + 1 & -2\zeta^2 + \zeta - 1 & 0 \\ 0 & 0 & -2\zeta^2 - 1 \end{array}\right) \nonumber \\
A^1 = \left(\begin{array}{ccc} 0 & 0 & 2\alpha \\ 0 & 0 & -\alpha(2\zeta^2 + 1) \\ 2\alpha & -\alpha(2\zeta^2 +1) & 0 \end{array}\right) && B^1 = \left(\begin{array}{ccc} 0 & 0 & 2 \\ 0 & 0 & -2\zeta^2 - 1 \\ 6\alpha & 3\alpha(2\zeta^2 + 1) & 0 \end{array}\right) \nonumber \\
C'^0 = \left(\begin{array}{ccc} 6(-2\zeta^2 + \zeta - 1) & 6\zeta & 0 \\ 6(2\zeta^2 - \zeta + 1) & 0 & 0 \\ 0 & 0 & -\alpha^{-1}(2\zeta^2 + 1) \end{array}\right) && C^1 = \left(\begin{array}{ccc} 0 & 0 & 6 \\ 0 & 0 & 3(2\zeta^2 + 1) \\ 6 & 3(2\zeta^2 + 1) & 0 \end{array}\right) \nonumber
\end{eqnarray}
together with
$$
D^0\,\,{\rm arbitrary},\qquad D'^{0} = \tfrac{1}{2}\,{\rm Tr}\,B'^{0} = \tfrac{1}{2}(-5\zeta^2 + 4\zeta - 5),\qquad D^1 = 0
$$
define a quantum Airy structure based on the Lie algebra $\mathfrak{l}_{-2}$. It has
$$
\dim H^i_{L}(V,\mathcal{D}_{V}) = \left\{\begin{array}{rcl} 1 & & i = 0 \\ 2 & & i = 1 \\ 1 & & i = 2 \end{array}\right. \,.
$$
The stabilizer $\mathcal{G}_{V}(L)$ is trivial. $D^0$ is the only deformation of this quantum Airy structure.
\end{lemma}
\noindent \textbf{Proof.} One can check by direct computation that the assignments in this Lemma indeed satisfies the desired commutation relations. We have however opted to provide some details explaining how we were led to this quantum Airy structure, in the hope that it may lead to the construction of other quantum Airy structures by similar techniques.
The main task is to find $(A,B,C)$ defining a classical Airy structure, \textit{i.e.} such that
$$
[\tilde{\varrho}_{1}(e_0),\tilde{\varrho}_{1}(e_1)] = \tilde{\varrho}_{1}(e_1),\qquad [\tilde{\varrho}_{1}(e_0),\tilde{\varrho}_1(e_0')] = q\tilde{\varrho}_{1}(e_0'),\qquad [\tilde{\varrho}_{1}(e_0'),\tilde{\varrho}_{1}(e_1)] = 0\,,
$$
with $\tilde{\rho}_{1}$ of the form \eqref{rho1ti}. As our goal is not to find all possible solutions, we are going to postulate certain properties, which eventually lead to a solution.
\vspace{0.1cm}
\noindent \textbf{1st postulate.} $L_0,L_0'$ have degree $0$, $L_1$ has degree $1$. \\
\noindent \textbf{2nd postulate.} $C^0 = 0$ --- for generic operators this can be achieved by a $\mathcal{G}_{V}$ transformation.
\vspace{0.1cm}
The matrices then take the form
\begin{align}\label{block}
&B^0=\left(\begin{array}{c|c}
\tilde{B} & 0 \\
\hline
0 & b_{33}
\end{array} \right), \quad B'^0=\left(\begin{array}{c|c}
\tilde{B}' & 0 \\
\hline
0 & b'_{33}
\end{array} \right), \quad C'^0=\left(\begin{array}{c|c}
2\tilde{C}' & 0 \\
\hline
0 & 2c'_{33}
\end{array} \right), \quad A^0=\left(\begin{array}{c|c}
2\tilde{A} & 0 \\
\hline
0 & 2a_{33}
\end{array} \right), \nonumber\\
& A'^0=\left(\begin{array}{c|c}
2\tilde{A}' & 0 \\
\hline
0 & 2a'_{33}
\end{array} \right), \quad B_1=\left(\begin{array}{c|c}
0 & {\mathbf v}^{\text{T}} \\
\hline
{\mathbf u} & 0
\end{array} \right), \quad C^1=\left(\begin{array}{c|c}
0 & 2{\mathbf c}^{\text{T}} \\
\hline
2{\mathbf c} & 0
\end{array} \right), \quad A^1=\left(\begin{array}{c|c}
0 & 2{\mathbf a}^{\text{T}} \\
\hline
2{\mathbf a} & 0
\end{array} \right), \\
& \tilde{A}^{\text{T}}=\tilde{A},\ (\tilde{A}')^{\text{T}}=\tilde{A}', \ (\tilde{C}')^{\text{T}}=\tilde{C}'\quad{\hbox{and}}\quad a_{12}=a'_{11},\ a_{22}=a'_{12},\ a_{33}={\mathbf a}_1,\
a'_{33}={\mathbf a}_2.\nonumber
\end{align}
where $\tilde{A},\tilde{A}',\tilde{C},\tilde{C}'$ are $2 \times 2$ matrices, and $\mathbf{a},\mathbf{c},\mathbf{u},\mathbf{v}$ are row vectors of size two. The relation \eqref{f2B} between the structure constants of the Lie algebra and antisymmetric part of $B$ results in
\begin{equation}
b'_{1,i}-b_{2,i}=q\delta_{i,2}, \ i=1,2;\ \ {\mathbf v}_2=b'_{33}, \ \ {\mathbf v}_1=b_{33}+1.
\label{n3.1}
\end{equation}
We first address the relation $[\tilde{\varrho}_{1}(0),\tilde{\varrho}_1(e'_0)]=q\tilde{\varrho}(e_0')$. It results in the following relations
\begin{align}
&-\tilde{C}'\tilde{B}-\tilde{B}^{\text{T}}\tilde{C}'=q\tilde{C}',\qquad -2c'_{33}b_{33}=qc'_{33}\,,\label{L001}\\
& [\tilde{B},\tilde{B}']-4\tilde{A}\tilde{C}'=q\tilde{B}',\qquad -4c'_{33}a_{33}=q b'_{33}\,,\label{L002}\\
& \tilde{B}\tilde{A}'+\tilde{A}'\tilde{B}^{\text{T}}-\tilde{B}'\tilde{A}-\tilde{A}(\tilde{B}')^{\text{T}}=q\tilde{A}',\qquad 2b_{33}a'_{33}-2b'_{33}a_{33}=q a'_{33}\,.\label{L003}
\end{align}
\vspace{0.1cm}
\noindent \textbf{3rd postulate.} We assume $C'^0$ is nondegenerate, \textit{i.e.} $\det \tilde{C}'\ne 0$ and $c'_{33}\ne 0$.
\vspace{0.1cm}
Then, from the second equation in (\ref{L001}), we have that $b_{33}=-\tfrac{q}{2}$. From (\ref{L003}) we then obtain
$$
b'_{33}a_{33}=-qa'_{33}.
$$
The first equation in (\ref{L001}) implies that $\tilde{B} = S-\tfrac{q}{2}\mathbb {\rm Id}$, where the matrix $S$ satisfies the equation
\begin{equation}
\label{STCeq} \tilde{C}'S=-S^{\text{T}}\tilde{C}',
\end{equation}
which admits a one-parameter family of solutions. So, we express $\tilde{B}$ in terms of $\tilde{C}'$ and the parameter $t$
\begin{equation}
\label{B} \tilde{B}=-\tfrac{q}{2} {\rm Id} + t\left(\begin{array}{cc}
-c'_{12} & c'_{11} \\
-c'_{22} & c'_{12}
\end{array}\right)\,.
\end{equation}
From (\ref{n3.1}) we then have $b'_{11}=-tc'_{22}$, $b'_{12}=tc'_{12}+\tfrac{q}{2}$, $\mathbf v_1=1-\tfrac{q}{2}$, and $\mathbf v_1=b'_{33}$. Due to general scaling transformations in $\mathcal{G}_{V}$ for even and odd variables, we have two \emph{free parameters}: $t$ and $a_{33}$.
\vspace{0.1cm}
\noindent \textbf{4th postulate} $t = 1$ -- we however keep $a_{33} = 2\alpha$ arbitrary.
We now solve (\ref{L002}) with respect to $\tilde{A}$. We first rewrite it in terms of $S$
$$
[S,\tilde{B}']-4\tilde{A}\tilde{C}'=q\tilde{B}'.
$$
Because $-S(\tilde{C}')^{-1}=(\tilde{C}')^{-1}\tilde{S}^{\text{T}}$ from \eqref{STCeq}, multiplying by $(\tilde{C}')^{-1}$ from the right, we obtain that
\begin{equation}
4\tilde{A}=S\tilde{B}'(\tilde{C}')^{-1}+\tilde{B}'(\tilde{C}')^{-1}S^{\text{T}}-q\tilde{B}'(\tilde{C}')^{-1}.
\label{A->SBC}
\end{equation}
Introducing $R:=\tilde{B}'(\tilde{C}')^{-1}$, we obtain that $4\tilde{A}=SR+RS^{\text{T}}-qR$. Another useful information can be extracted from the fact that
$\tilde{A}$ is symmetric: splitting $R$ into symmetric and skew-symmetric parts, $R=R_{\text{sym}}+R_{\text{asym}}$, we observe that
$$
SR_{\text{asym}}+R_{\text{asym}}S^{\text{T}}-qR_{\text{asym}}=0,
$$
wherefrom, accounting for an explicit form of $S$, the first two terms are mutually canceled, and we obtain that for $q\ne 0$,
$$
R_{\text{asym}}=0,
$$
that is,
\begin{equation}
\tilde{B}'(\tilde{C}')^{-1}=(\tilde{C}')^{-1}(\tilde{B}')^{\text{T}},\ \ \hbox{or}\ \ \tilde{C}'\tilde{B}'=(\tilde{B}')^{\text{T}}\tilde{C}',
\label{L00BC}
\end{equation}
which in the component form results in the condition
\begin{equation}
0=c'_{12}(c'_{11}+c'_{22})-c'_{11}+c'_{12}b'_{22}-c'_{22}b'_{21}\,.
\label{eq-2}
\end{equation}
Because $\tilde{B}=S-\tfrac{q}{2}{\rm Id}$, equation (\ref{L003}) in terms of the matrices $S$, $\tilde{A}$, $\tilde{A}'$, and $\tilde{B}'$ reads
$$
S\tilde{A}'+\tilde{A}'S^{\text{T}}-\tilde{B}'\tilde{A}-\tilde{A}(\tilde{B}')^{\text{T}}=2q\tilde{A}'.
$$
We solve this equation as follows. It follows from this equation that
\begin{equation}
\tilde{A}'[S^{\text{T}}-q{\rm Id}]=\tilde{B}'\tilde{A}+S_A,\qquad S_A=\left(\begin{array}{cc}
0 & z \\
-z & 0
\end{array}\right)
\label{A-A}
\end{equation}
for some $z \in \mathbb{C}$.
Consider now the equation $[\tilde{\varrho}_{1}(e_0),\tilde{\varrho}_{1}(e_1)]=\tilde{\varrho}_{1}(e_1)$. Substituting the form (\ref{block}) for the respective matrices, we obtain
\begin{align}
&(\tilde{B}-(b_{33} + 1){\rm Id}){\mathbf v}^{\text{T}}-4\tilde{A}{\mathbf c}^{\text{T}}=0, \label{L011}\\
&{\mathbf u}(\tilde{B}-(b_{33} + 1){\rm Id})-4a_{33}{\mathbf c}=0, \label{L012}\\
& {\mathbf c}\bigl[ \tilde{B}+(b_{33} + 1){\rm Id}\bigr]=0, \label{L013}\\
&\bigl[\tilde{B}+(b_{33} - 1){\rm Id}\bigr]{\mathbf a}^{\text{T}}-\tilde{A}{\mathbf u}^{\text{T}}-a_{33}{\mathbf v}^{\text{T}}=0. \label{L014}
\end{align}
\vspace{0.1cm}
\noindent\textbf{5th postulate.} $\mathbf c\ne 0$.
\vspace{0.1cm}
Then, $\mathbf{c}$ has to be in the kernel of the matrix $\bigl[ S+ (1-q){\rm Id}\bigr]$.
From the form of the matrix $S$ we have that
\begin{equation}
\det(S+\lambda{\rm Id})=\lambda^2-[c'_{12}]^2+c'_{11}c'_{22}.
\label{DET}
\end{equation}
\vspace{0.1cm}
\noindent \textbf{6th postulate.} $A_{33}\ne 0$, and $\pm 1$ are not eigenvalues of $S$. From \eqref{DET}, this implies $q \notin \{0,2\}$.
\vspace{0.1cm}
The above system of equations become
\begin{align}
&[S-{\rm Id}]{\mathbf v}^{\text{T}}=4\tilde{A}{\mathbf c}^{\text{T}}, \label{L011a}\\
&-{\mathbf u}[S+ {\rm Id}]=4a_{33}{\mathbf c}, \label{L012a}\\
& {\mathbf c}\bigl[S+(1-q){\rm Id}\bigr]=0, \label{L013a}\\
&[S-(1+q){\rm Id}]{\mathbf a}^{\text{T}}=\tilde{A}{\mathbf u}^{\text{T}}+a_{33}{\mathbf v}^{\text{T}}. \label{L014a}
\end{align}
Acting by $[S+(1-q){\rm Id}]$ on (\ref{L012a}) from the left, we obtain zero on the both sides; because $S+ {\rm Id}$ is non-degenerate by our 6th postulate, and commutes with $[S+(1-q){\rm Id}]$, the vector $\mathbf u$ is also a zero vector of $[S+(1-q){\rm Id}]$ and it is therefore nonzero itself and proportional to $\mathbf c$, so
\begin{equation}
\mathbf u=-\tfrac{4a_{33}}{q}\mathbf c.
\label{L01b}
\end{equation}
We next transform (\ref{L014a}) using (\ref{L01b}) and (\ref{L011a}) we have that
$$
\tilde{A}{\mathbf u}^{\text{T}}=-\frac{4a_{33}}{q}\tilde{A}{\mathbf c}^{\text{T}}=-\frac {a_{33}}{q}[S-{\rm Id}]{\mathbf v}^{\text{T}}
$$
and (\ref{L014a}) becomes
\begin{equation}
[S-(q+1){\rm Id}]\Bigl[{\mathbf a}^{\text{T}}+\frac{a_{33}}{q}{\mathbf v}^{\text{T}}\Bigr]=0.
\end{equation}
\vspace{0.1cm}
\noindent \textbf{7th postulate.} $\det(S-(q+1){\rm Id})\ne 0$.
\vspace{0.1cm}
Then
\begin{equation}
\label{LLL}\mathbf a=-\tfrac{a_{33}}{q}\mathbf v=
\left(-\tfrac{a_{33}}{q}\mathbf v_1,\ -\tfrac{a_{33}}{q}\mathbf v_2 \right)= \left(-\tfrac{a_{33}}{q}\Bigl(1-\tfrac {q}{2}\Bigr),\ +a'_{33} \right).
\end{equation}
The condition $a'_{33}=\mathbf a_2$ is satisfied automatically, whereas the condition $a_{33}=\mathbf a_1$ implies
\begin{equation}
\boxed{q=-2.}
\label{b-2}
\end{equation}
From now on, we set $q=-2$ in all further calculations. Note that thus the chosen $q$ implies the non-degeneracy of the above determinants $\det(S+\lambda {\rm Id})$ with $\lambda =-1,1$, and with $\lambda=1+q=-1$.
Because $S+3{\rm Id}$ has a nonzero kernel, we have that
\begin{equation}
\det (S+3{\rm Id})=9-[c'_{12}]^2+c'_{11}c'_{22}=0.
\label{eq-1}
\end{equation}
We now express $a_{ij}$ from the formula (\ref{A->SBC}). After a few computations we obtain
\begin{align}
4a_{11}&= -\tfrac{1}{9} \Bigl[2(1-c'_{12})\bigl( -[c'_{22}]^2 -[c'_{12}]^2 +c'_{12}\bigr)+2c'_{11}\bigl( c'_{12}(c'_{11}+c'_{22})-c'_{11}\bigr) \Bigr], \label{eq-3}\\
4a_{12}&=-\tfrac{1}{9} \Bigl[c'_{11}(-c'_{12}b'_{21}+c'_{11}b'_{22}) -c'_{22}\bigl( -[c'_{22}]^2 -[c'_{12}]^2 +c'_{12}\bigr)+2\bigl( c'_{12}(c'_{11}+c'_{22})-c'_{11}\bigr) \Bigr], \label{eq-4}\\
4a_{22}&=-\tfrac{1}{9} \Bigl[2(c'_{12}+1)(-c'_{12}b'_{21}+c'_{11}b'_{22}) -2c'_{22}\bigl( c'_{12}(c'_{11}+c'_{22})-c'_{11}\bigr) \Bigr]. \label{eq-5}
\end{align}
Finally, let us consider the last remaining condition $[\tilde{\varrho}_1(e_0'),\tilde{\varrho}_{1}(e_1)]=0$. We obtain the system
\begin{align*}
&[\tilde{B}'-b'_{33}{\rm Id}]{\mathbf v}^{\text{T}}=4\tilde{A}'{\mathbf c}^{\text{T}}-4c'_{33}{\mathbf a}^{\text{T}},\\
&{\mathbf u}[\tilde{B}'-b'_{33}{\rm Id}]=4\mathbf a \tilde{C}'-4a'_{33}{\mathbf c},\\
&{\mathbf c}\bigl[\tilde{B}'+b'_{33}{\rm Id}\bigr]=\mathbf v \tilde{C}'+\mathbf u c'_{33},\\
&[\tilde{B}'+b'_{33}{\rm Id}]{\mathbf a}^{\text{T}}=\tilde{A}'{\mathbf u}^{\text{T}}+a'_{33}{\mathbf v}^{\text{T}}.
\end{align*}
Because $\mathbf a+\tfrac{a_{33}}{q}\mathbf v=0$, $2\mathbf u=4a_{33}\mathbf c$, $q=-2$, and $-4a_{33}c'_{33}=qb'_{33}$, the above
system simplifies dramatically. Two out of four equations become redundant, and the only nontrivial equations that survive are
\begin{align}
& \mathbf u \tilde{B}'=4\mathbf a \tilde{C}', \label{L11a}\\
& \tilde{B}' {\mathbf a}^{\text{T}}=\tilde{A}' {\mathbf u}^{\text{T}}. \label{L11b}
\end{align}
We add to this system equation (\ref{L014a}), which, upon implementing (\ref{LLL}) becomes
\begin{equation}
\mathbf a[S^{\text{T}}-{\rm Id}]=\mathbf u \tilde{A}.
\label{MMM}
\end{equation}
\vspace{0.1cm}
\noindent \textbf{8th postulate.} $\det B'\ne 0$.
\vspace{0.1cm}
Recall that $\mathbf u[S+3{\rm Id}]=\mathbf c[S+3{\rm Id}]=0$. From equations (\ref{L11a}) and
(\ref{MMM}), we obtain that
$$
\mathbf a\bigl[ S^{\text{T}}-{\rm Id} -4\tilde{C}'(\tilde{B}')^{-1}\tilde{A}\bigr]=0.
$$
Substituting (\ref{A->SBC}), we obtain
\begin{align*}
&\mathbf a\Big( S^{\text{T}}- {\rm Id} -\tilde{C}'(\tilde{B}')^{-1}\bigl[S\tilde{B}'(\tilde{C}')^{-1}+\tilde{B}'(\tilde{C}')^{-1}S^{\text{T}}+2\tilde{B}'(\tilde{C}')^{-1}\bigr]\Big) \\
=&\mathbf a \bigl[-3{\rm Id}-\tilde{C}'(\tilde{B}')^{-1}S\tilde{B}'(\tilde{C}')^{-1}\bigr]=\mathbf a \tilde{C}'(\tilde{B}')^{-1}[-3{\rm Id}-S]\tilde{B}'(\tilde{C}')^{-1}\\
=&-\tfrac{1}{4} \mathbf u[S+3{\rm Id}]\tilde{B}'(\tilde{C}')^{-1}=0,
\end{align*}
by (\ref{L013}), so the condition (\ref{L11a}) is satisfied automatically.
From (\ref{L11a}) and (\ref{L11b}), we obtain that
\begin{equation}
\mathbf a\bigl[ (\tilde{B}')^{\text{T}} -4\tilde{C}'(\tilde{B}')^{-1}\tilde{A}'\bigr]=0.
\label{QQQ}
\end{equation}
Because $\tilde{A}'[2{\rm Id}+S^{\text{T}}]=\tilde{B}'\tilde{A}+S_{A}$, see (\ref{A-A}), multiplying from the right by $[2{\rm Id}+S^{\text{T}}]$,
we can perform the chain of transformations -- where we have that $\tilde{B}'(\tilde{C}')^{-1}$ is symmetric --
\begin{align*}
0&=\mathbf a\Big((\tilde{B}')^{\text{T}}[2{\rm Id}+S^{\text{T}}] -4\tilde{C}'\tilde{A} -4\tilde{C}'(\tilde{B}')^{-1}S_A\Big)\\
&=\mathbf a \tilde{C}'(\tilde{B}')^{-1}\Big(\tilde{B}'\bigl[(\tilde{C}')^{-1}(\tilde{B}')^{\text{T}}\bigr][2{\rm Id}+S^{\text{T}}] -4\tilde{B}'\tilde{A} -4S_A\Big)
\\
&=\mathbf a \tilde{C}'(\tilde{B}')^{-1}\Big((\tilde{B}')^2(\tilde{C}')^{-1}[2{\rm Id}+S^{\text{T}}] -\tilde{B}'\bigl[S\tilde{B}'(\tilde{C}')^{-1}+\tilde{B}'(\tilde{C}')^{-1}S^{\text{T}}+2\tilde{B}'(\tilde{C}')^{-1}\bigr] -4S_A\Big)\\
&=\tfrac{1}{4} \mathbf u \bigl[-\tilde{B}'S\tilde{B}'(\tilde{C}')^{-1}-4S_A\bigr] \\
&=\tfrac{1}{4} \mathbf u \bigl[-\tilde{B}'S-4S_A\tilde{C}'(\tilde{B}')^{-1}\Big)\tilde{B}'(\tilde{C}')^{-1},
\end{align*}
that is, both $S+3{\rm Id}$ and $\tilde{B}'S+4S_A\tilde{C}'(\tilde{B}')^{-1}$ share the same null vector $\mathbf u$. This imposes two more constraints on
the matrix elements of $\tilde{B}'S\tilde{B}'+4S_A\tilde{C}'$. Note that $S_A\tilde{C}'=-z S^{\text{T}}$, so both columns of the matrix $\tilde{B}'S\tilde{B}'-4zS^{\text{T}}$
must be proportional
to $\big(3-c'_{12},-c'_{22}\big)^{{\rm T}}$.
\vspace{0.1cm}
\noindent \textbf{9th postulate.} The row vector $(3-c'_{12},c'_{22})$ is non-zero.
\vspace{0.1cm}
Hence we obtain
\begin{equation}
-c_{22}'\bigl(-(c'_{22})^2+(9-c'_{12})b'_{21}+4zc'_{12}\bigr) =\bigl(b'_{21}c'_{12}c'_{22}+b'_{22}(c'_{22})^2+[b'_{21}]^2 c'_{11}+b'_{21}b'_{22}c'_{12}-4zc'_{11}\bigr)(3-c'_{12})
\label{eq-6}
\end{equation}
and
\begin{align}
&-c_{22}'\bigl(c'_{22}(c'_{12}-1)+(9-c'_{12})b'_{22}+4zc'_{22}\bigr) \nonumber\\
&\qquad\qquad=\bigl(-b'_{21}c'_{12}(c'_{12}-1)-b'_{22}c'_{22}(c'_{12}-1)+b'_{21}b'_{22} c'_{11}+[b'_{22}]^2c'_{12}-4zc'_{12}\bigr)(3-c'_{12}).
\label{eq-7}
\end{align}
The last set of relations comes from (\ref{A-A}). Here, we use that the matrix elements of $\tilde{A}$ and $\tilde{A}'$ are not independent. We obtain four equations:
\begin{align}
& (3-2c'_{12})a_{12}+c'_{11}a_{22}+c'_{22}a_{11}=0, \label{eq-8}\\
& (-c'_{12}+5)a_{22}+c'_{11}a'_{22}-b'_{21}a_{11}-b'_{22}a_{12}=0, \label{eq-9}\\
& (c'_{12}+2)a'_{22}-c'_{22}a_{22}-b'_{21}a_{12}-b'_{22}a_{22}=0, \label{eq-10}\\
& z=(2c'_{12}-3)a_{22}+c'_{11}a'_{22}-c'_{12}a_{12}. \label{eq-11}
\end{align}
We now have eleven (inhomogeneous and, in principle, nonlinear) equations on the ten variables $a_{11}$, $a_{12}$, $a_{22}$, $a'_{22}$, $b'_{21}$, $b'_{22}$, $z$, $c'_{11}$, $c'_{22}$, and $c'_{12}$. Equations (\ref{eq-3})--(\ref{eq-5}) and (\ref{eq-11}) are substitutions for $a_{11}$, $a_{12}$, $a_{22}$, and $z$. Equations (\ref{eq-2}) and (\ref{eq-8}) constitute a linear system determining $b'_{21}$ and $b'_{22}$. We then determine $a'_{22}$ from (\ref{eq-9}) and $c'_{22}$ from the determinant equation (\ref{eq-1}) and substitute all the obtained expressions into the remaining three nonlinear equations (\ref{eq-10}), (\ref{eq-6}) and (\ref{eq-7}) on two variables $c'_{11}$ and $c'_{12}$ using \textsc{Maple}\footnote{We acknowledge a valuable help of Misha Shapiro at this part of the work.}. It turns out that \eqref{eq-10} then factorizes into two factors, one rather large factor and the another is $9-(c'_{12})^2-(c'_{11})^2$. For the huge factor, its resultant with equations (\ref{eq-6}) and (\ref{eq-7}) is empty, whereas for the second factor we obtain a nonempty set of solutions. We now describe these solutions.
Let $9-(c'_{12})^2-(c'_{11})^2=0$. Then, using (\ref{eq-1}), either $c'_{11}=0$, which results in inconsistencies in all the above calculations, or
$$
c'_{11}+c'_{22}=0,
$$
which we assume in what follows. The substitutions then give
\begin{equation}
a_{11}=-\tfrac{1}{9} (5c'_{12}-9),\quad a_{12}=\tfrac{2}{9} c'_{11}, \quad a_{22}=-\tfrac{1}{9} (c'_{12}-3)
\end{equation}
and
\begin{equation}
b'_{21}=1-\tfrac{c'_{12}}{3},\quad b'_{22}=\tfrac{c'_{11}}{3},\quad a'_{22}=0.
\end{equation}
Equation (\ref{eq-10}) is satisfied identically, expressing $z$ out of (\ref{eq-11}), we obtain
$$
z=-\tfrac{1}{9} (2c'_{12}-3)(c'_{12}-3)-\tfrac{2}{9} c'_{12}c'_{11},
$$
and substituting all the above quantities into the matrix $\tilde{B}'S\tilde{B}'-4zS^{\text{T}}$, we obtain that all matrix elements are multiplied by the same factor:
$$
\tilde{B}'S\tilde{B}'-4zS^{\text{T}}=\beta \left(\begin{array}{rr} -c'_{12} & c'_{11}\\ c'_{11} & c'_{12}\end{array}\right),\qquad
\beta = 2(c'_{12}-3)^2+2c'_{12}c'_{11}.
$$
So, the only possibility for this matrix to be degenerate is to set $\beta=0$. Then this matrix vanishes and all vectors are its null vectors. We therefore have that
\begin{equation}
c'_{11}=-\frac{(c'_{12}-1)^2}{c'_{12}},
\label{c11}
\end{equation}
and since we require $c'_{12}-3\ne 0$, using (\ref{eq-1}) we obtain a cubic equation determining $c'_{12}$
\begin{equation}
(3-c'_{12})^3=(c'_{12}+3)[c'_{12}]^2\ \ \hbox{or}\ \ 2[c'_{12}]^3-6[c'_{12}]^2+27 c'_{12}-27=0.
\label{c12}
\end{equation}
This equation admits one real root $c'_{12}=1+\Bigl[1+\frac34 \sqrt{78} \Bigr]^{1/3}+\Bigl[1-\frac34 \sqrt{78}\Bigr]^{1/3}\approx 1.1898$ and two complex conjugate roots $c'_{12}\approx 0.9051 \pm 3.2445{\rm i}$. If we set $c'_{12} = 3\zeta$, we find that $\zeta$ are the roots of
\begin{equation}
\label{Proot} P(\zeta) = 2\zeta^3 - 2\zeta^2 + 3\zeta - 1
\end{equation}
Collecting all previous results and calling $\alpha = a_{33}$, we find that all coefficients of the matrices $(A,B,C)$ as explicit elements in $\mathbb{Q}(\zeta) + \alpha.\mathbb{Q}(\zeta)$. We can use \eqref{Proot} to express elements of $\mathbb{Q}(\zeta)$ as degree two polynomials in $\zeta$. The result is the one announced in Lemma~\ref{Ldim3}. Note that since the matrix $\tilde{B}'S\tilde{B}'-4zS^{\text{T}}$ is null, from (\ref{QQQ}) we have another nice relation
$$
\tilde{A}'=\tfrac{1}{4} \tilde{B}'(\tilde{C}')^{-1}(\tilde{B}')^{\text{T}}.
$$
We have thus found a classical Airy structure -- as can be directly checked. Lemma~\ref{basrem} tells us the quantum Airy structures which projects to this classical Airy structure are
$$
D^i = \tfrac{1}{2}{\rm Tr}\,B^i + \Delta^i,
$$
where $\Delta$ is an arbitrary elements of $V^*$ such that $\Delta$ is zero on the space of commutators in the Lie algebra. As the latter is spanned by $L_0'$ and $L_1$, we deduce that $\Delta'^0 = \Delta^1 = 0$ while $\Delta^0$ remains arbitrary.
These quantum Airy structures determine an (adjoint) module structure of $V$ on $\mathcal{D}_{V}$, and the cohomology spaces $H^p(V,\mathcal{D}_{V})$ for $p \in \{0,1,2\}$ can be computed with Lemma~\ref{cod3}. We are indeed in a three-dimensional situation with a codimension $1$ ideal $\mathfrak{i}$ spanned by $(x_1,x_2) = (L_1,L_0')$, the extra generator being $y = L_0 \in \mathfrak{q}$, and commutation relations \eqref{comut} determined by
$$
a = 0,\qquad \left(\begin{array}{cc} a_{11} & a_{12} \\ a_{21} & a_{22} \end{array}\right) = {\rm diag}(-1,2)
$$
The result is that $\dim H^0_{L}(V,\mathcal{D}_{V}) = 1$ corresponding to the constant in $\mathcal{D}_{V}$, $\dim H^1 = 2$ with one generator corresponding to deformation by the constant $\Delta^0$, and $\dim H^2 = 1$. \hfill $\Box$
\section{Four classes of examples from geometry}
In the remaining of the paper, we consider examples of quantum Airy structures which have a stronger geometric flavor -- in relation with topological quantum field theories and enumerative geometry.
\label{S4}
\subsection{From Frobenius algebras (2d TQFTs)}
\label{S4Frob}
\subsubsection{Quantum Airy structures}
Let $\mathbb{A}$ be a Frobenius algebra (not necessarily unital), \textit{i.e.} a finite-dimensional vector space together with a commutative, associative product $\mathbb{A} \otimes \mathbb{A} \rightarrow \mathbb{A}$ and a linear form $\varphi\,:\,\mathbb{A} \rightarrow \mathbb{C}$ such that the pairing $\mathbb{A} \otimes \mathbb{A} \rightarrow \mathbb{C}$ defined by $\langle v_1,v_2 \rangle = \varphi(v_1v_2)$ is non-degenerate. We recall that
\begin{lemma} \cite{Abrams}
\label{defFrob} If the product is given, all the other Frobenius algebra structures on $\mathbb{A}$ are of the form $\tilde{\varphi}(v) := \varphi(vu)$ for some invertible element $u \in \mathbb{A}$.
\end{lemma}
Let $(e_{i})_{i \in I}$ be a basis, and $(e_{i}^*)_{i \in I}$ the basis such that
$$
\langle e_{i},e_{j}^* \rangle = \delta_{i,j}.
$$
\begin{proposition}
\label{FrobABCD} For any $\theta_{A},\theta_{B},\theta_{C} \in \mathbb{A}$
\begin{eqnarray}
A_{j,k}^{i} & = & \varphi(e^*_{i}e^*_{j}e^*_{k}\theta_{A}). \nonumber \\
B_{j,k}^{i} & = & \varphi(e^*_{i}e_{j}^*e_{k}\theta_{B}). \nonumber \\
C_{j,k}^{i} & = & \varphi(e^*_{i}e_{j}e_{k}\theta_{C}). \nonumber
\end{eqnarray}
and arbitrary $D^{i}$ defines a quantum Airy structure on $V = \mathbb{A}$. It has $f^{i}_{j,k} = 0$, \textit{i.e.} the underlying Lie algebra is abelian.
\end{proposition}
\noindent \textbf{Proof.} Commutativity (resp. invariance) of the product shows that $B^{i}_{j,k} = B^{j}_{i,k}$ (resp. $A$ is fully symmetric). Therefore, we just need to check the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations of Section~\ref{S22}. By definition of the dual basis and the pairing we remark that
\begin{equation}
\label{propair} \forall v_1,v_2 \in \mathbb{A},\qquad v_1v_2 = \varphi(v_1v_2e_{a})\,e_{a}^*.
\end{equation}
We recall the matrix notations $X^{i} = (X^{i}_{j,k})_{j,k}$. Using invariance, associativity and commutativity of the product, we find
\begin{equation}
(B^{i}B^{j})_{k,\ell} = \varphi(e_{i}^*e_{k}^*e_{a}\theta_{B})\varphi(e_{j}^*e_{a}^*e_{\ell}\theta_{B}) = \varphi(e_{i}^*e_{j}^*e_{k}^*e_{\ell}\theta^2_{B}),
\end{equation}
and likewise
\begin{equation}
(B^{i}A^{j})_{k,\ell} = \varphi(e_{i}^*e_{k}^*e_{j}^*e_{\ell}\theta_{B}\theta_{A}),\qquad
(C^{i}B^{j})_{k,\ell} = \varphi(e_{i}^*e_{k}e_{j}^*e_{\ell}\theta_{C}\theta_{B}). \nonumber
\end{equation}
These expressions are completely symmetric under permutation of $i$ and $j$. So, pairs of terms in the \textbf{BB-AC}, \textbf{BC} and \textbf{BA} relations cancel each other. \hfill $\Box$
As we see from the proof, this example provides a rather trivial solution of the three relations in the sense that the three terms in each of them are already symmetric in $i$ and $j$.
Up to a change of basis in $\mathbb{A}$, we can and will assume in the remaining of this section that $(e_i)_{i \in I}$ is an orthonormal basis. We will sometimes need the element
$$
H := e_{a}^2.
$$
Let us make a last observation, in case $\theta_{A}$ is invertible. Using the symmetry \eqref{SymU} with $u_{i,j} = \varphi(e_{i}e_{j}\nu)$ and choosing $\nu = -\theta_B/\theta_{A}$, we find an equivalent quantum Airy structure with $\tilde{B} = 0$ and
\begin{eqnarray}
\tilde{A}^i_{j,k} & = & \varphi(e_ie_je_k\theta_{A}). \nonumber \\
\tilde{C}^i_{j,k} & = & \varphi\big(e_ie_je_k(\theta_{C} - \theta_{A}^{-1}\theta_{B}^2)\big). \nonumber \\
\tilde{D}^i & = & D^i - \tfrac{1}{2}\varphi(e_iH\theta_{B}). \nonumber
\end{eqnarray}
In the special case $\theta_{B}^2 = \theta_{A}\theta_{C}$, we have $\tilde{C} = 0$. On the contrary, if $\theta_{B}^2 \neq \theta_{A}\theta_{C}$, we cannot \textit{a priori} get rid of $\tilde{C}$ with the group action while keeping $\tilde{B} = 0$.
We can illustrate this phenomenon in a simple way when $\mathbb{A}$ has dimension $1$. In this case, we can perform a rescaling of $x$ (this transformation is in $\mathcal{G}_{V}$) to achieve $\tilde{\theta}_{A} = 1$, and conjugation by an exponential of the Laplacian as above to achieve $\tilde{\theta}_{B} = 0$. However, when $\theta_{B}^2 - \theta_{A}\theta_{C} \neq 0$, we do not have further freedom to modify $\tilde{\theta}_{C}$. Thus, $\theta_{B}^2 - \theta_{A}\theta_{C}$ is a continuous parameter of deformation for $\mathcal{G}_{V}$-orbits. If we allowed $\hbar$-rescalings -- this is not in the group $\mathcal{G}_{V}$ -- we could in fact achieve $\tilde{\theta}_{C} = 0$ or $1$, so this deformation parameter coincides with $\hbar$-rescalings.
The action by the group of invertible elements of $\mathbb{A}$ deforming the Frobenius structure according to Lemma~\ref{defFrob} also provides a family of commuting flows on the moduli space of quantum Airy structures on the abelian Lie algebra $\mathbb{A}$, although we do not make statements about the independence of these flows.
The classical Airy structure is in this case simply given by the infinitesimal symplectomorphisms $({\cal L}_i)_{i \in I}$, where
\begin{eqnarray}
{\cal L}_i e_j^* & = & - \langle \theta_{B}\cdot e_i^* \cdot e_j,e_a^* \rangle e_a^* - \langle \theta_{C}\cdot e_i^* \cdot e_j,e_a \rangle e_a \nonumber \\
\mathcal{L}_{i} e_j & = & -\langle \theta_{A}\cdot e_i^* \cdot e_j^*,e_a^* \rangle e_a^* + \langle \theta_{B}\cdot e_i^* \cdot e_j^*,e_a \rangle e_a \nonumber
\end{eqnarray}
The commutativity of the product directly implies that the commutation of the ${\cal L}_i'$s together with the full symmetry of $A$ and symmetry of $B$ with respect to the two first indices.
\subsubsection{Partition function for $\dim\,\mathbb{A} = 1$}
\label{1dZ}
Quantum Airy structures on a vector space of dimension $1$ are just differential operators of the form
$$
L = \hbar \partial_{x} - \tfrac{\theta_{A} x^2}{2} - \hbar \theta_{B} x\partial_{x} - \tfrac{\theta_C}{2}\partial_{x}^2 - \hbar D,
$$
where $\theta_{A},\theta_{B},\theta_{C},D$ are scalars. The differential equation $L\cdot Z = 0$ can be solved by elementary means to obtain the partition function. We have to distinguish several cases, all related by limiting procedures, which we will not discuss.
\vspace{0.2cm}
\noindent\boxed{\textit{$\theta_{C} = 0$}}
We have $Z(x) = \exp\big(\tfrac{S_0(x)}{\hbar} + S_1(x)\big)$ with
$$
S_0(x) = \tfrac{\theta_{A}}{2\theta_{B}^3}\Big(-\operatorname{ln}(1 - \theta_{B}x) - \theta_{B}x - \tfrac{\theta_{B}^2x^2}{2}\big),\qquad S_1(x) = -\tfrac{D}{\theta_{B}}\operatorname{ln}(1 - \theta_{B}x).
$$
The answer can be obtained combinatorially, as a corollary to Proposition~\ref{Cequal0}. This formula still makes sense when $\theta_{B} = 0$, it becomes $Z(x) = \exp\big(\tfrac{\theta_{A}x^3}{6\hbar} + Dx\big)$.
\vspace{0.2cm}
\noindent \boxed{\textit{$\theta_{C} \neq 0$ \textit{and} $\theta_{B}^2 = \theta_{A}\theta_{C}$}}
We find $Z(x) = \mathfrak{Z}(x)/\mathfrak{Z}(0)$ with
\begin{eqnarray}
\mathfrak{Z}(x) & = & \exp\Big\{\tfrac{1}{\hbar}\big(\tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}}\big)\Big\}\,{\rm Bi}\Big(\tfrac{1 - 2\theta_{B}x + \hbar\theta_{C}(\theta_{B} - 2D)}{(2\theta_{B}\theta_{C}\hbar)^{2/3}}\Big) \nonumber \\
& = & c_{\hbar}\,\exp\Big\{\tfrac{1}{\hbar}\big(\tfrac{x}{\theta_{C}}- \tfrac{\theta_{B}x^2}{2\theta_{C}}\big)\Big\} \int \mathrm{d} t\,\exp\Big(\tfrac{1}{\hbar}\big(\tfrac{1 - 2\theta_{B}x + \hbar\theta_{C}(\theta_{B} - 2D)}{(2\theta_{B}\theta_{C})^{2/3}}\,t - \tfrac{t^3}{3}\big)\Big). \nonumber
\end{eqnarray}
where $c_{\hbar}$ only depends on $\hbar$. Here ${\rm Bi}(z)$ is the Bairy function, which is the solution of the differential equation ${\rm Bi}''(z) = z{\rm Bi}(z)$, whose asymptotics at $z \rightarrow \infty$ is
$$
{\rm Bi}(z) = \frac{e^{\frac{2}{3}z^{\frac{3}{2}}}}{\pi^{\frac{1}{2}}z^{\frac{1}{4}}}\big(1 + o(1)\big).
$$
In particular, $Z(x) = \exp\big(\tfrac{S_0(x)}{\hbar} + S_1(x) + O(\hbar)\big)$ with
\begin{eqnarray}
S_0(x) & = & \frac{(1 - 2\theta_{B}x)^{\frac{3}{2}} - 1 + \tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}}}{3\theta_{B}\theta_{C}} \nonumber \\
S_1(x) & = & \tfrac{\theta_{B} - 2D}{2\theta_{B}}\big((1 - 2\theta_{B}x)^{\frac{1}{2}}- 1\big) - \tfrac{1}{4}\operatorname{ln}(1 - 2\theta_{B}x).
\end{eqnarray}
The integral above is a formal integral, \textit{i.e.} it is to be evaluated by expansion around the (unique) saddle point which realizes $F_{0,3} = \theta_{A}$. However, choosing the corresponding steepest descent contour offers the possibility to define $Z(x)$ here as an entire function of $x$. This is also what the Bairy function does. We retrieve the Taylor coefficients $F_{g,n}$ by expanding ${\rm Bi}(z)$ at $z \rightarrow \infty$. Elementary properties of the full asymptotic expansion of the Bairy function are collected in Appendix~\ref{Bairyapp}.
\vspace{0.2cm}
\noindent \boxed{\textit{$\theta_{C} \neq 0$ and $\sigma^2 := \theta_B^2 - \theta_{A}\theta_{C} \neq 0$}}
We find that $Z(x) = \mathfrak{Z}(x)/\mathfrak{Z}(0)$ with
\begin{eqnarray}
\label{Whity} \mathfrak{Z}(x) & = & \exp\Big\{\tfrac{1}{\hbar}\big(\tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}}\big) - \tfrac{1}{2}\operatorname{ln}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)\Big\}\,\mathcal{W}_{M}\Big(\tfrac{1}{4}\big(\tfrac{\theta_{A}}{\hbar\sigma^3} + \tfrac{2D - \theta_{B}}{\sigma}\big)\,;\,\tfrac{1}{4}\,;\,\tfrac{\theta_{B}^2}{\hbar\sigma^3\theta_{C}}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)^2\Big) \\
& = & c_{\hbar}\,\exp\Big\{\tfrac{1}{\hbar}\big(\tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}}\big) + \operatorname{ln}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)\Big\} \int_{-\frac{1}{2}}^{\frac{1}{2}} \mathrm{d} t\,\exp\Big\{\tfrac{\theta_{B}^2}{\hbar\sigma^3\theta_{C}}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)^2\,t\Big\}(\tfrac{1}{4} - t\big)^{-\frac{1}{4}}\bigg(\frac{\frac{1}{2} - t}{\frac{1}{2} + t}\bigg)^{\frac{1}{4}\big(\frac{\theta_{A}}{\hbar \sigma^3} + \frac{2D - \theta_{B}}{\sigma}\big)}. \nonumber
\end{eqnarray}
Here $\mathcal{W}_{M}(\mu;\nu;z)$ is the Whittaker-M function, solving the differential equation
$$
\Big\{\partial_{z}^2 + \big(-\tfrac{1}{4} + \tfrac{\mu}{z} + \tfrac{\frac{1}{4} - \nu^2}{z^2}\big)\Big\}\mathcal{W}_{M}(\mu;\nu;z) = 0.
$$
It is an entire function of $z$, but only its asymptotics near $z \rightarrow \infty$ matter to obtain the $F_{g,n}$s. These asymptotics can be found by saddle point analysis in the integral formula, which leads to $Z(x) = \exp\big(\tfrac{S_0(x)}{\hbar} + S_1(x) + O(\hbar)\big)$ and
\begin{eqnarray}
S_0(x) & = & \tfrac{\theta_{B}}{2\sigma^2\theta_{C}}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)(1 - 2\theta_{B}x + \sigma^2x^2)^{\frac{1}{2}} + \tfrac{\theta_{A}}{2\sigma^3}\,{\rm argth}\bigg(\tfrac{\sigma}{\theta_{B}}\,\tfrac{(1 - 2\theta_{B}x + \sigma^2x^2)^{\frac{1}{2}}}{1 - \tfrac{\sigma^2 x}{\theta_{B}}}\bigg) + \tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}} + c_0. \nonumber \\
S_1(x) & = & -\tfrac{1}{4}\operatorname{ln}(1 - 2\theta_{B}x + \sigma^2x^2) + \tfrac{2D - \theta_{B}}{2\sigma}\,{\rm argth}\bigg(\tfrac{\sigma}{\theta_{B}}\,\tfrac{(1 - 2\theta_{B}x + \sigma^2x^2)^{\frac{1}{2}}}{1 - \tfrac{\sigma^2x}{\theta_{B}}}\bigg) + c_1. \nonumber
\end{eqnarray}
This formula still makes sense for $\theta_{B} = 0$, the main simplification being that the variable of the Whittaker function becomes $z \leftarrow \frac{\sigma x^2}{\hbar\theta_{C}}$. In fact, as a result of Proposition~\ref{Propmain}, $F_{g,n}$ is a polynomial in $\theta_{A},\theta_{B},\theta_{C}$ and $D$. The first values are given in Appendix~\ref{Whitapp}.
\subsubsection{Partition function in general}
In this paragraph, $\mathbb{A}$ is assumed unital. When $\dim \mathbb{A} > 1$, we are going to show that the partition function for the quantum Airy structure of Proposition~\ref{FrobABCD} is computed by a higher-dimensional version of the integral formulas of the previous paragraph. We heavily rely on the commutativity of the product in $\mathbb{A}$. We denote
$$
D := D^ae_{a} \in \mathbb{A}.
$$
An easy rewriting of Proposition~\ref{Cequal0} in terms of the linear form $\varphi$ leads to
\begin{lemma}
\label{Cequal0Frob} If $\theta_{C} = 0$, then we have that $Z = \exp\big(\tfrac{S_0(x)}{\hbar} + S_1(x)\big)$ with
\begin{eqnarray}
S_0(x) & = & \varphi\Big(\tfrac{\theta_A}{2\theta_{B}^3}\big(-\operatorname{ln}(1 - \theta_{B}x) - \theta_Bx - \tfrac{\theta_B^2x^2}{2}\big)\Big). \nonumber \\
S_1(x) & = & \varphi\Big(-\tfrac{D}{\theta_{B}}\operatorname{ln}(1 - \theta_Bx)\Big). \nonumber
\end{eqnarray}
This formula also makes sense when $\theta_{B}$ is not invertible.\hfill $\Box$
\end{lemma}
When $\theta_{C} \neq 0$, we will check the formulas below by direct computation with help of Dyson-Schwinger equations. Other proofs could be obtained exploiting the action of $\mathcal{G}_{V}$ on the partition function of Lemma~\ref{Cequal0Frob}.
\begin{proposition}
\label{AiryZFrob} If $\theta_{C}$ is invertible, and $\theta_{B}^2 = \theta_{A}\theta_{C}$, we have $Z(x) = \mathfrak{Z}(x)/\mathfrak{Z}(0)$ with
\begin{equation}
\mathfrak{Z}(x) = {\rm Cste}_{\hbar}\,\exp\Big\{\tfrac{1}{\hbar}\varphi\big(\tfrac{x}{\theta_{C}}- \tfrac{\theta_{B}x^2}{2\theta_{C}}\big)\Big\} \int \,\exp\Big\{\tfrac{1}{\hbar}\varphi\big(\tfrac{1 - 2\theta_{B}x + \hbar\theta_{C}(\theta_{B} - 2D)}{(2\theta_{B}\theta_{C})^{2/3}}\,t - \tfrac{t^3}{3}\big)\Big\}\mathrm{d} t,
\end{equation}
where $\mathrm{d} t$ is the Lebesgue measure on the linear coordinates on $\mathbb{A}$ with respect to an orthonormal basis.
\end{proposition}
\begin{proposition}
\label{WhitZFrob} If $\theta_{C} \neq 0$ and $\sigma = \theta_B^2 - \theta_{A}\theta_{C}$ is invertible, we have $Z(x) = \mathfrak{Z}(x)/\mathfrak{Z}(0)$ with
\begin{eqnarray}
\mathfrak{Z}(x) & = & \exp\Big\{\tfrac{1}{\hbar}\varphi\big(\tfrac{x}{\theta_{C}} - \tfrac{\theta_{B}x^2}{2\theta_{C}}\big)t + H\,\operatorname{ln}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big)\Big\} \int \exp\Big\{\varphi\Big(\tfrac{\theta_{B}^2}{\hbar\sigma^3\theta_{C}}\big(1 - \tfrac{\sigma^2x}{\theta_{B}}\big) \nonumber \\
&& + \tfrac{1}{4}\big(\tfrac{\theta_{A}}{\hbar \sigma^3} + \tfrac{2D - H\theta_{B}}{\sigma} - H\big)\operatorname{ln}\big(\tfrac{1}{2} - t\big) + \tfrac{1}{4}\big(-\tfrac{\theta_{A}}{\hbar\sigma^3} - \tfrac{2D - H\theta_{B}}{\sigma} - H\big)\operatorname{ln}\big(\tfrac{1}{2} + t\big)\Big)\Big\}\mathrm{d} t, \nonumber
\end{eqnarray}
and we recall that $H = e_{a}^2$. This formula also makes sense when $\theta_{B}$ is not invertible.
\end{proposition}
The intermediate cases in Proposition~\ref{AiryZFrob}, when $\theta_{C} \neq 0$, but not invertible, or in Proposition~\ref{WhitZFrob} when $\sigma \neq 0$, but not invertible, will not be discussed, since they can be obtained by limiting procedures. For the same reason, we will only write a proof of Proposition~\ref{WhitZFrob}. It is in fact a good exercise for the reader to repeat the scheme of this proof in the specific (and simpler) case of Proposition~\ref{AiryZFrob}.
\vspace{0.2cm}
\noindent \textbf{Proof.} We consider a formal integral of the form
$$
\mathbb{I}(x) = \exp\Big\{\varphi\big(\alpha_1x + \tfrac{\alpha_2x^2}{2} + \nu\operatorname{ln}(1 - \beta x)\big)\Big\} \int \exp\Big\{\varphi\big(\gamma(1 - \beta x)^2t + \tau_1\operatorname{ln}(\tfrac{1}{2} - t) + \tau_2\operatorname{ln}(\tfrac{1}{2} + t)\big)\Big\}\mathrm{d} t,
$$
where $\alpha_1,\alpha_2,\nu,\gamma,\beta,\tau_1,\tau_2 \in \mathbb{A}$ are considered as parameters. If $f\,:\,\mathbb{A} \rightarrow \mathbb{K}$, we denote
$$
\langle f(t) \rangle := \frac{\int \,f(t)\,\exp\Big\{\varphi\big(\gamma(1 - \beta x)^2 t + \tau_1\operatorname{ln}(\tfrac{1}{2} - t) + \tau_2\operatorname{ln}(\tfrac{1}{2} + t)\big)\Big\}\mathrm{d} t}{\int \,\exp\Big\{\varphi\big(\gamma(1 - \beta x)^2 t + \tau_1\operatorname{ln}(\tfrac{1}{2} - t) + \tau_2\operatorname{ln}(\tfrac{1}{2} + t)\big)\Big\}\mathrm{d} t},
$$
which implicitly depends on $x$. This definition can be extended by linearity to a function $f\,:\,\mathbb{A} \rightarrow \mathbb{A}$. We also introduce $\varphi_i(v) := \varphi(e_iv)$ for $v \in \mathbb{A}$.
Let us compute the action of the differential operators of the quantum Airy structure on $\mathbb{I}(x)$.
\begin{eqnarray}
-\mathbb{I}^{-1}\,\hbar \partial_{x_i} \mathbb{I} & = & \hbar \varphi_i\big( - \alpha_1 - \alpha_2x + \tfrac{\nu\beta}{1 - \beta x} + 2\gamma\beta \langle t \rangle\big). \nonumber \\
\mathbb{I}^{-1}\tfrac{1}{2}\varphi(e_ie_ae_b\theta_{A})x_ax_b\mathbb{I} & = & \varphi_i(\tfrac{x^2}{2}\theta_{A}). \nonumber \\
\mathbb{I}^{-1} \hbar \varphi(e_ie_ae_b\theta_{B}) x_a \partial_{x_b} \mathbb{I} & = & \hbar\varphi(e_ixe_b\theta_{B})\varphi\Big(e_b\big(\alpha_1 + \alpha_2x - \frac{\nu\beta}{1 - \beta x} - 2\gamma\beta(1 - \beta x)x \langle t \rangle\big)\Big) \nonumber \\
& = & \hbar \varphi_i\Big(\theta_{B}x\big(\alpha_1 + \alpha_2x - \tfrac{\nu\beta}{1 - \beta x} - 2\gamma\beta(1 - \beta x)x \langle t \rangle\big)\Big). \nonumber
\end{eqnarray}
In the last line, we exploited the property \eqref{propair} of the pairing in an orthonormal basis, \textit{i.e.} such that $e_i^* = e_i$. In a similar fashion
\begin{eqnarray}
&& \mathbb{I}^{-1}\,\tfrac{\hbar^2}{2}\varphi(e_ie_ae_b\theta_{C})\partial_{x_a}\partial_{x_b}\mathbb{I} \nonumber \\
& = & \tfrac{\hbar^2}{2}{\varphi(e_ie_ae_b\theta_{C})\bigg\{\varphi\Big(e_a\big(\alpha_1 + \alpha_2x - \frac{\nu\beta}{1 - \beta x}\big)\Big)\varphi\Big(e_b\big(\alpha_1 + \alpha_2 - \frac{\nu\beta}{1 - \beta}x}\big)\Big) + \varphi\big(\alpha_2e_ae_b - \tfrac{\nu\beta^2}{(1 - \beta x)^2}e_ae_b\big) \nonumber \\
&& - 4\varphi\Big(e_a\big(\alpha_1 + \alpha_2x - \tfrac{\nu\beta}{1 - \beta x}\big)\Big)\varphi\big(2\gamma\beta(1 - \beta x)e_b\langle t \rangle \big) \nonumber \\
&& + \Big\langle 4\varphi\big(\gamma\beta(1 - \beta x)e_{a}t\big)\varphi\big(\gamma\beta(1 - \beta x)e_{b}t\big) \Big\rangle + 2\varphi\big(2\gamma\beta^2e_{a}e_{b}\langle t \rangle\big)\bigg\} \nonumber \\
& = & \frac{\hbar^2}{2}\varphi_i\bigg(\theta_{C}\Big\{\big(\alpha_1 + \alpha_2x - \tfrac{\nu\beta}{1 - \beta x}\big)^2 + \big(\alpha_2 - \tfrac{\nu\beta^2}{(1 - \beta x)^2}\big)H - 4\big(\alpha_1 + \alpha_2 x - \tfrac{\nu\beta}{1 - \beta x}\big)\gamma\beta(1 - \beta x)\langle t \rangle \nonumber \\
&& + 4\gamma^2\beta^2(1 - \beta x)^2\langle t^2 \rangle + 2\gamma\beta^2 H \langle t \rangle\Big\}\bigg), \nonumber
\end{eqnarray}
and we observe that $\mathbb{I}^{-1} \hbar D^i \mathbb{I} = \varphi_i(D)$. Summing all the terms yields an expression for $\mathbb{I}^{-1} L_i \mathbb{I}$, which involves in particular $\langle t^2 \rangle$.
We now write the Dyson-Schwinger relations satisfied by the averages $\langle \cdot \rangle$. They express the fact that (formal) integrals are invariant under change of variable. Here we use the infinitesimal change of variable $t \rightarrow t + \epsilon \big(\tfrac{1}{4} - t^2)e_j$ for a given $j$. We remark that $\mathrm{d} t \rightarrow \mathrm{d} t\big(1 - \epsilon \varphi(2tHe_j) + O(\epsilon^2)\big)$, and invariance of the integral at order $\epsilon$ gives the exact formula
$$
\Big\langle \varphi\Big(e_j\Big\{-2tH + \gamma(1 - \beta x)^2\big(\tfrac{1}{4} -t^2\big) - \tau_1\big(\tfrac{1}{2} + t) + \tau_2\big(\tfrac{1}{2} - t\big)\Big\}\Big)\Big\rangle = 0.
$$
This formula can equivalently be derived by integration by parts. As it holds for any $j$, we also have the identity in $\mathbb{A}$
$$
-2H\langle t \rangle + \gamma(1 - \beta x)^2\big(\tfrac{1}{4} -\langle t^2 \rangle\big) - \tau_1\big(\tfrac{1}{2} + \langle t \rangle) + \tau_2\big(\tfrac{1}{2} - \langle t \rangle \big) = 0.
$$
We use this identity to eliminate $\langle t^2 \rangle$ from $-\mathbb{I}^{-1} L_i \mathbb{I}$. After tedious but straightforward algebra, the terms can be collected as follows
\begin{eqnarray}
-\mathbb{I}^{-1}L_i\mathbb{I} & = & \varphi_i\bigg( \tfrac{\hbar^2\theta_{C}}{2}(\alpha_1^2 + \alpha_2H + 2\nu\alpha_2) - \hbar\alpha_1 + \hbar\nu\theta_{B} + D + \tfrac{\hbar^2\theta_{C}}{2}\big(\tfrac{\gamma^2\beta^2}{2} + \gamma\beta^2(\tau_2 - \tau_1)\big) \nonumber \\
&& + x\big\{- \hbar\alpha_2 + \hbar \alpha_1\theta_{B} + \hbar^2\theta_{C}\alpha_1\alpha_2 - \hbar^2\theta_{C}\gamma^2\beta^3\big\} + x^2\big\{\tfrac{\theta_{A}}{2} + \hbar\theta_{B}\alpha_2 + \tfrac{\hbar^2\theta_{C}}{2}(\alpha_2^2 + \gamma^2\beta^4)\big\} \nonumber \\
&& + \tfrac{\nu\beta}{1 - \beta x}\big\{\hbar\beta - \hbar\theta_{B} - \hbar^2\theta_{C}(\alpha_1\beta + \alpha_2)\big\} + \tfrac{\hbar^2\theta_{C}\nu\beta^2}{2(1 - \beta x)^2}(\nu - H) \nonumber \\
&& + 2\langle t \rangle \Big\{ \hbar\gamma\beta + \hbar^2\theta_{C}\gamma\beta\big(-\alpha_1 + \nu\beta - \beta(\tau_1 + \tau_2 + 2H)\big)\Big\} \nonumber \\
&& + x\,\gamma\beta\big\{-\beta - \hbar\theta_{B} + \hbar^2\theta_{C}(\beta\alpha_1 - \alpha_2)\big\} + x^2\,\hbar\gamma\beta^2\big\{\hbar\theta_{C}\alpha_2 + \theta_{B}\big\}\Big)\bigg). \nonumber
\end{eqnarray}
Imposing that the coefficients -- inside $\varphi_i$ -- of $1$, $x$, $x^2$, $(1 - \beta x)^{-1}$, $(1 - \beta x)^{-2}$, $\langle t \rangle$, $\langle t \rangle x$ and $\langle t \rangle x^2$ separately vanish uniquely fixes
$$
\alpha_1 = \frac{1}{\hbar\theta_{C}},\qquad \alpha_2 = -\frac{\theta_{B}}{\hbar\theta_{C}},\qquad \beta = \frac{\sigma^2}{\theta_{B}},\qquad \gamma = \frac{\theta_{B}^2}{\hbar \sigma^3 \theta_{C}},
$$
and
$$
\nu = H,\qquad \tau_1 = \tfrac{1}{4}\big(\tfrac{\theta_{A}}{\hbar\sigma^3} + \tfrac{2D - H\theta_{B}}{\sigma} - H\big),\qquad \tau_2 = -\tfrac{1}{4}\big(\tfrac{\theta_{A}}{\hbar\sigma^3} + \tfrac{2D - H\theta_{B}}{\sigma} + H\big).
$$
We then have found a common solution of $L_i Z = 0$ for all $i \in I$, and one can check by saddle point analysis that, for the above choice of parameters, it is indeed of the form
$$
\mathbb{I}(x) = \exp\bigg(\sum_{\substack{g \geq 0 \\ n \geq 1}} \sum_{i_1,\ldots,i_n} \tfrac{\hbar^{g - 1}}{n!}\,F_{g,n}(i_1,\ldots,i_n)\,x_{i_1}\cdots x_{i_n}\bigg),\qquad F_{0,1}(i) = F_{0,2}(i,j) = 0.
$$
Invoking the uniqueness of such a solution (Proposition~\ref{Propmain}) concludes the proof.
\hfill $\Box$
\subsubsection{Example: TQFT partition function}
\label{TQFTS4}
It is well known that unital Frobenius algebras are in one-to-one correspondence with two-dimensional topological quantum field theories \cite{Abrams}. In this context, the product on $\mathbb{A}$ comes from the map $\mathbb{A}^{\otimes 2} \rightarrow \mathbb{A}$ that the TQFT attaches to a pair of pants. Let us choose $\theta_{A} = \theta_{B} = \theta_{C} = 1$ in Proposition~\ref{FrobABCD}. We notice that, up to dualization $F_{0,3}(i,j,k) = A^{i}_{j,k}$ represents the product. This is the manifestation of a more general, easy fact. Let $\Sigma_{g,n}$ is a topological surface with $n$ boundaries oriented inward, and denote $\mathcal{F}(\Sigma_{g,n}) \in {\rm Hom}(\mathbb{A}^{\otimes},\mathbb{C})$ the amplitude assigned to $\Sigma_{g,n}$ by the TQFT $\mathcal{F}$. Using the TQFT glueing rules, it can be computed (or defined) as follows. Take a pair of pants decomposition of $\Sigma_{g,n}$. Take the tensor product over all pairs of pants of the maps $\mathcal{F}(\Sigma_{0,3}): \mathbb{A}^{\otimes 3} \rightarrow \mathbb{C}$, and apply the pairing $\mathbb{A} \otimes \mathbb{A} \rightarrow \mathbb{C}$ on the factors of $\mathbb{A}$ corresponding to coinciding boundary components of the pair of pants. The result is a multilinear map $\mathbb{A}^{\otimes n} \rightarrow \mathbb{C}$. Owing to the associativity, commutativity and invariance of the product, the result does not depend on the choice of pair of pants decomposition.
\begin{lemma}
\label{TQFTpart} Assume $\mathbb{A}$ unital. Then the partition function corresponding to the quantum Airy structure of Proposition~\ref{FrobABCD} with the choice of $D = \tfrac{H}{2}$ and $\theta_{A} = \theta_{B} = \theta_{C} = 1$, gives the amplitudes of the two-dimensional TQFT corresponding to $\mathbb{A}$, e.g. for $2g - 2 + n > 0$
\begin{equation}
\label{TQFTAiry} F_{g,n} = |\mathfrak{G}_{g,n}|\,\mathcal{F}(\Sigma_{g,n}).
\end{equation}
Here $\mathfrak{G}_{g,n}$ is the set of graphs described in the legend of Figure~\ref{TRgraph}, and its weighted cardinality
$$
|\mathfrak{G}_{g,n}(1)| := \sum_{G \in \mathfrak{G}_{g,n}(1)} \frac{1}{|{\rm Aut}\,G|}
$$
is computed by the generating series
$$
\exp\bigg(\sum_{g \geq 0} \sum_{n \geq 1} \tfrac{\hbar^{g - 1}}{n!}\,|\mathfrak{G}_{g,n}|\bigg) = \exp\Big\{\tfrac{1}{\hbar}\big(x - \tfrac{x^2}{2}\big)\Big\}\,{\rm Bi}\big((2\hbar)^{-\frac{2}{3}}(1 - x)\big).
$$
\end{lemma}
\textbf{Proof.} We obtain $\Sigma_{1,1}$ by glueing two boundaries of the same pair of pants. The TQFT assigns to this $\mathcal{F}(\Sigma_{1,1})(e_{i}) = \varphi(e_{i}e_{a}e_{a}) = \varphi(e_iH) = 2D^i = 2F_{1,1}(i)$. This coincides with \eqref{TQFTAiry} as the unique graph in $\mathfrak{G}_{g,n}(1)$ has a symmetry factor $\tfrac{1}{2}$. Unfolding \eqref{TRForm} using the fact that $B$ and $C$ here represent the product (maybe up to identification $\mathbb{A} \simeq \mathbb{A}^*$ with the pairing), we deduce that $F_{g,n}(i_1,\ldots,i_n)$ is the sum over $\mathfrak{G}_{g,n}(1)$ of the TQFT amplitude computed with a pair of pants decomposition canonically defined by the graph, and evaluated on $e_{i_1} \otimes \cdots \otimes e_{i_n}$. The fact that this amplitude is independent of the pair of pants decomposition leads to the factorization of the (automorphism-weighted) number of graphs in $\mathfrak{G}_{g,n}(1)$. \hfill $\Box$
It is clear from the proof, that if one chooses for $\theta_{A},\theta_{B},\theta_{C}$ arbitrary scalars, one will rather obtain $F_{g,n} = G_{g,n}(\theta_{A},\theta_{B},\theta_{C}) \mathcal{F}(\Sigma_{g,n})$ where $G_{g,n}(\theta_{A},\theta_{B},\theta_{C})$ are the Taylor coefficients of the partition function of the 1-dimensional quantum Airy structure $\hbar\partial_{x} - \tfrac{\theta_{A}x^2}{2} - \hbar\theta_{B}x\partial_{x} - \tfrac{\hbar^2\theta_{C}}{2}\partial_{x}^2 - \tfrac{\hbar}{2}$, see Section~\ref{1dZ}.
\subsection{From non-commutative Frobenius algebras}
\label{S4NCFrob}
Consider now a non-commutative Frobenius algebra $\mathbb{A}$. It is defined as a Frobenius algebra except that we drop the commutativity axiom, and impose that $\varphi([v_1,v_2]) = 0$ for any $v_1,v_2 \in \mathbb{A}$. Here we denote $[\cdot,\cdot]$ the commutator, and $\{\cdot,\cdot\}$ the anticommutator. We choose to work directly with an orthonormal basis $(e_i)_{i \in I}$.
\begin{proposition}
\label{NCFrob} Let $\lambda_{A},\lambda_{B},\lambda_{C}$ be central elements in $\mathbb{A}$, such that $\lambda_{B}^2 + \lambda_{A}\lambda_{C} = 0$. The assignments
\begin{eqnarray}
A^{i}_{j,k} & = & \varphi\big(\lambda_{A}\{e_j,e_k\}e_i\big), \nonumber \\
B^i_{j,k} & = & \varphi\big(\lambda_{B}[e_i,e_j]e_k\big), \nonumber \\
C^i_{j,k} & = & \varphi\big(\lambda_{C}\{e_i,e_j\}e_k\big), \nonumber
\end{eqnarray}
and $D = D^ae_{a} \in \mathbb{A}$ such that $\lambda_{B}D$ lies in the orthogonal complement of $[\mathbb{A},\mathbb{A}]$, define a quantum Airy structure.
\end{proposition}
As $\lambda_{X}$ for $X \in \{A,B,C\}$ are central, we can move them freely around inside $\varphi(\cdots)$. Then, due to invariance of the product, $A$ and $C$ are fully symmetric, and $B^i_{j,k}$ is fully antisymmetric under permutations of $(i,j,k)$ and equal to $\tfrac{1}{2}f_{i,j}^k$. We remark that, if $\lambda_{B} = \tfrac{1}{2}$, the Lie algebra structure on $V = \mathbb{A}$ determined by the quantum Airy structure coincides with the Lie algebra structure of $\mathbb{A}$ given by the commutator.
\noindent \textbf{Proof.} Using the property $\varphi(xe_a)e_a = x$ of the pairing and the orthonormality of the basis $(e_i)_{i \in I}$, we compute
$$
B^i_{j,a}B^a_{k,\ell} = \varphi\big(\lambda_{B}[e_i,e_j]e_a\big)\varphi\big(\lambda_{B}[e_a,e_k]e_{\ell}\big) = \varphi\big(\lambda_{B}^2[[e_i,e_j],e_k]e_{\ell}\big).
$$
With similar manipulations, expanding commutators/anticommutators and using cyclic invariance of $\varphi$ to move $e_{\ell}$ to the last position, we find that
\begin{eqnarray}
&& B^i_{j,a}B^a_{k,\ell} + B^i_{k,a}B^j_{a,\ell} + C^i_{\ell,a}A^j_{a,k} - (i \leftrightarrow j) \nonumber \\
& = & \varphi\big(2\lambda_{B}^2(e_ie_je_k - e_je_ie_k - e_ke_ie_j + e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{B}^2(e_je_ie_k - e_ie_ke_j - e_je_ke_i + e_ke_ie_j - e_ie_je_k + e_je_ke_i + e_ie_ke_j - e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{A}\lambda_{C}(e_je_ke_i + e_ke_je_i + e_ie_je_k + e_ie_ke_j - e_ie_ke_j - e_ke_ie_j - e_je_ie_k - e_je_ke_i)e_{\ell}\big) \nonumber \\
& = & \varphi\big( (\lambda_B^2 + \lambda_{A}\lambda_{C})(e_ie_je_k - e_je_ie_k - e_ke_ie_j + e_ke_je_i)e_{\ell}\big). \nonumber
\end{eqnarray}
Likewise, we compute that
\begin{eqnarray}
&& B^i_{j,a}C^a_{k,\ell} + C^i_{k,a}B^j_{a,\ell} + C^i_{\ell,a}B^j_{a,k} - (i \leftrightarrow j) \nonumber \\
& = & \varphi\big(2\lambda_{B}\lambda_{C}(e_ie_je_k - e_je_ie_k + e_ke_ie_j - e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{B}\lambda_{C}(e_je_ie_k - e_ie_ke_j + e_je_ke_i - e_ke_ie_j - e_ie_je_k + e_je_ke_i - e_ie_ke_j + e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{B}\lambda_{C}( e_ke_je_i - e_je_ke_i + e_ie_ke_j - e_ie_je_k - e_ke_ie_j + e_ie_ke_j - e_je_ke_i + e_je_ie_k)e_{\ell}\big) \nonumber \\
& = & 0
\end{eqnarray}
and
\begin{eqnarray}
&& B^i_{j,a}A^a_{k,\ell} + B^i_{k,a}A^{j}_{a,\ell} + B^i_{\ell,a}A^j_{a,k} - (i \leftrightarrow j) \nonumber \\
& = & \varphi\big(2\lambda_{B}\lambda_{A}(e_ie_je_k - e_je_ie_k + e_ke_ie_j - e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{B}\lambda_{C}(e_je_ie_k - e_je_ke_i + e_ie_ke_j - e_ke_ie_j - e_ie_je_k + e_ie_ke_j - e_je_ke_i + e_ke_je_i)e_{\ell}\big) \nonumber \\
&& + \varphi\big(\lambda_{B}\lambda_{C}(e_ke_je_i - e_ie_ke_j + e_je_ke_{i} - e_ie_je_k - e_ke_ie_j + e_je_ke_i - e_ie_ke_j + e_je_ie_k)e_{\ell}\big) \nonumber \\
& = & 0. \nonumber
\end{eqnarray}
Therefore, when $\lambda_{B}^2 + \lambda_{A}\lambda_{C} = 0$, we have a classical Airy structure. The last statement about $D$ is a consequence of Lemma~\ref{trL}, noticing that $D^{{\rm ref}} = \tfrac{1}{2}{\rm tr}\,B^i = 0$ since $B^i_{j,k} = -B^i_{k,j}$. \hfill $\Box$
If $\lambda_{A}$ is invertible, the quantum Airy structures of Proposition~\ref{NCFrob} are transformed by the symmetries \eqref{SymU} with $u_{a,b} = \varphi\big(\tfrac{\lambda_{B}}{2\lambda_{A}}e_{a}e_{b}\big)$ into
\begin{eqnarray}
\tilde{A}^i_{j,k} & = & \varphi\big(\mu_{A}\{e_i,e_j\}e_k\big), \nonumber \\
\tilde{B}^i_{j,k} & = & \varphi\big(\mu_{B}e_je_je_k\big),\nonumber \\
\label{CzeroAiry} \tilde{C} & = & 0,
\end{eqnarray}
where the new parameters are related to the old ones by
$$
\mu_{A} = \lambda_{A},\qquad \mu_{B} = 2\lambda_{B}.
$$
If $\mathbb{A}$ happens to be commutative, we retrieve particular cases of the quantum Airy structure of Section~\ref{S4Frob}, namely the one with $(\theta_{A},\theta_{B},\theta_{C}) = (2\lambda_{A},0,2\lambda_{C})$ for the quantum Airy structure of Proposition~\ref{NCFrob}, and the one with $(\theta_{A},\theta_{B},\theta_{C}) = (2\mu_{A},\mu_B,0)$ for \eqref{CzeroAiry}. Generically they fit in the case $\theta_B^2 - \theta_{A}\theta_{C} \neq 0$.
If we assume $-\lambda_{A} = \lambda_{B} = \lambda_{C} = 1$, the associated infinitesimal symplectomorphisms $({\cal L}_i)_{ i \in I}$ have a particularly nice form
$$
{\cal L}_i (e_j,0) = ([e_i,e_j],0) + (0,\{e_i,e_j\}),
$$
and
$$
{\cal L}_i (0,e_j) = (\{e_i,e_j\},0) - (0,[e_i,e_j]),
$$
relative to the orthonormal basis $(e_i)$. By direct computation, one can check that $[{\cal L}_i,{\cal L}_j] = 2 \phi([e_i,e_j] e_a) {\cal L}_a$.
We already computed in Proposition~\ref{Cequal0} the partition function for quantum Airy structures having $C = 0$, and we get the following expression using the pairing $\varphi$.
\begin{lemma}
The partition function $\tilde{Z}$ of the quantum Airy structure \eqref{CzeroAiry} is $\tilde{Z}(x) = \exp\big(\tfrac{\tilde{S}_0(x)}{\hbar} + \tilde{S}_1(x)\big)$ with
\begin{eqnarray}
\tilde{S}_0(x) & = & \varphi\Big(\tfrac{\mu_{A}}{\mu_{B}^3}\big(-\operatorname{ln}(1 - \mu_{B}x) - \mu_{B}x - \tfrac{\mu_B^2x^2}{2}\big)\Big), \nonumber \\
\tilde{S}_1(x) & = & \varphi\big(-\tfrac{\tilde{D}}{\mu_{B}}\operatorname{ln}(1 - \mu_{B}x)\big). \nonumber
\end{eqnarray} \hfill $\Box$
\end{lemma}
Transforming back to the initial quantum Airy structure, we deduce an integral formula for the expression for its partition function.
\begin{corollary}
The partition function of the quantum Airy structure of Proposition~\ref{NCFrob} is $Z(x) = \mathfrak{Z}(x)/\mathfrak{Z}(0)$ with
$$
\mathfrak{Z}(x) = \int \,\exp\Big\{\varphi\Big(\tfrac{\lambda_{A}}{\hbar\lambda_B}\big(-\tfrac{x^2}{2} - \tfrac{x}{2\lambda_B} + \tfrac{1}{4}\big(t - 2x - \tfrac{1}{2\lambda_{B}}\big)^2 - \tfrac{\operatorname{ln}(1 - 2\lambda_Bt)}{8\lambda_B^2}\big) - \tfrac{D}{2\lambda_B}\operatorname{ln}(1 - 2\lambda_B t)\Big)\Big\}\mathrm{d} t.
$$
\hfill $\Box$
\end{corollary}
\subsection{From loop spaces}
\label{S4Loop}
We consider $V = \mathbb{C}[[z]]$. We index a basis of $V$ by non-negative integers $k$. If $f \in \mathbb{C}[z^{-1},z]].\mathrm{d} z$ (germ of meromorphic $1$-forms), we denote
$$
\varphi(f) = \mathop{\rm Res}\limits_{z \rightarrow 0} f(z).
$$
Let $(\xi_k)_{k \geq 0}$ be a linearly independent family of germs of meromorphic $1$-forms, and $(\xi^*_{k})_{k \geq 0}$ be a linearly independent family of germs of functions (elements of $V$), such that
$$
\varphi(\xi_k \xi^*_{\ell}) = 0.
$$
Let $\theta \in \mathbb{C}[z^{-1},z]].(\mathrm{d} z)^{-1}$. Inspired by the Frobenius algebra example, we declare
\begin{eqnarray}
A^i_{j,k} & := & \varphi(\xi_i^*\,\mathrm{d}\xi_j^*\,\mathrm{d}\xi_k^*\,\theta), \nonumber \\
B^i_{j,k} & := & \varphi(\xi_i^*\,\mathrm{d}\xi_j^*\,\xi_k\,\theta), \nonumber \\
\label{Cr}C^i_{j,k} & := & \varphi(\xi_i^*\,\xi_j\,\xi_k\,\theta),
\end{eqnarray}
for some $\theta$ yet to be fixed.
\begin{proposition}
\label{Loop1} Let $u_{k,\ell} = u_{\ell,k}$ be scalars indexed by integers $k,\ell \geq 0$, and choose
\begin{equation}
\label{xikforms} \xi_{k} = \Big(\frac{k + 1}{z^{k + 2}} + \sum_{\ell \geq 0} u_{k,\ell}\,z^{\ell}\Big)\mathrm{d} z,\qquad \xi_k^* = \frac{z^{k + 1}}{k + 1},\qquad \theta = \sum_{r \in \mathbb{Z}} t_{r}z^{r}(\mathrm{d} z)^{-1},
\end{equation}
where $t_r$ are formal parameters. Then, $A$ is fully symmetric. Further, the triple $(A,B,C)$ given by \eqref{Cr} defines a classical Airy structure if and only if $\theta \in z^{-1}\mathbb{C}[[z]].(\mathrm{d} z)^{-1}$. In this case, $(A,B,C,D)$ defines a quantum Airy structure iff
\begin{equation}
\label{Deqnloop}\forall i \geq 1,\qquad \sum_{r \geq -1} t_{r}D^{i + r} = 0,
\end{equation}
\end{proposition}
The restriction on $\theta$ in this proposition implies that $A = 0$. Therefore, the partition function has $F_{0,n} = 0$ for all $n \geq 1$. The possibility to have a non-trivial partition function stems from the possible non-zero choices of $D$ satisfying \eqref{Deqnloop}. Lemma~\ref{Lemm1} in Appendix \ref{App4} gives vanishing rules for the $F_{g,n}$ of this quantum Airy structure. In particular, the case $t_{-1} \neq 0$ is not really interesting, as the partition function in this case is $Z = 1$, \textit{i.e.} all $F_{g,n}$ vanish.
We further find
$$
B^i_{j,k} = \sum_{r \geq -1} \frac{k + 1}{i + 1}t_{r}\,\delta_{i + j + r,k}.
$$
So, we get the commutation relations
\begin{equation}
\label{Lieloop} [\tilde{L}_i,\tilde{L}_j] = \sum_{r \geq -1} (j - i)t_{r}\tilde{L}_{i + j + r},\qquad \tilde{L}_i := (i + 1)L_i.
\end{equation}
This is just another form of a subalgebra of the Virasoro Lie algebra $\mathfrak{Vir}$. Recall that $\mathfrak{Vir}$ is the Lie algebra defined by generators $(\mathcal{L}_{i})_{i \in \mathbb{Z}}$ satisfying the commutation relations
$$
[\mathcal{L}_i,\mathcal{L}_{j}] = (i - j)\mathcal{L}_{i + j}.
$$
\begin{lemma}
\label{VirL} If $\theta(z) = t_{r_0}z^{r_0} + O(z^{r_0 + 1})$ with $t_{r_0} \neq 0$, define $\tfrac{z^{r_0}}{\theta(z)} = \sum_{k \geq 0} \tau_{k}\,z^{k}$, and for $n \geq r_0$
$$
\hat{L}_n = -\sum_{k \geq 0} \tau_{k}\tilde{L}_{n + k - r_0}\,.
$$
Then, for all $m,n \geq r_0$, we have that $[\hat{L}_{m},\hat{L}_{n}] = (m - n)\hat{L}_{m + n}$.
\end{lemma}
The proof is a straightforward computation and is omitted.
We can formulate a $\mathbb{Z}_{2}$-symmetric version of Proposition~\ref{Loop1}.
\begin{proposition}
\label{Loop2} Let $u_{k,\ell} = u_{\ell,k}$ be scalars indexed by integers $k,\ell \geq 0$, and choose
$$
\xi_{k} = \bigg(\frac{(2k + 1)}{z^{2k + 2}} + \sum_{\ell \geq 0} u_{k,\ell}\,z^{2\ell}\bigg)\,\mathrm{d} z,\qquad \xi_{k}^* = \frac{z^{2k + 1}}{2k + 1},\qquad \theta = \sum_{s \in \mathbb{Z}} t_{s} z^{2s}(\mathrm{d} z)^{-1}.
$$
and assume $\theta \in \mathbb{C}[z^{-2},z^2]].(\mathrm{d} z)^{-1}$. Then, $A$ is fully symmetric. Further, the triple $(A,B,C)$ given by \eqref{Cr} defines a classical Airy structure if and only if $\theta \in z^{-2}\mathbb{C}[[z^2]].(\mathrm{d} z)^{-1}$. In this case, $(A,B,C,D)$ defines a quantum Airy structure iff
\begin{equation}
\label{DDD}\forall k \geq 0,\qquad \sum_{s \geq 0} t_{s - 1}D^{s + k} = \frac{t_{-1}^2}{8}\,\delta_{k,0} + \frac{t_{-1}t_{0} + u_{0,0}t_{-1}^2}{4}\,\delta_{k,1}.
\end{equation}
\end{proposition}
The restriction on $\theta$ in this proposition implies that
\begin{eqnarray}
\label{Aexp} A^i_{j,k} & = & t_{-1}\delta_{i,j,k,0}. \\
\label{Bexp}B^i_{j,k} & = & \sum_{s \geq -1} \frac{2k + 1}{2i + 1}t_{s}\,\delta_{i + j + s,k} + t_{-1}u_{0,0}\delta_{i,j,k,0}.
\end{eqnarray}
In particular, if $t_{-1} = 0$, then $A = 0$ hence $F_{0,n} = 0$ for all $n$. Lemma~\ref{Lemm2} in Appendix \ref{App4} gives more general vanishing rules for the $F_{g,n}$'s of this quantum Airy structure.
The commutation relations deduced from \eqref{Bexp} are
$$
[\tilde{L}_i,\tilde{L}_j] = \sum_{s \geq -1} (j - i)t_{s} \tilde{L}_{i + j + s},\qquad \tilde{L}_i := \frac{i + 1}{2}\,L_i.
$$
They are the same as \eqref{Lieloop}, so we obtain again a sub-Lie algebra of $\mathfrak{Vir}$.
\vspace{0.2cm}
Specializing from formal parameters $t_r$ to complex-valued parameters, we obtain that $\theta(z) = \sum_{r \geq -1} t_{r}z^{r}(\mathrm{d} z)^{-1}$ in Proposition~\ref{Loop1} -- and $\theta(z) = \sum_{s \geq -1} t_{s}z^{2s}(\mathrm{d} z)^{-1}$ in Proposition \ref{Loop2} define quantum Airy structures. Conversely, if $\theta$ contains higher negative powers, $(A,B,C,D)$ cannot be a quantum Airy structure for generic parameters $(t_r)_{r \in \mathbb{Z}}$ -- but we do not rule out neither confirm the existence of non-generic $\theta$ for which these formulas define quantum Airy structures.
In particular, if we assume $D^k = 0$ for $k \geq 2$, \eqref{DDD} is satisfied iff
$$
\label{DssoL}D^0 = \frac{t_{0} + u_{0,0}t_{-1}}{8},\qquad D^1 = \frac{t_{-1}}{24}\,.
$$
Note that, for fixed $a \geq 0$
$$
B^i_{a,a} = (2a + 1)\big(\delta_{i,0}t_{0} + \delta_{i,1}\tfrac{t_{-1}}{3}\big) + \delta_{i,a,0}t_{-1}u_{0,0}\,.
$$
Therefore, ${\rm Tr}\,B^i$ is not well-defined, and even after zeta regularisation of the sum $\sum_{a \geq 0} (2a + 1)$, the expression $\tfrac{1}{2} ``\sum_{a} B^i_{a,a}"$ does not reproduce \eqref{DssoL}.
\noindent \textbf{Proof of Proposition \eqref{Loop1}.} First, we observe that the result for
\begin{equation}
\label{refxi} \xi_{k} = \xi_{k}^{(0)} := \frac{k + 1}{z^{k + 2}}\,\mathrm{d} z
\end{equation}
implies the general result, since the $(A,B,C)$ for general
$$
\xi_{k} = \xi_{k}^{(0)} + \sum_{\ell \geq 0} u_{k,\ell} \mathrm{d} \xi_{\ell}^*
$$
in \eqref{xikforms} is obtained from the $(A,B,C)$ for \eqref{refxi} by the symmetries \eqref{SymU}. For the choice \eqref{refxi}, unfortunately, we do not know an elegant proof of the result, thus we will proceed by direct computation.
We start with the warm-up case $\theta(z) = z^{r}(\mathrm{d} z)^{-1}$ for some integer $r \in \mathbb{Z}$, which yields
\begin{eqnarray}
A^i_{j,k} & := & \frac{1}{i + 1}\,\delta_{i + j + k + r + 2,0}. \nonumber \\
B^i_{j,k} & := & \frac{k + 1}{i + 1}\,\delta_{i + j + r,k}. \nonumber \\
C^i_{j,k} & := & \frac{(j + 1)(k + 1)}{i + 1}\,\delta_{i + r,j + k + 2}. \nonumber
\end{eqnarray}
In the following computations, it is implicit that the index $a$ is summed over, and it is important to keep in mind that indices are always $\geq 0$. We compute
\begin{eqnarray}
B^{i}_{j,a}B^{a}_{k,\ell} & = & \frac{\ell + 1}{(i + 1)(j + 1)}\,(j + 1)\,\delta_{i + j + r \geq 0} \delta_{i + j + k + 2r,\ell}. \nonumber \\
B^{i}_{k,a}B^{j}_{a,\ell} & = & \frac{\ell + 1}{(i + 1)(j + 1)}\,(i + k + r + 1)\,\delta_{i + k + r \geq 0}\delta_{i + j + k + 2r,\ell}. \nonumber \\
C^{i}_{\ell,a}A^{j}_{a,k} & = & \frac{\ell + 1}{(i + 1)(j + 1)}\,(i + r - \ell - 1)\,\delta_{i + r \geq \ell + 2}\,\delta_{i + j + k + 2r,\ell}. \nonumber
\end{eqnarray}
Therefore, the left-hand side of the \textbf{BA} relation reads
\begin{eqnarray}
& \frac{\ell + 1}{(i + 1)(j + 1)}\delta_{i + j + k + 2r,\ell}&\Big\{ (j - i)\delta_{i + j + r \geq 0} + (i + k + r +1)\delta_{i + k + r \geq 0} - (j + k + r + 1)\delta_{j + k + r \geq 0} \nonumber \\
&& + (i + r - \ell - 1)\delta_{i + r \geq \ell + 2} - (j + r - \ell - 1)\delta_{i + r \geq k + 2}\Big\} \nonumber \\
=& \frac{\ell + 1}{(i + 1)(j + 1)}\delta_{i + j + k + 2r,\ell}&\Big\{(i + r)(2\delta_{i + j + r < 0} - \delta_{i + k + r + 1,0} - \delta_{\ell + 1,i + r}) \nonumber \\
&& + k(\delta_{i + j + r < 0} - \delta_{i + k + r + 1,0}) + \ell(-\delta_{i + j + r < 0} + \delta_{\ell + 1,i + r})\Big\},
\end{eqnarray}
where we exploited the constraint in the delta function prefactor to get rid of all $j$'s, and we used $\delta_{A} = 1 - \delta_{\overline{A}}$ where $\overline{A}$ is the negation of $A$. The coefficients of $\delta_{i + k + r + 1,0}$ cancel each other, and likewise for $\delta_{\ell + 1,i + r}$. We are left with a non-zero multiple of
\begin{equation}
\label{BBend} \mathrm{L.H.S.}\,\,\mathrm{of}\,\,\textbf{BA} = (i - j)\delta_{i + j + r < 0}\delta_{i + j + k + 2r,\ell}.
\end{equation}
For non-negative indices $i,j$, the set $\{i + j + r < 0\}$ is empty if and only if $r \geq 0$. For $r = -1$, it consists only of $(i,j) = (0,0)$, but the prefactor $(i - j)$ vanish in this case, so \eqref{BBend} also vanish identically if $r = -1$. When $r \leq -2$, one can find non-negative $(i,j,k,\ell)$ for which \eqref{BBend} is non-zero. Therefore, the \textbf{BB-AC} relation holds if and only if $r \geq -1$.
Next, we compute
\begin{eqnarray}
B^{i}_{j,a}A^{a}_{k,\ell} & = & \frac{\delta_{i + j + k + \ell + 2r + 2,0}}{(i + 1)(j + 1)}\,(j + 1)\delta_{i + j + r \geq 0}\,, \nonumber \\
B^{i}_{k,a}A^{j}_{a,\ell} & = & \frac{\delta_{i + j + k + \ell + 2r + 2,0}}{(i + 1)(j + 1)}\,(i + k + r +1)\,\delta_{i + k + r \geq 0}\,, \nonumber \\
B^{i}_{\ell,a}A^{j}_{a,k} & = & \frac{\delta_{i + j + k + \ell + 2r + 2,0}}{(i + 1)(j + 1)}\,(i + \ell + r + 1)\,\delta_{i + \ell + r \geq 0}. \nonumber
\end{eqnarray}
Applying the same principles used in the previous computation, the left-hand side of the \textbf{BA} relation reads
\begin{equation}
\label{BAend} \mathrm{L.H.S.}\,\,\mathrm{of}\,\,\textbf{BA} = \frac{\delta_{i + j + k + \ell + 2r + 2,0}}{(i + 1)(j + 1)}\,(i-j)\,\delta_{i + j + r + 1 \leq 0}.
\end{equation}
A similar analysis shows that \eqref{BAend} is identically zero (\textit{i.e.} the \textbf{BA} relation is satisfied) if and only if $r \geq -1$.
We also compute
\begin{eqnarray}
B^{i}_{j,a}C^{a}_{k,\ell} & = & \frac{(k + 1)(\ell + 1)}{(i + 1)(j + 1)}\,(j + 1)\delta_{i + j + 2r,k + \ell + 2}\,\delta_{i + j + r \geq 0}, \nonumber \\
C^{i}_{k,a}B^{j}_{a,\ell} & = & \frac{(k + 1)(\ell + 1)}{(i + 1)(j + 1)}\delta_{i + j + 2r,k + \ell + 2}\,\delta_{i + r \geq k +2}, \nonumber \\
C^{i}_{\ell,a}B^{j}_{a,k} & = & \frac{(k + 1)(\ell + 1)}{(i + 1)(j + 1)}\delta_{i + j + 2r,k + \ell + 2}\,\delta_{i + r \geq \ell + 2}, \nonumber
\end{eqnarray}
and we obtain that the left-hand side of the \textbf{BC} relation reads
\begin{equation}
\label{BCend} \mathrm{L.H.S.}\,\,\mathrm{of}\,\,\textbf{BC} = \frac{(k + 1)(\ell + 1)}{(i + 1)(j + 1)}\,\delta_{i + j + 2r,k + \ell + 2}\,(i - j)\delta_{i + j + r + 1 \leq 0}.
\end{equation}
Here, we see that under the condition $i + j + 2r = k + \ell + 2$, we have $i + j + r + 1 \geq 1 - r$. Therefore, if $r \leq 0$, \eqref{BCend} vanishes identically. But $r \geq 0$ also implies that $i + j + r + 1 \leq 0$ cannot be satisfied for non-negative $(i,j)$. Hence, \eqref{BCend} vanishes identically (\textit{i.e.} the \textbf{BC} relation is satisfied) for any $r \in \mathbb{Z}$.
Let us now assume $r \geq -1$ and analyze the \textbf{D} relation. As we have $A = 0$, it takes the form, for all $i > j \geq 0$.
$$
\sum_{a \geq 0} \frac{(j - i)(a + 1)}{(i + 1)(j + 1)}\,\delta_{i + j + r,a}\,D^{a} = 0
$$
which is equivalent to the system of equations $D^{k + r} = 0$ for all $k \geq 1$.
Now we consider the case
$$
\theta = \sum_{r \geq -1} t_r\,z^{r}(\mathrm{d} z)^{-1}.
$$
This results in a decomposition
$$
X = \sum_{r \geq -1} t_r\,{}^{r}X,\qquad X \in \{A,B,C\}.
$$
As $(t_{r})_{r \in \mathbb{Z}}$ are considered as formal parameters, while checking the three quadratic relations it is enough to check the coefficient of the monomial $t_{r}t_{s}$ with $r \neq s$. Indeed the coefficient of $t_{r}^2$ already displays the relations which $({}^{r}A,{}^{r}B,{}^{r}C)$ satisfy according to Step 1. We should therefore check that
\begin{eqnarray}
{}^{r}B^{i}_{j,a}\cdot {}^{s}B^{a}_{j,k} + {}^{s}B^i_{j,a}\cdot {}^{s}B^a_{j,k} + {}^{r} B^i_{k,a}\cdot {}^{s}B^{j}_{a,\ell} + {}^{s}B^i_{k,a}\cdot {}^{r} B^j_{a,\ell} + \big({}^{r}C^i_{\ell,a}\cdot {}^{s} A^j_{a,k} + {}^{s}C^{i}_{\ell,a}\cdot {}^{r} C^j_{a,k}\big) - (i \leftrightarrow j) = 0,
\end{eqnarray}
and likewise for the two other relations. This is again checked by direct computations, which are very similar to the previous ones, thus omitted. Let us examine in this context the \textbf{D} relation. We get
$$
\forall i > j \geq 0,\qquad (i - j) \sum_{r \geq -1} t_{r} D^{i + j + r} = \sum_{r,s \geq -1} t_rt_s \frac{\delta_{i + j + r + s,0}}{2}(\sigma_{i + r - 2} - \sigma_{j + r - 2}).
$$
One can check that the right-hand side always vanish. Therefore, the \textbf{D} relation is equivalent to
\begin{equation}
\label{epsconstraint} \forall i \geq 1,\qquad \sum_{r \geq -1} t_r D^{r + i} = 0.
\end{equation}
Therefore, if $t_{k} = 0$ for $k \in \{0,\ldots,r_0\}$, $(D^i)_{i = 0}^{r_0}$ can be chosen arbitrarily.
\hfill $\Box$
\noindent \textbf{Proof of Proposition~\ref{Loop2}} Due to the symmetry \eqref{SymU} it is sufficient to consider
$$
\xi_k := \frac{(2k +1)}{z^{2k + 2}}\,\mathrm{d} z.
$$
If we denote $\eta_k$ the $1$-forms used in Proposition~\ref{Loop1}, note that $\xi_k = \eta_{2k}$ is just a subset of the $(\eta_k)_{k \geq 0}$. So, the case $\theta \in \mathbb{C}[[z]](\mathrm{d} z)^{-1}$ is covered by the previous Theorem with $r = 2s$ (excluding the statement about the \textbf{D} relation), by restricting to even indices. Indeed, although we have to sum over all indices $a$ to check the relations, the terms with odd $a$ are always zero since $X^{p}_{q,m} = 0$ whenever $p + q + m \neq 0\,\,{\rm mod}\,\,2$ for any $X \in \{A,B,C\}$, and two of the indices are not summed over and always even by our restriction. This is not so for the \textbf{D} relation, because the indices $a$ and $b$ which are summed over appear in the same symbol.
The case $\theta \in \big(t_{-1}z^{-2} + \mathbb{C}[[z]]\big)(\mathrm{d} z)^{-1}$ needs special care, as the relations \textbf{BA} and \textbf{BB-AC} failed in the proof of Proposition~\ref{Loop1} where indices of any parity were allowed.
We first focus on the warm-up case $\theta(z) = z^{-2}(\mathrm{d} z)^{-1}$, for which
\begin{eqnarray}
A_{j,k}^i & = & \delta_{i,j,k,0}. \nonumber \\
B_{j,k}^i & = & \frac{2k + 1}{2i + 1}\,\delta_{i + j,k + 1}. \nonumber \\
C_{j,k}^i & = & \frac{(2j + 1)(2k + 1)}{2i + 1}\,\delta_{i,j + k + 2}. \nonumber
\end{eqnarray}
We just have to check that the following specialization of \eqref{BBend} and \eqref{BAend}
$$
2(i - j)\delta_{i + j < 1}\delta_{i + j + k,\ell + 1} = 0,\qquad 2(i - j)\delta_{i + j + k + \ell + 1,0}\delta_{i + j < 1} = 0.
$$
Checking the first one is the same as checking \eqref{BBend} for $r = -1$ in the previous proof. And it is obvious that the expression is always $0$, as it forces $i = j = 0$ and thus $k + \ell + 1 = 0$. So, the three relations hold for $s = -1$ as well. The \textbf{D} relation reads
$$
\delta_{i + j + s \geq 0}(i - j)(2(i + j + s) + 1)D^{i + j + s} = \frac{\delta_{i + j + 2s,0}}{4}(\tilde{\sigma}_{i + s - 1} - \tilde{\sigma}_{j + s - 1}),
$$
with
$$
\tilde{\sigma}_m = \sum_{a + b = m} (2a + 1)(2b + 1).
$$
For $t_{-1} = 0$ we arrive to the same result, \textit{i.e.} $D^k$ is arbitrary for $k = 0,\ldots,r$ and $D^k = 0$ for $r \geq k + 1$. For $t_{-1} \neq 0$, we obtain
$$
D^k = \frac{\delta_{k,1}}{8},
$$
where the non-zero value is fixed by $i = 2,j = 0$ and the fact that $\tilde{\sigma}_{0} = 1$.
The proof for the more general case
$$
\theta(z) = \sum_{s \geq -1} t_{s}\,z^{2s}(\mathrm{d} z)^{-1}
$$
is very close to the one in absence of $\mathbb{Z}_{2}$-symmetry, thus omitted. Let us only examine the \textbf{D} relation in this case. We get
$$
\forall i > j \geq 0,\qquad (i - j)\sum_{s \geq -1} (2(i + j + s) + 1)t_s D^{i + j + s} = \frac{t_{-1}}{4}(t_{-1}\delta_{i,2} + t_{0}\delta_{i,1}),
$$
which is equivalent to
\begin{eqnarray}
\sum_{s \geq -1} t_{s}(2s + 3)D^{s+1} & = & \frac{t_{-1}t_{0}}{4}\,, \nonumber \\
\sum_{s \geq -1} t_{s}(2s + 5)D^{s+2} & = & u_{0,0}\frac{t_{-1}^2}{8}\,, \nonumber \\
\sum_{s \geq -1} t_{s}(2s + 2k + 1)D^{s + k} & = & 0, \qquad \forall k \geq 3\,. \nonumber
\end{eqnarray}
\hfill $\Box$
\subsection{Loop space of Frobenius algebras}
\label{S444}
Let $\mathbb{A}$ be a (commutative) Frobenius algebra and recall the notations of Section~\ref{S4Frob}. We choose an orthonormal basis $(e_{\alpha})_{\alpha}$ of $\mathbb{A}$, i.e $e_{\alpha} = e^*_{\alpha}$. Set $\mathcal{V} := \mathbb{A}[[z]]$. The proofs of Propositions~\ref{FrobABCD} and \ref{Loop1}-\ref{Loop2} can easily be adapted to this setting. If $f \in \mathbb{A}[z^{-1},z]].(\mathrm{d} z)$, we define
$$
\Phi(f) := \mathop{\rm Res}\limits_{z \rightarrow 0}\, \phi(f(z))
$$
using the linear form $\varphi\,:\,\mathbb{A} \rightarrow \mathbb{C}$ provided by the Frobenius structure. Let $\xi_{i,\alpha}$ be linearly independent family of elements of $\mathcal{V}$, indexed by $\alpha$ and integers $i \geq 0$.
We declare
\begin{eqnarray}
A^{(i,\alpha)}_{(j,\beta),(k,\gamma)} & = & \Phi(\xi_{i,\alpha}^*\,\mathrm{d}\xi_{j,\beta}^*\,\mathrm{d}\xi_{k,\gamma}^*\,\theta), \nonumber \\
B^{(i,\alpha)}_{(j,\beta),(k,\gamma)} & = & \Phi(\xi_{i,\alpha}^*\,\mathrm{d}\xi_{j,\beta}^*\,\xi_{k,\gamma}\,\theta), \nonumber \\
\label{CrcCC} C^{(i,\alpha)}_{(j,\beta),(k,\gamma)} & = & \Phi(\xi_{i,\alpha}^*\,\xi_{j,\beta}\,\xi_{k,\gamma} \theta),
\end{eqnarray}
where $\theta \in \mathbb{A}[z^{-1},z]].(\mathrm{d} z)^{-1}$.
\begin{proposition}
\label{ThmLoopTQFT}Let $v_{(k,\alpha),(\ell,\beta)} = v_{(\ell,\beta),(k,\alpha)}$ be scalars indexed by basis elements $\alpha,\beta$ of $\mathbb{A}$ and integers $k,\ell \geq 0$. Assume
$$
\theta = \sum_{\substack{r \geq -1 \\ \alpha}} t_{r,\alpha} z^{r}(\mathrm{d} z)^{-1}e_{\alpha}
$$
for some scalars $t_{r,\alpha}$, and choose
$$
\xi_{k,\alpha} = \bigg(\frac{(k + 1)e_{\alpha}\mathrm{d} z}{z^{k + 1}} + \sum_{\substack{\ell \geq 0 \\ \beta}} v_{(k,\alpha),(\ell,\beta)}\,z^{\ell}e_{\beta}\bigg)\,\mathrm{d} z,\qquad \xi^*_{k,\alpha} = \frac{z^{k + 1}}{k + 1}\,e_{\alpha}.
$$
Then, the triple $(A,B,C)$ given by \eqref{CrcCC} defines a classical Airy structure. Further, $(A,B,C,D)$ is a quantum Airy structure iff
$$
\forall i \geq 1\,\,\forall \alpha_1,\alpha_2 ,\qquad \sum_{\substack{r \geq -1 \\ \alpha,\beta}} \varphi(e_{\alpha_1}e_{\alpha_2}e_{\alpha}e_{\beta})\,t_{r,\beta}D^{(r + i,\alpha)} = 0.
$$
\end{proposition}
\begin{proposition}
\label{ThmLoopZ2TQFT} Let $v_{(k,\alpha),(\ell,\beta)} = v_{(\ell,\beta),(k,\alpha)}$ be scalars indexed by basis elements $\alpha,\beta$ of $\mathbb{A}$ and integers $k,\ell \geq 0$. Choose
$$
\xi_{k,\alpha} = \bigg(\frac{(2k + 1)e_{\alpha}}{z^{2k + 2}} + \sum_{\substack{\ell \geq 0 \\ \beta}} v_{(k,\alpha),(\ell,\beta)}\,z^{2\ell}e_{\beta}\bigg)\,\mathrm{d} z,\qquad \theta = \sum_{\substack{s \geq -1 \\ \alpha}} t_{s,\alpha} z^{2s}(\mathrm{d} z)^{-1}e_{\alpha}.
$$
Then, $(A,B,C)$ given by \eqref{CrcCC} defines a classical Airy structure. Further, $(A,B,C,D)$ then defines a quantum Airy structure iff, for any $\alpha_1,\alpha_2$
\begin{eqnarray}
\sum_{\substack{s \geq -1 \\ \alpha,\beta}} \varphi(e_{\alpha_1}e_{\alpha_2}e_{\alpha}e_{\beta})(2s + 3)t_{s,\alpha}D^{(s + 1,\beta)} & = & \sum_{\alpha,\beta} \varphi(e_{\alpha_1}e_{\alpha_2}He_{\alpha}e_{\beta})\,\frac{t_{-1,\alpha}t_{-1,\beta}}{8}\,, \nonumber \\
\sum_{\substack{s \geq -1 \\ \alpha,\beta}} \varphi(e_{\alpha_1}e_{\alpha_2}e_{\alpha}e_{\beta})\,(2s + 5)t_{s,\alpha}D^{(s + 2,\beta)} & = & \sum_{\alpha,\beta} \bigg( \varphi(e_{\alpha_1}e_{\alpha_2}He_{\alpha}e_{\beta})\,\frac{t_{-1,\alpha}t_{0,\beta}}{4} \nonumber \\
&& + \sum_{\gamma,\epsilon} \varphi(e_{\alpha_1}e_{\alpha_2}e_{\alpha}e_{\beta} e_{\gamma}e_{\epsilon})\,u_{(0,\gamma),(0,\epsilon)} \,\frac{t_{-1,\alpha} t_{-1,\beta}}{2}\bigg)\,, \nonumber \\
\label{fdsugn}\sum_{\substack{s\geq -1 \\ \alpha,\beta}} \varphi(e_{\alpha_1}e_{\alpha_2}e_{\alpha}e_{\beta})(2s + 2i + 1)t_{s,\alpha}D^{(s + i,\beta)} & = & 0, \qquad \forall i \geq 3\,,
\end{eqnarray}
where $H := \sum_{\alpha} e_{\alpha}^2$.
\end{proposition}
The proofs combine the two aspects of the proofs given in Sections~\ref{S4Frob} and \ref{S4Loop}, and hence are omitted. They rely on the fact that, each of the three terms in the three relations are already symmetric in $i$ and $j$ in the Frobenius algebra case. Note that, if we assume $\mathbb{A}$ is semi-simple and $(e_{\alpha})_{\alpha}$ is an orthonormal basis such that $e_{\alpha}e_{\beta} = \delta_{\alpha\beta}e_{\alpha}$, the constraints on $D$ can be rewritten, for all $\alpha$ as follows
\begin{eqnarray}
\sum_{s \geq -1} (2s + 3)t_{s,\alpha}\,D^{(s + 1,\alpha)} & = & \frac{t_{-1,\alpha^2}}{8}, \nonumber \\
\sum_{s \geq -1} (2s + 5)t_{s,\alpha}\,D^{(s + 2,\alpha)} & = & \frac{t_{-1,\alpha}t_{0,\alpha}}{4} + u_{(0,\alpha),(0,\alpha)}\frac{t_{-1,\alpha}^2}{2}, \nonumber \\
\sum_{s \geq -1} (2s + 2i + 1)t_{s,\alpha}\,D^{(s + i,\alpha)} & = & 0. \nonumber
\end{eqnarray}
Let us describe this classical Airy structure in the language of Section~\ref{SSymp}. On $W = \mathbb{A}((z^{-1}))$ we consider the following symplectic form
$$
\omega(f(z),g(z)) = \mathop{\rm Res}\limits_{z \to 0} \varphi\big(f(z)\mathrm{d} g(z)\big).
$$
Let $W = V^* \oplus V$ be a polarization of $W$, where $V$ (resp. $V^*$) has basis $\left(\xi_{k,\alpha}(z)\right)_{k,\alpha}$, respectively $\big(\xi_{k,\alpha}^*(z) := e_\alpha{z^{k+1} \over k+1}\big)_{k,\alpha}$, such that
$$
\forall (k,l) \in \mathbb{N}^2 \, , \; \omega(\xi_{k,\alpha}^*(z), \xi_{l,\beta}(z)) = \delta_{k,l} \delta_{\alpha,\beta}.
$$
One defines a classical Airy structure given by the set of the infinitesimal symplectomorphisms $\left({\cal L}_{k,\alpha}\right)_{k \in \mathbb{N}}$, where
$$
\forall k \in \mathbb{N} \, , \; \forall f \in W \, , \; {\cal L}_{k,\alpha} f(z) = \xi_{k,\alpha}^*(-z) \, {\theta(-z) \, \mathrm{d} f(-z) } ,
$$
and $\theta(z) \in z^{-1} \mathbb{A}[[z]] \cdot (\mathrm{d} z)^{-1}$.
\begin{corollary}
The operators $ \left({\cal L}_k\right)_{k \geq 0}$ together with the above orthonormal basis of $V$ and $V^*$ define a classical Airy structure.
\end{corollary}
\section{Relation with the topological recursion of \cite{EOFg}}
\label{TRcomp}
\subsection{Comparison}
\label{comparTR}
The original setting of \cite{EOFg} for the topological recursion is the data of a spectral curve, \textit{i.e.}
\begin{itemize}
\item[$\bullet$] a branched covering with simple ramification points $x\,:\,\Sigma \rightarrow \Sigma'$, where $\Sigma'$ is an open subset of $\mathbb{P}^1$.
\item[$\bullet$] a meromorphic function $y$ on $\Sigma$, such that the zeroes of $\mathrm{d} y$ are distinct from the zeroes of $\mathrm{d} x$. We set $\omega_{0,1} := y\mathrm{d} x$.
\item[$\bullet$] a symmetric bidifferential $\omega_{0,2}$ on $\Sigma^2$ with a double pole at coinciding points and the following behavior in local coordinates
$$
\omega_{0,2}(p_1,p_2) = \frac{\mathrm{d} z(p_1) \mathrm{d} z(p_2)}{(z(p_1) - z(p_2))^2} + O(1).
$$
Such an object is sometimes called a \emph{fundamental bidifferential of the second kind} on $\Sigma$.
\end{itemize}
We denote $\mathfrak{r} \subset \Sigma$ the set of the ramification points, \textit{i.e.} the zeros of $\mathrm{d} x$. As they are simple, we can find around each $r \in \mathfrak{r}$ a local coordinate $z$ such that
$$
x(p) = x(r) + \tfrac{z(p)^2}{2}.
$$
Let $U \subseteq \Sigma$ be the disjoint union of small enough neighborhoods of the ramification points, in which $\iota\,:\,z \mapsto -z$ is a well-defined holomorphic involution. By construction $x \circ \iota = x$. We introduce the recursion kernel
$$
K(p_0,p) = \frac{1}{2}\,\frac{\int_{\iota(p)}^{p} \omega_{0,2}(\cdot,p_0)}{\omega_{0,1}(p) - \omega_{0,1}(\iota(p))}.
$$
It is defined for $(p_0,p) \in \Sigma \times U$ as a $1$-form in $p_0$ and the inverse of a $1$-form in $p$. For $2g - 2 + n > 0$, \cite{EOFg} defines by induction
\begin{equation}
\label{oldTR}\omega_{g,n}(p_1,I) := \sum_{r \in \mathfrak{r}} \mathop{\rm Res}\limits_{p \rightarrow r} K(p_1,p)\bigg\{\omega_{g - 1,n + 1}(p,\iota(p),I) + \sum_{\substack{g' + g'' = g \\ J' \sqcup J'' = I}}^{*} \omega_{g',1 + |J'|}(p,J')\omega_{g'',1+|J''|}(\iota(p),J'')\bigg\},
\end{equation}
where $I = \{p_2,\ldots,p_n\}$, and $\sum^*$ means that we exclude the terms of the form $\omega_{0,1}\omega_{g,n}$ from the sum. Although \eqref{oldTR} gives a special role to the variable $p_1$, \cite{EOFg} proves inductively that $\omega_{g,n}(p_1,\ldots,p_n)$ is symmetric under permutation of all the $p_i$'s, therefore for $L = \{p_1,\ldots,p_{\ell}\}$ an unordered $\ell$-tuple of variables in $\Sigma$, the notation $\omega_{h,\ell}(L)$ makes sense. Manifestly \eqref{oldTR} produces differential forms whose only poles are located at the ramification points. In other words
$$ \omega_{g,n} \in H^0(\Sigma^{\times n}, K^{\boxtimes n}(\star \mathfrak{r}))^{\mathfrak{S}_n}.$$
In \cite{EInter}, it is shown that the $\omega_{g,n}$ can be decomposed on a suitable family of meromorphic $1$-forms. To be self-contained we review this proof, and make explicit the recursion following from \eqref{oldTR} for the coefficients of the decomposition.
\begin{definition}
For $k \geq 0$ and $r \in \mathfrak{r}$, we define for $p_0 \in \Sigma$ the meromorphic $1$-form
\begin{equation}
\label{xiforms} \xi_{k,r}(p_0) := \mathop{\rm Res}\limits_{p \rightarrow r} \Big(\int_{r}^{p} \omega_{0,2}(p_0,\cdot)\Big)\,\frac{(2k + 1)\mathrm{d} z(p)}{z(p)^{2k + 2}}.
\end{equation}
We also define, for $p_0$ in a neighborhood of $r$ in $\Sigma$
$$
\xi_{k,r}^*(p_0) := \frac{z^{2k + 1}(p_0)}{2k + 1},\qquad \theta(p_0) := \frac{-2}{\omega_{0,1}(p_0) - \omega_{0,1}(\iota(p_0))},
$$
and if $p_0$ is in a neighborhood of $r_0 \neq r$, we define $\xi_{k,r}^*(p_0) := 0$.
\end{definition}
\begin{lemma}
\label{LemmaFogi}
For $2g - 2 + n > 0$, there exists a unique decomposition with finitely many non-zero terms
\begin{equation}
\label{Fogi}\omega_{g,n}(p_1,\ldots,p_n) = \sum_{\substack{r_1,\ldots,r_n \in \mathfrak{r} \\ k_1,\ldots,k_n \geq 0}} W_{g,n}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots & k_n \end{smallmatrix}\right]\,\prod_{i = 1}^n \xi_{k_i,r_i}(p_i).
\end{equation}
\end{lemma}
More precisely, one can show \cite{EInter} that the coefficients in \eqref{Fogi} with $\sum_{i} k_i > 3g - 3 + n$ vanish. For completeness, we also give a proof in Appendix~\ref{App4}.
We can now compare with the quantum Airy structure of Section~\ref{S444}. We take $\mathbb{A} = \bigoplus_{r \in \mathfrak{r}} \mathbb{C}$ as the sum of trivial $1$-dimensional Frobenius algebras, and we let $V = \mathbb{A}[[z]]$ be the vector space with a basis indexed by $k \geq 0$ and $r \in \mathfrak{r}$. As we assumed $y$ is holomorphic and $\mathrm{d} y$ has no zero at $\mathfrak{r}$, we deduce that $\theta(p)$ has an expansion for $p \rightarrow r$ of the form
\begin{equation}
\label{om10exp}\theta(p) = \sum_{m \geq -1} t_{m,r}\,z^{2m}(p)\,(\mathrm{d} z(p))^{-1}.
\end{equation}
According to Proposition~\ref{ThmLoopZ2TQFT},
\begin{eqnarray}
A^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1,r_1}^*(p)\,\mathrm{d}\xi_{k_2,r_2}^*(p)\,\mathrm{d} \xi_{k_3,r_3}^*(p)\,\theta(p), \nonumber \\
B^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1,r_1}^*(p)\,\mathrm{d} \xi_{k_2,r_2}^*(p)\,\xi_{k_3,r_3}(p)\,\theta(p), \nonumber \\
C^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1,r_1}^*(p)\,\xi_{k_2,r_2}(p)\,\xi_{k_3,r_3}(p)\,\theta(p), \nonumber \\
D^{(k,r)} & = & \delta_{k,0}\big(\tfrac{t_{-1,r}}{2}\,\phi_{0,2}\left[\begin{smallmatrix} r & r \\ 0 & 0 \end{smallmatrix}\right] + \tfrac{t_{0,r}}{8}\big)+ \delta_{k,1}\tfrac{t_{-1,r}}{24}
\label{qAIryTR}
\end{eqnarray}
is a quantum Airy structure. Here, $\varphi_{0,2}\left[\begin{smallmatrix} r & r \\ 0 & 0 \end{smallmatrix}\right]$ is a scalar, which corresponds to the constant term in the expansion of $\omega_{0,2}$ near $(p_1,p_2) = (r,r)$ in local coordinates $(z(p_1),z(p_2))$. One can then check that indeed $D$ is a solution of the \textbf{D} relation in the form \eqref{fdsugn}. Substituting the expansion (\ref{Fogi}) in the residue formula (\ref{oldTR}) gives a recursion for the $W_{g,n}$, which is identical to KS recursion \eqref{TRForm} for the Taylor coefficients $F_{g,n}$ of this quantum Airy structure. Since we can check (see the proof below) that the initial data are the same, this leads to
\begin{proposition}
\label{FWid} For $2g - 2 + n > 0$, $F_{g,n}\big((k_1,r_1),\ldots,(k_n,r_n)\big)$ computed by KS topological recursion for the quantum Airy structure \eqref{qAIryTR} and $W_{g,n}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots & k_n \end{smallmatrix}\right]$ computed by the topological recursion of \cite{EOFg}, agree.
\end{proposition}
\noindent \textbf{Proof.} We start by a preliminary study of the recursion kernel. If we expand $\omega_{0,2}(p_1,p_2)$ in local coordinates when $p_i$ is in a neighborhood of $r_i \in \mathfrak{r}$ we get that
\begin{equation}
\label{om20exp} \omega_{0,2}(p_1,p_2) = \frac{\delta_{r_1,r_2}\mathrm{d} z(p_1)\mathrm{d} z(p_2)}{(z(p_1) - z(p_2))^2} + \sum_{\ell_1,\ell_2 \geq 0} \phi_{0,2}\left[\begin{smallmatrix} r_1 & r_2 \\ \ell_1 & \ell_2 \end{smallmatrix}\right]\,z^{\ell_1}(p_1)z^{\ell_2}(p_2) \mathrm{d} z(p_1)\mathrm{d} z(p_2).
\end{equation}
We find the following expansion for \eqref{xiforms} when $p_0 \rightarrow r_0$ for some $r_0 \in \mathfrak{r}$
$$
\xi_{k,r}(p_0) = \frac{(2k + 1)\delta_{r,r_0}}{z^{2k + 2}(p_0)} + (2k + 1) \sum_{\ell \geq 0} \phi_{0,2}\left[\begin{smallmatrix} r & r_0 \\ 2k & \ell \end{smallmatrix}\right]\,z^{\ell}(p_0)\,\mathrm{d} z(p_0).
$$
In particular, $\xi_{k,r}(p_0)$ has a pole of order $2k + 2$ at $p_0 = r$, and is holomorphic elsewhere. We also find for $p$ in a neighborhood of $r$
$$
\tfrac{1}{2} \int_{\iota(p)}^{p}\omega_{0,2}(\cdot,p_0) = \sum_{\substack{k \geq 0 \\ r \in \mathfrak{r}}} \xi_{k,r}(p_0)\,\xi_{k,r}^*(p)
$$
under the condition $|z(p_0)| > |z(p)|$ when $p_0$ is in the neighborhood of $r$. Here we have used that
$$
\frac{1}{(z(p_0) - z(p))^2} = \sum_{\ell \geq 0} \frac{(\ell + 1)\,z^{\ell}(p)}{z^{\ell + 2}(p_0)}
$$
for $p_0,p$ in the neighborhood of the same $r$. To perform the residue computation in \eqref{oldTR}, we will need the expansion of the recursion kernel $K(p_0,p)$ around $p \rightarrow r$
\begin{equation}
\label{Klocal} K(p_0,p) = -\tfrac{1}{2}\sum_{k \geq 0} \xi_{k,r}(p_0)\,\xi_{k,r}^*(p)\,\theta(p),\qquad \xi_{k,r}^*(p) \in O(z(p)^{2k - 1}).
\end{equation}
\vspace{0.2cm}
Let us start by computing $\omega_{0,3}$.
$$
\omega_{0,3}(p_1,p_2,p_3) = \sum_{r \in \mathfrak{r}} \mathop{\rm Res}\limits_{p \rightarrow r} K(p_1,p)\{\omega_{0,2}(p,p_2)\omega_{0,2}(\iota(p),p_3) + \omega_{0,2}(\iota(p),p_2)\omega_{0,2}(p,p_3)\big\}. \nonumber \\
$$
Since $\theta(p) \in O\big(z^{-2}(p)(\mathrm{d} z)^{-1}\big)$, $K(p_1,p)$ has a simple pole at $p = r$. So, the residue selects the coefficient of $(\mathrm{d} z(p))^2$ in $\big\{\cdots \big\}$, and as $\mathrm{d} \xi_{k,r}(p) = z^{k}(p)\mathrm{d} z(p)$, we find that
\begin{equation}
\label{ome03} \omega_{0,3}(p_1,p_2,p_3) = \delta_{k_1,k_2,k_3,0} \sum_{r \in \mathfrak{r}} \bigg(\prod_{i = 1}^3 \mathrm{d}\xi_{0,r}(p_i)\bigg) \mathop{\rm Res}\limits_{p \rightarrow r} \xi_{0,r}^*(p)\mathrm{d}\xi_{0,r}^*(p)\mathrm{d}\xi_{0,r}^*(p)\theta(p).
\end{equation}
The factor $-\tfrac{1}{2}$ in front of \eqref{Klocal} disappeared as \eqref{ome03} has two terms with equal contribution, and the $\iota(p)$ in one of the factor $\omega_{0,2}$ results into a minus sign in the local coordinate $z$. We therefore have proved \eqref{Fogi} for $(g,n) = (0,3)$, and can identify the coefficients $W_{0,3}\left[\begin{smallmatrix} r_1 & r_2 & r_3 \\ k_1 & k_2 & k_3 \end{smallmatrix}\right]$ with $A^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)}$ introduced in \eqref{qAIryTR} -- these coefficients vanish unless $k_1 = k_2 = k_3 = 0$ and $r_1 = r_2 = r_3$.
\vspace{0.2cm}
Likewise we compute $\omega_{1,1}$. Examining the local behavior at ramification points, we find
\begin{eqnarray}
\omega_{1,1}(p_1) & = & \sum_{r \in \mathfrak{r}} \mathop{\rm Res}\limits_{p \rightarrow r} K(p_1,p)\omega_{0,2}(p,\iota(p)) \nonumber \\
& = & \sum_{r \in \mathfrak{r}} \sum_{k = 0}^{1} \xi_{k,r}(p_1) \mathop{\rm Res}\limits_{p \rightarrow r} \tfrac{z(p)^{2k + 1}}{2(2k + 1)}\Big(\tfrac{t_{-1,r}}{z^2(p)} + t_{0} + O(z^2(p))\Big)\Big(\tfrac{1}{4z^2(p)} + \phi_{0,2}\left[\begin{smallmatrix} r & r \\ 0 & 0 \end{smallmatrix}\right] + O(z^2(p))\Big) \nonumber \\
& = & \sum_{r \in \mathfrak{r}} \big(\tfrac{t_{-1,r}}{2}\,\phi_{0,2}\left[\begin{smallmatrix} r & r \\ 0 & 0 \end{smallmatrix}\right] + \tfrac{t_{0,r}}{8}\big)\xi_{0,r}(p_1) + \tfrac{t_{-1,r}}{24}\,\xi_{1,r}(p_1),
\end{eqnarray}
which proves \eqref{Fogi} for $(g,n) = (1,1)$ with $W_{1,1}\left[\begin{smallmatrix} r \\ k \end{smallmatrix}\right] = D^{(k,r)}$ given in \eqref{qAIryTR}.
\vspace{0.2cm}
Now let $2g - 2 + n > 2$ and assume the claim of Lemma~\ref{LemmaFogi} has been established for all $(g',n')$ such that $2g' - 2 + n' < 2g - 2 + n$. Let $I = \{p_2,\ldots,p_n\}$ an unordered $(n - 1)$-uple of variables in $\Sigma$. In equation \eqref{oldTR} for $\omega_{g,n}(p_1,I)$, we denote $\omega_{g,n}^{B}$ the sum of terms in the right-hand side involving $\omega_{0,2}\omega_{g,n - 1}$, and $\omega_{g,n}^{C}$ the sum of all the other terms. We have that
\begin{equation}
\label{omegagnB}\omega_{g,n}^{B}(p_1,I) = \sum_{i = 2}^n \mathop{\rm Res}\limits_{p \rightarrow \mathfrak{r}} K(p_1,p)\big(\omega_{0,2}(p,p_i)\omega_{g,n - 1}(\iota(p),I\setminus\{p_i\}) + \omega_{0,2}(\iota(p),p_i)\omega_{g,n - 1}(p,I\setminus\{p_i\})\big).
\end{equation}
As $K(p_1,p)$ is invariant under $p \rightarrow \iota(p)$, the two terms give an equal contribution. The form \eqref{Fogi} of $\omega_{g,n - 1}$ by the induction hypothesis implies that
$$
\omega_{g,n - 1}(p,I\setminus\{p_i\}) = \tfrac{1}{2}\big(\omega_{g,n - 1}(p,I\setminus\{p_i\}) - \omega_{g,n - 1}(\iota(p),I\setminus\{p_i\})\big) + O(z(p)\mathrm{d} z(p)).
$$
As $\omega_{0,2}(p,p_i)$ is holomorphic near $p \rightarrow r$, we deduce that replacing it with its odd part in \eqref{omegagnB} does not change the residue
$$
\omega_{g,n}^B(p_1,I) = \sum_{i = 2}^n \mathop{\rm Res}\limits_{p \rightarrow \mathfrak{r}} K(p_1,p)\,\tfrac{1}{2}\big(\omega_{0,2}(\iota(p),p_i) - \omega_{0,2}(p,p_i)\big)\omega_{g,n - 1}(p,I\setminus\{p_i\}).
$$
We substitute in this formula, for $p$ in the neighborhood of $r$
$$
\tfrac{1}{2}\big(\omega_{0,2}(\iota(p),p_i) - \omega_{0,2}(p,p_i)\big) = -\sum_{k \geq 0} \xi_{k,r}(p_i)\mathrm{d}\xi_{k,r}^*(p),
$$
and the decomposition \eqref{Fogi} for $\omega_{g,n - 1}$. The result for $\omega_{g,n}^B$ decomposes like \eqref{Fogi} with coefficients
$$
W_{g,n}^{B}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots &k_n \end{smallmatrix}\right] = \sum_{i = 2}^n \sum_{k',r'} B^{(k_1,r_1)}_{(k_i,r_i),(k',r')}\,W_{g,n - 1}\left[\begin{smallmatrix} r' & r_{2} & \cdots & \widehat{r_i} & \cdots & r_n \\ k' & k_2 & \cdots & \widehat{k_i} & \cdots & k_n \end{smallmatrix}\right],
$$
where
$$
B^{(k_1,r_1)}_{(k_i,r_i),(k',r')} = \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1,r_1}^*(p)\,\mathrm{d} \xi_{k_i,r_i}^*(p)\,\xi_{k',r'}(p)\,\theta(p)
$$
as given in \eqref{qAIryTR}. Due to the local behavior of the integrand, $B^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)}$ vanishes when $r_1 \neq r_2$, or when $r_1 = r_2 \neq r_3$ and $k_1 + k_2 > 0$, or when $r_1 = r_2 = r_3$ and $k_1 + k_2 \geq k_3 + 1$. In particular these selection rules imply that there are finitely many non-zero $W_{g,n}^B$'s.
Let us turn to
$$
\omega_{g,n}^{C}(p_1,I) = \mathop{\rm Res}\limits_{p \rightarrow \mathfrak{r}} K(p_1,p)\bigg\{\omega_{g - 1,n + 1}(p,\iota(p),I) + \sum_{\substack{g' + g'' = g \\ J' \sqcup J'' = I}}^{**} \omega_{g',1+|J'|}(p,J')\omega_{g'',1 + |J''|}(\iota(p),J'')\bigg\},
$$
where $\sum^{**}$ excludes the terms of the form $\omega_{0,1}\omega_{g,n}$ or $\omega_{0,2}\omega_{g,n - 1}$. By induction hypothesis, we can directly substitute the decomposition \eqref{Fogi} for all the $\omega$'s involved in the left-hand side. We find that $\omega_{g,n}^{C}$ has a decomposition again of the form \eqref{Fogi}, with coefficients
\begin{eqnarray}
W_{g,n}^{C}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots & k_n \end{smallmatrix}\right] & = & \tfrac{1}{2} \sum_{\substack{k',k'' \geq 0 \\ r',r'' \in \mathfrak{r}}} C^{(k_1,r_1)}_{(k',r'),(k'',r'')}\bigg(W_{g - 1,n + 1}\left[\begin{smallmatrix} r' & r'' & r_2 & \cdots & r_n \\ k' & k'' & k_2 & \cdots & k_n \end{smallmatrix}\right] \nonumber \\
\label{WgnC} && + \sum_{\substack{g' + g'' = g \\ J' \sqcup J'' = \{2,\ldots,n\}}}^{**} W_{g',1+|J'|}\left[\begin{smallmatrix} r' & (r_j)_{j \in J'} \\ k' & (k_j)_{j \in J'} \end{smallmatrix}\right]W_{g'',1+ |J''|}\left[\begin{smallmatrix} r'' & (r_j)_{j \in J''} \\ k'' & (k_j)_{j \in J''} \end{smallmatrix}\right] \bigg),
\end{eqnarray}
where
$$
C^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} = \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1,r_1}^*(p)\,\xi_{k_2,r_2}(p)\,\xi_{k_3,r_3}(p)\,\theta(p)
$$
as given in \eqref{qAIryTR}. Due to the local behavior of the integrand, $C^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} = C^{(k_1,r_1)}_{(k_3,r_3),(k_2,r_2)}$ vanishes when $r_2,r_3 \neq r_1$ and $k_1 > 0$, or when $r_2 = r_1 \neq r_3$ and $k_1 \geq k_2 + 2$, or when $r_1 = r_2 = r_3$ and $k_1 \geq k_2 + k_3 + 3$. In particular, this implies that \eqref{WgnC} contains only finitely many non-zero terms. We therefore have justified that $\omega_{g,n} = \omega_{g,n}^{B} + \omega_{g,n}^{C}$ has the form \eqref{Fogi}, and proved Lemma~\ref{LemmaFogi} by induction.
Since we have checked $F_{0,3} = A = W_{0,3}$ and $F_{1,1} = D = W_{1,1}$, and the recursive rules to build the $W_{g,n}$'s agree with the KS topological recursion \eqref{TRForm} for the $F_{g,n}$'s, this entails Proposition~\ref{FWid}. \hfill $\Box$
More explicitly, in terms of coefficients of expansion of $\omega_{0,1}$ in \eqref{om10exp} and $\omega_{0,2}$ in \eqref{om20exp}, the relevant quantum Airy structure is
\begin{eqnarray}
A^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & = & \delta_{k_1,k_2,k_3,0}\delta_{r_1,r_2,r_3}t_{-1,r_1}, \nonumber \\
B^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & = & \frac{2k_3 + 1}{2k_1 + 1}\,\delta_{r_1,r_2}\bigg(\delta_{r_2,r_3}t_{k_3 - k_2 - k_1,r_1} + \delta_{k_1,k_2,0}\,\varphi_{0,2}\left[\begin{smallmatrix} r_3 & r_1 \\ 2k_3 & 0 \end{smallmatrix}\right]\bigg),\nonumber\\
C^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & = & \frac{(2k_3 + 1)(2k_2 + 1)}{2k_1 + 1}\bigg(\delta_{r_1,r_2,r_3}\,t_{1 + k_2 + k_3 - k_1,r_1} + \sum_{m = 0}^{1 + k_3 - k_1} \delta_{r_1,r_3}\,\varphi_{0,2}\left[\begin{smallmatrix} r_2 & r_1 \\ 2k_2 & 2m \end{smallmatrix}\right] t_{k_3 - k_1 - m,r_1} \nonumber \\
&& + \sum_{m = 0}^{1 + k_2 - k_1} \delta_{r_1,r_2}\,\varphi_{0,2}\left[\begin{smallmatrix} r_3 & r_1 \\ 2k_3 & 2m \end{smallmatrix}\right] t_{k_2 - k_1 - m,r_1} + \delta_{k_1,0}\,\varphi_{0,2}\left[\begin{smallmatrix} r_2 & r_1 \\ 2k_2 & 0 \end{smallmatrix}\right]\varphi_{0,2}\left[\begin{smallmatrix} r_3 & r_1 \\ 2k_3 & 0 \end{smallmatrix}\right] t_{-1,r_1}\bigg). \nonumber
\end{eqnarray}
\subsection{The point of view of Givental quantization of Lagrangian cones}
\label{SJUING}
One of the applications of the original topological recursion formalism is the study of Frobenius manifolds/cohomological field theories. In this setup, \cite{DBOSS} established that the topological recursion applied to a specific local spectral curve is equivalent to Givental's quantization formalism \cite{GiventalQuad} for computing the ancestor potential of a semi-simple cohomological field theory. This correspondence was obtained by a direct comparison of the result of the topological recursion and of Givental reconstruction procedure. In this section, we revisit this equivalence from the point of view of quantization of Givental's Lagrangian cone \cite{CoatesGivental,Giventalcone}, giving it a stronger geometric explanation. We first review the Lagrangian cone formalism, following Coates and Givental.
Let $V$ be a finite dimensional vector space equipped with a bilinear form $(\cdot,\cdot)$ and a distinguished vector $\mathbf{1}$, and let $\mathcal{W}:=V((z^{-1}))$ be the corresponding loop space equipped with the symplectic form\footnote{This is not the same symplectic form as in the example of Section~\ref{S444}. One can go from one to the other by a Laplace transform.} $\tilde{\omega}$ defined by
$$
\forall (f,g) \in \mathcal{W}^2,\qquad \tilde{\omega}(f,g) := {1 \over 2 {\rm i}\pi} \oint \left(f(-z),g(z)\right) \, \mathrm{d}\,z.
$$
Consider the polarization $\mathcal{W} = \mathcal{V}_+ \oplus \mathcal{V}_-$ where $\mathcal{V}_+ := V[[z]]$ and $\mathcal{V}_-:= z^{-1} V[[z^{-1}]]$. Then the symplectic form gives an identification $({\cal W},\tilde{\omega}) \simeq (T^*{\cal V}_+,\tilde{\omega})$.
Parametrizing elements $q$ of ${\cal V}_+$ by an infinite dimensional vector ${\bf t} :=(t_k)_{k\in \mathbb{N}}$,
$$
q(z) := \sum_{k=0}^\infty (t_k - \delta_{k,1}\mathbf{1}) z^k,
$$
one defines the graph ${\cal L}_{\cal F}$ of the derivative of a function ${\cal F}({\bf t})$ on ${\cal V}_+$ by
$$
{\cal L}_{\cal F}:=\{(p,q) \in T^*{\cal V}_+\,\,:\,\,\, p = \mathrm{d}_q {\cal F}({\bf t})\}.
$$
As a formal germ around $q = -z$, this defines a Lagrangian submanifold of $T^*{\cal V}_+$ and hence of $({\cal{W}},\tilde{\omega})$.
An interesting choice for such functions are genus $0$ free energies coming from a CohFT, \textit{i.e.} functions $F_0(\mathbf{t})$ satisfying the following three axioms. It is convenient to state them by choosing a basis $(e_{\nu})_{\nu = 1}^{d}$ of $V$ and denoting $g_{\mu,\nu} := (e_{\mu},e_{\nu})$ and by $(g^{\mu,\nu})_{\mu,\nu}$ its inverse matrix. For $k \geq 0$, we denote $t_k := \sum_{\nu = 1}^{d} t_k^{\nu} e_{\nu}$. With these notations, the three axioms defining a genus 0 free energy read as follows.
\begin{itemize}
\item The \emph{dilaton equation}, which states that $F_0$ is homogenous of degree 2
\begin{equation}
\label{DE}
2 F_0({\bf t}) = \sum_{k \in \mathbb{N}} \sum_{\nu = 1}^d t_{k}^\nu {\partial F_0({\bf t}) \over \partial t_{k}^\nu}.
\end{equation}
\item The \emph{string equation}, which decomposes the action of ${\partial \over \partial t_{0}^1}$ (the unit vector field)
\begin{equation}
\label{SE}
{\partial F_0 \over \partial t_{0}^1} = {1 \over 2} (t_0,t_0) + \sum_{k=0}^\infty \sum_{\nu = 1}^d t_{k+1}^\nu {\partial F_0 \over \partial t_k^\nu}.
\end{equation}
\item The \emph{topological recursion relations}\footnote{This set of equations is different from the topological recursion of \cite{EOFg}. It is unfortunate that both names coincide.}
\begin{equation}
\label{TRR}
\forall (\alpha,\beta,\gamma) \in \{1,\ldots,d\}^3\,\,\,\forall (k,l,m) \in \mathbb{N}^3 \, , \; {\partial^3 F_0 \over \partial t_{k+1}^\alpha \, \partial t_{l}^\beta \, \partial t_{m}^\gamma} = \sum_{\mu,\nu=1}^d {\partial^2 F_0 \over \partial t_{k}^\alpha \, \partial t_{0}^\mu } g^{\mu,\nu } {\partial^3 F_0 \over \partial t_{0}^\nu \, \partial t_{l}^\beta \, \partial t_{m}^\gamma} .
\end{equation}
\end{itemize}
The Lagrangians ${\cal L}_{F_0}$ defined by such functions have a very nice characterization.
\begin{theorem} \cite{Giventalcone}
\label{th.Givental.cone}
$F_0$ satisfies \mathop{\fam0 eq}\nolimits{DE}, \mathop{\fam0 eq}\nolimits{SE} and \mathop{\fam0 eq}\nolimits{TRR} if and only if ${\cal L}_{F_0}$ is a Lagrangian cone with vertex at the origin and such that its tangent spaces $L$ satisfy $zL = L$.
\end{theorem}
In addition, Givental described a large group of symmetries of the set of such cones.
\begin{theorem}\cite{Giventalcone}
The twisted loop group $G_{tw}$, consisting of elements $M(z) \in {\rm End}(V)[[z^{-1}]]$ such that $M^*(-z) M(z) = {\rm Id}$ preserves the class of cones of Theorem \ref{th.Givental.cone}.
\end{theorem}
Note that the condition $M^*(-z) M(z) = {\rm Id}$ implies that $M(z)$ defines a symplectomorphism of ${\cal W}$. The set of tangent spaces to such a cone carries the structure of a Frobenius manifold $\mathcal{M}$. For instance, this applies to (and was motivated by the application to) the genus $0$ descendent or ancestor potentials of Gromov-Witten theory of a complex projective variety \cite{CoatesGivental}, to genus $0$ correlation functions of cohomological field theories \cite{KMCohFT} and to quantum $K$-theory \cite{CoatesGivental}.
Conversely, if $\mathcal{M}$ is a semi-simple Frobenius manifold, there is a notion of a descendent (resp. an ancestor) potential $F_0^{\mathcal{M}}(\mathbf{t})$, which satisfy the axioms above. Here $(t^{\nu})_{\nu}$ are local flat coordinates on $\mathcal{M}$, and we fix a point $m \in \mathcal{M}$ to identify $V := T_{m}\mathcal{M}$. We denote as before $(t^{\nu}_{k})_{k \geq 0}$ the linear coordinates on $V[[z]]$. Dubrovin -- see e.g. \cite{Dubro0} -- associates to $\mathcal{M}$ a Riemann-Hilbert problem on $\mathbb{P}^1$. Its solution is an element $M_{v}(z)$ of the associated loop space depending on a point $m$ of the Frobenius manifold\footnote{This element and its factorization are unique if the Frobenius manifold admits an Euler vector field. Otherwise, one needs to fix the diagonal ambiguity by some other geometric condition.}, and it admits a Birkhoff factorization
$$
M_{v}(z) = M_{v,\infty}(z)^{-1} M_{v,0}(z),
$$
where $M_{v,0}(z)$ (resp. $M_{v,\infty}(z)$) is analytic and invertible for $|z|<1$ (resp. $1<|z|\leq\infty$). Combining Givental's analysis of the action of the twisted loop group \cite{Giventalcone} and Teleman's classification of semi-simple Frobenius manifolds \cite{Teleman}, one can conclude that the cone defined by the graph of the genus zero descendent (resp. ancestor) potential of a ${\rm dim}\,V = N$ semi-simple Frobenius manifold is obtained by the action of the symplectomorphism $\gamma(v) M_{v}(z)$ (resp. $M_{v,0}(z)$) on the cone ${\cal L}_N$ corresponding to the trivial theory of type $A_1^{\times N}$ where $\gamma(v)$ is a suitably chosen normalization factor.
Finally, Givental reconstruction procedure proved by Teleman through its classification can be expressed as the following quantization result.
\begin{theorem}\cite{Giventalcone,Teleman}
If ${\rm dim}\,V = N$, then the descendant (resp. ancestor) potential of a semi-simple Frobenius manifold is obtained by quantizing the cone obtained by the action of the symplectomorphism $\gamma(v) M_{v}(z)$ (resp. $M_{v,0}(z)$) on the cone ${\cal L}_N$ corresponding to the trivial theory of type $A_1^{\times N}$ where $\gamma(v)$ is a suitably chosen normalization factor.
\end{theorem}
In order to be more explicit, let us describe ${\cal L}_N$. Let $F_0^{{\rm KdV}}(\mathbf{t})$ be the genus 0 potential of the Gromov-Witten theory of a point, \textit{i.e.} the genus 0 part of the logarithm of the partition function of the quantum Airy structure of Proposition~\ref{ThmLoopZ2TQFT} with all $v_{(k,i),(l,j)}$ vanishing and $\theta(z) = z^{-2}\cdot 1_{V}$. Then, after the identification by the dilaton shift $q_{k,i} = t_{k}^i-\delta_{k,1}$, one has indeed
$$
{\cal L}_N:= \Big\{(p,q) \in T^*{\cal V}\,\,:\,\,\quad p = \sum_{i=1}^N \mathrm{d}_{q}F_0^{{\rm KdV}}({\bf t}^{i})\Big\}.
$$
The full partition function of this quantum Airy structure is the matrix Airy function of \cite{Konts}.
Because $M_{v,0}(z)$ is analytic for $|z|>1$, the quantum structure corresponding to the ancestor's Lagrangian cone is obtained by the action of the operator $\exp\big(\tfrac{\hbar}{2}\sum_{k,\ell,i,j} u_{(k,i),(\ell,j)} \partial_{(k,i)} \partial_{(\ell,j)} \big)$ on the quantum Airy structure built from ${\cal L}_N$ where
$$
{M_{v,0}^*(z_1) M_{v,0}(z_2) - {\rm Id} \over z_1+z_2} := \sum_{k,\ell\geq 0} (-1)^{k+\ell} U_{k,l} z_1^k z_2^{\ell},\qquad U_{k,\ell} e_i := \sum_{j=1}^N u_{(k,i),(\ell,j)} e_j.
$$
This transformation preserves Airy structures and the topological recursion gives a way to compute the ancestor potential. Further, the local spectral curve is fixed by the symplectomorphism $M_{v,0}(z)$.
The action of the Givental's twisted loop group on Lagragian cones is easily seen to coincide with the action of symplectomorphisms on Airy structures defined on ${\cal W}$. This leads to the equivalence of Givental quantization procedure and quantization of the corresponding Airy structure.
\begin{corollary}
The ancestor potential of a semi-simple Frobenius manifold is obtained from the quantum Airy structure defined on the loop space ${\cal W}$ from the action of the symplectomorphism $M_{v,0}(z)$ on the trivial quantum Airy structure whose partition function is the matrix Airy function.
This quantization procedure is equivalent to Givental's quantization of the corresponding Lagrangian cone.
\end{corollary}
On the other hand, the action of the quantization of $M_{v,\infty}(z)^{-1}$ taking the ancestor potential to the descendant potential is just the multiplication of the partition function by $\exp\big( \tfrac{\hbar}{2} \sum_{k,\ell,i,j} s_{(k,i),(\ell,j)} x_{(k,i)} x_{(\ell,j)}\big)$ where
$$
{M_{v,\infty}^*(z_1) M_{v,\infty}(z_2) - {\rm Id} \over1/ z_1+1/z_2} := \sum_{k,\ell\geq 0} S_{k,\ell} z_1^{-k} z_2^{-\ell},\qquad S_{k,\ell} e_i := \sum_{j=1}^N s_{(k,i),(\ell,j)} e_j.
$$
Even though it is a simple transformation, let us remark that this does not preserve the Airy structure property, since it adds a linear term in the $x$'s. This explains why the topological recursion does not directly provide the descendant potential, except in the cases where $M_{v,\infty}$ is trivial.
\section{A topological recursion without branched covers}
\label{SWithout}
In this section, we explain the simple observation that the quantum Airy structures of Proposition~\ref{ThmLoopTQFT} can be realized by a new variant of the topological recursion of \cite{EOFg}. We take as initial data
\begin{itemize}
\item[$\bullet$] a Riemann surface $\Sigma$.
\item[$\bullet$] a meromorphic $1$-form $\omega_{0,1}$ on $\Sigma$.
\item[$\bullet$] a fundamental bidifferential of the second kind $\omega_{0,2}$ on $\Sigma$.
\item[$\bullet$] a finite subset $\mathfrak{r} \subset \Sigma$, such that $\omega_{0,1}$ has at most simple zeroes at $\mathfrak{r}$ -- this allows poles of $\omega_{0,1}$ at $\mathfrak{r}$.
\item[$\bullet$] a meromorphic $1$-form $\omega_{1,1}$ on $\Sigma$, such that, for any $r \in \mathfrak{r}$, $(x(p) - x(r))^2\frac{\omega_{1,1}(z)}{\omega_{0,1}(p)}$ is holomorphic around $p \rightarrow r$.
\end{itemize}
We define a recursion kernel
$$
\tilde{K}(p_0,p) = \frac{\int_{r}^{p} \omega_{0,2}(\cdot,p_0)}{\omega_{0,1}(p)},
$$
and for $2g - 2 + n > 0$ and $(g,n) \neq (1,1)$, we make the inductive definition
\begin{equation}
\label{TRsans} \omega_{g,n}(p_1,I) = \sum_{r \in \mathfrak{r}} \mathop{\rm Res}\limits_{p \rightarrow r} \tilde{K}(p_0,p)\bigg\{\omega_{g - 1,n + 1}(p,p,I) + \sum_{\substack{g' + g'' = g \\ J' \sqcup J'' = I}}^* \omega_{g',1+|J'|}(p,J') \omega_{g'',1 + |J''|}(p,J'')\bigg\}.
\end{equation}
We have to include $\omega_{1,1}$ in the initial data since $\omega_{0,2}(p,p)$, which would appear in \eqref{TRsans} for $(g,n) = (1,1)$, does not make sense due to the double pole at coinciding point. As in Section~\ref{comparTR}, one can prove that the $\omega_{g,n}$ decompose on a basis of $1$-forms.
\begin{definition}
For $k \geq 0$ and $r \in \mathfrak{r}$, we define for $p_0 \in \Sigma$ the meromorphic $1$-form
$$
\xi_{k,r}(p_0) := \Big(\mathop{\rm Res}\limits_{p \rightarrow r} \int_{r}^{p_0} \omega_{0,2}(\cdot,p_0)\Big)\frac{(k + 1)\mathrm{d} x(p)}{(x(p) - x(r))^{k + 1}}.
$$
We also define, for $p_0$ in a neighborhood of $r$ in $\Sigma$
$$
\xi_{k,r}^*(p_0) := \frac{(x(p) - x(r))^{k + 1}}{k + 1},\qquad \theta(p) := \frac{1}{\omega_{0,1}(p)},
$$
and if $p_0$ is in a neighborhood of $r_0 \neq r$, we define $\xi_{k,r}^*(p_0) := 0$.
\end{definition}
\begin{lemma}
\label{PPP0}For $2g - 2 + n > 0$, there exists a unique decomposition with finitely many non-zero terms
\begin{equation}
\label{Fogisans}\omega_{g,n}(p_1,\ldots,p_n) = \sum_{\substack{r_1,\ldots,r_n \in \mathfrak{r} \\ k_1,\ldots,k_n \geq 0}} W_{g,n}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots & k_n \end{smallmatrix}\right]\,\prod_{i = 1}^n \xi_{k_i,r_i}(p_i).
\end{equation} \hfill $\Box$
\end{lemma}
The assumption on $\omega_{0,1}$ guarantees that $\theta(p)$ for $p \rightarrow r$ has an expansion of the form
$$
\theta(p) = \sum_{m \geq -1} t_{m,r}z^{m}(p)\,(\mathrm{d} x(p))^{-1}.
$$
According to Proposition~\ref{ThmLoopTQFT}, we have a quantum Airy structure given by
\begin{eqnarray}
A^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & 0, \nonumber \\
B^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1}^*(p)\,\mathrm{d} \xi_{k_2,r_2}^*(p)\,\xi_{k_3,r_3}(p)\,\theta(p), \nonumber \\
C^{(k_1,r_1)}_{(k_2,r_2),(k_3,r_3)} & := & \mathop{\rm Res}\limits_{p \rightarrow r_1} \xi_{k_1}^*(p)\,\xi_{k_2,r_2}(p)\,\xi_{k_3,r_3}(p)\,\theta(p), \nonumber \\
\label{KSsf} D^{(k_1,r_1)} & := & W_{1,1}\left[\begin{smallmatrix} r_1 \\ k_1 \end{smallmatrix}\right],
\end{eqnarray}
and by comparison of KS-TR and the recursive relation for $W_{g,n}$'s ensuing from \eqref{TRsans} we obtain that
\begin{proposition}
\label{PPP1}For $2g - 2 + n > 0$, $F_{g,n}\big((k_1,r_1),\ldots,(k_n,r_n)\big)$ computed by KS topological recursion for the quantum Airy structure \eqref{KSsf}, and $W_{g,n}\left[\begin{smallmatrix} r_1 & \cdots & r_n \\ k_1 & \cdots & k_n \end{smallmatrix}\right]$, agree. \hfill $\Box$
\end{proposition}
We omit the proof of Lemma~\ref{PPP0} and Proposition~\ref{PPP1}, as it is similar to Lemma~\ref{LemmaFogi} and Proposition~\ref{FWid}, in fact simpler due to the absence of the involution. Note that the assumption made on $\omega_{1,1}$ is equivalent to the \textbf{D} relation.
\section{Dynamics on (colored) Young diagrams}\label{s:TR}
\label{SYoung}
\def\Col#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\end{pspicture}
}
\def\ColOne#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny$1$}}
\end{pspicture}
}
\def\ColP#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny $p$}}
\end{pspicture}
}
\def\Colg#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\end{pspicture}
}
\def\ColgOne#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny$1$}}
\end{pspicture}
}
\def\ColgP#1{
\begin{pspicture}(0.2,1)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny $p$}}
\end{pspicture}
}
\def\Cols#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\end{pspicture}
}
\def\ColsOne#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny$1$}}
\end{pspicture}
}
\def\ColsP#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe(0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny $p$}}
\end{pspicture}
}
\def\Colsg#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\end{pspicture}
}
\def\ColsgOne#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny$1$}}
\end{pspicture}
}
\def\ColsgP#1{
\begin{pspicture}(0.2,0.6)
\multiput(0,0)(0,0.2){#1}{\psframe[fillstyle=solid,fillcolor=lightgray](0,0)(0.2,0.2)}
\rput(0,-0.2){\makebox(0,0)[lb]{\tiny $p$}}
\end{pspicture}
}
\subsection{Setting}
In this section we show that a quantum Airy structure on $V = \mathbb{C}[[z]]$ (or $V = \mathbb{C}^{d}[[z]]$, or $z\mathbb{C}^{d}[[z^2]]$, etc.) gives a recursion on (colored) Young diagrams, which are in correspondence with the monomials that can appear in the Taylor expansion of the partition function. We first formulate abstractly the recursion on Young diagrams, and relate it to quantum Airy structure in Proposition~\ref{PYoung} below. Please see also \cite{ACNP}, where this dynamics on Young diagrams in some special cases was given.
Let $d$ be a positive integer. For a Young diagram $\lambda = (\lambda_1 \geq \cdots \geq \lambda_{\ell(\lambda)})$, we consider $\lambda_i$ as column heights. We denote ${\rm Col}(\lambda)$ the set of columns, and $|\lambda|$ be the number of boxes.
\begin{definition}
A \emph{$d$-coloring} of a Young diagram $\lambda$ is a map ${\rm Col}(\lambda) \rightarrow \{1,\ldots,d\}$. A \emph{column type} is an ordered pair $(k,\alpha) \in \mathbb{N}_{> 0} \times \{1,\ldots,d\}$, where $k$ is the height and $\alpha$ is the color. We denote ${\rm Aut}\,\lambda$ the group of permutations of columns respecting column types.
\end{definition}
We use the notation $N_{k,\alpha}(\lambda)$ for the number of columns of type $(k,\alpha)$, hence
$$
|\lambda| = \sum_{k \geq 1} \sum_{\alpha = 1}^d N_{k,\alpha}(\lambda) k,\qquad |{\rm Aut}\,\lambda| = \prod_{\substack{k \geq 1 \\ 1 \leq \alpha \leq d}} N_{k,\alpha}(\lambda)!.
$$
\begin{definition}
Let $\mathcal{Y}_{g,n}^{(d)}$ be the set of $d$-colored Young diagrams $\lambda$ such that
\begin{equation}
\label{boundsY}\ell(\lambda) = n \qquad {\rm and}\qquad |\lambda| \leq (2g - 2 + n)r,
\end{equation}
and $\tilde{\mathcal{Y}}_{g,n}^{(d)}$ the set of such $d$-colored Young diagrams together with the choice of a column type -- remembered by a label ``1".
\end{definition}
We denote $s\,:\,\tilde{\mathcal{Y}}_{g,n}^{(d)} \rightarrow \mathcal{Y}_{g,n}^{(d)}$ the map which forgets the label ``1''. We have an injective linear map
$$
\begin{array}{rcccc}
S & : & {\mathbb C}[\mathcal{Y}_{g,n}^{(d)}] & \longrightarrow & {\mathbb C}[\tilde{\mathcal{Y}}_{g,n}^{(d)}] \\
& & \lambda & \longmapsto & \sum_{\tilde{\lambda} \in s^{-1}(\lambda)} \tilde{\lambda},
\end{array}
$$
that is, each Young diagram is mapped to the sum (linear combination with unit coefficients) of the same Young diagrams differing only by placing the label ``1'' on all types of columns present in this diagram. We now define two unary operations on diagrams from $\mathcal{Y}_{g,n}^{(d)}$ which results in two linear maps
$$
\Delta_{{\rm B}}\,:\, {\mathbb C}[\mathcal{Y}_{g,n}^{(d)}] \longrightarrow {\mathbb C}[\tilde{\mathcal{Y}}_{g,n+1}^{(d)}],\qquad \Delta_{{\rm C}}\,:\,{\mathbb C}[\mathcal{Y}_{g-1,n+1}^{(d)}] \longrightarrow {\mathbb C}[\tilde{\mathcal{Y}}_{g,n}^{(d)}],
$$
and a binary operation on ordered pairs of colored Young diagrams, which results in the bilinear map
$$ \Delta_{{\rm C}}^{(2)}\,:\, {\mathbb C}[\mathcal{Y}_{g_1,n_1}^{(d)}] \otimes {\mathbb C}[\mathcal{Y}_{g_2,n_2}^{(d)}] \longrightarrow {\mathbb C}[\tilde{\mathcal{Y}}_{g_1+g_2,n_1+n_2-1}^{(d)}].$$
The data for our recursion will be either finite or semi-infinite complex tensors $B = (B^{(k_1,\alpha_1)}_{(k_2,\alpha_2),(k_3,\alpha_3)})$ and $C = (C^{(k_1,\alpha_1)}_{(k_2,\alpha_2),(k_3,\alpha_3)})$, where $(k_i,\alpha_i) \in \mathbb{N}_{> 0} \times \{1,\ldots,d\}$ characterize the possible column types. We assume in the semi-infinite case that the entries of $B$ vanish whenever $k_1 + k_2 > k_3 + r$, and the entries of $C$ vanish whenever $k_1 > k_2 + k_3 + r$. This guarantees that all sums appearing below are finite. The bound \eqref{boundsY} on the number of boxes of our Young diagrams are tailored to this property of $B$ and $C$.
\subsection{The operations}
The first unary operation $\Delta_{{\rm B}}\,:\,{\mathbb C}[\mathcal{Y}_{g,n}^{(d)}]\to {\mathbb C}[\tilde{\mathcal{Y}}_{g,n + 1}^{(d)}]$
is defined by the following rule. It is a sum over $\alpha_3 \in \{1,\ldots,d\}$ followed by a sum over all possible column types in $\lambda$ of color $\alpha_3$. The terms of this sum are obtained by replacing a column of the selected type with two new columns of colors $\alpha_1$ and $\alpha_2$
\begin{equation}
\begin{pspicture}(0.2,0)
\rput(0.1,0.4){$\Delta_{{\rm B}}\biggl(\,\,\,\,\,\,\,\,\,\,$}
\end{pspicture}
\Col4
\begin{pspicture}(0.5,0)
\rput(-.1,0.9){\makebox(0,0)[cb]{\tiny $\alpha_3$}}
\put(0,0){\psline{<->}(0.1,0)(0.1,0.8)}
\rput(0.2,0.4){\makebox(0,0)[lc]{$k_3$}}
\end{pspicture}
\begin{pspicture}(0.5,0)
\rput(0.1,0.4){\makebox(0,0)[lc]{$\biggr)$}}
\end{pspicture}
=\sum_{k_1,k_2 \geq 1\atop k_1+k_2\le k_3+r} \sum_{\alpha_1,\alpha_2 = 1}^{d} B_{(k_2,\alpha_2),(k_3,\alpha_3)}^{(k_1,\alpha_1)} N_{k_2,\alpha_2}(\lambda)
\begin{pspicture}(1,1)
\put(0,0){\psline{<->}(0.9,0)(0.9,0.4)}
\rput(0.8,0.2){\makebox(0,0)[rc]{$k_1$}}
\rput(1.1,0.5){\makebox(0,0)[cb]{\tiny $\alpha_1$}}
\end{pspicture}
\ColOne2\,\,
\Col3
\begin{pspicture}(1,1)
\put(0,0){\psline{<->}(0.1,0)(0.1,0.6)}
\rput(0.2,0.3){\makebox(0,0)[lc]{$k_2$}}
\rput(-.1,0.65){\makebox(0,0)[cb]{\tiny $\alpha_2$}}
\end{pspicture}.
\label{rule1}
\end{equation}
In this operation, we place a label ``1'' on the column type $(k_1,\alpha_1)$ in the resulting Young diagram.
For the second unary operation $\Delta_{{\rm C}}\,:\,{\mathbb C}[\mathcal{Y}_{g-1,n+1} ]\to {\mathbb C}[\tilde{\mathcal{Y}}_{g,n}]$, we proceed analogously, except now we sum over ordered pairs $((k_2,\alpha_2),(k_3,\alpha_3))$ of column types in $\lambda$. The terms are obtained by replacing the ordered pair of column of each type
\begin{equation}
\begin{pspicture}(0.2,1)
\rput(0.1,0.1){$\Delta_{{\rm C}}\biggl(\,\,\,\,\,\,\,\,\,\,\,$}
\end{pspicture}
\begin{pspicture}(0.5,1)
\put(0,0){\psline{<->}(0.4,0)(0.4,0.4)}
\rput(0.3,0.2){\makebox(0,0)[rc]{\tiny$k_2$}}
\end{pspicture}
\underbracket[1pt]{\Col2\,\,\Col3}
\begin{pspicture}(0.5,1)
\put(0,0){\psline{<->}(0.1,0)(0.1,0.6)}
\rput(-.1,0.7){\makebox(0,0)[cb]{\tiny $\alpha_3$}}
\rput(-.45,0.7){\makebox(0,0)[cb]{\tiny $\alpha_2$}}
\rput(0.2,0.3){\makebox(0,0)[lc]{\tiny$k_3$}}
\end{pspicture}
\rput(0.1,0.1){\makebox(0,0)[lc]{$\biggr)$}}
\quad =\sum_{1 \leq k_1 \leq k_2 + k_3 + r} \sum_{\alpha_1 = 1}^d \tfrac{1}{2}\,C_{(k_2,\alpha_2),(k_3,\alpha_3)}^{(k_1,\alpha_1)}
\ColOne6
\begin{pspicture}(1.4,1)
\rput(-.1,1.3){\makebox(0,0)[cb]{\tiny $\alpha_1$}}
\put(0,0){\psline{<->}(0.1,0)(0.1,1.2)}
\rput(0.2,0.6){\makebox(0,0)[lc]{$k_1$}}
\end{pspicture},
\label{PP1}
\end{equation}
with a new column of color $\alpha_1$.
The binary operation $ \Delta_{{\rm C}}^{(2)}\,: \,{\mathbb C}[\mathcal{Y}_{g_1,n_1}] \times {\mathbb C}[\mathcal{Y}_{g_2,n_2}] \rightarrow{\mathbb C}[\tilde{\mathcal{Y}}_{g_1+g_2,n_1+n_2-1}]$ is obtained in a similar way to $\Delta_{{\rm C}}$, but fusing the two Young diagrams. More precisely, $ \Delta_{{\rm C}}^{(2)}(\lambda , \lambda')$ is the sum over all column types $(k_2,\alpha_2)$ in $\lambda$ and column types $(k_3,\alpha_3)$ in $\lambda'$, of the following contribution. We erase one column of the selected type in $\lambda$ and $\lambda'$, fuse the two Young diagrams, and insert a column of type $(k_1,\alpha_1)$ with a label ``1'' and a weight
$$
\tfrac{1}{2}C^{(k_1,\alpha_1)}_{(k_2,\alpha_2),(k_3,\alpha_3)}.
$$
These terms are then summed over $(k_1,\alpha_1) \in \mathbb{N}_{> 0} \times \{1,\ldots,d\}$ such that $k_1 \leq k_2 + k_3 + r$ to give $\Delta_{{\rm C}}^{(2)}(\lambda , \lambda')$.
\subsection{Evaluation and relation to quantum Airy structures}
Let $\mathcal{S}^{(d)}$ be the $\mathbb{C}$-algebra of symmetric functions in infinite number of variables $x_{\alpha,i}$, $i \in \mathbb{N}_{\geq 0} $ and $\alpha \in \{1,\ldots,d\}$. The power sums $p_{k,\alpha} := \sum_{i \geq 0} x_{\alpha,i}^{k}$ give a linear basis for $\mathcal{S}^{(d)}$. If $\lambda$ is a Young diagram with a $d$-coloring we denote
$$
P_{\lambda} := \prod_{\substack{k \geq 1 \\ 1 \leq \alpha \leq d}} p_{k,\alpha}^{N_{k,\alpha}(\lambda)}\,.
$$
We define the linear evaluation map
$$
\begin{array}{rcccc} {\rm ev} & \,:\, & \mathbb{C}[\mathcal{Y}_{g,n}^{(d)}] & \longrightarrow & \mathcal{S}^{(d)} \\ && \lambda & \longmapsto & |{\rm Aut}\,\lambda|^{-1}\,P_{\lambda} \end{array}\,.
$$
The map ${\rm ev}$ is obviously injective. Let $\underline{k_1,\alpha_1|\,\cdots\,|k_{\ell},\alpha_{\ell}}$ be the Young diagram with columns of height $k_i$ and color $\alpha_i$. We remark that $\mathcal{S}^{(d)}$ is isomorphic -- \textit{via} Taylor expansions -- to the algebra of polynomial functions on
\begin{equation}
\label{Vyoung} V:= z\mathbb{C}^{d}[[z]] \cong \bigoplus_{\substack{k \geq 1 \\ 1 \leq \alpha \leq d}} \mathbb{C}.\underline{k,\alpha}.
\end{equation}
\begin{proposition}
\label{PYoung} Assume that $(A,B,C,D)$ defines a quantum Airy structure on $V$ given by \eqref{Vyoung}, and assume that $A^{(k_1,\alpha_1)}_{(k_2,\alpha_2),(k_3,\alpha_3)} = 0$ whenever $k_1 + k_2 + k_3 > r$, and $D^{(k,\alpha)}$ vanishes whenever $k > r$. Then there exists unique $\Omega_{g,n} \in {\mathbb C}[Y_{g,n}]$ indexed by $g \geq 0$ and $n \geq 1$ satisfying $2g-2+n>0$, such that
\begin{eqnarray}
\label{imfsg}\Omega_{0,3} & = & \sum_{(k_i,\alpha_i)_{i = 1}^3} A^{(k_1,\alpha_1)}_{(k_2,\alpha_2),(k_3,\alpha_3)}\,\,\underline{k_1,\alpha_1|k_2,\alpha_2|k_3,\alpha_3}\,,\label{A3} \\
\Omega_{1,1} & = & \sum_{k,\alpha} D^{(k,\alpha)}\,\,\underline{k,\alpha}\,, \label{EPS}
\end{eqnarray}
and for $2g + n \geq 0$
\begin{equation}
\label{imfsg2} S(\Omega_{g,n}) = \Delta_{{\rm B}}(\Omega_{g,n-1}) + \Delta_{{\rm C}}(\Omega_{g-1,n+1}) + \sum_{\substack{g_1+g_2 = g\\ n_1+n_2 = n}} \Delta_{{\rm C}}^{(2)}(\Omega_{g_1,n_1} ,\Omega_{g_2,n_2}).
\end{equation}
Moreover the coefficients of the partition function of the quantum Airy structure are $F_{g,n} = {\rm ev}(\Omega_{g,n})$.
Vice versa, if the dynamics of Young diagrams is governed by relations \eqref{rule1}, \eqref{PP1} endowed with the initial conditions
\eqref{A3} and \eqref{EPS} and we require the result of this action
to have the form $S(\Omega_{g,n})$ for some $\Omega_{g,n}\in {\mathbb C}[Y_{g,n}]$ for all $g$ and $n$, \textit{i.e.}, this result must belong to the image of the mapping $S$ for all $g$ and $n$, then the partition function of these correlation functions exists and is annihilated by the $L_i$ given by \eqref{Lform}.
\end{proposition}
Let us comment on this formalism. The operations $\Delta_{{\rm B}},\Delta_{{\rm C}}$ and $\Delta_{{\rm C}}^{(2)}$ introduce some asymmetry in the treatment of the column types, tracked by the label ``1''. The linear map $S$ discards this label by summing over all underlying colored Young diagrams. For given $(A,B,C,D)$, we would like to define a dynamic on (colored) Young diagrams by the formulae \eqref{imfsg}-\eqref{imfsg2} -- note that $B$ and $C$ enter in the definition of $\Delta_{{\rm B}}$, $\Delta_{{\rm C}}$ and $\Delta_{{\rm C}}^{(2)}$. However, at each step the right-hand side of \eqref{imfsg2} is an expression in terms of Young diagrams with a label ``1''. As $S$ is an injection, there is at most one expression in colored Young diagrams $\Omega_{g,n}$ satisfying \eqref{imfsg2}. Such a $\Omega_{g,n}$ does exists if and only if the right-hand side of \eqref{imfsg2} produces a function on labeled colored Young diagrams which lies in the image of the linear map $S$ -- \textit{i.e.} it gives a symmetric function on $V$ when evaluated. This is true only if the quadruple $(A,B,C,D)$ satisfies some conditions. The first part of Proposition~\ref{Explicitsym} shows that a sufficient condition for the right-hand side to be in the image of $S$ is that $(A,B,C,D)$ defines a quantum Airy structure. In this case, \eqref{imfsg}-\eqref{imfsg2} just mimicks, at the level of functions on Young diagrams, the recursive computation of the partition function of the quantum Airy structure. The proof is straightforward and thus omitted.
Let us prove the inverse statement of Proposition \ref{PYoung}.
\noindent {\bf Proof}.
For simplicity, we replace the multi-index $(k,\alpha)$ merely by $k$. The consideration below is general and does not depend on
details of the model. It is also insensitive to whether we are in a finite or infinite-dimensional situation.
We first identify the coefficients of $F_{g,n}$ with the Taylor coefficients at $\bs{\xi} = 0$ of a function $S_{g}(\bs{\xi})$
$$
F_{g,n}(k_1,\dots,k_n)=\left.\frac{\partial^n S_{g}}{\partial \xi_{k_1}\partial \xi_{k_2}\cdots \partial \xi_{k_n}}\right|_{\xi_i=0}.
$$
It is convenient to interpret $\Omega_{g,n}$ as symmetric differentials
$$
\Omega_{g,n}:=\sum_{\{k\}}F_{g,n}(k_1,\dots,k_n)\,\mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n}.
$$
We introduce an auxiliary object
$$
\overline{\Omega}:=\sum_{2g - 2 + n >0}\hbar^{2g - 2 + n} \sum_{\{k_i\}_{i=1}^n} \frac{\partial^n S_g}{\partial \xi_{k_1}\cdots \partial \xi_{k_n}}
\mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n},
$$
where \emph{we do not impose the constraint} $\xi_i=0$. We segregate the term proportional to $\mathrm{d}\xi_s \mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n}$ without a priori symmetrization with respect to the index $s$. The condition that the right-hand side is actually fully symmetric with respect to permutations of all indices including $s$ implies that it must be of the form
$$
\mathrm{d}\overline{\Omega}=\sum_{s} \sum_{2g - 2 + n >0}\hbar^{2g - 2 + n} \sum_{\{k_i\}_{i=1}^n} \frac{\partial^{n+1} S_{g}}{\partial \xi_s\partial \xi_{k_1}\cdots \partial \xi_{k_n}}\mathrm{d}\xi_s \mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n},
$$
which is fully symmetric by construction.
In the right-hand side we have several terms. Let us segregate the coefficients of $\mathrm{d}\xi_s \mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n}$. Our strategy is to push the whole collection of partial derivatives $\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}$ outside the action of the other operators. At the end of the calculations we set all $\xi_i=0$, thus obtaining the original TR relations.
\begin{itemize}
\item [(1)] In the term corresponding to $C^s_{q,p}$ we remove two differentials $\mathrm{d}\xi_{q}$ and $\mathrm{d}\xi_{p}$ and replace them by $\mathrm{d}\xi_s$. The corresponding term proportional to $\mathrm{d}\xi_s \mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n}$ reads
\begin{align*}
&C^s_{q,p}\sum_{{g_1+g_2=g\atop I\cup J=\{1,\dots,n\}}}\frac{\partial}{\partial \xi_{q}}\frac{\partial^{|I|}S_{g_1}}{\partial \xi_{k_{\alpha_1}}\cdots \partial \xi_{k_{\alpha_{|I|}}}} \frac{\partial}{\partial \xi_{p}}\frac{\partial^{|J|}S_{g_2}}{\partial \xi_{k_{\beta_1}}\cdots \partial \xi_{k_{\beta_{|J|}}}} +
\frac{\partial^2}{\partial \xi_{q}\partial \xi_{p}}\frac{\partial^{n}S_{g}}{\partial \xi_{k_1}\cdots \partial \xi_{k_n}}
\cr
&
=\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}\left[ C^s_{q,p}\sum_{g_1+g_2=g} \frac{\partial S_{g_1}}{\partial \xi_{q}}
\frac{\partial S_{g_2}}{\partial \xi_{p}} + C^s_{q,p} \frac{\partial^2 S_{g-1}}{\partial \xi_{q}\partial \xi_{p}} \right].
\end{align*}
\item[(2)] In the term corresponding to $B^s_{q,p}$ we erase $\mathrm{d}\xi_{p}$ and add $\mathrm{d}\xi_s$ and $\mathrm{d}\xi_{q}$ and multiply the result by $N_{q}$---the number of times the index $q$ appears in the set $\{k_1,\dots, k_n\}$. For the action of this operation not to vanish, the index $q$ has to be found at least once in the set
$k_1,\dots,k_n$, say, $q=k_a$ and the corresponding coefficient is $\tfrac{\partial^{n-1}S_{g}}{\partial \xi_{k_1}\cdots \widehat{\partial \xi_{k_a}}\cdots \partial \xi_{k_n}}$. Here, the hat denotes the omission of the corresponding term. In order to collect the set of partial derivatives with respect to all $\xi_{k_i}$, $i=1,\ldots, n$ we use the following trick. We write
\begin{equation}
N_{k_a}\frac{\partial^{n}S_{g}}{\partial \xi_{p} \partial \xi_{k_1}\cdots \widehat{\partial \xi_{k_a}}\cdots \partial \xi_{k_n}}
=\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}\left[\xi_{k_a}\frac {\partial S_{g}}{\partial \xi_{p}}\right] + N_{k_a} \xi_{k_a}
\frac{\partial^{n+1}S_{g}}{\partial \xi_{p} \partial \xi_{k_1}\cdots \partial \xi_{k_n}},
\label{TR-B1}
\end{equation}
where $N_{k_a}=N_q$ is exactly the proper coefficient appearing in the
TR relations (\ref{rule1}). The second term in the right-hand side of (\ref{TR-B1}) vanishes, when we impose the condition $\xi_i=0$ at the end of calculations. The remaining term reads
\begin{equation}
\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}\left[ \sum_{q,p} B^s_{q,p}\xi_{q}\frac {\partial S_{g}}{\partial \xi_{p}}\right].
\label{TR-B2}
\end{equation}
\item[(3)] The last two terms describe the lowest order terms of TR, not determined by recursion formulae, namely $F_{0,3}$ and $F_{1,1}$. For the first term, we just use that
$$
\frac{\partial^3 S_{0}}{\partial \xi_s\partial \xi_{q}\partial \xi_{p}}= A^s_{q,p} \frac{\partial^2 }{\partial \xi_{q}\partial \xi_{p}}\bigl[ \xi_{q}\xi_{p}\bigr] +O(\xi_i), \ \hbox{where}\ A^s_{q,p} =\left. \frac{\partial^3 S_{0}}{\partial \xi_s\partial \xi_{q}\partial \xi_{p}}\right|_{\xi_i = 0}
$$
and for the second we get that
$$
\frac{\partial S_{1}}{\partial \xi_s}= D_s +O(\xi_s) \ \hbox{where}\ D_s=\left. \frac{\partial S_{1}}{\partial \xi_s}\right|_{\xi_s=0}.
$$
\end{itemize}
Combining all terms together, we obtain the following statement.
The coefficient of $\hbar^{2g+n-2} \mathrm{d}\xi_s \mathrm{d}\xi_{k_1}\cdots\,\mathrm{d}\xi_{k_n}$ in the TR relations is given by the following expression
\begin{align}
&\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}\left[\hbar^{-1} \frac{\partial S_{g}}{\partial \xi_s}-\sum_{q,p} B^s_{q,p}\xi_{q}\frac{\partial S_{g}}{\partial \xi_{p}}-\sum_{q,p} C^s_{q,p}\biggl( \sum_{g_1+g_2=g}\frac{\partial S_{g_1}}{\partial \xi_{q}}
\frac{\partial S_{g_2}}{\partial \xi_{p}}+\frac{\partial^2 S_{g - 1}}{\partial \xi_{q}\partial \xi_{p}}\biggr)\right.\nonumber\\
&\qquad \left.\left.
-\delta_{2g+n,3}\Bigl(\sum_{q,p}A^s_{q,p}\xi_{q}\xi_{p} + D_s\Bigr)
\right] \right|_{\xi_i=0}=0.
\label{XXX}
\end{align}
Because relation (\ref{XXX}) holds for all sets of external partial derivatives $\frac {\partial^n} {\partial \xi_{k_1}\cdots \partial \xi_{k_n}}$ and the expression in square brackets depends neither on $n$ nor on the set $\{k_i\}_{i=1}^n$, whereas the quantities $S_{g}$ are defined to be formal power series in $\xi_i$, we conclude that this expression is identically zero for all $\xi_i$ in an open neighborhood of the set of initial values $\xi_i=0$. Thus the set of TR relations is equivalent to the set of $(A,B,C,D)$-differential constraints $L_s$ imposed on the partition function ${\mathcal Z}$. \hfill $\Box$
This proposition applies in particular to the quantum Airy structure of Proposition~\ref{ThmLoopTQFT} and \ref{ThmLoopZ2TQFT}, and \textit{a fortiori} to the one underlying the topological recursion of \cite{EOFg} according to Proposition~\ref{FWid} and its new, branched cover-free version Proposition~\ref{PPP1}.
\section{A list of problems}
\label{SConclu}
By way of conclusion, we collect a few problems opened throughout the article -- a disjoint list of problems was put forward in \cite{KSTR}.
\begin{problem}
Complete the classification of finite-dimensional quantum Airy structures based on semi-simple Lie algebras.
\end{problem}
This is likely to be a case study of the candidate modules listed in Proposition~\ref{calsssls}. One may wonder in particular whether $\mathfrak{sl}_{2}(\mathbb{C})$ is the only simple Lie algebra supporting a quantum Airy structure, and if not, if the resulting classification has a geometric meaning.
\begin{problem}
What is the field theoretic meaning of the quantum Airy structure/partition function attached to a non-commutative Frobenius manifold?
\end{problem}
In the commutative case, we have found that the partition function computes the 2d TQFT amplitudes (Lemma~\ref{TQFTpart}). It is desirable to have a similar interpretation, maybe involving open-closed 2d TQFTs \cite{Lazaroiu,Mooreopen} -- these are indeed in correspondence with pairs of commutative and non-commutative Frobenius algebras together with some morphisms between them \cite{Nataclass}. Independently, one may wonder if the amplitudes of open-closed 2d TQFTs can be computed from quantum Airy structures.
We gave in Section~\ref{SS4} the basic definitions of moduli spaces of classical and quantum Airy structures. The geometry of these spaces is worth studying. In particular, the translations define commuting flows (although maybe not independent) on them. In this direction, one may wonder if those spaces carry integrable systems.
\begin{problem}
Study the algebraic geometry of the moduli spaces of classical and quantum Airy structures.
\end{problem}
The Lagrangian cones studied by Givental and Coates are quadratic Lagrangians in a symplectic space $T^*V[[z]]$, therefore seem to be close to the setting for classical Airy structures. For those cones related to semi-simple Frobenius manifolds, one could use the action of the twisted loop group to bring such Lagrangians in the form of a standard Lagrangian cone which is known to describe an Airy structure (whose partition function is a product of the matrix Airy function of \cite{Konts}), and therefore the original cone does correspond to a classical Airy structure. This is however rather indirect.
\begin{problem}
Can one associate directly to Givental's Lagrangian cones a classical Airy structure (or some generalization, \textit{e.g.} dropping the assumption that the differential operators are at most quadratic), without the semi-simplicity assumption?
\end{problem}
\begin{problem}
Given a quantum Airy structure on $V$, when does $S_0$ defines the structure of a (germ of a) Frobenius manifolds at $0$ on $V_0$? Or, more generally, if $V$ is infinite-dimensional and contains a distinguished finite-dimensional subspace $V_0$, when does the restriction of $S_0$ to $V_0$ defines a (germ of a) Frobenius manifolds at $0$ on $V_0$?
\end{problem}
We have checked that the $S_0$ of the quantum Airy structure of Proposition~\ref{FrobABCD} on $V =$ a Frobenius algebra, does define the prepotential of a germ of a Frobenius manifold at $0$ in $V$. We also know that this is true for the quantum Airy structure corresponding (see Section~\ref{comparTR}) to TR for compact spectral curves. In this case $S_0$ is the prepotential of the Hurwitz space equipped with its usual Frobenius structure \cite{Dubrovin,TRFROBN}. It would be interesting to know whether this is still true for the loop space examples of Section~\ref{S444} especially when $\mathbb{A}$ is non semi-simple.
The setting of quantum Airy structure is very much restricted to the case of a symplectic space isomorphic to $T^*V$. Yet, some works indicate that TR should be related to quantization of moduli spaces, which are curved K\"ahler manifolds.
\begin{problem}
Can one construct interesting families of quantum Airy structure from (curved) symplectic manifolds, or from K\"ahler manifolds?
\end{problem}
\newpage
|
\section{Introduction}
Superstring theories are expected for an ultimate unified theory of
particle physics including gravitational interactions.
One of their remarkable features is that
superstring theories are defined in ten-dimensional (10D) spacetime and
predict the presence of extra dimensions of space
for theoretical consistencies.
We usually consider that
the extra six-dimensional (6D) space is compactified
in order to describe our universe.
In such string compactifications, one of challenging tasks is
to realize a chiral spectrum
in their four-dimensional (4D) effective theories,
because they must be consistent with the standard model (SM) or
some extensions such as
minimal supersymmetric standard model (MSSM).
For the purpose,
D-brane models are attractive because they can lead to
various gauge groups with generations of chiral fermions
\cite{Bachas:1995ik,Berkooz:1996km,Blumenhagen:2000wh,
Aldazabal:2000dg,Angelantonj:2000hi}, and
several D-brane models were proposed
realizing suitable 4D chiral spectra as zero-modes of
open strings on intersecting D-branes~\cite{Ibanez:2001nd,Cremades:2002qm,
Honecker:2004kb, Cvetic:2001nr}.
For the last decade, similar model building was
actively attempted in their T-dual picture, that is,
in the framework of IIB strings with magnetized D-branes,
and it was found that viable three-generation models can be obtained
\cite{Cremades:2004wa,Abe:2008sx,Abe:2015yva}.
In particular, in a concrete model proposed in Ref.~\cite{Abe:2012fj},
a semi-realistic flavor structure of the quarks and the leptons
including their hierarchical masses and mixing angles was obtained,
and furthermore, a spectrum of the supersymmetric particles and
the Higgs bosons was calculated to verify its consistency
with experimental results.
Another one of the key issues in the string compactifications is
stabilization of moduli fields which are massless scalar modes
originating from extra components of the higher-dimensional gravitational
fields and n-form fields.
Moduli stabilization is necessary to stabilize the extra compact space,
and that is also significant in particle and
cosmological phenomenologies.
In these decades, several moduli stabilization mechanisms are proposed
in the framework of superstring theories.
We will discuss the moduli stabilization, concentrating
on type IIB compactifications in this paper
to associate them with magnetized D-brane models
(Moduli stabilizations with the magnetic fluxes were discussed
in Refs.~\cite{Antoniadis:2004pp,Antoniadis:2005nu,Antoniadis:2006eu}).
We find three types of dynamical variables to be stabilized,
dilaton field, complex structure moduli and K\"ahler moduli fields.
Basically, in IIB string theories,
we can introduce nontrivial fluxes for 3-form field strengths
to stabilize the dilaton and complex structure moduli fields~\cite{Gukov:1999ya,Giddings:2001yu}.
In the presence of the 3-form fluxes turned on, however,
the potential for the K\"ahler moduli keeps flat at the tree level,
and there remain some flat directions
even when $\alpha'$-corrections and string 1-loop corrections
are taken into account.
We usually expect that
those flat directions of K\"ahler moduli fields are
stabilized by nonperturbative effects somehow.
In D-brane models, one of computable nonperturbative effects is
D-brane instantons~\cite{Blumenhagen:2006xt,
Ibanez:2006da, Ibanez:2007rs, Cvetic:2007ku,Blumenhagen:2009qh},
which we call Euclidean-branes (E-branes) in the present paper.
That is D-branes localized at a point on 4D Minkowski spacetime
but has a nonzero volume on the extra compact space.
Thus, they are possible
to yield a superpotential for the K\"ahler moduli and the dilaton field.
Besides that, gaugino condensations of hidden D-branes
are also computable nonperturbative effects
to yield the superpotential
of the moduli fields, but we will focus on the former one in this paper.
In most of previous works\footnote{
There are several studies of moduli stabilizations in D-brane models, see
Refs.~\cite{Blumenhagen:2005tn,Blumenhagen:2007sm,
Cicoli:2011qg,Cicoli:2012bi}},
D-brane model building for the visible sector and
the moduli stabilization is discussed independently from each other.
Such a scenario can be justified under the situation that
the visible sector is irrelevant to the sector to stabilize moduli.
For example, if the SM sector is localized at a certain point on the 6D compact space
and the dynamics to stabilize moduli originates from the sector on cycles far away from
the SM-localized point, those would be independent.
However, if the SM sector and the moduli-stabilizing sector occupy
at a similar place in the 6D compact space, they would affect each other.
Indeed, it is not trivial that
the instanton effect yields a superpotential suitable for
the moduli stabilization such as $W \sim A e^{-aT}$, where $T$ is the modulus,
and that in practice depends on configurations of D-branes
for the visible sector.
This is due to the fact that
one needs to integrate over the instanton zero-modes
to obtain nonperturbative superpotentials.
We can realize the superpotential successfully when
there is only a single E-brane wrapping $O(1)$-cycles without D-branes.
On the other hand,
in association with D-branes,
there appear open string zero-modes between the E-branes and the D-branes.
When they can not be soaked up by fermionic integration, the nonperturbative superpotential vanishes.
Furthermore, even if zero-modes are successfully soaked up, the superpotential including matter fields can be induced
as $W \sim (\Phi_1 \Phi_2 \cdots ) e^{-aT}$, but not the pure moduli term $W \sim A e^{-aT}$.
Such moduli-dependent terms with matter fields would be important to realize the right-handed Majorana neutrino masses and
$\mu$-terms of the Higgs fields in MSSM
\cite{Blumenhagen:2006xt,Ibanez:2006da,Ibanez:2007rs,Cvetic:2007ku,Blumenhagen:2009qh,Kobayashi:2015siy,Kobayashi:2016ovu}.
However, such moduli-dependent terms with matter fields are not suitable for moduli stabilizations.
We are thus required to study distributions of the zero-modes
for each brane configuration and confirm that
no harmful fermionic zero-modes remain
to incorporate the moduli stabilizations with the D-brane models.
In this paper, we study
the moduli stabilization due to the E-branes
in association with concrete magnetized
D-brane models for the visible sector
in type IIB orientifolds.
We assume the 3-form fluxes to stabilize
the dilaton and the complex structure
moduli fields preserving supersymmetry (SUSY), which allow us to
concentrate on the K\"ahler moduli stabilization\footnote{
Strictly speaking, we assume that
the 3-form fluxes do not change the toroidal geometry so much,
and blow-up moduli fields are set to zero.}.
In those models,
we will also turn on the ``magnetic'' fluxes for
worldvolume gauge field strength of the D-branes
in order to realize the flavor structure of the SM.
These magnetic fluxes classically produce
moduli depending Fayet-Iliopoulos (FI) terms.
We will find supersymmetric vacua with a certain ratio of
the VEVs of the moduli fields,
that means the D-term potential can stabilize the K\"ahler
moduli fields except for one flat direction.
In order to stabilize the flat direction,
we introduce E-branes and investigate the zero-mode structure
in the D-brane models .
This paper is organized as follows.
In section \ref{sec:d9}, we first review the
magnetized $T^6/Z_2\times Z_2'$ orbifolds in 10D SYM theories, which correspond to
the low-energy effective field theory of D9-brane systems,
which explains an essence of magnetized orbifold models.
Consequently, we propose several concrete models based on
the Pati-Salam gauge group.
Two types of E-branes are possible to give
stable brane configurations in association with D9-branes.
In the rest of the section, we study both the instanton effects
to find several brane configurations with which
the instanton effects work successfully and
the induced superpotential stabilizes the moduli field.
In section \ref{sec:D7-brane},
we perform a similar analysis with D7-brane models of the visible sector
where the Pati-Salam gauge group is broken by the magnetic fluxes
to realize a more realistic spectrum.
Section~\ref{sec:con} is devoted to
conclusions and discussions.
In Appendix A, we discuss the zero-mode structure in T-dual picture.
\section{D9-brane models}
\label{sec:d9}
We study mixed configurations of magnetized D-branes
and E-branes to
construct models with all the moduli fields stabilized.
In this section, we focus on
Pati-Salam models based on a stack of eight D9-branes as
the SM sector. These are the simplest but semi-realistic
magnetized orbifold models. First we briefly review the 10D SYM theories
compactified on magnetized orbifold which are low-energy effective field theories of magnetized D9-branes.
In the theories,
we can find several semi-realistic models based
on the Pati-Salam gauge group.
Finally, we will investigate E-brane's effects in the D9-brane systems,
which generate nonperturbative superpotential and stabilize
the moduli fields.
Note that any configuration of E-branes can appear and we have to
take into account all the possible E-branes.
Some of them have no effects in low-energy effective field theory, but
a certain E-brane can have nonperturbative moduli terms such as $W \sim Ae^{-aT}$.
We are interested in such E-brane effects.
\subsection{Review of magnetized orbifolds in 10D SYM theories}
We give an overview on magnetized orbifold in 10D SYM theories.
In this paper, we consider three 2-tori, $T^2\times T^2\times T^2$,
as an extra compact space, denoting their coordinates by
$z_i$ and $\bar{z}_i$ ($i=1,2,3$).
The 10D SYM theories can be described in the formulation of
4D $\mathcal N=1$ superspace,
focusing on a 4D $\mathcal N=1$ SUSY out of full $\mathcal N=4$ SUSY of
the 10D SYM theories~\cite{Marcus:1983wb}.
This was developed in compactifications of $T^2\times T^2\times T^2$
with magnetic fluxes in Ref~\cite{Abe:2012ya}.
10D SYM theories consist of 10D vector and Majorana-Weyl spinor
fields, which are decomposed into 4D vector, complex scalar and Weyl
spinor fields. These 4D fields form 4D $\mathcal N=1$ supermultiplets.
As a result, field contents of the theories are expressed by
a vector superfield $V$ and three chiral superfields $\Phi_i$.
Note that they are in adjoint representations of
gauge symmetry of the SYM theories.
In the following, we consider $U(N)$ SYM theories as
effective field theories of one stack of $N$ D9-branes.
We introduce Abelian magnetic fluxes in the $U(N)$ theories,
which are parametrized by $N\times N$ diagonal matrices as
\begin{equation*}
M^{(i)}={\rm diag}\,(m_1^{(i)},m_2^{(i)},\ldots,m_N^{(i)}),
\end{equation*}
where $i$ runs over $1,2,3$ corresponding to three $T^2$.
When $m_n^{(i)}$ takes nondegenerate values $U(N)$ gauge group is
broken down.
For example, suppose the simplest case as follows,
\begin{equation}
M^{(i)}={\rm diag}\,(m_a^{(i)},\ldots,m_a^{(i)},m_b^{(i)},
\ldots,m_b^{(i)}), \label{eq:simp}
\end{equation}
where $m_a^{(i)}\neq m_b^{(i)}$.
Then, these magnetic fluxes break the gauge group as
$U(N)\rightarrow U(N_a)\times U(N_b)$.
In this gauge symmetry breaking, we express the superfields as
\begin{equation}
\Phi_i\rightarrow
\begin{pmatrix}
\Phi_i^{aa}&\Phi_i^{ab}\\
\Phi_i^{ba}&\Phi_i^{bb}
\end{pmatrix}\label{eq:decomp},
\end{equation}
where diagonal and off-diagonal entries are in adjoint and bifundamental
representations of the unbroken gauge group $U(N_a)\times U(N_b)$, respectively.
On this magnetized background, zero-mode equations for $\Phi_j^{ab}$
on the $i$-th $T^2$ are given by
\begin{eqnarray}
\left[\bar\partial_{\bar i}+\frac{\pi}{2\im\tau_i}
(m_a^{(i)}-m_b^{(i)})z_i\right]\Phi_j^{ab}&=&0\qquad({\rm for}~~i=j),
\label{eq:zeroii}\\
\left[\partial_{i}-\frac{\pi}{2\im\tau_i}
(m_a^{(i)}-m_b^{(i)})\bar z_{\bar i}\right]\Phi_j^{ab}&=&0
\qquad({\rm for}~~i\neq j), \label{eq:zeroij}
\end{eqnarray}
where $\tau_i$ is a complex structure of the $i$-th $T^2$.
For $i=j$, that has $m_a^{(i)}-m_b^{(i)}$ degenerate zero-modes
when $m_a^{(i)}-m_b^{(i)}$ is positive,
while its conjugate one $\Phi_j^{ba}$ has no zero-modes because of
$m_b^{(i)}-m_a^{(i)}<0$.
Thus, the magnetic fluxes produce
generations of chiral fermions in 4D effective theories.
This is almost the same for $i\neq j$, except for
a relative sign in Eq.~(\ref{eq:zeroij}), and
$|m_a^{(i)}-m_b^{(i)}|$ degenerate zero-modes are produced for
$\Phi_j^{ab}$ when $m_a^{(i)}-m_b^{(i)}$ is negative.
Next we study $ Z_2$ orbifolding in this magnetized SYM theories.
Let us consider $ Z_2$ orbifolding which acts on the first and
the second $T^2$, that is,
\begin{equation*}
(z_1,z_2,z_3)\rightarrow (-z_1,-z_2,z_3).
\end{equation*}
On this orbifold, the superfields have to transform as
\begin{eqnarray}
V(z_1,z_2,z_3) &\rightarrow& +PV(-z_1, -z_2, z_3)P^{-1},\nonumber\\
\Phi_1(z_1,z_2,z_3) &\rightarrow& - P\Phi_1(-z_1, -z_2, z_3)P^{-1},
\nonumber\\
\Phi_2(z_1,z_2,z_3) &\rightarrow& -P\Phi_2(-z_1, -z_2, z_3) P^{-1},
\nonumber\\
\Phi_3(z_1,z_2,z_3) &\rightarrow& + P\Phi_3(-z_1, -z_2, z_3)P^{-1},
\label{eq:Z_2transform}
\end{eqnarray}
where projection operator $P$ is an $N\times N$
matrix satisfying $P^2=\bm 1$.
In accordance with these transformation laws, each entry of
Eq.~(\ref{eq:decomp}) is assigned into either $ Z_2$ even or odd mode.
This $ Z_2$ projection reduces the number of
the degenerate zero-modes induced by the magnetic fluxes,
as shown in Table~\ref{tb:numzero}~\cite{Abe:2008fi}.
We can also introduce discrete Wilson lines,
and the number of zero-modes depends on values of discrete Wilson lines \cite{Abe:2013bca}.
Here, for simplicity, we restrict ourselves to models with vanishing Wilson lines.
\begin{table}[t]
\center
\begin{tabular}{ccccccccc}
$|M|$&$0$& $1$ &$2$ &$3$ &$4$ &$5$ &$2n$&$2n+1$ \\\hline
Even& $1$ &$1$ &$2$ &$2$ &$3$ &$3$ &$n+1$&$n+1$\\
Odd& $0$ &$0$ &$0$ &$1$ &$1$ &$2$ &$n-1$&$n$\\
\end{tabular}
\caption{The number of active zero-modes on
the magnetized orbifold is shown, where $M$ represents an
effective magnetic flux (That corresponds to $m_a^{(i)}-m_b^{(i)}$ in
Eqs.~(\ref{eq:zeroii}) and (\ref{eq:zeroij}).).}
\label{tb:numzero}
\end{table}
It is most important that the Abelian magnetic fluxes
generically induce the FI-term
for trivial $U(1)$ parts of
unbroken gauge subgroups. For instance,
in the case of Eq.~(\ref{eq:simp}),
there appear the FI-terms with the following
parameters in diagonal parts $U(1)_a \times U(1)_b$ of
$U(N_a)$ and $U(N_b)$,
\begin{eqnarray*}
\xi_a&=&\frac{1}{\mathcal{A}^{(1)}}m_a^{(1)}+
\frac{1}{\mathcal{A}^{(2)}}m_a^{(2)}
+\frac{1}{\mathcal{A}^{(3)}}m_a^{(3)},
\label{eq:fia}\\
\xi_b&=&\frac{1}{\mathcal{A}^{(1)}}m_b^{(1)}+
\frac{1}{\mathcal{A}^{(2)}}m_b^{(2)}
+\frac{1}{\mathcal{A}^{(3)}}m_b^{(3)},
\label{eq:fib}
\end{eqnarray*}
where $\mathcal{A}^{(i)}$ is the area of the $i$-th $T^2$.
When setting $\mathcal{A}^{(i)}$ for $\xi_a$ and $\xi_b$ to vanish,
we can find a supersymmetric vacuum with unbroken $U(N_a)$ and $U(N_b)$
gauge symmetries\footnote{
Magnetized supersymmetric
vacua with broken $U(N_a)$ and $U(N_b)$ can also exist
when charged fields develop their nonvanishing VEV in D-flat directions.
This was discussed in Ref.~\cite{Abe:2016jsb}}.
This means that some of the K\"ahler moduli fields
are stabilized by the D-term potential at the supersymmetric vacuum.
In the present case, only the ratios of $\mathcal{A}^{(i)}$
are completely determined
unless $m_a^{(1)}=m_a^{(2)}=m_a^{(3)}=0$ and/or
$m_b^{(1)}=m_b^{(2)}=m_b^{(3)}=0$, and thus, only
a linear combination of the three K\"ahler moduli remains massless.
There exists one flat direction even when we consider more complicated
configurations of the magnetic fluxes to get three or more
unbroken gauge subgroups.
The aim of this paper is to stabilize
this remaining massless moduli field by nonperturbative superpotential
originating from E-branes.
\subsection{Pati-Salam Models based on D9-branes}
\label{susec:PSmodel}
We construct Pati-Salam models based on a stack of eight D9-branes, whose
low-energy effective field theory is
10D $U(8)$ SYM theory.
In the rest of this paper, we consider
$ Z_2\times Z'_2$ orbifolding
to eliminate harmful zero-modes, which acts as
\begin{eqnarray}
Z_2: (z_1,z_2,z_3)&\rightarrow& (-z_1,-z_2,z_3),\nonumber\\
Z'_2: (z_1,z_2,z_3)&\rightarrow& (z_1,-z_2,-z_3). \label{eq:z2z2}
\end{eqnarray}
Under these $ Z_2$ and $ Z'_2$ symmetries,
the superfields transform properly (see, Eq.~(\ref{eq:Z_2transform}))
with projection operators $P$ and $P'$, respectively.
For later convenience
we define the following matrix
\begin{equation}
P_{\alpha\beta\gamma}=\begin{pmatrix}
\alpha\times {\bf 1}_4 & 0 &0\\
0 & \beta\times {\bf 1}_2 &0\\
0 & 0 & \gamma\times {\bf 1}_2
\end{pmatrix}, \label{eq:pabc}
\end{equation}
where $\alpha,\beta$ and $\gamma$ take $+1$ or $-1$ and
${\bf 1}_n$ denotes $(n \times n)$ unit matrix.
Orbifolding with projection operator of this form must
respect the Pati-Salam gauge group.
In the $U(8)$ SYM theories, magnetic fluxes are represented by
three $8\times 8$ matrices.
It is convenient to parameterize them as,
\begin{eqnarray*}
M^{(1)}&=&{\rm diag}\,(0,0,0,0,X,X,-Y,-Y)+a\times {\bm 1}_8,\\
M^{(2)}&=&{\rm diag}\,(0,0,0,0,-1,-1,0,0)+b\times {\bm 1}_8,\\
M^{(3)}&=&{\rm diag}\,(0,0,0,0,0,0,1,1)+c\times {\bm 1}_8,
\end{eqnarray*}
where $a,b,c\in\mathbb{Z}$ and $X,Y\in\mathbb{N}$.
Note that, the 4D effective theories are independent of $a,~b$ and $c$
within the D9-brane sector
except for the FI-parameters.
They will play a significant role in association with E-branes.
These magnetic fluxes break the $U(8)$ gauge group down to
the Pati-Salam gauge group, $U(4)_C\times U(2)_L\times U(2)_R$ up to $U(1)$ factors,
and produce the FI-terms for
diagonal parts of them as
\begin{eqnarray*}
\xi_C&=&\frac{1}{\mathcal{A}^{(1)}}a +
\frac{1}{\mathcal{A}^{(2)}}b
+\frac{1}{\mathcal{A}^{(3)}}c,
\label{eq:fiC}\\
\xi_L&=&\frac{1}{\mathcal{A}^{(1)}}(a+X) +
\frac{1}{\mathcal{A}^{(2)}}(b-1)
+\frac{1}{\mathcal{A}^{(3)}}c,
\label{eq:fiL}\\
\xi_R&=&\frac{1}{\mathcal{A}^{(1)}}(a-Y) +
\frac{1}{\mathcal{A}^{(2)}}b
+\frac{1}{\mathcal{A}^{(3)}}(c+1).
\label{eq:fiR}
\end{eqnarray*}
These FI-parameters vanish when
\begin{equation}
\mathcal{A}^{(1)}/\mathcal{A}^{(2)}=X,\qquad
\mathcal{A}^{(1)}/\mathcal{A}^{(3)}=Y,\qquad
a+Xb+Yc=0.
\label{eq:D_cond}
\end{equation}
At supersymmetric vacua with the Pati-Salam gauge group,
this implies that two of the three K\"ahler moduli
are stabilized by the D-term.
On this magnetized orbifold with $P'=P_{+--}$ (see, Eq.~(\ref{eq:pabc})),
there remain the following zero-modes,
\begin{equation*}
\Phi_1=\begin{pmatrix}
0 & 0 &0\\
0 & 0 &H\\
0 & 0 & 0
\end{pmatrix},\quad
\Phi_2=\begin{pmatrix}
0 & Q_{L}& 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{pmatrix},\quad
\Phi_1=\begin{pmatrix}
0 & 0 & 0\\
0 & 0 & 0\\
Q_R & 0 & 0
\end{pmatrix},
\end{equation*}
where three rows and columns correspond to
$U(4)_C$, $U(2)_L$ and $U(2)_R$.
We can find degenerate zero-modes in bifundamental representation
$(1,{\bm 2},\bar{\bm 2})$,
$({\bm 4},\bar{\bm 2},1)$ and
$(\bar{\bm 4},1,{\bm 2})$,
which can be identified with the Higgs fields $H$,
the left-handed matter fields $Q_L$ and right-handed matter fields $Q_R$, respectively.
Their degeneracy, that is, the number of generations,
is determined by $X$, $Y$ and $Z_2$ projection operator $P$.
Three-generation magnetized orbifold models based on
the Pati-Salam gauge group were systematically
studied in Ref.~\cite{Abe:2008sx}.
According to that, we summarize
all possible ans\"atze of $(X,\,Y,\,P)$ for realizing
the three generations of the quarks
and the leptons in Table \ref{tab:numH}.
\begin{table}[th]
\centering
\begin{tabular}{ccccc}
\hline
& $X$ & $Y$ & $P$ & \# of Higgs
\\
\hline
No.1 & 4 & 4 & $P_{+-+}$ & 5\\
No.2 & 5 & 5 & $P_{+-+}$ & 6\\
No.3 & 7 & 7 & $P_{++-}$ & 8\\
No.4 & 8 & 8 & $P_{++-}$ & 9\\
No.5 & 4 & 5 & $P_{+-+}$ & 5\\
No.6 & 7 & 8 & $P_{++-}$ & 8\\
No.7 & 4 & 7 & $P_{+--}$ & 5\\
No.8 & 4 & 8 & $P_{+--}$ & 5\\
No.9 & 5 & 7 & $P_{+--}$ & 5\\
No.10 & 5 & 8 & $P_{+--}$ & 6\\
\hline
\end{tabular}
\caption{They are all possible sets of $X$, $Y$ and $P$
to realize the three-generation structure of the SM.
One can exchange the values of $X$ and $Y$ in configuration 5 and 6.
In configuration 7-10, we have to replace the projection operator
by $P_{+++}$ when exchanging $X$ and $Y$. }
\label{tab:numH}
\end{table}
In these models,
a reasonable mechanism to realize hierarchical masses and
mixing angles works, which leads to
a semi-realistic spectrum without fine tunnings
for parameters~\cite{Abe:2014vza}.
Note that, there are other configurations to realize the three generations,
but they have a phenomenological difficulty in textures of Yukawa matrices
and we have omitted them here.
It is remarkable that zero-modes cannot remain in diagonal entries of the above matrices which correspond to open string moduli fields. That is, open string moduli are completely stabilized. The idea of this open string moduli stabilization would be a T-dual picture to intersecting D-branes wrapping rigid cycles \cite{Blumenhagen:2005tn}.
One may expect that
magnetized backgrounds with more complicated gauge symmetry breaking,
e.g., $U(8)\rightarrow U(3)_C \times U(1)_\ell \times U(2)_L \times U(2)_R$,
lead to a new class of three-generation models.
In that case, however, a nonvanishing FI-term inevitably appears
within the 10D $U(8)$ SYM theories~\cite{Abe:2014vza}.
We will propose such a model with all the vanishing FI parameters
on the basis of D7-brane systems in section \ref{sec:D7-brane}.
\subsection{Nonperturbative Superpotential : E1-branes}
\label{sec:e1d9}
We study E-branes in D9-brane models.
In general, there can be various E-branes generating superpotential.
Here, we focus only on E-brane configurations which contribute to the moduli stabilization.
In the presence of D9-branes,
two types of E-branes are possible to lead to a stable brane system;
E1-branes wrapping two-cycles
and E5-branes.
E-branes generically have $O(N)$ or $USp(N)$ gauge groups, and
only the $O(N)$-type instantons can generate the superpotential.
In the present setup, we can choose discrete torsions to obtain the $O(N)$-type instantons \cite{Blumenhagen:2005tn},
and we assume that the discreet torsions are tuned on suitably in this paper.
These instantons can induce superpotential of the form
\begin{equation}
W_{np}=\sum_{i} A_i e^{-a_i T_i} +A_Se^{-S},\label{eq:supapo}
\end{equation}
where $T_i$ and $S$ are K\"ahler moduli and dilaton superfields, respectively.
Coefficients $A_i$ and $A_S$ depend on complex moduli fields,
which are supposed to be stabilized by the 3-form fluxes and replaced by their VEVs.
In the present case, they are given by
\begin{equation*}
T_i=e^{-\phi} {\mathcal A}^{(i)}+i\int_{T^{2}} C_2,\qquad
S=e^{-\phi} {\mathcal A}^{(1)}{\mathcal A}^{(2)}{\mathcal A}^{(3)}
+i\int_{T^6}C_6,
\end{equation*}
where $C_2$ and $C_6$ are RR-forms and $\phi$ is the 10D dilaton field.
The SUSY condition (\ref{eq:D_cond}) stabilizes
two directions of $T_i$.
It is important that
this superpotential changes or vanishes if there exist open string zero-modes
between the D9-branes and the E-branes.
We have to study configurations of these branes
in order to eliminate such harmful zero-modes.
First we study E1-branes, which wrap one of the three $T^2$ and
are collapsed at a fixed point on the other $T^2$.
A single E1-brane has an $O(1)$ gauge symmetry and is
to generate the superpotential for the K\"ahler moduli
(the first term of Eq.~(\ref{eq:supapo}))
as long as there is no extra zero-mode.
A zero-mode configuration of E1/D9 systems is equivalent to that of a system
consisting of
D9-branes and an unfluxed D5-brane wrapping
the $i$-th $T^2$.
Such a D-brane system contains a six-dimensional $\mathcal N=1$
hypermultiplet as D5-D9 (or E1-D9) open strings.
Naming these D9-branes ``D9$_A$'',
we can represent the hypermultiplet
by using two 4D $\mathcal N=1$ superfields
as ($\Phi^{AE}_j, \Phi^{EA}_k$) ($i\neq j\neq k\neq i$) in
the superfield description (see, Ref.~\cite{Abe:2015jqa}).
Note that superscripts $AE$ and $EA$ reflect their gauge transformation laws,
and they are (anti-)fundamental representation of $U(N)$ gauge group of
the D9-branes.
They are affected by the magnetic fluxes of the D9-branes,
and thus a chiral spectrum with generation structure is
produced in this E1-D9 sector,
in the same way as D9-brane sector.
The transformation law of these chiral superfields
under the $Z_2$ and $Z_2'$ orbifolding
is given in a way similar to the D9-brane fields, e.g.,
\begin{equation*}
\Phi^{AE}_1 \rightarrow - P \Phi^{AE}_1 P_E^{-1},\qquad
\Phi^{AE}_1 \rightarrow + P' \Phi^{AE}_1 P_E^{'-1},
\end{equation*}
where we can set $P_E$ and $P_E'$ to $\pm1$.
Note that all of E1-branes with $P_E = \pm 1$ and $P_E'=\pm1$ can appear and
we have to take into account all the possible E1-branes including projections, $P_E = \pm 1$ and $P_E'=\pm1$.
However, some of them do not induce nonperturbative terms and others induce nonperturbative terms
such as (\ref{eq:supapo}) as well as nonperturbative terms with matter fields.
We are interested in E1-branes with proper orbifold parities, $P_E$ and $P_E'$, which can induce
(\ref{eq:supapo}).
When the superfield has a different subscript,
the overall signs can be changed.
Their wavefunctions can be even or odd functions
on the $i$-th $T^2$.
On the other $T^2$, however,
they cannot survive the orbifold projection when they are
assigned into odd mode,
because they are localized at a fixed point of the $T^2$ and
their wavefunctions must be given by a delta function.
We study how to find the E1-brane configurations where
all the harmful
massless modes are eliminated, taking
an E1-brane wrapping the third $T^2$ as an example.
For the purpose, it is satisfactory to
investigate a zero-mode configuration of
$\Phi^{AE}_1$ and $\Phi^{EA}_2$.
They transform under the $Z_2$ symmetry as
\begin{equation*}
\Phi^{AE}_1 \rightarrow - P \Phi^{AE}_1 P_E^{-1},\qquad
\Phi^{EA}_2 \rightarrow - P_E\Phi^{EA}_2P^{-1}.
\end{equation*}
They cannot have zero-modes
when they are assigned into $Z_2$ odd mode
on the first and the second $T^2$ as discussed above.
Thus, for $P=P_{+++}$, we can eliminate
all the components of $\Phi^{AE}_1$ and $\Phi^{EA}_2$ by
$P_E=+1$.
Even when $P\neq P_{+++}$, it is possible to eliminate them as follows.
In the Pati-Salam models, both of them have eight components, which are
classified into three parts by their gauge representations,
i.e., $U(4)_C$, $U(2)_L$ and $U(2)_R$.
A proper choice for $P_E$ can forbid the charged zero-modes
in two of the three parts.
Seen from Table~\ref{tb:numzero}, we can eliminate the remaining ones
when the absolute values of their effective magnetic fluxes are
less than three and they are assigned into $Z_2'$ odd mode on the
third $T^2$.
We can always find $P_E'$ and $c$ which realize such a situation,
satisfying Eq.~(\ref{eq:D_cond}).
Thus, it is always possible for the E1-brane
to generate the nonperturbative superpotential.
One can easily confirm that E1-branes wrapping the first or
the second $T^2$ can also induce nonperturbative terms to stabilize the moduli.
We examine the stabilization of the moduli field
minimizing its potential.
We expect to obtain the following nonperturbative
superpotential,
\begin{equation}
W=Ae^{-2\pi T_3}+W_0.\label{eq:tstabi}
\end{equation}
We assumed that nonperturbative term due to E1-brane wrapping the third $T^2$ is dominant.
Even when other terms are dominant, the following discussion is the same.
A constant term $W_0$ is also necessary for the moduli stabilization,
and we will discuss its origin later.
In toroidal compactifications of type IIB with O5/O9 planes,
the K\"ahler potential for the moduli fields
is given by
\begin{equation*}
K_0=-\log(S+\bar S)-\sum_{i=1}^3\log(T_i+\bar{T_i})
-\sum_{i=1}^3\log(U_i+\bar{U_i}).
\end{equation*}
Setting ${\rm Re\,}T_i=\tau_i$ and ${\rm Im\,}T_i=0$,
we get the F-term potential
\begin{equation*}
V_F=\frac{\pi Ae^{-4\pi\tau_3}}{\tau_1 \tau_2}\left(A+2\pi A\tau_3+W_0e^{2\pi \tau_3}\right).
\end{equation*}
Minimizing this potential, we find a supersymmetric minimum,
\begin{equation*}
\frac{W_0}A=-(1+4\pi\tau_3)e^{-2\pi \tau_3},
\end{equation*}
where $\tau_3$ is stabilized.
In this case, one sees that
a legitimate value of $\langle\tau_3\rangle$ implies
a quite small value of $W_0$, indeed,
$\langle\tau_3\rangle=1$ requires $W_0/A\sim 10^{-2}$.
The origin of such a small $W_0$ will be discussed in the next subsection.
\subsection{Nonperturbative Superpotential : E5-branes}
\label{sec:E5}
We perform a study similar to the previous subsection for E5-branes.
The number of zero-modes in D9-E5 open strings
can be counted in the same way as a mixed configuration of
the magnetized D9-branes and an additional D9-brane with no magnetic fluxes.
Although it is difficult in D9/E5 systems
to give a setup to eliminate all the harmful zero-modes systematically,
we show a reasonable setup to generate the nonperturbative superpotential
to be incorporated in a wide class of the Pati-Salam models shown
in Table~\ref{tab:numH}.
First we set $b=-1$ and $c=+1$, which implies $a=X-Y$
for the vanishing D-terms (see, Eq.~(\ref{eq:D_cond})).
That is, the magnetic fluxes in the Pati-Salam sector are given by
\begin{eqnarray*}
M^{(1)}&=&{\rm diag}\,(X-Y,X-Y,X-Y,X-Y,2X-Y,2X-Y,X-2Y,X-2Y),\\
M^{(2)}&=&{\rm diag}\,(-1,-1,-1,-1,-2,-2,-1,-1),\\
M^{(3)}&=&{\rm diag}\,(1,1,1,1,1,1,2,2).
\end{eqnarray*}
When $2X-Y\neq0$ and $X-2Y\neq0$,
an association of chirality projections due to the magnetic fluxes
and $Z_2'$ orbifold projections with $P_E'=-1$ eliminates
open string zero-modes charged under $U(2)_L$ and $U(2)_R$.
The remaining ones, which are
(anti-)fundamentals in the $U(4)_C$ gauge group,
can also be eliminated by $Z_2$ orbifolding with a suitable choice for $P_E$,
if $0<|X-Y|<3$.
Thus we can always provide configurations of D9/E5 systems
to generate
the nonperturbative superpotential for the dilaton superfield
(the second term of Eq.~(\ref{eq:supapo})),
when $X$ and $Y$ satisfy
\begin{equation*}
2X-Y\neq0,\qquad X-2Y\neq0,\qquad 0<|X-Y|<3.
\end{equation*}
That is, models 5,\,6 and 9 shown in Table~\ref{tab:numH}
are available (We can exchange the values of $X$ and $Y$ as discussed there).
In particular,
we find that some of these models can be associated with
an E5-brane and an E1-brane simultaneously
(e.g., $X=7$, $Y=5$ and $P=P_{+++}$).
In this case,
we can obtain the nonperturbative superpotential
\begin{equation}
W=A_Ee^{-2\pi T_3}+A_Se^{-S}.\label{eq:wst}
\end{equation}
We have assumed the presence of supersymmetric 3-form fluxes to
stabilize the dilaton satisfying $\langle W_{\rm 3-form}\rangle=0$.
The superfield $S$ can be replaced by its VEV, and then
the effective superpotential is equivalent to Eq.~(\ref{eq:tstabi}),
that is,
\begin{equation*}
W_0=A_Se^{-\langle S\rangle}.
\end{equation*}
From this expression, it is found that
a reasonable value of $\langle S\rangle$ induces a
sufficiently small $W_0$ which is required for the above
successful moduli stabilization.
Thus, all the moduli fields can be stabilized
in the framework of magnetized D-branes by
an interplay of the two instanton effects.
\section{D7-brane models}
\label{sec:D7-brane}
In this section, we consider another model based on D7-branes, instead of
the Pati-Salam models based on D9-branes\footnote{
The model discussed in this section was proposed in Ref.~\cite{horie}}.
\subsection{MSSM-like model}
We consider an MSSM-like model on the basis of
two stacks of four D7-branes which we denote by
D7$_A$-branes and D7$_B$-branes with a configuration
shown in Table \ref{tab:D7system}.
\begin{table}[t]
\begin{center}
\begin{tabular}{cccc}
&$T^2$ &$T^2$&$T^2$\\\hline
${\rm D7}_A$& \checkmark & $\times$ & \checkmark \\
${\rm D7}_B$& \checkmark & \checkmark & $\times$
\end{tabular}
\end{center}\vspace{-1.5em}
\caption{The configuration of two stacks of D7-branes is shown.
A symbol ``\checkmark'' means that D-branes wrap $T^2$, and
another one ``$\times$'' expresses that
D-branes are localized at a fixed point on $T^2$. }\label{tab:D7system}
\end{table}
An effective field theory of D7-branes
is derived from a 10D SYM theory,
and the superfield description of that was formulated
in Ref.~\cite{Abe:2015jqa}.
One of the three chiral superfields $\Phi_i$ contained in 10D SYM theories
turns to a position moduli field there.
In the present case of the mixed D7-brane system,
there also appears a hyper multiplet corresponding to
open string modes between the D7$_A$- and D7$_B$-branes,
which is denoted by two chiral superfields $\Phi_2^{AB}$ and $\Phi_3^{BA}$.
Thus, this system consists of the following chiral superfields,
\begin{equation*}
\Phi_1^A,\quad \tilde\Phi_2^A,\quad \Phi_3^A,\quad \Phi_1^B,
\quad \Phi_2^B,\quad \tilde\Phi_3^B,\quad \Phi_2^{AB},\quad
\Phi_3^{BA}.
\end{equation*}
The first three superfields are in the $U(4)_A$ adjoint representation,
and the next three are in the $U(4)_B$ adjoint one.
The last two are bifundamental representation of $U(4)_A\times U(4)_B$.
The tilde represents that the superfield turns to be a position moduli
of the corresponding D7-branes.
In this section, we again consider $T^2\times T^2\times T^2$ as the
extra compact space with $Z_2\times Z_2'$ orbifolding.
These $Z_2\times Z_2'$ act on the three $T^2$ in the same way as
in the previous section, and
the transformation laws of the superfields are determined by
their subscript and
four $4\times 4$ projection matrices, $P_A$, $P_B$, $P'_A$
and $P'_B$.
Note that, active D7-brane fields must be assigned into even mode on
$T^2$ where the D7-brane is localized as a point because
such a point-like localization implies a wavefunction of delta function.
In particular,
D7-D7 open strings, $\Phi_2^{AB}$ and $\Phi_3^{BA}$,
have to be assigned into even mode on the second and the third $T^2$
in order to survive the orbifold projections.
We introduce the magnetic fluxes
in this D7$_A$/D7$_B$ brane system
as follows,
\begin{equation*}
M_A^{(1)}=\begin{pmatrix}
-5 \times{\bf 1}_3 & 0 \\
0 & -4\times{\bf 1}_1
\end{pmatrix},\qquad
M_A^{(3)}=\begin{pmatrix}
5\times{\bf 1}_3 & 0 \\
0 & 4\times{\bf 1}_1
\end{pmatrix},
\label{eq:E7-1}
\end{equation*}
\begin{equation*}
M_B^{(1)}=\begin{pmatrix}
0 \times{\bf 1}_3 & 0 \\
0 & -12\times{\bf 1}_2
\end{pmatrix},\qquad
M_B^{(2)}=\begin{pmatrix}
0 \times{\bf 1}_2 & 0 \\
0 & 1\times{\bf 1}_2
\end{pmatrix}.
\label{eq:E7-2}
\end{equation*}
These magnetic fluxes break
$U(4)_A\times U(4)_B \rightarrow U(3)_C\times U(1)_l \times U(2)_L \times U(2)_R$.
One remarkable feature of this model is breaking of the $U(4)_C$
gauge symmetry of the Pati-Salam models.
This means that the quarks and the leptons
can have a distinguished difference in their flavor structure.
The flux-induced FI-terms vanish in all the
unbroken gauge subgroups when
\begin{equation}
\mathcal{A}^{(1)}/\mathcal{A}^{(2)}=12\quad{\rm and}\quad
\mathcal{A}^{(1)}/\mathcal{A}^{(3)}=1. \label{eq:d7fi}
\end{equation}
Setting the projection operators as $P_A=P_B=P_A'={\bm 1}_4$
and $P_B'=-{\bm 1}_4$, we find the following zero-mode structure,
\begin{equation*}
\Phi_1^B=\begin{pmatrix}
0 &H\\
0 & 0
\end{pmatrix},~~
\Phi_2^{AB}=\begin{pmatrix}
Q_L & 0\\
L_L & 0
\end{pmatrix},~~
\Phi_3^{BA}=\begin{pmatrix}
0 & 0 \\
Q_R & L_R
\end{pmatrix},~~
\end{equation*}
and $\Phi_1^A, \Phi_2^A, \Phi_3^A, \Phi_2^B$ and $\Phi_3^B$ have no zero-mode.
We can identify
$H$, $Q_L$, $Q_R$, $L_L$ and $L_R$ with
the Higgs fields, the left-handed quarks, the right-handed quarks,
the left-handed leptons and the right-handed leptons of the MSSM,
respectively.
All of the position and Wilson-line moduli fields are stabilized in this model as well as in the D9-models.
\subsection{Nonperturbative Superpotential : E3-branes}
In the present D7-brane system,
there are two types of E-branes keeping the whole brane system stable;
E3-branes and E(-1)-branes.
When there are no open string zero-modes interplaying
the D-branes and the E-branes,
these instantons generate the nonperturbative
superpotential (Note again that, we have assumed discrete torsions tuned on to obtain
$O(N)$-type E-branes.),
\begin{equation}
W_{{\rm np}}=\sum_{i} A_i e^{-a_i T_i} +A_S e^{-S}. \label{eq:d7super}
\end{equation}
In the IIB orientifold with $O3/O7$-planes, $T_i$ and $S$ are given by,
($i\neq j\neq k\neq i$)
\begin{equation*}
T_i=e^{-\phi} {\mathcal A^{(j)}}{\mathcal A^{(k)}}+ i\int_{T^4}C_4,
\qquad S=e^{-\phi} + i C_0,
\end{equation*}
and again, we see that two of the K\"ahler moduli fields are stabilized
by Eq.~(\ref{eq:d7fi}).
We first discuss an E3-brane wrapping two of three $T^2$
and localized at a fixed point on the other one,
which has an $O(1)$ gauge symmetry and generates the first term
superpotential of Eq.~(\ref{eq:d7super}), without extra zero-modes.
Generic E3/D7 systems are classified into two cases.
One is the case
when the E-branes and the D-branes wrap the same $T^4=T^2\times T^2$ and
localized at fixed points on the last $T^2$. In this case,
it is easy to eliminate E3-D7 open string zero-modes,
because the two stacks of the branes can be sequestered spatially when
the two stacks are localized at different fixed points.
In the other case,
we have to study the zero-mode distribution in detail for each model.
Recall again that any E3-brane including all the possible positions and orbifold parities
can appear and we have to take into account all the possibilities.
However, we are interested only in E3-brane configurations to lead moduli-dependent superpotential terms.
In the present D7-brane system,
there are two stacks of D7-branes which wrap the different four directions
of extra compact space.
An additional E-brane can be sequestered from one stack by
a localization at different fixed points,
but there exist
massless open strings between the E-brane and the other stack of D7-branes
to be eliminated by the orbifold projection.
Let us consider an E3-brane which wraps the first and the second $T^2$
and is localized at a ``vacant'' fixed point on the third $T^2$.
E3-D7$_B$ zero-modes cannot appear, but
there are E3-D7$_A$ open strings denoted by $\Phi_2^{AE}$ and $\Phi_3^{EA}$.
Fortunately, we can eliminate them easily as follows.
They transform under the $Z_2'$ symmetry ($P_A'=+{\bm 1}_4$) as
\begin{equation*}
\Phi_2^{AE}\rightarrow -\Phi_2^{AE}P_E',\qquad
\Phi_3^{EA}\rightarrow -P_E^{'-1}\Phi_3.
\end{equation*}
Thus they all can be assigned into $Z_2'$ odd mode by $P_E'=+1$
and are eliminated as we wanted,
because wavefunctions of these open strings must be
an even function on the second and the third $T^2$
as explained above.
As a result we obtain the superpotential
\begin{equation*}
W=A_3e^{-2\pi T_3}.
\end{equation*}
We will see that
this stabilizes the moduli in association
with an additional E(-1)-brane in the following subsection.
\subsection{Nonperturbative Superpotential : E(-1)-branes}
It is much easier to find an E(-1)-brane configuration generating
the superpotential for the dilaton superfield.
The E(-1)-brane is an instanton localized completely
at a point on the whole compact space.
Thus we can trivially sequester the E(-1)-brane from the D7-brane
system in order not to produce the harmful zero-modes,
unless four fixed points of a $T^2/Z_2(Z'_2)$ are
occupied by multiple stacks of D7-branes.
One can straightforwardly see that the D7-brane system admits
an E(-1)-brane and an E3-brane simultaneously
and superpotential~(\ref{eq:wst}) is generated.
Similarly to the D9-brane systems,
the second term of Eq.~(\ref{eq:wst}) produces a small constant term
in the superpotential,
and the K\"ahler moduli field is stabilized
with a moderate value of the VEV.
\section{Conclusion and Discussion}
\label{sec:con}
We have studied the nonperturbative superpotential induced
by E-branes in semi-realistic D-brane models based on
the toroidal orbifolds.
We have considered two types of D-brane models for the visible sector.
One is based on a stack of eight D9-branes, where
the magnetic fluxes and the orbifold projection yield the
Pati-Salam gauge group with the three generations of the quarks and
the leptons.
In the models, magnetic fluxes generate FI-terms, which depend on the K\"ahler moduli,
and those fix the ratio among three K\"ahler moduli.
Furthermore, we have found that
an E1-brane and an E5-brane generate the superpotential for the dilaton
and the K\"ahler moduli, respectively.
The dilaton is replaced by its VEV in the nonperturbative superpotential
because we have assumed the 3-form fluxes to stabilize that.
That gives rise to the sufficient small constant term, and as a result,
the K\"ahler moduli field is stabilized with
a moderately large value of the VEV.
The other D-brane model is derived from
the two stacks of the D7-branes.
In this model, the moduli-dependent FI-terms can fix the ratio of three K\"ahler moduli.
On top of that, we have found
that an E3-brane and an E(-1)-brane successfully generate
the superpotential and stabilize the moduli.
In our study, we have found
some constraints on the magnetic fluxes and the orbifold parities
for realizing the moduli stabilization, and
it is quite nontrivial that there exists
a successful configuration of D-branes and E-branes.
In this paper,
we have studied moduli stabilizations with only the visible sector.
The vacuum is the supersymmetric vacuum with negative energy.
We need SUSY breaking and uplifting the vacuum energy to almost zero energy.
Thus, towards more realistic models, we should also
consider a hidden sector for SUSY breaking.\footnote{See e.g. for explicit construction of the SUSY breaking
sector \cite{Abe:2016zgq}.}
In that case, we have to care about open string zero-modes
between the E-branes and the hidden D-branes,
because the moduli stabilizing superpotential vanishes if
there appears an extra zero-mode.
Besides the open string zero-modes, we expect that
there are several important interplays between
the SUSY breaking and the moduli stabilization.
It seems that such an extension to contain the SUSY breaking sector
is a very challenging task towards realistic D-brane models.
Nonperturbative effects due to E-branes are applied to other
phenomenological issues than the moduli stabilization.
Another challenging task of D-brane models is
to obtain Majorana mass terms
and supersymmetric Higgs mass term ($\mu$-term).
We are able to consider additional E-branes to generate
these mass terms~\cite{Blumenhagen:2006xt,Ibanez:2006da,Ibanez:2007rs,Cvetic:2007ku,Blumenhagen:2009qh,Kobayashi:2015siy,Kobayashi:2016ovu}.
It is also an attractive prospect to try that
in the D-brane models shown in this paper.
\section*{Acknowledgments}
H.~A. was supported by JSPS KAKENHI Grant Number JP16K05330.
T.~K. was supported in part by JSPS KAKENHI Grant Number JP26247042.
K.~S. was supported by Waseda University Grant for Special Research Projects No.~2016B-200.
S.~U. was supported by JSPS KAKENHI Grant Number JP15J02107.
|
\section*{Appendix}
In this appendix we give the detailed equations for the implementation of nonequilibrium DMFT for the two-band Hubbard model with $e_g$ orbitals. We start from the Hamiltonian Eq.~(6) and (7) of the main text, which describes electrons on two orbitals of $e_g$ symmetry at each site, and adopt the parametrization of the orbitals as described in Ref.~\cite{Nussinov2015}, where orbital $1$ and $2$ correspond to orbitals $d_{x^2-y^2}$ and $d_{3z^2-r^2}$, respectively. We introduce the orbital spinor (omitting site and spin indices for simplicity)
\begin{align}
\hat \psi=
\begin{pmatrix}
c_{1}
\\
c_{2}
\end{pmatrix}
\equiv
\begin{pmatrix}
c_{x^2-y^2}
\\
c_{3z^2-r^2}
\end{pmatrix}.
\end{align}
The Wannier orbitals transform among each other like the atomic $e_g$ orbitals, which have wave functions $\phi_{x^2-y^2}(\vec{r})=\phi(r)(x^2-y^2)$ and $\phi_{3z^2-r^2}(\vec{r})=\phi(r)(3z^2-r^2)/\sqrt{3}$, with some radial part $\phi(r)$. By forming linear combinations of $d_{x^2-y^2}$ and $d_{3z^2-r^2}$, one arrives at the other basis states $d_{y^2-z^2}$, $d_{z^2-x^2}$, $d_{3x^2-r^2}$, and $d_{3y^2-r^2}$. These linear combinations can be conveniently represented as rotations in orbital-pseudospin space, generated by ($\sigma_{1,2,3}$ denote the Pauli matrices)
\begin{align}
\hat R(\theta)=e^{i\hat \sigma_2\theta/2 }=
\begin{pmatrix}
\cos(\theta/2) & \sin(\theta/2)
\\
- \sin(\theta/2) & \cos(\theta/2)
\end{pmatrix}.
\end{align}
Rotations by $\theta= 4\pi/3$ give,
\begin{align}
\hat R
\big(
\tfrac43\pi
\big)
\hat \psi
=
\begin{pmatrix}
-\frac{1}{2} c_{x^2-y^2}+\frac{\sqrt{3}}{2}c_{3z^2-r^2}
\\
-\frac{\sqrt{3}}{2} c_{x^2-y^2}-\frac{1}{2}c_{3z^2-r^2}
\end{pmatrix}
=
\begin{pmatrix}
c_{z^2-x^2}
\\
c_{3y^2-r^2}
\end{pmatrix},
\label{rotation1}
\end{align}
where the last equality can be checked by evaluating the operators on the one-particle states, e.g.,
$\langle 0 | -\frac{1}{2} c_{x^2-y^2}+\frac{\sqrt{3}}{2} c_{3z^2-r^2} | \vec{r}\rangle$
$= -\frac{1}{2} \phi_{x^2-y^2}(\vec{r})^* +\frac{\sqrt{3}}{2} \phi_{3z^2-r^2}(\vec{r})^*$
$=\phi(r)^* [-\frac{1}{2} (x^2-y^2) +\frac{\sqrt{3}}{2} (2z^2-x^2-y^2)/\sqrt{3}]$
$=\phi_{z^2-x^2}(\vec{r})^* = \langle 0 | c_{z^2-x^2} |\vec{r}\rangle$. In summary, successive application of $\hat R(4 \pi/3)$ corresponds to a permutation of the orbitals $xyz\to zxy \to yzx \to xyz$ \cite{Nussinov2015}.
With the orbital spinor, the hopping is written as
\begin{align}
T=
-
\sum_{ j\sigma}
\sum_{a=x,y,z}
\Big(
e^{i\phi_a(t)}\,
\hat \psi_{j+\hat e_a,\sigma}^\dagger
\,\hat v_a\, \hat\psi_{j,\sigma} + h.c.
\Big),
\end{align}
where $\hat e_a$ denotes the unit vector along a bond direction $a$, $\phi_a(t)=\phi_{j+\hat e_a,j}$ is the Peierls phase along the bond ($j+\hat e_a,j$), and $\hat v_a$ is a $2\times2$ matrix. As stated in the main text, we focus on a cubic environment, in which the only nonvanishing matrix element for hopping along the $z$ axis is between $d_{3z^2-r^2}$ orbitals,
\begin{align}
\hat v_z
=
\begin{pmatrix}
0 & 0
\\
0 & t_0
\end{pmatrix}.
\end{align}
To obtain the hopping along the $x$-bond, we first rotate the lattice around the $y$ axis, which maps $x\to z $ and $z\to -x$, so that the
hopping along the $x$-bond is then written as
\begin{align}
\begin{pmatrix}
c_{z^2-y^2}^\dagger & c_{3x^2-r^2}^\dagger
\end{pmatrix}
\begin{pmatrix}
0 & 0
\\
0 & t_0
\end{pmatrix}
\begin{pmatrix}
c_{z^2-y^2}
\\
c_{3x^2-r^2}
\end{pmatrix}.
\label{vxrotated}
\end{align}
Analogous to Eq.~\eqref{rotation1} we get
\begin{align}
\hat R\big(
-\tfrac43\pi
\big)
\begin{pmatrix}
c_{x^2-y^2}
\\
c_{3z^2-r^2}
\end{pmatrix}
=
-\hat \sigma_3
\begin{pmatrix}
c_{z^2-y^2}
\\
c_{3x^2-r^2}
\end{pmatrix},
\end{align}
so that Eq.~\eqref{vxrotated} (and the corresponding equation for the $y$-bond) becomes
$\hat \psi^\dagger \hat v_{x/y}\hat \psi$,
with (the sign $\sigma_3$ cancels)
\begin{align}
\label{hoppx}
\hat v_x
&=
\hat R\big(\tfrac43\pi\big)
\hat v_z \hat R(-\tfrac43\pi\big)
=
\frac{t_0}{4}
\begin{pmatrix}
3 & -\sqrt{3}
\\
-\sqrt{3} & 1
\end{pmatrix},
\\
\label{hoppy}
\hat v_y
&=
\hat R(-\tfrac43\pi\big) \hat v_z \hat R(\tfrac43\pi\big)
=
\frac{t_0}{4}
\begin{pmatrix}
3 & \sqrt{3}
\\
\sqrt{3} & 1
\end{pmatrix}.
\end{align}
In DMFT, we compute the $2\times2$ contour-ordered Green's function
\begin{align}
\hat G(t,t')
&=
-i \langle
T_\mathcal{C}
\hat \psi_{\sigma}(t)
\hat \psi_{\sigma}^\dagger(t')
\rangle
\\
&\equiv
\begin{pmatrix}
G_{11}(t,t') & G_{12}(t,t')
\\
G_{21}(t,t') & G_{22}(t,t')
\end{pmatrix}.
\end{align}
(For an introduction into the Keldysh formalism for contour-ordered Green's functions and to nonequilibrium dynamical mean-field theory, see Ref.~\cite{aoki2014rev}.) The DMFT impurity action is given by ($H_{loc}=V$ is the local interaction)
\begin{align}
\label{action}
\mathcal{S}
=
&-i \!\int_{\mathcal{C}} dt\,H_{loc}(t)
-i \sum_\sigma
\!\int_{\mathcal{C}} \!\!dtdt'\,
\hat\psi^\dagger_\sigma(t)
\hat \Delta(t,t') \hat\psi_\sigma(t'),
\end{align}
with a self-consistently determined hybridization matrix $\hat \Delta$,
and $\hat G$ is determined by
$
\hat G
(t,t')
=-i
\text{tr}
[T_{\mathcal{C}}
e^{\mathcal{S}}
\hat \psi_\sigma(t)
\hat \psi_\sigma^\dagger(t')
]/\mathcal{Z}
$.
For the DMFT simulation, we focus on a bipartite Bethe lattice in the limit of infinite coordination number $6Z\to\infty $, in which each site has $Z$ bonds attached with rescaled hopping $-e^{\eta i\phi_a(t)}\hat v_a/\sqrt{6Z}$ for each of the $6$ combinations $\eta=\pm$, $a=x,y,z$. This can be envisioned as the simple limit of an infinitely coordinated lattice with a local cubic environment. As in the single band case, results can be expected to be qualitatively the same as for a three-dimensional cubic lattice with the bandwidth $4t_0$. Denoting by $\hat G_{A,B}$ the Green's function on the two sublattices $A,B$ of the bipartite lattice, the DMFT self-consistency is given by
\begin{align}
\hat \Delta_A(t,t')
&=
\frac{1}{6}\sum_{a=x,y,z}
\Big(
e^{i\phi_a(t)} \hat v_a\hat G_B(t,t') \hat v_a e^{-i\phi_a(t')}
\nonumber\\
&\,\,\,\,\,\,\,+
e^{-i\phi_a(t)} \hat v_a \hat G_B(t,t') \hat v_a e^{i\phi_a(t')}
\Big)
\label{bethesc}
\end{align}
(the derivation uses the cavity approach, analogous to the single-band case \cite{Georges1996}).
The strong-coupling limit of the $e_g$ Hubbard model is characterized by a entangled spin and orbital dynamics. Here we focus on the case where the spin is fully polarized. We can omit the spin index, and the local Hamiltonian reduces to
\begin{align}
\label{hlocKK}
H_{loc} = \sum_{j}\big[ (U-3J_H)n_{j, 1}n_{j, 2} -\mu(n_{j, 1}+n_{j, 2})\big].
\end{align}
The model thus becomes equivalent to a spin-less single-band model with orbital-dependent hopping, and we use the standard strong-coupling expansion \cite{eckstein2010} to solve the impurity model in the strongly interacting Mott regime. We allow for Ne\'el-type anti-ferro-orbital sublattice symmetry breaking, and look for solutions in which the Green's function on site $B$ is obtained from site $A$ by a $\pi$ rotation in orbital space,
\begin{align}
\label{NeelAFO}
\hat G_B(t,t') = \hat R(\pi) \hat G_A(t,t') \hat R(-\pi),
\end{align}
which closes the DMFT equations.
It is instructive to verify the symmetries for the DMFT equations \eqref{action}, \eqref{bethesc}, \eqref{hlocKK}, \eqref{NeelAFO}: One can show that in equilibrium (i.e., for $\phi_a=0$) the solution is rotationally invariant around the $\sigma_2$ axis in orbital pseudospin, i.e., if $\hat G$ is a solution, $\hat G(\theta)\equiv \hat R(\theta) \hat G \hat R(-\theta)$ is a solution as well for all $\theta$. To show this we rotate the spinors,
$\hat \psi(\theta)\equiv R(\theta) \hat \psi$, so that $
\hat G(\theta)(t,t')
=-i
\text{tr}
[T_{\mathcal{C}}
e^{\mathcal{S}}
\hat \psi(\theta)(t)
\hat \psi(\theta)^\dagger(t')
]/Z$.
The interaction is rotationally invariant,
$H_{loc}[\hat\psi^\dagger,\hat\psi]=H_{loc}[\hat\psi^\dagger(\theta),\hat\psi(\theta)]$, and the hybridization can be written as
$-i \int_{\mathcal{C}} dtdt'
\hat\psi^\dagger(\theta)(t)
\hat R(\theta)
\hat \Delta(t,t')\hat R(-\theta)
\hat\psi(\theta)(t')$. We can see that $\hat G(\theta)(t,t')$ is a solution of the DMFT equations if the rotated hybridiyation function
$\hat R(\theta)
\hat \Delta_A(t,t')
\hat R(-\theta)
$ satisfies the self-consistency \eqref{bethesc} with the rotated Green's function, i.e.,
\begin{align}
\hat R(\theta) \Big[
\sum_{a=x,y,z}
\,\, &v_a \hat G_B(t,t') \hat v_a
\Big]
\hat R(-\theta)
\nonumber
\\
&\stackrel{!}{=}
\sum_{a=x,y,z}
\,\,\hat v_a \hat R(\theta) \hat G_B(t,t') \hat R(-\theta) \hat v_a,
\end{align}
which can be written as
\begin{align}
I(\theta)
&\equiv
\sum_{a=x,y,z}
\!\!\!\!
\,\hat v_a(\theta) \hat G_B(t,t') \hat v_a(\theta)
\stackrel{!}{=}I(0),
\end{align}
with $\hat v_z(\theta)= \hat R(-\theta) \hat v_z \hat R(\theta)$. Using Eq.~\eqref{hoppx} and \eqref{hoppy}, we get
\begin{align}
\label{III}
I(\theta)=
\sum_{n=-2,0,2}
\!\!\!
\hat v_z(\theta+\frac{2n\pi}{3}) \hat G_B(t,t') \hat v_z(\theta+\frac{2n\pi}{3}).
\end{align}
Explicit evaluation gives
\begin{align}
\hat v_z(\phi) \hat G \hat v_z(\phi)
&=
G_{11}
\begin{pmatrix}
s^4
&
-s^3c
\\
-s^3c
&
s^2 c^2
\end{pmatrix}
+
G_{22}
\begin{pmatrix}
s^2 c^2
&
-sc^3
\\
-sc^3
&
c^4
\end{pmatrix}\nonumber\\&
+
(G_{12}+G_{21})
\begin{pmatrix}
-s^3c
&
-s^2c^2
\\
-s^2c^2
&
-sc^3
\end{pmatrix}
,
\label{Grot}
\end{align}
with $s\equiv \sin(\phi/2)$ and $c\equiv \cos(\phi/2)$.
After summation in Eq.~\eqref{III} we get,
\begin{align}
I(\theta)=
\frac{1}{8}
\begin{pmatrix}
9G_{11}+3G_{22} &
-3(G_{12}+G_{21})
\\
-3(G_{12}+G_{21})
&
3G_{11}+9G_{22}
\end{pmatrix},
\end{align}
independent of $\theta$. This implies that the DMFT solutions will be rotationally invariant in equilibrium, which is a consequence of the spacial mean-field character of the equations: In the real lattice, the mean-field pseudospin solution is rotationally invariant, and the rotational invariance is broken to the lattice point group only by the order-by-disorder mechanism, which takes into account fluctuations of the order parameter around the long range order (spin waves), which are not correctly captured in DMFT \cite{Nussinov2015}.
|
\subsubsection*{Conventions} I work over the ground field $\mathbb{C}$, though this assumption can be weakened. Varieties $X$ are assumed quasi-projective, with bounded derived category of coherent sheaves denoted by $D(X)$. The variety of hyperplanes in a vector space $V$ is denoted by $\mathbb{P}V$.
\section{{Deformation algebras for $3$-folds}}
The theorem below applies noncommutative deformations to study derived symmetries of $3$-folds. Given smooth $3$-folds $X$ and $X'$ related by a flop, Bridgeland \cite{BriFlop} constructs certain canonical Fourier--Mukai equivalences
\begin{equation*}\label{equation flops}\begin{tikzpicture}
\node (A) at (0,0) {$\phantom{.}D(X)$};
\node (B) at (3,0) {$D(X').$};
\node (C) at (1.5,0) {$\scriptstyle\sim$};
\draw[arrow,transform canvas={yshift=+0.8ex}] (A) -- node[above] {$\scriptstyle\mathsf{F}$} (B);
\draw[arrow,transform canvas={yshift=-0.6ex}] (B) -- node[below] {$\scriptstyle\mathsf{F'}$} (A);
\end{tikzpicture}
\vspace{-0.25cm}
\end{equation*}
These equivalences are not mutually inverse: the theorem explains this using deformations of curves on $X$.
Consider a $3$-fold $Y$\ with an isolated rational singular point~$p$, and a resolution $f\colon X \to Y$ of this singularity, with one-dimensional exceptional locus with components~$C_i$ for $i=1, \dots, n$.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}
\node at (0,0) {\begin{tikzpicture}[scale=0.6]
\coordinate (T) at (1.9,2);
\coordinate (B) at (2.1,1);
\coordinate (L) at (1.8,1.4);
\coordinate (R) at (2.8,1.5);
\draw[line width=0.5pt] (T) to [bend left=25] (B);
\draw[line width=0.5pt] ($(L)+(0.15,0)$) to [bend left=25] ($(R)+(0.15,0)$);
\draw[line width=0.5pt] ($(L)+(0.1,0.15)$) to [bend left=25] ($(R)+(0.1,0.2)$);
\draw[line width=0.5pt] ($(L)+(0.05,0.3)$) to [bend left=25] ($(R)+(0.05,0.4)$);
\draw[rounded corners=5pt,line width=0.75pt] (0.5,0) -- (1.5,0.3)-- (3.6,0) -- (4.3,1.5)-- (4,3.2) -- (2.5,2.7) -- (0.2,3) -- (-0.2,2)-- cycle;
\node at (-0.5,2.5) {$X$};
\end{tikzpicture}};
\node at (0,-2) {\begin{tikzpicture}[scale=0.6]
\filldraw (2.1,0.75) circle (1.5pt);
\draw[rounded corners=5pt,line width=0.75pt] (0.5,0) -- (1.5,0.15)-- (3.6,0) -- (4.3,0.75)-- (4,1.6) -- (2.5,1.35) -- (0.2,1.5) -- (-0.2,1)-- cycle;
\node at (-0.5,0.5) {$Y$};
\node (label) at (2.6,0.75) {$p$};
\end{tikzpicture}};
\draw[->] (0.25,-1) -- node[color=black,left]{$f$} (0.25,-1.5);
\end{tikzpicture}
\end{center}
\caption{Contraction $f$ for Theorem~\ref{theorem flop flop}.}
\end{figure}
\begin{theorem}\label{theorem flop flop} \cite{DW1,DW2,DW3} Noting that the subvarieties $C_i$ of $X$\! are projective lines, we have that:
\begin{enumerate}
\item\label{theorem flop flop 1} there exists a $\mathbb{C}^n$-algebra $A$ which represents the functor of noncommutative deformations of the sheaves~$\mathcal{O}_{C_i}(-1)$ on $X$.
\end{enumerate}
\noindent Write $E$ for the corresponding universal sheaf on $X$. If the contraction~$f$ corresponds to a flop of $X$, then:
\begin{enumerate}[resume]
\item\label{theorem flop flop 2} there is a Fourier--Mukai autoequivalence $\twistLocal$ of~$D(X)$, fitting into a distinguished triangle of functors
\begin{equation*}\label{triangle} \mathbb{R}\!\operatorname{Hom}_X(E,-) \underset{A}{\overset{\mathbb{L}}{\otimes}} E \longrightarrow \operatorname{Id}_{D(X)} \longrightarrow \twistLocal \longrightarrow; \end{equation*}
\item\label{theorem flop flop 3} there is a natural isomorphism of functors
\[\twistLocal \cong \left( \mathsf{F}' \circ \mathsf{F} \right)^{-1}\!.\]
\end{enumerate}
\end{theorem}
In the simplest flopping situation, where $f$ contracts a $(-1,-1)$-curve, the autoequivalence $\twistLocal$ is a spherical twist in the sense of Seidel--Thomas~\cite{ST}. For a contraction of a $(-2,0)$-curve, it is a generalized spherical twist as first constructed by Toda~\cite{Toda}, who furthermore established the conclusion of Theorem~\ref{theorem flop flop}(\ref{theorem flop flop 3}) in this case.
\begin{remark} The noncommutative deformation theory used here relies on work of Laudal~\cite{Laudal}, Eriksen~\cite{Eriksen}, E.\ Segal~\cite{Segal}, and Efimov--Lunts--Orlov~\cite{ELO1}.
\end{remark}
\begin{remark} The algebra $A$ above, and similar noncommutative deformation algebras, have now been applied in settings including: enumerative geometry of curves on $3$-folds by Toda and Hua--Toda~\cite{TodaGV,HuaT}; flops of families of curves in higher dimensions by Bodzenta and Bondal~\cite{BB}; construction of autoequivalences and exceptional objects by Kawamata~\cite{Kawamata}; and new braid-type groups of derived symmetries of $3$-folds by the author and Wemyss~\cite{DW3}.
\end{remark}
\begin{remark} The full statement of Theorem~\ref{theorem flop flop} does not require $X$ to be smooth: I leave details to the references.
\end{remark}
\section{General results}
The following theorem from~\cite{DW4} gives a sheafy analogue of the deformation algebra $A$, applicable in higher dimensions. For a birational contraction $f\colon X \to Y$ satisfying the assumption below, we define a sheaf of algebras $\mathcal{A}$ on $Y$ which is supported on the locus over which $f$ is not an isomorphism. We furthermore construct an associated autoequivalence of $D(X)$.
\begin{assumption} Suppose that $f\colon X \to Y$ is a contraction with $\operatorname{dim} X \geq 2$, and that either:
\begin{enumerate}[label=(\alph*),ref=\alph*]
\item\label{assumption 1} the variety $X$ has an $f$\!-relative tilting generator with summand $\mathcal{O}_X$, where $f$ is crepant, and \,$Y$\!~is Gorenstein;
\end{enumerate}
or, alternatively,
\begin{enumerate}[label=(\alph*),ref=\alph*,resume]
\item\label{assumption 2} the fibres of $f$ have dimension at most one.
\end{enumerate}
\end{assumption}
\begin{remark} The tilting generator assumption from~(\ref{assumption 1}) is satisfied in a range of situations, including symplectic resolutions of quotient singularities as established by Bezrukavnikov and Kaledin~\cite{BK}, and contractions with fibres of dimension at most two under conditions of Toda and~Uehara~\cite{TodaUehara}.
\end{remark}
Write $Z$ for the locus in $Y$\! over which $f$ is not an isomorphism.
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}
\node at (0,0) {\begin{tikzpicture}[scale=0.6]
\coordinate (T) at (1.9,2);
\coordinate (B) at (2.1,1);
\coordinate (sT) at (1.9,2.3);
\coordinate (sL) at (1.7,1.5);
\coordinate (sR) at (2.35,1.5);
\coordinate (sB) at (2.2,0.7);
\draw[rounded corners=7pt,line width=0.5pt] (sT) -- (sR) -- (sB) -- (sL) -- cycle;
\foreach \x/\y/\h in {0.31/0.02/0.08,0.51/0.03/0.06,0.71/0.04/0.04,0.91/0.05/0.02}{
\draw[line width=0.5pt] ($(T)+(\x,\y)+(0.02,\h)$) to [bend left=25] ($(B)+(\x,\y)+(0.02,-\h)$);
\draw[line width=0.5pt] ($(T)+(-\x,-\y)+(0.02,\h)$) to [bend right=25] ($(B)+(-\x,-\y)+(0.02,-\h)$);
}
\draw[dashed,rounded corners=6pt,line width=0.75pt] (0.5,0.3) -- (1.5,0.3)-- (3.6,0) -- (4.3,1.5)-- (3.8,2.7) -- (2.5,2.7) -- (0.3,2.8) -- (-0.2,1.8)-- cycle;
\node at (-0.5,2.5) {$X$};
\end{tikzpicture}};
\node at (0,-2) {\begin{tikzpicture}[scale=0.6]
\filldraw (2.1,0.81) circle (1pt);
\draw (1,0.65) to [bend left=5] (3,0.85);
\draw[dashed,rounded corners=5pt,line width=0.75pt] (0.5,0.15) -- (1.5,0.15)-- (3.6,0) -- (4.3,0.75)-- (3.8,1.35) -- (2.5,1.35) -- (0.3,1.4) -- (-0.2,0.9)-- cycle;
\node at (-0.5,0.5) {$Y$};
\node (label) at (3.4,0.85) {$Z$};
\end{tikzpicture}};
\draw[->] (0.25,-0.9) -- node[color=black,left]{$f$} (0.25,-1.4);
\end{tikzpicture}
\end{center}
\caption{Contraction $f$ for Theorem~\ref{theorem dw4 simple version}.}
\end{figure}
\begin{theorem}\label{theorem dw4 simple version} \cite{DW4} Under the assumption above, there is a sheaf of algebras~$\mathcal{A}$ on~$Y$\!, inducing an object $\mathcal{E}$ of $D(X)$, such that:
\begin{enumerate}
\item\label{theorem dw4 simple version 1} the support of $\mathcal{A}$ is $Z$.
\end{enumerate}
For points $z$ of $Z$ such that $f^{-1}(z)$ is one-dimensional with components~$C_i$, then:
\begin{enumerate}[start=2]
\item\label{theorem dw4 simple version 3} the completion $\mathcal{A}_z$ is an algebra which prorepresents the functor of noncommutative deformations of the sheaves $\mathcal{O}_{C_i}(-1)$ on $X$, up to Morita equivalence;
\item\label{theorem dw4 simple version 4} the restriction of\, $\mathcal{E}$ to the formal fibre of $f$\! over $z$ is a sheaf, namely the universal family corresponding to the prorepresenting object given in~(\ref{theorem dw4 simple version 3}), up to summands of finite sums of sheaves.
\end{enumerate}
If the following hold:
\begin{enumerate}[label=(\roman*),ref=\roman*]
\item\label{theorem dw4 simple version assumptions 1} the contraction $f$ is crepant,
\item\label{theorem dw4 simple version assumptions 2} the base $Y$ is complete locally a hypersurface at each point of~$Z$,
\end{enumerate}
and either $\operatorname{codim} Z \geq 3$ or, alternatively,
\begin{enumerate}[label=(\roman*),ref=\roman*,resume]
\item\label{theorem dw4 simple version assumptions 3} the sheaf $\mathcal{A}$ is Cohen--Macaulay, and
\item\label{theorem dw4 simple version assumptions 4} the object $\mathcal{E}$ is perfect,
\end{enumerate}
then:
\begin{enumerate}[start=4]
\item\label{theorem dw4 simple version 2} there is a Fourier--Mukai autoequivalence~$\twistGlobal$ of~$D(X)$, fitting into a distinguished triangle of functors
\begin{equation*}\label{triangle 2} f^{-1} \mathbb{R}f_* \mathbb{R}\mathcal{H}om_X(\mathcal{E},-) \!\!\underset{f^{-1} \mathcal{A}}{\overset{\mathbb{L}}{\otimes}}\! \mathcal{E} \longrightarrow \operatorname{Id}_{D(X)} \longrightarrow \twistGlobal \longrightarrow.
\end{equation*}
\end{enumerate}
\end{theorem}
I indicate the construction of the sheaf of algebras $\mathcal{A}$, and explain how it allows us to prove Theorem~\ref{theorem dw4 simple version}. Under the assumption above, we have an $f$-relative tilting generator $\mathcal{O}_X \oplus N$, either by assertion in case~(\ref{assumption 1}), or by a theorem of Van~den~Bergh~\cite{VdB} in case~(\ref{assumption 2}). Let $\mathcal{T}$ denote the relative endomorphism algebra of $\mathcal{O}_X \oplus N$, a~sheaf of algebras on~$Y$. We establish that
\[\mathcal{T} = f_* \mathcal{E}nd_X (\mathcal{O}_X \oplus N) \cong \mathcal{E}nd_Y f_* (\mathcal{O}_X \oplus N). \]
This allows us to make the following definition for $\mathcal{A}$. This is a sheafy version of a construction of the algebra $A$ from our previous work~\cite{DW1}.
\begin{definition}\cite{DW4} Let $\mathcal{A} = \mathcal{T} / \mathcal{I}$, a sheaf of algebras on $Y$\!, where $\mathcal{I}$ is the ideal of sections of $\mathcal{T}$ which factor, at each stalk, through a sum of copies of $\mathcal{O}_Y$.
\end{definition}
The prorepresenting property of $\mathcal{A}$ in Theorem~\ref{theorem dw4 simple version}(\ref{theorem dw4 simple version 3}) is then proved as a sheafy version of the representing property of $A$ in Theorem~\ref{theorem flop flop}(\ref{theorem flop flop 1}). The object $\mathcal{E}$ of $D(X)$ is defined as the image of $\mathcal{A}$ under an appropriate tilting equivalence: I refer to \cite[Section~3]{DW4} for a precise statement. Theorem~\ref{theorem dw4 simple version}(\ref{theorem dw4 simple version 4}) show that this $\mathcal{E}$ has a universal property which is a sheafy version of the universal property of $E$ in Theorem~\ref{theorem flop flop}. We also have the following.
\begin{proposition}\cite{DW4} The support of $\mathcal{E}$ is contained in the exceptional locus of~$f$.
\end{proposition}
The construction of a Fourier-Mukai autoequivalence~$\twistGlobal$ in Theorem~\ref{theorem dw4 simple version}(\ref{theorem dw4 simple version 2}) generalizes the construction of~$\twistLocal$ from Theorem~\ref{theorem flop flop}(\ref{theorem flop flop 2}). In particular, we have the following.
\begin{remark}When $Z$ is a point, the autoequivalence~$\twistGlobal$ reduces to the autoequivalence~$\twistLocal$ appearing in Theorem~\ref{theorem flop flop}(\ref{theorem flop flop 2}).
\end{remark}
\begin{remark}{} Although the tilting generator $\mathcal{O}_X \oplus N$, and thence the sheaf of algebras $\mathcal{A}$, is not canonically defined (see for instance the construction of Van~den~Bergh in~\cite{VdB}) it seems that the autoequivalence $\twistGlobal$ may be canonical, given a choice of contraction~$f$. For instance, given two different tilting generators related by duplication of summands we obtain Morita equivalent sheaves of algebras $\mathcal{A}$, and thence isomorphic autoequivalences~$\twistGlobal$. \end{remark}
\begin{remark}
For $f$ a flopping contraction, it would be interesting to establish when $\twistGlobal$ is related to a flop-flop functor, as in Theorem~\ref{theorem flop flop}(\ref{theorem flop flop 3}).
\end{remark}
\begin{remark}
It is tempting to speculate that the `tilting' condition in the requirement for an $f$-relative tilting generator in~(\ref{assumption 1}) may be relaxed by upgrading~$\mathcal{A}$ to an appropriate sheaf of differential graded algebras.
\end{remark}
I record the following $3$-fold setting where the assumptions of Theorem~\ref{theorem dw4 simple version} may be established.
\begin{theorem}\label{theorem dw4 3fold irred version}\cite{DW4} With $\operatorname{dim} X = 3$, assume that
\begin{enumerate}[label=(\roman*),ref=\roman*]
\item the contraction $f$ is crepant,
\item the base $Y$ is complete locally a hypersurface at each point of~$Z$,
\item the exceptional fibres of $f$ are irreducible curves.
\end{enumerate}
Then the assumptions of Theorem\! \ref{theorem dw4 simple version} hold, and there exists an associated autoequivalence $\twistGlobal$ of $D(X)$.
\end{theorem}
\begin{figure}[h]
\begin{center}
\begin{tikzpicture}
\node at (0,0) {\begin{tikzpicture}[scale=0.6]
\coordinate (T) at (1.9,2);
\coordinate (TM) at (2.12-0.02,1.5-0.15);
\coordinate (BM) at (2.12-0.09,1.5+0.2);
\coordinate (B) at (2.1,1);
\draw[line width=1pt] (T) to [bend left=25] (B);
\foreach \x/\y in {0.2/0.02,0.4/0.04,0.6/0.06,0.8/0.065,1/0.07}{
\draw[line width=0.5pt] ($(T)+(\x,0)+(0.02,\y)$) to [bend left=25] ($(B)+(\x,0)+(0.02,\y)$);
\draw[line width=0.5pt] ($(T)+(-\x,0)+(-0.02,-\y)$) to [bend left=25] ($(B)+(-\x,0)+(-0.02,-\y)$);}
\draw[rounded corners=5pt,line width=0.75pt] (0.5,0) -- (1.5,0.3)-- (3.6,0) -- (4.3,1.5)-- (4,3.2) -- (2.5,2.7) -- (0.2,3) -- (-0.2,2)-- cycle;
\node at (-0.5,2.5) {$X$};
\end{tikzpicture}};
\node at (0,-2) {\begin{tikzpicture}[scale=0.6]
\filldraw (2.1,0.77) circle (1pt);
\draw (1,0.65) to [bend right=5] (2,0.75);
\draw (2,0.75) to [bend left=5] (3,0.85);
\draw[rounded corners=5pt,line width=0.75pt] (0.5,0) -- (1.5,0.15)-- (3.6,0) -- (4.3,0.75)-- (4,1.6) -- (2.5,1.35) -- (0.2,1.5) -- (-0.2,1)-- cycle;
\node at (-0.5,0.5) {$Y$};
\node (label) at (3.4,0.85) {$Z$};
\end{tikzpicture}};
\draw[->] (0.25,-1) -- node[color=black,left]{$f$} (0.25,-1.5);
\end{tikzpicture}
\end{center}
\caption{Contraction $f$ for Theorem~\ref{theorem dw4 3fold irred version}.}
\end{figure}
\newpage
\section{Springer resolution example}\label{section springer}
For an example in which the theory of the previous section applies to a contraction with higher-dimensional fibres, consider the Springer resolution of the variety of singular $d$-by-$d$ matrices. Namely, for a vector~space~$V$ of dimension~$d$ with $d \geq 2$, take the singular cone
\[ Y = \big\{ M \in \operatorname{End} V \mathrel{\big|} \det M = 0 \,\big\}, \]
which is a Gorenstein hypersurface. It has a resolution by
\[ X = \big\{ (M,H) \in \operatorname{End} V \times \mathbb{P} V \mathrel{\big|} \operatorname{Im} M \subseteq H \,\big\} \]
whose natural projection\! $f$ to $\operatorname{End} V$ surjects onto $Y$. This resolution~$f$ is crepant. Its exceptional fibres lie over points $M$ in $Y$ with $\operatorname{rk} M < d-1$, and are projective spaces of dimension $d-1-\operatorname{rk} M$.
A tilting generator for $X$ has been constructed by Buchweitz, Leuschke, and Van~den~Bergh \cite{BLV}, so that we are in the setting of Assumption~(\ref{assumption 1}). Conditions~(\ref{theorem dw4 simple version assumptions 1}) and~(\ref{theorem dw4 simple version assumptions 2}) of Theorem~\ref{theorem dw4 simple version} are noted above, and
$\operatorname{codim} Z \geq 3$, so we can apply Theorem~\ref{theorem dw4 simple version}(\ref{theorem dw4 simple version 2}) to obtain an autoequivalence $\twistGlobal$ of $D(X)$.
\begin{remark}For $d=2$, the variety $X$ is just a $3$-fold resolving a conifold~$Y$ with exceptional fibre a $(-1,-1)$-curve, and we are in the setting of Theorem~\ref{theorem flop flop}.
\end{remark}
\begin{remark}
In joint work with E.~Segal~\cite{DonSeg1}, I studied the resolution $f$, along with more general resolutions where $Y$ is the variety of $d$-by-$d$ matrices of rank at most~$r$ for $0 < r < d$. We constructed certain `Grassmannian twist' autoequivalences of the corresponding $D(X)$ by quite different methods: it~would be interesting to compare these with $\twistGlobal$.
\end{remark}
\begin{remark} The sheaf of algebras $\mathcal{A}$ for this example may be computed from the presentation of the endomorphism algebra $\mathcal{T}$ in~\cite{BLV}.
\end{remark}
\section{Conjectures}\label{section conjectures}
In the setting of a $3$-fold flopping contraction as in Theorem~\ref{theorem flop flop}, we made a conjecture~\cite[Conjecture~1.4]{DW1} that the complete local neighbourhood of the $3$-fold $Y$ near the singularity~$p$ is determined, up to isomorphism, by~the deformation algebra $A$. This conjecture is clear in the following simple cases, namely the two kinds of flopping curve for which $A$ is commutative, but remains open more generally.
\newpage\begin{enumerate}
\item Contractions of $(-1,-1)$-curves. In this case $A=\mathbb{C}$, and the completion of $Y$ at $p$ is necessarily a conifold singularity.
\item Contractions of $(-2,0)$-curves. Here $A \cong \mathbb{C}[\varepsilon]/\varepsilon^w$ with $w\geq 2$, where~$w$ is the width invariant of Reid~\cite{Pagoda}, and the completion of~$Y$ at $p$ is determined by $w$.
\end{enumerate}
Hua and Toda subsequently proposed an $A_\infty$ version of the conjecture~\cite[Conjecture~5.3]{HuaT} in which $A$ is endowed with the structure of an $A_\infty$-algebra. They established their conjecture for contractions to weighted homogeneous hypersurface singularities~\cite[Theorem~5.5]{HuaT}, and it has now been settled in general by Hua~\cite{Hua}. A key idea in these works is that the deformation algebra~$A$ may be viewed as a noncommutative analogue of the Milnor algebra, and that the $A_\infty$ structure on it allows recovery of the Milnor algebra along with enough information to apply a version of the Mather--Yau theorem.
\begin{remark}
It would be interesting to extend these results to higher dimensions, and to non-isolated singularities. For instance, it is natural to ask whether the the complete local neighbourhood of the variety $Y$\! near the non-isomorphism locus $Z$ is determined by $(Z,\mathcal{A})$, potentially along with some appropriate $A_\infty$~structure.\end{remark}
\newpage
\renewcommand{\baselinestretch}{1.15}
|
\section{Introduction}
Let $F$ be a nonarchimedean local field of characteristic zero and odd residual characteristic. Let $G = \G(F)$ be the $F$-points of a connected reductive group defined over $F$. Let $\theta$ be an $F$-involution (order two $F$-automorphism) of $\G$ and let $H = \G^\theta(F)$ be the group of $\theta$-fixed points in $G$. The quotient $H \backslash G$ is a $p$-adic symmetric space. An irreducible representation of $G$ that occurs in the discrete spectrum of $H\backslash G$ (an irreducible subrepresentation of $L^2(Z_GH\backslash G)$, where $Z_G$ is the centre of $G$) is called a relative discrete series (RDS) representation.
In this paper, we construct an infinite family of RDS representations for $H\backslash G$, that do not appear in the discrete spectrum of $G$, in two cases:
\begin{enumerate}
\item The linear case: $G = \GL_{2n}(F)$ and $H = \GL_n(F) \times \GL_n(F)$.
\item The Galois case: $G = \GL_n(E)$ and $H = \GL_n(F)$, where $E/F$ is a quadratic extension.
\end{enumerate}
In a more general setting, Murnaghan has constructed relatively supercuspidal (\textit{cf.}~\Cref{defn-rsc}) representations that are not supercuspidal \cite{murnaghan2011a, murnaghan2016-pp}. Her construction is also via parabolic induction from representations of $\theta$-elliptic Levi subgroups (\textit{cf.}~\Cref{theta-elliptic-defn}) and provided the initial motivation for this work. We obtain a special case of Murnaghan's results in our setting (\textit{cf.}~\Cref{cor-sc-main-thm}); however, we apply completely different methods.
The study of harmonic analysis on $H\backslash G$ is of interest due to connections with non-vanishing of global period integrals, functoriality and poles of $L$-functions (see, for instance, \cite{jacquet--rallis1996, kable2004, feigon--lapid--offen2012}). For example, often $\GL_n(F) \times \GL_n(F)$-distinction of a representation of $\GL_{2n}(F)$ is equivalent to the existence of a nonzero Shalika model (\textit{cf.}~\Cref{rmk-lin-Shalika}).
In addition, $H\backslash G$ is a spherical variety and its study fits into the general framework of \cite{sakellaridis--venkatesh2017}. In a broad sense, the work of Sakellaridis--Venkatesh lays the formalism and foundations of a relative Langlands program (\textit{cf.} \cite{prasad2015-pp}). The aim of such a program is to fully understand the link between global automorphic period integrals and local harmonic analysis.
For certain $p$-adic spherical varieties $X$, Sakellaridis--Venkatesh give an explicit Plancherel formula describing $L^2(X)$ up to a description of the discrete spectrum \cite[Theorem 6.2.1]{sakellaridis--venkatesh2017}.
In addition, the results of \cite{sakellaridis--venkatesh2017} include a description of the discrete spectrum of $p$-adic symmetric spaces in terms of toric families of RDS. In fact, they show that \cite[Conjecture 9.4.6]{sakellaridis--venkatesh2017} is true for strongly factorizable spherical varieties.
The decomposition of the norm on the discrete spectrum $L^2_{\operatorname{disc}}(X)$ provided by this result is not necessarily a direct integral, the images of certain intertwining operators packaged with the toric families of RDS may be non-orthogonal.
On the other hand, Sakellaridis and Venkatesh believe that their conjectures on $\X$-distinguished Arthur parameters may give a canonical choice of mutually orthogonal toric families of RDS that span $L^2_{\operatorname{disc}}(X)$ \cite[Conjectures 1.3.1 and 16.2.2]{sakellaridis--venkatesh2017}.
However, Sakellaridis--Venkatesh do not give an explicit description of the RDS required to build the toric families. For this reason, an explicit construction of RDS for $p$-adic symmetric spaces is a step towards completing the picture of the discrete spectrum of $H\backslash G$. Kato and Takano \cite{kato--takano2010} have shown that any $H$-distinguished discrete series representation of $G$ is a RDS. Thus, we are interested in constructing RDS representations of $G$ that are in the complement of the discrete spectrum of $G$.
We state our main theorem below, after giving the necessary definitions.
A $\theta$-stable Levi subgroup $L$ of $G$ is $\theta$-elliptic if it is not contained in any proper $\theta$-split parabolic subgroup, where a parabolic subgroup $P$ is $\theta$-split if $\theta(P)$ is opposite to $P$.
An element $g\in G$ is said to be $\theta$-split if $\theta(g) = g^{-1}$ and a $\theta$-stable subset $Y$ of $G$ is $\theta$-split if every element $y\in Y$ is $\theta$-split.
An $F$-torus $S$ is $(\theta,F)$-split if it is both $F$-split and $\theta$-split.
A representation $\tau$ of a Levi subgroup $L$ of $G$ is regular if for every non-trivial element $w\in N_G(L)/L$ we have that the twist ${}^w \tau = \tau ( w^{-1} (\cdot) w)$ is not equivalent to $\tau$.
Let $n\geq 2$ (respectively, $n\geq 4$) and let $G$ be equal to $\GL_{2n}(F)$ (respectively, $\GL_{n}(E)$) and let $H$ be equal to $\GL_{n}(F) \times \GL_n(F)$ (respectively, $\GL_n(F)$).
Let $A_0$ be a $\theta$-stable maximal $F$-split torus of $G$ containing a fixed maximal $(\theta, F)$-split torus $S_0$.
Let $L_0 = C_G( (A_0^\theta)^\circ)$ be a minimal $\theta$-elliptic Levi subgroup of $G$ containing $A_0$.
We make a particular choice $\Delta^{ell}$ of simple roots for the root system $\Phi(G,A_0)$ (\textit{cf.}~$\S$\ref{sec-theta-ell-levi}).
The $F$-split component of the centre of $L_0$ is determined by a proper nonempty subset $\Delta^{ell}_{\min}$ of $\Delta^{ell}$.
A Levi subgroup $L$ is a standard-$\theta$-elliptic Levi subgroup if $L$ is standard with respect to $\Delta^{ell}$ and contains $L_0$.
The main result of this paper is the following.
\begin{thm*}[\Cref{main-RDS-thm}]
Let $\Omega^{ell} \subset {\Delta^{ell}}$ be a proper subset such that $\Omega^{ell}$ contains $\Delta^{ell}_{\min}$.
Let $Q = Q_{\Omega^{ell}}$ be the proper $\Delta^{ell}$-standard parabolic subgroup associated to the subset $\Omega^{ell}$. The parabolic subgroup $Q$ is $\theta$-stable and has $\theta$-stable standard-$\theta$-elliptic Levi subgroup $L = L_{\Omega^{ell}}$.
Let $\tau$ be a regular $L^{\theta}$-distinguished discrete series representation of $L$.
The parabolically induced representation $\pi = \iota_Q^G\tau$ is an irreducible $H$-distinguished relative discrete series representation of $G$.
Moreover, $\pi$ is in the complement of the discrete series of $G$.
\end{thm*}
\begin{rmk}
The representations constructed in \Cref{main-RDS-thm} are induced from discrete series, and are therefore tempered (generic) representations of $G$. In particular, one observes that $L^2(Z_GH\backslash G)$ contains the $H$-distinguished discrete series representations of $G$ \cite{kato--takano2010}, as well as certain tempered representations that do not appear in $L^2(Z_G\backslash G)$.
\end{rmk}
Outside of two low-rank examples considered by Kato--Takano \cite[$\S$5.1-2]{kato--takano2010}, \Cref{main-RDS-thm} provides the first construction of a family of non-discrete relative discrete series representations.
In \Cref{cor-infinite-family}, we show that there are infinitely many equivalence classes of representations of the form constructed in \Cref{main-RDS-thm}.
However, we do not prove that our construction exhausts the discrete spectrum in these two cases (\textit{cf.}~\Cref{rmk-exhaustion}). It appears that the major obstruction to showing that an irreducible $H$-distinguished representation of $G$ is not relatively discrete is establishing non-vanishing of the invariant forms $r_P\lambda$ defined by Lagier and Kato--Takano \cite{lagier2008,kato--takano2008} (\textit{cf.}~$\S$\ref{sec-r-P-lambda}).
In the author's PhD thesis, a similar (more restricted) construction is carried out for the case $G = \GL_{2n}(E)$ and $H = \mathbf U_{E/F}$ a quasi-split unitary group \cite[Theorem 5.2.22]{smith-phd2017}. It is work in progress to extend the construction to arbitrary symmetric quotients of the general linear group. Some modification will be required for this generalization; the representations constructed in \Cref{main-RDS-thm} are generic and no such representation can be distinguished by the symplectic group \cite{heumos--rallis1990}. It is expected that the Speh representations form the discrete spectrum of $\mathbf{Sp}_{2n}(F) \backslash \GL_{2n}(F)$ \cite{offen--sayag2007, sakellaridis--venkatesh2017}.
We now give an outline of the content of the paper. In \Cref{sec-notation}, we establish notation and our conventions in the linear and Galois cases. In \Cref{sec-parabolic-inv}, we review basic results on tori and parabolic subgroups relevant to the study of harmonic analysis on $H \backslash G$. Here we introduce the notion of a $\theta$-elliptic Levi subgroup. \Cref{sec-rep-theory}, contains a review of Kato--Takano's generalization of Casselman's Criterion, preliminaries on distinguished representations and some results on the exponents of induced representations. The most important results in this section are \Cref{reduction-standard-split}, \Cref{red-to-ind-exp} and \Cref{non-dist-gen-eig-sp}. In \Cref{sec-structure-lin-Gal}, we give explicit descriptions of the tori, parabolic subgroups, and simple roots needed for our work in the linear and Galois cases (\textit{cf}.~\Cref{max-theta-split,ind-subgp-prop}).
The main result, \Cref{main-RDS-thm}, is stated and proved in \Cref{sec-main-theorem}; however, several preliminary results required for the proof are deferred until \Cref{sec-exp-dist-pi-N}. In \Cref{sec-inducing-data}, we briefly survey the literature on distinguished discrete series representations in the linear and Galois cases. In addition, we establish the existence of infinite families of inducing representations in \Cref{lem-L-theta-dist}, from which we can deduce \Cref{cor-infinite-family}. Finally, in \Cref{sec-exp-dist-pi-N}, we assemble the technical results, on the exponents and distinction of Jacquet modules, required to prove \Cref{main-RDS-thm}. The main results of the final section are \Cref{prop-reduce-Cas-ind,no-dist-unit-exp}.
\section{Notation and conventions}\label{sec-notation}
Let $F$ be a nonarchimedean local field of characteristic zero and odd residual characteristic.
Let $\of$ be the ring of integers of $F$ with prime ideal $\prm$.
Let $E$ be a quadratic Galois extension of $F$.
Fix a generator $\varepsilon$ of the extension $E/F$ such that $E = F(\varepsilon)$.
Let $\sigma \in \Gal(E/F)$ be a generator of the Galois group of $E$ over $F$.
Let $\G$ be a connected reductive group defined over $F$ and let $G = \G(F)$ denote the group of $F$-points.
Let $e$ be the identity element of $G$.
We let $Z_G$ denote the centre of $G$ while $A_G$ denotes the $F$-split component of the centre of $G$.
As is the custom, we will often abuse notation and identify an algebraic group defined over $F$ with its group of $F$-points.
When the distinction is to be made, we will use boldface to denote the algebraic group and regular typeface to denote the group of $F$-points.
For any $F$-torus $\mathbf{A}$ of $\G$, we let $A^1$ denote the group of $\of$-points $A^1=\mathbf A (\of)$.
Let $P$ be a parabolic subgroup of $G$ and let $N$ be the {unipotent radical} of $P$.
The modular character of $P$ is given by $\delta_P(p) = | \det \Ad_{\style n}(p) |$, for all $p\in P$, where $\Ad_{\style n}$ denotes the adjoint action of $P$ on the Lie algebra $\style n$ of $N$ \cite{Casselman-book}.
Let $\theta$ be an $F$-involution of $\G$, that is, an order-two automorphism of $\G$ defined over $F$.
Define $\Hbf=\G^\theta$ to be the closed subgroup of $\theta$-fixed points of $\G$.
The quotient $H\backslash G$ is a $p$-adic symmetric space.
\begin{defn}\label{defn-involution-action}
We say that an involution $\theta_1$ of $G$ is $G$-conjugate (or $G$-equivalent) to another involution $\theta_2$ if there exists $g\in G$ such that $\theta_1 = \Int g^{-1} \circ \theta_2 \circ \Int g$,
where $\Int g$ denotes the inner $F$-automorphism of $\G$ given by $\Int g(x) = g x g^{-1}$, for all $x\in \G$.
We write $g\cdot \theta$ to denote the involution $\Int g^{-1} \circ \theta \circ \Int g$.
\end{defn}
Let $\GL_n$ denote the general linear group of $n$ by $n$ invertible matrices. As is customary, we denote the block-upper triangular parabolic subgroup of $\GL_n$, corresponding to a partition $(\underline{m}) = (m_1,\ldots, m_k)$ of $n$,
by $\mathbf P_{(\underline{m})}$, with block-diagonal Levi subgroup $\mathbf M_{(\underline{m})} \cong \prod_{i=1}^k \GL_{m_i}$ and unipotent radical $\mathbf N_{(\underline{m})}$.
We use $\diag(a_1,a_2, \ldots, a_n)$ to denote an $n \times n$ diagonal matrix with entries $a_1,\ldots, a_n$.
For any $g,x \in G$, we write ${}^g x = gxg^{-1}$. For any subset $Y$ of $G$, we write ${}^g Y = \{ {}^g y : y \in Y\}$.
Let $C_G(Y)$ denote the centralizer of $Y$ in $G$ and let $N_G(Y)$ be the normalizer of $Y$ in $G$.
Given a real number $r$ we let $\floor{r}$ denote the greatest integer that is less than or equal to $r$.
We use $\widehat {(\cdot)}$ to denote that a symbol is omitted. For instance, $\diag(\widehat{a_1}, a_2, \ldots, a_n)$ may be used to denote the diagonal matrix $\diag(a_2,\ldots, a_n)$.
\subsection{The linear case}\label{sec-notation-lin}
In the linear case, we set $G = \GL_n(F)$, where $n\geq 4$ is an even integer.
Let $\theta$ denote the inner involution of $G$ given by conjugation by the matrix
\begin{align*}
w_\ell = \left( \begin{matrix} & & 1 \\ & \iddots & \\ 1 & & \end{matrix} \right),
\end{align*}
that is, for any $g\in \G$, we have
\begin{align*}
\theta(g) &= \Int w_\ell (g) = w_\ell g w_\ell^{-1}.
\end{align*}
The element $w_\ell$ is diagonalizable over $F$; in particular, there exists $x_\ell \in \GL_n(F)$ such that
\begin{align}\label{eq-w-ell-diag}
x_\ell w_\ell x_\ell^{-1} = \diag(1_{n/2}, -1_{n/2}),
\end{align}
where $1_{n/2}$ denotes the ${n/2}\times {n/2}$ identity matrix.
It follows that $H = x_\ell^{-1} M_{(n/2,n/2)} x_\ell$, where $M_{(n/2,n/2)}$ is the standard Levi subgroup of $G$ of type $(n/2,n/2)$. Thus, in the linear case, $H = \G^\theta(F)$ is isomorphic to $\GL_{n/2}(F)\times \GL_{n/2}(F)$.
\subsection{The Galois case}\label{sec-notation-Gal}
In the Galois case, for $n\geq 4$, we let $\G = R_{E/F}\GL_n$ be the restriction of scalars of $\GL_n$ with respect to $E/F$. We identify the group $G$ of $F$-points with $\GL_n(E)$.
The non-trivial element $\sigma$ of the Galois group of $E$ over $F$ gives rise to an $F$-involution $\theta$ of $G$ given by coordinate-wise Galois conjugation
\begin{align*}
\theta ( (a_{ij}) ) = (\sigma(a_{ij})),
\end{align*}
where $(a_{ij}) \in G$.
In the Galois case, we have that $H = \G^\theta(F)$ is equal to $\GL_n(F)$.
\subsection{Choices of particular group elements and supplementary involutions}\label{sec-choice-elts}
For a positive integer $r$, we'll write $\G_r$ for $\GL_r$ with $F$-points $G_r$ in the linear case, and similarly for $R_{E/F} \GL_r$ with $F$-points $G_r \simeq \GL_r(E)$, in the Galois case.
Write $J_r$ for the $r\times r$-matrix in $G_r$ with unit anti-diagonal
\begin{align*}
J_r = \left( \begin{matrix} & & 1 \\ & \iddots & \\ 1 & & \end{matrix} \right).
\end{align*}
Note that $w_\ell=J_n$.
In the linear case, $\theta_r$ will denote the inner involution $\Int J_r$ of $\G_r$ with fixed points $\Hbf_r$.
In the Galois case, we let $\theta_r$ denote the $F$-involution of $\G_r$ given by coordinate-wise Galois conjugation; then $H_r = \GL_r(F)$ is the group of $F$-points of the $\theta_r$-fixed subgroup of $\G_r$.
In the Galois case, for any positive integer $r$, there exists $\gamma_r \in G_r$ such that $\gamma_r^{-1} \theta_r(\gamma_r) = J_r \in H_r$.
For instance, if $r$ is even, then we may take
\begin{align*
\gamma_r = \left( \begin{matrix}
1 & & & & &1 \\
& \ddots & & & \iddots & \\
& & 1 & 1 & & \\
& & -\varepsilon & \varepsilon & & \\
& \iddots & & & \ddots & \\
-\varepsilon & & & & & \varepsilon
\end{matrix} \right),
\end{align*}
where $E = F(\varepsilon)$, and if $r$ is odd, then we set
\begin{align*
\gamma_r = \left( \begin{matrix}
1 & & && & &1 \\
& \ddots & && & \iddots & \\
& & 1 &0& 1 & & \\
& & 0& 1 & 0& & \\
& & -\varepsilon &0& \varepsilon & & \\
& \iddots & && & \ddots & \\
-\varepsilon & & & & & & \varepsilon
\end{matrix} \right).
\end{align*}
Define $\gamma = \gamma_n \in G$ and note that $w_\ell = J_n = \gamma^{-1}\theta(\gamma)$ is an order-two element of $H$.
In the Galois case, we define a second involution $\vartheta$ of $\G$, that is $G$-conjugate to $\theta$, by declaring that $\vartheta = \gamma \cdot \theta$ (\textit{cf.}~\Cref{defn-involution-action}). Explicitly,
\begin{align}\label{eq-vartheta-defn}
\vartheta (g) = \gamma^{-1} \theta(\gamma g \gamma^{-1}) \gamma,
\end{align}
for any $g\in \G$.
Since $w_\ell=\gamma^{-1}\theta(\gamma)$ is $\theta$-fixed, we have that
\begin{align}\label{Galois-involution-relation}
\vartheta = \Int w_\ell \circ \theta = \theta \circ \Int w_\ell.
\end{align}
Similarly, for any positive integer $r$, we define
\begin{align}\label{vartheta-r-defn}
\vartheta_r = \gamma_r \cdot \theta_r = \Int J_r \circ \theta_r = \theta_r \circ \Int J_r.
\end{align}
In both cases, define $w_+ \in \GL_n(F) \subset \GL_n(E)$ to be the permutation matrix corresponding to the permutation of $\{1,\ldots, n\}$ given by
\begin{align*
&\left\{\begin{array}{lll}
2i-1 & \mapsto i & 1 \leq i \leq \floor{n/2}+1 \\
2i & \mapsto n+1-i & 1 \leq i \leq \floor{n/2}
\end{array}
\right. & (n \ \text{odd}),
\end{align*}
when $n$ is odd, and when $n$ is even by
\begin{align*
&\left\{\begin{array}{lll}
2i-1 & \mapsto i & 1 \leq i \leq n/2 \\
2i & \mapsto n+1-i & 1 \leq i \leq n/2
\end{array}
\right. &(n \ \text{even}).
\end{align*}
Remember that in the linear case we'll always assume that $n$ is even.
Finally, define
\begin{align}\label{w-zero}
w_0 &= \left\{\begin{array}{lll}
w_+ & \text{in the linear case:} \ G = \GL_n(F), n\geq 4 \ \text{even}, \\
{}^\gamma w_+= \gamma w_+ \gamma^{-1} & \text{in the Galois case:} \ G = \GL_n(E), \ \text{any} \ n\geq 4.
\end{array}
\right.
\end{align}
\section{Symmetric spaces and associated parabolic subgroups}\label{sec-parabolic-inv}
For now, we work in general and let $G$ be an arbitrary connected reductive group over $F$, with $\theta$ and $H$ as in \Cref{sec-notation}.
An element $g\in G$ is said to be $\theta$-split if $\theta(g) = g^{-1}$.
A subtorus $S$ of $G$ is $\theta$-split if every element of $S$ is $\theta$-split.
\subsection{Tori and root systems relative to involutions}\label{sec-tori-involution}
An $F$-torus $S$ contained in $G$ is $(\theta,F)$-split if $S$ is both $F$-split and $\theta$-split.
Let $S_0$ be a maximal $(\theta,F)$-split torus of $G$.
By \cite[Lemma 4.5(\rm{iii})]{helminck--wang1993}, there exists a $\theta$-stable maximal $F$-split torus $A_0$ of $G$ that contains $S_0$.
Let $\Phi_0 = \Phi(G,A_0)$ be the root system of $G$ with respect to $A_0$. Let $W_0$ be the Weyl group of $G$ with respect to $A_0$. Since $A_0$ is $\theta$-stable, there is an action of $\theta$ on the $F$-rational characters $X^*(A_0)$ of $A_0$.
Explicitly, given $\chi \in X^*(A_0)$, we have
\begin{align*}
(\theta \chi) (a) = \chi (\theta(a)),
\end{align*}
for all $a \in A_0$.
Moreover, $\Phi_0$ is stable under the action of $\theta$ on $X^*(A_0)$.
Let $\Phi_0^\theta$ denote the subset of $\theta$-fixed roots in $\Phi_0$.
\begin{defn}\label{defn-theta-base}
A base $\Delta_0$ of $\Phi_0$ is called a $\theta$-base if for every positive root $\alpha \in \Phi_0^+$ with respect to $\Delta_0$ that is not fixed by $\theta$, we have that $\theta(\alpha) \in \Phi_0^-$.
\end{defn}
As shown in \cite{helminck1988}, a $\theta$-base of $\Phi_0$ exists. Let $\Delta_0$ be a $\theta$-base of $\Phi_0$.
Let $p: X^*(A_0) \rightarrow X^*(S_0)$ be the surjective homomorphism defined by restricting the $F$-rational characters on $A_0$ to the subtorus $S_0$. The kernel of the map $p$ is the submodule $X^*(A_0)^\theta$ of $X^*(A_0)$ consisting of $\theta$-fixed $F$-rational characters.
The restricted root system of $H \backslash G$ (relative to our choice of $(A_0,S_0,\Delta_0)$) is defined to be
\begin{align*
\overline \Phi_0 = p(\Phi_0)\setminus \{0\} = p(\Phi_0 \setminus \Phi_0^\theta).
\end{align*}
The set $\overline \Phi_0$ coincides with the set $\Phi(G,S_0)$ of roots with respect to $S_0$ and this is a (not necessarily reduced) root system by \cite[Proposition 5.9]{helminck--wang1993}.
The set
\begin{align*
\overline \Delta_0 = p(\Delta_0) \setminus \{0\} = p(\Delta_0 \setminus \Delta_0^\theta)
\end{align*}
is a base for the restricted root system $\overline \Phi_0$.
Indeed, the linear independence of $\overline \Delta_0$ follows from the fact that $\Delta_0$ is a $\theta$-base and that $\ker p = X^*(A_0)^\theta$.
Given a subset $\overline \Theta \subset \overline \Delta_0$, define the subset
\begin{align*
[\overline\Theta ] = p^{-1}(\overline\Theta ) \cup \Delta_0^\theta
\end{align*}
of $\Delta_0$.
Subsets of $\Delta_0$ of the form $[\overline\Theta ]$, for $\overline \Theta \subset \overline \Delta_0$, are said to be $\theta$-split.
The maximal $\theta$-split subsets of $\Delta_0$ are of the form
$[\overline\Delta_0 \setminus\{\bar\alpha\}]$,
where $\bar\alpha \in \overline\Delta_0$.
\subsection{Parabolic subgroups relative to involutions}\label{sec-pblc-rel-inv}
Given a subset $\Theta$ of $\Delta_0$, one may canonically associate a $\Delta_0$-standard parabolic subgroup $P_\Theta$ of $G$ and a standard choice of Levi subgroup.
Let $\Phi_\Theta $ be the subsystem of $\Phi_0$ generated by the simple roots $\Theta$.
Let $\Phi_\Theta ^+$ be the set of positive roots determined by $\Theta$.
The unipotent radical $N_\Theta $ of $P_\Theta $ is generated by the root groups $N_\alpha$, where $\alpha \in \Phi_0^+\setminus \Phi_\Theta ^+$.
The parabolic subgroup $P_\Theta $ admits a Levi factorization $P_\Theta = M_\Theta N_\Theta $, where $M_\Theta $ is the centralizer in $G$ of the $F$-split torus
\begin{align*}
A_\Theta = \left( \bigcap_{\alpha \in \Theta } \ker \alpha \right)^\circ.
\end{align*}
Here $(\cdot)^\circ$ indicates the Zariski-connected component of the identity.
In fact, $A_\Theta $ is the $F$-split component of the centre of $M_\Theta $ and $\Phi_\Theta $ is the root system $\Phi(M_\Theta, A_0)$ of $A_0$ in $M_\Theta $.
\begin{rmk}
When considering standard parabolic subgroups $P_\Theta $, associated to $\Theta \subset \Delta_0$, we will always work with the Levi factorization $P_\Theta = M_\Theta N_\Theta $, where $M_\Theta = C_G(A_\Theta )$ is the standard Levi subgroup of $P_\Theta $.
\end{rmk}
Let $M$ be any Levi subgroup of $G$.
The $(\theta,F)$-split component of $M$ is the largest $(\theta,F)$-split torus $S_M$ contained in the centre of $M$.
In fact, we have that $S_M$ is the connected component (of the identity) of the subgroup of $\theta$-split elements in the $F$-split component $A_M$, that is,
\begin{align}\label{eq-S-M-defn}
S_M = \left( \{ x \in A_M : \theta(x) = x^{-1}\} \right)^\circ.
\end{align}
A parabolic subgroup $P$ of $G$ is called $\theta$-split if $\theta(P)$ is opposite to $P$. In this case, $M = P\cap \theta(P)$ is a $\theta$-stable Levi subgroup of both $P$ and $\theta(P)=P^{op}$.
Given a $\theta$-split subset $\Theta \subset \Delta_0$, the $\Delta_0$-standard parabolic subgroup $P_\Theta = M_\Theta N_\Theta$ is $\theta$-split.
Any $\Delta_0$-standard $\theta$-split parabolic subgroup arises from a $\theta$-split subset of $\Delta_0$ \cite[Lemma 2.5(1)]{kato--takano2008}.
Let $S_\Theta $ denote the $(\theta,F)$-split component of $M_\Theta $.
We have that
\begin{align}\label{eq-std-theta-split-torus}
S_\Theta = \left( \{ s \in A_\Theta : \theta(s) = s^{-1} \} \right)^\circ = \left( \bigcap_{ \bar\alpha \in p(\Theta)} \ker (\bar\alpha: S_0 \rightarrow F^\times ) \right)^\circ.
\end{align}
For the second equality in \eqref{eq-std-theta-split-torus}, see \cite[$\S1.5$]{kato--takano2010}.
For any $0 < \epsilon \leq 1$, define
\begin{align}\label{eq-split-dominant-part}
S_\Theta ^-(\epsilon) = \{ s \in S_\Theta : |\alpha(s)|_F \leq \epsilon, \ \text{for all} \ \alpha \in \Delta_0 \setminus \Theta \}.
\end{align}
We write $S_\Theta ^-$ for $S_\Theta ^-(1)$ and refer to $S_\Theta ^-$ as the dominant part of $S_\Theta$.
By \cite[Theorem 2.9]{helminck--helminck1998}, the subset $\Delta_0^\theta$ of $\theta$-fixed roots in $\Delta_0$ determines the $\Delta_0$-standard minimal $\theta$-split parabolic subgroup $P_0 = P_{\Delta_0^\theta}$. By \cite[Proposition 4.7(\rm{iv})]{helminck--wang1993}, the minimal $\theta$-split parabolic subgroup $P_0$ has standard $\theta$-stable Levi $M_0 = C_G(S_0)$.
Let $P_0 = M_0 N_0$ be the standard Levi factorization of $P_0$.
\begin{lem}\label{levi-theta-split}
Let $P$ be a $\theta$-split parabolic subgroup with $\theta$-stable Levi $M = P \cap \theta(P)$. The Levi subgroup $M$ is equal to the centralizer in $G$ of its $(\theta,F)$-split component $S_M$.
\end{lem}
\begin{proof}
The lemma follows immediately from \cite[Lemma 4.6]{helminck--wang1993}.
\end{proof}
\begin{lem}\label{levi-theta-split-2}
If $M$ is the centralizer in $G$ of a non-central $(\theta,F)$-split torus $S$, then $M$ is the Levi subgroup of a proper $\theta$-split parabolic subgroup of $G$.
\end{lem}
\begin{proof}
Let $S_0$ be a maximal $(\theta,F)$-split torus of $G$ containing $S$ and $A_0$ a $\theta$-stable maximal $F$-split torus of $G$ containing $S_0$.
The subgroup $M = C_G(S)$ is a $\theta$-stable Levi subgroup of $G$ since $S$ is a $\theta$-stable $F$-split torus.
Since $S$ is not central in $G$, $M$ is a proper Levi subgroup.
Moreover, since $S$ is contained in $S_0$, we have that $M_0$ is contained in $M$.
Let $P = MN_0$. Note that $P$ is a closed subgroup containing $P_0$; therefore, $P$ is a proper parabolic subgroup of $G$ with Levi subgroup $M$.
It remains to show that $P$ is $\theta$-split. Since $P_0$ is $\theta$-split, we have that $\theta(N_0) = N_0^{op}$ is the opposite unipotent radical of $N_0$.
Since $M$ is $\theta$-stable, it follows that $\theta(P) = MN_0^{op}$ and this is the parabolic opposite to $P$.
\end{proof}
The minimal $\theta$-split parabolic subgroups of $G$ are not always $H$-conjugate \cite[Example 4.12]{helminck--wang1993}. On the other hand, the following result holds.
\begin{lem}[{\cite[Lemma 2.5]{kato--takano2008}}]\label{KT08-lem-2.5}
Let $S_0 \subset A_0$, $\Delta_0$, and $P_0 = M_0N_0$ be as above.
\begin{enumerate}
\item Any $\theta$-split parabolic subgroup $P$ of $G$ is conjugate to a $\Delta_0$-standard $\theta$-split parabolic subgroup by an element $g \in (\Hbf \mathbf M_0)(F)$.
\item If the group of $F$-points of the product $(\Hbf \mathbf M_0)(F)$ is equal to $HM_0$, then any $\theta$-split parabolic subgroup of $G$ is $H$-conjugate to a $\Delta_0$-standard $\theta$-split parabolic subgroup.
\end{enumerate}
\end{lem}
Let $P = MN$ be a $\theta$-split parabolic subgroup and choose $g\in (\Hbf \mathbf M_0)(F)$ such that $P = gP_\Theta g^{-1}$ for some $\Delta_0$-standard $\theta$-split parabolic subgroup $P_\Theta $.
Since $g\in (\Hbf \mathbf M_0)(F)$ we have that $g^{-1}\theta(g) \in \mathbf M_0(F)$.
Thus, we may take $S_M = g S_\Theta g^{-1}$.
For a given $\epsilon >0$, one may extend the definition \eqref{eq-split-dominant-part} of $S_\Theta ^-$ to the non-$\Delta_0$-standard torus $S_M$.
Indeed, we may set
$S_M^-(\epsilon) = g S_\Theta ^-(\epsilon) g^{-1}$
and we define $S_M^- = S_M^-(1)$ with $S_M^1 = S_M(\of)$, as above.
We give the following definition, following the terminology of Murnaghan \cite{murnaghan2016-pp}, in analogy with the notion of an elliptic Levi subgroup.
\begin{defn}\label{theta-elliptic-defn}
A $\theta$-stable Levi subgroup $L$ of $G$ is $\theta$-elliptic if and only if $L$ is not contained in any proper $\theta$-split parabolic subgroup of $G$.
\end{defn}
We note the following simple lemma, which follows immediately from \Cref{theta-elliptic-defn}.
\begin{lem}\label{contain-theta-elliptic}
If a $\theta$-stable Levi subgroup $L$ of $G$ contains a $\theta$-elliptic Levi subgroup, then $L$ is $\theta$-elliptic.
\end{lem}
The following characterization of the $\theta$-elliptic property is also useful.
\begin{lem}\label{theta-elliptic-centre}
A $\theta$-stable Levi subgroup $L$ is $\theta$-elliptic if and only if $S_L = S_G$.
\end{lem}
\begin{proof}
If $L=G$, then there is nothing to do.
Without loss of generality, $L$ is a proper subgroup of $G$.
Suppose that $L$ is $\theta$-elliptic.
We have that $A_G$ is contained in $A_L$ and it follows that $S_G$ is contained in $S_L$.
If $S_L$ properly contains $S_G$, then $L$ is contained in the Levi subgroup $M = C_G(S_L)$.
By Lemma \ref{levi-theta-split-2}, $M$ is a Levi subgroup of a proper $\theta$-split parabolic subgroup.
It follows that $L \subset M$ is contained in a proper $\theta$-split parabolic subgroup. This contradicts the fact that $L$ is $\theta$-elliptic, so we must have that $S_L = S_G$.
On the other hand, suppose that $S_L$ is equal to $S_G$.
Argue by contradiction, and suppose that $L$ is contained in a proper $\theta$-split parabolic $P=MN$ with $\theta$-stable Levi subgroup $M=P\cap\theta(P)$.
We have that $L = \theta(L)$ is contained in $M$ and $S_L \subset S_M$. By Lemma \ref{levi-theta-split}, $M$ is the centralizer of its $(\theta,F)$-split component $S_M$.
Since $M$ is a proper Levi subgroup of $G$, we have that $S_M$ properly contains $S_G$.
However, by assumption $S_L=S_G$ is the largest $(\theta,F)$-split torus of contained $L$ and this is impossible.
We conclude that $L$ must be $\theta$-elliptic.
\end{proof}
The next proposition appears in \cite{murnaghan2016-pp}.
\begin{prop}\label{theta-elliptic-theta-stable}
Let $Q$ be a parabolic subgroup of $G$. If $Q$ admits a $\theta$-elliptic Levi factor $L$, then $Q$ is $\theta$-stable.
\end{prop}
\begin{proof}
The subgroup $L$ is $\theta$-stable by definition.
For any root $\alpha$ of $A_L$ in $G$, one can show that $\theta \alpha = \alpha$.
It follows that the unipotent radical of $Q$, hence $Q$, must be $\theta$-stable.
\end{proof}
\section{Distinguished representations and the Relative Casselman's Criterion}\label{sec-rep-theory}
\subsection{Distinguished representations}
All of the representations that we consider are on complex vector spaces.
A representation $(\pi,V)$ of $G$ is smooth if for every $v\in V$ the stabilizer of $v$ in $G$ is an open subgroup. A smooth representation $(\pi,V)$ of $G$ is admissible if, for every compact open subgroup $K$ of $G$, the subspace of $K$-invariant vectors $V^K$ is finite dimensional. A smooth one-dimensional representation of $G$ is a quasi-character of $G$. A character of $G$ is a unitary quasi-character. Let $(\pi,V)$ be a smooth representation of $G$. If $\omega$ is a quasi-character of $Z_G$, then $(\pi,V)$ is an $\omega$-representation if $\pi$ has central character $\omega$.
Let $\chi$ be a quasi-character of $H$. We also let $\pi$ denote its restriction to $H$.
\begin{defn}\label{defn-dist}
The representation $\pi$ is $(H,\chi)$-distinguished if the space $\Hom_H(\pi,\chi)$ is nonzero.
If $\chi$ is the trivial character of $H$, then we refer to $(H,1)$-distinguished representations simply as $H$-distinguished.
\end{defn}
Of course, in \Cref{defn-dist}, the subgroup $H = G^\theta$ may be replaced by any closed subgroup of $G$; however, we're only concerned with the symmetric subgroup setting.
As a first observation, we record the following lemma.
\begin{lem}\label{sub-quotient2}
Let $(\pi,V)$ be a finitely generated admissible representation of $G$. If $(\pi,V)$ is $H$-distinguished, then there exists an (irreducible) $H$-distinguished sub-quotient of $(\pi,V)$.
\end{lem}
The next lemma shows that distinction, relative to an involution $\theta$, depends only on the equivalence class of $\theta$ under the right action of $G$ on the set of involutions (\textit{cf.}~\Cref{defn-involution-action}).
\begin{lem}\label{orbit-dist}
The subgroup $G^{g\cdot \theta}$, of $g\cdot \theta$-fixed points in $G$, is $G$-conjugate to $G^\theta$. Precisely, we have $G^{g\cdot \theta} = g^{-1} (G^\theta) g$.
Moreover, a smooth representation $(\pi, V)$ of $G$ is $G^\theta$-distinguished if and only if $\pi$ is $G^{g\cdot \theta}$-distinguished.
\end{lem}
\begin{proof}
Let $h \in G^\theta$, then we have that
\begin{align*}
g\cdot \theta (g^{-1} h g) & = g^{-1} \theta( gg^{-1} h g g^{-1}) g
= g^{-1} \theta( h ) g
= g^{-1} h g
\end{align*}
so $g^{-1} h g$ is $g\cdot\theta$-fixed. It follows that $g^{-1} (G^\theta) g \subset G^{g\cdot \theta}$.
Since conjugation by $g$ is an automorphism of $G$, it follows that $G^{g\cdot \theta} = g^{-1}( G^\theta )g$.
Let $\lambda$ be a nonzero element of $\Hom_{G^\theta}(\pi,1)$. Define $\lambda ' = \lambda \circ \pi(g)$ and observe that $\lambda'$ is a nonzero $G^{g\cdot \theta}$-invariant linear functional on $V$.
It follows that the map $\lambda \mapsto \lambda \circ \pi(g)$ is a bijection from $\Hom_{G^\theta}(\pi,1)$ to $\Hom_{G^{g\cdot \theta}}(\pi,1)$ with inverse $\lambda ' \mapsto \lambda' \circ \pi(g^{-1})$.
In particular, $\pi$ is ${G^\theta}$-distinguished if and only if $\pi$ is $G^{g\cdot \theta}$-distinguished.
\end{proof}
\subsection{Relative matrix coefficients}\label{sec-matrix-coeff}
Let $(\pi,V)$ be a smooth $H$-distinguished representation of $G$.
Let $\lambda \in \Hom_H(\pi,1)$ be a nonzero $H$-invariant linear form on $V$ and let $v$ be a nonzero vector in $V$.
In analogy with the usual matrix coefficients, define a complex-valued function $\varphi_{\lambda,v}$ on $G$ by
$\varphi_{\lambda,v}(g) = \ip{\lambda}{\pi(g)v}$.
We refer to the functions $\varphi_{\lambda,v}$ as relative matrix coefficients (with respect to $\lambda$) or as $\lambda$-relative matrix coefficients.
Since $\pi$ is a smooth representation, the relative matrix coefficient $\varphi_{\lambda,v}$ lies in $C^\infty(G)$, for every $v\in V$.
In addition, since $\lambda$ is $H$-invariant, the functions $\varphi_{\lambda,v}$ descend to well-defined functions on the quotient $H\backslash G$.
In analogy with the classical case, one makes the following definitions.
\begin{defn}\label{defn-rsc}
The representation $(\pi,V)$ is said to be
\begin{enumerate}
\item $(H,\lambda)$-relatively supercuspidal, or relatively supercuspidal with respect to $\lambda$, if and only if all of the $\lambda$-relative matrix coefficients are compactly supported modulo $Z_GH$.
\item $H$-relatively supercuspidal if and only if $\pi$ is $(H,\lambda)$-relatively supercuspidal for every $\lambda \in \Hom_H(\pi,1)$.
\end{enumerate}
\end{defn}
Let $\omega$ be a unitary character of $Z_G$ and further suppose that $\pi$ is an $\omega$-representation.
\begin{defn}
The representation $(\pi,V)$ is said to be
\begin{enumerate}
\item $(H,\lambda)$-relatively square integrable, or relatively square integrable with respect to $\lambda$, if and only if all of the $\lambda$-relative matrix coefficients are square integrable modulo $Z_GH$.
\item $H$-relatively square integrable if and only if $\pi$ is $(H,\lambda)$-relatively square integrable for every $\lambda \in \Hom_H(\pi,1)$.
\end{enumerate}
\end{defn}
Notice that we must also take the quotient of $G$ by the (noncompact) centre $Z_G$ in order to make sense of compactly supported (respectively, square integrable) functions on $H \backslash G$. Moreover, to integrate relative matrix coefficients over $Z_GH\backslash G$ we need a $G$-invariant measure on the quotient $Z_GH \backslash G$. The centre $Z_G$ of $G$ is unimodular since it is abelian.
The fixed point subgroup $H$ is also reductive (\textit{cf.}~\cite[Theorem 1.8]{digne--michel1994}) and thus unimodular.
It follows that there exists a $G$-invariant measure on the quotient $Z_GH \backslash G$ by \cite[Proposition 12.8]{Robert-book}.
\begin{note}
When $H$ is understood, we refer to $H$-relatively supercuspidal (respectively, $H$-relatively square integrable) representations simply as relatively supercuspidal (respectively, relatively square integrable).
\end{note}
\begin{defn}
If $(\pi,V)$ is an irreducible subrepresentation of $L^2(Z_GH\backslash G)$, then we say that $(\pi,V)$ occurs in the discrete spectrum of $H\backslash G$.
In this case, we say that $(\pi,V)$ is a relative discrete series (RDS) representation.
\end{defn}
\subsection{Parabolic induction and Jacquet restriction}
Let $P$ be a parabolic subgroup of $G$ with Levi subgroup $M$ and unipotent radical $N$.
Given a smooth representation $(\rho, V_\rho)$ of $M$ we may inflate $\rho$ to a representation of $P$, also denoted $\rho$, by declaring that $N$ acts trivially.
We define the representation $\iota_P^G \rho$ of $G$ to be the induced representation $\ind_P^G (\delta_P^{1/2} \otimes \rho)$.
We refer to the functor $\rho \mapsto \iota_P^G\rho$ as (normalized) parabolic induction. When it is more convenient (\textit{cf.}~\Cref{obs-isom-type-RDS}, $\S$\ref{sec-inducing-data}-\ref{sec-exp-dist-pi-N}), we also use the Bernstein--Zelevinsky notation for parabolic induction on general linear groups \cite{bernstein--zelevinsky1977, zelevinsky1980}.
Let $(\pi,V)$ be a smooth representation of $G$. Let $(\pi_N, V_N)$ denote the Jacquet module of $\pi$ along $P$, normalized by $\delta_P^{-1/2}$.
Explicitly, if $V(N) = \spn\{ \pi(n)v - v : n\in N , v\in V\}$, then $V_N = V / V(N)$ and $\pi_N(p)(v+V(N)) = \delta_P^{-1/2}(p) \pi(p) v + V(N)$,
for all $p\in P$, $v+V(N) \in V_N$.
Since $\delta_P$ is trivial on $N$, we see that $(\pi_N,V_N)$ is a representation of $P$ on which $N$ acts trivially.
We will regard $(\pi_N, V_N)$ as a representation of the Levi factor $M \cong P/ N$ of $P$.
We will now give a statement of the Geometric Lemma \cite[Lemma 2.12]{bernstein--zelevinsky1977}, which is a fundamental tool in our work and the study of induced representations in general. First, we recall two results on double-coset representatives.
\begin{lem}[{\cite[Proposition 1.3.1]{Casselman-book}}]\label{lem-nice-reps}
Let $\Theta$ and $\Omega$ be subsets of $\Delta_0$.
The set
\begin{align*
[W_\Theta \backslash W_0 /W_\Omega] = \{ w \in W_0 : w\Omega, w^{-1}\Theta \subset \Phi^+ \}
\end{align*}
provides a choice of Weyl group representatives for the double-coset space $P_\Theta \backslash G / P_\Omega$.
\end{lem}
\begin{prop}[{\cite[Proposition 1.3.3]{Casselman-book}}]\label{casselman1-3-3}
Let $\Theta,\Omega \subset \Delta_0$ and let $w \in [W_\Theta \backslash W_0 /W_\Omega]$.
\begin{enumerate}
\item The standard parabolic subgroup $P_{\Theta \cap w \Omega}$ is equal to $\left(P_\Theta \cap {}^w P_\Omega \right){}^wN_\Omega$.
\item The unipotent radical of $P_{\Theta \cap w \Omega}$ is generated by ${}^wN_\Omega$ and $N_\Theta \cap {}^w N_{\emptyset}$, where $N_{\emptyset}$ is the unipotent radical of the minimal parabolic subgroup corresponding to $\emptyset \subset \Delta_0$.
\item The standard Levi subgroup of $P_{\Theta \cap w \Omega}$ is $M_{\Theta \cap w \Omega} = M_\Theta \cap w M_\Omega w^{-1}$.
\item The subgroup $P_\Theta \cap {}^wM_\Omega$ is a $w\Omega$-standard parabolic in $M_{w\Omega}={}^wM_\Omega$ with unipotent radical $N_\Theta \cap {}^wM_\Omega$ and standard Levi subgroup $M_{\Theta \cap w \Omega}=M_\Theta \cap {}^w M_\Omega$.
\end{enumerate}
\end{prop}
When applying the Geometric Lemma along two standard parabolic subgroups $P_\Theta$ and $P_\Omega$, associated to $\Theta,\Omega \subset \Delta_0$, we will always use the choice of ``nice" representatives $[W_\Theta \backslash W_0 / W_\Omega]$ for the double-coset space $P_\Theta \backslash G / P_\Omega$.
\begin{lem}[The Geometric Lemma]\label{geom-lem}
Let $P_\Omega$ and $P_\Theta$ be two $\Delta_0$-standard parabolic subgroups of $G$.
Let $\rho$ be a smooth representation of $M_\Omega$.
There is a filtration of the space of the representation $(\iota_{P_\Omega}^G \rho)_{N_\Theta}$ such that the associated graded object is isomorphic to the direct sum
\begin{align*
\bigoplus_{w \in [W_\Theta \backslash W_0 / W_\Omega]} \iota_{M_{\Theta}\cap {}^wP_\Omega}^{M_\Theta} \left( ({}^w\rho)_{N_\Theta \cap{}^w M_\Omega} \right).
\end{align*}
\end{lem}
We write $\mathcal F_{\Theta}^w(\rho) = \mathcal F_{N_\Theta}^w(\rho)$ to denote the representation $\iota_{M_{\Theta}\cap {}^wP_\Omega}^{M_\Theta} \left( ({}^w\rho)_{N_\Theta \cap{}^w M_\Omega} \right)$ of $M_\Theta$.
\subsection{Distinction of induced representations}
\Cref{lem-hom-injects} is well known and follows from an explicit version of Frobenius Reciprocity due to Bernstein and Zelevinsky \cite[Proposition 2.29]{bernstein--zelevinsky1976}. Let $Q = LU$ be a $\theta$-stable parabolic subgroup with $\theta$-stable Levi factor $L$ and unipotent radical $U$. Note that the identity component of $Q^\theta = L^\theta U^\theta$ is a parabolic subgroup of $H^\circ$, with the expected Levi decomposition (\textit{cf.}~\cite{helminck--wang1993}, \cite[Lemma 3.1]{gurevich--offen2015}). Let $\mu$ be a positive quasi-invariant measure on the (compact) quotient $Q^\theta \backslash H$ \cite[Theorem 1.21]{bernstein--zelevinsky1976}.
\begin{lem}\label{lem-hom-injects}
Let $\rho$ be a smooth representation of $L$ and let $\pi = \iota_Q^G \rho$.
The map $\lambda \mapsto \lambda^G$ is an injection of $\Hom_{L^\theta}( \delta_Q^{1/2}\rho, \delta_{Q^\theta})$ into $\Hom_H(\pi,1)$, where $\lambda^G$ is given explicitly by
\begin{align*
\ip{\lambda^G}{\phi} &= \int_{Q^\theta \backslash H} \ip{\lambda}{\phi(h)} \ d\mu(h)
\end{align*}
for any function $\phi$ in the space of $\pi$.
\end{lem}
\begin{cor}\label{cor-hom-injects}
If $\delta_Q^{1/2}$ restricted to $L^\theta$ is equal to $\delta_{Q^\theta}$, then map $\lambda \mapsto \lambda^G$ is an injection of $\Hom_{L^\theta}(\rho, 1)$ into $\Hom_H(\pi,1)$. In particular, if $\rho$ is $L^\theta$-distinguished, then $\pi$ is $H$-distinguished.
\end{cor}
\begin{proof}
Observe that $\Hom_{L^\theta}(\delta_Q^{1/2}\rho, \delta_{Q^\theta}) = \Hom_{L^\theta}(\rho, \delta_Q^{-1/2}\vert_{L^\theta}\delta_{Q^\theta})$.
\end{proof}
Alternatively, the $H$-invariant linear form on $\pi = \iota_Q^G \rho$ may be understood to arise from the closed orbit in $Q \backslash G / H$ via the Mackey theory.
\subsection{Invariant linear forms on Jacquet modules}\label{sec-r-P-lambda}
Let $(\pi,V)$ be an admissible $H$-distinguished representation of $G$. Let $\lambda$ be a nonzero element of $\Hom_H(\pi,1)$. Let $P$ be a $\theta$-split parabolic subgroup of $G$ with unipotent radical $N$ and $\theta$-stable Levi component $M= P \cap \theta(P)$.
One may associate to $\lambda$ a canonical $M^\theta$-invariant linear form $r_P\lambda$ on the Jacquet module $(\pi_N, V_N)$.
The construction of $r_P\lambda$, via Casselman's Canonical Lifting \cite[Proposition 4.1.4]{Casselman-book}, was discovered independently by Kato--Takano and Lagier.
We refer the reader to \cite{kato--takano2008, lagier2008} for the details of the construction. We record the following result (\textit{cf.}~\cite[Proposition 5.6]{kato--takano2008}).
\begin{prop}[Kato--Takano, Lagier]\label{rPlambda-prop}
Let $(\pi,V)$ be an admissible $H$-distinguished representation of $G$.
Let $\lambda \in \Hom_H(\pi,1)$ be nonzero and let $P$ be a $\theta$-split parabolic subgroup of $G$ with unipotent radical $N$ and $\theta$-stable Levi component $M= P \cap \theta(P)$.
\begin{enumerate}
\item The linear functional $r_P\lambda: V_N \rightarrow \C$ is $M^\theta$-invariant.
\item The mapping $r_P: \Hom_H(\pi,1) \rightarrow \Hom_{M^\theta}(\pi_N, 1)$, sending $\lambda$ to $r_P\lambda$, is linear.
\end{enumerate}
\end{prop}
Kato and Takano use the invariant forms $r_P\lambda$ to provide the following characterization of relatively supercuspidal representations \cite[Theorem 6.2]{kato--takano2008}.
\begin{thm}[Kato--Takano]\label{kt08-rsc-thm}
Let $(\pi,V)$ be an admissible $H$-distinguished representation of $G$ and let $\lambda$ be a nonzero $H$-invariant linear form on $V$. Then, $(\pi,V)$ is $(H,\lambda)$-relatively supercuspidal if and only if $r_P\lambda = 0$ for every proper $\theta$-split parabolic subgroup $P$ of $G$.
\end{thm}
\subsection{Exponents and the Relative Casselman's Criterion}\label{sec-exp-and-rel-cass}
Let $(\pi, V)$ be a finitely generated admissible representation of $G$. Recall that $A_G$ denotes the $F$-split component of the centre of $G$.
Let $\chi$ be a quasi-character of $A_G$. For $n\in \N$, $n\geq 1$, define the subspace
\begin{align*
V_{\chi, n} = \{ v \in V : (\pi(z) - \chi(z))^n v = 0 \ \text{for all} \ z\in A_G \},
\end{align*}
and set
\begin{align*
V_{\chi, \infty} = \bigcup_{n=1}^\infty V_{\chi, n}.
\end{align*}
Each $V_{\chi, n}$ is a $G$-stable subspace of $V$ and $V_{\chi, \infty}$ is the generalized eigenspace in $V$ for the $A_G$-action on $V$ by the eigencharacter $\chi$.
By \cite[Proposition 2.1.9]{Casselman-book}, we have that
\begin{enumerate}
\item $V$ is a direct sum $\displaystyle V = \bigoplus_{\chi} V_{\chi,\infty}$, where $\chi$ ranges over quasi-characters of $A_G$, and
\item since $V$ is finitely generated, there are only finitely many $\chi$ such that $V_{\chi,\infty} \neq 0$. Moreover, there exists $n\in \N$ such that $V_{\chi,\infty} = V_{\chi,n}$, for each $\chi$.
\end{enumerate}
Let $\Exp_{A_G}(\pi)$ be the (finite) set of quasi-characters of $A_G$ such that $V_{\chi,\infty} \neq 0$. The quasi-characters that appear in $\Exp_{A_G}(\pi)$ are called the exponents of $\pi$.
The second item above implies that $V$ has a finite filtration such that the quotients are $\chi$-representations, for $\chi \in \Exp_{A_G}(\pi)$.
From this last observation, we obtain the following lemma.
\begin{lem}\label{exp-irred-subq}
The characters $\chi$ of $A_G$ that appear in $\Exp_{A_G}(\pi)$ are precisely the central quasi-characters of the irreducible subquotients of $\pi$.
\end{lem}
Note that he same analysis as above can be carried out for any closed subgroup $Z$ of $Z_G$, i.e., we can consider the generalized $Z$-eigenspaces in $V$. Moreover, we have the following.
\begin{lem}\label{closed-subgp-centre}
Let $Z_1 \supset Z_2$ be two closed subgroups of the centre $Z_G$ of $G$. The map of exponents $\Exp_{Z_1}(\pi) \rightarrow \Exp_{Z_2}(\pi)$ defined by restriction of quasi-characters is surjective.
\end{lem}
\begin{proof}
Let $\chi \in \Exp_{Z_2}(\pi)$. By assumption, there exists a nonzero vector $v \in V_{\chi, \infty}$.
In particular, there is an irreducible subquotient of $V_{\chi, \infty}$, hence of $(\pi,V)$, where $Z_2$ acts by the character $\chi$. On this irreducible subquotient, by Schur's Lemma, the subgroup $Z_1$ must act by some extension $\widehat\chi$ of $\chi$.
By \Cref{exp-irred-subq}, $\widehat\chi$ must occur in $\Exp_{Z_1}(\pi)$.
\end{proof}
For our purposes, we're interested in the exponents of parabolically induced representations.
\begin{lem}\label{red-to-ind-exp}
Let $P = MN$ be a parabolic subgroup of $G$, let $(\rho, V_\rho)$ be an finitely generated admissible representation of $M$ and let $\pi = \iota_P^G\rho$.
The quasi-characters $\chi \in \Exp_{A_G}(\pi)$ are the restriction to $A_G$ of characters $\mu$ of $A_M$ appearing in $\Exp_{A_M}(\rho)$.
\end{lem}
\begin{proof}
Without loss of generality, assume that $P=MN$ is a proper parabolic subgroup of $G$.
Given $a\in A_G$, we have that $\delta_P(a) = 1$, since $a$ is central in $G$. It follows that for any $f\in V$ we have
\begin{align}\label{action-A_G}
\left(\pi(a)f \right)(g) & = f(ga)
= f(ag)
= \delta_P^{1/2}(a) \rho(a) f(g)
= \rho(a) f(g),
\end{align}
for all $a\in A_G$ and $g\in G$.
Suppose that $\chi \in \Exp_{A_G}(\pi)$ and $f\in V_{\chi,\infty}$ is nonzero. Fix $g_0 \in G$ such that $w_0 = f(g_0)$ is nonzero.
There exists $n\in \N$, $n\geq 1$ such that $f\in V_{\chi,n}$.
More precisely, $(\pi(a) - \chi(a))^n f = 0_V$, for all $a\in A_G$, where $0_V: G \rightarrow V_\rho$ is the zero function.
By induction and using \eqref{action-A_G}, we see that
\begin{align*}
0 & = [(\pi(a) - \chi(a))^n f](g_0) = (\rho(a) - \chi(a))^n (f(g_0)) = (\rho(a) - \chi(a))^n w_0,
\end{align*}
for any $a\in A_G$.
That is, $w_0 \in (V_\rho)_{\chi,\infty}$ and $(V_\rho)_{\chi,\infty}$ is nonzero; moreover, $\chi \in \Exp_{A_G}(\rho)$.
By \Cref{closed-subgp-centre}, the map $\Exp_{A_M}(\rho) \rightarrow \Exp_{A_G}(\rho)$ defined by restriction is surjective.
In particular, there exists $\mu \in \Exp_{A_M}(\rho)$ such that $\chi$ is equal to the restriction of $\mu$ to $A_G$.
\end{proof}
Let $(\pi,V)$ be a finitely generated admissible representation of $G$. Let $P=MN$ be a parabolic subgroup of $G$ with Levi factor $M$ and unipotent radical $N$.
It is a theorem of Jacquet that $(\pi_N, V_N)$ is also finitely generated and admissible (\textit{cf.}~\cite[Theorem 3.3.1]{Casselman-book}). Applying \cite[Proposition 2.1.9]{Casselman-book}, we obtain a direct sum decomposition
\begin{align*}
V_N = & \bigoplus_{\chi \in \Exp_{A_M}(\pi_N)} (V_N)_{\chi,\infty}
\end{align*}
where the set $\Exp_{A_M}(\pi_N)$ of quasi-characters of $A_M$, such that $(V_N)_{\chi,\infty} \neq 0$, is finite.
The quasi-characters of $A_M$ appearing in $\Exp_{A_M}(\pi_N)$ are called the exponents of $\pi$ along $P$.
Suppose, in addition, that $(\pi,V)$ is $H$-distinguished. Fix a nonzero $H$-invariant form $\lambda$ on $V$.
For any closed subgroup $Z$ of the centre of $G$, Kato and Takano \cite{kato--takano2010} define
\begin{align}\label{relative-exponent}
\Exp_Z(\pi,\lambda) & = \{ \chi \in \Exp_Z(\pi) : \lambda \vert_{V_{\chi,\infty}} \neq 0 \},
\end{align}
and refer to the set $\Exp_Z(\pi,\lambda)$ as exponents of $\pi$ relative to $\lambda$.
The next result is \cite[Theorem 4.7]{kato--takano2010}, which is a key ingredient used in the proof of our main result.
\begin{thm}[The Relative Casselman's Criterion, Kato--Takano]\label{rel-casselman-crit}
Let $\omega$ be a unitary character of $Z_G$. Let $(\pi,V)$ be a finitely generated admissible $H$-distinguished $\omega$-representation of $G$.
Fix a nonzero $H$-invariant linear form $\lambda$ on $V$. The representation $(\pi,V)$ is $(H,\lambda)$-relatively square integrable
if and only if the condition
\begin{align}\label{rel-casselman}
|\chi(s)| &< 1 & \text{for all} \ \chi \in \Exp_{S_M}(\pi_N, r_P\lambda) \ \text{and all} \ s\in S_M^- \setminus S_GS_M^1
\end{align}
is satisfied for every proper $\theta$-split parabolic subgroup $P=MN$ of $G$.
\end{thm}
\begin{rmk}
Note that the Relative Casselman's Criterion reduces to Casselman's Criterion in the group case: $G = G' \times G'$ and $H = \Delta G \cong G'$ is the diagonal subgroup.
\end{rmk}
\begin{cor}[Kato--Takano]
If $(\pi,V)$ is an $H$-distinguished discrete series representation of $G$, then $\pi$ is $H$-relatively square integrable.
\end{cor}
The next two lemmas let us prove \Cref{reduction-standard-split}, which allows us to reduce to checking the Relative Casselman's Criterion along maximal $\Delta_0$-standard parabolic subgroups (under an additional assumption).
\begin{lem}\label{exponent-bijection}
Let $P = MN$ be a proper ${\theta}$-split parabolic subgroup of $G$. Assume that $P$ is $H$-conjugate to a $\Delta_0$-standard $\theta$-split parabolic subgroup $P_\Theta$.
If $P = hP_\Theta h^{-1}$, where $h\in H$, then there is a bijection
\begin{align*}
\Exp_{S_\Theta}(\pi_{N_\Theta}, r_{P_\Theta} \lambda) & \longleftrightarrow \Exp_{S}(\pi_{N}, r_{P} \lambda) \\
\chi' & \mapsto {}^h\chi',
\end{align*}
with inverse given by $\chi \mapsto {}^{h^{-1}}\chi$.
\end{lem}
\begin{proof}
The bijection between $\Exp_{S_\Theta}(\pi_{N_\Theta})$ and $\Exp_S(\pi_N)$ is automatic from the equality $S = h S_\Theta h^{-1}$ (and holds for $h \in (\Hbf \mathbf{M}_0)(F)$). Using that $h\in H$ and $H$-invariance of $\lambda$, one can show that: If $r_{P_\Theta}\lambda$ is nonzero on $(V_{N_\Theta})_{\chi', \infty}$, then $r_P\lambda$ is nonzero on $(V_N)_{\chi,\infty}$, where $\chi = {}^h\chi'$.
\end{proof}
\begin{lem}\label{lem-red-max-std}
Assume that any ${\theta}$-spilt parabolic subgroup $P$ of $G$ is $H$-conjugate to a $\Delta_0$-standard $\theta$-split parabolic. If condition \eqref{rel-casselman} holds for all $\Delta_0$-standard ${\theta}$-split parabolic subgroups of $G$, then the condition \eqref{rel-casselman} holds for all ${\theta}$-split parabolic subgroups of $G$.
\end{lem}
\begin{proof}
Let $P = MN$ be a proper ${\theta}$-split parabolic subgroup of $G$. By assumption, there exists $h\in H$ and a ${\theta}$-split subset $\Theta \subset {\Delta_0}$ such that $P = h P_\Theta h^{-1}$. In particular, the $({\theta},F)$-split component $S$ of $P$ is equal to $hS_\Theta h^{-1}$; moreover, $S^- = h S_\Theta^- h^{-1}$. Let $\chi \in \Exp_{S}(\pi_{N}, r_{P} \lambda)$, by \Cref{exponent-bijection}, there exists $\chi ' ={}^{h^{-1}}\chi \in \Exp_{S_\Theta}(\pi_{N_\Theta}, r_{P_\Theta} \lambda)$.
Let $s \in S^- \setminus S^1 S_{\Delta_0}$, then $s' = h^{-1} s h \in S_\Theta^- \setminus S_\Theta^1 S_{\Delta_0}$. It follows that
\begin{align*}
|\chi(s)| & = |\chi(h s' h^{-1})| = |\chi' (s')| <1,
\end{align*}
where the final inequality holds by the assumption that \eqref{rel-casselman} holds for $P_\Theta$.
\end{proof}
\begin{prop}\label{reduction-standard-split}
Let $\pi$ be an $H$-distinguished representation of $G$ and let $\lambda$ be a nonzero $H$-invariant linear form on the space of $\pi$.
Assume that any ${\theta}$-spilt parabolic subgroup $P$ of $G$ is $H$-conjugate to a $\Delta_0$-standard $\theta$-split parabolic.
Then $\pi$ is $(H,\lambda)$-relatively square integrable if and only if the condition \eqref{rel-casselman} holds for all $\Delta_0$-standard maximal ${\theta}$-split parabolic subgroups of $G$.
\end{prop}
\begin{proof}
Apply \Cref{lem-red-max-std}, \cite[Lemma 4.6]{kato--takano2010} and \Cref{rel-casselman-crit}.
\end{proof}
The next result is key in our application of \Cref{rel-casselman-crit}. It allows us to ignore ``bad" exponents relative to $\lambda$, as long as the appropriate subquotients are not distinguished.
\begin{prop}\label{non-dist-gen-eig-sp}
Let $(\pi, V)$ be a finitely generated admissible representation of $G$. Let $\chi \in \Exp_{Z_G}(\pi)$ and assume that none of the irreducible subquotients of $(\pi,V)$ with central character $\chi$ are $H$-distinguished. Then for any $\lambda \in \Hom_H(\pi,1)$, the restriction of $\lambda$ to $V_{\chi,\infty}$ is equal to zero, i.e., $\lambda \vert_{V_{\chi,\infty}} \equiv 0$.
\end{prop}
\begin{proof}
Suppose, by way of contradiction, that $\lambda \vert_{V_{\chi,\infty}}\neq 0$. Then $(\pi\vert_{V_{\chi,\infty}}, V_{\chi,\infty})$ is an admissible finitely generated $H$-distinguished representation of $G$. By \Cref{sub-quotient2}, some irreducible subquotient $(\rho, {V_\rho})$ of $V_{\chi,\infty}$ must be $H$-distinguished. However, the representation $(\rho, {V_\rho})$ is also an irreducible subquotient of $(\pi,V)$ and has central character $\chi$. By assumption, no such $(\rho, {V_\rho})$ can be $H$-distinguished; therefore, we must have that $\lambda \vert_{V_{\chi,\infty}}$ is identically zero.
\end{proof}
\section{Tori and parabolic subgroups: The linear and Galois cases}\label{sec-structure-lin-Gal}
\begin{rmk}
For the remainder of the paper we work in the linear and Galois cases.
Refer to Sections \ref{sec-notation-lin} and \ref{sec-notation-Gal} for notation.
\end{rmk}
\subsection{Tori and root systems relative to $\theta$}\label{sec-tori-roots}
In the linear case, let $A_0$ be the diagonal maximal $F$-split torus of $G$.
Note that $A_0$ is $\theta$-stable.
Let $S_0$ be the $(\theta,F)$-split component of $A_0$.
It is straightforward to check that
\begin{align*
S_0 &= \{\diag(a_1,\ldots, a_{\frac{n}{2}}, a_{\frac{n}{2}}^{-1}, \ldots, a_1^{-1}) : a_i \in F^\times, 1\leq i \leq \textstyle\frac{n}{2}\}.
\end{align*}
Moreover, $S_0$ is a maximal $(\theta,F)$-split torus of $G$.
Indeed, it is readily verified that the upper-triangular Borel subgroup of $G$ is a minimal $\theta$-split parabolic subgroup with Levi subgroup $A_0$. It follows from \cite[Proposition 4.7(\rm{iv})]{helminck--wang1993} that $S_0$ is a maximal $(\theta,F)$-split torus of $G$ contained in $A_0$.
In the Galois case, the torus $\mathbf T$ obtained as the restriction of scalars of the diagonal torus of $\GL_n$ is a maximal non-split $F$-torus of $\G$.
We identify $T = \mathbf{T}(F)$ with the diagonal matrices in $\GL_n(E)$.
Define $T_0 = {}^\gamma T$, where $\gamma$ is described in $\S$\ref{sec-choice-elts}. Then $A_0={}^\gamma A_T$ is the $F$-split component of $T_0$.
The tori $T$, $A_T$, $T_0$ and $A_0$ are all ${\theta}$-stable.
As above, and using \eqref{Galois-involution-relation}, it is readily verified that
\begin{align*
S_0 = \{ {}^\gamma \diag(a_1,\ldots, a_{\floor{\frac{n}{2}}}, \widehat 1, a_{\floor{\frac{n}{2}}}^{-1}, \ldots, a_1^{-1} ) : a_i \in F^\times, 1\leq i \leq \floor{\textstyle\frac{n}{2}}\},
\end{align*}
is a maximal $(\theta,F)$-split torus of $G$ contained in $A_0$.
In both cases, let $\Phi_0 = \Phi(G,A_0)$ be the set of roots of $G$ relative to $A_0$.
Explicitly, in the linear case, we have
\begin{align*
\Phi_0 = \{ \epsilon_i - \epsilon_j : 1\leq i\neq j \leq n\},
\end{align*}
where $\epsilon_i \in X^*(A_0)$ is the $i$\textsuperscript{th} coordinate ($F$-rational) character of $A_0$.
Let
\begin{align*
\Delta_0 = \{\epsilon_i - \epsilon_{i+1} : 1 \leq i \leq n-1\}
\end{align*}
be the standard base of $\Phi_0$.
The set $\Phi_0^+$ of positive roots (determined by $\Delta_0$) is
\begin{align*}
\Phi_0^+ = \{ \epsilon_i - \epsilon_j : 1\leq i < j \leq n\}.
\end{align*}
In the Galois case, we relate $\Phi_0$ to another collection of roots, those relative to $A_T$.
Let $\Phi = \Phi(G, A_T)$ be the root system of $G$ with respect to $A_T$ with standard base $\Delta$.
We observe that $\Phi_0 = {}^\gamma \Phi$, where given a root $\beta \in \Phi$ we have
\begin{align*}
({}^\gamma\beta)(a) = \beta ({}^{\gamma^{-1}}a) = \beta (\gamma^{-1} a \gamma),
\end{align*}
for $a\in A_0$. Moreover, $\Delta_0 = {}^\gamma \Delta$ is a base for $\Phi_0$ and it is clear that
\begin{align*
{\Phi}_0^+ = \{ {}^\gamma (\epsilon_i - \epsilon_j) : 1\leq i < j \leq n\},
\end{align*}
where, as above, $\epsilon_i$ is the $i$\textsuperscript{th}-coordinate ($F$-rational) character of the diagonal $F$-split torus $A_T$.
It is elementary to verify the following.
\begin{lem}\label{lem-Delta-0-theta-base}
The set of simple roots $\Delta_0$ of $\Phi_0$ is a ${\theta}$-base for $\Phi_0$. In addition, the subset of ${\theta}$-fixed roots in $\Phi_0$ is empty.
\end{lem}
\begin{cor}\label{cor-min-std-split-pblc}
The Borel subgroup $P_\emptyset = P_0 = M_0 N_0$ corresponding to $\emptyset \subset \Delta_0$ is a minimal $\theta$-split parabolic subgroup of $G$.
\end{cor}
\begin{proof}
The subset $\Delta_0^\theta$ is a minimal $\theta$-split subset of $\overline{\Delta}_0$; therefore, the parabolic $P_{\Delta_0^\theta}$ is a minimal standard $\theta$-split parabolic subgroup \cite{kato--takano2010}.
Since $\Delta_0^\theta=\emptyset$, we have $P_{\Delta_0^\theta} = P_\emptyset = P_0$. In the linear case, $M_0 = A_0$ and in the Galois case $M_0 = C_G(A_0) = T_0$.
\end{proof}
Following $\S$\ref{sec-tori-involution}, since $\Delta_0^\theta = \emptyset$, the restricted root system is just the image of $\Phi_0$ under the restriction map
$p: X^*(A_0) \rightarrow X^*(S_0)$. That is, we have $\overline \Phi_0 = p(\Phi_0)$ and $\overline \Delta_0 = p(\Delta_0)$.
Explicitly, in the linear case,
\begin{align*
\overline \Delta_0 = \left\{\bar\epsilon_i - \bar\epsilon_{i+1} : 1\leq i \leq \textstyle\frac{n}{2}-1\right\} \cup \left\{2\bar\epsilon_{\frac{n}{2}}\right\},
\end{align*}
where $\bar\epsilon_i \in X^*(S_0)$ is the $i$\textsuperscript{th} coordinate character of $S_0$ given by
\begin{align*}
\bar\epsilon_i (\diag(a_1,\ldots,a_{\frac{n}{2}},a_{\frac{n}{2}}^{-1},\ldots, a_1^{-1}))= a_i.
\end{align*}
Similarly in the Galois case, we have that
\begin{align*
\overline \Delta_0= \left\{{}^\gamma\bar\epsilon_i - {}^\gamma\bar\epsilon_{i+1} : 1\leq i \leq \floor{\textstyle\frac{n}{2}}-1\right\} \cup \{\bar\alpha\},
\end{align*}
where $\bar \alpha = {}^\gamma\bar\epsilon_{\floor{\frac{n}{2}}}$ when $n$ is odd, and $\bar \alpha = 2{}^\gamma\bar\epsilon_{\frac{n}{2}}$ when $n$ is even.
The following result is now an immediate consequence of \Cref{lem-Delta-0-theta-base}.
\begin{lem}\label{lem-max-split-subsets}
For $1\leq k \leq \floor{\frac{n}{2}}$, let $\Theta _k$ denote the $\floor{\frac{n}{2}}$ maximal $\theta$-split subsets of $\Delta_0$.
In the linear case, for $1\leq k \leq n/2-1$
\begin{align*
\Theta _k & = [\overline \Delta_0 \setminus \{\bar\epsilon_k - \bar\epsilon_{k+1}\}]
= \Delta_0 \setminus \{ \epsilon_k - \epsilon_{k+1}, \epsilon_{n-k} - \epsilon_{n-k+1}\}
\end{align*}
and
\begin{align*
\Theta _{\frac{n}{2}} & = [\overline \Delta_0 \setminus \left\{2\bar\epsilon_{\frac{n}{2}} \right\}]
= \Delta_0 \setminus \left\{\epsilon_{\frac{n}{2}} - \epsilon_{\frac{n}{2}+1}\right\}.
\end{align*}
Respectively, in the Galois case, for $1\leq k \leq \floor{\frac{n}{2}}-1$
\begin{align*
\Theta _k & = [\overline \Delta_0\setminus \{{}^\gamma\bar\epsilon_k - {}^\gamma\bar\epsilon_{k+1}\}]
= {\Delta_0} \setminus \{ {}^\gamma(\epsilon_k - \epsilon_{k+1}), {}^\gamma(\epsilon_{n-k} - \epsilon_{n-k+1})\}
\end{align*}
and
\begin{align*
\Theta _{\floor{\frac{n}{2}}} & = [\overline \Delta_0\setminus \{\bar\alpha \}]
= {\Delta_0} \setminus p^{-1}\{\bar\alpha\},
\end{align*}
where $\bar \alpha = {}^\gamma\bar\epsilon_{\floor{\frac{n}{2}}}$ when $n$ is odd, and $\bar \alpha = 2{}^\gamma\bar\epsilon_{\frac{n}{2}}$ when $n$ is even.
\end{lem}
\begin{note}
When $n$ is odd
\begin{align*}
\Theta _{\floor{\frac{n}{2}}} &= {\Delta_0} \setminus \left\{{}^\gamma\left(\epsilon_{\floor{\frac{n}{2}}} - \epsilon_{\floor{\frac{n}{2}}+1}\right), {}^\gamma\left(\epsilon_{\floor{\frac{n}{2}}+1} - \epsilon_{\floor{\frac{n}{2}}+2}\right)\right\},
\end{align*}
and when $n$ is even
\begin{align*}
\Theta _{\frac{n}{2}} = {\Delta_0} \setminus \left\{ {}^\gamma\left(\epsilon_{\frac{n}{2}} - \epsilon_{\frac{n}{2}+1}\right)\right\}.
\end{align*}
\end{note}
The next proposition follows immediately from \Cref{lem-max-split-subsets}.
\begin{prop}\label{max-theta-split}
The $\Delta_0$-standard maximal ${\theta}$-split parabolic subgroups of $G$ are:
\begin{align*
P_k& := P_{\Theta _k} = \left \{ \begin{array}{lll} P_{(k,n-2k,k)}, & 1 \leq k \leq \textstyle\frac{n}{2}-1, & \text{in the linear case} \\
{}^\gamma P_{(k,n-2k,k)}, & 1 \leq k \leq \floor{\frac{n}{2}}-1, & \text{in the Galois case}
\end{array} \right.
\end{align*}
and
\begin{align*
P_{\floor{\frac{n}{2}}} & := P_{\Theta _{\floor{\frac{n}{2}}}} = \left \{ \begin{array}{lll} P_{\left(\frac{n}{2},\frac{n}{2}\right)}, & \text{in the linear case, since $n$ is even} \\
{}^\gamma P_{\left(\frac{n}{2},\frac{n}{2}\right)}, & \text{in the Galois case when $n$ is even} \\
{}^\gamma P_{\left(\floor{\frac{n}{2}},1,\floor{\frac{n}{2}}\right)}, & \text{in the Galois case when $n$ is odd}.
\end{array} \right.
\end{align*}
\end{prop}
\begin{rmk}\label{rmk-theta-split-parabolic-notation}
In both cases, $P_k = M_k N_k$, where $M_k=M_{\Theta_k}$ is the standard Levi factor and $N_k$ is the unipotent radical of $P_k$. We write $A_k$ for the $F$-split component and $S_k$ for the $({\theta},F)$-split component of $M_k$.
\end{rmk}
In preparation for our proof of \Cref{no-dist-unit-exp}, here we determine the $\theta$-fixed points of the Levi subgroups $M_k$, $1\leq k \leq \floor{\frac{n}{2}}$.
Recall from \eqref{eq-vartheta-defn} that, in the Galois case, $\vartheta$ is the involution $\gamma\cdot {\theta} = \Int w_\ell \circ {\theta} = {\theta} \circ \Int w_\ell$.
The following is a special case of \Cref{orbit-dist} and also holds for $\theta_r$ and $\vartheta_r$ (\textit{cf.}~\eqref{vartheta-r-defn}).
\begin{lem}\label{theta--theta--fixed}
Assume that we are in the Galois case.
An element of $G$ of the form ${}^\gamma x$ is ${\theta}$-fixed (respectively ${\theta}$-split) if and only if $x$ is $\vartheta$-fixed (respectively $\vartheta$-split).
\end{lem}
\begin{prop}\label{prop-M-Theta-fixed-pts}
Let $1\leq k \leq \floor{\frac{n}{2}}$.
In the linear case, the group $M_k^\theta$ of $\theta$-fixed points in $M_k = M_{(k,n-2k,k)}$ is equal to
\begin{align}\label{H-bullet}
H_{(k,n-2k,k)} = \left \{ \begin{array}{ll} \left\{ \diag\left( A, B, \theta_k(A)\right) : A \in G_k,\ B \in H_{n-2k} \right \}, &\text{if} \ k \neq n/2 \\
\left\{ \diag\left( A, \theta_k(A)\right) : A \in G_k \right \}, &\text{if} \ k=n/2
\end{array} \right..
\end{align}
In the Galois case, $M_k = {}^\gamma M_{(k,n-2k,k)}$ and $M_k^\theta$ is $G$-conjugate to $H_{(k,n-2k,k)}$. Explicitly, we have that
\begin{align*}
H_{(k,n-2k,k)} & = \gamma_{(k,n-2k,k)} M_{(k,n-2k,k)}^{\vartheta}\gamma_{(k,n-2k,k)}^{-1}
\end{align*}
and
\begin{align*}
M_k^{\theta} & = \gamma M_{(k,n-2k,k)}^{\vartheta}\gamma^{-1} = \gamma\gamma_{(k,n-2k,k)}^{-1} H_{(k,n-2k,k)} \gamma_{(k,n-2k,k)}\gamma^{-1},
\end{align*}
where $\gamma_{(k,n-2k,k)}=\diag(\gamma_k, \gamma_{n-2k}, \gamma_k) \in M_{(k,n-2k,k)}$.
\end{prop}
\begin{proof}
Let $1\leq k \leq \floor{\frac{n}{2}}$ and let $M_\bullet = M_{(k,n-2k,k)}$, respectively $M_{(n/2,n/2)}$ when $n$ is even and $k=n/2$. In the linear case, $M_k = M_\bullet$ while, in the Galois case, $M_k = {}^\gamma M_\bullet$.
The statement in the linear case follows from the proof in the Galois case (note the relationship between $\theta$ and $\vartheta$, \textit{cf.}~\eqref{Galois-involution-relation}).
Without loss of generality, we work in the Galois case and assume that $k < n/2$.
By \Cref{theta--theta--fixed}, we have $M_k^\theta = {}^\gamma \left(M_\bullet^{\vartheta}\right)$.
Let ${}^\gamma m \in M_k$ where $m \in M_{\bullet}$. Explicitly, we have $m = \diag(A,B,C)$, where $A,C \in G_k$ and $B \in G_{n-2k}$.
One may verify that
\begin{align*
\vartheta(m)&= \Int w_\ell \circ {\theta} (m) = \diag(\vartheta_k (C), \vartheta_{n-2k}(B), \vartheta_k(A)).
\end{align*}
It follows that
\begin{align*
M_\bullet^{\vartheta}= \left\{ \diag\left( A, B, \vartheta_k(A)\right) : A \in G_k,\ B \in G_{n-2k},\ B = \vartheta_{n-2k}(B) \right \}.
\end{align*}
Conjugating $M_\bullet^{\vartheta}$ by the element $\gamma_\bullet=\diag(\gamma_k, \gamma_{n-2k}, \gamma_k)$ in $M_\bullet$ and applying \Cref{theta--theta--fixed}, we obtain that $M_\bullet^{\vartheta}$ is $M_\bullet$-conjugate (and $F$-isomorphic to) the subgroup $H_\bullet$.
\end{proof}
The next result will allow us to apply \Cref{reduction-standard-split}.
\begin{lem}\label{split-parabolic-conjugate}
Any $\theta$-split parabolic subgroup $P$ of $G$ is $H$-conjugate to a $\Delta_0$-standard $\theta$-split parabolic subgroup $P_\Theta$, for some $\Theta \subset \Delta_0$.
\end{lem}
\begin{proof}
One can check that the degree-one Galois cohomology of $\mathbf A_0 \cap \Hbf$ (respectively, $\mathbf T_0 \cap \Hbf$) over $F$ is trivial.
By a standard argument, we have that $( \Hbf \mathbf A_0)(F) = HA_0$ (respectively, $(\Hbf\mathbf T_0)(F) = HT_0$).
The proposition follows from \Cref{cor-min-std-split-pblc} and \cite[Lemma 2.5(2)]{kato--takano2008}.
\end{proof}
\subsection{A class of $\theta$-elliptic Levi subgroups and $\theta$-stable parabolic subgroups}\label{sec-theta-ell-levi}
The next two lemmas may be readily verified by hand.
\begin{lem}\label{min-rel-ell-Levi}
The Levi subgroup $L_0=C_G\left((A_0^\theta)^\circ\right)$ of $G$ is $\theta$-elliptic and $A_0^\theta = (A_0^\theta)^\circ = A_{L_0}$.
Moreover, $L_0$ is minimal among $\theta$-elliptic Levi subgroup of $G$ that contain $A_0$.
\end{lem}
\begin{proof}
First, we observe that since $(A_0^\theta)^\circ$ is $\theta$-stable, the Levi subgroup $L_0$ is ${\theta}$-stable.
It is immediate that the maximal $F$-split torus $A_0$ is contained in $L_0$ (since $A_0$ is abelian).
Now, we show that $L_0$ is $\theta$-elliptic.
First, note that the $(\theta,F)$-split component $S_G$ of $G$ is the trivial group.
Indeed, in the linear case, $\theta$ is inner and we have that $A_G \cong F^\times$ is pointwise $\theta$-fixed.
It follows from \eqref{eq-S-M-defn} that $S_G = \left( \{\pm e \} \right)^\circ = \{e\}$.
Again, in the Galois case, ${\theta}$ acts trivially on the $F$-split component of the centre $A_G$ of $G$ and $S_G = \{e\}$.
In both the linear and Galois cases, it is readily verified that the $F$-split component of the centre of $L_0$ is equal to $(A_0^\theta)^\circ$, that is, $A_{L_0}=(A_0^\theta)^\circ$. Moreover, in both cases, we have that $(A_0^\theta)^\circ = A_0^\theta$.
In particular, $A_{L_0}$ is contained in $H$ and it follows that $S_{L_0} = \{e\}$ (\textit{cf.}~\cite[$\S1.3$]{helminck1997}).
By Lemma \ref{theta-elliptic-centre}, $L_0$ is $\theta$-elliptic.
Finally, we prove that $L_0$ is minimal among $\theta$-elliptic Levi subgroups containing $A_0$. Suppose that $L \subset L_0$ is a proper Levi subgroup of $L_0$ that contains $A_0$. We argue that $L$ cannot be $\theta$-elliptic. Since $L$ is proper in $L_0$, we have that $A_{L_0} = (A_0^\theta)^\circ$ is a proper sub-torus of $A_L$. Following \cite[$\S1.3$]{helminck1997}, we have an almost direct product $A_L = (A_L^\theta)^\circ S_L$. Observe that, since $A_L \subset A_0$, we have
\begin{align*}
(A_L^\theta)^\circ = (A_L \cap A_0^\theta)^\circ = (A_L \cap A_{L_0})^\circ = A_{L_0}.
\end{align*}
Since $S_G =S_{L_0} = \{e\} \subset A_{L_0} = (A_L^\theta)^\circ$ and $A_L = (A_L^\theta)^\circ S_L$ properly contains $A_{L_0}$, it must be the case that $S_G$ is a proper subtorus of $S_L$; in particular, $S_L$ is non-trivial. It follows from Lemma \ref{theta-elliptic-centre} that $L$ is not $\theta$-elliptic and this completes the proof.
\end{proof}
\begin{lem}\label{Weyl-gp-conjugate}
In the Galois case, conjugation by $\gamma$ maps $N_{G}(A_T)$ to $N_{G}(A_0)$ and induces an explicit isomorphism of the Weyl group $W_T = W(G,A_T)$, with respect to $A_T$, with the Weyl group $W_0 = W(G,A_0)$, with respect to $A_0$. Moreover, we identify $W_T$ with the group of permutation matrices in $G$ (isomorphic to the symmetric group $\style S_n$) and $W_0$ with the $\gamma$-conjugates of the permutation matrices.
\end{lem}
In both the linear and Galois cases, define
\begin{align*
{\Delta^{ell}} = w_0 \Delta_0,
\end{align*}
where $w_0$ is defined in \eqref{w-zero}.
Since $\Delta^{ell}$ is a Weyl group translate of $\Delta_0$, we have that $\Delta^{ell}$ is a base of $\Phi_0$.
In both cases, set
\begin{align*
\Delta_{odd} = \{\epsilon_{i}- \epsilon_{i+1} : i\ \text{is odd}\},
\end{align*}
and in the Galois case further denote
\begin{align*
\Delta_{0,odd} = {}^\gamma \Delta_{odd} \subset \Delta_0 = {}^\gamma \Delta.
\end{align*}
In the linear case, the define the subset $\Delta^{ell}_{\min}$ of ${\Delta^{ell}}$ by
\begin{align*
{\Delta^{ell}_{\min}} = w_0 \Delta_{odd}
\end{align*}
and in the Galois case, define
\begin{align*
{\Delta^{ell}_{\min}} = w_0 \Delta_{0,odd}.
\end{align*}
In both cases, the subset $\Delta^{ell}_{\min}$ is exactly the subset of $\Delta^{ell}$ that cuts out the torus $A_0^{\theta}$ from $A_0$.
In particular,
\begin{align}\label{A0-ell-roots}
A_0^{\theta} = A_{{\Delta^{ell}_{\min}}} = \left( \bigcap_{\beta \in {\Delta^{ell}_{\min}}} \ker (\beta: A_0 \rightarrow F^\times) \right)^\circ,
\end{align}
and
\begin{align*}
L_0 = L_{{\Delta^{ell}_{\min}}} =C_{G}\left(A_{{\Delta^{ell}_{\min}}}\right).
\end{align*}
Write $\Delta_{even} = \{ \epsilon_i - \epsilon_{i+1} : 2\leq i\leq n-1,\ i \ \text{is even} \}$.
\begin{defn}\label{defn-std-theta-ell-levi}
A Levi subgroup $L$ of $G$ is a standard-$\theta$-elliptic Levi subgroup if and only if $L$ is $\Delta^{ell}$-standard and contains $L_0$.
\end{defn}
The next proposition characterizes the inducing subgroups in \Cref{main-RDS-thm}.
\begin{prop}\label{ind-subgp-prop}
Let $\Omega^{ell} \subset {\Delta^{ell}}$ such that $\Omega^{ell}$ contains $\Delta^{ell}_{\min}$.
\begin{enumerate}
\item The ${\Delta^{ell}}$-standard parabolic subgroup $Q = Q_{\Omega^{ell}}$, associated to $\Omega^{ell}$, is ${\theta}$-stable.
\item In particular, the unipotent radical $U = U_{\Omega^{ell}}$ is $\theta$-stable.
\item The Levi subgroup $L=L_{\Omega^{ell}} = C_G(A_{\Omega^{ell}})$ is a standard-$\theta$-elliptic Levi of $G$.
\item The modular function $\delta_Q$ of $Q$ satisfies
$\left. \delta_Q^{1/2} \right\vert_{L^{\theta}} = \delta_{Q^{\theta}}.$ \label{ind-subgp-prop-modular}
\item We have that
\begin{align}\label{L-isom-class}
L &\cong \prod_{i=1}^k G_{m_i} &\text{and}& &L^{\theta} &\cong \prod_{i=1}^k H_{m_i}
\end{align}
where $\sum_{i=1}^k m_i = n$, such that when $n$ is odd exactly one $m_i$ is odd, and when $n$ is even all of the $m_i$ are even.
\end{enumerate}
\end{prop}
\begin{proof}
Since $\Delta_{\min}^{ell} \subset \Omega^{ell}$, it follows from \eqref{A0-ell-roots} that $A_L = A_{\Omega^{ell}}$ is contained in $A_{L_0} = A_0^\theta$.
It follows that $L_0$ is contained in $L$ and $L$ is a standard-$\theta$-elliptic Levi subgroup by \Cref{contain-theta-elliptic} and \Cref{min-rel-ell-Levi}.
By \Cref{theta-elliptic-theta-stable}, $Q$ is a $\theta$-stable parabolic subgroup with $\theta$-stable unipotent radical $U = U_{\Omega^{ell}}$.
In the Galois case, the statement about modular functions is \cite[Lemma 5.5]{gurevich--offen2015}, a proof is given in \cite[Lemma 2.5.1]{lapid--rogawski2003}. In the linear case, one may compute the modular functions by hand to verify the desired equality. We omit the straightforward computation.
Finally, we explicitly describe both $L$ and $L^\theta$. Note that, in the Galois case, $\gamma$ centralizes $A_T^\vartheta$ and by \Cref{theta--theta--fixed} we see that $A_0^\theta = \gamma A_T^\vartheta \gamma^{-1} = A_T^\vartheta$.
In both cases, it follows that $A_0^\theta$ is equal to the $w_+$-conjugate of the $F$-split torus
\begin{align*}
A_{(2,\ldots,2,\widehat1)} = \{ \diag(a_1,a_1,a_2,a_2, \ldots, a_{\floor{\frac{n}{2}}},a_{\floor{\frac{n}{2}}}, \widehat{a}) : a, a_i \in F^\times \},
\end{align*}
corresponding to the partition $(2,\ldots,2,\widehat1)$ of $n$.
Precisely, $A_0^\theta = w_+ A_{(2,\ldots,2,\widehat1)} w_+^{-1}$; moreover, it follows that $L_0 = w_+ M_{(2,\ldots,2,\widehat1)} w_+^{-1}$.
Furthermore, we can realize $\Omega^{ell} = w_0 \Omega$, where $\Omega \subset \Delta_0$ contains $\Delta_{odd}$, respectively $\Delta_{0,odd}$.
In the linear case, $\Omega^{ell} = w_+ \Omega$ since $w_0 = w_+$, while in the Galois case, $\Omega^{ell} = w_0 \Omega = w_+ \gamma \underline \Omega$, where $\underline \Omega = {}^{\gamma^{-1}} \Omega \subset \Delta$ contains $\Delta_{odd}$.
It follows that $L = L_{\Omega^{ell}}$ is the $w_+$-conjugate of a block diagonal Levi subgroup $M_{(m_1,\ldots, m_k)}$ that contains $M_{(2,\ldots,2,\widehat1)}$.
We have now established the first isomorphism in \eqref{L-isom-class}:
\begin{align*}
L & = w_+ M_{(m_1,\ldots, m_k)} w_+^{-1} \cong \prod_{i=1}^k G_{m_i},
\end{align*}
where the partition $(m_1,\ldots, m_k)$ of $n$ is refined by $(2,\ldots,2,\widehat1)$.
In particular, when $n$ is even, each $m_i$ is even and when $n$ is odd, exactly one $m_j$ is odd.
Let $l = w_+ m w_+^{-1} \in L$, where $m \in M=M_{(m_1,\ldots, m_k)}$. To determine $L^\theta$, we treat the linear and Galois cases separately.
Starting in the linear case, we see that $l$ is $\theta$-fixed if and only if $m$ is fixed by the involution $\theta_+ = w_+ \cdot \theta = \Int({w_+^{-1}w_\ell w_+})$.
The element $w_+^{-1}w_\ell w_+$ is the permutation matrix
\begin{align*}
w_+^{-1}w_\ell w_+ = \left( \begin{matrix} \begin{matrix}0 & 1 \\ 1 &0 \end{matrix} & & \\
& \ddots& \\
& & \begin{matrix} 0& 1 \\ 1 & 0\end{matrix} \end{matrix} \right),
\end{align*}
which lies in $M_{(2,\ldots,2)} \subset M$.
We observe that, since each $m_i$ is even, $\theta_+$ acts on the $i$\textsuperscript{th}-block $G_{m_i} = \GL_{m_i}(F)$ of $M$ as conjugation by
\begin{align*}
w_{m_i} = \left( \begin{matrix} \begin{matrix}0 & 1 \\ 1 &0 \end{matrix} & & \\
& \ddots& \\
& & \begin{matrix} 0& 1 \\ 1 & 0\end{matrix} \end{matrix} \right) \in G_{m_i}.
\end{align*}
Moreover, $w_{m_i}$ is $G_{m_i}$-conjugate to $J_{m_i}$ (\textit{cf.}~$\S$\ref{sec-choice-elts}).
It follows that $\theta_+$ acting on $M$ is $M$-equivalent to the product involution $\theta_{m_1} \times \ldots \times \theta_{m_k}$;
therefore, by \Cref{orbit-dist} we have
\begin{align*}
L^{\theta} & = w_+ M^{\theta_+} w_+^{-1} \cong M^{\theta_+} \cong \prod_{i=1}^k (G_{m_i})^{\theta_{m_i}} = \prod_{i=1}^k H_{m_i},
\end{align*}
where the second isomorphism is given by conjugation by an element of $M$.
In the Galois case, note that $w_+$ is ${\theta}$-fixed and $M$ is ${\theta}$-stable.
Then $l = w_+ m w_+^{-1}$ is ${\theta}$-fixed if and only if $m$ is ${\theta}$-fixed.
It follows that $L^{\theta} = w_+ M^{\theta} w_+^{-1}$ and we have that
\begin{align*}
L^{\theta} = w_+ M^{\theta} w_+^{-1} \cong M^{\theta} = \prod_{i=1}^k (G_{m_i})^{{\theta}_{m_i}}= \prod_{i=1}^k H_{m_i},
\end{align*}
as claimed.
\end{proof}
\section{Construction of relative discrete series: The main theorem}\label{sec-main-theorem}
\begin{defn}\label{defn-regular}
A smooth representation $\tau$ of a Levi subgroup $L$ of $G$ is regular if for every non-trivial element $w \in N_G(L)/L$ we have that ${}^w \tau \ncong \tau$, where ${}^w \tau = \tau \circ \Int w^{-1}$.
\end{defn}
For general linear groups, we can immediately translate \Cref{defn-regular} into the following result.
\begin{lem}\label{pairwise-inequiv}
Let $(m_1,\ldots, m_k)$ be a partition of $n$.
Let $\tau_i$ be an irreducible admissible representation of $G_{m_i}$, for $1 \leq i \leq k$.
The representation $\tau_1 \otimes \ldots \otimes \tau_k$ of $M_{(m_1,\ldots, m_k)}$ is regular if and only if $\tau_i \ncong \tau_j$, for all $1\leq i \neq j \leq k$.
\end{lem}
Now we come to the main result of the paper.
\begin{thm}\label{main-RDS-thm}
Let $Q=LU$ be a proper $\Delta^{ell}$-standard $\theta$-stable parabolic subgroup of $G$ with standard-$\theta$-elliptic Levi factor $L$ and unipotent radical $U$.
Let $\tau$ be a regular $L^\theta$-distinguished discrete series representation of $L$.
The parabolically induced representation $\pi = \iota_Q^G \tau$ is irreducible and $H$-relatively square integrable.
\end{thm}
\begin{proof}
By assumption, $\tau$ is unitary and regular; therefore, $\pi$ is irreducible by a result of Bruhat \cite{bruhat1961} (\textit{cf.}~\cite[Theorem 6.6.1]{Casselman-book}).
Since $\tau$ is $L^\theta$-distinguished, $\pi$ is $H$-distinguished by \Cref {ind-subgp-prop}\eqref{ind-subgp-prop-modular} and \Cref{cor-hom-injects}.
Let $\lambda$ denote a fixed nonzero $H$-invariant linear form on $\pi$.
By \Cref{prop-mult-one}, $\lambda$ is unique up to scalar multiples.
To complete the proof, it remains to show that $\pi$ satisfies the Relative Casselman's Criterion.
By \Cref{split-parabolic-conjugate} and \Cref{reduction-standard-split}, it is sufficient to verify that the condition \eqref{rel-casselman} is satisfied for every $\Delta_0$-standard maximal ${\theta}$-split parabolic subgroup.
By assumption, $Q = Q_{\Omega^{ell}}$ for some proper subset $\Omega^{ell} = w_0 \Omega$ of $\Delta^{ell}$ containing $\Delta^{ell}_{\min}$, where $\Omega \subset \Delta_0$. Let $P_\Theta$ be a maximal $\Delta_0$-standard $\theta$-split parabolic subgroup. It follows from \Cref{lem-nice-reps} that the set $[W_\Theta \backslash W_0 / W_\Omega]\cdot w_0^{-1}$ provides a ``nice" choice of representatives for the double-coset space $P_\Theta \backslash G / Q$.
By the Geometric Lemma (\textit{cf.}~\Cref{geom-lem}) and \Cref{exp-irred-subq}, the exponents of $\pi$ along $P_\Theta$ are given by the union
\begin{align*}
\Exp_{A_\Theta}(\pi_{N_\Theta}) = \bigcup_{y\in [W_\Theta \backslash W_0 / W_\Omega]\cdot w_0^{-1}} \Exp_{A_\Theta} (\mathcal F_\Theta^y(\tau)),
\end{align*}
where the exponents on the right-hand side are the central characters of the irreducible subquotients of $\mathcal F_\Theta^y(\tau)$.
By \Cref{closed-subgp-centre}, the map from $\Exp_{A_\Theta} (\mathcal F_\Theta^y(\tau))$ to $\Exp_{S_\Theta}(\pi_{N_\Theta}, r_{P_\Theta} \lambda)$ defined by restriction of characters is surjective.
Set $y=ww_0^{-1}$, where $w\in [W_\Theta \backslash W_0 / W_\Omega]$.
If ${}^w M_\Omega \subseteq M_\Theta$, then the parabolic subgroup $P_\Theta\cap{}^wM_\Omega$ of ${}^wM_\Omega$ is equal to ${}^wM_\Omega$. The containment ${}^w M_\Omega \subseteq M_\Theta$ occurs in two cases:
\begin{enumerate}
\item[(A)] when ${}^w M_\Omega = M_\Theta$, which occurs if and only if $w\in [W_\Theta \backslash W_0 / W_\Omega] \cap W(\Theta, \Omega)$, where $W(\Theta, \Omega) = \{ w \in W_0 : w\Omega = \Theta \}$, and
\item[(B)] when $w\in [W_\Theta \backslash W_0/ W_\Omega]$ is such that ${}^wM_\Omega \subsetneq M_\Theta$ is a proper Levi subgroup of $M_\Theta$.
\end{enumerate}
In both cases (A) and (B), $\mathcal F_\Theta^y(\tau)$ is not $M_\Theta^{\theta}$-distinguished by \Cref{no-dist-unit-exp} and the unitary exponents $\chi_{\Theta,y}$ of $\mathcal F_\Theta^y(\tau)$ do not contribute to $\Exp_{S_\Theta}(\pi_{N_\Theta}, r_{P_\Theta} \lambda)$ by \Cref{non-dist-gen-eig-sp} (\textit{cf.}~\eqref{relative-exponent}).
Otherwise, if ${}^w M_\Omega \nsubseteq M_\Theta$, we have that $P_\Theta\cap {}^w M_\Omega$ is a proper parabolic subgroup of ${}^w M_\Omega$.
By \Cref{prop-reduce-Cas-ind}, we have that
\begin{align*}
|\chi(s)|_F < 1, \ \text{for all} \ \chi \in \Exp_{S_\Theta}(\mathcal F_\Theta^y(\tau)), \ \text{and all} \ s \in S_\Theta^- \setminus S_\Theta^1 S_{\Delta_0}.
\end{align*}
In particular, \eqref{rel-casselman} holds for all exponents $\chi \in \Exp_{S_\Theta}(\pi_{N_\Theta}, r_{P_\Theta} \lambda)$ relative to $ \lambda$ along all maximal $\Delta_0$-standard $\theta$-split parabolic subgroups $P_\Theta$.
By \Cref{rel-casselman-crit}, we conclude that $\pi$ is $(H,\lambda)$-relatively square integrable.
\end{proof}
\begin{obs}\label{obs-isom-type-RDS}
A representation $\pi$ of $G$ is $H$-distinguished if and only if $\pi$ is ${}^gH$-distinguished; in particular, the property of distinction only depends on the $G$-conjugacy class of $H$ (or the $G$-equivalence class of $\theta$, \textit{cf.}~\Cref{orbit-dist}).
Thus, taking into account \Cref{ind-subgp-prop} and \Cref{pairwise-inequiv}, we may rephrase \Cref{main-RDS-thm} as follows:
\begin{enumerate}
\item Assume that $n$ is even. Let $(m_1, \ldots, m_k)$ be a partition of $n$ such that each $m_i$ is even. Let $\tau_1, \ldots, \tau_k$ be pairwise inequivalent $H_{m_i}$-distinguished discrete series representations of $G_{m_i}$. The parabolically induced representation $\tau_1 \times \ldots \times \tau_k$ is an irreducible $H$-distinguished relative discrete series representation of $G$.
\item If $n$ is odd, then we must be in the Galois case. Let $(m_1, \ldots, m_k)$ be a partition of $n$ such that exactly one $m_l$ is odd, and all other $m_i$ are even. Let $\tau_1, \ldots, \tau_k$ be pairwise inequivalent $\GL_{m_i}(F)$-distinguished discrete series representations of $\GL_{m_i}(E)$. The parabolically induced representation $\tau_1 \times \ldots \times \tau_k$ is an irreducible $\GL_n(F)$-distinguished relative discrete series representation of $\GL_n(E)$.
\end{enumerate}
\end{obs}
\begin{cor}\label{cor-rds-not-ds}
Let $\pi = \iota_Q^G\tau$ be as in \Cref{main-RDS-thm}.
The representation $\pi$ is a relative discrete series representation that does not lie in the discrete spectrum of $G$.
\end{cor}
\begin{proof}
By \Cref{main-RDS-thm}, $\pi$ is irreducible and $H$-relatively square integrable; therefore, $\pi$ is a relative discrete series. Since $\pi = \iota_Q^G\tau$, where $Q$ is proper in $G$, it follows from the work of Zelevinsky \cite{zelevinsky1980} that $\pi$ does not occur in the discrete spectrum of $G$.
\end{proof}
\begin{rmk}\label{rmk-exhaustion}
At present, the author does not know if the construction outlined in \Cref{main-RDS-thm} exhausts all non-discrete relative discrete series in the linear and Galois cases. In order to show that a representation is not $(H,\lambda)$-relatively square integrable, it is necessary to show that $r_P\lambda$ is non-vanishing on the generalized eigenspace corresponding to an exponent $\chi \in \Exp_{S_M}(\pi_N,r_P\lambda)$ that fails the condition \eqref{rel-casselman}. The non-vanishing of $r_P\lambda$ is obscured by the nature of the construction of the form via Casselman's Canonical Lifting. Due to this lack of precise information, we cannot exclude the possibility that certain representations are RDS. For instance, it may be possible to relax the regularity condition imposed in \Cref{main-RDS-thm}, which is essential in the proof of \Cref{no-dist-unit-exp}. At this time, the author does not have a method to remove the assumption of regularity from \Cref{no-dist-unit-exp}, due to a lack of information regarding the support of the $r_P\lambda$.
\end{rmk}
By \cite[Theorem 6.2]{kato--takano2008} and the proof of \Cref{main-RDS-thm},
one may obtain the following.
\begin{cor}\label{cor-sc-main-thm}
Let $Q = LU$ be as in \Cref{main-RDS-thm}.
If $\tau$ is a regular $L^\theta$-distinguished supercuspidal representation of $L$, then $\pi = \iota_Q^G\tau$ is $H$-relatively supercuspidal.
\end{cor}
\begin{rmk}
Note that \Cref{cor-sc-main-thm} can be obtained by more direct methods; see, for instance, the work of Murnaghan \cite{murnaghan2016-pp} for such results in a more general setting.
\end{rmk}
\section{Distinguished discrete series: Known results and inducing data}\label{sec-inducing-data}
In this section, we survey the known results on distinguished discrete series representations in the linear and Galois cases. Our ultimate goal is to prove \Cref{prop-infinite-reg-dist-non-sc-ds} and thus \Cref{cor-infinite-family}.
First, we note that, in the linear case, multiplicity-one is due to Jacquet and Rallis \cite{jacquet--rallis1996}.
In the Galois case, multiplicity-one is due to Flicker \cite{flicker1991}.
\begin{prop}[Jacquet--Rallis, Flicker]\label{prop-mult-one}
Let $\pi$ be irreducible admissible representation of $G$. If $\pi$ is $H$-distinguished, then $\Hom_H(\pi,1)$ is one-dimensional.
\end{prop}
Jacquet and Rallis \cite{jacquet--rallis1996} also prove the next proposition.
\begin{prop}[Jacquet--Rallis]\label{lin-dist-property}
Let $M$ be a maximal Levi subgroup of $\GL_m(F)$, where $m\geq 2$.
Let $\pi$ be an irreducible admissible representation of $\GL_m(F)$.
If $\pi$ is $M$-distinguished, then $\pi$ is equivalent to its contragredient $\widetilde \pi$.
\end{prop}
Flicker \cite{flicker1991} uses the methods of \cite{gelfand--kazhdan1975} to prove the following result.
\begin{prop}[Flicker]\label{Gal-dist-property}
Let $\pi$ be an irreducible admissible representation of $\GL_m(E)$, where $m\geq 2$. If $\pi$ is $\GL_m(F)$-distinguished, then $\pi \cong {}^{\theta}\widetilde\pi$.
\end{prop}
\subsection{Distinguished discrete series in the linear case}
In this subsection, unless otherwise noted, we let $G = \GL_n(F)$, where $n\geq 2$ is even, and let $H=\GL_{n/2}(F) \times \GL_{n/2}(F)$.
\begin{rmk}\label{rmk-lin-Shalika}
It is known that an irreducible square integrable representation $\pi$ of $G$ is $H$-distinguished if and only if $\pi$ admits a Shalika model.
It was shown by Jacquet and Rallis \cite{jacquet--rallis1996} that if $\pi$ is an irreducible admissible representation of $G$ that admits a Shalika model, then $\pi$ is $H$-distinguished.
For irreducible supercuspidal representations, the converse appears as \cite[Theorem 5.5]{jiang--nien--qin2008}.
Independently, Sakellaridis--Venkatesh and Matringe proved the converse result for relatively integrable and relatively square integrable representations by the technique of ``unfolding" \cite[Example 9.5.2]{sakellaridis--venkatesh2017}, \cite[Theorem 5.1]{matringe2014a}.
In fact, Sakellaridis--Venkatesh prove that there is an equivariant unitary isomorphism between $L^2(H\backslash G)$ and $L^2(S\backslash G)$, where $S$ is the Shalika subgroup.
Several analogous global results appear in \cite{gan--takeda2010}.
\end{rmk}
Let $\pi$ be a discrete series representation of $\GL_m(F)$, $m \geq 2$. Denote by $L(s,\pi\times\pi)$ the local Rankin--Selberg convolution $L$-function.
It is well known that $L(s,\pi\times\pi)$ has a simple pole at $s=0$ if and only if $\pi$ is self-contragredient \cite{jacquet--piatetski-shaprio--shalika1983}.
By \cite[Lemma 3.6]{shahidi1992}, we have a local identity
\begin{align}\label{eq-RS-L-fcn-factors}
L(s,\pi\times \pi) = L(s,\pi, \wedge^2) L(s,\pi, \Sym^2),
\end{align}
where $L(s,\pi, \wedge^2)$, respectively $L(s,\pi, \Sym^2)$, denotes the exterior square, respectively symmetric square, $L$-function of $\pi$ defined via the Local Langlands Correspondence (LLC).
It is also well known, see \cite{bump2004, kewat--raghunathan2012} for instance, that $L(s,\pi, \wedge^2)$ cannot have a pole when $m$ is odd.
\begin{thm}[{\cite[Proposition 6.1]{matringe2014a}}]\label{thm-matringe-ext-square}
Suppose that $\pi$ is a square integrable representation of $G$, then $\pi$ is $H$-distinguished
if and only if the exterior square $L$-function $L(s,\pi,\wedge^2)$ has a pole at $s=0$.
\end{thm}
\begin{rmk}
It is now known that for all discrete series, and when $n$ is even all irreducible generic representations, the Jacquet--Shalika and Langlands--Shahidi local exterior square $L$-functions agree with with the exterior square $L$-function defined via the LLC \cite[Theorem 4.3 in $\S$4.2]{henniart2010}, \cite[Theorems 1.1 and 1.2]{kewat--raghunathan2012}.
\end{rmk}
Let $\rho$ be an irreducible unitary supercuspidal representation of $\GL_r(F)$, $r\geq 1$. For an integer $k\geq 2$, write $\st(k,\rho)$ for the unique irreducible (unitary) quotient of the parabolically induced representation $\nu^{\frac{1-k}{2}} \rho \times \nu^{\frac{3-k}{2}} \rho \times \ldots \times \nu^{\frac{k-1}{2}} \rho$ of $\GL_{kr}(F)$ (\textit{cf.}~\cite[Proposition 2.10, $\S$9.1]{zelevinsky1980}), where $\nu(g) = |\det(g)|_F$, for any $g \in \GL_r(F)$.
The representations $\st(k,\rho)$ are often called {generalized Steinberg representations} and they are exactly the nonsupercuspidal discrete series representations of $\GL_{kr}(F)$ \cite[Theorem 9.3]{zelevinsky1980}.
The usual Steinberg representation $\st_n$ of $\GL_n(F)$ is obtained as $\st(n,1)$.
Note that $\st({k_1},\rho_1)$ is equivalent to $\st({k_2},\rho_2)$ if and only if $k_1 = k_2$ and $\rho_1$ is equivalent to $\rho_2$ \cite[Theorem 9.7(b)]{zelevinsky1980}.
\begin{thm}[{\cite[Theorem 6.1]{matringe2014a}}]\label{thm-matringe-lin}
Suppose that $n= kr$ is even. Let $\rho$ be an irreducible supercuspidal representation of $\GL_r(F)$.
Let $\pi = \st(k,\rho)$ be a generalized Steinberg representation of $G$.
\begin{enumerate}
\item If $k$ is odd, then $r$ must be even, and $\pi$ is $H$-distinguished
if and only if $L(s, \rho, \wedge^2)$ has a pole at $s=0$ if and only if $\rho$ is $\GL_{r/2}(F)\times \GL_{r/2}(F)$-distinguished
\item If $k$ is even, then $\pi$ is $H$-distinguished
if and only if $L(s, \rho, \Sym^2)$ has a pole at $s=0$. \label{thm-matringe-lin-even}
\end{enumerate}
\end{thm}
The following is a corollary of \cite[Proposition 10.1]{murnaghan2011a} and \cite[Theorem 1.3]{Hakim--Murnaghan2002}.
\begin{thm}[Murnaghan, Hakim--Murnaghan]\label{thm-lin-HM}
For any even integer $n \geq 2$, there exist
\begin{enumerate}
\item infinitely many equivalence classes of $H$-distinguished irreducible (unitary) tame supercuspidal representations of $G$, and \label{HM-dist}
\item infinitely many equivalence classes of self-contragredient irreducible (unitary) tame supercuspidal representations of $G$ that are not $H$-distinguished. \label{HM-not-dist}
\end{enumerate}
\end{thm}
Finally, we note the following.
\begin{prop}\label{prop-lin-even-dist-ds}
For any even integer $n \geq 2$, there are infinitely many equivalence classes of $H$-distinguished discrete series representations of $G$.
Moreover,
\begin{enumerate}
\item if $n=2$, there are exactly four $H$-distinguished twists of the Steinberg representation $\st_2$ of $G$; \label{prop-lin-even-dist-ds--2}
\item if $n=4$, there are exactly four $H$-distinguished twists of the Steinberg representation $\st_4$ of $G$, and there are infinitely many equivalence classes of $H$-distinguished generalized Steinberg representations of $G$ of the form $\st(2,\rho)$. \label{prop-lin-even-dist-ds--4}
\item if $n\geq 6$, there are infinitely many equivalence classes of $H$-distinguished nonsupercuspidal discrete series representations of $G$. \label{prop-lin-even-dist-ds--geq6}
\end{enumerate}
\end{prop}
\begin{proof}
The main statement follows from \Cref{thm-lin-HM}.
\begin{enumerate}
\item The Steinberg representation $\st_2$ has trivial central character and so it is $H$-distinguished \cite{prasad1993}.
A twist $\chi \otimes \st_2$ of $\st_2$ by a quasi-character $\chi$ of $F^\times$ has trivial central character if and only if $\chi$ is trivial on $(F^\times)^2$.
In particular, $\chi$ must be a quadratic (unitary) character and, since $F$ has odd residual characteristic, there are four distinct such characters.
\item By \cite[Corollary 8.5(\rm{ii})]{gan--takeda2010}, a twisted Steinberg representation $\chi\otimes\st_4$ admits a Shalika model if and only if $\chi$ is trivial on $(F^\times)^2$.
In this case, $\chi\otimes\st_4$ is $H$-distinguished (\textit{cf.}~\Cref{rmk-lin-Shalika}).
By \Cref{thm-lin-HM}\eqref{HM-not-dist}, there exist infinitely many classes of self-contragredient supercuspidal representations $\rho$ of $G_2$ that are not $H_2$-distinguished. In particular, given such a $\rho \cong \wt\rho$, the Rankin--Selberg $L$-fucntion $L(s,\rho\times\rho)$ has a pole at $s=0$ \cite{jacquet--piatetski-shaprio--shalika1983}; however, by \Cref{thm-matringe-ext-square}, $L(s,\rho,\wedge^2)$ does not have a pole at $s=0$. It follows from \eqref{eq-RS-L-fcn-factors} that $L(s,\rho, \Sym^2)$ has a pole at $s=0$.
The claim follows from \Cref{thm-matringe-lin}\eqref{thm-matringe-lin-even}.
\item The last statement is an immediate consequence of \Cref{thm-matringe-lin,thm-lin-HM}.
\end{enumerate}
\end{proof}
\subsection{Distinguished discrete series in the Galois case}
In this subsection, unless otherwise noted, let $G = R_{E/F} \GL_n(F)$, where $n\geq 2$.
We identify $G$ with $\GL_n(E)$.
Let $H = \GL_n(F)$ be the subgroup of Galois fixed points in $G$.
Let $\eta: E^\times \rightarrow \C^\times$ be an extension to $E^\times$ of the character $\eta_{E/F}:F^\times \rightarrow \C$ associated to $E/F$ by local class field theory.
The following result is due to Anandavardhanan--Rajan \cite[Section 4.4]{anandavardhanan--rajan2005}, and also appears as \cite[Theorem 1.3]{anandavardhanan2008} and \cite[Corollary 4.2]{matringe2009a}.
\begin{thm}[Anandavardhanan--Rajan]\label{thm-matringe-Gal}
Let $\rho$ be an irreducible supercuspidal representation of $\GL_r(E)$, then the generalized Steinberg representation $\pi = \st(k,\rho)$ of $\GL_{kr}(E)$ is $\GL_{kr}(F)$-distinguished if and only if $\rho$ is $(\GL_r(F),\eta_{E/F}^{k-1})$-distinguished.
\end{thm}
The next result is due to Prasad for $n=2$ and Anandavardhanan--Rajan for $n\geq 3$.
\begin{thm}[{\cite{prasad1992}},{\cite[Theorem 1.5]{anandavardhanan--rajan2005}}]\label{thm-Gal-st}
Let $\chi$ be a quasi-character of $F^\times$ and identify $\chi$ with the quasi-character $\chi \circ \det$ of $H$.
The Steinberg representation $\st_n$ of $G$ is $(H,\chi)$-distinguished if and only if:
\begin{enumerate}
\item $n$ is odd and $\chi = 1$, or \label{thm-Gal-st-odd}
\item $n$ is even and $\chi = \eta_{E/F}$. \label{thm-Gal-st-even}
\end{enumerate}
\end{thm}
\begin{cor}\label{cor-Gal-st-2}
If $n=2$, then the twist $\eta\otimes \st_2$ of the Steinberg representation $\st_2$ of $G$ is $H$-distinguished.
\end{cor}
The following is a corollary of \cite[Proposition 10.1]{murnaghan2011a} and \cite[Theorem 1.1]{Hakim--Murnaghan2002}.
\begin{thm}[Hakim--Murnaghan]\label{thm-Gal-HM}
There are infinitely many equivalence classes of
\begin{enumerate}
\item irreducible (unitary) supercuspidal representations of $G$ that are $H$-distinguished.
\item irreducible (unitary) supercuspidal representations of $G$ that are $(H, \eta_{E/F})$-distinguished.
\end{enumerate}
\end{thm}
The next result is an immediate consequence of \Cref{thm-matringe-Gal,thm-Gal-st,thm-Gal-HM}.
\begin{cor}\label{prop-Gal-infinite-non-sc-dist-ds}
Let $G = \GL_n(E)$ and let $H = \GL_n(F)$.
\begin{enumerate}
\item If $n \geq 4$ is not equal to an odd prime, then there are infinitely many equivalence classes of $H$-distinguished nonsupercuspidal discrete series representations of $G$.
\item If $n$ is equal to an odd prime, then the Steinberg representation $\st_n$ of $G$ is a nonsupercuspidal $H$-distinguished discrete series.
\end{enumerate}
\end{cor}
\begin{proof}
Assume that $n \geq 4$ is not an odd prime. Then $n = kr$ for two integers $k,r \geq 2$. Note that $\eta_{E/F}$ is a quadratic character; in particular, if $k$ is even, then $\eta_{E/F}^{k-1} = \eta_{E/F}$ and if $k$ is odd, then $\eta_{E/F}^{k-1} = 1$. By \Cref{thm-Gal-HM}, there are infinitely many equivalence classes of irreducible supercuspidal representations $\rho$ of $G_r$ that are $(G_r, \eta_{E/F}^{k-1})$-distinguished. By \Cref{thm-matringe-Gal} and \cite[Theorem 9.7(b)]{zelevinsky1980}, there are infinitely many equivalence classes of generalized Steinberg representations of $G$ of the form $\st(k,\rho)$ and that are $H$-distinguished.
Of course, the generalized Steinberg representations $\st(k,\rho)$ are nonsupercuspidal discrete series representations.
The second statement follows from \Cref{thm-Gal-st}\eqref{thm-Gal-st-odd}.
\end{proof}
\subsection{The inducing representations in \Cref{main-RDS-thm}}
For the remainder of the paper, fix a proper $\Delta^{ell}$-standard $\theta$-stable parabolic subgroup $Q = Q_{\Omega^{ell}}$, for some proper subset $\Omega^{ell} \subset {\Delta}^{ell}$ containing $\Delta^{ell}_{\min}$.
As in \Cref{ind-subgp-prop}, the subgroup $Q$ admits a standard-$\theta$-elliptic Levi subgroup $L = L_{\Omega^{ell}}$ and unipotent radical $U= U_{\Omega^{ell}}$.
The next lemma is straightforward to verify by using the description of $L^\theta$ given in (the proof of) \Cref{ind-subgp-prop} and \Cref{orbit-dist}. The multiplicity-one statement follows from \Cref{prop-mult-one}.
\begin{lem}\label{lem-L-theta-dist
Let $L \cong \prod_{i=1}^k G_{m_i}$ be a standard-$\theta$-elliptic Levi subgroup of $G$. Let $\tau \cong \bigotimes_{i=1}^k \tau_i$ be an irreducible admissible representation of $L$ where each $\tau_i$ is an irreducible admissible representation of $G_{m_i}$, $1\leq i \leq k$.
\begin{enumerate}
\item Then $\tau$ is $L^\theta$-distinguished if and only if $\tau_i$ is $H_{m_i}$-distinguished for all $1\leq i \leq k$.
\item If $\tau$ is $L^\theta$-distinguished, then $\Hom_{L^\theta}(\tau,1)$ is one-dimensional.
\end{enumerate}
\end{lem}
\begin{prop}\label{prop-infinite-reg-dist-non-sc-ds}
Let $L$ be a standard-$\theta$-elliptic Levi subgroup of $G$.
There exist infinitely many equivalence classes of regular non-supercuspidal $L^\theta$-distinguished discrete series representations of $L$.
\end{prop}
\begin{proof}
By assumption $n\geq 4$ and $n$ is always taken to be even in the linear case.
We have that $L$ is isomorphic to a product $\prod_{i=1}^k G_{m_i}$ of smaller general linear groups, for some partition $(m_1,\ldots, m_k)$ of $n$.
Let $\tau_i$ be an irreducible admissible representation of $G_{m_i}$, $1 \leq i \leq k$.
By \Cref{pairwise-inequiv}, the representation $\tau_1 \otimes \ldots \otimes \tau_k$ of $L$ is regular if and only if $\tau_i \ncong \tau_j$ for all $1\leq i \neq j \leq k$.
Moreover, $\tau_1 \otimes \ldots \otimes \tau_k$ is supercuspidal (square integrable) if and only if every $\tau_i$ is supercuspidal (square integrable).
It is sufficient to prove that for any relevant partition of $n$ (\textit{cf.} \Cref{ind-subgp-prop}), there exist pairwise inequivalent $H_{m_i}$-distinguished discrete series representations $\tau_i$, such that at least one $\tau_i$ is not supercuspidal.
In the linear case, by \Cref{ind-subgp-prop}, each $m_i \geq 2$ is even. By \Cref{thm-lin-HM}, there are infinitely many equivalence classes of $H_{m_i}$-distinguished irreducible supercuspidal representations of $G_{m_i}$.
By \Cref{prop-lin-even-dist-ds}, there exists at least one non-supercuspidal $H_{m_i}$-distinguished discrete series representation of $G_{m_i}$ (infinitely many when $m_i \geq 4$). It follows from \Cref{lem-L-theta-dist} that there exist infinitely many equivalence classes of regular non-supercuspidal $L^\theta$-distinguished discrete series representations of $L$.
In the Galois case, by \Cref{ind-subgp-prop}, at most one $m_i$ is odd. Without loss of generality, assume that $m_k$ is odd.
By \Cref{thm-Gal-st}\eqref{thm-Gal-st-odd}, the Steinberg representation $\st_{m_k}$ of $G_{m_k}$ is a non-supercuspidal $H_{m_k}$-distinguished discrete series.
By \Cref{thm-Gal-HM}, there are infinitely many equivalence classes of $H_{m_i}$-distinguished irreducible supercuspidal representations of $G_{m_i}$.
By \Cref{cor-Gal-st-2} and \Cref{prop-Gal-infinite-non-sc-dist-ds}, there exists at least one non-supercuspidal $H_{m_i}$-distinguished discrete series representation of $G_{m_i}$ (infinitely many when $m_i\geq 4$ is not an odd prime).
It follows from \Cref{lem-L-theta-dist} that there exist infinitely many equivalence classes of regular non-supercuspidal $L^\theta$-distinguished discrete series representations of $L$.
\end{proof}
\begin{cor}\label{cor-infinite-family}
There are infinitely many equivalence classes of $H$-distinguished relative discrete series representations of $G$ of the form constructed in \Cref{main-RDS-thm}. In particular, there are infinitely many classes of such representations where the discrete series $\tau$ is not supercuspidal.
\end{cor}
\begin{proof}
This is immediate from \Cref{prop-infinite-reg-dist-non-sc-ds} and \cite[Theorem 9.7(b)]{zelevinsky1980}.
\end{proof}
\section{Computation of exponents and distinction of Jacquet modules}\label{sec-exp-dist-pi-N}
We work under the hypotheses of Theorem \ref{main-RDS-thm} and use the notation of its proof.
In order to discuss Casselman's Criterion for the inducing data of $\pi=\iota_Q^G\tau$ we use the following notation.
If $\Theta_1 \subset \Theta_2 \subset \Delta_0$, then we define
\begin{align*}
A_{\Theta_1}^- = \{ a \in A_{\Theta_1} : |\alpha(a)|\leq 1,\ \text{for all} \ \alpha \in \Delta_0\setminus {\Theta_1}\}
\end{align*}
and
\begin{align*}
A_{\Theta_1}^{-\Theta_2} = \{ a \in A_{\Theta_1} : |\beta(a)|\leq 1,\ \text{for all} \ \beta \in \Theta_2 \setminus {\Theta_1}\}.
\end{align*}
The set $A_{\Theta_1}^-$ is the dominant part of $A_{\Theta_1}$ in $G$, while $A_{\Theta_1}^{-\Theta_2}$ is the dominant part of $A_{\Theta_1}$ in $M_{\Theta_2}$.
In both the linear and Galois cases, we have that $Q = w_0 P_\Omega w_0^{-1}$ is a Weyl group conjugate of a $\Delta_0$-standard parabolic subgroup $P_\Omega$, where $\Omega^{ell} = w_0 \Omega$. We also have that $L = w_0 M_\Omega w_0^{-1}$.
If $\tau_0$ is a representation of $M_\Omega$, then $\tau = {}^{w_0}\tau_0$ is a representation of $L$.
Let $P_\Theta$ be a $\Delta_0$-standard maximal $\theta$-split parabolic subgroup corresponding to a maximal proper $\theta$-split subset $\Theta \subset \Delta_0$. Below $y = ww_0^{-1}$ is a ``nice" representative of a double-coset in $P_\Theta \backslash G / Q$, where $w\in [W_\Theta \backslash W_0/ W_\Omega]$ (\textit{cf.}~\Cref{lem-nice-reps}).
\begin{lem}\label{red-to-Cas-ind-data}
The exponents of $\mathcal F_\Theta^y(\tau)$ are the restriction to $A_\Theta$ of the exponents of ${}^w \tau_0$ along the parabolic subgroup $P_\Theta\cap {}^w M_\Omega$ of ${}^w M_\Omega$.
If $\tau = {}^{w_0} \tau_0$ is a discrete series representation of $L$, and the parabolic subgroup $P_\Theta\cap {}^w M_\Omega$ of ${}^w M_\Omega$ is proper, then for any exponent $\chi \in \Exp_{A_\Theta}(\mathcal F_\Theta^y(\tau))$ the inequality $|\chi(a)|_F < 1$ is satisfied for every $a \in A_{\Theta \cap w\Omega}^{-w\Omega} \setminus A_{\Theta \cap w\Omega}^1 A_{w\Omega}$.
\end{lem}
\begin{proof}
This is a special case of \Cref{red-to-ind-exp} and the usual Casselman's Criterion (\textit{cf.}~\cite[Theorem 6.5.1]{Casselman-book}) applied to the discrete series representation ${}^w\tau_0$ of ${}^w M_\Omega$.
\end{proof}
\begin{lem}\label{unitary-exponents}
Assume that $\tau$ is a regular unitary irreducible admissible representation of $L$.
If $y= w w_0^{-1}$, where $w\in [W_\Theta \backslash W_0/ W_\Omega]$, is such that ${}^w M_\Omega \subset P_\Theta$, then $\mathcal F_\Theta^y(\tau)$ is irreducible and the central character $\chi_{\Theta,y}$ of $\mathcal F_\Theta^y(\tau)$ is unitary.
\end{lem}
\begin{proof}
If ${}^w M_\Omega \subset P_\Theta$, then $P_\Theta\cap {}^w M_\Omega = {}^w M_\Omega$, $N_\Theta \cap {}^w M_\Omega = \{e\}$, and ${}^wM_\Omega \subset M_\Theta$.
It follows that the representation $({}^w \tau_0)_{N_\Theta\cap {}^w M_\Omega}$ is equal to ${}^w \tau_0$ and it is irreducible and unitary.
Moreover, since $\tau$ is a regular representation of $L$, it follows that ${}^w \tau_0$ is regular as a representation of ${}^w M_\Omega$ regarded as a Levi subgroup of $M_\Theta$.
By \cite[Theorem 6.6.1]{Casselman-book}, the representation $\mathcal F_\Theta^y(\tau)$ is irreducible and unitary.
By the irreducibility of $\mathcal F_\Theta^y(\tau)$, the only exponent is its central character $\chi_{\Theta,y}$.
Since $\mathcal F_\Theta^y(\tau)$ is unitary, the character $\chi_{\Theta,y}$ of $A_\Theta$ is unitary.
\end{proof}
\begin{rmk}\label{rmk-two-cases-lin-Gal}
Recall that $W(\Theta, \Omega) = \{ w \in W_0 : w\Omega = \Theta \}$.
We find ourselves in the situation of \Cref{unitary-exponents} in two cases:
\begin{itemize}
\item Case (A): $w\in [W_\Theta \backslash W_0 / W_\Omega] \cap W(\Theta, \Omega)$, if and only if ${}^w M_\Omega = M_\Theta$,
\item Case (B): $w\in [W_\Theta \backslash W_0/ W_\Omega]$ is such that ${}^wM_\Omega \subsetneq M_\Theta$ is a proper Levi subgroup of $M_\Theta$.
\end{itemize}
\end{rmk}
In order to apply the Relative Casselman's Criterion \ref{rel-casselman-crit}, using \Cref{red-to-Cas-ind-data}, we need the following technical fact.
\begin{lem}\label{technical-torus-nbhd}
Let $\Omega$ be a proper subset of ${\Delta_0}$ such that $\Omega^{ell} = w_0\Omega$ contains $\Delta^{ell}_{\min}$. Let ${\Theta}$ be a maximal ${\theta}$-split subset of ${\Delta_0}$. Let $w\in [W_\Theta \backslash W_0 / W_\Omega]$ such that $M_{\Theta\cap w\Omega} = M_\Theta \cap wM_\Omega w^{-1}$ is a proper Levi subgroup of $M_{w\Omega} = w M_\Omega w^{-1}$. Then we have the containment:
\begin{align*
S_\Theta^- \setminus S_\Theta^1 S_{\Delta_0} \subset A_{\Theta\cap w\Omega}^{-w\Omega} \setminus A_{\Theta\cap w\Omega}^1A_{w\Omega}.
\end{align*}
\end{lem}
\begin{proof}
Recall that for any $F$-torus $\mathbf A$ we write $A^1$ for the $\of$-points of $\mathbf A$.
First, $S_\Theta$ is contained in $A_\Theta$ and since $\Theta\cap w\Omega$ is a subset of $\Theta$, we have that
\begin{align*}
\mathbf A_\Theta & = \left( \bigcap_{\alpha\in{\Theta}} \ker \alpha \right)^\circ
\subset \left( \bigcap_{\alpha\in\Theta\cap w\Omega} \ker \alpha \right)^\circ
= \mathbf A_{\Theta\cap w\Omega}.
\end{align*}
At the level of $F$-points, we have $A_\Theta \subset A_{\Theta\cap w\Omega}$, and similarly for the integer points
$A_\Theta^1 \subset A_{\Theta\cap w\Omega}^1$. It follows that $S_\Theta \subset A_{\Theta\cap w\Omega},$ and $S_\Theta^1 \subset A_{\Theta\cap w\Omega}^1$. Also, we have $S_{\Delta_0} \subset A_{\Delta_0} \subset A_\Omega$, and since $S_{\Delta_0}=S_G$ is central in $G$, we have $S_{\Delta_0} = wS_{\Delta_0} w^{-1} \subset wA_\Omega w^{-1} = A_{w\Omega}$.
We now observe that $A_\Theta^- \subset A_{\Theta\cap w\Omega}^{-w\Omega}$ and in particular that $S_\Theta^- \subset A_{\Theta\cap w\Omega}^{-w\Omega}$; it is clear that $S_\Theta^- \subset A_\Theta^-$. Note that $w\Omega$ is a base for the root system of $M_{w\Omega}$ relative to the maximal $F$-split torus ${A_0}$. Suppose that $a \in A_\Theta^-$, then $|\alpha(a)| \leq 1$, for all $\alpha \in {\Delta_0} \setminus {\Theta}$.
Moreover, since $a\in A_\Theta$ we have that $|\alpha(a)| = 1$, for $\alpha \in {\Theta}$ as well. Let $\beta \in w\Omega \setminus (\Theta\cap w\Omega)$, then $\beta =w\alpha$ for some $\alpha \in \Omega$. Since $w\in [W_\Theta \backslash W_0 / W_\Omega]$, we have that $\beta = w\alpha \in \Phi_0^+$. Write $\beta = \sum_{\epsilon \in {\Delta_0}} c_\epsilon \cdot \epsilon$,
where $c_\epsilon \geq 0$, $c_\epsilon \in \Z$. Then we have that
\begin{align*}
|\beta(a)| = \left\vert \prod_{\epsilon \in {\Delta_0}} \epsilon(a)^{c_\epsilon} \right \vert = \prod_{\epsilon \in {\Delta_0}} |\epsilon(a)|^{c_\epsilon} \leq 1,
\end{align*}
since $|\epsilon(a)|\leq 1$, for all $\epsilon \in {\Delta_0}$, and $c_\epsilon \geq 0$. In particular, $a \in A_{\Theta\cap w\Omega}^{-w\Omega}$.
Putting this together, we see that $S_\Theta^1 S_{\Delta_0} \subset S_\Theta^- \cap A_{\Theta\cap w\Omega}^1A_{w\Omega}$; therefore, to prove the desired result, it suffices to prove the opposite inclusion.
It is at this point that we specialize to the two explicit cases. By assumption ${\Theta} = \Theta_k$, for some $1\leq k \leq \floor{\frac{n}{2}}$, as in \Cref{max-theta-split}.
Suppose that $s\in S_\Theta^- \cap A_{\Theta\cap w\Omega}^1A_{w\Omega}$. We want to show that $s \in S_\Theta^1 S_{\Delta_0}$.
Notice that $S_{\Delta_0} = \{ e\} $; therefore, it is sufficient to prove that $s\in S_\Theta^1$.
By assumption, $s = tz$ where $t\in A_{\Theta\cap w\Omega}^1$ and $z \in A_{w\Omega}$. Since $w\in [W_\Theta \backslash W_0 / W_\Omega]$, we have that $w\Omega \subset \Phi_0^+$;
moreover, by the assumption that $M_{\Theta\cap w\Omega}$ is a proper Levi subgroup of $M_{w\Omega}$,
we have that ${\Theta} \cap w\Omega \subsetneq w\Omega$ is a proper subset.
It follows that $w\Omega$ cannot be contained in $\Phi_\Theta^+$.
Moreover, there exists $\alpha \in w\Omega \setminus ({\Theta} \cap w\Omega)$ such that $\alpha \in \Phi_0^+$ and $\alpha \notin \Phi_\Theta^+$.
In the Galois case, there is a unique expression
$\alpha = \sum_{j=1}^{n-1} c_j {}^\gamma (\epsilon_j -\epsilon_{j+1})$, where $c_j \in \Z$ and $c_j \geq 0$,
such that, since ${\Theta} = \Theta_k$ and $\alpha \notin \Phi_\Theta^+$, at least one of $c_k$ or $c_{n-k}$ is nonzero ($c_{n/2} \neq 0$, when $n$ even, $k=n/2$).
In the linear case, $\gamma$ doesn't appear in the expression for $\alpha$.
First observe that
$\alpha(s) = \alpha(t)\alpha(z) = \alpha(t) \in \of^\times$,
since $z\in A_{w\Omega}$ and $t\in A_{\Theta\cap w\Omega}^1 = A_{\Theta\cap w\Omega}(\of)$.
On the other hand,
in the Galois case, writing $s$ explicitly as $s = {}^\gamma s'$, where
\begin{align*}
s' & =
\left\{ \begin{array}{ll} \diag(\underbrace{a,\ldots,a}_{k}, \underbrace{1\ldots,1}_{\floor{\frac{n}{2}}-2k},\underbrace{a^{-1},\ldots, a^{-1}}_{k}), & 1\leq k \leq \floor{\frac{n}{2}} \vspace{0.25cm}\\
\diag(\underbrace{a,\ldots,a}_{n/2}, \underbrace{a^{-1},\ldots, a^{-1}}_{n/2}), & n \ \text{even}, k=n/2
\end{array} \right.
\end{align*}
(with $s = s'$ and $n$ even in the linear case). Applying $\alpha$ to $s$, we have
\begin{align*}
\alpha(s) & = \left(\sum_{j=1}^{n-1} c_j {}^\gamma (\epsilon_j -\epsilon_{j+1})\right) ({}^\gamma s')
= (\epsilon_k- \epsilon_{k+1}) (s')^{c_k} (\epsilon_{n-k}- \epsilon_{n-k+1}) (s')^{c_{n-k}}
= a^{c_k} a^{c_{n-k}}
\end{align*}
So in the Galois case, $\alpha(s) = a^c \in \of^\times$, where $c = c_k + c_{n-k}$ (or $c=2 c_{\frac{n}{2}}$ when $n$ is even and $k = n/2$), and similarly in the linear case.
Then we have $|a|_F^c = 1$ for $c$ a positive integer
so $|a|_F = 1$. In particular, we have that $a\in \of^\times$, and $s\in S_\Theta^1 = S_\Theta(\of)$, as desired.
\end{proof}
\begin{prop}\label{prop-reduce-Cas-ind}
If $y=ww_0^{-1} \in [W_\Theta \backslash W_0 / W_\Omega]\cdot w_0^{-1}$ is such that
$P_\Theta\cap {}^w M_\Omega$ is a proper parabolic subgroup of ${}^w M_\Omega$, then the exponents $\chi \in \Exp_{S_\Theta}(\mathcal F_\Theta^y(\tau))$ of $\mathcal F_\Theta^y(\tau)$ satisfy
$$|\chi(s)|_F < 1,$$ for all $s \in S_\Theta^- \setminus S_\Theta^1 S_{\Delta_0}$.
\end{prop}
\begin{proof}
First, by \Cref{technical-torus-nbhd}, we have that $S_\Theta^- \setminus S_\Theta^1 S_{\Delta_0} \subset A_{\Theta\cap w\Omega}^{-w\Omega} \setminus A_{\Theta\cap w\Omega}^1A_{w\Omega}$.
By \Cref{closed-subgp-centre}, any exponent $\chi \in \Exp_{S_\Theta}(\mathcal F_\Theta^y(\tau))$ is the restriction to $S_\Theta$ of an exponent $\widehat\chi \in \Exp_{A_\Theta}(\mathcal F_\Theta^y(\tau))$; therefore, the result follows from \Cref{red-to-Cas-ind-data}.
\end{proof}
Finally, we study $M_\Theta^\theta$-distinction of the Jacquet module $\pi_{N_\Theta}$.
In preparation for this, we characterized the $\theta$-fixed points of the standard Levi $M_r=M_{\Theta_r}$ of the maximal $\theta$-split parabolic subgroups $P_r=P_{\Theta_r}$ in \Cref{prop-M-Theta-fixed-pts}.
The characterization is in terms of the groups $M_{(r,n-2r,r)}$ and $H_{(r,n-2r,r)}$.
First, we note \Cref{H-bullet-dist}, which characterizes $H_{(r,n-2r,r)}$-distinction in the Galois case. We omit the elementary verification (and the obvious modification when $n$ is even and $r=n/2$).
\begin{lem}\label{H-bullet-dist}
Assume that we are in the Galois case. Fix an integer $1\leq r \leq \floor{\frac{n}{2}}$ and assume that $r \neq n/2$.
Let $\pi_1\otimes \pi_2 \otimes \pi_3$ be an irreducible admissible representation of $M_{(r,n-2r,r)}$. Then $\pi_1\otimes \pi_2 \otimes \pi_3$ is $H_{(r,n-2r,r)}$-distinguished if and only if $\pi_2$ is $H_{n-2r}$-distinguished and $\pi_3 \cong {}^{\theta_k} \widetilde\pi_1$.
\end{lem}
\begin{prop}\label{no-dist-unit-exp}
Let $\tau \cong \bigotimes_{i=1}^k \tau_i$ be an irreducible admissible regular representation of $L$.
Assume that $\tau$ is $L^{\theta}$-distinguished.
Let $P_\Theta$ be a maximal $\Delta_0$-standard parabolic subgroup corresponding to a maximal $\theta$-split subset $\Theta$ of $\Delta_0$.
Let $y = ww_0^{-1}$, where $w \in [W_\Theta \backslash W_0 / W_\Omega]$.
In both Case (A) and Case (B), $\mathcal F_\Theta^y(\tau)$ cannot be $M_\Theta^{\theta}$-distinguished.
\end{prop}
\begin{proof}
By assumption, we are in either Case (A) or Case (B) of \Cref{rmk-two-cases-lin-Gal}. In particular, we have that $y = ww_0^{-1}$, where $w\in [W_\Theta \backslash W_0 / W_\Omega]$, such that
\begin{align*}
\mathcal F_\Theta^y(\tau) & = \iota_{M_\Theta \cap {}^yQ}^{M_\Theta} ({}^y \tau) = \iota_{M_\Theta \cap {}^w M_\Omega}^{M_\Theta} ({}^w \tau_0)
\end{align*}
and
\begin{align*}
\tau_0 = {}^{w_0^{-1}} \tau = {}^\gamma \left( \bigotimes_{i=1}^k \tau_i \right)
\end{align*}
is the representation of $M_\Omega$ corresponding to the representation $\tau$ of $L = L_{\Omega^{ell}} = w_0 M_\Omega w_0^{-1}$.
Where $\gamma$ simply doesn't appear in the linear case.
By \Cref{lem-L-theta-dist}, $\tau$ is $L^\theta$-distinguished if and only if each $\tau_i$ is $H_{m_i}$-distinguished for all $1 \leq i \leq k$.
By \Cref{lem-max-split-subsets}, $\Theta$ is equal to $\Theta_r$ for some $1 \leq r \leq \floor{\frac{n}{2}}$ and $M_\Theta = M_{\Theta_r} =M_r$.
Without loss of generality $r < n/2$.
By \Cref{max-theta-split}, in the linear case $M_r = M_{(r,n-2r,r)}$ and in the Galois case $M_r = {}^\gamma M_{(r,n-2r,r)}$.
We again use the shorthand $M_\bullet = M_{(r,n-2r,r)}$ and $H_\bullet = H_{(r,n-2r,r)}$.
The $\theta$-fixed point subgroup of $M_r$ is described in \Cref{prop-M-Theta-fixed-pts}. In the linear case, we have that
\begin{align*}
M_r^\theta =H_\bullet = \{ \diag(A, B, \theta_r(A)) : A \in G_r, B \in H_{n-2r} \},
\end{align*}
while in the Galois case, we have
\begin{align*}
M_r^\theta = \gamma \gamma_\bullet^{-1} H_\bullet \gamma_\bullet \gamma^{-1},
\end{align*}
where $\gamma_\bullet = \gamma_{(r,n-2r,r)} = \diag(\gamma_r,\gamma_{n-2r},\gamma_r) \in M_{(r,n-2r,r)}$.
By \Cref{orbit-dist}, in the Galois case, $\mathcal F_\Theta^y(\tau)$ is $M_\Theta^{\theta}$-distinguished if and only if ${}^{\gamma^{-1}}\mathcal F_\Theta^y(\tau)$ is $H_\bullet$-distinguished.
Without loss of generality, we complete the proof in the Galois case.
To obtain the proof in the linear case replace the application of \Cref{Gal-dist-property} with \Cref{lin-dist-property}.
\textbf{Case (A).}
Suppose that ${\Theta}$ and $\Omega$ are associate and $w\in [W_\Theta \backslash W_0 / W_\Omega] \cap W(\Theta, \Omega)$. Then $M_\Theta = {}^w M_\Omega$ and $\mathcal F_\Theta^y(\tau) = {}^w\tau_0 = {}^{w\gamma}(\tau_1\otimes \tau_2 \otimes \tau_3)$, where $\tau_1,\tau_3$ are representations of $G_r$ and $\tau_2$ is a representation of $G_{n-2r}$. By convention, $\gamma^{-1}w \gamma$ is a permutation matrix (\textit{cf.}~\Cref{Weyl-gp-conjugate}).
Moreover, ${}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau)) = {}^{\gamma^{-1}w \gamma}(\tau_1\otimes \tau_2 \otimes \tau_3)$ is equal to $\tau_{x(1)}\otimes \tau_{x(2)} \otimes \tau_{x(3)}$, for some compatible permutation $x$ of $\{1,2,3\}$.
By \Cref{H-bullet-dist}, ${}^{\gamma^{-1}}\mathcal F_\Theta^y(\tau)$ is $H_\bullet$-distinguished if and only if $\tau_{x(3)} \cong {}^{\theta_r}\widetilde{\tau_{x(1)}}$ and $\tau_{x(2)}$ is $H_{n-2r}$-distinguished. By \Cref{Gal-dist-property}, each $\tau_i$ satisfies $\tau_i \cong {}^{\theta_{m_i}}\widetilde{\tau_i}$. However, since $\tau$ is regular, \Cref{pairwise-inequiv} implies that the $\tau_i$ are pairwise inequivalent. In particular, we have that
\begin{align*}
{}^{\theta_r}\widetilde{\tau_{x(1)}} \cong \tau_{x(1)} \ncong \tau_{x(3)};
\end{align*}
therefore, ${}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau))$ is not $H_\bullet$-distinguished and $\mathcal F_\Theta^y(\tau)$ is not $M_\Theta^{\theta}$-distinguished.
\textbf{Case (B).}
Suppose that $w\in [W_\Theta \backslash W_0 / W_\Omega]$ is such that $wM_\Omega w^{-1} \subset M_\Theta$ is a proper Levi subgroup.
In this case, $\mathcal F_\Theta^y(\tau) = \iota_{M_\Theta\cap{}^wP_\Omega}^{M_\Theta} {}^w \tau_0$ and
$\tau_0$ is a regular irreducible unitary representation. By \cite[Theorem 6.6.1]{Casselman-book}, the representation ${}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau))$ is an irreducible unitary representation of $M_\bullet$.
Indeed, $M_\Theta = {}^\gamma M_\bullet$ and $M_\Omega = {}^\gamma M$, with $M_\bullet = M_{(r,n-2r,r)}$ and $M=M_{(m_1,\ldots,m_k)}$;
moreover, $w' = {}\gamma^{-1} w \gamma$ is an element of $[W_{M_\bullet} \backslash W / W_{M}]$ and conjugates $M$ into $M_\bullet$.
Writing $P$ for $P_{(m_1,\ldots,m_k)}$, we have
\begin{align*}
{}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau)) & = {}^{\gamma^{-1}}\left(\iota_{M_\Theta\cap{}^wP_\Omega}^{M_\Theta} {}^w \tau_0\right)
= {}^{\gamma^{-1}}\left(\iota_{{}^\gamma M_\bullet \cap{}^{w\gamma}P}^{{}^\gamma M_\bullet} {}^{w\gamma} (\tau_1\otimes\ldots \otimes \tau_k)\right)
\cong \iota_{M_\bullet \cap{}^{w'}P}^{M_\bullet} {}^{w'} (\tau_1\otimes\ldots \otimes \tau_k)
\end{align*}
and this representation is isomorphic to $\pi_1 \otimes \pi_2 \otimes \pi_3$,
where $\pi_1, \pi_2$ are irreducible admissible representations of $G_r$ and $\pi_2$ is an irreducible admissible representation of $G_{n-2r}$.
Since $w' \in [W_{M_\bullet} \backslash W / W_{M}]$, by \Cref{casselman1-3-3},
the group $M_\bullet \cap{}^{w'}P$ is a parabolic subgroup of $M_\bullet$
(a product of parabolic subgroups on each block of $M_\bullet$).
It follows that each of the $\pi_j$, $j=1,2,3$, are irreducibly induced representations of the form $\tau_{a_1} \times \ldots \times \tau_{a_r}$, for some subset of the representations $\{\tau_1, \ldots, \tau_k\}$. Again, by \Cref{H-bullet-dist} ${}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau))$ is $H_\bullet$-distinguished if and only if $\pi_2$ is $G_{n-2r}$-distinguished and $\pi_3 \cong {}^{\theta_r} \widetilde{\pi_1}$. Suppose that $\pi_1 = \tau_{a_1} \times \ldots \times \tau_{a_{l}}$, and $\pi_3 = \tau_{b_1} \times \ldots \times \tau_{b_{s}}$, then we have that
\begin{align*}
\widetilde \pi_1 & \cong \widetilde {\tau_{a_1}} \times \ldots \times \widetilde{\tau_{a_{l}}}
\cong {}^{\theta_{m_{a_1}}}{\tau_{a_1}} \times \ldots \times {}^{\theta_{m_{a_l}}}{\tau_{a_{l}}}
\cong {}^{\theta_r} \pi_1
\end{align*}
Moreover, we have that $\pi_1 \cong {}^{\theta_r} \widetilde \pi_1$.
Since $\tau$ is regular, by \Cref{pairwise-inequiv}, the discrete series $\tau_i$ are pairwise inequivalent; therefore, by \cite[Theorem 9.7(b)]{zelevinsky1980}, we have that $\pi_1 \ncong \pi_3$.
That is, we have ${}^{\theta} \widetilde{\pi_1} \cong \pi_1 \ncong \pi_3$.
In particular, ${}^{\gamma^{-1}}(\mathcal F_\Theta^y(\tau))$ is not $H_\bullet$-distinguished, and $\mathcal F_\Theta^y(\tau)$ is not $M_\Theta^{\theta}$-distinguished.
\end{proof}
\subsection*{Acknowledgements}
The author would like to thank his doctoral advisor, Fiona Murnaghan, for her support and guidance throughout the work on this project. Thank you to U.~K.~Anandavardhanan and Yiannis Sakellaridis for helpful comments.
Thank you to Shaun Stevens for pointing out an error in a previous version of \Cref{prop-lin-even-dist-ds}\eqref{prop-lin-even-dist-ds--4}. Finally, thank you to the anonymous referee for many helpful suggestions that improved this article.
\bibliographystyle{amsplain}
|
\section{Introduction}
Let $\mathbb{N}$ denote the set of nonnegative integers and $\mathbb{Z}$ be the set of integers. For $A,B\subseteq \mathbb{Z}$ and $k \in \mathbb{Z}$, let $A+B=\{a+b:~a\in A,~b\in B\}$ and $kA = \{ka: a \in A\}$.
If $A+B=\mathbb{Z}$, then $A$ is called an additive complement to
$B$ in $\mathbb{Z}$. If no proper subset of $A$ is a complement to
$B$, then $A$ is called a minimal complement to $B$ in
$\mathbb{Z}$.
It is easy to see that if $A\subseteq \mathbb{Z}$ is a (minimal) complement to $B \subseteq \mathbb{Z}$, then $A$ is also a (minimal) complement to $B+d$, $d\in \mathbb{Z}$, where $B+d=\{ b+d:b\in B\} $.
In 2011, Nathanson \cite{nath} proved the following theorem.
\noindent{\bf Nathanson's theorem} {\rm (See \cite[Theorem
8]{nath}).} {\em Let $W$ be a nonempty, finite set of integers.
In $\mathbb{Z}$, every complement to $W$ contains a minimal
complement to $W$.}
In the same paper, Nathanson also posed the following problem.
\noindent{\bf Problem} {\rm (See \cite[Problem 11]{nath}).} Let
$W$ be an infinite set of integers. Does there exist a minimal
complement to $W$? Does there exist a complement to $W$ that does not
contain a minimal complement?
For the second part of the above problem, in 2012, Chen and Yang \cite{chenyang} gave two infinite sets $W_1$ and $W_2$ of integers such that there exists a complement to $W_1$ that does not contain a minimal complement and every complement to $W_2$ contains a minimal complement.
For the first part of the above problem, in 2012, Chen and Yang \cite{chenyang} proved the following results.
\noindent{\bf Theorem A} (See \cite[Theorem 1]{chenyang}).~{\em
Let $W$ be a set of integers with $\inf W=-\infty$ and $\sup
W=+\infty$. Then there exists a minimal complement to $W$.}
By Theorem A, now we only need to consider the cases $\inf W>-\infty$
or $\sup W<+\infty$. Without loss of generality, we may assume
that $\inf W>-\infty$.
\noindent{\bf Theorem B} (See \cite[Theorem 2]{chenyang}).~{\em
Let $W=\{1=w_1<w_2<\cdots\}$ be a set of integers and
$$\overline{W}=(\mathbb{Z} \cap (0, +\infty)) \setminus W=\{\overline{w_1}
<\overline{w_2}<\cdots\}.$$
(a) If $\limsup_{i\rightarrow +\infty}(w_{i+1}-w_i)=+\infty$, then
there exists a minimal complement to $W$.
(b) If $\lim_{i\rightarrow
+\infty}(\overline{w_{i+1}}-\overline{w_i})=+\infty$, then there
does not exist a minimal complement to $W$.}
Let $W=\cup_{k=0}^{\infty} [10^k,2\times 10^k]$. Then it is clear
that both $\limsup_{i\rightarrow +\infty}(w_{i+1}-w_i)=+\infty$
and $\limsup_{i\rightarrow
+\infty}(\overline{w_{i+1}}-\overline{w_i})=+\infty$ hold. Hence
$\lim_{i\rightarrow
+\infty}(\overline{w_{i+1}}-\overline{w_i})=+\infty$ in Theorem B
(b) cannot be changed to $\limsup_{i\rightarrow
+\infty}(\overline{w_{i+1}}-\overline{w_i})=+\infty$.
In this paper, we will give further results on Nathanson's problem
and deal with some sets $W$ do not satisfy the conditions of
Theorem B.
First we give some definitions. Let $S\subseteq \mathbb{N}$. Denote by $S\mod m$ the set of residues of $S$ modulo $m$, i.e.,
$$S\mod m =\{ r: r\in \{0,1,\dots ,m-1\}, r\equiv s~(\text{mod}~m)\text{ for some }s\in S\}.$$
Let $X\subseteq \mathbb{N}$. If
there exists a positive integer $T$ such that $x+T\in X$ for all
$x\in X$, then we call $X$ {\em{periodic with period $T$}}. If
$X\cup C$ is a periodic set for some finite set $C\subseteq
\mathbb{N}$, then we call $X$ {\em{quasiperiodic}}. If there exists
a positive integer $T$ such that $x+T\in X$ for all sufficiently
large integers $x\in X$, then we call $X$ {\em {eventually
periodic with period $T$}}. Clearly, a periodic set must be
quasiperiodic and a quasiperiodic set must be eventually periodic.
If $W$ is eventually periodic with $|\mathbb{N}\setminus W| = +\infty$, then both $\lim_{i\rightarrow
+\infty}(w_{i+1}-w_i)<+\infty$ and $\lim_{i\rightarrow
+\infty}(\overline{w_{i+1}}-\overline{w_i})<+\infty$ hold. Hence
$W$ does not satisfy the conditions of Theorem B.
Suppose that $W$
is an eventually periodic set and $m$ is a positive
period. By shifting a number, we may assume that $W$ has the
following structure:
\begin{eqnarray}\label{def1}W=(m\mathbb{N}+X_m)\cup
Y^{(0)}\cup Y^{(1)},\end{eqnarray}
where $X_m\subseteq \{0, 1, \dots{}, m-1\}, Y^{(0)}\subseteq \mathbb{Z}^{-},Y^{(1)}$ are finite sets with
$Y^{(0)}~\text{mod}~ m\subseteq
X_m$ and
$(Y^{(1)}~\text{mod}~m) \cap X_m = \emptyset$.
For example, if $W=\{2,4,7,8,9,12,13,17,18,22,23,27,28,\ldots\}$, then by shifting a number $5$, we may assume that
$$W=\{-3,-1,2,3,4,7,8,12,13,17,18,22,23,\ldots\}.$$ Hence $m=5$, $X_m=\{2,3\}$, $Y^{(0)}=\{-3\}$, $Y^{(1)}=\{-1,4\}$.
In this paper, we study that what conditions are needed to ensure the existence of a minimal complement to $W$. First we prove a sufficient condition.
\begin{theorem}\label{thm3} Let $W$ be defined in \eqref{def1}. If there exists a minimal complement
to $W$, then there exists $C\subseteq \{0,1,\dots ,m-1\}$ such that the
following two conditions hold:
(a) $C+(X_m\cup Y^{(1)}) \mod m=\{0,1,\dots ,m-1\}$;
(b) For any $c\in C$, there exists $y\in Y^{(1)}$ such that
$c+y\not \equiv c'+x~(\text{mod}~m)$ for any $c'\in C$ and $x\in X_m$.
\end{theorem}
\begin{remark}\label{remark2} By the proof of Theorem \ref{thm3},
we know that Theorem \ref{thm3} also holds when $Y^{(1)}$ is an
infinite set with $|Y^{(1)}\cap \mathbb{Z}^{-}|<+\infty$.
\end{remark}
Let $m=3$, $X_m=\{0\}$, $Y^{(1)}\subseteq 3\mathbb{N}+1$. By
Theorem \ref{thm3}, we have the following corollary.
\begin{corollary}\label{coro1} Let $Y\subseteq 3\mathbb{N}+1$ and
$W=3\mathbb{N}\cup Y$. Then there does not exist a minimal
complement to $W$.
\end{corollary}
\begin{remark}\label{remark2} We can choose an infinite set
$Y$ in Corollary \ref{coro1} such that $W$ is not eventually
periodic. Hence, there exists an infinite, not eventually periodic
set $W\subseteq \mathbb{N}$ such that $w_{i+1}-w_i\in \{1,2,3\}$
for all $i$, and there does not exist a minimal complement to $W$.
\end{remark}
\begin{remark}\label{remark1} If $W\subseteq \mathbb{N}$ is a
quasiperiodic set, then $Y^{(1)}=\emptyset$ and the condition (b)
in Theorem \ref{thm3} does not hold. Hence there does not exist a
minimal complement to $W$.
\end{remark}
In the next step we prove a necessary condition.
\begin{theorem}\label{thm4} Let $W$ be defined in \eqref{def1}. Suppose that there exists $C\subseteq
\{0,1,\dots ,m-1\}$ such that the following two conditions hold:
(a) $C+(X_m\cup Y^{(1)})\mod m=\{0,1,\dots ,m-1\}$;
(b) For any $c\in C$, there exists $y\in Y^{(1)}$ such that $c+y\not \equiv c'+x~(\text{mod}~m)$ for
any $c'\in C \setminus \{c\}$ and $x\in X_m \cup Y^{(1)}$.
Then there exists a minimal complement to $W$.
\end{theorem}
By Theorems \ref{thm3} and \ref{thm4}, we have the following
corollary.
\begin{corollary} Let $W=(m\mathbb{N}+X_m)\cup Y^{(0)}\cup \{a\},$
where $X_m\subseteq \{0,1,\dots ,m-1\},~(Y^{(0)}\mod m)\subseteq X_m$ and $a\not \equiv x~(\text{mod}~m)$ if $x\in X_m$. Then there exists a minimal complement to $W$
if and only if there exists a subset $C\subseteq \{0,1,\dots ,m-1\}$ such that:
(a) $C+(X_m\cup\{ a\})\mod m =\{0,1,\dots ,m-1\}$;
(b) For any $c\in C$, $c+a\not \equiv c'+x~(\text{mod}~m)$, where $c'\in C \setminus \{c\}$ and $x\in X_m$.
\end{corollary}
We see that Theorems \ref{thm3} and \ref{thm4} transfer
Nathanson's problem into a finite modulo version when $W$ is an eventually periodic set.
In the next theorem, we give a sufficient and necessary condition, but we
cannot bound the module.
\begin{theorem}\label{thm5} Let W be defined in (1). There exists a minimal complement to $W$
if and only if there exists $T\in \mathbb{Z}^{+},m\mid T$,
and $C\subseteq \{0,1,\dots ,T-1\}$ such that
(a) $C+(X_T\cup Y^{(1)})\mod T=\{0,1,\dots ,T-1\}$, where $X_{T} = \cup_{i=0}^{\frac{T}{m}-1}\{im+X_{m}\}$;
(b) for any $c\in C$, there exists $y\in Y^{(1)}$ such that
$c+y\not \equiv c'+x~(\text{mod}~T)$ for any $c'\in C \setminus \{c\}$ and $x\in X_T\cup Y^{(1)}$.
\end{theorem}
Finally, as a complement to Remark \ref{remark2}, we give the
following theorem.
\begin{theorem}\label{thm2} There exists an infinite, not
eventually periodic set $W\subseteq \mathbb{N}$ such that
$w_{i+1}-w_i\in \{1,2\}$ for all $i$ and there exists a minimal
complement to $W$.
\end{theorem}
Now we pose two problems for further research.
\begin{problem} We know that Theorem \ref{thm3} also holds when
$Y^{(1)}$ is infinite. Is Theorem \ref{thm4} also true when
$Y^{(1)}$ is infinite?
\end{problem}
\begin{problem} Does there exist an explicit formula for the upper bound of $T$ in Theorem 3
using $m,Y^{(0)}$ and $Y^{(1)}$?
\end{problem}
\section{Proofs}
\begin{proof}[Proof of Theorem \ref{thm3}] Suppose that $D$ is a
minimal complement to $W$. For $i\in \{0,1,\ldots,m-1\}$, let
$D_i=\{d\in D:~d\equiv i~(\text{mod}~m)\}$ and
$$C=\{j:~0\le j\le m-1~\text{and}~|D_j\cap
\mathbb{Z}^{-}|=+\infty\}.$$ For any $t\in
\{0,1,\ldots,m-1\}\setminus C$, the set $\{d\in D:~d\equiv
t~(\text{mod}~m)\}+W$ does not contain any sufficiently small
negative integers. It follows from $D+W=\mathbb{Z}$ that
$C+W~\text{mod}~m =\{0,1,\dots ,m-1\}$. That is, $C+(X_m\cup Y^{(1)})~\text{mod}~m=\{0,1,\dots ,m-1\}$.
Next we shall prove (b). Suppose that there exists $c\in C$ such
that for any $y\in Y^{(1)}$ there exist $c'\in C$ and $x\in X_m$ with $c+y\equiv c'+x~(\text{mod}~m)$. We take an integer $d\in D$ such that $d\equiv
c~(\text{mod}~m)$ and we shall prove that $D\setminus \{d\}$ is
also a complement to $W$. For any integer $n$, write $n=d'+w$,
where $d'\in D$ and $w\in W$.
Case 1. $d'\not=d$. Then $n=d'+w\in (D\setminus \{d\})+W$.
Case 2. $d'=d$.
Subcase 2.1. $(\{w\}~\text{mod}~m) \subseteq X_m$. In this case, there exists
a positive integer $k_0$ such that $w+km\in W$ for all integers
$k\ge k_0$. Since $|D_c\cap \mathbb{Z}^{-}|=+\infty$, it follows
that there exists an integer $k\ge k_0$ such that $d-km\in D$.
Hence $n=(d-km)+(w+km)$, where $d-km\in D\setminus \{d\}$ and
$w+km\in W$. That is, $n\in (D\setminus \{d\})+W$.
Subcase 2.2. $w\in Y^{(1)}$. Since $\{c+y\}~\text{mod}~m\subseteq C+X_m ~\text{mod}~m$
for any $y\in Y^{(1)}$ and $d\equiv c~(\text{mod}~m),~w\in
Y^{(1)}$, it follows that ~$\{n\}~\text{mod}~m=\{d+w\}~\text{mod}~m\subseteq C+X_m ~\text{mod}~m$. Hence
there exist a $c'\in D$ with $c'~(\text{mod}~m)\in C$ and $x\in W$
with $x~\text{mod}~m\in X_m$ such that $n\equiv
~c'+x~(\text{mod}~m)$. We choose a sufficiently large integer $k$
such that $c'-km\in D$,~$c'-km\not=d$ and $x+km\in W$. Hence
$n=(c'-km)+(x+km)$, where $c'-km\in D \setminus \{d\}$ and $x+km\in W$.
Hence, $(D\setminus \{d\})+W=\mathbb{Z}$ which contradicts the fact that $D$
is a minimal complement. Therefore, (b) holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm4}] Let
$C_1=C+X_m~\text{mod}~m$,~$C_2=\{0,1,\dots ,m-1\}\setminus C_1$,
$$C'=\{d\in\mathbb{Z}:~d\equiv c~(\text{mod}~m)~\text{for some}~c\in C\},$$
$$C_1'=\{d\in\mathbb{Z}:~d\equiv c~(\text{mod}~m)~\text{for some}~c\in C_1\},$$
$$C_2'=\{d\in\mathbb{Z}:~d\equiv c~(\text{mod}~m)~\text{for some}~c\in C_2\}.$$
By (a), we have $C'+W=\mathbb{Z}$. Since $C+X_m~\text{mod}~m=C_1$, it follows
that $$C'+(W\setminus
Y^{(1)})~\text{mod}~m=C+X_m~\text{mod}~m=C_1,$$ and so
$\left(C'+(W\setminus Y^{(1)})\right)\cap C_2'=\emptyset$. It follows from (b) that $C_2' \ne \emptyset$. Noting
that $C'+W=\mathbb{Z}$, we have $C'+Y^{(1)}\supseteq C_2'$. Since
$Y^{(1)}$ is a finite set, by the proof of Nathanson's theorem
(See \cite[Theorem 4, page 2015]{nath}), there exists $D'\subseteq C'$ such
that $D'+Y^{(1)}\supseteq C_2'$ and for any $d\in D'$,
$$(D'\setminus \{d\})+Y^{(1)}\not\supseteq C_2'.$$
Next we shall prove that $D'$ is a minimal complement to $W$.
For $i\in C$, let $D_i'=\{d\in D':d\equiv i~(\text{mod}~m)\}$. First we prove that $|D_i'\cap \mathbb{Z}^{-}|=+\infty$ for all $i\in
C$. Suppose that there exists a $j\in C$ such that $|D_j'\cap
\mathbb{Z}^{-}|<+\infty$. By (b), there exists a $y\in Y^{(1)}$ such
that $j+y\not \equiv c+x~(\text{mod}~m)$, where $c\in C\setminus \{j\}$, $x\in X_m\cup Y^{(1)}$ and so
$$D'+Y^{(1)}\not\supseteq \{d\in \mathbb{Z}:d\equiv j+y~(\text{mod}~m)\}.$$
Noting that $(\{j+y\}~\text{mod}~m) \not\subseteq C+X_m ~\text{mod}~m=C_1$, we have $(\{j+y\}~\text{mod}~m)\subseteq
C_2$. It follows that $D'+Y^{(1)}\not\supseteq
C_2'$, a contradiction. Hence, $|D_i'\cap \mathbb{Z}^{-}|=+\infty$
for all $i\in C$.
Next we prove that $D^{'}$ is a complement.
For any integer $n\in C_1'$, by $C+X_m ~\text{mod}~m=C_1$, there exists $c\in C$
and $x\in X_m$ such that $n\equiv c+x~(\text{mod}~m)$. Since
$|D_c'\cap \mathbb{Z}^{-}|=+\infty$, there exists a sufficiently
small negative integer $d\in D_c'$ such that $n-d>0$. The congruences $n\equiv c+x ~(\text{mod}~m)$ and $d\equiv c ~(\text{mod}~m)$ imply that $n-d\equiv x ~(\text{mod}~m)$. Hence, $n-d\in m\mathbb{N}+X_m$ and so
$$n=d+(n-d)\in D_c'+(m\mathbb{N}+X_m)\subseteq D'+W.$$
Hence $C_1'\subseteq D'+W.$ On the other hand, $D'+W\supseteq
D'+Y^{(1)}\supseteq C_2'$. Therefore, $D'+W=\mathbb{Z}$.
Finally, we prove that $D^{'}$ is a minimal complement.
For any $d\in D'$,
we have $$\Big((D'\setminus \{d\})+(W\setminus
Y^{(1)})~\text{mod}~m\Big) \subseteq C+X_m ~\text{mod}~m=C_1.$$ It follows that $$\Big((D'\setminus
\{d\})+(W\setminus Y^{(1)})\Big)\cap C_2'=\emptyset,$$ and so
$(D'\setminus \{d\})+W\not\supseteq C_2'$. Hence $(D'\setminus
\{d\})+W\not=\mathbb{Z}$.
Therefore, $D'$ is a minimal complement to $W$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm5}] Assume that the set $W$ satisfies the conditions of Theorem 3.
Applying Theorem 2 with $m = T$, it follows that $W$ has a
minimal complement.
Suppose that $W$ has a minimal complement $E$. We
will prove that there exist a positive integer $T$ and a set $C
\subseteq \{0,1, \dots{}, T-1\}$ satisfying the conditions of Theorem 3.
For $0 \le i < m$, let
\[
E^{-}_{i} = \{e: e < 0,e\in E, e \equiv i~(\text{mod}~m)\}.
\]
Let $0 \le i_{1} < i_{2} < \dots{} < i_{t} < m$ be the sequence of
indices with $|E^{-}_{i_{j}}| = \infty$. It is clear that there
exists an integer $N_{0}$ such that, $e \in E$ and $e \le
N_{0}$ imply that $e \in E_{i_{j}}$ for some $i_{j}$. It follows from Theorem 1 that $Y^{(0)} \cup Y^{(1)} \ne \emptyset$. Let
\[
y_{+} = \text{max}\{y: y \in Y^{(0)} \cup Y^{(1)}\},
\]
\[
y_{-} = \text{min}\{y: y \in Y^{(0)} \cup Y^{(1)}\},
\]
and $y_{0} = \max\{y_{+},-y_{-},y_{+}-y_{-}\}$. Let $\chi_{E}(k)$ denote the
characteristic function of the set $E$, i.e.,
\[
\chi_{E}(k) = \left\{
\begin{aligned}
1 \textnormal{, if } k \in E; \\
0 \textnormal{, if } k \notin E.
\end{aligned} \hspace*{3mm}
\right.
\]
Let $A = N_{0} + \text{min}\{0, y_{-}\}$. We consider the
following vectors:
\begin{align*}
\textbf{v}_{A} & = \left( \chi_{E}(A-y_0),\chi_{E}(A-y_0+1), \dots{}, \chi_{E}(A+y_0)\right), \\
\textbf{v}_{A-m} & = \left( \chi_{E}(A-m-y_0),\chi_{E}(A-m-y_0+1), \dots{}, \chi_{E}(A-m+y_0)\right), \\
&\mathrel{\makebox[\widthof{=}]{\vdots}} \\
\textbf{v}_{A-im} & = \left( \chi_{E}(A-im-y_0),\chi_{E}(A-im-y_0+1), \dots{}, \chi_{E}(A-im+y_0)\right) ,\\
&\mathrel{\makebox[\widthof{=}]{\vdots}}
\end{align*}
It is clear that there are infinitely many vectors
$\textbf{v}_{A}, \textbf{v}_{A-m}, \dots{}, \textbf{v}_{A-im},
\dots{},$ each of them has $2y_{0}+1$ coordinates, which are $0$ or
$1$. Since there are at most $2^{2y_{0}+1}$ different vectors,
by the pigeon hole principle, there exist a vector $\textbf{v}$
and an infinite sequence $0\le k_{1} < k_{2} <
\cdots{}$ such that $\textbf{v}_{A-k_{i}m} = \textbf{v}$ for all positive integer $i$.
Define $L = A-k_{1}m$ and choose a sufficiently large integer $k_{i}$ such that
$k_{i}m - k_{1}m \ge y_{0}$
and $[A-k_{i}m, A-k_{1}m[ \cap E_{i_{j}} \ne \emptyset$ for every index $i_{j}$. Let $K = A - k_{i}m$,
$T = L-K$ and
\[
C = \{l: K \le l < L, ~ l \in E\}~\text{mod}~T.
\]
Now we shall prove that such an integer $T$ and a set $C$ satisfy the conditions of Theorem 3.
By definitions, $K$ and $L$ have the following properties.
\begin{equation}
L \le N_{0} + \text{min}\{0, y_{-}\},
\end{equation}
\begin{equation}
L-K\ge y_{0},
\end{equation}
\begin{equation}
m \mid L - K,
\end{equation}
\begin{equation}
\chi_{E}(K+i) = \chi_{E}(L+i)~ for~
-y_{0} \le i \le +y_{0},
\end{equation}
\begin{equation}
[K, L[ \cap E_{i_{j}} \ne \emptyset~\text{for all}~i_j.
\end{equation}
First we prove that $C + (X_{T} \cup Y^{(1)})~\text{mod}~T =
\{0,1,\ldots, T-1\}$, where $X_{T} = \bigcup_{i=0}^{\frac{T}{m}-1}\{im+X_{m}\}$.
For any integer $l$ with $K \le l < L$, there exist $e \in E$ and $w \in W$ such that $l = e + w$.
As $w \ge \text{min}\{0, y_{-}\}$,
it follows from (2) that $e = l - w < L - \text{min}\{0, y_{-}\}
\le N_{0}$, and so $e \in E_{i_{j}}^{-}$ for some integer $i_j$.
Suppose that $w \in Y^{(0)} \cup Y^{(1)}$. Then we have $y_{-}
\le w \le y_{+}$ and $K-y_{+}
\le e =l-w< L-y_{-}$. We have three cases.
Case 1. $K - y_{+} \le e < K$. Noting that
$$K-y_0\le K-y_+\le e<K<K+y_0,$$
by (5), we have $e+(L-K)\in E$. By $K - y_{+} \le e < K$ and (3), it follows that
$K\le L-y_+\le e+(L-K)<L$. Let $c=e+(L-K)$. Then $c~\text{mod}~T \in C$ and $l\equiv c+w~(\text{mod}~T)$.
Hence $l~\text{mod}~T\in C + (X_{T} \cup Y^{(1)})~\text{mod}~T$.
Case 2. $K \le e < L$. It follows that $e~\text{mod}~T\in C$ and so
$l~\text{mod}~T \in C + (X_{T} \cup Y^{(1)})~\text{mod}~T$.
Case 3. $L \le e < L - y_{-}$. Noting that
$$L-y_0\le L\le e<L-y_{-}\le L+y_0,$$
by (5), we have $e - (L - K) \in E$. Since $L \le e < L - y_{-}$ and (3), it follows that
$K \le e - (L - K) < K - y_{-}<L$. Let $c=e-(L-K)$. Then $c~\text{mod}~T \in C$ and
$l\equiv c+w~(\text{mod}~T)$. Hence $l~\text{mod}~T\in C + (X_{T} \cup Y^{(1)})~\text{mod}~T$.
Suppose that $w \in T\mathbb{N} + X_{T}$. Since $w \ge 0$ and $e = l - w < L \le N_{0}$, we have $e \in E_{i_{j}}^{-}$ for some integer
$i_j$, and so $e \equiv i_{j}~(\text{mod}~m)$. It follows from (6) that
there exists an integer $e^{'}$ such that $e^{'} \in E_{i_{j}}^{-}$, $K \le
e^{'} < L$ and $e^{'} \equiv i_{j}~(\text{mod}~m)$. Let $w
\equiv x~(\text{mod}~m)$, where $x\in X_m$. Obviously, $l
\equiv e + w \equiv i_{j} + x \equiv e^{'} + x~(\text{mod}~m)$. By
(4) there exists an integer $u$ with $0 \le u < \frac{T}{m}$
such that $l \equiv e^{'} + um + x~(\text{mod}~T)$. Thus $l~\text{mod}~T\in C + (X_{T} \cup Y^{(1)})~\text{mod}~T$.
Therefore, $C + (X_{T} \cup Y^{(1)})~\text{mod}~T=\{0,1,\ldots,T-1\}$.
In the next step we show that the second condition of Theorem 3
holds. For any integer $e\in E$ with $K \le e < L$, there
exists a $w \in W$ such that $e + w \ne e^{'} + w^{'}$ for any $e' \ne
e$, $e'\in E$ and $w'\in W$.
If $w \in (m\mathbb{N} + X_{m}) \cup Y^{(0)}$, by $e<L\le N_0$ and $e\in E$,
then there
exists a positive integer $s$ such that $e + w = (e - sm) + (w +
sm)$, where $e - sm \in E$ and $w + sm \in W$. This is a contradiction.
Now we may assume that $w \in Y^{(1)}$. It is enough to prove
that $e + w \not\equiv e^{'} + w^{'}~(\text{mod}~T)$ for any $e^{'}(\not=e)\in E$, $K \le
e^{'} < L$ and $w^{'} \in X_{T} \cup (Y^{(1)}~\text{mod}~T)$. Suppose that such $e'$ and $w'$
exist, i.e., $e+w\equiv e'+w'~(\text{mod}~T)$.
If $w^{'} \in X_{T}$, then $e + w = e^{'} + w^{'}
+ tT$ for some integer $t$. Hence there exists a positive integer $s$ such that
$e + w = (e^{'} - sm) + (w^{'} + sm + tT)$, where
$e^{'} - sm \in E$ and $w^{'} + sm + tT \in W$. This is a contradiction.
Now we assume $w^{'} \in Y^{(1)}$. Then $K + y_{-} \le e +
w, e^{'} + w^{'} < L + y_{+}$. By $T=L-K\ge y_0\ge y_+-y_-$, it follows that either $e + w = e^{'} +
w^{'}$ or $e + w = e^{'} + w^{'} +T$ or $e + w = e^{'} +
w^{'} - T$.
Case 1. $e + w = e^{'} + w^{'}$. Then we have $e =
e^{'}$ and $w = w^{'}$, a contradiction.
Case 2. $e + w = e^{'} +
w^{'} +T$. It follows that $K + y_{-} \le e^{'} + w^{'} < K +
y_{+}$, and so $$K-y_0\le K + y_{-} - y_{+} \le e^{'} < K + y_{+} -
y_{-}\le K+y_0.$$
By (5), we have $e^{'} + T\in E$, and then $e + w$ has another representation $(e^{'} +T) + w^{'}$.
Therefore $w
= w^{'}$ and $e = e^{'} + T$, which contradicts with $K \le
e, e^{'} < L$.
Case 3. $e + w = e^{'} + w^{'} - T$. Then we have
$L + y_{-} \le e^{'} + w^{'} < L + y_{+}$. Thus $L-y_0\le L +
y_{-} - y_{+} \le e^{'} < L + y_{+} - y_{-}\le L+y_0$. It follows from
(5) that $e^{'} - T \in E$ which implies that $e + w =
(e^{'} - T) + w^{'}$. Therefore $w = w^{'}$ and $e =
e^{'} -T$, which is a contradiction because $K \le e, e^{'}
< L$.
The proof of Theorem 3 is completed.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm2}]
By induction we can construct
$\{d_i\}_{i=1}^{\infty},~\{W_i\}_{i=1}^{\infty}$ and
~$\{c_i\}_{i=1}^{\infty}$ such that
(i) $d_1=-1,~W_1=\{1,2,\ldots,12\},~c_1=-3$;
(ii) $d_i$ is the largest negative integer $\not\in
W_{i-1}+\{c_1,c_2,\ldots,c_{i-1}\}$ for $i\ge 2$;
(iii) $c_i<d_i+2c_{i-1}$ for all $i\ge 2$;
(iv) for $i\ge 2$, let $W_i=W_{i-1}\cup
\left([-2c_{i-1},-2c_i-1]\setminus \cup_{j=1}^{i-1}
\{-c_i+d_{j}\}\right).$
Let $W=\cup_{i=1}^{\infty}W_i$ and $C=\{c_i\}_{i=1}^{\infty}$.
Now we prove that $C$ is a minimal complement to $W$.
First we prove $d_{i+1}-d_i\le -2$ for all integers $i\ge 1$.
Clearly $d_2=-3,~d_2-d_1=-2$. Suppose that $d_{i+1}-d_i\le -2$ for
all integers $i<k~(k\ge 2)$. Since
$$d_k=(d_k-c_k)+c_k,\quad d_k-1=(d_k-1-c_k)+c_k,$$
$$-2c_{k-1}\le d_k-1-c_k<d_k-c_k<-c_k+d_{k-1},$$
it follows that $d_k-c_k,~d_k-1-c_k\in W_k$ and then
~$d_k,d_k-1\in W_k+\{c_k\}$. Hence $d_{k+1}\le d_k-2$. By (iv), we
have $w_{j+1}-w_j\in \{1,2\}$. Since $d_k\rightarrow -\infty$, by (ii) we
have $(-\infty,9]\subseteq W+C$. For any integer $n\ge 10$, there
exists an $i$ such that $-c_{i-1}\le n<-c_i$. Hence
$$-c_i+d_1<-c_{i-1}-c_i\le n-c_i<-2c_i,$$
and so $n-c_i\in W_i$, that is, $n\in W_i+\{c_i\}$. Therefore,
$W+C=\mathbb{Z}$.
Next, we prove that the complement $C$ is minimal. For any
positive integer $i$, we write $d_i=c+w$ with $c\in C$ and $w\in
W$. Now we shall prove that $c=c_i$. By (iv), we have
$d_i-c_j\not\in W$ for all integers $j>i$. Hence $c\not=c_j$ for
all integers $j>i$. Since $-2c_{i-1}$ is the minimal value of
$W\setminus W_{i-1}$ and for any positive integers $j\le i-1$,
$d_i-c_j\le d_i-c_{i-1}<-2c_{i-1}$, it follows that
$d_i-c_j\not\in W\setminus W_{i-1}$ for all positive integer $j\le
i-1$. Noting that $d_i\not\in W_{i-1}+\{c_1,\ldots,c_{i-1}\}$, we
have $d_i\not\in W+\{c_1,c_2,\ldots,c_{i-1}\}$. Hence $c=c_i$.
Therefore, $C$ is a minimal complement to $W$. Furthermore, by
(iii), we can choose suitable $c_i$ such that $W$ is infinite and
not eventually periodic.
\end{proof}
\section{Acknowledgement} This work was done during the third author visiting to Budapest
University of Technology and Economics. He would like to thank Dr.
S\'andor Kiss and Dr. Csaba S\'{a}ndor for their warm hospitality. We would like
to thank the anonymous referee very much for the detailed comments.
|
\section{Introduction}
The \emph{probabilistic truth table task} was introduced by two independent studies at the beginning of this millennium \cite{evans03,oberauer03}. It can be seen as one of the starting points of the new (probability-based) paradigm psychology of reasoning \cite<see, e.g.>{baratgin14,elqayam16,oaksford07,over09,pfeifer13b,pfeifer14}, which started to replace the the old (classical logic-based) paradigm psychology of reasoning. The probabilistic truth table task was constructed to investigate how people interpret conditionals (i.e., indicative sentences of the form \emph{If $A$, then $C$}). As its name suggests, the task consists in inferring the degree of belief in a conditional based on probabilistic information attached to the truth table cases. What are truth table cases? Let $A$ and $C$ denote two propositions (i.e., sentences for which it makes sense to assign the truth values \emph{true} or \emph{false}) like \emph{a fair die is rolled} and \emph{the side of the die shows an even number}, respectively. The four truth table cases induced by $A$ and $C$ are: $A \wedge C$, $A \wedge \neg C$, $\neg A \wedge C$, and $\neg A \wedge \neg C$,
where $\wedge$ denotes conjunction (``and'') and $\neg$ denotes negation. Classical logic is bivalent, involving only the truth values \emph{true} and \emph{false}. Therefore, the conditional defined in classical logic (i.e., the \emph{material conditional}, see Table~\ref{TAB:ttables}) is either true or false. The \emph{conditional event} $C|A$ involved in conditional probability ($p(C|A)$), however, is not bivalent, as it is \emph{void} (or undetermined) if its antecedent $A$ is false (see Table~\ref{TAB:ttables}). Therefore, it cannot be represented by the means of classical logic.
Table~\ref{TAB:ttables} presents the truth conditions of the most important psychological interpretations for adult reasoning about indicative conditionals (i.e., conditional event, conjunction, and material conditional). Moreover, it presents the biconditional and biconditional event interpretations, which were reported in developmental psychological studies \cite<see, e.g.>{barrouillet15}.
Psychological evidence for the conditional event interpretation was already observed within the old paradigm psychology of reasoning \cite<see, e.g.,>{wason72}. The response pattern, which is consistent with the conditional event interpretation was seen as irrational (dubbed ``defective truth table''), as it violates the semantics of the material conditional. The material conditional used to be the gold standard of reference for the meaning of indicative conditionals in the old paradigm. However, within the new paradigm psychology of reasoning this response pattern is, of course, perfectly rational \cite{over17,pfeifertulkki17}.
\begin{table}[!ht]
\begin{center}
\caption{Truth tables for the material conditional $A \supset C$,
the conjunction $A\wedge C$, the biconditional $A\equiv C$, the biconditional event $C||A$ (i.e., $A\wedge C|A\vee C$), and the conditional event $C|A$. }
\label{TAB:ttables}
\vskip 0.12in
\begin{tabular}{ccccccc}\hline
$A$ & $C$ & $A \supset C$& $A\wedge C$& $A\equiv C$&$C||A$&$C|A$\\\hline
true &true &true &true &true & true&true \\
true &false &false &false &false & false&false \\
false &true &true &false &false& false &void \\
false &false &true &false &true&void & void \\\hline
\end{tabular}
\end{center}
\end{table}
While classical truth table tasks require to respond truth values, \emph{probabilistic} truth table tasks require to respond degrees of belief. Following \citeA{pfeifer13b}, we interpret the task as a probability logical inference problem. Specifically, it is formalised as a probability logical argument with assigned degrees of belief in the four truth table cases as its \emph{premises} and the degree of belief in a conditional as its \emph{conclusion}. The inference problem consists in propagating the uncertainties of the premises to the conclusion. As an example, consider the conditional probability interpretation of the ($p(C|A)$) conditional \emph{If $A$, then $C$} as the conclusion. This argument scheme is formalised by:
\begin{quote}
\begin{itemize}
\item[($\mathcal{A}$)]From $p(A \wedge C)=x_1$, $p(A \wedge \neg C)=x_2$, $p(\neg A \wedge C)=x_3$, and $p(\neg A \wedge \neg C)=x_4$, infer $p(C|A)=x_1/(x_1+x_2)$.
\end{itemize}
\end{quote}
In argument scheme~($\mathcal{A}$), the fraction $x_1/(x_1+x_2)$ is the probability propagation rule for the conditional probability. For different interpretations of the conditional, the probability propagation rules differ. Table~\ref{TAB:ProbProp} presents the corresponding probability propagation rules of the interpretations given in Table~\ref{TAB:ttables}.
\begin{table}[!ht]
\begin{center}
\caption{Probability propagation rules for the different interpretations of \emph{If $A$, then $C$}. The premise set $\{p(A \wedge C)=x_1, p(A \wedge \neg C)=x_2, p(\neg A \wedge C)=x_3, p(\neg A \wedge \neg C)=x_4\}$ entails the respective conclusion (see also Table~\ref{TAB:ttables}).}
\label{TAB:ProbProp}
\vskip 0.12in
\begin{tabular}{ll}\hline
Interpretation &Conclusion \\\hline
Material conditional&$p(A \supset C)=x_1+x_3+x_4$\\
Conjunction& $p(A\wedge C)=x_1$\\
Biconditional & $p(A\equiv C)=x_1+x_4$\\
Biconditional event& $p(C||A)=x_1/(x_1+x_2+x_3)$\\
Conditional event& $p(C|A)=x_1/(x_1+x_2)$ \\\hline
\end{tabular}
\end{center}
\end{table}
The main empirical result of classical probabilistic truth table tasks is that the dominant responses are consistent with the conditional event interpretation of conditionals. Moreover, if people solve the task several times, some people shift to the conditional event interpretation during the course of the experiment \cite<see, e.g.,>{fugard11a,pfeifer15a}.
A key feature of classical probabilistic truth table tasks is that they
present \emph{complete probabilistic knowledge} w.r.t.\ the truth table cases: all (point-valued) probabilities $x_1, \dots x_4$ are available to the participant \cite<see, e.g.,>{evans03,oberauer03,fugard11a,pfeifer15a}. When all probabilities involved in argument scheme~($\mathcal{A}$) are given as point values, for example, it is possible to infer a \emph{precise} (point-valued) probability of $C|A$. Of course, $x_3$ and $x_4$ are irrelevant for calculating $p(C|A)$ in this case. However, as full probabilistic information is usually not available in everyday life, we argue for \emph{investigating incomplete probabilistic knowledge}. If $x_1$ in argument scheme~($\mathcal{A}$) is only available as an \emph{imprecise} (i.e., interval-valued) probability, i.e., $x_1' \leq p(A\wedge C) \leq x_1''$, the probability of the conclusion of~($\mathcal{A}$) is also imprecise, i.e.,
$x_1'/(x_1'+x_2) \leq p(C|A) \leq x_1''/(x_1''+x_2).$
Table~\ref{TAB:interpretations} (see below) presents a numerical illustration of the different interpretations of conditionals in the probabilistic truth table task with imprecise premises. Incomplete probabilistic knowledge has not been investigated within the probabilistic truth table task paradigm yet \cite<for an exception, where only indicative conditionals were investigated, see>{pfeifer13b}.
With only few exceptions \cite<i.e.,>{over07b,pfeifer15a}, the important classes of causal conditionals and counterfactuals have not been investigated yet within the probabilistic truth table task paradigm. Causal conditionals are characterized by connecting cause (i.e., the conditional's antecedent) and effect (i.e., the conditional's consequent), like \emph{If you take aspirin, your headache will disappear}. Counterfactuals are conditionals in subjunctive mood, where the grammatical structure signals that the antecedent is false. For instance, \emph{If you were to take aspirin, your headache would disappear}, which signals that you had not taken aspirin yet. This example is a counterfactual version of the above described causal conditional. Of course, there are also non-causal versions of counterfactuals, like \emph{If the side of a card were to show an ace, it would show spades}.
Formally, the meaning of counterfactuals can be
modeled by the \emph{coherence-based probability logic} of nested conditionals
\cite{GiSa13c,gilio14,2016:SMPS1,GOPSsubm}. In a nutshell, a counterfactual is interpreted as a ``conditional (if $B$, then $C$) conditionalized on the factual statement ($A$)'', where the factual statement logically contradicts the antecedent of the conditional (i.e., $A \wedge B$ is logically false). Specifically, let
$A$, $B$, and $C$ denote three events, where $A$ is the factual statement which logically contradicts the antecedent $B$ of the conditional event $C|B$. It has been shown that the \emph{prevision of the conditional random quantity}
$(C|B)|A$ equals the \emph{conditional probability of the conditional event} $C|B$ \cite[see Example 1,
p. 225]{GiSa13c}. Thus, the counterfactual \emph{If $A$ were the case,
$C$ would be the case} can be modeled by the degree of belief in the
conditional random quantity $(C|A)|\neg A$ which equals to $p(C|A)$
(i.e., $Prevision((C|A)|\neg A)=p(C|A)$). Therefore, we hypothesize that the participants' degrees of belief in counterfactuals are equal to corresponding conditional
probabilities.
To our knowledge, \emph{abductive conditionals} have not yet been investigated in the probabilistic truth table task paradigm. Abductive conditionals can be conceived as reversed causal conditionals, characterized as follows: the effect is located in the conditional's antecedent and the cause is located in the conditional's consequent. For example, \emph{If your headache disappeared, then you took aspirin}. Abductive inferences are also known as inferences to the best explanation \cite<for philosophical and psychological overviews on abduction see, e.g.,>[respectively]{sep-abduction, lombrozo12}. Like indicative and causal conditionals, abductive conditionals can be formulated in indicative and in subjunctive mood.
The aim of the present study is to help to fill the above mentioned research gaps. Specifically, we aim to shed light on the following questions using probabilistic truth table tasks under incomplete probabilistic knowledge:
\begin{itemize}
\item Are there reasoning strategies for inferring lower and upper bounds in the context of incomplete probabilistic knowledge?
\item Is the conditional event interpretation dominant for abductive, causal, and non-causal counterfactuals?
\end{itemize}
\section{Method}
\paragraph{Materials and Design}
The task material consisted of 18 pen and paper tasks, preceded by four examples explaining the answer format. The task sequence consisted of nine different tasks that were presented twice in the same random order (i.e., task T10 is a repetition of task T1). The tasks were designed to test how participants infer about uncertain conditional sentences in four different situations (see Table~\ref{Design}). For the first two conditions we used a non-causal scenario in both indicative and counterfactual moods. For the other two conditions we used two variations of a causal scenario in counterfactual mood; inference from causes to effects (causal) and inference from effects to causes (abductive). The material was adapted from probabilistic truth table tasks, which involved precise premises \cite{fugard11a,pfeifer15a}.
\begin{table}[!ht]
\begin{center}
\caption{Experimental conditions C1--C4 defined by the types and formulations of conditionals, and sample sizes.}
\label{Design}
\vskip 0.12in
\begin{tabular}{cllc}
\hline
& Type & Formulation & Sample \\
\hline
C1 & non-causal & indicative & ($n_1=20$) \\
C2 & non-causal & counterfactual & ($n_2=20$) \\
C3 & causal & counterfactual & ($n_3=20$) \\
C4 & abductive & counterfactual & ($n_4=20$) \\
\hline
\end{tabular}
\end{center}
\end{table}
For the \emph{non-causal scenario} we used a vignette story about a six-sided die. The story describes that the die was randomly thrown so that the participants did not know which of the sides ended up facing upwards. The sides of the die were illustrated as six squares. Each side had an image of a black or white geometric figure. In tasks T1, T2, T10, and T11 all sides of the die were shown. To introduce incomplete probabilistic knowledge we presented ``covered'' sides in the rest of the tasks. Covered sides were indicated by a question mark. Figure~\ref{FIG:SampleDice} shows an example of how we presented the six sides of a die.
\begin{figure}[ht]
\begin{center}$
\ifx\JPicScale\undefined\def\JPicScale{0.2}\fi
\psset{unit=\JPicScale mm}
\psset{linewidth=0.3,dotsep=1,hatchwidth=0.3,hatchsep=1.5,shadowsize=1,dimen=middle}
\psset{dotsize=0.7 2.5,dotscale=1 1,fillcolor=black}
\psset{arrowsize=1 2,arrowlength=1,arrowinset=0.25,tbarsize=0.7 5,bracketlength=0.15,rbracketlength=0.15}
\begin{array}{cccccc}
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\rput{0}(20,20){\psellipse[fillstyle=solid](0,0)(10,-10)}
\end{pspicture}
&
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\pspolygon[fillstyle=solid](10,30)(30,30)(30,10)(10,10)
\end{pspicture}
&
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\pspolygon[fillstyle=solid](10,30)(30,30)(30,10)(10,10)
\end{pspicture}
&
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\pspolygon[](10,10)(30,10)(30,30)(10,30)
\end{pspicture}
&
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\pspolygon[](10,10)(30,10)(30,30)(10,30)
\end{pspicture}
&
\begin{pspicture}(0,0)(40,40)
\pspolygon[](0,0)(40,0)(40,40)(0,40)
\put(10,10){\makebox(20,20)[cc]{?}}
\end{pspicture}
\end{array}$
\end{center}\vspace{-.4cm}
\caption{Example of presented sides of a die (non-causal task). Covered sides are indicated by the question mark.}
\label{FIG:SampleDice}
\end{figure}
Next, the participants were presented with the question ``How sure can you be that the following sentence holds?'' (\emph{Kuinka varma voit olla siit\"a, ett\"a seuraava lause pit\"a\"a paikkaansa?}). The target sentences were highlighted with a frame to make the scope of the question clear, for example:
\begin{center}
\noindent
\fbox{
\parbox{.43\textwidth}{\footnotesize {\bf If} the figure on the
upward facing side of the die is a \emph{circle},\newline {\bf then} the figure is \emph{black}.
}
}
\end{center}
The answer format had two sets of tick boxes in a ``x out of y'' format for responding interval-valued degrees of belief. The two response boxes were labeled accordingly (``at least'' and ``at most''). It was explained in the introduction to give point valued responses by marking the same numbers in both response boxes (i.e., lower and upper bounds coincide).
See Figure~\ref{FIG:Tickbox} for an example.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=.3\textwidth]{Example_answerformat.eps}
\end{center}\vspace{-.4cm}
\caption{Upper bound response (``at most 4 out of 6'').}
\label{FIG:Tickbox}
\end{figure}
The target sentence in the non-causal tasks was formulated in terms of ``If $A$, then $C$''. In all non-causal tasks the antecedent mentioned a form and the consequent mentioned a color. After completing each task, the participants were asked to rate their confidence in the correctness of their response on a 10-step rating scale from ``fully confident that your answer is incorrect" to ``fully confident that your answer is correct".
Apart from the following two differences, the counterfactual task version was identical to the indicative version of the task: (i) we added a factual statement which contradicted the antecedent of the target sentence and (ii) the target sentence was formulated in subjunctive mood. ``The form of an upward-facing side of the die is a cube'' is an example of a factual statement and the corresponding target sentence is: ``if the figure on the upward facing side of the die were a circle, then the figure would be black'' (\emph{Jos yl\"osp\"ain osoittavan kyljen kuvio ol\underline{isi} ympyr\"a, niin se kuvio olisi musta}). The suffix \emph{-isi} in the Finnish original indicates the counterfactual mood.
For the \emph{causal} and \emph{abductive} conditions, the tasks were structurally identical to the (non-causal) dice-scenario. However, instead of dice, a vignette story about drugs and their effects created a causal scenario. In the vignette
story, six patient reports were shown to the participants. The patient reports were illustrated as six rectangles having a name of a fictional drug and a result of the medication (either ``diminishes symptoms'' or ``no impact on the symptoms'').
We used question marks on some patient reports (like in the dice scenario) to introduce incomplete probabilistic knowledge. Figure~\ref{FIG:ExampleReports} shows an example of the patient reports. This example contains the same probabilistic information as the die-example in Figure~\ref{FIG:SampleDice}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=.45\textwidth] {ReportsTask3.eps}
\end{center}\vspace{-.5cm}
\caption{Example of presented patient reports (causal / abductive tasks; see also Figure~\ref{FIG:SampleDice}).}
\label{FIG:ExampleReports}
\end{figure}
In the causal version of the task material the antecedent denotes the name of a drug and the consequent denotes an effect. In the abductive version the order was reversed: first an effect was presented and then a drug was named. In this way the tasks called for either \emph{causal inferences} from causes to effects, or \emph{abductive inferences} from effects to causes. As the material was formulated in counterfactual mood, we added a factual statement to each task, which contradicted the antecedent of the target sentence.
\paragraph{Participants and Procedure}
Eighty students from the University of Helsinki (Finland) participated in the experiment. The students were native Finnish speakers with no previous academic training in logic or probability. Each participant was tested individually. The paper and pencil tasks were followed by a short structured interview about how the participants had interpreted the target tasks. Participants were paid 15{\EUR} for their participation.
\section{Results and Discussion}
We performed Fisher's exact tests to investigate whether the four different versions of the task booklets had an impact on the participants' degrees of belief in the respective target sentences. After p-value corrections for multiple significance tests, we did not observe significant differences between the four conditions and we therefore pooled the data.
Table~\ref{Results1} shows the percentages of responses according to the different interpretations of conditionals. All tasks differentiate between the conditional probability, the conjunction, and the material conditional interpretation. A subset of the tasks differentiates between biconditional and biconditional event interpretations as well.
Conditional probability interpretations marked with indices, however, were patterns identified from the data and were not anticipated during the construction of the task material. Therefore, not all tasks differentiate among all interpretations. Table~\ref{TAB:interpretations} shows the normative answers for each interpretation for the example task discussed in the previous section (see Figure~\ref{FIG:SampleDice}). Both, lower and upper bound responses, had to match the normative lower and upper bounds for the categorization of the responses in Table~\ref{Results1}. Since each response box enables 42 different ``$X$ out of $Y$'' responses, and since both, lower and upper bound responses needed to match for the classification, the \emph{a priori} chance for guessing an interpretation was very low (i.e., $1/(42^2) = 0.0006$).
\begin{table}[!ht]
\begin{center}
\caption{Example of predicted responses where the task consists in inferring the degree of belief in the conditional ``{\bf If} the figure on the
upward facing side of the die is a \emph{circle}, {\bf then} the figure is \emph{black}'' (i.e., the conclusion), based on the die presented in Figure~\ref{FIG:SampleDice} (i.e., Figure~\ref{FIG:SampleDice} contains the premises). The index $\overline{l}$ (resp., $\overline{u}$) denotes conditional probability responses where the covered sides are ignored for inferring the lower (resp., upper) bound response. These response types are the ``halfway lower'' and ``halfway upper'' interpretations, respectively. $\overline{lu}$ denotes conditional probability responses where covered sides are ignored for inferring both bound responses, i.e., the ``halfway both interpretation''. See also Table~\ref{TAB:ProbProp}.}
\label{TAB:interpretations}
\vskip 0.12in
\begin{tabular}{lcc}
\hline
Interpretation & \multicolumn{2}{c}{Predictions}\\
& at least & at most \\
\hline
$p(\text{black }|\text{ circle})$ &1 out of 2 & 2 out of 2 \\
$p(\text{black }|\text{ circle})_{\overline{l}}$ & 1 out of 1 & 2 out of 2\\
$p(\text{black }|\text{ circle})_{\overline{u}}$ & 1 out of 2 & 1 out of 1 \\
$p(\text{black }|\text{ circle})_{\overline{lu}}$ & 1 out of 1 & 1 out of 1\\
$p(\text{circle }\wedge\text{ black})$ & 1 out of 6 & 2 out of 6 \\
$p(\text{circle }\supset\text{ black})$ & 5 out of 6 & 6 out of 6 \\
$p(\text{circle }\equiv\text{ black})$ & 3 out of 6 & 4 out of 6 \\
$p(\text{circle }||\text{ black})$ & 1 out of 4 & 2 out of 4 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[!ht]
\begin{center}
\caption{Percentages of responses from all four conditions and all 18 tasks (T1--T18; $N=80$). ``Grouped $p(\cdot|\cdot)$'' denotes the sum of all conditional probability responses, including those marked with the indices. The halfway interpretations (indices $\overline{l}$ and $\overline{u}$) and the numerical predictions are explained in Table~\ref{TAB:interpretations}. ``- -'' denotes cases where different conditional probability interpretations cannot be individuated (i.e., in the point value tasks). Similarly, {``[-~-]''} denotes cases where biconditional and biconditional event interpretations cannot be distinguished from the other interpretations. The interpretations are explained in Table~\ref{TAB:ProbProp}. ``$^{\star}$'' denotes psychological main interpretations.}
\label{Results1}
\vskip 0.12in
\begin{tabular}{lcccccc}
\hline
Interpretation & T1 & T2 & T3 & T4 & T5 & T6 \\
\hline
$[p(\cdot|\cdot)]^{\star}$ & [46] & [55] & [15] & [19] & [24] & [24] \\
$[p(\cdot|\cdot)_{\overline{l}}]$ & [- -] & [- -] & [5] & [13] &[18] & [11] \\
$[p(\cdot|\cdot)_{\overline{u}}]$ & [- -] & [- -] & [23] & [10] & [13] & [11] \\
$[p(\cdot|\cdot)_{\overline{lu}}]$ & [- -] & [- -] & [0] & [3] & [1] &[0] \\
Grouped $p(\cdot|\cdot)$& 46 & 55 & 43 & 44 & 55 & 46\\[.2em
$p(\cdot\wedge\cdot)^{\star}$ & 29 & 28 & 34 & 39 & 34 & 31 \\
$p(\cdot\supset\cdot)^{\star}$ & 1 & 0 & 0 & 0 & 0 & 1 \\
$p(\cdot \equiv \cdot)$ & [- -] & [- -] & 1 & [- -] & [- -] & 0 \\
$p(\cdot ||\cdot)$ & [- -] & [- -] & 3 & [- -] & [- -] & 0 \\
Other & 24 & 18 & 20 & 18 & 11 & 21 \\
\hline
& T7 & T8 & T9 & T10 & T11 & T12 \\
\hline
$[p(\cdot|\cdot)]^{\star}$ & [24] & [28] & [26] & [44] & [55] & [25] \\
$[p(\cdot|\cdot)_{\overline{l}}]$ & [10] & [15] & [10] & [- -] & [- -] & [9] \\
$[p(\cdot|\cdot)_{\overline{u}}]$ & [16] & [8] & [10] & [- -] & [- -] & [23] \\
$[p(\cdot|\cdot)_{\overline{lu}}]$ & [0] & [0] & [0] & [- -] & [- -] &[0] \\
Grouped $p(\cdot|\cdot)$& 46 & 55 & 43 & 44 & 55 & 46\\[.2em
$p(\cdot\wedge\cdot)^{\star}$ & 34 & 29 & 33 & 26 & 29 & 30 \\
$p(\cdot\supset\cdot)^{\star}$ & 0 & 0 & 0 & 1 & 0 & 0 \\
$p(\cdot \equiv \cdot)$ & [- -] & [- -] & [- -] & [- -] & [- -] & 0 \\
$p(\cdot ||\cdot)$ & [- -] & [- -] & [- -] & [- -] & [- -] & 0 \\
Other & 16 & 21 & 21 & 18 & 18 & 14 \\
\hline
& T13 & T14 & T15 & T16 & T17 & T18 \\
\hline
$[p(\cdot|\cdot)]^{\star}$ & [34] & [33] & [29] & [26] & [31] & [31] \\
$[p(\cdot|\cdot)_{\overline{l}}]$ & [9] & [13] & [11] & [10] & [18] & [13] \\
$[p(\cdot|\cdot)_{\overline{u}}]$ & [11] & [10] &[11] & [15] & [8] & [11] \\
$[p(\cdot|\cdot)_{\overline{lu}}]$ & [0] & [0] & [1] & [3] & [0] &[0] \\
Grouped $p(\cdot|\cdot)$& 46 & 55 & 43 & 44 & 55 & 46\\[.2em
$p(\cdot\wedge\cdot)^{\star}$ & 28 & 30 & 26 & 29 & 25 & 28 \\
$p(\cdot\supset\cdot)^{\star}$ & 0 & 0 & 0 & 0 & 0 & 0 \\
$p(\cdot \equiv \cdot)$ & [- -] & [- -] & 0 & [- -] & [- -] & [- -] \\
$p(\cdot ||\cdot)$ & [- -] & [- -] & 3 & [- -] & [- -] & [- -] \\
Other & 19 & 15 & 19 & 18 & 19 & 18 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[!ht]
\begin{center}
\includegraphics[width=.4\textwidth] {MeanConfidence.eps}
\end{center}\vspace{-.45cm}
\caption{Mean confidence values for tasks T1--T18 by condition. C1--C4 denote the four condition as defined in Table~\ref{Design}.}
\label{FIG:Confidence}
\end{figure}
The task material was designed such that the normative predictions of the three main psychological interpretations of conditionals (i.e., conditional probability, conjunction, and material conditional) were unique for each task. During the analysis we identified three further response strategies related to the conditional probability interpretation. In what we call \emph{halfway lower} interpretation (denoted by $p(\cdot |\cdot)_{\overline{l}}$) the upper bound is the same as in conditional probability, but the lower bound response differs in that the covered sides (i.e., sides marked with question mark) are ignored. \emph{Halfway upper} interpretation (denoted by $p(\cdot |\cdot)_{\overline{u}}$) is the same, but in reverse order. In a \emph{halfway both} interpretation the covered sides are ignored for both bound responses. As these answer strategies are in a sense partial versions of the conditional probability interpretation, we combined their results with conditional probability answers into \emph{grouped conditional probability}. Notice that the tasks T1,T2, T10 and T11 with full information (i.e., no question marks) have the same value for lower and upper bound answers. Therefore the halfway responses could not be distinguished from the conditional probability answers in these tasks.
Of all 1440 responses, 32.1\% were consistent with conditional event responses, 29.9\% were consistent with conjunction responses, and 0.2\% were consistent with material conditional responses. Like the material conditional, also the biconditional and the biconditional event response frequencies play a neglectable r\^ole in the data (0\%--3\% in the four tasks T3, T6, T12, T15 where these interpretations were differentiated). The predominant response strategy in point-valued tasks (T1, T2, T10 and T11) was consistent with the conditional probability interpretation. In nine out of 14 tasks with incomplete probabilistic information (i.e., tasks involving ``covered'' sides or patient reports) the predominant answer strategy was conjunction. We also observed shifts of interpretation towards conditional probability: comparing the first three tasks with incomplete information (i.e., T3--T5) to the last three (i.e., T16--T18), the number of conditional probability answers increased from 19\% to 30\%, and conjunction answers decreased from 35\% to 27\%. This replicates shifts of interpretations reported in the literature \cite{fugard11a,pfeifer13b}.
However, when all the conditional probability response types are grouped together, the resulting set of response strategies is clearly the predominant one in all tasks. 51.5\% of all answers are consistent with the grouped conditional probability responses. 18.1\% were ``other'' responses, that is, responses that did not fit the grouped conditional probability, conjunction, biconditional, biconditional event, or material conditional. Thus, in total 81.9\% of the data can be modeled by the investigated hypotheses concerning the interpretation of conditionals.
Compared to a previous study that investigated non-causal indicative conditionals under incomplete probabilistic information \cite[i.e., similar tasks as in condition C1]{pfeifer13b}, our results show lower level of conditional event responses (compared to the previous 65.6\%), and higher levels of conjunction responses (compared to the previous 5.6\%). The material conditional responses were similar (previously 0.3\%). \citeA{pfeifer15a} investigated conditionals under complete probabilistic knowledge and used similar tasks as in our conditions C1, C2 and C3. These authors also reported higher levels of conditional probability answers and lower levels of conjunction responses, while material conditional responses were similarly low. The lower levels of conditional probability responses may be explained by the apparent higher complexity of the tasks used in the present experiment. The tasks are more complex (i) because of the combination of using counterfactuals as target sentences in three of four conditions and (ii) because of imprecise probabilities in the premises (i.e., incomplete probabilistic information).
The tendency to give answers that partially coincide with conditional probability has also been found in a previous study which tested non-causal cases in indicative mood with similar task material as we used for condition C1 \cite{pfeifer13b}. In that study our halfway lower-interpretation is referred to as ``halfway conditional event strategy''. However, in the present study we found two completely new strategies: the halfway upper- and the halfway both-interpretation. The halfway upper-interpretation is particularly interesting, as it explains 12.8\% of the total 1120 responses, slightly more than the halfway lower response strategy (i.e., 11.6\%). Halfway conditional probability responses might unburden the working memory load by ignoring the covered sides \cite<see also>{pfeifer13b}.
Figure~\ref{FIG:Confidence} shows the results of the confidence ratings. We performed analyses of variance to investigate impacts of the different conditions. After Holm-Bonferroni corrections we observed significant differences within the three tasks T1, T2 and T10. The corresponding p-values were 0.006, 0.01, and 0.008. In each of these tasks---as well as in all other tasks---the condition C4 had lower mean confidence values compared to the other conditions. The lower confidence may be because of the apparent higher difficulty of the task material in condition C4 for two reasons: first, the target sentence was a counterfactual. Because of the inconsistency between the factual statement and the antecedent, many participants reported counterfactuals as puzzling in the post-test interview. Second, abductive tasks required ``backward'' inference from effects to causes and are incongruent with the more natural \emph{if cause, then effect}-direction. In general, backward inferences are known to be harder to draw compared to forward inferences \cite{evans81}.
\section{Concluding Remarks}
We investigated how people reason about conditionals under incomplete probabilistic knowledge. The novel features in our test design were comparisons of \emph{causal} and \emph{abductive} scenarios, as well as \emph{counterfactuals} under \emph{incomplete} probabilistic knowledge.
One of our main findings is that the dominant response is consistent
with the conditional event interpretation of conditionals
among all four groups. Moreover, we discovered two major answer strategies, \emph{halfway upper} and \emph{halfway lower conditional event responses}, which can be understood as strategies to unburden the working memory load
Inferentialist accounts of conditionals propose that there should be an inferential relation between the antecedent and the consequent \cite<see, e.g.,>{douven2016}. Thus, when conditionals with inferential relations (e.g., causal or abductive ones) are compared with conditionals where no apparent inferential relation exists (like in our conditions C1 and C2), one would expect significant differences. Our data, however, do not support this inferentialist hypothesis.
The results of our paper broaden the area of inferences where conditional probability seems to be the best predictor for how people reason.
We have shown that coherence-based probability logic provides a formalization of the meaning of counterfactuals and provides a rationality framework for reasoning under complete and incomplete probabilistic knowledge. This suggests that it may also be suitable for subclasses of causal reasoning like abductive reasoning, which is important in the studies on (scientific) explanation and learning.
{\small\paragraph{Acknowledgments}
This research was supported by the DFG project
PF~740/2-2 (awarded to Niki Pfeifer) as part of the Priority Program ``New Frameworks of Rationality'' (SPP1516).}
|
\section{Introduction and main results}
In his famous book {\em New thoughts on Besov spaces}, page 151, Jaak Peetre posed the problem
to determine the set of all pointwise multipliers $M (B^s_{p,q}(\R))$ of the Besov space
$B^s_{p,q}(\R)$ in case $s>d/p$. Now, more than 40 years later, we are able to present the
complete solution to this problem.
To describe this we need to introduce a few related classes of functions.
Here we are forced to distinguish between the cases $p \le q$, $q<p< \infty$ and $p=\infty$.
First we deal with $p \le q$.
\begin{definition} \label{unif}
Let $\psi \in C_0^\infty (\R)$ be a nonnegative nontrivial function.
Let $0< p,q \le \infty$ and $s\in \re$. Then
$B^s_{p,q}(\R)_{\unif}$ denotes the collection of all tempered distributions
$f \in \cs' (\R)$ such that
\beqq
\| \, f\, | B^s_{p,q}(\R)_{\unif}\|:= \sup_{\lambda \in \R} \| \, f(\, \cdot \, )\, \psi (\, \cdot\, -\lambda) \, | B^s_{p,q}(\R)\|
<\infty \, .
\eeqq
\end{definition}
The spaces $ B^s_{p,q}(\R)_{\unif}$ are quasi-Banach spaces independent of the choice of $\psi$ (in the sense of equivalent quasi-norms).
\begin{theorem}\label{b-main}
Let $0<p\leq q\leq \infty$, and $s>d/p$. Then
\be\label{ws-02}
M(B^s_{p,q}(\R))=B^s_{p,q}(\R)_{\unif}
\ee
in the sense of equivalent quasi-norms.
\end{theorem}
In proving this theorem we will make use of the characterization of Besov spaces by differences in a way
similar to Strichartz in his paper \cite{Str-67}, see also the monographs of Maz'ya and Shaposnikova \cite{MS1}, \cite{MS2}.
\begin{remark}
\rm
There is a large number of references dealing with pointwise multipliers for Besov or Lizorkin-Triebel spaces.
Here we selected only those which include characterizations of $M (B^s_{p,q}(\R))$.
\begin{itemize}
\item Strichartz \cite{Str-67} proved \eqref{ws-02} for $p=q=2$. In fact, he was dealing with the more
general case of Bessel potential spaces
$H^s_p (\R)$, $s>d/p$, but $B^s_{2,2} (\R) = H^s_2 (\R)$ (in the sense of equivalent norms).
His main tool were consisting in characterizations of $H^s_p (\R)$ by differences.
\item
Peetre \cite{Pe}, page 151, proved \eqref{ws-02} for $1 \le p=q \le \infty $. He used a method nowadays called paramultiplication, which consists
in a clever decomposition of the product in the Fourier image.
\item
Maz'ya and Shaposnikova, see \cite[Theorems 4.1.1, 5.3.1, 5.3.2, 5.4.1]{MS2}, proved \eqref{ws-02} for $1 \le p=q < \infty $.
Also these authors worked with characterizations by differences.
\item
Netrusov \cite{Net} proved characterizations of $M(B^s_{p,q}(\R))$ in cases $0<p=q\le 1$ and\\
$0 <p\le 1$, $q=\infty$ in Fourier analytic terms. This has been the first contribution to the case $p \neq q$.
\item Sickel \cite{Si} proved characterizations of $M(B^s_{p,p}(\R))$ in terms of capacities for all $p$, $0 <p< \infty$, and all
$s> d/p$.
The used method here is again paramultiplication in connection with the Fourier analytic description of the spaces.
\item
Smirnov and S. \cite{SS} have shown the identity \eqref{ws-02} in case $1\le p,q\le \infty$
by using atomic characterizations of Besov spaces.
\item
Triebel \cite[Proposition 2.22]{Tr06} proved a new characterization of $M(B^s_{p,p}(\R))$, where either $0 <p \le \infty$ and $s> d/p$
or $0 < p \le 1$ and $s=d/p$.
\end{itemize}
\end{remark}
Now we turn to the slightly more complicated case $q <p$.
Here we have to introduce spaces with a different type of localization.
\begin{definition}\label{multdef}
Let $\psi \in C_0^\infty (\R)$ be a nonnegative function satisfying
\be\label{ws-03}
\sum_{\mu \in \Z} \psi (x-\mu) = 1 \qquad \mbox{for all}\quad x \in \R\, .
\ee
Let $s > 0$, $0 < p,q\le \infty$, $m\in \N$ and $s < m \leq s + 1$.
By using $\psi_\mu (\, \cdot \, ):= \psi (\, \cdot\, -\mu)$, $\mu \in \Z$,
the space $M^s_{p,q}(\R)$ is the collection of all $f \in L_1^{\ell oc}(\R)$ such that
\beq\label{ws-04}
\| \, f\, |M^s_{p,q}(\R)\|& := &
\sup_{\|\, \{C_{\mu}\}_\mu\, |\ell_p (\Z)\|\le 1}\Bigg\{\Big\|\, f (\, \cdot\, )\, \Big(\sum_{\mu\in \Z} C_{\mu} \psi_{\mu}(\, \cdot \, )\Big)
\, \Big|L_p(\R)\Big\|^q
\\
&& + \quad
\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h| < 2^{-k}} \sum_{\mu\in \Z} |C_{\mu}|^p \big\|\Delta_h^m(\psi_{\mu}\, f )(\cdot)
\big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q} <\infty \, .
\nonumber
\eeq
\end{definition}
We fix $\nu$.
By choosing $C_\mu := \delta_{\mu,\nu}$, $\mu \in \Z$, it is easily seen that
the right-hand side in \eqref{ws-04} reduces to the quasi-norm of $\psi_{\nu}\, f $ in $B^s_{p,q}(\R)$, see Proposition \ref{diff}.
This observation yields
\beqq
M^s_{p,q}(\R)\hookrightarrow B^s_{p,q}(\R)_{\unif}\, .
\eeqq
It is not difficult to see that this embedding is proper in case $q <p$ (all details will be given below).
We have the following nice characterization of $M(B^s_{p,q} (\R))$.
\begin{theorem}\label{nec1}
Let $0< q < p < \infty$ and $s>d/p$. Let $\psi_\mu$, $\mu \in \Z$, be as in Definition \ref{multdef}.
Then $f \in M(B^s_{p,q}(\R))$ if and only if $\sum_{\mu \in \Z} C_\mu \psi_\mu f$ belongs to $B^s_{p,q}(\R)$
for all $\{C_\mu\}_\mu \in \ell_p (\Z)$ and
\beqq
\sup_{\| \{C_\mu\}_\mu \, |\ell_p(\Z) \|\le 1} \, \Big\|\sum_{\mu \in \Z} C_\mu \psi_\mu f\Big|B^s_{p,q}(\R)\Big\| < \infty \, .
\eeqq
\end{theorem}
The second main result of our paper can now formulated as follows.
\begin{theorem}\label{final}
Let $ 0< q <p < \infty$ and $s>d/p$.
Then we have
\beqq
M(B^s_{p,q}(\R))=M^s_{p,q}(\R)
\eeqq
in the sense of equivalent quasi-norms.
\end{theorem}
We supplement our findings with the more easy case of $p=\infty$.
\begin{theorem}\label{final2}
Let $s>0$ and $0<q\leq \infty$. Then it holds $M(B^s_{\infty,q}(\R))=B^s_{\infty,q}(\R)$ in the sense of equivalent quasi-norms.
\end{theorem}
\begin{remark}
\rm
(i) Theorem \ref{final} seems to be a novelty. We are not aware of any additional reference in this direction.
\\
(ii) The simple characterization of $M(B^s_{\infty,q}B(\R))$ has been known, see \cite{SS}, but \cite{SS} was never published.
\\
(iii) Of course, Theorem \ref{final2} and Theorem \ref{b-main} overlap in case $p=q=\infty$.
In this context it is of certain interest to notice that
$M(B^s_{\infty,q}(\R)) = B^s_{\infty,q}(\R)_\unif$ if and only if $q=\infty$.
\end{remark}
Finally we collect results on the limiting situation $s=d/p>0$.
\begin{theorem}\label{final3}
(i) Let $0<p = q\leq 1$ and $s=d/p$. Then
\beqq
M(B^s_{p,p}(\R))=B^s_{p,p}(\R)_{\unif}
\eeqq
holds in the sense of equivalent quasi-norms.
\\
(ii) Let $ 0< q <p \le 1$ and $s=d/p$.
Then we have
\beqq
M(B^s_{p,q}(\R))=M^s_{p,q}(\R)
\eeqq
in the sense of equivalent quasi-norms.
\\
(iii) Let $\psi_\mu$, $\mu \in \Z$, be as in Definition \ref{multdef}.
Under the same restrictions as in (ii) \\
$f \in M(B^s_{p,q}(\R))$ if and only if $\sum_{\mu \in \Z} C_\mu \psi_\mu f$ belongs to $B^s_{p,q}(\R)$
for all $\{C_\mu\}_\mu \in \ell_p (\Z)$ and
\beqq
\sup_{\| \{C_\mu\}_\mu \, |\ell_p(\Z) \|\le 1} \, \Big\|\sum_{\mu \in \Z} C_\mu \psi_\mu f\Big|B^s_{p,q}(\R)\Big\| < \infty \, .
\eeqq
\end{theorem}
\begin{remark}\label{abcd}
\rm
(i) In Corollary \ref{abc} below we shall prove that in the Banach space case Theorem \ref{b-main} and
Theorem \ref{final3} cover all cases where we have the coincidence $M(B^s_{p,q}(\R))=B^s_{p,q}(\R)_{\unif}$.
\\
(ii) Let $\Omega $ be an open, nontrivial and bounded subset of $\R$.
Define $B^s_{p,q} (\Omega)$ as the collection of all distributions $g \in \mathcal{D}'(\Omega)$ such that there exists
some $f \in B^s_{p,q} (\R)$ satisfying $f(\varphi) = g (\varphi)$ for all $\varphi \in \mathcal{D}(\Omega)$.
Equipped with the quotient norm
\[
\|\, g \, | B^s_{p,q} (\Omega)\|:= \inf \Big\{ \|\, f \, | B^s_{p,q} (\R)\| ~: \quad f_{|_\Omega} =g \Big\}
\]
$B^s_{p,q} (\Omega)$ becomes a quasi-Banach space.
Intrinsic characterizations, e.g., in the spirit of Proposition \ref{diff}, are known in case of a Lipschitz boundary, we refer to Dispa \cite{Di} and Triebel \cite[4.1.4]{Tr06}.
Essentially as a consequence of Theorem \ref{algebra} below it follows
\[
M(B^s_{p,q}(\Omega)) = B^s_{p,q} (\Omega)
\]
if $0 < p,q\le \infty$ and either $s>d/p$ or $s=d/p>0$ and $0 < q \le 1$.
Hence, the difficulties in determining $M(B^s_{p,q}(\R))$ are connected with the unboundedness
of $\R$ and the difficult localization properties of Besov spaces, see Proposition \ref{s<p,q<u}.
\end{remark}
\subsection*{A short overview on further results in case $s\le d/p$}
As a service for the reader we finish this section with an overview about the knowledge on $M(B^s_{p,q}(\R))$ in case $s\le d/p$.
\begin{itemize}
\item Let $1\le p < \infty$ and $0< s \le d/p$. Then $M(B^s_{p,p}(\R))$ has been characterized by
Maz'ya and Shaposnikova, see \cite[Theorems 4.1.1, 5.3.1, 5.3.2, 5.4.1]{MS2}.
\item
Let $0 <p =q \le 1$ and $d \, ( \frac 1p -1)< s \le d/p $. Then $M(B^s_{p,p}(\R))$ has been characterized by
Netrusov \cite{Net}. Here we wish to mention that the description of $M(B^s_{p,p}(\R))$ given by Netrusov looks different compared to
Theorem \ref{final3}.
\item
Let $0 <p \le 1$ and $d\, (\frac 1p -1)< s \le d/p$. Then $M(B^s_{p,\infty}(\R))$ has been characterized by
Netrusov \cite{Net}.
\item
Let $0 <p=q < \infty$ and $d\, \max (0, \frac 1p -1)< s \le d/p$. Then $M(B^s_{p,p}(\R))$ has been characterized by
Sickel \cite{Si}.
\item
Let $p=\infty$ and $s=0$. In \cite{KS} the spaces $M(B^s_{\infty,1}(\R))$ and $M(B^s_{\infty,\infty}(\R))$
have been characterized.
\item
Triebel \cite[Theorem 2.25]{Tr06} has found a new characterization of $M(B^s_{p,p}(\R))$, $0< p \le 1$ and $s> d(\frac 1p - 1)$,
in terms of the quantity
\[
\sup_{\mu \in \Z} \sup_{j \in \N_0} \, \| \psi_\mu (\, \cdot\, ) \, f(2^{-j}\cdot\,)\, |B^s_{p,p} (\R) \|
\]
where $\psi_\mu$ is defined as in Definition \ref{multdef}, see also Schneider and Vybiral \cite{SV}.
\end{itemize}
The paper will be organized as follows.
In Section \ref{def} we collect all what we need about the function spaces under consideration.
This will be followed by a short section including basic properties of pointwise multipliers.
The next Section is devoted to the proof of Theorem \ref{b-main} including some limiting cases
with $s=d/p$.
In Section \ref{main4} we deal with the proof of Theorem \ref{final}.
\subsection*{Notation}
As usual, $\N$ denotes the natural numbers, $\N_0=\N\cup\{0\}$,
$\zz$ denotes the integers,
$\re$ the real numbers,
and $\C$ the complex numbers. For a real number $a$ we put $a_+ := \max(a,0)$. The letter $d\in \N, \ d>1,$ is always reserved for the underlying dimension in $\R$ and $\Z$.
If $X$ and $Y$ are two (quasi-)normed spaces, the (quasi-)norm
of an element $x$ in $X$ will be denoted by $\|x\,|\,X\|$.
The symbol $X \hookrightarrow Y$ indicates that the
identity operator is continuous. For two sequences $a_n$ and $b_n$ we will write $a_n \lesssim b_n$ if there exists a
constant $c>0$ such that $a_n \leq c\,b_n$ for all $n$. We use $a_n \asymp b_n$ if $a_n \lesssim b_n$ and $b_n
\lesssim a_n$.
Let $\cs(\R)$ be the Schwartz space of all complex-valued rapidly decreasing infinitely differentiable functions on $\R$.
The topological dual, the class of tempered distributions, is denoted by $\cs'(\R)$ (equipped with the weak topology).
The Fourier transform on $\cs(\R)$ is given by
\[
\cf \varphi (\xi) = (2\pi)^{-d/2} \int_{\R} \, e^{-ix \xi}\, \varphi (x)\, dx \, , \qquad \xi \in \R\, .
\]
The inverse transformation is denoted by $\cfi $.
We use both notations also for the transformations defined on $\cs'(\R)$\,.
\section{Besov spaces}\label{def}
General references for Besov spaces are, e.g., the monographs of Nikol'skij \cite{Ni}, Peetre \cite{Pe} and
Triebel \cite{Tr83}, \cite{Tr92}, \cite{Tr06}.
To introduce Besov spaces for the full range of parameters we make use of Fourier analysis, a way,
originally introduced by Peetre \cite{Pe} and later propagated also by Triebel \cite{Tr83}, \cite{Tr92}, \cite{Tr06}.
\subsection{Definition and basic properties}
Let $\varphi_0 \in C_0^{\infty}({\R})$ be a non-negative function such that
$\varphi_0(x) = 1$ if $|x|\leq 1$ and $ \varphi_0 (x) =0$ if $|x|\geq 3/2$.
For $k\in \N$ we define
\beqq
\varphi_k(x) = \varphi_0(2^{-k}x)-\varphi_0(2^{-k+1}x) ,\qquad\ x \in \R\, .
\eeqq
Because of
\[
\sum_{k=0}^\infty \varphi_k(x) = 1\, , \qquad x\in \R\, ,
\]
and
\[
\supp \varphi_k \subset \big\{x\in \R: \: 2^{k-1}\le |x|\le 3 \cdot 2^{k-1}\big\}\, , \qquad k \in \N\, ,
\]
we call the system $(\varphi_k)_{k\in \N_0 }$ a smooth dyadic decomposition of unity on $\R$.
Clearly,
by the Paley-Wiener-Schwarz theorem,
\be\label{ws-09}
f_k (x):= \cfi[\varphi_{k}\, \cf f](x)\, , \qquad x \in \R\, , \quad k \in \N_0\, .
\ee
is a smooth function for all $f\in \cs'(\R)$.
\begin{definition}\label{def-do}
Let $(\varphi_k)_{k\in \N_0 }$ be the above system. Let $0< p,q\leq \infty$ and $s \in \re$. Then $ B^{s}_{p,q}(\re^d)$ is the
collection of all tempered distributions $f \in \mathcal{S}'(\R)$
such that
\beqq
\|\, f \, |B^s_{p,q}(\R)\|_{\varphi_0} :=
\bigg(\sum\limits_{k=0}^{\infty} 2^{k s q}\, \big\|\, \cfi[\varphi_{k} \cf f](\, \cdot \, )\,
\big|L_p(\re^d)\big\|^q\bigg)^{1/q} <\infty
\eeqq
(with usual modification if $q= \infty$).
\end{definition}
Clearly, Besov spaces are quasi-Banach spaces independent of the generator $\varphi_0 $ (in the sense of equivalent quasi-norms).
Therefore we will write $\|\, f \, |B^s_{p,q}(\R)\|$ instead of $\|\, f \, |B^s_{p,q}(\R)\|_{\varphi_0}$.
\\
For us, embeddings into $L_1^{\ell oc} (\R)$ and $L_\infty (\R)$ will be of some interest.
\begin{lemma}\label{emb1}
Let $0 < p,q\le \infty$ and $s \in \re$.
\\
{\rm (i)} Let $s>d\, \max (0, \frac 1p -1)$. Then
$B^s_{p,q} (\R)$ is continuously embedded into $L_{\max(1,p)} (\R)$.
\\
{\rm (ii)} The Besov space $B^s_{p,q} (\R)$ is continuously embedded into $L_\infty (\R)$ if and only if
either $s> d/p$ or $ s=d/p$ and $0< q \le 1$.
\end{lemma}
\begin{remark}\label{emb2}
\rm
(i) As it is well-known, $B^s_{p,q} (\R) \hookrightarrow L_\infty (\R)$ if and only if
$B^s_{p,q} (\R) \hookrightarrow C (\R)$. Here $C(\R)$ denotes the Banach space of all uniformly continuous functions on $\R$ equipped with the supremum norm.
\\
(ii) The embeddings described in Lemma \ref{emb1} have a certain history.
We would like to mention at least Grisvard, Peetre, Golovkin, Stein, Zygmund, Besov, Il'yin and Brudnij.
For the Banach space case we refer to the supplement written by Lizorkin in the russian edition \cite[Suppl.~1.7]{TL}
of Triebel's monograph \cite{Tr83} for more details.
The general quasi-Banach case has been considered in \cite{sitr}.
\end{remark}
\subsection{Tools - the characterization by differences and some inequalities}
For us the characterization of Besov spaces by differences will be more important.
In case of a multivariate function $f:\R\to \C$, $m \in \N$, $x, h \in \R$, we put
\[
\Delta_{h}^{m} f(x):= \sum_{\ell =0}^{m} (-1)^{m -\ell} \, \binom{m}{\ell} \, f(x + \ell h )\, .
\]
The related modulus of smoothness is defined as
\[
\omega_m (f,t)_p := \sup_{|h|<t} \|\, \Delta_{h}^{m} f\, |L_p (\R)\|\, , \qquad t>0\, .
\]
\begin{proposition}\label{diff}
Let $0< p,q\leq \infty$, $s> d\, \max(0, \frac 1p-1)$ and $s <m$ for some natural number $m$. Then the Besov space
$B^s_{p,q}(\R)$ is a collection of all $f\in L_p(\R)$ such that
\beqq
\|f|B^s_{p,q}(\R)\|_m := \|\, f\, |L_p(\R)\| + \bigg( \sum_{k=0}^{\infty} \, \big(2^{ks}\, \omega_m (f, 2^{-k})\big)^q\bigg)^{1/q} < \infty.
\eeqq
Furthermore, $\|\, \cdot \, |B^s_{p,q}(\R)\|_m $ and $\|\, \cdot\, |B^s_{p,q}(\R)\|$ are equivalent on
$L_{\max(1,p)} (\R)$ for any admissible $m$.
\end{proposition}
\begin{remark}
\rm
The restriction $s> d\, \max(0, \frac 1p-1)$ is natural in such a context. Since
$B^s_{p,q}(\R)$ contains singular distributions if $s < d\, \max(0, \frac 1p-1)$
a characterization as in Proposition \ref{diff} becomes impossible.
The version stated in Proposition \ref{diff} is a direct consequence of Theorem 2.5.12 in \cite{Tr83}
using the monotonicity of $\omega_m (f,t)_p $ with respect to $t$.
\end{remark}
The following construction of a maximal function is essentially due to Peetre, but based on earlier work of Fefferman and Stein.
Let $a>0$ and $b >0$ be fixed. For $f \in L_1^{\ell oc} (\R)$ we define the Peetre maximal function $P_{b,a}f$ by
\beqq
P_{b,a}f(x) := \sup\limits_{z\in \R} \frac{|f(x-z)|}{ 1+|b z |^{a}}\, , \qquad x \in \R\, .
\eeqq
\begin{proposition}\label{peetremax}
Let $0< p \leq\infty$ and define $\Omega :=\{x: |x|\leq b \}$ for some $b>0$. Let further $a>d/p$.
Then there exists a positive constant $C$, independent of $b$, such that
\beqq
\big\| P_{b,a}f \big|L_p(\R)\big\|\leq C\, \|f |L_p(\R)\|
\eeqq
holds for all $f \in L_{\max(1,p)}(\R)$ with $\supp (\gf f)\subset \Omega$.
\end{proposition}
For a proof we refer to \cite[Thm.~1.4.1]{Tr83}.
A very useful relation between Peetre maximal function and differences is given by the following
lemma, see \cite{U1} and also \cite[page 102]{Tr83}.
\begin{lemma}\label{1dim-1}
Let $a>0$ and $m \in \N$.
Then there exists a constant $C$
such that
\beqq
|\Delta^m_hf(x)| \leq C\, \max (1,|bh|^a)\, \min(1,|bh|^m)\, P_{b,a}f(x)\,.
\eeqq
holds for all $b > 0$, all $h,x\in \R$ and all $f\in \cs'(\R)$ satisfying $\supp(\gf f) \subset \{x: |x|\leq b \}$.
\end{lemma}
Later we shall need also the following modification.
\begin{lemma}\label{help}
Let $a>0$ and
$\psi \in C_0^{k}(\R)$ for some $k \in \N$. Then, if $m\in \N$, $m \le k$, there exists a constant $C$ such that
\beqq
|\Delta^m_h (\psi \, \cdot \, f)(x)| \leq C \, \max\{1,|bh|^a\}\min\{1,|bh|^m\}P_{b,a}f(x)
\eeqq
holds for all $x,h\in \R$ and all $f\in \cs'(\R)$ such that $\supp (\gf f) \subset
\{x: |x|\leq b \}$, $b >0$. Here $C$ can be chosen independent of $b$.
\end{lemma}
\begin{proof}
The proof in one dimension can be found in \cite{NUU}. The general case follows by the same type of argument.
\end{proof}
\section{Pointwise multipliers}\label{main}
\subsection{Some generalities on pointwise multipliers}\label{main1}
For a quasi-Banach space $X$ of functions we shall call a function $f$ a pointwise multiplier
if $f \, \cdot \, g \in X$ holds for all $g \in X$
(this is includes, of course, that the operation $g \mapsto f \, \cdot \, g$ must be well defined for all $g\in X$).
If $X \hookrightarrow L_p (\Omega)$ for some $p$ (here $\Omega$ is a domain in $\R$),
as a consequence of the Closed Graph Theorem, we obtain that the liner operator
$T_f : ~ g \mapsto f \, \cdot \, g$, associated to such a pointwise multiplier, must be continuous in $X$,
see \cite[p.~33]{MS2}. As usual we put
\[
\|\, T_f\, |\cl (X) \|:= \sup_{\|g|X\|\le 1} \, \|f \, \cdot \, g\, |X \|\, .
\]
The collection of all pointwise multipliers for a given space $X$ will be denoted by $M(X)$ and equipped with the quasi-norm
\[
\| \, f \, |M(X)\|:= \|\, T_f\, |\cl (X) \|\, .
\]
Of course, beside the natural interpretation of $f\, \cdot \, g$ as a pointwise product of functions, one can define
$f\, \cdot \, g$ on certain subsets of $\cs'(\R) \times \cs'(\R)$, in particular on $\cs(\R) \times \cs'(\R)$.
Those extensions of the product will not play an important role within this paper. For that reason we will skip details and refer to
Johnsen \cite{Jo1} and the monograph \cite[4.2]{RS}.
Besov spaces are translation invariant, i.e., $\|\, g(\, \cdot \, - \mu)\, |B^s_{p,q}(\R)\|=\|\, g\, |B^s_{p,q}(\R)\|$
for all $\mu \in \R$. This implies that the associated multiplier space $M(B^s_{p,q}(\R))$ is translation invariant as well.
Let $f \in M(B^s_{p,q}(\R))$.
Let $\psi \in C_0^\infty (\R)$ be a nonnegative nontrivial function.
Since $C_0^\infty (\R) \subset B^s_{p,q}(\R)$ we conclude that $\psi (\, \cdot \, -\mu) \, \cdot \, f(\, \cdot \, ) \in B^s_{p,q}(\R)$
for all $\mu \in \R$.
By assumption and translation invariance of $B^s_{p,q}(\R)$ we know that
\[
\|\psi (\, \cdot \, -\mu) \, \cdot \, f(\, \cdot \, )\, |B^s_{p,q}(\R)\| \le
\|\, f\, |M(B^s_{p,q}(\R))\|
\, \|\psi \, |B^s_{p,q}(\R)\|\, ,
\]
which proves $M(B^s_{p,q}(\R)) \hookrightarrow B^s_{p,q}(\R)_\unif$.
\begin{lemma}\label{mult}
Let $0 <p,q\le \infty $ and $s\in \re$. Then it follows
$M(B^s_{p,q}(\R)) \hookrightarrow B^s_{p,q}(\R)_\unif$.
\end{lemma}
\subsection{On the algebra property of Besov spaces}\label{main2}
We shall call a quasi-Banach space of functions $X$ an algebra with respect to pointwise multiplication
(for short a multiplication algebra) if $f\, \cdot \, g \in X$ for all $f,g\in X$ and there exists a constant $c$ such that
\[
\|f \, \cdot \, g\, |X \| \le c \, \|\, f \, |X \|\, \| \, g\, |X \|
\]
holds for all $f,g \in X$.
\begin{theorem}\label{algebra}
Let $0<p,q\leq \infty$ and $s\in \re$. Then $B^s_{p,q}(\R)$ is a multiplication algebra if and only if one of the
following conditions is satisfied
\begin{itemize}
\item $s>d/p$;
\item $0<p<\infty$, $s=d/p$ and $q\leq 1$.
\end{itemize}
\end{theorem}
\begin{remark}\rm
The {\em if}-part with $p\ge 1$ has been proved for the first time 1970 in Peetre \cite{Pe1}.
But he had called this assertion well-known in \cite{Pe1}. The extension to the quasi-Banach case
has been obtained by Triebel \cite{Tr1}, \cite[2.6.2]{Tr78}. There also necessity of the above conditions
is shown.
However, Triebel had overlooked that $B^0_{\infty, q} (\R)$, $0< q \le 1$, is not an algebra.
For this correction we refer to \cite[4.6.4,~4.8.3]{RS}.
\end{remark}
\subsection{Localized Besov spaces}\label{lokal}
This subsection has preparatory character.\\
Let $\psi$ be a non-negative $C_0^{\infty}(\R)$ function such that \eqref{ws-03} is satisfied.
As above we put $\psi_{\mu}(x):=\psi(x-\mu)$, $\mu\in \Z,\ x\in \R$.
\begin{definition}\label{def-unif}
Let $0<p,q\leq \infty$ and $s \in \re$.
\\
{\rm (i)} $B^{s,\ell oc}_{p,q}(\R)$ denotes the collection of all $g \in \cs'(\R)$ such that
$ \varphi \cdot g \, \in B^s_{p,q}(\R)$ for all $\varphi \in C_0^\infty (\R)$.
\\
{\rm (ii)}
Let $0 < u \leq \infty$. Then $B^{s}_{p,q,v}(\R)$ is the collection of all $ g\in B^{s,\ell oc}_{p,q}(\R)$ such that
\beqq
\| \, g\, |B^s_{p,q,v}(\R)\| := \bigg(\sum_{\mu\in \Z}\| \psi_{\mu}f|B^s_{p,q}(\R)\|^v \bigg)^{1/v} <\infty
\eeqq
with the usual modification in case $u=\infty$.
\end{definition}
\begin{remark}
\rm
Obviously, the spaces $B^s_{p,q,\infty}(\R)$ and $B^s_{p,q}(\R)_{\unif}$ coincide.
In addition we wish to mention that the classes $B^{s}_{p,q,v}$ are quasi-Banach spaces,
independent of $\psi$ in the sense of equivalent quasi-norms.
\end{remark}
\begin{proposition}\label{s<p,q<u}
Let $0<p,q,v\leq \infty$ and $s>d\max(0, \frac 1p-1)$.
\\
{\rm (i)} The embedding $B^s_{p,q,v}(\R)\hookrightarrow B^s_{p,q}(\R)$ holds if and only if $v \leq \min(p,q)$.
\\
{\rm (ii)} The embedding $B^s_{p,q}(\R)\hookrightarrow B^s_{p,q,v}(\R)$ holds if and only if $v\geq \max(p,q)$.
\end{proposition}
\begin{proof}
{\it Step 1.} Sufficiency.\\
{\it Substep 1.1.} Proof of (i). Let $v \leq \min(p,q)$ and let $m$ be a natural number with $m>s$.
We suppose $g \in B^s_{p,q,v}(\R)$.
In case $0< v < \infty$ this implies
$\psi_\mu g \in L_{r} (\R)$ with $r:= \max(1,p)$, see Lemma \ref{emb1}, and at least formally
\beq\label{ws-07}
\Big\| \sum_{\mu \in \Z} \psi_{\mu}g \big | L_r(\R)\Big\|
& \lesssim &
\bigg(\sum_{\mu \in \Z} \big\| \psi_{\mu} g \big| L_r(\R)\big\|^r\bigg)^{1/r}\nonumber \\
& \lesssim & \bigg(\sum_{\mu \in \Z} \big\| \psi_{\mu} g \big| L_r(\R)\big\|^v\bigg)^{1/v}
\lesssim \| \, g \, | B^s_{p,q,v}(\R)\|\, .
\eeq
However, convergence of $\sum_{\mu \in \Z} \psi_{\mu}g $ in $L_r (\R)$ can be derived from \eqref{ws-07} as well
(because of $v<\infty$). This implies that $\sum_{\mu \in \Z} \psi_{\mu}g$ is a regular distribution. Hence,
the following argument makes sense
\beqq
&& \hspace{-0.7cm}
\Big\|\sum_{\mu\in \Z} \psi_{\mu}g \Big|B^s_{p,q}(\R) \Big\|^v_m \\
& \lesssim &
\bigg(\sum_{\mu \in \Z} \big\| \psi_{\mu}g \big | L_p(\R)\big\|^p\bigg)^{v/p} + \Bigg( \sum_{k=0}^{\infty} \bigg\{2^{sk }
\sup_{|h|<2^{-k}}\Big\| \sum_{\mu\in \Z} |\Delta_{h}^m(\psi_{\mu}g)(\cdot)|\Big| L_p(\R )\Big\|\bigg\}^q\Bigg)^{v/q}
\\
&\lesssim & \|\, g\, |B^s_{p,q,v}(\R)\|^v + \Bigg( \sum_{k=0}^{\infty} \bigg\{2^{skv} \sup_{|h|<2^{-k}}
\Big\| \Big( \sum_{\mu\in \Z} |\Delta_{h}^m(\psi_{\mu}g)(\cdot)|\Big)^v\Big| L_{p/v}(\R )\Big\|\bigg\}^{q/v}\Bigg)^{v/q} .
\eeqq
Since $ \Delta_h^m(\psi_{\mu}g)(x) \equiv 0$ if $|x-\mu|>c$ for some appropriate positive $c$ (independent of $\mu$) we get
\beqq
\bigg( \sum_{\mu\in \Z} |\Delta_{h}^m(\psi_{\mu}g)(x)|\bigg)^v \leq C \sum_{\mu\in \Z} |\Delta_{h}^m(\psi_{\mu}g)(x)|^v
\eeqq
with a constant $C$ independent of $g$ and $x$.
Inserting this into the previous inequality and applying triangle inequality with respect to $L_{p/v} (\R)$ first, afterwards
with respect to $\ell_{q/v}$,
we obtain
\beqq
\Big\|\, \sum_{\mu\in \Z} \psi_{\mu}g \, \Big|B^s_{p,q}(\R)\Big\|^v \lesssim \sum_{\mu\in \Z}\big\| \psi_{\mu}g\, \big|B^s_{p,q}(\R) \big\|^v
\, .
\eeqq
Because of $g = \sum_{\mu\in \Z} \psi_{\mu}g $ in $L_r (\R)$, see \eqref{ws-07}, we
have coincidence also almost everywhere and therefore also in $B^s_{p,q}(\R)$ by using Proposition \ref{diff}.
Hence we conclude
\beqq
\big\|\,g \, \big|B^s_{p,q}(\R)\big\| \lesssim \| \, g\, |B^s_{p,q,v}(\R) \|
\, .
\eeqq
This implies sufficiency in (i) in case $0< v< \infty$.
Now we consider the case $ v=p=q = \infty$. For each $x\in \R$ we choose $\mu\in \Z$ such that $x\in \supp\psi_\mu$.
Denote $$\Omega_{\mu}:= \big\{\nu\in \Z: \dist(\supp\psi_\mu,\supp\psi_\nu)\leq m\big\}$$
and observe that the cardinality of the sets $\Omega_\mu$ is uniformly bounded in $\mu$.
For $|h|<1$, \eqref{ws-03} and Proposition \ref{diff} in case $p=q=\infty$ yield
\beqq
|h|^{-s}|\Delta_h^mg(x)|&\leq& \sum_{\nu\in\Omega_\mu}|h|^{-s}|\Delta_h^m (g\psi_\nu)(x)|\\
&\leq& \sum_{\nu\in \Omega_\mu}|h|^{-s}\| \Delta_h^m (g\psi_\nu)(\cdot)|C(\R)\| \leq C\, \|\, g\, |B^s_{\infty,\infty,\infty}(\R)\|\, ,
\eeqq
which implies
\beqq
\|\, g\, |B^s_{\infty,\infty}(\R)\|_m = \| \, g\, |C(\R)\| + \sup_{|h|<1}\sup_{x\in \R}|h|^{-s}| \Delta_h^mg(x)| \lesssim
\|\, g\, |B^s_{\infty,\infty,\infty}(\R)\|\, .
\eeqq
{\it Substep 1.2.} Proof of (ii). Let $g \in B^s_{p,q} (\R)$ and assume that $v\geq \max(p,q)$.
By Lemma \ref{emb1} we conclude that $g \in L_p (\R)$. It follows
\be\label{ws-10}
\bigg(\sum_{\mu \in \Z} \big\| \psi_{\mu}g \big | L_p(\R)\big\|^v\bigg)^{1/v} \leq \bigg(\sum_{\mu \in \Z} \big\| \psi_{\mu}g \big | L_p(\R)\big\|^p\bigg)^{1/p}
\asymp \| \, g\, |L_p(\R)\| \lesssim \|\, g\, |B^s_{p,q}(\R)\|.
\ee
Next we consider the term
\[
A(g):= \Bigg\{\sum_{\mu\in \Z}\bigg(\sum_{k=0}^{\infty}2^{skq}\sup_{|h| < 2^{-k}}\|\Delta_h^m(\psi_{\mu}g)(\cdot) |L_p(\R)\|^q \bigg)^{v/q}\Bigg\}^{1/v} \, .
\]
Since $v\ge q$ it follows
\[
A(g)\le
\Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}}\|\Delta_h^m(\psi_{\mu}g)(\cdot) |L_p(\R)\|^v \bigg)^{q/v}\Bigg\}^{1/q}.
\]
For the next step of our estimate we shall
use the convention that $\varphi_\ell \equiv 0$ if $\ell < 0$ and define $g_{\ell} :=\gf^{-1}(\varphi_{\ell}\gf g)$, $\ell \in \zz$, see \eqref{ws-09}.
It follows
\be \label{decom}
\psi_{\mu} g =\psi_{\mu} \sum_{\ell \in \zz} \gf^{-1}(\varphi_{k+\ell} \gf g) = \sum_{\ell \in \zz} \psi_{\mu} g_{k+\ell} \, ,
\ee
which is valid in $L_p (\R)$, see Lemma \ref{emb1}. Hence, using the monotonicity of the $\ell_r$-norms, we find
\beq \label{k-02}
A(g) &\leq &
\Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}}
\Big\|\sum_{\ell\in \zz} |\Delta_h^m(\psi_{\mu}g_{k+\ell})(\cdot) |\Big|L_p(\R)\Big\|^v \bigg)^{q/v}\Bigg\}^{1/q}
\nonumber
\\
& \leq & \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}}\Big\|\sum_{\ell\in \zz} |\Delta_h^m(\psi_{\mu}g_{k+\ell})(\cdot)
|\Big|L_p(\R)\Big\|^p \bigg)^{q/p}\Bigg\}^{1/q}\, .
\eeq
Temporarily we assume $p\leq 1$. By means of $|a+b|^p \le |a|^p + |b|^p$, $a,b \in \re$, this yields
\[
A(g) \leq \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\ell\in \zz}\sum_{\mu\in \Z} \sup_{|h| < 2^{-k}}\big\|\Delta_h^m(\psi_{\mu}g_{k+\ell})(\cdot)
|L_p(\R)\big\|^p \bigg)^{q/p}\Bigg\}^{1/q} \, .
\]
Recall from Substep 1.1, that $\Delta_h^{m}(\psi_{\mu}g_{k+\ell})(x)=0 $ if $|x-\mu|>c$. In case $|x-\mu|<c$ we apply Lemma \eqref{help} with $\ell<0$ and obtain
\beqq
\sup_{|h| < 2^{-k}} \,
\big|\Delta_h^{m}(\psi_{\mu}g_{k+\ell})(x)\big| \leq C 2^{m\ell} P_{2^{k+\ell},a} g_{k+\ell}(x)\, , \qquad x \in \R\, ,
\eeqq
which leads to
\beqq
\sum_{\mu\in \Z} \sup_{|h| < 2^{-k}} \big\| \Delta_h^{m}(\psi_{\mu}f_{k+\ell})(\cdot) \big| L_p(\R) \big\|^p & \lesssim &
2^{m\ell p} \, \big\| P_{2^{k+\ell},a} g_{k+\ell}(\cdot) \big|L_p(\R) \big\|
\\
& \lesssim & 2^{m\ell p} \, \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^p
\eeqq
as long as we choose $a>d/p$, see Proposition \ref{peetremax}.
In case $\ell \ge 0$ we use the obvious elementary inequality
\beqq
\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}} \big\| \Delta_h^{m}(\psi_{\mu}g_{k+\ell})(\cdot) \big| L_p(\R) \big\|^p
\lesssim \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^p.
\eeqq
Altogether we have found
\be \label{Pfa}
\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}} \big\| \Delta_h^{m}(\psi_{\mu}g_{k+\ell})(\cdot) \big| L_p(\R) \big\|^p
\lesssim \min\big\{1 , 2^{m\ell p}\big\} \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^p
\ee
for all $\ell\in \zz$. Inserting this into \eqref{k-02} we obtain
\beqq
A(g)
\lesssim \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\ell\in \zz}
\min\big\{1 , 2^{m\ell p}\big\} \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^p \bigg)^{q/p}\Bigg\}^{1/q}.
\eeqq
Now, if $q/p\leq 1$, we conclude
\beq\label{ws-11}
A(g) & \lesssim &
\bigg\{ \sum_{\ell\in \zz}\min\big\{1 , 2^{m\ell q}\big\} \sum_{k=0}^{\infty}2^{skq} \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^q \bigg\}^{1/q}
\nonumber
\\
& \leq & \bigg\{ \sum_{\ell\in \zz}\min\big\{1 , 2^{m\ell q}\big\}2^{-s\ell q} \sum_{k=0}^{\infty}2^{s(k+\ell)q} \big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^q \bigg\}^{1/q}
\nonumber
\\
& \lesssim & \| \, g \, |B^s_{p,q}(\R)\|
\eeq
since $s < m$.
If $q/p >1$ we use the triangle inequality in $\ell_{q/p}$ and find
\beq\label{ws-12}
A(g)\lesssim \Bigg\{\sum_{\ell\in \zz} \min\big\{1 , 2^{m\ell p}\big\} \bigg(\sum_{k=0}^{\infty}2^{skq}
\big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\|^q \bigg)^{p/q}\Bigg\}^{1/p} \lesssim \| \, g\, |B^s_{p,q}(\R)\| .
\eeq
Summarizing, the embedding $B^s_{p,q}(\R)\hookrightarrow B^s_{p,q,v}(\R)$ in case $0 < p \le 1$ follows from
\eqref{ws-10}, \eqref{ws-11} and \eqref{ws-12}.\\
Now we turn to the case $p>1$. Inequality \eqref{k-02} yields
\beqq
A(g)
& \lesssim & \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\mu\in \Z}\Big(\sum_{\ell\in \zz}\sup_{|h| < 2^{-k}}
\big\| \Delta_h^m(\psi_{\mu}g_{k+\ell})(\cdot) \big|L_p(\R)\big\|\Big)^p \bigg)^{q/p}\Bigg\}^{1/q} \\
& \lesssim & \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\ell\in \zz}\Big(\sum_{\mu\in \Z}\sup_{|h| < 2^{-k}}
\big\| \Delta_h^m(\psi_{\mu}g_{k+\ell})(\cdot) \big|L_p(\R)\big\|^p\Big)^{1/p} \bigg)^{q }\Bigg\}^{1/q} \\
& \lesssim & \Bigg\{\sum_{k=0}^{\infty}2^{skq}\bigg(\sum_{\ell\in \zz}\min\big\{1 , 2^{m\ell }\big\}
\big\| g_{k+\ell}(\cdot) \big|L_p(\R) \big\| \bigg)^{q }\Bigg\}^{1/q} ,
\eeqq
see \eqref{Pfa}. Next we divide into two cases: $q\geq 1$ and $q<1$. Then we can continue as before and obtain
$B^s_{p,q}(\R)\hookrightarrow B^s_{p,q,v}(\R)$ also in case $1 \le p \le \infty$.\\
{\it Step 2.} Necessity.\\
{\em Substep 2.1.} The $p$-dependence.
Without loss of generality we assume that $\psi\equiv 1$ on $[-\delta,\delta]^d$ for some
$0<\delta<1$ and $\supp\psi \subset [-1,1]^d$. Let $\phi\in C_0^{\infty}(\R)$ be a nontrivial function
such that $\supp \phi\subset [-\delta,\delta]^d $.
Then we define a sequence $\{\mu_\ell\}_{\ell=0}^{n}$ with $\mu_\ell:=(4m\ell,0,\ldots,0)\in \Z$ and a sequence of functions
\[
g_n := \sum_{\ell=0}^{n} C_{\ell}\, \phi(x-\mu_\ell)\, , \qquad x \in \R\, ,
\]
where the real numbers $C_\ell>0$, $\ell=1,\, \ldots\, ,n$, will be chosen later on.
Elementary calculations, based on the use of $\|\, \cdot \, |B^s_{p,q} (\R)\|_m$, yield
\beqq
\| \, g_n\, |B^s_{p,q,v}(\R)\| \asymp \Big(\sum_{\ell=0}^n C_{\ell}^v\Big)^{1/v} \qquad \text{and} \qquad \|\, g_n\, |B^s_{p,q}(\R)\| \asymp \Big(\sum_{\ell=0}^n C_{\ell}^p\Big)^{1/p}.
\eeqq
with hidden constants independent of $n$ and $(C_\ell)_\ell$.
This implies the relations of $u$ to $p$.
\\
{\em Substep 2.2.} The $q$-dependence. What concerns the $q$-dependence it is not longer convenient to work with differences.
We will switch to wavelets. Wavelet bases in Besov spaces are a
well-developed concept. We refer to the monographs of Meyer \cite{me},
Wojtasczyk \cite{woj} and Triebel \cite{Tr06,t08} for the general
$d$-dimensional case (for the one-dimensional case see also the book of Kahane and Lemarie-Rieuseut
\cite{kl}). Let $\phi$ be a compactly supported sufficiently smooth scaling function and
let
$\tilde\psi_1,\,\ldots,\,\tilde\psi_{2^d-1}$ be associated wavelets, all defined on $\R$.
In addition we assume that $\supp \tilde{\psi}_1 \subset [-T,T]$ for some $T>0$.
For $j\in \N_0$, $k\in \Z$, and $i=1,\ldots, 2^d-1$, we shall make use of the standard abbreviations in this context:
\[
\phi_{0,k}(x):= \phi(x-k) \qquad \mathrm{and}\qquad
\tilde\psi_{i,j,k}(x):= 2^{jd/2}\, \tilde\psi_i(2^jx-k),\qquad x\in\R.
\]
Then any $g \in B^s_{p,q}(\R)$ admits an unique representation
\begin{equation}\label{4.22}
g = \sum_{k\in\zz^d} a_k \, \phi_{0,k} + \sum_{i=1}^{2^d-1} \sum_{j=0}^\infty
\sum_{k\in\Z} a_{i,j,k}\, \tilde\psi_{i,j,k}
\end{equation}
in $\cs'(\R)$, where
\[
a_k := \langle g,\,\phi_{0,k}\rangle \qquad \mbox{and} \qquad
a_{i,j,k}:= \langle g,\,\tilde\psi_{i,j,k}\rangle \, .
\]
Moreover,
\be\label{ws-13}
\|\, g \, |{B^s_{p,q}(\R)}\| \asymp
\bigg(\sum_{k\in\Z} |a_k|^p\bigg)^{1/p} + \Bigg[\sum_{i=1}^{2^d-1}\sum_{j=0}^\infty \, 2^{j(s+d/2)q}
\bigg(\sum_{k\in\Z} 2^{-jd}\, |a_{i,j,k}|^p\bigg)^{q/p}\Bigg]^{1/q},
\ee
in the sense of equivalent quasi-norms; see, e.\,g., \cite[Theorem 1.20]{t08}.
Now we define
\[
g_{\alpha, \mu} := \sum_{j=1}^\infty \alpha_j \, \tilde\psi_{1,j,\mu_j}
\]
for some sequence $\alpha := (\alpha_j)_j$ of positive numbers and some sequence $\mu:= (\mu_j)_j \subset \Z$ to be chosen later on.
It follows from \eqref{ws-13}
\be\label{ws-14}
\|\, g_{\alpha,\mu} \, |{B^s_{p,q}(\R)}\| \asymp
\bigg(\sum_{j=1}^\infty \, 2^{j\big(s+d (\frac 12 - \frac 1p)\big)q} \, |\alpha_j|^q\bigg)^{1/q}
\ee
with hidden constants independent of $\alpha$ and $\mu$.
Without loss of generality we may assume that
our function $\psi$ used in the definition of $B^s_{p,q,v} (\R)$ satisfies
$\psi\equiv 1$ on $[-\delta,\delta]^d$ for some
$0<\delta<1$ and $\supp\psi \subset [-1,1]^d$.
Next we choose a natural number $M$ such that
\[
\supp \tilde\psi_{1,j,0} \subset [-\delta,\delta]^d \qquad \mbox{for all}\quad j \ge M\, .
\]
Then it is easily checked that we get
\be\label{ws-15a}
\|\, g_{\alpha,\mu} \, |{B^s_{p,q,v}(\R)}\| \asymp
\bigg(\sum_{j=1}^\infty \, 2^{j\big(s+d (\frac 12 - \frac 1p)\big)v} \, |\alpha_j|^v \bigg)^{1/v}
\ee
for all sequences $\alpha$ satisfying $\alpha_1 = \ldots = \alpha_M =0$ and
$\mu$ chosen as $\mu_j:=(j\, 2^j,0,\ldots,0)$.
Based on \eqref{ws-14} and \eqref{ws-15a}, the relations of $q$ to $v$ follow.
\end{proof}
\begin{remark}
\rm
Probably Bourdaud \cite{Bou} was the first who had considered the classes $B^s_{p,q,v}(\R)$ with $p \neq q$ and $v \neq \infty$.
He already investigated
the embeddings in Proposition \ref{s<p,q<u} in case $v=p$.
\end{remark}
For convenience of the reader we state one obvious but important consequence of Proposition \ref{s<p,q<u}.
\begin{corollary}\label{klar}
Let $0 <p \le \infty$ and $s>d\max(0, \frac 1p-1)$. Then $B^s_{p,p}(\R)=B^s_{p,p,p}(\R)$ in the sense of equivalent quasi-norms.
\end{corollary}
\begin{remark}
\rm
Corollary \ref{klar} is well-known, we refer to Peetre \cite[page 150]{Pe} ($p \ge 1$),
Maz'ya and Shaposnikova \cite[Lem.~3.1.1.9]{MS1}, \cite[Prop.~4.2.6]{MS2} ($p \ge 1$),
and Triebel \cite[2.4.7]{Tr92} (general case).
\end{remark}
\subsection{Proof of Theorem \ref{b-main}} \label{main3}
The heart of the matter consists in the following proposition.
\begin{proposition}\label{wichtig}
Let $0<p\leq q\leq \infty$ and $s>d/p$. Then there exists a constant $C>0$ such that
\be \label{k-01}
\|\, fg\, |B^s_{p,q}(\R)\| \leq C\, \| g\,|\, B^s_{p,q}(\R)\| \, \|\, f\,|\, B^s_{p,q}(\R)_{\unif} \|
\ee
holds for all $g\in B^s_{p,q}(\R)$ and all $f\in B^s_{p,q}(\R)_{\unif}$.
\end{proposition}
\begin{proof}
Let $m-1\leq s < m$. For technical reasons we shall estimate $\|\, fg\, |B^s_{p,q}(\R)\|_{2m} $.
Let $\psi, \phi \in C_0^{\infty}(\R)$ are chosen such that the following holds:
$\psi$ is nontrivial and satisfies \eqref{ws-03} and $\phi \equiv 1$ on $\supp \psi $. As explained
in the proof of Proposition \ref{s<p,q<u} we have
\beqq
\| \, fg\, |B^s_{p,q}(\R)\|_{2m} =\Big\| \sum_{\mu\in \Z} \phi_{\mu} g \psi_{\mu} f \, \Big|B^s_{p,q}(\R)\Big\|_{2m}\, .
\eeqq
Clearly,
\beq \label{ws-15}
\Big \| \sum_{\mu\in \Z} \phi_{\mu} g \psi_{\mu}f \Big| L_p(\R)\Big\| & \leq &
\Big\| \sum_{\mu \in \Z} |\phi_{\mu}g | \cdot \| \psi_{\mu} f\, |L_\infty(\R)\|\, \Big | L_p(\R)\Big\|
\nonumber
\\
& \lesssim & \|\, g \, |L_p(\R)\| \cdot \, \sup_{\mu\in \Z}\, \|\, \psi_{\mu}f\, |L_\infty(\R)\|
\nonumber
\\
& \lesssim & \| \, g\, |B^s_{p,q}(\R)\| \, \|\, f\, |B^s_{p,q}(\R)_{\unif}\|
\eeq
where we used in the last step the embedding $B^s_{p,q}(\R) \hookrightarrow L_\infty (\R)$, see Lemma \ref{emb1}.
Next we need some identities for differences.
Note that if $F,G:\ \R \to \C$ are two functions and $n \in \N$ we have
\beqq
\Delta_h^{n}(F\, \cdot \, G)(x) = \sum_{j=0}^{n} \binom{n}{j} \, \Delta_h^{n-j}F(x+j h)\, \Delta_h^{j}G(x),\qquad x,h\in \R \, .
\eeqq
This can be proved by induction on $n$.
Making use of this formula we obtain
\be\label{ws-15b}
|\Delta_h^{2m} (fg)(x)| \leq \sum_{u=0}^{2m} \binom{2m}{u} \sum_{\mu\in \Z}
\big| \Delta_h^{2m-u}(\phi_{\mu}g)(x+uh)\Delta_h^u(\psi_{\mu}f)(x)\big|,\ \ \ h,x\in \R.
\ee
Let $|h|\le 1$.
Since $ \Delta_h^u (\phi_{\mu}g)(x+uh) \equiv \Delta_h^u (\psi_{\mu} f)(x) \equiv 0$ if $|x-\mu|>c$ for some appropriate positive $c$
(independent of $\mu$) this yields
\beqq
S_u: & = & \Bigg\{\sum_{k=0}^{\infty} 2^{ksq} \sup_{|h|< 2^{-k}}
\Big\|\sum_{\mu\in \Z} \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot+uh)\Delta_h^u(\psi_{\mu}f)(\cdot) \Big|L_p(\R)
\Big\|^q \Bigg\}^{1/q}
\\
& \lesssim & \Bigg\{\sum_{k=0}^{\infty}\, 2^{ksq} \sup_{|h|< 2^{-k}}\, \bigg( \sum_{\mu\in \Z}\big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot+uh)
\Delta_h^u(\psi_{\mu}f)(\cdot) \big|L_p(\R) \big\|^p\bigg)^{q/p} \Bigg\}^{1/q}
\eeqq
for any $u$, $0 \le u \le 2m$. To estimate $S_u$ we have to distinguish into different cases.\\
{\it Step 1.} The case $0 \le u < m$. Recall, $B^s_{p,q} (\R) \hookrightarrow L_\infty (\R)$ under the given restrictions, see Lemma \ref{emb1}. It follows
\beqq
\big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot +uh)&& \hspace{-0.9cm}\Delta_h^u(\psi_{\mu}f)(\cdot) \big|L_p(\R) \big\|
\\
& \leq & \big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot+uh) \big|L_p(\R) \big\|\cdot \big\|\Delta_h^u(\psi_{\mu}f)(\cdot) \big| L_\infty (\R)\big\|
\\
& \lesssim & \big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot+uh) \big|L_p(\R) \big\|\cdot \big\| \psi_{\mu}f \, \big|\, L_\infty(\R)\big\|\\
& \leq & \big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot) \big|L_p(\R) \big\| \cdot \| \, f\, |\, B^s_{p,q }(\R)_{\unif}\|\, ,
\eeqq
where in the last step we performed a change of variable. Consequently, we obtain
\beqq
S_u \lesssim \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp}\sup_{|h|< 2^{-k}} \sum_{\mu\in \Z}
\big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot) \big|L_p(\R)
\big\|^p\bigg)^{q/p} \Bigg\}^{1/q} \cdot \| \, f \, |B^s_{p,q}(\R)_{\unif}\|\, .
\eeqq
Now we deal with the term $\{\, \ldots \, \}^{1/q}$ on the right-hand side. As in proof of Proposition \ref{s<p,q<u},
see formula \eqref{decom}, we use the decomposition
\beqq
\phi_{\mu} g = \sum_{\ell \in \zz} (\phi_{\mu} g_{k+\ell})\,.
\eeqq
This yields
\beqq
\Bigg\{\sum_{k=0}^{\infty} && \hspace{-0.6cm}
\bigg(2^{ksp}\sup_{|h|< 2^{-k}} \sum_{\mu\in \Z} \big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot) \big|L_p(\R)
\big\|^p\bigg)^{q/p} \Bigg\}^{1/q}
\\
& \le & \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp}\sup_{|h|< 2^{-k}} \sum_{\mu\in \Z}
\Big\| \sum_{\ell \in \zz} |\Delta_h^{2m-u}(\phi_{\mu}g_{k+\ell})(\cdot)|\, \Big|L_p(\R)
\Big\|^p\bigg)^{q/p} \Bigg\}^{1/q} \, .
\eeqq
Since $2m-u \ge m$ we can proceed as in Substep 1.2, proof of Proposition \ref{s<p,q<u}, starting at formula \eqref{k-02}.
As a result we find
\be\label{ws-17}
S_u\lesssim \| \, g \, | B^s_{p,q}(\R)\| \, \| f| B^s_{p,q}(\R)_{\unif}\|
\ee
for all $u$, $0 \le u \le m$.
\\
{\it Step 2.} The case $m < u \le 2 m$. We have
\beqq
\big\| \Delta_h^{2m-u}(\phi_{\mu}g)(\cdot+uh) \Delta_h^u(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|
& \leq &
\big\| \Delta_h^{2m-u}(\phi_{\mu}g)\big|L_\infty(\R) \big\| \, \big\|\Delta_h^u(\psi_{\mu}f) \big| L_p(\R)\big\|
\\
& \lesssim & \big\| \phi_{\mu}g \big|L_\infty(\R) \big\| \, \big\|\Delta_h^u(\psi_{\mu}f) \big| L_p(\R)\big\|\, .
\eeqq
Inserting this into $S_u$ we find
\be\label{ws-23}
S_u \leq \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h|< 2^{-k}}
\sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^p \, \big\|\Delta_h^u(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|^p
\bigg)^{q/p} \Bigg\}^{1/q}.
\ee
Using the triangle inequality in $L_{q/p}$ we arrive at
\beqq
S_u & = & \Bigg\{\sum_{\mu\in \Z} \bigg( \big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^q \sum_{k=0}^{\infty} 2^{ksq}
\sup_{|h|< 2^{-k}} \big\|\Delta_h^u(\psi_{\mu}f) \big| L_p(\R)\big\|^q \bigg)^{p/q} \Bigg\}^{1/p}
\\
&\leq & \bigg\{\sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^p \,
\big\|\, \psi_{\mu}f\, |B^s_{p,q}(\R)\|^p \bigg\}^{1/p}
\\
& \leq &\bigg\{\sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^p \bigg\}^{1/p} \,
\big\|\, f\, |B^s_{p,q}(\R)_{\unif}\|.
\eeqq
Since $s>d/p$, there exists $\varepsilon>0$ such that $s-\varepsilon>d/p$.
This implies $B^{s-\varepsilon}_{p,p}(\R)\hookrightarrow L_\infty(\R)$, see Lemma \ref{emb1}. Hence we obtain the estimate
\be\label{ws-19}
\bigg( \sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^p \bigg)^{1/p}
\lesssim \bigg( \sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|B^{s-\varepsilon}_{p,p}(\R) \big\|^p \bigg)^{1/p}
\asymp \| \, g\, |B^{s-\varepsilon}_{p,p}(\R)\|
\ee
where we used Corollary \ref{klar} with respect to $B^{s-\varepsilon}_{p,p}(\R)$.
The elementary embedding $B^s_{p,q}(\R) \hookrightarrow B^{s-\varepsilon}_{p,p}(\R)$ implies
\be\label{ws-18}
S_u\lesssim \| \, g\,|\, B^s_{p,q}(\R)\| \, \|\, f\,|\, B^s_{p,q}(\R)_{\unif} \|
\ee
also in this situation.
Summarizing, \eqref{ws-15}, \eqref{ws-15b}-\eqref{ws-18}, prove the claim \eqref{k-01}.
\end{proof}
\begin{remark}
\rm
Let us mention that the inequality \eqref{ws-19} is not true in the limiting case $s=d/p$, $p < q\leq 1$.
From the Sobolev-type embedding $B^{d/p}_{p,q}(\R) \hookrightarrow B^{d/q}_{q,q}(\R)$, see \cite[2.7.1]{Tr83}, we derive
\beqq
\bigg( \sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^q \bigg)^{1/q}
\lesssim \bigg( \sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|B^{d/q}_{q,q}(\R) \big\|^q \bigg)^{1/q}
\lesssim \| \, g\, |B^{d/p}_{p,q}(\R)\|\, .
\eeqq
The exponent $q$ is best possible, i.e.,
if
\[
\bigg( \sum_{\mu\in \Z} \big\| \phi_{\mu}g \big|L_\infty (\R) \big\|^v \bigg)^{1/v} \lesssim \| \, g\, |B^{d/p}_{p,q}(\R)\|
\]
holds for all $g \in B^{d/p}_{p,q}(\R)$, then $v \ge q$ follows.
\end{remark}
\begin{proposition}\label{limes}
Let $0 <p \leq 1$ and $s=d/p$. Then there exists a constant $C>0$ such that
\beqq
\|\, fg\, |B^{d/p}_{p,p}(\R)\| \leq C\, \| g\,|\, B^{d/p}_{p,p}(\R)\| \, \|\, f\,|\, B^{d/p}_{p,p}(\R)_{\unif} \|
\eeqq
holds for all $g\in B^s_{p,p}(\R)$ and all $f\in B^s_{p,p}(\R)_{\unif}$.
\end{proposition}
\begin{proof}
Let $\phi$ and $\psi$ be as in proof of Proposition \ref{wichtig}.
Employing Corollary \ref{klar} and the algebra property of $B^{d/p}_{p,p}(\R)$, see Theorem \ref{algebra},
we get
\beqq
\| \, fg\,|\, B^s_{p,p}(\R)\| & \asymp & \bigg(\sum_{\mu \in \Z}
\| \, (\psi_{\mu} f )(\phi_\mu g)\,|\, B^s_{p,p}(\R)\|^p \bigg)^{1/p}
\\
& \lesssim & \bigg(\sum_{\mu \in \Z} \| \, \psi_\mu f \,|\, B^s_{p,p}(\R)\|^p
\| \, \phi_\mu g\,|\, B^s_{p,p}(\R)\|^p \bigg)^{1/p}
\\
& \lesssim & \|\, f \, |B^s_{p,p}(\R)_\unif\|\,
\bigg(\sum_{\mu \in \Z} \| \, \phi_\mu g\,|\, B^s_{p,p}(\R)\|^p \bigg)^{1/p}
\\
& \lesssim & \|\, f \, |B^s_{p,p}(\R)_\unif\|\,\, \| \, g\,|\, B^s_{p,p}(\R)\|\, ,
\eeqq
which proves the claim.
\end{proof}
\begin{remark}
\rm
As already mentioned above this result can be found in Netrusov \cite{Net} and Triebel \cite[Proposition 2.22]{Tr06}.
\end{remark}
\noindent
{\bf Proof of Theorem \ref{b-main}.}
From Proposition \ref{wichtig} we derive
$ B^s_{p,q}\R)_\unif \hookrightarrow M(B^s_{p,q}(\R))$, whereas from
Lemma \ref{mult}
$M(B^s_{p,q}(\R)) \hookrightarrow B^s_{p,q}(\R)_{\unif}$ follows.
\subsection{Proofs of Theorems \ref{final} and \ref{nec1}} \label{main4}
\begin{lemma}\label{unif-L}
Let $0<p,q\leq \infty$ and $s >d(1/p-1)_+$. Then we have
\beqq
B^s_{p,q}(\R)\hookrightarrow M^s_{p,q}(\R)\hookrightarrow B^s_{p,q}(\R)_{\unif}.
\eeqq
\end{lemma}
\begin{proof}
The right-hand embedding has been explained just before Theorem \ref{final}.
Also the left-hand embedding is easily seen. From the trivial inequality
\beqq
\| g|M^s_{p,q}(\R)\|\leq
\|g|L_p(\R)\|+ \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp}
\sup_{|h| < 2^{-k}} \sum_{\mu\in \Z} \big\|\Delta_h^m(\psi_{\mu}g)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\eeqq
and the same argument as used in Substep 1.2 of the proof of Proposition \ref{s<p,q<u}, see formula \eqref{k-02},
the left-hand embedding follows.
\end{proof}
\noindent
{\bf Proof of Theorem \ref{final}}.
{\it Step 1.} Let $m$ be a natural number such that $m-1 < s < m$. We claim that
\beqq
\| \, fg\, |B^s_{p,q}(\R)\|_{2m} \leq C\, \|\, f\, |M^s_{p,q}(\R)\| \, \| \, g\, |B^s_{p,q}(\R)\|
\eeqq
holds for all $f\in M^s_{p,q}(\R)$ and $g\in B^s_{p,q}(\R)$. As a first step of the proof we observe that
in \eqref{ws-17} we did not use the condition $p\le q$. Hence we can apply \eqref{ws-17} for all terms $S_u$ with $0 \le u \le m$.
Since $M^s_{p,q}(\R)\hookrightarrow B^s_{p,q}(\R)_{\unif}$ this is sufficient for us. It remains to deal with the case $ m < u \le 2m$.
Starting point for us is formula \eqref{ws-23}
\[
S_u \leq \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h| < 2^{-k}} \sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^p
\cdot \big\|\Delta_h^m(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}.
\]
Now we choose
\[
C_{\mu}:= c \, \frac{ \| \phi_{\mu}g |L_\infty(\R) \| }{ \|\, g\, |B^{s}_{p,q}(\R) \| }\, , \qquad \mu\in \Z,
\]
where $c$ will be fixed later on.
Of course, here we assume that $\|\, g \, |B^{s}_{p,q}(\R) \|>0$, otherwise there is nothing to prove.
With a proper choice of $c$, the sequence $\{C_{\mu} \}_{\mu\in \Z}$ belongs to the set $\ell_p^+$ because of
\[
\sum_{\mu \in \Z} C_\mu^p = c^p \, \sum_{\mu \in \Z} \frac{ \| \phi_{\mu}g |L_\infty(\R) \|^p }{ \|\, g\, |B^{s}_{p,q}(\R) \|^p}
\le 1\, ,
\]
see \eqref{ws-19}. Notice that $c$ can be chosen independent of $g$.
Hence, we conclude that
\beqq
S_u \lesssim \big\| \, g\, \big|B^s_{p,q}(\R)\big\|\cdot \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h| < 2^{-k}}
\sum_{\mu\in \Z} C_{\mu}^p \big\|\Delta_h^m(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\eeqq
which proves $M^s_{p,q}(\R)\hookrightarrow M(B^s_{p,q}(\R))$.\\
{\it Step 2.} Let $f \in M(B^s_{p,q}(\R))$ and $\{C_{\mu}\}_{\mu}\in \ell_p (\Z)$.
We subdivide $\Z$ into a finite number of disjoint sets $\Omega_\ell$, $\ell=1,\ldots,n$, such that
\[
\dist\big(\supp \psi_{\mu_i}, \supp \psi_{\mu_j}\big)\geq 2m\, ,
\]
for all $\mu_i,\mu_j\in \Omega_\ell$, $\mu_i \not=\mu_j$, and all $\ell=1,\ldots,n$.
With
\[
P_\mu := \{x \in \R: \: \dist (\supp \psi_\mu,x)\le m\}\, , \qquad \mu \in \Z\, ,
\]
we find
\[
\sum_{\mu\in \Omega_\ell} |C_{\mu}|^p \int_{P_\mu} |\Delta_h^m(\psi_{\mu}f)(x) |^p dx = \int_{\R} \Big|\Delta_h^m \Big(
\sum_{\mu\in \Omega_\ell} C_{\mu} \psi_{\mu} f \Big)(x) \Big|^p dx \,, \qquad |h|< 1.
\]
This implies
\beq \label{k-06}
\Bigg\{\sum_{k=0}^{\infty}&& \hspace{-0.7cm}\bigg(2^{ksp}
\sup_{|h| < 2^{-k}} \sum_{\mu\in \Z} |C_{\mu}|^p \big\|\Delta_h^m(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\nonumber
\\
&\lesssim & \sum_{\ell=1,\ldots,n} \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h| < 2^{-k}} \sum_{\mu\in \Omega_\ell}
|C_{\mu }|^p \big\|\Delta_h^m(\psi_{\mu }f)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\nonumber
\\
& \lesssim & \sum_{\ell=1,\ldots,n} \Big\| \sum_{\mu \in\Omega_\ell} C_{\mu } \psi_{\mu }f \, \Big|B^s_{p,q}(\R)\Big\|.
\eeq
Next, we make use of the definition of $M(B^s_{p,q}(\R))$
\be \label{k-05}
\Big\| f \sum_{\mu\in \Omega_\ell } C_{\mu } \psi_{\mu }\Big|B^s_{p,q}(\R)\Big\| \lesssim
\big\| f\big| M(B^s_{p,q}(\R)) \big\| \cdot \Big\| \sum_{\mu \in \Omega_\ell} C_{\mu } \psi_{\mu }\Big|B^s_{p,q}(\R)\Big\| .
\ee
A simple calculation leads to
\beqq
&& \hspace{-0.7cm} \Big\| \sum_{\mu \in \Omega_\ell} C_{\mu }\psi_{\mu }\Big|B^s_{p,q}(\R)\Big\|_m
\\
&\lesssim & \Big\| \sum_{\mu \in \Omega_\ell} C_{\mu } \psi_{\mu }\Big|L_{p}(\R)\Big\| +
\Bigg\{\sum_{k=0}^{\infty} 2^{ksq} \sup_{|h| < 2^{-k}} \bigg(\sum_{\mu\in \Omega_\ell} |C_{\mu}|^p
\big\|\Delta_h^m\psi_{\mu} (\cdot) \big|L_p(\R) \big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\\
&\lesssim & \bigg( \sum_{\mu \in \Omega_\ell} |C_{\mu }|^p \|\, \psi\, |L_{p}(\R)\|^p\bigg)^{1/p} +
\Bigg\{\sum_{k=0}^{\infty} 2^{ksq} \sup_{|h| < 2^{-k}} \bigg(\sum_{\mu\in \Z} |C_{\mu}|^p
\big\|\Delta_h^m\psi (\cdot) \big|L_p(\R) \big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\\
& \lesssim & \|\, \{C_\mu\}_\mu\, |\ell_p (\Z)\|\, \big\| \psi\big|B^s_{p,q}(\R)\big\|.
\eeqq
Inserting this into \eqref{k-05} we find
\be\label{ws-31}
\sup_{\|\, \{C_\mu\}_\mu\, |\ell_p (\Z)\| \le 1} \,
\Big\| f \sum_{\mu\in \Omega_\ell } C_{\mu } \psi_{\mu }\Big|B^s_{p,q}(\R)\Big\| \lesssim \big\|\, f\, \big| M(B^s_{p,q}(\R)) \big\|
\ee
for all $\ell=1,\ldots,n$.
Consequently, we obtain $\big\| f\big| M^s_{p,q}(\R) \big\| \lesssim \big\| f\big| M(B^s_{p,q}(\R)) \big\| $. The proof is complete.
\hspace*{\fill} \rule{3mm}{3mm}
\\
In Step 2 of the preceding proof we did not use the restrictions in $p,q$ and $s$. We only used the possibility to describe
the quasi-norm of $B^s_{p,q} (\R)$ by differences as explained in Proposition \ref{diff}. This yields the following.
\begin{lemma}\label{nec}
Let $ 0< p,q \le \infty$ and $s> d (\frac 1p -1)_+$. Then we have the continuous embedding
\beqq
M(B^s_{p,q}(\R)) \hookrightarrow M^s_{p,q}(\R)\, .
\eeqq
\end{lemma}
\noindent
{\bf Proof of Theorem \ref{nec1}}.
{\em Step 1.}
Let $f\in M(B^s_{p,q}(\R))$ and $(C_\mu)_\mu \in \ell_p (\Z)$.
Then \eqref{ws-31} yields what we need.
\\
{\em Step 2.} Let $f$ be a function such that
$\sum_{\mu \in \Z} C_\mu \psi_\mu f$ belongs to $B^s_{p,q}(\R)$
for all $\{C_\mu\}_\mu \in \ell_p (\Z)$ and
\beqq
\sup_{\| \{C_\mu\}_\mu \, |\ell_p(\Z) \|\le 1} \, \Big\|\sum_{\mu \in \Z} C_\mu \psi_\mu f\Big|B^s_{p,q}(\R)\Big\| < \infty \, .
\eeqq
By chosing $\{C_\mu\}_\mu$ appropriate it is immediate that
$f \in B^s_{p,q}(\R)_\unif$.
On the one hand it follows
\beqq
\Big\|\, f \, \sum_{\mu\in \Z} C_{\mu} \psi_{\mu}
\, \Big|L_p(\R)\Big\| \lesssim \Big\|f \, \sum_{\mu\in \Z} C_{\mu} \psi_{\mu} \Big|B^s_{p,q}(\R)\Big\| \le C_f < \infty
\eeqq
for all sequences $\{C_\mu\}_\mu $, $\|\, \{C_\mu\}_\mu\, |\ell_p (\Z)\| \le 1$.
On the other hand \eqref{k-06} yields
\beqq
\Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp}
\sup_{|h| < 2^{-k}} \sum_{\mu\in \Z} && \hspace{-0.7cm} |C_{\mu}|^p
\big\|\Delta_h^m(\psi_{\mu}f)(\cdot) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\\
& \lesssim & \sum_{\ell=1,\ldots,n} \Big\| \sum_{\mu \in\Omega_\ell} C_{\mu }\psi_{\mu }f \, \Big|B^s_{p,q}(\R)\Big\|
\lesssim C_f <\infty
\eeqq
for all sequences $\{C_\mu\}_\mu $, $\|\, \{C_\mu\}_\mu\, |\ell_p (\Z)\| \le 1$.
In view of the definition of $M^s_{p,q}(\R)$, combined with Theorem \ref{final}, we conclude that $f\in M(B^s_{p,q}(\R))$.
\hspace*{\fill} \rule{3mm}{3mm}
\subsection{Proof of Theorem \ref{final2} and Theorem \ref{final3}} \label{main5}
\noindent
{\bf Proof of Theorem \ref{final2}}. Because of $s>0$ the space $B^s_{\infty,q}(\R)$ forms an algebra with respect
to pointwise multiplication, see Theorem \ref{algebra}, i.e., $B^s_{\infty,q}(\R) \hookrightarrow M( B^s_{\infty,q}(\R))$.
On the other hand, the function $g \equiv 1$ belongs to $B^s_{\infty,q}(\R)$.
This implies that a pointwise multiplier $f$ has to belong to $B^s_{\infty,q}(\R)$ as well.
\hspace*{\fill} \rule{3mm}{3mm}
\\
\noindent
{\bf Proof of Theorem \ref{final3}}.
Part (i) follows from Proposition \ref{limes} and Lemma \ref{mult}.
Now we turn to (ii).
Let $s=d/p$ and $q\leq \min(1,p)$.
The needed modifications of Step 1 of the proof of Theorem \ref{final}
are based on the estimate
\beqq
\sum_{\mu\in \Z}\big\| \phi_{\mu}f \big|L_\infty(\R)\big\|^p
\lesssim \sum_{\mu\in \Z}\big\| \phi_{\mu}f \big|B^{s}_{p,q}(\R)\big\|^p \lesssim \big\|f\big|B^{s}_{p,q}(\R)\big\|^p,
\eeqq
where we first employed Lemma \ref{emb1} and afterwards Proposition \ref{s<p,q<u}.
This yields sufficiency. Necessity is a consequence of Lemma \ref{nec}.
\\
Finally, the arguments from the proof of Theorem \ref{nec1} carry over to this limiting situation described in (iii).
\hspace*{\fill} \rule{3mm}{3mm}
\subsection{Localized Besov spaces as subspaces of $M(B^s_{p,q}(\R))$ and consequences} \label{main6}
\begin{proposition}\label{loc}
Let $0<q<p\leq\infty$ and $0<v\leq \infty$. Let either $s >d/p$ or $s =d/p$ and $q\leq 1$.
Then we have $$B^s_{p,q,v}(\R)\hookrightarrow M(B^s_{p,q}(\R))$$
if and only if $1/v\geq 1/q-1/p$.
\end{proposition}
\begin{proof}
{\it Step 1.} Sufficiency. Proposition \ref{s<p,q<u} yields that it is enough to consider the case $1/v=1/q-1/p$.
Let $s<m\leq s+1$. We shall prove that
\beqq
\|\, fg\, |B^s_{p,q}(\R)\|_{2m} \leq C\, \|\, g\, |B^s_{p,q}(\R)\| \, \|\, f\, |B^s_{p,q,v}(\R)\|
\eeqq
holds for all $g\in B^s_{p,q}(\R)$ and $f\in B^s_{p,q,v}(\R)$.
From now on we shall follow the proof of Proposition \ref{wichtig}. Observe $B^s_{p,q,v}(\R) \hookrightarrow B^s_{p,q,\infty}(\R)=
B^s_{p,q}(\R)_\unif$.
Since the condition $p\leq q$ is not needed in Step 1 and Substep 2.1 of the proof of Proposition \ref{wichtig},
it is enough to deal with the case $ m < u \le 2m$. From \eqref{ws-23}, $q< p$ and Proposition \ref{diff} we derive that
\beqq
S_u &\leq & \Bigg\{\sum_{k=0}^{\infty}\bigg(2^{ksp} \sup_{|h| < 2^{-k}} \sum_{\mu\in \Z}
\big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^p \cdot \big\|\Delta_h^u(\psi_{\mu}f) \big| L_p(\R)\big\|^p \bigg)^{q/p} \Bigg\}^{1/q}
\\
&\leq & \Bigg\{\sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^q\sum_{k=0}^{\infty} 2^{ksq} \sup_{|h| < 2^{-k}}
\big\|\Delta_h^u(\psi_{\mu}f)\big| L_p(\R)\big\|^q \Bigg\}^{1/q}
\\
&\leq & \Bigg\{\sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^q \cdot \big\| \, \psi_{\mu}f\, \big| B^s_{p,q}(\R)\big\|^q \Bigg\}^{1/q}.
\eeqq
Now H\"older's inequality with $\frac{1}{v}+\frac{1}{p}=\frac 1q$, Lemma \ref{emb1} and Corollary \ref{klar} yield
\beqq
S_u
&\leq &
\Bigg\{\sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|L_\infty(\R) \big\|^p\Bigg\}^{1/p}
\Bigg\{ \sum_{\mu\in \Z} \big\| \, \psi_{\mu}f \, \big| B^s_{p,q}(\R)\big\|^v\Bigg\}^{1/v}
\\
&\lesssim & \bigg(\sum_{\mu\in \Z}\big\| \phi_{\mu}g \big|B^s_{p,p}(\R) \big\|^p\bigg)^{1/p} \, \big\| \, f\,\big| B^s_{p,q,v}(\R)\big\|\\
&\lesssim & \|\, g\, |B^s_{p,q}(\R)\| \, \| \, f\, |B^s_{p,q,v}(\R)\|.
\eeqq
\noindent
{\it Step 2.} Necessity.
We shall employ the same type of arguments as in Substep 2.2 of the proof of Proposition \ref{s<p,q<u}, see in particular \eqref{4.22} and \eqref{ws-13}.
Then we choose
\[
f_{M,N,\alpha}(x) := \sum_{j=M}^{N} \, \alpha_j\, \tilde{\psi}_{1,j,\mu_j} (x)\, , \qquad \mu_j := (4^j, 0, \ldots , 0)
\]
for some sequence $\alpha:= (\alpha_j)_j$, $N$ and $M$ (to be fixed later).
Define $\gamma_j := 2^{j\big(s + d(\frac 1 2 - \frac 1 p)\big)} |\alpha_j|$, $j \ge M$.
Then, by making use of the same conventions as in \eqref{ws-15a}, we obtain
\[
\| \, f_{M,N,\alpha} \, | B^s_{p,q} (\R) \| \asymp \bigg( \sum_{j=M}^N 2^{j\big(s + d(\frac 1 2 - \frac 1 p)\big)q} |\alpha_j|^q \bigg)^{1/q} =
\bigg( \sum_{j=M}^N |\gamma_j|^q \bigg)^{1/q}
\]
and
\[
\| \, f_{M,N,\alpha} \, | B^s_{p,q,v} (\R) \| \asymp \bigg( \sum_{j=M}^N 2^{j\big(s + d(\frac 1 2 - \frac 1 p)\big)v} |\alpha_j|^v \bigg)^{1/v}
= \bigg( \sum_{j=M}^N |\gamma_j|^v \bigg)^{1/v} \, .
\]
Defining
\[
g_{M,N}(x) := \sum_{j =M}^N \, \, \psi (x-2^{-j}\mu_j), \qquad x \in \R\, ,
\]
we conclude
\[
\| \, g_{M,N} \, | B^s_{p,q} (\R) \| \asymp \, (N-M)^{1/p} \, .
\]
The lower bound is trivial (it even holds for $\| \, g_{M,N} \, | L_{p} (\R) \|$). For the proof of the upper bound
one uses the information on the supports of the functions $\psi (\, \cdot \, -2^{-j}\mu_j)$ and Proposition \ref{diff}.
Observe, all hidden constants are independent of $M,N$ and $\alpha$.
By construction we have the identity $f_{M,N,\alpha} \cdot g_{M,N} = f_{M,N,\alpha}$.
Hence, $B^s_{p,q,v} (\R) \subset M(B^s_{p,q}(\R))$ implies
\\
$B^s_{p,q,v} (\R) \hookrightarrow M(B^s_{p,q}(\R))$ and therefore
\beqq
\| \, f_{M,N,\alpha} \, | B^s_{p,q}(\R) \| & = & \| \, f_{M,N,\alpha} \, \cdot \, g_{M,N} \, | B^s_{p,q}(\R) \|
\\
&\le & \, c \, \| \,f_{M,N,\alpha} \, | B^s_{p,q,v} (\R)\|
\, \, \| \, g_{M,N} \, | B^s_{p,q}(\R)\|
\eeqq
for some $c>0$. Consequently
\[
\bigg( \sum_{j=M}^N \, |\gamma_j|^q \bigg)^{1/q} \le c' \,
\bigg( \sum_{j=M}^N \, |\gamma_j|^v \bigg)^{1/v} \,
\big(N-M\big)^{1/p}
\]
holds for some $c'$ independent of $N$, $M$ and $\alpha$.
Choosing $\gamma_j =1 $ for all $j$ the necessity of $\frac 1 v \ge \frac 1 q - \frac 1 p $ follows.
\end{proof}
\begin{remark} \rm
Proposition \ref{loc} in case $1\leq q<p \leq \infty$ can be found in \cite{SS}.
\end{remark}
Proposition \ref{loc} has an interesting consequence. From Lemma \ref{mult} we know that always
$M(B^s_{p,q}(\R)) \hookrightarrow B^s_{p,q}(\R)_\unif$ holds.
Now we ask for coincidence.
\begin{corollary}\label{abc}
Let $1\le p,q \le \infty$ and $s\in \re$.
Then
$M(B^s_{p,q}(\R))$ coincides with $B^s_{p,q} (\R)_\unif$ if and only if either $1\leq p\leq q\leq \infty$ and $s>d/p$ or $p=q=1$ and $s=d$.
\end{corollary}
\begin{proof}
{\em Step 1.} Necessity.
From $M(B^s_{p,q}(\R))= B^s_{p,q} (\R)_\unif$ we conclude $B^s_{p,q}(\R) \hookrightarrow M(B^s_{p,q} (\R))$. Now Theorem \ref{algebra}
yields either $s>d/p$ or $0< p<\infty$, $0 < q \le 1$ and $s=d/p$.
Next we employ Proposition \ref{loc}. Hence, if $s>d/p$ then $B^{s}_{p,q} (\R)_\unif \not\subset M(B^s_{p,q}(\R))$ if $0 < q < p \le \infty$.
Now we turn to $s=d/p$. Applying Proposition \ref{loc} once again, we find
$B^{d/p}_{p,1} (\R)_\unif \not\subset M(B^{d/p}_{p,1}(\R))$ if $1 < p \le \infty$.
\\
{\em Step 2.} Sufficiency. This follows immediately from Corollary \ref{klar} and Theorem \ref{b-main}.
\end{proof}
\noindent
{\bf Proof of Remark \ref{abcd}}.
Theorem \ref{algebra} yields
\[
\|f \, \cdot \, g\, |B^s_{p,q} (\Omega) \| \le c \, \|\, f \, |B^s_{p,q} (\Omega) \|\, \| \, g\, |B^s_{p,q} (\Omega) \|
\]
whenever this inequality is true on $\R$.
Hence, $B^s_{p,q} (\Omega) \hookrightarrow M(B^s_{p,q} (\Omega))$ under the restrictions of Theorem \ref{algebra}.
On the other hand, the function $f\equiv 1$ belongs to all spaces $B^s_{p,q} (\Omega)$, since $\Omega$ is bounded.
Hence, $f \in B^s_{p,q} (\Omega)$ is also a necessary condition for $f$ to belong to $M(B^s_{p,q} (\Omega))$.
\hspace*{\fill} \rule{3mm}{3mm}
|
\section{Introduction}
\subsection{The Riemann-Siegel $\vartheta (t)$ Function}
The Riemann-Siegel $\vartheta$ function is defined by
\begin{equation}
\begin{array}{cl}
\vartheta (t) & = - \frac{i}{2} \left( \ln \Gamma \left( \frac{1}{4} +
\frac{i t}{2} \right) - \ln \Gamma \left( \frac{1}{4} - \frac{i t}{2}
\right) \right) - \frac{\ln (\pi) t}{2} \text{}\\
& = \arg \left( \Gamma \left( \frac{1}{4} + \frac{i t}{2} \right) \right)
- \frac{\ln (\pi) t}{2}
\end{array}
\end{equation}
Let
\[ \tilde{\vartheta} (t) = \frac{t \ln \left( \frac{t}{2 \pi e} \right)}{2} -
\frac{\pi}{8} \]
be the approximate $\vartheta$ function where the $\Gamma$ function has been
replaced with its Stirling approximation.
\begin{equation}
\Gamma (s) \simeq \sqrt{2 \pi} s^{s - \frac{1}{2}} e^{- s}
\end{equation}
The $\vartheta (t)$ function is not invertible but the inverse of its
approximation $\tilde{\vartheta} (t)$ is defined by a linear function of the
Lambert W function given by
\begin{equation}
\tilde{\vartheta}^{- 1} (t) = \frac{\pi + 8 t}{4 W \left( \frac{\pi + 8 t}{8
\pi e} \right)}
\end{equation}
Let $\tmop{frac} (x) = \left\{ \begin{array}{ll}
x - \lfloor x \rfloor & x \geqslant 0\\
x - \lceil x \rceil & x < 0
\end{array} \right. \forall x \in \mathbbm{R}$ be the function which gives the
fractional part of a real number by subtracting either the floor $\lfloor x
\rfloor$ or the ceiling $\lceil x \rceil$ of $x$ from $x$, depending upon its
sign. Furthermore, let
\begin{equation}
S (t) = \arg \left( \zeta \left( \frac{1}{2} + \tmop{it} \right) \right) =
\lim_{\varepsilon \rightarrow 0} \frac{1}{2} ((S (\rho + i
\varepsilon) - S (\rho - i \varepsilon))
\end{equation}
be the argument of $\zeta$ on the critical line.
\begin{conjecture}
\label{ec}The \tmverbatim{exact equation} for the imaginary part of the
$n$-th zero of $\zeta \left( \frac{1}{2} + i t \right)${\cite[Equation
20]{z0t}}
\begin{equation}
\vartheta (t_n) + S (t_n) = \left( n - \frac{3}{2} \right) \pi \label{ee}
\end{equation}
has a solution for each integer $n \geqslant 1$ where $t_n$ enumerate the
zeros of $Z$ on the real line and the zeros of $\zeta$ on the critical line
\begin{equation}
\zeta \left( \frac{1}{2} + i t_n \right) = 0 \forall n \in \mathbbm{Z}^+
\end{equation}
where $\mathbbm{Z}^+$ denotes the positive integers. {\cite[Equation
14]{z0t}}
\end{conjecture}
\subsection{The Gram Points}
The $n$-th Gram point is defined to be the solution of the equation
\begin{equation}
\vartheta (t) = (n - 1) \pi
\end{equation}
A very accurate approximation $\tilde{g} (n)$ to the Gram points is $g (n)$ is
found by inverting $\tilde{\vartheta} (t)$ to get the exact solution
\begin{equation}
\begin{array}{cl}
\tilde{g} (n) & = \{ t : \tilde{\vartheta} (t) - (n - 1) \pi = 0 \}\\
& = \left\{ t : \left( \frac{t \ln \left( \frac{t}{2 \pi e} \right)}{2} -
\frac{\pi}{8} \right) - (n - 1) \pi = 0 \right\}\\
& = \frac{(8 n - 7) \pi}{4 W \left( \frac{8 n - 7}{8 e} \right)}\\
& = g (n) + O (\delta_n)
\end{array}
\end{equation}
where $W$ is the Lambert W function, and the approximation bounds $\delta_n$
when $n = 1$ is $\delta_1 = 0.00223698 \ldots$, followed by $\delta_2 =
0.00137812 \ldots$ and decreases monotincally with increasing $n$, that is,
$\delta_{n + 1} < \delta_n$. \ The inverse of $\tilde{g} (n)$ is given by
\begin{equation}
\begin{array}{ll}
\tilde{g}^{- 1} (n) & = \{ t : \tilde{g} (n) = 0 \}\\
& = \frac{t \ln \left( \frac{t}{2 \pi e} \right)}{2 \pi} + \frac{7}{8}
\end{array}
\end{equation}
Now define the infinite sequence of functions indexed by $n \in \mathbbm{Z}^+$
\begin{equation}
\begin{array}{cl}
T_n (t) & = 1 + \lfloor \tilde{g}^{- 1} (n) \rfloor - n\\
& = 1 + \lfloor \frac{t \ln \left( \frac{t}{2 \pi e} \right)}{2 \pi} +
\frac{7}{8} \rfloor - n
\end{array}
\end{equation}
Near each ``bad'' Gram point there will be a corresponding zero on the
critical line which has an argument not on the principal branch. The function
$T_n (t)$ determines how many multiples of $\pi$ to add or subtract to $-
\frac{1}{2} - \lfloor \frac{\vartheta (s)}{\pi} \rfloor$ so that it agrees
with the argument of $\zeta$ at a zero on the critical line where it is
discontinuous, having the value $\lim_{\varepsilon \rightarrow 0} \frac{1}{2}
((S (\rho + i \varepsilon) - S (\rho - i \varepsilon))$ when $\zeta
(\rho) = 0$.
\begin{definition}
Let
\begin{equation}
\begin{array}{ll}
S_n (t_{}) & = \pi \left( \frac{1}{2} - \tmop{frac} \left(
\frac{\vartheta (t)}{\pi} \right) - T_n (t) \right)\\
& = \pi \left( \frac{1}{2} - \tmop{frac} \left( \frac{\vartheta
(t)}{\pi} \right) - (\lfloor \tilde{g}^{- 1} (n) \rfloor - n + 1)
\right)\\
& =_{} \pi \left( \frac{3}{2} - \tmop{frac} \left( \frac{\vartheta
(t)}{\pi} \right) - (\lfloor \tilde{g}^{- 1} (n) \rfloor - n) \right)\\
& =_{} \pi \left( \frac{3}{2} - \tmop{frac} \left( \frac{\vartheta
(t)}{\pi} \right) + \left( \lfloor \frac{t \ln \left( \frac{t}{2 \pi e}
\right)}{2 \pi} + \frac{7}{8} \rfloor - n \right) \right)
\end{array}
\end{equation}
\end{definition}
Let $s_{\vartheta} (t) = \frac{\vartheta (t)}{| \vartheta (t) |}$ be the sign
of $\vartheta (t)$.
\begin{conjecture}
\label{c}The argument $S (t)$ of $\zeta \left( \frac{1}{2} + i t \right)$ at
the n-th non-trivial zero \ $\zeta \left( \frac{1}{2} + i t_n \right) = 0
\forall n \geqslant 1$ on the critical strip is equal to $s_{\vartheta} (t)
S_n (t)$, that is
\begin{equation}
\begin{array}{cl}
S (t_n) & = s_{\vartheta} (t) S_n (t_n) = \frac{1}{2} (\lim_{t
\rightarrow t^-_n} S (t_n) + \lim_{t \rightarrow t^+_n} S (t_n))\\
& = \frac{\vartheta (t_n)}{| \vartheta (t_n) |}_{} \pi \left(
\frac{3}{2} - \tmop{frac} \left( \frac{\vartheta (t_n)}{\pi} \right) +
\left( \lfloor \frac{t \ln \left( \frac{t_n}{2 \pi e} \right)}{2 \pi} +
\frac{7}{8} \rfloor - n \right) \right)
\end{array} \label{een}
\end{equation}
\end{conjecture}
\begin{theorem}
If Conjecture \ref{c} is true then Conjecture \ref{ec} is true and, due to
$\tmop{Lemma}$ \ref{fl}, so is Conjecture \ref{RH}, the Riemann hypothesis.
\end{theorem}
\begin{proof}
If $s_{\vartheta} (t) S_n (t_n)$=$S (t_n)$ then $S (t_n)$ is well-defined
$\forall n \geqslant 1$ since $s_{\vartheta} (t) S_n (t_n)$ is well-defined
$\forall n \geqslant 1$.
\end{proof}
\begin{figure}[h]
\resizebox{5.5in}{2.5in}{\includegraphics{liftedTwisted.eps}}
\caption{$S_n (t)$ and the imaginary parts of the roots of $\zeta \left(
\frac{1}{2} + i t_n \right)$ marked with verticle lines just touching its
touching its corresponding pair of neighboring curves $S_n (t)$ for $n = 0
\ldots 14$}
\end{figure}
\begin{figure}[h]
\resizebox{5.5in}{2.5in}{\includegraphics{liftedTwistedAroundFirstBadGramPoints.eps}}
\caption{$S_n (t)$ and the imaginary parts of the roots of $\zeta \left(
\frac{1}{2} + i t_n \right)$ marked with verticle lines just touching its
touching its corresponding pair of neighboring curves $S_n (t)$ for $n = 120
\ldots 140$ which includes two ``bad'' Gram points at $n = 126$ and $n =
134$}
\end{figure}
\begin{theorem}
$\arg \left( \zeta \left( \frac{1}{2} + i g (n) \right) \right) = 0 \forall
n \in \mathbbm{Z}^+$
\end{theorem}
\begin{proof}
The argument of any positive number $x$ with $\tmop{Im} (x) = 0$ is equal to
$0$ and $\tmop{Im} \left( \zeta \left( \frac{1}{2} + i g (n) \right) \right)
= 0$
\end{proof}
\begin{conjecture}
$S (t) - f_0 (t) \in \{ - 1, 0, 1 \} \forall t \in \mathbbm{R}$. That is
\begin{equation}
\tmop{frac} \left( \frac{\vartheta (t)}{\pi} \right) + \frac{1}{\pi} \arg
\left( \zeta \left( \frac{1}{2} + \tmop{it} \right) \right) \in \{ - 1, 0,
1 \} \forall 0 < t \in \mathbbm{R}
\end{equation}
\end{conjecture}
\begin{figure}[h]
\resizebox{5.5in}{3.5in}{\includegraphics{meet1.eps}}
\caption{Illustration of convergence around the first zero}
\end{figure}
\section{Appendix}
\subsection{Transcendental Equations Satisifed By The Nontrivial Riemann
Zeros}
\begin{definition}
The \tmverbatim{critical line} is the line in the complex plane defined by
$\tmop{Re} (t) = \frac{1}{2}$.
\end{definition}
\begin{definition}
The \tmverbatim{critical strip} is the strip in the complex plane defined by
$0 < \tmop{Re} (t) < 1$.
\end{definition}
\begin{definition}
The \tmverbatim{asymptotic equation} for the $n$-th zero of the Hardy $Z$
function
\begin{equation}
\frac{t_n}{2 \pi} \ln \left( \frac{t_n}{2 \pi t} \right) + S (t_n) = n -
\frac{11}{8} \label{ae}
\end{equation}
{\cite[Equation 20]{z0t}}
\end{definition}
\begin{theorem}
\label{le}If the limit
\begin{equation}
\lim_{\delta \rightarrow 0^+} \arg \left( \zeta \left( \frac{1}{2} +
\delta + i t \right) \right)
\end{equation}
is exists and is well-defined $\forall t$ then the left-hand side of
Equation (\ref{ae}) is well-defined $\forall t$, and due to monotonicity,
there must be a unique solution for every $n \in \mathbbm{Z}^+$.
{\cite[II.A]{z0t}}
\end{theorem}
\begin{corollary}
The number of solutions of Equation (\ref{ae}) over the interval $[0, t]$ is
given by
\begin{equation}
N_0 (t) = \frac{t}{2 \pi} \ln \left( \frac{t}{2 \pi e} \right) +
\frac{7}{8} + S (t) + O (t^{- 1}) \label{N0}
\end{equation}
\end{corollary}
which counts the number of zeros \tmverbatim{on the critical line}.
\begin{conjecture}
\label{RH}(The Riemann hypothesis) All solutions $t$ of the equation
\begin{equation}
\zeta (t) = 0
\end{equation}
besides the trivial solutions $t = - 2 n$ with $n \in \mathbbm{Z}^+$ have
real-part $\frac{1}{2}$, that is, $\tmop{Re} (t) = \frac{1}{2}$ when $\zeta
(t) = 0$ and $t \neq - 2 n$.
\end{conjecture}
\begin{definition}
The Riemann-von-Mangoldt formula makes use of Cauchy's argument principle to
count the number of zeros \tmverbatim{inside the critical strip} $0 <
\tmop{Im} (\rho_n) < t$ where $\zeta (\sigma + i \rho_n)$ with $0 < \sigma <
1$
\begin{equation}
N (t) = \frac{t}{2 \pi} \ln \left( \frac{t}{2 \pi e} \right) + \frac{7}{8}
+ S (t) + O (t^{- 1})
\end{equation}
and this definition does not depend on the Riemann hypothesis(Conjecture
\ref{RH}). This equation has exactly the same form as the asymptotic
Equation \ref{ae}. {\cite[Equation 15]{z0t}}
\end{definition}
\begin{lemma}
\label{fl}If the exact Equation (\ref{ee}) has a unique solution for each $n
\in \mathbbm{Z}^+$ then Conjecture \ref{RH}, the Riemann hypothesis,
follows.
\end{lemma}
\begin{proof}
If the exact equation has a unique solution for each $n$, then the zeros
obtained from its solutions on the \tmverbatim{critical line} can be counted
since they are enumerated by the integer $n$, leading to the counting
function $N_0 (t)$ in Equation (\ref{N0}). The number of solutions obtained
on the \tmverbatim{critical line} would saturate counting function of the
number of solutions on the \tmverbatim{critical strip} so that $N (t) = N_0
(t)$ and thus all of the non-trivial zeros of $\zeta$ would be enumerated in
this manner. If there are zeros off of the critical line, or zeros with
multiplicity $m \geqslant 2$, then the exact Equation (\ref{ee}) would fail
to capture all the zeros on the critical strip which would mean $N_0 (t) < N
(t)$. \ {\cite[IX]{z0t}}
\end{proof}
|
\section{Introduction}
\label{sec.intro}
Density ratio estimation (DRE) \cite{Nguyen2010, Huang2007,Sugiyama2012} is an important tool
in various branches of machine learning and statistics.
Due to its ability of directly modelling the differences between two
probability density functions, DRE finds its applications in change
detection \cite{Kawahara2012,Fazayeli2016}, two-sample test
\cite{Wornowizki2016} and outlier detection \cite{azmandian2012local, Smola2009}. In
recent years, a sampling framework called Generative Adversarial Network
(GAN) (see e.g., \cite{Goodfellow2014,Nowozin2016}) uses the density
ratio function to compare artificial samples from a generative distribution and real samples from
an unknown distribution. DRE has also been widely discussed in statistical literatures for adjusting non-parametric density estimation \cite{Efron1996}, stabilizing the estimation of heavy tailed distribution \cite{Fithian2015} and fitting multiple distributions at once \cite{Fokianos2004}.
However, as a density ratio function can grow unbounded,
DRE can suffer from robustness and stability issues:
a few corrupted points may completely mislead the
estimator (see Figure \ref{fig.mnchange} in Section \ref{sec.label} for example). Considering a density ratio $p(x)/q(x)$, a point $x$ that is extremely far away from the high density region of $q$ may have an almost infinite ratio value and DRE results can be \emph{dominated} by such points.
This makes DRE performance very sensitive to rare pathological data or small modifications of the dataset. Here we give two examples:
\paragraph{Cyber-attack}
In change detection applications, a density ratio $p(x)/q(x)$ is used to determine how the data generating model differs between $p$ and $q$.
Consider a ``hacker''
who can spy on our data may just inject a few data points in $p$
which are extremely far away from the high-density region of $q$. This would result excessively large $p(x)/q(x)$ tricking us to believe there is a
significant change from $q(x)$ to $p(x)$, even if there is no change at all. If the generated outliers are also far away from the high density region of $p(x)$, we end up with a very different density ratio function and the original parametric pattern in the ratio is ruined. We give such an example in Section \ref{sec.label}.
\paragraph{Volatile Samples}
The change of external environment may be responded in unpredictable ways.
It is possible that a small portion of samples react more ``aggressively'' to the change than the others. These samples may be skewed and show very high density ratios, even if the change of distribution is relatively mild when these volatile samples are excluded.
For example, when testing a new fertilizer, a small number of plants may fail to adapt, even if the vast majority of crops are healthy.
Overly large density ratio values can cause further troubles when the ratio is used to weight samples. For example, in the domain adaptation setting, we may reweight samples from one task and reuse them in another task. Density ratio is a natural choice of such ``importance weighting'' scheme \cite{Sugiyama2008,Shimodaira2000}.
However, if one or a few samples have extremely high ratio, after renormalizing, other samples will have almost zero weights and have little impact to the learning task.
Several methods have been proposed to solve this problem. The relative
density ratio estimation \cite{Yamada2013} estimates a ``biased'' version of density ratio
controlled by a mixture parameter $\alpha.$ The relative density
ratio is always upper-bounded by $\frac{1}{\alpha}$,
which can give a more robust estimator.
However,
it is not clear how to de-bias such an estimator
to recover the true density ratio function.
\cite{Smola2009} took a more direct approach. It estimates a \emph{thresholded}
density ratio by setting up a tolerance $t$ to the density
ratio value. All likelihood ratio values bigger than $t$
will be clipped to $t$. The estimator was derived from Fenchel duality
for $f$-divergence \cite{Nguyen2010}. However,
the optimization for the estimator is not convex
if one uses log-linear models.
The formulation also relies on the non-parametric approximation of the
density ratio function (or the log ratio function)
making the learned model
hard to interpret. Moreover, there is no intuitive way to directly control the proportion of ratios that are thresholded.
Nonetheless, the concept studied in our paper
is inspired by this pioneering work.
In this paper, we propose a novel method based on a ``trimmed Maximum
Likelihood Estimator'' \cite{Neykov1990, Hadi1997}. This idea relies on a specific
type of density ratio estimator (called log-linear KLIEP) \cite{Tsuboi2009}
which can be written as a maximum likelihood formulation. We simply
``ignore'' samples that make the empirical likelihood take exceedingly
large values. The trimmed density ratio estimator can be formulated as a convex optimization
and translated into a weighted M-estimator. This helps us develop
a simple subgradient-based algorithm that is guaranteed to reach the
global optimum.
Moreover, we shall prove that in addition to recovering the correct density
ratio under the outlier setting, the estimator can also obtain a ``corrected''
density ratio function under a truncation setting. It ignores
``pathological'' samples and recovers density ratio only using ``healthy'' samples.
Although trimming will usually result a more robust estimate of the density ratio function, we also point out that it should not be abused. For example, in the tasks of two-sample test, a diverging density ratio might indicate interesting structural differences between two distributions.
In Section \ref{sec:Preliminary:-Trimmed-Maximum},
we explain some preliminaries on trimmed maximum likelihood estimator.
In Section \ref{sec:Trimmed-Density-Ratio}, we introduce a trimmed
DRE.
We solve it using a convex formulation whose optimization
procedure is explained in Section \ref{sec:Optimization}. In Section
\ref{sec:-Consistency-in}, we prove the estimation error
upper-bound with respect to a sparsity inducing regularizer. Finally, experimental results are shown in Section \ref{sec.label} and we conclude our work in Section \ref{sec.concl}.
\section{Preliminary: Trimmed Maximum Likelihood Estimation\label{sec:Preliminary:-Trimmed-Maximum}}
Although our main purpose is to estimate the density ratio, we first
introduce the basic concept of \emph{trimmed estimator} using density
functions as examples. Given $n$ samples drawn from a distribution
$P$, i.e., $X:=\left\{ \boldsymbol{x}^{(i)}\right\} _{i=1}^{n}\stackrel{\mathrm{i.i.d.}}{\sim} P,\boldsymbol{x}\in\mathbb{R}^{d}$,
we want to estimate the density function $p(\boldsymbol{x})$. Suppose
the true density function is a member of \emph{exponential family}
\cite{Pitman1936},
\begin{align}
p(\boldsymbol{x};\boldsymbol{\theta})=\exp\left[\langle\boldsymbol{\theta},\boldsymbol{f}(\boldsymbol{x})\rangle-\log Z(\boldsymbol{\theta})\right],~~Z(\boldsymbol{\theta})=\int q(\boldsymbol{x})\exp\langle\boldsymbol{\theta},\boldsymbol{f}(\boldsymbol{x})\rangle d\boldsymbol{x}\label{eq:expfamily}
\end{align}
where $\boldsymbol{f}(\boldsymbol{x})$ is the sufficient statistics,
$Z(\boldsymbol{\theta})$ is the normalization function and $q(\boldsymbol{x})$
is the base measure.
Maximum Likelihood Estimator (MLE) maximizes the
empirical likelihood over the entire dataset. In contrast, a \emph{trimmed
} MLE only maximizes the likelihood over a \emph{subset} of samples
according to their likelihood values (see e.g., \cite{Hadi1997,Vandev1998}).
This paradigm can be used to derive a popular outlier detection method, one-class Support Vector Machine (one-SVM) \cite{Schoelkopf2000}.
The derivation is crucial to the development of our trimmed density ratio estimator in later sections.
Without loss of generality, we can
set the log likelihood function as $\log p(\boldsymbol{x}^{(i)};\boldsymbol{\theta})-\tau_{0},$
where $\tau_{0}$ is a constant.
As samples corresponding to high likelihood values are likely to be inliers, we can trim all samples whose likelihood is bigger than $\tau_{0}$
using a clipping function $[\cdot]_{-}$, i.e.,
$
\hat{{\boldsymbol{\theta}}}=\arg\max_{{\boldsymbol{\theta}}}\sum_{i=1}^{n}[\log p(\boldsymbol{x}^{(i)};\boldsymbol{\theta})-\tau_{0}]_{-},\label{eq:trimMLE}
$
where $[\ell]_{-}$ returns $\ell$ if $\ell\le0$ and $0$ otherwise.\emph{
}This optimization has a \emph{convex }formulation:
\begin{align}
\min_{\boldsymbol{\theta},\boldsymbol{\epsilon}\ge0}\langle\boldsymbol{\epsilon},\boldsymbol{1}\rangle,~~\text{ s.t. }\forall i,\log p\left(\boldsymbol{x}^{(i)};\boldsymbol{\theta}\right)\ge\tau_{0}-\epsilon_{i},\label{eq:trimmedMLE}
\end{align}
where $\boldsymbol{\epsilon}$ is the slack variable measuring the
difference between $\log p\left(\boldsymbol{x}^{(i)};\boldsymbol{\theta}\right)$
and $\tau_{0}$. However, formulation \eqref{eq:trimmedMLE} is not
practical since computing the normalization term $Z(\boldsymbol{\theta})$
in \eqref{eq:expfamily} is intractable for a general \textbf{$\boldsymbol{f}$}
and it is unclear how to set the trimming level $\tau_{0}$. Therefore we ignore the normalization term and introduce other control
terms:
\begin{align}
\min_{\boldsymbol{\theta},\boldsymbol{\epsilon}\ge0,\tau\ge0}\frac{1}{2}\|\boldsymbol{\theta}\|^{2}-\nu\tau+\frac{1}{n}\langle\boldsymbol{\epsilon},\boldsymbol{1}\rangle~~\text{ s.t. }\forall i,\langle\boldsymbol{\theta},\boldsymbol{f}(\boldsymbol{x}^{(i)})\rangle\ge\tau-\epsilon_{i}.\label{eq:onesvm}
\end{align}
The $\ell_{2}$ regularization term is introduced to avoid $\boldsymbol{\theta}$
reaching unbounded values.
A new hyper parameter $\nu\in(0,1]$ replaces $\tau_{0}$ to control
the number of trimmed samples. It can be proven using KKT conditions that at most $1-\nu$
fraction of samples are discarded (see e.g., \cite{Schoelkopf2000}, Proposition 1 for details).
Now we have reached the standard formulation
of one-SVM.
This trimmed estimator ignores the large likelihood
values and creates a focus only on the low density region. Such a
trimming strategy allows us to discover ``novel'' points or outliers
which are usually far away from the high density area.
\section{Trimmed Density Ratio Estimation\label{sec:Trimmed-Density-Ratio}}
In this paper, our main focus is to derive a \emph{robust} density
ratio estimator following a similar trimming strategy. First, we briefly
review the a density ratio estimator \cite{Sugiyama2012} from the
perspective of Kullback-Leibler divergence minimization.
\subsection{Density Ratio Estimation (DRE)}
For two sets of data
$
X_{p}:=\{\boldsymbol{x}_{p}^{(1)},\dots,\boldsymbol{x}_{p}^{(n_{p})}\}\stackrel{\mathrm{i.i.d.}}{\sim} P,X_{q}:=\{\boldsymbol{x}_{q}^{(1)},\dots,\boldsymbol{x}_{q}^{(n_{q})}\}\stackrel{\mathrm{i.i.d.}}{\sim} Q,
$
assume both the densities $p(\boldsymbol{x})$ and $q(\boldsymbol{x})$
are in exponential family \eqref{eq:expfamily}. We know
$
\frac{p(\boldsymbol{x};\boldsymbol{\theta}_{p})}{q(\boldsymbol{x};\boldsymbol{\theta}_{q})}\propto\exp\left[\langle\boldsymbol{\theta}_{p}-\boldsymbol{\theta}_{q},\boldsymbol{f}(\boldsymbol{x})\rangle\right].
$
Observing that the data $\boldsymbol{x}$ only interacts with the
parameter $\boldsymbol{\theta}_{p}-\boldsymbol{\theta}_{q}$ through
$\boldsymbol{f}$ , we can keep using $\boldsymbol{f}(\boldsymbol{x})$
as our sufficient statistic for the density ratio model, and merge
two parameters $\boldsymbol{\theta}_{p}$ and $\boldsymbol{\theta}_{q}$
into one single parameter $\boldsymbol{\delta}=\boldsymbol{\theta}_{p}-\boldsymbol{\theta}_{q}$.
Now we can model our density ratio as
\begin{align}
r(\boldsymbol{x};\boldsymbol{\delta}):=\exp\left[\langle\boldsymbol{\delta},\boldsymbol{f}(\boldsymbol{x})\rangle-\log N(\boldsymbol{\delta})\right],~N(\boldsymbol{\delta}):=\int q(\boldsymbol{x})\exp\langle\boldsymbol{\delta},\boldsymbol{f}(\boldsymbol{x})\rangle d\boldsymbol{x},\label{eq:ratio_model}
\end{align}
where $N(\boldsymbol{\delta})$ is the normalization term that guarantees
$\int q(\boldsymbol{x})r(\boldsymbol{x};\boldsymbol{\delta})d\boldsymbol{x}=1$
so that $q(\boldsymbol{x})r(\boldsymbol{x};\boldsymbol{\delta})$
is a valid density function and is normalized over its domain.
Interestingly, despite the parameterization (changing from $\boldsymbol{\theta}$
to $\boldsymbol{\delta}$), \eqref{eq:ratio_model} is exactly the
same as \eqref{eq:expfamily} where $q(\boldsymbol{x})$ appeared
as a base measure. The difference is, here, $q(\boldsymbol{x})$ is
a \emph{density function} from which $X_{q}$ are drawn so that $N(\boldsymbol{\delta})$
can be approximated accurately from samples of $Q$. Let us define
\begin{align}
\label{eq:ratiomodel2}
\hat{r}(\boldsymbol{x};\boldsymbol{\delta}):=\exp\left[\langle\boldsymbol{\delta},\boldsymbol{f}(\boldsymbol{x})\rangle-\log\widehat{N}(\boldsymbol{\delta})\right],~\widehat{N}(\boldsymbol{\delta}):=\frac{1}{n_{q}}\sum_{j=1}^{n_{q}}\exp\left[\langle\boldsymbol{\delta},\boldsymbol{f}(\boldsymbol{x}_{q}^{(j)})\rangle\right].
\end{align}
Note this model can be computed for any $\boldsymbol{f}$ even if
the integral in $N(\boldsymbol{\delta})$ does not have a closed form
.
In order to estimate $\boldsymbol{\delta}$, we minimize the Kullback-Leibler
divergence between $p$ and $q\cdot r_{\boldsymbol{\delta}}$:
\begin{align}
\min_{\boldsymbol{\delta}}\mathrm{KL}\left[p|q\cdot r_{\boldsymbol{\delta}}\right] & =\min_{\boldsymbol{\delta}}\int p(\boldsymbol{x})\log\frac{p(\boldsymbol{x})}{q(\boldsymbol{x})r(\boldsymbol{x};\boldsymbol{\delta})}d\boldsymbol{x}=c-\max_{\boldsymbol{\delta}}\int p(\boldsymbol{x})\log r(\boldsymbol{x};\boldsymbol{\delta})d\boldsymbol{x}\nonumber \\
& \approx c-\max_{\boldsymbol{\delta}}\frac{1}{n_{p}}\sum_{i=1}^{n_{p}}\log\hat{r}(\boldsymbol{x}_{p}^{(i)};\boldsymbol{\delta})\label{eq:MLEOBJ}
\end{align}
where $c$ is a constant irrelevant to $\boldsymbol{\delta}$. It
can be seen that the minimization of KL divergence boils down to \emph{maximizing
log likelihood} \emph{ratio} over dataset $X_{p}$.
Now we have reached the log-linear Kullback-Leibler Importance Estimation
Procedure (log-linear KLIEP) estimator \cite{Tsuboi2009,Liu2016a}.
\subsection{Trimmed Maximum Likelihood Ratio}
As stated in Section \ref{sec.intro}, to rule out the influences of large density ratio, we trim samples with large likelihood
ratio values from \eqref{eq:MLEOBJ}. Similarly to one-SVM in \eqref{eq:trimmedMLE},
we can consider a trimmed MLE
$
\hat{\boldsymbol{\delta}}=\arg\max_{{\boldsymbol{\delta}}}\sum_{i=1}^{n_{p}}[\log\hat{r}(\boldsymbol{x}_{p}^{(i)};\boldsymbol{\delta})-t_{0}]_{-}\label{eq:hingeMLE}
$
where $t_{0}$ is a threshold above which the likelihood ratios are
ignored. It has a convex formulation:
\begin{align}
\min_{\boldsymbol{\delta},\boldsymbol{\epsilon}\ge\boldsymbol{0}}\langle\epsilon,\boldsymbol{1}\rangle,~~\text{s.t. }\forall\boldsymbol{x}_{p}^{(i)}\in X_{p},\log\hat{r}(\boldsymbol{x}_{p}^{(i)};\boldsymbol{\delta})\ge t_{0}-\epsilon_{i}.\label{eq:objsimple}
\end{align}
\eqref{eq:objsimple} is similar to \eqref{eq:trimmedMLE} since we
have only replaced $p(\boldsymbol{x};\boldsymbol{\theta})$ with $\hat{r}(\boldsymbol{x};\boldsymbol{\delta})$.
However, the ratio model $\hat{r}(\boldsymbol{x};\boldsymbol{\delta})$
in \eqref{eq:objsimple} comes with a tractable normalization term
$\hat{N}$ while the normalization term $Z$ in $p(\boldsymbol{x};\boldsymbol{\theta})$
is in general intractable.
Similar to \eqref{eq:onesvm}, we can directly control the trimming
quantile via a hyper-parameter $\nu$:
\begin{align}
\min_{\boldsymbol{\delta},\boldsymbol{\epsilon}\ge\boldsymbol{0},t\ge0}\frac{1}{n_{p}}\langle\epsilon,\boldsymbol{1}\rangle-\nu\cdot t+\lambda R(\boldsymbol{\delta}),~~\text{s.t. }\forall\boldsymbol{x}_{p}^{(i)}\text{\ensuremath{\in}}X_{p},\log\hat{r}(\boldsymbol{x}_{p}^{(i)};\boldsymbol{\delta})\ge t-\epsilon_{i}\label{eq:objquantile}
\end{align}
where $R(\boldsymbol{\delta})$ is a convex regularizer. \eqref{eq:objquantile}
is\emph{ }also convex, but it has $n_{p}$ number of \emph{non-linear}
constraints and the search for the global optimal solution can be time-consuming.
To avoid such a problem, one could derive and solve the dual problem
of \eqref{eq:objquantile}. In some applications,
we rely on the primal parameter structure (such as sparsity) for model
interpretation, and feature engineering. In Section \ref{sec:Optimization}, we translate \eqref{eq:objquantile} into an
equivalent form so that its solution is obtained via a subgradient
ascent method which is guaranteed to converge to the global optimum.
One common way to construct a convex robust estimator is using a Huber loss \cite{Huber1964}.
Although the proposed trimming technique rises from a different setting, it shares the same guiding principle with Huber loss: avoid assigning dominating values to outlier likelihoods in the objective function.
In Section \ref{sec:poofproposvm} in the supplementary material, we show the relationship between trimmed DRE and binary Support Vector Machines \cite{Scholkopf2001,Cristianini2000}.
\section{Optimization\label{sec:Optimization}}
The key to solving \eqref{eq:objquantile} efficiently is reformulating
it into an equivalent $\max\min$ problem.
\begin{prop}
\label{prop:minma}Assuming $\nu$ is chosen such that $\hat{t}>0$
for all optimal solutions in \eqref{eq:objquantile}, then $\hat{\boldsymbol{\delta}}$
is an optimal solution of \eqref{eq:objquantile} if and only if it
is also the optimal solution of the following $\max \min$ problem:
\begin{align}
\max_{\boldsymbol{\delta}}\min_{\boldsymbol{w}\in\left[0,\frac{1}{n_{p}}\right]^{n_{p}},\langle\boldsymbol{1},\boldsymbol{w}\rangle=\nu}\mathcal{L}(\boldsymbol{\delta},\boldsymbol{w})-\lambda R(\boldsymbol{\delta}),\; & \mathcal{L}(\boldsymbol{\delta},\boldsymbol{w}):=\sum_{i=1}^{n_{p}}w_{i}\cdot\log\hat{r}(\boldsymbol{x}_{p}^{(i)};\boldsymbol{\delta}).\label{eq:min_max}
\end{align}
\end{prop}
The proof is in Section \ref{subsec:minma} in the supplementary material. We define $(\hat{\boldsymbol{\delta}},\hat{\boldsymbol{w}})$ as a
saddle point of \eqref{eq:min_max}:
\begin{align}
\label{eq:saddlepoint}
\nabla_{\boldsymbol{\delta}}\mathcal{L}(\hat{\boldsymbol{\delta}},\hat{\boldsymbol{w}})-\nabla_{\boldsymbol{\delta}}\lambda R(\hat{\boldsymbol{\delta}})={\boldsymbol{0}},\hat{\boldsymbol{w}}\in\arg\min_{\boldsymbol{w}\in[0,\frac{1}{n_p}]^{n_{p}},\langle\boldsymbol{w},\boldsymbol{1}\rangle=\nu}\mathcal{L}(\hat{\boldsymbol{\delta}},\boldsymbol{w}),
\end{align}
where the second $\nabla_{\boldsymbol{\delta}}$ means the subgradient
if $R$ is sub-differentiable.
Now the ``trimming'' process of our estimator can be clearly seen
from \eqref{eq:min_max}: The $\max$ procedure estimates a density
ratio given the currently assigned weights $\boldsymbol{w}$, and
the $\min$ procedure trims the large log likelihood ratio values
by assigning corresponding $w_{i}$ to $0$ (or values smaller
than $\frac{1}{n_{p}}$).
For simplicity, we only consider the cases where $\nu$ is a multiple of $\frac{1}{n_p}$.
Intuitively, $1-\nu$ is the proportion of likelihood ratios that are trimmed thus $\nu$ should not be greater than 1.
Note if we set $\nu=1$, \eqref{eq:min_max}
is equivalent to the standard density ratio estimator \eqref{eq:MLEOBJ}.
Downweighting outliers while estimating the model parameter $\boldsymbol{\delta}$
is commonly used by robust estimators (See e.g., \cite{Cleveland1979,Suykens2002}).
The search for $(\hat{\boldsymbol{\delta}},\hat{\boldsymbol{w}})$
is straightforward. It is easy to solve with respect to $\boldsymbol{w}$
or $\boldsymbol{\delta}$ while the other is fixed: given a parameter
$\boldsymbol{\delta}$, the optimization with respect to $\boldsymbol{w}$
is a linear programming and \emph{one of} the extreme optimal solutions is
attained by assigning weight $\frac{1}{n_{p}}$
to the elements that correspond to the $\nu n_{p}$-smallest log-likelihood
ratio $\log\hat{r}(\boldsymbol{x}^{(i)},\boldsymbol{\delta})$. This
observation leads to a simple ``gradient ascent and trimming'' algorithm
(see Algorithm \ref{alg}). In Algorithm \ref{alg},
\[
\nabla_{\boldsymbol{\delta}}\mathcal{L}(\boldsymbol{\delta},\boldsymbol{w})=\frac{1}{n_{p}}\sum_{i=1}^{n_{p}}w_{i}\cdot\boldsymbol{f}(\boldsymbol{x}_{p}^{(i)})-\nu\cdot\sum_{j=1}^{n_{q}}\frac{e^{(j)}}{\sum_{k=1}^{n_{q}}e^{(k)}}\boldsymbol{f}(x_{q}^{(j)}),~~e^{(i)}:=\exp(\langle\boldsymbol{\delta},\boldsymbol{f}(x_{q}^{(i)})\rangle).\label{eq:subgra}
\]
\begin{algorithm}[t]
\begin{algorithmic}
\STATE Input: $X_{p},X_{q},\nu$ and step sizes $\{\eta_{\mathrm{it}}\}_{\mathrm{it=1}}^{\mathrm{it}_\mathrm{max}}$;
Initialize $\boldsymbol{\delta}_{0},\boldsymbol{w}_{0}$, Iteration
counter: $\mathrm{it}=0$, Maximum number of iterations: $\mathrm{it}_{\mathrm{max}}$,
Best objective, parameter pair $(O_{\mathrm{best}}=-\infty,\boldsymbol{\delta}_{\mathrm{\mathrm{best}}},\boldsymbol{w}_{\mathrm{best}})$
.
\WHILE{not converged and $\mathrm{it}\le\mathrm{it}_{\mathrm{max}}$}
\STATE Obtain a sorted set $\left\{ \boldsymbol{x}_{p}^{(i)}\right\} _{i=1}^{n_{p}}$
so that $\log\hat{r}(\boldsymbol{x}_{p}^{(1)};\boldsymbol{\delta}_{\mathrm{it}})\le\log\hat{r}(\boldsymbol{x}_{p}^{(2)};\boldsymbol{\delta}_{\mathrm{it}})\cdots\le\log\hat{r}(\boldsymbol{x}_{p}^{(n_{p})};\boldsymbol{\delta}_{\mathrm{it}}).$
\STATE $w_{\mathrm{it}+1,i}=\frac{1}{n_{p}},\forall i\le\nu n_{p}.~~w_{\mathrm{it+1},i}=0,\text{otherwise.}$
\STATE Gradient ascent with respect to $\boldsymbol{\delta}$: $\boldsymbol{\delta}_{\mathrm{it}+1} =\boldsymbol{\delta}_{\mathrm{it}}+\eta_{\mathrm{it}}\cdot\nabla_{\boldsymbol{\delta}}[\mathcal{L}(\boldsymbol{\delta}_{\mathrm{it}},\boldsymbol{w}_{\mathrm{it}+1})-\lambda R(\boldsymbol{\delta}_{\mathrm{it}})],\:$
\STATE
$
O_{\mathrm{best}} =\max(O_{\mathrm{best}},\mathcal{L}(\boldsymbol{\delta}_{\mathrm{it+1}},\boldsymbol{w}_{\mathrm{it}+1}))
$
and update $({\boldsymbol{\delta}}_\mathrm{best}, {\boldsymbol{w}}_\mathrm{best})$ accordingly.
$~~\mathrm{it}=\mathrm{it}+1.$
\ENDWHILE
Output: $(\boldsymbol{\delta}_{\mathrm{best}},\boldsymbol{w}_{\mathrm{best}})$
\end{algorithmic}
\caption{Gradient Ascent and Trimming \label{alg}}
\end{algorithm}
In fact, Algorithm \ref{alg} is a subgradient method \citep{Boyd2014,Nedic2009},
since the optimal value function of the inner problem of \eqref{eq:min_max} is not
differentiable at some $\boldsymbol{\delta}$ where the inner problem
has multiple optimal solutions. The subdifferential of the optimal value of the inner problem with respect to ${\boldsymbol{\delta}}$ can be a \emph{set} but Algorithm \ref{alg} only computes a subgradient
obtained using the extreme point solution $\boldsymbol{w}_{\mathrm{it+1}}$
of the inner linear programming. Under mild conditions, this subgradient ascent approach will converge to
optimal results with diminishing step size rule and $\mathrm{it}\rightarrow\infty$.
See \citep{Boyd2014} for details.
Algorithm \ref{alg} is a simple gradient ascent procedure and can be implemented by deep learning softwares such as Tensorflow\footnote{\url{https://www.tensorflow.org/}} which benefits from the GPU acceleration. In contrast, the original problem \eqref{eq:objquantile}, due to its heavily constrained nature, cannot be easily programmed using such a framework.
\section{Estimation Consistency in High-dimensional Settings \label{sec:-Consistency-in}}
\begin{figure}
\centering
\subfloat[Outlier Setting. Blue and red points are i.i.d.]{
\includegraphics[width=.49\textwidth]{illus1}
\label{fig:illus1}
}
\subfloat[Truncation Setting. There are no outliers.]{
\label{fig:illus2}
\includegraphics[width=.49\textwidth]{illus2}
}
\caption{Two settings of theoretical analysis.}
\end{figure}
In this section, we show how the estimated parameter $\hat{\boldsymbol{\delta}}$ in \eqref{eq:saddlepoint} converges to the ``optimal parameters''
$\boldsymbol{\delta}^{*}$
as both sample size and dimensionality goes to infinity
under the ``outlier'' and ``truncation'' setting respectively.
In the \textbf{outlier setting} (Figure \ref{fig:illus1}),
we assume $X_p$ is contaminated by outliers and all ``inlier'' samples in $X_p$ are i.i.d.. The outliers are injected into our dataset $X_p$ after looking at our inliers. For example, hackers can spy on our data and inject fake samples so that our estimator exaggerates the degree of change.
In the \textbf{truncation setting}, there are no outliers. $X_p$ and $X_q$ are i.i.d. samples from $P$ and $Q$ respectively. However, we have a subset of ``volatile'' samples in $X_p$ (the rightmost mode on histogram in Figure \ref{fig:illus2}) that are pathological and exhibit large density ratio values.
In the theoretical results in this section, we focus on analyzing the performance of our estimator for high-dimensional data assuming the number of non-zero elements in the optimal ${\boldsymbol{\delta}}^*$ is $k$ and use the $\ell_1$ regularizer, i.e., $R(\boldsymbol{\theta})=\|\boldsymbol{\theta}\|_{1}$ which induces sparsity on $\hat{{\boldsymbol{\delta}}}$.
The proofs
rely on a recent development \cite{Yang2015,yang2016high}
where a ``weighted'' high-dimensional estimator was studied.
We also assume the optimization of ${\boldsymbol{\delta}}$ in \eqref{eq:min_max} was conducted
within an $\ell_{1}$ ball of width $\rho$, i.e., $\mathrm{Ball}(\rho)$,
and $\rho$ is wisely chosen so that the optimal parameter $\boldsymbol{\delta}^{*}\in\mathrm{Ball}(\rho)$.
The same technique was used in previous works \cite{Loh2015,Yang2015}.
\paragraph{Notations:}
We denote ${\boldsymbol{w}}^* \in \mathbb{R}^{n_p}$ as the ``optimal'' weights depending on ${\boldsymbol{\delta}}^*$ and our data.
To lighten the notation, we shorten the \emph{log} density ratio model as $z_{\boldsymbol{\delta}}(\boldsymbol{x}):=\log r(\boldsymbol{x};\boldsymbol{\delta}), \hat{z}{}_{\boldsymbol{\delta}}(\boldsymbol{x}):=\log\hat{r}(\boldsymbol{x};\boldsymbol{\delta})$
The proof of Theorem \ref{thm:main}, \ref{thm:main-1} and \ref{thm:main-1-1} can be found in Section \ref{subsec:theorem1}, \ref{sec:proofcol1} and \ref{sec:proofoftruncation} in supplementary materials.
\subsection{A Base Theorem\label{subsec:Preparations}}
Now we provide a base theorem giving
an upperbound of $\|\hat{{\boldsymbol{\delta}}}-{\boldsymbol{\delta}}^{*}\|$.
We state this theorem only with
respect to an arbitrary pair $({\boldsymbol{\delta}}^{*},{\boldsymbol{w}}^{*})$ and
the pair is set properly later in Section \ref{sec:outliersetting} and \ref{sec:truncationsetting}.
We make a few regularity conditions on samples from $Q$. Samples of $X_{q}$ should be well behaved in terms of log-likelihood
ratio values.
\begin{assumption}
\label{assu:.boundedrq}$\exists0<c_{1}<1,1<c_{2}<\infty$
$
\forall\boldsymbol{x}_{q}\in X_{q},\boldsymbol{u}\in\mathrm{Ball}(\rho),c_{1}\le\exp\langle\boldsymbol{\delta}^{*}+\boldsymbol{u},\boldsymbol{x}_{q}\rangle\le c_{2}
$
and collectively $c_{2}/c_{1}=C_{r}$.
\end{assumption}
We also assume the Restricted Strong Convexity (RSC) condition on the covariance of ${\boldsymbol{X}}_q$, i.e., $\mathrm{cov}({\boldsymbol{X}}_q) = \frac{1}{n_q} ({\boldsymbol{X}}_q - \frac{1}{n_q}{\boldsymbol{X}}_q \bold1) ({\boldsymbol{X}}_q - \frac{1}{n_q}{\boldsymbol{X}}_q \bold1)^\top$. Note this property has been verified for various different design
matrices ${\boldsymbol{X}}_{q}$, such as Gaussian or sub-Gaussian (See, e.g., \cite{Raskutti2010,Rudelson2013}).
\begin{assumption}
\label{assu:RSC} RSC condition of $\mathrm{cov}({\boldsymbol{X}}_q)$
holds for all $\boldsymbol{u}$, i.e.,
there exists $\kappa'_{1} > 0$ and $c>0$ such that $\boldsymbol{u}^{\top}\mathrm{cov}({\boldsymbol{X}}_q)\boldsymbol{u}\ge\kappa'_{1}\|\boldsymbol{u}\|^{2}-\frac{c}{\sqrt{n_{q}}}\|\boldsymbol{u}\|_{1}^{2}\text{ with high probability. }$
\end{assumption}
\begin{thm}
In addition to Assumption \ref{assu:.boundedrq} and \ref{assu:RSC},
there exists coherence between parameter $\boldsymbol{w}$ and
$\boldsymbol{\delta}$ at a saddle point $(\hat{{\boldsymbol{\delta}}},\hat{{\boldsymbol{w}}})$:
\begin{align}
\langle\nabla_{{\boldsymbol{\delta}}}\mathcal{L}(\hat{{\boldsymbol{\delta}}},\hat{{\boldsymbol{w}}})-\nabla_{\boldsymbol{\delta}}\mathcal{L}(\hat{{\boldsymbol{\delta}}},\boldsymbol{w}^{*}),\hat{\boldsymbol{u}}\rangle\ge-\kappa_{2}\|\hat{\boldsymbol{u}}\|^{2}-\tau_{2}(n,d)\|\hat{\boldsymbol{u}}\|_{1},\label{eq:ass2}
\end{align}
where $\hat{{\boldsymbol{u}}}:=\hat{{\boldsymbol{\delta}}}-{\boldsymbol{\delta}}^{*}$, $\kappa_2 > 0$ is a constant and $\tau_2(d,n) > 0$. It can be
shown that if \[\textstyle{
\lambda_{n}\ge2\max\left[\|\nabla_{\boldsymbol{\delta}}\mathcal{L}(\boldsymbol{\delta}^{*},\boldsymbol{w}^{*})\|_{\infty},\frac{\rho\nu c}{2C_{r}^{2}\sqrt{n_{q}}},\tau_{2}(n,d)\right]}
\]
and $\nu\kappa'_{1}>2C_{r}^{2}\kappa_{2}$, where $c>0$ is a constant determined by RSC condition, we are guaranteed that $\|\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\|\le\frac{C_{r}^{2}}{(\nu\kappa'_{1}-2C_{r}^{2}\kappa_{2})}\cdot\frac{3\sqrt{k}\lambda_{n}}{2}$ with probability converging to one.
\label{thm:main}
\end{thm}
The condition \eqref{eq:ass2} states that if we swap $\hat{{\boldsymbol{w}}}$ for ${\boldsymbol{w}}^*$, the change of the gradient $\nabla_{\boldsymbol{\delta}}\mathcal{L}$ is limited.
Intuitively, it shows that our estimator \eqref{eq:min_max} is not ``picky'' on ${\boldsymbol{w}}$: even if we cannot have the optimal weight assignment ${\boldsymbol{w}}^*$, we can still use ``the next best thing'', $\hat{{\boldsymbol{w}}}$ to compute the gradient which is close enough.
We later show how \eqref{eq:ass2} is satisfied.
Note if $\|\nabla_{\boldsymbol{\delta}}\mathcal{L}(\boldsymbol{\delta}^{*},\boldsymbol{w}^{*})\|_{\infty}$,
$\tau_{2}(n,d)$ converge to zero as $n_{p},n_{q},d\to\infty$,
by taking $\lambda_{n}$ as such, Theorem 1 guarantees the consistency
of $\hat{\boldsymbol{\delta}}$.
In Section \ref{subsec:Consistency-under-Outlier} and \ref{subsec:trimmedsetting},
we explore two different settings of $(\boldsymbol{\delta}^{*},{\boldsymbol{w}}^{*})$
that make $||\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\|$
converges to zero.
\subsection{Consistency under Outlier Setting\label{subsec:Consistency-under-Outlier}}
\label{sec:outliersetting}
\paragraph{Setting:}
Suppose dataset $X_{p}$ is the union of two disjoint sets $G$ (Good points) and
$B$ (Bad points) such that $G\stackrel{\mathrm{i.i.d.}}{\sim} p({\boldsymbol{x}})$
and $\min_{j\in B}z_{\boldsymbol{\delta}^{*}}(\boldsymbol{x}_{p}^{(j)})>\max_{i\in G}z_{\boldsymbol{\delta}^{*}}(\boldsymbol{x}_{p}^{(i)})$ (see Figure \ref{fig:illus1}). Dataset $X_{q} \stackrel{\mathrm{i.i.d.}}{\sim} q({\boldsymbol{x}})$ does \emph{not
}contain any outlier. We set $\nu=\frac{|G|}{n_{p}}.$ The optimal
parameter $\boldsymbol{\delta}^{*}$ is set such that $p({\boldsymbol{x}})=q({\boldsymbol{x}})r({\boldsymbol{x}};\boldsymbol{\delta}^{*})$. We set
$\boldsymbol{w}_{i}^{*}=
\frac{1}{n_{p}}, \forall \boldsymbol{x}_{p}^{(i)}\text{ \ensuremath{\in}G}$ and 0 otherwise.
\paragraph{Remark:} Knowing the inlier proportion $|G|/n_{p}$ is a strong assumption. However it is only imposed for theoretical analysis. As we show in Section \ref{sec.label}, our method works well even if $\nu$ is only a rough guess (like $90\%$). Loosening this assumption will be an important future work.
\begin{assumption}
\label{ass.ratio.lip-1}
$
\forall\boldsymbol{u}\in\mathrm{Ball}(\rho),\sup_{\boldsymbol{x}}\left|\hat{z}{}_{\boldsymbol{\delta}^{*}+\boldsymbol{u}}(\boldsymbol{x})-\hat{z}_{\boldsymbol{\delta}^{*}}(\boldsymbol{x})\right|\le C_{\mathrm{lip}}\|\boldsymbol{u}\|_{1}.
$
\end{assumption}
This assumption says that the log density ratio model is Lipschitz continuous
around its optimal parameter ${\boldsymbol{\delta}}^{*}$ and hence there is a limit
how much a log ratio model can deviate from the optimal model under a small perturbation ${\boldsymbol{u}}$.
As our estimated weights $\hat{w}_i$ depends on the relative ranking of $\hat{z}_{\hat{{\boldsymbol{\delta}}}} ({\boldsymbol{x}}_{p}^{(i)})$, this assumption implies that the relative ranking between two points will remain unchanged under a small perturbation ${\boldsymbol{u}}$ if they are far apart.
The following theorem shows that if we have enough clearance between ``good''and ``bad samples'', $\hat{{\boldsymbol{\delta}}}$ converges to the optimal parameter ${\boldsymbol{\delta}}^*$.
\begin{thm}
\label{cor:outlier}
In addition to Assumption \ref{assu:.boundedrq}, \ref{assu:RSC} and
a few mild technical conditions (see Section \ref{sec:proofcol1} in the supplementary material), Assumptions \ref{ass.ratio.lip-1} holds. Suppose $\min_{j\in B}z_{\boldsymbol{\delta}^{*}}(\boldsymbol{x}_{p}^{(j)})-\max_{i\in G}z_{\boldsymbol{\delta}^{*}}(\boldsymbol{x}_{p}^{(i)})\ge3C_{\mathrm{lip}}\rho,$
$\nu=\frac{|G|}{n_{p}}, n_q = \Omega(|G|^2)$.
If
$
\lambda_{n}\ge2\cdot\max\left(\sqrt{\frac{K_{1}\log d}{|G|}},\frac{\rho \nu c}{2C_{r}^{2}\sqrt{n_{q}}}\right),
$
where $K_{1}>0,c>0$ are constants, we are
guaranteed that $||\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\|\le\frac{C_{r}^{2}}{\nu\kappa'_{1}}\cdot3\sqrt{k}\lambda_{n}$
with probability converging to 1. \label{thm:main-1}
\end{thm}
It can be seen that $\|\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\| =O\left(\sqrt{{\log d}/{\min(|G|, n_q)}}\right)$ if $d$ is reasonably large.
\subsection{Consistency under Truncation Setting\label{subsec:trimmedsetting}}
\label{sec:truncationsetting}
In this setting, we do not assume there are outliers in the observed data.
Instead, we examine
the ability of our estimator recovering the density ratio up to a
certain quantile of our data. This ability is especially useful when the
behavior of the tail quantile is more volatile and makes the standard estimator \eqref{eq:MLEOBJ} output
unpredictable results.
\paragraph{Notations:}
Given $\nu \in (0,1]$, we call $t_{\nu}(\boldsymbol{\delta})$ is the $\nu$-th quantile
of $z_{\boldsymbol{\delta}}$ if $P\left[z_{\boldsymbol{\delta}}<t_{\nu}(\boldsymbol{\delta}))\right]\le\nu$
and $P\left[z_{\boldsymbol{\delta}}\le t_{\nu}(\boldsymbol{\delta}))\right]\ge\nu.$
In this setting, we consider $\nu$ is fixed by a user thus we drop
the subscript $\nu$ from all subsequent discussions. Let's define
a truncated domain: $\overline{X}(\boldsymbol{\delta})=\left\{ \boldsymbol{x}\in\mathbb{R}^{d}|z_{\boldsymbol{\delta}}(\boldsymbol{x})<t(\boldsymbol{\delta})\right\} $,
$\overline{X}^{p}(\boldsymbol{\delta})=X_{p}\cap\overline{X}(\boldsymbol{\delta})$
and $\overline{X}^{q}(\boldsymbol{\delta})=X_{q}\cap\overline{X}(\boldsymbol{\delta})$. See Figure \ref{fig:illus2} for a visualization of $t({\boldsymbol{\delta}})$ and $\overline{X}({\boldsymbol{\delta}})$ (the dark shaded region).
\paragraph{Setting:}
Suppose dataset $X_{p}\stackrel{\mathrm{i.i.d.}}{\sim} P$ and $X_{q}\stackrel{\mathrm{i.i.d.}}{\sim} Q$. Truncated densities
$\overline{p}_{\boldsymbol{\delta}}$ and $\overline{q}_{\boldsymbol{\delta}}$
are the unbounded densities $p$ and $q$ restricted only on the truncated domain
$\overline{X}(\boldsymbol{\delta})$. Note that the truncated densities
are dependent on the parameter $\boldsymbol{\delta}$ and $\nu$.
We show that under some assumptions, the parameter $\hat{\boldsymbol{\delta}}$
obtained from \eqref{eq:min_max} using a fixed hyperparameter $\nu$
will converge to the $\boldsymbol{\delta}^{*}$such that $\overline{q}_{\boldsymbol{\delta}^{*}}(\boldsymbol{x})r(\boldsymbol{x};\boldsymbol{\delta}^{*})=\overline{p}_{\boldsymbol{\delta}^{*}}(\boldsymbol{x})$.
We also define the ``optimal'' weight assignment
$
w_{i}^{*} =
\frac{1}{n_{p}}, \forall i,\boldsymbol{x}_{p}^{(i)}\in\overline{X}(\boldsymbol{\delta}^{*})
$
and 0 otherwise.
Interestingly, the constraint in \eqref{eq:min_max}, $\langle \boldsymbol{w}^{*}, {\boldsymbol{1}} \rangle = \nu $ may \emph{not} hold, but our analysis in this section suggests we can always find
a pair $(\hat{{\boldsymbol{\delta}}},\hat{\boldsymbol{w}})$ in the feasible
region so that $\|\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\|$
converges to 0 under mild conditions.
We first assume the log density ratio model
and its CDF is Lipschitz continuous.
\begin{assumption}
\label{ass.ratio.lip}
\begin{equation}
\forall\boldsymbol{u}\in\mathrm{Ball}(\rho),\sup_{\boldsymbol{x}}\left|\hat{z}{}_{\boldsymbol{\delta}^{*}+\boldsymbol{u}}(\boldsymbol{x})-\hat{z}_{\boldsymbol{\delta}^{*}}(\boldsymbol{x})\right|\le C_{\mathrm{lip}}\|\boldsymbol{u}\|.\label{eq:lip}
\end{equation}
Define $T(\boldsymbol{u},\epsilon):=\left\{ \boldsymbol{x}\in\mathbb{\mathbb{R}^{\mathit{d}}}\mid\;|z_{\boldsymbol{\delta}^{*}}({\boldsymbol{x}})-t(\boldsymbol{\delta}^{*})|\le2C_{\mathrm{lip}}\|\boldsymbol{u}\|+\epsilon\right\} $
where $0<\epsilon\le1$. We assume $\forall\boldsymbol{u}\in\mathrm{Ball}(\rho),0<\epsilon\le1$
\[
P\left[\boldsymbol{x}_{p}\in T(\boldsymbol{u},\epsilon)\right]\le C_{\mathrm{CDF}}\cdot\|\boldsymbol{u}\|+\epsilon.
\]
\end{assumption}
In this assumption, we define a ``zone'' $T(\boldsymbol{u},\epsilon)$
near the $\nu$-th quantile $t(\boldsymbol{\delta}^{*})$ and assume
the CDF of our ratio model is upper-bounded over this region. Different from Assumption \ref{ass.ratio.lip-1}, the RHS of \eqref{eq:lip}
is with respect to $\ell_{2}$ norm of ${\boldsymbol{u}}$. In the following assumption, we assume regularity on $P$ and $Q$.
\begin{assumption}
\label{assu:regP}
$\forall {\boldsymbol{x}}_q \in \mathbb{R}^{d}, \|\boldsymbol{f}(\boldsymbol{x}_{q})\|_{\infty}\le C_{q}$ and $\forall\boldsymbol{u}\in\mathrm{Ball}(\rho),\forall\boldsymbol{x}_{p}\in T(\boldsymbol{u},1),\|\boldsymbol{f}(\boldsymbol{x}_{p})\|_{\infty}\le C_{p}$.
\end{assumption}
\begin{thm}
\label{cor:truncationsetting}
In addition Assumption \ref{assu:.boundedrq} and \ref{assu:RSC} and other mild assumptions (see Section \ref{sec:proofoftruncation} in the supplementary material), Assumption \ref{ass.ratio.lip} and
\ref{assu:regP} hold. If $1\ge\nu\ge\frac{8C_{\mathrm{CDF}}\sqrt{k}C_{p}C_{r}^{2}}{\kappa'_{1}},n_q = \Omega(|\overline{X}{}^{p}(\boldsymbol{\delta}^{*})|^2)$,
\[\textstyle{
\lambda_{n}\ge2\max\left[\sqrt{\frac{K'_{1}\log d}{|\overline{X}{}^{p}(\boldsymbol{\delta}^{*})|}}+\frac{2C_{r}^{2}C_{q}|X_{q}\backslash\overline{X}^{q}(\boldsymbol{\delta}^{*})|}{n_{q}},\frac{2L\cdot C_{p}}{\sqrt{n_{p}}},\frac{\rho \nu c}{2C_{r}^{2}\sqrt{n_{q}}}\right],
}
\]
where $K'_{1}>0, c>0$ are constants,
we are guaranteed that $||\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\|\le\frac{4C_{r}^{2}}{\nu\kappa'_{1}}\cdot3\sqrt{k}\lambda_{n}$
with high probability.\label{thm:main-1-1}
\end{thm}
It can be seen that $\|\hat{\boldsymbol{\delta}}-\boldsymbol{\delta}^{*}\| = O\left(\sqrt{{\log d}/{\min(|\overline{X}^{p}(\boldsymbol{\delta}^{*})|, n_q)}}\right)$ if $d$ is reasonably large and $|X_{q}\backslash\overline{X}^{q}(\boldsymbol{\delta}^{*})|/{n_{q}}$ decays fast.
\section{Experiments}
\label{sec.label}
\subsection{Detecting Sparse Structural Changes between Two Markov Networks (MNs) \cite{Liu2016a}}
In the first experiment\footnote{Code can be found at \url{http://allmodelsarewrong.org/software.html}}, we learn changes between two Gaussian MNs under the outlier setting.
The ratio between two Gaussian MNs can be parametrized as $p({\boldsymbol{x}})/q({\boldsymbol{x}}) \propto \exp(-\sum_{i,j \le d} \Delta_{i,j} x_i x_j)$, where $\Delta_{i,j} := \Theta^{p}_{i,j} - \Theta^{q}_{i,j}$ is the difference between precision matrices.
We generate 500 samples as $X_p$ and $X_q$ using two randomly structured Gaussian MNs.
One point $[10, \dots, 10]$ is added as an outlier to $X_p$. To induce sparsity, we set $R({\boldsymbol{\Delta}}) = \sum_{i,j=1, i\le j}^{d}{|\Delta_{i, j}|}$ and fix $\lambda = 0.0938$. Then run DRE and TRimmed-DRE to learn the sparse \emph{differential} precision matrix ${\boldsymbol{\Delta}}$ and results are plotted on Figure \ref{fig:mnchange2} and \ref{fig:mnchange1}\footnote{Figures are best viewed in color.} where the ground truth (the position $i,j, \Delta_{i,j}^* \neq 0$) is marked by red boxes. It can be seen that the outlier completely misleads DRE while TR-DRE performs reasonably well. We also run experiments with two different settings ($d=25, d=36$) and plot True Negative Rate (TNR) - True Positive Rate (TPR) curves. We fix $\nu$ in TR-DRE to 90\% and compare the performance of DRE and TR-DRE using DRE without any outliers as gold standard (see Figure \ref{fig:TNRTPR}). It can be seen that the added outlier makes the DRE fail completely while TR-DRE can almost reach the gold standard. It also shows the price we pay: TR-DRE does lose some power for discarding samples. However, the loss of performance is still acceptable.
\begin{figure}
\centering
\subfloat[$\hat{{\boldsymbol{\Delta}}}$ obtained by DRE, $d = 20$, with one outlier.]{
\label{fig:mnchange2}\includegraphics[width = .33\textwidth]{changeD2}}
\subfloat[$\hat{{\boldsymbol{\Delta}}}$ obtained by TR-DRE, $\nu = 90\%$, with one outlier.]{\label{fig:mnchange1}\includegraphics[width = .33\textwidth]{changeD1}}
\subfloat[TNR-TPR plot, $\nu=90\%$]{\label{fig:TNRTPR}\includegraphics[width = .33\textwidth]{robust_kliep}}
\caption{Using DRE to learn changes between two MNs. We set $R(\cdot) = \|\cdot\|_1 $ and $ f(x_i, x_j) = x_i x_j$. \label{fig.mnchange}}
\end{figure}
\subsection{Relative Novelty Detection from Images}
\begin{figure}
\centering
\subfloat[Dataset]{\label{fig:dataset}\includegraphics[width = .16\textwidth]{combine}}
\subfloat[$\nu = 97\%$]{\label{fig:detected97}\includegraphics[width = .16\textwidth]{detected97}}
\subfloat[$\nu = 90\%$]{\label{fig:detected90}\includegraphics[width = .16\textwidth]{detected90}}
\subfloat[$\nu = 85\%$]{\label{fig:detected85}\includegraphics[width = .16\textwidth]{detected85}}
\subfloat[TH-DRE]{\label{detectedsmola}\includegraphics[width = .16\textwidth]{detectedsmola}}
\subfloat[one-SVM]{\label{detectedosvm}\includegraphics[width = .16\textwidth]{detectedlibsvm}}
\caption{Relative object detection using super pixels. We set $R(\cdot) = \|\cdot\|^2$, $ {\boldsymbol{f}}({\boldsymbol{x}})$ is an RBF kernel.}
\end{figure}
In the second experiment,
we collect four images (see Figure \ref{fig:dataset}) containing three objects with a textured background: a pencil, an earphone and an earphone case. We create data points from these four images using sliding windows of $48 \times 48$ pixels (the green box on the lower right picture on Figure \ref{fig:dataset}). We extract 899 features using MATLAB HOG method on each window and construct an 899-dimensional sample. Although our theorems in Section \ref{sec:-Consistency-in} are proved for linear models, here ${\boldsymbol{f}}({\boldsymbol{x}})$ is an RBF kernel using all samples in $X_p$ as kernel basis. We pick the top left image as $X_p$ and using all three other images as $X_q$, then run TR-DRE, THresholded-DRE \cite{Smola2009}, and one-SVM.
In this task, we select high density ratio super pixels on image $X_p$.
It can be expected that the super pixels containing the pencil will exhibit high density ratio values as they did not appear in the reference dataset $X_q$ while super pixels containing the earphone case, the earphones and the background, repeats similar patches in $X_q$ will have lower density ratio values.
This is different from a conventional novelty detection, as a density ratio function help us capture only the relative novelty.
For TR-DRE, we use the trimming threshold $\hat{t}$ as the threshold for selecting high density ratio points.
It can be seen on Figure \ref{fig:detected97}, \ref{fig:detected90} and \ref{fig:detected85}, as we tune $\nu$ to allow more and more high density ratio windows to be selected, more relative novelties are detected: First the pen, then the case, and finally the earphones, as the lack of appearance in the reference dataset $X_q$ elevates the density ratio value by different degrees.
In comparison, we run TH-DRE with top 3\% highest density ratio values thresholded, which corresponds to $\nu = 97\%$ in our method. The pattern of the thresholded windows (shaded in red) in Figure \ref{detectedsmola} is similar to Figure \ref{fig:detected97} though some parts of the case are mistakenly shaded.
Finally, one-SVM with $3\%$ support vectors (see Figure \ref{detectedosvm}) does not utilize the knowledge of a reference dataset $X_q$ and labels all salient objects in $X_p$ as they corresponds to the ``outliers'' in $X_p$.
\section{Conclusion}
\label{sec.concl}
We presents a robust density ratio estimator based on the idea of
trimmed MLE.
It has a convex formulation
and the optimization can be easily conducted using a subgradient ascent method.
We also investigate its theoretical property through an equivalent
weighted M-estimator whose $\ell_{2}$ estimation error bound was
provable under two high-dimensional, robust settings.
Experiments confirm the
effectiveness and robustness of the our trimmed estimator.
\section*{Acknowledgments}
We thank three anonymous reviewers for their detailed and helpful comments. Akiko Takeda thanks Grant-in-Aid for Scientific Research (C), 15K00031. Taiji Suzuki was partially supported by MEXT KAKENHI (25730013, 25120012, 26280009 and 15H05707), JST-PRESTO and JST-CREST. Song Liu and Kenji Fukumizu have been supported in part by MEXT Grant-in-Aid for Scientific Research on Innovative Areas (25120012).
\bibliographystyle{plain}
|
\section{Introduction}
The nonlinear Schr\"{o}dinger equation (NLSE) is a generic
fundamental mathematical model with numerous applications in
science and technology. In particular, the NLSE describes a
path-average propagation of light in fibre-optic systems that is
the backbone of the modern global communication networks.
The NLSE is an example of a fundamental
model of nonlinear physics which can be integrated by the Inverse
Scattering Transform (IST) method~\cite{zakharov1972exact,AKNS}.
The IST method is one of the greatest achievements of
mathematical physics in the 20-th century (see e.g.~\cite{zakharov1972exact,AKNS,ZMNP,Newell,AS,hasegawa1995solitons,lamb1980elements}
and references therein). In recent years (especially in optical communications) the
IST method is also referred to as the Nonlinear Fourier Transform (NFT) stressing the similarity to the
conventional Fourier transform and the ability of the
IST/NFT to present solutions of the nonlinear evolution equation
on the basis of non-interacting modes called scattering data
or (in the NFT notation) \textit{nonlinear spectrum}.
One specific, albeit highly important, application of the NLSE is in optical communications, where
it is derived as a path-average (over periodic variations of power due to loss and gain) model governing
the signal propagation along the transmission line~\cite{hasegawa1995solitons,kivshar2003optical,mollenauer2006solitons} (here we use the normalised version):
\begin{equation}\label{NLSE}
\mathrm i\frac{\pt q}{\pt z} + \frac{1}{2}\frac{\pt^2 q}{\pt t^2} + |q|^2q=0.
\end{equation}
Here $q(z,t)$ is an optical field envelope, $z$ is a propagation
distance along the optical fibre and $t$ is the retarded time.
The general solution of the NLSE is presented by the superposition of solitary (localised in time) waves
corresponding to the discrete (solitonic) part of the nonlinear spectrum and dispersive waves associated with the
continuous part of the nonlinear spectrum. Recent advances in coherent optical communication allowing information coding both over amplitude and phase
have made it possible to reconsider relatively old ideas of using the soliton solution of the NLSE~\cite{hasegawa1995solitons,mollenauer2006solitons} and nonlinear spectrum eigenvalues for the transmission of information~\cite{HN} in the new context. The recent surge of interest in nonlinear transmission techniques is in particular due to the observation that conventional (linear) data transmission techniques are facing serious challenges induced by the nonlinear properties of the optical fibre communication channels (an excellent overview is given in~\cite{Rene1,Rene2}). This calls for the development of new nonlinear techniques of signal coding, transmission and processing.
The traditional soliton transmission has been recently reassessed in~\cite{PDTPRL,yushko2014coherent} in the context of coherent communication and the use of soliton phase for data transmission. Moreover, a great deal of interest has been sparked recently by the application of the powerful IST/NFT methods in optical communications (see e.g.~\cite{NFT1,yousefi2014information,NFT3,NFT4,NFT5,NFT6,NFT7} and references therein, we simply are not able to review here all relevant papers that have been published recently in this fast growing field).
The efficiency of numerical algorithms for data encoding/decoding is critically important
in the digital telecommunication networks. For instance, in wireless communication, the success and popularity of the orthogonal frequency division multiplexing
(OFDM) method is due to the exceptional computational performance and high spectral efficiency
of the fast Fourier transform (FFT)~\cite{shieh2009ofdm}.
The success of the practical implementation of the nonlinear IST/NFT techniques will be defined by the availability of the fast and super-fast NFT methods~\cite{belai2007efficient,frumin2015efficient,NFT5} and the stability of algorithms with respect to noise impact. The IST/NFT technique is relatively new compared to conventional methods and the currently available numerical algorithms of information encoding/decoding using solitonic signal are still far from the efficiency required in practical hardware implementation.
Here we propose to use the kernel of the Gelfand-Levitan-Marchenko equations (GLME) to encode information, in particular, we demonstrate that the OFDM scheme can be applied in an efficient way. To create signal at the beginning of the transmission line as well as to recover the encoded kernel at the end of the line, here we use the efficient Toeplitz inner-bordering (TIB) numerical scheme of inverse and direct scattering transform (which was recently introduced by Frumin and co-authors~\cite{belai2007efficient,frumin2015efficient}) and the exact soliton solution known from the IST theory \cite{zakharov1972exact,AKNS}.
For the discrete nonlinear spectrum, we propose a soliton orthogonal frequency division multiplexing (SOFDM) technique that is
based on the choice of identical imaginary parts of N-soliton solution eigenvalues, corresponding to equidistant soliton frequencies making it similar to the conventional OFDM scheme and allowing the use of the efficient fast Fourier transform algorithm to recover the data.
We also demonstrate how the concept of the OFDM can be applied for the continuous spectrum kernel~\cite{NFT3}. The important advantage of using coding over kernel of the GLME is the possibility of controlling signal parameters by utilising the time domain window functions in the modulated kernel.
\section{$N$-soliton solutions of the NLSE for $N$-symbol block transmission}
In the traditional soliton transmission a single (soliton) pulse is used as an information carrier
sent over a time slot allocated to one symbol in a given spectral channel~\cite{hasegawa1995solitons,mollenauer2006solitons}. Transmission in this case is affected by the soliton interactions, and/or is restricted in the spectral efficiency, because a separate soliton occupies a small fraction of the symbol duration time.
A great deal of attention has recently been placed on the potential use of the discrete nonlinear eigenvalues in fibre-optic channels.
Most of the current studies of discrete nonlinear eigenvalue communications are limited to exploring different solitonic waveforms (forming a transmitted alphabet) in a single symbol time slot. To avoid interaction between neighbouring symbols a long guard interval is typically used to suppress inter-symbol interactions, thus, limiting spectral efficiency of such burst-mode transmission.
We propose here to use the well-known~\cite{zakharov1972exact,AKNS}
general analytical $N$-soliton solutions of the NLSE (N-SS) (see the Supplemental Material~\cite{SupM}) for $N$-symbol block modulation and coding. In a block transmission technique the information symbols are arranged in the blocks separated by some known symbols. Application of the N-SS allows one to simultaneously code information over $N$ symbol time intervals. Four soliton parameters in principle offer a possibility of four-dimensional modulation/coding per soliton/symbol.
Over the interval of $N$ symbols, $N$-soliton solutions can offer $4 \times N$ degrees of freedom.
Recall that single soliton solutions reads:
\begin{equation}\label{1-SS}
q^{(1)}(z,t) = 2\beta \frac{\exp[-2i\omega t - 2i (\omega^2-\beta^2) z + i \theta]}{\cosh(2\beta t + 4\omega\beta z - \delta t)}\,.
\end{equation}
Here, obviously, $2 \beta$ corresponds to soliton amplitude, $2 \omega$ is soliton frequency, $\theta$ is pulse phase
and $ \delta t/(2 \beta)$ defines soliton timing position. These four parameters can be used for coding of information, i.e. amplitude, frequency, phase and pulse position modulations, leading to various high-level modulation formats. Note that interactions between solitons are automatically accounted for in the $N$-soliton solution. Therefore, in N-SS coding there is no issue of soliton interactions that occur when solitons are treated as separate entities.
The N-SS is defined by its scattering data/nonlinear spectrum: two sets of $N$ complex constants. The first set corresponds to the complex eigenvalues of
solitons $\xi_k = i\beta_k+\omega_k$, $k=1,...,N$. As discussed above, the imaginary part $\beta_k>0$ defines corresponding (with the index $k$) soliton amplitude and the real part $\omega_k$ is related to the soliton frequency (and corresponding group speed). The second set is given by the complex numbers $c_k=C_k \exp(i\theta_k)$ with real parameters $C_k$ and $\theta_k$. For the well separated solitons, parameters $C_k$ define timing positions of solitons in the following way: $\delta t_k = \ln[C_k/(2\beta_k)]$, while parameters $\theta_k$ define soliton phases. Based on the structure of the solitonic scattering data, the possible data coding of N-SS form two natural groups classified as: \textit{amplitude-frequency} modulation and \textit{pulse position-phase} modulation. In general, there are $4 \times N$ free parameters that can be used for modulation.
The generation of modulated (i.e. encoded) N-SS signal at the transmitter requires an algorithmic realisation of IST/NFT in the encoder. Here, to find the N-SS we use the standard factorisation of the GLME, that leads to the well known exact formulae (see the Supplemental Material~\cite{SupM}). Alternatively, the N-SS can be obtained by algebraic versions of IST such as the Zakharov-Shabat dressing
method~\cite{zakharov1974scheme}, Darboux transformation~\cite{matveev1991darboux}, method of
Hirota~\cite{hirota1973exact} and by the IST TIB algorithm. Note, that all these approaches are numerically unstable at large $N$, that limits their applications (see the Supplemental Material~\cite{SupM}). The kernel $\Omega(z,t)$ of the GLME for the N-SS has the following form:
\begin{equation}\label{kernel}
\Omega^{(N)}(t,z) = \sum_{k=1}^{N} c_k(z)e^{-i\xi_k t}\,.
\end{equation}
The sum in~Eq.(\ref{kernel}) is similar to the Fourier series,
however, the "frequencies" $\xi_k$, in general, are complex numbers. Formally,
the N-SS can be written as IST/Inverse NFT of the kernel~(\ref{kernel}):
\begin{equation}\label{N-SS IST}
q^{(N)}(z,t) = IST[\Omega^{(N)}(z,t)]\,.
\end{equation}
In what follows, for the sake of simplicity, the index $(N)$ will be
omitted. We assume that coding/modulation is applied at
$z=0$ (encoder) and decoding/demodulation (decoder) at $z=L$. The
IST/NFT method links the nonlinear spectrum $z=0$ and $z=L$ by the
following simple phase shift:
\begin{equation}\label{phase correction}
c_k(L) = c_k(0)\exp(-2i\xi_k^2 L), \,\,\,\,\, k=(1,...,N)\,.
\end{equation}
Expressions~(\ref{N-SS IST}) and~(\ref{phase correction})
formally solve the problem of compensation of signal distortions in the communication channels
described by the NLSE. We believe, that formula~(\ref{kernel}) can offer some advantages for
encoding/decoding operations compared to the traditionally used formula~(\ref{N-SS IST}).
Our central idea is to use the N-SS
kernel~(\ref{kernel}) for the modulation of the information data. In this case the
number of numerical operations at the decoder is reduced. In the proposed scheme
the decoding operation requires to recover only the kernel~(\ref{kernel}) at $z=0$ by solving the direct scattering
problem and by application of a simple transformation~(\ref{phase
correction}). Moreover, as we will demonstrate, the analogy of the
N-SS kernel~(\ref{kernel}) with the Fourier series allows to
introduce the OFDM scheme for the discrete spectrum.
\section{Solitonic OFDM method}
In this section we introduce a soliton OFDM (SOFDM) technique in which the GLME kernel can be efficiently used for encoding/decoding $2\times N$ position-phase parameters. The key idea can be understood from the expression for the N-soliton kernel~(\ref{kernel}).
The kernel would be similar to the conventional OFDM in case of real $\xi_k$.
Therefore, we impose special conditions on the complex soliton parameters $\xi_k$. Namely, we consider
N-SS with eigenvalues $\xi_n = \omega_n + iA$. In this case
solitons have equal amplitudes, but different equidistantly selected frequencies. Thus, the
GLME kernel~(\ref{kernel}) at the beginning of the line is given
by the finite Fourier series multiplied by $e^{At}$:
\begin{equation}\label{SOFMkernel}
\Omega(0,t) = e^{At} \sum_{k=1}^{N} c_k e^{-i \omega_k t}\,.
\end{equation}
This greatly simplifies the processing of such signals.
Now, without loss of generality, we consider modulation over phase $\theta_n$,
while the pulse positions $\delta_n$ are left unmodulated. As a particular
example of the SOFDM encoding we consider
$\tilde{N}$-phase-shift keying ($\tilde{N}$-PSK) modulated N-SS.
To apply the SOFDM over the finite time slot $T$ we introduce the
discrete time grid:
\begin{equation}
t_m = (m-1)T/N, \,\,\, m=1,...,N.
\end{equation}
The orthogonality of Fourier harmonics is given by the following
condition:
\begin{equation}\label{Orthogonality}
t_m \omega_n = 2\pi (m-1)(n-1)/N, \,\,\, m,n=1,...,N\,.
\end{equation}
Similar to the standard OFDM, the FFT makes it possible to determine
parameters of signal modulation $c_n$ by $O(N \ln N)$ arithmetic
operations:
\begin{equation}
c_n = FFT[\Omega(0,t_m) \exp(-A t_m)]\,.
\end{equation}
To compute the scattering data from the received signal $q(t,L)$ one can use any available algorithm of the direct NFT. Here, without loss of generality, we use the direct TIB algorithm calculating the entire signal kernel in time domain by solving the GLME (see Supplemental Material~\cite{SupM} and ~\cite{frumin2015efficient}). The kernel contains all scattering data information: soliton eigenvalue positions (corresponding to amplitude and frequency modulations), pulse positions and phases ($4\times N$ parameters). Here we focus only on a phase modulation to illustrate the proposed concept. The eigenvalue modulation is also possible, but it faces challenges in terms of efficiency and stability (see Supplemental Material~\cite{SupM}).
For illustration purposes we choose the minimum possible number of time samples $M=N$. But
actually, the value of $M$ depends on the algorithm of the direct
scattering transform at the receiver and usually $M > N$.
At the transmitter, we use the IFFT to obtain the
kernel~(\ref{SOFMkernel}) from the given data encoded by the phases $c_n$ and
then solve the inverse scattering problem as described in the Supplemental Material~\cite{SupM}
to generate the input optical N-SS signal $q(0,t_m)$.
We test the SOFDM method in numerical simulations of data
transmission by the use of quaternary phase-shift keying (QPSK)
modulated $6$-SS. In Fig.~\ref{Fig_1} (left) we present an example of
a 6-soliton signal at the beginning and at the end of the transmission line of length $L=2000$ km.
Using the direct TIB method we recover the encoded kernel that presented in Fig.~\ref{Fig_1} (right).
To avoid signal expansion, we arrange the solitons in order of descending velocity so that the slowest soliton occupies
the first position in the signal while the fastest soliton starts propagation from the signal end.
However, we would like to stress that the practical implementation of the solitonic OFDM scheme requires further development of fast noise-stable methods for solving the direct scattering problem that we consider in the Discussion section.
\begin{figure}[h]
\includegraphics[width=1.65in]{EPS/Fig1a.eps}
\includegraphics[width=1.65in]{EPS/Fig1b.eps}
\caption{\label{Fig_1} LEFT: The modulus of 6-soliton solution at
the beginning ($q_0$, red dashed line) and at the end ($q_1$, blue solid line) of the NLSE governed transmission line.
\\
RIGHT: 6-soliton normalized kernel of the GLME
equations at the beginning of the transmission line:
$|\Omega(0,t)exp(-At)|$. Blue solid line -- the exact normalized
kernel encoded by the SOFDM method with QPSK, red dashed line -- the same
kernel restored by the use of direct TIB method.}
\end{figure}
\section{Kernel coding of the continuous nonlinear spectrum OFDM}
The kernel of the GLME for the continuous spectrum is presented in the following form:
\begin{equation}\label{CSOFMkernel}
\Omega^{(R)}(0,t) = R(0,t) = \frac{1}{2\pi}\int_{-\infty}^{\infty}r_{\omega}e^{-i\omega t} d\omega \,.
\end{equation}
Here $r_{\omega}$ is the reflection coefficient of the Zakharov-Shabat system for the given signal $q^{(R)}(0,t)$ while $R(0,t)$ is commonly referred to as a signal response function. Similar to the discrete spectrum case, the IST links the continuous spectrum at $z=0$ and $z=L$ by the following relation:
\begin{equation}\label{phase correction CS}
r_{\omega}(L) = r_{\omega}(0)\exp(-2i\omega^2 L)\,.
\end{equation}
The general idea to apply the OFDM scheme to the continuous nonlinear spectrum was previously considered in the framework of the so-called "nonlinear inverse synthesis" method~\cite{NFT3,le2015nonlinear}. This approach, whilst promising, has an important challenge -- how to control the signal characteristics in the time domain? Indeed, the reflection coefficient $r(\omega)$ which was chosen for encoding information is nontrivially coupled with the signal via IST.
Here, we propose to apply the additional window transformation to the kernel $\Omega^{(R)}(0,t)$ in the time domain, as a method of controlling signal parameters. For IST-based schemes the strong localisation of signal in time slots is highly critical to avoid nonlinear interactions between neighbouring symbol intervals. Bearing in mind the linear limit ($q^{(R)}(z,t) \ra 2\Omega^{(R)}(z,t)$) we conclude, that well localised (in time domain) signals should correspond to the localised kernel at least in a weakly nonlinear case. We have examined different window transformation functions known from the linear communication theory (see, for instance~\cite{jenkins1968spectral}) and found that the excellent signal localisation in time is achieved for window functions with smooth polynomial fronts.
Figure~\ref{Fig_2} demonstrates signal generation at the beginning of the transmission line. We start from 16 Fourier harmonics encoded using the OFDM $8$-PSK scheme. Then we apply the window transformation $f(t)$ similar to the well known Lorentzian function:
\begin{equation}\label{Lorenz}
F(t) = \frac{\tilde{A}}{ [\Gamma(t-t_0)]^2 + 1}\,,
\end{equation}
to localise signal in time slot. Here $\tilde{A},\, \Gamma$ are the parameters of the window transformation corresponding to the characteristic amplitude and width of the modulated kernel and $t_0$ corresponds to the centre of the time slot. We also add to the window transformation function~(\ref{Lorenz}) exponentially decaying tails which do not affect the general shape of the signal, but helps to cancel interactions between neighbouring bursts. Finally, we find the signal profile using the inverse TIB method (Fig.~\ref{Fig_2}, right, red).
\begin{figure}[h]
\includegraphics[width=1.65in]{EPS/Fig2a.eps}
\includegraphics[width=1.65in]{EPS/Fig2b.eps}
\caption{\label{Fig_2} LEFT: 16 encoded kernel harmonics (grey, solid line) and window function (black, dashed line). Parameters of the window transformation function~(\ref{Lorenz}) are the following: $\tilde{A}=15,\, \Gamma=20$.
\\
RIGHT: The double kernel (blue, solid line) obtained by multiplying the encoded harmonics by the window function and corresponding signal obtained by the use of the inverse TIB algorithm (red, dashed line). The inset picture shows the absolute value of signal Fourier spectrum, frequency index $n$ is defined in ~(\ref{Orthogonality}))}.
\end{figure}
Next, we study the dependence of signal shape on the parameters $\tilde{A}$ and $\Gamma$ of the modulated kernel. We have found that varying the kernel window function parameters allows us to control the characteristics of the generated signal without affecting the information content as illustrated in Fig.~\ref{Fig_3}. Figure~\ref{Fig_4} presents the results of numerical modelling of a burst mode signal transmission in the NLSE channel (with noise) with a total propagation distance L of 1000km and SNR=19.7 dB. The modelling was performed using the standard split-step method with adding noise at each numerical spatial step that corresponds to the distributed noise model (see e.g.~\cite{Rene1,Rene2}). We have found that the direct TIB algorithm remains stable for considered SNR values. Interestingly, better decoding results can be achieved using only the first (left) part of the pulse. This is not surprising since the direct TIB algorithm successively recover the kernel from the left to the right end of the signal. Thus, the right part of the recovered kernel is additionally affected by noise distortions accumulated during the calculation of the left kernel part.
\begin{figure}[h]
\includegraphics[width=1.65in]{EPS/Fig3a.eps}
\includegraphics[width=1.65in]{EPS/Fig3b.eps}
\caption{\label{Fig_3}
Dependence of signal from the parameters of kernel window transformation function~(\ref{Lorenz}).}
\end{figure}
\begin{figure}[h]
\includegraphics[width=1.65in]{EPS/Fig4a.eps}
\includegraphics[width=1.65in]{EPS/Fig4b.eps}
\caption{\label{Fig_4}
LEFT: Propagation of the burst mode signal in the model NLSE channel with L=1000km and SNR=19.7 dB. Blue solid lines correspond to the signal at the beginning of the line, red dashed lines show the signal at the end of the line. Parameters of of the window transformation function are the same as in the Fig.~\ref{Fig_2}. The bottom picture corresponds to the encoded (blue, solid lines) and decoded using the direct TIB algorithm (red, dashed lines) kernel harmonics for the central burst interval.
\\
RIGHT: Constellation diagram for the central burst interval (statistics on a $10^3$ randomly encoded initial signals).}
\end{figure}
\section{Discussion and Conclusion}
In this work, we proposed and examined new approaches to coding information over the kernel of the GLME. We have considered both the discrete (solitonic) and continuous part of scattering data. We demonstrated that application of the direct TIB method allowes one to recover the most stable part of the kernel, that is an advantage in the presence of distributed noise.
We have proposed, to the best of our knowledge for the first time, to use the general N-soliton solution of the NLSE for simultaneous coding of $N$ symbols
involving $4\times N$ coding parameters, instead of separate $N$ solitons. As a particular sub-class of the general schemes we examined a soliton orthogonal frequency division multiplexing technique that is based on the choice of identical imaginary parts of N-soliton solution eigenvalues, corresponding to equidistant soliton frequencies making it similar to the conventional OFDM scheme. This allows us to use the efficient fast Fourier transform algorithm to recover the data. We would like to point out that efficient implementation of numerical recovery of solitonic kernels by solving GLME requires the development of numerical algorithms, which are stable against additive noise.
For the continuous spectrum we have tested stability of the direct TIB method against the additive noise and proposed to use the localised kernel in the time domain to control properties of the corresponding generated signal. The latter can be considered as a novel realisation of the "nonlinear inverse synthesis" method~\cite{NFT3,le2015nonlinear}.
We demonstrated that the mathematical properties of the NLSE can be used for introducing fundamentally novel (compared to the linear communication theory)
methods for coding and detection of signal setting foundation for the nonlinear communication theory.
\section{Acknowledgments and Contributions}
This work was supported by the UK EPSRC Programme Grant UNLOC EP/J017582/1 and Grant of the Ministry of Education and Science of the Russian Federation (agreement No. 14.B25.31.0003). Part of the work described in the section "Kernel coding of the continuous nonlinear spectrum OFDM" was supported by the Russian Science Foundation (Grant No. 14-22-00174).
All Authors (L.F., A.G., S.K.T.) proposed key ideas, participated in the discussion of results and contributed equally to this work. Numerical modelling was performed by A.G. and L.F.
\bibliographystyle{apsrev4-1}
\section{Numerical approaches for inverse and direct scattering transform}
In this paragraph we briefly overview and remind basic information concerning numerical Inverse Scattering Transform (IST) which is used in the main text of our paper~\cite{PAPER}. First, we write down the Gelfand-Levitan-Marchenko equations (GLME) in the standard form for the left scattering problem at fixed distance (e.g. $z=0$):
\begin{eqnarray}\nonumber
A^*_1(t,s) + \int_{-s}^t A_2(t,\tau) \Omega(s+\tau) d\tau = 0,
\\\nonumber
-A^*_2(t,s) + \int_{-s}^t A_1(t,\tau) \Omega(s+\tau) d\tau + \Omega(t+s)= 0,
\\\nonumber
\Omega(t) \equiv \Omega(z=0,t),
\\
-t \leqslant s < t, \,\,\,\,\,\,\,\,\,\,\,\,\,\,\, 0 \leqslant t \leqslant T_s.
\label{GLME}
\end{eqnarray}
Here $A_1(t,s)$ and $A_2(t,s)$ are the auxiliary complex functions that links together the kernel $\Omega$ and solution $q$ of the NLSE via the GLME~(\ref{GLME}) and the following relation:
\begin{equation}\label{N-SS}
q(z=0,t) = -2 A_2^*(t,t) \,.
\end{equation}
The propagation problem is solved by the use of a simple formulae for scattering data evolution (see Eq.~(5) and Eq.~(11) in~\cite{PAPER}).
In the general case numerical solution of an integral equation requires $\sim M^3$ operations (recall that $M$ is the number of signal discretisation points). To reconstruct the whole signal $q(t_m)$ we need to perform this procedure at all points of the discrete grid (formula.~(7) in~\cite{PAPER}) and, thus, the total cost $\sim M^4$ operations, that is not feasible for practical numerical implementation.
In this work we use the efficient Toeplitz inner-bordering (TIB) numerical scheme for both the inverse and direct scattering transform. Indeed, as it was shown Frumin and co-authors (reference [23] in~\cite{PAPER}) the GLME~(\ref{GLME}) can be rewritten in the Toeplitz form by applying a simple transformation:
\begin{eqnarray}\nonumber
u(t,x) = A_1(t,t-x)\,,
\\
v(t,y) = -A_2^*(t,y-t)\,.
\end{eqnarray}
Now the GLME contains Toeplitz-type kernel $\Omega(y-x)$:
\begin{eqnarray}\nonumber
u(t,x) - \int_{-x}^{2t} \Omega^*(y-x) v(t,y) dy = 0,
\\
v(t,y) + \int_{0}^y \Omega(y-x) u(t,x) dx + \Omega(y)= 0\,,
\label{Toeplitz GLME}
\end{eqnarray}
and, as a result, the numerical TIB IST takes only $M^2$ operations (see details in reference [23] in~\cite{PAPER}). Moreover, recently Frumin and co-authors have demonstrated (reference [24] in~\cite{PAPER}) that the TIB algorithm can be reversely applied to the GLME~(\ref{Toeplitz GLME}), i.e. it allows to find the kernel $\Omega(t_m)$ from the known signal $q(t_m)$. Again, the required number of numerical operations is $M^2$. The numerical schemes and details can be found in references [23,24] in~\cite{PAPER}.
Here~\cite{PAPER} we apply both inverse and direct TIB algorithm to the continuous spectrum signals. For the discrete spectrum case we apply only direct TIB method to recover the kernel, meanwhile to create signal at the beginning of the transmission line we use exact N-SS, described in the next paragraph.
\subsection{$N$-soliton solutions of the NLSE}
For the discrete spectrum kernel (see formula (3) in~\cite{PAPER}) factorization of the GLME~(\ref{GLME}) leads to the system of linear algebraic equations (see, for instance the monograph of Lamb~\cite{lamb1980elements}). Then, the N-SS can be found in the following exact form:
\begin{equation}\label{N-SS}
q^{(N)}(z=0,t) = -2 \bra{\mathbf{\Psi}(t)}(\widehat{\mathbf{E}} +
\widehat{\mathbf{M}}(t)^*
\widehat{\mathbf{M}}(t))^{-1}\ket{\mathbf{\Phi}(t)}\,.
\end{equation}
Here $\widehat{\mathbf{E}}$ is $N\times N$ identity matrix,
\begin{eqnarray}\label{N-SS details}
\bra{\mathbf{\Psi}(t)} = \bra{c_1 e^{-i\xi_1
t},...,c_Ne^{-i\xi_1 t}}\,,
\\\nonumber
\bra{\mathbf{\Phi}(t)} = \bra{e^{-i\xi_1 t},...,e^{-i\xi_1
t}}\,,
\\\nonumber
\widehat{\mathbf{M}}_{k,j}(t) = c_j \frac{e^{i (\xi^*_k-\xi_j)
t}}{\xi^*_k-\xi_j}\,,
\end{eqnarray}
and parameters $c_k$ are defined in~\cite{PAPER}.
To the best of our knowledge all the existing discrete spectrum numerical IST algorithms are unstable at large $N$, that can be understood by looking at the exact N-SS formulae~(\ref{N-SS}),(\ref{N-SS details}). Indeed, the eigenvalues $\xi_k$ are complex and, thus, the matrix $\widehat{\mathbf{E}} +\widehat{\mathbf{M}}(t)^*\widehat{\mathbf{M}}(t)$ in~(\ref{N-SS}) may become ill-conditioned at large $|t|$. In such cases we use the arbitrary precision arithmetics to obtain accurate N-SS signal. Recently, A.A.Gelash and D.S. Agafontsev found that numerical realisation of the Zakharov-Shabat dressing method can be stably used up to $N \sim 32$ soliton solutions~\cite{gelash2016uniform}. Application of the dressing method to our kernel-based approach is a nontrivial task, however we believe that this can be an interesting direction for future research.
\section{Parametric kernel decoding}
In this paragraph we discuss the N-SS kernel general parametric encoding/decoding schemes involving $4\times N$ coding parameters.
Let us write the N-SS kernel (formula~(3) in~\cite{PAPER}) as a time series on the discrete grid (see formula.~(7) in~\cite{PAPER}):
\begin{eqnarray}\label{kernel sampling}
\Omega_m \equiv \Omega(t_m) =
\\\nonumber
=\sum_{k=1}^{N} c_k e^{-i\xi_k t_m} = \sum_{k=1}^{N} c_k e^{-i\xi_k T(m-1)} = \sum_{k=1}^{N} c_k z_k^{m-1}\,.
\end{eqnarray}
Parameters $z_k=\exp(-i\xi_k T)$ in~(\ref{kernel sampling}) are defined by the soliton eigenvalues $\xi_k$ and by the value of time slot $T$. Here, we again choose the minimum possible number of time samples $M=N$. Then, for the decoding problem we obtain system of equations with the Vandermonde matrix:
\begin{equation}\label{vandermonde}
\left(\begin{array}{cccc}z_1^0, & z_2^0, & ..., & z_N^0 \\z_1^1, & z_2^1, & ..., & z_N^1 \\... & ... & ... & ... \\z_1^{N-1}, & z_2^{N-1}, & ..., & z_N^{N-1},\end{array}\right)
\left(\begin{array}{c}c_1 \\c_2 \\... \\c_N\end{array}\right)=
\left(\begin{array}{c}\Omega_1 \\\Omega_2 \\... \\\Omega_N\end{array}\right)\,.
\end{equation}
Now we consider both \textit{position-phase} modulation and \textit{amplitude-frequency} modulation of the N-SS kernel and discuss numerical problems that occur in general case.
\subsection{Position-phase modulation}
Suppose we know the eigenvalues $\xi_k$ and hence the parameters
$z_k = \exp(-i\xi_k)$. The decoding problem is to find the
parameters $c_k$ by the measured kernel samples $\Omega_i$, that
can be done by solving system~(\ref{vandermonde}). However, the
straightforward numerical algorithm based, for example, on Gauss
elimination in a general case is extremely challenging since the Vandermonde matrix~(\ref{vandermonde})
exponentially fast becomes ill-conditioned with the increase of
$N$~\cite{tyrtyshnikov1994bad}. On the other hand, the Vandermonde
matrix belongs to the class of structured matrices for which the
effective numerical algorithms have been
developed~\cite{blahut2010fast}. By applying the effective
matrix inversion algorithm the kernel decoding can be
performed using $N^2$ operations, however, the numerical stability
restricts $N$ by around $\sim 60$ harmonics (see, for example~\cite{gohberg1997fast}).
The inversion of the Vandermonde matrix becomes numerically stable at any $N$ when $z_k$ are the
complex $k$-th roots of unity. For the N-SS kernel this is possible only when $\xi_k$ have
identical imaginary parts (that can be moved to the right part of the matrix system~(\ref{vandermonde})), i.e. in the case presented by formula (6) in~\cite{PAPER}.
The additional harmonics orthogonality condition (formula.~(8) in~\cite{PAPER}) allows us to use the FFT/IFFT algorithms instead of matrix inversion operations, that motivated us to focus on this elegant encoding scheme~\cite{PAPER}.
\subsection{Amplitude-frequency modulation}
Another possibility is to use the eigenvalues $\xi_k$ as the carriers of
information. They have to be found from the measured kernel samples
$\Omega_m$, while the shift-phase parameters $c_k$ are all known and are not used for coding of information. The
parametric approach based on the Prony's method (see, for
example~\cite{marple1987digital}, chapter 11) uses the following
\textit{master polynomial}
\begin{equation}\label{polynomial}
\phi(z) = \prod_{n=1}^N (z-z_k)^n = \sum_{n=0}^N a_n z^n, \,\,\,\,\,\, a_0=1\,,
\end{equation}
with the complex roots $z_k$. The coefficients $a_n$ of the
polynomial~(\ref{polynomial}) can be determined by solving the Toeplitz
system of equations:
\begin{equation}\label{toeplitz}
\left(\begin{array}{cccc}\Omega_N, & \Omega_{N-1}, & ..., & \Omega_1 \\ \Omega_{N+1}, & \Omega_N, & ..., & \Omega_2 \\... & ... & ... & ... \\ \Omega_{2N-1}, & \Omega_{2N-2}, & ..., & \Omega_N,\end{array}\right)
\left(\begin{array}{c}a_1 \\a_2 \\... \\a_N\end{array}\right)=
\left(\begin{array}{c}\Omega_{N+1} \\\Omega_{N+2} \\... \\\Omega_{2N}\end{array}\right)\,.
\end{equation}
Numerical solution of the problem~(\ref{toeplitz}) can be obtained
by the use of Levinson-Durbin-Trench algorithm through the
$O(N^2)$ arithmetic operations~\cite{blahut2010fast}. However, the
subsequent roots finding of the master polynomial $\phi(z)$ is the
hard numerical problem for the large number of samples $N$. For
example, the well known factorization algorithm of Jenkins and
Traub becomes numerically unstable at $N \sim 100$
\cite{jenkins1972algorithm}.
We note, that in the case of continuous spectrum
kernel (formula (10) in~\cite{PAPER}) the corresponding Vandermonde matrix
can be always stably inverted since it becomes Fourier matrix.
We conclude that the general (parametric) N-SS kernel decoding requires matrix inversion and/or finding roots of the polynomial in the decoder. Although, the advanced numerical algorithms with a relatively small number of operations $\sim N^2$ can be exploited, their stability against large number of harmonics and additive noise requires a separate comprehensive analysis that is beyond the scope of this Letter.
\bibliographystyle{apsrev4-1}
|
\section{Introduction and background}
\subsection{Background}
Network coding technique has significantly improved the performance of communication networks, and has been studied extensively in the past two decades. Index coding problem (ICP) can be considered as a special case of network coding problem \cite{Ahl}. ICP has emerged as an important topic of recent research due to its applications in many of the practically relevant problems including that in satellite networks, topological interference management, wireless caching and cache enabled cloud radio access networks for 5G cellular systems.
\par The noiseless index coding problem with side information was first studied in \cite{Birk} as an Informed-Source Coding-On-Demand (ISCOD) problem, in which a central server (sender) wants to broadcast data blocks to a set of clients (receivers) which already has a proper subset of the data blocks. The problem is to minimize the data that must be broadcast, so that each receiver can derive its required data blocks. Consider the case of a sender with $n$ messages denoted by the set $\mathcal{X} = \{x_1,x_2,...,x_n\}$, $x_i ~\in~ \mathbb{F}_{q}$, $\mathbb{F}_{q}$ is a field with $q$ elements, which it broadcasts as coded messages, to a set of $m$ receivers, $\mathcal{R}=\{R_1,R_2,...,R_m\}$. Each receiver $R_i ~\in~\mathcal{R}$ wants a subset $\mathcal{W}_{i}$ of the messages, knows a priori a proper subset $\mathcal{K}_{i}$ of the messages, where $\mathcal{W}_{i}~\cap~\mathcal{K}_{i}=\phi$, and is identified by the pair $(\mathcal{W}_{i},\mathcal{K}_{i})$. The noiseless index coding problem is to find the smallest number of transmissions required and is specified by $(\mathcal{X},\mathcal{R})$. The set $\mathcal{K}_{i}$ is referred to as the side information available to the receiver $R_i$.
\begin{defn}
An index code (IC) for a given ICP $(\mathcal{X},\mathcal{R})$ is defined by an encoding function,
$g: \mathbb{F}^{n}_{q} \rightarrow \mathbb{F}^{l}_{q}, $
and a set of $m$ decoding functions
$\mathcal{D}_i: \mathbb{F}^{l}_{q} \times \mathbb{F}^{|\mathcal{K}_{i}|}_{q} \rightarrow \mathbb{F}^{|\mathcal{W}_{i}|}_{q}, $ $\forall~ i~ \in~ \{1,2,...,m\}$
corresponding to the $m$ receivers, such that,
\begin{equation*}
\mathcal{D}_{i}(g(\textbf{x}),\mathcal{K}_{i}) = \mathcal{W}_{i},~~~\forall~ \textbf{x}~ \in~ \mathbb{F}^{n}_{q},~~~\forall~i~ \in~ \{1,2,...,m\}.
\end{equation*}
\end{defn}
\par In this paper, we consider ICP over binary field ($q=2$). The integer $l$, as defined above is called the length of the index code. For noiseless broadcast channels, an index code of minimum length is called an optimal index code \cite{Ong}, \cite{Yossef}.
Even though this is interesting theoretically, index coding over noisy channels is more practical. Noisy index coding over a binary symmetric channel was considered in \cite{Anp1}, \cite{Anp2}. Binary transmission of index coded bits were assumed.
For this set up the problem of identifying the number of optimal index codes possible for a given ICP is important and that was studied in \cite{Kav1}, \cite{Kav2}.
\par A special case of ICP over Gaussian broadcast channel, based on multidimensional QAM constellation with $2^n$ points, where every receiver demands all messages (which it does not have) from the source, was considered in \cite{Lpn}.
The case of noisy index coding over AWGN broadcast channel, along with minimum Euclidean distance decoding, was studied in \cite{Anj1}, where the receivers demand a subset of messages as defined in \cite{Birk}.
An algorithm to map the broadcast vectors to PSK signal points so that the receiver with maximum side information gets maximum PSK side information coding gain, was also proposed. The algorithm assumes that an index code is given and is applicable only for one specific order of priority (in the non increasing order of amount of side information) among the receivers. Minimum Euclidean distance of the effective broadcast signal set seen by a receiver, was considered as the basic parameter which decides the message error probability of the receiver and the proposed algorithm tries to maximize the minimum Euclidean distance.
\par In this paper, we discuss the maximum likelihood (ML) decoder for index coded PSK modulation. Further, we study the case in which the length of the index code is specified for the ICP but not necessarily the index code. The receivers can have any a priori defined arbitrary order of priority among themselves. For a chosen priority order, we consider all possible index codes, to obtain the mappings to appropriate PSK constellation which will result in the best message error performance in terms of PSK index coding gain (PSK-ICG, defined in Section \ref{isd_section}) of the receivers, respecting the defined order of priority.
\subsection{Our Contribution}
Consider a noisy index coding problem with $n$ messages, over $\mathbb{F}_2$ which uses an AWGN broadcast channel for transmission. For the ICP $(\mathcal{X},\mathcal{R})$, consider index codes of length $N$, $N<n$, which will generate $2^N$ broadcast vectors (elements of $\mathbb{F}^N_2$). The broadcast vectors are mapped to $2^N$-PSK signal points, so that $2^N$-PSK modulation can be used, to minimize the bandwidth requirement. Note that, transmitting one $2^N$-PSK signal point instead of $N$ binary bits (as in noiseless index coding), results in $N/2$ fold saving in bandwidth.
\par Our contributions are summarized below:
\begin{itemize}
\item We derive a decision rule for maximum likelihood decoding which gives the best message error performance, for any receiver $R_i$, for a given index code and mapping.
\item We show that, at very high SNR, the message error performance of the receiver employing ML decoder, depends on the minimum inter-set distance (defined in Section \ref{isd_section}). The mapping which maximized the minimum inter-set distance is optimal for the best message error performance at high SNR.
\item For the ICP $(\mathcal{X},\mathcal{R})$, when the receivers are arranged in the decreasing order of priority, we propose an algorithm to find (index code, mapping) pairs, each of which gives the best message error performance for the receivers, for the given order of priority. Using any one of the above (index code, mapping) pairs, the highest priority receiver achieves the maximum possible gain (PSK-ICG) that it can get using any IC and any mapping for $2^N$-PSK constellation, at very high SNR. Given that the highest priority receiver achieves its best performance, the next highest priority receiver achieves its maximum gain possible and so on in the specified order of priority.
\end{itemize}
\section{Preliminaries and notation}
\par Let $[n] \triangleq \{1,2,...,n\}$. For a vector $\mathbf{z}=(z_1~z_2~...z_n) \in \mathbb{F}^n_{2}$ and a subset $B=\{i_1,i_2,...,i_b\}$ of $[n]$ (for any integer $b, 1 \leq b \leq n$), where $i_1<i_2<...<i_b$, $\mathbf{z}_B$
denote the vector $(z_{i_1}~z_{i_2}~...~z_{i_b})$.
\par We consider the noisy index coding problem over $\mathbb{F}_2$ with a single sender having a set of messages $\mathcal{X} = \{x_1,x_2,...,x_n\}$, $x_i\in\mathbb{F}_{2}$, and a set of $m$ receivers, $\mathcal{R}=\{R_1,R_2,...,R_m\}$, where each receiver $R_i$ is identified by $(\mathcal{W}_{i},\mathcal{K}_{i})$, the want set and the known set. Let, $\mathcal{I}_{i} \triangleq \{j : x_j \in \mathcal{K}_i\}$ be the set of indices corresponding to the known set. It is sufficient to consider the case where each receiver demands only one message. If there is a receiver which demands more than one message, it can be considered as $|\mathcal{W}_{i}|$ equivalent receivers each demanding one message and having the same side information. Each $R_i$, $i \in [m]$ wants the message $x_{f(i)}$, where $f:[m] \rightarrow [n]$ and $x_{f(i)} \notin \mathcal{K}_{i}$, $\forall i \in [m]$.
\par For the given ICP, we consider scalar linear index codes of length $N$ (not necessarily the minimum or optimum length), such that the set of all broadcast vectors gives $\mathbb{F}_2^N.$ Let $L$ be an $n~\times~N$ encoding matrix for one such index code, $\mathcal{C}$. Let $\mathbf{x}=(x_1~x_2~ ...~x_n)$ and $\mathbf{y} =(y_1~y_2~...y_N)$ denote the message vector and the broadcast vector respectively, where $\mathbf{y} = \mathbf{x} L.$
\begin{exmp} \label{exmp:eg1}
Consider the following ICP with $n=m=5$ and $\mathcal{W}_{i}=x_i, ~\forall i \in \{1,2,...,5\}$. The side information available with the receivers is as follows: $\mathcal{K}_1= \{x_2, x_3\}, ~\mathcal{K}_2= \{x_3, x_4, x_5\},~ \mathcal{K}_3= \{x_2, x_4, x_5\},~\mathcal{K}_4= \{x_5\},~\mathcal{K}_5= \{x_4\}$.
\par For this ICP we can choose a scalar linear index code of length $N=3$, as given by the following encoding matrix $L$.
\[
L=
\begin{bmatrix}
1 & 1 & 0 \\
0 & 1 & 0 \\
0 & 1 & 0 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix}
\]. \\
The index coded bits are given by \\
$(y_1~y_2~y_3)=(x_1~x_2~x_3~x_4~x_5)L$ as $y_1=x_1+x_4+x_5,~ y_2=x_1+x_2+x_3+x_4+x_5,~ y_3=x_4+x_5$.
\end{exmp}
\par Instead of using $N$ BPSK transmissions, the $N$ index coded bits of $\mathbf{y}$ are sent as a signal point from a $2^N$-PSK signal set, over an AWGN channel, to save bandwidth \cite{Anj1}. In this paper we consider index coded $2^N$-PSK modulation for a chosen $N$ and so when we refer to index codes of length $N$, we consider only those index codes for which the set of all broadcast vectors is $\mathbb{F}^N_{2}$. Let the chosen $2^N$-PSK signal set be denoted as $\mathcal{S}=\{\mathbf{s_{1}},\mathbf{s_{2}},...,\mathbf{s_{2^N}}\}$. Assume that for the index code $\mathcal{C}$ a mapping scheme specifies the mapping of $\mathbb{F}^N_{2}$ to the signal set $\mathcal{S}$. All receivers are assumed to know the encoding matrix, $L$ for the index code $\mathcal{C}$.
\par Let $\mathbf{a}_i \in \mathbb{F}^{|\mathcal{K}_i|}_{2}$ be a realization of $\mathbf{x}_{\mathcal{I}_{i}}$. As each receiver $R_i$ knows some messages (from its side information), $R_i$ needs to consider only a subset of $\mathbb{F}^N_{2}$ for decoding and this subset is called the {\it effective broadcast vector set}.
\begin{defn}
For a chosen index code based on the encoding matrix $L$, the {\it effective broadcast vector set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ is defined by,
\begin{equation*}
\begin{aligned}
\mathcal{C}_L(\mathbf{a}_i) &\triangleq \{\mathbf{y}\in\mathbb{F}^N_2 : \mathbf{y}= \mathbf{x} L, \mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i, x_j \in \mathbb{F}_2, j \in [n] \setminus \mathcal{I}_{i}\}.
\end{aligned}
\end{equation*}
\end{defn}
The corresponding set of signal points in $2^N$-PSK constellation is referred to as the {\it effective broadcast signal set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ and is denoted by $\mathcal{S}_L(\mathbf{a}_i)$. For a chosen index code, all effective broadcast signal sets and effective broadcast vector sets seen by $R_i$ are of the same size ($|\mathcal{S}_L(\mathbf{a}_i)|=|\mathcal{S}_L(\mathbf{a}'_i)|=|\mathcal{C}_L(\mathbf{a}_i)|=|\mathcal{C}_L(\mathbf{a}'_i)|$ where $\mathbf{a}_i,\mathbf{a}'_i \in \mathbb{F}^{|\mathcal{K}_i|}_{2}$).
Half the number of broadcast vectors in an effective broadcast vector set corresponds to $x_{f(i)}=0$ and the remaining half corresponds to $x_{f(i)}=1$. So, we can partition an effective broadcast vector set into two subsets as defined below.
\begin{defn}
The {\it 0-effective broadcast vector set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ is defined by,
\begin{equation*}
\begin{aligned}
\mathcal{C}_{L0}(\mathbf{a}_i) \triangleq \{ &\mathbf{y}~\in~\mathbb{F}^N_2 : \mathbf{y}= \mathbf{x} L, \mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i, x_{f(i)}=0, x_j \in \mathbb{F}_2, \\ &j \in [n] \setminus (\mathcal{I}_{i} \cup \{f(i)\})\}.
\end{aligned}
\end{equation*}
\end{defn}
The corresponding set of signal points in $2^N$-PSK constellation is referred to as the {\it 0-effective broadcast signal set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ and is denoted as $\mathcal{S}_{L0}(\mathbf{a}_i)$.
\begin{defn}
The {\it 1-effective broadcast vector set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ is defined by,
\begin{equation*}
\begin{aligned}
\mathcal{C}_{L1}(\mathbf{a}_i) \triangleq \{ &\mathbf{y}~\in~\mathbb{F}^N_2 : \mathbf{y}= \mathbf{x} L, \mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i, x_{f(i)}=1, x_j \in \mathbb{F}_2, \\ &j \in [n] \setminus (\mathcal{I}_{i} \cup \{f(i)\})\}.
\end{aligned}
\end{equation*}
\end{defn}
The corresponding set of signal points in $2^N$-PSK constellation is referred to as the {\it 1-effective broadcast signal set} seen by $R_i$ for $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$ and is denoted as $\mathcal{S}_{L1}(\mathbf{a}_i)$.
The effective broadcast vector sets, 0-effective broadcast vector sets and 1-effective broadcast vector sets seen by $R_2$ for the IC in Example \ref{exmp:eg1} is given in Table \ref{table:tbl1}. It is clear that, two different realizations of $\mathbf{x}_{\mathcal{I}_{i}}$ may have the same effective broadcast vector set. However, 1-effective broadcast vector set for a particular realization of $\mathbf{x}_{\mathcal{I}_{i}}$ may become the 0-effective broadcast vector set of another realization of $\mathbf{x}_{\mathcal{I}_{i}}$ and vice versa. But the way in which the effective broadcast vector set gets partitioned will be the same. For example consider the case of $\mathcal{C}_L(011)$ and $\mathcal{C}_L(100)$ in Table \ref{table:tbl1}.
\begin{table}[h]
\caption{Effective broadcast vector sets and its partitions (seen by $R_2$) for the IC in Example \ref{exmp:eg1}.}
\centering
\begin{tabular}{p{0.5cm}|p{3.4cm}|p{1.6cm}|p{1.6cm}}
\hline
\hline
~~$\mathbf{a}_2$ & ~~~~~~~~~~~$\mathcal{C}_L(\mathbf{a}_2)$ & ~~~~$\mathcal{C}_{L0}(\mathbf{a}_2)$ & ~~~~$\mathcal{C}_{L1}(\mathbf{a}_2)$\\
\hline
$(000)$ & $\{(000),(010),(110),(100)\}$ & $\{(000),(110)\}$ & $\{(010),(100)\}$ \\
$(001)$ & $\{(111),(101),(001),(011)\}$ & $\{(111),(001)\}$ & $\{(101),(011)\}$ \\
$(010)$ & $\{(111),(101),(001),(011)\}$ & $\{(111),(001)\}$ & $\{(101),(011)\}$ \\
$(011)$ & $\{(000),(010),(110),(100)\}$ & $\{(000),(110)\}$ & $\{(010),(100)\}$ \\
$(100)$ & $\{(000),(010),(110),(100)\}$ & $\{(010),(100)\}$ & $\{(000),(110)\}$ \\
$(101)$ & $\{(111),(101),(001),(011)\}$ & $\{(101),(011)\}$ & $\{(111),(001)\}$ \\
$(110)$ & $\{(111),(101),(001),(011)\}$ & $\{(101),(011)\}$ & $\{(111),(001)\}$ \\
$(111)$ & $\{(000),(010),(110),(100)\}$ & $\{(010),(100)\}$ & $\{(000),(110)\}$ \\ [1ex]
\hline
\end{tabular}
\label{table:tbl1}
\end{table}
\begin{exmp} \label{exmp:eg2}
Consider the following ICP with $n=m=6$ and $\mathcal{W}_{i}=x_i, ~\forall i \in \{1,2,...,6\}$. The side information available with the receivers is as follows: $\mathcal{K}_1= \{x_2, x_3, x_4, x_5, x_6\}, ~\mathcal{K}_2= \{x_1, x_3, x_4, x_5\},~ \mathcal{K}_3= \{x_2, x_4, x_6\},~\mathcal{K}_4= \{x_1, x_6\},~\mathcal{K}_5= \{x_3\},~\mathcal{K}_6= \{\} $.
\par For this ICP we can choose a scalar linear index code of length $N=4$, based on encoding matrix $L$, with $y_1=x_1+x_4,~ y_2=x_2+x_3,~ y_3=x_5,~y_4=x_6$. \\
Then, the effective broadcast vector sets of $R_2$ for four different realization of $\mathbf{x}_{\mathcal{I}_{2}}$ are as given below.
\begin{itemize}
\item $\mathcal{C}_L(0000)=\{(0000),(0100),(0001),(0101)\}$
\item $\mathcal{C}_L(0001)=\{(0010),(0110),(0011),(0111)\}$
\item $\mathcal{C}_L(0010)=\{(1000),(1100),(1001),(1101)\}$
\item $\mathcal{C}_L(0011)=\{(1010),(1110),(1011),(1111)\}$.
\end{itemize}
\end{exmp}
Suppose an IC based on an encoding matrix $L$, and an effective broadcast vector set, $\mathcal{C}_L(\mathbf{a}_i)$ of $R_i$ are given. $\mathcal{C}_L(\mathbf{a}_i)$ can be partitioned into 0-effective broadcast vector set and 1-effective broadcast vector set as follows:
\begin{itemize}
\item Identify an $\mathbf{x}$ such that $\mathbf{x} L \in \mathcal{C}_L(\mathbf{a}_i)$. Let the corresponding realization of $\mathbf{x}_{\mathcal{I}_{i}}$ be $\mathbf{a}_i$ .
\item For $\mathbf{a}_i$, partition $\mathcal{C}_L(\mathbf{a}_i)$ into $\mathcal{C}_{L0}(\mathbf{a}_i)$ and $\mathcal{C}_{L1}(\mathbf{a}_i)$
\end{itemize}
The partitioning of $\mathcal{C}_L(\mathbf{a}_i)$ is illustrated in the following example. Consider the ICP given in Example \ref{exmp:eg2}. Suppose the effective broadcast vector set, $\mathcal{C}_L(\mathbf{a}_2)=\{(0000),(0100),(0001),(0101)\}$ of $R_2$ needs to be partitioned into $\mathcal{C}_{L0}(\mathbf{a}_2)$ and $\mathcal{C}_{L1}(\mathbf{a}_2)$. Choose $\mathbf{x}=(110100)$ such that $\mathbf{y}= \mathbf{x} L=(0100) \in \mathcal{C}_L(\mathbf{a}_2)$, and then $\mathbf{a}_2=(1010)$. Note that $y_2=x_2+x_3$, $x_3 \in \mathcal{K}_{2}$, $R_2$ wants $x_2$, and from $\mathbf{a}_2=(1010)$, $x_3=0$. So $y_2=x_2$ and only two broadcast vectors, $(0000)$ and $(0001)$ in $\mathcal{C}_L(1010)$ has $y_2=0$. So $\mathcal{C}_{L0}(1010)=\{(0000),(0001)\}$. Similarly, $\mathcal{C}_{L1}(1010)=\{(0100),(0101)\}$. It should be noted that for some other choice of $\mathbf{x}$ with $x_3=1$, we may get $\mathcal{C}_{L0}(\mathbf{a}_2)=\{(0100),(0101)\}$ and $\mathcal{C}_{L1}(\mathbf{a}_2)=\{(0000),(0001)\}$. We are only interested in partitioning the effective broadcast vector set into two subsets such that all broadcast vectors in each subset correspond to the same value of $x_{f(i)}$.
\begin{figure*}
\begin{center}
\includegraphics[scale=1]{isd1.eps}
\caption{8-PSK mapping and inter-set distance for $R_1$ in Example \ref{exmp:eg1}.}
\label{figure:fig1}
\end{center}
\end{figure*}
\section{Maximum likelihood decoder}
In this section we derive a decision rule for the maximum likelihood decoder for the receiver $R_i$. We follow an approach similar to the one used in \cite{Vit}.
\par Let $\mathcal{M}$ be the map from $\mathbb{F}^N_{2}$ to the signal set $\mathcal{S}$. The received vector $\mathbf{r}$ is given by
\begin{equation*}
\mathbf{r}=\mathcal{M}(\mathbf{x} L)+\mathbf{w}
\end{equation*}
where $\mathbf{w}=(w_1 ~w_2)$; $w_1$ and $w_2$ are independent Gaussian variables with zero mean and variance $N_0/2$. The conditional probability density of $\mathbf{r}$ given that $\mathcal{M}(\mathbf{x} L)$ is transmitted (likelihood function) is
\begin{multline}
p(\mathbf{r}|\mathcal{M}(\mathbf{x} L)) = \frac{1}{(\pi N_0)}\exp{\left(-\frac{\|\mathbf{r}-\mathcal{M}(\mathbf{x} L)\|^2}{N_0}\right)}. \label{eqawgn}
\end{multline}
Consider the decoder for a receiver $R_i$. The minimum error probability decoder should make a decision $x'_{f(i)}$ on the desired message $x_{f(i)}$ based on the received vector $\mathbf{r}$ and the side information $\mathbf{x}_{\mathcal{I}_{i}}$, minimizing the probability of error. Given $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$, when $x_{f(i)}=0$ the probability of error in this decision is $\textbf{Pr}(x_{f(i)}=1|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r})$ and that when $x_{f(i)}=1$ is $\textbf{Pr}(x_{f(i)}=0|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r})$.
To minimize the error probability, the decision $x'_{f(i)}=0$ is taken if
\begin{equation}
\begin{aligned}
\textbf{Pr}(x_{f(i)}=0|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}) \geq \textbf{Pr}(x_{f(i)}=1|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}) \label{eq1}
\end{aligned}
\end{equation}
and the decision $x'_{f(i)}=1$ is taken if
\begin{equation}
\begin{aligned}
\textbf{Pr}(x_{f(i)}=0|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}) < \textbf{Pr}(x_{f(i)}=1|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}). \label{eq2}
\end{aligned}
\end{equation}
Combining (\ref{eq1}) and (\ref{eq2}), and ignoring ties, the decision rule can be written as
\begin{equation}
\begin{aligned}
\textbf{Pr}(x_{f(i)}=0|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}) \underset{0}{\overset{1}\lessgtr} \textbf{Pr}(x_{f(i)}=1|\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i,\mathbf{r}). \label{eq3}
\end{aligned}
\end{equation}
Using Bayes rule in (\ref{eq3}), we obtain the decision rule in terms of the likelihood functions as
\begin{multline*}
\frac{p(\mathbf{r}|x_{f(i)}=0,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i)\textbf{Pr}(x_{f(i)}=0)}{p(\mathbf{r})} \underset{0}{\overset{1}\lessgtr} \\ \frac{p(\mathbf{r}|x_{f(i)}=1,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i)\textbf{Pr}(x_{f(i)}=1)}{p(\mathbf{r})},
\end{multline*}
which implies
\begin{multline}
p(\mathbf{r}|x_{f(i)}=0,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i)\textbf{Pr}(x_{f(i)}=0) \underset{0}{\overset{1}\lessgtr} \\
p(\mathbf{r}|x_{f(i)}=1,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i)\textbf{Pr}(x_{f(i)}=1). \label{eq5}
\end{multline}
$\mathcal{S}_{L0}(\mathbf{a}_i)$, the 0-effective broadcast signal set seen by $R_i$ (for $\mathbf{a}_i$), is the set of all signal points corresponding to broadcast vectors with $x_{f(i)}=0$ and $\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i$. Therefore,
\begin{equation}
\begin{aligned}
p(\mathbf{r}|x_{f(i)}=0,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i) &= p(\mathbf{r}|\mathcal{S}_{L0}(\mathbf{a}_i)). \label{eqeff0}
\end{aligned}
\end{equation}
Similarly,
\begin{equation}
\begin{aligned}
p(\mathbf{r}|x_{f(i)}=1,\mathbf{x}_{\mathcal{I}_{i}}=\mathbf{a}_i) &= p(\mathbf{r}|\mathcal{S}_{L1}(\mathbf{a}_i)). \label{eqeff1}
\end{aligned}
\end{equation}
Assuming that all the messages take values 0 or 1 with equal probability, from (\ref{eq5}), (\ref{eqeff0}) and (\ref{eqeff1}) we obtain the decision rule as
\begin{equation}
\sum_{k:s_k \in \mathcal{S}_{L0}(\mathbf{a}_i) } p(\mathbf{r}|s_k) \underset{0}{\overset{1}\lessgtr} \sum_{k:s_k \in \mathcal{S}_{L1}(\mathbf{a}_i) } p(\mathbf{r}|s_k). \label{eqml2}
\end{equation}
From (\ref{eqawgn}) and (\ref{eqml2}),
\begin{multline*}
\sum_{k:s_k \in \mathcal{S}_{L0}(\mathbf{a}_i)} {\left(\frac{1}{(\pi N_0)}\exp{\left(-\frac{\|\mathbf{r}-s_k\|^2}{N_0}\right)}\right)}
\underset{0}{\overset{1}\lessgtr} \\
\sum_{k:s_k \in \mathcal{S}_{L1}(\mathbf{a}_i)} {\left(\frac{1}{(\pi N_0)}\exp{\left(-\frac{\|\mathbf{r}-s_k\|^2}{N_0}\right)}\right)}.
\end{multline*}
Thus we obtain the ML decision rule as,
\begin{multline}
\sum_{k:s_k \in \mathcal{S}_{L0}(\mathbf{a}_i)}{\left( \exp{\left(-\frac{\|\mathbf{r}-s_k\|^2}{N_0}\right)}\right)}
\underset{0}{\overset{1}\lessgtr} \\
\sum_{k:s_k \in \mathcal{S}_{L1}(\mathbf{a}_i)}{\left( \exp{\left(-\frac{\|\mathbf{r}-s_k\|^2}{N_0}\right)}\right)}. \label{eqml3}
\end{multline}
It is clear that the ML decoder decision is based on the Euclidean distance of all signal points in 0-effective broadcast signal set to the received vector $\mathbf{r}$ relative to that of the signal points in 1-effective broadcast signal set. This indicates that, to reduce the message error probability, the signal points in 0-effective broadcast signal set and 1-effective broadcast signal set must be as separated as possible in terms of Euclidean distance.
\section{Inter-set distance and PSK Index Coding Gain} \label{isd_section}
\begin{defn}
{\it Inter-set distance} of an effective broadcast signal set seen by a receiver $R_i$ is the minimum among the Euclidean distances between a signal point in the 0-effective broadcast signal set and a signal point in the 1-effective broadcast signal set.
\begin{equation*}
\begin{aligned}
d_{IS}(\mathcal{S}_L(\mathbf{a}_i)) \triangleq \min\{d(\mathbf{s_a},\mathbf{s_b}):&\mathbf{s_a} \in \mathcal{S}_{L0}(\mathbf{a}_i), \mathbf{s_b} \in \mathcal{S}_{L1}(\mathbf{a}_i)\}
\end{aligned}
\end{equation*}
where $d(\mathbf{s_a},\mathbf{s_b})$ denotes the Euclidean distance between $2^N$-PSK signal points, $\mathbf{s_a}$ and $\mathbf{s_b}$.
\end{defn}
A labeled 8-PSK constellation which can be used for the ICP discussed in Example \ref{exmp:eg1} is shown in Fig. \ref{figure:fig1}(a) and the inter-set distance of the effective broadcast signal set seen by $R_1$ for $\mathbf{x}_{\mathcal{I}_{i}}=(00)$, $\mathcal{S}_L(00)$ is shown in Fig. \ref{figure:fig1}(b). For this example, $\mathcal{S}_L(00)=\{\mathbf{s_1},\mathbf{s_2},\mathbf{s_5},\mathbf{s_{6}}\}$, $\mathcal{S}_{L0}(00)=\{\mathbf{s_1},\mathbf{s_{2}}\}$ and $\mathcal{S}_{L1}(00)=\{\mathbf{s_5},\mathbf{s_6}\}$.
\begin{defn}
For a given index code and mapping, the minimum inter-set distance for a receiver $R_i$, denoted by $d^{(i)}_{IS,min}$, is defined as the minimum of the inter-set distances among all the effective broadcast signal sets seen by $R_i$.
\begin{equation*}
d^{(i)}_{IS,min} \triangleq \min\{d_{IS}(\mathcal{S}_L(\mathbf{a}_i)):\mathbf{a}_i~\in~\mathbb{F}^{|\mathcal{K}_i|}_2\}
\end{equation*}
\end{defn}
In the case of Example \ref{exmp:eg1}, the minimum inter-set distance for $R_1$ is shown in Fig. \ref{figure:fig1}(b) and \ref{figure:fig1}(c) for two different mappings. Clearly, the mapping shown in Fig. \ref{figure:fig1}(b) has a larger minimum inter-set distance for $R_1$.
\par In (\ref{eqml3}), the term with the signal point closest to $\mathbf{r}$ is dominant in the summations at very high SNR. The decoder makes an error if the broadcasted signal point is in 0-effective broadcast signal set but $\mathbf{r}$ is closest to a signal point in 1-effective broadcast signal set or vice versa. The probability of this event is more when the minimum inter-set distance is less. At high SNR, this error is dominant and so an optimal mapping for the best message error performance must maximize the minimum inter-set distance. Among the mappings which has the same minimum inter-set distance, the one which has more second minimum inter-set distance will perform better and so on.
\begin{defn}
The PSK Index Coding Gain (PSK-ICG) of a receiver $R_i$, for a given IC and mapping is defined as
\begin{equation*}
g_{i} \triangleq 20~log\left(\frac{d^{(i)}_{IS,min}}{d_{min,n}}\right)
\end{equation*}
where $d^{(i)}_{IS,min}$ is the minimum inter-set distance for $R_i$ and $d_{min,n}$ is the minimum Euclidean distance between any two signal points in a $2^n$-PSK constellation.
\end{defn}
\section{Mapping based on inter-set distances}
For mapping, it is more appropriate to consider the minimum inter-set distance than to consider the minimum Euclidean distance of the effective broadcast signal sets. For example, consider the mappings given in Fig. \ref{figure:fig1}(b) and Fig. \ref{figure:fig1}(c). With the mapping shown in Fig. \ref{figure:fig1}(b), the minimum inter-set distance for $R_1$ is more but the minimum Euclidean distance of its effective broadcast signal sets is less, compared to that with the mapping shown in Fig. \ref{figure:fig1}(c). The simulation results (discussed in Section \ref{sim_section}) show that $R_1$ performs better with the mapping in Fig. \ref{figure:fig1}(b) than with the mapping in Fig. \ref{figure:fig1}(c).
\par For the given ICP and $2^N$-PSK constellation, when the receivers are arranged in the decreasing order of priority, we propose an algorithm which maximizes the minimum inter-set distance, to find (index code, mapping) pairs, each of which gives the optimal message error performance for the receivers, for the given order of priority. Assume that the decreasing order of priority for the receivers is $(R_1,R_2,...,R_m)$. Here optimality is based on minimum inter-set distance and is in the following sense:\\
\begin{itemize}
\item No other mapping of $2^N$-PSK constellation for any index code, can give PSK-ICG $>g_{1}$ for $R_1$.
\item Any mapping for any index code which gives the PSK-ICG $g_i$ for receiver $R_i$, $i \in \{1,2,...,j-1\}$ cannot give a PSK-ICG $>g_j$ for $R_j$, $j \leq m$.
\end{itemize}
It may so turn out that maximizing the gain of a receiver $R_i$, minimizes the gain that can be achieved by a lower priority receiver $R_j$. With this mapping it is not necessary that a higher priority receiver will get higher PSK-ICG compared to that of the lower priority receivers. The PSK-ICG achieved by a receiver $R_j$ depends on its priority, $\mathcal{W}_j$, $\mathcal{K}_j$, $\mathcal{W}_i$ and $\mathcal{K}_i$ $\forall i$ such that $R_i$ is a higher priority receiver than $R_j$.
In the following subsections, we explain the mapping algorithm and then illustrate it with examples.
\subsection{Mapping Algorithm}
Without loss of generality, assume that the decreasing order of priority among the receivers is $(R_1,R_2,...,R_m)$.
For a given index code based on encoding matrix $L$, optimal mapping for a receiver $R_i$ is obtained as follows:
\begin{enumerate}
\item Find all effective broadcast vector sets for $\mathbf{a}_i \in \mathbb{F}^{|\mathcal{K}_i|}_{2}$ . These sets partition $\mathbb{F}_2^N$.
\item Consider an effective broadcast vector set, $\mathcal{C}_L(\mathbf{a}_i)$.
\item \label{map-begn} Partition the effective broadcast vector set into 0-effective broadcast vector set ($\mathcal{C}_{L0}(\mathbf{a}_i)$) and 1-effective broadcast vector set ($\mathcal{C}_{L1}(\mathbf{a}_i)$).
\item All the broadcast vectors in $\mathcal{C}_{L0}(\mathbf{a}_i)$ must be mapped to adjacent signal points. Let the set of signal points corresponding to $\mathcal{C}_{L0}(\mathbf{a}_i)$ be $\mathcal{S}_{L0}(\mathbf{a}_i)$.
\item \label{map-end} All the broadcast vectors in $\mathcal{C}_{L1}(\mathbf{a}_i)$ must be mapped to signal points diametrically opposite to signal points in $\mathcal{S}_{L0}(\mathbf{a}_i)$. This will result in a mapping with broadcast vectors in $\mathcal{C}_{L1}(\mathbf{a}_i)$ mapped to adjacent signal points.
\item Repeat steps \ref{map-begn} to \ref{map-end} by considering the remaining effective broadcast vector sets one by one.
\end{enumerate}
For a receiver $R_i$, when we compare the optimal mappings for two different index codes, the code which has less $|\mathcal{S}_L(\mathbf{a}_i)|$ will perform better as the minimum inter-set distance will be more (note that for both the index codes we do optimal mapping). \\
The mapping algorithm is explained below. Index codes are identified using the corresponding encoding matrices.
\begin{enumerate}
\item The algorithm starts by considering $\mathbb{L}_N$, the set of all index codes of length $N$, for the given ICP. For $R_i$ define,
\begin{equation*}
\eta_i \triangleq \min_{L \in \mathbb{L}_N}{|\mathcal{S}_L(\mathbf{a}_i)|}
\end{equation*}
\item Find $\eta_1$. If $\eta_1<2^N$ proceed to step \ref{alg-t-nonzero} with $i=1$.
\item If $\eta_1=2^N$, $R_1$ sees the full $2^N$-PSK constellation as the effective broadcast signal set. In such a case all mappings for all the index codes will give same PSK-ICG for $R_1$ with $d^{(1)}_{IS,min}$ same as the minimum Euclidean distance between any two points of $2^N$-PSK constellation. In such a case, any mapping for any index code is optimum for $R_1$. Then consider the next highest priority receiver, $R_2$ and continue until a receiver $R_i$ for which $\eta_i<2^N$ is found. If $\eta_i=2^N$ for all receivers, do an arbitrary mapping and exit. Now consider the case where a receiver $R_i$ for which $\eta_i<2^N$ is found.
\item \label{alg-t-nonzero} Let $\{L:|\mathcal{S}_L(\mathbf{a}_i)|=\eta_i\}$ be $\{L_1,L_2,...,L_{n_{L,i}}\}$. For each $L_j, j \in \{1,2,...,{n_{L,i}}\}$, find optimal mappings for $R_i$. Let there be $n_{\mathcal{M},i}$ optimal mappings for each index code and denote the mappings corresponding to index code $L_j$ as $\mathcal{M}_{j1},\mathcal{M}_{j2},...,\mathcal{M}_{jn_{\mathcal{M},i}}$. Define $\mathcal{O}$, the set of ordered pairs as,
\begin{equation*}
\begin{aligned}
\mathcal{O} \triangleq \{&(L_1,\mathcal{M}_{11}),(L_1,\mathcal{M}_{12}),...,(L_1,\mathcal{M}_{1n_{\mathcal{M},i}}),\\
&(L_2,\mathcal{M}_{21}),(L_2,\mathcal{M}_{22}),...,(L_2,\mathcal{M}_{2n_{\mathcal{M},i}}),...,\\
&(L_{n_{L,i}},\mathcal{M}_{{n_{L,i}}1}),(L_{n_{L,i}},\mathcal{M}_{{n_{L,i}}2}),...,\\
&(L_{n_{L,i}},\mathcal{M}_{{n_{L,i}}n_{\mathcal{M},i}})\}.
\end{aligned}
\end{equation*}
The set $\mathcal{O}$ contains all the (index code, mapping) pairs which give the maximum gain possible for $R_i$. Now from this set, identify the pairs which give maximum gain for $R_{i+1}$. For this, choose the pairs which have maximum $d^{(i+1)}_{IS,min}$. Now consider these pairs as set $\mathcal{O}$ and continue until the last receiver $R_m$ is considered and the pairs which have maximum $d^{(m)}_{IS,min}$ is obtained. These are the (index code, mapping) pairs which are optimal.
\end{enumerate}
\subsection{Illustration of Mapping Algorithm}
We illustrate the mapping algorithm given as Algorithm $\ref{alg:a1}$ with an example.
\par Consider the ICP given in Example \ref{exmp:eg1}. Assume that the decreasing order of priority is $(R_1,R_2,R_3,R_4,R_5)$.
Let $N=3$ which is also the length of the optimal index code in this case. Since this is a single unicast ICP, we can find all index codes by considering fitting matrices \cite{Yossef} of rank 3. There are a total of 32 such matrices. For each of these 32 matrices, choose any 3 independent rows as a basis for the row space. So we obtain 32 row spaces (which represents 32 index codes for the given ICP). Of these, only six row spaces are distinct. From Corollary 1 in \cite{Kav2}, the number of index codes possible with the optimal length $c$ for a single-unicast IC problem is given by $\frac{\mu}{c!}\prod^{c-1}_{i=0}(2^c-2^i)$ where $\mu$ is the number of distinct row spaces of c-ranked fitting matrices.
\par For the example under consideration, there are a total of 168 index codes (28 index codes for each distinct row space). So, $\mathbb{L}_3$ contains 168 index codes. $\eta_1 = \min_{L \in \mathbb{L}_3}{|\mathcal{S}_L(\mathbf{a}_1)|}=4$. There are 84 index codes with $\eta_1=4$ and 32 optimal mappings for $R_1$, for each of these index codes. The set $\mathcal{O}$ has $32*84=2688$ (index code, mapping) pairs which are optimal for $R_1$. One such $(L,\mathcal{M})$ pair has the index code as given in Example \ref{exmp:eg1} and mapping as given in Fig.\ref{figure:fig1}(b).
Consider $R_2$. After all pairs in $\mathcal{O}$ are considered, the maximum value possible for $d^{(2)}_{IS,min}=1.414$ and there are 336 pairs which are optimal for $R_2$. Now consider $R_3$. All 336 pairs gives the same $d^{(3)}_{IS,min}=1.414$. For $R_4$ and $R_5$ all pairs have same minimum inter-set distance and these 336 pairs give the (index code, mapping) pairs which are optimal for the ICP considered. For illustration, four such $(L,\mathcal{M})$ pairs are given below. Index code based on encoding matrix $L$ is given in the form of $(y_1,y_2,y_3)$. $\mathcal{M}$ is given as an ordered list of eight integers, representing the decimal equivalent of the 3-tuple, which is mapped to $(\mathbf{s_1},\mathbf{s_2},...,\mathbf{s_{8}})$ where $(\mathbf{s_1},\mathbf{s_2},...,\mathbf{s_{8}})$ are $2^N$-PSK signal points as shown in Fig. \ref{figure:fig1}.
\begin{itemize}
\item $(\{x_1, x_2+x_3,x_4+x_5\},(0,1,2,3,4,5,6,7))$
\item $(\{x_1, x_2+x_3,x_4+x_5\},(0,1,6,7,4,5,2,3))$
\item $(\{x_1, x_2+x_3,x_1+x_4+x_5\},(0,1,2,3,5,4,7,6))$
\item $(\{x_1, x_1+x_2+x_3,x_4+x_5\},(0,1,2,3,6,7,4,5))$
\end{itemize}
\begin{algorithm}
\caption{Algorithm to find optimal (index code, mapping) pairs for a given ICP.} \label{alg:a1}
\begin{algorithmic}[1]
\State $i\gets 1$
\State Find $\eta_i = \min_{L \in \mathbb{L}_N}{|\mathcal{S}_L(\mathbf{a}_i)|}$ \label{step2}
\If {($\eta_i=2^N$)}
\State $i\gets i+1$ \label{step3}
\If {$(i>m)$}
\State Do an arbitrary mapping and Exit.
\Else
\State Goto \ref{step2}
\EndIf
\Else
\begin{itemize}
\item Consider the set of index codes $\{L_1,L_2,...,L_{n_{L,i}}\}=\{L:|\mathcal{S}_L(\mathbf{a}_i)|=\eta_i\}$
\item Find $\mathcal{O}$, the set of all (index code, optimal mappings) pairs for $R_i$.
\end{itemize}
\State $i\gets i+1$ \label{step12}
\If {$(i > m)$}
\State Output $\mathcal{O}$ and Exit.
\Else
\State Choose any $(L,\mathcal{M}) \in \mathcal{O}$
\State $\mathcal{O}^{i} \gets \{(L,\mathcal{M})\}$. Find $\delta=d^{(i)}_{IS,min}$.
\State $\mathcal{O} \gets \mathcal{O} \setminus \{(L,\mathcal{M})\}$ \label{step17}
\If {$(\mathcal{O}=\{\})$}
\State $\mathcal{O} \gets \mathcal{O}^{i}$
\State Goto \ref{step12}
\Else
\State Consider any $(L,\mathcal{M}) \in \mathcal{O}$. Find $d^{(i)}_{IS,min}$.
\If {($d^{(i)}_{IS,min}>\delta$)}
\State $\mathcal{O}^{i} \gets \{(L,\mathcal{M})\}$, $\delta=d^{(i)}_{IS,min}$. Goto \ref{step17}.
\Else
\If {($d^{(i)}_{IS,min}=\delta$)}
\State $\mathcal{O}^{i} \gets \mathcal{O}^{i} \cup \{(L,\mathcal{M})\}$. Goto \ref{step17}.
\Else
\State Goto \ref{step17}.
\EndIf
\EndIf
\EndIf
\EndIf
\EndIf
\end{algorithmic}
\end{algorithm}
\textit{Claim 1:} Algorithm $\ref{alg:a1}$ guarantees that no other mapping of $2^N$-PSK constellation for any index code, can give PSK-ICG $>g_{1}$ for $R_1$.
\begin{proof}
The coding gain (PSK-ICG) achieved by a receiver is maximized when the minimum inter-set distance is maximum. Consider an index code of length $N$. Using Algorithm $\ref{alg:a1}$, for each of the effective broadcast signal sets of the highest priority receiver, the broadcast vectors in 0-effective broadcast vector set are always mapped to adjacent points. Similarly, the broadcast vectors in 1-effective broadcast vector set are always mapped to adjacent points. These sets of points are placed diametrically opposite to each other. Thus, the minimum inter-set distance is maximized for the chosen index code and the mapping is optimal.
\par When we compare the message error performance of $R_1$ with respect to different possible index codes, the code which has less $|\mathcal{S}_L(\mathbf{a}_i)|$ performs better. Index codes with minimum $|\mathcal{S}_L(\mathbf{a}_i)|$ are only considered for mapping in Algorithm $\ref{alg:a1}$. So, the pairs considered by Algorithm $\ref{alg:a1}$ has index codes with minimum $|\mathcal{S}_L(\mathbf{a}_i)|$ and mappings which are optimal. No other mapping of $2^N$-PSK constellation for any index code of length $N$, can give PSK-ICG $>g_{1}$ for $R_1$.
\end{proof}
\textit{Claim 2:} Algorithm $\ref{alg:a1}$ guarantees that, any mapping for any index code which gives the PSK-ICG $g_i$ for receivers $R_i$, $i \in \{1,2,...,j-1\}$ cannot give a PSK-ICG $>g_j$ for $R_j$, $j \leq m$.
\begin{proof}
Algorithm $\ref{alg:a1}$ finds all (index code, mapping) pairs which are optimal for $R_1$. In the next step, among these pairs, which ever gives the maximum gain for $R_2$ are chosen. So, given that $R_1$ has the same PSK-ICG, it is not possible to find another pair for which $R_2$ performs better. Same argument extends to other receivers as well.
\end{proof}
\par Algorithm $\ref{alg:a1}$ can also be used to obtain optimal (index code, mapping) pairs for a given set of index codes of length $N$. In this case the algorithm must be run by considering the given set of index codes instead of all possible index codes of length $N$. This can be illustrated using the ICP given in Example \ref{exmp:eg2}. Assume that the decreasing order of priority is $(R_1,R_2,R_3,R_4,R_5,R_6)$. Let $N=4$ and assume that only one index code as given in Example \ref{exmp:eg2} need to be considered (the given set of index codes is a singleton set).
Consider the highest priority receiver $R_1$. Obtain the effective broadcast vector sets seen by $R_1$ for $\mathbf{a}_1 \in \mathbb{F}^5_2$ and partition these sets. The effective broadcast vector sets and its partitions for $R_1$ are given in Table \ref{table:tbl2}. For any realization of $\mathbf{x}_{\mathcal{I}_{1}}=\mathbf{a}_1$ which is not listed in Table \ref{table:tbl2}, the effective broadcast vector set is same as one of the effective broadcast vector sets given in the table.
\begin{table}[ht]
\caption{Effective broadcast vector sets and its partitions (seen by $R_1$) for the IC in Example \ref{exmp:eg2}.}
\centering
\begin{tabular}{c c c c}
\hline
\hline
$\mathbf{a}_1$ & $\mathcal{C}_L(\mathbf{a}_1)$ & $\mathcal{C}_{L0}(\mathbf{a}_1)$ & $\mathcal{C}_{L1}(\mathbf{a}_1)$\\
\hline
$(00000)$ & $\{(0000),(1000)\}$ & $\{(0000)\}$ & $\{(1000)\}$ \\
$(00001)$ & $\{(0001),(1001)\}$ & $\{(0001)\}$ & $\{(1001)\}$ \\
$(00010)$ & $\{(0010),(1010)\}$ & $\{(0010)\}$ & $\{(1010)\}$ \\
$(00011)$ & $\{(0011),(1011)\}$ & $\{(0011)\}$ & $\{(1011)\}$ \\
$(01000)$ & $\{(0100),(1100)\}$ & $\{(0100)\}$ & $\{(1100)\}$ \\
$(01001)$ & $\{(0101),(1101)\}$ & $\{(0101)\}$ & $\{(1101)\}$ \\
$(01010)$ & $\{(0110),(1110)\}$ & $\{(0110)\}$ & $\{(1110)\}$ \\
$(01011)$ & $\{(0111),(1111)\}$ & $\{(0111)\}$ & $\{(1111)\}$ \\
[1ex]
\hline
\end{tabular}
\label{table:tbl2}
\end{table}
There are $645120$ optimal mappings for $R_1$. The set $\mathcal{O}$ has $645120$ (index code, mapping) pairs which are optimal for $R_1$, with the index code being the same for all the pairs. Consider $R_2$. After all pairs in $\mathcal{O}$ are considered, the maximum value possible for $d^{(2)}_{IS,min}=1.847$ and there are 128 pairs which are optimal for $R_2$. Now consider $R_3$. There are 24 pairs which are optimal with $d^{(3)}_{IS,min}=0.765$. For $R_4$ there are 16 optimal pairs with minimum inter-set distance $d^{(4)}_{IS,min}=0.765$. For $R_5$ and $R_6$ all these pairs give the same minimum inter-set distance. These 16 pairs are the optimal mappings for the IC considered. One of these mappings is given in Fig. \ref{figure:fig2}(a).
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.7]{eg2MapComparison.eps}
\caption{Two 16-PSK mappings for the IC in Example \ref{exmp:eg2}.}
\label{figure:fig2}
\end{center}
\end{figure}
\section{Simulation results} \label{sim_section}
We have considered the ICP given in Example \ref{exmp:eg1} and used Algorithm \ref{alg:a1} to obtain all optimal (index code, mapping) pairs. One such pair, $(L_1,\mathcal{M}_1)$ has the index code as given in Example \ref{exmp:eg1} and mapping as given in Fig. \ref{figure:fig1}(b). We compared this optimal mapping with another mapping $\mathcal{M}_2$ (shown in Fig. \ref{figure:fig1}(c)) which is not optimal for the same index code, $L_1$. The pair $(L_1,\mathcal{M}_2) \notin \mathcal{O}$, the output set obtained from the execution of the algorithm. We obtained the message error probability of the receivers for the two different mappings, by simulation. The first mapping, ($\mathcal{M}_1$) used Algorithm \ref{alg:a1} and the second mapping ($\mathcal{M}_2$) used an algorithm based on maximizing the minimum Euclidean distances \cite{Anj1}. Simulation results are given in Fig. \ref{figure:sim1}.
\begin{figure}
\includegraphics[clip,scale=0.45]{ml_v11_4e.eps}
\caption{Simulation results comparing the performance of receivers for two different mappings (Example \ref{exmp:eg1}).}
\label{figure:sim1}
\end{figure}
The performance of receivers $R_1$, $R_2$ and $R_3$ is significantly better with $\mathcal{M}_1$ than with $\mathcal{M}_2$ at high SNR. The minimum inter-set distances are more for $\mathcal{M}_1$ (Fig.\ref{figure:fig1}(b)) than for $\mathcal{M}_2$ (Fig.\ref{figure:fig1}(c)). For receivers $R_4$ and $R_5$, the minimum inter-set distances are same for both the mappings.
\par We have carried out simulation based studies to compare the performance of the receivers for the ICP and the IC given in Example \ref{exmp:eg2} for two different mappings as given in Fig. \ref{figure:fig2}. The mapping ($\mathcal{M}_1$) given in Fig. \ref{figure:fig2}(a) used Algorithm \ref{alg:a1} and the mapping ($\mathcal{M}_2$) given in Fig \ref{figure:fig2}(b) used the algorithm based on maximizing the minimum Euclidean distances \cite{Anj1}. Simulation results are given in Fig. \ref{figure:sim2}.
\begin{figure}
\includegraphics[clip,scale=0.45]{v12_0_2_MLeg2e.eps}
\caption{Simulation results comparing the performance of receivers for two different mappings (Example \ref{exmp:eg2}).}
\label{figure:sim2}
\end{figure}
For $R_1$, $R_3$, $R_4$, $R_5$ and $R_6$, the minimum inter-set distances and hence the performances are the same for both the mappings. But the performance of receiver $R_2$ is significantly better with $\mathcal{M}_1$ than with $\mathcal{M}_2$ at high SNR.
\par The simulation results indicate the effectiveness of the algorithm based on minimum inter-set distances (Algorithm \ref{alg:a1}) for mapping the broadcast vectors to PSK signal points.
It should be noted that Algorithm \ref{alg:a1} does not guarantee that all the receivers will perform better or as good as that with any other algorithm. It is possible that, a mapping based on some algorithm (say, Algorithm 2) gives a better performance to a receiver $R_j$ than that with Algorithm \ref{alg:a1}. But then there will be a receiver $R_i$ which performs better with Algorithm \ref{alg:a1} than with Algorithm 2, where $R_i$ is a higher priority receiver than $R_j$. In other words, Algorithm \ref{alg:a1} attempts to maximize the gain achieved by the receivers by considering the receivers in the given order of priority. This is further illustrated in Example \ref{exmp:eg3}.
\begin{exmp} \label{exmp:eg3}
Consider the following ICP with $n=m=5$ and $\mathcal{W}_{i}=x_i, ~\forall i \in \{1,2,...,5\}$. The side information available with the receivers is as follows: $\mathcal{K}_1= \{x_2, x_3, x_4, x_5\}, ~\mathcal{K}_2= \{x_1, x_4, x_5\},~ \mathcal{K}_3= \{x_1, x_4\},~\mathcal{K}_4= \{x_2\},~\mathcal{K}_5= \{\}$.
\par For this ICP a scalar linear index code of length $N=4$ (not optimal), is specified as $y_1=x_1+x_2,~ y_2=x_3,~ y_3=x_4,~ y_4=x_5$. Assume that the decreasing order of priority is given as $(R_1,R_2,R_3,R_4,R_5)$.
\end{exmp}
Using Algorithm \ref{alg:a1}, optimal mappings for the specified IC is obtained, of which one mapping ($\mathcal{M}_1$) is given in Fig. \ref{figure:fig3}(a). Another mapping $\mathcal{M}_2$, is found by using the algorithm based on maximizing the minimum Euclidean distances \cite{Anj1} and is given in Fig. \ref{figure:fig3}(b).
Simulation results comparing the performance of the receivers for these two mappings are given in Fig. \ref{figure:sim3}.
\begin{figure}
\begin{center}
\includegraphics[clip,scale=0.7]{eg3MapComparison.eps}
\caption{Two 16-PSK mappings for the IC in Example \ref{exmp:eg3}.}
\label{figure:fig3}
\end{center}
\end{figure}
\begin{figure}
\includegraphics[clip,scale=0.45]{v12_0_3_MLeg4e.eps}
\caption{Simulation results comparing the performance of receivers for two different mappings (Example \ref{exmp:eg3}).}
\label{figure:sim3}
\end{figure}
It is clear from Fig. \ref{figure:sim3} that, $R_2$ performs better with $\mathcal{M}_1$ than with $\mathcal{M}_2$. But $R_3$, which is of lower priority than $R_2$, has better performance with $\mathcal{M}_2$.
\section{Discussion}
In this paper we have considered index coded PSK modulation and have derived a decision rule for the ML decoder which minimizes the message error probability of the receivers, for a given ICP. We have introduced the concept of inter-set distances and illustrated its importance in noisy index coding problems. It was also shown that at high SNR the dominant factor which decides the message error is the minimum inter-set distance and so an optimal mapping must maximize the minimum inter-set distance.
\par Subsequently, we have considered the problem of finding optimal (index code, mapping) pairs across all possible mappings for all possible index codes of length $N$, for a chosen $2^N$-PSK modulation. This problem was not addressed so far in literature. The algorithm which is proposed for a given ICP, can find (index code, mapping) pairs, each of which gives the best PSK-ICG for the receivers, for any given order of priority.
\par Finding all index codes of a chosen length (greater than or equal to the optimal length) for a given ICP is in general NP hard. If it is too complex to find all the index codes, the algorithm can be executed by considering a chosen set of index codes. But the complexity of the proposed algorithm increases exponentially with the length of the index code.
\section*{Acknowledgment}
This work was supported partly by the Science and Engineering Research Board (SERB) of Department of Science and Technology (DST), Government of India, through J.C. Bose National Fellowship to B. Sundar Rajan.
|
\section{Introduction}
\label{sec:intro}
One branch of analysis on non-Archimedean local fields $F$ deals with functions from (subsets of) $F^n$ to ${\mathbb C}$.
In this paper, we develop a framework which makes it possible to carry out this kind of analysis
uniformly in the local field $F$, in a similar sense as algebraic geometry works in a field-independent way.
This extends the work pursued in \cite{CGH}, \cite[Appendix B]{ShinTemp}, \cite{CGH4}, and
\cite[Section 4]{GordonHales}.
One of our motivations is the program initiated by Hales to reformulate in a field-independent way the
entire theory of complex admissible
representations of reductive groups over local fields, \cite{Hales,Hales1}. We note that
in retrospect, one can see an inkling of the possibility of such a reformulation
already in one of the earliest books on the subject, see
\cite{GelfandGraevPyat}, p.~121.
This paper develops the framework in which
this program can be carried out. Another motivation is to enable further applications in the line of e.g.~\cite{CHL,YGordon,CCGordonS,CGH2,ShinTemp,GordonHales}.
In particular, in Section \ref{sec:app}, we develop uniform in $p$ bounds for the Fourier transforms of orbital integrals, normalized
by the discriminant, discussed in more detail below.
Using our framework has several direct implications. The most striking one is probably that
for any property that can be expressed uniformly in the field using the formalism, one has
a transfer principle like the one of Ax--Koshen/Ershov \cite{AK1,Ersov}, i.e., whether the property holds can be transferred between
local fields of positive and mixed characteristic (provided that the residue field characteristic is big enough).
Another direct implication is that if some value
(obtained using the formalism) can be bounded for each local field $F$ independently, then
one obtains some precise information on how the bound depends on $F$, for free. An example of this approach, applied to Fourier transforms of orbital integrals, appears in Section \ref{sec:app}.
By ``doing analysis uniformly in several fields'', we mean that the analytic operations are carried out on certain abstract objects
which can be specialized to every specific field, yielding the familiar concrete objects. We take a very naive
approach to this: we simply define an abstract object to be the collection of its specializations. We do however
require that this collection is given in some uniform way (to be made precise below).
As an example, one kind of objects we work with are definable sets in the sense
of model theory. According to the above approach, we consider a definable set as a collection $(X_F)_F$ of sets $X_F \subset F^n$, all given by the same formula,
where $F$ runs over the local fields we are interested in.
All our uniformity results are valid for all local fields whose characteristic is $0$ or sufficiently big.
For this reason, it makes sense to think of two objects $(X_F)_F$ and $(X'_F)_F$ as being the same if they differ only
for $F$ of small positive characteristic. (Formally defining the objects this way would have been possible but unhandy.)
Note that in particular, we obtain uniformity results even for local fields which are arbitrarily ramified extensions of ${\mathbb Q}_p$, for every $p$. This is new, compared
to most previous papers, and it builds on \cite{CHallp}.
The central objects of study are ``${\mathscr C}^{\mathrm{exp}}$-functions'', as in \cite{CHallp}, also called ``functions of ${\mathscr C}^{\mathrm{exp}}$-class'' or ``of motivic exponential class''. (Those generalize the ``motivic exponential functions'' of \cite{CGH,CGH3,CGH4}, based on the ``motivic constructible exponential functions'' of \cite{CLexp}.) A ${\mathscr C}^{\mathrm{exp}}$-function $f$ is a uniformly given collection of
functions $f_F$ from, say, $F^n$ to ${\mathbb C}$ for every local field $F$.
As an example, given a polynomial $g \in {\mathbb Z}[x]$, we obtain a ${\mathscr C}^{\mathrm{exp}}$-function by setting $f_F(x) := |g(x)|_F$,
where $|\cdot|_F$ is the norm on the local field $F$ (in fact, this function lies in a smaller class of ${\mathscr C}$-functions -- a similar class without the exponentials, also defined below). Another example can be built by
composing a nontrivial additive character on $F$
with a polynomial. In general,
the class of ${\mathscr C}^{\mathrm{exp}}$-functions is defined as the ${\mathbb C}$-algebra generated by those and some other functions,
which involve the valuation on $F$, certain exponential sums over the residue field, and additive characters $F \to {\mathbb C}^\times$.
(The word ``exponential'' refers to the presence of those additive characters, which behave like exponentiation.)
The class of ${\mathscr C}^{\mathrm{exp}}$-functions has intentionally been defined in such a way that it is closed under integration in the following sense.
Given any ${\mathscr C}^{\mathrm{exp}}$-function $f = (f_F)_F$, say in $m+n$ variables,
there exists a ${\mathscr C}^{\mathrm{exp}}$-function $g = (g_F)_F$ in the first $m$ variables such that for every local field $F$ of characteristic $0$ or $\gg1$,
we have
\[
\int_{y\in F^n} f_F(x,y) |dy| = g_F(x) \qquad\text{for all } x \in F^m.
\]
(Here, the integral is a usual Lebesgue-integral with respect to a suitably normalized Haar measure on $F^n$; for the moment, we assume that $f_F(x,\cdot)$ is integrable for each $F$ and each $x$.)
In this sense, we consider $g$ as a ``field-independent integral of $f$''.
In this paper, we prove that a whole zoo of analytic operations on
${\mathscr C}^{\mathrm{exp}}$-functions can be done in a similar, field-independent way.
For an operation which turns functions into other functions, being able to carry it out field-independently in this paper means
that the class of ${\mathscr C}^{\mathrm{exp}}$-functions is closed under the operation, in the same sense as it is closed under integration as above.
Another type of question that we want to be able to answer field-independently is
whether a function has
a given (analytic) property (like being integrable, being continuous, being locally constant, existence of various kinds of limits).
This really becomes meaningful only in a parametrized version:
Given a ${\mathscr C}^{\mathrm{exp}}$-function $f$ in, say, $m + n$ variables, we want to understand how the set
\[X_F := \{x \in F^m \mid f(x,\cdot) \text{ has the desired property}\}\]
depends on $F$.
Typically, the collection $X = (X_F)_F$ is not a definable set; however for most of the properties we will consider, we will prove that
$X = (X_F)_F$ is what we call a ``${\mathscr C}^{\mathrm{exp}}$-locus'': $X_F = \{x \in F^m \mid g_F(x) = 0\}$ for some ${\mathscr C}^{\mathrm{exp}}$-function $g$.
Those loci already appeared in \cite{CGH}; in the present paper, they play a central role.
Even though ${\mathscr C}^{\mathrm{exp}}$-loci are not definable sets, they are almost as flexible; for example, the collection of ${\mathscr C}^{\mathrm{exp}}$-loci is closed under positive boolean combinations,
under the $\forall$-quantifier and under various kinds of for-almost-all quantifiers. This flexibility is inherited by our formalism:
It suffices to prove that ${\mathscr C}^{\mathrm{exp}}$-functions and ${\mathscr C}^{\mathrm{exp}}$-loci are closed under very few
``fundamental'' operations, to then obtain
a multitude of other operations essentially for free (i.e., by combining the
fundamental ones in various ways). Accordingly, this paper contains a few long and deep proofs (of the fundamental operations),
from which then many other results follow easily.
As stated at the beginning of this introduction, our formalism yields Ax--Kochen/Ershov like transfer principles.
More precisely, one has a general transfer principle for arbitrary conditions that can be expressed in terms of ${\mathscr C}^{\mathrm{exp}}$-loci.
(This follows directly from the transfer principle for equalities in \cite{CLexp}.)
This means that we obtain a transfer principle for every analytic condition which can be expressed uniformly
in the sense of this paper. For example, given a
${\mathscr C}^{\mathrm{exp}}$-function $f$, it only depends on the residue field of $F$ whether $f_F$ is continuous/integrable/bounded/constant/locally constant/has a limit at $0$/etc., provided that the residue field characteristic is big enough.
\medskip
The outline of the paper is as follows. Section~\ref{sec:basic-def} contains the basic definitions that
are necessary to understand all results; in particular, ${\mathscr C}^{\mathrm{exp}}$-functions and
${\mathscr C}^{\mathrm{exp}}$-loci are defined.
In Section~\ref{sec:locbasicop}, we recall some results from previous papers needed in this
paper. Here, Proposition~\ref{locbasicop} plays a key role, listing the most important operations
under which the collection of ${\mathscr C}^{\mathrm{exp}}$-loci is closed.
In the remainder of Section~\ref{sec:intro}, we recall some more results from \cite{CGH,CHallp} (in
particular about integrability), we introduce one new basic ``operation'' on ${\mathscr C}^{\mathrm{exp}}$-loci (namely eventual behavior;
Section~\ref{sec:eventual}), and we give some first example applications of the locus formalism.
The remaining uniform analytic operations are grouped by topic in the next two sections,
as follows.
Section~\ref{sec:bounds} contains the results about bounds and suprema. Given a bounded ${\mathscr C}^{\mathrm{exp}}$-function $f$
(i.e., such that $f_F$ is bounded for each $F$), we obtain results about
the dependence of the bound on $F$,
and also, for families of functions,
the dependence of the bound on the family parameter.
A statement about a bound on $f(x)$ is the same as a statement about $\sup_x |f(x)|$,
so it would be natural to first prove that the class of
${\mathscr C}^{\mathrm{exp}}$-functions is closed under taking suprema, a result which would also be useful for many other purposes.
Unfortunately, given
a real-valued ${\mathscr C}^{\mathrm{exp}}$-function $f(x,y)$, the supremum $\sup_y f(x,y)$ is, in general, not a
${\mathscr C}^{\mathrm{exp}}$-function in $x$. However, what we obtain instead is that the supremum
is ``approximately'' a ${\mathscr C}^{\mathrm{exp}}$-function. For all applications of the supremum
in this paper, this turns out to be strong enough.
The main topic of Section~\ref{sec:limits} is limit behavior: pointwise limits, uniform limits and limits
with respect to the $L^p$-norm for various $p$. As applications, we prove results about continuity and about Fourier transforms.
Finally, in Section~\ref{sec:app}, we give an application of the results of Section~\ref{sec:bounds} to Fourier transforms of orbital integrals. The Fourier transform of an orbital integral of a reductive Lie algebra ${\mathfrak g}$ is itself a locally integrable function on ${\mathfrak g}$, and it is known that when normalized by the square root of the discriminant, this function is bounded on compact sets. We prove that for a definable family of definable compact sets $\omega_\lambda$, the bound is of the form $q_F^{a+b\|\lambda\|}$, where
$q_F$ is the cardinality of the residue field of $F$, and $a$ and $b$ are constants that depend only on ${\mathfrak g}$ and the formulas defining the sets $\omega_\lambda$.
This systematic study of bounds for ${\mathscr C}^{\mathrm{exp}}$-functions grew out of the similar study for motivic functions without the exponential that was initiated in order to answer a question that arose in an application of the Trace Formula,
\cite[Appendix B]{ShinTemp}.
Though we do not have such an application in mind for the Fourier transforms of orbital integrals,
we use this opportunity to establish this bound, both for possible future applications, and to illustrate applicability of this machinery.
In particular, now that the orbital integrals are proved to be motivic exponential functions (see Lemma~\ref{lem:can.orb} and the preceding paragraph), all the results
of Section \ref{sec:limits} on the speed of convergence are applicable in particular, to families of orbital integrals, giving for free the estimates by a negative power of $q_F$ for all kinds of sequences of orbital integrals that are known to converge to zero. We do not pursue any specific results of this kind in this paper, but note that they should follow from the results of
Section \ref{sec:limits} and Lemma \ref{lem:can.orb} automatically. We hope that these results will be useful in applications of the Trace Formula in the spirit of \cite{ShinTemp}.
\subsection{Summary of the results}
\label{sec:summary}
Spread over the paper, we give many examples of results that can be deduced
from the basic ones using the formalism.
For the convenience of the reader, we here give a summary of all results (including those that were known before).
Those marked with $^\star$ are ``fundamental'' ones, i.e., those which need real work;
those without $^\star$ are the ones deduced using the formalism.
\smallskip
\noindent
\textbf{Operations on ${\mathscr C}^{\mathrm{exp}}$-functions:}
Applying any of the following operations to a ${\mathscr C}^{\mathrm{exp}}$-function yields a ${\mathscr C}^{\mathrm{exp}}$-functions again:
\nopagebreak
\noindent
\begin{tabular}{@{}ll}
Integral & Theorem~\ref{thm:mot.int.}$^\star$
\\
Bound & Theorem~\ref{thm:fam}
\\
Approximate supremum & Theorem~\ref{thm:fam:gen}$^\star$
\\
Pointwise limit & Theorems~\ref{limits}, \ref{limits:basic}$^\star$
\\
Continuous extension & Proposition~\ref{prop.cont.ext}
\\
Uniform and $L^p$-limit & Theorem~\ref{L^pcom}$^\star$
\\
Fourier transform & Theorem~\ref{stab:four:II}
\end{tabular}
\medskip
\noindent
\textbf{Properties of ${\mathscr C}^{\mathrm{exp}}$-functions:} The following properties of ${\mathscr C}^{\mathrm{exp}}$-functions
are given by ${\mathscr C}^{\mathrm{exp}}$-loci (2nd column) and can be transferred in the Ax--Kochen/Ershov way (3rd column):
\nopagebreak
{\small
\noindent
\begin{tabular}{@{}lll}
Constant & Corollary~\ref{cor.const} & Corollary~\ref{trans.cons}
\\
Locally constant & Corollary~\ref{cor.loc.const} & Corollary~\ref{trans.loc.cons}
\\
$L^1$-Integrable & Theorem~\ref{thm:mot.int.}$^\star$ & \cite[Thm 4.4.1]{CGH}
\\
Bounded & Theorem~\ref{thm:fam}$^\star$ & \cite[Thm 4.4.2]{CGH}
\\
Limit is $0$ & Theorems~\ref{thm:limits0}, \ref{thm:limits0:gen} & Corollary~\ref{trans:lim}
\\
\raggedright
(Pointwise) limit exists & Thms~\ref{limits}, \ref{limits:gen}, \ref{limits:basic}$^\star$ & Corollary~\ref{trans:lim}
\\
Continuous & Corollary~\ref{cor.cont} & Corollary~\ref{trans:lim}
\\
Limits exist: uniform, $L^2$, $L^\infty$ & Theorem~\ref{L^pcom}$^\star$ & Corollary~\ref{trans:funlim}
\\
$L^1$-, $L^2$-, $L^\infty$-integrable & Corollary~\ref{cor.finLp} & Corollary~\ref{trans:Lplim}
\\
$L^\infty$-norm equal to $0$ & Corollary~\ref{cor.almost} & Corollary~\ref{trans:Lplim}
\end{tabular}
}
\medskip
\noindent
\textbf{Operations on ${\mathscr C}^{\mathrm{exp}}$-loci and properties of ${\mathscr C}^{\mathrm{exp}}$-loci:}
Applying any of the following operations to a ${\mathscr C}^{\mathrm{exp}}$-locus yields a ${\mathscr C}^{\mathrm{exp}}$-locus again:
\nopagebreak
\noindent
\begin{tabular}{@{}ll}
Boolean combinations, $\forall$-quantification & Proposition~\ref{locbasicop}$^\star$
\\
Eventual behavior & Proposition~\ref{for:all:large:mu}$^\star$
\\
Almost everywhere quantification & Corollary~\ref{cor.almost}
\end{tabular}
\medskip
\noindent
\textbf{Consequences of being a ${\mathscr C}^{\mathrm{exp}}$-function / a ${\mathscr C}^{\mathrm{exp}}$-locus:}\nopagebreak
\noindent
\begin{tabular}{@{}ll}
Transfer principle for ${\mathscr C}^{\mathrm{exp}}$-loci & Theorem~\ref{thm.trans}$^\star$
\\
Bounds are uniform in the field & Theorems~\ref{thm:presburger-fam}, \ref{thm:fam}
\\
Speed of convergence of pointwise limits & Thms~\ref{thm:limits0}, \ref{thm:limits0:gen}, \ref{limits:basic}$^\star$
\\
$L^p$-convergence $\Rightarrow$ pointwise convergence & Lemma~\ref{Lp-vs-pointwise}$^\star$
\end{tabular}
\medskip
\noindent
\textbf{Applications to Fourier transforms of orbital integrals:}\\
\nopagebreak
{\small
\begin{tabular}{@{}ll}
Orbital integrals are of ${\mathscr C}$-class
(up to a ${\mathscr C}$-constant) & Lemma \ref{lem:can.orb}
\\
Fourier transforms of orbital integrals are of ${\mathscr C}^{\mathrm{exp}}$-class & Lemma \ref{lem:can.orb}
\\
Uniform bound for Fourier transforms of orbital integrals & Theorem \ref{thm:orb.int.bound}
\end{tabular}
}
\subsection{Basic definitions: functions and loci of ${\mathscr C}^{\mathrm{exp}}$-class}
\label{sec:basic-def}
The main object of study in this paper is the class of ``${\mathscr C}^{\mathrm{exp}}$-functions''; by incorporating all finite field extensions of ${\mathbb Q}_p$ for all $p$ as in \cite{CHallp}, they generalize the ``motivic exponential functions'' of \cite[Section 2]{CGH4} and \cite[Section 2]{CGH3}, all based on \cite{CLexp}.
We note that this class of functions differs also in another way from the class of motivic constructible exponential functions originally
defined in \cite{CLexp}: instead of the abstract motivic functions of \cite{CLexp}, we consider collections of functions, uniformly in local fields. This avoids issues related to what is called \emph{null-functions} in \cite{Casselman-Cely-Hales}.
We recall this framework and fix our terminology.
\begin{defn}[Local fields]\label{AO}
Let ${\mathrm{Loc}}$ be the collection of all non-Archimedean local fields $F$ equipped with
a uniformizer $\varpi_F \in F$ for the valuation ring of $F$.
(So more formally, elements of ${\mathrm{Loc}}$ are pairs $(F, \varpi_F)$.)
Here, by a non-Archimedean local field we mean a finite extension of ${\mathbb Q}_p$ or of ${\mathbb F}_p\mathopen{(\!(} t\mathopen{)\!)}$ {for any prime $p$}.
Given an integer $M$, let ${\mathrm{Loc}}_{M}$ be the collection of $(F,\varpi_F) \in {\mathrm{Loc}}$ such that
$F$ has characteristic either $0$ or at least $M$.
We will use the notation ``for all $F \in {\mathrm{Loc}}_{\gg 1}$'' to mean ``there exists an $M$ such that for all $F \in {\mathrm{Loc}}_{M}$''.
\end{defn}
Note that in various previous papers, the characteristic of the residue field $k_F$ is required to be sufficiently big.
In contrast, results in this paper only require $F \in {\mathrm{Loc}}_{\gg 1}$, which allows the characteristic of $k_F$ to be arbitrary if the characteristic of $F$ is zero. That this is sufficient builds on the results of \cite{CHallp}. On the other hand, it still seems far out of reach to treat local fields $F$ of small positive characteristic, given that
the model theory of such fields is not understood.
\begin{notn}[Sorts]
Given a local field $F \in {\mathrm{Loc}}$, we sometimes write $\mathrm{VF}_F$ for the underlying set of $F$, we write ${\mathcal O}_F$ for the valuation ring of $F$, ${\mathcal M}_F$ for the maximal ideal, ${\rm RF}_F$ for the residue field and $q_F$ for
the number of elements of ${\rm RF}_F$. The value group (even though always being ${\mathbb Z}$) will sometimes be denoted by $\mathrm{VG}_F$.
Moreover, we introduce a notation for the \emph{residue rings:} For positive integers $n$, set ${\rm RF}_{n,F} := {\mathcal O}_F/n{\mathcal M}_F$
(where $n{\mathcal M}_F = \{na \mid a \in {\mathcal M}_F\}$).
We write $\operatorname{ord}_F\colon \mathrm{VF}_F \to \mathrm{VG}_F \cup \{\infty\}$ for the valuation map, $\operatorname{res}_F\colon {\mathcal O}_F \to {\rm RF}_F$ for the residue map and more generally
$\operatorname{res}_{n,F}\colon {\mathcal O}_F \to {\rm RF}_{n,F}$ for the canonical projections.
We use ``$W \subset F^*$'' (or ``$W \subset \mathrm{VF}_F^*$'') as a shorthand notation for $W$ being a subset of $F^n$ for some $n \ge 0$,
and likewise ``$W \subset {\mathbb Z}^*$'' (or ``$W \subset \mathrm{VG}_F^*$'') and ``$W \subset \@ifnextchar_{\RFss@}{\RFss@@{}}_{F}$'',
the latter meaning that $W$ is a subset of a product $\prod_{i=1}^\ell {\rm RF}_{n_i,F}$, for some $\ell$ and $n_1, \dots, n_\ell$.
\end{notn}
Note that $n{\mathcal M}_F$ is always a power of the ideal ${\mathcal M}_F$, but which power it is depends on $n$ and on $F$. Namely, for $r$ the unique non-negative integer such that the order of $\varpi_F ^r$ equals $\operatorname{ord} n$, one has $n{\mathcal M}_F = {\mathcal M}_F^{1+r}$.
In particular, we have $n{\mathcal M}_F = {\mathcal M}_F$ and hence ${\rm RF}_{n,F} = {\rm RF}_F$ whenever the residue field characteristic of $F$ does not divide $n$. This somewhat technical definition
of the sorts is necessary to obtain the desired uniformity in $F$. In particular, since any formula uses only finitely many sorts,
this implies that it suffices to exclude finitely many $p$ to achieve that the formula does not use any residue ring at all (except ${\rm RF}_F$).
This fits to the papers that exclude small residue characteristic but do not need residue rings.
We use the same generalized Denef--Pas language ${\mathcal L}_{\mathrm {DP}}$ as in \cite[Section~2.2]{CHallp} and consider each $F \in {\mathrm{Loc}}$ as a structure in that language.
It is defined as follows.
\begin{defn}\label{defn.lang}
The language ${\mathcal L}_{\mathrm {DP}}$
has sorts $\mathrm{VF}$ (the valued field), $\mathrm{VG}$ (the value group) and ${\rm RF}_n$ for each $n \ge 1$ (the residue rings), with
the ring language on $\mathrm{VF}$ and on each ${\rm RF}_n$, the ordered abelian group language on $\mathrm{VG}$, the valuation map
$\operatorname{ord}\colon \mathrm{VF} \to \mathrm{VG} \cup \{\infty\}$ and generalized angular component maps ${\overline{\rm ac}}_n\colon \mathrm{VF} \to {\rm RF}_n$ sending $x$ to $\operatorname{res}_n(\varpi_F^{-\operatorname{ord}(x)}x)$.
See \cite[Section~2.2]{CHallp} for more details.
\end{defn}
For some applications, it may be useful to have additional constants in the language (e.g.
for the elements of a fixed subring of $F$, or for the uniformizer $\varpi_F$).
Using some standard techniques from model theory,
all of our results can be suitably reformulated in such an enriched language;
see Appendix~\ref{sec:const} for details.
An ${\mathcal L}_{\mathrm {DP}}$-formula uniformly yields sets for all $F \in {\mathrm{Loc}}$.
As stated in the introduction, it will be convenient to consider a definable set
as the collection of the sets it actually defines in the fields $F$:
\begin{defn}[Definable sets and functions]\label{defset}
A collection
$X = (X_F)_{F \in {\mathrm{Loc}}_{M}}$ of subsets $X_F\subset \mathrm{VF}_F^*\times {\rm RF}^*_{*,F}\times \mathrm{VG}_F^*$ for some $M$ is called a \emph{definable set} if there is an
${\mathcal L}_{\mathrm {DP}}$-formula $\varphi$ such that $X_F = \varphi(F)$ for each
$F$ in ${\mathrm{Loc}}_{M}$.
For definable sets $X$ and $Y$, a collection $f = (f_F)_{F \in{\mathrm{Loc}}_{M}}$ of functions $f_F:X_F\to Y_F$ for some $M$ is called a \emph{definable function} and denoted by $f:X\to Y$ if the collection of graphs of the $f_F$ is a definable set. (For this to make sense, we assume the $M$ of $f$ to be at least as big as the ones of $X$ and $Y$.)
\end{defn}
In reality, we are only interested in definable sets and functions for ``big $M$'',
i.e., we think of $X$ and $X'$ as being equal if
\begin{equation}\label{eq.defset}
X_F = X'_F \text{ for all } F \in {\mathrm{Loc}}_{\gg1}.
\end{equation}
However, for practical reasons, it is often easier
to work with representatives instead of introducing this equivalence relation formally.
Nevertheless, we will often implicitly enlarge $M$, for example calling $(X_F)_{F}$ definable
if $(X_F)_{F\in{\mathrm{Loc}}_{M}}$ is definable for big enough $M$.
Our definition of definable sets is the one which will be most convenient for this paper,
but note that there are various other possibilities. For example, if one were to
define a definable set to be the collection indexed only by the local fields of characteristic $0$,
one could recover (\ref{eq.defset}) by the Ax--Kochen/Ershov transfer principle.
We apply the typical set-theoretical notation to definable sets $X, Y$, e.g.\ writing
$X \subset Y$ (if $X_F \subset Y_F$ for each $F \in {\mathrm{Loc}}_{\gg 1}$), $X \times Y$, and so on, which may increase $M$ if necessary.
More generally, we will often omit the indices $F$ when the intended meaning is clear,
writing e.g. ``$\{x \in X \mid \text{some condition written without indices $F$}\}$'' to mean the corresponding collection of subsets of the $X_F$;
also, by ``for all $x \in X$, $f(x) = g(x)$'' we mean ``for all $F \in {\mathrm{Loc}}_{\gg 1}$ and all $x \in X_F$, $f_F(x) = g_F(x)$''.
Using definable functions and sets as building blocks, we introduce
``functions of ${\mathscr C}$-class''; those will then be generalized to ``functions of ${\mathscr C}^{\mathrm{exp}}$-class''.
\begin{defn}[Functions of ${\mathscr C}$-class]\label{motfun}
Let $X = (X_F)_{{F\in {\mathrm{Loc}}_{M}}}$ be a definable set.
A collection $H = (H_F)_F$ of functions $H_F:X_F\to{\mathbb R}$ is called \emph{a function of ${\mathscr C}$-class} (or simply a \emph{${\mathscr C}$-function}) on $X$ if
there exist integers
$N$, $N'$, and $N''$, nonzero integers $a_{i\ell}$, definable functions $\alpha_{i}:X\to {\mathbb Z}$ and $\beta_{ij}:X\to {\mathbb Z}$,
and definable sets $Y_i\subset X\times \@ifnextchar_{\RFss@}{\RFss@@{}}$ such that for all $F\in {\mathrm{Loc}}_{ {M}}$ and all $x\in X$
\begin{equation}\label{eq.motfun}
H(x)=\sum_{i=1}^N \# Y_{i,x} \cdot q^{\alpha_{i}(x)} \cdot \big( \prod_{j=1}^{N'} \beta_{ij}(x) \big) \cdot \big( \prod_{\ell=1}^{N''} \frac{1}{1-q^{a_{i\ell}}} \big),
\end{equation}
where $Y_{i,x} = \{y\in \prod_{t=1}^{\ell_i} {\rm RF}_{n_{i,t},F} \mid (x,y)\in Y_{i}\}$, for some $\ell_i$ and $n_{i,t}$.
We write ${\mathscr C}(X)$ to denote the ring of ${\mathscr C}$-functions on $X$.
\end{defn}
In this definition, we already omitted lots of indices $F$ from the notation. Equation (\ref{eq.motfun}) e.g.\ really means
\[
H_F(x)=\sum_{i=1}^N \# Y_{i,F,x} \cdot q_F^{\alpha_{iF}(x)} \cdot \big( \prod_{j=1}^{N'} \beta_{ijF}(x) \big) \cdot \big( \prod_{\ell=1}^{N''} \frac{1}{1-q_F^{a_{i\ell}}} \big).
\]
(Note that each $Y_{i,F,x}$ is a finite set, so $\# Y_{i,F,x}$ makes sense.)
\begin{defn}[Additive characters]\label{psiu}
For any local field $F$, let ${\mathcal D}_F$ be the set of the additive characters $\psi$ on $F$ that are trivial on the maximal ideal ${\mathcal M}_F$ of ${\mathcal O}_F$ and
nontrivial on ${\mathcal O}_F$.
\end{defn}
Expressions involving additive characters of $p$-adic fields often give rise to exponential sums; that's what the ``exp'' stands for
in the definition below.
\begin{defn}[Functions of ${\mathscr C}^{\mathrm{exp}}$-class]\label{expfun}
Let $X = (X_F)_{ {F\in {\mathrm{Loc}}_{M}}}$ be a definable set.
A collection $H = (H_{F,\psi})_{F,\psi}$ of functions $H_{F,\psi}:X_F\to{\mathbb C}$ for $F\in {\mathrm{Loc}}_{M}$ and $\psi\in {\mathcal D}_F$ is called \emph{a function of ${\mathscr C}^{\mathrm{exp}}$-class}
(or a \emph{${\mathscr C}^{\mathrm{exp}}$-function}) on $X$ if
there exist integers $N>0$ and $n_i\geq 1$, functions $H_i=(H_{iF})_F$ in ${\mathscr C}(X)$, definable sets $Y_i\subset X\times \@ifnextchar_{\RFss@}{\RFss@@{}}$ and definable functions $g_i:Y_i\to \mathrm{VF}$ and $e_i:{Y_i}\to {\rm RF}_{n_i}$ for $i=1,\ldots,N$, such that for all $F\in {\mathrm{Loc}}_{M}$, all $\psi\in {\mathcal D}_F$ and all $x\in X_F$
\begin{equation}\label{fexp}
H_{F,\psi}(x)=\sum_{i=1}^N H_{iF}(x)\left( \sum_{y \in Y_{i,F,x}}\psi\left( g_{iF}(x,y) + \frac{e_{iF}(x,y)}{n_i} \right)\right),
\end{equation}
where $\psi(a+\frac{v}{n})$ for $a\in F$ and $v\in {\rm RF}_{n,F}$, by abuse of notation, is defined as $\psi(a+\frac{u}{n})$ with $u$ any element in ${\mathcal O}_F$ such that $\operatorname{res}_n(u)=v$, which is well defined since $\psi$ is constant on ${\mathcal M}_F$.
We write ${\mathscr C}^{\mathrm{exp}}(X)$ to denote the ring of ${\mathscr C}^{\mathrm{exp}}$-functions on $X$.
\end{defn}
In the same way as we omit indices $F$, we will also omit indices $\psi$ from the notation,
for example writing ``$\forall x\in X\colon H(x) = H'(x)$'' to mean
``$\forall F\in {\mathrm{Loc}}_{\gg1},\psi\in {\mathcal D}_F,x\in X_F\colon H_{F,\psi}(x) = H_{F,\psi}'(x)$''.
Also, we will freely consider collections $X = (X_F)_{F\in {\mathrm{Loc}}_{M}}$ as collections
indexed by $F$ and $\psi$ not depending on $\psi$.
Loci of vanishing of ${\mathscr C}^{\mathrm{exp}}$-functions will play a key role in this paper.
To some extent, we will be able to apply the same reasoning as we do to the definable sets.
We develop some terminology for that.
\begin{defn}[Loci]\label{locset}
A \emph{locus of ${\mathscr C}$-class} (or a ${\mathscr C}$-locus) is a zero set $X$ of a function $f \in {\mathscr C}(\mathrm{VF}^*\times \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*)$,
i.e., $X = (X_F)_{F \in {\mathrm{Loc}}_{M}}$, where $X_F = \{x \in \mathrm{VF}_F^* \times \@ifnextchar_{\RFss@}{\RFss@@{}}_F \times \mathrm{VG}_F^* \mid f_F(x) = 0\}$.
Similarly, a \emph{locus of ${\mathscr C}^{\mathrm{exp}}$-class} is a zero set of a function $f \in {\mathscr C}^{\mathrm{exp}}(\mathrm{VF}^*\times \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*)$.
(The latter one is a collection indexed by $F$ and $\psi$).
\end{defn}
Note that any definable set is also a ${\mathscr C}$-locus, since characteristic functions of definable sets are of ${\mathscr C}$-class.
The converse, however, is not true; for example, $X = \{(x,y) \in \mathrm{VG}^2 \mid q^x - y = 0\}$ is a ${\mathscr C}$-locus, but it is not definable.
Moreover, definable sets never depend on $\psi$.
Sometimes, it is notationally more convenient to refer to the condition describing a set instead of the set itself:
\begin{notn}[Conditions]\label{loccond}
Fix some integers $M,n,\ell,m_1, \dots, m_\ell,r$ and consider a collection $P = (P_{F,\psi})_{F \in {\mathrm{Loc}}_{M},\psi \in {\mathcal D}_F}$ of
conditions $P_{F,\psi}(x)$ on elements $x \in \mathrm{VF}_F^n \times {\rm RF}_{m_1,F} \times \dots \times {\rm RF}_{m_\ell,F} \times \mathrm{VG}^r$.
We call $P$ a \emph{definable condition} if the collection of sets $X = \{x \in \mathrm{VF}^* \times \@ifnextchar_{\RFss@}{\RFss@@{}} \times \mathrm{VG}^*
\mid P(x) \text{ holds}\}$ is definable.
Analogously, we call $P$ a condition of ${\mathscr C}$-class or of ${\mathscr C}^{\mathrm{exp}}$-class, respectively, if $X$
is a locus of the corresponding class. Again, we also say ${\mathscr C}$-condition or ${\mathscr C}^{\mathrm{exp}}$-condition for short.
\end{notn}
Thus a ${\mathscr C}^{\mathrm{exp}}$-condition is a family $P = (P_{F,\psi})_{F,\psi}$ of conditions given by a ${\mathscr C}^{\mathrm{exp}}$-function $f$, namely:
$P_{F,\psi}(x)$ holds if and only if $f_{F,\psi}(x) = 0$ (for all appropriate $F$, $\psi$, $x$).
Note that in the case $n = \ell = r = 0$, each $P_{F,\psi}$ is a condition on elements of a one point set
so it makes sense to ask whether $P_{F,\psi}$ holds without specifying any element.
(Definable such $P$ are essentially just first order sentences.)
\begin{rem}\label{remove-exp}
In the remainder of this section and in all of Sections~\ref{sec:bounds} and \ref{sec:limits}, literally every result that is stated for ${\mathscr C}^{\mathrm{exp}}$-functions and ${\mathscr C}^{\mathrm{exp}}$-loci is also valid if one replaces all occurrences of ${\mathscr C}^{\mathrm{exp}}$ by ${\mathscr C}$, i.e., if no input object of a result uses the additive character $\psi$, then $\psi$ does not appear in the output objects either. (This is obvious from the proofs.)
\end{rem}
\subsection{The locus formalism, transfer, and constancy}
\label{sec:locbasicop}
By definition of first order formulas, the class of definable conditions is closed under finite boolean combinations and under quantification.
Results from \cite{CGH,CHallp} state that this is partially also true for the new classes of conditions introduced in Notation~\ref{loccond}:
Each of them is closed under finite positive boolean combinations and under universal quantification.
Note however that unlike definable conditions, they are not closed under negation, and neither under
existential quantification. For ${\mathscr C}^{\mathrm{exp}}$-conditions this follows from Example~\ref{ex.noExist}, and for ${\mathscr C}$-conditions it follows from Example \ref{ex:existq}.
The following proposition summarizes the positive results for further reference:
\begin{prop}[Basic operations on Loci]\label{locbasicop}
In the following,
$x$ runs over a definable set $X$, and $y$ runs over a definable set $Y$.
\begin{enumerate}
\item Any definable condition is a ${\mathscr C}^{\mathrm{exp}}$-condition.
\item If $P(x)$ and $Q(x)$ are ${\mathscr C}^{\mathrm{exp}}$-conditions on $x$, then so are $P(x) \wedge Q(x)$ and $P(x) \vee Q(x)$.
\item If $P(x, y)$ is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$ and $y$, then $\forall y\colon P(x,y)$ is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$.
\item A ${\mathscr C}^{\mathrm{exp}}$-condition $P(x)$ stays a ${\mathscr C}^{\mathrm{exp}}$-condition when considered as a condition on $x$ and $y$ that is independent of $y$.
\end{enumerate}
\end{prop}
Again, we are omitting indices. For example, ``$\forall y\colon P(x,y)$'' is the family $R = (R_{F,\psi})_{F,\psi}$ of conditions, where,
for $F\in {\mathrm{Loc}}_{\gg1},\psi\in {\mathcal D}_F,x\in X_F$, $R_{F,\psi}(x)$ holds if and only if $P_{F,\psi}(x,y)$ holds
for all $y \in Y_F$.
\begin{proof}[Proof of Proposition~\ref{locbasicop}]
(1) is just a reformulation of the fact that definable sets are loci.
(2) is just a simple manipulation of the ${\mathscr C}^{\mathrm{exp}}$-functions corresponding to $P$ and $Q$; see
Corollary 3.5.4 of \cite{CGH}.
(3) is the (Iva)-part of Theorem 4.4.2 of \cite{CHallp} (cf.\ also Theorem 4.3.2 of \cite{CGH}).
(4) follows from the corresponding statement for ${\mathscr C}^{\mathrm{exp}}$-functions, i.e., that such functions in $x$
can also be considered as functions in $x$ and $y$.
\end{proof}
Another way of stating Proposition~\ref{locbasicop} is that any (syntactically correct) mathematical expression built out of ${\mathcal L}_{\mathrm {DP}}$-formulas,
``$f = 0$'' (for $f$ a ${\mathscr C}^{\mathrm{exp}}$-function), $\wedge$, $\vee$ and $\forall$ defines a ${\mathscr C}^{\mathrm{exp}}$-condition.
(Note that (4) is implicitly used when we write something like $P(x, y)\wedge Q(x, z)$, which is a condition on $x,y,z$.)
The ${\mathscr C}^{\mathrm{exp}}$-functions $f$ in turn can be given by any expression of the form as in Definition~\ref{motfun} or \ref{expfun}.
Such expressions can be considered as a generalization of first order formulas, and they provide
a short and convenient way to prove that new conditions are of ${\mathscr C}^{\mathrm{exp}}$-class.
Several of the main results of this paper (and also of \cite{CGH}) just state that certain
additional operations can be used in expressions defining ${\mathscr C}^{\mathrm{exp}}$-conditions.
As an example of how this formalism works, we prove:
\begin{cor}[Constancy]\label{cor.const}
Let $f$ be in ${\mathscr C}^{\mathrm{exp}}(X\times Y)$, for some definable sets $X$ and $Y$.
Then the set of $x \in X$ for which the function $y \mapsto f(x, y)$ is constant is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\end{cor}
\begin{proof}
That $f(x,\cdot)$ is constant can be written as ${\mathscr C}^{\mathrm{exp}}$-condition as follows:
\[\forall y_1, y_2 \in Y\colon f(x,y_1) = f(x,y_2).\]
\end{proof}
To illustrate our conventions and formalism once more, let us give the following extra explanation of the above proof:
$f(x,y_1) - f(x,y_2)$ is a ${\mathscr C}^{\mathrm{exp}}$-function in $x$, $y_1$, $y_2$, so this difference being zero defines a ${\mathscr C}^{\mathrm{exp}}$-condition
on $x$, $y_1$, $y_2$, namely the family
$P = (P_{F,\psi})_{F,\psi}$, where $P_{F,\psi}(x, y_1, y_2)$ holds if and only if
$f_{F,\psi}(x,y_1) - f_{F,\psi}(x,y_2) = 0$. By Proposition~\ref{locbasicop} (3),
\[
P'(x, y_1) \;:\Longleftrightarrow\; \forall y_2\colon P(x, y_1, y_2)
\]
is a ${\mathscr C}^{\mathrm{exp}}$-condition, and using Proposition~\ref{locbasicop} (3) once more,
so is
\[
P''(x) \;:\Longleftrightarrow\; \forall y_1\colon P'(x, y_1) \iff \forall y_1,y_2\colon P(x, y_1, y_2).
\]
This $P''(x)$ corresponds to a ${\mathscr C}^{\mathrm{exp}}$-locus $Z = \{x \in X \mid P''(x)\}$, which, by definition of $P''$, is the one desired by the corollary:
For every $F, \psi$, we have $x \in Z_{F, \psi}$ if and only if $f_{F,\psi}(x,y_1) - f_{F,\psi}(x,y_2) = 0$ holds for all $y_1, y_2 \in Y_F$
(which just means that $f_{F,\psi}(x,\cdot)$ is constant).
One motivation to introduce ${\mathscr C}^{\mathrm{exp}}$-conditions is that
for them, one has a transfer principle analogous to the Ax--Kochen/Ershov Theorem for first order sentences:
Whether a ${\mathscr C}^{\mathrm{exp}}$-condition holds only depends on the residue field, provided that the residue field characteristic is big enough.
More precisely, we have the following:
\begin{thm}[Transfer]\label{thm.trans}
Suppose that $P(x)$ is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x \in X$, for some definable set $X$.
Then there exists an $M$ such that for all $F \in {\mathrm{Loc}}$ with
residue field characteristics at least $M$, the following holds.
If $P_{F,\psi}(x)$ holds for all $\psi \in {\mathcal D}_{F}$ and all $x \in X_F$, then for any
$F' \in {\mathrm{Loc}}$ whose residue field is isomorphic to the one of $F$,
$P_{F',\psi}(x)$ holds for all $\psi \in {\mathcal D}_{F'}$ and all $x \in X_{F'}$.
\end{thm}
\begin{proof}
This is Proposition~9.2.1 of \cite{CLexp}.
\end{proof}
Note that as a ``basic transfer result'', it would be enough to have Theorem~\ref{thm.trans} for 0-ary
${\mathscr C}^{\mathrm{exp}}$-conditions. Indeed, Theorem~\ref{thm.trans} for general $P(x)$ is obtained by applying transfer
to the 0-ary condition
\[
\forall x \in X\colon P(x),
\]
which is of ${\mathscr C}^{\mathrm{exp}}$-class by Proposition~\ref{locbasicop} (3).
This transfer principle can be combined with many other results from this paper.
As a first example, we obtain:
\begin{cor}[Transfer for constancy]\label{trans.cons}
Let $X$ and $Y$ be definable sets, and let $f$ be in ${\mathscr C}^{\mathrm{exp}}(X\times Y)$.
Then there exists $M$ such that, for any $F\in {\mathrm{Loc}}$ of residue field characteristic $\ge M$,
the truth of the following statement only depends on (the isomorphism class of) the residue field of $F$:
\[
\text{For all $\psi\in{\mathcal D}_F$ and all $x\in X_F$, }
y\mapsto f_{F,\psi} (x,y) \text{ is constant on } Y_F.
\]
\end{cor}
\begin{proof}
By Corollary~\ref{cor.const}, $f(x,\cdot)$ being constant is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$, so Theorem~\ref{thm.trans} applies.
\end{proof}
It might be tempting to replace our ${\mathscr C}^{\mathrm{exp}}$-loci by a larger class which is additionally closed
under existential quantification, and maybe even under negation (thus turning it into a class of definable sets in a
first order language).
However, the following example shows that this would destroy the transfer principle.
\begin{example}\label{ex.noExist}
Transfer (in the style of Theorem~\ref{thm.trans}) does not hold for the statement
\begin{equation}\label{eq:no-trans}
\forall x \in \mathrm{VF}_F\colon \exists y \in {\mathcal O}_F\colon \psi_F(x) = \psi_F(y).
\end{equation}
Indeed: The image $\psi({\mathcal O}_F)$ consists of the $p$-th roots of unity, where $p$ is the residue characteristic
of $F$. If $F$ itself has characteristic $p$, then the entire image $\psi(F)$ consists only of the $p$-th roots of
unity, whereas if $F$ has characteristic $0$, then $\psi(F)$ contains all $p^r$-th roots of unity for all $r > 1$.
Thus (\ref{eq:no-trans}) holds if and only if $F$ has positive characteristic.
\end{example}
\begin{example}\label{ex:existq}
We give an example that, in general, ${\mathscr C}$-loci are not closed under existential quantification.
If ${\mathscr C}$-loci would be closed under existential quantification, then in particular
$Y = \{y \in \mathrm{VG} \mid \exists x \in \mathrm{VG}\colon q^x - y = 0\}$ would be a ${\mathscr C}$-locus, i.e., $Y = (Y_F)_F$ with
$Y_F = \{q_F^n \mid n \in {\mathbb N}\}$. However, for any $H = (H_F)_F \in {\mathscr C}(\mathrm{VG})$ and any $F$, the zero set of $H_F$
is ultimately periodic, i.e., there exist $N, e \in {\mathbb N}$ such that for $\mu > N$, whether $H_F(\mu)$ is zero or not only depends on $\mu \mod e$.
Indeed, this follows from Proposition~\ref{repar}: For $\mu$ big enough and in a fixed congruence class modulo some suitable $e$,
we have \[
H_F(\mu) = \sum_i c_{i}\mu^{a_i}q_F^{b_i\mu}
\]
for finitely many non-negative integers $a_i$, rational $b_i$ and complex $c_i$, and such a function is either constant equal to zero (namely when all $c_i$ are zero)
or has only finitely many zeros,
e.g. by \cite[Lemma~2.1.8]{CGH}.
\end{example}
\subsection{Eventual behavior and local constancy}
\label{sec:eventual}
The first new building block for ${\mathscr C}^{\mathrm{exp}}$-conditions provided by this paper is
the following proposition about eventual behavior.
(One should not confuse this result with Theorem 14 of \cite{GordonHales} where $\mu$ is not allowed to depend on $x$.)
\begin{prop}[Eventual behavior]\label{for:all:large:mu}
If $P(x, \mu)$ is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$ running over a definable set $X$ and on $\mu$ running over $\mathrm{VG}$, then the condition $Q(x)$ given by
\begin{equation}\label{eq:event}
Q(x) \iff \exists \mu_0\in \mathrm{VG}\colon \forall \mu \ge \mu_0\colon P(x, \mu)
\end{equation}
is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$. Moreover, $\mu_0$ can be chosen to depend definably on $x$, i.e.,
there exists a definable function $\mu_0:X\to \mathrm{VG}$ such that for each $F\in {\mathrm{Loc}}_{\gg1}$, each $\psi\in {\mathcal D}_F$ and each
$x\in X_F$, if $Q_{F,\psi}(x)$ holds, then
$P_{F,\psi}(x,\mu)$ holds for all $\mu\geq \mu_{0,F}(x)$.
\end{prop}
Recall that (\ref{eq:event}) defines a condition depending not only on $x$, but implicitly also on $F$ and $\psi$,
that is, for any $F\in {\mathrm{Loc}}_M$, any $\psi\in {\mathcal D}_F$, and any $x\in X_F$, the condition $Q_{F,\psi}(x)$ holds if and only if there is an integer $\mu_0$ (possibly depending on $F,\psi,x$) such that for all integers $\mu$ with $\mu \ge \mu_0$ one has $P_{F,\psi}(x, \mu)$.
Thus the first part of the proposition states that in expressions defining ${\mathscr C}^{\mathrm{exp}}$-conditions (as explained below Proposition~\ref{locbasicop}),
we are also allowed to use the quantifier ``$\forall \lambda \gg 1$'' (meaning for all sufficiently big $\lambda \in \mathrm{VG}$).
The proof of this proposition (below) essentially follows a reasoning often used in \cite{CGH}, whose
main ingredient is that the dependence of $P$ on $\mu$ is controlled by Presburger-definable data.
However, given that we work in a context allowing arbitrary ramification, this becomes true only after
introducing additional $\@ifnextchar_{\RFss@}{\RFss@@{}}$-variables. The following proposition summarizes several of the
results from \cite{CHallp} we shall need.
\begin{prop}\label{repar}
Let $X$ and $U\subset X\times \mathrm{VG}$ be definable and let $H$ be a ${\mathscr C}^{\mathrm{exp}}$-function on $U$. Then there exists
a definable bijection $\sigma\colon U \to U_{\rm par} \subset \@ifnextchar_{\RFss@}{\RFss@@{}} \times X \times \mathrm{VG}$ commuting with the projection to $X \times \mathrm{VG}$
and a finite definable partition of $U_{\rm par}$ such that we have the following for each part $A_j$.
We set $H_{\rm par} := H \circ \sigma^{-1} \in {\mathscr C}^{\mathrm{exp}}(U_{\rm par})$.
\begin{enumerate}
\item The set $A_j$ is a Presburger Cell over $\@ifnextchar_{\RFss@}{\RFss@@{}} \times X$, i.e.
\[
A_j = \{(z,\mu) \in B_j \times \mathrm{VG} \mid \alpha_{j}(z) \leq \mu \le \beta_{j}(z) \wedge \mu \equiv e_j\bmod n_j \}
\]
where $e_j$ and $n_j > 0$ are integers, $B_j \subset \@ifnextchar_{\RFss@}{\RFss@@{}} \times X$
is the projection of $A_j$, $\alpha_j$ is either a definable function $B_j\to\mathrm{VG}$ or constant equal to $-\infty$ and
$\beta_j$ is either a definable function $B_j\to\mathrm{VG}$ or constant equal to $+\infty$.
\item If $\alpha_j \ne -\infty$, there exists an integer $N \ge 1$ and a definable function
$\alpha^0_j \colon X \to \mathrm{VG}$ such that $\alpha_j(\xi,x) - \alpha^0_j(x) \in [0, \operatorname{ord}(N)]$ for all $(\xi,x) \in B_j$;
and similarly for $\beta_j$. (And without loss, $N$ can be taken the same for all $\alpha_j$ and $\beta_j$.)
\item
There are finitely many functions $c_{ij}$ in ${\mathscr C}^{\mathrm{exp}}(B_j)$ and distinct pairs $(a_{ij},b_{ij})$ with $a_{ij}$ nonnegative integers and $b_{ij}$ rational numbers such that for all $F\in {\mathrm{Loc}}_{ \gg 1}$, for each $\psi \in {\mathcal D}_F$ and each $(z,\mu)\in A_{j,F}$,
\begin{equation}\label{eq.repar}
H_{{\rm par}, F,\psi} (z,\mu) = \sum_{i} c_{ij,F,\psi}(z) \mu^{a_{ij}} q_F^{b_{ij}\mu}.
\end{equation}
\end{enumerate}
(The map $\sigma$ is called a reparameterization.)
\end{prop}
\begin{proof}
Let $f_i\colon U \to \mathrm{VG}$ be the definable functions into the value group appearing in the definition of $H$.
Apply \cite[Corollary~5.2.3]{CHallp} to $U$ and to those functions $f_i$.
This yields a reparameterization $\sigma\colon U \to U_{\rm par} \subset \@ifnextchar_{\RFss@}{\RFss@@{}} \times X \times \mathrm{VG}$ and a
finite definable partition
of $U_{\rm par}$ such that
each $f_{i,\rm par} := f_i \circ \sigma^{-1}$ is linear over $\@ifnextchar_{\RFss@}{\RFss@@{}} \times X$ on each part $A_j$,
i.e., $f_{i,\rm par}(z, \cdot)\colon (A_j)_z \to \mathrm{VG}$ is linear for each fixed $z \in \@ifnextchar_{\RFss@}{\RFss@@{}} \times X$, with the coefficient of $\mu$ independent of $F$ and $z$.
This already implies (3).
By refining the partition using Presburger cell decomposition, we obtain (1). Finally, we
apply \cite[Corollary~5.2.4]{CHallp} to the functions $\alpha_j$ and $\beta_j$ bounding the cells.
This yields a partition of each $B_j$ such that, after using that partition to refine the partition $A_j$,
(2) holds.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{for:all:large:mu}]
It is enough to find a definable $\mu_0$ as in the ``moreover'' part. Indeed, using such a $\mu_0$,
$Q$ can be expressed as
\begin{equation}\label{eq:falm}
Q(x) \iff \forall \mu \ge \mu_0(x)\colon P(x, \mu),
\end{equation}
which is a ${\mathscr C}^{\mathrm{exp}}$-condition by Proposition~\ref{locbasicop} (1), (2) and (3) (since it can be written as
$\forall \mu \in \mathrm{VG}\colon (\mu < \mu_0(x) \vee P(x, \mu))$).
Let $H$ in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG})$ be a function such that $P$ holds iff $H = 0$.
To obtain $\mu_0$, we apply Proposition~\ref{repar} (with $U = X \times \mathrm{VG}$); we also use the notation from there.
We may reduce to the case where $H_{\rm par}$ is non-zero only on a single part $A_j$. Indeed,
the function $H_j(x, \mu)$ which is equal to $H(x,\mu)$ iff $\sigma(x, \mu) \in A_j$ and $0$ otherwise
is of ${\mathscr C}^{\mathrm{exp}}$-class, and we can define $\mu_0$ to be the maximum of the $\mu_{0,j}$ corresponding to the $H_j$.
If the upper bound $\beta_{j}$ defining $A_j$ is not $+\infty$, then using
Proposition~\ref{repar} (2) we can take $\mu_0(x) := \beta^0_j(x) + \operatorname{ord}(N) + 1$,
so now assume $\beta_{j} = +\infty$. In that case, we claim that we can take $\mu_0(x) = 0$.
More precisely, we claim that for each $x$, the set $C_x := \{\mu \mid H(x,\mu) \ne 0\}$ either is empty or has no upper bound.
From
\[
H_{{\rm par}} (z,\mu) = \sum_{i} c_{ij}(z) \mu^{a_{ij}} q^{b_{ij}\mu}
\qquad\text{for } (z,\mu) \in A_j
\]
(see (\ref{eq.repar})) and using that all pairs $(a_{ij},b_{ij})$ are distinct,
we obtain that for each $z \in B_j$, the set
\[C'_z := \{\mu \in \mathrm{VG} \mid (z,\mu) \in A_j \wedge H_{\rm par}(z,\mu) \ne 0\}\]
is empty if $c_{ij}(z) = 0$ for each $i$, and unbounded otherwise. Now the claim follows, using
that $C_x$ is a union of sets of the form $C'_z$.
\end{proof}
Using Proposition~\ref{for:all:large:mu}, we can turn global ${\mathscr C}^{\mathrm{exp}}$-properties of functions
into local ones. As an example, we prove several variants of: local constancy is
a property of ${\mathscr C}^{\mathrm{exp}}$-class.
\begin{cor}[Local constancy]\label{cor.loc.const}
Let $f$ be in ${\mathscr C}^{\mathrm{exp}}(X\times Y)$, for some definable sets $X$ and $Y\subset \mathrm{VF}^n$ for some $n > 0$. Then we have the following
(where for each $F$, on $Y_F\subset F^n$ we put the topology induced by the one on $F^n$).
\begin{enumerate}
\item The set of $x \in X$ for which the function $y \mapsto f(x, y)$ is locally constant is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item The set of $(x,y) \in X \times Y$ such that the function $f(x, \cdot)$ is constant on a neighborhood of $y$
is a ${\mathscr C}^{\mathrm{exp}}$-locus.
Moreover, the radius of constancy can be chosen definably,
i.e., there exists a definable function $\varphi:X\times Y\to \mathrm{VG}$ such that
for all $F \in {\mathrm{Loc}}_{\gg1}$, for all $\psi \in {\mathcal D}_F$,
for all $x \in X_F$ and for all $y \in Y_F$, if $f_{F,\psi}(x, \cdot)$ is constant on a neighborhood of $y$, then it is constant on the intersection of $Y_F$ with the ball
of valuative radius $\varphi_F(x, y)$ around $y$ in $F^n$.
\end{enumerate}
\end{cor}
\begin{proof}
(1) follows from (2) and Proposition~\ref{locbasicop}, since $f(x, \cdot)$ being locally constant can be expressed by
\[
\forall y \in Y\colon \text{$f(x,\cdot)$ is constant on a neighborhood of $y$}.
\]
(2) That $f(x, \cdot)$ is constant on a neighborhood of $y$ can be expressed by
\[
\forall \lambda \gg 1\colon f(x,\cdot) \text{ is constant on $Y\cap B_\lambda(y)$},
\]
where $B_\lambda(y)$ is the ball around $y$ of valuative radius $\lambda$, which again
is a condition of ${\mathscr C}^{\mathrm{exp}}$-class by Propositions~\ref{locbasicop} and \ref{for:all:large:mu} (and, if one wants,
Corollary~\ref{cor.const}).
Moreover, the witness $\mu_0$ for the quantifier ``$\forall \lambda \gg 1$'' provided by Proposition~\ref{for:all:large:mu}
is the desired definable function $\varphi$.
\end{proof}
In particular, we obtain a transfer principle for local constancy:
\begin{cor}[Transfer for local constancy]\label{trans.loc.cons}
Corollary~\ref{trans.cons} still holds if one replaces
``constant'' by ``locally constant''.
\end{cor}
\begin{proof}
Apply Theorem~\ref{thm.trans} to ``$f(x, \cdot)$ is locally constant'', which is a ${\mathscr C}^{\mathrm{exp}}$-property (about $x$)
by Corollary~\ref{cor.loc.const}.
\end{proof}
\subsection{Integration}
We put the Haar measure on $F$ so that ${\mathcal O}_F$ has measure $1$; on
${\rm RF}_{m,F}$ and on ${\mathbb Z}$, we use the counting measure and for $X\subset \mathrm{VF}^*\times \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*$ we use the measure on $X_F$ induced by the product measure on $\mathrm{VF}_F^*\times \@ifnextchar_{\RFss@}{\RFss@@{}}_F\times {\mathbb Z}^*$. Likewise, we put the discrete topology on ${\mathbb Z}$ and on $\@ifnextchar_{\RFss@}{\RFss@@{}}_F$, the valuation topology on $F$, the product topology on $F^*\times \@ifnextchar_{\RFss@}{\RFss@@{}}_F\times {\mathbb Z}^*$, and the subset topology on $X_F$.
Maybe the most important aspect of ${\mathscr C}^{\mathrm{exp}}$-functions is that they have nice and natural properties related to integration, see e.g.~the following theorem stating stability under integration, generalizing Theorem 9.1.4 of \cite{CLexp}, by including also $p$-adic fields of small residue field characteristic.
\begin{thm}[{Integration; \cite[Theorems 4.1.1, 4.4.2]{CHallp}}]\label{thm:mot.int.}
Let $f$ be in ${\mathscr C}^{\mathrm{exp}}(X\times Y)$, for some definable sets $X$ and $Y$.
\begin{enumerate}
\item The set of $x$ such that $y\mapsto f(x,y)$ is $L^1$-integrable over $Y$ is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item There exists a function $I$ in ${\mathscr C}^{\mathrm{exp}}(X)$ such that
$$I(x) = \int_{y\in Y}f(x,y)\,|dy|$$
whenever $y\mapsto f(x,y)$ is $L^1$-integrable over $Y$. In more detail,
for every $F \in {\mathrm{Loc}}_{\gg1}$, every $\psi \in {\mathcal D}_F$ and
every $x \in X_F$, if $y \mapsto f_{F,\psi}(x, y)$ is $L^1$-integrable, then $I_{F,\psi}(x) = \int_{y\in Y_{F,\psi}}f_{F,\psi}(x,y)\,|dy|$.
\end{enumerate}
\end{thm}
Note that Theorem~\ref{thm:mot.int.} immediately implies a generalized version, where the
set we are integrating over is allowed to depend on $x$: If $f$ is a
${\mathscr C}^{\mathrm{exp}}$-function on $W \subset X \times Y$ (and $W_x \subset Y$ denotes the fiber of $W$ at $x \in X$), then
$\int_{W_x} f(x, y)\,|dy| = \int_{Y} f(x, y)\cdot 1_W(x,y)\,|dy|$, where $1_W$ is the characteristic function of $W$.
Thus by applying the theorem to $f(x, y)\cdot 1_W(x,y)$, we obtain that $L^1$-integrability of $f(x,\cdot)$ on
$W_x$ is a ${\mathscr C}^{\mathrm{exp}}$-condition, and the value of that integral (where it exists) is equal to $I(x)$ for some function $I$ of ${\mathscr C}^{\mathrm{exp}}$-class.
In \cite{CGH}, we proved that the condition that $f(x,\cdot)$ is locally integrable is also a ${\mathscr C}^{\mathrm{exp}}$-condition of the corresponding class.
This can be deduced from Theorem~\ref{thm:mot.int.} in a similar way as for local constancy (Corollary~\ref{cor.loc.const}).
However note also that since integrability on all small balls is equivalent to integrability on all balls, one doesn't even
need Proposition~\ref{for:all:large:mu}.
\section{Bounds and approximate suprema}
\label{sec:bounds}
\subsection{Results about bounds and approximate suprema}
One result of \cite{CGH} (generalized to the present context in \cite{CHallp}) is that a ${\mathscr C}^{\mathrm{exp}}$-function being bounded is a ${\mathscr C}^{\mathrm{exp}}$-condition
(see Theorem~\ref{thm:fam} (1)). In this section, we describe how bounds depend on the field $F$ and on additional parameters;
this generalizes results from Appendix B of \cite{ShinTemp}.
We give three main results on bounds, in increasing strength and complexity, with Theorem \ref{thm:fam:gen} containing the core result, with the hardest and longest proof of this paper.
\begin{thm}[Bounds]\label{thm:presburger-fam}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(W \times \mathrm{VG}^n)$,
where $W$ is a definable set and $n\geq 0$. Then
there exist an integer $b$ and a definable $d \in \mathrm{VG}$ such that for all $F\in {\mathrm{Loc}}_{ \gg1}$, all $\psi \in {\mathcal D}_F$ and all $\lambda\in {\mathbb Z}^n$, the following holds.
If the function
$$
w\mapsto |H_{F,\psi} (w,\lambda) |_{{\mathbb C}}
$$
is bounded on $W_F$, then one actually has
\begin{equation}\label{upper}
| H_{F,\psi} (w,\lambda) |_{{\mathbb C}} \le q_F^{ b \|\lambda\| + d} \mbox{ for all } w \in W_F,
\end{equation}
where $\|\lambda\| = \sum_{i=1}^n |\lambda_i|_{\mathbb C}$ and where $|\cdot|_{\mathbb C}$ is the norm on ${\mathbb C}$.
\end{thm}
By a definable $d \in \mathrm{VG}$, we mean a definable set $d=(d_F)_{F\in{\mathrm{Loc}}_M}$ where $d_F$ is a singleton in $\mathrm{VG}_F$ for each $F$, where we identify $d_F$ with the integer $d_F^0$ with $d_F=\{d_F^0\}$. Since we do not use parameters in our language until the appendix, a definable $d\in\mathrm{VG}$ can be bounded by $a+\operatorname{ord} (c)$ for some integers $a$ and $c \ge 1$ depending on $d$ but not on $F$, that is, $d_F\leq a+\operatorname{ord}_F (c)$ for each $F$; indeed, this follows from quantifier elimination \cite[Theorem 5.1.1]{CHallp}. If one is moreover only interested in $F$ of sufficiently big residue characteristic, then
one may assume $\operatorname{ord}_F (c)$ to be zero, so that the exponent in (\ref{upper}) becomes $b\|\lambda\| + a$ for some integers $a,b$ not depending on $F$.
In the appendix, $d$ can be more general depending on the constants added to ${\mathcal L}_{\mathrm {DP}}$, and on the requirements (axioms) put on the interpretations of the constant symbols.
We observe that in the case with $n=0$, and still using that we do not use parameters,
Theorem \ref{thm:presburger-fam} yields that
for any ${\mathscr C}^{\mathrm{exp}}$-function $H$ on $W$, there exist integers $a, c \ge 1$ such that
for any $F$ (in ${\mathrm{Loc}}_{\gg1}$) and any $\psi$ for which $H_{F,\psi}$ is bounded, the bound can be taken to be $q_F^{a + \operatorname{ord}_F(c)}$.
Note also that $\operatorname{ord}_F(c)$ is relevant only when one is interested in uniformity in $F$ for small residue characteristic and
arbitrary (but finite) ramification.
Part (2) of the following theorem generalizes Theorem~\ref{thm:presburger-fam}
by allowing the domain of $\lambda$ to be an arbitrary definable set $X$ (instead of just $\mathrm{VG}^n$).
In part (1), we recall the result from \cite{CGH,CHallp} about boundedness, since it fits nicely with part (2).
\begin{thm}[Bounds]\label{thm:fam}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(W \times X)$,
where $W$ and $X$ are definable sets.
\begin{enumerate}
\item
The set of $x \in X$ such that the function $w\mapsto | H(w,x) |_{{\mathbb C}}$ is bounded is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item
There exists a definable function
$\alpha:X\to\mathrm{VG}$ such that for every $F \in {\mathrm{Loc}}_{ \gg 1}$, every $\psi \in {\mathcal D}_F$ and every
$x \in X_F$, if the function
$$
w\mapsto | H_{F,\psi}(w,x) |_{{\mathbb C}}
$$
is bounded on $W_F$, then one has
$$
| H_{F,\psi}(w,x) |_{{\mathbb C}} \le q_F^{\alpha_F(x)} \mbox{ for all } w \in W_F.
$$
\end{enumerate}
\end{thm}
This theorem implies Theorem~\ref{thm:presburger-fam} as follows.
\begin{proof}[Proof of Theorem \ref{thm:presburger-fam}]
Applying Theorem \ref{thm:fam} (2) with $X = \mathrm{VG}^n$ yields a definable function $\alpha:X\to\mathrm{VG}$. To obtain Theorem~\ref{thm:presburger-fam},
we need to know that we can replace $\alpha_F(\lambda)$ by $b\|\lambda\| + d_F$ for some definable $d$ and some integer $b$.
Note that when allowing big ramification, a definable function $\alpha$ is not necessarily piecewise linear.
However, by \cite[Corollary~5.2.5]{CHallp}, and up to working piecewise with finitely many definable pieces $X_i \subset X$, it differs from a linear definable function $\ell\colon X_i \to \mathrm{VG}$ by at most $\operatorname{ord}(c)$ for some integer $c \ge 1$. More precisely, for every $F \in {\mathrm{Loc}}_{\gg 1}$ and every $\lambda \in X_{i,F}$, $|\alpha_F(\lambda) - \ell_F(\lambda)| \le \operatorname{ord}_F(c)$,
where $c$ does not depend on $F$.
Moreover, we may assume that the coefficients of $\ell_F$ (which are rational numbers) are independent of $F$ and the constant term is given by a definable constant $d \in \mathrm{VG}$.
Thus on each piece $X_i$, we obtain a bound of the form $b\|\lambda\| + d$. Finally we take maxima of the so-obtained finite collection of $b$ and $d$, yielding a bound of the desired form.
\end{proof}
In general, the supremum over $w$ of $| H(w,x) |_{{\mathbb C}}$ is not a ${\mathscr C}^{\mathrm{exp}}$-function in $x$.
However, it can be reasonably well approximated by the square root of a ${\mathscr C}^{\mathrm{exp}}$-function.
The precise statement is the following.
\begin{thm}[Approximate suprema]\label{thm:fam:gen}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(W \times X)$,
where $W$ and $X$ are definable sets.
Then there exist a definable $d \in \mathrm{VG}$ and a function $G$ in ${\mathscr C}^{\mathrm{exp}}(X)$
taking non-negative real values such that
the following holds: For each $F \in {\mathrm{Loc}}_{\gg1}$, each $\psi \in D_F$ and each $x \in X_F$ for which
the function
\begin{equation*
w\mapsto H_{F,\psi}(w,x)
\end{equation*}
is bounded on $W_F$, one has
\begin{equation}\label{eq:sq1}
\sup_{w\in W_F} |H_{F,\psi}(w,x)|_{\mathbb C}^2 \leq G_{F,\psi}(x) \le q_F^{d_F} \sup_{w\in W_F} |H_{F,\psi}(w,x)|_{\mathbb C}^2.
\end{equation}
\end{thm}
The remark about the definable $d\in\mathrm{VG}$ just below Theorem \ref{thm:presburger-fam} also applies to the $d$ of Theorem \ref{thm:fam:gen}.
We do not believe that Theorem \ref{thm:fam:gen} would hold without squaring, i.e.,
with $\sup_{w\in W} |H(w,x) |_{{\mathbb C}}$ instead of $\sup_{w\in W} |H(w,x) |^2_{{\mathbb C}}$.
Indeed, using a one-point set as $W$, this would imply that the absolute value of a
${\mathscr C}^{\mathrm{exp}}$-function can be approximated by a ${\mathscr C}^{\mathrm{exp}}$-function, which already seems
unlikely for the function $H \in {\mathscr C}^{\mathrm{exp}}(\mathrm{VG}^2)$, $H(x,y) = x^2 - y$.
Theorem \ref{thm:fam:gen} implies Theorem~\ref{thm:fam} as follows.
\begin{proof}[Proof of Theorem \ref{thm:fam}]
(1) is the Bdd-part of \cite[Theorem~4.4.2]{CHallp}.
(2) Apply Theorem \ref{thm:fam:gen} to $|H|^2_{{\mathbb C}}$ to obtain a (real-valued) function $G$ in ${\mathscr C}^{\mathrm{exp}}(X)$ such that
$\sup_{w\in W} |H(w,x) |^2_{{\mathbb C}}\leq G(x)$ whenever $H(\cdot,x)$ is bounded.
It remains to show that any $G$ can be bounded by $q^{\alpha(x)}$
for some definable $\alpha$ (as functions depending on $F \in {\mathrm{Loc}}_{\gg1}, \psi \in {\mathcal D}_F, x \in X_F$).
(We will more generally bound $|G(x)|_{{\mathbb C}}$ for arbitrary $G \in {\mathscr C}^{\mathrm{exp}}(X)$, i.e., not necessarily real valued.)
Using the definition of ${\mathscr C}^{\mathrm{exp}}$, one easily reduces to the case $G \in {\mathscr C}(X)$. Indeed, by that definition,
\[
G(x)=\sum_{i=1}^N G_{i}(x)\Big( \sum_{y \in Y_{i,x}}\psi\big( g_{i}(x,y) { + e_{i}(x,y)/n_i } \big)\Big)
\]
for some $G_i\in {\mathscr C}(X)$, some definable sets $Y_i \subset X \times \@ifnextchar_{\RFss@}{\RFss@@{}}$ and some definable functions $g_i:X\to \mathrm{VF}$ and $e_i:X'\to {\rm RF}_{n_i}$.
The complex norm of the inner sum is bounded by a product of some $\#{\rm RF}_{m_{ij}} = q^{1+\operatorname{ord}(m_{ij})}$ for each $i$, so it remains to bound the $|G_i(x)|_{{\mathbb C}}$.
So now suppose $G \in {\mathscr C}(X)$. By definition of ${\mathscr C}$, we have
\[
G(x)=\sum_{i=1}^N \# Y_{i,x} \cdot q^{\alpha_{i}(x)} \cdot \big( \prod_{j=1}^{N'} \beta_{ij}(x) \big) \cdot \big( \prod_{\ell=1}^{N''} \frac{1}{1-q^{a_{i\ell}}} \big)
\]
for some definable $Y_i \subset X \times \@ifnextchar_{\RFss@}{\RFss@@{}}$, $\alpha_i, \beta_{ij}\colon X \to {\mathbb Z}$ and some non-zero integers $a_{i\ell}$.
Clearly, each of the factors of the terms of the sum can be bounded as desired (using again $\#{\rm RF}_{m} = q^{1+\operatorname{ord}(m)}$), and hence so can the sum.
\end{proof}
\subsection{Proof of Theorem \ref{thm:fam:gen}}\label{theproofs}
The proof of Theorem \ref{thm:fam:gen}
is based on two deep
results, which we develop first:
one which allows us to reduce to functions living on $\@ifnextchar_{\RFss@}{\RFss@@{}}$ and $\mathrm{VG}$, and
another one which allows us to find approximate suprema of functions on $\mathrm{VG}$.
A key result of \cite{CGH} is its Proposition~4.5.8; the following is the generalization
of that result to the present context given in \cite{CHallp}. (Here, for simplicity, we
state a slightly weakened version.)
\begin{prop}[{\cite[Proposition~4.6.1]{CHallp}}]
\label{IML}
Let $m\geq 0$ be an integer, let $X$ and $U \subset X\times \mathrm{VF}^m$ be definable,
and let $H$ be in ${\mathscr C}^{\mathrm{exp}}(U)$. Write $x$ for variables running over $X$ and $y$ for variables running over $\mathrm{VF}^m$.
Then there exist integers $N\geq 1$, $d\geq 0$, a definable surjection $\varphi:U\to V\subset X\times \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*$ over $X$, definable functions $h_{i}:U\to \mathrm{VF}$, and functions $G_{i}$ in ${\mathscr C}^{\mathrm{exp}}(V)$ for $i=1,\ldots,N$,
such that the following conditions hold for each $F\in {\mathrm{Loc}}_{\gg1}$ and each $\psi\in {\mathcal D}_F$;
we omit the indices $F,\psi$ to keep the notation lighter.
\begin{enumerate}
\item One has
$$
H(x,y) = \sum_{i=1}^{N}G_{i}(\varphi(x,y))\psi(h_{i}(x,y));
$$
\item if one sets, for $(x,r)\in V$,
$$
U_{x,r} := \{y\in U_{x}\mid \varphi(x,y)=(x,r)\}
$$
and
$$
W_{x,r}:=\{y \in U_{x,r}\mid
\max_{i} | G_{i}(x,r)|_{{\mathbb C}} \leq |H(x,y)|_{{\mathbb C}} \},
$$
then
\[
\operatorname{Vol}(U_{x,r}) \leq q^d \cdot \operatorname{Vol}(W_{x,r}) < +\infty,
\]
where the volume $\operatorname{Vol}$ is taken with respect to the Haar measure on $\mathrm{VF}^m$.
\end{enumerate}
Here, as usual, $U_{x}$ is the set of $y\in \mathrm{VF}^m$ such that $(x,y)$ lies in $U$, and, for the map $\varphi$ to ``be over $X$'' means that $\varphi$ makes a commutative diagram with the natural maps to $X$.
\end{prop}
Note that (2) in particular implies that if $U_{x,r}$ has positive measure, then
\begin{equation}\label{eq.max<sup}
\max_{i} | G_{i}(x,r)|_{{\mathbb C}} \leq \sup_{y\in U_{x,r}}|H(x,y)|_{{\mathbb C}}.
\end{equation}
By virtue of Proposition \ref{IML}, one can simplify the domain of a ${\mathscr C}^{\mathrm{exp}}$-function, while preserving most of its behavior on growth and size, as follows.
\begin{cor}\label{from.exp.to.e.IML}
Let $W$, $X$, and $U\subset W\times X$ be definable sets and let $H$ be in ${\mathscr C}^{\mathrm{exp}}(U)$.
Then there exists a nonnegative integer $N$, a definable surjection
$$
\varphi:U\to V\subset X\times \@ifnextchar_{\RFss@}{\RFss@@{}} \times \mathrm{VG}^*
$$
over $X$ and a non-negative real valued function $\tilde H$ in ${\mathscr C}^{\mathrm{exp}}(V)$ such that
for all $F\in {\mathrm{Loc}}_{ \gg1}$, $\psi \in {\mathcal D}_F$, $x\in X_{F}$ and $v \in V_{F,x}$, one has
\begin{equation}\label{eq.tildeH}
\frac{1}{N} \tilde H_{F,\psi}(v) \leq \sup_{\substack{ w \in W_F\\ \varphi_F(w,x)=v } } |H_{F,\psi}(w,x)|^2_{{\mathbb C}} \leq N \tilde H_{F,\psi}(v),
\end{equation}
where $V_x$ is the fiber of $V$ above $x$, and where by the right hand inequality, we in particular mean that the supremum is finite.
\end{cor}
\begin{proof}[Proof of Corollary~\ref{from.exp.to.e.IML}]
Write $W\subset \mathrm{VF}^m\times W_0$ for some $m$ and some $W_0 \subset \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*$. If $m=0$ there is nothing to prove. We proceed by induction on $m$.
Apply Proposition~\ref{IML} to $H$ (where the $X$ from the proposition is $X \times W_0$), yielding integers $N,d$, a definable surjection $\varphi: U\to V\subset X\times \@ifnextchar_{\RFss@}{\RFss@@{}} \times \mathrm{VG}^*$ and $G_i$ in ${\mathscr C}^{\mathrm{exp}}(V)$ for $i=1,\ldots,N$.
We may assume that each fiber of $\varphi$ has positive measure, by treating
the fibers of measure zero separately, using induction on $m$;
in particular, (\ref{eq.max<sup}) holds, i.e., $\max_i |G_i(v)|_{{\mathbb C}} \le \sup_{w, \varphi(w,x)=v} |H(w, x)|_{{\mathbb C}}$.
Now set
$$
\tilde H(v) := \sum_{i=1}^{N} |G_{i}(v)|_{\mathbb C}^2 \quad \mbox{for $v$ in } V.
$$
Then $\tilde H$ is as desired (with this $N$): The left hand inequality of (\ref{eq.tildeH}) of follows from (\ref{eq.max<sup}),
and the right hand inequality follows from $H(w, x) = \sum_i G_i(\varphi(w, x))\psi(h_i(w,x))$ and $|\psi(h_i(w,x))| = 1$.
\end{proof}
We now come to the second ingredient to the proof of Theorem \ref{thm:fam:gen}.
Suppose that $H$ is a bounded ${\mathscr C}^{\mathrm{exp}}$-function with domain in the value group, say $H \in {\mathscr C}^{\mathrm{exp}}(W)$ for some definable $W \subset \mathrm{VG}$.
To obtain an approximate maximum of $H$,
we will choose a finite subset $W_0 \subset W$ such that $H$ takes its maximum on $W_0$, up to some factor $m$. We will need to be able to
do this in families, in such a way that $m$ and $\#W_0$ do not depend on the parameters, and the elements of $W_0$ depend definably on the parameters.
After various simplifications, what we end up needing are the following two lemmas:
Lemma~\ref{lem:Z-ubd} which is used in the case where $W$ is infinite, and Lemma~\ref{lem:Z-bd} which is used in the case where $W$ is finite but growing in size (when varying the family members).
\begin{lem}\label{lem:Z-ubd}
Suppose that for $i = 1, \dots, k$, we have $a_i \in {\mathbb N}$ and $b_i \in {\mathbb Z}$. Then there exist positive integers $m$ and $\ell$
such that the following holds for every tuple $c = (c_1, \dots, c_k) \in {\mathbb C}^k$ and every $q \ge 2$.
Suppose that the function
\[
h(w) := \sum_i c_i w^{a_i} q^{b_i w}
\]
is bounded on ${\mathbb N}$; then
we have \[\sup_{w \in {\mathbb N}} |h(w)| \le m\cdot \max_{0 \le w \le \ell} |h(w)|.\]
\end{lem}
\begin{lem}\label{lem:Z-bd}
Suppose that for $i = 1, \dots, k$, we have $a_i \in {\mathbb N}$ and $b_i \in {\mathbb Z}$. Then there exist positive integers $m$ and $\ell$, and
Presburger definable functions $w_1, \dots, w_\ell\colon {\mathbb N} \to {\mathbb N}$ with $w_i(t) \le t$ for all $i$ and $t$
and such that the following holds for every tuple $c = (c_1, \dots, c_k) \in {\mathbb C}^k$, every $t \in {\mathbb N}$ and every $q \ge 2$.
Consider the function
\[
h(w) := \sum_i c_i w^{a_i} q^{b_i w} \quad\mbox{for } w\in{\mathbb N};
\]
then we have \[\max_{0 \le w \le t} |h(w)| \le m\cdot \max_{1 \le i \le \ell} |h(w_i(t))|.\]
\end{lem}
We will prove both lemmas together.
\begin{proof}[Proof of Lemmas \ref{lem:Z-ubd} and \ref{lem:Z-bd}]
We start working on Lemma~\ref{lem:Z-bd}. First note that we may impose lower bounds on $t$, by taking some more Presburger functions covering the whole interval when $t$ is smaller; we will do this whenever convenient.
We will treat three special cases, namely when all $b_i$ are negative, when all $b_i$ are positive, and when all $b_i$ are zero.
This will then be put together to obtain the result for arbitrary $b_i$. For the latter to work, we will prove slightly stronger statements in the special cases, namely the following three claims.
\medskip
Claim $-1$: If all $b_i$ are negative, then there exists $\ell \in {\mathbb N}$ such that for every $c = (c_i)_i$ and every $q \ge 2$, there exists $w_{-1} \in \{0, \dots, \ell-1\}$ such that
\begin{equation}
\tag{$*_{-1}$}
|h(w)| \le q^{(\ell-w)/2} \cdot |h(w_{-1})| \qquad \text{for every } w \ge 0
\end{equation}
\medskip
Claim $0$: If all $b_i$ are zero, then there exist $\ell,m \in {\mathbb N}$ such that for every $c = (c_i)_i$, every $q\ge 2$ and every sufficiently big $t$, there exists $w_0 \in \{s, 2s, \dots, (\ell - 1)s\}$, where $s = \lfloor \frac{t}{\ell} \rfloor$ such that
\begin{equation}
\tag{$*_{0}$}
|h(w)| \le m \cdot |h(w_0)| \qquad \text{for every } 0 \le w \le t
.
\end{equation}
\medskip
The third claim follows from Claim $-1$ by replacing $h(w)$ by $h(t-w)$:
Claim $1$: If all $b_i$ are positive, then there exists $\ell \in {\mathbb N}$ such that for every $c = (c_i)_i$, every $t$ and every $q \ge 2$, there exists $w_{1} \in \{t-\ell+1, \dots, t\}$ such that
\begin{equation}
\tag{$*_{1}$}
|h(w)| \le q^{(\ell-t+w)/2} \cdot |h(w_{1})| \qquad \text{for every } w \le t
.
\end{equation}
\medskip
Before we prove those claims, here is how they imply Lemma~\ref{lem:Z-bd}.
Write $h(w)$ as a sum $h_{-1}(w) + h_{0}(w) + h_{1}(w)$, according to the sign of the $b_i$, and
apply the corresponding claims to $h_{-1}$, $h_0$ and $h_1$.
All the possible values of $w_{-1}, w_0, w_1$ appearing in the three claims are Presburger functions in $t$; these are the functions we use to conclude the lemma, so we need
to prove that for any tuple $c$, there exists a $\nu \in \{-1, 0, 1\}$ such that $\max_w |h(w)| \le m'\cdot |h(w_\nu)|$
for some constant $m'$. The idea for this is that using that the bound on $h_{-1}$ is decreasing and the one on $h_{1}$ is increasing, we can deduce that for one of the $\nu = -1,0,1$, $|h_\nu(w_\nu)|$ dominates the other two summands of $h$ at $w_\nu$; we then can bound
$\max_{w} |h(w)|$ in terms of $h(w_\nu)$. Here are the details:
We may assume that the values $\ell$ obtained from Claims~$-1$ and 1 are the same.
By imposing a lower bound on $t$, we can ensure that
\begin{equation}\label{eq:t-big}
\ell + 4 \le w_0 \le t - \ell - 4.
\end{equation}
Indeed, we have $w_0 \ge s > t/\ell - 1$, so imposing $t \ge \ell\cdot(\ell + 5)$ implies the left hand inequality;
the right hand inequality is obtained in a similar way using $w_0 \le (\ell - 1)s \le t - t/\ell$.
It is sufficient to bound $\max_{\ell \le w \le t-\ell} |h(w)|$, since all $w$ outside of this interval are equal to one of the definable functions anyway.
Let $\nu \in \{-1, 0, 1\}$ be such that
$B_\nu$ is maximal among
\[
B_{-1} := |h_{-1}(w_{-1})|,\quad
B_0 := 3m\cdot |h_{0}(w_{0})|,\quad
B_1 := h_{1}(w_{1}).
\]
We will prove that for this choice of $\nu$, we have
\begin{equation}\label{eq:h-hnu}
h_{\nu}(w_{\nu}) \le 3 h(w_{\nu}).
\end{equation}
This then implies, for $\ell \le w \le t-\ell$:
\begin{align*}
|h(w)| \le\,\,& |h_{-1}(w)| + |h_{0}(w)| + |h_{1}(w)|
\\
\overset{\hskip-6ex(*_{-1}), (*_{0}), (*_{1})\hskip-6ex}{\le}\,\,&
\hskip4ex |h_{-1}(w_{-1})| + m|h_{0}(w_{0})| +
|h_{1}(w_{1})|
\\
\le\,\,& m' \cdot |h_\nu(w_\nu)|
\le 3m'\cdot |h(w_\nu)|
\end{align*}
for some suitable multiple $m'$ of $m$ ($m' = 7m$ works independently of $\nu$);
this implies the lemma. Let us now prove
(\ref{eq:h-hnu}).
Suppose first that $\nu = -1$.
Then we have
\begin{equation}\label{eq.ineq}
\begin{aligned}
|h_0(w_{-1})| &\overset{(*_{0})}{\le} m\cdot |h_{0}(w_{0})|
\le \textstyle{\frac13} |h_{-1}(w_{-1})| \qquad\text{and}
\\
|h_1(w_{-1})| &\overset{(*_{1})}{\le} q^{(\ell-t+w_{-1})/2}\cdot |h_{-1}(w_{-1})| \le \textstyle{\frac13}
|h_{-1}(w_{-1})|,
\end{aligned}
\end{equation}
where the last inequality follows from $t - \ell \overset{(\ref{eq:t-big})}{\ge} \ell + 8 \ge w_{-1} + 9$.
This yields
\begin{equation}\label{eq.inecomp}
\begin{aligned}
|h(w_{-1})| &= |h_{-1}(w_{-1}) + h_{0}(w_{-1}) + h_{1}(w_{-1})|\\
&\ge |h_{-1}(w_{-1})| - |h_{0}(w_{-1})| - |h_{1}(w_{-1})|\\
&\overset{(\ref{eq.ineq})}{\ge} |h_{-1}(w_{-1})|\cdot (1 - \frac13 - \frac13)
\end{aligned}
\end{equation}
and hence $|h_{-1}(w_{-1})| \le 3|h(w_{-1})|$.
The case $\nu = 1$ works in exactly the same way.
Finally, suppose $\nu = 0$. Then we get
\[
|h_{-1}(w_0)| \overset{(*_{-1})}{\le}
q^{(\ell-w_0)/2} \cdot |h(w_{-1})|
\le \textstyle{\frac13}|h(w_{-1})|,
\]
using $w_0 \ge \ell + 4$. Analogously, we get
$|h_{1}(w_0)| \le \frac13|h(w_{-1})|$,
and hence, by the same computation as in (\ref{eq.inecomp}),
$|h_{-1}(w_{0})| \le 3|h(w_{0})|$.
This finishes the proof that the three claims imply Lemma~\ref{lem:Z-bd}.
\medskip
Before proving the claims, we carry out similar (but simpler) arguments for Lemma~\ref{lem:Z-ubd}.
Again, we write
$h(w)$ as a sum $h_{-1}(w) + h_{0}(w) + h_{1}(w)$, according to the sign of the $b_i$. Since the conclusion of Lemma~\ref{lem:Z-ubd} only speaks about tuples $c$ for which $h$ is bounded,
we may assume that $h_0$ is constant and $h_1$ vanishes entirely.
We apply Claim~$-1$ to $h_{-1}$, we don't need Claim~$1$,
and instead of using Claim~0, we simply use $w_0 := \ell + 4$ (for the $\ell$ from Claim~$-1$). Then the same arguments as for Lemma~\ref{lem:Z-bd} yield
\[
\max_{w} |h(w)| \le m'\cdot \max_{w \le \ell + 1} |h(w)|.
\]
(This time, we do a case distinction on which of
$B_{-1} := |h_{-1}(w_{-1})|$ and $B_0 := 2\cdot |h_{0}(w_{0})|$ is bigger.)
Thus for both lemmas, it remains to prove the three claims. More precisely, it suffices to prove Claims 0 and $-1$, since Claim 1 is equivalent to Claim $-1$.
\medskip
\emph{Proof of Claim 0}.
The function $h(w) = \sum_i c_i w^{a_i}$ is a polynomial of degree $d := \max_i a_i$.
Let $V$ be the vector space of all polynomials of degree $d$ and consider the map
\[
\phi\colon V \to {\mathbb C}^{d+1},
f \mapsto \big(f(\frac{1}{2d+4}), f(\frac{2}{2d+4}), \dots, f(\frac{d+1}{2d+4})\big).
\]
Set $S := \{f \in V \mid \|\phi(f)\|_\infty = 1\}$,
i.e., the preimage of the unit sphere in ${\mathbb C}^{d+1}$ with respect to the maximum norm.
Since $\phi$ is injective, $S$ is compact, so the maximum
\[
m := \max\{|f(\lambda)| \mid f \in S, 0 \le \lambda \le 1\}
\]
exists. From this, we deduce that Claim 0 holds for $\ell = d + 2$.
Indeed, set $f(\lambda) := h(\alpha\lambda)$, where
\[
\alpha := (2d+4)\cdot s = 2(d+2) \lfloor \frac{t}{d+2}\rfloor \ge 2(d+2) (\frac{t}{d+2} - 1).
\]
By imposing $t \ge 2(d+2)$, we obtain $\alpha \ge t$, and hence $\alpha^{-1}w \in [0,1]$
for $0 \le w \le t$. Thus
\[
|h(w)| = |f(\alpha^{-1} w)| \le m\cdot \|\phi(f)\|_\infty
= m\cdot \max_{j=1, \dots, d+1} |h(js)|,
\]
which is what we had to show.
\medskip
\emph{Proof of Claim $-1$}.
Set $A := \max_i a_i$, $B := \max_i -b_i$ and $I = \{0, 1, \dots, A\} \times \{-1, -2, \dots, -B\}$. We can write $h(w)$ as
\[
h(w) = \sum_{(a,b) \in I} c_{a,b}w^a q^{bw}.
\]
Let $c = (c_{a,b})_{(a,b) \in I}$ be the tuple of all coefficients.
Set $\ell := (A+1)\cdot B$ and $x := (h(0), \dots, h(\ell-1)) \in {\mathbb C}^\ell$.
We will find an $N$ depending only on $A$ and $B$ (but not on $c$)
such that
\begin{equation}\label{eq:c1-xinfty}
\|c\|_1 \le q^N\cdot \|x\|_\infty,
\end{equation}
where $\|\cdot\|_1$ and $\|\cdot\|_\infty$ are the usual norms on ${\mathbb C}^\ell$. This then implies, for all $w \ge 0$,
\[
\begin{array}{rc@{}c@{}c}
|h(w)| \le& \|c\|_1 &\,\,\cdot\,\,& \max_{(a,b) \in I} w^a q^{bw}\\
\le& q^N \max_{w' < \ell} |h(w')| &\,\,\cdot\,\,& w^A q^{-w}\\
\le&
\multicolumn{3}{l}{q^{(\ell' - w)/2} \cdot \max_{w' < \ell'} |h(w')|,}
\end{array}
\]
where the last inequality is ensured by choosing $\ell'$ big enough, namely such
that $N + A \log_q w \le \ell'/2 + w/2$ holds for all $w \ge 0$ and $q \ge 2$.
Thus (\ref{eq:c1-xinfty}) implies Claim~$-1$.
To obtain (\ref{eq:c1-xinfty}), instead of bounding $\|c\|_1$, we may as well bound $\|c\|_\infty$ (since they
differ at most by a factor $\ell$). We have $x = Zc$, where $Z = (z_{w,(a,b)})_{0\le w<\ell, (a,b) \in I}$
is the matrix with coefficients
\[
z_{w,(a,b)} = w^aq^{bw}.
\]
Suppose for the moment that $Z$ is invertible. Then, by definition of the operator norm $\|Z^{-1}\|_\infty$,
we have $\|c\|_\infty \le \|Z^{-1}\|_\infty \cdot \|x\|_\infty$, so it suffices to find an
upper bound on $\|Z^{-1}\|_\infty$ of the form $q^N$,
where $N$ only depends on $A$ and $B$.
(Note that the entire matrix $Z$ only depends on $A$, $B$ and $q$.)
Up to a constant factor (namely $\ell$),
$\|Z^{-1}\|_\infty$ is bounded by the maximum of the absolute values of the entries of $Z^{-1}$.
These entries are of the form $\det(Z')/\det(Z)$, where $Z'$ is a minor of $Z$. Since
$\det(Z')$ is a polynomial in the entries of $Z$ and those entries are bounded (even independently of $q$,
since all exponents $bw$ are non-positive),
we have an upper bound on $|\det(Z')|$, and it remains to find a lower bound on $|\det(Z)|$ of the form $q^{-N}$;
this at the same time will prove that $Z$ is invertible.
Considering $q$ as an indeterminate, we have $\det(Z) \in {\mathbb Z}[q^{-1}] \subseteq {\mathbb Z}[q, q^{-1}]$.
It is enough to prove that $\det(Z) \ne 0$ when considered as such a Laurent polynomial.
Indeed, this then implies that for $q$ big enough, we have $|\det(Z)| > \alpha\cdot q^{-N}$
for some fixed $\alpha > 0$, where $-N$ is the least negative power of $q$ appearing in $\det(Z)$.
We have
\begin{equation}\label{eq:detZ}
\det(Z) = \sum_{\sigma} \operatorname{sgn}(\sigma)\prod_{(a,b) \in I} z_{\sigma(a,b),(a,b)}
\end{equation}
where $\sigma$ runs over all bijections $I \to \{0, \dots, \ell - 1\}$.
(We fix an order on $I$ for $\operatorname{sgn}(\sigma)$ and the sign of $\det(Z)$ to be well-defined.)
Each summand of the sum (\ref{eq:detZ}) is a monomial in $q$, namely of degree
\[
d(\sigma) := \sum_{(a,b) \in I}b\sigma(a,b) = \sum_{-B \le b \le -1} b\cdot\sum_{0 \le a \le A} \sigma(a,b).
\]
Let $d_0$ be the smallest (i.e., most negative) degree in $q$ appearing among the summands in (\ref{eq:detZ}).
To prove $\det(Z) \ne 0$,
we will prove that the sum $R$ of the monomials of degree $d_0$ in $q$ is non-zero.
Let us write $\{0, \dots, \ell - 1\}$ as a disjoint union of $B$ many intervals $J_b$ of length $A$:
\[
J_b := \{n \in {\mathbb Z} \mid (A+1)(-1-b) \le n \le (A+1)(-1-b) + A\},
\]
for $b = -1, \dots, -B$.
The degree $d(\sigma)$ is minimal if and only if,
for every $(a,b), (a', b') \in I$ with $b < b'$, we have $\sigma(a,b) > \sigma(a', b')$.
This is equivalent to $\sigma(\cdot, b)$ sending $\{0, \dots, A\}$ to $J_b$ for every $b$.
Write $S_b$ for the set of bijections $\{0, \dots, A\} \to J_b$.
Then the sum of the monomials of minimal degree in (\ref{eq:detZ}) is (maybe up to sign)
\begin{align*}
R &= \sum_{\sigma_{-1} \in S_{-1}} \dots \sum_{\sigma_{-B} \in S_{-B}} \operatorname{sgn}(\sigma_{-1})\cdots\operatorname{sgn}(\sigma_{-B})
\prod_{(a,b) \in I} z_{\sigma_b(a),(a,b)}\\
&= \sum_{\sigma_{-1} \in S_{-1}} \dots \sum_{\sigma_{-B} \in S_{-B}}
\prod_b
\bigg(\operatorname{sgn}(\sigma_{b})
\prod_{a} z_{\sigma_b(a),(a,b)}\bigg)\\
&=
\prod_b\sum_{\sigma_{b} \in S_{b}}
\bigg(\operatorname{sgn}(\sigma_{b})
\prod_a z_{\sigma_b(a),(a,b)}\bigg)
\\
&= q^{d_0} \prod_b
\sum_{\sigma_b \in S_b}\bigg( \operatorname{sgn}(\sigma_b)\prod_a \sigma_b(a)^a\bigg).
\end{align*}
Each factor in the product over $b$ is a Vandermonde determinant
(corresponding to a polynomial of degree $A$ evaluated at each element of $J_b$) and hence non-zero.
Thus $R \ne 0$, which is what remained to be proven.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:fam:gen}]
By Corollary \ref{from.exp.to.e.IML}, it is enough to consider
a non-negative real-valued $H \in {\mathscr C}^{\mathrm{exp}}(W \times X)$ for $W\subset \@ifnextchar_{\RFss@}{\RFss@@{}}\times \mathrm{VG}^*$ and find a
non-negative real-valued $G \in {\mathscr C}^{\mathrm{exp}}(X)$ and a definable $d \in \mathrm{VG}$ such that
\begin{equation}\label{eq:sq:prf}
\sup_{w \in W} H(w, x) \le G(x) \le q^{d} \sup_{w \in W} H(w, x).
\end{equation}
It will be handy to consider a slightly more general situation, namely where $H$ lives on a definable
subset $U \subset W \times X$ and where the suprema run over $w \in U_x$.
By a recursive procedure, it is enough to only consider the situations where $W\subset {\rm RF}_n$ for some $n$ or $W \subset \mathrm{VG}$.
In the first case, we set
$$
G(x) := \sum_{w\in U_x} H(w,x),
$$
(which lies in ${\mathscr C}^{\mathrm{exp}}(X)$ by Theorem~\ref{thm:mot.int.}).
This $G(x)$ is obviously at least as big as $\sup_{w \in U_x}H(w,x)$, and it exceeds the supremum
at most by a factor $\#{\rm RF}_n = q^{\operatorname{ord}(n) + 1}$, so it is as desired.
In the case $W \subset \mathrm{VG}$, the idea is to use a rectilinearization result as in \cite{CPres} to reduce to
Lemmas \ref{lem:Z-ubd} and \ref{lem:Z-bd}, though to deal with our setting allowing highly ramified fields,
we need rectilinearization in the form stated in \cite[Proposition~5.2.6]{CHallp}.
The details are as follows.
By \cite[Proposition~5.2.3]{CHallp}, after possibly introducing new residue ring variables
(which we can later get rid of again as explained above), and after a finite partition of $U$,
we may suppose that ``$H$ has linear ingredients'', i.e., all functions $U \to \mathrm{VG}$ involved in the definition of $H$ depend linearly on $w$.
Moreover, by \cite[Proposition~5.2.6]{CHallp}
we may suppose that either $U_x = {\mathbb N}$ for all $x$ or $U_x = \{w \in \mathrm{VG} \mid 0 \le w \le \alpha(x)\}$
for some definable function $\alpha:X\to \mathrm{VG}$; this involves introducing more new residue ring variables,
another finite partition of $U$, and applying a bijection which is linear over $X$.
(The linearity of this bijection ensures that $H$ still has linear ingredients.)
That $H$ has linear ingredients means that there exist $k$ and $a_i,b_i$ ($1 \le i \le k$) such that
for any $F$, $\psi$ and $x$, the map $w\mapsto H_{F,\psi}(w,x)$ is of the form
\[
h(w) := \sum_i c_i w^{a_i} q^{b_i w}
\]
for some $c_i$ depending on $F$, $\psi$ and $x$. Thus we can either apply
Lemma \ref{lem:Z-ubd} (if $U_x = {\mathbb N}$) or Lemma \ref{lem:Z-bd} with $t = \alpha_F(x)$. In
both cases, we obtain an integer $m \ge 1$ and finitely many definable functions $w_1, \dots, w_\ell\colon X \to \mathrm{VG}$
with $w_i(x) \in U_x$ such that
\begin{equation}\label{lemZresult}
\sup_{w \in U_x}H(w,x) \le m\cdot\max_{1\le i\le \ell} H(w_i(x),x)
\end{equation}
for all $x$. (In the $U_x = {\mathbb N}$ case, we take $w_i$ to be constant equal to $i - 1$.) Now set
$$
G(x) := m \sum_{i=1}^\ell H(w_i(x),x).
$$
Then we have
\[
\sup_{w \in U_x} H(w, x) \overset{(\ref{lemZresult})}{\le} G(x) \le m\cdot \ell\cdot \sup_{w \in W} H(w, x).
\]
So this $G$ is as desired, given that $m\cdot \ell$ can easily be bounded by an integer power of $q$.
\end{proof}
\section{Limits of ${\mathscr C}^{\mathrm{exp}}$-functions}
\label{sec:limits}
This section contains various results about limits, namely: Existence of limits in various contexts is given by
${\mathscr C}^{\mathrm{exp}}$-conditions, and the limit itself is a ${\mathscr C}^{\mathrm{exp}}$-function in the parameters.
We consider both, classical pointwise limits, and $L^p$-limits. Even though we obtain
a whole variety of different results, it should be noted that certain other, subtly different
statements do not seem to hold; see the remark after Theorem~\ref{limits}.
\subsection{Pointwise limits and continuity}
The first result states that limits being $0$ is a ${\mathscr C}^{\mathrm{exp}}$-condition and that convergence cannot
be arbitrarily slow.
\begin{thm}[$0$-limits]\label{thm:limits0}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG}^n)$,
where $X$ is a definable set and where $n\geq 0$.
\begin{enumerate}
\item
The set $\{x \in X \mid \lim\limits_{\lambda,\ \|\lambda\|\to +\infty} |H (x,\lambda) |_{{\mathbb C}} = 0\}$ is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item
There exists a rational number $r<0$ and a definable function $\alpha: X\to \mathrm{VG}$, such that for each $F\in {\mathrm{Loc}}_{ \gg1}$, for each $\psi \in {\mathcal D}_F$ and each $x\in X_F$, the following holds.
If one has
\begin{equation}\label{limits:upper000}
\lim\limits_{\lambda,\ \|\lambda\|\to +\infty} |H_{F,\psi} (x,\lambda) |_{{\mathbb C}} =0 ,
\end{equation}
then one actually has
\begin{equation*
| H_{F,\psi} (x,\lambda) |_{{\mathbb C}} \le q_F^{ r \|\lambda\|} \mbox{ for all $\lambda \in {\mathbb Z}^n$ with $\| \lambda \|>\alpha_F(x)$}.
\end{equation*}
\end{enumerate}
\end{thm}
The above result also holds for more general kinds of limits:
\begin{thm}[$0$-limits]\label{thm:limits0:gen}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times Y)$,
where $X$ and $Y$ are definable sets, and let $\gamma:Y\to \mathrm{VG}_{\ge 0}$ be a surjective definable function.
\begin{enumerate}
\item
The set $\{x \in X \mid \lim\limits_{y\in Y,\ \gamma(y)\to +\infty} |H (x,y) |_{{\mathbb C}} = 0\}$ is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item
There exists a rational number $r<0$ and a definable function $\alpha: X\to \mathrm{VG}$ such that for all $F\in {\mathrm{Loc}}_{ \gg 1}$, for each $\psi \in {\mathcal D}_F$ and each $x\in X_F$, the following holds.
If one has
\begin{equation}\label{limits:upper0}
\lim\limits_{y\in Y_F,\ \gamma_F(y)\to +\infty} |H_{F,\psi} (x,y) |_{{\mathbb C}} =0 ,
\end{equation}
then one actually has
\begin{equation}\label{limits:upper}
| H_{F,\psi} (x,y) |_{{\mathbb C}} \le q_F^{ r \gamma_F(y) } \mbox{ for all $y\in Y_F$ with $\gamma_F(y)>\alpha_F(x)$}.
\end{equation}
\end{enumerate}
\end{thm}
For some $F, \psi, x \in X_F$, the limits appearing in Theorems~\ref{thm:limits0} and \ref{thm:limits0:gen} might not even exist.
(The theorems do not assume that they do.)
The next result states that existence of limits is also a ${\mathscr C}^{\mathrm{exp}}$-condition and that
if a limit exists, then its value is given by a ${\mathscr C}^{\mathrm{exp}}$-function.
\begin{thm}[Limits]\label{limits}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VG}^n)$ for some definable set $X$.
\begin{enumerate}
\item
The set $\{x \in X \mid \lim\limits_{\|\lambda\|\to +\infty} H (x,\lambda) \text{ exists in ${\mathbb C}$}\}$ is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\item
There exists $G$ in ${\mathscr C}^{\mathrm{exp}}(X)$ such that the following holds for all $F\in {\mathrm{Loc}}_{ \gg1}$, for each $\psi \in {\mathcal D}_F$ and each $x\in X_F$.
$$
\mbox{If }
\lim_{\|\lambda\|\to+\infty} H_{F,\psi} (x,\lambda) \mbox{ exists in ${\mathbb C}$},
$$
then
$$
G_{F,\psi}(x) = \lim_{\|\lambda\|\to+\infty} H_{F,\psi} (x, \lambda).
$$
\end{enumerate}
\end{thm}
Again, we can ask for a version of this for more general limits. We obtain:
\begin{thm}[Limits]\label{limits:gen}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times Y)$,
where $X$ and $Y$ are definable sets, and let $\gamma:Y\to \mathrm{VG}_{\ge 0}$ be a surjective definable function. Then
the set $\{x \in X \mid \lim\limits_{y\in Y,\ \gamma(y)\to +\infty} H (x,y) \text{ exists in }{\mathbb C}\}$ is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\end{thm}
Note that surprisingly, in this generality it is not straightforward to prove that the value of the limit is of ${\mathscr C}^{\mathrm{exp}}$-class,
and we even doubt that this is true. It seems that such a result would require some better understanding of the number of rational
points of families of varieties over finite fields. On the other hand, it should be possible to obtain the generalization
in a slightly altered context, namely after adding function symbols for Skolem functions to the sorts $\@ifnextchar_{\RFss@}{\RFss@@{}}$, or, after adding more denominators like in the rational motivic functions of \cite{Kien:rational}.
From Theorem~\ref{limits:gen}, one easily deduces that continuity is a ${\mathscr C}^{\mathrm{exp}}$-condition.
In the following, given a definable set $Z \subset \mathrm{VF}^n$, we write $\bar Z$ for its topological closure
(i.e., $\bar{Z}_F$ is the topological closure of $Z_F$ in $F^n$ for each $F$), which is also definable.
\begin{cor}[Continuity]\label{cor.cont}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times Y)$, where $X$ and $Y$ are definable sets and $Y \subset \mathrm{VF}^n$.
Then
\[\{(x,y) \in X \times Y \mid H(x, \cdot) \text{ is continuous at } y\}\]
and
\begin{align*}
\{(x,y) \in X \times (\bar{Y}\setminus Y) \mid \,&H(x, \cdot) \text{ has an extension to $Y\cup \{y\}$,}\\
&\text{continuous at $y$} \}
\end{align*}
are ${\mathscr C}^{\mathrm{exp}}$-loci.
\end{cor}
\begin{proof}
The conditions at $(x,y)$ can be expressed respectively as follows:
\[
\begin{cases}
\lim_{y' \in Y, \operatorname{ord}(y' - y) \to +\infty} (H(x, y') - H(x,y)) = 0; \\
\lim_{y' \in Y, \operatorname{ord}(y' - y) \to +\infty} H(x, y') \text{ exists}.
\end{cases}
\]
Now use Theorem~\ref{thm:limits0:gen} and Proposition~\ref{locbasicop}.
\end{proof}
Note that if $H \in {\mathscr C}^{\mathrm{exp}}(Y)$ is continuous on $Y$ and if $\lim_{y' \in Y, y' \to y} H(x, y')$ exists for each $y\in\bar{Y}\setminus Y$, then the unique continuous extension of $H$ is not known by us to be of ${\mathscr C}^{\mathrm{exp}}$-class, for the same reasons as explained after Theorem~\ref{limits:gen}.
However, it is of ${\mathscr C}^{\mathrm{exp}}$-class in many slightly more restrictive cases. The most general setting in which this can be proven would probably be
very technical, so in the following, we just prove it under reasonable assumptions.
\begin{prop}[Continuous extension]\label{prop.cont.ext}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times Y)$, where $X$ and $Y$ are definable sets and $Y \subset \mathrm{VF}^n$ and suppose that $\bar Y$ is clopen (i.e., $\bar Y_F \subset F^n$ is open and closed for every $F \in {\mathrm{Loc}}_{\gg 1}$).
Then $H$ can be extended to a ${\mathscr C}^{\mathrm{exp}}$-function $\bar H$ on $X \times \bar Y$ such that for every $(x,y) \in X \times \bar Y$,
if $\lim_{y' \in Y, y' \to y} H(x, y')$ exists, then $\bar H(x,\cdot)$ is continuous at $y$.
Namely, with all indices, for every
$F \in {\mathrm{Loc}}_{\gg1}$, for every $\psi \in {\mathcal D}_F$ and for every $(x,y) \in X_F \times \bar Y_F$,
if $\lim_{y' \in Y_F, y' \to y} H_{F,\psi}(x, y')$ exists, then $\bar H_{F,\psi}(x,\cdot)$ is continuous at $y$.
\end{prop}
\begin{proof}
We can simply define
\[
\bar H(x,y) := \lim_{\lambda \to \infty} q^{-n\lambda} \int_{B_\lambda(y) \cap Y} H(x,y') dy',
\]
where $B_\lambda(y) \subset \mathrm{VF}^n$ is the closed ball of valuative radius $\lambda$ around $y$ (considered as a definable set).
More precisely, we use Theorems~\ref{thm:mot.int.} and \ref{limits} to find an $\bar H$ such that
for every $F, \psi$ and every $x \in X_F, y \in Y_F$, we have
\[
\bar H_{F,\psi}(x,y) = \lim_{\lambda \to \infty} q_F^{-n\lambda} \int_{B_{F,\lambda}(y) \cap Y_F} H_{F,\psi}(x,y') dy',
\]
whenever the integral and the limit exist, where $B_{F,\lambda}(y)$ is the concrete closed ball in $F^n$ given by $\lambda \in {\mathbb Z}$ and $y \in F^n$.
Since $\bar Y_F$ is clopen, for sufficiently
big $\lambda$, we have $B_{F,\lambda}(y) \subset \bar Y_F$, so that we are just averaging over this ball (where $\bar Y_F\setminus Y_F$ has measure zero by dimension considerations).
If $H_{F,\psi}(x,\cdot)$ is continuous at $y$, then the integral exists for large $\lambda$, and continuity also implies that the limit of those average values exists
and is equal to $\lim_{y' \in Y_F, y' \to y} H_{F,\psi}(x, y')$, as desired.
\end{proof}
As usual (namely, by transfer for ${\mathscr C}^{\mathrm{exp}}$-conditions) the above theorems imply corresponding transfer principles for limits, as follows.
\begin{cor}[Transfer principle for limits and continuity]\label{trans:lim}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times Y)$,
where $X$ and $Y$ are definable sets, and let $\gamma:Y\to \mathrm{VG}$ be a definable function.
Then there exists $M$ such that, for any $F\in {\mathrm{Loc}}$ of residue characteristic $\ge M$,
the truth of each of the following statements depends only on (the isomorphism class of) the residue field of $F$;
here, in the 3rd and 4th statement, we assume $Y \subset \mathrm{VF}^n$ for some $n$.
\[
\lim\limits_{ y\in Y_F,\ \gamma_F(y) \to +\infty} H_{F,\psi} (x,y) =0 \text{ for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$};
\]
\[
\lim\limits_{ y\in Y_F,\ \gamma_F(y) \to +\infty} H_{F,\psi} (x,y) \text{ exists for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$};
\]
\[
H_{F,\psi}(x,\cdot) \text{ is continuous on $Y_F$ for all $x \in X_F$};
\]
\[
H_{F,\psi}(x,\cdot) \text{ has a continuous extension to ${\bar Y}_F$ for all $x \in X_F$.}
\]
\end{cor}
The proofs of Theorems \ref{thm:limits0:gen} and \ref{limits} use Theorem \ref{thm:fam:gen} in a crucial way, to reduce to the case
of a limit over a single $\mathrm{VG}$-variable; other proofs, relying on Proposition \ref{IML} above and the rectilinearization result given by Proposition~5.2.6 of \cite{CHallp}, may also be thinkable. Here is the single $\mathrm{VG}$-variable version of Theorems \ref{thm:limits0:gen} and \ref{limits} (formulated in a more concise way):
\begin{thm}[Limits]\label{limits:basic}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG}_{\ge0})$, where $X$ is a definable set. Then we have the following.
\begin{enumerate}
\item The condition ``$\lim_{\lambda \to +\infty} H(x,\lambda)$ exists'' is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$.
\item There exists a function $G \in {\mathscr C}^{\mathrm{exp}}(X)$, a definable function $\alpha\colon X \to \mathrm{VG}_{\ge0}$ and a rational number $r < 0$
such that for every $F \in {\mathrm{Loc}}$, every $\psi \in {\mathcal D}_F$ and every $x \in X_F$, if the limit $\lim_{\lambda \to +\infty} H_{F,\psi}(x,\lambda)$ exists, then
\begin{equation}\label{upper-basic}
|H_{F,\psi}(x,\lambda) - G_{F,\psi}(x)|_{\mathbb C} \le q_F^{ r \cdot\lambda }
\end{equation}
for every $\lambda \ge \alpha_F(x)$. In particular, the limit is equal to $G_{F,\psi}(x)$.
\end{enumerate}
\end{thm}
This theorem implies Theorems~\ref{thm:limits0} -- \ref{limits:gen} as follows.
\begin{proof}[Proof of Theorems \ref{thm:limits0} and \ref{thm:limits0:gen}]
Theorem \ref{thm:limits0} follows from \ref{thm:limits0:gen} (using $Y = \mathrm{VG}^n$ and $\gamma(\lambda) = \|\lambda\|$), so we now prove
Theorems \ref{thm:limits0:gen}.
We may suppose that $Y=Y_0\times \mathrm{VG}_{\ge0}$ and that $\gamma$ is the coordinate projection to the $\mathrm{VG}$-variable.
By Theorem~\ref{thm:fam:gen}, there exists a real-valued function $H'$ in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG}_{\ge0})$
such that $H'(x,\lambda)$ differs from $\sup_{y_0\in Y_0} |H(x,y_0,\lambda)|^2_{\mathbb C}$ by at most a factor
of the form $q^{d}$, for some definable $d \in \mathrm{VG}$.
This implies that we can without loss replace $H$ by $H'$, considering the limit $\lim_{\lambda \to +\infty} H'(x,\lambda)$;
in this version, the Theorem follows pretty directly from Theorem~\ref{limits:basic}.
Indeed, that the limit is $0$ can be expressed by the condition ``$\lim_{\lambda \to +\infty} H'(x,\lambda)$ exists
and $G(x) = 0$'', which is of ${\mathscr C}^{\mathrm{exp}}$-class, where $G$ is the function obtained by applying Theorem~\ref{limits:basic} to $H'$,
and Claim~(\ref{limits:upper}) follows from (\ref{upper-basic}).
\end{proof}
\begin{proof}[Proof of Theorems \ref{limits} and \ref{limits:gen}]
Define $H'$ in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG}_{\ge0} \times Y \times Y)$ by
\[
H'(x, \lambda, y_1, y_2) =
\begin{cases}
H(x, y_1) - H(x, y_2) & \text{if } \gamma(y_1), \gamma(y_2) \ge \lambda\\
0 & \text{otherwise}.
\end{cases}
\]
By applying Theorem~\ref{thm:fam:gen} to $H'$, we find $G'$ in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VG}_{\ge0})$
such that \change
}
this $G'(x,\lambda)$ differs from $\sup_{y_1, y_2} |H(x, y_1) - H(x, y_2)|_{\mathbb C}^2$ by at most a factor
not depending on $\lambda$, where $y_1$ and $y_2$ both run over $\{y \in Y \mid \gamma(y) \ge \lambda\}$.
In particular we have that $\lim_{\lambda \to +\infty}G'(x,\lambda) = 0$ iff $H(x, \cdot)$
is a Cauchy sequence in the sense of the $\gamma$-limit. Thus the limit $\lim\limits_{y\in Y,\ \gamma(y)\to +\infty} H (x,y)$
exists if and only if $\lim_{\lambda \to +\infty}G'(x,\lambda) = 0$, which is a ${\mathscr C}^{\mathrm{exp}}$-condition by Theorem~\ref{limits:basic}.
This finishes the proof of Theorem~\ref{limits:gen} and of Theorem~\ref{limits} (1).
To obtain a function $G$ as desired in Theorem~\ref{limits} (2),
we can simply take the limit of $H(x, \lambda)$ along any sequence of $\lambda$ with $\|\lambda\| \to \infty$,
so we obtain it e.g.\ by applying Theorem~\ref{limits:basic} (2) to
$$
\lim_{\mu \to +\infty} H(x, \mu, 0, \dots, 0),
$$
where $\mu$ runs over $\mathrm{VG}_{\ge 0}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{limits:basic}]
We apply Proposition~\ref{repar} to $H$, yielding (using the same notation as there)
a reparameterization $\sigma\colon U = X \times \mathrm{VG}_{\ge0} \to U_{\rm par} \subset \@ifnextchar_{\RFss@}{\RFss@@{}} \times X \times \mathrm{VG}_{\ge0}$
and a partition of $U_{\rm par}$ into sets
\begin{equation}\label{lb.A}
A_j = \{(z,\mu) \in B_j \times \mathrm{VG} \mid \alpha_{j}(z) \leq \mu \le \beta_{j}(z) \wedge \mu \equiv e_j\bmod n_j \}
\end{equation}
(for some definable $B_j \subset \@ifnextchar_{\RFss@}{\RFss@@{}} \times X$, $\alpha_j, \beta_j\colon B_j \to \mathrm{VG} \cup \{\pm \infty\}$ and some integers $e_j \ge 0, n_j > 0$) such that for $(z, \mu) \in A_j$, we have
\begin{equation}\label{lb.H}
H(\sigma^{-1}(z,\mu)) =
H_{{\rm par}} (z,\mu) = \sum_{i} c_{ij}(z) \mu^{a_{ij}} q^{b_{ij}\mu}
\end{equation}
for some $a_{ij} \in {\mathbb N}$, $b_{ij} \in {\mathbb Q}$ and some $c_{ij} \in {\mathscr C}^{\mathrm{exp}}(B_{ij})$.
Denote the preimage of $A_j$ under $\sigma$ by
\[
A'_j := \sigma^{-1}(A_j) \subset U.
\]
For any fixed $F$ and $x \in X_F$, the fibers $A'_{j,F,x} = \{\lambda \in {\mathbb N} \mid (x,\lambda) \in A'_{j,F}\}$ form a partition of ${\mathbb N}$.
The limit $\lim_{\lambda \to +\infty}H_{F,\psi}(x,\lambda)$ exists if and only if, for each $j$ for which $A'_{j,F,x}$ is unbounded,
the corresponding restricted limit $\lim_{\lambda \to +\infty, \lambda \in A'_{j,F,x}}H_{F,\psi}(x,\lambda)$ exists and moreover all those limits are equal.
In particular, those $A_j$ for which $\beta_j \ne +\infty$ are irrelevant. Indeed, recall that for each $j$, either $\beta_j$ is constant $+\infty$, or, by
Proposition~\ref{repar}~(2), $A'_{j,F,x}$ is bounded for all $F,x$.
Therefore, in the remainder of the proof, we only consider those parts $A_j$ (and $A'_j$) for which $\beta_j = +\infty$.
After shrinking $U_{\rm par}$ to the union of those parts (and shrinking $U$ accordingly and restricting $H$ to $U$),
we can entirely get rid of the reparameterization, by replacing each $A_j$ by its preimage $A'_j$.
To see this, what we need to check is that we have analogues of (\ref{lb.A}) and (\ref{lb.H}) for those $A'_j$.
Those analogues can be obtained, provided that the restriction of $\sigma$ to $A'_j$
is of the form $\sigma(x, \mu) = (\tau(x), \mu)$ for some bijection $\tau \colon B'_j \subset X \to B_j$,
where $B'_j$ is the projection of $A'_j$ to $X$. Indeed, such a $\tau$ is automatically definable (using $\sigma$)
and hence (\ref{lb.A}) and (\ref{lb.H}) for $A'_j$ can be obtained by pre-composing $\alpha_j$ and $c_{ij}$ with $\tau$.
Since $\sigma$ is a reparameterization,
the fiber $(A'_j)_x = \{\mu \in \mathrm{VG} \mid (x, \mu) \in A'_j\}$ is equal
to a disjoint union of certain fibers $(A_j)_{\xi,x} = \{\mu \in \mathrm{VG} \mid (\xi,x,\mu) \in A_j\}$,
and what we need to check is that $(A'_j)_x$ is actually equal to a single fiber $(A_j)_{\tau(x)}$.
This follows from the fact that different $(A_j)_{\xi,x}$, $(A_j)_{\xi',x}$ cannot be disjoint, by
(\ref{lb.A}): both fibers contain all big $\mu \in \mathrm{VG}$ satisfying $\mu \equiv e_j \bmod n_j$.
Thus the reparameterization is indeed unnecessary, so from
now on we assume that
\[
A_j = \{(x,\mu) \in B_j \times \mathrm{VG} \mid \alpha_{j}(x) \leq \mu \wedge \mu \equiv e_j\bmod n_j \}
\]
for some $B_j \subset X$ and that, for $(x,\mu) \in A_j$, we have
\[
H(x,\mu) = \sum_{i} c_{ij}(x) \mu^{a_{ij}} q^{b_{ij}\mu}.
\]
We assume without loss that $a_{0j} = b_{0j} = 0$, and we
let $I$ be the collection of those $(i,j)$ satisfying either $b_{ij}>0$, or, $b_{ij}=0$ and $a_{ij}>0$.
Since for each fixed $j$ the pairs $(a_{ij},b_{ij})$ are distinct, given $x \in X$,
the limit exists if and only if, jointly,
\begin{equation}\label{limex1}
c_{ij}(x)=0 \mbox{ for all those $(i,j) \in I$ satisfying } x\in B_{j},
\end{equation}
and
\begin{equation}\label{limex2}
c_{0j}(x) = c_{0k}(x)
\text{ for all those $j$, $k$ satisfying $x\in B_{j}$, $x\in B_{k}$.}
\end{equation}
Each of these equations is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$, so their conjunction is also a ${\mathscr C}^{\mathrm{exp}}$-condition (by Proposition~\ref{locbasicop}), hence finishing the proof of (1).
When the limit exists, it is equal to the following ${\mathscr C}^{\mathrm{exp}}$-function:
$$
G(x) := c_{0j}(x)
$$
for any $j$ with $x\in B_{j}$ (e.g.\ the smallest $j$ for some linear ordering of the $j$).
It remains to find $r < 0$ and $\alpha\colon X \to \mathrm{VG}$. To this end, we may from now on without loss suppose that
all $b_{ij}$ are negative (by subtracting $G(x)$, and by ignoring those $x$ for which the limit does not exist).
In particular, the limit is $0$ if it exists.
We need to choose $r$ and $\alpha$ in such a way that we have
\begin{equation}\label{r-alpha}
|\sum_{i} c_{ij}(x) \lambda^{a_{ij}} q^{b_{ij}\lambda}|_{\mathbb C} \le q^{r\lambda}
\quad \text{for all $\lambda \ge \alpha(x)$.}
\end{equation}
Let $a_0$ and $b_0$ be the maximum of all $a_{ij}$ and of all $b_{ij}$, respectively, and set $r := b_0/2$. Then to obtain (\ref{r-alpha}), it
suffices to prove
\[
N \cdot (\max_{ij} |c_{ij,F,\psi}(x)|_{\mathbb C})\cdot \lambda^{a_0} \le q_F^{-r\lambda}
\]
for all $F, \psi, x$, where $N$ is the number of summands in (\ref{r-alpha}).
By the same argument as in the proof of Theorem~\ref{thm:fam} (or by actually applying Theorem~\ref{thm:fam} with $W$ being a singleton),
we obtain a bound of the form $|c_{ij,F,\psi}(x)|_{{\mathbb C}} \le q_F^{g_F(x)}$ for some definable $g\colon X \to \mathrm{VG}$.
From this, one can easily obtain an $\alpha$ as desired (e.g. choosing it such that $g_F(x) \le -r\alpha_F(x)/2$
and $N\lambda^{a_0} \le q_F^{-r\lambda/2}$ for $\lambda \ge \alpha_F(x)$).
\end{proof}
\subsection{Limits of functions}
In this section, we consider various types of limits of sequences of functions. For simplicity we restrict to functions
on $\mathrm{VF}^n$ (though the notions would also make sense on other definable sets, using the counting measures on $\mathrm{VG}$ and on ${\rm RF}_*$,
and the proofs also go through in this generalized setting). We ask the usual two questions: Does the limit exist and if yes, what is it?
Those results will be used to study Fourier transforms of $L^2$ functions in Section \ref{sec:Fourier}.
By uniform convergence of a sequence of functions $f_\mu$, $\mu \in {\mathbb N}$, we mean convergence with respect to the sup norm, and
by $L^p$-convergence (for $p \in {\mathbb R}_{\ge1} \cup \{\infty\}$), we mean convergence with respect to the usual $L^p$-norm.
(In particular, $L^{\infty}$-convergence means uniform convergence except on a subset of measure zero.) Note that we do not require
the individual functions $f_\mu$ to be $L^p$-integrable; instead, $L^p$-convergence towards a limit function $g$ means that
the $L^p$-norm of the difference $f_\mu - g$ exists for $\mu$ sufficiently big and that it goes to $0$.
\begin{thm}[Limits of sequences of functions]\label{L^pcom}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VF}^n\times \mathrm{VG}_{\ge0})$ for some definable set $X$ and some $n \ge 0$.
Given $F\in {\mathrm{Loc}}_{ \gg1}$, $\psi \in {\mathcal D}_F$ and $x\in X_F$, consider the sequence of functions (on $F^n$)
\begin{equation}\label{eq.seq}
y \mapsto H_{F,\psi}(x, y, \mu),
\end{equation}
indexed by integers $\mu\geq 0$.
\begin{enumerate}
\item
For each of the notions of convergence listed below, the set of $x \in X$ such that the sequence (\ref{eq.seq}) of functions converges in that sense is a ${\mathscr C}^{\mathrm{exp}}$-locus.
The notions of convergence are: pointwise convergence, uniform convergence, $L^2$-convergence, $L^\infty$-convergence.
\item
There exists a $G$ in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VF}^n)$ such that the following holds for all $F\in {\mathrm{Loc}}_{ \gg1}$, for each $\psi \in {\mathcal D}_F$ and each $x\in X_F$.
If the sequence (\ref{eq.seq}) of functions converges pointwise or uniformly or in $L^p$-norm for any $p \in {\mathbb R}_{\ge 1} \cup \{\infty\}$,
then the function
$$
y\mapsto G_{F,\psi}(x,y)
$$
is the limit of this sequence.
\end{enumerate}
\end{thm}
We believe that also part (1) of this theorem is true for $L^p$-convergence for any $p \in {\mathbb R}_{\ge 1} \cup \{\infty\}$. However, the proof would be more involved (even in the case $p = 1$).
From the theorem, we deduce various related results (some of which would also have simple direct proofs).
\begin{cor}[$L^p$-integrability]\label{cor.finLp}
Fix $p \in \{1, 2, \infty\}$ and let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VF}^n)$. Then the set of $x \in X$ for which
$H(x,\cdot)$ is $L^p$-integrable is a ${\mathscr C}^{\mathrm{exp}}$-locus.
\end{cor}
\begin{proof}
The case $p = 1$ is exactly Theorem~\ref{thm:mot.int.} (1). (We repeat the result for $p=1$ here only for completeness.)
For $p = 2, \infty$, we obtain the corollary by applying Theorem~\ref{L^pcom} to $H'(x,y,\mu) := q^{-\mu}H(x,y)$. Indeed, for any fixed $F, \psi, x$,
the $L^p$-limit $\lim_{\mu \to \infty} H'_{F,\psi}(x, \cdot, \mu)$ exists (and is $0$) if and only if $H_{F,\psi}(x,\cdot)$ is $L^p$-integrable.
\end{proof}
\begin{cor}[The ``almost everywhere'' quantifier]\label{cor.almost}
Suppose that $P(x, y)$ is a ${\mathscr C}^{\mathrm{exp}}$-condition, where $x$ runs over a definable set $X$ and $y$ runs over $\mathrm{VF}^n$.
Then
\[
P(x,y) \text{ holds for almost all } y \in \mathrm{VF}^n
\]
is a ${\mathscr C}^{\mathrm{exp}}$-condition on $x$. (Here, by ``almost all'', we mean that the complement has measure $0$.)
\end{cor}
\begin{proof}
Choose $H \in {\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n)$ such that $H(x,y)$ is $0$ if and only if $P(x,y)$ holds,
and set $H'(x,y,\mu) := q^{\mu} H(x,y)$. Then for any $F, \psi, x, y$, the sequence $H'_{F,\psi}(x,y,\mu)$ diverges for $\mu \to \infty$ if and only
$P_{F,\psi}(x,y)$ does not hold.
Thus $P_{F,\psi}(x,y)$ holds for almost all $y$ if and only if the $L^\infty$-limit of $H'_{F,\psi}(x,\cdot,\mu)$ exists,
so that the corollary follows by applying Theorem~\ref{L^pcom} to $H'(x,y,\mu)$.
\end{proof}
As usual, we also obtain some transfer results:
\begin{cor}[Transfer principles for limits of functions]\label{trans:funlim}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n \times \mathrm{VG}_{\ge 0})$ for some definable set $X$ and some $n \ge 0$.
Given $F\in {\mathrm{Loc}}_{ \gg1}$, $\psi \in {\mathcal D}_F$ and $x\in X_F$, consider the sequence of functions (on $F^n$)
\begin{equation}\label{eq.seq.trans}
y \mapsto H_{F,\psi}(x, y, \mu),
\end{equation}
indexed by integers $\mu\geq 0$.
Fix a notion of convergence among: pointwise convergence / uniform convergence / convergence in $L^2$-norm / in $L^\infty$-norm.
Then there exists $M$ such that, for any $F\in {\mathrm{Loc}}$ of residue characteristic $\ge M$,
the truth of each of the following statements depends only on (the isomorphism class of) the residue field of $F$.
\[
\text{The sequence (\ref{eq.seq.trans}) converges for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$};
\]
\[
\text{The sequence (\ref{eq.seq.trans}) converges to $0$ for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$}.
\]
\end{cor}
Concerning transfer of convergence to $0$, let $G$ be obtained from $H$ as in Theorem~\ref{L^pcom} (2). Then
convergence to $0$ can be expressed by a ${\mathscr C}^{\mathrm{exp}}$-condition as follows (using Theorem~\ref{L^pcom} (1) and Proposition~\ref{locbasicop}):
\[
\forall x\colon \big(\text{the sequence $H(x,\cdot,\mu)$ converges}
\wedge \forall y\colon G(x, y) = 0\big)
\]
\begin{cor}[Transfer principles for $L^p$-norms]\label{trans:Lplim}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n)$ for some definable set $X$ and some $n \ge 0$.
Then there exists $M$ such that, for any $F\in {\mathrm{Loc}}$ of residue characteristic $\ge M$,
the truth of each of the following four statements depends only on (the isomorphism class of) the residue field of $F$.
\[
\text{$H_{F,\psi}(x, \cdot)$ has finite $L^1$-/$L^2$-/$L^\infty$-norm
for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$};
\]
\[
\text{$H_{F,\psi}(x, \cdot)$ is $0$ on almost all of $F^n$,
for all $x\in X_F$ and all $\psi\in{\mathcal D}_F$}.
\]
\end{cor}
Let us now prove Theorem~\ref{L^pcom}.
One ingredient to the second part is that existence of the $L^p$-limit implies
almost everywhere existence of the point-wise limit. This is not true in general, but it is true for sequences of functions of ${\mathscr C}^{\mathrm{exp}}$-class.
Here is the precise statement:
\begin{lem}\label{Lp-vs-pointwise}
Fix $p \in {\mathbb R}_{\ge1} \cup \{+\infty\}$ and let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n\times \mathrm{VG}_{\ge0})$ for some definable set $X$.
Also fix $F \in {\mathrm{Loc}}_{\gg1}$, $\psi \in {\mathcal D}_F$ and $x \in X_F$.
If the sequence of functions $H_{F,\psi}(x, \cdot, \lambda)$ (on $F^n$) converges in $L^p$-norm for $\lambda \to \infty$,
then $H_{F,\psi}(x, y, \lambda)$ converges for almost all $y \in F^n$.
\end{lem}
(Without the assumption of $H$ being of ${\mathscr C}^{\mathrm{exp}}$-class, one can only find a subsequence which converges for almost all $y$.)
\begin{proof}[Proof of Lemma~\ref{Lp-vs-pointwise}]
We apply the same reasoning as in the proof of Theorem~\ref{limits:basic}, using the same notations (except that
our set $X \times \mathrm{VF}^n$ plays the role of the set $X$ from the proof of Theorem~\ref{limits:basic}), namely:
We find a reparameterization $\sigma\colon U = X \times \mathrm{VF}^n \times \mathrm{VG}_{\ge 0} \to U_{\rm par}$, a partition of $U_{\rm par}$
into finitely many $A_j$, we can disregard those $A_j$ that have an upper bound $\beta_j \ne +\infty$, and after that,
we can get rid of the reparameterization again in the same way as for Theorem~\ref{limits:basic}.
Fix $F$, $\psi$ and $x \in X_F$ for the entire remainder of the proof, and
suppose that there exists a set $Y \subset F^n$ of positive measure such that
$H_{F,\psi}(x, y, \cdot)$ does not converge for any $y \in Y$. For each such $y \in Y$, either (\ref{limex1}) or (\ref{limex2}) fails.
Thus, up to shrinking $Y$, either
(a) there exists an
$A_j$ with $\{x\} \times Y \subset B_j$ such that $c_{ij,F,\psi}(x,y) \not= 0$ for all $y$ in $Y$, where $(i,j) \in I$, or
(b)
there exist
$A_j$, $A_k$ with $\{x\} \times Y \subset B_j \cap B_k$ such that $c_{0j,F,\psi}(x,y) \ne c_{0k,F,\psi}(x,y)$ for all $y$ in $Y$.
If (a) holds, we fix $j$ and choose, among the different $i$ for which (a) holds, the one
corresponding to the dominant term, i.e., such that $b_{ij}$ is maximal, and among equal $b_{ij}$, the one such that $a_{ij}$
is maximal. Then for big $\lambda \equiv e_j \mod n_j$, the $L^p$-norm of $H_{F,\psi}(x, \cdot, \lambda)$ restricted to $Y$ is eventually larger than
half of the
$L^p$-norm of the dominant term $c_{ij,F,\psi}(x,y)\lambda^{a_{ij}}q_F^{b_{ij}\lambda}$ on $Y$
and hence diverges.
Now suppose that no $A_j$ as in (a) exists. Then we are in case (b), and $A_j$ and $A_k$ correspond to two different sub-sequences
of $H$, which, on $Y$, converge to two different functions, namely $c_{0j,F,\psi}(x,\cdot)$ resp.\ $c_{0k,F,\psi}(x,\cdot)$.
Again, this contradicts $L^p$-convergence of $H_{F,\psi}(x, \cdot, \cdot)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{L^pcom}]
We first prove (2). Using Theorem~\ref{limits}, we find a function
$G \in {\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n)$ such that
$$
G_{F,\psi}(x,y) = \lim_{\lambda\to\infty}H_{F,\psi}(x,y,\lambda)
$$
for every $F, \psi$, $x \in X_F$ and $y \in F^n$ for which that limit exists.
In particular, if the sequence $H_{F,\psi}(x,\cdot,\lambda)$ converges pointwise (or uniformly),
then the limit is $G_{F,\psi}(x,\cdot)$. If $H_{F,\psi}(x,\cdot,\lambda)$ converges in $L^p$-norm,
then $H_{F,\psi}(x,y,\lambda)$ converges for almost all $y$ by Lemma~\ref{Lp-vs-pointwise},
and for those $y$ for which it does, the limit is $G_{F,\psi}(x,y)$. Therefore, $G_{F,\psi}(x,\cdot)$
is also the $L^p$-limit.
We now consider (1) with the various notions of convergence. By replacing $H(x,y,\lambda)$
by $H(x,y,\lambda) - G(x,y)$ (with $G$ obtained from Part (2)), we may assume that if $\lim_{\lambda\to\infty}H_{F,\psi}(x,y,\lambda)$ exists,
then it is equal to $0$.
Pointwise convergence: This can be expressed as
\[
\forall y\colon \lim_{\lambda\to\infty}H_{F,\psi}(x,y,\lambda) \text{ exists},
\]
which is a ${\mathscr C}^{\mathrm{exp}}$-condition by Theorem~\ref{limits:basic}.
Uniform convergence: Apply Theorem~\ref{thm:limits0:gen} to $H$, where $\gamma\colon Y \times \mathrm{VG}_{\ge0} \to \mathrm{VG}$
just sends $(y,\lambda)$ to $\lambda$.
$L^2$-convergence: The function
\[
(x,\lambda) \mapsto \int_{\mathrm{VF}^n}|H(x,y,\lambda)|^2_{\mathbb C} dy
\]
is of ${\mathscr C}^{\mathrm{exp}}$-class
(since the square of the absolute value can be obtained by multiplying with the complex conjugate, and then using Theorem~\ref{thm:mot.int.})
so this going to 0 is a ${\mathscr C}^{\mathrm{exp}}$-condition by Theorem~\ref{limits:basic}.
$L^\infty$-convergence:
By Theorem 4.4.3 of \cite{CHallp}
${\mathscr C}^{\mathrm{exp}}$-functions are locally constant almost everywhere; more precisely,
there exists a definable set $Z \subset X \times \mathrm{VF}^n \times \mathrm{VG}$ such that for every $F$, $\psi$, $x$ and $\lambda$,
the set $F^n \setminus Z_{x,\lambda}$ has dimension less than $n$ and $H_{F,\psi}(x,\cdot,\lambda)$ is locally constant on
$Z_{x,\lambda}$. Define $H' \in {\mathscr C}^{\mathrm{exp}}(X \times \mathrm{VF}^n \times \mathrm{VG})$ to be equal to $H$ on $Z$ and equal to $0$ outside of $Z$. Then
$H'$ and $H$ have the same behavior concerning $L^\infty$-convergence, but $H'$ has the additional property
that $L^\infty$-convergence is equivalent to uniform convergence; this has already been dealt with above.
\end{proof}
\subsection{Stability under Fourier Transformation for $L^2$-functions of ${\mathscr C}^{\mathrm{exp}}$-class}\label{sec:Fourier}
That the Fourier transform of an $L^1$-function of ${\mathscr C}^{\mathrm{exp}}$-class is again of ${\mathscr C}^{\mathrm{exp}}$-class follows directly from closedness under
integration; see \cite[Corollary 4.3.1]{CHallp} for a version in the current context. For $L^2$-functions of ${\mathscr C}^{\mathrm{exp}}$-class,
this can be obtained using our formalism as follows.
It seems that this result is new even for fixed $F$. It compares to a real counterpart about stability under $L^2$ Fourier transformation of Theorem 8.10 of \cite{CCMRS}.
\begin{thm}[Stability under $L^2$ Fourier transformation]\label{stab:four:II}
Let $H$ be in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VF}^m)$ for some $m\geq 0$ and some definable set $X$. Then there exists ${\mathcal F}_{/X}(H)$ in ${\mathscr C}^{\mathrm{exp}}(X\times \mathrm{VF}^m)$ such that the following holds for all $F$ in ${\mathrm{Loc}}_{\gg 1}$, for all $\psi\in {\mathcal D}_F$ and for each $x\in X_F$.
\quote{If the map $H_{F,\psi,x} : y\mapsto H_{F,\psi}(x,y)$ is $L^2$-integrable on $F^m$, then $z\mapsto {\mathcal F}_{/X}(H)_{F,\psi}(x,z)$ is the Fourier transform of $H_{F,\psi,x}$. }
\end{thm}
\begin{proof}
The Fourier transform of an $L^2$-function $H_{F,\psi}(x,\cdot)$ is defined as the $L^2$-limit of the sequence of functions
\[
G_{F,\psi}(x,\cdot,\lambda)\colon z \mapsto \int_{y\in F^m, \operatorname{ord}_F(y) > -\lambda} H_{F,\psi} (x,y) \psi ( y\cdot z ) |dy|,
\]
indexed by $\lambda \in {\mathbb N}$ (where $z$ runs over $F^m$ and $y \cdot z$ is the scalar product).
By Theorem \ref{thm:mot.int.} (and the comment below that theorem), $G$ is of ${\mathscr C}^{\mathrm{exp}}$-class, so the limit is of ${\mathscr C}^{\mathrm{exp}}$-class by Theorem~\ref{L^pcom}.
\end{proof}
Note that exactly the same proof also works for $L^p$-functions for $1 \le p \le 2$.
\section{An application to orbital integrals}\label{sec:app}
Some of the motivation for this paper came from harmonic analysis on reductive groups, and we conclude it by an illustration: we apply the powerful techniques developed above to Fourier transforms of orbital integrals.
\subsection{Preliminaries}
Let ${\bf G}$ be a connected reductive algebraic group
over a local field $F$, with Lie algebra ${\mathfrak g}$.
We need to represent ${\bf G}(F)$ and ${\mathfrak g}(F)$ as members of a family of definable sets, and we do it as in
\cite[\S 2]{GordonHales}. Let us briefly review this setting and the notation.
\subsubsection{Classification of reductive groups}\label{subsub:groups}
We start with the \emph{fixed choices} as in \cite[\S 2.1]{GordonHales}.
The fixed choices are constructed so that
for each Galois extension of local fields $L/F$ whose Galois group $\operatorname{Gal}(L/F)=\Gamma$ matches the data from the fixed choice (see item (1) below) (and with residue characteristic of $F$ not $2$ or $3$),
a fixed choice $\Sigma$ determines (uniquely up to isomorphism) a split reductive group ${\bf G}^{\ast\ast}$ defined over $F$
and an action of $\operatorname{Gal}(L/F)$ on ${\bf G}^{\ast\ast}(L)$.
A fixed choice $\Sigma$ consists of:
\begin{enumerate}
\item An abstract finite group $\Gamma$ with an enumeration of its elements, a fixed subgroup $I$, and a distinguished element $\sigma_1$.
Later when we associate a split reductive algebraic $F$-group ${\bf G}^{\ast\ast}$ with our fixed choice, the group $\Gamma$ is interpreted as the Galois group of a Galois extension of $F$
with $I$ being the inertia subgroup, and $\sigma_1$ is interpreted as a generator of the Galois group of the maximal unramified sub-extension;
\item a Dynkin diagram with an action of $\Gamma$ (this is interpreted as the Dynkin diagram of ${\bf G}^{\ast\ast}$);
\item Two abstract lattices of the rank determined by the Dynkin diagram with action of $\Gamma$ (these are interpreted as the character and co-character lattices of a split maximal torus ${\bf T}^{\ast\ast}$ in ${\bf G}^{\ast\ast}$).
\end{enumerate}
Each fixed choice $\Sigma$ yields a finite set of isomorphism classes of connected reductive groups ${\bf G}$ defined over $F$,
as described in \cite[\S 6]{GordonHales}. Those isomorphism classes corresponding to $\Sigma$ are distinguished by the elements of a definable parameter set $Z_\Sigma$ (which we will call the ``cocycle space''). An element of $(Z_\Sigma)_{{F}}$ encodes (a) an extension $L/{F}$ over which ${\bf G}$ splits
(which has to have $\operatorname{Gal}(L/F)\simeq \Gamma$ with the inertia subgroup mapping to $I$) and (b) the Galois cocycle that determines ${\bf G}$ as a twist of ${\bf G}^{\ast\ast}$.
We observe that if ${\bf G}$ is unramified over $F$, then {${\bf G}$ itself} is also determined uniquely by the fixed choices.
The precise details of this construction will not be important here, and we refer the reader to \cite[\S 2]{GordonHales} for the details.
Naturally, not all fixed choices of $\Gamma$, $I$, etc. give rise to a family of reductive groups (e.g. $\Gamma$ has to be solvable in order to be a Galois group of an extension of local fields, etc.). From here onwards, we are only interested in fixed choices that do satisfy the compatibility conditions that ensure they give rise to a family of non-empty reductive groups.
Here it is important for us to note that even though the set $Z_{F}$ is infinite for all ${F}$, there are only finitely many isomorphism classes of connected reductive groups corresponding to the same fixed choices, as apparent by
inspection from \cite[\S 6]{GordonHales}, and these isomorphism classes are parameterized independently of ${F}$.
Indeed, in the case (1) of \emph{loc.cit.}, there are $d$ possible invariants in the Brauer group ${\mathbb Q}/{\mathbb Z}$ with denominator $d$; in the cases (2) and (3), excluding type $D_4$, there are two possible ramified quadratic extensions when $p>2$; if ${\bf G}$ contains any factors of type $D_4$, there are still finitely many possibilities, corresponding to distinct degree $3$ ramified extensions with Galois group $S_3$.
In summary, we have
\begin{prop}\label{prop:groups}(\cite[\S2.2.2]{GordonHales}, \cite[Theorem 4]{gordon-roe})
For every fixed choice $\Sigma$ there exists $M>0$, a definable set $Z_\Sigma$,
and definable families
$\underline{\bG}}%{{\mathcal G} \to Z_{\Sigma}$ and $\underline{\fg}}%{g \to Z_\Sigma$ such that for every local field $F$ of residue characteristic greater than $M$
the following holds:
\begin{quote}
for $z\in ({Z_\Sigma})_F$, the set $\underline{\bG}}%{{\mathcal G}_{F,z}$ is the set ${\bf G}_z(F)$ of $F$-points of a connected reductive group
${\bf G}_z$ with absolute root datum determined by $\Sigma$
and the set $\underline{\fg}}%{g_{F,z}$ is the set of $F$-points ${\mathfrak g}_z(F)$ of the Lie algebra of ${\bf G}_z$.
\end{quote}
\end{prop}
In the following, we deal with the sets ${\bf G}_z(F)$, ${\mathfrak g}_z(F)$, etc., while implicitly thinking of them as elements of the definable families
$\underline{\bG}}%{{\mathcal G} \to Z$ and $\underline{\fg}}%{g \to Z$.
In particular,
by a ``definable subset $\omega \subset {\mathfrak g}_z(F)$'',
we really mean a definable subset $\underline{\omega} \subset \underline{\fg}}%{g$, where $\omega := \underline{\omega}_{F,z}$.
\subsubsection{Orbital integrals}\label{sub:orb.int}
Our goal is to provide uniform estimates for Fourier transforms of the orbital integrals on the Lie algebra.
Let $X\in {\mathfrak g}({F})$. An orbital integral at $X$ is the distribution
on ${C_c^\infty}({\mathfrak g}({F}))$, defined by
$$\mu_X(f)=\int_{{\bf G}({F})/C_{{\bf G}}(X)}f(\operatorname{Ad}(g^{-1}) X) d^\ast x,$$
where $C_{{\bf G}}(X)=\{g\in {\bf G}({F})\mid \operatorname{Ad}(g)X=X\}$ is the centralizer of $X$, and
$d^\ast x$ is a ${\bf G}({F})$-invariant measure on ${\bf G}({F})/C_{{\bf G}}(X)$.
Note that such a measure is unique up to a constant multiple, and is uniquely determined by the normalization of the Haar measures on ${\bf G}({F})$ and on $C_{{\bf G}}(X)$.
The fact that orbital integrals converge and thus are well-defined distributions (which is a non-trivial question when $X$ is not regular semi-simple) was proved by Deligne and Ranga Rao for the fields ${F}$ of characteristic zero, and
by G. McNinch for ${F}$ of sufficiently large positive characteristic.
In this paper, however, we focus on regular semisimple elements $X$. We will denote the set of regular semisimple elements in ${\mathfrak g}({F})$ by ${\mathfrak g}({F})^\mathrm{rss}$.
By definition, the centralizer $C_{{\bf G}}(X)$ of a regular semisimple element is a maximal torus. There are finitely many conjugacy classes of maximal tori in ${\bf G}({F})$, with the upper bound on their number depending only on the fixed choices determining ${\bf G}^{\ast\ast}$, so in order to pin down the normalization of the measures on all regular semi-simple orbits it suffices to pick a Haar measure on ${\bf G}({F})$ and on the (finitely many) representatives of conjugacy classes of the maximal tori.
The classical statements, quoted below, on the boundedness of the Fourier transforms of orbital integral are independent of these normalizations of the measures.
However, the main point of this section is to discuss how these bounds depend on the field ${F}$, and therefore we need a consistent normalization of measures.
Following e.g.\ Kottwitz' note on bounds for orbital integrals \cite[Appendix A]{ShinTemp}, we choose on ${\bf G}({F})$ and on all its maximal tori the so-called canonical measures, which assign volume $1$ to the canonical parahoric subgroup in the sense of Gross.
With this normalization of measures, for $X\in {\mathfrak g}({F})^\mathrm{rss}$, let $\mu_X$ denote the orbital integral at $X$, considered as a distribution on $C_c^\infty({\mathfrak g}({F}))$.
\subsubsection{Fourier transform}
Given an additive character $\psi$ of $F$, and a non-degenerate ${\bf G}(F)$-invariant bilinear form $B$ on ${\mathfrak g}(F)$, the Fourier transform for a
Schwartz-Bruhat function $f\in {C_c^\infty}({\mathfrak g}(F))$ is defined as
$$\hat f(Y) =\int_{{\mathfrak g}(F)} f(X)\psi(B(X,Y)) \, dX.$$
Later we will allow the character $\psi$ to vary through the collection ${\mathcal D}_F$ of level zero characters satisfying {Definition} \ref{psiu}.
We will also need the bilinear form $B$ to be a definable function of $X$ and $Y$.
Such a form exists for all $F$ of characteristic zero or sufficiently large positive characteristic by
\cite{adler-roche}, see also \cite[\S 4.1]{CGH2}.
\subsubsection{The bounds on Fourier transforms of orbital integrals}
Let $D(X)$ be the discriminant of $X$,
$$
D(X)=
\prod_{\alpha\in \Phi}\alpha(X),
$$
where $\Phi$ is the {root system} of ${\bf G}$.
It is a well-known theorem of Harish-Chandra that for any $X\in {\mathfrak g}({F})$, the orbital integral at $X$ (as a distribution on the space of locally constant compactly supported functions on ${\mathfrak g}({F})$) is represented by a function $\widehat\mu_X$, defined and locally constant on the set of regular semisimple elements ${\mathfrak g}({F})^\mathrm{rss}$; this function, extended by zero to all of ${\mathfrak g}({F})$, is locally integrable on
${\mathfrak g}({F})$. Moreover, the function
$|D(Y)|^{1/2}\widehat\mu_X(Y)$ is locally bounded on ${\mathfrak g}({F})$.
Following \cite{kottwitz:clay}, we will call such a function a ``nice'' function on
${\mathfrak g}({F})$.
R. Herb generalized the boundedness result, also
making it uniform in $X$ in some sense \cite{herb}.
To state Herb's theorem precisely, we
first need a definition.
Let ${\mathfrak h}$ be a Cartan subalgebra of ${\mathfrak g}({F})$, let $T$ be the torus in ${\bf G}({F})$ such that
${\mathfrak h}= \operatorname{Lie}(T)$, and let $A$ be the split component of $T$.
Fix an arbitrary open compact subgroup $K$ of ${\bf G}({F})$, and
define a function on ${\mathfrak h}({F})\times {\mathfrak g}({F})$ by
\begin{equation}\label{eq:Phi:0}
\Phi({\mathfrak g}, X,Y):= |D(X)|^{1/2}|D(Y)|^{1/2}\int_{{\bf G}({F})/A}\int_K \psi(B(kY, xX))\,dk dx^\ast,
\end{equation}
where $dx^\ast$ is the quotient measure obtained from the Haar measures on ${\bf G}({F})$ and on $A$, and $dk$ is the normalized Haar measure on $K$.
This definition is due to Harish-Chandra, and in his definition the normalization of the measure on $A$ does not matter. We note that $T/A$ is compact; we have already chosen a normalization of the measure on $T$. Let us normalize the measure on $A$ so that the volume of $T/A$ is $1$.
This gives a collection of measures $dx^\ast$ on ${\bf G}({F})/A_i$, where $A_i$ are the
split components of the representatives of the conjugacy classes of maximal tori in
${\bf G}({F})$.
For any fixed collection of normalizations of measures on ${\bf G}({F})/A_i$, consistent with the normalization of measures on the maximal tori,
Harish-Chandra proved the following
\begin{lem}(\cite[Lemma 7.9]{hc:queens}) Let ${\mathfrak h}$ and ${\mathfrak h}_1$ be two Cartan subalgebras in ${\mathfrak g}({F})$. For $X\in {\mathfrak h}^\mathrm{rss}$, $Y\in {\mathfrak h}_1^\mathrm{rss}$,
$$\Phi({\mathfrak g}, X,Y)= |D(X)|^{1/2}|D(Y)|^{1/2}\widehat\mu_X(Y).$$
\end{lem}
R. Herb proved that the function $\Phi({\mathfrak g}, X,Y)$ is uniformly bounded on compact sets:
\begin{thm}(\cite[Theorem 1.3]{herb})\label{thm:herb}
Let ${\mathfrak h}$ be a Cartan subalgebra of ${\mathfrak g}$, and let $\omega$ be a compact subset of the set of regular elements in ${\mathfrak h}$.
Then
$$
\sup_{X\in \omega, Y\in {\mathfrak g}^\mathrm{rss}}|\Phi({\mathfrak g}, X, Y)|<\infty.
$$
\end{thm}
Combining this theorem with Harish-Chandra's lemma, one obtains that Fourier transforms of regular semisimple orbital integrals are uniformly bounded on compact sets, in a given Lie algebra over a given field.
\subsection{Uniformity: the questions}
Our goal is to answer a series of related natural questions:
\begin{enumerate}
\item How does the bound in Theorem \ref{thm:herb} vary for a family of compact sets $\omega_\lambda$?
\item Given a definable set $\omega$, how does the bound in Theorem \ref{thm:herb}
depend on the field ${F}$?
More specifically, do the same bounds apply for the fields ${F}$ of positive characteristic and of characteristic zero with isomorphic residue fields? And,
\item Given a definable set $\omega$, how does the bound depend on the cardinality of the residue field as ${F}$ varies through the family of all completions of a given global field?
\end{enumerate}
\subsection{Uniform bounds for Fourier transforms of regular semisimple orbital integrals}
Here we prove two results, analogous to Theorems 1 and 2 of
\cite[Appendix B]{ShinTemp}, for the Fourier transforms of regular semisimple orbital integrals, which answer the above questions.
\begin{thm}\label{thm:orb.int.bound}
Let $\Sigma$ be a fixed choice as in \S \ref{subsub:groups}. Let $Z_\Sigma$ be the corresponding cocycle space, and let
$\underline{\bG}}%{{\mathcal G}\to Z_{\Sigma}$
be the definable family of reductive groups as in Proposition \ref{prop:groups}, with $\underline{\fg}}%{g\to Z_\Sigma$ the corresponding family of Lie algebras.
Let $\{\omega_\lambda\}_{\lambda\in {\mathbb Z}^n}$ be a definable family of subsets
of ${\mathfrak g}$, understood as in the remark after Proposition \ref{prop:groups}.
Then
\begin{enumerate}
\item There exist constants $M>0$, $a$ and $b$ that depend only on $\Sigma$ and on the {formula} defining
$\{\omega_\lambda\}_{\lambda\in {\mathbb Z}^n}$, such that for each non-Archimedean local field ${F}$
with residue characteristic at least $M$, the following holds. Let ${\bf G}$ be any connected reductive algebraic group over
${F}$ corresponding to the fixed choice $\Sigma$.
Then
for all $X\in \omega_\lambda$ with $\|\lambda\|\le \kappa$, for all $Y\in {\mathfrak g}({F})^\mathrm{reg}$,
\begin{equation}\label{eq:bound}
|D(X)|^{1/2}|D(Y)|^{1/2}|\widehat\mu_X(Y)|\le
{q_{F}}^{a+b\kappa}.
\end{equation}
\item Moreover, if some constants $a$ and $b$ provide a bound for all fields of sufficiently large positive characteristic, then there exist constants $M$ and $N$ (as above, depending only on the data defining the group and the family of subsets $\{\omega_\lambda\}$) such that for all local fields of characteristic zero and residue characteristic larger than $M$,
the bound
\begin{equation}
|D(X)|^{1/2}|D(Y)|^{1/2}|\widehat\mu_X(Y)|\le
N{q_{F}}^{a+b\kappa}.
\end{equation}
holds for all $X\in \omega_\lambda$ with $\|\lambda\|\le \kappa$, for all $Y\in {\mathfrak g}({F})^\mathrm{reg}$, and vice versa, i.e., with positive characteristic and characteristic zero swapped.
\end{enumerate}
\end{thm}
In order to prove this theorem, we need a lemma establishing that orbital integrals \emph{with respect to the canonical measure} are motivic.
The fact that orbital integrals are motivic functions was first proved for \emph{semisimple orbital integrals} in \cite{CHL}, but only for unramified reductive groups. This result was extended to all orbital integrals and all connected
reductive groups in \cite{CGH2}, but the measure on the orbits used there was not necessarily the canonical measure.
Further, this lemma was needed for all semisimple elements, not necessarily regular, in \cite[Appendix B]{ShinTemp}, but at that time we could only prove it up to a uniformly bounded factor (denoted by $i_M$ in \emph{loc. cit.}).
Now we include the proof in full generality for completeness, even though in this paper we need the statement only for regular semisimple elements.
First, we need one more notation. We observe that the connected component of a centralizer of a semisimple element in ${\mathfrak g}(F)$ is the set of $F$-points of a connected reductive algebraic group, and there is a finite list of possible root data associated with such groups, cf. \cite[\S A.2]{ShinTemp}. In particular, there are finitely many fixed choices $\Sigma_i$ giving rise to split forms
${\bf M}^{\ast\ast}$ of such centralizers, and for each
$\Sigma_i$, a corresponding cocycle space $Z_i'$ parameterizing all their forms.
We let $Z'$ be the disjoint union of all these cocycle spaces (considered as a definable set); this yields a definable family $\underline{\bM}}%{\cM \to Z'$ which has the property that all centralizers of elements of ${\mathfrak g}(F)$ arise as ${\bf M}_{z'}(F) := \underline{\bM}}%{\cM_{F,z'}$ for some $z' \in Z'_F$.
We also note that the condition on the pair $(X, z')\in {\mathfrak g}(F)\times Z'_F$ stating that the centralizer of $X$ is isomorphic to
${\bf M}_z'$ as reductive groups over $F$ is a definable condition.
\begin{lem}\label{lem:can.orb}
Let $\underline{\bG}}%{{\mathcal G}\to Z_\Sigma$ be as in Theorem \ref{thm:orb.int.bound} above, and let $\underline{\bM}}%{\cM\to Z'$ be a family of subgroups of ${\bf G}_z$ as above.
Then there exists an integer $M>0$ and a function
$c^{{\bf G}}$ in ${\mathscr C}(Z)$, such that for any definable family $\{f_a\}_{a\in S}$ of test functions on ${\mathfrak g}_z$
of ${\mathscr C}^{\mathrm{exp}}$-class
there exists a function $H(X, a, z,z')$ in
${\mathscr C}^{\mathrm{exp}}(\underline{\fg}}%{g\times S\times Z\times Z')$ such that for
all {local fields of residue characteristic greater than $M$}, the following holds.
For every $z\in (Z_\Sigma)_F$, let ${\bf G}_z$ be the corresponding connected reductive group with Lie algebra ${\mathfrak g}_z$.
Let $X \in {\mathfrak g}_z(F)$ be a semisimple element
with centralizer isomorphic to ${\bf M}_{z'}$ for some $z'\in Z'_F$.
Then
$$\mu_X(f_a)=\frac{1}{{c^{{\bf G}}_F}(z)}H_F(X, a,z,z').$$
We recall from \S \ref{sub:orb.int} that $\mu_X$ denotes the orbital integral at $X$ with respect to the canonical measure.
\end{lem}
\begin{rem} We note that everywhere in this lemma, we could replace ${\mathscr C}^{\mathrm{exp}}$-class with ${\mathscr C}$-class, see Remark \ref{remove-exp}.
\end{rem}
\begin{proof}
Recall that we deal with the sets ${\bf G}_z(F)$, ${\mathfrak g}_z(F)$, etc., while thinking of them as elements of a definable family:
${\bf G}_z(F) = \underline{\bG}}%{{\mathcal G}_{F,z}$, for any $z\in (Z_\Sigma)_F$.
Since the centralizer of $X$ is a definable set (with its defining formulas using $X$ as a parameter), we see that
the condition that the centralizer of $X$ is isomorphic to ${\bf M}_{z'}(F)$ for some $z'\in Z'_F$
gives a definable subset of $\underline{\fg}}%{g\times Z'$. In particular, there exists a definable family of
subsets
${\mathfrak g}_{z, z'}$ of $\underline{\fg}}%{g$,
such that
$$({{\mathfrak g}_{z, z'}})_F =\{X\in {\mathfrak g}_z(F)\mid C_{{\bf G}_z}(X)\cong {\bf M}_{z'} \text{ as reductive groups over $F$}\},$$
where $C_{{\bf G}_z}(X)$ denotes the centralizer of $X$ under the adjoint action of ${\bf G}_z(F)$.
By \cite[Theorem 6]{gordon-roe}, there exists a function $c^{\bf G}\in {\mathscr C}(Z)$, and a family of motivic measures
$\lambda_z$ on the groups ${\bf G}_z$ such that the canonical measure $dg$ on ${\bf G}_z(F)$
satisfies
$$c_G(z) dg = d\lambda_z(g).$$
Now, if $\frac{dg}{dt}$ is the quotient measure on the orbit of $X$, where $dt$ is the canonical measure on
$C_{{{\bf G}}_z(F)}(X)$, then we can represent it as
$$\frac{dg}{dt} = \frac{c^{{\bf M}}(z'){d\lambda_z}}{c^{\bf G}(z)d\lambda_{z'}}=
\frac1{c^{\bf G}(z)}\frac{c^{{\bf M}}(z'){d\lambda_z}}{d\lambda_{z'}}.$$
Consider the quotient measure $\frac{c^{{\bf M}}(z'){d\lambda_z}}{d\lambda_{z'}}$.
It is a quotient of two motivic measures, and thus it is motivic by exactly the same argument as the one of \cite[Lemmma 14.15]{ShinTemp}.
The statement of the lemma follows.
\end{proof}
\subsubsection{Proof of Theorem \ref{thm:orb.int.bound}} Let $X\in {\mathfrak h}(K)$ be an arbitrary regular element, and let $T=G_X$ be its centralizer (which is a torus independent of $X$, with $\operatorname{Lie}(T)={\mathfrak h}$).
First, note that the statement of Theorem \ref{thm:herb} is independent of the normalization of the measure on $T$.
More precisely, it is independent of the normalization of the measure on $A$, the maximal split subtorus of $T$. We can choose such normalization that the volume of the compact torus $T/A$ is equal to $1$.
Then we have
\begin{equation}\label{eq:Phi}
\Phi({\mathfrak g}, X, Y)=|D(X)|^{1/2}|D(Y)|^{1/2}\int_{{\bf G}({F})/T}\int_K \psi(B(kY, xX))\,dk dx^\ast,
\end{equation}
where $dx^\ast$ is the resulting quotient measure.
In particular, for the choice of measure $dx^\ast$ that corresponds to the canonical measures (as above) on
${\bf G}({F})$ and on $T$, we obtain, by Theorem \ref{thm:herb} that $\Phi({\mathfrak g}, X, Y)$ is bounded on $\omega\times{\mathfrak g}^\mathrm{rss}$.
By Lemma \ref{lem:can.orb}, and since $X \mapsto |D(X)|^{1/2}$ is a
function of ${\mathscr C}$-class (generalized to allow for $\sqrt{q}$ as in \cite[B.3.1]{CGH2}),
the right-hand side of (\ref{eq:Phi}) is ``of ${\mathscr C}^{\mathrm{exp}}$-class up to ${\mathscr C}$-constant'',
i.e., we can write (\ref{eq:Phi}) in the form
$$\Phi({\mathfrak g}_{z}, {\mathfrak h}_{z'}, X, Y)= \frac{1}{{c^{{\bf G}}_F}(z)}H_F(X, Y, z, z'),$$
where $H$ is a ${\mathscr C}^{\mathrm{exp}}$-class function (also generalized as in \cite[B.3.1]{CGH2}),
$z\in Z$, and
the parameter $z'$
determines the isomorphism class of the centralizer of $X$. Since here $X$ is assumed to be regular, its centralizer is a torus, and $z'$
corresponds to the choice of the Cartan subalgebra ${\mathfrak h}$ that contains $X$.
Let us consider the function
$$\tilde H(X, Y, z, z', \lambda):= {\bf 1}_{\omega_\lambda}H(X, Y, z, z').$$
By Theorem \ref{thm:herb}, we know that for every $\lambda\in {\mathbb Z}^n$ and every pair $(z, z')\in Z_F\times Z'_F$, the function
$\Phi({\mathfrak g}_z, {\mathfrak h}_{z'}, X, Y)$ is bounded (as a function of $X$ and $Y$) on $\omega_\lambda\times {\mathfrak g}$.
Next we use the observation from \S \ref{subsub:groups} that for every fixed choice there are, in fact, finitely many
possible values of $z, z'$ that give rise to distinct Lie algebras. Hence, for every $\lambda$, the function
$\tilde H(z, z', X, Y, \lambda)$ is bounded.
Then by Theorem \ref{thm:presburger-fam}, we have
\begin{equation}
|\tilde H_F(X, Y, z, z',\lambda)|< q^{a +b\|\lambda\|},
\end{equation}
where the integers $a$ and $b$ depend only on the fixed choices, which completes the proof of Part (1).
Indeed, since we are working in sufficiently large residue field characteristic, the remark just below Theorem \ref{thm:presburger-fam}
allows us to use an integer $a$ instead of a definable $d\in \mathrm{VG}$ in the exponent.
{\bf Part (2).} Follows immediately from \cite[Theorem 3.2]{CGH4}, with a single function $H$ from Part (1) on the left, and
$G(X, Y, z, z',\lambda)= q_F^{a +b\|\lambda\|}$.
\medskip
Finally, in the spirit of \cite[Appendix B]{ShinTemp}, we state an easy corollary that might be useful.
\begin{thm}
Let $G$ be a connected reductive algebraic group over a number field ${\mathbf F}$.
Let $\{\omega_\lambda\}_{\lambda\in {\mathbb Z}^n}$, be a definable family of definable subsets
of ${\mathfrak g}({\mathbf F}_v)$.
Then there exist constants $a_G$ and $b_G$ that depend only on the
global model of $G$ and the formulas defining the subsets $\omega_\lambda$
such that for all {$\lambda \in {\mathbb Z}^n$} with $\|\lambda\|\le \kappa$, for
all but finitely many places $v$,
for all $X\in \omega_\lambda$ with $\|\lambda\|\le \kappa$, for all $Y\in {\mathfrak g}({\mathbf F}_v)^\mathrm{rss}$,
$$
|D(X)|^{1/2}|D(Y)|^{1/2}|\widehat\mu_X(Y)|
\le q_v^{a_G+b_G\kappa} $$
where $q_v$ is the cardinality of the residue field of ${\mathbf F}_v$.
\end{thm}
\begin{proof} Since there are finitely many possibilities for the root data of the groups $G_v$ as $v$ varies over the finite
places of $\mathbf F$, this theorem immediately follows from Theorem \ref{thm:orb.int.bound}, exactly in the same way as Theorem
14.1 of \cite{ShinTemp} follows from Theorem 14.2.
\end{proof}
|
\part*{Introduction}
From the viewpoint of the minimal model program, complex projective manifolds $X$ should be birationally classified according to the sign of the canonical class $K_X$. It is natural to ask how far we can lift the positivity (or the negativity) of $K_X$ to the cotangent sheaf $\Omega_X^1$. The following two theorems were due to Miyaoka (See \cite[Corollary 6.4]{Miy87}).
\begin{thm}
\label{thm-Miya}
Let $X$ be a complex projective manifold such that $K_X$ is pseudoeffective. Then the sheaf $\Omega_X^1$ is generically nef. That is, $\Omega_X^1|_C$ is a nef vector bundle for any Mehta-Ramanathan-general curve $C$.
\end{thm}
\begin{thm}
\label{thm-Miya2}
Let $X$ be a complex projective manifold of dimension $n$ such that $K_X$ is nef. Then for any ample divisors $H_1,...,H_{n-2}$, we have $$c_2(\Omega_X^1)\cdot H_1\cdot \cdots \cdot H_{n-2}\geqslant 0,$$ where $c_2$ stands for the second Chern class.
\end{thm}
A Mehta-Ramanathan-general curve $C$ is the complete intersection of $n-1$ sufficiently ample divisors in general positions, where $n$ is the dimension of $X$. In this note, we are interested in complex projective varieties with nef anticanonical class $-K_X$. These varieties have been studied by many mathematicians, \textit{e.g.} Demailly-Peternell-Schneider (\cite{DPS93},\cite{DPS94}, \cite{DPS96}, \cite{DPS01}), Zhang (\cite{Zhang96}, \cite{Zhang05}), P\u aun (\cite{Pau97}, \cite{Pau12}), Peternell (\cite{Pet12}), Campana-Demailly-Peternell (\cite{CDP15}), Cao-H\"oring (\cite{CH17}, \cite{CH17b}), Cao (\cite{Cao16})... Based on their works, we prove the following theorem
\begin{thm}
\label{thm-main}
Let $X$ be a complex projective variety of dimension at least $2$ with $\mathbb{Q}$-factorial log canonical singularities. Assume that that $-K_X$ is nef. Then the reflexive tangent sheaf $T_X$ is generically nef. That is, $T_X|_C$ is a nef vector bundle for any Mehta-Ramanathan-general curve $C$.
\end{thm}
The statement above was conjectured by Peternell in \cite[Conjecture 1.5]{Pet12}. If we assume that $C$ is the intersection of $n-1$ sufficiently ample divisors of the same class, then Cao (see \cite[Theorem 1.2]{Cao13}) and Guenancia (see \cite[Theorem C]{Guenan16}) proved the nefness of $T_X|_C$ independently by analytic methods. Our approach is more algebraic though.
We note that Theorem \ref{thm-main} does not hold if we only assume that $-K_X$ is pseudoeffective. The following example was due to Demailly, Peternell and Schneider (see \cite[Example 4.14]{DPS01}). Let $Y$ be a curve of genus $g\geqslant 2$ and let $L$ be a line bundle on $Y$ of degree smaller than $2-3g$. Let $X=\mathbb{P}_Y(\mathscr{O}_Y\oplus L)$ and let $f:X\to Y$ be the natural projection. Then, on the one hand, $-K_X$ is effective. On the other hand, we have a natural surjective morphism $T_X\to p^*T_Y$. Since the $\Omega^1_Y$ is ample, the vector bundle $T_X|_C$ is not nef if $C\subseteq X$ is a general very ample divisor.
We also remark that Theorem \ref{thm-main} does not hold if we replace Mehta-Ramanathan-general curves by movable curves. For example, Boucksom, Demailly, P\u aun and Peternell showed that if $X$ is a projective K3-surface or a projective Calabi-Yau threefold, then there is a dominant family of curves $(C_t)_{t\in T}$ such that $T_X|_{C_t}$ is not nef for general $t\in T$ (see \cite[Theorem 7.7]{BDPP13}).
For movable curves classes, we prove the following theorem, which implies Theorem \ref{thm-main} by Mehta-Ramanathan theorem (see \cite[Theorem 6.1]{MR82}).
\begin{thm}
\label{thm-main-mov}
Let $X$ be a complex projective variety with $\mathbb{Q}$-factorial log canonical singularities. Assume that $-K_X$ is nef. Let $\alpha$ be a movable class of curves. Then for any non-zero torsion-free quotient sheaf $Q$ of $T_X$, we have $\alpha\cdot c_1(Q)\geqslant 0$, where $c_1$ stands for the first Chern class.
\end{thm}
As a corollary, we obtain the following theorem, which was proved by Xie in the case of smooth threefolds (see \cite[Theorem 1.2]{Xie04}).
\begin{cor}
\label{cor-c2}
Let $X$ be a normal complex projective variety of dimension $n$ with nef anticanonical class $-K_X$. Assume that $X$ has $\mathbb{Q}$-factorial log canonical singularities and is smooth in codimension $2$. Then for any nef divisors $H_1,...,H_{n-2}$, we have $$c_2(T_X)\cdot H_1\cdot \cdots \cdot H_{n-2}\geqslant 0.$$
\end{cor}
There are two main ingredients for the proof of Theorem \ref{thm-main-mov}. The first one is the following theorem (see also Proposition \ref{prop-CP-alg-foliation-sing} for a singular version) by Campana and P\u aun on algebraicity of foliations (See \cite[Theorem 1.1]{CP15}).
\begin{thm}
\label{thm-CP-foli}
Let $X$ be a projective manifold. Let $\alpha$ be any movable curve class. Let $\mathcal{F}\subset T_X$ be a foliation. Assume that the slope, with respect to $\alpha$, of any non-zero torsion-free quotient of $\mathcal{F}$ is strictly positive. Then $\mathcal{F}$ has algebraic leaves. That is, $\mathcal{F}$ is the foliation induced by some rational dominant map $f:X\dashrightarrow Y$. Moreover, general leaves of $\mathcal{F}$ are rationally connected.
\end{thm}
Another crucial theorem is the following one, which is a refined version of a theorem of Chen and Zhang (see \cite[Main Theorem]{CZ13}).
\begin{thm}
\label{thm-gen-positive-base-anticanonical}
Let $(X,D)$ be a projective $\mathbb{Q}$-factorial log canonical pair with $-(K_X+D)$ nef. Let $f:X\dashrightarrow Y$ be a rational dominant map with $0<\dim Y< \dim X$. Let $\mathcal{F}$ be the foliation induced by $f$. Then $ K_{\mathcal{F}} - K_X-D_{ver} $ is pseudo-effective, where $D_{ver}$ is the vertical part of $D$ over $Y$, and $K_{\mathcal{F}}$ is the canonical class of $\mathcal{F}$.
\end{thm}
We recall that, if $f: X \dashrightarrow Y $ is a rational dominant map between two varieties and if $\Delta$ is a prime divisor in $X$, then $\Delta$ is said to be horizontal over $Y$ if its strict transform in the graph of $f$ dominates $Y$. Otherwise, $\Delta$ is said to be vertical over $Y$.
Let us sketch the proof of the Theorem \ref{thm-main-mov}. Our idea is inspired by Peternell's proof in the case of rational surfaces (see \cite[Theorem 5.9]{Pet12}). We assume by contradiction that there is some movable class $\alpha$ and some torsion-free quotient $T_X\to Q$ such that $\alpha\cdot c_1(Q) <0$. Then we can find a suitable subsheaf, namely $\mathcal{F}$, in the Harder-Narasimhan semistable filtration of $T_X$ such that $\alpha\cdot c_1(T/\mathcal{F})$ is negative, and that $\mathcal{F}$ is a foliation on $X$. In particular, $$\alpha\cdot K_{\mathcal{F}}< \alpha\cdot K_X.$$ Moreover, by using Theorem \ref{thm-CP-foli}, we can show that $\mathcal{F}$ is induced by a rational dominant map $f:X\dashrightarrow Y$. Then from Theorem \ref{thm-gen-positive-base-anticanonical}, we obtain that $$C\cdot K_{\mathcal{F}}\geqslant C\cdot K_X.$$ This is a contradiction.
An orbifold version of Theorem \ref{thm-Miya} was established by Campana and P\u aun (see \cite[Theorem 2.1]{CP15a}): if $(X,\Delta)$ is a projective $\mathbb{Q}$-factorial log canonical pair with $K_X+D$ pseudoeffective, then the orbifold cotangent sheaf $\Omega^1(X,\Delta)$ is $\pi$-generically nef for any adapted Kawamata finite cover $\pi:Z\to X$. By using the orbifold version of Theorem \ref{thm-CP-foli} of Campana and P\u aun (see \cite[Theorem 1.4]{CP15}), we can also deduce the following orbifold version of Theorem \ref{thm-main}.
\begin{thm}
\label{thm-main-orbifold}
Let $(X,\Delta)$ be a complex projective $\mathbb{Q}$-factorial log canonical pair such that $-(K_X+\Delta)$ is nef. Then the orbifold tangent sheaf $T(X,\Delta)$ is $\pi$-generically nef for any adapted Kawamata finite cover $\pi:Z\to X$.
\end{thm}
It is natural to ask under which conditions the positivities in the theorems above are strict. In the second part of the paper, we prove the the following two theorems.
\begin{thm}
\label{thm-gen-ample}
Let $X$ be a smooth projective complex manifold of dimension $n\geqslant 2$ with nef anticanonical class $-K_X$. Then the following properties are equivalent:
\begin{enumerate}
\item $X$ is rationally connected;
\item $T_X$ is generically ample, that is, $T_X|_C$ is an ample vector bundle for any Mehta-Ramanathan-general curve $C$.
\end{enumerate}
\end{thm}
The next theorem was pointed out to the author by Junyan Cao, and the idea of the proof goes back to Andreas H\"oring (see \cite[Proposition 4.6]{Cao13}
\begin{thm}
\label{thm-classification}
Let $X$ be a smooth complex projective manifold of dimension $n$ with nef anticanonical class $-K_X$. Then the following properties are equivalent:
\begin{enumerate}
\item there are ample divisors $H_1,...,H_{n-2}$ such that $$c_2(T_X)\cdot H_1\cdot \cdots \cdot H_{n-2} = 0;$$
\item there is a finite \'etale cover $\tilde{X}\to X$ such that $\tilde{X}$ is either isomorphic to an abelian variety or isomorphic to a $\p^1$-bundle over an abelian variety.
\end{enumerate}
\end{thm}
This theorem answers a question of Yau (see \cite[Problem 66]{Yau93}) in the projective case. We note that one could not expect the $\p^1$-bundle to be trivial, see for example \cite[Example 3.5]{DPS94}.
Throughout this paper, we will work over $\mathbb{C}$, the field of complex numbers. \\
\noindent \textbf{Acknowledgment.} The author would like to express his gratitude to St\'ephane Druel and Burt Totaro for reading the preliminary version of this paper and warm encouragement. He is grateful to Junyan Cao for pointing out the application Theorem \ref{thm-classification} to him. He would also like to thank Jun Li, Chen Jiang, Claire Voisin, Yuan Wang and Jian Xiao for general discussions.
\part{Positivity of tangent sheaves}
\section{Slope semistability and Harder-Narasimhan filtrations}
\label{Slope semistability and Harder-Narasimhan filtrations}
In this section, we will study some properties on Harder-Narasimhan semistable filtrations. Let $X$ be a normal projective variety and let $\alpha$ be a movable curve class. Assume that either $X$ is $\mathbb{Q}$-factorial or $\alpha$ is the class of a complete intersection of basepoint-free divisors. Then for any torsion-free coherent sheaf $E$ with positive rank on $X$, the slope of $E$ with respect to $\alpha$ is the number $$\mu_{\alpha}(E)=\frac{\alpha \cdot c_1(E) }{\mathrm{rank}\, E}.$$ We would like to refer to \cite[Appendix A]{GKP14} for more details on slope notions on singular spaces. The maximal slope is defined as follows, $$\mu_{\alpha,max}(E)=\sup \{\mu_{\alpha}(F) \ |\ F\ \mbox{is a non-zero saturated subsheaf of } E\}.$$ The supremum is in fact a maximum (see \textit{e.g.} \cite[Proposition A.2]{GKP14}). The sheaf $E$ is called $\alpha$-semistable (or just semistable if there is no ambiguity) if $\mu_{\alpha,max} (E)= \mu_{\alpha}(E)$. If $E$ is not semistable, then there is a unique maximal subsheaf $F$ of $E$ such that $\mu_{\alpha,max} (E)= \mu_{\alpha}(F)$. This $F$ is called the maximal destabilizing subsheaf and is automatically semistable.
There is a unique filtration, called the Harder-Narasimhan semistable filtration, of saturated subsheaves, $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r=E$$ such that $E_i/E_{i-1}$ is the maximal destabilizing subsheaf of $E/E_{i-1}$ for all $i\in \{1,..., r\}$ and that the sequence $(\mu_{\alpha}(E_i/E_{i-1}))_{i\in \{1,..., r\}}$ is strictly decreasing.
The minimal slope (see \textit{e.g.} \cite[Definition 2.3]{CP15}) is defined as follows, $$\mu_{\alpha,min}(E)=\inf \{\mu_{\alpha}(Q) \ |\ Q\ \mbox{is a non-zero torsion-free quotient sheaf of } E\}.$$ This infimum is also a minimum. Indeed, we have $\mu_{\alpha,min}(E)= \mu_{\alpha}(E/E_{r-1})$, where $E_{r-1}$ is the saturated subsheaf defined in the Harder-Narasimhan semistable filtration above (see \cite[Proposition 1.3]{CP11}).
We will use the following lemma in the proof of Theorem \ref{thm-main}.
\begin{lemma}
\label{lem-invol-subsheaf}
Let $X$ be a normal projective $\mathbb{Q}$-factorial variety and let $\alpha$ be a movable curve class in $X$. Let $E$ be a torsion-free sheaf on $X$ such that $\mu_{\alpha}(E)\geqslant 0$ and $\mu_{\alpha,min}(E) <0$. Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r=E$$ with $r\geqslant 2$ be the Harder-Narasimhan semistable filtration with respect to $\alpha$. Then there is some $k\in \{1,...,r-1\}$ such that $\mu_{\alpha}(E/E_k) <0$ and $\mu_{\alpha,min}(E_k) = \mu_{\alpha}(E_k/E_{k-1})>0$.
\end{lemma}
\begin{proof}
We have $\mu_{\alpha}(E/E_{r-1})=\mu_{\alpha,min}(E)<0$. Let $k$ be the smallest integer in $\{0,...,r-1\}$ such that $\mu_{\alpha}(E/E_{k})<0$. Since $\mu_{\alpha}(E)\geqslant 0$, we know that $k\geqslant 1$. We consider the following exact sequence $$0\to E_{k}/E_{k-1}\to E/E_{k-1} \to E/E_{k} \to 0.$$ By the definition of $k$, we have $\mu_{\alpha}(E/E_{k-1})\geqslant 0$. Thus $\mu_{\alpha}(E_k/E_{k-1})>0$.
\end{proof}
\section{Foliations and relative tangent sheaves}
Let $X$ be a normal variety of dimension at least $2$ and let $T_X=(\Omega_X^1)^*$ be the reflexive tangent sheaf. A foliation $\mathcal{F}$ on $X$ is a non-zero saturated subsheaf of $T_X$ which is closed under Lie brackets. The canonical class $K_{\mathcal{F}}$ of $\mathcal{F}$ is a Weil divisor such that $$\mathscr{O}_{X}(- K_{\mathcal{F}}) \cong \mathrm{det}\, \mathcal{F},$$ where $\mathrm{det}\, \mathcal{F}$ is the reflexive hull of the top wedge product of $\mathcal{F}$. We say that $\mathcal{F}$ has algebraic leaves if the dimension of the Zariski closure of a general leaf of $\mathcal{F}$ is equal to the rank of $\mathcal{F}$.
Typical examples of foliations are relative tangent sheaves as follows. Consider a rational dominant map $f:X\dashrightarrow Y$ between normal varieties. Assume that $\dim\, Y< \dim\, X$. Let $V$ be the smooth locus of $Y$. Let $U$ be a non-empty smooth open subset of $X$ such that $f|_U$ is regular and $f(U)\subset V$. The relative tangent sheaf $T_{U/V}$ of $f|_U:U\to V$ is defined as the kernel of the natural morphism $$\mathrm{d}f|_U:T_U\to f^*T_V.$$ There is a unique saturated subsheaf $T_{X/Y}$ of the reflexive tangent sheaf $T_X$ such that $T_{X/Y}|_U=T_{U/V}$. We call $T_{X/Y}$ the relative tangent sheaf of $f:X\dashrightarrow Y$. It is a foliation on $X$. We note that a foliation on $X$ has algebraic leaves if and only if it is induced by some rational dominant map as above (see \textit{e.g.} \cite[Lemma 3.2]{AD13}).
The following proposition is a singular version of Theorem \ref{thm-CP-foli}.
\begin{prop}
\label{prop-CP-alg-foliation-sing}
Let $X$ be a projective normal $\mathbb{Q}$-factorial variety. Let $\alpha$ be a movable curve class. Assume that $\mathcal{F}$ is a saturated subsheaf of $T_X$ such that
\begin{enumerate}
\item $\mu_{\alpha,min}(\mathcal{F}) > 0$,
\item $2\mu_{\alpha,min}(\mathcal{F}) > \mu_{\alpha,max} (T_X/\mathcal{F}).$
\end{enumerate}
Then $\mathcal{F}$ is a foliation and has algebraic leaves. Moreover, general leaves of $\mathcal{F}$ are rationally connected.
\end{prop}
\begin{proof}
Let $r:X'\to X$ be a resolution of singularities, and let $\alpha'=r^*\alpha$ be the numerical pull-back such that $$\alpha' \cdot \beta' =\alpha \cdot r_*\beta'$$ for any divisor class $\beta'$ on $X'$ (see \cite[Construction A.15]{GKP14}). Since $\alpha$ is movable, so is $\alpha'$ by \cite[Lemma A.17]{GKP14}. If $\mathscr{G}'$ and $\mathscr{G}$ are torsion-free sheaves on $X'$ and $X$ respectively such that $r_*\mathscr{G}'$ is isomorphic to $\mathscr{G}$ in codimension $1$, then $$\mu_{\alpha'}(\mathscr{G}') = \mu_{\alpha}(\mathscr{G}), \ \mu_{\alpha',max}(\mathscr{G}') = \mu_{\alpha,max}(\mathscr{G}),\ \mathrm{and}\ \mu_{\alpha',min}(\mathscr{G}') = \mu_{\alpha,min}(\mathscr{G}).$$ Therefore, if $\mathcal{F}'$ be the saturated subsheaf of $T_{X'}$ induced by $\mathcal{F}$, then $$\mu_{\alpha',min}(\mathcal{F}') =\mu_{\alpha,min}(\mathcal{F}) > 0,$$ and $$2\mu_{\alpha',min}(\mathcal{F}') = 2 \mu_{\alpha,min}(\mathcal{F}) > \mu_{\alpha,max} (T_{X}/\mathcal{F})=\mu_{\alpha',max} (T_{X'}/\mathcal{F}').$$ Hence $\mathcal{F}'$ is a foliation and has algebraic leaves by \cite[Theorem 1.4]{CP15}. Moreover, general leaves of $\mathcal{F}'$ are rationally connected. The proposition then follows from the property that $\mathcal{F}\cong (r_*\mathcal{F}')^{**}$.
\end{proof}
\section{Proof of Theorem \ref{thm-gen-positive-base-anticanonical}}
In this section, we will prove Theorem \ref{thm-gen-positive-base-anticanonical}. It follows from the following theorem, which is a special case of a theorem of Druel (see \cite[Proposition 4.1]{Dru15})
\begin{thm}
\label{thm-pseuf}
Let $f:X\to Y$ be a surjective morphism between normal projective $\mathbb{Q}$-factorial varieties. Let $\Delta$ be an effective $\mathbb{Q}$-divisor in $X$. Assume that the pair $(F,\Delta|_F)$ is log canonical, where $F$ is a general fiber of $f$. If there is some positive integer $m$ such that $m(K_X+\Delta)$ is Cartier and that $h^0(F, \mathscr{O}_F(m(K_X+\Delta)|_F))>0$, then $K_{\mathcal{F}}+\Delta$ is pseudo-effective, where $\mathcal{F}$ is the foliation induced by $f$.
\end{thm}
\begin{proof}[{Proof of Theorem \ref{thm-gen-positive-base-anticanonical}}]
There is a log resolution $\pi:Z\to X$ of $(X,D)$ such that the induced map $g:Z\to Y$ is a morphism. By blowing up $Y$ and $Z$ if necessary, we may assume that $Y$ is smooth.
We write $$K_Z+D_Z\sim_{\mathbb{Q}} \pi^*(K_X+D)+E,$$ where $D_Z$ and $E$ are effective $\mathbb{Q}$-divisors without common components. Moreover $E$ has $\pi$-exceptional support. Since $(X,D)$ is log canonical, so is the pair $(Z,D_Z)$.
\centerline{
\xymatrix{
Z \ar[d]^{g} \ar[r]^{\pi} & X\\
Y &
}
}
Let $L$ be an ample divisor in $Z$ and let $\delta >0$ be a rational number. Then $-\pi^*(K_X+D)+\delta L$ is ample. We can then choose a smooth irreducible $\mathbb{Q}$-divisor $$A\sim_{\mathbb{Q}}-\pi^*(K_X+D)+\delta L$$ such that $(Z,D_Z+A)$ is a log canonical. We have $$K_{Z}+ D_Z+A \sim_{\mathbb{Q}} E+\delta L.$$ Let $\Delta=D_Z+A-\pi_*^{-1}D_{ver}$. Then $\Delta$ is effective, and we have \begin{equation}
K_{Z}+ \Delta \sim_{\mathbb{Q}} E+\delta L-\pi_*^{-1}D_{ver}. \tag{*}
\end{equation}
Let $G$ be a general fiber of $g$. We claim that $m(K_{Z}+\Delta)|_G$ has non-zero global sections for large enough and sufficiently divisible integer $m$. Indeed, by (*), we have $$(K_{Z}+\Delta)|_G \sim_{\mathbb{Q}} (E+\delta L-\pi_*^{-1}D_{ver})|_G \sim_{\mathbb{Q}} (E+\delta L)|_G.$$ The right-hand-side above is a big divisor. Hence $m(K_{Z}+\Delta)|_G$ has non-zero global sections for large enough and sufficiently divisible integer $m$.
By Theorem \ref{thm-pseuf}, we obtain that $K_{\mathcal{G}}+\Delta$ is pseudoeffective, where $\mathcal{G}$ is the foliation induced by $g$. Hence
\begin{eqnarray*}
E+\delta L+(K_{\mathcal{G}}-K_Z-\pi_*^{-1}D_{ver}) &\sim_{\mathbb{Q}}& E +\delta L - \pi_*^{-1}D_{ver} +(K_{\mathcal{G}}-K_Z)\\
&\sim_{\mathbb{Q}}& K_{Z}+\Delta + (K_{\mathcal{G}}-K_Z)\\
& \sim_{\mathbb{Q}}& K_{\mathcal{G}}+\Delta
\end{eqnarray*}
is pseudoeffective.
Since this is true for arbitrary $\delta>0$, we obtain that $$E+(K_{\mathcal{G}}-K_Z-\pi_*^{-1}D_{ver})$$ is pseudoeffective.
Thus $$K_{\mathcal{F}}-K_X- D_{ver}=\pi_*(E+(K_{\mathcal{G}}-K_Z-\pi_*^{-1}D_{ver}))$$ is pseudoeffective.
\end{proof}
\section{Proofs of Theorem \ref{thm-main-mov}, \ref{thm-main} and Corollary \ref{cor-c2}}
\begin{proof}[{Proof of Theorem \ref{thm-main-mov}}]
Assume the opposite. Then $\mu_{\alpha,min}(T_X) < 0$. In particular, $T_X$ is not $\alpha$-semistable. Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r=T_X$$ with $r\geqslant 2$ be the Harder-Narasimhan semistable filtration.
Then by Lemma \ref{lem-invol-subsheaf}, there is some $k\in \{1,...,r-1\}$ such that $\mu_{\alpha}(T_X/E_k) <0$ and $$\mu_{\alpha,min}(E_k) = \mu_{\alpha}(E_k/E_{k-1})>0.$$ We have the following inequality,
\begin{eqnarray*}
2\mu_{\alpha,min}(E_k) &=& 2\mu_{\alpha}(E_k/E_{k-1}) > \mu_{\alpha}(E_k/E_{k-1}) \\
&>& \mu_{\alpha}(E_{k+1}/E_{k}) = \mu_{\alpha,max}(T_X /E_{k}).
\end{eqnarray*} Hence $E_k$ is a foliation and has algebraic leaves by Proposition \ref{prop-CP-alg-foliation-sing}. We denote $E_k$ by $\mathcal{F}$. Then $$\alpha\cdot K_{\mathcal{F}} < \alpha\cdot K_X $$ for $\mu_{\alpha}(T_X/E_k) <0$.
Since $\mathcal{F}$ has algebraic leaves, there is a rational dominant map $f:X\dashrightarrow Y$ such that $\mathcal{F}$ is induced by $f$. Moreover, since $\mathcal{F}$ is a non-zero proper subsheaf, we have $0 < \dim Y<\dim X$. Hence Theorem \ref{thm-gen-positive-base-anticanonical} shows that $$\alpha\cdot K_{\mathcal{F}} \geqslant \alpha\cdot K_X.$$ This is a contradiction.
\end{proof}
\begin{proof}[{Proof of Theorem \ref{thm-main}}]
Let $C$ be a Mehta-Ramanathan-general curve. Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r=T_X$$ be the Harder-Narasimhan filtration with respect to the class $\alpha$ of $C$. Then, by Mehta-Ramanathan Theorem (see \cite[Theorem 6.1]{MR82}), the restriction $$0=E_0|_C\subsetneq E_1|_C \subsetneq \cdots \subsetneq E_r|_C=T_X|_C$$ is the Harder-Narasimhan filtration for $T_X|_C$. In particular, we have $$\mu_{min}(T_X|_C)=\mu_{\alpha, min}(T_X)\geqslant 0.$$ Thus $T_X|_C$ is nef.
\end{proof}
\begin{proof}[{Proof of Corollary \ref{cor-c2}}] By Theorem \ref{thm-main-mov}, $T_X$ is generically $(H_1,...,H_{n-2})$-semipositive. Since $-K_X$ is nef, Miyaoka inequality (see \cite[Theorem 6.1]{Miy87}) shows that $$c_2(T_X)\cdot H_1\cdot \cdots \cdot H_{n-2}\geqslant 0.$$
\end{proof}
\section{An orbifold version of generic nefness}
\label{An orbifold version of generic nefness}
In this section, we will prove Theorem \ref{thm-main-orbifold}. We would like to refer to \cite[Section 5]{CP15} for detailed notions of orbifolds. Let $(X,D)$ be a projective $\mathbb{Q}$-factorial log canonical pair of dimension $n$. Let $\pi:Z\to X$ be a Kawamata finite cover adapted to $(X,D)$. Let $H_1,...,H_{n-1}$ be ample divisors in $X$. Then the orbifold cotangent sheaf $\Omega^1(X,D)$ (respectively the orbifold tangent sheaf $T(X,D)$) is said to be $\pi$-generically semipositive with respect to $H_1,...,H_{n-1}$ if for any non-zero torsion-free quotient $Q$ of $\pi^*\Omega^1(X,D)$ (respectively of $\pi^*T(X,D)$), we have
$$c_1(Q) \cdot \pi^* H_1\cdot \cdots \cdot \pi^* H_{n-1}\geqslant 0.$$
We say that $\Omega^1(X,D)$ (respectively $T(X,D)$) is $\pi$-generically nef if it is $\pi$-generically semipositive with respect to any ample divisors $H_1,...,H_{n-1}$ in $X$. In particular, we note that if $D$ is an integral divisor and if $\pi$ is the identity map, then $\pi$-generic nefness is the same as generic nefness by Mehta-Ramanathan theorem (see \cite[Theorem 6.1]{MR82}).
In order to prove Theorem \ref{thm-main-orbifold}, we will need the following version of \cite[Theorem 1.4]{CP15} for singular spaces.
\begin{thm}
\label{thm-CP-alg-foliation-sing-orbifold}
Let $(X,D)$ be a projective $\mathbb{Q}$-factorial log canonical pair. Let $\pi:Z\to X$ be a finite cover adapted to $(X,D)$. Let $H_1, ... , H_{n-1}$ be very ample divisors in $X$ and let $\alpha$ be the class of $\pi^*H_1\cdot \cdots \cdot \pi^*H_{n-1}$. Assume that there is a saturated subsheaf $F$ of $\pi^*T(X,D)$ such that
\begin{enumerate}
\item $F$ is $G$-invariant, where $G$ is the Galois group of $\pi$.
\item $\mu_{\alpha,min}(F) > 0$,
\item $2\mu_{\alpha,min}(F) > \mu_{\alpha,max} (\pi^*T(X, D)/F).$
\end{enumerate}
Then the saturation of $F$ in $\pi^*T_X$ defines an algebraic foliation $\mathcal{F}$ on $X$. Moreover, $F$ is the saturation of $\pi^*\mathcal{F}\cap \pi^*T(X,D)$ in $\pi^*T(X,D)$.
\end{thm}
\begin{proof}
Let $r:X'\to X$ be a log resolution of $(X,D)$ which is an isomorphism over the smooth locus $U$ of $(X,D)$. Let $Z'$ be the normalization of $Z\times_X X'$. Then the natural morphism $\pi':Z'\to X'$ is an adapted finite cover of $(X',D')$, where $D'=r_*^{-1}D$. Let $U'=r^{-1}(U)$, $V=\pi^{-1}(U)$ and $V'=\pi'^{-1}(U')$.
\centerline{
\xymatrix{
Z' \ar[d]_{\pi'} \ar[r] & Z \ar[d]^{\pi}\\
X' \ar[r]^{r} & X
}
}
There is a unique $G$-invariant saturated subsheaf $F'$ of $\pi'^*T(X',D')$ such that $F'|_{V'}$ is isomorphic to $F|_V$. Let $$\alpha' = (\pi'\circ r)^*H_1\cdot \cdots \cdot (\pi'\circ r)^*H_{n-1}.$$ We have $\mu_{\alpha',min}(F')= \mu_{\alpha,min}(F) > 0$ and $$\mu_{\alpha' ,max} (\pi'^*T(X', D')/F') = \mu_{\alpha,max} (\pi^*T(X, D)/F).$$ Hence by \cite[Theorem 1.4]{CP15}, the saturation of $F'$ in $\pi'^*T_{X'}$ defines an algebraic foliation $\mathcal{F}'$ on $X'$. Moreover, $F'$ is the saturation of $\pi'^*\mathcal{F}'\cap \pi'^*T(X',D')$ in $\pi'^*T(X',D')$ (see \cite[Corollary 5.10]{CP15}). Let $\mathcal{F}$ be the saturation of the natural image of $r_*\mathcal{F}'$ in $T_X$. Then $\mathcal{F}$ is an algebraic foliation and $F$ is the saturation of $\pi^*\mathcal{F}\cap \pi^*T(X,D)$ in $\pi^*T(X,D)$.
\end{proof}
Now we will prove Theorem \ref{thm-main-orbifold}.
\begin{proof}[{Proof of Theorem \ref{thm-main-orbifold}}]
The proof is similar to the one of Theorem \ref{thm-main}. Assume the opposite.
Then there are very ample divisors $H_1,...,H_{n-1}$ such that $T(X,D)$ is not $\pi$-generically semipositive with respect to $H_1,...,H_{n-1}$. Let $\alpha$ be the class of $\pi^*H_1\cdot \cdots \cdot \pi^*H_{n-1}$.
By applying Lemma \ref{lem-invol-subsheaf} to $\pi^*T(X,D)$, we can find a saturated subsheaf $F$ of $\pi^*T(X,D)$ such that $\mu_{\alpha,min}(F)>0$ and that $ \mu_{\alpha}(\pi^*T(X,D)/F) <0$. From the uniqueness of Harder-Narasimhan filtration, we know that $F$, as a component in the Harder-Narasimhan filtration, is invariant under the Galois group of $\pi$. Moreover, as in the proof of Theorem \ref{thm-main-mov}, we have $$2\mu_{\alpha,min}(F) > \mu_{\alpha, max}(\pi^*T(X,D)/F).$$ By Theorem \ref{thm-CP-alg-foliation-sing-orbifold}, the saturation of $F$ in $(\pi^*T_X)^{**}$ defines an algebraic foliation $\mathcal{F}$ on $X$. Assume that $\mathcal{F}$ is the relative tangent sheaf of some dominant rational map $f:X\dashrightarrow Y$.
Then, on the one hand, by Theorem \ref{thm-gen-positive-base-anticanonical}, we have $$\alpha\cdot \pi^*(K_{\mathcal{F}}-K_{X}-D_{ver})\geqslant 0,$$ where $D_{ver}$ is the vertical part of $D$ over $Y$.
On the other hand, by Theorem \ref{thm-CP-alg-foliation-sing-orbifold}, the sheaf $F$ is the saturation of the intersection $\pi^*\mathcal{F} \cap \pi^*T(X,D)$ in $\pi^*T(X,D)$. By \cite[Prop. 2.17]{Clau15}, we have $$\mathrm{det}\, F \cong \mathscr{O}_Z(\pi^*(-K_{\mathcal{F}}-D_{hor})),$$ where $D_{hor}$ is the horizontal part of $D$ over $Y$. Thus $$\mathrm{det}\, (\pi^*T(X,D)/F) \cong \mathscr{O}_Z(\pi^*(K_{\mathcal{F}}-K_{X}-D_{ver})),$$ and we have $$ \alpha\cdot c_1(\pi^*T(X,D)/F) = \alpha\cdot \pi^*(K_{\mathcal{F}}-K_{X}-D_{ver}) \geqslant 0.$$ Since $\mu_{\alpha}(\pi^*T(X,D)/F) <0$, this is a contradiction.
\end{proof}
\part{Discussion on equality conditions}
\section{Proof of Theorem \ref{thm-gen-ample}}
We will prove Theorem \ref{thm-gen-ample} in this section.
\begin{lemma}
\label{lem-vanishing}
Let $X$ be a smooth projective variety and let $\delta$ be a cycle of pure dimension $k$. Assume that for any ample divisors $H_1,...,H_k$, we have $\delta\cdot H_1\cdot \cdots \cdot H_k\geqslant 0$. Then the following two conditions are equivalent:
\begin{enumerate}
\item there are ample divisors $H_1,...,H_k$ such that $\delta\cdot H_1\cdot \cdots \cdot H_k= 0$;
\item for any ample divisors $H_1,...,H_k$, we have $\delta\cdot H_1\cdot \cdots \cdot H_k= 0$;
\item for any nef divisors $H_1,...,H_k$, we have $\delta\cdot H_1\cdot \cdots \cdot H_k= 0$.
\end{enumerate}
\end{lemma}
\begin{proof}
Since every nef divisor is a limit of ample ($\mathbb{Q}$-)divisors, by continuity, we see that (2) implies (3). Hence we only need to prove that (1) implies (2). Let $H_1,...,H_k$ be ample divisors such that $\delta\cdot H_1\cdot \cdots \cdot H_k= 0$, and let $A_1,..., A_k$ be any ample divisors. We need to prove that $\delta\cdot A_1\cdot \cdots \cdot A_k= 0$.
Let $m>0$ be a natural number such that $mH_1-A_1$ is still an ample divisor. Then we have
\begin{eqnarray*}
0 &\leqslant& \delta\cdot (mH_1-A_1)\cdot \cdots \cdot H_k \\
&=& m(\delta\cdot H_1\cdot \cdots \cdot H_k) - (\delta\cdot A_1\cdot H_2 \cdot \cdots \cdot H_k)\\
&=& -\delta\cdot A_1\cdot H_2\cdot \cdots \cdot H_k \\
&\leqslant& 0.
\end{eqnarray*} Thus $\delta\cdot A_1\cdot H_2\cdot \cdots \cdot H_k =0.$ By repeating this procedure $k-1$ more times, we can obtain that $\delta\cdot A_1\cdot \cdots \cdot A_k =0.$
\end{proof}
Now we can prove Theorem \ref{thm-gen-ample}.
\begin{proof}[{Proof of Theorem \ref{thm-gen-ample}}]
The case when $n=1$ is trivial. We assume from now on that $n\geqslant 2$. First we assume that $X$ is rationally connected. Assume by contradiction that $T_X$ is not generically ample. Then there is a Mehta-Ramanathan-general curve $C$ such that $T_X|_C$ is nef but not ample. By \cite[Theorem 2.4]{Har71}, there is a non-zero quotient bundle of $T_X|_C$ of degree zero. Thus, by Mehta-Ramanathan theorem (see \cite[Theorem 6.1]{MR82}), we have $\mu_{\alpha,min}(T_X)=0$, where $\alpha$ is the class of $C$. This implies that there is a surjective morphism $T_X\to Q$ such that $Q$ is a non-zero torsion-free sheaf and that $\alpha \cdot c_1(Q)=0.$ Since a quotient bundle of a nef bundle is still nef, a quotient torsion-free sheaf of a generically nef sheaf is also generically nef. Hence $Q$ is generically nef and $c_1(Q)\cdot H_1\cdot \cdots \cdot H_{n-1} \geqslant 0$ for any ample divisors $H_1,...,H_{n-1}$. Thus by Lemma \ref{lem-vanishing}, $c_1(Q)\cdot H_1\cdot \cdots \cdot H_{n-1} = 0$ for any ample divisors $H_1,...,H_{n-1}$. This shows that $c_1(Q)$ is numerically zero. Since $X$ is rationally connected, it is simply connected. Therefore, we have $\mathrm{det}\, Q \cong \mathscr{O}_X.$ The injective morphism $Q^*\to \Omega_X^1$ then induces a non-zero morphism $\mathscr{O}_X\to \Omega_X^k$, where $k$ is the rank of $Q$. This is a contradiction, since $h^0(X,\Omega_X^k)=0$ by \cite[Corollary IV.3.8]{Kol96}.
Now we assume that $X$ is not rationally connected. Then, by \cite[Corollary 1]{Zhang05}, there is a dominant rational map $f:X\dashrightarrow Y$ such that $Y$ is smooth with Kodaira dimension $\kappa(Y)=0$ and that the general fibers of $f$ are proper and rationally connected. Since $X$ is not rationally connected, $Y$ has positive dimension $d$. There is some positive integer $m$ such that $h^0(Y, \mathscr{O}_Y(mK_Y))\neq 0$. Thus $h^0(X, (\Omega_X^1)^{\otimes md}) \neq 0.$ Let $C$ be a Mehta-Ramanathan-general curve and let $\alpha$ be the class of $C$. Then $\mu_{\alpha, max}(\Omega_X^1) \geqslant 0$ by \cite[Corollary 5.11]{CP11}. This implies that $\mu_{\alpha, min}(T_X) \leqslant 0$. Hence $T_X|_C$ is not ample by Mehta-Ramanathan theorem (see \cite[Theorem 6.1]{MR82}).
\end{proof}
\begin{rem}
\label{rem-sing-nef-rc-gn}
Theorem \ref{thm-gen-ample} does not hold without assuming the smoothness of $X$. For example, let $G$ be the group $\mathbb{Z}/2\mathbb{Z}$ and let $E$ be an elliptic curve with an action of $G$ such that $E/G=\p^1$. We also endow $\p^1$ with the canonical action of $G$ ($g.[a:b]=[a:-b]$ if $g$ is the generator of $G$). Then $G$ acts on the product $E\times \p^1$ diagonally and the quotient $E\times \p^1 \to (E\times \p^1)/G=X$ is \'etale in codimension $1$. In particular, $X$ has canonical singularities and $-K_X$ is nef. In addition, as in \cite[Remark and Question 3.8]{GKP14}, we have $h^0(X, ((\Omega_X^1)^{\otimes 2})^{**}) > 0.$ This implies that $\mu_{\alpha, min}(T_X) \leqslant 0$ for any ample class $\alpha$.
\end{rem}
\section{Equality conditions of Miyaoka inequality}
We recall that a non-zero torsion-free sheaf $E$ on a projective manifold $X$ of dimension $n$ is said to be generically $(H_1,...,H_{n-2})$-semipositive for some ample divisors $H_1,...,H_{n-2}$ if for each nef divisor $D$, we have $\mu_{\alpha,min}(E)\geqslant 0$, where $\alpha$ is the class of $D\cdot H_1\cdot \cdots \cdot H_{n-2}$. We remark that if $E$ is generically nef, then it is generically $(H_1,...,H_{n-2})$-semipositive for any ample divisors $H_1,...,H_{n-2}$.
In \cite[Theorem 6.1]{Miy87}, Miyaoka proved that if $c_1(E)$ is nef and if $E$ is generically $(H_1,...,H_{n-2})$-semipositive, then $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}\geqslant 0$. As a consequence, if $c_1(E)$ is nef and if $E$ is generically nef, then $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}\geqslant 0$ for any ample divisors $H_1,...,H_{n-2}$.
In this section, we will study the equality conditions of these inequalities. We will assume that $E$ is generically nef and that $c_1(E)$ is nef. By Lemma \ref{lem-vanishing}, the equality $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}= 0$ holds for some ample divisors $H_1,...,H_{n-2}$ if and only if $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}= 0$ for any ample divisors $H_1,...,H_{n-2}$. Thus, in order to study the equality conditions, we may assume that $c_2(E)\cdot H^{n-2}=0$ for some ample divisor $H$.
Our idea is to look into the details in Miyaoka's proof, and study every inequality inside. We will discuss following the numerical dimension $\nu(c_1(E))$ of $c_1(E)$. Recall that the numerical dimension $\nu$ of a nef divisor $N$ is the largest integer such that $N^{\nu+1} \equiv 0$.
\subsection*{Preparatory Lemmas} We will first collect some useful elementary results for this section.
\begin{lemma}
\label{lem-2form-1}
Let $\mathbf{q}(\cdot,\cdot)$ be a non-degenerated symmetric bilinear form of signature $(1,m)$ on a real vector space $V$. Let $\vec{x},\vec{y}\in V$ be two vectors. Assume that $\mathbf{q}(\vec{x},\vec{x})=0$, $\mathbf{q}(\vec{x},\vec{y})=0$ and $\mathbf{q}(\vec{y},\vec{y})=0$. Then $\vec{x}$ and $\vec{y}$ are linear dependent.
\end{lemma}
\begin{proof}
If $m=0$, then $\mathbf{q}$ is definite and $\vec{x}=\vec{y}=\vec{0}$. We assume then that $m>0$. By Sylvester theorem, there is an orthogonal basis $(\vec{e},\vec{e}_1,...,\vec{e}_m)$ of $V$ such that $\mathbf{q}(\vec{e},\vec{e})=1$ and that $\mathbf{q}(\vec{e}_i,\vec{e}_i)=-1$ for all $i=1,...,m$.
Let $(a,b_1,...,b_m)$ and $(a',b_1',...,b_m')$ be the coordinates of $\vec{x}$ and $\vec{y}$ respectively. Then by assumption, we have $$a^2=b_1^2+\cdots b_m^2,\ a'^2=b_1'^2+\cdots b_m'^2,\ \mathrm{and}\ aa'=b_1b_1'+\cdots b_mb_m'.$$ Thus we have $$(aa')^2=(b_1b_1'+\cdots +b_mb_m')^2=(b_1^2+\cdots b_m^2)(b_1'^2+\cdots b_m'^2).$$ From the equality condition of Cauchy inequality, this shows that $(b_1,...,b_m)$ and $(b_1',...,b_m')$ are linearly dependent. We then deduce that $(a,b_1,...,b_m)$ and $(a',b_1',...,b_m')$ are linearly dependent.
\end{proof}
The following lemma might be well-known to experts. For reader's convenience, we recall briefly the proof here.
\begin{lemma}
\label{lem-c_2-decreasing}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a torsion-free sheaf on $X$. Then for any ample divisors $H_1,...,H_{n-2}$, we have $$c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2} \geqslant c_2(E^{**})\cdot H_1\cdot \cdots \cdot H_{n-2}.$$ Moreover, the equality holds if and only if $E$ is locally free in codimension $2$.
\end{lemma}
\begin{proof}
If $n=2$, then the lemma follows from \cite[Lemma 10.9]{Meg92}. We assume then that $n\geqslant 2$. We may assume that $H_1,...,H_{n-2}$ are effective sufficiently ample divisors in general positions. Let $S$ be their intersection. Since $X$ is smooth, there is a finite free resolution of $E$ as follows, $$0\to F_k\to\cdots\to F_0\to E.$$ Since $S$ is in general position, we may assume that $E|_S$ is still torsion-free, and that $$0\to F_k|_S\to\cdots\to F_0|_S\to E|_S$$ is again a free resolution. Hence $$c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2} =c_2(E|_S).$$
By the same argument, we may assume that $E^{**}|_S$ is still reflexive and is isomorphic to $(E|_S)^{**}$. Moreover, we may also assume that $$c_2(E^{**})\cdot H_1\cdot \cdots \cdot H_{n-2} = c_2(E^{**}|_S) = c_2((E|_S)^{**}).$$ By \cite[Lemma 10.9]{Meg92}, we have $$c_2((E|_S)^{**}) \geqslant c_2(E|_S),$$ and the equality holds if and only if $E|_S$ is locally free. Since $S$ is in general position, the sheaf $E|_S$ is locally free if and only $E$ is locally free in codimension $2$. This completes the proof of the lemma.
\end{proof}
\subsection*{Case of $\nu(c_1(E)) \geqslant 2$} We will consider the case when $c_1(E)$ has numerical dimension at least $2$, and will show that $E$ is an extension of a torsion-free sheaf with numerically trivial first Chern class by an invertible sheaf. To this end, we only need the weaker condition that $E$ is generically $(H_1,...,H_{n-2})$-semipositive for some ample divisors $H_1,...,H_{n-2}$. We first prove the following lemma.
\begin{lemma}
\label{lem-c2-quotient=0}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf on $X$ which is generically $(H_1,...,H_{n-2})$-semipositive for some ample divisors $H_1,...,H_{n-2}$. Assume that $c_1(E)$ is nef and $$c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}=0.$$ Let $0\to F\to E\to Q\to 0$ be an exact sequence of non-zero torsion-free sheaves. If $c_1(Q)\equiv 0$, then $F$ is generically $(H_1,...,H_{n-2})$-semipositive, and $$c_2(Q)\cdot H_1\cdot \cdots \cdot H_{n-2}=c_2(F)\cdot H_1\cdot \cdots \cdot H_{n-2}=0.$$
\end{lemma}
\begin{proof}
Since $c_1(Q)\equiv 0$, we have $c_1(E/G)\equiv c_1(F/G)$ for any saturated subsheaf $G$ of $F$. This implies that $F$ is generically $(H_1,...,H_{n-2})$-semipositive and $c_1(F)\equiv c_1(E)$ is nef. We have $$c_2(E)=c_2(F)+c_2(Q)+c_1(F)\cdot c_1(Q)\equiv c_2(F)+c_2(Q).$$ We note that $Q$ is also generically $(H_1,...,H_{n-2})$-semipositive. By Miyaoka inequality (see \cite[Theorem 6.1]{Miy87}), $c_2(F) \cdot H_1\cdot \cdots \cdot H_{n-2} \geqslant 0$ and $c_2(Q) \cdot H_1\cdot \cdots \cdot H_{n-2} \geqslant 0.$ We obtain that $$0= c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}=c_2(F) \cdot H_1\cdot \cdots \cdot H_{n-2} +c_2(Q) \cdot H_1\cdot \cdots \cdot H_{n-2} \geqslant 0.$$ This implies that $$c_2(Q)\cdot H_1\cdot \cdots \cdot H_{n-2}=c_2(F)\cdot H_1\cdot \cdots \cdot H_{n-2}=0.$$
\end{proof}
\begin{prop}
\label{prop-HN-E-ne-0-nd>1}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf of rank at least $2$ on $X$ which is generically $(H_1,...,H_{n-2})$-semipositive for some ample divisors $H_1,...,H_{n-2}$. Assume that $c_1(E)$ is nef with numerical dimension at least $2$. Let $\alpha$ be the class of $c_1(E)\cdot H_1\cdots H_{n-2}$. If $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}=0$, then the Harder-Narasimhan filtration of $E$ with respect to $\alpha$ is of the form $$0\subsetneq E_1 \subsetneq E,$$ such that
\begin{enumerate}
\item $E_1$ is an invertible sheaf such that $c_1(E_1)\equiv c_1(E)$;
\item $c_1(E/E_1)\equiv 0;$
\item $c_2(E/E_1)\cdot H_1\cdot \cdots \cdot H_{n-2}=0.$
\end{enumerate}
\end{prop}
\begin{proof}
Since $c_1(E)$ is nef and has numerical dimension at least $2$, $c_1(E)^2\cdot H_1 \cdot \cdots \cdot H_{n-2}>0.$ In particular, the class $\alpha$ is not zero. Since $c_2(E)\cdot H_1\cdot \cdots \cdot H_{n-2}=0$, and since $E$ has rank at least $2$, Bogomolov-Gieseker inequality for semistable sheaves (see \cite[Corollary 4.7]{Miy87}) shows that $E$ is not $\alpha$-semistable. Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r=E $$ be the Harder-Narasimhan semistable filtration. We note that (1) implies (2), and hence (3) by Lemma \ref{lem-c2-quotient=0}.
Therefore, we only need to prove that the filtration has length $r=2$ and satisfies property (1).
Without loss of generality, we may assume that $H_1,...,H_{n-2}$ are effective very ample divisors in general position. Let $S$ be the intersection surface of $H_1,...,H_{n-2}$. For simplicity, we let $G_i=E_i/E_{i-1}$ and $r_i=\mathrm{rank}\, G_i$ for $i=1,...,r$. We have
\begin{eqnarray*}
2c_2(E)\cdot S &=& (\sum_{i=1}^r 2c_2(G_i) + \sum _{1\leqslant i < j \leqslant r} 2c_1(G_i) c_1 (G_j))\cdot S\\
&= & (\sum_{i=1}^r 2c_2(G_i) + c_1(E)^2 -\sum _{ i =1}^r c_1(G_i)^2 )\cdot S.
\end{eqnarray*}
Since each $G_i$ is $\alpha$-semistable, Bogomolov-Gieseker inequality (see \cite[Corollary 4.7]{Miy87}) shows that
\begin{eqnarray*}
2c_2(E)\cdot S &\geqslant& (\sum_{i=1}^r \frac{r_i-1}{r_i} c_1(G_i)^2 + c_1(E)^2 -\sum _{ i =1}^r c_1(G_i)^2)\cdot S\\
&=& ( c_1(E)^2 -\sum _{ i =1}^r \frac{1}{r_i} c_1(G_i)^2)\cdot S.
\end{eqnarray*}
Since $c_1(E)^2\cdot S >0$, Hodge index theorem on $S$ shows that $$(c_1(E)^2\cdot S )(c_1(G_i)^2 \cdot S)\leqslant ( c_1(E) \cdot c_1(G_i) \cdot S)^2.$$ Therefore,
\begin{equation*}\label{eq:hodge:index} \tag{$*$}
c_1(G_i)^2 \cdot S \leqslant \frac{ (c_1(E) \cdot c_1(G_i) \cdot S)^2}{c_1(E)^2\cdot S }.
\end{equation*}
We let $$a_i = \frac{ c_1(G_i) \cdot c_1(E) \cdot S }{r_i c_1(E)^2\cdot S}=\frac{\mu_{\alpha}(G_i)}{c_1(E)^2\cdot S}.$$ Then by the definition of Harder-Narashimhan filtration, we see that $a_1>\cdots > a_r$. Moreover, we note that $\sum_{i=1}^r r_i a_i=1$ and $a_i\geqslant 0$ for all $i$. Hence $a_i\leqslant 1$ for all $i$. The inequality (\ref{eq:hodge:index}) becomes $$c_1(G_i)^2 \cdot S \leqslant r_i^2a_i^2 c_1(E)^2\cdot S.$$ Therefore, we have
\begin{eqnarray*}
2c_2(E)\cdot S &\geqslant& ( c_1(E)^2 -\sum _{ i =1}^r \frac{1}{r_i} c_1(G_i)^2)\cdot S \\
&\geqslant& ( c_1(E)^2 -\sum _{ i =1}^r \frac{1}{r_i} \cdot r_i^2a_i^2 c_1(E)^2 )\cdot S \\
&=& ( c_1(E)^2 -\sum _{ i =1}^r r_ia_i^2 c_1(E)^2 )\cdot S\\
&=&(1- \sum _{ i =1}^r r_ia_i^2) c_1(E)^2\cdot S\\
&\geqslant& (1- \sum _{ i =1}^r r_ia_ia_1) c_1(E)^2\cdot S\\
&=& (1- a_1) c_1(E)^2\cdot S.
\end{eqnarray*}
By assumption, $c_2(E)\cdot S=0$ and $c_1(E)^2\cdot S>0$. Since $a_1\leqslant 1$, the inequality above shows that $a_1=1$. We recall that $\sum_{i=1}^r r_ia_i=1$, $a_i\geqslant 0$ for all $i$ and $a_1>\cdots > a_r$. Hence we can only have $r=2$, $r_1=1$ and $a_2=0$.
It remains to prove that $c_1(E/E_1)\equiv 0$. On the one hand, we have $$c_1(E/E_1)\cdot c_1(E) \cdot S = c_1(E)^2\cdot S - c_1(E_1)\cdot c_1(E)\cdot S = 0.$$ Since $c_2(E_1)=0$, we obtain that
\begin{eqnarray*}
0 = c_2(E) \cdot S &=& (c_1(E_1)\cdot c_1(E/E_1) + c_2(E/E_1))\cdot S \\
&=& (c_1(E)\cdot c_1(E/E_1)-c_1(E/E_1)^2+c_2(E/E_1))\cdot S\\
&=& (-c_1(E/E_1)^2+c_2(E/E_1))\cdot S.
\end{eqnarray*}
Since $E/E_1=G_2$ is semistable, Bogomolov-Gieseker inequality (see \cite[Corollary 4.7]{Miy87}) shows that $$c_2(E/E_1)\cdot S \geqslant \frac{r_2-1}{2r_2} c_1(E/E_1)^2 \cdot S.$$ Therefore, we have $$0=(-c_1(E/E_1)^2+c_2(E/E_1))\cdot S \geqslant (-1+\frac{r_2-1}{2r_2}) c_1(E/E_1)^2 \cdot S,$$ which implies that $c_1(E/E_1)^2 \cdot S \geqslant 0.$
On the other hand, from Lefschetz theorem, we see that the symmetric bilinear form $\mathbf{q}(\delta, \delta')=\delta \cdot \delta' \cdot S$ defined on $N^1(X)$, the space of real numerical divisors classes, is non-degenerated. By Hodge index theorem, $\mathbf{q}$ has exactly one positive eigenvalue. We also have $\mathbf{q}(c_1(E),c_1(E))>0$, and $\mathbf{q}(c_1(E),c_1(E/E_1))=0.$ Hence, by Sylvester theorem, $$c_1(E/E_1)^2\cdot S= \mathbf{q}(c_1(E/E_1),c_1(E/E_1))\leqslant 0,$$ and the equality holds if and only if $c_1(E/E_1)\equiv 0$. This completes the proof of the proposition.
\end{proof}
\subsection*{Case of $\nu(c_1(E)) =1$}
Next we will consider the case when $\nu(c_1(E)) = 1$. In this case, $E$ is semistable with respect to the class $c_1(E)\cdot H^{n-2}$, and the Harder-Narasimhan filtration with respect to this class does not provide any further information. In order to obtain more information, we will use curve classes of the form $(c_1(E)+ \epsilon H)^{n-1}$ with $\epsilon>0$. We will prove the following results.
\begin{prop}
\label{prop-HN-E-ne-nd=1}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf such that $c_1(E)$ is nef with numerical dimension $1$, and that $E$ is generically nef. Assume that $c_2(E)\cdot H^{n-2}=0$ for some ample divisor $H$. Then there is a filtration $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ such that
\begin{enumerate}
\item for each $\epsilon>0$ small enough, it is the Harder-Narasimhan filtration with respect to the class $$\alpha_\epsilon=(c_1(E)+\epsilon H)^{n-1}.$$
\item $c_1(E_k/E_{k-1})$ is nef and numerically proportional to $c_1(E)$ for any $k=1,...,r$;
\item $c_2(E_k/E_{k-1})\cdot (c_1(E)+\epsilon H)^{n-2}=0$ for any $k=1,...,r$ and any $\epsilon>0$ small enough.
\end{enumerate}
\end{prop}
The proof of the proposition consists of several lemmas. The existence of common Harder-Narasimhan filtration follows from the following lemma.
\begin{lemma}
\label{lem-common-filtration}
Let $X$ be a normal projective $\mathbb{Q}$-factorial variety of dimension $n\geqslant 2$. Let $D_1,...,D_{n-1}$ be nef divisors and let $H$ be an ample divisors. Let $E$ be a torsion-free sheaf on $X$. Then there is a filtration $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ such that for any $\epsilon>0$ small enough, it is the Harder-Narasimhan filtration with respect to the class of $(D_1+{\epsilon} H) \cdot \cdots \cdot (D_{n-1}+{\epsilon} H)$.
\end{lemma}
\begin{proof}
See \cite[Lemma 6.5]{KMMc04}.
\end{proof}
We also have the following variant.
\begin{lemma}
\label{lem-common-filtration-stable}
Let $X$ be a normal projective $\mathbb{Q}$-factorial variety of dimension $n\geqslant 2$. Let $D_1,...,D_{n-1}$ be nef divisors and let $H$ be an ample divisors. Let $E$ be a torsion-free sheaf on $X$ which is semistable with respect to the class $\alpha_{{\epsilon}} $ of $(D_1+{\epsilon}H) \cdot \cdots \cdot (D_{n-1}+{\epsilon} H)$ for all ${\epsilon}>0$ small enough. Then there is a filtration $$0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ such that it is a Jordan-H\"older filtration with respect to the class $\alpha_{{\epsilon}}$ for all $\epsilon>0$ small enough.
\end{lemma}
We remind that a Jordan-H\"older filtration for a semistable sheaf is a filtration such that each quotient $E_k/E_{k-1}$ is stable.
\begin{proof}
The proof is similar to the one of \cite[Lemma 6.5]{KMMc04}. By induction, it is enough to construct the first term $E_1$ in the filtration. That is, it is enough to find a saturated non-zero subsheaf which is $\alpha_{\epsilon}$-stable for all $\epsilon>0$ small enough.
For each $\epsilon>0$ small enough, we fix a non-zero $\alpha_{{\epsilon}}$-stable saturated subsheaf $F_{\epsilon} \subseteq E$ of smallest rank with $\mu_{\alpha_{\epsilon}}(F_{\epsilon}) = \mu_{\alpha_{\epsilon}}(E)$. We note that, for all positive $\epsilon$ and $\delta$ small enough, $$\mu_{\alpha_{\epsilon}}(F_\delta)=\sum_{k=0}^{n-1}( \frac{1}{\mathrm{rank} \, F_\delta} \cdot \sum_{ \substack{i_1+\cdots +i_{n-1}=n-1-k, \\ 1\geqslant i_1,...,i_{n-1}\geqslant 0}} c_1(F_{\delta}) \cdot H^kD_1^{i_1}\cdots D_{n-1}^{i_{n-1}}) \epsilon^k.$$
For each $\delta>0$ small enough, we can denote by $P_{\delta}$ the polynomial function such that $\mu_{\alpha_\epsilon}(F_\delta)=P_\delta(\epsilon)$ for all $\epsilon>0$ small enough.
Since the sets of rational numbers $$\{ c_1(F) \cdot H^kD_1^{i_1}\cdots D_{n-1}^{i_{n-1}} \ |\ F \mbox{ is a subsheaf of }E \}$$ are bounded from above, and since the set $\{\mathrm{rank} \, F_{\epsilon}\ |\ \epsilon>0\}$ is finite, by \cite[Lemma 6.4]{KMMc04}, there is some $\eta>0$ small such that $$P_\eta(\epsilon) \geqslant P_\delta (\epsilon)$$ for all $\epsilon,\delta >0$ small enough. Moreover, the equality holds for some $\epsilon,\delta>0$ small enough if and only if $P_\eta=P_\delta$.
By definition of $F_\epsilon$, we have $$P_{\epsilon}(\epsilon) = \mu_{\alpha_\epsilon}(F_\epsilon)\geqslant \mu_{\alpha_\epsilon} (F_\eta) = P_{\eta}(\epsilon)$$ for all $\epsilon>0$ small enough. Therefore, for all $\epsilon>0$ small enough, we have $P_\epsilon=P_\eta$. As a consequence, for all $\epsilon, \delta >0$ small enough,
\begin{equation*}
{\mu_{\alpha_\epsilon}}(F_\epsilon) = P_\epsilon(\epsilon)=P_\eta(\epsilon)=P_\delta(\epsilon)= \mu_{\alpha_\epsilon}(F_\delta). \tag{**}
\end{equation*}
Let $E_1=F_{\gamma}$ for some $\gamma>0$ small enough such that $\mathrm{rank} \, F_{\gamma}\leqslant \mathrm{rank}\, F_{\epsilon}$ for all $\epsilon>0$ small enough. We claim that $E_1$ is $\alpha_{\epsilon}$-stable for all $\epsilon>0$ small enough. Indeed, by the equation (**), we have $$\mu_{\alpha_{\epsilon}}(E_1) = \mu_{\alpha_{\epsilon}}(F_{\gamma})= \mu_{\alpha_{\epsilon}}(F_{\epsilon}) = \mu_{\alpha_{\epsilon}}(E).$$ Since $F_{\epsilon}$ is a $\alpha_{{\epsilon}}$-stable subsheaf of smallest rank, and since $\mathrm{rank}\, E_1\leqslant \mathrm{rank}\, F_{\epsilon}$, we obtain that $E_1$ is ${\alpha_{\epsilon}}$-stable.
\end{proof}
The second and the third property in Proposition \ref{prop-HN-E-ne-nd=1} are consequences of the two lemmas below.
\begin{lemma}
\label{lem-filtration-c2}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf. Let $ H_1,...,H_{n-2}$ be ample divisors and let $D$ be a nef divisor such that the class $\alpha$ of $D\cdot H_1\cdot \cdots \cdot H_{n-2}$ is not zero. Assume that
\begin{enumerate}
\item[(i)] $c_1(E)^2\cdot H_1 \cdot \cdots \cdot H_{n-2}=c_2(E)\cdot H_1 \cdot \cdots \cdot H_{n-2}=0;$
\item[(ii)] there is a filtration of saturated subsheaves $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ such that $E_k/E_{k-1}$ is $\alpha$-semistable and that $$c_1(E_k/E_{k-1})^2\cdot H_1 \cdot \cdots \cdot H_{n-2} \leqslant 0 $$ for each $k=1,...,r$.
\end{enumerate}
Then each $E_k/E_{k-1}$ is locally free in codimension 2 and
\begin{eqnarray*}
c_1(E_k/E_{k-1})^2 \cdot H_1 \cdot \cdots \cdot H_{n-2} &=& c_2(E_k/E_{k-1}) \cdot H_1 \cdot \cdots \cdot H_{n-2} \\
&=&c_2((E_k/E_{k-1})^{**}) \cdot H_1 \cdot \cdots \cdot H_{n-2} \\
&=& 0.
\end{eqnarray*}
\end{lemma}
\begin{proof}
Let $S=H_1 \cdot \cdots \cdot H_{n-2}$. Then we have $$0=2c_2(E)\cdot S = (\sum_{k=1}^r 2 c_2(E_k/E_{k-1}) + \sum_{1\leqslant j < k \leqslant r} 2 c_1(E_j/E_{j-1}) \cdot c_1(E_{k}/E_{k-1})) \cdot S.$$ By Lemma \ref{lem-c_2-decreasing}, we obtain that $$ 0 \geqslant (\sum_{k=1}^r 2 c_2((E_k/E_{k-1})^{**}) + \sum_{1\leqslant j < k \leqslant r} 2 c_1(E_j/E_{j-1}) \cdot c_1(E_{k}/E_{k-1})) \cdot S.$$ Since each $(E_k/E_{k-1})^{**}$ is also $\alpha$-semistable, by Bogomolov-Gieseker inequality (see \cite[Corollary 4.7]{Miy87}), we have
\begin{eqnarray*}
0&\geqslant& (\sum_{k=1}^r \frac{d_k-1}{d_k} c_1(E_k/E_{k-1})^2 + \sum_{1\leqslant j < k \leqslant r} 2 c_1(E_j/E_{j-1}) \cdot c_1(E_{k}/E_{k-1})) \cdot S\\
&=&(c_1(E)^2-\sum_{k=1}^r\frac{1}{d_k}c_1(E_k/E_{k-1})^2)\cdot S \\
&=& -\sum_{k=1}^r(\frac{1}{d_k}c_1(E_k/E_{k-1})^2 \cdot S) \\
&\geqslant & 0,
\end{eqnarray*}
where $d_k$ is the rank of $E_k/E_{k-1}$. Therefore, all of the inequalities above are equalities. We conclude hence $$ c_1(E_k/E_{k-1})^2 \cdot S =0, $$ and $$c_2(E_k/E_{k-1}) \cdot S =c_2((E_k/E_{k-1})^{**}) \cdot S = \frac{d_k-1}{2d_k} c_1(E_k/E_{k-1})^2 \cdot S =0,$$ for each $k=1,...,r$. From the equality condition in Lemma \ref{lem-c_2-decreasing}, we also obtain that $E_k/E_{k-1}$ is locally free in codimension $2$.
\end{proof}
\begin{lemma}
\label{lem-HN-E-ne-nd=1}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf such that $c_1(E)$ is nef with numerical dimension $1$ and that $c_2(E)\cdot H^{n-2}=0$ for some ample an ample divisor $H$.
Assume that there is a filtration of saturated subsheaves $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ which satisfies the following two conditions:
\begin{enumerate}
\item[(i)] for each $\epsilon>0$ small enough and each $k=1,...,r$, the sheaf $E_k/E_{k-1}$ is semistable with respect to the class $$\alpha_\epsilon=(c_1(E)+\epsilon H)^{n-1};$$
\item[(ii)] for each $\epsilon>0$ small enough and each $k=1,...,r$, $$\mu_{\alpha_\epsilon}(E_k/E_{k-1})\geqslant 0.$$
\end{enumerate}
Then $c_1(E_k/E_{k-1})$ is nef and numerically proportional to $c_1(E)$ for any $k=1,...,r$ .
\end{lemma}
\begin{proof}
We can assume that $H$ is very ample. Let $S$ be the surface cut out by general members of the linear system of $H$.
Since $c_1(E)$ has numerical dimension $1$, we have $c_1(E)^i\cdot H^{n-i}=0$ for $i\geqslant 2$.
Thus, for each $\epsilon>0$, $$\alpha_\epsilon = \binom{n-1}{2}\epsilon^{n-2} \cdot (c_1(E) + \frac{\epsilon}{\binom{n-1}{2}} H )\cdot S.$$ Let $\beta_{\eta}$ be the curve class $$\beta_\eta = (c_1(E)+\eta H)\cdot S$$ for any $\eta >0$.
Then for any $\eta>0$ small enough and any $k=1,...,r$, condition (i) implies that $E_k/E_{k-1}$ is semistable with respect to $\beta_\eta$ and condition (ii) implies that $\mu_{\beta_\eta}(E_k/E_{k-1})\geqslant 0.$ In particular, if we let $\eta$ go to zero, then we obtain $$c_1(E_k/E_{k-1})\cdot c_1(E) \cdot S \geqslant 0$$ for any $k=1,...,r$. Since $c_1(E)=\sum_{k=1}^r c_1(E_k/E_{k-1})$ and $c_1(E)^2\cdot S = 0$, this shows that $$c_1(E_k/E_{k-1})\cdot c_1(E) \cdot S=0. $$
From Lefschetz theorem,
the symmetric bilinear form $\mathbf{q}(\delta, \delta')=\delta \cdot \delta' \cdot S$ defined on $N^1(X)$ is non-degenerated, where $N^1(X)$ is the space of real numerical divisors classes. By Hodge index theorem, $\mathbf{q}$ has exactly one positive eigenvalue. Since $c_1(E)\not\equiv 0$, the condition $$\mathbf{q}(c_1(E_k/E_{k-1}),c_1(E))=\mathbf{q}(c_1(E),c_1(E))=0$$ implies that $\mathbf{q}(c_1(E_k/E_{k-1}),c_1(E_k/E_{k-1}))\leqslant 0$ by Sylvester theorem.
Since each $E_k/E_{k-1}$ is semistable with respect to the class $\alpha_\epsilon$ for $\epsilon>0$ small enough, by Lemma \ref{lem-filtration-c2}, we obtain that $$\mathbf{q}(c_1(E_k/E_{k-1}),c_1(E_k/E_{k-1}))=0.$$ Thus, by Lemma \ref{lem-2form-1}, $c_1(E_k/E_{k-1})$ is numerically proportional to $c_1(E)$ for any $k=1,...,r$. Moreover, it is nef since $\mu_{\alpha_\epsilon}(E_k/E_{k-1}) \geqslant 0$ for $\epsilon>0$ small enough.
\end{proof}
Now we can deduce Proposition \ref{prop-HN-E-ne-nd=1}.
\begin{proof}[{Proof of Proposition \ref{prop-HN-E-ne-nd=1}}]
Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ be a common Harder-Narasimhan filtration with respect the classes $\alpha_\epsilon$ for all $\epsilon>0$ small enough (see Lemma \ref{lem-common-filtration}). Then this filtration satisfies property (1). Since $E$ is generically nef, $\mu_{\alpha_\epsilon, min }(E) \geqslant 0$ for any $\epsilon>0$. Thus for each $\epsilon>0$ small enough and each $k=1,...,r$, we have $$\mu_{\alpha_\epsilon}(E_k/E_{k-1}) \geqslant \mu_{\alpha_\epsilon}(E_r/E_{r-1})=\mu_{\alpha_\epsilon, min }(E) \geqslant 0.$$ Property (2) then follows from Lemma \ref{lem-HN-E-ne-nd=1}. In particular, $c_1(E_k/E_{k-1})^2\cdot H^{n-2} = 0 $ for each $k=1,...,r$. Thus property (3) follows from Lemma \ref{lem-filtration-c2}.
\end{proof}
Next we will study each semistable component $E_k/E_{k-1}$ in Proposition \ref{prop-HN-E-ne-nd=1}. We note that equality holds in the Bogomolov-Gieseker inequality for $E_k/E_{k-1}$. For such sheaves, we will prove the following proposition. Our main ingredient is a theorem of Bando-Siu on stable reflexive sheaves (see \cite[Corollary 3]{BS94}).
\begin{prop}
\label{prop-BG-equality-cond}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a torsion-free sheaf and let $H$ be an ample divisor. Assume that
\begin{enumerate}
\item[(i)] $E$ is semistable with respect to the class $\alpha_\epsilon = (c_1(E)+ \epsilon H)^{n-1}$ for all ${\epsilon}>0$ small enough;
\item[(ii)] $c_1(E)$ is nef with numerical dimension $1$;
\item[(iii)] $ c_2(E) \cdot (c_1(E)+ \epsilon H)^{n-2}=0$ for all $\epsilon>0$ small enough.
\end{enumerate}
Then there is a filtration $$0 =E_0 \subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ such that
\begin{enumerate}
\item $E_k/E_{k-1}$ is $\alpha_{\epsilon}$-stable for all $\epsilon>0$ small enough;
\item $c_1(E_k/E_{k-1})$ is nef and numerically proportional to $c_1(E)$;
\item $ E_k/E_{k-1} $ is locally free in codimension $2$, and
$$c_2( E_k/E_{k-1}) \cdot (c_1(E)+\epsilon H)^{n-2}=c_2((E_k/E_{k-1})^{**}) \cdot (c_1(E)+\epsilon H)^{n-2}=0$$ for $\epsilon>0$ small enough;
\item $(E_k/E_{k-1})^{**}$ is locally free and projectively flat.
\end{enumerate}
As a consequence, there is an integral divisor $D$ such that $c_1(E)\equiv lD$, where $l= \mathrm{rank} \, E$.
\end{prop}
\begin{proof}
Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ be a common Jordan-H\"older filtration with respect the classes $\alpha_\epsilon$ for all $\epsilon>0$ small enough (see Lemma \ref{lem-common-filtration-stable}). Then this filtration satisfies property (1). Since $E$ is $\alpha_\epsilon$-semistable for all $\epsilon>0$ small enough and since $c_1(E)$ is nef, we have $\mu_{\alpha_\epsilon}(E_k/E_{k-1})\geqslant 0$ for any $k=1,...,r$. Property (2) then follows from Lemma \ref{lem-HN-E-ne-nd=1}. Thus $c_1(E_k/E_{k-1})^2\equiv 0$ for any $k=1,...,r$ for $c_1(E)$ has numerical dimension $1$. Since $c_2(E)\cdot (c_1(E)+\epsilon H)^{n-2}=0$ for $\epsilon>0$ small enough, we can apply Lemma \ref{lem-filtration-c2} to obtain property (3).
Since $(E_k/E_{k-1})^{**}$ is also $\alpha_\epsilon$-stable, by \cite[Theorem 3]{BS94}, it admits an admissible Einstein-Hermitian metric. We note that property (3) implies that equality holds in the Bogomolov-Gieseker inequality (see \cite[Corollary 4.7]{Miy87}) for $(E_k/E_{k-1})^{**}$. Thus, by \cite[Corollary 3]{BS94}, $(E_k/E_{k-1})^{**}$ is locally free and projectively flat. This proves property (4).
Since $(E_k/E_{k-1})^{**}$ is projectively flat, $c_1(E_k/E_{k-1})$ is divisible by the rank of $E_k/E_{k-1}$, that is, there is some integral divisor $D_k$ such that $$c_1(E_k/E_{k-1})= \mathrm{rank} \, (E_k/E_{k-1}) \cdot D_k.$$ We have $$c_1(E)=\sum_{k=1}^{r} c_1(E_k/E_{k-1})=\sum_{k=1}^{r}\mathrm{rank} \, (E_k/E_{k-1}) \cdot D_k.$$ We note that the $D_k$ are nef and numerically proportional to $c_1(E)$ by property (2). Since $E$ is $\alpha_\epsilon$-semistable, the $D_k$ must be numerically equivalent to each other.
Thus if we let $D=D_1$, then $c_1(E)\equiv lD$. This completes the proof of the proposition.
\end{proof}
\subsection*{Case of $\nu(c_1(E))= 0$} We will finish this section with the case when $\nu(c_1(E)) = 0$. We note that $c_1(E)$ is semistable with respect to the class $H^{n-1}$ in this case. We will prove the following proposition.
\begin{prop}
\label{prop-HN-E-ne-nd=0}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$. Let $E$ be a non-zero torsion-free sheaf such that $c_1(E)\equiv 0$ and that $E$ is generically nef. Let $$0=E_0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ be a Jordan-H\"older filtration with respect to the class $\alpha=H^{n-1}$. Then
\begin{enumerate}
\item $c_1(E_k/E_{k-1})\equiv 0$.
\end{enumerate}
If we assume further that $c_2(E)\cdot H^{n-2}=0$ for some ample divisor $H$, then
\begin{enumerate}
\item[(2)] $ E_k/E_{k-1} $ is locally free in codimension $2$, and
$$c_2( E_k/E_{k-1}) \cdot H^{n-2}=c_2((E_k/E_{k-1})^{**}) \cdot H^{n-2}=0;$$
\item[(3)] $(E_k/E_{k-1})^{**}$ is a flat locally free sheaf.
\end{enumerate}
\end{prop}
\begin{proof}
For property (1), by induction, it is enough to prove that $c_1(E/E_{1})\equiv 0$. Since $E$ is generically nef, so is $E/E_1$. Hence $c_1(E/E_1)\cdot H_1\cdot \cdots \cdot H_{n-1}\geqslant 0$ for any ample divisor $H_1,...,H_{n-1}.$ Since $\mu_{\alpha}(E/E_1)=\mu_{\alpha}(E)=0$, we have $c_1(E/E_1)\cdot H^{n-1}=0$. Hence by Lemma \ref{lem-vanishing}, we obtain that $c_1(E/E_1)\equiv 0.$
Property (2) follows from Lemma \ref{lem-filtration-c2}. It remains to prove property (3). Since $(E_k/E_{k-1})^{**}$ is also $\alpha_\epsilon$-stable, by \cite[Theorem 3]{BS94}, it admits an admissible Einstein-Hermitian metric. Property (2) implies that equality holds in the Bogomolov-Gieseker inequality (see \cite[Corollary 4.7]{Miy87}) for $(E_k/E_{k-1})^{**}$. Thus, by \cite[Corollary 3]{BS94}, $(E_k/E_{k-1})^{**}$ is locally free and projectively flat. Since $c_1(E_k/E_{k-1})\equiv 0$, by \cite[Lemma 4.4.12]{Kob87}, we obtain that $(E_k/E_{k-1})^{**}$ is flat.
\end{proof}
\section{Proof of Theorem \ref{thm-classification}}
We will finish the proof of Theorem \ref{thm-classification} in this section. We will need the following proposition.
\begin{prop}
\label{prop-c2=0-q<quotient}
Let $X$ be a smooth projective variety of dimension $n\geqslant 2$ with nef anticanonical class $-K_X$. Assume that $c_2(T_X)\cdot H^{n-2}=0$ for some ample divisor $H$. Assume further that there is some non-zero torsion-free quotient $T_X\to Q$ such that $c_1(Q)\equiv 0$ and $\mathrm{rank}\, Q=k$. Then the augmented irregularity $\tilde{q}(X)\geqslant k$.
\end{prop}
We recall that the irregularity of a smooth projective variety $X$ is $q(X)=h^1(X,\mathscr{O}_X)$. It is equal to the dimension of the Albanese variety of $X$. The augmented irregularity is defined as $$\tilde{q}(X) = \sup \{q(\tilde{X})\ | \ \tilde{X}\to X \mbox{ is a finite {\'e}tale cover} \}.$$ We will first prove that $\tilde{q}(X)$ is not zero under the assumption of Proposition \ref{prop-c2=0-q<quotient}.
\begin{lemma}
\label{lem-c2=0-q=0}
Under the condition of Proposition \ref{prop-c2=0-q<quotient}, we have $\tilde{q}(X)\neq 0.$
\end{lemma}
\begin{proof}
Assume by contradiction that $\tilde{q}(X)=0$. Then by \cite[Theorem 2]{Pau97}, $X$ has finite fundamental group. By replacing $X$ by some finite \'etale cover if necessary, we may assume that $X$ is simply connected.
By Theorem \ref{thm-main}, $T_X$ is generically nef. Thus so is the quotient $Q$. From Lemma \ref{lem-c2-quotient=0}, we deduce that $c_2(Q)\cdot H^{n-2}=0$. By Proposition \ref{prop-HN-E-ne-nd=0}, there is some non-zero torsion-free quotient $Q\to G$ such that $G^{**}$ is a flat locally free sheaf. Since $X$ is simply connected, this implies that $G^{**}$ is isomorphic to a direct sum of copies of $\mathscr{O}_X$. We obtain then an injective morphism $\mathscr{O}_X\to \Omega_X^1.$ This shows that $q(X)=h^0(X,\Omega_X^1)\geqslant 1$, which is a contradiction.
\end{proof}
Now we will prove Proposition \ref{prop-c2=0-q<quotient}.
\begin{proof}[{Proof of Proposition \ref{prop-c2=0-q<quotient}}]
By replacing $X$ by some finite \'etale cover if necessary, we may assume that the Albanese morphism $f:X\to A$ induces an isomorphism of fundamental groups (see \cite[Theorem 2]{Pau97}). Then \cite[Theorem 1.2]{Cao16} implies that $f$ is an isotrivial fibration and every fiber of $f$ is simply connected. As a consequence, we have $$\tilde{q}(X)=q(X)=\mathrm{dim}\, A=q.$$
Assume by contradiction that the proposition does not hold. Then we must have $n\geqslant k > q$. There is an exact sequence $$0\to T_{X/A} \to T_X \to f^*T_A.$$ Let $F$ be a general fiber of $f$ and let $i:F\to X$ be the natural injection. Then $\tilde{q}(F)=0$ as $F$ is simply connected.
If $\mathrm{dim}\, F=1$, then we can only have $n=k=q+1$. In particular, $T_X=Q$ and $c_1(X)\equiv 0$. This shows that $F$ is an elliptic curve, which is a contradiction. Hence we have $\mathrm{dim}\, F \geqslant 2$.
By restricting the exact above to $F$, we can obtain an exact sequence $$0\to T_F \to T_{X}|_F \to (\mathscr{O}_F)^{q}\to 0.$$ In particular, we have $c_2(T_F)=i^*c_2(T_X)$. Since $F$ is numerically equivalent to the complete intersection of nef divisors, by Corollary \ref{cor-c2} and Lemma \ref{lem-vanishing}, the vanishing condition on $c_2(T_X)$ implies that $$c_2(T_F)\cdot (H|_F)^{n-q-2}=c_2(T_X)\cdot H^{n-q-2}\cdot F=0.$$ We note that $F$ is a smooth projective manifold with nef anticanonical class $-K_F$. Hence $T_F$ is generically nef by Theorem \ref{thm-main}, and so is $T_X|_F$.
Let $Q|_F\to Q'$ be the quotient whose kernel is the torsion part of $Q|_F$. Since $Q$ is torsion-free, it is locally free in codimension $1$. Moreover, since $F$ is a general fiber, we may assume that $$c_1(Q')=i^* c_1(Q) \equiv 0.$$
Since $k>q$, the induced morphism $T_F\to Q'$ is non-zero. Let $R$ be its image. We claim that $c_1(R)\equiv 0.$ Let $\beta=(H|_F)^{n-1-q}$ be a curve class in $F$. Then $ \mu_{\beta}(R)\geqslant 0$ for $T_F$ is generically nef. We also note that $Q'$ is generically nef for it is a torsion-free quotient of $T_X|_F$. Thus $$\mu_{\beta, max}(Q')\geqslant \mu_{\beta, min}(Q')\geqslant 0.$$ However, since $c_1(Q')\equiv 0$, we must have $$\mu_{\beta, max}(Q')= \mu_{\beta, min}(Q') = \mu_{\beta}(Q')= 0.$$ Thus the condition $ \mu_{\beta}(R)\geqslant 0$ then implies that $\mu_{\beta}(R)=0$ as $R$ is a subsheaf of $Q'$. Since $R$ is generically nef, this implies that $c_1(R)\equiv 0$ by Lemma \ref{lem-vanishing}. Since $R$ is a quotient of $T_F$, by Lemma \ref{lem-c2=0-q=0}, we obtain that $\tilde{q}(F)>0.$ This is a contradiction as $F$ is simply connected.
\end{proof}
Now we can conclude Theorem \ref{thm-classification}.
\begin{proof}[{Proof of Theorem \ref{thm-classification}}]
Since the second Chern class of an abelian variety is zero, we see that the property (2) implies (1). We will prove that (1) implies (2). The case when $n=1$ is trivial. We assume from now on that $n\geqslant 2$. If $K_X\equiv 0$, then by Beauville decomposition theorem (see \cite[Th\'eor\`eme 1]{Bea83}), there is finite \'etale cover $X'\to X$ such that $X'$ is isomorphic to a product of irreducible holomorphic symplectic manifolds, Calabi-Yau manifolds and an abelian variety. The vanishing condition (1) on $c_2(T_X)$ then implies that $X'$ is an abelian variety.
We will now assume that $K_X\not\equiv 0$. Then $X$ is uniruled. By replacing $X$ by some finite \'etale cover if necessary, we may assume that the Albanese morphism $f:X\to A$ induces an isomorphism of fundamental groups (see \cite[Theorem 2]{Pau97}). Then \cite[Theorem 1.2]{Cao16} implies that $f$ is an isotrivial fibration and every fiber of $f$ is simply connected. As a consequence, we have $$\tilde{q}(X)=q(X)=\mathrm{dim}\, A=q.$$ Since $X$ is uniruled, $f$ is not an isomorphism. Moreover, general fibers of $f$ are uniruled.
Let $H$ be an ample divisor. By Lemma \ref{lem-vanishing}, the property (1) in the theorem implies that $c_2(T_X)\cdot H^{n-2}=0$. Thanks to Theorem \ref{thm-main-mov}, we know that $T_X$ is generically $(H,...,H)$-semipositive.
First we assume that $K_X^2\cdot H^{n-2}\neq 0,$ that is, $-K_X$ has numerical dimension at least $2$. Then by Proposition \ref{prop-HN-E-ne-0-nd>1}, there is a torsion-free quotient $T_X\to Q$ such that $c_1(Q)\equiv 0$ and $\mathrm{rank} \, Q=n-1$. Hence by Proposition \ref{prop-c2=0-q<quotient}, the augmented irregularity of $X$ satisfies $q \geqslant n-1$. Since we have assume that $f:X\to A$ is not an isomorphism, we have $q=n-1$. Then $f:X\to A$ is a $\p^1$-bundle.
Now we assume that $K_X^2\cdot H^{n-2} = 0.$ Since $-K_X\not\equiv 0$, this means that $-K_X$ has numerical dimension $1$. Let $$0\subsetneq E_1 \subsetneq \cdots \subsetneq E_r= E$$ be the common Harder-Narasimhan filtration with respect to $$(-K_X+\epsilon H)^{n-1}$$ for all $\epsilon >0$, as in Proposition \ref{prop-HN-E-ne-nd=1}. By Proposition \ref{prop-BG-equality-cond}, for each $k=1,...,r$, there is some nef divisor $D_k$ such that $c_1(E_k/E_{k-1}) \equiv l_kD_k$, where $l_k = \mathrm{rank} \, (E_k/E_{k-1})$. Since $c_1(E_k/E_{k-1})$ is numerically proportional to $c_1(T_X)$ for all $k=1,...,r$ by Proposition \ref{prop-HN-E-ne-nd=1}, each $D_k$ is also numerically proportional to $-K_X$. We have $$-K_X \equiv \sum_{k=1}^r l_kD_k.$$
We will discuss in two cases. In the first case, we assume that
\begin{enumerate}
\item[-] either there is some $k$ such that $l_k\geqslant 2$ and $D_k\not\equiv 0$,
\item[-] or there are $k\neq k'$ such that $D_k\not \equiv 0$ and $D_{k'}\not \equiv 0$.
\end{enumerate}
Since $X$ is uniruled, there is an elementary contradiction $g:X\to Y$ of some $K_X$-negative extremal ray $R$. The assumption above implies that the length of $R$ is at least $2$. Moreover, since the numerical dimension of $-K_X$ is $1$, and since $-K_X$ is ample on every fiber of $g$, we obtain that every fiber of $g$ has dimension at most $1$. By \cite[Theorem 1.1]{Wis91}, we have $\mathrm{dim}\, Y< n$. Since every fiber of $g$ has dimension at most $1$, we obtain that $\mathrm{dim}\, Y= n-1$. In addition, $g$ is a conic bundle and $Y$ is smooth by \cite[Theorem 3]{Ando84}. Since the length of $R$ is at least $2$, we obtain that $g$ is smooth. This implies that $g_*(K_X^2)\equiv -4K_Y$ (see \cite[Section 4.11]{Miyanishi83}), and thus $K_Y\equiv 0$.
Since $g$ has relative dimension $1$ and since $K_Y\equiv 0$, we have $c_2(T_X)\equiv g^*c_2(T_Y)$. The vanishing condition on $c_2(T_X)$ then shows that $c_2(T_Y)\cdot H'_1\cdots H'_{n-3}=0$ for any ample divisors $H'_1,...,H'_{n-3}$ on $Y$ if $\mathrm{dim}\, Y\geqslant 2$. Hence, there is a finite \'etale cover $Y'\to Y$ such that $Y'$ is an abelian variety. As a consequence, $X\times_Y Y'$ is a $\p^1$-bundle over an abelian variety.
We will now study the remaining case. The decomposition $-K_X \equiv \sum_{k=1}^r l_kD_k$ then satisfies both of the following two conditions:
\begin{enumerate}
\item[(i)] for all $k$, if $D_k\not\equiv 0$, then $l_k=1$;
\item[(ii)] there is at most one $k$ such that $D_k\not \equiv 0$.
\end{enumerate}
From the definition of Harder-Narasimhan filtration, the condition (ii) implies that the filtration has length $r\leqslant 2$. Since $K_X\not\equiv 0$, we can only have $r=2$ and $c_1(E/E_1)\equiv 0$. The condition (i) then implies that $\mathrm{rank}\, E_1=1$. Thus, by Proposition \ref{prop-c2=0-q<quotient}, the augmented irregularity of $X$ is at least $n-1$. We conclude then $f:X\to A$ is a $\p^1$-bundle.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Resonances as $L\rightarrow\infty$ when $V_{2,L}(x) = e^{-cL}\delta(x-L)$}\label{mudelta}
In this section we determine the asymptotic positions of the resonances of $H_L=-d^2/dx^2 + V_1(x) + e^{-cL}\delta(x-L)$for large $L$. The analysis is similar to the previous section, so we will omit many details. {At the end of this section we make some comments about how things change when instead of $ e^{-cL}\delta(x-L)$ we use $\mu(L) V_2(x-L)$ where $V_2$ }{ satisfies the same hypotheses as before} {and $\mu(L)$ can be $e^{-cL}$ or decrease faster or slower than an exponential. These comments are based on calculations presented in an Appendix.}
{For now we consider the moving delta function.} The functions $f_1(k)$ and $f_2(k)$ are defined by \eqref{eqn:f} as before. But now $f_2$ depends on $L$. We can compute explicitly that
$$
f_2(k) = \frac{e^{-cL} - 2ik}{e^{-cL}}.
$$
The zero of $f_2(k)$ gives the position of the single resonance of $-d^2/dx^2 + e^{-cL}\delta(x)$ at $k=-ie^{-cL}/2$. There are no reflectionless points for $V_2=\delta$ and therefore no poles for $f_2$. Equation \eqref{eqn:R} can now be written
\begin{equation}\label{eqn:R2}
e^{(2ik-c)L} = f_1(k) (e^{-cL} - 2ik)
\end{equation}
We find resonance free regions as before. Let
$$
U_1(a,A) = \{k\in{\mathbb C}_- : \Im(k) > -c/2 + a, |f_1(k)| > 1/A \hbox{\ and\ }|k| > 1/A\}
$$
for $a,A>0$. Now we are excluding neighbourhoods of resonances of $H$, since these occur at the zeros of $f_1$.
Then we let
$$
U_2(a,A) = \{k\in{\mathbb C}_- : \Im(k) < -c/2 - a, |f_1(k)| < A \hbox{\ and\ }|k| < A\},
$$
again for $a,A>0$. Here we are excluding neighbourhoods of reflectionless points of $H$ as before.
\begin{proposition}\label{leaving2}
There exists $L_0$ such that $U_1(a,A)\cap{\cal R}(L) = U_2(a,A)\cap{\cal R}(L) =\emptyset$ when $L>L_0$.
\end{proposition}
\begin{proof}
If $k\in U_1(a,A)$ is a resonance then using \eqref{eqn:R2} we find
$$
e^{-2aL}\ge |e^{(2ik-c)L}| = |f_1(k) (e^{-cL} - 2ik)|\ge \frac{1}{A}\left(\frac{2}{A}-e^{-cL}\right)
$$
If $k\in U_2(a,A)$ is a resonance then
$$
e^{2aL} \le |e^{(2ik-c)L}| = |f_1(k) (e^{-cL} - 2ik)|\le A(1+2A)
$$
both these inequalities are impossible for $L>L_0$ for some $L_0(a,A,c)$.
\end{proof}
{\begin{proposition}\label{converging}
If $k$ is a resonance of $H$ with $ \text{Im} (k) > -c/2$, then $k$ is a limit of resonances of $H_L$ as $L \to \infty$. If $k$ is a reflectionless point of $H$ with $ \text{Im} (k) < -c/2$ then $k$ is a limit of resonances of $H_L$ as $L \to \infty$. The resonances of $H$ with $\text{Im} (k) <-c/2$ are not limits of resonances of $H_L$. The resonances of $H_L$ accumulate densely on the line $\text{Im} (k) = -c/2$.
\end{proposition}}
{We do not explicitly give the proof of Proposition \ref {converging} but it will be clear as we treat the six cases below. }
Retracing our steps {from Section \ref{no coupling constant}}, we now find all the resonances near some point $k_0$ in the closed lower half plane. In view of \ref{leaving2} we consider six cases: $k_0 = 0$, $\Im(k_0)>-c/2$ and $f_1(k_0)=0$, $\Im(k_0)=-c/2$ and $0<|f_1(k_0)|<\infty$, $\Im(k_0)=-c/2$ and $f_1(k_0)=0$, $\Im(k_0)=-c/2$ and $k_0$ is a pole of $f_1$ and $\Im(k_0)<-c/2$ and $k_0$ is a pole of $f_1$. In each case {(except for $k_0 = 0$ where we will do more)} we will identify the approximate resonances in terms of the log or the Lambert $W$ function but omit the error estimates, which can be carried out as in the previous section. The error estimates say that near $k_0$ every resonance has a nearby approximate resonance and vice versa.
\bigskip
\noindent{\bf Case 1: $k_0 = 0$}\\
\medskip
We know that $|f_1(0)| \ge 1$ and that generically $f_1(0) =1$. But in fact there are situations where $f_1$ has a pole at $0$. So write equation (\ref{eqn:R2}) as
$$e^{2ikL} = \frac{g(k)}{k^p}(1-2ike^{cL})$$ where $g$ is analytic in a neighborhood of $|k| \le \delta < c/2$, $p$ is a non-negative integer and $|g(k)| \ge c_0 > 0$ for $|k|\le \delta$. Choose $\epsilon \in (2\delta, c)$ and let $|k|\le\delta$ . Then if $|k| > e^{-(c-\epsilon)L}$ the left side of (\ref{eqn:R2})
has modulus $\le e^ {2\delta L}$ while the modulus of the right side is $\ge \frac{c_0}{\delta^p} (2e^{\epsilon L} - 1)$ which is false for large $L$ if $k$ satisfies (\ref{eqn:R2}). Thus for large $L$ any solution to (\ref{eqn:R2}) with $|k| \le \delta$ satisfies $|k| \le e^{-(c-\epsilon)L}$. Iterating this argument once more with an improved estimate on the modulus of the left side of (\ref{eqn:R2}) shows that for large $L$ any solution to (\ref{eqn:R2}) with $|k| \le \delta$ satisfies $|k| \le R e^{-cL}$ with $R = 1/2 + c_0^{-1}$.
Let $\xi = ke^{cL}$ and $$f(\xi) = (2i)^{-1}\big(1 - e^{2ikL}(\xi e^{-cL})^p (g(\xi e^{-cL})^{-1} \big ).$$ Then (\ref{eqn:R2}) becomes $f(\xi) = \xi$. $R$ has been chosen so that it is clear that $f$ maps the closed ball $\overline{B(0,R)}$ into itself. In this closed ball we can see that $|f'(\xi)| \le C_1 Le^{-cL}$ so that for large $L$ there is a unique solution to (\ref{eqn:R2}) for $|k| \le \delta$. If $f_1(0) = 1$ then this is the $0$ solution. Thus generically $k=0$ is the only solution to (\ref{eqn:R2}) for $|k| \le \delta$. If $0 < |f_1(0) -1| <\infty$, let $\xi_1 = (2i)^{-1}(1-f_1(0)^{-1})$ and compute that $\xi_2 = f(\xi_1) = \xi_1 + O(Le^{-cL})$. Then if $\xi^* =f(\xi^*)$ we have $|\xi^* - \xi_2| = |f(\xi^*) - f(\xi_1)| \le \max_{t\in [01]} |f'( (1-t)\xi^* + t \xi_1)||\xi^* - \xi_1| = O(Le^{-cL})$. Thus if $0< |1-f_1(0)| < \infty$ the unique solution to (\ref{eqn:R2}) with $|k| \le \delta$ is $k = (2i)^{-1}(1-f_1(0)^{-1})e^{-cL} + O(Le^{-2cL})$. Finally suppose $p\ge 1$. Then $\xi^* = f(\xi^*) = 1/2i + O(e^{-pcL }) $ so the unique solution to (\ref{eqn:R2}) with $|k| \le \delta$ is $k = (2i)^{-1}e^{-cL} + O( e^{-(p+1)cL})$.\\
To summarize: In the generic case where $f_1(0) = 1$ the unique solution with $|k| < \delta$ is $k=0$. \\
If $|1-f_1(0)| <\infty$ then the unique solution with $|k| < \delta$ satisfies
$$ k = (2i)^{-1}(1-f_1(0)^{-1})e^{-cL} + O(Le^{-2cL})$$
and if $f_1$ has a pole of order $p$ at $0$ the unique solution with $|k| < \delta$ satisfies
$$k = (2i)^{-1}e^{-cL} + O( e^{-(p+1)cL}).$$
The case where $k_0=0$ and $f_1$ has a pole at $k_0$ is different from {all} the other cases with poles {in Section \ref {no coupling constant} and below}, because there is only one resonance near $k_0$ even when $p > 1$.
\bigskip
\noindent{\bf Case 2: $\Im(k_0)>-c/2$ and $f_1(k_0)=0$}\\
\medskip
These are the resonances of $V_1$ above the line $\Im(k)=-c/2$. Write $f_1(k) = (k-k_0)^p g(k)$ where $g(k)$ is analytic at $k_0$ and $g(k_0)\ne 0$. Then \eqref{eqn:R2} becomes
$$
(k-k_0)^p e^{(-2ik+c)L} = \frac{1}{g(k)(e^{-cL}-2ik)}.
$$
Let $\lambda =2 L/p$ and $\omega = e^{2\pi i/p}$. Let $G(k)$ be a fixed branch of $(g(k)(e^{-cL}-2ik))^{1/p}$ analytic near $k_0$. Then the equation above is equivalent to the $p$ equations
$$
(k-k_0) e^{(-ik+c/2)\lambda} = \frac{\omega^l}{G(k)},
$$
for $l=0\ldots p-1$, which we rewrite as
$$
-i\lambda(k-k_0) e^{-i(k-k_0)\lambda} = \frac{-i\lambda\omega^le^{(ik_0-c/2)\lambda}}{G(k)}.
$$
Solutions to these equations are the same as solutions to
$$
k = k_0 - \frac{1}{i\lambda}W_j\left(\frac{-i\lambda\omega^le^{(ik_0-c/2)\lambda}}{G(k)} \right)
$$
for $l=0\ldots p-1$ and $j\in{\mathbb Z}$. Since the argument of $W_j$ is exponentially small, we are in the situation where only $j=0$ results in an approximate resonance. So we have $p$ approximate resonances
$$
k_{0,l} = k_0 - \frac{1}{i\lambda}W_0\left(\frac{-i\lambda\omega^le^{(ik_0-c/2)\lambda}}{G(k_0)} \right)
$$
close to $k_0$.
\bigskip
\noindent{\bf Case 3: $\Im(k_0)=-c/2$ and $0<|f_1(k_0)|<\infty$}\\
In this case we can write \eqref{eqn:R2} as
$$
e^{2i(k-k_0)L} = e^{(c-2ik_0)L}f_1(k)(e^{-cL}-2ik) = e^{-2i\Re(k_0)L}f_1(k)(e^{-cL}-2ik)
$$
and take the $\log$. This results in approximate resonances at
$$
k_j = k_0 + \frac{1}{2iL}\log(e^{-2i\Re(k_0)L}f_1(k_0)(e^{-cL}-2ik_0) + \frac{j\pi}{L}.
$$
\bigskip
\noindent{\bf Case 4: $\Im(k_0)=-c/2$ and $f_1(k_0)=0$}\\
\medskip
This is like Case 2 except that $i(k_0-c/2)=i\Re(k_0)$ is purely imaginary so we no longer have exponential decay in $\lambda$ in the argument of the Lambert $W$ function. Instead we have linear growth which means that all the $W_j$'s will contribute approximate resonances near $k_0$. Thus, the approximate resonances are
$$
k_{j,l} = k_0 - \frac{1}{i\lambda}W_j\left(\frac{-i\lambda\omega^le^{i\Re(k_0)\lambda}}{G(k_0)} \right)
$$
\bigskip
\noindent{\bf Case 5: $\Im(k_0)=-c/2$ and $k_0$ is a pole of $f_1$}\\
\medskip
Cases 5 and 6 are related analogously to Cases 2 and 4. When $\Im(k_0)=-c/2$ we will obtain approximate resonances
$$
k_{j,l} = k_0 + \frac{1}{i\lambda}W_j\left(i\lambda\omega^le^{-i\Re(k_0)\lambda}G(k_0) \right)
$$
where $f_1(k)=(k-k_0)^{-p}g(k)$ and $G(k)$ and $\omega$ are defined as before.
\bigskip
\noindent{\bf Case 6: $\Im(k_0)<-c/2$ and $k_0$ is a pole of $f_1$}\\
\medskip
These are the reflectionless points for $V_1$ below the line $\Im(k)=-c/2$. Close to these points we will get $p$ approximate resonances
$$
k_{0,l} = k_0 + \frac{1}{i\lambda}W_0\left(i\lambda\omega^le^{(c/2-ik_0)\lambda}G(k_0) \right)
$$
{We now replace $e^{-cL}\delta(x-L)$ with $\mu V_2(x-L)$. We will take $\mu = \mu(L)$ with $\lim_{L \to \infty} \mu(L) = 0$. We take $V_2$ as before with compact support in $(0,x_1)$. From the Appendix in Section 9, The function $f_2$ for small coupling, we find}{
\begin{equation}\label{eqn:muV2}
\mu f_2(k,\mu) = -2ik/I(k) + O(\mu).
\end{equation}}
where $I(k) = \int _{\mathbb R} V_2(x) e^{2ikx} dx$. The $O(\mu)$ term is uniform on compact subsets of $k$ which do not contain $k=0$.
{Here we have indicated explicitly that $f_2$ depends on $\mu$ as well as $k$. This expression should be compared to the expression for $f_2$ obtained above when $\mu V_2(x-L)$ is replaced by $e^{-cL}\delta(x-L)$, namely
$$e^{-cL}f_2(k) = -2ik + e^{-cL}$$}
{With $\mu V_2$ instead of $e^{-cL}\delta$, (\ref{eqn:R2}) becomes
$$\mu e^{2ikL} = f_1(k)\mu f_2(k,\mu)$$}
{It is clear from (\ref{eqn:muV2}) that if $\mu(L) \to 0 $ faster than any exponential, then for any compact set $K \subset \mathbb C \setminus\{0\}$, all resonances of $H_L = H + \mu(L) V_2(x-L)$ either leave $K$ or converge to resonances of $H$ in $K$. In fact if $k_0$ is a resonance of $H$ (thus a zero of $f_1$), there is a resonance of $H_L$ converging to $k_0$.}
{Consider now a compact subset $K$ of $ \{k \ne 0: I(k) \ne 0, \text{Im} k < -a < 0\} $. (We assume $ I(k) \ne 0$ and $k \ne 0$ for simplicity.) Then if $\mu(L) \to 0$ slower than any exponential, the resonances of $H_L$ either leave $K$ or converge to a reflectionless point of $H$. Resonances of $H_L$ converge to the real axis in a similar way as is discussed in the case that $\mu(L) = 1.$ (For example if $f_1(k)$ is analytic at $k_0 \in \mathbb R$ the resonances near $k_0$ are uniformly displaced by approximately $-(2iL)^{-1}\log \mu(L)$ but the separation between them is still approximately $\pi/L$. Again we assume $I(k_0) \ne 0$ and $k_0 \ne 0$ .)}\\
{If $\mu(L) = e^{-cL}$, the analysis is very similar to that given earlier in this Section but instead of equation (\ref{eqn:R2}) we have
$$e^{(2ik - c)L} = [f_1(k) (-2ik)/I(k)](1 + O(e^{-cL}).$$
We have assumed we are in a neighborhood of a point $k_0 \ne 0$ where $I(k_0) \ne 0$. The basic phenomenology is the same as with $e^{-cL}\delta(x-L)$. We omit the proofs.}
\section{Resonances as $L\rightarrow\infty$ when $V_{2,L}(x) = V_2(x-L)$}\label{no coupling constant}
In this section we determine the asymptotic positions of the resonances of $H_L=-d^2/dx^2 + V_1(x) + V_2(x-L)$ for large $L$.
Let ${\cal R}(L)$ denote the set of solutions to \eqref{eqn:R} where $f(k)$ is the function given by \eqref{eqn:f} and let ${\mathbb C}_-=\{k\in{\mathbb C}:\Im(k)<0\}$. Then ${\cal R}(L)\cap {\mathbb C}_-$ is the set of resonances. However, {generically,} ${\cal R}(L)$ contains $k=0$ even if the \textrm Green's function\ has no pole there.
We will locate the asymptotic positions of the points in ${\cal R}(L)$. A consequence of our estimates is:
\begin{theorem} \label{thm:counting}
Let $N(L)=\#\Big({\cal R}(L)\cap \big([a,b]\times [-ic,0]\big)\Big)$ for $0\le a<b<\infty$ and $c>0$. We then have
$$\lim _{L \to \infty} N(L)/L = (b-a)/\pi$$
If $[a,b]$ contains no poles of $f$ then there is a constant $c_1> 0$ so that the number of resonances in $[a,b] \times [ -ic_2,-ic_1 L^{-1} ]$ is bounded uniformly in $L$ as $L \to \infty$ for any $c_2 > 0$. If $[a,b]$ contains a pole of $f$ there is a $c'_1> 0$ so that the number of resonances in $[a,b] \times [ -ic_2, -ic'_1 L^{-1}\log L]$ is bounded uniformly in $L$ as $L \to \infty$ for any $c_2 > 0$. If $[a,b]$ contains no poles of $f$ of order higher than $p$ then we can take $c'_1 = p/2$.
\end{theorem}
To begin, we show that the sets
$$
U(a,A) = \{k\in{\mathbb C}_- : \Im(k) < -a \hbox{\ and\ }|f(k)| < A\}
$$
for $a,A>0$ are resonance free for $L$ sufficiently large. Recall that reflectionless points are poles of $f$.
So the sets $U(a,A)$ exclude a strip below the real axis together with some neighbourhoods of the reflectionless points.
\begin{proposition}\label{leaving}
There exists $L_0$ such that $U(a,A)\cap{\cal R}(L) = \emptyset$ when $L>L_0$.
\end{proposition}
\begin{proof} Comparing the moduli of the left and right sides of \eqref{eqn:R},
we find that any resonance in $U(a,A)$ satisfies
\[
e^{2La}<|e^{2iLk}| = |f(k)| < A,
\]
so that $L<\log(A)/(2a)$. This proves the proposition with $L_0=\log(A)/(2a)$.
\end{proof}
We will use the following notation. The disk centered at $k_0$ with radius $r$ is
$$B(k_0,r)=\{k\in{\mathbb C}: |k-k_0|<r\}.$$
For $f$ analytic near $k_0$ with $f(k_0)\ne 0$
$$
\Omega(f,k_0,\epsilon)
=\left\{k\in{\mathbb C} : \left|1-\dfrac{f(k)}{f(k_0)}\right|<\epsilon\right\}.
$$
Now let us find all the resonances in a neighbourhood of $k_0$ in the closed {lower} half plane $\overline{{\mathbb C}}_-$ for $L$ large. Given proposition \ref{leaving} we need only consider $k_0\in{\mathbb R}$ and $k_0$ a pole of $f(k)$. We will distinguish three cases.
\bigskip
\noindent{\bf Case 1: $k_0\in{\mathbb R}$ and $f$ is analytic at $k_0$}\\
Let $\log$ be any branch of the logarithm {and $r>0$ such that $f(k) \ne 0$ for $k \in B(k_0,r)$. }Then
$$
{\cal R}(L) {\cap B(k_0,r)} =\left\{k { \in B(k_0,r)}:k=\dfrac{1}{2iL}\big(\log(f(k)) + 2\pi ij\big) \hbox{\ for some\ } j\in\mathbb Z\right\}.
$$
We call $k_j$ an approximate resonance for $L$ near $k_0$ if
$$
k_j = \dfrac{1}{2iL}\big(\log(f(k_0)) + 2\pi ij\big)
$$
for some $j\in{\mathbb Z}$, and denote the set of these by ${\cal A}(k_0,L)$. Notice that $f(k_0)\ne 0$, since $k_0\in{\mathbb R}$ implies $|f(k_0)|\ge 1$.
\begin{proposition} \label{prop:nopole}
Suppose $f(k)$ is analytic near $k_0\in{\mathbb R}$. Let $\epsilon \in(0,1/2)$. Then there exists $\delta>0$ and $L_0$ such that for all $L>L_0$
\begin{enumerate}[label=(\roman*)]
\item If $k_j\in B(k_0,\delta) \cap {\cal A}(k_0,L)$ then there is exactly one $k\in{\cal R}(L)$ with $|k-k_j|<\epsilon/L$.
\item If $k\in B(k_0,\delta) \cap {\cal R}(L)$ then there is a $k_j\in {\cal A}(k_0,L)$ with $|k-k_j|<\epsilon/L$.
\end{enumerate}
\end{proposition}
\begin{proof}
Let $\log$ be a branch of the logarithm such that $\log(f(k))$ is analytic for $k\in\Omega(f, k_0,\epsilon)$. This is possible because if $k\in\Omega(f, k_0,\epsilon)$ then $|f(k_0)-f(k)| < (1/2)|f(k_0)|$. Thus $f(k)$ lies in a disk of radius $|f(k_0)|/2$ centred at $f(k_0)$, and we can find a branch cut that misses this disk. {Choose $0 < \delta \le r/2 $ small enough so that $B(k_0, 2\delta) \subset \Omega(f, k_0,\epsilon)$}. For $j\in{\mathbb Z}$, define
$$
\phi_j(k) = k - \dfrac{1}{2iL}\big(\log(f(k) + 2\pi ij\big)
$$
so that {$k \in {\cal R}(L) \cap B(k_0, 2\delta)$} if and only if {$k\in B(k_0,2\delta)$ and} $\phi_j(k)=0$ for some $j\in{\mathbb Z}$.
Let $L_0=2/\delta$ and assume $L > L_0$. If $k_j$ is an approximate resonance for $L$ in $B(k_0,\delta)$ then $B(k_j,\epsilon/L)\subset B(k_0,2\delta)$ so both $\phi_j$ and $\xi_j(k) = k-k_j$ are analytic in a neighbourhood of {$\overline{B(k_j,\epsilon/L)}$}. Clearly $|\xi_j(k)| = \epsilon/L$ for $k$ on the boundary of $B(k_j,\epsilon/L)$.
On the other hand
\begin{align*}
|\xi(k)-\phi_j(k)| &=\dfrac{1}{2L}\left|\log\left(\dfrac{f(k)}{f(k_0)}\right)\right|\\
&=\dfrac{1}{2L}\left|\log\left(1+\dfrac{f(k)}{f(k_0)} -1\right)\right|\\
&\le \dfrac{1}{2L}\dfrac{\left|\dfrac{f(k)}{f(k_0)} -1\right|}{1-\left|\dfrac{f(k)}{f(k_0)} -1\right|}\\
&< \dfrac{1}{2L}\dfrac{\epsilon}{1-\epsilon}\\
&<\epsilon/L.\\
\end{align*}
Here we used that $|\log(1+z)|\le |z|/(1-|z|)$ for $|z|<1$ and that $\epsilon<1/2$. Thus Rouch\'e's theorem implies that $\xi$ and $\phi_j$ have the same number of zeros in $B(k_j,\epsilon/L)$, namely one. This proves (i).
To prove (ii) we suppose that we have a resonance $k\in B(k_0,\delta)\cap{\cal R}(L)$. Then
$$
k = \dfrac{1}{2iL}\big(\log(f(k) + 2\pi ij\big)
$$for some $j\in{\mathbb Z}$. For this $j$ define
$$
k_j= \dfrac{1}{2iL}\big(\log(f(k_0) + 2\pi ij\big).
$$
Then, since $B(k_0,\delta)\subset\Omega(f, k_0,\epsilon)$.
$$
|k-k_j| = \dfrac{1}{2L}\left|\log\left(\dfrac{f(k)}{f(k_0)}\right)\right|<\epsilon/L
$$
as before.
\end{proof}
\bigskip
\setlength{\fboxsep}{10pt}
\fbox{\begin{minipage}{40em}
We will illustrate the approximations with $V_1 = \delta(x+1)+\delta(x)$ and $V_2=\delta(x)+\delta(x-1)$. Then
$$
f(k)=\frac{(-e^{2ik}+1-4ik-4k^2)^2}{(-e^{2ik}+1+2ik(-e^{2ik}-1))^2}.
$$
There are infinitely many poles on the real axis but none in ${\mathbb C}_-$. So all the resonances will be converging to the real axis. We pick a point that is not a pole, say $k_0=3$ (blue dot) and compute the the exact (red circle) and approximate (blue crosses) resonances. Here we have taken $L=30$.
\hskip 20pt
\begin{center}
\includegraphics[width=8cm]{approx1.pdf}
\end{center}
\end{minipage}}
\bigskip\bigskip
\noindent{\bf Case 2: $k_0\in{\mathbb R}$ and $k_0$ is a pole of $f$}.
Write $f(k)=\dfrac{g(k)}{(k-k_0)^p}$ where $p$ is a positive integer, $g$ is analytic at $k_0$ and $g(k_0)\ne 0$. Then { for $k$ near $k_0$ ,}$k\in {\cal R}(L)$ whenever
$$
e^{2iLk}=\dfrac{g(k)}{(k-k_0)^p}
$$
or, equivalently,
\begin{equation}\label{eq:resp}
(k-k_0)^p e^{2iL(k-k_0)}= e^{-2iLk_0}g(k)
\end{equation}
Choose a branch of $G(k)$ of $g^{1/p}(k)$ that is analytic in a neighbourhood of $k_0$. To simplify notation define
$\lambda = 2L/{p}$ and $\omega=e^{2\pi i /p}$. Then \eqref{eq:resp} holds if and only if
\begin{equation}\label{eq:resproot}
i\lambda(k-k_0) e^{i\lambda (k-k_0)}= i\lambda e^{-i\lambda k_0}G(k)\omega^l
\end{equation}
for some $l\in \{0,1,\ldots, p-1\}$.
The solutions $z$ of $ze^z = w$ are (by definition) branches of the Lambert $W$ function. It follows from \eqref{eq:resproot} that $k$ is a resonance for {$H_L$} if and only if
$$
k=k_0 + \frac{1}{i\lambda}W(A(k,k_0,l,\lambda))
$$
for some branch $W$ of the Lambert $W$ function, where
$$
A(k,k_0,l,\lambda) = i\lambda e^{-i\lambda k_0}\omega^lG(k),
$$
We want to mention that recently in [Sa] the Lambert $W$ function was used to calculate shape resonances induced by two delta function barriers.
Let $W_j(z)$ for $j\in{\mathbb Z}$ be the branches of the Lambert $W$ function defined in [CGHJK]. Define approximate resonances
\begin{equation}\label{approxres2}
k_{j,l}=k_0 + \frac{1}{i\lambda}W_j(A(k_0,k_0,l,\lambda))
\end{equation}
for $j\in{\mathbb Z}$ and $l\in \{0,1,\ldots, p-1\}$. We will denote the set of approximate resonances near $k_0$ by ${\cal B}(k_0,\lambda)$.
\begin{proposition}\label{asympt1}
Suppose $f(k)$ has a pole of order $p$ at $k_0\in{\mathbb R}$. Let $\epsilon \in(0,1/2)$. Then there exists $\delta>0$ and $\lambda_0$ such that for all $\lambda>\lambda_0$
\begin{enumerate}[label=(\roman*)]
\item If $k_{j,l}\in B(k_0,\delta) \cap {\cal B}(k_0,\lambda)$ then there is exactly one $k\in{\cal R}(L)$ with $|k-k_{j,l}|<\epsilon/\lambda$.
\item If $k\in B(k_0,\delta) \cap {\cal R}(L)$ then there is a $k_{j,l}\in {\cal B}(k_0,\lambda)$ with $|k-k_{j,l}|<\epsilon/\lambda$.
\end{enumerate}
\end{proposition}
Before proving this proposition we collect some facts about the branches of the Lambert $W$ function. For the branches $W_j$, the branch points are at $-1/e$ for $j=0$, at $0$ and $-1/e$ for $j=1,-1$ and at $0$ for $|j|>1$. The branch cuts are on the negative real axis.
The branches $W_j$ have expansions of the form
$$
W_j(z) = {\rm Log}_j(z) -\log({\rm Log}_j(z)) + \sum_{k=0}^\infty\sum_{m=1}^\infty c_{k,m}\frac{\log({\rm Log}_j(z))^m}{{\rm Log}_j(z)^{k+m}}
$$
where $\log$ is the principal branch and ${\rm Log}_j(z) = \log(z) + 2\pi ij$. These expansions are convergent for large $|z|$. For $j\ne 0$ the expansions converge near $z=0$ too.
For any branch $W=W(z)$ we have $e^{2|W|} \ge |We^W| = |z|$ so
\begin{equation}\label{eq:Wlb}
|W(z)| \ge (1/2)\log|z|
\end{equation}
The derivative of any branch in its region of analyticity can be found by implicit differentiation. We obtain
\begin{equation}\label{eq:Wderiv}
W'(z) = \frac{W(z)}{(1+W(z))z}
\end{equation}
\begin{lemma}\label{Wbound}
Let $a_0\in\mathbb C$ and $\epsilon\in(0,1/2)$ with $(1-\epsilon/4)|a_0| > e^3$. Let $B$ denote the disk $B(a_0,\epsilon|a_0|/4)$. For any $j\in\mathbb Z$ there is a branch $W_*$ of the Lambert $W$ function analytic in $B$ such that $W_*(a_0)=W_j(a_0)$. For any $a\in B$ we have
\begin{equation}\label{West}
|W_*(a) - W_*(a_0)| < \epsilon.
\end{equation}
\end{lemma}
\begin{proof}
Any $a\in B$ satisfies $|a| \ge (1-\epsilon/4)|a_0| > e^3$. Thus $B$ does not contain the possible the branch points of the $W_j$ (i.e., $0$ and $-1/e$). So the analytic continuation $W_*$ of $W_j$ near $a_0$ will be analytic and single valued in $B$.
For $a\in B$, let $[a_0,a]$ denote the straight line path from $a_0$ to $a$. Then we may estimate
\begin{align*}
|W_*(a) - W_*(a_0)| &= \left|\int_{[a_0,a]}W'_*(z) dz\right|\\
&\le \sup_{z\in B}\left| W'(z)\right ||a-a_0|\\
&\le \sup_{z\in B} \frac{|W_*(z)|}{|1+W_*(z)||z|}\epsilon|a_0|/4\\
&\le \sup_{z\in B}\left(1 + \frac{1}{|W_*(z)|-1}\right)\frac{\epsilon|a_0|/4}{(1-\epsilon/4)|a_0|}\\
&\le \sup_{z\in B}\left(1 + \frac{1}{\log|z|/2-1}\right)\frac{\epsilon}{(4-\epsilon)}\\
\end{align*}
Now we use that $|z|>(1-\epsilon/4)|a_0|>e^3$ which implies $\log|z| > 3$ and $\epsilon<1/2$ to conclude
$$
|W_*(a) - W_*(a_0)| \le \frac{3\epsilon}{4-\epsilon} < \epsilon.
$$
\end{proof}
\begin{proof} (of Proposition \ref{asympt1})
Given $\epsilon\in (0,1/2)$, choose $\delta$ sufficiently small so that $G(k)$ is analytic in $B(k_0,2\delta)$ and $B(k_0,2\delta)\subset \Omega(G,k_0,\epsilon/4)$. Let $\lambda_0 = \max\{e^3/((1-\epsilon/4)|G(k_0)|), \epsilon/\delta\}$.
Assume that $k_{j,l}$ is an approximate resonance given by \eqref{approxres2} for some $j\in\mathbb Z$ and $l\in\{0,\ldots,p-1\}$ and that $k_{j,l}\in B(k_0,\delta)$. We must show that there exists a resonance in $k\in{\cal R}(L)$ where $L=p\lambda/2$ with $|k-k_{j,l}|<\epsilon/\lambda$ whenever $\lambda > \lambda_0$.
To simplify notation, denote $A(k,k_0,l,\lambda)$ by $A(k)$. Note that $k\in \Omega(G,k_0,\epsilon/4)$ implies that $A(k)\in B(A(k_0),\epsilon |A(k_0)|/4)$. In addition, $|A(k_0)|=\lambda|G(k_0)|$ so that $\lambda > \lambda_0$ implies $(1-\epsilon/4)|A(k_0)| > e^3$. Thus we may apply Lemma \ref{Wbound} with $a_0=A(k_0)$ to conclude that
$$
|W_*(A(k))-W_*(A(k_0))| < \epsilon
$$
for $k\in B(k_0,2\delta)$ and $\lambda > \lambda_0$.
Now define $\xi_{j,l}(k)=k-k_{j,l}$ and $\phi_{j,l}(k)=k-k_0-(i\lambda)^{-1}W_*(A(k))$. We wish to apply Rouch\'e's theorem on $B(k_{j,l},\epsilon/\lambda)$.
Since $k_{j,l}\in B(k_0,\delta)$ and $\epsilon/\lambda < \delta$ for $\lambda > \lambda_0$ it follows that for $\lambda > \lambda_0$, $B(k_{j,l},\epsilon/\lambda)\subseteq B(k_0,2\delta)$. On the boundary of $B(k_{j,l},\epsilon/\lambda)$ we have $|\xi_{j,l}(k)|=\epsilon/\lambda$. On the other hand, for all $k\in B(k_0,2\delta)$ we also have
\begin{align*}
|\xi_{j,l}(k)-\phi_{j,l}(k)|
&= \frac{1}{\lambda}\left|W_*(A(k)) -W_j(A(k_0))\right| \\
&= \frac{1}{\lambda}\left|W_*(A(k)) -W_*(A(k_0))\right| \\
&< \epsilon/\lambda,\\
\end{align*}
provided $\lambda > \lambda_0$. Then Rouch\'e's theorem says that there is exactly one solution of $\phi_{j,l}(k)=0$, that is, exactly one resonance, in
$B(k_{j,l},\epsilon/\lambda)$. This proves (i).
To prove (ii) we note that every resonance can be written $k = k_0 + (i\lambda)^{-1}W_j(A(k))$ for some some $j\in\mathbb Z$ and $l\in\{0,\ldots,p-1\}$.
If we know that $k\in B(k_0,\delta)$, with $\delta$ chosen as above, then for $\lambda > \lambda_0$ we will have $|W_*(A(k))-W_*(A(k_0))| < \epsilon$ as before. This implies that if we let $k_{j,l}$ be the approximate resonance $k_{j,l}=k_0 + (i\lambda)^{-1}W_j(A(k_0))$, then $|k-k_{j,l}| < \epsilon/\lambda$.
\end{proof}
\bigskip
\setlength{\fboxsep}{10pt}
\fbox{\begin{minipage}{40em}
Here are the computed resonances (red circles) and approximate resonances ($l=0$ blue and $l=1$ green) near the double pole $k_0\sim 4.81584231784594$.
\begin{center}
\includegraphics[width=10cm]{approx2.pdf}
\end{center}
\end{minipage}}
\bigskip\bigskip
We can make a further approximation of the approximate resonances that does not involve the Lambert $W$ function. Using the first two terms of the expansion for $W_j$ gives
$$
k_{j,l} \sim k_0 + \frac{1}{i\lambda}\Big(\log(A_0) + 2\pi i j + \log(\log(A_0) + 2\pi i j)\Big)
$$
to an accuracy of $O\left(\dfrac{\log\log(\lambda)}{\lambda\log(\lambda)}\right)$. Here $A_0=A(k_0,k_0,l,\lambda)=i\lambda e^{-i\lambda k_0}\omega^lG(k_0)$ as before.
For $x\in{\mathbb R}$, let $[x]$ denote the integer contained in $(x-1/2,x+1/2]$ define $\Phi(k_0,l,\lambda)\in(-1/2,1/2]$ as
$$
\Phi(k_0,l,\lambda)=\frac{-\lambda k_0 + \varphi_0}{2\pi} + \frac{l}{p} + \frac{1}{4} - \left[\frac{-\lambda k_0 + \varphi_0}{2\pi} + \frac{l}{p} + \frac{1}{4} \right].
$$
where $\varphi_0=\arg(G(k_0))$. Then we have
$$
A(k_0,k_0,l,\lambda)=\lambda|G(k_0)|e^{2\pi i\Phi(k_0,l,\lambda)}
$$
and
$$
\log(A_0) + 2\pi i j = \log(\lambda |G(k_0)|) +2\pi i(j+\Phi(k_0,l,\lambda))
$$
so that
\begin{align*}
\Re(k_{j,l}) &\sim k_0
+ \frac{2\pi}{\lambda}(j+\Phi(k_0,l,\lambda))
-\frac{1}{\lambda} \arctan\left(\frac{2\pi (j+\Phi(k_0,l,\lambda))}{\log(\lambda |G(k_0)|)}\right)\\
\end{align*}
and
\begin{align*}
\Im(k_{j,l}) &\sim
-\frac{\log(\lambda |G(k_0)|)}{\lambda}
+\frac{1}{2\lambda}\log\left( \log(\lambda |G(k_0)|)^2 + 4\pi^2(j+\Phi(k_0,l,\lambda))^2 \right)
\end{align*}
So we get that the real parts have a spacing of $\sim 2\pi/\lambda=p\pi/L$, i.e., $p$ times wider than in case 1. But there are $p$ sequences corresponding to $l=0,\ldots,p-1$ so the counting remains the same. The imaginary parts are most negative when $j=0$, which corresponds to real parts close to $k_0$.
\bigskip
\setlength{\fboxsep}{10pt}
\fbox{\begin{minipage}{40em}
Here the approximate resonances (diamonds) are compared to the first two terms in their expansions (crosses).
\vskip -2.5cm\
\begin{center}
\includegraphics[width=9cm]{approx3.pdf}
\end{center}
\end{minipage}}
\bigskip\bigskip
\noindent{\bf Case 3: $\Im k_0 < 0$ and $k_0$ is a pole of $f$}.
\bigskip
As in Case 2, {a point $k$ near $k_0$ } is a resonance if and only if it satisfies
$$
k=k_0 + \frac{1}{i\lambda}W(A(k,k_0,l,\lambda))
$$
for some branch $W$ of the Lambert $W$ function, with
$$
A(k,k_0,l,\lambda) = i\lambda e^{-i\lambda k_0}\omega^lG(k),
$$
where as before, $G(k)$ is a branch of $g(k)^{1/p}$ and $\omega=e^{2\pi i/p}$.
We may define approximate resonances as before as
$$
k_{j,l} = k_0 + \frac{1}{i\lambda}W(A(k_0,k_0,l,\lambda)).
$$
The difference is that now $|A(k,k_0,l,\lambda)|=\lambda e^{-|\Im k_0|\lambda}|G(k)|$ is exponentially small in $\lambda$. This means that we must consider the behaviour of branches $W_j(z)$ of the Lambert W function near $z=0$. The principal branch $W_0(z)$ is analytic in a disk of radius $1/e$ about $z=0$ with leading behaviour $W_0(z) = z + O(z^2)$.
We will use that there is a constant $C$ such that for $|z|<1/10$,
\begin{equation}\label{w0taylor}
|W_0(z)-z| \le C |z|^2.
\end{equation}
The other branches $W_j(z)$ for $j\ne 0$ all behave like $\log(z)$ and are given by the convergent expansions above. The differing behaviours result in only the $p$ approximate resonances with $j=0$ being associated to resonances near $k_0$.
Let ${\cal C}(k_0,\lambda) = \{k_{0,l}, l=0\ldots p-1\}$ denote these approximate resonances.
\begin{proposition}
Suppose $f(k)$ has a pole of order $p$ at $k_0\in{\mathbb C}_-$. Let $\epsilon \in(0,1/2)$. Then there exists $\delta>0$ and $\lambda_0$ such that for all $\lambda>\lambda_0$
\begin{enumerate}[label=(\roman*)]
\item If $k_{0,l}\in B(k_0,\delta)\cap {\cal C}(k_0,\lambda)$ then there is a $k \in {\cal R}(L)$ with $|k-k_{0,l}|<\epsilon e^{-|\Im(k_0)|\lambda}$.
\item If $k\in B(k_0,\delta)\cap {\cal R}(L)$ then there is a $k_{0,l}\in {\cal C}(k_0,\lambda)$ with $|k-k_{0,l}|<\epsilon e^{-|\Im(k_0)|\lambda}$.
\end{enumerate}
\end{proposition}
\begin{proof}
Choose $\delta$ small enough to ensure that $G(k)$ is analytic and $|G(k)-G(k_0)|\le \epsilon/2$ for $k\in B(k_0,2\delta)$ and $\delta < |\text {Im} k_0|$. Choose $\lambda_0$ sufficiently large so that $\lambda>\lambda_0$ implies
$$
\lambda e^{-|\Im k_0|\lambda}(|G(k_0)|+1)<\frac{1}{10},\quad
e^{-|\Im k_0|\lambda}<\delta,\quad
\lambda e^{-|\Im k_0|\lambda}2C(|G(k_0)|+1)^2<\epsilon/2,
$$
where $C$ is the constant in \eqref{w0taylor}.
Then for $k\in B(k_0,2\delta)$ and $\lambda>\lambda_0$,
$$
|A(k,k_0,l,\lambda)|=\lambda e^{-|\Im k_0|\lambda}|G(k)|\le \lambda e^{-|\Im k_0|\lambda}(|G(k_0)|+1)<\frac{1}{10}
$$
This implies that $A(k,k_0,l,\lambda)$ lies in the region of analyticity for $W_0$ and that $|W_0(A)-A| \le C |A|^2$ for both $A=A(k,k_0,l,\lambda)$ and $ A=A(k_0,k_0,l,\lambda)$.
Now assume that we are given $k_{0,l}\in{\cal C}(k_0,\lambda)\cap B(k_0,\delta)$. We will apply Rouch\'e's theorem to
$$\phi(k) = k - k_0 - \frac{1}{i\lambda}W_0(A(k,k_0,l,\lambda))$$
and
$$\xi(k) =k-k_{0,l} = k - k_0 - \frac{1}{i\lambda}W_0(A(k_0,k_0,l,\lambda)).$$
on $B(k_{0,l},\epsilon e^{-|\Im(k_0)|\lambda})$.
Clearly $\xi(k)$ is analytic on this ball and
$|\xi(k)| =\epsilon e^{-|\Im(k_0)|\lambda}$ on the boundary. On the other hand, given that $k_{0,l}\in B(k_0,\delta)$ we find that points $k\in B(k_{0,l},\epsilon e^{-|\Im(k_0)|\lambda})$ obey
$$
|k-k_0| \le |k-k_{0,l}| + |k_{0,l}-k_0|\le \epsilon e^{-|\Im(k_0)|\lambda} + \delta <2\delta
$$
provided $\lambda>\lambda_0$.
So $k\in B(k_0,2\delta)$. This implies that $\phi(k)$ is analytic in neighbourhood of $B(k_{0,l},\epsilon e^{-|\Im(k_0)|\lambda})$.
Let $A(k)$ denote $A(k,k_0,l,\lambda)$. For $k\in B(k_{0,l},\epsilon e^{-|\Im(k_0)|\lambda})$ and $\lambda>\lambda_0$ we have
\begin{align*}
|\phi(k)-\xi(k)| &= \frac{1}{\lambda} |W_0(A(k))-W_0(A(k_0))|\\
&\le\frac{1}{\lambda}\left( |A(k)-A(k_0)| + C\left(|A(k)|^2+|A(k_0)|^2\right)\right)\\
& \le \frac{1}{\lambda} \Big(\lambda e^{-|\Im(k_0)|\lambda}|G(k)-G(k_0)|+ 2C(|G(k_0)|+1)^2\lambda^2 e^{-2|\Im(k_0)|\lambda}\Big)\\
&\le \left( \frac{\epsilon}{2} +2C(|G(k_0)|+1)^2 \lambda e^{-|\Im(k_0)|\lambda}|\right)e^{-|\Im(k_0)|\lambda}\\
&<\epsilon e^{-|\Im(k_0)|\lambda}\\
\end{align*}
Now (i) follows from Rouch\'e's theorem.
Now suppose that $k$ is a resonance. Then
$
k=k_0 + \frac{1}{i\lambda}W_j(A(k,k_0,l,\lambda))
$
for some $j,l$. When $j=0$ our previous estimate shows
$$
|k-k_{0,j}| \le \frac{1}{\lambda} |W_0(A(k))-W_0(A(k_0))|<\epsilon e^{-|\Im(k_0)|\lambda}
$$
provided $k\in B(k_0,\delta)$ and $\lambda>\lambda_0$. So (ii) holds in this case.
When $j\ne 0$ the values of $W_j$ are bounded away from zero and we have
$$
e^{\Re(W_j(z))}=\frac{|z|}{|W_j(z)|} \le C |z|
$$
Since $|A(k,k_0,l,\lambda)| < C(k_0) \lambda e^{-|\Im k_0|\lambda}|$ we conclude
$$
\Re(W_j(A(k,k_0,l,\lambda)))<-|\Im k_0|\lambda+\ln(C(k_0)\lambda).
$$
This implies that
$$
|k-k_0|\ge\frac{1}{\lambda}|\Re(W_j(A(k,k_0,l,\lambda)))|\ge |\Im k_0|
-\ln(C(k_0)\lambda)/\lambda > \delta
$$
for $\lambda$ large enough. Thus (ii) holds vacuously.
\end{proof}
\begin{proof} \textbf{of Theorem \ref{thm:counting}}
The counting part of the proof follows from the formulas for the approximate resonances in Case 1) and Case 2) and from Propositions \ref{prop:nopole} and \ref{asympt1}. Much of what we show below follows from the estimates above but we give a simple self-contained proof. Suppose $f$ is analytic at the point $k_0 \in \mathbb{R}$. Then if $\text{Im} k < -c_1/L$ and $k$ is a resonance near $k_0$ we have
$$|f(k)| = e^{-2\text{Im} k L} > e^{2c_1}$$ which is false for $c_1$ large enough and $k$ in a fixed $L$ - independent neighborhood of $k_0$. If $k_0$ is a pole of $f$ of order $p$ then $|f(k)| \le c|k-k_0|^{-p}$ in a neighborhood of $k_0$. If $k$ is a resonance in this neighborhood with $\text{Im} k < - (p/2)L^{-1} \log L$ then
$$ |f(k)| = e^{-2\text{Im} k L} > e^{p\log L} = L^p .$$ But $$ |f(k)| \le c/|k-k_0|^p < c(2p^{-1})^p (L/\log L)^p,$$ a contradiction for large $L$.
\end{proof}
\section{Resonances and long lived states}
What is the physical meaning of the resonances that are crowding together and approaching a horizonal line? It is not completely clear to us, but we offer the following discussion.
If $k_0$ is a resonance with resonant energy $k_0^2=\lambda_0-i\delta_0$, $\delta_0 > 0$, and $\delta_0$ is small, then $k_0$ should correspond to a long lived resonant state. What is meant by this is open to discussion. A possible definition is a normalized state $\varphi$ satisfying
\begin{equation}\label{resstate}
\sup_{|t|\le T}\left |\left\langle \varphi, e^{-itH}\varphi\right\rangle - e^{-it\lambda-|t|\delta}\right| \le \epsilon
\end{equation}
with possible conditions on $T$, $\epsilon$ and the support (or variance) of $\varphi$ { in $x$ - space}.
Lavine [L] showed that if $k_0=\sigma_0-i\epsilon_0$ is a resonance and $\varphi$ is the (exponentially growing) outgoing solution of \eqref{eqn:SSE}, cut off so its support matches that of the potential, and normalized, then \eqref{resstate} holds with $T=\infty$ and with $\epsilon=O(\epsilon_0/\sigma_0)\log(\sigma_0/\epsilon_0)$ for small $\epsilon_0/\sigma_0$ with explicit constants. Lavine proved the result in the context of half line scattering. A version for the whole line that applies the Hamiltonians we consider is given in Appendix 2.
Lavine's result gives us a resonant state for each of the resonances in our examples that are crowding {onto} the real axis. But since the supports of the potentials in our examples have width $L$, the resulting resonant states, which have the same support, are not well localized. Non-localized states satisfying \eqref{resstate} can exist even when $H$ has potential $V=0$ and no resonances at all, as was pointed out by Lavine [L]. We need only find $\varphi$ such that the spectral measure $d\mu_\varphi$ is close to the Lorenzian $d\mu_L$ corresponding to $\lambda_0-i\delta_0=k_0^2$. We can't do this exactly because the spectral measure $d\mu_\varphi$ is supported in $[0,\infty)$ and $d\mu_L$ has non-zero density everywhere on ${\mathbb R}$. But by making $\lambda_0$ large, we can make $\epsilon$ as small as we like. But the resulting $\varphi$ is very spread out. {Thus in search of a physical meaning for these resonances which are densely crowding on the real axis one might try to superpose them to obtain states which shuttle back and forth between the supports of $V_1$ and $V_{2,L}$ a few times before eventually the whole wave function is transmitted through the potentials.}
Does an estimate like \eqref{resstate} ever signal the existence of a resonance or eigenvalue?
If $\epsilon=0$ and $T=\infty$ then the answer is yes. In this case the spectral measure $d\mu_\varphi$ satisfies $\int_{\mathbb R} e^{its}d\mu_\varphi(s) = \left\langle \varphi, e^{-itH}\varphi\right\rangle = e^{-it\lambda_0-|t|\delta_0}$ which implies that $d\mu_\varphi = d\mu_L$, where $d\mu_L$ is the Lorenzian with parameters $\lambda_0$ and $\delta_0$.
Since the spectral measure is bounded below, this is impossible unless $\delta_0=0$. In this limiting case $d\mu_L$
is a delta function supported on $\{\lambda_0\}$. Thus $\varphi$ is an eigenfunction of $H$ with eigenvalue $\lambda_0$.
If $\epsilon>0$, then an estimate like \eqref{resstate} even with a localized $\varphi$ need not signal the existence of resonances near $k_0$. Consider $H=-d^2/dx^2 + V_1$ for a potential $V_1$ supported in $[-1,1]$. Lavine's theorem gives us a resonant state $\varphi$ also supported in $[-1,1]$ satisfying \eqref{resstate} with $T=\infty$ and $\epsilon$ depending only on $k_0$. The estimate \eqref{resstate} will continue to hold, with slightly adjusted constants if we mollify $\varphi$ to put it in $\cal S$. But then, Theorem \ref{feelingV_L-2} implies that \eqref{resstate} also continues to hold when $H$ is replaced with $H_L=H+V_2(x-L)$, although we must change $T$ from $\infty$ to $cL^{1-\delta}$. As $L$ increases this resonant state lives for an increasingly long time without becoming delocalized. At the same time, the resonances of $H_L$ are moving away from $k_0$.
\section{Appendix: The function $f_2$ for small coupling}
When $H$ is a Hamiltonian of the form $-d^2/dx^2 + V_1(x) + \mu V_{2}(x-L)$ the function $f_2(k,\mu)$ defined in (\ref{eqn:f}) will depend on $\mu$. In this appendix we establish the small $\mu$ behaviour of this function. Since $V_1$ is not involved in the definition of $f_2$ we drop the subscript on $V_2$.
\begin{proposition}\label{f2smallmu}
Let $V(x) = V_{0}(x) + \sum_{i=1}^N\alpha_i\delta(x-x_i)$ where $0=x_1<x_2<\cdots<x_N=x_f$, $V_{0}(x)$ is continuous and $\mathop{\rm supp}(V_0)\subseteq [0,x_f]$. Suppose that $\widehat V(-2ik) \ne 0$ {where $\widehat V (k) = \int V(x) e^{-ikx} dx$}.
Then the function $f_2(k,\mu)$ defined in (\ref{eqn:f}) sastisfies
\begin{align*}
\mu f_2(k,\mu ) &=\frac{ -2ik}{\widehat V(-2ik) } +
\mu\left(\frac{\widehat V(0)}{\widehat V(-2ik)}\right.\\
&+\left.\frac{1}{2ik\widehat V(-2ik)^2}\int_0^{x_f}\int_{t_1}^{x_f}V(t_1)V(t_2)(-e^{ikt_1})\sin((t_1-t_2)k)e^{ikt_2}dt_1dt_2\right) + O(\mu^2)
\end{align*}
where the $O(\mu^2)$ term is uniform for $k$ in a bounded set bounded away from zero.
\end{proposition}
To establish this we let $\psi_2(x,k,\mu)$ be the solution of $-\psi'' +\mu V(x)\psi = k^2\psi$ with $\psi(x)=e^{ikx}$ for $x\ge x_N$. We must compute the expression in (\ref{eqn:f}).
The vector $\Psi = \twomat{\psi(x)}{\psi'(x)}$ satisfies the first order system
$$
\Psi' =\twomat{0&1}{V_2(x) - k^2&0} \Psi
$$
away from the $x_i$. At the points $x_i$ we have
\be\label{deltajump}
\Psi(x_i+) = \twomat{1&0}{\alpha_i&1}\Psi(x_i-).
\end{equation}
We now write
$$
\psi_2(x,k,\mu) = a(x,k,\mu)e^{ikx} + b(x,k,\mu)e^{-ikx}
$$
where
$$
a'(x,k,\mu)e^{ikx} + b'(x,k,\mu)e^{-ikx} = 0.
$$
and define
$$X(x)=X(x,k,\mu)=\twomat{ a(x,k,\mu)}{b(x,k,\mu)}.$$
Then
$$
\Psi(x) = \twomat{e^{ikx} &e^{-ikx} }{ike^{ikx} &-ike^{-ikx}}X(x).
$$
and a short calculation shows that away from the points $x_i$, $X$
satisfies the first order system
$$
X'(x) = \frac{V(x)}{2ik}A(x,k)X(x).
$$
where
$$
A(x,k) =\twomat{1&e^{-2ikx}}{-e^{2ikx}&-1}
$$
The change in $X$ as $x$ passes a point $x_i$ is given by
\be\label{onestep}
X(x_i-) = \left(I - \frac{\alpha_i}{2ik}A(x_i,k)\right)X(x_i+).
\end{equation}
The solution $X$ that satisfies the final condition $X_f=X(x_f)=\twomat{1}{0}$ solves the integral equation
\begin{align}\label{inteq}
X(x) = X_f - \int_{x}^{x_{f}}\frac{\mu V(t)}{2ik}A(t,k)X(t)dt.
\end{align}
provided that for a piecewise continuous vector function $Y(x)$ that may have jump discontinuities at the points $x_i$, we interpret
$$
\int_{x_i-}^{x_i+} \delta(x-x_i) Y(x)dx = Y(x_i+).
$$
The solution $X(x)=\twomat{ a(x,k,\mu)}{b(x,k,\mu)}$ to \eqref{inteq} then determines $f_2$ via
$$
f_2(k,\mu) = -\frac{a(0-,k,\mu)}{b(0-,k,\mu)}.
$$
We will need an a priori bound on $X(x)$. Starting with $\|X(x_N+)\|=\|X_f\| = 1$ we use \eqref{onestep} to estimate
\begin{align*}
\|X(x_N-)\| &\le \left(1+\frac{\mu |\alpha_N| M(k)}{2|k|}\right)\|X(x_N+)\| \\
&\le \exp\left({\frac{\mu |\alpha_N| M(k)}{2|k|}}\right)
\end{align*}
where $M(k) = \sup_{[x\in[0,x_f]}\|A(k,x)\| = 2\cosh(x_f\Im(k))$.
Then, using Gronwall's inequality, it is not hard to see that
$$
\|X(x_{N-1}+)\|\le \exp\left({\frac{\mu \|V_0\|_\infty(x_{N}-x_{N-1}) M(k)}{2|k|}}\right)\|X(x_N-)\|.
$$
Continuing like this, we arrive at the bound
\be\label{apriori}
\sup_{x\in[0,x_f]}\|X(x)\| \le \exp\left({\frac{\mu M(k) (\|V_0\|_\infty x_f+ \sum|\alpha_j| )}{2|k|}}\right) = C_1(M(k)/k).
\end{equation}
To solve \eqref{inteq} we iterate the equation to obtain
\begin{align*}
X(x) &= X_f \\
&+\sum_{j=1}^{m}(-1)^j \int_{t_{j-1}}^{x_{f}}\cdots \int_{t_1}^{x_{f}}\int_{x}^{x_{f}}\frac{\mu^j V(t_1)\cdots V(t_{m-1})}{(2ik)^j}A(t_1,k)\cdots
A(t_j,k)X_fdt_1\cdots dt_j+ E_m,\\
\end{align*}
where
$$
E_m=(-1)^{m+1}\int_{t_{m}}^{x_{f}}\cdots \int_{t_1}^{x_{f}} \int_{x}^{x_{f}}\frac{\mu^{m+1}V(t_1)\cdots V(t_{m+1})}{(2ik)^{m+1}}A(t_1,k)\cdots
A(t_m,k)X(t_{m+1})dt_1\cdots dt_{m+1}.
$$
The matrix $A(x,k)$ has rank one and can be written $A(x,k)=a(x,k)b(x,k)^T$ where $a(x,k)=\twomat{e^{-ikx}}{-e^{ikx}}$ and $b(x,k)=\twomat{e^{ikx}}{e^{-ikx}}$. Using that
$b(t_i,k)^Ta(t_{i+1},k)=2i\sin((t_i-t_{i+1})k)$ we can write
$$
\frac{1}{(2ik)^{m+1}}A(t_1,k)\cdots
A(t_m,k) =
\frac{1}{(2ik)}a(t_1,k)\frac{\sin((t_1-t_{2})k)}{k}\cdots\frac{\sin((t_{m-1}-t_{m})k)}{k}
b(t_m,k)^T.
$$
so that
$$
\left|\frac{1}{(2ik)^{m+1}}A(t_1,k)\cdots A(t_m,k)\right|
\le \frac{1}{2|k|}C_2(k)^m
$$
where $C_2(k)$ is bounded for $|k|$ in a bounded set.
This bound together with $\|a(x,k)\|\|b(x,k)\|\le M(k)$ leads to
\begin{align*}
\|E_m\| &\le \frac{\mu^{m+1}}{2|k|} \int_{t_{m}}^{x_{f}}\cdots \int_{0}^{x_{f}}|V(t_1)|\cdots |V(t_{m+1})| C_2(k)^m M(k) C_1(M(k)/|k|)dt_1\cdots dt_{m+1}\\
&\le \frac{\mu^{m+1}}{2|k|(m+1)!}\|V\|_1^{m+1} C_2(k)^m C_1(M(k)/|k|)
\end{align*}
{where $\|V\|_1 = \|V_0\|_1 + \sum_j |\alpha_j|$}. This shows that the series for $X$ converges for all $\mu$.
Proposition \ref{f2smallmu} follows from computing the first few terms in this series.
\section{Introduction}
A compactly supported potential $V(x)$ in one dimension with high barriers will give rise to shape resonances close to the real axis. These resonance energies $z_j$ and their associated resonant states $\phi_j$ determine the time evolution under $H=-d^2/dx^2 + V$ of an initial state $\phi$ whose energy and position are suitably localized near $\Re z_i$ and $\mathop{\rm supp}(V)$.
Roughly speaking, for $t\ge 0$ we expect such a $\phi$ to obey
\[
\langle \phi, e^{-itH}\phi \rangle \sim \sum_{j=1}^n |\langle \phi, \phi_j \rangle|^2 e^{-itz_j}.
\]
What happens when a Dirichlet condition is imposed at $x=L$?
In a recent paper, Herbst and Mavi [HM] answered this question when $V$ is a sum of two delta functions. Call the Hamiltonian with the additional Dirichlet condition $H_L$. It seems reasonable to expect that for a long time (roughly the time it takes $\phi$ to travel a distance $L$ under the free time evolution) the delta function should have little effect. In this time interval, the time evolution under $H_L$ should be close to that under $H$ and so continue to be determined by the shape resonances. This turns out to be true. Given this, it is reasonable to expect that the resonances of $H_L$ should converge to those of $H$ as $L\rightarrow\infty$. This turns out to be false in a spectacular way. Instead of converging to the resonances of $H$, the resonances of $H_L$ cluster together and move toward the real axis, leaving every compact set in the lower half plane.
The present work began with the observation that by modifying $V$ slightly, it is possible that {\it some} of the resonances of $H_L$ may not move towards the real axis, but instead converge to points in the open lower half plane. However, these points are not resonances of $H$! Moreover, numerical experiments suggested that there are special points on the real axis {which} seem to repel the approaching resonances.
To begin we will consider one dimensional Hamiltonians of the form $H_L=-d^2/dx^2 + V_1(x) + V_2(x-L)$, where $V_1$ and $V_2$ are compactly supported.
We give asymptotic formulas for the positions of the resonances of $H_L$ as $L\rightarrow\infty$. The special points, both those in the lower half plane where resonances may approach and those on the real axis where the imaginary parts are repelled, turn out to be energies at which {a certain} reflection coefficient of the (meromorphic continuation of the) scattering matrix for either $V_1$ or $V_2$ vanishes. We will call these reflectionless points. (That $V_1$ and $V_2$ play a symmetrical role here is perhaps not so surprising considering that $H_L=-d^2/dx^2 + V_1(x) + V_2(x-L)$ {has the same resonances as} $-d^2/dx^2 + V_1(x+L) + V_2(x)$.) We will prove three asymptotic formulas, two for resonances near a point in ${\mathbb R}$, depending on whether the point is reflectionless, and one for resonances near a reflectionless point in the lower half plane.
Next we examine what happens when we try to improve the convergence of $H_L$ to $H$ by adding an exponentially decreasing coupling constant. We consider $H_L=-d^2/dx^2 + V_1(x) +e^{-cL} \delta(x-L)$ and give asymptotic formulas for resonance positions near $k_0$ in six cases: three for $k_0$ on the line $\Im(k_0)=-c/2$ depending on whether $k_0$ is a resonance for $H_1$, a reflectionless point for $H_1$ or neither, one for $k_0$ a reflectionless point for $V_1$ below the line $\Im(k)=-c/2$, one for $k_0$ a resonance for $V_1$ above the line $\Im(k)=-c/2$. In addition, {there may be} one resonance of $H_L$ converging to zero. {As we will explain, similar phenomena occur when $e^{-cL} \delta(x-L)$ is replaced with $\mu(L) V(x-L)$ when $\mu(L) = e^{-cL}$ and $V$ is continuous with compact support. In fact the exponentially decreasing coupling is the critical rate of decrease. If it is slower than exponential the resonances of $H$ all disappear while if it is faster than any exponential, the resonances of $H_L$ converge to those of $H$.}
{For movies showing the large $L$ behavior of resonances of $H_L$ for both constant and decreasing coupling see the accompanying files resmovies.html, resc00.mp4 and resc16.mp4.}
Finally, we show that for long times (depending on $L$) the time evolution of a state $\phi$ under $H$ is close to the time evolution under $H_L$.
The lost resonances in the title of this paper refer to the original resonances {of $H$} which disappear and are not evident in the resonance set of $H_L$ as $L\rightarrow\infty$, even though they show up in the description of the time evolution of resonant states under $H_L$.
The found resonances are the resonances which are crowding together and moving toward the real axis {(or the line $\text {Im} (k) = -c/2$)} as $L$ becomes large.
\input hamres
\input asymptotic
\input asymptotic2
\input lost1
\input discussion
\input lost2
\input lavinebound
\input f2
\input biblio
\vfill\eject
\end{document}
\begin{center}{\em Not to be included in the public version}\end{center}
\input commutatoreqn
\input tV_2
\input variation
\end{document}
\section{Hamiltonian and resonances}
We will consider Hamiltonians defined as self-adjoint realizations of $H=-d^2/dx^2 + V$ acting in $L^2({\mathbb R} )$, where the potential $V$ has compact support. For simplicity, and since our goal is not to treat the most general potentials, assume that
$$V(x)=V_0(x) + \sum_{i=1}^N \alpha_i \delta(x-x_i),$$
where $V_0$ is continuous with compact support.
The resolvent $(H-z)^{-1}$ for $z\in {\mathbb C} \backslash ([0,\infty) \cup \{\hbox{eigenvalues}\})$ has an integral kernel given by the \textrm Green's function
\[
G(x,y;k) = \frac{1}{W(k)}\begin{cases} {\psi_1}(x;k){\psi_2}(y;k) &x\le y\\{\psi_1}(y;k) {\psi_2}(x;k) &y\le x\\\end{cases}
\]
where $\Im k > 0$, $k^2=z$ and ${\psi_1}$ and ${\psi_2}$ are solutions to
\begin{equation}\label{eqn:SSE}
-\psi'' + V(x)\psi = k^2\psi,
\end{equation}
satisfying ${\psi_1}(x;k)=e^{-ikx}$ for $x\le\inf\mathop{\rm supp}(V)$ and ${\psi_2}(x;k)=e^{ikx}$ for $x\ge\sup\mathop{\rm supp}(V)$. The Wronskian
\[
W(k) = \det\left(\twovec{{\psi_1}(x,k) & {\psi_2}(x,k) }{{\psi_1}'(x,k) & {\psi_2}'(x,k) }\right)
\]
does not depend on $x$.
When the potential includes a non-zero delta function term $\sum_{i=1}^N \alpha_i \delta(x-x_i)$ then \eqref{eqn:SSE} is interpreted to mean that $\psi(x)$ is a classical $C^2$ solution in the open intervals between the points $x_i$, while at the points $x_i$, $\psi(x)$ is continuous and $\psi'(x)$ has left and right limits that are related by
\[
\psi'(x_i+) = \psi'(x_i-)+\alpha_i\psi(x_i).
\]
This condition ensures that $\psi$ is (locally) in the domain of $H$, {(that is $\phi \psi $ is in the domain of $H$ for all $\phi \in C_0^{\infty}(\mathbb R))$}. (See [AR] for the meaning of this equation when $V$ is a finite Borel measure).
The solutions ${\psi_1}(x;k)$ and ${\psi_2}(x;k)$ are defined for any complex $k$. For fixed $x$ they are entire functions of $k$. The derivatives ${\psi_1}'(x;k)$ and ${\psi_2}'(x;k)$ for fixed $x$ ({or} the right and left limits when $x=x_i$) are also entire functions of $k$.
Given these analyticity properties, we see that the \textrm Green's function\ has a meromorphic continuation to all of ${\mathbb C}$ with poles at the zeros of $W(k)$.
The zeros of $W(k)$ can be characterised as those $k$ values for which ${\psi_1}(x;k)$ is a multiple of ${\psi_2}(x;k)$.
When this happens for $\Im(k)>0$ the solution ${\psi_1}$ is an eigenfunction so that $E=k^2$ is an eigenvalue of $H$.
This can only occur at points $k$ on the positive imaginary axis.
\begin{definition} The {\it resonances} of $H$ are poles in the continuation of the \textrm Green's function\ in $k$ to the lower half plane. Equivalently, they are zeros of $W(k)$ in the lower half plane.
\end{definition}
Another characterisation of resonances is that they are values of $k$ such that for some constants $a$ and $b$, ${\psi_1}(x;k) = a e^{ikx}$ for large positive $x$ and ${\psi_2}(x;k) = b e^{-ikx}$ for large negative $x$.
\begin{definition}
The {\it reflectionless points} are values of $k$ such that for some constants $c$ and $d$, ${\psi_1}(x;k) = c e^{-ikx}$ for large positive $x$ {or} ${\psi_2}(x;k) = d e^{ikx}$ for large negative $x$. They are the values of $k$ where {a} reflection coefficient in the {(meromorphically continued)} scattering matrix vanishes.
\end{definition}
With these definitions resonances and reflectionless points are $k$ values. The corresponding complex energies are at $z=k^2$ on the second sheet obtained by analytic continuation across the positive real axis.
Resonances are values of $k$ for which ${\psi_1}(x,k)$ is a multiple of ${\psi_2}(x,k)$.
We now analyze this condition when $V(x)=V_1(x)+V_2(x-L)$ with $\mathop{\rm supp}(V_1)\subseteq [x_0,0]$ and $\mathop{\rm supp}(V_2)\subseteq [0,x_1]$. For the moment we also assume that $k\ne 0$. Since $V(x)=0$ for $x\in(0,L)$, any solution $\psi(x;k)$ of \eqref{eqn:SSE} is a linear combination $\psi(x;k)=c_1 e^{-ikx} + c_2 e^{-ikx}$ for $x$ in this interval. This implies that for any solution $\psi(x;k)$
\[
\twovec{\psi(L;k) }{\psi'(L-;k) } = \twovec{\cos(kL)&\sin(kL)/k}{-k\sin(kL)&\cos(kL)}\twovec{\psi(0;k) }{\psi'(0+;k) }.
\]
Since $k\ne 0$, $\twovec{\phantom{-}ik&1}{-ik&1}$ is invertible and we have
\[
\twovec{\phantom{-}ik&1}{-ik&1}\twovec{\cos(kL)&\sin(kL)/k}{-k\sin(kL)&\cos(kL)}\twovec{\phantom{-}ik&1}{-ik&1}^{-1}
=\twovec{e^{ikL}&0}{0&e^{-ikL}}
\]
so that
\begin{equation}\label{eqn:PP}
\twovec{\phantom{-}ik\psi(L;k) +\psi'(L-;k)}{-ik\psi(L;k) +\psi'(L-;k)}=\twovec{e^{ikL}&0}{0&e^{-ikL}}\twovec{\phantom{-}ik\psi(0;k) +\psi'(0+;k)}{-ik\psi(0;k) +\psi'(0+;k)}.
\end{equation}
Now suppose that $k$ is a non-zero resonance. We apply \eqref{eqn:PP} to ${\psi_1}$ and use the fact that ${\psi_2}$ is a multiple of ${\psi_1}$. This yields
\begin{equation}\label{eqn:R}
e^{2ikL} = f(k)
\end{equation}
where
\begin{equation}\label{eqn:f}
f = f_1 f_2, \quad
f_1(k) =\left(\frac{-ik{\psi_1}(0;k) +{\psi_1}'(0+;k)}{\phantom{-}ik{\psi_1}(0;k) +{\psi_1}'(0+;k)}\right), \quad
f_2(k) =\left(\frac{\phantom{-}ik{\psi_2}(L,k) +{\psi_2}'(L-;k)}{-ik{\psi_2}(L,k) +{\psi_2}'(L-;k)}\right).
\end{equation}
On the other hand, if \eqref{eqn:R} holds with $f=f_1f_2$ given by \eqref{eqn:f}, then \eqref{eqn:PP} holds with ${\psi_2}$ on the right and ${\psi_1}$ on the left, up to a non-zero multiple. Equation \eqref{eqn:PP} also holds with ${\psi_1}$ on both sides. Thus we find that ${\psi_1}$ and ${\psi_2}$ satisfy the same initial condition up to a multiple. This implies that ${\psi_1}(x;k)$ is a multiple of ${\psi_2}(x;k)$ which means that $k$ is a resonance. Thus the non-zero resonances for $V_1(x)+V_2(x-L)$ can be characterized as solutions $k$ in the lower half plane of \eqref{eqn:R}, where $f$ is given by \eqref{eqn:f}.
Notice that $f(0)=1$ provided $\psi_1' (0+;0)$ and $\psi_2' (L-;0)$ are not zero. This is generically true when{ $V_1$ and $V_2$ are} not zero. Thus, $k=0$ is generically a solution to \eqref{eqn:R}, even if the \textrm Green's function\ does not have a pole at $0$.
For example, if $V_1(x)=V_2(x)=\alpha\delta(x)$ then $f(k)=(\alpha-2ik)^2/\alpha^2$ while $W(k) = 2(\alpha -ik) + i\alpha^2(1-e^{2ikL})/2k$. Here $W(0)=L\alpha^2 + 2\alpha$ is non-zero, so $k=0$ is not a pole of the \textrm Green's function. However $f(0)=1$ in this example so that $e^{2i0L}=f(0)$.
Notice that ${\psi_1}(0;k) = {\psi_1}(0;k;V_1)$ (the third argument denotes the relevant potential) while ${\psi_2}(L;k)={\psi_2}(0;k;V_2)$. This implies that $f(k)$ is independent of $L$. Moreover $f(k)$ has a pole if ${\psi_2}'(L-;k;V_2)=ik{\psi_2}(L;k;V_2)$ or ${\psi_1}'(0+;k;V_1)=-ik{\psi_1}(0;k;V_1)$. The first condition implies that the solution ${\psi_2}(x;k;V_2)$ is a multiple of $e^{ikx}$ both for $x<0$ and $x>x_1$. Thus $k$ is a reflectionless point for $V_2$. Similarly, the second condition holds if ${\psi_1}(x;k;V_1)$ is a multiple of $e^{-ikx}$ both for $x<x_0$ and $x>0$, that is, if $k$ is a reflectionless point for $V_1$. This shows that the poles of $f(k)$ occur at the reflectionless points of either $V_1$ or $V_2$. Similarly, the zeros of $f(k)$ occur at the resonances of either $V_1$ of $V_2$. {However if $k$ is a resonance of $V_1$ and at the same time a reflectionless point of $V_2$ (or vice versa), $k$ will in general be neither a pole nor a zero of $f$.}
We will need one more property of $f(k)$, namely that $|f(k)| > 1$ when {$k \in \mathbb R$}. To see this we define
$$
w_1(x;k)={\psi_1}'(x;k)/{\psi_1}(x;k),
$$
which should be thought of as an affine co-ordinate for the complex direction of $\twovec{{\psi_1}(x;k) }{{\psi_1}'(x,k) }$ in ${\mathbb C}^2$. We allow $w_1=\infty$ to include the direction where ${\psi_1}=0$. We have $w_1(x;k) = -ik$ for $x\le x_0$. For $x$ between the points $x_i$,
$w_1(x;k)$ solves the Ricatti equation
$$
w_1'(x;k) = V(x) - k^2 - w_1^2(x;k).
$$
At the points $x_i$, $w_1(x;k)$ jumps to the right by $\alpha_i$ with the convention that $\infty+\alpha_i=\infty$. For a better understanding of what happens near a point where $w_1=\infty$ define $u_1(x;k)=1/w_1(x;k)$. Then $u_1$ satisfies the equation $u_1' = 1+u_1^2(k^2-V(x))$ which is well defined near $u_1=0$. When $k$ and $w_1$ are both real, the right side of the Ricatti equation is real. Moreover, the jumps at the points $x_i$ are also in the real direction. This implies that if $w_1(x;k)$ is real for some $x$ it must be real for all $x$ and cannot satisfy the initial condition. Therefore for $k > 0$ $w_1(x;k)$ must stay in the lower half plane. For $k >0$ the fractional linear transformation $w\mapsto (w-ik)/(w+ik)$ maps the lower half plane to the exterior of the unit disk. Thus for $k >0$,
$$
\left|\frac{w_1(0;k)-ik}{w_1(0;k)+ik}\right| = \left|\frac{-ik{\psi_1}(0;k) +{\psi_1}'(0+;k)}{ik{\psi_1}(0;k) +{\psi_1}'(0+;k)}
\right| >1
$$
A similar argument using $w_2(x;k)={\psi_2}'(x;k)/{\psi_2}(x;k)$ shows
$$
\left|
\frac{ik{\psi_2}(L,k) +{\psi_2}'(L-;k)}{-ik{\psi_2}(L,k) +{\psi_2}'(L-;k)}
\right|>1
$$
as well. We see that $\bar{f}(k) = f(-\bar{k})$ which implies $|f(k)| > 1$ for $k < 0$.
The inequality $|f(k)|>1$ for $k \in \mathbb{R}\setminus \{0\}$
implies that \eqref{eqn:R} has no real non-zero solutions since $|e^{2iLk}|=1$ for these $k$ values.
Summarizing, we have proved the following proposition.
\begin{proposition} Let $f(k)$ be defined by \eqref{eqn:f}. A {nonzero} complex number $k$ solves \eqref{eqn:R}
if and only if
\begin{enumerate}
\item $k$ is on the positive imaginary axis and $k^2$ is an eigenvalue, or
\item $k$ is a resonance in the open lower half plane.
\end{enumerate}
The function $f(k)$ is a meromorphic function with poles at the reflectionless points of either $V_1$ or $V_2$ and zeros at the resonances of either $V_1$ or $V_2$. When {$k \in {\mathbb R}\backslash\{0\}$}, $ |f(k)| >1$. We have $\bar{f}(k) = f(-\bar{k})$ and thus if $k$ is a resonance then so is $-\bar k$.
\end{proposition}
\section{Appendix: Lavine's bound}
\newcommand{\jap}[1]{{\langle#1\rangle}}
In this appendix we present Lavine's bound on the time evolution of resonant states.
\begin{theorem} Let $H=-d^2/dx^2 + V$ acting in $L^2({\mathbb R})$, where $V$ satisfies the hypothesis of Section 2 with $\mathop{\rm supp}(V)\subseteq [-r,r]$. Suppose that $k_0$ is a resonance and let $\psi$ be the corresponding outgoing solution of
\begin{equation}\label{req}
-\psi''(x) + (V(x) - k_0^2)\psi(x) = 0.
\end{equation}
Recall that outgoing means that $\psi(x)$ is a multiple of $e^{-ikx}$ for $x\le -r$ and a multiple of $e^{ikx}$ for $x\ge r$. Write $k_0 = \sigma_0 - i\epsilon_0$ and $k_0^2 = \lambda_0-i\delta_0$ with $\sigma_0,\epsilon_0,\delta_0 > 0$. Let $\chi=\chi_{[-r,r]}$ be the indicator function for $[-r,r]$ and define $\varphi=\chi\psi$. Then
\begin{equation}\label{lavineest}
\left|\langle{\varphi,\left(e^{-itH} - e^{-it\lambda_0-\delta_0|t|}\right)\varphi}\rangle\right| \le C(k_0)\|\varphi\|^2
\end{equation}
where
$$
C(k_0)=\left (\frac{1}{5}\log(1+ (\frac{\sigma_0}{2\epsilon_0})^2) + 1\right )\frac{6\epsilon_0}{\sigma_0}
$$
\end{theorem}
\begin{remark} The constant $C$ is $O(({\epsilon_0}/{\sigma_0})\log(\sigma_0/\epsilon_0))$ for small ${\epsilon_0}/{\sigma_0}$, and is independent of $V$, $r$ and $t$. No spectral cutoff is required for $\varphi$, so the bound implies that $\varphi$ is almost orthogonal to bound states if $\epsilon_0$ is small. Lavine's result is for the half line. To generalize this to the whole line we rewrite his bounds in terms of interior and exterior Dirichlet to Neumann maps and use that these maps commute.
\end{remark}
Following Lavine, we begin the proof of this theorem by introducing the Lorenzian
distribution given by
$$
d\mu_L(\lambda) ={{1}\over{\pi}}\Im (\lambda_0-i\delta_0-\lambda)^{-1} d\lambda.
$$
The Fourier transform of this distribution is $e^{-it\lambda_0-\delta_0|t|}$.
Thus, if $\varphi\in L^2({\mathbb R})$ has spectral measure $d\mu_\varphi(\lambda)$,
we find that
\begin{equation}\label{ftformula}
\langle{\varphi,\left(e^{-itH} - e^{-it\lambda_0-\delta_0|t|}\right)\varphi}\rangle
= \int_{\mathbb R} e^{-it\lambda} \left(d\mu_\varphi(\lambda)-\|\varphi\|^2d\mu_L(\lambda)\right).
\end{equation}
This means that we will be able to approximate the time evolution of $\varphi$
if the spectral measure $\mu_\varphi$ is well approximated by $\|\varphi\|^2\mu_L$. In fact, it is enough to make this approximation on an interval $I$ outside which $\mu_L(\lambda)$ is small.
\begin{proposition}\label{lavineprop}
Suppose that for some interval $I\subset {\mathbb R}$ we have
\begin{equation}\label{diffest}
\int_I \left|d\mu_\varphi(\lambda) - \|\varphi\|^2d\mu_L(\lambda)\right| \le \epsilon_1\|\varphi\|^2
\end{equation}
and
\begin{equation}\label{lorentzest}
\int_{{\mathbb R}\backslash I} d\mu_L(\lambda) \le \epsilon_2.
\end{equation}
Then
$$
\left|\langle{\varphi,\left(e^{-itH} - e^{-it\lambda_0-\delta_0|t|}\right)\varphi}\rangle\right|
\le 2(\epsilon_1+\epsilon_2)\|\varphi\|^2
$$
\end{proposition}
\begin{proof}
From \eqref{ftformula} we find
\begin{align*}
\left|\jap{\varphi,\left(e^{-itH} - e^{-it\lambda_0-\delta_0|t|}\right)\varphi}\right|
&\le \int_I \left|\left(d\mu_\varphi(\lambda)-\|\varphi\|^2d\mu_L(\lambda)\right)\right|
+ \int_{{\mathbb R}\backslash I}\|\varphi\|^2d\mu_L(\lambda) + \int_{{\mathbb R}\backslash I}d\mu_\varphi(\lambda) \\
&\le (\epsilon_1 + \epsilon_2)\|\varphi\|^2 + \int_{{\mathbb R}\backslash I}d\mu_\varphi(\lambda)\\
\end{align*}
Since the measures $d\mu_\varphi$ and $\|\varphi\|^2d\mu_L$ both have total mass $\|\varphi\|^2$,
\begin{align*}
\int_{{\mathbb R}\backslash I}d\mu_\varphi(\lambda)
&=\|\varphi\|^2-\int_{I}d\mu_\varphi(\lambda)\\
&=\|\varphi\|^2-\|\varphi\|^2\int_{I}d\mu_L(\lambda) + \int_{I}\left(\|\varphi\|^2d\mu_L(\lambda)-d\mu_\varphi(\lambda)\right)\\
&\le \|\varphi\|^2\int_{{\mathbb R}\backslash I}d\mu_L(\lambda) + \int_{I}\left|d\mu_\varphi(\lambda)-\|\varphi\|^2d\mu_L(\lambda)\right|\cr
&\le (\epsilon_1+\epsilon_2)\|\varphi\|^2.\\
\end{align*}
This completes the proof.
\end{proof}
Recall that the resolvent $R(k)=(H-k^2)^{-1}$, initially defined for $\Im(k)>0$, has an integral kernel given by the\ \textrm Green's function\ $G(x,y;k)$. This representation defines the limit $\jap{\varphi, R(k+i0)\varphi}$ onto the real line (and the continuation beyond). We denote the continuation by $\jap{\varphi, R(k)\varphi}$ even when the operator $R(k)$ is not defined. With our assumptions on $V$, the spectral measure $d\mu_\varphi(\lambda)$ is absolutely continuous for $\lambda\in[0,\infty)$ with density $\pi^{-1}\Im(\jap{\varphi, (H-\lambda - i0)^{-1}\varphi})$. Thus the crucial quantity to be estimated in \eqref{diffest} can be written
\begin{align*}
\int_I \left|d\mu_\varphi(\lambda) - \|\varphi\|^2d\mu_L(\lambda)\right|
&=
{{1}\over{\pi}}\int_I\left|\Im\left[
\jap{\varphi, (H-\lambda - i0)^{-1} \varphi} - \|\varphi\|^2 (\lambda_0-i\delta_0-\lambda)^{-1}
\right]\right|d\lambda\\
&={{1}\over{\pi}}\int_{\sqrt{I}}\left|\Im
\jap{\varphi,\left[R(k) - {{1}\over{k_0^2-k^2}}\right]\varphi}
\right|2kdk\numberthis \label{resolventest}\\
\end{align*}
Here we are assuming that $I\subseteq [0,\infty)$.
We now compute another expression for the integrand.
\begin{proposition} Let $\psi$ be the outgoing solution to \eqref{req}. For $\Im(k)>0$ and $\chi\in C_0^\infty({\mathbb R})$ we have
\begin{equation}\label{coe2}
R(k)\chi\psi = {{1}\over{k_0^2-k^2}}\chi\psi - {{1}\over{k_0^2-k^2}}R(k)[H,\chi]\psi,
\end{equation}
and if $\chi_1\in C_0^\infty$ with $\chi_1[H,\chi]=0$ and $k$ is real, then
\begin{equation}\label{coe3}
\jap{\chi_1\psi,\left[R(k) - {{1}\over{k_0^2-k^2}}\right]\chi\psi}
={{1}\over{\left|k_0^2-k^2\right|^2}}\jap{[H,\chi_1]\psi,R(k)[H,\chi]\psi}
\end{equation}
\end{proposition}
\begin{proof}
Equation \eqref{coe2} follows from
$$
(H-k^2)\chi\psi = (k_0^2-k^2)\chi\psi +[H,\chi]\psi.
$$
To prove \eqref{coe3} we use \eqref{coe2} (twice) and $\chi_1[H,\chi]=0$ to write, for $\Im(k)>0$,
\begin{align*}
\jap{\chi_1\psi,\left[R(k) - {{1}\over{k_0^2-k^2}}\right]\chi\psi}
&={{-1}\over{k_0^2 - k^2}}\jap{\chi_1\psi,R(k)[H,\chi]\psi}\\
&={{-1}\over{k_0^2 - k^2}}\jap{R(-\overline k)\chi_1\psi,[H,\chi]\psi}\\
&={{1}\over{(k_0^2 - k^2)(\overline k_0^2-k^2)}}
\jap{-\chi_1\psi+R(-\overline k)[H,\chi_1]\psi,[H,\chi]\psi}\\
&={{1}\over{(k_0^2 - k^2)(\overline k_0^2-k^2)}}
\jap{[H,\chi_1]\psi,R(k)[H,\chi]\psi}\\
\end{align*}
Here we used that $R(k)^* = R(-\overline k)$. Now we take the limit on the real axis.
\end{proof}
We will use the following integration by parts formula
\begin{proposition} Let $\psi$ be the outgoing solution to \eqref{req} and $\varphi=\chi_{[-r,r]}\psi$. Then
\begin{equation}
\left\|\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}\right\|^2 = 2\epsilon_0\|\varphi\|^2
\end{equation}
\end{proposition}
\begin{proof}
\begin{align*}
2i\delta_0\|\varphi\|^2
&=2i\delta_0\|\psi\|^2_{L^2([-r,r])}\\
&= \jap{(\lambda_0-i\delta_0)\psi,\psi}_{L^2([-r,r])} - \jap{\psi,(\lambda_0-i\delta_0)\psi}_{L^2([-r,r])}\\
&= \jap{(-d^2/dx^2+V)\psi,\psi}_{L^2([-r,r])} - \jap{\psi,(-d^2/dx^2+V)\psi}_{L^2([-r,r])}\\
&= -\overline\psi'(x)\psi(x)\Big|_{x=-r}^r + \overline\psi(x)\psi'(x)\Big|_{x=-r}^r\\
\end{align*}
This implies
\begin{equation}
\delta_0\|\varphi\|^2 = \Im\left(\left\langle\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix},\begin{bmatrix}-\psi'(-r)\\\phantom{-}\psi'(r)\\\end{bmatrix}\right\rangle\right).
\end{equation}
Since $\psi$ is outgoing, $\begin{bmatrix}-\psi'(-r)\\\phantom{-}\psi'(r)\\\end{bmatrix}=ik_0\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}$. Since $k_0=\sigma_0-i\epsilon_0$ and $\delta_0=2\epsilon_0\sigma_0$ this gives the result.
\end{proof}
We now introduce the interior and exterior Dirichlet to Neumann maps. Given $\begin{bmatrix}a\\b\\\end{bmatrix}\in{\mathbb C}^2$, solve $-\psi'' + (V-k^2)\psi = 0$ on $[-r,r]$ with $\psi(-r)=a$ and $\psi(r)=b$. This is possible as long as $k^2$ is not a Dirichlet eigenvalue of $-d^2/dx^2 + V$. Then the interior Dirichlet to Neumann map is defined to be
$$
\Lambda_k\begin{bmatrix}a\\b\\\end{bmatrix}=\begin{bmatrix}-\psi'(-r)\\\psi'(r)\\\end{bmatrix}.
$$
When $k\in{\mathbb R}$, $\Lambda_k$ is a real symmetric $2\times 2$ matrix (when it is defined).
The exterior Dirichlet to Neumann map is defined similarly, except we now find the outgoing solution to $-\psi''-k^2\psi=0$ on $(-\infty,-r] \cup [r,\infty)$, again with $\psi(-r)=a$ and $\psi(r)=b$. The solution is simply $ae^{-ik(x+r)}$ on $(-\infty,-r]$ and $be^{ik(x-r)}$ on $[r,\infty)$ so the exterior Dirichlet to Neumann map is
$$
\Omega_k\begin{bmatrix}a\\b\\\end{bmatrix}=\begin{bmatrix}-\psi'(-r)\\\psi'(r)\\\end{bmatrix} = ik\begin{bmatrix}a\\b\\\end{bmatrix}.
$$
So we see that $\Omega_k = ikI$ where $I$ is the $2\times 2$ identity matrix.
Now let $r' > r$ and introduce the matrices
$$
{\bf G}(r,r',k) = \begin{bmatrix}G(-r,-r';k)&G(-r,r';k)\\G(r,-r';k)&G(r,r';k)\\\end{bmatrix}
$$
which are the restrictions of the resolvent to ``spheres''. Since the point $-r'$ is outside the interval $[-r,r]$, the functions $G(x,-r';k)$ and $G(x,r';k)$ solve the equation $(-d^2/dx^2+V(x)-k^2)G(x,\pm r';k)=0$ in the interval $[-r,r]$. This implies that
$$
\begin{bmatrix}-G_x(-r,-r';k)&-G_x(-r,r';k)\\G_x(r,-r';k)&G_x(r,r';k)\\\end{bmatrix} = \Lambda_k {\bf G}(r,r',k).
$$
Similarly
$$
\begin{bmatrix}-G_y(-r,-r';k)& G_y(-r,r';k)\\-G_y(r,-r';k)&G_y(r,r';k)\\\end{bmatrix} = {\bf G}(r,r',k)\Omega_k^T ={\bf G}(r,r',k)\Omega_k
$$
and
$$
\begin{bmatrix}G_{x,y}(-r,-r';k)&-G_{x,y}(-r,r';k)\\-G_{x,y}(r,-r';k)&G_{x,y}(r,r';k)\\\end{bmatrix} = \Lambda_k {\bf G}(r,r',k)\Omega_k.
$$
Now we take the limit as $r'\rightarrow r$ and define
$${\bf G}(k)= \lim_{r'\rightarrow r}{\bf G}(r,r',k).$$
\begin{proposition} The matrix ${\bf G}(k)$ has the representation
$$
{\bf G}(k) = (\Lambda_k - \Omega_k)^{-1}
$$
\end{proposition}
\begin{proof}
$\Lambda_k {\bf G}(r,r';k)$ is given by the formula above. To compute $\Omega_k{\bf G}(r,r';k)$ introduce the exterior Dirichlet\ \textrm Green's function\ $G_0(x,y,k)$ on $(-\infty,-r] \cup [r,\infty)$. This can be computed explicitly.
Then $G(x,\pm r';k)-G_0(x,\pm r';k)$ is an outgoing solution to $(-d^2/dx^2-k^2)G(x,\pm r';k)=0$ for $x\in(-\infty,-r] \cup [r,\infty)$. This implies that
$$
\Omega_k {\bf G}(r,r';k) = \begin{bmatrix}-G_x(-r,-r';k)+G_{0,x}(-r,-r';k)&-G_x(-r,r';k)+G_{0,x}(-r,r';k)\\G_x(r,-r';k)-G_{0,x}(r,-r';k)&G_x(r,r';k)-G_{0,x}(r,r';k)\\\end{bmatrix}
$$
so that
\begin{equation}\label{dirgreen}
(\Lambda_k - \Omega_k){\bf G}(r,r';k) = \begin{bmatrix}-G_{0,x}(-r,-r';k)&-G_{0,x}(-r,r';k)\\G_{0,x}(r,-r';k)& G_{0,x}(r,r';k)\\\end{bmatrix}
\end{equation}
Since $(-\infty,-r] \cup [r,\infty)$ is disconnected, $G_{0,x}(r,-r';k)=G_{0,x}(-r,r';k)=0$ To compute $G_{0,x}(r,r';k)$ let $f(x)$ be an outgoing solution of $(-d^2/dx^2 - k^2)f=0$ on $[r,\infty)$. Explicitly $f(x)=f(r)e^{ik(x-r)}$. Then, since $(-d^2/dx^2 - k^2)G_0(x,r';k)=\delta(r-r')$, we have that for $R>r'$
\begin{align*}
f(r') &= \int_r^R \big((-d^2/dx^2 - k^2)G_0(x,r';k)\big) f(x)dx\\
&=-G_{0,x}(x,r';k)f(x)\Big|_{x=r}^R + G_{0}(x,r';k)f_x(x)\Big|_{x=r}^R + \int_r^R G_0(x,r';k)\big((-d^2/dx^2 - k^2)f(x)\big) dx\\
&=-G_{0,x}(R,r';k)f(R) + G_{0,x}(r,r';k)f(r)+G_{0}(R,r';k)f_x(R)-G_{0}(r,r';k)f_x(r)\\
&=-ikG_{0}(R,r';k)f(R) + G_{0,x}(r,r';k)f(r)+G_{0}(R,r';k)ikf(R) - 0\\
&=G_{0,x}(r,r';k)f(r)\\
\end{align*}
Here we used that $G_0(x,r';k)$ and $f(x)$ are both outgoing at $x=R$, and $G_{0}(r,r';k)=0$. Now take $r'\rightarrow r$ to conclude $\lim_{r'\rightarrow r}G_{0,x}(r,r';k)=1$. Similary $\lim_{r'\rightarrow r}G_{0,x}(-r,-r';k)=-1$.
So the right side of \eqref{dirgreen} converges to $I$ as $r'\rightarrow r$. This completes the proof.
\end{proof}
\begin{remark}
This proposition and the preceeding formulas can also be proven using the explicit representation for the\ \textrm Green's function\ in terms of ${\psi_1}$ and ${\psi_2}$. The above proof has the advantage that it generalizes to higher dimensions. However, in higher dimensions $\Omega_k$ and $\Lambda_k$ no longer commute, unless, for example, $V$ is radial.
\end{remark}
Now we return to \eqref{coe3} and rewrite the right side. We have
\[
\jap{[H,\chi_1]\psi,R(k)[H,\chi]\psi} = \jap{(D_x\chi_1' + \chi_1'D_x)\psi,R(k)(D_x\chi' + \chi'D_x)\psi}
\]
Here $D_x$ is the operator of differentiation by $x$. Let $r'>r$ and choose $\chi_1$ to be a smooth cutoff to $[-r,r]$ and $\chi$ a smooth cutoff to $[-r',r']$ chosen so that the derivatives $\chi_1'$ and $\chi'$ have disjoint support. Then $\chi_1'(x) = \delta_{-r}(x) - \delta_r(x)$ and $\chi'(x) = \delta_{-r'}(x) - \delta_{r'}(x)$ where $\delta_{-r}(x), \delta_r(x), \delta_{-r'}(x),\delta_{r'}(x)$ are approximate delta functions with disjoint support. Now we write the right side as a double integral with\ \textrm Green's function s in the integrand. Then integrate by parts, until the derivatives are hitting either the\ \textrm Green's function\ or $\psi$. Then we let take the limit as the approximate delta functions become exact delta functions. Finally we let $r'\rightarrow r$. A typical term in this calculation looks like
\begin{align*}
&\jap{D_x(\delta_{-r}(x) - \delta_r(x))\psi,R(k)(\delta_{-r'}(x) - \delta_{r'}(x))D_x\psi }\\
&=-\int\int (\delta_{-r}(x) - \delta_r(x))\overline\psi(x) G_x(x,y;k)(\delta_{-r'}(y) - \delta_{r'}(y))D_y\psi(y) dy dx\\
&=\left\langle\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}, \Lambda_k{\bf G}(k)\begin{bmatrix}\psi'(-r)\\-\psi'(r)\\\end{bmatrix}\right\rangle\\
&=\left\langle\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}, -\Lambda_k{\bf G}(k)\Omega_{k_0}\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}\right\rangle\\
\end{align*}
In this way we arrive at
\begin{align*}
{{1}\over{\left|k_0^2-k^2\right|^2}}&\jap{[H,\chi_1]\psi,R(k)[H,\chi]\psi}\\
&={{1}\over{\left|k_0^2-k^2\right|^2}}\left\langle\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix},(\Lambda_k-\Omega_{k_0}^*)(\Lambda_k-\Omega_k)^{-1}(\Omega_k-\Omega_{k_0})\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}\right\rangle
\end{align*}
Recall that $\Omega_k=ik$. This leads to
\begin{align*}
\|(\Lambda_k-\Omega_{k_0}^*)(\Lambda_k-\Omega_k)^{-1}(\Omega_k-\Omega_{k_0})\| &=
\|(\Omega_k-\Omega_{k_0}) +(\Omega_k-\Omega_{k_0}^*)(\Lambda_k-\Omega_k)^{-1}(\Omega_k-\Omega_{k_0})\|\\
&\le |k-k_0|\left(1 + \frac{|k+\overline k_0|}{|k|}\right)\\
\end{align*}
So returning to \eqref{resolventest} with $I =[(\sigma_0/2)^2, (3\sigma_0/2)^2]$, $\sqrt{I}=[\sigma_0/2, 3\sigma_0/2]$ we find
\begin{align*}
\int_I \left|d\mu_\varphi(\lambda) - \|\varphi\|^2\mu_L(\lambda)\right|
&={{1}\over{\pi}}\int_{I}\left|\Im
\jap{\varphi,\left[R(k) - {{1}\over{k_0^2-k^2}}\right]\varphi}
\right|dk^2\\
&\le {{1}\over{\pi}}\int_I \frac{|k-k_0|}{|k_0^2-k^2|^2}\left(1 + \frac{|k+\overline k_0|}{k}\right)\left\|\begin{bmatrix}\psi(-r)\\\psi(r)\\\end{bmatrix}\right\|^2dk^2\\
&\le {{4\epsilon_0}\over{\pi}}\int_{\sqrt{I}}\frac{k+|k+k_0|}{|k_0+k|^2}\frac{1}{|k-k_0|}dk\; \|\varphi\|^2\\
\end{align*}
We have
$$\frac{4\epsilon_0}{\pi} \int_{\sqrt{I}} \frac{k + |k+k_0|}{|k_0 + k|^2 |k-k_0|} dk \le \frac{4\epsilon_0}{\pi} \int_{\sqrt{I}} \frac{k + |k+\sigma_0|}{|\sigma_0 + k|^2 |k-k_0|} dk$$
The derivative of $\frac{2k+\sigma_0}{(k+\sigma_0)^2} $ is negative so in this term we replace $k$ by $\sigma_0/2$. Thus
$$\frac{4\epsilon_0}{\pi} \int_{\sqrt{I}} \frac{k + |k+k_0|}{|k_0 + k|^2 |k-k_0|} dk \le \frac{32}{9\pi}(\epsilon_0/\sigma_0)\int_{\sqrt{I}} \frac{dk}{|k-k_0|} \le \frac{32}{9\pi}(\epsilon_0/\sigma_0)\int_{-1/2}^{1/2} \frac{dx}{\sqrt{x^2 + (\epsilon_0/\sigma_0)^2}}$$
$$ \le \frac{64}{9\pi}(\epsilon_0/\sigma_0)\log \left( \frac{\sigma_0}{2\epsilon_0} + \sqrt {1 + (\frac{\sigma_0}{2\epsilon_0})^2}\right)$$
The other estimate is for $\int_{\mathbb{R} \setminus I}d\mu_L(\lambda)$. We first estimate the contribution from $[0,\infty)\setminus I$ which is
$$\pi^{-1} \int_{[0,\infty)\setminus {\sqrt{I}}}|\text {Im} \frac{1}{k^2-k_0^2}|2kdk = \pi^{-1} \int_{[0,\infty)\setminus {\sqrt{I}}}|\text {Im} (\frac{1}{k-k_0} + \frac{1}{k+k_0})|dk$$
$$ = \epsilon_0\pi^{-1} \int_{[0,\infty)\setminus {\sqrt{I}}}(\frac{1}{|k-k_0|^2} - \frac{1}{|k+k_0|^2})dk \le \epsilon_0\pi^{-1} \int_{[0,\infty)\setminus {\sqrt{I}}}\frac{1}{|k-k_0|^2} dk$$
$$ \le \epsilon_0\pi^{-1} \int _ {[0,\infty)\setminus {\sqrt{I}}}\frac {1}{(k-\sigma_0)^2 }dk = \frac{3 \epsilon_0}{\pi \sigma_0}.$$
Next we integrate over $(-\infty,0)$.
$$\pi^{-1} \int_{-\infty}^0|\text{Im} (\lambda - k_0^2)^{-1}|d\lambda = \pi^{-1} \int_0^{\infty}|\text{Im} (-k^2 - k_0^2)^{-1}|2kdk = \pi^{-1} \int_0^{\infty}|\text{Im}( \frac{1}{k+ik_0} + \frac{1}{k-ik_0})|dk $$
$$ = \sigma_0\pi^{-1} \int_0^{\infty}( \frac{1}{|k-ik_0|^2} - \frac{1}{|k+ik_0|^2})dk = 4 \sigma_0\epsilon_0\pi^{-1} \int_0^{\infty}( \frac{k}{|k-ik_0|^2|k+ik_0|^2} dk$$
$$\le 4 \sigma_0\epsilon_0\pi^{-1} \int_0^{1+\epsilon_0}\frac{1}{2\sigma_0} \frac{1}{(k-\epsilon_0)^2 + \sigma_0^2} dk + 4 \sigma_0\epsilon_0\pi^{-1} \int_{1+\epsilon_0}^{\infty}\frac{1}{\sigma_0} \frac{1}{(k-\epsilon_0)^2 + \sigma_0^2} dk $$
$$\le \frac{4\epsilon_0}{\sigma_0\pi}( \pi/2 + \pi/2) = \frac{4\epsilon_0}{\sigma_0}$$
Summing we get for the total error bound
$$\left (\frac{64}{9\pi}\log \left( \frac{\sigma_0}{2\epsilon_0} + \sqrt {1 + (\frac{\sigma_0}{2\epsilon_0})^2}\right) + (4+3/\pi) \right)\frac{\epsilon_0}{\sigma_0}$$
$$ <\left (\frac{1}{5}\log(1+ (\frac{\sigma_0}{2\epsilon_0})^2) + 1\right )\frac{6\epsilon_0}{\sigma_0}$$
This completes the proof of Lavine's estimate.
\section{Appendix: Proof of Theorem \ref{feelingV_L-2}}
Our proof of Theorem \ref{feelingV_L-2} uses the basic equation (\ref{basic equation}). Thus we need to bound $||V_{2,L}e^{-isH}\psi||$.
For this we need a propagation estimate to show that $e^{-isH} \psi(x)$ remains small for $x \ge L$ for times $s$ for which the velocities in $\psi$ do not allow propagation of $\psi $ from around the origin where it is localized initially to $x \ge L$. We first make an energy cut-off for $\psi$ so that we have an effective cut-off in the velocity. We finally estimate the remainder using the fact that $\psi$ is also well localized in energy.
Let $\tilde {h}_j, j = 1,2$ be two real non-negative $C^{\infty}(\mathbb R)$ functions such that $\tilde{h}^2_1 + \tilde{h}^2_2 = 1$ and $\tilde{h}_1(t) = 1$ for $-\infty < t < 1$ and $\tilde{h}_1(t) = 0$ for $ 1+\epsilon < t < \infty$ with $\epsilon > 0$. Let $k_j(t) = \tilde {h}_j(t/K)$ where we will later choose $K$ to increase with $L$. Let $\psi_K = k_1(H)\psi$ and $\psi_{K,t} = e^{-itH}\psi_K$. Let $h_j(t) = \tilde {h}_j(t/(1+\epsilon)K)$
Let $v = 2(1+\epsilon)\sqrt K$. ($2\sqrt {(1+\epsilon)K}$ is the maximum possible velocity in $\psi_K$ if we ignore $V_1$). Let $\langle x \rangle = \sqrt{|x|^2 + 1}$.
\begin{proposition}
For any $\alpha > 0$ and $t\ge 0$
\begin{equation} \label{propest}
||1_{\{|x| > v(1+ 2\epsilon)(t+1)\ge 0 \} }\langle x \rangle^{\alpha/2} \psi_{K,t}|| \le (1+\epsilon^{-1})^{\alpha}||\langle x \rangle^{\alpha/2} \psi_{K}|| + C_{\alpha}(K)
\end{equation}
where $\lim_ {K \to \infty} C_{\alpha}(K) = 0$.
\end{proposition}
An estimate of this form appears in \cite {S} where Skibsted proves several propagation estimates for N-particle problems. The above estimate is essentially in \cite {S} but for an $h_1$ which vanishes in a neighborhood of the thresholds of the problem ($0$ in our case). This restriction was necessary for other propagation estimates but not for this one (as Skibsted knew \cite { S1}). Control of the constants which occur in the proof of the result is essential for us and this needs a bit more care than is present in the proof of \cite {S} . For these reasons and for completeness we present a proof of the proposition (following the method in \cite {S}). We do not assume that we are in one dimension.
\begin{proof}
We first introduce the propagation observable: $A(\tau) = \tau v - \langle x \rangle $, where $\tau = t + 1$. For each $\alpha \ge 0$ we will need a function $$g(u, \tau) = -(-u)^{\alpha} \chi(u/\tau)$$ where $\chi$ satisfies the following: $u\chi'(u) + \alpha\chi(u)= \tilde \chi^2(u)$ with $\chi' \le 0$, $\tilde\chi \ge 0$, and $\tilde \chi \in C^{\infty}(\mathbb R)$, $\chi(u) = 1, u \le -2\epsilon v, \chi(u) = 0, u \ge -\epsilon v$ for $\epsilon > 0$. Such a $\chi$ can be shown to exist for any $\alpha \ge 0$ and $\epsilon v > 0$.
We calculate
$$\partial g(u,\tau)/ \partial u = (-u)^{\alpha - 1}[\alpha \chi(u/\tau) + (u/\tau)\chi'(u/\tau)],$$
$$\partial g(u,\tau)/\partial \tau = (-u)^{\alpha + 1}\tau^{-2} (-\chi'(u/\tau)).$$
If $u = u(t)$, then
$$(d/dt) g(u(t),\tau) = (\partial g(u,\tau)/ \partial u )du/dt + \partial g(u,\tau)/\partial \tau $$ \\
which is positive if $du/dt \ge 0$. If $u(t) = \tau v - \langle x(t) \rangle $ with $x(t)$ the orbit of a classical particle with Hamiltonian $p^2 + V_1(x)$ then $du/dt = v - \frac{x(t)}{\langle x(t) \rangle }\cdot 2p$ which is positive if $v \ge |2p|$. This would lead to the useful estimate
$$( \langle x(t) \rangle - (t+1)v)^{\alpha} \chi(v- \langle x(t) \rangle /(t+1)) \le ( \langle x(0) \rangle - v)^{\alpha} \chi(v- \langle x(0) \rangle )$$
Quantum mechanically we are interested in
$$ (d/dt) e^{itH} g(A(\tau), \tau) e^{-itH} = e^{itH}\Big( i[H, g(\tau v - \langle x \rangle, \tau) ] + v\partial g(A(\tau),\tau)/\partial u + \partial g(A(\tau),\tau) /\partial \tau \Big)e^{-itH}.$$
The idea is to show that $$d/dt(\psi_{K,t}, g(A(\tau), \tau) \psi_{K,t}) \ge I(t) $$ where $I(t)$ is integrable on $[0,\infty)$. Then it would follow that
$$ (\psi_{K,t}, (-A(\tau))^{\alpha} \chi(A(\tau)/\tau) \psi_{K,t}) \le (\psi_K, (-A(1))^{\alpha} \chi(A(1)) \psi_K) + \int_0^{\infty}|I(t)|dt$$
We have
$$ i[H, g(\tau v - \langle x \rangle, \tau) ] = - p\cdot (\partial g(A(\tau),\tau)/ \partial u) x/\langle x \rangle - (\partial g(A(\tau),\tau)/ \partial u)( x/\langle x\rangle) \cdot p$$
Let $f(u, \tau) = (-u)^{(\alpha - 1)/2} \tilde \chi(u/\tau)$.
$$ i[H, g(\tau v - \langle x \rangle, \tau) ] + v\partial g(A(\tau),\tau)/\partial u = - (p \cdot (x/\langle x\rangle)\partial g/\partial u+ \partial g/\partial u \cdot (x/\langle x\rangle) \cdot p) + v\partial g/\partial u $$
$$ = -(p \cdot \hat x f(A(\tau),\tau)^2+ f(A(\tau),\tau)^2\hat x\cdot p) + v f(A(\tau),\tau)^2$$
$$ = f(v - (\hat x\cdot p + p \cdot\hat x))f$$
where $\hat x = x/\langle x \rangle $. We have $-(\hat x\cdot p + p \cdot\hat x)) \le p^2/a + a$ so taking $a = v/2$ we have
$$ i[H, g(\tau v - \langle x \rangle, \tau) ] + v\partial g(A(\tau),\tau)/\partial u \ge f(v/2 -2 p^2/v) f$$
$$ = 2fV_1f/v + f(v/2 - 2v^{-1} H)f$$
Because $h_1^2(H)H \le (1+\epsilon)^2Kh_1(H)^2$,\ $h_1(H)^2(v/2 - 2v^{-1}H) \ge 0$ and thus
$$ f(v/2 - 2v^{-1} H)f \ge f h_2 (H)(v/2 - 2v^{-1} H)h_2(H)f \ge -2v^{-1} fh_2(H) H h_2(H)f$$ and since $k_1h_2 = 0$,
$$ d/dt (\psi_{K,t}, g(A(\tau),\tau)\psi_{K,t}) \ge 2v^{-1}(\psi_{K,t}, f^2 V_1(x)\psi_{K,t}) -2v^{-1} (\psi_{K,t} [f,h_2(H)]H[h_2(H), f]\psi_{K,t}). $$
Note that the derivatives of $h_2$ have compact support. For $n>1$ we have (see \cite{S} but note typos and different definition of $\ad$ - we define $\ad_H(B) = [H,B]$ and $\ad_H^j = (\ad_H)^j$):
$$[h_2(H),f(A(\tau),\tau)] = \sum_{j=1}^{n-1} (-1)^{j+1}h_2^{(j)}(H) \ad_{H}^j(f)/j! + \int_{-\infty}^{\infty} \widehat{ h_2^{(n)}}(w)e^{iwH} R^r_{n,H,f}(w)dw$$
$$ = \sum _{j=1}^{n-1}\ad_{H}^j(f) h_2^{(j)}(H)/j! + \int_{-\infty}^{\infty} \widehat{ h_2^{(n)}}(w) R^l_{n,H,f}(w)e^{iwH}dw.$$
where
$$R^r_{n,H,f}(w) = \int_0^w (s-w)^{n-1}w^{-n}e^{-isH}\ad_{H}^n(f) e^{isH}ds /(n-1)!/\sqrt{2\pi}. $$ and
$$ R^l_{n,H,f}(w) = \int_0^w (w-s)^{n-1}w^{-n}e^{isH}\ad_{H}^n(f) e^{-isH}ds /(n-1)!/\sqrt{2\pi}.$$
Note that $h_2^{(j)}(H) \psi_{K,t} = 0$ so that only the remainder terms in the commutators contribute. We have
$$\ad^n_{H}(f) = [H,[H,...,[H,f]..] = \sum_{|\gamma| \le n, |\beta|= 2n - |\gamma|} c_{\gamma,\beta}p^{\gamma} D_x^{\beta}f = \sum_{|\gamma| \le n, |\beta|= 2n - |\gamma|} c'_{\gamma,\beta}D_x^{\beta}f p^{\gamma}$$
$$D_x^{\beta} f(u,\tau) = \sum_{|\gamma| \le |\beta| -1} c_{\gamma}(x) \langle x \rangle^{-|\gamma|} D_u^{|\beta| - |\gamma|}f$$ where $c_{\gamma}(x)$ is a bounded smooth function with bounded derivatives.
$$(\partial/\partial u)^{\lambda} f(u,\tau) = \sum_{m \le \lambda} c_m (-u)^{(\alpha - 1)/2 -(\lambda -m)} \tau^{-m} \tilde \chi^{(m)}(u/\tau)$$
In the support of $\tilde \chi^{(m)}(u/\tau)$ for $m > 0$ we have $(-u/\langle x \rangle)^{|\gamma|} \le (2\epsilon v \tau/(\tau(v+\epsilon v))^{|\gamma|} = (2\epsilon/(1+\epsilon))^{|\gamma|}$ while in the support of $\tilde \chi (u/\tau)$ we have $0 \le -u/\langle x \rangle \le 1$.
It follows that with an $\epsilon$ dependent $c_{\beta}$
$$ |D_x^{\beta}f| \le c_{\beta}[(-u)^{(\alpha -1)/2 - |\beta|}\tilde \chi(u/\tau) +( \epsilon v) ^{(\alpha - 1)/2 }\tau^{(\alpha - 1)/2 - |\beta| }]$$ We assume $n > (\alpha - 1)/2 + 2$. Thus since $|\beta| \ge n$ we have $$ |D_x^{\beta}f| \le c_{\beta}'( \epsilon v)^{(\alpha - 1)/2 )}\tau ^{(\alpha - 1)/2 - |\beta|}.$$ We have
$$||k_1(H)p^{\gamma} D_x^{\beta}f || \le CK^{|\gamma|/2}(\epsilon v\tau) ^{(\alpha - 1)/2 - |\beta|}$$ and
$$||k_1(H)\ad^n_{H}(f)|| \le C\sum _{|\beta| \ge n}K^{n + (\alpha -1)/4- |\beta|/2}\tau ^{(\alpha - 1)/2 - |\beta|}$$
We find $\int_{-\infty}^{\infty} |\widehat{h_2^{(n)}}(w)|dw = CK^{-n}$. Then
$$||k_1(H)[f(A(\tau),\tau),h_2(H)]|| \le C\sum _{|\beta| \ge n}K^{(\alpha -1)/4- |\beta|/2}\tau ^{(\alpha - 1)/2 - |\beta|}$$
We use this for the left commutator in $ (\psi_{K,t}, [f,h_2(H)]H[h_2(H), f]\psi_{K,t})$. For the term $H[h_2(H), f]\psi_{K,t}$ we use $[h_2(H), f]k_1(H) = - \int_{-\infty}^{\infty} \widehat{ h_2^{(n)}}(w) R^l_{n,H,f}(w)e^{iwH}dwk_1(H)$
so that
$$||H[h_2(H), f]k_1(H)|| \le C\sum _{|\beta| \ge n}K^{1 +(\alpha -1)/4 - |\beta|/2}\tau^{(\alpha - 1)/2 - |\beta|} + C\sum _{|\beta| \ge n+1}K^{1 + (\alpha - 1)/4 - |\beta|/2}\tau ^{(\alpha - 1)/2 - |\beta|} .$$ It follows that with $K \ge 1$
$$v^{-1}||k_1(H)[f,h_2(H)]H[h_2(H),f]k_1(H)|| \le C K^{(\alpha +2 - 2n)/2} \tau^{\alpha - 1 - 2n}.$$ Since we have already taken $2n > \alpha +3$ we have
$$ (\psi_{K,t}, (-A(\tau))^{\alpha} \chi(A(\tau)/\tau) \psi_{K,t}) \le (\psi_K, (-A(1))^{\alpha} \chi(A(1)) \psi_K) + v^{-1}\int_0^{\infty}(\psi_{K,t}, |V_1(x)|f^2\psi_{K,t} )dt + C'(K)$$
where $\lim_ {K\to \infty} C'(K) = 0$. Since $V_1$ is bounded with compact support $|V_1(x)|f^2 \le c_N (v\tau)^{-N}$ for any $N$ and thus
$$ (\psi_{K,t}, (-A(\tau))^{\alpha} \chi(A(\tau)/\tau) \psi_{K,t}) \le (\psi_K, (-A(1))^{\alpha} \chi(A(1)) \psi_K) + C''(K) \le ||\langle x \rangle ^{\alpha/2}\psi_K||^2 + C''(K)$$
where $\lim_ {K\to \infty} C''(K) = 0$.
Since in the support of $ \chi(A(\tau)/\tau), \langle x \rangle -v\tau \ge \epsilon \langle x \rangle/(1+\epsilon)$ we have
$$ (\psi_{K,t}, \langle x \rangle^{\alpha} \chi(A(\tau)/\tau) \psi_{K,t}) \le (1+ \epsilon^{-1})^{\alpha}||\langle x \rangle ^{\alpha/2}\psi_K||^2 + C'''(K).$$ In addition $\chi(A(\tau)/\tau)) = 1$ if $\langle x \rangle \ge (1+2\epsilon)v\tau$ which gives Proposition \ref{propest}.
\end{proof}
\begin{proof} of Theorem \ref{feelingV_L-2}
Now back to what we want to bound: $$\int_0^t ||V_{2,L}(x)e^{-isH}\psi ||ds \le ||V_{2,L}||_{\infty}\Big ( t || k_2(H)\psi|| +L^{-\alpha/2}\int_0^t ||1_{\{\langle x \rangle \ge L\}}\langle x \rangle ^{\alpha/2}e^{-isH} \psi_K|| ds \Big).$$
We have used $1-k_1 \le k_2$. Since $\psi \in \mathcal{S}(\mathbb R)$, $||k_2(H)\psi|| \le c_N K^{-N}$.
We have
$$||1_{\{\langle x \rangle \ge L\}}\langle x \rangle ^{\alpha/2}e^{-isH}\psi_K||
\le ||1_{\{\langle x \rangle \ge (1+ 2\epsilon) v (s+1)\}}\langle x \rangle ^{\alpha/2}e^{-isH} \psi_K|| $$
where we take $(1+ 2\epsilon) v(t+1)\le L$ and $0 \le s \le t$. It follows from Proposition \ref{propest} that
$$||1_{\{\langle x \rangle \ge L\}}\langle x \rangle ^{\alpha/2}e^{-isH}k_1(H) \psi|| \le (1 + \epsilon^{-1})^{\alpha} ||\langle x \rangle ^{\alpha/2}e^{-isH}k_1(H) \psi|| + C_{\alpha}(K)$$ where $\lim _{K \to \infty} C_{\alpha}(K) =0$.
Thus we have
$$\int_0^t ||V_{2,L}(x)e^{-isH}\psi ||ds \le t(c_NK^{-N} + cL^{-\alpha/2}(||\langle x \rangle ^{\alpha/2}k_1(H)\psi|| +1).$$
We take $K= L^{2\delta}$ for $0 < \delta << 1$ and thus $t \le (1+\epsilon)^{-2} L/ \sqrt {4K} = cL^{1-\delta}$.
Since $\alpha$ is arbitrary we have
$$\int_0^t ||V_{2,L}(x)e^{-isH}\psi ||ds \le c_nL^{-n}$$
as long as $t\le cL^{1-\delta}$ and we can prove that $||\langle x \rangle ^{\alpha/2}k_1(H)\psi||$ is bounded as $K\to \infty$. To deal with $||\langle x \rangle ^{\alpha/2}k_1(H)\psi||$ write the square $$(\psi,k_1(H)\langle x \rangle ^{\alpha}k_1(H)\psi) = (\psi, k_1(H)^2\langle x \rangle ^{\alpha}\psi) + (\psi,k_1(H)[\langle x \rangle ^{\alpha},k_1(H)]\psi).$$ We use
$$[k_1(H),f] = \sum_{j=1}^{n-1}(-1)^{j+1} k_1^{(j)}(H) \ad_{H}^j(f)/j! + \int_{-\infty}^{\infty} \widehat{ k_1^{(n)}}(w)e^{iwH} R^r_{n,H,f}(w)dw$$ where $f = \langle x \rangle ^{\alpha}$.
We have
$$\ad^j_{H}(\langle x \rangle^{\alpha} ) = [H,[H,...,[H,\langle x \rangle ^{\alpha}]..] = \sum_{|\gamma| \le j, |\beta|= 2j - |\gamma|} c_{\gamma,\beta}p^{\gamma} D_x^{\beta}\langle x \rangle ^{\alpha}$$ and thus
$$|| k_1^{(j)}(H)\ad^j_{H}(\langle x \rangle^{\alpha} )\psi|| \le CK^{-j}K^{j/2}||\langle x \rangle^{\alpha} \psi||.$$ Finally $$\int _{-\infty}^{\infty}|| k_1(H)R^r_{n,H,f}|| \ |\widehat{k_1^{(n)}}(w)| dw \le CK^{-n}K^{n/2}$$
if $n\ge \alpha$.
This gives Theorem \ref{feelingV_L-2}.
\end{proof}
|
\section{Introduction}
The layered transition metal oxides have received special attention in condensed matter physics research because of their rich physical properties, including ion intercalation for potential battery applications, superconductivity, and intercalant-sensitive magnetic phase transitions.~\cite{Terasaki1997, Schaak2003, Lee2006, Shu2016, Weller2009, Berthelot2011, Medarde2013, Shu2013, Linden2010} The group IA ions that are sandwiched between the transition metal oxide layers act as passive charge reservoirs to influence the electronic structure. $P2$-type Na$_2$Ni$_2$TeO$_6$ has been shown to exhibit high ionic conductivity at room temperature for potential applications as a separator in the Na-ion battery.~\cite{Berthelot2012, Evstigneeva2011,Gupta2013} The mixed edge-sharing (Te/Ni)O$_6$ octahedra in each layer create a unique Te-centered NiO$_6$ honeycomb network, which leads to a complex potential field profile that is used to drive the Na ion diffusion in 2D.
X-ray diffraction studies have shown that the crystal structure of Na$_2$Ni$_2$TeO$_6$ can be indexed with a space group $P6_3/mcm$ satisfactorily, and a 3D antiferromagnetic (AFM) spin ordering of T$_N$$\sim$27 K has been proposed; however, the impact of Na ion diffusion on the crystal and spin structure has never been explored in detail, in contrast to its sister compound Na$_x$CoO$_2$ which has been studied extensively and shown rich physical insights.\cite{Lee2006, Shu2008, Shu2016, Weller2009, Berthelot2011, Medarde2013, Shu2013, Evstigneeva2011, Sankar2014} It is highly desirable to investigate the 2D structure of the Na-layer that reflects the potential field generated by the (Te/Ni)-O honeycomb sublattice and its relationship to the magnetic structure of Ni spins. In particular, due to the diffusive nature of intercalated Na ions, a unique dynamic magneto-phonon coupling of the system could be revealed through the detailed analysis of the Na ion distribution and crystal structure change as a function of temperature. Below we present an integrated synchrotron X-ray and high-resolution neutron powder diffraction study that clearly reveals an in-plane chiral circular pattern within the Na-layers. An inverse Fourier transform (iFT) technique was applied to provide real space information from the diffraction data for a more accurate Na ion site assignment on structure refinement. Comparing to the calculated bond valence sum (BVS) map of Na, the Na-ion diffusion in 2D is found to follow an effective translational path of alternating chirality change.\\
\begin{figure*}
\begin{center}
\includegraphics[width=5.5in]{fig-crystru}
\end{center}
\caption{\label{fig-crystru} (a) The crystal structure of Na$_2$Ni$_2$TeO$_6$ of space group $P6_3/mcm$ with three major Na sites (Na1, Na2, and Na3) assigned following Ref.~\onlinecite{Evstigneeva2011}. (b) The 2D honeycomb network is composed of edge-shared NiO$_6$ with Te (also edge-shared TeO$_6$ octahedron) sitting in each honeycomb center, with Na sites in the neighboring layer shown as circles. The electron density map obtained from the iFT-assisted X-ray diffraction at z=0.25 is shown on the left. The effective magnetic exchange couplings $J_1$ and $J_2$ are indicated.}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=3.5in]{fig-NPD}
\end{center}
\caption{\label{fig-NPD} Observed (crosses) and fitted (solid lines) high-resolution neutron powder-diffraction patterns taken at 300 K, assuming a Hexagonal crystal structure with the $P6_3/mcm$ space group. The differences between the calculated and observed patterns are plotted at the bottom. The solid vertical lines mark the calculated positions of the Bragg reflections of the proposed crystalline structure. }
\end{figure}
\section{Experimental details and theoritical calculations}
The experimental details of the polycrystalline sample preparation and single crystal growth of Na$_2$Ni$_2$TeO$_6$ have been reported in our previous work.~\cite{Sankar2014} The preliminary X-ray diffraction measurements were performed on a Bruker D8 ADVANCE diffractometer employing Cu K$_\alpha$ radiation, and Synchrotron X-ray was performed at NSRRC using BL01C2 beam line. Neutron powder diffraction data between 2 to 450 K were collected on the high-resolution neutron powder diffractometer (HRPD) ECHIDNA [monochromator: Ge (331), $\lambda = 2.439 $ \AA$ $] and the high-intensity powder diffractometer (HIPD) WOMBAT [monochromator: Ge (113), $\lambda = 2.9502 \AA $], installed at the OPAL reactor in ANSTO, Australia. Approximately $\sim$7 grams of the powder sample was loaded into a cylindrical vanadium-can which gave rise to no measurable background diffraction peaks. Neutron scattering experiments on powder and single crystal samples were also conducted using the NG-5 triple-axis spectrometer SPINS at the NIST Center for Neutron Research [$\lambda=4.702 \AA$]. Cooled beryllium filter was used to reduce the higher-order neutron contaminations. The magnetic susceptibility measurement was performed using a SQUID-VSM magnetometer (Quantum Design USA). Chemical analysis was performed using Electron Probe X-Ray Microanalyzer (EPMA) taken with a JEOL JXA-8200 analyzer. Bond valence sum (BVS) was calculated using the program 3DBVSMAPPER.\cite{Sale2012} The nuclear or electron density maps were obtained employing the General Structure Analysis System (GSAS) or FullProf programs. Starting with the profile refinement of neutron or X-ray diffraction pattern, the electron density distribution can be obtained by calculating the inverse Fourier transform of the structure factors. The scattering density $\rho({\bf r}) = \frac{1}{V} \sum_{H}F_{obs}({\bf H}) exp\{-2\pi i(\bf H.\bf r)\}$ was calculated using the program $Fourier$ (GSAS) or $GFourier$ (FullProf), where V is the volume of the unit cell, H is a reciprocal space vector, \textbf{r} is a vector position inside the unit cell, and $F_{obs}({\bf H}) = \vert F({\bf H})\vert exp\{i\phi({\bf H})\}$, where $\vert F({\bf H})\vert$ is the square root of integrated intensities taken directly from the measurement, and the calculated phase $\phi({\bf H})$ derived from the structural model by the GSAS or FullProf program is used to perform the Fourier transform.
Theoretical calculations were performed based on first-principles density functional theory (DFT) with generalized gradient approximation (GGA).\cite{Perdew1996} To describe the electron-electron correlation associated with the 3d states of Ni, the GGA plus on-site coulomb repulsion (GGA +U) calculations are performed with an effective U$_{eff}$ = (U - J) = 6.0 eV.~\cite{Dudarev1998, Jain2011} We have used the frozen-core full-potential projector-augmented wave (PAW) method~\cite{Blochl1994,Kresse1999} as implemented in the Vienna ab initio simulation package (VASP).\cite{Kresse1993} The wave functions are expressed in a plane wave basis set with an energy cutoff of 400 eV, and the self-consistent field energies are converged up to 10$^{-5}$ eV. For our calculations, we have used four metastable magnetic configuraions to estimate the magnetic couplings $J_1$, $J_2$ and $J_3$ of Ni atoms of interatomic distances 3.00194, 5.19951 and 5.56422~\AA (the three shortest interatomic distances) respectively. To explore the magnetic ground state, a supercell of size 4 times the size of primitive cell is considered, and there are 16 Ni atoms in the unit cell. In the present calculations, we used the tetrahedron method with $Bl\ddot{o}chl$ corrections for the Brillouin zone integration with a $\Gamma$-centered Monkhorst-Pack k-point mesh of $(9\times 9\times 7)$. The magnetic couplings $J_1$, $J_2$ and $J_3$ among the Ni atoms are expressed in terms of the spin Heisenberg Hamiltonian, $H=E_0-\sum J_{ij} {\bf S_i\cdot S_j}$, where $J_{ij}$ is the exchange interaction parameter between the Ni atoms at site $i$ and site $j$, and ${\bf S_i}$ (${\bf S_j}$) is the unit vector that represents the direction of the local magnetic moment at site $i~(j)$. For an antiferromagnetic interaction, $J<0$ is assumed, and for a ferromagnetic interaction, $J>0$ is assumed, and the constant E$_0$ contains all of the spin-independent interactions.
\begin{table}
\caption{List of the refined structural parameters of Na$_2$Ni$_2$TeO$_6$ at 300 K, based on the assigned Na positions similar to those reported in the literature by Evstigneeva \textit{et al.}, and the iFT-assisted refinement from this study (lower), where $B_{iso}$ represents the isotropic temperature parameter and M represents the multiplicity. }
\centering
\begin{tabular}{c rrrrrr}
\hline\hline
\multicolumn{7}{c}{\textbf{Na$_2$Ni$_2$TeO$_6$}} \\
\multicolumn{7}{c}{Hexagonal $P6_3/mcm$ space group ( No. 193)} \\
\multicolumn{7}{c}{T=300 K, a = b = 5.2036(3) $ \AA$, c = 11.1387(3)$ \AA$} \\
\hline
Atom & x & y & z & M & $B_{iso}$ ($ \AA$$^2$) & Occupancy\\
\hline
Te & 0 & 0 & 0 & 2a & 0.19(3) & 1.000 \\
Ni & $\frac{2}{3}$ & $\frac{1}{3}$ & 0 & 4c & 0.30(2) & 0.997(3) \\
O & 0.3112(2) & 0.3112(2) & 0.5944(1) & 12b & 0.57(4) & 1.002(3) \\
Na1 & 0.6306(3) & 0.0 & $\frac{1}{4}$ & 6g & 4.98(5) & 0.508(4) \\
Na2 & $\frac{1}{3}$ & $\frac{2}{3}$ & $\frac{1}{4}$ & 4c & 3.12(3) & 0.203(2) \\
Na3 & 0 & 0.0 & $\frac{1}{4}$ & 2a & 2.88(3) & 0.073(1) \\ \\[0.2ex]
\multicolumn{7}{c}{$\chi$$^2$ = 5.654, $R_p$ = 6.92\%, $R_{wp}$ = 8.81\% } \\
\hline\hline
\multicolumn{7}{c}{\textbf{Na$_2$Ni$_2$TeO$_6$}} \\
\multicolumn{7}{c}{Hexagonal $P6_3/mcm$ space group ( No. 193)} \\
\multicolumn{7}{c}{T=300 K, a = b = 5.2039(1) $ \AA$, c = 11.1383(4)$ \AA$} \\
\hline
Atom & x & y & z & M & $B_{iso}$ ($ \AA$$^2$) & Occupancy\\
\hline
Te & 0 & 0 & 0 & 2a & 0.18(5) & 1.000 \\
Ni & $\frac{2}{3}$ & $\frac{1}{3}$ & 0 & 4c & 0.28(4) & 1.003(2) \\
O & 0.3112(2) & 0.3112(2) & 0.5943(1) & 12b & 0.57(6) & 0.992(2) \\
Na1a & 0.3538(4) & 0.0 & $\frac{1}{4}$ & 6g & 2.28(4) & 0.012(2) \\
Na1b & 0.415(3) & 0.072 & $\frac{1}{4}$ & 6g & 2.28(4) & 0.051(3) \\
Na1c & 0.6250(5) & 0.0 & $\frac{1}{4}$ & 6g & 6.74(7) & 0.038(2) \\
Na1d & 0.6921(5) & 0.065 & $\frac{1}{4}$ & 6g & 6.74(7) & 0.165(2) \\
Na1e & 0.6252(2) & 0.1042(3) & $\frac{1}{4}$ & 12j & 0.91(2) & 0.018(1) \\
Na2a & $\frac{2}{3}$ & $\frac{1}{3}$ & $\frac{1}{4}$ & 4c & 1.44(3) & 0.103(3) \\
Na2b & 0.5639(4) & 0.1872(1) & $\frac{1}{4}$ & 12j & 0.91(1) & 0.013(4) \\
Na2c & 0.6881(2) & 0.2190(4) & $\frac{1}{4}$ & 12j & 3.71(2) & 0.014(2) \\
Na3a & 0 & 0.0 & $\frac{1}{4}$ & 2a & 1.60(4) & 0.043(1) \\
Na3b & 0.8570(4) & 0.0 & $\frac{1}{4}$ & 6g & 4.37(3) & 0.012(3) \\ \\[0.2ex]
\multicolumn{7}{c}{$\chi$$^2$ = 3.854, $R_p$ = 5.42\%, $R_{wp}$ = 6.83\% } \\
\hline
\end{tabular}
\end{table}
\section{Results and Discussion}
\subsection{Crystal structure}
The crystalline structure of Na$_2$Ni$_2$TeO$_6$ can be viewed as a 2D honeycomb network that is composed of edge-sharing NiO$_6$ octahedra with TeO$_6$ in each honeycomb center, and the Na monolayer is sandwiched between two (Ni/Te)-O layers with mirrored orientations along the $a$-axis in each unit cell, as shown in Fig.~\ref{fig-crystru}. Based on the X-ray diffraction results, Evstigneeva \textit{et al.} indexed the diffraction pattern with a space group of $P6_3/mcm$ using three crystallographic sites of Na1 at $6g$ (0.35, 0, $\frac{1}{4}$), Na2 at $4c$ ($\frac{1}{3}$, $\frac{2}{3}$, $\frac{1}{4}$), and Na3 at $2a$ (0, 0, $\frac{1}{4}$), with corresponding occupancies of 44\%, 21\%, and 25\%, respectively.~\cite{Evstigneeva2011} In this study, neutron diffraction was applied to elucidate the Na ion positions and their corresponding occupancies more precisely. The high-resolution neutron powder diffraction (NPD) data were refined with the Rietveld method using the General Structure Analysis System (GSAS) program.\cite{Larson2004,Rietveld1969}
The refinements were performed assuming a hexagonal crystal structure with space group $P6_3/mcm$. Figure~\ref{fig-NPD} displays the observed (crosses) and calculated (solid lines) diffraction patterns taken at 300 K, with differences plotted at the bottom. The nuclear density of the Na layer at z = 0.25 is represented with a map that takes a 0.2 $\AA$ thick cut from the 3D real space view, which is re-constructed from the diffraction data via an inverse Fourier transform (iFT), as displayed in Fig.~\ref{fig-NDmap}(a)-(c). It is noted that the original assignments of the Na1, Na2 and Na3 sites do not match the iFT-assisted real space crystal structure of the Na layer perfectly. A better site assignment should reflect the additional Na sites that are revealed by the iFT technique, as indicated by the quintuplet splitting of Na1(a-e), the triplet splitting of Na2(a-c), and the doublet splitting of Na3(a-b). All corresponding occupancies are summarized in the Table I. The newly assigned Na sites are overlaid to match the nuclear density mapping more satisfactorily, as shown in Fig.~\ref{fig-NDmap}(a).
It must be noted that these newly identified Na sites from the iFT-assisted NPD data refinement do not offer additional 3D symmetry breaking from the original space group $P6_3/mcm$. In addition, similar iFT-assisted X-ray diffraction data analysis showing electron density distribution (inset of Fig.~\ref{fig-crystru}(b)) is not sufficiently sensitive to resolve the Na sites of low occupancy, which makes current iFT-assisted NPD data analysis a unique and necessary method for the accurate site refinement of the temperature and potential field-sensitive diffusive Na ions. \\
\begin{figure*}
\begin{center}
\includegraphics[width=6.5in]{fig-NDmap}
\end{center}
\caption{\label{fig-NDmap}(color online) (a) The nuclear density map for Na$_2$Ni$_2$TeO$_6$ at z = 0.25 is overlaid with Na sites obtained from the iFT-assisted NPD data refinement, which indicates additional Na1-quintuplet, Na2-triplet, and Na3-doublet splittings are required for the space group $P6_3/mcm$, and the corresponding site occupancies are summarized in Table I. The corresponding 3D contour maps at layers (b) z = 0.25 and (c) z = 0.75 are compared. The nuclear densities of Na1a-Na1d sites are shown with opposite handedness surrounding the Na2a center between the neighboring Na-layers also. (d) Similar handedness can also be identified with space group $P1$ without any presumed symmetry in the hexagonal system, as shown for the z = 0.25 plane via iFT-assisted data analysis using the same NPD data. }
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=5.5in]{fig-diffusion}
\end{center}
\caption{\label{fig-diffusion} (a) The nuclear density map of the z=0.25 layer cut for Na$_2$Ni$_2$TeO$_6$ at 300 K, obtained using iFT-assisted NPD data via space group $P6_3/mcm$, where the calculated BVS iso-surface (in brown color) of Na ions is overlaid on top. (b) The temperature dependence of the integrated Na1-Na2-Na3 site occupancies. Solid lines are a guide to the eye. }
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=5.5in]{fig-OrderPara}
\end{center}
\caption{\label{fig-OrderPara}(a) CM and ICM AFM peaks observed in the powder sample of Na$_2$Ni$_2$TeO$_6$ below 28 K. (b) Thermal variations of ICM and CM peaks indicate an onsets of $T_N$$\sim$27.5 K. (c) The d$\chi$/dT and C$_p$/T plots also indicate an anomaly near $\sim$27.5 K. }
\end{figure*}
\subsection{Sodium distribution and diffusion}
Following the iFT-assisted NPD data analysis, the experimental nuclear density mapping revealed a circular pattern of well-defined handedness in the Na layer, as illustrated in Fig.~\ref{fig-NDmap}(a)-(c). Chiral circular pattern can be identified from the nuclear density distribution of the Na1(a-e) sites surrounding the Na2a center in circles showing alternating handedness (counter-clockwise and clockwise). The observed 2D chiral pattern has also been confirmed by the electron density mapping using X-ray diffraction with the similar iFT-assisted data analysis [inset of Fig.~\ref{fig-crystru}(b)], wherein the electron density (not the nuclear density) distribution of the unresolved Na1 sites indeed shows an in-plane (2D) chiral pattern similar to that identified by the iFT-assisted NPD data [Fig.~\ref{fig-NDmap}(a)]. Such chiral pattern in 2D does not break the 3D symmetry, but its role in 2D symmetry cannot be dismissed completely. For example, such unique local chiral element is found in chiral molecules distributed in 2D. \cite{Elemans2009} The unique chiral element (or named chiral asymmetry) would reduce the 2D symmetry from \textit{p3m1} (No.14) to \textit{p3} (No.13) within the classification of seventeen 2D (Wallpaper) space groups.
Although a unique potential field reversal is expected between the (Ni/Te)-O layers following the symmetry operation with the space group $P6_3/mcm$, such an in-plane chiral pattern cannot be assigned unambiguously with the parameters of anisotropic B-factor alone, as indicated by the B-factors listed in Table I. It is noted that distinct chiral pattern can also be identified by applying the iFT-assisted NPD data analysis using the primitive space group $P1$ (No.1) without adding presumed symmetry operation for a hexagonal system, as shown in Fig.~\ref{fig-NDmap}(d), i.e., the observed chirality is intrinsic and not an artifact generated by the choice of space group used in the iFT operation.
In view of Na$_2$Ni$_2$TeO$_6$ as a layered compound with Na ions intercalated in the van der Waals (vdW) gap, the nuclear density distribution shown in Fig.~\ref{fig-NDmap} could be interpreted as the time-averaged Na ion distribution influenced by the electric potential field, and the potential field is constructed by the neighboring (Ni/Te)-O layers under thermal fluctuation. This assumption has been supported by the calculated Na ion diffusion path from the molecular dynamic simulations performed earlier.\cite{Sau2015} The iso-surface of the Na bond valence sum (Na-BVS) map shown in Fig.~\ref{fig-diffusion} is calculated from the difference between the experimentally calculated Na-BVS and the ideal valence of +1($\pm$0.2) for the Na ion.\cite{Sale2012} The BVS iso-surface network is continuous across the Na1(c-e) and Na2(a-c) multiplet sites within Na-layer, but leaves out the Na1(a-b) and Na3(a-b) sites completely, as shown in Fig.~\ref{fig-diffusion}(a), which suggests that a Na-ion diffusion path in 2D is established following the Na1(c-e)-Na2(a-c) sites in a circular manner. A similar Na diffusion path has also been reported earlier in Na$_{0.7}$CoO$_2$ and Na$_3$[Ti$_2$P$_2$O$_{10}$F] via the mean-square displacements and thermal ellipsoid calculations respectively.\cite{Medarde2013, Ma2014}
It is instructive to examine the Na distribution as a function of temperature for the Na site occupancies in detail, as shown in Fig.~\ref{fig-diffusion}(b). Na1(c-e) shows nearly twice the integrated occupancy than that of Na2(a-c) at 300 K, but there is an additional weight shift of approximately 2$\%$ of the occupancy from Na2(a-c) to Na1(c-e) between $\sim$220-150 K. On the other hand, the occupancies of Na1(a-b) and Na3(a-b) remain low and constant in the same temperature range. The actual Na diffusion path is thus hinted by the gradual weight shift of the occupancy between Na1(c-e) and Na2(a-c) experimentally, in agreement with the prediction of BVS calculation. In addition, it is suggested that Na diffusion occurs above $\simeq$220 K but falls to the relative ground state below $\simeq$150 K, which is consistent to the freezing phenomenon of Na observed in the layered Na$_x$CoO$_2$ below $\sim$150 K via $^{23}$Na NMR.\cite{Weller2009, Schulze2008}
Since the Na1(c-e)$\leftrightarrow$Na2(a-c) diffusion path (Fig.~\ref{fig-diffusion}) forms closed loops in 2D close packing, a purely rotational diffusion is likely. In order to introduce an effective translational diffusion, the possible purely rotational diffusion can be avoided by coupling the rotation to an orientational diffusion, as demonstrated by the chiral diffusion for a system of rotary nanomotors.\cite{Nourhani2013} The revealed chiral circular pattern in 2D must help on making Na$_2$Ni$_2$TeO$_6$ a compound of high ionic conductivity.
\subsection{Magnetic phase transitions}
The AFM phase transition for Na$_2$Ni$_2$TeO$_6$ was verified with elastic neutron scattering using both powder and single crystal samples through the trace of superlattice peaks as a function of the temperature, as shown in Fig.~\ref{fig-OrderPara}(a). At T=5 K, a commensurate (CM) superlattice peak at (0.5 0 1) and a broad incommensurate (ICM) superlattice peak near $\textbf{k}$=(0.47 0.44 0.28) can be identified in the powder sample. While the observed ICM peak has a shoulder close to a possible CM peak of (0.5 0 0) and the ICM peak cannot be deconvoluted with confidence due to the limited instrumental resolution, the peak intensities are plotted as a function of temperature to show that both CM and ICM peaks have same onset at T$_N$ of $\sim$27.5 K, as shown in Fig.~\ref{fig-OrderPara}(b), which has also been identified by the distinct cusp shown in the $d\chi/dT$ and $C_p$/T plots [see Fig.~\ref{fig-OrderPara}(c)]. Our preliminary single crystal studies indicated that the T$_N$ is extremely sensitive to the excess Na content, e.g., T$_N$$\sim$22K has been identified in a single crystal sample of Na content $\sim$2.16.\cite{Karna} A separate spin-polarized neutron experiment is required to elucidate the relationship among the Na content and ICM/CM spin structures further.
We have tentatively explored the CM superlattice peak at (0.5 0 1) using a separate single crystal sample, as shown by the H-L contour plot of the (0.5 0 1) peak intensity [inset of Fig.~\ref{fig-OrderPara}(a)] at 5 K. The critical exponent of $\beta$ = 0.202 fitted from the powder sample ICM peak intensity with (1-T/T$_N$)$^{2\beta}$ [Fig.~\ref{fig-OrderPara}(b)] suggests the 2D nature of the observed phase transition, which is consistent with the evolution of a short-range exchange coupling prior to the 3D long-range spin ordering, as also revealed by the broad $\chi$(T) peak near $\sim$30-40 K right above T$_N$ (lower inset of Fig.~\ref{fig-lattice}).
\subsection{Magnetic structure calculations}
\begin{figure*}
\begin{center}
\includegraphics[width=5.5in]{fig-SpinStru}
\end{center}
\caption{\label{fig-SpinStru}(color online)(a) The proposed commensurate AFM spin structure in 2D, where solid and empty circles represent spins up and down. Zigzag chains with doubled periodicity along the $a$-direction is indicated by the dashed lines, which can also be viewed as antiferromagnetically coupled FM zigzag chains extended along the (110) direction. (b)-(e) Four possible spin configurations for the theoretical ground state calculation. The proposed spin configuration shown in (b) is found to be identical and can be described equivalently using spin dimers in linear chain along the $a$- or $b$-direction after a 60$^\circ$ axis transformation within the $ab$-plane, as shown in (a). }
\end{figure*}
Judging from the $\chi_{\|c}$(T$<$T$_N$) drop to indicate the on-site spin anisotropy along the $c$-direction,\cite{Sankar2014} we have considered four possible magnetic configurations of the ground state to evaluate the three spin exchange parameters based on the first-principle density functional theory with a generalized gradient approximation (GGA),\cite{Perdew1996} including an antiferromagnetically coupled FM zigzag chain structure (AF1), a $q$=0 spin configuration for each honeycomb subnetwork (AF2), an A-type AFM spin structure similar to that found in Na$_{0.82}$CoO$_2$ (AF3),\cite{Bayrakci2005} and a FM spin structure were calculated, as illustrated in Fig.~\ref{fig-SpinStru} (b)-(e). The explicit relations between the energy and exchange coupling parameters for the proposed configurations are: $E_{AF1}$ = $E_0$ - 8$J_1$ + 16$J_2$ + 16$J_3$, $E_{AF2}$ = $E_0$ + 8$J_1$ - 16$J_2$ + 16$J_3$, $E_{AF3}$ = $E_0$ - 24$J_1$ - 48$J_2$ + 16$J_3$, and $E_{FM}$ = $E_0$ - 24$J_1$ - 48$J_2$ - 16$J_3$. Considering the in-plane nearest neighbor couplings of $J_1$ (Ni-O-Ni) through the superexchange route and the next nearest neighbor coupling $J_2$ (Ni-O...O-Ni) through the super-superexchange route [Fig.~\ref{fig-crystru}(b)], and an inter-plane coupling of $J_3$, the magnetic ground state is calculated to be the spin configuration of the antiferromagnetically coupled FM zigzag chains (AF1), as illustrated in Fig.~\ref{fig-SpinStru}(a). The calculated values for the ground state are found to be $E_0$ = -451.37 eV/unit cell, $J_1$ = 0.15 meV, $J_2$ = -1.16 meV, and $J_3$ = -0.08 meV. As a check for the validity of calculations, a mean-field estimate of Weiss temperature $\Theta$ based on the calculated exchange coupling parameters gives $\Theta$$\sim$$-$26 K, being in good agreement with the experimental value of $-$32 K obtained from the Curie-Weiss law fitting.\cite{Anderson}
It is noted that the definition of a Ni S=1 spin chain in the 2D honeycomb network is ambiguous when three equivalent zigzag chains can be defined along the (1 0 0), (0 1 0) and (1 1 0) directions of a hexagonal system. According to the experimental and theoretical informations collected to date, including the identification of the AFM peak ($\frac{1}{2}$ 0 1), the confirmed 3D AFM spin ordering, the spin anisotropy along the $c$-direction,~\cite{Sankar2014} and the theoretically calculated ground state configurations [Fig.~\ref{fig-SpinStru}(b)-(e)] of dominant next nearest neighbor AFM coupling $J_2$ with weak nearest neighbor FM coupling $J_1$, a commensurate AFM spin structure is proposed and shown in Fig.~\ref{fig-SpinStru}(a). This proposed spin structure satisfies all of the experimental and theoretical conditions for the confirmed AFM long-range spin ordering, with a doubled lattice size along the equivalent $a$- and $b$-directions of the hexagonal lattice.
Although the AFM long-range spin ordering is expected to coincide with the lattice size doubling along the $c$-direction of opposite handedness in the Na-layer, strong ICM peaks were identified in the powder sample [Fig.~\ref{fig-OrderPara}(a)], which could have resulted from the incommensurate modulation of the AFM spin arrangement of a 3D helical spin ordering. While the chiral symmetry is closely related to the Coulomb field that is revealed by the nuclear density distribution of the Na-layer, the observed incommensurability could be closely related to the oxygen and/or Na nonstoichiometry. Preliminary chemical analysis using electron probe microanalysis (EPMA) suggests that the studied powder sample has an oxygen vacancy level about $\sim$0.02 per formula unit. A careful neutron study on a series of single crystal samples with controlled oxygen and Na nonstoichiometry is expected and in progress. \\
\subsection{Spin-phonon-electronic coupling}
\begin{figure}
\begin{center}
\includegraphics[width=3.5in]{fig-lattice}
\end{center}
\caption{\label{fig-lattice} The temperature dependence of the lattice parameters \textit{a} and \textit{c} for Na$_2$Ni$_2$TeO$_6$. The linear temperature dependence is found to be above $\sim$220 K, and the ZTE phenomenon is found to be below $\sim$ 100 K. The upper left inset shows anomalies in the specific heat and resistivity of Na$_2$Ni$_2$TeO$_6$ near $\sim$220 K. The lower right inset shows the temperature dependence of magnetic susceptibilities for powder Na$_2$Ni$_2$TeO$_6$ measured in the applied field of H$_a$ = 1T. The solid line represents the Curie-Weiss law fitting. }
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=2.5in]{fig-strain}
\end{center}
\caption{\label{fig-strain} A portion of the fitted (a) X-ray and (b) neutron diffraction patterns showing $l$$\neq$0 peaks broadening and $l$=0 peaks with FWHM within the instrumental resolution. All Cu-$K_{\alpha 2}$ for the X-ray has been filtered out. The calculated corresponding strain is tabulated in Table II.}
\end{figure}
The lattice parameters $a$ and $c$ for Na$_2$Ni$_2$TeO$_6$ measured as a function of the temperature are shown in Fig.~\ref{fig-lattice}. No structural symmetry change or lattice distortion could be identified from the high-resolution neutron powder diffraction patterns between 3 and 450 K. The thermal expansion was essentially isotropic above $\sim$220 K, as indicated by the linear temperature dependence. The calculated thermal expansion coefficients are $\alpha_a$ = 9.78$\times$ 10$^{-6}$ K$^{-1}$ along the \textit{a}-axis and $\alpha_c$ = 2.62$\times$ 10$^{-5}$ K$^{-1}$ along the \textit{c}-axis for data above $\sim$220 K. The $\alpha_c$ is approximately 2.68 times higher than $\alpha_a$, which indicates that the inter-layer coupling is considerably weaker than that of the intra-layer coupling, as expected for a layered compound with vdW gaps.
The temperature range for the occurrence of the zero thermal expansion (ZTE) phenomenon below $\sim$100 K is found close to the onset that $\chi$(T) starts to deviate from the Curie-Weiss law (lower right inset of Fig.~\ref{fig-lattice}), indicating both the anharmonic phonon contribution and the short-range spin exchange coupling are required as the precursor to the 3D long-range spin ordering below T$_N$$\sim$27.5 K. These findings are consistent with a picture of the dominant spin-phonon coupling under reduced thermal fluctuation. The onset of linear thermal expansion near $\sim$220 K was also found near the temperature range where the Na1/Na2 occupancies start to switch weights [Fig.~\ref{fig-diffusion}(b)], which suggests that the subtle in-plane chirality must be closely related to the electron-phonon coupling indicated by the cusp of specific heat C$_p$ and the onset of resistivity increase near $\sim$220 K (upper left inset of Fig.~\ref{fig-lattice}). Clearly, it is the diffusive nature of the Na ions makes the observation of all these correlated phenomena possible.
\subsection{2D chirality and 3D spin ordering}
\begin{table}[t]
\caption{Strain calculated for various diffraction peaks shown in the neutron diffraction pattern. Strain is estimated using Fullprof suite from the integral peak width $\beta$ and the $d$-spacing as $\frac{1}{2}$$\beta$d. }
\centering
\begin{tabular}{c rrr}
\hline\hline
HKL & ${HG}_o$ & ${HG}_i$ & Strain
\\ [0.5ex]
\hline
(1 0 0) & 0.4025 & 0.4267 & Resolution limited \\
(1 0 2) & 0.4232 & 0.4111 & 21.7251 \\
(3 0 0) & 0.3141 & 0.3390 & Resolution limited \\
(1 1 6) & 0.7227 & 0.3383 & 22.0218 \\
(3 0 2) &0.4909 &0.3498 &11.0855 \\
(1 1 2) &0.4116 &0.3762 &16.8185 \\
(2 1 1) & 0.3426 & 0.3375 &6.7922 \\[1ex]
\hline
\end{tabular}
\label{tab:afm}
\end{table}
We believe that the hidden chirality in Na layer is resulted from the electric potential gradient created by the neighboring Ni-O layers, which is supported by the detected significant (\textit{h, k, $l$$\neq$0}) peak broadening due to strain calculated using FullProf Suite~\cite{Rodriguez1993} in Table II, in contrast to the narrow (\textit{h, k, $l$=0}) peaks which are within the instrumental resolution (Fig.~\ref{fig-strain}). The different type of bonding between the weaker inter-layer van der Waals interaction and the stronger intra-layer covalent bonding could be responsible for the tolerable anisotropic displacement of oxygen atoms under thermal fluctuation, and the detected strain reflects the required potential gradient for the Na ion diffusion following the Fick's second law.\cite{Shu2008} Above all, the chirality displayed by the Na ion distribution is hidden in the regular refinement that considers the statistical atomic position only.
\begin{table}
\caption{ DFT calculations of exchange couplings $J_1$, $J_2$,$J_3$ for three different types of Na occupancy distribution are compared.}
\scalebox{0.9}{
\begin{tabular}{c c c c }
\hline \hline
\multicolumn{4}{c}{\textbf{$J_1$/$J_2$/$J_3$}} \\
\hline
Occupancies &100\%Na2 &62.5\%Na1,37.5\%Na2 & 100\%Na1 \\
\hline
$J_i$ (meV) & 0.15/-1.16/-0.08 & -0.11/-1.12/-0.07 & -0.06/-1.13/-0.05 \\
$J_i$ (K) & 1.70/-13.47/-0.87 & -1.31/-12.96/-0.83 & -0.66/-13.08/-0.54 \\
\hline \hline
\end{tabular}
}
\label{tab-occu}
\end{table}
In order to test the impact of Na chiral distribution on the spin ordering, DFT calculations of exchange couplings $J_1$-$J_2$-$J_3$ for systems having three different site occupancy assignments are compared, from calculation using the refined occupancies of 62.5$\%$Na1/37.5$\%$Na2 to the fully occupied Na2 and Na1, as shown in Table~\ref{tab-occu}. It is found that the dominant $J_2$ is not affected significantly by the occupancy change, but the weak $J_1$ and $J_3$ couplings are reduced accordingly. Based on these simulated calculations, we propose that the observed Na-layer 2D chiral distribution via weighted multiplet splitting cannot affect the magnetic couplings significantly.
While the in-plane antiferromagnetically coupled spin chains [see Fig.~\ref{fig-SpinStru}(a)] may break the mirror symmetry of the honeycomb network via an expected Peierls-like AFM inter-chain coupling, current neutron and X-ray structure analysis failed to detect the expected mirror symmetry breaking, instead, an in-plane chirality was uniquely identified in the Na-layer. It is likely that the 2D chiral distribution pattern identified in the Na-layer may induce the structure relaxation via phonon softening only without breaking the original 3D crystal symmetry.
\section{Conclusions}
In summary, we have demonstrated that the iFT-assisted powder neutron diffraction refinement technique is crucial to provide an accurate site refinement for the diffusive Na ions in the layered material like Na$_2$Ni$_2$TeO$_6$. A chiral distribution pattern hidden in the Na-layer is found to be intimately related to its diffusion behavior and the spin structure of Ni arranged in a honeycomb network.\\
\\
\\
|
\section{Introduction}
Deformations of smooth hypersurfaces provide examples of great interest and importance in the theory of variation of Hodge structures, especially because of the generic Torelli theorem, see \cite{VO2}, Chapter 6. In a recent thesis \cite{Zhao}, Y. Zhao considers deformations of nodal surfaces in the $3$-dimensional complex projective space $\mathbb{P}^3$ and shows that the infinitesimal Torelli theorem still holds.
Let $S=\mathbb{C}[x_0,\cdots, x_n]=\bigoplus_{d=0}^\infty S_d$ be the graded ring of polynomials and let $f\in S_d$ be a homogeneous polynomial of degree $d$. Denote by $X_f: f=0$ the hypersurface in $\mathbb{P}^n$ defined by $f$. Moreover, let
$$
J(f)=(\frac{\partial f}{\partial x_0},\cdots, \frac{\partial f}{\partial x_n})
$$
be the graded ideal generated by the first derivatives of $f$, also called the Jacobian ideal of $f$. We consider the following map
\begin{equation}\label{eq: phi}
\varphi:\quad (S/J(f))_d\to \text{\rm Hom}((S/J(f))_{d-n-1}, (S/J(f))_{2d-n-1}),\qquad [P]\mapsto([Q]\mapsto [PQ]).
\end{equation}
As a matter of fact, Y. Zhao \cite{Zhao} proves the infinitesimal Torelli theorem by showing that the map $\varphi$
is injective when $n=3$ and $X_f$ is a {\bf nodal} surface. This result can be extended to higher dimensional cases.
\begin{thm}\label{main thm}
Assume $n\geq 3$ is an integer and $d\geq n+1$. Let $f\in S$ be a homogeneous polynomial of degree $d$ such that $X_f: f=0$ is a nodal hypersurface in $\mathbb{P}^n$. Then the map $\varphi$ is injective.
\end{thm}
As it is proved in \cite{Zhao}, Chapter 3, Example 3.1.3, $(S/J(f))_d$ parameterizes the equivalence classes of deformations of the pair $(\mathbb{P}^n, X_f)$. Alternatively, let $GL=GL(n+1,\mathbb{C})$ be the general linear group of rank $n+1$. Then $GL$ acts on $S_d$ by coordinate transformations and for any $f\in S_d$, the tangent space at $f$ of the orbit $GL\cdot f$ is given by $J(f)_d$, see \cite{D87}, Chapter 4, Formula (4.16). It follows that $(S/J(f))_d$ can be seen as the set of directions in $S_d$ that are transversal to the orbit $GL\cdot f$ at $f$. In addition, any smooth analytic subset $\mathcal{U}\subseteq S_d$ can be seen as a family of hypersurfaces in $\mathbb{P}^n$. If $f\in\mathcal{U}$ and $T_f\mathcal{U}\cap J(f)_d=\{0\}$, then we call $\mathcal{U}$ an {\bf effective} deformation of $f$. From this point of view, $(S/J(f))_d$ is the maximal set of effective deformations of $f$.
Now let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ and let $n(f)$ be the number of nodes in $X_f$. Then we have a moduli space, denoted by $\mathfrak{B}_f\subseteq S_d$, parameterizing all nodal hypersurfaces in $\mathbb{P}^n$ having exactly $n(f)$ nodes. By the discussion following Corollary 3.8 in \cite{D92}, Chapter 1, we have that $\mathfrak{B}_f$ is a constructible subvariety of $S_d$ and the topological type of $(\mathbb{P}^n, X_g)$ is locally trivial for $g\in \mathfrak{B}_f$. Moreover, for any $g$ lying in the connected component of $\mathfrak{B}_f$ containing $f$, $(\mathbb{P}^n, X_g)$ is topologically equivalent to $(\mathbb{P}^n, X_f)$.
Now assume $\mathcal{U}\subseteq\mathfrak{B}_f$ is a connected {\bf smooth} subvariety and $f\in\mathcal{U}$. For any $g\in\mathcal{U}$, $X_g$ is homeomorphic to $X_f$ by the local topological triviality of the pair $(\mathbb{P}^n,X_g)$. So there is a natural identification $H^{n-1}_0(X_g)\cong H^{n-1}_0(X_f)$, where $H^{n-1}_0(X_g)$ is the primitive cohomology of $X_g$ defined by $H^{n-1}_0(X_g)=\text{Coker}(H^{n-1}(\mathbb{P}^n)\to H^{n-1}(X_g))$. In particular, $\dim H^{n-1}_0(X_g)$ is constant for $g\in\mathcal{U}$.
Moreover, $H^{n-1}_0(X_g)$ has a natural mixed Hodge structure, since $X_g$ is a singular algebraic variety, see \cite{PS}, Part II, Chapter 5. It turns out that $\dim F^{n-1}H^{n-1}_0(X_g)$ and $\dim F^{n-2}H^{n-1}_0(X_g)$ are also constant for $g\in\mathcal{U}$ (in most cases), see Corollary \ref{cor: dimension} below. Thus, we have the following well-defined map
\begin{equation}\label{eq: mP}
\mathcal{P}:\qquad \mathcal{U}\ni g\mapsto (F^{n-1}H^{n-1}_0(X_g), F^{n-2}H^{n-1}_0(X_g))\in\mathcal{F}
\end{equation}
where $\mathcal{F}$ is the corresponding flag manifold of subspaces of $H^{n-1}_0(X_f)$.
By relating the primitive cohomology with the graded pieces of the algebra $S/J(f)$ and applying Theorem \ref{main thm}, we prove the following, as a generalization of \cite{Zhao}, Chapter 3.
\begin{thm}\label{main cor}
Assume $n\geq 3$ is odd or $n\geq 6$ is even. Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$ and let $\mathcal{U}\subseteq \mathfrak{B}_f$ be a smooth subvariety of $\mathfrak{B}_f$ which gives an effective deformation of $X_f$. Then the map $\mathcal{P}$ above is well-defined and the differential $d\mathcal{P}$ is injective at $f$.
\end{thm}
Thus, loosely speaking, the infinitesimal Torelli theorem also holds for nodal hypersurfaces.
Note that for smooth hypersurfaces, the {\bf generic} Torelli theorem holds, see \cite{VO2}, Part II, Chapter 6, Section 6.3.2, and it remains an interesting question whether this is also the case for nodal hypersurfaces. Recall that in the proof of the generic Torelli theorem for smooth hypersurfaces, the essential part is to show that a generic homogeneous polynomial can be reconstructed from its Jacobian ideal, which also holds for nodal hypersurfaces by Theorem 1.1 in \cite{ZW}, because a generic $f$ of degree $d>3$ with the associated hypersurface $X_f$ having a fixed number of nodes is not of Sabastiani-Thom type, which is the only exception for $f$ not to be reconstructed from $J(f)$; another key ingredient in the smooth case is the symmetriser lemma, which is still open for nodal hypersurfaces.\\
The author would like to thank an anonymous referee, whose remarks make the exposition of this paper improved.
\section{Syzygies of the Jacobian ideal}\label{sec: syzygies}
Let $K^\bullet(f)$ be the Koszul complex of $\frac{\partial f}{\partial x_0},\cdots, \frac{\partial f}{\partial x_n}$ with the natural grading $\deg(x_j)=1$ and $\deg(dx_j)=1$:
$$
K^\bullet(f):\quad 0\to \Omega^0\to\Omega^1\to\cdots\to\Omega^{n+1}\to 0
$$
where $\Omega^1=\sum_{i=0}^n Sdx_i$ and $\Omega^p=\bigwedge^p\Omega^1$, and the differentials are given by the wedge product with $df=\sum_{i=0}^n\frac{\partial f}{\partial x_i}dx_i$.
The homogeneous component of the cohomology group $H^n(K^\bullet(f))_{n+r}$ describes the syzygies
$$
\sum_{j=0}^n a_j \frac{\partial f}{\partial x_j}=0
$$
with $a_j\in S_r$ modulo the trivial syzygies generated by
$$
\frac{\partial f}{\partial x_i}\frac{\partial f}{\partial x_j}+\biggl(-\frac{\partial f}{\partial x_i}\biggr)\biggl(\frac{\partial f}{\partial x_j}\biggr)=0,\qquad i<j.
$$
We may restate the main result in \cite{D13} or Theorem 9 in \cite{DS14} in the following form.
\begin{lem}\label{lem: syzygies}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d>2$ and $n\geq 3$, then
$H^n(K^\bullet(f))_m=0$ for any
$$
m\leq \frac{nd-1}{2}.
$$
\end{lem}
Let $f_s\in S_d$ be such that $X_{f_s}: f_s=0$ is a smooth hypersurface. It is well-known that $\dim (S/J(f_s))_k$ depends only on $n,d$ and $k$, see \cite{D87}, Chapter 7, Proposition 7.22. In the introduction part of \cite{D13R}, the following two notions are given:
$$
ct(X_f)=\max\{q\quad:\quad \dim (S/J(f))_k=\dim (S/J(f_s))_k\text{ for all }k\leq q\}
$$
and
$$
mdr(X_f)=\min\{q\quad:\quad H^n(K^\bullet(f))_{q+n}\neq 0\}.
$$
They have the following relation
$$
ct(X_f)=mdr(X_f)+d-2,
$$
see loc. cit.. We have the following.
\begin{lem}\label{lem: dimension}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$ and $n\geq 3$, then
$$
\dim (S/J(f))_k=\dim (S/J(f_s))_k,\quad k=d-n-1, 2d-n-1.
$$
In particular, $\dim (S/J(f))_k$ does not depend on the concrete equation of the polynomial $f$ for $k=d-n-1, 2d-n-1$.
\end{lem}
\begin{proof}
We only need to check that $2d-n-1\leq ct(X_f)$. Indeed, by Lemma \ref{lem: syzygies}, we immediately have
$$
ct(X_f)=mdr(X_f)+d-2\geq\biggl(\frac{nd-1}{2}-n\biggr)+d-2>2d-n-1,
$$
where the last inequality follows from $n\geq 3$ and $d\geq n+1$.
\end{proof}
\subsection{Proof of Theorem \ref{main thm}}
To prove Theorem \ref{main thm}, we first prove the following.
\begin{lem}\label{lem: in}
Assume $X_f: f=0$ is a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$ and $n\geq 3$.
Let $G\in S_t$ such that $t<2d-n-1$ and $Gx_j\in J(f)$ for all $j=0,\cdots, n$, then $G\in J(f)$.
\end{lem}
\begin{proof}
Assume
\begin{equation}\label{eq: 1}
Gx_i=\sum_{k=0}^n H_{ik}\frac{\partial f}{\partial x_k},\quad i=0,\cdots, n,
\end{equation}
with $H_{ik}\in S_{t+2-d}, i,k=0,\cdots, n$, then
$$
0=x_i(x_jG)-x_j(x_iG)=\sum_{k=0}^n(x_iH_{jk}-x_jH_{ik})\frac{\partial f}{\partial x_k}.
$$
Note that
$$
t+3-d+n\leq(2d-n-2)+3+n-d=d+1\leq\frac{nd-1}{2},
$$
so by Lemma \ref{lem: syzygies}, we get $x_iH_{jk}-x_jH_{ik}\in J(f)$ for all $i,j,k=0,\cdots, n$ while all these polynomials have degree $t+3-d<(2d-n-1)+3-d=d-n+2\leq d-1$, so they must all vanish identically; in particular,
$$
x_iH_{jk}-x_jH_{ik}=0,\quad i\neq j
$$
thus, $x_i|H_{ik}$. It follows that $G\in J(f)$ as desired.
\end{proof}
{\it Proof of Theorem \ref{main thm}: }
We first remark that Theorem \ref{main thm} holds when $d=n+1$. In fact, in this case, $J(f)_{d-n-1}=J(f)_0=0$ and $(S/J(f))_{d-n-1}=S_0=\mathbb{C}$ consists of constants. Since $1\in (S/J(f))_{d-n-1}$ and $\varphi([P])(1)=[P]$, one sees easily that $\varphi$ is injective.
Thus, in the sequel of the proof, we will focus on the case $d>n+1$.
Aiming at a contradiction, we assume that there exists $P\in S_d\setminus J(f)_d$ such that $\varphi([P])=0$.
Then there exists a $Q\in S_l, 0\leq l<d-n-1$ such that $PQ\notin J(f)$ and $l$ is chosen to be maximal. By the maximality of $l$, we have $(PQ)x_j\in J(f)$ for all $j=0,\cdots, n$. Note that $PQ\in S_{l+d}$ and $l+d<2d-n-1$, hence by Lemma \ref{lem: in}, $PQ\in J(f)$, contradiction.
\section{Hodge theory for nodal hypersurfaces}\label{sec: hodge}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$ and $n\geq 3$. The cohomology groups under consideration below all have $\mathbb{C}$ as coefficients unless otherwise explicitly pointed out.
By the Lefschetz hyperplane theorem for singular varieties (see \cite{Ha}), we have
$$
H^i(X_f)=H^i(\mathbb{P}^n),\qquad i<n-1,
$$
and
$$
H^{n-1}(\mathbb{P}^n)\to H^{n-1}(X_f)
$$
is injective. Let
$$
H^{n-1}_0(X_f)=\text{Coker}(H^{n-1}(\mathbb{P}^n)\to H^{n-1}(X_f)),
$$
be the primitive cohomology of $X_f$. Then $H^{n-1}_0(X_f)$ admits a mixed Hodge structure. Moreover, let $U_f=\mathbb{P}^n\setminus X_f$ be the complement of $X=X_f$, then $H^n(U_f)$ also admits a mixed Hodge structure and $H^n(U_f)$ and $H^{n-1}_0(X_f)$ are closely related.
\subsection{Relation between $H^*(U_f)$ and $H^*(X_f)$}
Let $X_f^*$ be the smooth locus of $X_f$ and let
$$
H^{n-1}_0(X_f^*)=\text{Coker}(H^{n-1}(\mathbb{P}^n)\to H^{n-1}(X_f^*)).
$$
Then $H^{n-1}_0(X_f^*)$ has a natural mixed Hodge structure. Moreover, as is shown in \cite{D92}, Chapter 6, Corollary 3.11, there is a natural residue isomorphism
\begin{equation}\label{eq: UX}
\overline{R}_f:\quad H^n(U_f)\xrightarrow{\sim} H^{n-1}_0(X_f^*)
\end{equation}
which is also an isomorphism of mixed Hodge structures of type $(-1,-1)$.
Let $i: X_f^*\to X_f$ be the inclusion. We have the naturally induced homomorphisms in cohomology
$$
i^*: H^{n-1}(X_f)\to H^{n-1}(X_f^*)
$$
and
$$
i^*_0: H^{n-1}_0(X_f)\to H^{n-1}_0(X_f^*).
$$
Moreover, $i^*, i^*_0$ are also morphisms of mixed Hodge structures. Our discussion will be divided into two cases, regarding whether $n$ is odd or even.
\subsubsection{Case 1: $n$ is odd }
When $n$ is odd, the variety $X_f$ is a $\mathbb{Q}$-homology manifold, i.e., for any point $x\in X_f$, $H^i(X_f, X_f\setminus\{x\},\mathbb{Q})=\mathbb{Q}$ if $i=2n$ and $0$ otherwise. Moreover, we have the following claim.
\begin{claim}
$i^*$ and $i_0^*$ are both isomorphisms.
\end{claim}
\begin{proof}
Indeed, we have a long exact sequence of mixed Hodge structures with respect to the pair $(X_f,X_f^*)$:
\begin{equation}\label{eq: revise1}
\rightarrow H^{n-1}(X_f,X_f^*)\to H^{n-1}(X_f)\xrightarrow{i^*} H^{n-1}(X_f^*)\to H^{n}(X_f,X_f^*)
\end{equation}
Let $x_i, i=1,\cdots, r$ be all the nodes in $X_f$, then $X_f^*=X_f\setminus\{x_1,\cdots, x_r\}$, and furthermore, by the excision theorem
$$
H^{n-1}(X_f,X_f^*)=\bigoplus_{i=1}^rH^{n-1}(X_f,X_f\setminus\{x_i\})=0,
$$
since $X_f$ is a $\mathbb{Q}$-homology manifold and $n-1\neq 0,2n$ for $n\geq 3$. Similarly, $H^n(X_f,X_f^*)=0$. Thus, it follows from \eqref{eq: revise1} that $i^*$ and $i^*_0$ are both isomorphisms.
\end{proof}
Note that the weights of $H^{n-1}(X_f)$ are $\leq n-1$ since $X_f$ is compact while the weights of $H^{n-1}(X_f^*)$ are $\geq n-1$ since $X_f^*$ is smooth (see \cite{PS}, p. 131, Table 5.1), hence both $H^{n-1}(X_f^*)$ and $H^{n-1}(X_f)$ have pure Hodge structures of weight $n-1$ and it follows from the isomorphism \eqref{eq: UX} that $H^n(U_f)$ has a pure Hodge structure of weight $n+1$.
Let
$$
R_f=(i_0^*)^{-1}\circ\overline{R}_f:\quad H^n(U_f)\to H^{n-1}_0(X_f).
$$
Then $R_f$ is an isomorphism of mixed Hodge structures of type $(-1,-1)$. It follows that we have isomorphisms
$$
R_f:\quad F^pH^n(U_f)\xrightarrow{\sim}F^{p-1}H^{n-1}_0(X_f)
$$
for all $p$. In particular, there are isomorphisms
\begin{equation}\label{hodge1}
R_f:\quad Gr_F^{p+1}H^n(U_f)\xrightarrow{\sim}Gr_F^pH^{n-1}_0(X_f),\quad p=n-1,n-2.
\end{equation}
\subsubsection{Case 2: $n$ is even }
When $n$ is even, $X_f$ is no longer a $\mathbb{Q}$-homology manifold. However, there is still an explicit description of the relations between $H^n(U_f)$ and $H^{n-1}_0(X_f)$. Note that in this case $H^{n-1}(\mathbb{P}^n)=0$ and thus
$$
H^{n-1}_0(X_f)=H^{n-1}(X_f),\quad H^{n-1}_0(X_f^*)=H^{n-1}(X_f^*)
$$
and $i^*=i^*_0$. Moreover, there exists an exact sequence of mixed Hodge structures
\begin{equation}\label{eq: exact}
\cdots\to H^{n-1}(X_f,X_f^*)\to H^{n-1}(X_f)\xrightarrow{i^*} H^{n-1}(X_f^*)\to H^n(X_f,X_f^*)\to\cdots.
\end{equation}
To make use of this exact sequence, we first give the following claim.
\begin{claim}
For $k=n-1,n$, $H^k(X_f,X_f^*)$ has a pure Hodge structure of type $(\rho_k,\rho_k)$ for some $\rho_k\in\mathbb{N}$.
\end{claim}
\begin{proof}
Let $a_1,\cdots, a_m$ be the nodes in $X_f$ and $B_i\ni a_i, i=1,\cdots,m$ be a small ball in $\mathbb{P}^n$ around $a_i$ such that $B_i\cap B_j=\emptyset$ for $i\neq j$.
By the excision theorem and conic structure theorem (see \cite{D92}, Chapter 1, Theorem 5.1),
$$
H^k(X_f,X_f^*)=\bigoplus_{i=1}^mH^k(B_i\cap X_f, B_i\cap X_f\setminus\{a_i\})\simeq\bigoplus_{i=1}^m H^{k-1}(K_i),\quad k=n-1,n
$$
where $K_i$ is the link of $X_f$ around $a_i$ ($i=1,\cdots, m$).
For each $i$, $K_i$ has the homotopy type of the unit sphere bundle of tangent bundle of $S^{n-1}$. Indeed, locally around $a_i$, $X_f$ is defined as $z_1^2+\cdots+z_n^2=0$, where $(z_1,\cdots, z_n)$ is the local coordinate system of $\mathbb{P}^n$ centered at $a_i$. Then $K_i$ can be described as
$$
K_i=\{(z_1,z_2,\cdots,z_n)\in\mathbb{C}^n\quad :\quad \sum_{j=1}^nz_j^2=0,\ \text{and}\ \sum_{j=1}^n|z_j|^2=\epsilon^2\ \}
$$
where $\epsilon>0$ is small. Let
$$
z_j=\frac{\epsilon}{\sqrt{2}}(v_j+\sqrt{-1}w_j),\quad j=1,\cdots,n
$$
and
$$
v=(v_1,\cdots, v_n), w=(w_1,\cdots, w_n),
$$
then
$$
K_i=\{(v,w)\in\mathbb{R}^n\times\mathbb{R}^n\quad :\quad |v|^2=|w|^2=1\ \text{and}\ \langle v,w\rangle=0\ \},
$$
which is the unit sphere bundle of tangent bundle of $S^{n-1}$.
It follows that
$$
H^{k-1}(K_i)=\mathbb{C},\quad k=n-1,n.
$$
Note also that $H^{k-1}(K_i)=H^k(B_i\cap X_f,B_i\cap X_f\setminus\{a_i\})=H^k(X_f,X_f\setminus\{a_i\})$ admits a natural mixed Hodge structure. In particular,
$$
1=\dim H^{k-1}(K_i)=\sum_{w\in\mathbb{N}}\dim Gr^W_wH^{k-1}(K_i)=\sum_{w\in\mathbb{N}}\sum_{p+q=w}\dim (Gr^W_wH^{k-1}(K_i))^{p,q}
$$
where $Gr^W_wH^{k-1}(K_i)$ is a pure Hodge structure of weight $w$ and
$$
Gr^W_wH^{k-1}(K_i)=\bigoplus_{p+q=w}(Gr^W_wH^{k-1}(K_i))^{p,q}
$$
is the Hodge decomposition. By the Hodge symmetry, we have
$$
(Gr^W_wH^{k-1}(K_i))^{p,q}=\overline{(Gr^W_wH^{k-1}(K_i))^{q,p}}.
$$
It follows that there exists $\rho_{k,i}\in\mathbb{N}$ such that
$$
Gr^W_wH^{k-1}(K_i)=0,\qquad w\neq 2\rho_k
$$
and
$$
(Gr^W_{2\rho_{k,i}}H^{k-1}(K_i))^{p,q}=0,\qquad p\neq q
$$
and
$$
\dim (Gr^W_{2\rho_{k,i}}H^{k-1}(K_i))^{\rho_{k,i},\rho_{k,i}}=1.
$$
In particular, $H^{k-1}(K_i)$ is pure of type $(\rho_{k,i},\rho_{k,i})$.
Note that the mixed Hodge structure on $H^{k-1}(K_i)$ depends only on the local structure of $X_f$ around $a_i$ (see \cite{Df83}, Theorem 3.4). Since all the $a_i$'s are nodes, $H^{k-1}(K_i)$ is naturally isomorphic to $H^{k-1}(K_j)$ as mixed Hodge structures for any $i,j$, hence there exists $\rho_k\in\mathbb{N}$ such that
$$
\rho_{k,1}=\rho_{k,2}=\cdots=\rho_{k,m}=\rho_k,
$$
and thus $H^k(X_f, X_f^*)$ is pure of type $(\rho_k,\rho_k)$ for $k=n-1,n$.
\end{proof}
By Proposition (C28) in \cite{D92}, Appendix C (see also \cite{Df83}, Proposition 3.8) , it follows that $2\rho_{n-1}\leq n-2$. Thus,
$$
Gr_F^pH^{n-1}(X_f,X_f^*)=0,\quad p=n-2,n-1.
$$
Moreover, by the discussions above Example 3.18 in \cite{D92}, Chapter 6, $H^{n-1}(K_i)$ has weight $n$, namely, $2\rho_n=n$, and thus for $n\geq 6$
$$
Gr_F^pH^n(X_f,X_f^*)=0,\quad p=n-2,n-1.
$$
Therefore, it follows from \eqref{eq: exact} that we have an isomorphism
$$
i_0^*:\quad Gr_F^{n-1}H^{n-1}_0(X_f)=F^{n-1}H^{n-1}_0(X_f)\xrightarrow{\sim}Gr_F^{n-1}H^{n-1}_0(X_f^*)=F^{n-1}H^{n-1}_0(X_f^*)
$$
for $n\geq 4$. Furthermore, we have isomorphisms
$$
i_0^*:\quad Gr_F^{n-2}H^{n-1}_0(X_f)\xrightarrow{\sim}Gr_F^{n-2}H^{n-1}_0(X_f^*)
$$
and
$$
i_0^*:\quad F^{n-2}H^{n-1}_0(X_f)\xrightarrow{\sim}F^{n-2}H^{n-1}_0(X_f^*)
$$
for $n\geq 6$; but for $n=4$, we only have injections
$$
i_0^*:\quad Gr_F^{n-2}H^{n-1}_0(X_f)\hookrightarrow Gr_F^{n-2}H^{n-1}_0(X_f^*).
$$
and
$$
i_0^*:\quad F^{n-2}H^{n-1}_0(X_f)\hookrightarrow F^{n-2}H^{n-1}_0(X_f^*).
$$
Using the residue isomorphism \eqref{eq: UX}, we denote
$$
F^{n-1}(U_f,X_f)=\overline{R}_f^{-1}\biggl(i_0^*(F^{n-2}H^{n-1}_0(X_f))\biggr)\subseteq F^{n-1}H^n(U_f)
$$
for $n\geq 4$ (and $n$ is even). Then clearly, $F^{n-1}(U_f,X_f)=F^{n-1}H^n(U_f)$ for $n\geq 6$.
We still denote by $\overline{R}_f$ its restriction to $F^{n-1}(U_f,X_f)$. Then
$$
i_0^*:\quad F^{n-2}H^{n-1}_0(X_f)\to \overline{R}_f(F^{n-1}(U_f,X_f))
$$
is an isomorphism and we have an isomorphism
$$
R_f=(i_0^*)^{-1}\circ\overline{R}_f:\quad F^{n-1}(U_f,X_f)\xrightarrow{\sim} F^{n-2}H^{n-1}_0(X_f).
$$
\subsubsection{Conclusion }
In conclusion, no matter whether $n$ is even or odd, we always have isomorphisms
\begin{equation}\label{eq: hodgen-1}
R_f:\quad Gr_F^nH^n(U_f)\xrightarrow{\sim}Gr_F^{n-1}H^{n-1}_0(X_f)
\end{equation}
and
\begin{equation}\label{eq: hodgen-2}
R_f:\quad F^{n-1}(U_f,X_f)/F^nH^n(U_f)\xrightarrow{\sim}Gr_F^{n-2}H^{n-1}_0(X_f)
\end{equation}
where $F^{n-1}(U_f,X_f)=\overline{R}_f^{-1}\biggl(i_0^*(F^{n-2}H^{n-1}_0(X_f))\biggr)$ is a subspace of $F^{n-1}H^n(U_f)$ containing $F^nH^n(U_f)$; and $R_f=(i_0^*)^{-1}\circ\overline{R}_f$.
\subsection{Cohomology of $X_f$}
Denote by
$$
\Omega=\sum_{i=0}^n(-1)^ix_idx_0\wedge\cdots\wedge dx_{i-1}\wedge\widehat{dx_i}\wedge dx_{i+1}\wedge\cdots\wedge dx_n
$$
where $\widehat{(\ )}$ means that the term is omitted. As is shown in \cite{D92}, Chapter 6, any cohomology class in $F^pH^n(U_f)$ can be represented by a form
$$
\omega(h)=\frac{h\Omega}{f^{n-p+1}}
$$
with $h\in S_{(n-p+1)d-n-1}$. Hence, by \eqref{eq: hodgen-1}, we see that any element in $Gr_F^{n-1}H^{n-1}_0(X_f)$ can be represented by
$$
R_f([\frac{h_1\Omega}{f}])
$$
with $h_1\in S_{d-n-1}$ and similarly, by \eqref{eq: hodgen-2}, any element in $Gr_F^{n-2}H^{n-1}_0(X_f)$ can be represented by
$$
R_f([\frac{h_2\Omega}{f^2}])
$$
with $h_2\in S_{2d-n-1}$.
Such results agree with \cite{DSW}, Theorem 2.2, where the following formulae are given
$$
Gr_F^nH^n(U_f)=(S/J(f))_{d-n-1},\quad Gr_F^{n-1}H^n(U_f)=(S/J(f))_{2d-n-1},
$$
for $n>3$ and for $n=3$,
$$
Gr_F^nH^n(U_f)=(S/J(f))_{d-n-1},\quad Gr_F^{n-1}H^n(U_f)=(I(f)/J(f))_{2d-n-1},
$$
where $I(f)$ is the saturation of $J(f)$, which is also equal to the radical of $J(f)$ for a nodal hypersurface (see \cite{D13R}, Remark 2.2).
Putting all the discussions above in this section together, we obtain the following.
\begin{prop}\label{prop: hodge}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$. Then
\begin{enumerate}[\rm(i)]
\item when $n\geq 3$, there is an isomorphism
$$
\Lambda_f:\quad (S/J(f))_{d-n-1}\to Gr_F^{n-1}H^{n-1}_0(X_f),\quad \Lambda_f(h_1)=R_f([\frac{h_1\Omega}{f}]),
$$
\item when $n>4$, there is an isomorphism
$$
\Lambda_f:\quad (S/J(f))_{2d-n-1}\to Gr_F^{n-2}H^{n-1}_0(X_f),\quad \Lambda_f(h_2)=R_f([\frac{h_2\Omega}{f^2}]),
$$
\item when $n=3$, there is an isomorphism
$$
\Lambda_f:\quad (I(f)/J(f))_{2d-n-1}\to Gr_F^{n-2}H^{n-1}_0(X_f),\quad \Lambda_f(h_2)=R_f([\frac{h_2\Omega}{f^2}]),
$$
\item when $n=4$, there is an isomorphism
$$
\Lambda_f:\quad S'/J(f)_{2d-n-1}\to Gr_F^{n-2}H^{n-1}_0(X_f),\quad \Lambda_f(h_2)=R_f([\frac{h_2\Omega}{f^2}]),
$$
where $S'\subseteq S_{2d-n-1}$ is a vector subspace containing $J(f)_{2d-n-1}$ obtained via
$$
S'/J(f)_{2d-n-1}=\omega^{-1}(F^{n-1}(U_f,X_f)/F^nH^n(U_f))
$$
where $\omega$ is the isomorphism
$$
\omega:\quad (S/J(f))_{2d-n-1}\to Gr_F^{n-1}H^n(U_f),\qquad \omega(h_2)=[\frac{h_2\Omega}{f^2}]
$$
established in \cite{DSW}, Theorem 2.2, and $F^{n-1}(U_f,X_f)$ is obtained in \eqref{eq: hodgen-2}.
\end{enumerate}
In all the formulae above, $R_f$ denotes the residue map.
\end{prop}
As a corollary, we have the following.
\begin{cor}\label{cor: dimension}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$. Then
\begin{enumerate}[\rm(i)]
\item if $n\geq 3$, the dimension
$$
\dim F^{n-1}H^{n-1}_0(X_f)
$$
depends only on $n,d$.
\item if $n\geq 3$ is odd or $n\geq 6$ is even, the dimension
$$
\quad \dim F^{n-2}H^{n-1}_0(X_f)
$$
depends only on $n,d$ and possibly the number of nodes in $X_f$.
\end{enumerate}
\end{cor}
\begin{proof}
Note that
$$
\dim F^{n-1}H^{n-1}_0(X_f)=\dim Gr_F^{n-1}H^{n-1}_0(X_f)
$$
and
$$
\dim F^{n-2}H^{n-1}_0(X_f)=\dim Gr_F^{n-1}H^{n-1}_0(X_f)+\dim Gr_F^{n-2}H^{n-1}_0(X_f).
$$
If $n>4$, the results follow from Proposition \ref{prop: hodge} and Lemma \ref{lem: dimension}, and the dimensions depend only on $n,d$.
When $n=3$, $X_f$ is a $\mathbb{Q}$-homology manifold and the Hodge numbers of $X_f$ depend only on $n,d$ and the number of nodes in $X_f$, see also \cite{D96}.
\end{proof}
\section{Variations of mixed Hodge structures}\label{sec: var}
Let $X_f: f=0$ be a nodal hypersurface in $\mathbb{P}^n$ of degree $d\geq n+1$. When $n$ is odd, assume $n\geq 3$ while when $n$ is even, assume $n\geq 6$.
\subsection{Topological triviality}
Recall that $\mathfrak{B}_f\subseteq S_d$ parameterizes all nodal hypersurfaces with the same number of nodes as $X_f$. Let $\mathcal{U}\subseteq\mathfrak{B}_f$ be a contractible smooth subvariety containing $f$ such that it gives an effective deformation for $X_f$. Set
$$
\mathfrak{X}_{\mathcal{U}}=\{(x,g)\in\mathbb{P}^n\times\mathcal{U}\quad:\quad x\in X_g\quad \}
$$
which can be seen as the union of all nodal hypersurfaces parameterized by $\mathcal{U}$.
Then by the First Thom Isotopy Lemma (see \cite{D92}, Chapter 1, Section 3), there is a homeomorphism $\Phi$ satisfying the following commutative diagram
$$
\xymatrix{
(\mathbb{P}^n\times\mathcal{U},\mathfrak{X}_{\mathcal{U}})\ar[rr]^{\Phi}\ar[rd]_{p_1} & & (\mathbb{P}^n, X_f)\times\mathcal{U}\ar[ld]^{p_2}\\
& \mathcal{U} &
}
$$
where $p_1,p_2$ are natural projections. In fact, $\Phi$ can be obtained by integrating some well-controlled stratified vector field; for a proof, see \cite{Ma}. From now on, we fix such a homeomorphism.
In particular, for any $g\in\mathcal{U}$, there is a canonical homeomorphism $\Phi_g:\mathbb{P}^n\to\mathbb{P}^n$, which induces homeomorphisms $\Phi_{g,X}: X_f\to X_g$ and $\Phi_{g,U}: U_f\to U_g$ with $\Phi_f=\text{\rm Id}$.
Moreover, we have an induced isomorphism of groups
$$
\Phi_{g,X}^*:\quad H^{n-1}_0(X_g)\xrightarrow{\sim}H^{n-1}_0(X_f).
$$
Hence $\dim H^{n-1}_0(X_g)$ is constant for $g\in\mathcal{U}$.
In addition, by Corollary \ref{cor: dimension}, under our assumption on $n$, the dimensions
$$
\dim F^{n-1}H^{n-1}_0(X_g), \quad \dim F^{n-2}H^{n-1}_0(X_g)
$$
are constant with respect to $g\in\mathcal{U}$.
Via the identification $\Phi_{g, X}^*: H^{n-1}_0(X_g)\xrightarrow{\sim} H^{n-1}_0(X_f)$, it follows that $(F^{n-1}H^{n-1}_0(X_g), F^{n-2}H^{n-1}_0(X_g))$ can be identified with $(\Phi_{g,X}^*F^{n-1}H^{n-1}_0(X_g), \Phi_{g,X}^*F^{n-2}H^{n-1}_0(X_g))$, which are two subspaces of $H^{n-1}_0(X_f)$ of fixed dimension. Therefore, we have the well-defined map as in \eqref{eq: mP}
\begin{equation*}
\mathcal{P}:\qquad \mathcal{U}\ni g\mapsto (\Phi_{g,X}^*F^{n-1}H^{n-1}_0(X_g), \Phi_{g,X}^*F^{n-2}H^{n-1}_0(X_g))\in\mathcal{F}
\end{equation*}
where $\mathcal{F}$ is the following flag manifold
\begin{eqnarray*}
\mathcal{F}&=&\{(E_1,E_2)\quad:\quad E_1\subseteq E_2\text{ are vector subspaces of }H^{n-1}_0(X_f)\text{ and }\\
& &\dim E_1=\dim F^{n-1}H^{n-1}_0(X_f)\text{ and }\dim E_2=\dim F^{n-2}H^{n-1}_0(X_f)\}.
\end{eqnarray*}
When $n$ is odd, all the Hodge numbers of $X_g$ are constant for $g\in\mathcal{U}$, and $\mathcal{P}$ is just two components of the period map in the theory of variation of Hodge structures, see \cite{VO1}, Part III, Chapter 10.
\subsection{Infinitesimal deformation}
Now we consider the differential of $\mathcal{P}$. Note that a component of $d\mathcal{P}_f$ is the map
$$
d\mathcal{P}_f:\quad T_f\mathcal{U}\to \text{\rm Hom}(F^{n-1}H^{n-1}_0(X_f), H^{n-1}_0(X_f)/F^{n-1}H^{n-1}_0(X_f));
$$
for the properties of tangent spaces of flag manifolds, we refer to \cite{VO1}, Part III, Chapter 10 and for analogous treatments for smooth hypersurfaces, see \cite{VO2}, Part II, Chapter 6. Recall that Proposition \ref{prop: hodge} implies that any element in $F^{n-1}H^{n-1}_0(X_f)$ is of the form
$$
\omega(h_1)=R_f([\frac{h_1\Omega}{f}]).
$$
The following holds.
\begin{lem}\label{lem: dp}
For $h\in T_f\mathcal{U}\subseteq S_d$, we have
$$
d\mathcal{P}_f(h)(\omega(h_1))=d\mathcal{P}_f(h)\biggl(R_f([\frac{h_1\Omega}{f}])\biggr)=R_f([-\frac{hh_1\Omega}{f^2}]) \Mod F^{n-1}H^{n-1}_0(X_f)
$$
\end{lem}
Its proof is a little lengthy and we postpone it to the end of this section; instead, we first derive Theorem \ref{main cor} from Lemma \ref{lem: dp}.
\subsection{Proof of Theorem \ref{main cor} }
From Lemma \ref{lem: dp} and Proposition \ref{prop: hodge}, the image of $d\mathcal{P}_f$ is contained in
\begin{align*}
&\text{\rm Hom}(F^{n-1}H^{n-1}_0(X_f), F^{n-2}H^{n-1}_0(X_f)/F^{n-1}H^{n-1}_0(X_f))\\
=&\text{\rm Hom}(Gr_F^{n-1}H^{n-1}_0(X_f), Gr_F^{n-2}H^{n-1}_0(X_f)).
\end{align*}
Moreover, we get the following commutative diagram
\begin{equation}\label{eq: square}
\xymatrix{
T_f\mathcal{U}\ar[r]^--{d\mathcal{P}_f}\ar[d]^{i_1} & \text{\rm Hom}(Gr_F^{n-1}H^{n-1}_0(X_f), Gr_F^{n-2}H^{n-1}_0(X_f))\ar[d]^{i_2}\\
(S/J(f))_d\ar[r]^--{\varphi} & \text{\rm Hom}((S/J(f))_{d-n-1}, (S/J(f))_{2d-n-1})
}
\end{equation}
where $\varphi$ is given in \eqref{eq: phi}. $i_1$ is the composite $T_f\mathcal{U}\subseteq S_d\to S_d/J(f)_d$, which is injective since $\mathcal{U}$ is an effective deformation. $i_2$ is defined as follows: for $\eta\in\text{\rm Hom}(Gr_F^{n-1}H^{n-1}_0(X_f), Gr_F^{n-2}H^{n-1}_0(X_f))$ and $h_1\in(S/J(f))_{d-n-1}$,
$$
i_2(\eta)(h_1)=-\Lambda_f^{-1}\biggl(\eta\bigl(\Lambda_f(h_1)\bigr)\biggr),
$$
where $\Lambda_f$ is the isomorphism given in Proposition \ref{prop: hodge}.
By Theorem \ref{main thm}, $\varphi$ is injective, hence $\varphi\circ i_1$ is injective. Thus it follows from \eqref{eq: square} that $d\mathcal{P}_f$ is injective, hence Theorem \ref{main cor} follows.
\begin{rk}
The result is probably also true for $n=4$. We exclude this case because we do not know whether the dimension $\dim F^{n-2}H^{n-1}_0(X_g)$ or equivalently $\dim Gr_F^{n-2}H^{n-1}_0(X_g)$ is constant for $g\in\mathcal{U}$ in this case.
\end{rk}
\subsection{Proof of Lemma \ref{lem: dp}} The proof is almost the same as that in \cite{VO2}, Part II, Chapter 6 where variations of smooth hypersurfaces are considered. However, to avoid any possible confusion, we give the details here.
From the topological triviality of the family $X_g, g\in\mathcal{U}$, it follows that there exists a small contractible neighbourhood $\mathcal{N}\ni f$ in $\mathcal{U}$, such that for any $g\in \mathcal{N}$, $X_g$ is a deformation retract of
$$
\mathcal{X}_{\mathcal{N}}:=\bigcup_{g\in\mathcal{N}}X_g\subseteq\mathbb{P}^n.
$$
Set
$$
U_{\mathcal{N}}=\mathbb{P}^n\setminus\ X_{\mathcal{N}}.
$$
Then $U_{\mathcal{N}}$ is a deformation retract of $U_g$ for every $g\in\mathcal{N}$. Let for $g\in\mathcal{N}$
$$
\tau_g:\quad U_{\mathcal{N}}\hookrightarrow U_g
$$
be the natural inclusion, then the induced homomorphism in cohomology
$$
\tau_g^*:\quad H^n(U_g)\to H^n(U_{\mathcal{N}})
$$
is an isomorphism.
The differential $d\mathcal{P}_f$ can be computed as follows:
for any $h\in T_f\mathcal{U}\subseteq S_d$, choose a curve $g(t): (-\epsilon,\epsilon)\to\mathcal{N}\subseteq\mathcal{U}$ such that $g(0)=f$ and $\frac{d g}{dt}(0)=h$. For any element in $F^{n-1}H^{n-1}_0(X_f)$ of the form
$$
\omega(h_1)=R_f([\frac{h_1\Omega}{f}]),
$$
let
$$
\omega_t(h_1)=R_{g(t)}([\frac{h_1\Omega}{g(t)}])
$$
give an element of $F^{n-1}H^{n-1}_0(X_{g(t)})$. Then
$$
d\mathcal{P}_f(h)(\omega(h_1))=\frac{d}{d t}\biggl|_{t=0}\Phi_{g(t),X}^*(\omega_t(h_1))\quad\Mod F^{n-1}H^{n-1}_0(X_f).
$$
We have
\begin{eqnarray*}
\frac{d}{d t}\biggl|_{t=0} \Phi_{g(t),X}^*(\omega_t(h_1))&=&\frac{d}{dt}\biggl|_{t=0}R_f\biggl(\Phi_{g(t),U}^*([\frac{h_1\Omega}{g(t)}])\biggr)\\
&=&R_f\biggl(\frac{d}{dt}\biggl|_{t=0}\Phi_{g(t),U}^*([\frac{h_1\Omega}{g(t)}])\biggr),
\end{eqnarray*}
where $\Phi_{g,U}^*$ is the homomorphism induced by the map $\Phi_{g,U}: U_f\to U_g$. Note that $\Phi_{g(t),U}^*: H^n(U_{g(t)})\to H^n(U_f)$ is equal to the composition
$$
H^{n}(U_{g(t)})\xrightarrow{\tau_{g(t)}^*}H^n(U_{\mathcal{N}})\xrightarrow{(\tau_f^*)^{-1}} H^n(U_f).
$$
Hence,
\begin{eqnarray*}
\frac{d}{d t}\biggl|_{t=0} \Phi_{g(t),X}^*(\omega_t(h_1))&=&R_f\biggl(\frac{d}{dt}\biggl|_{t=0}(\tau_{f}^*)^{-1}\tau_{g(t)}^*[\frac{h_1\Omega}{g(t)}]\biggr)\\
&=&R_f\biggl((\tau_{f}^*)^{-1}\frac{d}{dt}\biggl|_{t=0}[\tau_{g(t)}^*\frac{h_1\Omega}{g(t)}]\biggr).\\
\end{eqnarray*}
Note that $\tau_{g(t)}^*$ acting on forms is a restriction map, it follows that
\begin{eqnarray*}
\frac{d}{dt}\biggl|_{t=0}[\tau_{g(t)}^*\frac{h_1\Omega}{g(t)}]&=&\frac{d}{dt}\biggl|_{t=0}[\frac{h_1\Omega}{g(t)}\biggl|_{U_\mathcal{N}}]\\
&=&[\frac{d}{dt}\biggl|_{t=0}\frac{h_1\Omega}{g(t)}\biggl|_{U_\mathcal{N}}]\\
&=&[-\frac{hh_1\Omega}{f^2}\biggl|_{U_\mathcal{N}}].
\end{eqnarray*}
Therefore,
\begin{eqnarray*}
\frac{d}{d t}\biggl|_{t=0}\Phi_{g(t),X}^*(\omega_t(h_1))
&=&R_f\biggl((\tau_{f}^*)^{-1}[-\frac{hh_1\Omega}{f^2}\biggl|_{U_\mathcal{N}}]\biggr)\\
&=&R_f([-\frac{hh_1\Omega}{f^2}]).\\
\end{eqnarray*}
Now the proof of Lemma \ref{lem: dp} is complete.
\begin{rk}
To prove Theorem \ref{main cor}, it is essential for us to obtain a diagram like \eqref{eq: square}. In fact, when Y. Zhao \cite{Zhao} proves the infinitesimal Torelli theorem for nodal surfaces, he uses such a diagram implicitly; however, he does not give any proofs. We believe that a detailed proof is indeed needed and this is a special reason why our discussions above always include the case $n=3$.
\end{rk}
|
\section{Introduction}
\label{sec-introduction}
\subsection{Badly approximable vectors}
\label{subsec-badly-approximable-vectors}
Given a positive integer $n$, the weighted version of Dirichlet's approximation theorem says the following:
\begin{theorem}[Dirichlet's Theorem, 1842]
A vector $\vect{r} = (r_1 , \dots, r_n)$ is called a $n$-dimensional weight if $r_i \geq 0$ for $i=1,\dots, n$ and
\[r_1 + \cdots + r_n = 1.\]
For any $n$-dimensional weight $\vect{r} = (r_1,\dots, r_n)$, the following statement holds. For any vector $\vect{x} = (x_1, \dots , x_n) \in \mathbb{R}^n$ and any $N > 1$, there exists an integer vector $\vect{p} = (p_1, \dots, p_n , q) \in \mathbb{Z}^{n+1}$such that $0 < |q| \leq N$ and
\[ |q x_i + p_i| \leq N^{-r_i} \text{, for } i =1,\dots , n.\]
\end{theorem}
This theorem is the starting point of study in simutaneous Diophantine approximation. Using this theorem, one can easily show the following:
\begin{corollary}
\label{cor:dirichlet-diophantine-approximation}
For any vector $\mathbf{x} = (x_1, \dots, x_n) \in \mathbb{R}^n$, there are infinitely many integer vectors $\vect{p} = (p_1,\dots, p_n , q)\in \mathbb{Z}^{n+1}$ with $q \neq 0$ satisfying the following:
\begin{equation}
\label{equ:weighted-diophantine}
|q|^{r_i} |q x_i + p_i| \leq 1 \text{ for } i = 1 ,\dots, n.
\end{equation}
\end{corollary}
\par For almost every vector $\mathbf{x} \in \mathbb{R}^n$, the above corollary remains true if we replace $1$ with any smaller constant $c>0$ on the right hand side of \eqref{equ:weighted-diophantine}. The exceptional vectors are called $\vect{r}$-weighted badly approximable vectors. We give the formal definition as follows:
\begin{definition}
\label{def:badly-approximable}
Given a $n$-dimensional weight $\vect{r} = (r_1, \dots, r_n)$, a vector $\mathbf{x} \in \mathbb{R}^n$ is called $\vect{r}$-weighted badly approximable if there exists a constant $c >0$ such that for any $\vect{p} =(p_1, \dots, p_n, q) \in \mathbb{Z}^{n+1}$ with $q \neq 0$,
\[
\max_{1 \leq i \leq n} |q|^{r_i} |q x_i + p_i| \geq c .
\]
\end{definition}
For a $n$-dimensional weight $\vect{r}$, let us denote the set of $\vect{r}$-weighted badly approximable vectors in $\mathbb{R}^n$ by $\mathrm{\mathbf{Bad}}(\vect{r})$. In particular, $\mathrm{\mathbf{Bad}}(1)$ denotes the set of badly approximable numbers.
\par The study of the size of $\mathrm{\mathbf{Bad}}(\vect{r})$ has a long history and is active in both number theory and homogeneous dynamical systems. It is well known that the Lebesgue measure of
$\mathrm{\mathbf{Bad}}(\vect{r})$ is zero for any $n$-dimensional weight $\vect{r}$. However, people have shown that $\mathrm{\mathbf{Bad}}(\vect{r})$ has full Hausdorff dimension, cf. \cite{Jarnik}, \cite{Schmidt-Game}, \cite{Pollington-Velani} and \cite{Klein-Weiss-Modified-Schmidt}.
\par For the intersection of sets of different weighted badly approximable vectors, Wolfgang M. Schmidt makes the following famous conjecture in 1982:
\begin{conjecture}[Schmidt's Conjecture, see \cite{Schmidt}]
\label{conj:schmidts-conjecture}
\[\mathrm{\mathbf{Bad}}( 1/3, 2/3) \cap \mathrm{\mathbf{Bad}}( 2/3, 1/3) \neq \emptyset.\]
\end{conjecture}
In 2011, Badziahin, Pollington and Velani \cite{BPV} settle this conjecture by showing the following: for any countable collection of $2$-dimensional weights $\{(i_t, j_t): t \in \mathbb{N}\}$, if $\liminf_{t \to \infty} \min\{i_t, j_t\} >0$, then
\[\dim_H\left(\bigcap_{t =1}^{\infty} \mathrm{\mathbf{Bad}}( i_t, j_t )\right) = 2.\]
Here $\dim_H(\cdot)$ denotes the Hausdorff dimension of a set. An \cite{An} later strengthens their result by removing the condition on the weights. In fact, An proves the following much stronger result: for any $2$-dimensional weight $(r_1, r_2)$, $\mathrm{\mathbf{Bad}}(r_1, r_2)$ is $(24\sqrt{2})^{-1}$-winning. Here a set is called $\alpha$-winning if it is a winning set for Schmidt's $(\alpha, \beta)$-game for any $\beta \in (0,1)$. This statement implies that any countable intersection of sets of weighted badly approximable vectors is $\alpha$-winning. The reader is refered to \cite{Schmidt-Game} for more details of Schmidt's game.
\par For $n \geq 3$, Beresnevich \cite{Beresnevich} proves the following theorem:
\begin{theorem}[see~{\cite[Corollary 1]{Beresnevich}}]
\label{thm:beresnevich-thm}
Let $n \geq 2$ be an integer and $\mathcal{U} \subset \mathbb{R}^n$ be an analytic and non-degenerate submanifold in $\mathbb{R}^n$. Here a submanifold is called non-degenerate if it is not contained in any hyperplane of $\mathbb{R}^n$. Let $W$ be a finite or countable set of $n$-dimensional weights such that $\inf_{\vect{r} \in W}\{ \tau(\vect{r})\} >0$ where $\tau(r_1, \dots, r_n) := \min\{ r_i : r_i >0\}$ for an $n$-dimensional weight $(r_1, \dots, r_n)$. Then
\[ \dim_H \left( \bigcap_{\vect{r} \in W} \mathrm{\mathbf{Bad}}(\vect{r}) \cap \mathcal{U} \right) = \dim \mathcal{U}. \]
\end{theorem}
\subsection{Notation}
\label{subsec-notation}
\par In this paper, we will fix the following notation.
\par For a set $\mathcal{S}$, let $|\mathcal{S}|$ denote the cardinality of $\mathcal{S}$. For a measurable subset $E \subset \mathbb{R}$, let $m(E)$ denote its Lebesgue measure.
\par For a matrix $M$, let $M^{\mathrm{T}}$ denote its transpose. For integer $k >0$, let $\mathrm{I}_k$ denote the $k$-dimensional identity matrix.
\par Let $\|\cdot\|$ denote the supremum norm on $\mathbb{R}^n$ and $\mathbb{R}^{n+1}$. Let $\|\cdot\|_2$ denote the Euclidean norm on $\mathbb{R}^n$ and $\mathbb{R}^{n+1}$. For $\vect{x} \in \mathbb{R}^{n+1}$ (or $\in \mathbb{R}^n$) and $r >0$, let $B(\vect{x}, r)$ denote the closed ball in $\mathbb{R}^{n+1}$ (or $\mathbb{R}^n$) centered at $\vect{x}$ of radius $r$, with respect to $\|\cdot\|$. For every $i = 1, \dots , n+1$, there is a natural supremum norm on $\bigwedge^i \mathbb{R}^{n+1}$. Let us denote it by $\|\cdot\|$.
\par Throughout this paper, when we say that $c$ is a constant, we always mean that $c$ is a constant only depending on the dimension $n$. For quantities $A$ and $B$, let us use $A \ll B$ to mean that there is a constant $C>0$ such that $A \leq C B$. Let $A \asymp B$ mean that $A \ll B$ and $B \ll A$. For a quantity $A$, let $O(A)$ denote a quantity which is $\ll A$ or a vector whose norm is $\ll A$.
\subsection{Main results}
\label{subsec-main-result}
In this paper, we will strengthen Theorem \ref{thm:beresnevich-thm} by removing the condition on weights:
\begin{theorem}
\label{thm:goal}
Let $n \geq 2$ be an integer and $\mathcal{U} \subset \mathbb{R}^n$ be an analytic and non-degenerate submanifold in $\mathbb{R}^n$. Let $W$ be a finite or countable set of $n$-dimensional weights. Then
\[ \dim_H \left( \bigcap_{\vect{r} \in W} \mathrm{\mathbf{Bad}}(\vect{r}) \cap \mathcal{U} \right) = \dim \mathcal{U}. \]
\end{theorem}
\par By the reduction argument in \cite{Beresnevich}, to prove the above theorem, it suffices to prove the theorem for analytic curves:
\begin{theorem}
\label{thm:main-thm}
Let $\varphi: I = [a,b] \to \mathbb{R}^n$ be an analytic and non-degenerate curve in $\mathbb{R}^n$. Let $W$ be a finite or countable set of $n$-dimensional weights. Then
\[ \dim_H \left( \bigcap_{\vect{r} \in W} \mathrm{\mathbf{Bad}}(\vect{r}) \cap \varphi(I) \right) = 1. \]
\end{theorem}
\par By \cite{Beresnevich}, Theorem \ref{thm:goal} has the following corollary:
\begin{corollary}
\label{cor:double-intersection}
Let $m,n \in \mathbb{N}$, $B$ be a ball in $\mathbb{R}^m$, $W$ be a finite or countable set of $n$-dimensional weights and $\mathcal{F}_n(B)$ be a finite family of analytic non-degenerate maps $\vect{f} : B \to \mathbb{R}^n$. Then
\[ \dim_H \left( \bigcap_{\vect{f} \in \mathcal{F}_n(B)} \bigcap_{\vect{r} \in W} \vect{f}^{-1}(\mathrm{\mathbf{Bad}}(\vect{r}))\right) = m.\]
\end{corollary}
\par Compared with \cite{Beresnevich}, in this paper, we study this problem through homogeneous dynamics and prove Theorem \ref{thm:main-thm} using the linearization technique.
\subsection{Bounded orbits in homogeneous spaces}
\label{subsec-bounded-orbits-homogeneous-space}
\par Let us briefly recall the correspondence between Diophantine approximation and homogeneous dynamics. The reader may see \cite{Dani}, \cite{Klein_Mar} and \cite{Klein_Weiss} for more details.
\par Let $G = \mathrm{SL}(n+1, \mathbb{R})$, and $\Gamma = \mathrm{SL}(n+1, \mathbb{Z})$. The homogeneous space $X = G/\Gamma$ can be identified with the space of unimodular lattices in $\mathbb{R}^{n+1}$. The point $g\Gamma$ is identified with the lattice $g\mathbb{Z}^{n+1}$. For $\epsilon >0$, let us define
\[K_{\epsilon}:=\left\{\Lambda \in X: \Lambda \cap B(\vect{0}, \epsilon) = \{\vect{0}\}\right\}.\]
It is well known that every $K_{\epsilon}$ is a compact subset of $X$ and every compact subset of $X$ is contained in some $K_{\epsilon}$.
\par For a weight $\vect{r}=(r_1, \dots, r_n)$, let us define the diagonal subgroup $A_{\vect{r}} \subset G$ as follows:
\[A_{\vect{r}} := \left\{ a_{\vect{r}}(t) := \begin{bmatrix}e^{r_1 t} & & & \\ & \ddots & & \\ & & e^{r_n t} & \\ & & & e^{-t}\end{bmatrix}: t \in \mathbb{R} \right\}.\]
\par For $\mathbf{x} \in \mathbb{R}^n$, let us denote
\[V(\mathbf{x}) := \begin{bmatrix} \mathrm{I}_n & \mathbf{x} \\ & 1 \end{bmatrix}.\]
\begin{proposition}
\label{prop:correspondence-diophantine-dynamics}
$\mathbf{x} \in \mathrm{\mathbf{Bad}}(\vect{r})$ if and only if $\{a_{\vect{r}}(t)V(\mathbf{x})\mathbb{Z}^{n+1}: t >0\}$ is bounded.
\end{proposition}
\begin{proof}
\par The proof is well known and standard. We give the proof here for completeness.
\par On the one hand, if $\mathbf{x} \in \mathrm{\mathbf{Bad}}(\vect{r})$, then there exists a constant $c >0$ such that for any integer vector
\[\vect{p} = (p_1, \dots, p_n, q)^{\mathrm{T}}\]
with $q \neq 0$, we have that $\max_{1\leq i \leq n} |q|^{r_i}|q x_i + p| \geq c$. Now let us consider the lattice
$\Lambda(t) = a_{\vect{r}}(t) V(\mathbf{x}) \mathbb{Z}^{n+1}$. We claim that for any $t >0$, every nonzero vector in $\Lambda(t)$ has norm at least $c$. In fact, for any nonzero integer vector $\vect{p} = (p_1, \dots, p_n, q)^{\mathrm{T}}$, we have that
\[a_{\vect{r}}(t) V(\mathbf{x}) \vect{p} = (e^{r_1 t} (qx_1 + p_1), \dots , e^{r_n t}(qx_n + p_n), e^{-t}q)^{\mathrm{T}}.\]
If $q = 0$, then the claim is obvious since
\[ \|a_{\vect{r}}(t) V(\mathbf{x}) \vect{p}\| = \max_{1\leq i\leq n} e^{r_i t} |p_i| >1.\]
Let us assume that $q \neq 0$. For $e^t < |q|$, we have that
\[\|a_{\vect{r}}(t) V(\mathbf{x}) \vect{p}\| \geq e^{-t}|q| > 1. \]
For $e^t \geq |q|$, we have that
\[
\begin{array}{cl}
\|a_{\vect{r}}(t) V(\mathbf{x}) \vect{p}\| & \geq \max_{1\leq i \leq n} e^{r_i t} |qx_i + p_i| \\
& \geq \max_{1\leq i \leq n} q^{r_i} |qx_i + p_i| \geq c.
\end{array}
\]
This proves one direction of the statement.
\par On the other hand, if $\{a_{\vect{r}}(t) V(\mathbf{x}) \mathbb{Z}^{n+1}: t >0\}$ is bounded, we want to show that $\mathbf{x} \in \mathrm{\mathbf{Bad}}(\vect{r})$. In fact, there exists a constant $c >0$ such that $a_{\vect{r}}(t) V(\mathbf{x}) \mathbb{Z}^{n+1} \in K_{c}$ for all $t >0$. Then for any integer vector $\vect{p} =(p_1, \dots, p_n, q)^{\mathrm{T}}$ with $q \neq 0$, we have that
\[\|a_{\vect{r}}(t)V(\mathbf{x}) \vect{p}\| \geq c, \text{ for any } t >0.\]
Let $t = t_0$ such that $e^{t_0} = 2 |q| /c$. Then the above inequality tells that
\[\max_{1\leq i \leq n} \xi^{r_i}|q|^{r_i}|q x_i + p_i| \geq c, \text{ where } \xi = 2/c.\]
Let $\epsilon = \max_{1\leq i \leq n} \xi^{r_i}$, then the above inequality implies that
\[\max_{1\leq i \leq n} |q|^{r_i}|q x_i + p_i| \geq c\epsilon^{-1} .\]
\par This proves the other direction.
\end{proof}
Therefore our main theorem is equivalent to saying that for any analytic submanifold $\mathcal{U} \subset \mathbb{R}^n$ and any countable collection of one-parameter diagonal subgroups $\{A_{\vect{r}_s}: s \in \mathbb{N}\}$, the set of $\mathbf{x} \in \mathcal{U}$ such that
\[\{a_{\vect{r}_s}(t)V(\mathbf{x})\mathbb{Z}^{n+1}: t>0\}\]
is bounded for all $s \in \mathbb{N}$ has full Hausdorff dimension.
\par The study of bounded trajectories under the action of diagonal subgroups in homogeneous spaces is a fundamental topic in homogeneous dynamics and has been active for decades. The basic set up of this type of problems is the following. Let $G$ be a Lie group and $\Gamma \subset G$ be a nonuniform lattice in $G$. Then $X = G/\Gamma$ is a noncompact homogeneous space. Let $A = \{ a(t) : t \in \mathbb{R}\}$ be a one-dimensional diagonalizable subgroup and let $\mathrm{\mathbf{Bd}}(A)$ be the set of $x \in X$ such that $A^{+}x$ is bounded in $X$, where $A^+ := \{a(t): t >0\}$. Then one can ask whether $\mathrm{\mathbf{Bd}}(A)$ has full Hausdorff dimension. For a submanifold $\mathcal{U} \subset X $, one can also ask whether $\mathrm{\mathbf{Bd}}(A)\cap \mathcal{U}$ has Hausdorff dimension $\dim \mathcal{U}$.
\par In 1986, Dani \cite{Dani-Bounded-Orbits} studies the case where $G$ is a semisimple Lie group with $\mathbb{R}$-rank one. In this case, he proves that for any non-quasi-unipotent one parameter subgroup $A \subset G$, $\mathrm{\mathbf{Bd}}(A)$ has full Hausdorff dimension. His proof relies on Schmidt's game. In 1996, Kleinbock and Margulis \cite{Klein-Margulis-Bounded} study the case where $G$ is a semisimple Lie group and $\Gamma$ is a irreducible lattice in $G$. In this case, they prove that $\mathrm{\mathbf{Bd}}(A)$ has full Hausdorff dimension for any non-quasi-unipotent subgroup $A$. Their proof is based on the mixing property of the action of $A$ on $X$. Recently, An, Guan and Kleinbock study the case where $G = \mathrm{SL}(3,\mathbb{R})$ and $\Gamma = \mathrm{SL}(3,\mathbb{Z})$. They prove that for any countable one-parameter diagonalizable one-parameter subgroups $\{F_s : s \in \mathbb{N}\}$, the intersection $\bigcap_{s =1 }^{\infty} \mathrm{\mathbf{Bd}}(F_s)$ has full Hausdorff dimension. Their proof closely follows the argument in the work of An \cite{An} and uses a variantion of Schmidt's game.
\subsection{The linearization technique}
\label{subsec-main-tools-proof}
\par In \cite{Beresnevich}, the proof relies on a very careful study of the distribution of integer points in $\mathbb{R}^{n+1}$ and the argument is elementary.
\par In this paper, we study this problem through homogeneous dynamics and tackle the technical difficulties using the linearization technique. We study the Diophantine properties using homogeneous dynamics. It turns out that in order to get the Hausdorff dimension, it is crucial to study distributions of long pieces of unipotent orbits in the homogeneous space $G/\Gamma$. To be specific, for a particular long piece $C$ of a unipotent orbit, we need to estimate the length of the part in $C$ staying outside a large compact subset $K$ of $G/\Gamma$. In homogeneous dynamics, the standard tool to study this type of problem is the linearization technique. The linearization technique is a standard and powerful technique in homogeneous dynamics. Using the linearization technique, we can transform a problem in dynamical systems to one in linear representations. Then we can study this problem using tools and results in representation theory.
\par Let us briefly describe the technical difficulty when we apply the linearization technique. Let $\mathcal{V}$ be a finite dimensional linear representation of $\mathrm{SL}(n+1,\mathbb{R})$ with a norm $\|\cdot\|$ and $\Gamma(\mathcal{V}) \subset \mathcal{V}$ be a fixed discrete subset of $\mathcal{V}$. Let $U = \{u(r): r \in \mathbb{R}\}$ be a one dimensional unipotent subgroup of $G$. Given a large number $ T > 1$, our main task is to estimate the measure of $r \in [ - T, T]$ such that there exists $v \in \Gamma(\mathcal{V})$ such that $\|u(r)v\| \leq \epsilon$ where $\epsilon>0$ is a small number. By Dani non-divergence theorem (see \cite{Dani}), the measure is very small compared with $T$ given that for any such $v \in \Gamma(\mathcal{V})$
\[ \max\{ \|u(r)v\|: r \in [-T, T] \} \geq \rho \]
where $\rho >0$ is some fixed number. The difficulty is that there exists some $v \in \Gamma(\mathcal{V})$, such that
\[ \max\{ \|u(r)v\|: r \in [-T, T] \} \leq \rho. \]
Let us call such intervals $T$-bad intervals. In this paper, we will use the representation theory to get some nice properties of such $v$'s. We then use these nice properties to show that in a longer interval, say $[-T^2, T^2]$, the number of $T$-bad intervals is $\ll T^{1-\mu}$ for some constant $\mu >0$. This result is sufficient to prove Theorem \ref{thm:main-thm}.
\par In this paper, $\mathcal{V}$ is the canonical representation of $\mathrm{SL}(n+1, \mathbb{R})$ on $\bigwedge^i \mathbb{R}^{n+1}$ and $\Gamma(\mathcal{V}) = \bigwedge^i \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$ where $i = 1, \dots, n$.
\par The main technical results in this paper are proved in \S \ref{sec:count-dangerous-intervals}, \S \ref{subsec-dangerous-case} and \S \ref{subsec-extremely-dangerous-case}.
\par We refer the reader to \cite{Ratner}, \cite{Margulis-Tomanov}, \cite{Mozes_Shah}, \cite{Shah_1}, \cite{Shah_2} and \cite{Lindenstrauss-Margulis} for more applications of the linearization technique.
\subsection{The organization of the paper}
\label{subsec-organization-of-paper}
\par The paper is organized as follows:
\begin{itemize}
\item In \S \ref{sec-preliminaries}, we will recall some basic facts on Diophantine approximation, linear representations and lattices in $\mathbb{R}^{n+1}$.
\item In \S \ref{sec:cantor-like-construction}, we will recall a theorem on computing the Hausdorff dimension of Cantor like sets. We will also construct a Cantor-like covering of the set of weighted badly approximable points.
\item In \S \ref{sec:count-dangerous-intervals}, we will prove two technical results on counting lattice points. This is one of main technical contributions in this paper. Our proof relies on the linearization technique and $\mathrm{SL}(2, \mathbb{R})$ representations.
\item In \S \ref{sec-proof-main-result}, we will finish the proof of Theorem \ref{thm:main-task}. Our proof relies on the Kleinbock-Margulis non-divergence theorem (Theorem \ref{thm:non-divergence-kleinbock-margulis}) and the linearization technique.
\end{itemize}
\medskip
\par \noindent {\bf Acknowledgements.} The author would like to thank Professor Elon Lindenstrauss and Professor Barak Weiss for sharing many insightful ideas on this problem. He also thanks Professor Shahar Mozes for helpful conversations on this problem. He appreciates their encouragements during the process of this work. He also wants to thank Professor Jinpeng An for some early discussion on related problems.
\section{Preliminaries}
\label{sec-preliminaries}
\subsection{Dual form of approximation}
\label{subsec-dual-form}
\par We first recall the following equivalent definition of $\mathrm{\mathbf{Bad}}(\vect{r})$:
\begin{lemma}[see~{\cite[Lemma 1]{Beresnevich}}]
\label{lm:equivalent-definition-badly-approximable}
Let $\vect{r} = (r_1, \dots, r_n) \in \mathbb{R}^n$ be a weight and $\mathbf{x} \in \mathbb{R}^n$. The following statements are equivalent:
\begin{enumerate}
\item $\mathbf{x} \in \mathrm{\mathbf{Bad}}(\vect{r})$.
\item There exists $c >0$ such that for any integer vector $(p_1, \dots, p_n ,q)$ such that $q \neq 0$, we have that
\[ \max_{1 \leq i \leq n} |q|^{r_i}|q x_i + p_i| \geq c.\]
\item There exists $c >0$ such that for any $N \geq 1$, the only integer solution $(a_0, a_1, \dots, a_n)$ to the system
\[
\begin{array}{rr}
|a_0 + a_1 x_1 + \cdots + a_n x_n| < c N^{-1} , & |a_i| < N^{r_i} \text{ for all } 1 \leq i \leq n
\end{array}
\]
is $a_0 = a_1 = \cdots = a_n =0$.
\end{enumerate}
\end{lemma}
\begin{proof}
The reader is referred to \cite{Mahler}, \cite[Appendix]{BPV} and \cite[Appendix A]{Beresnevich} for the proof.
\end{proof}
\par Later in this paper we will use the third statement as the definition of $\mathrm{\mathbf{Bad}}(\vect{r})$.
\par Given a weight $\vect{r}=(r_1, \dots, r_n)$, let us define
\[ D_{\vect{r}} := \left\{ d_{\vect{r}}(t) := \begin{bmatrix}e^{t} & & & \\
& e^{-r_1 t} & & \\
& & \ddots & \\
& & & e^{-r_n t} \end{bmatrix}: t \in \mathbb{R} \right\}.\]
For $\mathbf{x} \in \mathbb{R}^n$, let us define
\[U(\mathbf{x}) := \begin{bmatrix} 1 & \mathbf{x}^{\mathrm{T}} \\ & \mathrm{I}_n \end{bmatrix}.\]
If we use the third statement in Lemma \ref{lm:equivalent-definition-badly-approximable} as the definition of $\mathrm{\mathbf{Bad}}(\vect{r})$, then it is easy to show that $\mathbf{x} \in \mathrm{\mathbf{Bad}}(\vect{r})$ if and only if $U(\mathbf{x})\mathbb{Z}^{n+1} \in \mathrm{\mathbf{Bd}}(D_{\vect{r}})$. In fact, the statement can be proved using the same argument as in the proof of Proposition \ref{prop:correspondence-diophantine-dynamics}.
\subsection{The canonical representation}
\label{subsec-canonical-representation-sln}
\par Let $V = \mathbb{R}^{n+1}$. There is a canonical representation of $G = \mathrm{SL}(n+1, \mathbb{R})$ on $V$. It induces a canonical representation of $G$ on $\bigwedge^i V$ for every $i=1,2,\dots, n$. For $g \in G$ and
\[\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge\nolimits^i V, \]
$g \vect{v} = (g \vect{v}_1) \wedge \cdots \wedge (g \vect{v}_i).$
\par For $i =1 , \dots, n$, let $\vect{e}_i \in \mathbb{R}^n$ denote the vector with $1$ in the $i$th component and $0$ in other components.
\par Let us fix a basis for $V$ as follows. Let $\vect{w}_{+} := ( 1, 0, \dots, 0)^{\mathrm{T}}$. For $i = 1, \dots, n$, let $\vect{w}_i := (0, \dots, 1, \dots, 0)^{\mathrm{T}}$ with $1$ in the $i+1$st component and $0$ in other components. Then $\{\vect{w}_+, \vect{w}_1, \dots, \vect{w}_n\}$ is a basis for $V$. Let $W$ denote the subspace of $V$ spanned by $\{\vect{w}_1, \dots, \vect{w}_n\}$. For $j= 2, \dots, n$, let $W_j$ the subspace of $W$ spanned by $\{\vect{w}_j, \dots, \vect{w}_n\}$.
\par Let us define
\[Z : = \left\{ z = \begin{bmatrix}1 & \\ & \mathfrak{k}(z)\end{bmatrix}: \mathfrak{k}(z) \in \mathrm{SO}(n) \right\}.\]
There is a canonical action of $\mathrm{SO}(n)$ on $\mathbb{R}^n$. For $\mathfrak{k} \in \mathrm{SO}(n)$ and $\vect{x} \in \mathbb{R}^n$, let us denote by $\mathfrak{k} \cdot \vect{x}$ the canonical action of $\mathfrak{k}$ on $\vect{x}$. It is straightforward to check that for $z = \begin{bmatrix}1 & \\ & \mathfrak{k}(z)\end{bmatrix} \in Z$ and $\mathbf{x} \in \mathbb{R}^n$,
\[z U(\mathbf{x}) z^{-1} = U(\mathfrak{k}(z)\cdot \mathbf{x}).\]
\par For any $\mathbf{x} \in \mathbb{R}^n$, $U(\mathbf{x})$ can be embedded into a subgroup $\mathrm{SL}(2, \mathbf{x})$ of $G$ isomorphic to $\mathrm{SL}(2,\mathbb{R})$. In this $\mathrm{SL}(2,\mathbb{R})$ copy, $U(\mathbf{x})$ corresponds to $\begin{bmatrix}1 & \|\mathbf{x}\|_2 \\ & 1\end{bmatrix}$. For $r >0$, let $\xi_{\mathbf{x}} (r) \in \mathrm{SL}(2 , \mathbf{x})$ denote the element corresponding to $\begin{bmatrix} r & \\ & r^{-1}\end{bmatrix} \in \mathrm{SL}(2,\mathbb{R})$.
\par Let us consider the representation of $\mathrm{SL}(2, \mathbf{x})$ on $V$.
\par Let us first consider the case $\mathbf{x} = \vect{e}_1$. Let us denote
\[U_1 := \{ u_1 (r ): = U(r \vect{e}_1): r \in \mathbb{R} \},\]
and
\[\Xi_1 : = \{ \xi_1 (r) : = \mathrm{diag} \{r , r^{-1},1, \dots, 1 \}: r > 0 \}.\]
It is easy to see that $\xi_1 (r) \vect{w}_+ = r \vect{w}_+$, $u_1 (r) \vect{w}_+ = \vect{w}_+$, $\xi_1 (r) \vect{w}_1 = r^{-1} \vect{w}_1$, $u_1(r)\vect{w}_1 = \vect{w}_1 + r \vect{w}_+$, and for any $\vect{w} \in W_2$, $\vect{w}$ is fixed by $\mathrm{SL}(2, \vect{e}_1)$.
\par For general $\mathbf{x} \in \mathbb{R}^n$, we have that $\mathbf{x} = \|\mathbf{x}\|_2 \mathfrak{k} \cdot \vect{e}_1$ for some $\mathfrak{k} \in \mathrm{SO}(n)$. Then
\[\mathrm{SL}(2,\mathbf{x}) = z(\mathfrak{k}) \mathrm{SL}(2, \vect{e}_1) z^{-1}(\mathfrak{k}) \]
where $z(\mathfrak{k}) = \begin{bmatrix}1 & \\ & \mathfrak{k}\end{bmatrix} \in Z$. In particular, we have that
\[U(\mathbf{x}) = z(\mathfrak{k}) u_1(\|\mathbf{x}\|_2)z^{-1}(\mathfrak{k})\]
and $\xi_{\mathbf{x}}(r) = z(\mathfrak{k}) \xi_1(r) z^{-1}(\mathfrak{k})$. Since $z(\mathfrak{k}) \vect{w}_+ = \vect{w}_+$ and $z(\mathfrak{k}) W = W$, we have that $\xi_{\mathbf{x}}(r) \vect{w}_+ = r \vect{w}_+$, $U(\mathbf{x}) \vect{w}_+ = \vect{w}_+$, $\xi_{\mathbf{x}}(r) z(\mathfrak{k}) \vect{w}_1 = r^{-1} \mathfrak{k} \cdot \vect{w}_1$, $U(\mathbf{x}) z(\mathfrak{k}) \vect{w}_1 = z(\mathfrak{k}) \vect{w}_1 + \|\mathbf{x}\|_2 \vect{w}_+$ and for any $\vect{w} \in z(\mathfrak{k})W_2$, $\vect{w}$ is fixed by $\mathrm{SL}(2, \mathbf{x})$.
\par Let us consider the action of $\mathrm{SL}(2,\mathbf{x})$ on $\bigwedge^i V$ for $i = 2,\dots, n$. Let us denote $\mathbf{x} = \|\mathbf{x}\|_2 \mathfrak{k} \cdot \vect{e}_1$ as above. For any $\vect{w} \in \bigwedge^{i-1} z(\mathfrak{k})W_2$, we have that
\[\xi_{\mathbf{x}}(r) ((z(\mathfrak{k})\vect{w}_1)\wedge \vect{w}) = r^{-1} ((z(\mathfrak{k})\vect{w}_1)\wedge \vect{w})\]
,
\[U(\mathbf{x})((z(\mathfrak{k})\vect{w}_1)\wedge \vect{w}) =(z(\mathfrak{k})\vect{w}_1)\wedge \vect{w} + \|\mathbf{x}\|_2 (\vect{w}_+ \wedge \vect{w}),\]
\[\xi_{\mathbf{x}}(r) (\vect{w}_+ \wedge \vect{w}) = r (\vect{w}_+ \wedge \vect{w})\]
and
\[U(\mathbf{x})(\vect{w}_+ \wedge \vect{w}) = \vect{w}_+ \wedge \vect{w}.\]
For any $\vect{w} \in \bigwedge^{i} z(\mathfrak{k})W_2$ and any $\vect{w}' \in \bigwedge^{i-2} z(\mathfrak{k})W_2$, we have that $\vect{w}$ and $\vect{w}_+ \wedge (z(\mathfrak{k})\vect{w}_1) \wedge \vect{w}'$ are fixed by $\mathrm{SL}(2,\mathbf{x})$.
\subsection{Lattices in $\mathbb{R}^{n+1}$}
\label{subsec-lattices-in-Rn+1}
\par In this subsection let us recall some basic facts on lattices and sublattices in $\mathbb{R}^{n+1}$.
\par For a discrete subgroup $\Delta$ of $\mathbb{R}^{n+1}$, let $\mathrm{Span}_{\mathbb{R}}(\Delta)$ denote the $\mathbb{R}$-span of $\Delta$.
\par Let $\Lambda \in X = G/\Gamma$ be a unimodular lattice in $\mathbb{R}^{n+1}$. For $i=1, \dots, n+1$, let
$\Lambda_i \subset \Lambda$ be a $i$-dimensional sublattice of $\Lambda$. We say that $\Lambda_i$ is primitive if
$\mathrm{Span}_{\mathbb{R}}(\Lambda_i)\cap \Lambda = \Lambda_i$.
\par Given a $i$-dimensional primitive sublattice $\Lambda_i$ of $\Lambda$, let us choose a basis $\{\vect{v}_1, \dots, \vect{v}_i\}$ of $\Lambda_i$. Let us denote $d(\Lambda_i) = \|\vect{v}_1 \wedge \cdots \wedge \vect{v}_i\|$.
\par For $ j = 1, \dots, i$, let
\[\lambda_j(\Lambda_i) := \inf\{r\geq 0: B(\vect{0}, r) \text{ contains at least } j \text{ linearly independent vectors of } \Lambda_i\}.\]
By the Minkowski Theorem (see \cite{Cassels}), we have the following:
\begin{equation}
\label{equ:minkowski-thm}
\lambda_1(\Lambda_i)\cdots \lambda_i(\Lambda_i) \asymp d(\Lambda_i).
\end{equation}
Moreover, there exists a basis (called Minkowski reduced basis) of $\Lambda_i$, $\{\vect{v}_j: j = 1, \dots, i\}$, such that $\|\vect{v}_j\| \asymp \lambda_j(\Lambda_i)$ for every $j = 1, \dots, i$.
\par We will need the following result on counting sublattices:
\begin{proposition}
\label{prop:counting-sublattices}
\par There exists a constant $N >1$ such that the following statement holds. For any $0 <\epsilon < \rho $ and any $i= 1, \dots, n$, let $\Lambda \in K_{\epsilon}$ be a unimodular lattice in $\mathbb{R}^{n+1}$ such that $\lambda_1(\Lambda) \geq \epsilon$. For $\rho >0$, let $\mathcal{C}_i(\Lambda, \rho)$ denote the collection of $i$-dimensional primitive sublattices $\Lambda_i$ of $\Lambda$ with $d(\Lambda_i) \leq \rho$. Then we have that,
\[|\mathcal{C}_i(\Lambda, 1)| \leq \epsilon^{-N} .\]
\end{proposition}
\begin{proof}
\par First note that there exists a constant $N_1 >1$ such that for any $i=1,\dots, n$ and $\rho >0$,
\[ |\mathcal{C}_i(\mathbb{Z}^{n+1}, \rho)| \leq \rho^{N_1}.\]
\par It is a standard fact that there exists a constant $N_2 >1$ such that for any $\Lambda \in K_{\epsilon}$, there eixsts $g \in \mathrm{SL}(n+1, \mathbb{R})$ with $\|g^{-1}\| \leq \epsilon^{-N_2}$ such that $\Lambda = g \mathbb{Z}^{n+1}$. Let us fix $\rho > \epsilon$ and $i = 1, \dots, n$. Then for any $\Lambda_i \in \mathcal{C}_i(\Lambda, 1)$, then we have that $g^{-1} \Lambda_i \subset \mathbb{Z}^{n+1}$ and
\[d(g^{-1} \Lambda_i) \leq \|g^{-1}\|^i d(\Lambda_i) \leq \epsilon^{-(n+1) N_2} .\]
Therefore, we have that
\[ |\mathcal{C}_i(\Lambda, 1)| \leq |\mathcal{C}_i(\mathbb{Z}^{n+1}, \epsilon^{-(n+1) N_2} )| \leq \epsilon^{-N} \]
where $N = N_1 N_2 (n+1)$.
\par This completes the proof.
\end{proof}
\section{A Cantor like construction}
\label{sec:cantor-like-construction}
\par In this section, we will introduce a Cantor like construction which will help us to compute Hausdorff dimension.
\begin{definition}[See~{\cite[\S 5]{Beresnevich}}]
\label{def:cantor-like-construction}
\par For an integer $R >0$ and a closed interval $J \subset \mathbb{R}$, let us denote by $\mathrm{\mathbf{Par}}_{R}(J)$ the collection of intervals obtained by dividing $J$ into $R$ equal closed subintervals. For a collection $\mathcal{I}$ of closed intervals, let us denote
\[\mathrm{\mathbf{Par}}_{R}(\mathcal{I}) := \bigcup_{I \in \mathcal{I}} \mathrm{\mathbf{Par}}_R(I).\]
A sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ of collections of closed intervals is called a $R$-sequence if for every $q \geq 1$, $\mathcal{I}_q \subset \mathrm{\mathbf{Par}}_R(\mathcal{I}_{q-1})$. For a $R$-sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ and $q \geq 1$, let us define $\hat{\mathcal{I}}_q := \mathrm{\mathbf{Par}}_R(\mathcal{I}_{q-1})\setminus \mathcal{I}_q$ and
\[\mathcal{K}(\{\mathcal{I}_q: q \in \mathbb{N}\}) := \bigcap_{q \in \mathbb{N}} \bigcup_{I_q \in \mathcal{I}_q} I_q .\]
Then every $R$-sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ gives a Cantor like subset $\mathcal{K}(\{\mathcal{I}_q\}_{q \in \mathbb{N}})$ of $\mathbb{R}$.
\par For $q \geq 1$ and a partition $\{\hat{\mathcal{I}}_{q, p}\}_{0 \leq p \leq q-1}$ of $\hat{\mathcal{I}}_q$, let us define
\[d_q(\{\hat{\mathcal{I}}_{q, p}\}_{0\leq p \leq q-1}):= \sum_{p=0}^{q-1} \left( \frac{4}{R}\right)^{q-p} \max_{I_p \in \mathcal{I}_p} F( \hat{\mathcal{I}}_{q,p}, I_p), \]
where $F( \hat{\mathcal{I}}_{q,p}, I_p) := |\{I_q \in \hat{\mathcal{I}}_{q,p}, I_q \in I_p\}|$. Let us define
\[d_q(\mathcal{I}_q) := \min_{\{\hat{\mathcal{I}}_{q, p}\}_{0\leq p \leq q-1}} d_q(\{\hat{\mathcal{I}}_{q, p}\}_{0\leq p \leq q-1}),\]
where $\{\hat{\mathcal{I}}_{q, p}\}_{0\leq p \leq q-1}$ runs over all possible partitions of $\hat{\mathcal{I}}_q$. Let us define
\[d(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) := \max_{q \in \mathbb{N}} d_q(\mathcal{I}_q).\]
\end{definition}
\begin{definition}[See~{\cite[\S 5]{Beresnevich}}]
\label{def:m-rich}
For $M >1$ and a compact subset $X \subset \mathbb{R}$, we say that $X$ is $M$-Cantor rich if for any $\epsilon >0$ and any integer $R \geq M$, there exists a $R$-sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ such that
\[\mathcal{K}(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) \subset X\]
and $d(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) \leq \epsilon$.
\end{definition}
\par Our proof relies on the following two theorems:
\begin{theorem}[See~{\cite[Theorem 6]{Beresnevich}}]
\label{thm:hausdorff-dimension-m-rich}
Any $M$-Cantor rich set $X$ has full Hausdorff dimension.
\end{theorem}
\begin{theorem}[See~{\cite[Theorem 7]{Beresnevich}}]
\label{thm:intersection-m-rich}
Let $I_0$ be a compact interval. Then any countable intersection of $M$-Cantor rich sets in $I_0$ is $M$-Cantor rich.
\end{theorem}
To show Theorem \ref{thm:goal}, it suffices to find a constant $M >1$ and show that for any weight $\vect{r}$, $\varphi^{-1}(\mathrm{\mathbf{Bad}}(\vect{r})\cap \varphi(I))$ is $M$-Cantor rich. We will determine $M>1$ later.
\begin{theorem}
\label{thm:main-task}
There exists a constant $M > 1$ such that for any weight $\vect{r}$, $\varphi^{-1}(\mathrm{\mathbf{Bad}}(\vect{r})\cap \varphi(I))$ is $M$-Cantor rich.
\end{theorem}
\par Our main task is to prove Theorem \ref{thm:main-task}.
\par Let us fix $R \geq M $. We will show that for any $\epsilon >0$, we can construct a $R$-sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ such that $\mathcal{K}(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) \subset \mathrm{\mathbf{Bad}}(\vect{r})$ and $d(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) < \epsilon$. We will follow the construction in \cite{Beresnevich} with some modifications.
\begin{StandingAssumption}
\label{standing-assumptions}
\par Let us make some assumptions to simplify the proof.
\begin{enumerate}[label=\textbf{A.\arabic*}]
\item \label{assumption-1} Without loss of generality, we may assume that $r_1 \geq r_2 \geq \cdots \geq r_n$. We may also assume that $r_n >0$. By \cite{Beresnevich}, if $r_n = 0$, we can reduce the problem to the $n-1$ dimensional case.
\item \label{assumption-2} Since
\[\varphi =(\varphi_1, \dots, \varphi_n)^{\mathrm{T}}: I \to \mathbb{R}^{n}\]
is analytic and nondegenerate, we may assume that for any $s \in I$ and any $i = 1 ,\dots, n$, $\varphi^{(1)}_i (s) \neq 0$. If this is not the case, we can choose a closed subinterval $I' \subset I$ satisfying this condition. Then since $I$ is closed, there exist constants $C_1 > c_1 >0$ such that for any $s \in I$ and any $i=1,\dots, n$, $c_1 \leq |\varphi^{(1)}_i (s)| \leq C_1$.
\end{enumerate}
\end{StandingAssumption}
\par Let us fix some notation. Let $m >0$ be a large integer which we will determine later. Let $ \kappa = R^{-m}$. Let $b >0$ be such that $b^{1+r_1}=R$. For $t >0$, let us denote
\[ g_{\vect{r}} (t) := \begin{bmatrix} b^{t} & & & \\ & b^{-r_1 t} & & \\ & & \ddots & \\ & & & b^{-r_n t} \end{bmatrix}.\]
For $i =1 , \dots, n$, let $\lambda_i = \frac{1+r_i}{1+r_1}$. Then we have that $1 = \lambda_1 \geq \lambda_2 \geq \cdots \geq \lambda_n$. Without loss of generality, we may assume that $ m(I) = 1$.
\par Let us define the $R$-sequence as follows. Let $\mathcal{I}_0 = \{ I \}$. Suppose that we have defined $\mathcal{I}_{q-1}$ for $q \geq 1$ and every $I_{q-1} \in \mathcal{I}_{q-1}$ is a closed interval of length $ R^{-q+1}$. Let us define $\mathcal{I}_{q}\subset \mathrm{\mathbf{Par}}_R(\mathcal{I}_{q})$ as follows. For any $I_{q} \in \mathrm{\mathbf{Par}}_R(\mathcal{I}_q)$, $I_{q} \in \hat{\mathcal{I}}_{q}$ if and only if there exists $s \in I_{q}$ such that $g_{\vect{r}} (q) U(\varphi(s)) \mathbb{Z}^{n+1} \notin K_{\kappa}$. That is to say, there exists $\vect{a} \in \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$ such that $\|g_{\vect{r}} (q) U(\varphi(s))\vect{a}\| \leq \kappa$. Let us define $\mathcal{I}_q = \mathrm{\mathbf{Par}}_R(\mathcal{I}_{q-1})\setminus \hat{\mathcal{I}}_q$. This finishes the construction of $\{\mathcal{I}_q\}_{q\in \mathbb{N}}$. It is easy to see that
\[ \mathcal{K}(\{\mathcal{I}_q\}_{q \in \mathbb{N}}) \subset \mathrm{\mathbf{Bad}}(\vect{r}). \]
\par We need to prove the following:
\begin{proposition}
\label{prop:distance-sequence-small}
For any $\epsilon >0$, there exists an integer $m >0$ such that the $R$-sequence $\{\mathcal{I}_q\}_{q \in \mathbb{N}}$ constructed as above with $\kappa = R^{-m}$ satisfies that
\begin{equation}
\label{equ:distance-sequence-small}
d(\{\mathcal{I}_q\}_{q\in \mathbb{N}}) \leq \epsilon.
\end{equation}
\end{proposition}
\par Let $N >1$ be the constant from Proposition \ref{prop:counting-sublattices}. Let us give the partition $\{\hat{\mathcal{I}}_{q,p}\}_{0\leq p \leq q-1}$ of $\hat{\mathcal{I}}_q$ for each $q \in \mathbb{N}$.
\begin{definition}
\label{def:finer-partition-Iq}
\par Let us fix a small constant $0 < \rho < 1 $. We will modify the choice of $\rho$ later in this paper according to the constants coming from our technical results. For $q \leq 10^6 n^4 N m$, let us define $\hat{\mathcal{I}}_{q, 0} := \hat{\mathcal{I}}_q$ and $\hat{\mathcal{I}}_{q,p} = \emptyset$ for other $p$'s.
\par For $q > 10^6 n^4 N m$ and $l = 2000 n^2 N m$, let $ p = q - 2l$. Let us define $\hat{\mathcal{I}}_{q,p}$ to be the collection of $I_q \in \hat{\mathcal{I}}_q$ with the following property: there exists $x \in I_q$ such that for any $j=1,\dots, n$ and any $\vect{w} = \vect{w}_1 \wedge \cdots \wedge \vect{w}_j \in \bigwedge^j \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$,
\[ \max \{g_{\vect{r}}(q) U(\varphi(x')) \vect{w} : x' \in [x-R^{-q+l}, x+R^{-q+l}] \} \geq \rho^j.\]
\par Let $\eta = \frac{1}{100n^2}$ and $\eta' = \frac{\eta}{1+r_1}$. For $q > 10^6 n^4 N m$ and $ 2000 n^2 N m \leq l \leq 2\eta' q$, let $p = q - 2l$. For $i = 1, \dots, n$, let us define $\hat{\mathcal{I}}_{q,p}(i)$ to be the collection of $I_q \in \hat{\mathcal{I}}_q$ satisfying that there exists $s \in I_q$ and $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus\{\vect{0}\}$ such that
\[ \|g_{\vect{r}}(q) U(\varphi(s'))\vect{v}\| \leq \rho^i,\]
for any $s' \in [s - R^{-q + l}, s + R^{-q +l}]$ and for any $j = 1, \dots, n$ and any $\vect{w} = \vect{w}_1 \wedge \cdots \wedge \vect{w}_j \in \bigwedge^j \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$,
\[\max\left\{ \|g_{\vect{r}}(q) U(\varphi(s')) \vect{w}\|:s' \in [s - R^{-q+l}, s+ R^{-q + l} ]\right\} \geq \rho^j.\]
Let us define $\hat{\mathcal{I}}_{q,p} = \bigcup_{i=1}^{n} \hat{\mathcal{I}}_{q,p}(i)$.
\par For $i = 1, \dots, n$, let us define $\hat{\mathcal{I}}_{q,0}(i)$ to be the collection of $I_q \in \hat{\mathcal{I}}_q$ satisfying that there exists $s \in I_q$ and $\vect{v}= \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1}\setminus\{\vect{0}\}$ such that
\[\max\left\{ \|g_{\vect{r}}(q) U(\varphi(s)) \vect{v}\|:s' \in [s - R^{-q (1-2\eta')}, s+ R^{-q(1-2\eta')} ]\right\} \leq \rho^i.\]
Let us define $\hat{\mathcal{I}}_{q,0} = \bigcup_{i=1}^{n} \hat{\mathcal{I}}_{q,0}(i)$.
\par Let us define $\hat{\mathcal{I}}_{q,p} := \emptyset$ for other $p$'s. It is easy to see that $\{\hat{\mathcal{I}}_{q,p}\}_{0 \leq p \leq q-1}$ is a partition of $\hat{\mathcal{I}}_q$, cf. \cite[Proposition 3]{Beresnevich}.
\end{definition}
\par Besides the definition of $\{\hat{\mathcal{I}}_{q,p}\}_{ 0 \leq p \leq q-1 }$, let us also introduce the notion of dangerous intervals and extremely dangerous intervals:
\begin{definition}
\label{def:dangerous-interval}
For $q \geq 10^6 n^4 N m$, $ 1000 n^2 N m \leq l \leq \eta' q $, and $\vect{a} \in \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$, the $(q,l)$-dangerous interval associated with $\vect{a}$, which is denoted by $\Delta_{q,l} (\vect{a})$, is a closed interval of the form $\Delta_{q,l}(\vect{a}) = [ x - R^{-q + l }, x + R^{-q +l}] \subset I$ such that $I_q \subset \Delta_{q,l}(\vect{a})$ for some $I_q \in \hat{\mathcal{I}}_q$ and
\[\max \{ \|g_{\vect{r}}(q) U(\varphi(s')) \vect{a}\|: s' \in \Delta_{q,l}(\vect{a}) \} = c \rho,\]
for some $c \in [ 1/2, 1]$.
\par For $q \geq 10^6 n^4 N m$ and $\vect{a} \in \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$, the $q$-extremely dangerous interval associated with $\vect{a}$, which is denoted by $\Delta_q (\vect{a})$, is a closed interval of the form $\Delta_q(\vect{a}) = [x - R^{-q + l'}, x + R^{-q + l'}]$ with $l' > \eta' q$ such that $I_q \subset \Delta_q(\vect{a})$ for some $I_q \in \hat{\mathcal{I}}_q$ and
\[\max \{ \|g_{\vect{r}} (q) U(\varphi(s')) \vect{a}\| : s' \in [x - 2 R^{-q + l'}, x + 2 R^{-q + l'}]\} = c \rho\]
for some $c \in [1/2,1]$.
\end{definition}
\begin{remark}
Note that for any $q \geq 10^6 n^4 N m$, there are only finitely many $\vect{a}$'s such that $\Delta_{q, l} (\vect{a})$ or $\Delta_q (\vect{a})$ exist.
\end{remark}
\section{Counting dangerous intervals}
\label{sec:count-dangerous-intervals}
\par In this section we will count dangerous intervals and extremely dangerous intervals.
\begin{proposition}
\label{prop:count-dangerous-intervals}
Let $q \geq 10^6 n^4 N m$, $ 1000 n^2 N m \leq l \leq \eta' q$ and $p = q - 2l $. For $I_p \in \mathcal{I}_p$, let
$\mathcal{D}_{q, l} (I_p)$ denote the collection of $(q,l)$-dangerous intervals which intersect $I_p$. Then for any $I_p \in \mathcal{I}_p$,
\[|\mathcal{D}_{q,l}(I_p)| \ll R^{(1- \frac{1}{10n}) l} .\]
\end{proposition}
\begin{proposition}
\label{prop:count-extremely-dangerous-intervals}.
Let $q \geq 10^6 n^4 N m$. Let $E_q \subset I$ denote the union of $q$-extremely dangerous intervals contained in $I$. Then
$E_q$ can be covered by a collection of $N_q$ closed intervals of length $\delta_q$ and
\[ N_q \leq \frac{K_0 (\rho^{n+1} b^{-\eta q})^{\alpha}}{\delta_q} \]
where $\delta_q = R^{-q(1-\eta')}$, $K_0>0$ is a constant, and $\alpha = \frac{1}{(n+1)(2n-1)}$.
\end{proposition}
\par Proposition \ref{prop:count-extremely-dangerous-intervals} is a rephrase of the following theorem due to Bernik, Kleinbock and Margulis:
\begin{theorem}[See~{\cite[Proposition 2]{Beresnevich}} and~{\cite[Theorem 1.4]{Bernik-Kleinbock-Margulis}}]
\label{thm:bernik-kleinbock-margulis}
Let $q > 10^6 n^4 N m$. Let us define $E_q \subset I$ to be the collection of $s \in I$ satisfying that there exists $\vect{a} =(a_0, a_1, \dots, a_n)^{\mathrm{T}} \in \mathbb{Z}^{n+1}\setminus\{\vect{0}\}$ such that $|a_i| < \rho b^{r_i q}$ for $i=1,\dots,n$, $|f(s)| < \rho b^{-q}$ and $|f'(s)| < b^{(r_1 - \eta)q}$ where
\[f(x) = a_0 + a_1 \varphi_1(x) + \cdots + a_n \varphi_n(x).\]
Then $E_q$ can be covered by a collection $\mathcal{E}_{q}$ of intervals such that
\[m(\Delta) \leq \delta_q \text{ for all } \Delta \in \mathcal{E}_{q},\]
and
\[|\mathcal{E}_{q}| \leq \frac{K_0 (\rho^{n+1} b^{-\eta q})^{\alpha}}{\delta_q},\]
where $K_0 >0$ is a constant, $\delta_q = R^{-q(1 - \eta')}$ and $\alpha = \frac{1}{(n+1)(2n-1)}$.
\end{theorem}
\par The theorem we quote here is the version used in \cite{Beresnevich} with some minor modifications. The original version proved in \cite{Bernik-Kleinbock-Margulis} is more general.
\begin{proof}[Proof of Proposition \ref{prop:count-extremely-dangerous-intervals}]
\par In fact, for every $q$-extremely dangerous interval $\Delta_q(\vect{a})$ where $l' \geq \eta' q$ and $\vect{a} = (a_0, a_1, \dots, a_n)^{\mathrm{T}}$, we have that
\begin{equation}
\label{equ:small-vector}
\|g_{\vect{r}}(q)U(\varphi(x))\vect{a}\| \leq \rho
\end{equation}
Let us fix $x \in \Delta_q(\vect{a})$. By direct computation, we have that
\[ g_{\vect{r}}(q) U(\varphi(x))\vect{a} = (v_0(x), v_1(x), \dots, v_n(x))^{\mathrm{T}} \]
where
\[v_0(x) = b^q (a_0 + a_1 \varphi_1 (x) + \cdots + a_n \varphi_n (x)),\]
and $v_i (x) = b^{-r_i q} a_i$ for $i =1, \dots, n$. Following the notation in Theorem \ref{thm:bernik-kleinbock-margulis}, let us denote
\[ f(x) = a_0 + a_1 \varphi_1 (x) + \cdots + a_n \varphi_n (x).\]
Then \eqref{equ:small-vector} implies that $|a_i| \leq \rho b^{r_i q}$ for $i=1, \dots, n$, and $|f(x)| \leq \rho b^{-q}$. Moreover, for any $x' \in [x- R^{-q(1-\eta')}, x + R^{-q(1-\eta')}]$, we have that
\[\|g_{\vect{r}}(q) U(\varphi(x'))\vect{a}\| \leq \rho.\]
By direct calculation, this implies that
\[|f(x')| \leq \rho b^{-q} \]
for any $x' \in [x - R^{-q(1-\eta')} , x + R^{-q(1-\eta')}]$. Let $x' = x + r R^{-q(1-\eta')}$ for some $r \in [ -1, 1]$. Then
\[
f(x') = f(x) + f'(x) r R^{-q(1-\eta')} + O(R^{-2q(1-\eta')}).
\]
Therefore, we have that for any $r \in [ -1,1]$,
\[
\begin{array}{rcl}
|f'(x) r R^{-q(1-\eta')}| & = & |f(x') - f(x) - O(R^{-2q(1-\eta')})| \\
& \leq & |f(x')| + |f(x)| + O(R^{-2q(1-\eta')}) \\
& \leq & \rho b^{-q} + \rho b^{-q} + \rho b^{-q} \leq b^{-q}.
\end{array}
\]
By letting $r = 1$, we have that
\[
|f'(x)| \leq R^{q(1-\eta')} b^{-q} = b^{q(r_1-\eta)}.
\]
The last equality above holds because $b^{1+r_1} = R$ and $\eta' = \frac{\eta}{1+r_1}$. This shows that
$x \in E_q$ for any $x \in \Delta_q(\vect{a})$, i.e., $\Delta_q(\vect{a}) \subset E_q$. Therefore, we have that
$D_q \subset E_q$. Then the conclusion follows from Theorem \ref{thm:bernik-kleinbock-margulis}.
\end{proof}
\par The rest of the section is devoted to the proof of Proposition \ref{prop:count-dangerous-intervals}. This is one of the main technical results of this paper.
\begin{proof}[Proof of Proposition \ref{prop:count-dangerous-intervals}]
\par Let us fix $I_p \in \mathcal{I}_p$. Let $I_p = [ x- R^{-q + 2l}, x + R^{-q + 2l}]$. We claim that we may assume that $\varphi(I_p)$ is a straight line. In fact, for any $x' \in I_p$, let us write $x' = x + r R^{-q + 2l}$ for some $r \in [-1, 1]$. By Taylor's expansion, we have that
\[\begin{array}{rcl}
g_{\vect{r}}(q)U(\varphi(x')) & = & g_{\vect{r}}(q) U(\varphi(x) + r R^{-q + 2l} \varphi^{(1)}(x) + O(R^{-2q + 4l})) \\
& = & g_{\vect{r}}(q) U(O(R^{-2q + 4l})) g_{\vect{r}}(-q) g_{\vect{r}}(q) U(\varphi(x) + r R^{-q + 2l} \varphi^{(1)}(x)) \\
& = & U(O(R^{-q + 4l})) g_{\vect{r}}(q) U(\varphi(x) + r R^{-q + 2l} \varphi^{(1)}(x)).
\end{array}\]
Since $l \leq \eta' q$, we have that $O(R^{-q + 4l})$ is exponentially small. Therefore, we may assume that $\varphi(x') = \varphi(x) + (x'-x) R^{-q + 2l} \varphi^{(1)}(s)$ for any $x' \in I_p$.
\par Let us take a typical $(q,l)$-dangerous interval $\Delta_{q,l}(\vect{a})$ that intersects $I_p$. Let us take $x \in \Delta_{q,l}(\vect{a})\cap I_p$ such that $x \in I_{q-1}$ for some $I_{q-1} \in \mathcal{I}_{q-1}$. It is easy to see that either $[x, x + R^{-q +l}] \subset \Delta_{q,l}(\vect{a})$, or $[x - R^{-q + l}, x] \subset \Delta_{q, l}(\vect{a})$. Without loss of generality, we may assume that $[x, x+R^{-q +l}] \subset \Delta_{q,l}(\vect{a})$. Let us write $\vect{a} = (a_0, a_1, \dots, a_n)^{\mathrm{T}} \in \mathbb{Z}^{n+1}\setminus\{\vect{0}\}$. For $x' \in [x, x + R^{-q +l}]$, let us denote
\[g_{\vect{r}}(q)U(\varphi(x'))\vect{a} = \vect{v}(x') = (v_0(x'), v_1(x'), \dots, v_n(x') )^{\mathrm{T}}. \]
Then we have that $\max\{ \|\vect{v}(x')\|: x' \in [x , x + R^{-q +l}] \} = c \rho $ for some $c \in [ 1/2, 1]$.
\par Recall that for $j =1, \dots, n$, $\lambda_j = \frac{1+ r_j}{1+ r_1}$. Let $1 \leq n' \leq n $ be the largest index $j$ such that $(1 -\lambda_j)q \leq l$.
\par For $x' \in [x , x + R^{-q + l}]$, let us write $x' = x + r R^{-q + l}$ for $r \in [0, 1]$. By our assumption we have that $\varphi(x') = \varphi(x) + r R^{-q + l} \varphi^{(1)}(x)$. Let us write
\[ \varphi^{(1)}(x) = (\varphi^{(1)}_1(x), \dots, \varphi^{(1)}_n(x))^{\mathrm{T}}.\]
By our standing assumption on $\varphi$ (Standing Assumption \ref{assumption-2}), we have that $c_1 \leq |\varphi^{(1)}_j(x)| \leq C_1$ for $j = 1, \dots, n$. By direct calculation, we have that
\[\begin{array}{rcl}
g_{\vect{r}}(q) U(\varphi(x'))\vect{a} & = & g_{\vect{r}}(q)U(\varphi(x') - \varphi(x)) g_{\vect{r}}(-q) g_{\vect{r}}(q) U(\varphi(x))\vect{a} \\
& = & g_{\vect{r}}(q)U(r R^{-q + l} \varphi^{(1)}(x)) g_{\vect{r}}(-q) \vect{v}(x) \\
& = & g_{\vect{r}}(q)U(r R^{-q + l} \varphi^{(1)}(x)) g_{\vect{r}}(-q) \vect{v}(x).
\end{array} \]
Let us write
\[ g_{\vect{r}}(q)U(r R^{-q + l} \varphi^{(1)}(x)) g_{\vect{r}}(-q) \vect{v}(x) = \vect{v}(x') = (v_0(x'), v_1(x'), \dots, v_n(x') )^{\mathrm{T}}, \]
where $x' = x + r R^{-q + l}$. By direct calculation, we have that
\[g_{\vect{r}}(q)U(r R^{-q +l} \varphi^{(1)}(x)) g_{\vect{r}}(-q) = U\left(r R^l \sum_{i=1}^n R^{-(1 - \lambda_i)q} \varphi^{(1)}_i(x) \vect{e}_i \right).\]
By our assumption, for $i \geq n' +1$, we have that $|r R^l R^{-(1-\lambda_i)q}| \leq 1$. Therefore, if we write
\[U\left( - r R^l \sum_{i=n'+1}^n R^{-(1 - \lambda_i)q} \varphi^{(1)}_i(x) \vect{e}_i \right)\vect{v}(x') = \vect{v}'(x') = (v'_0(x'), v'_1(x'), \dots, v'_n(x'))^{\mathrm{T}},\]
where $v'_0(x') = v_0(x') -r \sum_{i= n' +1}^n R^l R^{-(1-\lambda_i)q} \varphi^{(1)}_i(x) v_i(x')$ and
$v'_i(x') = v_i(x')$ for $i=1,\dots, n$. Then $|v'_0(x')| < C = (n+1)C_1 \rho $, and $|v'_i(x')| < \rho$ for $i = 1, \dots, n$. Let
\[\vect{h} = \sum_{i=1}^{n'} R^{-(1 - \lambda_i)q } \varphi^{(1)}_i(x) \vect{e}_i\]
and
\[\vect{h}_W = \sum_{i=1}^{n'} R^{-(1 - \lambda_i)q } \varphi^{(1)}_i(x) \vect{w}_i \in W.\]
Then $\|\vect{h}\|_2 = \|\vect{h}_W\|_2 \asymp 1$. For $ r \in [-1,1]$, and $x' = x + r R^{-q+l}$, our discussion above shows that
\[ U(r R^l \vect{h} ) \vect{v}(x) = (v'_0(x'), v'_1(x'), \dots, v'_n(x'))^{\mathrm{T}},\]
where $|v'_0(x')| < C $, and $|v'_i(x')| < \rho$ for $i=1,\dots, n$. Let $E_{n'}$ be the subspace of $\mathbb{R}^n$ spanned by $\{\vect{e}_1, \dots, \vect{e}_{n'}\}$ and $W'_{n'}$ be the subspace of $W$ spanned by $\{\vect{w}_1, \dots, \vect{w}_{n'}\}$. Then $\vect{h} \in E_{n'}$. Let $\mathfrak{k} \in \mathrm{SO}(n)$ be the element such that $\mathfrak{k}\cdot \vect{e}_1 = \vect{h}$, $\mathfrak{k}\cdot E_{n'} = E_{n'}$, and $\mathfrak{k} \cdot \vect{e}_i = \vect{e}_i$ for $i = n' +1 , \dots, n$. Let $z(\mathfrak{k}) = \begin{bmatrix}1 & \\ & \mathfrak{k} \end{bmatrix} \in Z$. It is easy to see that $z(\mathfrak{k}) \vect{w}_+ = \vect{w}_+$, $z(\mathfrak{k}) \vect{w}_1 = \vect{h}_W$, $z(\mathfrak{k}) W'_{n'} = W'_{n'}$, and $z(\mathfrak{k}) \vect{w}_i = \vect{w}_i$ for $i = n' +1, \dots, n$. By the definition of $z(\mathfrak{k})$ and our discussion in \S \ref{subsec-canonical-representation-sln}, we have that $U(\vect{h}) = z(\mathfrak{k}) U(\|\vect{h}\|_2\vect{e}_1)z^{-1}(\mathfrak{k})$. Therefore, we have that $U(\vect{h}) \vect{h}_W = \vect{h}_W + \|\vect{h}\|_2 \vect{w}_{+}$. Moreover, we have that $U(\vect{h}) \vect{w}_+ = \vect{w}_+$; for $i = 2, \dots, n'$, $U(\vect{h}) z(\mathfrak{k}) \vect{w}_i = z(\mathfrak{k})\vect{w}_i$; and for $i = n'+1, \dots, n$, $U(\vect{h}) \vect{w}_i = \vect{w}_i$. Let us write
\[\vect{v}(x) = a_+(x) \vect{w}_+ + \sum_{i=1}^{n'} a_i(x) z(\mathfrak{k})\vect{w}_i + \sum_{i= n' +1}^n a_i(x) \vect{w}_i.\]
Then the above discussion shows that
\[U(r R^l \vect{h})\vect{v}(x) = (a_+(x) + r R^l a_1(x)) \vect{w}_+ + \sum_{i=1}^{n'} a_i(x) z(\mathfrak{k})\vect{w}_i + \sum_{i= n' +1}^n a_i(x) \vect{w}_i.\]
By our previous argument, we have that there exists a constant $C >0$ such that $|a_i(x)| < C$ for $i = 1, \dots, n$ and
$|a_+(x) + r R^l a_1(x)| < C $ for any $r \in [ 0, 1]$. This implies that $|a_+(x)| < C $, and $|a_1(x)| < C R^{-l} $. Therefore, we have that $\vect{v}(x) \in z(\mathfrak{k}) ([-C, C] \times [-C R^{-l}, C R^{-l}] \times [-C, C]^{n-1})$.
\par Now let us estimate $|\mathcal{D}_{q,l}(I_p)|$.
\par Suppose that $\mathcal{D}_{q,l}(I_p) = \{\Delta_{q,l}(\vect{a}_u) : 1 \leq u \leq L\}$. For each $u = 1, \dots, L$, let us take $x_u \in \Delta_{q,l}(\vect{a}_u)\cap I_p$ such that $x_u \in I_{q-1, u}$ for some $I_{q-1,u} \in \mathcal{I}_{q-1}$. Let us denote
\[\vect{v}_u = g_{\vect{r}}(q)U(\varphi(x_u))\vect{a}_u.\]
Then by our previous argument, we have that
\[\vect{v}_u = a_{u, +} \vect{w}_+ + \sum_{i=1}^{n'} a_{u, i} z(\mathfrak{k})\vect{w}_i + \sum_{i=n'+1}^n a_{u,i} \vect{w}_i,\]
where $|a_{u, +}| < C $, $|a_{u,1}| < C R^{-l}$, and $ |a_{u, i} | < C $ for $ i = 2, \dots, n$.
\par Now let us consider $g_{\vect{r}}(q)U(\varphi(x_1))\vect{a}_u$. Let us write $x_u = x_1 - r R^{-q + 2l}$ for some $r \in [-1,1]$. Using our assumption that $\varphi(I_p)$ is a straight line, we have that
\[\begin{array}{cl}
& g_{\vect{r}}(q)U(\varphi(x_1))\vect{a}_u \\
= & g_{\vect{r}}(q) U(\varphi(x_1) - \varphi(x_u)) g_{\vect{r}}(-q) g_{\vect{r}}(q) U(\varphi(x_u))\vect{a}_u \\
= & g_{\vect{r}}(q) U(\varphi(x_1) - \varphi(x_u)) g_{\vect{r}}(-q) \vect{v}_u \\
= & g_{\vect{r}}(q) U(r R^{-q + 2l} \varphi^{(1)}(x)) g_{\vect{r}}(-q) \vect{v}_u \\
= & U\left(r R^{2l} \sum_{i= 1}^n R^{-(1-\lambda_i)q} \varphi^{(1)}_i (x) \vect{e}_i \right) \vect{v}_u.
\end{array}\]
Let us denote $\vect{h} = \sum_{i=1}^{n'} R^{-(1-\lambda_i)q} \varphi^{(1)}_i (x) \vect{e}_i$ as before. Then by our previous argument, we have that
\[\begin{array}{rcl}
g_{\vect{r}}(q)U(\varphi(x_1))\vect{a}_u & = & U(r R^{2l} \vect{h} + r R^{2l}\sum_{i = n' +1}^n R^{-(1-\lambda_i)q} \varphi^{(1)}(x) \vect{e}_i)\vect{v}_u. \\
& = & \left(a_{u, +} + r R^{2l} a_{u, 1} + r R^{2l}\sum_{i= n' +1}^n R^{-(1-\lambda_i)q} \varphi^{(1)}(x) a_{u, i} \right) \vect{w}_{+} \\
& & + \sum_{i=1}^{n'} a_{u, i} z(\mathfrak{k})\vect{w}_i + \sum_{i=n'+1}^n a_{u,i} \vect{w}_i.
\end{array}\]
Since $|a_{u,1}| \leq C R^{-l}$, and since for $i = n'+1 , \dots, n$, $(1-\lambda_i) q > l$, $|a_{u, i}| <C$, and $|\varphi^{(1)}_i(x)|\leq C_1$, we have that
\[\begin{array}{cl} & \left| a_{u, +} + r R^{2l} a_{u, 1} + r R^{2l}\sum_{i= n' +1}^n R^{-(1-\lambda_i)q} \varphi^{(1)}(x) a_{u, i} \right| \\
\leq & |a_{u, +}| + |r| R^{2l } |a_{u,l}| + |r| R^{2l } \sum_{i=n'+1}^n R^{-(1-\lambda_i)q} | \varphi^{(1)}(x)| |a_{u,i}| \\
\leq & C + R^{2l } C R^{-l} + R^{2l } \sum_{i = n' + 1}^n R^{-l} C_1 C \\
\leq & C + R^{2l } C R^{-l } + R^{2l } n R^{-l} C_1 C \\
\leq & C_2 R^{l} \\
\end{array}\]
where $ C_2 = 2C + n C_1 C >0$. This implies that for any $u = 1, \dots, L$, we have that
\[g_{\vect{r}}(q)U(\varphi(x_1))\vect{a}_u \in z(\mathfrak{k})([-C_2 R^l, C_2 R^l] \times [-C R^{-l}, C R^{-l}] \times [-C, C]^{n-1}).\]
Let us consider the range of $g_{\vect{r}}(q -l) U(\varphi(x_1))\vect{a}_u= g_{\vect{r}}(-l) g_{\vect{r}}(q)U(\varphi(x_1))\vect{a}_u$. Let us write $g_{\vect{r}}(-l) = d_2 (l) d_1(l)$ where
\[d_1(l) = \begin{bmatrix} b^{-l} & & & & \\ & b^{r_1 l} \mathrm{I}_{n_1} & & & \\ & & b^{r_{n' + 1} l} & & \\ & & & \ddots & \\
& & & & b^{r_n l}\end{bmatrix},\]
and
\[
d_2(l) = \begin{bmatrix} 1 & & & & & \\ & 1 & & & & \\ & & b^{-( r_1 - r_2)l} & & & \\ & & & \ddots & & \\
& & & & b^{-(r_1 - r_{n'})l} & \\ & & & & & \mathrm{I}_{n - n'} \end{bmatrix}.
\]
Then we have that
\[g_{\vect{r}}(q -l) U(\varphi(x_1))\vect{a}_u \in d_2(l) d_1(l) z(\mathfrak{k})([-C_2 R^l, C_2 R^l] \times [-C R^{-l}, C R^{-l}] \times [-C, C]^{n-1}).\]
By the definition of $z(\mathfrak{k})$, we have that $d_1(l) z(\mathfrak{k}) = z(\mathfrak{k}) d_1(l)$. Therefore, we have that
\[\begin{array}{cl} & d_1(l) z(\mathfrak{k})([-C_2 R^l, C_2 R^l] \times [-C R^{-l}, C R^{-l}] \times [-C, C]^{n-1}) \\
= & z(\mathfrak{k}) d_1 (l) ([-C_2 R^l, C_2 R^l] \times [-C R^{-l}, C R^{-l}] \times [-C, C]^{n-1}) \\
= & z(\mathfrak{k}) ([-C_2 b^{r_1 l} , C_2 b^{r_1 l}] \times [-C b^{-l} , C b^{-l} ]\times [-C b^{r_1 l} , C b^{r_1 l}]^{n_1 -1} \times \prod_{i=n'+1}^n [-C b^{r_i l} , C b^{r_i l}]) \\
\subset & z(\mathfrak{k})([-C_2 b^{r_1 l} , C_2 b^{r_1 l}] \times [-1 , 1 ]\times [-C b^{r_1 l} , C b^{r_1 l}]^{n' -1} \times \prod_{i=n'+1}^n [-C b^{r_i l} , C b^{r_i l}]). \end{array}\]
It is easy to see that
\[z(\mathfrak{k})([-C_2 b^{r_1 l} , C_2 b^{r_1 l}] \times [-1 , 1 ]\times [-C b^{r_1 l} , C b^{r_1 l}]^{n' -1} \times \prod_{i=n'+1}^n [-C b^{r_i l} , C b^{r_i l}])\]
can be covered by a collection $\mathcal{B}$ of $ O(b^{\lambda l})$ balls of radius $1$ where $\lambda = n' r_1 + \sum_{i = n' +1}^n r_i$. Then we have that
\[ \begin{array}{rcl} g_{\vect{r}}(q-l)U(\varphi(x_1))\vect{a}_u & \in & d_2(l)\bigcup_{B \in \mathcal{B}} B \\
& = & \bigcup_{B \in \mathcal{B}} d_2(l)B. \end{array}\]
Since $d_2(l)$ is a contracting map, for every $B \in \mathcal{B}$, there exists a ball $B'$ of radius $C$ such that
$d_2(l) B \subset B'$. Let $\mathcal{B}'$ denote the collection of all such $B'$'s. Then we have that
\[g_{\vect{r}}(q-l)U(\varphi(x_1))\vect{a}_u \in \bigcup_{B' \in \mathcal{B}'} B'.\]
Since $g_{\vect{r}}(q-l)U(\varphi(x_1))\vect{a}_u \in g_{\vect{r}}(q-l)U(\varphi(x_1))\mathbb{Z}^{n+1}$, we have that
\[g_{\vect{r}}(q-l)U(\varphi(x_1))\vect{a}_u \in \bigcup_{B' \in \mathcal{B}'} B' \cap \Lambda,\]
where $\Lambda = g_{\vect{r}}(q-l)U(\varphi(x_1))\mathbb{Z}^{n+1}$. By our assumption, $x_1 \in I_{q-1, 1}$ for some $I_{q-1, 1} \in \mathcal{I}_{q-1}$. This implies that $x_1 \in I_{q-l}$ for some $I_{q-l} \in \mathcal{I}_{q-l}$. Therefore, $\Lambda = g_{\vect{r}}(q-l) U(\varphi(x_1))\mathbb{Z}^{n+1} \in K_{\kappa}$, i.e., $\Lambda$ does not contain any nonzero vectors with norm $\leq \kappa$. Therefore, there exists a constant $C_4$ such that every ball of radius $1$ contains at most $C_4 \kappa^{-n-1} = C_4 R^{(n+1)m}$ points in $\Lambda$. Thus, we have that
\[\begin{array}{rcl} |\mathcal{D}_{q,l}(I_p)| = |\{g_{\vect{r}}(q-l)U(\varphi(x_1))\vect{a}_u : 1 \leq u \leq L\}| & \leq & \sum_{B' \in \mathcal{B}'} |B'\cap \Lambda| \\
& \leq & \sum_{B' \in \mathcal{B}'} C_4 R^{(n+1)m} \\
& \leq & C_5 b^{\lambda l + 4n m} \leq C_5 b^{(\lambda + \frac{1}{200n})l}, \end{array}\]
where $C_5 = C_3 C_4$ and $\lambda = n' r_1 + \sum_{i= n' +1}^n r_i$. Now let us estimate $\lambda$. In fact,
\[\begin{array}{rl}\lambda & = \sum_{i=1}^n r_i + \sum_{i=1}^{n'} (r_1 - r_i)\\
& = 1 + \sum_{i=1}^{n'} (r_1 - r_i) .\end{array}\]
By our assumption, for $i=1, \dots, n'$, we have that $r_1 - r_i \leq \frac{l}{q} \leq \frac{1}{100 n^2}$. Therefore, we have that
\[\lambda \leq 1 + n \frac{1}{100n^2} = 1+ \frac{1}{100 n}.\]
Thus, we have that
\[|\mathcal{D}_{q,l}(I_p)| \leq C_5 b^{(1+ \frac{1}{100n} + \frac{1}{200n})l} \leq C_5 R^{(1- \frac{1}{10n})l}.\]
The last inequality above holds because $b = R^{\frac{1}{1+r_1}} \leq R^{\frac{n}{n+1}}$.
\par This completes the proof.
\end{proof}
\section{Proof of the main result}
\label{sec-proof-main-result}
\par In this section we will finish the proof of Proposition \ref{prop:distance-sequence-small}. By our discussion in \S \ref{sec-introduction} and \S \ref{sec:cantor-like-construction}, Proposition \ref{prop:distance-sequence-small} implies Theorem \ref{thm:main-task} and Theorem \ref{thm:main-thm}.
\par The structure of the section is as follows. In the first subsection, we will prove Proposition \ref{prop:distance-sequence-small} for the case $q \leq 10^6 n^4 N m$. The second, third and fourth subsections are devoted to the proof for the case $q > 10^6 n^4 N m$. The key point is to estimate $F(\hat{\mathcal{I}}_{q, p}, I_p)$ for $I_p \in \mathcal{I}_p$. The second subsection deals with the case $p = q - 4000 n^2 N m$. The third subsection deals with the case $p = q - 2l$ where $ 2000 n^2 N m < l < 2 \eta' q$. The fourth subsection deals with the case $p = 0$.
\par The third and fourth subsections contain some technical results on the canonical representation of $\mathrm{SL}(n+1, \mathbb{R})$ on $\bigwedge^i V$ for $i = 2, \dots, n$. They are also main technical contributions of this paper.
\par Our basic tool is the following non-divergence theorem due to Kleinbock and Margulis:
\begin{theorem}[see~{\cite[Theorem 5.2]{Klein_Mar}}]
\label{thm:non-divergence-kleinbock-margulis}
There exist constants $C, \alpha >0$ such that the following holds: For any subinterval $J \subset I$ and any $t \geq 0$, if for any $i= 1,2,\dots, n$ and any $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$,
\[\max\{ \|g_{\vect{r}}(t)U(\varphi(x))\vect{v}\|: x \in J \} \geq \rho^i,\]
then for any $\epsilon >0$,
\[ m \left( \{x: g_{\vect{r}}(t) U(\varphi(x)) \mathbb{Z}^{n+1} \notin K_{\epsilon} \} \right) \leq C \left( \frac{\epsilon}{\rho} \right)^{\alpha} m(J). \]
\end{theorem}
\begin{remark}
\label{rmk:non-divergence}
The original version of Theorem \ref{thm:non-divergence-kleinbock-margulis} is more general. The version above is tailored for our setting. For the case where $\varphi(x)$ is a polynomial, the statement is proved by Dani \cite{Dani}.
\end{remark}
\par We will also need the following result due to Nimish Shah \cite{Shah2010} on $\mathrm{SL}(n+1, \mathbb{R})$ representations.
\begin{theorem}[see~{\cite[Proposition 4.9]{Shah2010}}]
\label{thm:shah-sln-representations}
Let $\mathcal{V}$ be any finite dimensional representation of $\mathrm{SL}(n+1, \mathbb{R})$ with a norm $\|\cdot\|$ and let $\vect{r}$ be any $n$-dimensional weight. There exists a constant $c >0$ such that for any nonzero vector $v \in \mathcal{V}$ and any $t \geq 0$,
\[ \max\{ \|g_{\vect{r}}(t)U(\varphi(x))v\| : x \in I \} \geq c \|v\|.\]
\end{theorem}
\begin{remark}
\label{rmk:shah-thm}
\par The exact statement in \cite[Proposition 4.9]{Shah2010} is different from the above version, but it easily implies the above statement.
\par The proof makes use of the fact that $\varphi$ is not contained in any proper affine subspace in $\mathbb{R}^n$.
\end{remark}
\par We may choose $0< \rho <1$ small enough such that $\rho^{n+1} < c$ where $c$ denotes the constant from Theorem \ref{thm:shah-sln-representations}.
\subsection{The case where $q$ is small}
\label{subsec-case-q-small}
\par In this subsection, let us assume that $q \leq 10^6 n^4 N m$. Then only $\hat{\mathcal{I}}_{q,0} = \hat{I}_q$ is nonempty.
\begin{proposition}
\label{prop:small-portion-q-small}
\[ F(\hat{\mathcal{I}}_{q, 0}, I) \ll R^{q - \alpha m}.\]
\end{proposition}
\begin{proof}
By Theorem \ref{thm:shah-sln-representations} and our assumption on $\rho$, we have that for any $i = 1, 2,\dots, n$ and $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus\{\vect{0}\}$,
\[\max\{ \|g_{\vect{r}}(q) U(\varphi(x)) \vect{v}\| : x \in I \} \geq c \|\vect{v}\| \geq c \geq \rho^i.\]
Then by Theorem \ref{thm:non-divergence-kleinbock-margulis}, we have that
\[ m(\{x \in I: g_{\vect{r}}(q)U(\varphi(x))\mathbb{Z}^{n+1} \notin K_{2\kappa} \}) \leq C \left(\frac{2\kappa}{\rho}\right)^{\alpha}. \]
\par On the other hand, it is easy to see that $g_{\vect{r}}(q)U(\varphi(I_q)) \mathbb{Z}^{n+1} \subset X \setminus K_{2 \kappa} $ for any $I_q \in \hat{\mathcal{I}}_q$. Therefore, we have that
\[
\begin{array}{rcl}
F(\hat{\mathcal{I}}_{q,0}, I) R^{-q} & = & m\left(\bigcup_{I_q \in \hat{\mathcal{I}}_q} I_q \right) \\
& = & m(\{x \in I: g_{\vect{r}}(q)U(\varphi(x))\mathbb{Z}^{n+1} \notin K_{2\kappa} \}) \leq C_6 \kappa^{\alpha} = C_6 R^{-\alpha m}
\end{array}
\]
where $C_6 = C \left( \frac{2}{\rho} \right)^{\alpha}$. This finishes the proof.
\end{proof}
Let us choose $M >1$ such that $M^{\alpha} > 1000^{10^6 n^4 N}$.
\begin{proof}[Proof of Proposition \ref{prop:distance-sequence-small} for $q \leq 10^6 n^4 N m$]
\par It suffices to show that
\[ \left( \frac{4}{R} \right)^q F(\hat{\mathcal{I}}_{q,0}, I) \]
can be arbitrarily small. In fact, by Proposition \ref{prop:small-portion-q-small}, we have that
\[
\begin{array}{rcl}
\left( \frac{4}{R} \right)^q F(\hat{\mathcal{I}}_{q,0}, I) & = & \left( \frac{4}{R} \right)^q O(R^{q - \alpha m}) \\
& = & O(\frac{4^q}{ R^{\alpha m}}) = O(\frac{4^{10^6 n^4 N m}}{R^{\alpha m}}) = O(\left(\frac{4}{1000}\right)^{10^6 n^4 N m}) .
\end{array}
\]
Then it is easy to see that $\left( \frac{4}{R} \right)^q F(\hat{\mathcal{I}}_{q,0}, I) \to 0$ as $m \to \infty$.
\par This completes the proof for $q \leq 10^6 n^4 N m$.
\end{proof}
\subsection{The generic case}
\par The rest of the section is devoted to the proof of Proposition \ref{prop:distance-sequence-small} for $q > 10^6 n^4 N m$. In the following subsections, we will estimate $F(\hat{\mathcal{I}}_{q,p}, I_p)$ for different $p$'s. In this subsection we will estimate $F(\hat{\mathcal{I}}_{q,p}, I_p)$ for $p = q - 4000 n^2 N m$. We call it the generic case.
\begin{proposition}
\label{prop:small-portion-q-large-p-large}
Let $q > 10^6 n^4 N m$ and $p = q- 4000 n^2 N m$. Then for any $I_p \in \mathcal{I}_p$, we have that
\[F(\hat{\mathcal{I}}_{q,p}, I_p) \ll R^{q-p - \alpha m} .\]
\end{proposition}
\begin{proof}
\par Let us fix $I_p \in \mathcal{I}_p$. Let $I_p = [a, b]$ and $I'_p = [a-R^{-q+ 2000 n^2 N m}, b+R^{-q+ 2000 n^2 N m}] $. It is easy to see that $I_p \subset I'_p$ and $m(I'_p) < 2 m(I_p)$.
\par If $F(\hat{\mathcal{I}}_{q,p}, I_p) = 0$, then the statement apparently holds.
\par Suppose $F(\hat{\mathcal{I}}_{q,p}, I_p) >0$, let us take $I_q \in \hat{\mathcal{I}}_{q,p}$ and $x \in I_q\cap I_p$.
Then for any $i = 1, \dots, n$ and $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1}\setminus\{\vect{0}\}$, we have that
\[\max\{\|g_{\vect{r}}(q) U(\varphi(x')) \vect{v}\|: x' \in [x- R^{-q+ 2000 n^2 N m}, x + R^{-q + 2000 n^2 N m}]\} \geq \rho^i.\]
It is easy to see that $[x- R^{-q+ 2000 n^2 N m}, x + R^{-q + 2000 n^2 N m}] \subset I'_p$. Therefore, we have that
\[\max\{ \|g_{\vect{r}}(q)U(\varphi(x')) \vect{v}\| : x' \in I'_p \} \geq \rho^i.\]
By Theorem \ref{thm:non-divergence-kleinbock-margulis}, we have that
\[m(\{ x \in I'_p: g_{\vect{r}}(q)U(\varphi(x))\mathbb{Z}^{n+1} \notin K_{2\kappa}\}) \leq C \left(\frac{2\kappa}{\rho}\right)^{\alpha} m(I'_p). \]
This implies that
\[m(\{ x \in I_p: g_{\vect{r}}(q)U(\varphi(x))\mathbb{Z}^{n+1} \notin K_{2\kappa}\}) \leq C \left(\frac{2\kappa}{\rho}\right)^{\alpha} m(I'_p) \leq 2 C \left(\frac{2\kappa}{\rho}\right)^{\alpha} m(I_p). \]
\par On the other hand, it is easy to see that $g_{\vect{r}}(q) U(\varphi(I_q))\mathbb{Z}^{n+1} \subset X \setminus K_{2\kappa}$ for any $I_q \in \hat{\mathcal{I}}_q$. Therefore we have that
\[\begin{array}{cl}
& F(\hat{\mathcal{I}}_{q, p}, I_p) R^{-q} \\
\leq & m(\{ x \in I_p: g_{\vect{r}}(q)U(\varphi(x))\mathbb{Z}^{n+1} \notin K_{2\kappa}\}) \\
\leq & 2 C \left( \frac{2\kappa}{\rho} \right)^{\alpha} m(I_p) \\
= & 2 C \left(\frac{2}{\rho}\right)^{\alpha} \kappa^{\alpha} R^{-p} = C_7 R^{-p - \kappa m}
\end{array}\]
where $C_7 = 2 C \left(\frac{2}{\rho}\right)^{\alpha}$. This proves the statement.
\end{proof}
\par By Proposition \ref{prop:small-portion-q-large-p-large}, we have that for $p = q - 4000 n^2 N m$ and any $I_p \in \mathcal{I}_p$, the following holds:
\begin{equation}
\label{equ:q-large-p-large}
\begin{array}{rcl}
\left(\frac{4}{R}\right)^{q-p} F(\hat{\mathcal{I}}_{q,p}, I_p) & \ll & \left(\frac{4}{R}\right)^{q-p} R^{q-p - \alpha m} \\
& = & \frac{4^{4000 n^2 N m}}{R^{\alpha m}} = \left( \frac{4}{1000}\right)^{4000 n^2 N m}.
\end{array}
\end{equation}
Then it is easy to see that $\left(\frac{4}{R}\right)^{q-p} F(\hat{\mathcal{I}}_{q,p}, I_p) \to 0$ as $ m \to \infty$.
\subsection{Dangerous case}
\label{subsec-dangerous-case}
\par In this subsection, we will consider the case where $2000 n^2 N m < l < 2\eta' q$ and $p = q - 2l$. We call this case the $(q,l)$-dangerous case.
\begin{proposition}
\label{prop:small-portion-dangerous-case}
For any $I_p \in \mathcal{I}_p$, we have that
\[ F(\hat{\mathcal{I}}_{q,p}, I_p) \ll R^{ q - p -\frac{l}{20n} }. \]
\end{proposition}
\par Let us recall that for $1000 n^2 N m < l' < \eta' q$, a $(q,l')$-dangerous interval $\Delta_{q,l'}(\vect{a})$ associated with a nonzero integer vector $\vect{a} \in \mathbb{Z}^{n+1}$ is a closed interval of the form
\[\Delta_{q,l'}(\vect{a}) = [ x - R^{-q + l'}, x + R^{-q + l'}]\]
such that $I_q \subset \Delta_{q,l'}(\vect{a})$ for some $I_q \in \hat{\mathcal{I}}_q$ and
\[ \max\{ \|g_{\vect{r}}(q)U(\varphi(x'))\vect{a}\|: x' \in \Delta_{q,l'}(\vect{a}) \} = c \rho \]
for some $c \in [ 1/2, 1]$.
\par The following lemma is crucial to prove Proposition \ref{prop:small-portion-dangerous-case} and is one of the main technical contributions of this paper:
\begin{lemma}
\label{lm:key-lemma}
For any $i = 1, \dots, n$ and $I_q \in \hat{\mathcal{I}}_{q,p}(i)$ intersecting $I_p$, either there exists a $(q,l')$-dangerous interval $\Delta_{q, l'} (\vect{a})$ containing $I_q$ for some $ l/2 \leq l' \leq l$, or there exists $x \in I_q$ and
\[\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge\nolimits^i \mathbb{Z}^{n+1}\setminus \{\vect{0}\}\]
such that if we write
\[ g_{\vect{r}}(q)U(\varphi(x))\vect{v} = \vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)} \]
where $\vect{w}^{(i-1)} \in \bigwedge^{i-1} W$ and $\vect{w}^{(i)} \in \bigwedge^i W$, then we have that $\|\vect{w}_+ \wedge \vect{w}^{(i-1)}\| = \| \vect{w}^{(i-1)} \| \leq \rho^i$ and $\|\vect{w}^{(i)}\| \leq \rho^i R^{-l/2}$.
\end{lemma}
\begin{proof}
\par If $i = 1$, then the first statement apparently holds. We may assume that $i \geq 2$.
\par By the definition of $\hat{\mathcal{I}}_{q,p}(i)$, there exists $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$ such that for any $x \in I_q$,
\[ \max\{ \|g_{\vect{r}}(q)U(\varphi(x')) \vect{v}\|: x' \in [x- R^{-q+l}, x+ R^{-q +l} ] \} = c \rho^i \]
for some $c \in [1/2, 1]$.
\par Without loss of generality, we may assume that the sublattice $L_i$ generated by $\{\vect{v}_1, \dots , \vect{v}_i \}$ is a primitive $i$-dimensional sublattice of $\mathbb{Z}^{n+1}$. Then $\Lambda_i = g_{\vect{r}}(q) U(\varphi(x)) L_i$ is a primitive $i$-dimensional sublattice of $\Lambda = g_{\vect{r}}(q) U(\varphi(x))\mathbb{Z}^{n+1}$. For simplicity, let us denote $g = g_{\vect{r}}(q)U(\varphi(x))$. Let us choose the Minkowski reduced basis $\{g\vect{v}'_1, \dots, g\vect{v}'_i\}$ of $\Lambda_i$. Since
\[ d(\Lambda_i) = \|g\vect{v}\| \leq \rho^i,\]
we have that $\|g\vect{v}'_1\| \leq \rho$ by the Minkowski Theorem.
\par Let us follow the argument in the proof of Proposition \ref{prop:count-dangerous-intervals}. Recall that
for $j = 1,\dots, n $, $\lambda_j = \frac{1+ r_j}{1+ r_1}$. Let $1 \leq n' \leq n$ be the largest index $j$ such that $(1- \lambda_j) q \leq l$. Let us write
\[\varphi(s) = (\varphi_1(s), \dots, \varphi_n(s))^{\mathrm{T}}.\]
By Standing Assumption \ref{assumption-2}, we have that $c_1 \leq |\varphi^{(1)}_i (s)| \leq C_1$ for any $i = 1, \dots, n$ and $s \in I$. Let $\vect{h} = \sum_{i=1}^{n'} R^{-(1-\lambda_i)q} \varphi^{(1)}(s)\vect{e}_i$. For any $x' \in [ x - R^{-q + l}, x + R^{-q + l}]$, let us write $x' = x + r R^{-q + l}$ where $ r \in [-1, 1]$. By the same argument as in the proof of Proposition \ref{prop:count-dangerous-intervals}, we have that
\[ g_{\vect{r}}(q)U(\varphi(x')) = U(O(1)) U(r R^l \vect{h}) g_{\vect{r}}(q)U(\varphi(x)) = U(O(1)) U(r R^l \vect{h}) g. \]
Therefore, we have that
\[ \| U(r R^l \vect{h}) g\vect{v} \| \leq \rho^i \]
for any $r \in [-1,1]$.
\par Following the notation in the proof of Proposition \ref{prop:count-dangerous-intervals}, let us denote
$\vect{h} = \mathfrak{k} \cdot \vect{e}_1$ for $\mathfrak{k} \in \mathrm{SO}(n)$ and
\[ z(\mathfrak{k}) = \begin{bmatrix}1 & \\ & \mathfrak{k} \end{bmatrix} \in Z.\]
For $ j = 1, \dots, i$, let us write
\[g \vect{v}'_j = a_+ (j) \vect{w}_+ + a_1(j) z(\mathfrak{k})\vect{w}_1 + \vect{w}'(j) \]
where $\vect{w}' (j) \in z(\mathfrak{k}) W_2$. Then
\[
\begin{array}{rcl}
g \vect{v} & = & (g\vect{v}'_1)\wedge \cdots \wedge (g\vect{v}'_i) \\
& = & \bigwedge_{j=1}^i (a_+ (j) \vect{w}_+ + a_1(j) z(\mathfrak{k})\vect{w}_1 + \vect{w}'(j) ) \\
& = & \vect{w}_+ \wedge (z(\mathfrak{k})\vect{w}_1) \wedge \left( \sum_{j <j'} \epsilon_{+,1}(j,j') a_+(j) a_1(j') \bigwedge_{k \neq j,j'} \vect{w}'(k) \right) \\
& & + \vect{w}_+ \wedge \left( \sum_{j=1}^i \epsilon_+(j) a_+(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) + (z(\mathfrak{k})\vect{w}_1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \\
& & + \bigwedge_{j=1}^i \vect{w}'(j)
\end{array}
\]
where $\epsilon_{+,1}(j,j'), \epsilon_+(j), \epsilon_1(j) \in \{\pm 1\}$ for every $j,j' \in \{1, \dots, i\}$. By our discussion in \S \ref{subsec-canonical-representation-sln} on the representation of $\mathrm{SL}(2, \vect{h})$ on $\bigwedge^i V$, we have that
\[
\begin{array}{rcl}
U( r R^l \vect{h})g \vect{v} & = & \vect{w}_+ \wedge (z(\mathfrak{k})\vect{w}_1) \wedge \left( \sum_{j <j'} \epsilon_{+,1}(j,j') a_+(j) a_1(j') \bigwedge_{k \neq j,j'} \vect{w}'(k) \right) \\
& & + \vect{w}_+ \wedge \left( \sum_{j=1}^i \epsilon_+(j) a_+(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \\
& & + r R^l \vect{w}_+ \wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \\
& & + (z(\mathfrak{k})\vect{w}_1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) + \bigwedge_{j=1}^i \vect{w}'(j) .
\end{array}
\]
Since $\|U(r R^l \vect{h}) g \vect{v}\| \leq \rho^i$ for any $r \in [-1,1]$, we have that
\[ \left\|\sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k)\right\| \leq \rho^i R^{-l}. \]
Let us consider the following two cases:
\begin{enumerate}
\item $|a_1(1)| \leq R^{-l/2}$.
\item $|a_1(1)| > R^{-l/2}$.
\end{enumerate}
\par Let us first suppose $|a_1(1)| \leq R^{-l/2}$. Note that $\|g\vect{v}'_1\| \leq \rho$. Then by repeating the calculation in the proof of Proposition \ref{prop:count-dangerous-intervals}, we conclude that
\[ \max\{ \|g_{\vect{r}}(q) U(\varphi(x')) \vect{v}'_1\| : x' \in [ x- R^{-q + l/2}, x + R^{-q + l/2} ] \} \leq \rho. \]
On the other hand, by our definition on $\hat{\mathcal{I}}_{q, p}(i)$, we have that
\[\max\{ \|g_{\vect{r}}(q)U(\varphi(x'))\vect{v}'_1\|: x' \in [ x - R^{-q+l}, x + R^{-q+l}] \} \geq \frac{1}{2}\rho. \]
This implies that $I_q \subset \Delta_{q,l'}(\vect{v}'_1)$ for some $l/2 \leq l' \leq l$. This proves the first part of the statement.
\par Now let us suppose $|a_1(1)| > R^{-l/2}$. Then we have that
\[
\begin{array}{rcl}
\epsilon_1(1) a_1(1) \bigwedge_{j=1}^i \vect{w}'(j) & = & \vect{w}'(1)\wedge \left( \epsilon_1(1) a_1(1)
\bigwedge_{k\neq 1} \vect{w}'(k) \right) \\
& = & \vect{w}'(1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right).
\end{array}
\]
Therefore, we have that
\[
\begin{array}{rcl}
|a_1 (1)| \left\|\bigwedge_{j=1}^i \vect{w}'(j)\right\| & = & \left\| \vect{w}'(1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \right\| \\
& \leq & \|\vect{w}'(1)\| \left\| \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right\| \\
& \leq & \rho \cdot \rho^i R^{-l} = \rho^{i+1} R^{-l}.
\end{array}
\]
Since $|a_1(1)| > R^{-l/2}$ and $\rho < 1$, we have that
\[ \left\|\bigwedge_{j=1}^i \vect{w}'(j)\right\| \leq \rho^{i} R^{-l/2}. \]
If we write
\[ g \vect{v} = \vect{w} \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)} \]
where $\vect{w}^{(i-1)} \in \bigwedge^{i-1} W$ and $\vect{w}^{(i)} \in \bigwedge^i W$, then
\[ \vect{w}^{(i)} = (z(\mathfrak{k})\vect{w}_1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) + \bigwedge_{j=1}^i \vect{w}'(j). \]
By our previous argument, we have that
\[ \|\vect{w}^{(i)}\| \leq \rho^i R^{-l/2}. \]
This proves the second part of the statement.
\end{proof}
\par The following lemma takes care of the second case of Lemma \ref{lm:key-lemma}.
\begin{lemma}
\label{lm:2nd-case-key-lemma}
Let $i \in \{2,\dots, n \}$. Let $\mathcal{D}_{q,p}(I_p, i)$ denote the collection of $I_q \in \hat{\mathcal{I}}_{q,p}$ intersecting $I_p$ and not contained in any $(q,l')$-dangerous interval for any $l/2 \leq l' \leq l$. Let
\[D_{q,p}(I_p, i) := \bigcup_{I_q \in \mathcal{D}_{q,p}(I_p, i)} I_q.\]
Then for any closed subinterval $J \subset I_p$ of length $R^{-q + (1+ \frac{1}{2n})l}$, we have that
\[ m(D_{q,p}(I_p, i) \cap J) \ll R^{-\frac{l}{20n}} m(J). \]
\end{lemma}
\begin{proof}
\par Let us fix a closed subinterval $J \subset I_p$ of length $R^{-q + (1+ \frac{1}{2n})l}$.
\par For any $x \in I_q \in \mathcal{D}_{q,p}(I_p, i)$, there exists $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$ such that
\[ \max \{ \|g_{\vect{r}}(q) U(\varphi(x'))\vect{v}\| : x' \in [ x - R^{-q+l}, x + R^{-q + l}] \} \leq \rho^i.\]
\par Let us denote the interval $[x - R^{-q+l}, x + R^{-q+l}]$ by $\Delta_{q,l}(\vect{v}, i)$. Then every $I_q \in \mathcal{D}_{q,l}(I_p,i)$ is contained in some $\Delta_{q,l}(\vect{v},i)$ and every $\Delta_{q,l}(\vect{v}, i)$ contains at most $O(R^l)$ different $I_q \in \mathcal{D}_{q,l}(I_p, i)$.
\par We will follow the notation used in the proof of Lemma \ref{lm:key-lemma}. Let $g = g_{\vect{r}}(q)U(\varphi(x))$, $\vect{h} = \mathfrak{k} \cdot \vect{e}_1$ and
\[z(\mathfrak{k}) = \begin{bmatrix} 1 & \\ & \mathfrak{k} \end{bmatrix} \in Z\]
be as in the proof of Lemma \ref{lm:key-lemma}.
. For $j = 1 , \dots, i$, let us write
\[\begin{array}{rcl}
g\vect{v}_j & = & a_+(j)\vect{w}_+ + a_1(j) z(\mathfrak{k})\vect{w}_1 + \vect{w}'(j) \\
& = & a_+(j)\vect{w}_+ + \vect{w}(j)
\end{array}\]
where $\vect{w}'(j) \in z(\mathfrak{k})W_2$ and $\vect{w}(j) = a_1(j) z(\mathfrak{k})\vect{w}_1 + \vect{w}'(j) \in W$. Then
\[
\begin{array}{rcl}
g \vect{v} & = & \vect{w}_+ \wedge (z(\mathfrak{k})\vect{w}_1) \wedge \left( \sum_{j <j'} \epsilon_{+,1}(j,j') a_+(j) a_1(j') \bigwedge_{k \neq j,j'} \vect{w}'(k) \right) \\
& & + \vect{w}_+ \wedge \left( \sum_{j=1}^i \epsilon_+(j) a_+(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \\
& & + (z(\mathfrak{k})\vect{w}_1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) + \bigwedge_{j=1}^i \vect{w}'(j) .
\end{array}
\]
By Lemma \ref{lm:key-lemma}, we have that
\[ \left\| (z(\mathfrak{k})\vect{w}_1)\wedge \left( \sum_{j=1}^i \epsilon_1(j) a_1(j) \bigwedge_{k \neq j} \vect{w}'(k) \right) \right\| \leq \rho^i R^{-l}\]
and
\[
\left\| \bigwedge_{j=1}^i \vect{w}'(j) \right\| \leq \rho^i R^{-l/2}.
\]
Let us take the collection of all possible $\Delta_{q,l}(\vect{v}, i)$'s intersecting $J$, say
\[ \{ \Delta_{q,l}(\vect{v}(M), i) = [ x(M) - R^{-q +l}, x(M) + R^{-q + l}] : M = 1, \dots, L \}. \]
For simplicity, let us denote $g(M) = g_{\vect{r}}(q) U(\varphi(x(M)))$ for $M = 1,\dots, L$. For $M = 1, \dots, L$, let us write
\[
g(M) \vect{v}(M) = \vect{w}_+ \wedge \vect{w}^{(i-1)}(M) + (z(\mathfrak{k})\vect{w}_1) \wedge (\vect{w}')^{ (i-1)}(M) + \vect{w}^{(i)}(M)
\]
where $\vect{w}^{(i-1)}(M) \in \bigwedge^{i-1}W$, $(\vect{w}')^{ (i-1)}(M) \in \bigwedge^{i-1} z(\mathfrak{k})W_2$ and $\vect{w}^{(i)}(M) \in \bigwedge^{i} z(\mathfrak{k})W_2$. By our previous discussion, we have that
\[ \left\| \vect{w}_+ \wedge \vect{w}^{(i-1)}(M) \right\| \leq \rho^i, \]
\[ \|(\vect{w}')^{ (i-1)}(M) \| = \left\| (z(\mathfrak{k})\vect{w}_1) \wedge (\vect{w}')^{ (i-1)}(M) \right\| \leq \rho^i R^{-l},\]
and
\[ \left\| \vect{w}^{(i)}(M) \right\| \leq \rho^i R^{-l/2}.\]
Now let us consider $g(1) \vect{v}(M)$. Let us write $x(1) - x(M) = r R^{-q + (1 + \frac{1}{2n})l}$ where $r \in [ -1, 1]$. By our previous discussion, we have that
\[
\begin{array}{rcl}
g(1) = g_{\vect{r}}(q) U(\varphi(x(1))) & = & U(O(1)) U(r R^{(1+ \frac{1}{2n})l} \vect{h}) g_{\vect{r}}(q) U(\varphi(x(M))) \\
& = & U(O(1)) U(r R^{(1+ \frac{1}{2n})l} \vect{h}) g(M).
\end{array}
\]
Therefore, we have that
\[
g(1)\vect{v}(M) = U(O(1)) U(r R^{(1+ \frac{1}{2n})l} \vect{h}) g(M) \vect{v}(M).
\]
\par It is easy to see that we can ignore the contribution of $U(O(1))$ and identify $g(1)\vect{v}(M)$ with
$U(r R^{(1+ \frac{1}{2n})l} \vect{h}) g(M) \vect{v}(M)$. Then we have that
\[
\begin{array}{rcl}
g(1)\vect{v}(M) & = & U(r R^{(1+ \frac{1}{2n})l} \vect{h}) g(M) \vect{v}(M) \\
& = & \vect{w}_+ \wedge \vect{w}^{(i-1)}(M) + r R^{(1+\frac{1}{2n})l} \vect{w}_+ \wedge (\vect{w}')^{ (i-1)}(M) \\
& & + (z(\mathfrak{k})\vect{w}_1) \wedge (\vect{w}')^{ (i-1)}(M) + \vect{w}^{(i)}(M) .
\end{array}
\]
\par Now let us look at the range of
\[g_{\vect{r}}(-l/2) g(1) \vect{v}(M) = g_{\vect{r}}(q-l/2) U(\varphi(x(1))) \vect{v}(M).\]
It is easy to see that $g_{\vect{r}}(-l/2)\vect{w}_+ = b^{-l/2} \vect{w}_+$, $\|g_{\vect{r}}(-l/2)z(\mathfrak{k})\vect{w}_1\| \leq b^{r_1 l/2} \|z(\mathfrak{k})\vect{w}_1\|$,
\[\|g_{\vect{r}}(-l/2)\vect{w}^{(i-1)}(M)\| \leq b^{l/2} \|\vect{w}^{(i-1)}(M)\|,\]
\[ \|g_{\vect{r}}(-l/2)(\vect{w}')^{ (i-1)}(M) \| \leq b^{(1-r_1)l/2} \|(\vect{w}')^{ (i-1)}(M) \|, \]
and
\[ \|g_{\vect{r}}(-l/2) \vect{w}^{(i)}(M)\| \leq b^{l/2} \| \vect{w}^{(i)}(M)\|. \]
Since
\[
\begin{array}{rcl}
g_{\vect{r}}(-l/2) g(1) \vect{v}(M) & = & b^{-l/2} \vect{w}_+ \wedge (g_{\vect{r}}(-l/2)\vect{w}^{(i-1)}(M)) \\ & & +
r R^{(1+\frac{1}{2n})l} b^{-l/2} \vect{w}_+ \wedge(g_{\vect{r}}(-l/2) (\vect{w}')^{ (i-1)}(M) ) \\
& & + (g_{\vect{r}}(-l/2)z(\mathfrak{k})\vect{w}_1) \wedge (g_{\vect{r}}(-l/2)(\vect{w}')^{ (i-1)}(M)) \\
& & + g_{\vect{r}}(-l/2) \vect{w}^{(i)}(M),
\end{array}
\]
we have that
\[
\begin{array}{rcl}
\|g_{\vect{r}}(-l/2) g(1) \vect{v}(M)\| & \leq & b^{-l/2} \|\vect{w}_+ \wedge (g_{\vect{r}}(-l/2)\vect{w}^{(i-1)}(M))\| \\
& & + R^{(1+\frac{1}{2n})l} b^{-l/2} \|\vect{w}_+ \wedge(g_{\vect{r}}(-l/2) (\vect{w}')^{ (i-1)}(M) )\| \\
& & + \|g_{\vect{r}}(-l/2)z(\mathfrak{k})\vect{w}_1\| \cdot \|g_{\vect{r}}(-l/2)(\vect{w}')^{ (i-1)}(M)\| \\
& & + \|g_{\vect{r}}(-l/2) \vect{w}^{(i)}(M)\| \\
& \leq & b^{-l/2} b^{l/2} \|\vect{w}^{(i-1)}(M))\| + R^{(1+\frac{1}{2n})l} b^{-l/2}b^{(1-r_1)l/2} \|(\vect{w}')^{ (i-1)}(M) \| \\
& & + b^{r_1 l/2} \|z(\mathfrak{k})\vect{w}_1\| \cdot b^{(1-r_1)l/2} \|(\vect{w}')^{ (i-1)}(M) \| + b^{l/2}\| \vect{w}^{(i)}(M)\| \\
& \leq & b^{-l/2} b^{l/2} \rho^i + R^{(1+\frac{1}{2n})l} b^{-l/2}b^{(1-r_1)l/2} \rho^i R^{-l} \\
& & + b^{r_1 l/2} b^{(1-r_1)l/2} \rho^i R^{-l} + b^{l/2} \rho^i R^{-l/2} \\
& \leq & \rho^i + \rho^i + \rho^i R^{-l/2} + \rho^i \leq 1.
\end{array}
\]
For $M = 1, \dots, L$, let $\Lambda_i(\vect{v}(M))$ denote the $i$-dimensional primitive sublattice of $\mathbb{Z}^{n+1}$ corresponding to $\vect{v}(M)$. We will apply Proposition \ref{prop:counting-sublattices} to estimate $L$. Thus, let us keep the notation used there. By the inequality above, we have that $g_{\vect{r}}(-l/2)g(1)\Lambda_i(\vect{v}(M)) \in \mathcal{C}_i(g_{\vect{r}}(-l/2)g(1)\mathbb{Z}^{n+1}, 1)$ for every $M = 1, \dots, L$. On the other hand, since
$x(1) \in I_q \in \hat{\mathcal{I}}_q$, we have that
\[ g_{\vect{r}}(-l/2)g(1)\mathbb{Z}^{n+1} = g_{\vect{r}}(q - l/2) U(\varphi(x(1)))\mathbb{Z}^{n+1} \in K_{\kappa}. \]
By Proposition \ref{prop:counting-sublattices}, we have that
\[ L \leq |\mathcal{C}_i(g_{\vect{r}}(-l/2)g(1)\mathbb{Z}^{n+1}, 1)| \leq \kappa^{-N} = R^{Nm}.\]
Therefore, we have that
\[
\begin{array}{rcl}
m(D_{q,p}(I_p, i) \cap J) & \ll & L R^{-q+ l} \leq R^{-q + l + N m} \\
& \leq & R^{-q + l + \frac{l}{100 n}} \leq R^{- \frac{l}{20n}} R^{-q + (1 + \frac{1}{2n})l} = R^{- \frac{l}{20n}} m(J).
\end{array} \]
\par This completes the proof.
\end{proof}
\par Lemma \ref{lm:2nd-case-key-lemma} easily implies the following:
\begin{corollary}
\label{cor:2nd-case-key-lemma}
Let us keep the notation as above. Then
\[m(D_{q, p} (I_p, i)) \ll R^{-\frac{l}{20n}} m(I_p).\]
\end{corollary}
\begin{proof}
\par The statement follows from Lemma \ref{lm:2nd-case-key-lemma} by dividing $I_p$ into subintervals of length $R^{-q + (1 + \frac{1}{2n})l}$.
\end{proof}
\par Now we are ready to prove Proposition \ref{prop:small-portion-dangerous-case}.
\begin{proof}[Proof of Proposition \ref{prop:small-portion-dangerous-case}]
\par Let us fix $I_p \in \mathcal{I}_p$. For every $l/2 \leq l' \leq l$, let us denote by $D_{q,l'}(I_p)$ denote the union of $(q, l')$-dangerous intervals intersecting $I_p$. By Proposition \ref{prop:count-dangerous-intervals}, we have that $m(D_{q,l'}(I_p)) = O\left(R^{-\frac{l'}{10n}}\right)m(I_p)$. Therefore, we have that
\[
\begin{array}{rcl}
m\left( \bigcup_{l/2 \leq l' \leq l} D_{q,l'}(I_p) \right) & \leq & \sum_{l/2 \leq l' \leq l} m( D_{q,l'}(I_p)) \\
& \ll & \sum_{l/2 \leq l' \leq l} R^{-\frac{l'}{10n}} m(I_p) \\
& \ll & R^{-\frac{l}{20n}} m(I_p).
\end{array}
\]
By Corollary \ref{cor:2nd-case-key-lemma}, we have that
\[ \begin{array}{rcl}
m \left(\bigcup_{i = 2}^{n} D_{q, p} (I_p, i)\right) & \leq & \sum_{i = 2}^{n} m(D_{q, p} (I_p, i)) \\
& \ll & \sum_{i=2}^{n} R^{-\frac{l}{20n}} m(I_p) \ll R^{-\frac{l}{20n}} m(I_p)
\end{array}\]
\par By Lemma \ref{lm:key-lemma}, we have that
\[I_q \subset \bigcup_{l/2 \leq l' \leq l} D_{q,l'}(I_p)\bigcup_{i = 2}^{n} D_{q, p} (I_p, i) \]
for any $I_q \in \hat{\mathcal{I}}_{q,p}$.
Therefore, we have that
\[ \begin{array}{rcl}
F(\hat{\mathcal{I}}_{q,p}, I_p) R^{-q} & \leq & m\left( \bigcup_{l/2 \leq l' \leq l} D_{q,l'}(I_p)\bigcup_{i = 2}^{n} D_{q, p} (I_p, i) \right) \\
& \leq & m\left( \bigcup_{l/2 \leq l' \leq l} D_{q,l'}(I_p) \right) + m \left(\bigcup_{i = 2}^{n} D_{q, p} (I_p, i)\right) \\
& \ll & R^{-\frac{l}{20n}} m(I_p) = R^{-p - \frac{l}{20n}}.
\end{array} \]
This proves that
\[ F(\hat{\mathcal{I}}_{q,p}, I_p) \ll R^{q-p - \frac{l}{20n}}. \]
\end{proof}
\par By Proposition \ref{prop:small-portion-dangerous-case}, we have that
\begin{equation}
\label{equ:dangerous-case}
\begin{array}{cl}
& \sum_{l = 2000 n^2 N m}^{2\eta' q} \left(\frac{4}{R}\right)^{2l} \max_{I_{q- 2l} \in \mathcal{I}_{q- 2l}} F(\hat{\mathcal{I}}_{q, q - 2l}, I_{q - 2l}) \\
\ll & \sum_{l = 2000 n^2 N m}^{2\eta' q} \left(\frac{4}{R}\right)^{2l} R^{2l - \frac{l}{20n}} \\
\leq & \sum_{l = 2000 n^2 N m}^{2\eta' q} \left( \frac{16}{1000}\right)^l \ll \left( \frac{16}{1000}\right)^{2000 n^2 N m}.
\end{array}
\end{equation}
From this it is easy to see that
\begin{equation}
\label{equ:dangerous-case-final}
\sum_{l = 2000 n^2 N m}^{2\eta' q} \left(\frac{4}{R}\right)^{2l} \max_{I_{q- 2l} \in \mathcal{I}_{q- 2l}} F(\hat{\mathcal{I}}_{q, q - 2l}, I_{q - 2l}) \to 0
\end{equation}
as $m \to \infty$.
\subsection{Extremely dangerous case}
\label{subsec-extremely-dangerous-case}
\par In this subsection we will estimate $F(\hat{\mathcal{I}}_{q,0}, I)$. We call this case the extremely dangerous case.
\begin{proposition}
\label{prop:small-portion-extremely-dangerous}
There exists a constant $\nu >0$ such that for any $q > 10^6 n^4 N m$, we have that
\[ F(\hat{\mathcal{I}}_{q,0}, I) \ll R^{(1-\nu) q}. \]
\end{proposition}
\par Similar to Lemma \ref{lm:key-lemma}, we have the following:
\begin{lemma}
\label{lm:key-lemma-extremely-dangerous}
\par For any $i = 1, \dots, n$ and $I_q \in \hat{\mathcal{I}}_{q,0}(i)$, either there exists a $q$-extremely dangerous interval $\Delta_q(\vect{a})$ such that $I_q \in \Delta_q(\vect{a})$, or there exists $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$ such that the following holds:
for any $x \in I_q$, if we write
\[g_{\vect{r}}(q)U(\varphi(x))\vect{v} = \vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)} \]
where $\vect{w}^{(i-1)} \in \bigwedge^{i-1} W$ and $\vect{w}^{(i)} \in \bigwedge^i W$, then $\|\vect{w}_+ \wedge \vect{w}^{(i-1)}\| \leq \rho^i$ and $\|\vect{w}^{(i)}\| \leq \rho^i R^{-\eta' q}$.
\end{lemma}
\begin{proof}
\par The proof is the same as the proof of Lemma \ref{lm:key-lemma}.
\end{proof}
\begin{definition}
\label{def:extremely-dangerous-second-case}
\par For $ i = 2, \dots, n$, let $\mathcal{D}_q(i) $ denote the collection of $I_q \in \hat{\mathcal{I}}_{q,0}(i)$ such that the second case in Lemma \ref{lm:key-lemma-extremely-dangerous} holds and let
\[ D_q(i) := \bigcup_{I_q \in \mathcal{D}_q(i)} I_q .\]
Moreover, for $I_q \in \mathcal{D}_q(i)$, let $\vect{v} =\vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$ be the vector given in the second case of Lemma \ref{lm:key-lemma-extremely-dangerous}. Then for $x \in I_q$, we can write
\[g_{\vect{r}}(q)U(\varphi(x))\vect{v} = \vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)} \]
as in the second case of Lemma \ref{lm:key-lemma-extremely-dangerous}. For $l \geq \eta' q$, let $\mathcal{D}'_{q,l}(i)$ denote the collection of $I_q \in \mathcal{D}_q(i)$ such that
\[ \rho^i R^{-l+1} \leq \|\vect{w}^{(i)}\| \leq \rho^i R^{-l},\]
and let
\[D'_{q,l}(i) : = \bigcup_{I_q \in \mathcal{D}'_{q,l}(i)} I_q .\]
\end{definition}
\begin{lemma}
\label{lm:key-lemma-extremely-dangerous-2nd-case}
There exists a constant $\nu >0$ such that for any $q > 10^6 n^4 N m$ and any $i= 2, \dots, n$, we have that
\[ m(D_q(i)) \ll R^{- \nu q}. \]
\end{lemma}
\begin{proof}
\par For any $ \eta' q \leq l \leq 2\eta' q $, using the same argument as in the proof of Lemma \ref{lm:2nd-case-key-lemma}, we can prove that
\[ m(D'_{q,l}(i)) \ll R^{-\frac{l}{20n}}. \]
Therefore, we have that
\[
\begin{array}{rcl}
m\left( \bigcup_{l = \eta' q}^{2\eta' q} D'_{q,l}(i) \right) & \leq & \sum_{l = \eta' q}^{2\eta' q} m(D'_{q,l}(i)) \\
& \ll & \sum_{l = \eta' q}^{2\eta' q} R^{-\frac{l}{20n}} \ll R^{-\frac{\eta' q}{ 20 n}}.
\end{array}
\]
\par Let us denote
\[\mathcal{D}'_q(i) := \bigcup_{l > 2\eta' q} \mathcal{D}'_{q,l}\]
and
\[D'_q(i) : = \bigcup_{I_q \in \mathcal{D}'_q(i)} I_q.\]
Then it is enough to show that
\[ m(D'_q(i)) \ll R^{-\nu q}.\]
\par For any $I_q \in \mathcal{D}'_q(i)$ and $x \in I_q$, there exists $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_i \in \bigwedge^i \mathbb{Z}^{n+1} \setminus \{\vect{0}\}$ such that if we write
\[ g_{\vect{r}}(q) U(\varphi(x))\vect{v} = \vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)} \]
where $\vect{w}^{(i-1)} \in \bigwedge^{i-1} W$ and $\vect{w}^{(i)} \in \bigwedge^i W$, then we have that
$\|\vect{w}_+ \wedge \vect{w}^{(i-1)}\| \leq \rho^i$ and $\|\vect{w}^{(i)}\| \leq \rho^i R^{-2\eta' q}$.
\par Recall that $\eta = (1+r_1)\eta'$. Let us deal with the following two cases separately:
\begin{enumerate}
\item $r_n \geq \frac{\eta}{n}$.
\item There exists $ 1 < n_1 \leq n $ such that for $r_i \geq \frac{\eta}{n} $ for $1 \leq i < n_1$ and
$r_i < \frac{\eta}{n}$ for $ n_1 \leq i \leq n$.
\end{enumerate}
\par Let us first deal with the first case. For this case, let us define
\[ g^{\eta}(t) := \begin{bmatrix} b^{-\eta t} & \\ & b^{\eta t/n}\mathrm{I}_n \end{bmatrix} \in \mathrm{SL}(n+1, \mathbb{R}) \]
and $g_{\vect{r}, \eta}(t) := g^{\eta}(t) g_{\vect{r}}(t)$. It is easy to see that
\[g^{\eta}(t) \vect{w}_+ = b^{-\eta t} \vect{w}_+ = R^{-\eta' t} \vect{w}_+,\]
and
\[g^{\eta}(t) \vect{w} = b^{\eta t/n} \vect{w} = R^{\eta' t/n} \vect{w}\]
for any $\vect{w} \in W$.
\par Then we have that
\[
\begin{array}{rcl}
\| g_{\vect{r}, \eta}(q)U(\varphi(x))\vect{v}\| & = & \|g^{\eta}(q)(\vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)})\| \\
& \leq & \|g^{\eta}(q)(\vect{w}_+ \wedge \vect{w}^{(i-1)})\| + \|g^{\eta}(q) \vect{w}^{(i)}\| \\
& = & b^{-\eta q(1- \frac{i-1}{n})} \|\vect{w}_+ \wedge \vect{w}^{(i-1)}\| + b^{\frac{\eta q i}{n}} \|\vect{w}^{(i)}\| \\
& \leq & b^{-\frac{\eta q}{n}} \rho^i + b^{\eta q} R^{-2\eta' q} \rho^i \leq R^{-\frac{\eta' q}{n}} \rho^i.
\end{array}
\]
By the Minkowski Theorem, the above inequality implies that the lattice $g_{\vect{r}, \eta}(q)U(\varphi(x))\mathbb{Z}^{n+1}$ contains a nonzero vector with norm $\leq R^{-\frac{\eta' q}{n^2}} \rho$. Therefore, for any $I_q \in \mathcal{D}'_q(i)$ we have that
\[ g_{\vect{r}, \eta}(q)U(\varphi(I_q))\mathbb{Z}^{n+1} \not\in K_{\sigma } \]
where $\sigma = R^{-\frac{\eta' q}{n^2}} \rho$. By Theorem \ref{thm:shah-sln-representations}, for any $j = 1, \dots, n$ and any $\vect{v} = \vect{v}_1 \wedge \cdots \wedge \vect{v}_j \in \bigwedge^j \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$, we have that
\[ \max\{ g_{\vect{r},\eta}(q) U(\varphi(x)) \vect{v}: x \in I \} \geq \rho^i.\]
Then by Theorem \ref{thm:non-divergence-kleinbock-margulis}, we have that
\[ m\left( \{ x \in I: g_{\vect{r},\eta}(q) U(\varphi(x))\mathbb{Z}^{n+1} \not\in K_{\sigma} \} \right) \ll \sigma^{\alpha} = R^{-\frac{\alpha\eta' q}{n^2}}.\]
This proves that
\[ m(D'_q (i)) \ll R^{-\frac{\alpha\eta' q}{n^2}} . \]
\par This finishes the proof for the first case.
\par Now let us take care of the second case. Let us denote
\[ \xi(t) := \begin{bmatrix} b^{-\beta t} & & & & & \\ & 1 & & & & \\ & & \ddots & & & \\ & & & b^{r_{n_1} t} & & \\ & & & & \ddots & \\ & & & & & b^{r_n t} \end{bmatrix} \in \mathrm{SL}(n+1, \mathbb{R}) \]
where $\beta = \sum_{j= n_1}^n r_j < \eta$ and
\[g' (t) := \xi(t) g_{\vect{r}}(t) = \begin{bmatrix} b^{\chi t} & & & & & & \\ & b^{-r_1 t} & & & & & \\ & & \ddots & & & & \\ & & & b^{- r_{n_1 - 1} t} & & & \\ & & & & 1 & & \\ & & & & & \ddots & \\ & & & & & & 1 \end{bmatrix}\]
where $\chi = \sum_{j=1}^{n_1 - 1} r_j$. Then it is easy to see that
\[\xi(t) \vect{w}_+ = b^{-\beta t} \vect{w}_+ , \]
\[\xi(t) \vect{w}_j = \vect{w}_j\]
for $j = 1, \dots, n_1 - 1$, and
\[\xi(t) \vect{w}_j = b^{r_j t} \vect{w}_j\]
for $j = n_1, \dots, n$.
Then we have that
\[
\begin{array}{rcl}
\|g'(q) U(\varphi(x)) \vect{v}\| & = & \|\xi(q) (\vect{w}_+ \wedge \vect{w}^{(i-1)} + \vect{w}^{(i)}) \| \\
& \leq & \|\xi(q)(\vect{w}_+ \wedge \vect{w}^{(i-1)})\| + \|\xi(q) \vect{w}^{(i)}\| \\
& \leq & \| \vect{w}_+ \wedge \vect{w}^{(i-1)} \| + b^{\beta q} \|\vect{w}^{(i)}\| \\
& \leq & \rho^i + b^{\beta q} R^{-2\eta' q} \rho^i \\
& \leq & \rho^i + b^{\eta q} R^{-2\eta' q} \rho^i \leq \rho^i + R^{-\eta' q} \rho^i < (2\rho)^i.
\end{array}
\]
Moreover, for any $x' \in \Delta(x):= [ x - R^{-q(1-2\eta')} , x + R^{-q(1-2\eta')}]$, we also have that
\[ \|g'(q) U(\varphi(x')) \vect{v}\| < (2\rho)^i. \]
Let $C > 0$ and $\alpha >0$ be the constants given in Theorem \ref{thm:non-divergence-kleinbock-margulis}. We can choose $0 < \rho < 1$ small enough at beginning such that there exists another small constant $0 < \rho_1 <1 $ with $C \left(\frac{2 \rho}{\rho_1}\right)^{\alpha} < \frac{1}{1000}$. Then by the Minkowski Theorem, the inequality above implies that for any $x' \in \Delta(x)$, the lattice $g'(q) U(\varphi(x'))\mathbb{Z}^{n+1}$ contains a nonzero vector of length $< 2\rho$. Let $\vect{v}_{x'} \in \mathbb{Z}^{n+1}\setminus \{\vect{0}\}$ be the vector such that
$\|g'(q)U(\varphi(x'))\vect{v}_{x'}\| < 2 \rho$. Let us write
\[\vect{v}_{x'} = (v_{x'}(0), v_{x'}(1),\dots, v_{x'}(n)).\]
Then for $j = n_1 , \dots, n$, we have that $|v_{x'}(j)| < 2\rho$. Therefore, $v_{x'}(j) = 0$ for any $j = n_1 , \dots, n$. In other words, $\vect{v}_{x'}$ is contained in the subspace spanned $\{\vect{w}_+ , \vect{w}_1, \dots, \vect{w}_{n_1 - 1}\}$. For notational simplicity, let us denote this subspace by $\mathbb{R}^{n_1}$ and denote the set of integer points contained in the subspace by $\mathbb{Z}^{n_1}$. Accordingly, let us denote by $\mathrm{SL}(n_1, \mathbb{R})$ the subgroup
\[ \left\{ \begin{bmatrix} X & \\ & \mathrm{I}_{n+1 - n_1} \end{bmatrix}: X \in \mathrm{SL}(n_1, \mathbb{R}) \right\} \subset \mathrm{SL}(n+1, \mathbb{R}) \]
and denote by $\mathrm{SL}(n_1, \mathbb{Z})$ the subgroup of integer points in $\mathrm{SL}(n_1,\mathbb{R})$. Note that $g'(g) \in \mathrm{SL}(n_1, \mathbb{R})$ and $U(\varphi(x'))$ can be considered as an element in $\mathrm{SL}(n_1,\mathbb{R})$ since it preserves $\mathbb{R}^{n_1}$. Then $\|g'(q)U(\varphi(x'))\vect{v}_{x'}\| < 2\rho$ implies that for any $x' \in \Delta(x)$, the lattice $g'(q)U(\varphi(x'))\mathbb{Z}^{n_1}$ contains a nonzero vector of length $< 2\rho$. Let $K_{2\rho}(n_1) \subset X(n_1) = \mathrm{SL}(n_1, \mathbb{R})/\mathrm{SL}(n_1, \mathbb{Z})$ denote the set of unimodular lattices in $\mathbb{R}^{n_1}$ which do not contain any nonzero vector of length $< 2 \rho$. Then the claim above implies that
\[ m(\{x' \in \Delta(x): g'(q) U(\varphi(x')) \mathbb{Z}^{n_1} \not\in K_{2\rho}(n_1) \} ) = m(\Delta(x)). \]
By Theorem \ref{thm:non-divergence-kleinbock-margulis}, there exist $j \in {1, \dots, n_1 -1}$ and $\vect{v}' = \vect{v}'_1 \wedge \cdots \wedge \vect{v}'_j \in \bigwedge^j \mathbb{Z}^{n_1} \setminus\{\vect{0}\}$ such that
\[ \max\{ \|g'(q)U(\varphi(x'))\vect{v}'\|: x' \in [x- R^{-q(1-2\eta')}, x+ R^{-q(1-2\eta')}] \} < \rho_1^j\]
since otherwise we will have that
\[m(\{x' \in \Delta(x): g'(q) U(\varphi(x')) \mathbb{Z}^{n_1} \not\in K_{2\rho}(n_1) \} ) \leq C \left(\frac{2 \rho}{\rho_1}\right)^{\alpha} m(\Delta(x)) < \frac{1}{1000} m(\Delta(x)).\]
This reduces the second case in dimension $n+1$ to the first case in dimension $n_1 < n+1$. Then we can finish the proof by induction.
\end{proof}
\par Now we are ready to prove Proposition \ref{prop:small-portion-extremely-dangerous}.
\begin{proof}[Proof of Proposition \ref{prop:small-portion-extremely-dangerous}]
\par Recall that in Theorem \ref{thm:bernik-kleinbock-margulis}, we denote by $E_q$ the union of all $q$-extremely dangerous intervals. By Lemma \ref{lm:key-lemma-extremely-dangerous}, we have that
\[I_q \subset E_q \bigcup \bigcup_{i=2}^n D_q(i). \]
By Theorem \ref{thm:bernik-kleinbock-margulis}, we have that
\[ m(E_q) \ll R^{-\nu q}\]
for some constant $\nu >0$. On the other hand, by Lemma \ref{lm:key-lemma-extremely-dangerous-2nd-case}, we have that
\[m(D_q(i)) \ll R^{-\nu q}\]
for any $i = 2, \dots, n$. Therefore, we have that
\[
\begin{array}{rcl}
F(\hat{\mathcal{I}}_{q,0}, I) R^{-q} & = & m\left(\bigcup_{I_q \in \hat{\mathcal{I}}_{q,0}} I_q\right) \\
& \leq & m\left( E_q \bigcup_{i=2}^n D_q(i) \right) \leq m(E_q) + \sum_{i=2}^n m(D_q(i)) \ll R^{-\nu q}.
\end{array}
\]
This completes the proof.
\end{proof}{}
\par Now we are ready to prove Proposition \ref{prop:distance-sequence-small} for $q > 10^6 n^4 N m$.
\begin{proof}[Proof of Proposition \ref{prop:distance-sequence-small} for $q > 10^6 n^4 N m$]
\par We can choose $M$ such that $M^{\nu} > 1000$. Recall that $ R \geq M$. By Proposition \ref{prop:small-portion-extremely-dangerous}, we have that
\begin{equation}
\label{equ:extremely-dangerous-case}
\begin{array}{rcl}
\left( \frac{4}{R} \right)^q F(\hat{\mathcal{I}}_{q,0}, I) & \ll & \left( \frac{4}{R} \right)^q R^{(1-\nu) q} = \left( \frac{4}{R^{\nu}} \right)^q <\left( \frac{4}{1000} \right)^q.
\end{array}
\end{equation}
\par Combining \eqref{equ:q-large-p-large}, \eqref{equ:dangerous-case} and \eqref{equ:extremely-dangerous-case}, we have that
\[ \sum_{p = 0}^{q-1} \left( \frac{4}{R}\right)^{q-p} \max_{I_p \in \mathcal{I}_p} F(\hat{\mathcal{I}}_{q,p}, I_p) \to 0 \]
as $m \to \infty$. This proves the statement.
\end{proof}
\medskip
|
\section{Introduction}
Describing accurately atomic vibrations
is crucial in many branches of physics and chemistry because
thermodynamic, transport, and superconducting properties of materials and molecules
as well as the spectra obtained in many spectroscopic techniques depend on
how atoms vibrate\cite{born-huang}.
The standard harmonic approximation provides the simplest description
of vibrations, which are also present at 0 K due to the quantum zero-point motion.
The harmonic approximation is based on the
expansion of the Born-Oppenheimer (BO) energy surface to the second
order around the ionic equilibrium positions{. It predicts well-defined
non-interacting quasi-particles (phonons) with infinite lifetime
and a temperature-independent energy spectrum.}
Within this approximation many {physical} effects cannot be described.
For example, finite values of {the} thermal conductivity
and temperature dependent effects, like {the thermal expansion in solids,
cannot be accounted for at the harmonic level}. Therefore, it is
of paramount importance to describe accurately the vibrations of atoms
beyond the harmonic approximation.
Anharmonic effects, i.e. effects due to {higher} orders in the energy surface expansion,
introduce interaction between phonons, thus finite scattering rates and finite
{lifetimes}.
Anharmonicity can be treated {at} different levels
{of theory}. The basic approach is to consider higher order terms
in the potential expansion as a small perturbation of the harmonic potential\cite{PhysRev.128.2589}.
However, the perturbative approach can be used under quite restrictive conditions{:
the displacements of the atoms must be within the range in which the
harmonic approximation is valid so that higher order terms are considerably smaller
than the harmonic potential. Unfortunately, there are
several cases in which a non-perturbative regime is reached. For example
when light atoms such as hydrogen are present~\cite{PhysRevLett.111.177002,PhysRevB.89.064302,PhysRevLett.114.157004,Errea81,
0953-8984-28-49-494001,PhysRevB.82.104504},
or when the system is close to a dynamical instability (a phase transition)
as in ferroelectrics or materials undergoing a charge-density wave (CDW)
instability~\cite{PhysRevLett.17.753,delaire614,0022-3719-13-19-018,
PhysRevLett.107.107403,PhysRevB.92.140303,PhysRevLett.86.3799,PhysRevB.86.155125,
doi:10.1080/00150199808009159,PhysRevB.80.241108,PhysRevLett.106.196406,PhysRev.140.A863, PhysRevB.63.144109,PhysRevB.14.4321,PhysRevB.92.094107}.
In these cases, a non-perturbative approach is required in order to account for
anharmonic effects~\cite{Errea237}.}
Anharmonic effects at {a non-perturbative level are commonly treated within molecular
dynamics (MD) simulations or methods based on
them~\cite{PhysRevLett.103.125902,PhysRevLett.112.058501,PhysRevB.42.11276,PhysRevB.87.174110,
PhysRevLett.110.105503,PhysRevLett.112.058501,PhysRevB.84.180301,PhysRevB.87.104111,PhysRevB.88.144301}.
However, these approaches are computationally expensive as long simulation times
are needed to obtain converged renormalized phonon energies and have an intrinsic limitation
because they are based on Newtonian dynamics,
which limits their application to temperatures above the Debye temperature}.
This problem can be overcome by path-integral molecular dynamics~\cite{RevModPhys.67.279},
but the {even greater computational cost that the method needs to
incorporate the quantum character of atomic vibrations} makes it challenging. To surmount these difficulties,
{several methods~\cite{PhysRevLett.106.165501,PhysRevB.92.054301,Needs,PhysRevB.89.064302,
PhysRevLett.111.177002,:/content/aip/journal/jcp/138/4/10.1063/1.4788977,:/content/aip/journal/jcp/137/14/10.1063/1.4754819,
PhysRevB.92.201205}
have been developed,
mainly inspired by the self-consistent harmonic approximation (SCHA)
devised by Hooton~\cite{doi:10.1080/14786440408520575}}.
The main idea of the SCHA is to use a variational principle, the Gibbs-Bogoliubov (GB) principle,
in order to approximate the free energy of the true ionic Hamiltonian
with the free energy {calculated with a trial harmonic density matrix for the same system}.
In particular, in the stochastic self-consistent harmonic approximation (SSCHA)~\cite{PhysRevB.89.064302,PhysRevLett.111.177002},
the free energy is explicitly minimized by using a conjugate-gradient algorithm
with respect to the independent coefficients of
the trial harmonic potential{. In the SSCHA}
the free energy and its gradient are evaluated through averages computed
with stochastic sampling of the configuration space and the importance sampling technique~\cite{PhysRevB.89.064302,PhysRevLett.111.177002}.
In that way the (approximated but {non-perturbative})
anharmonic free energy of the system is directly accessible.
The stochastic approach is particularly suited to be used in
conjunction with \emph{ab initio} calculations,
and it has been employed to study thermal anharmonic effects in several
compounds {such as hydrides and transition metal
dichalcogenides}~\cite{PhysRevLett.111.177002,PhysRevLett.114.157004,PhysRevB.89.064302,Errea81,
PhysRevB.93.174308,PhysRevB.92.140303,0953-8984-28-49-494001}.
{Within the SCHA the free energy as a
function of the average atomic positions, i.e. the \emph{centroids} positions,
can be estimated
for any temperature. These can be used to study
structural second order phase transitions like, for example, in some ferroelectric
and CDW phase transitions~\cite{PhysRevLett.17.753,delaire614,0022-3719-13-19-018,
PhysRevLett.107.107403,PhysRevB.92.140303,PhysRevLett.86.3799,PhysRevB.86.155125,
doi:10.1080/00150199808009159,PhysRevB.80.241108,PhysRevLett.106.196406,
PhysRev.140.A863, PhysRevB.63.144109,PhysRevB.14.4321,PhysRevB.92.094107}}.
In general, at any temperature the system is in equilibrium in the configuration which
{minimizes} the free energy.
{According to Landau's theory~\cite{landau-vol-5}},
in a second order phase transition the high temperature free energy minimum is in a high-symmetry phase.
As the temperature decreases, the
minimum
becomes less and less pronounced until {it becomes a saddle point
at {the} transition temperature $T_c$,} and,
on lowering the temperature further,
the equilibrium position moves continuously towards lower symmetry configurations, where the free
energy is {smaller} (Cfr.~Fig.~\ref{fig:phase_transition}).
In this scheme the observable to be studied {as a function of temperature}
is the second derivative of the free energy with respect to the centroids positions,
i.e. the {free energy curvature, in the high-symmetry phase.}
Indeed the Hessian in the high-symmetry phase is positive-definite at high-temperature, but lowering the temperature it
develops {first a null ($T=T_c$) and then negative ($T<T_c$) eigendirection,}
which indicates the instability distortion that lowers the free energy.
Using the SSCHA {code~\cite{PhysRevLett.111.177002,PhysRevB.89.064302}} it is possible
to compute the free energy for several centroids positions and, therefore,
to calculate numerically, by finite difference, the curvature in a
point. This has been done, for example, to study the {quantum} H-bond
symmetrization in the record superconductor H${}_3$S~\cite{Errea81}.
However, even if legitimate, this `brute force' approach to compute the free energy curvature
{is computationally demanding.
In fact, a careful calculation of the curvature by finite differences requires
small statistical noise, implying a great deal of calls to the total-energy-force
engine used. Moreover, it also
requires SSCHA calculations in the low-symmetry distorted phase,
which are always more statistically demanding because the number of
free parameters in the trial harmonic Hamiltonian is larger due to the
reduced symmetry.}
Motivated by these considerations, in this paper we {derive the general exact
analytic expression of the SCHA free energy curvature for a generic atomic configuration.
Our approach is similar to the one proposed by G\"otze and Michel in the
context of elastic constants of anharmonic crystals~\cite{GotzeZfurphys}.
We also present an expression of the SCHA free energy curvature
that only depends on atomic displacements and forces. The latter is suited
for a stochastic implementation in conjunction with any total-energy-force
engine. The method we are presenting here allows, thus,}
to compute the curvature of the SCHA free energy {for
a given structure straightforwardly
once the GB functional has been minimized within the SSCHA.
Since the method is much more efficient and precise than
any finite-difference approach~\cite{Errea81}, it is
especially suited to be used in
conjunction with first-principles calculations.}
Besides the practical achievements,
the curvature formula {described here} has also interesting conceptual consequences.
{Since the only physical observable given by the SCHA is the (approximated) free energy,
the effective SCHA quadratic matrix that minimizes the GB functional
must be understood just as an auxiliary quantity, even if its eigenvalues have been often used
to calculate renormalized anharmonic phonon spectra~\cite{PhysRevLett.111.177002,PhysRevLett.114.157004,PhysRevB.89.064302,Errea81,
PhysRevB.93.174308,PhysRevB.92.140303,0953-8984-28-49-494001,PhysRevLett.106.165501}}.
A significant parallelism can be traced with the Hartree-Fock approximation.
In that case, the Rayleigh-Ritz functional of the total energy is
{minimized} with trial Slater determinants describing {non-interacting} fermions.
The energy obtained is an approximation of the true energy of the system{, but, on the contrary,
the corresponding trial non-interacting many-body wave function and related single particle spectrum
do not have in general a physical meaning
(except that in some aspects, e.g. see Koopmans' theorem~\cite{1934Phy.....1..104K}). An analogous
situation occurs with density-function theory (DFT)
and the corresponding non-interacting electrons and energy bands~\cite{PhysRevLett.49.1691}}. In the same way,
the SCHA matrix (divided by the square root of the masses, in order to have the correct dimensions) cannot be considered
a {generalized dynamical matrix and, therefore, its temperature-dependent eigenstates do not
represent phonons renormalized by anharmonicity.
On the contrary, the free energy curvature (in the equilibrium position) divided by the square root of the masses
defines a proper anharmonic temperature-dependent generalization of the harmonic dynamical matrix, whose eigenstates represent
temperature-dependent anharmonic phonons.
Indeed, at variance with the SCHA matrix, which is positive-definite by construction,
the dynamical matrix based on
the free energy curvature can have negative eigenvalues, and
a softening in its spectrum corresponds to a system instability.}
The free-energy based dynamical matrix is a particularly important tool especially when
we consider crystalline systems. Indeed in that case, exploiting the lattice periodicity and the
Fourier interpolation technique, it allows to find structural
instabilities with any modulations in real space by performing calculations
only on a coarse grid of the Brillouin zone.
The theory based on the free energy curvature with respect to the centroids position is `static'
in the sense that there are no dynamical variables evolving
with time. However, here we also propose a minimal `dynamic' extension of the theory that resembles the
work by Goldman \emph{et al.}~\cite{PhysRevLett.24.1424}, which allows to study, in a full
non-perturbative way, the spectral properties of anharmonic phonons, and thus allows to have finite
scattering times and linewidths. This makes the theory able to interpret
the results of inelastic scattering processes between anharmonic phonons and external incident particles (typically neutrons),
as well as allowing the calculation of the thermal conductivity in strongly
anharmonic solids where the harmonic approximation breaks down.
Despite the proposed dynamic extension being based on an \emph{ansatz},
it is reasonable because it yields good results in two limits:
at the lowest perturbative level it reduces to the standard perturbation theory result and
in the static limit it predicts the same instabilities found with the free energy curvature.
The paper is structured as follows. In section~\ref{Sec:Self-consistent_harmonic_approximation},
we present the fundamentals of the SCHA method, we define the SCHA free energy as a function of
the centroids position, and we fix the notation used.
In section~\ref{sec:Second_order_phase_transition_and_curvature_of_the_free_energy}
we show how to analyze structural second order phase transitions through
the Hessian of the free energy with respect to the centroids position (i.e. the curvature), and
in section~\ref{Derivatives_of_F} we give the explicit expression of the free energy curvature.
In section~\ref{sec:phonons_in_the_SCHA}, on the basis of the results obtained,
we describe the temperature-dependent free-energy-based
generalization of the harmonic dynamical matrix. In section~\ref{sec:Diagrammatic_representation},
we express the theory developed so far in a diagrammatic way.
In section~\ref{sec:Stochastic_implementation}, we show how to implement the curvature formula in a
stochastic way and, in particular, how to take into account symmetries in order to
speed up the statistical convergence.
In section~\ref{sec:Perturbative_limit}, we find the lowest order perturbative limit of the results obtained.
In section~\ref{sec:Anstaz_for_a_dynamic_theory}, inspired by the results obtained, we propose the \emph{ansatz}
to obtain a dynamical extension of the theory. Finally, in section~\ref{sec:Numerical_test},
we perform numerical tests on a toy model based on the ferroelectric transition in
SnTe, with the double objective of demonstrating the correctness of our findings
and showing the power of the method.
In section~\ref{sec:conclusions}, we summarize our results and draw some final conclusions.
The paper is completed with several appendices including the proofs of all the
equations given in the manuscript and the details of the toy model.
\allowdisplaybreaks
\section{Self-consistent harmonic approximation}
\label{Sec:Self-consistent_harmonic_approximation}
We consider the quantum atomic free energy of crystal lattices and molecules.
For notation clarity, we derive the main results by using a real space formalism
for both cases. This means that in the case of periodic crystals, we are actually studying a supercell
with periodic boundary conditions. Later, we will explicitly consider
a crystalline case for the numerical example
and we will take advantage of translational lattice symmetry.
The dynamic of atomic degrees of freedom is determined by the
quantum Hamiltonian
\begin{equation}
H=\sum_{s=1}^{N_{\text{a}}}\sum_{\alpha=1}^3\frac{{p^2_{s,\alpha}}}{2M_s}+V(\boldsymbol{R})\,,
\label{eq:full_ham}
\end{equation}
where $N_{\text{a}}$ is the total number of atoms,
$s$ and $\alpha$ are the atom and Cartesian component indices, respectively, $M_s$ is the mass of the $s$-th atom,
$p_{s,\alpha}$ and $R^{s,\alpha}$ are the momentum and position operators, respectively, and $V(\boldsymbol{R})$ is the
Born-Oppenheimer potential,
where the bold letter $\boldsymbol{R}$ indicates $R^{s,\alpha}$ in component-free notation.
In what follows, we will use bold letters
in component-free notation also for other observables and higher order tensors with respect to the
$(s,\alpha)$ index.
Moreover, in order to simplify the notation, we will
use a single composite index $a=(s,\alpha)$ to indicate both
atom and Cartesian indices. Notice that double index can be used also for the masses by
defining $M_{s,\alpha}=M_s$.
Fixed the temperature $T$, the free energy $F$ of the ionic Hamiltonian is given by
the sum of the total energy and the entropic contribution
\begin{equation}
F = \mathrm{tr} \bigl[ \scalebox{1.1}{$\rho$} H\bigr] + \frac{1}{\beta}
\mathrm{tr} \bigl[ \scalebox{1.1}{$\rho$} \ln \scalebox{1.1}{$\rho$}\bigr] \,,
\label{true-free-energy}
\end{equation}
where $\beta = (k_BT)^{-1}$ , `tr' is the trace operation on the $N_{\text{a}}$
atom Hilbert space, and
\begin{equation}
\scalebox{1.1}{$\rho$} = e^{- \beta H} \big/ \mathrm{tr} [ e^{- \beta H}]
\end{equation}
is the equilibrium density matrix.
In systems comprising many interacting particles, calculating $F$
can represent a complicated task. Nevertheless, a quantum variational principle for the free energy
can be established.
By replacing the true density matrix $\scalebox{1.1}{$\rho$}$ in Eq.~\eqref{true-free-energy} with a generic density matrix
$\tilde{\rho}$, we can define the functional
\begin{equation}
\mathcal{F}\bigl[\tilde{\rho}\bigr] = \mathrm{tr} \bigl[ \tilde{\rho} H\bigr] +
\frac{1}{\beta}
\mathrm{tr} \bigl[ \tilde{\rho}\ln \tilde{\rho}\bigr]\,,
\label{var-free-energy}
\end{equation}
and the Gibbs-Bogoliubov (GB) variational principle~\cite{0022-3689-1-5-305} states that
\begin{equation}
F\le \mathcal{F}\bigl[\tilde{\rho}\bigr]\,.
\label{gibbs-bogoliubov}
\end{equation}
Obviously the equality holds when $\tilde{\rho}=\scalebox{1.1}{$\rho$}$.
In particular, the SCHA is obtained by restricting {the} trial density {matrix}
to equilibrium density matrices
\begin{equation}
\tilde{\rho} _{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}= e^{- \beta \widetilde{H}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}} \big/ \mathrm{tr} [ e^{- \beta \widetilde{H}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}]
\label{eq:rhotrial}
\end{equation}
for the same temperature of a general trial harmonic Hamiltonian $\widetilde{H}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$ for
the same particles~\cite{doi:10.1080/14786440408520575}.
The trial harmonic Hamiltonian is parametrized in terms of
the vector $\boldsymbol{\Rcal}$ of dimension $3N_{\text{a}}$ and the square positive-definite matrix $\boldsymbol{\varPhi}$ of order $3N_{\text{a}}$ as
\begin{subequations}\begin{align}
&\widetilde{H}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}=\sum_a\frac{{p^2_{a}}}{2M_a}+\widetilde{V}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}\label{eq:Htrial}\\
&\widetilde{V}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}=\frac{1}{2}\sum_{ab}\varPhi_{ab} (R-\mathcal{R})^{a} (R-\mathcal{R})^{b}\label{eq:Vtrial}\,.
\end{align}\end{subequations}
In what follows we consider only trial harmonic potentials $\widetilde{V}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$
that respect the symmetries of the system.
With $\Bavg{\mathds{O}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}}$
we indicate the average of an observable $\mathds{O}$ with respect to the density matrix $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$:
\begin{equation}
\Bavg{\mathds{O}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}}=\Tr{\mathds{O}\,\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}\,.
\label{eq:def_avg}
\end{equation}
In what follows it will be relevant to consider observables $\mathds{O}(\boldsymbol{R})$ that are function of the position only.
In that case, the average can be written as
\begin{equation}
\Bavg{\mathds{O}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}}=\int \mathds{O}(\boldsymbol{R})\,\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})\,d\boldsymbol{R}\,,
\label{eq:avg_pos_operator}
\end{equation}
where $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$ is the diagonal part of the density matrix $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$
in coordinate representation~\cite{PhysRevLett.111.177002,PhysRevB.89.064302}:
\allowdisplaybreaks[0]
\begin{align}
&\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})=\sqrt{\text{det}\,(\boldsymbol{\Upsilon}/2\pi)}\nonumber\\
&\qquad\qquad\times\text{exp}\biggl[-\frac{1}{2}\sum_{ab}\Upsilon_{ab}(R-\mathcal{R})^a(R-\mathcal{R})^b\biggr]\,.
\label{eq:dens_prob}
\end{align}
\allowdisplaybreaks
Here `det' indicates the determinant and $\boldsymbol{\Upsilon}$ is the symmetric matrix associated to $\boldsymbol{\varPhi}$ by
\begin{equation}
\Upsilon_{ab}=\sqrt{M_aM_b}\,\sum_{\mu} \frac{2\omega_{\mu}}{(1+2n_\mu)\hbar}\,e^a_{\mu}e^b_{\mu}\,,
\label{eq:def_Upsilon}
\end{equation}
where $\omega^2_{\mu}$ and $e^a_{\mu}$ are eigenvalues and corresponding eigenvectors of
$\varPhi_{ab}/\sqrt{M_aM_b}$, and $n_{\mu}=1/(e^{\beta\hbar\omega_{\mu}}-1)$
is the bosonic average occupation number associated to $\omega_{\mu}$.
Note that to have a normalizable distribution $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$,
$\boldsymbol{\Upsilon}$ and thus $\boldsymbol{\varPhi}$ must be positive-definite matrices, as specified above.
From Eqs.~\eqref{eq:avg_pos_operator}--\eqref{eq:dens_prob} we see that the average positions of the atoms
for the trial density matrix $\tilde{\rho} _{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$, namely the `centroids',
coincide with $\boldsymbol{\Rcal}$:
\begin{equation}
\Bavg{\boldsymbol{R}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}}=\boldsymbol{\Rcal}\,.
\end{equation}
According to the GB variational principle, the best approximation of the free energy within the SCHA is $F^{\scriptscriptstyle{(\text{SCHA})}}$, given by
\begin{equation}
F^{\scriptscriptstyle{(\text{SCHA})}}=\min_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}\mathcal{F}\bigl[\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}\bigr]
=\min_{\boldsymbol{\Rcal}}\Bigl(\min_{\boldsymbol{\varPhi}} \mathcal{F}\bigl[{\tilde{\rho}}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}\bigr]\Bigr)\,.
\label{eq:minFcal}
\end{equation}
With $F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal})$ we indicate the SCHA free energy for the centroids position $\boldsymbol{\Rcal}$,
\begin{equation}
F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal})=\min_{\boldsymbol{\varPhi}} \mathcal{F}\bigl[\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}\bigr]\,,
\label{eq:F(R)_def_1}
\end{equation}
and with $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ we indicate the configuration that minimizes $F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal})$:
\begin{equation}
F^{\scriptscriptstyle{(\text{SCHA})}}=\min_{\boldsymbol{\Rcal}}F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal})=F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})\,.
\label{eq:defin_Req}
\end{equation}
Therefore, $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ is the SCHA equilibrium configuration of the centroids at the considered temperature.
Given a configuration $\boldsymbol{\Rcal}$ for the centroids, we define the corresponding SCHA square matrix
$\boldsymbol{\Phi}(\boldsymbol{\Rcal})$ as the matrix that minimizes the functional $\mathcal{F}[\displaystyle\rhotrial_{\bRcal,\bvarPhi}]$ with respect to $\boldsymbol{\varPhi}$.
Therefore, from Eq.~\eqref{eq:F(R)_def_1} we have
\begin{equation}
F^{\scriptscriptstyle{(\text{SCHA})}}(\boldsymbol{\Rcal})=\mathcal{F}\bigl[\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}\bigr]\,.
\label{eq:F(R)_def_2}
\end{equation}
The SCHA matrix satisfies the following self-consistent equation (see Eq.~\eqref{eq:app_SCHA_rel}):
\begin{equation}
\Phi_{ab}(\boldsymbol{\Rcal})=\avg{\frac{\partial^2V}{\partial R^a\partial R^b}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}}}\,.
\label{eq:SCHA_matrix_def}
\end{equation}
Notice that, for clarity, we are using two different symbols for the generic trial matrix $\boldsymbol{\varPhi}$ (a `slanted' phi),
and for the SCHA matrix $\boldsymbol{\Phi}(\boldsymbol{\Rcal})$, the specific trial matrix that minimizes $\mathcal{F}[\displaystyle\rhotrial_{\bRcal,\bvarPhi}]$ with respect to $\boldsymbol{\varPhi}$
for a given centroids position $\boldsymbol{\Rcal}$.
In the rest of this paper, we will consider exclusively the SCHA approximation for the free energy. Therefore, in order to simplify the notation,
in what follows we can safely omit the superscript ${}^{\scriptscriptstyle{(\text{SCHA})}}$ without ambiguity:
$F$ will always refer to the SCHA free energy. Moreover, as guide and reference for the reader, we collect in table~\ref{tab:symbols} some
symbols used in the text with a concise description. Several symbols collected in the table will appear later in the course of the paper.
\iftoggle{tabella_simboli}{
\begin{table*}
\begin{ruledtabular}
\begin{tabular}{lll}
Symbol & Meaning & First use \\
\hline
$\boldsymbol{R}$ & Atomic position (canonical variable) & Eq.~\eqref{eq:full_ham} \\
$V(\boldsymbol{R})$ & Potential energy & Eq.~\eqref{eq:full_ham} \\
$\boldsymbol{\Rcal}$ & Centroids position (parameter) & Eq.~\eqref{eq:rhotrial} \\
$F(\boldsymbol{\Rcal})$ & SCHA free energy for the centroids $\boldsymbol{\Rcal}$ & Eq.~\eqref{eq:F(R)_def_1}\\
$\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$ & Minimum point of $V(\boldsymbol{R})$ & Eq.~\eqref{eq:Harmonic_dyn_mat}\\
$\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ & Minimum point of $F(\boldsymbol{\Rcal})$ &Eq.~\eqref{eq:defin_Req}\\
$\boldsymbol{\varPhi}$& Generic trial harmonic matrix (parameter) & Eq.~\eqref{eq:rhotrial}\\
$\tilde{\rho} _{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$ & Probability distribution of $\boldsymbol{R}$ for a given value of the parameters $\boldsymbol{\Rcal}$, $\boldsymbol{\varPhi}$ & Eq.~\eqref{eq:rhotrial}\\
$\boldsymbol{\Phi}(\boldsymbol{\Rcal})$& $\boldsymbol{\varPhi}$ that minimizes the SCHA free energy functional at a given $\boldsymbol{\Rcal}$& Eq.~\eqref{eq:F(R)_def_2}\\
$\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}(\boldsymbol{\Rcal})$ &
Average of the $n$-th derivative of $V(\boldsymbol{R})$ with the probability $\tilde{\rho} _{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}(\boldsymbol{R})$ & Eq.~\eqref{eq:SCHA_matrix_def_nth} \\
$\boldsymbol{\phi}$ & Second derivative of $V(\boldsymbol{R})$ in $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$& Eq.~\eqref{eq:Ham_harm}\\
$\!\!\!\overset{\,\,\,\scriptscriptstyle{(n)}}{\boldsymbol{\phi}}$ & $N$-th derivative of $V(\boldsymbol{R})$ in $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$ & Eq.~\eqref{eq:n-th_order_force_constant}\\
$\boldsymbol{D}^{\scriptscriptstyle{(F)}}$& Second derivative of $F(\boldsymbol{\Rcal})$ in $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$, divided by the square root of the masses & Eq.~\eqref{eq:SCHA_dyn_mat}\\
$\boldsymbol{D}^{\scriptscriptstyle{(S)}}$& Matrix $\boldsymbol{\Phi}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})$ divided by the square root of the masses & Eq.~\eqref{eq:def_D}\\
$\overset{\scriptscriptstyle{(n)}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}$& Tensor $\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})$ divided by the square root of the masses &Eq.~\eqref{eq:def_D_n}\\
$\boldsymbol{D}^{\scriptscriptstyle{(0)}}$& Matrix $\boldsymbol{\phi}$ divided by the square root of the masses & Eq.~\eqref{eq:Harmonic_dyn_mat} \\
$\overset{\scriptscriptstyle{(n)}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(0)}}$& Tensor $\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\phi}}$ divided by the square root of the masses & Eq.~\eqref{eq:def_dyn_harm_nth}\\
$\overset{\scriptscriptstyle{-1}}{\boldsymbol{\mathcal{A}}}$ & Inverse of the matrix $\boldsymbol{\mathcal{A}}$ & Eq.~\eqref{eq:relaz_GS_DS}\\
$\boldsymbol{G}_{\scriptscriptstyle{(S)}}$& Green function associated to $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$ & Eq.~\eqref{eq:relaz_GS_DS}\\
$\boldsymbol{G}_{\scriptscriptstyle{(0)}}$& Green function associated to $\boldsymbol{D}^{\scriptscriptstyle{(0)}}$ & Eq.~\eqref{eq:relaz_G0_D0}\\
\end{tabular}
\end{ruledtabular}
\caption{Collection of some symbols frequently used in the main text.
First column, the symbol used. Second column, a short description of the meaning. Third column,
first labeled equation where the symbol appears.}
\label{tab:symbols}
\end{table*}
}
\section{Structural second order phase transition and curvature of the free energy}
\label{sec:Second_order_phase_transition_and_curvature_of_the_free_energy}
In second order phase transitions involving the position of the atoms,
e.g. in ferroelectric and in charge-density wave phase transitions~\cite{PhysRevLett.17.753,delaire614,0022-3719-13-19-018,
PhysRevLett.107.107403,PhysRevB.92.140303,PhysRevLett.86.3799,PhysRevB.86.155125,
doi:10.1080/00150199808009159,PhysRevB.80.241108,PhysRevLett.106.196406,PhysRev.140.A863, PhysRevB.63.144109,PhysRevB.14.4321,
PhysRevB.92.094107}, we can
use the centroids $\boldsymbol{\Rcal}$ to define the order parameter, which is the
observable measured in diffraction experiments.
The (temperature-dependent) function $F(\boldsymbol{\Rcal})$ rules the
phase transitions. At each temperature, the system is in equilibrium in the
(temperature-dependent) configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$, where $F(\boldsymbol{\Rcal})$ has a minimum.
Therefore, in $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ the first derivative of $F(\boldsymbol{\Rcal})$ is zero,
$\partial F/\partial\mathcal{R}^a|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}}=0$, and the Hessian matrix of $F(\boldsymbol{\Rcal})$ (i.e. the curvature), $\partial^2 F/\partial\mathcal{R}^a\partial\mathcal{R}^b|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}}$,
is positive-definite.
Landau's theory of second order phase transitions~\cite{landau-vol-5} shows that above a certain critical temperature $T_c$
the equilibrium configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ is in a high-symmetry phase $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}$.
As $T$ decreases and approaches $T_c$ from above,
the minimum of $F(\boldsymbol{\Rcal})$ in $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}$ becomes less and less pronounced. At $T=T_c$,
$\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}$ becomes a saddle point, i.e. the Hessian of $F(\boldsymbol{R})$ in $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}$ develops at least one null eigenvalue,
which becomes negative by lowering further the temperature. At the same time,
the minimum point $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}(T)$, now depending on temperature, continuously deviates from
$\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}$ to different configurations having lower symmetry.
Since during the phase transition the equilibrium configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}(T)$ remains a continuous function of temperature,
these are also called ``continuous phase transitions''.
In Fig.~\ref{fig:phase_transition} we show an example of a typical second order phase transition.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{sec_ord_phase_trans.pdf}
\caption{(Color online) Example of a second order (i.e. continuous) phase transition, as described by Landau's theory.
$Q$ is a macroscopic, scalar, order parameter identifying a system configuration. We consider a situation in which symmetry $Q\rightarrow -Q$ holds.
$Q=0$ is a high-symmetry phase. $\Delta F(Q)= F(Q)-F(0)$ is the difference between the free energy of phase $Q$ and the
free energy of the high-symmetry phase, at a certain temperature $T$. For each $T$, $\Delta F(Q)$ has minimum in the equilibrium
configuration $Q_{\scriptscriptstyle{\text{eq}}}(T)$. Plots are in arbitrary units.
The free energy difference is an even polynomial $\Delta F(Q)=A_2(T)\,Q^2+A_4(T)\,Q^4+\mathcal{O}(Q^6)$, with
$A_4(T)>0$ and $A_2(T)$ that decreases from positive to negative values. $T_c$ is the transition temperature.
At $T>T_c$, the free energy has minimum in $Q_{\scriptscriptstyle{\text{eq}}}(T)=0$, i.e. $A_2(T)$ is positive.
At $T<T_c$, $A_2(T)$ is negative: $Q=0$ becomes a local maximum, whereas the minimum $Q_{\scriptscriptstyle{\text{eq}}}(T)$ acquires two opposite degenerate values,
different from zero. $Q_{\scriptscriptstyle{\text{eq}}}(T)$ is a continuous function even during the transition.
Upper panel: variation of the free energy $\Delta F$ as a function of the
order parameter $Q$ for three temperatures $T$, above, below and equal to the transition temperature $T_c$.
Bottom panel: value of the equilibrium order parameter $Q_{\scriptscriptstyle{\text{eq}}}$ as a function of the temperature $T$.}
\label{fig:phase_transition}
\end{figure}
In conclusion, at any temperature it is $\partial F/\partial\mathcal{R}^a|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}}=0$ and the phase transition is
characterized by the change of character of the Hessian matrix $\left.\partial^2F/\partial\mathcal{R}^a \partial\mathcal{R}^b\right|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{hs}}}}$:
at $T>T_c$ it is positive-definite, whereas as $T<T_c$
it develops at least one negative eigendirection indicating the distortion which decreases the free energy.
It follows that a method to estimate the transition temperature $T_c$ and the instability modes
of a second order phase transition can be obtained by computing the Hessian of $F(\boldsymbol{\Rcal})$
in the high-symmetry configuration
and studying its evolution as a function of temperature.
In the next section we will find explicit formulas for the first and second derivative of $F(\boldsymbol{\Rcal})$.
\section{Derivatives of $F(\boldsymbol{\Rcal})$}
\label{Derivatives_of_F}
From the definition of Eq.~\eqref{eq:F(R)_def_2}, we can calculate explicitly the derivatives
of $F(\boldsymbol{\Rcal})$ with respect to $\boldsymbol{\Rcal}$.
Here we present only the results, while the derivation is given in appendix~\ref{app:SCHA_proofs}.
For the first derivative we have the intuitive result (see~Eq.~\eqref{eq:app_first_der})
\begin{equation}
\frac{\partial F}{\partial\boldsymbol{\Rcal}}=\avg{\frac{\partial V}{\partial \boldsymbol{R}}}_{\ds{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}}}\,.
\label{eq:first_derivative_F}
\end{equation}
The derivative of the free energy is the average of the potential derivative. In other words,
the forces on the centroids are equal to the average of the mechanical forces on the atoms.
From Eq.~\eqref{eq:first_derivative_F}, for the equilibrium position $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$
defined in Eq.~\eqref{eq:defin_Req} it is
\begin{equation}
0=\avg{\frac{\partial V}{\partial \boldsymbol{R}}}_{\ds{\tilde{\rho}_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}},\boldsymbol{\Phi}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})}}}\,.
\label{eq:first_derivative_F_in_eq}
\end{equation}
For what follows it is convenient to define the $n$-th order SCHA tensor, which generalizes Eq.~\eqref{eq:SCHA_matrix_def}
to higher orders:
\begin{equation}
\overset{\scriptscriptstyle{(n)}}{\Phi}_{a_1\cdots a_n}(\boldsymbol{\Rcal})=\avg{\frac{\partial^nV}{\partial R^{a_1}\cdots\partial R^{a_n}}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}}}\,.
\label{eq:SCHA_matrix_def_nth}
\end{equation}
Notice that we did not use the superscript $\scriptstyle{(2)}$ for the square SCHA matrix.
The $n$-th order SCHA tensor has the same properties of the $n$-th order force constant (i.e. the $n$-th derivative of the potential). Notably,
it is invariant with respect to permutation of indices; it is invariant with respect to
all the symmetry operations (including lattice translations
in a crystal) associated to the configuration $\boldsymbol{\Rcal}$~\cite{RevModPhys.40.1};
and it satisfies the acoustic sum rule (ASR), i.e.
the sum over any atom index vanishes (see~Eq.~\eqref{eq:app_def_ASR}).
Deriving a second time Eq.~\eqref{eq:first_derivative_F} with respect to $\boldsymbol{\Rcal}$ we obtain
(see~Eqs.~\eqref{eq:app_der_int_0}--\eqref{eq:app_Theta_equation} )
\begin{align}
&\frac{\partial^2 F}{\partial \mathcal{R}^a \partial \mathcal{R}^b}=\Phi_{ab}+\sum_{c_1c_2c_3c_4}\overset{\scriptscriptstyle{(3)}}{\Phi}_{ac_1c_2}\Lambda^{c_1c_2c_3c_4}\overset{\scriptscriptstyle{(3)}}{\Phi}_{c_3c_4b}\nonumber\\
&\,\,\,+\sum_{\substack{c_1c_2c_3c_4\\d_1d_2d_3d_4}}\overset{\scriptscriptstyle{(3)}}{\Phi}_{ac_1c_2}\Lambda^{c_1c_2c_3c_4}\Theta_{c_3c_4d_1d_2}\Lambda^{d_1d_2d_3d_4}\overset{\scriptscriptstyle{(3)}}{\Phi}_{d_3d_4b}\,,
\label{eq:sec_der_F_1}
\end{align}
where
\begin{align}
\Lambda^{abcd}&=-\frac{\hbar^2}{8}\sum_{\nu\mu}
\frac{F(0,\omega_{\mu},\omega_{\nu})}{\omega_{\mu}\omega_{\nu}}\nonumber\\
&\mkern72mu\times\frac{e^a_{\nu}}{\sqrt{M_a}}\frac{e^b_{\mu}}{\sqrt{M_b}}\frac{e^c_{\nu}}{\sqrt{M_c}}\frac{e^d_{\mu}}{\sqrt{M_d}}\,.
\label{eq:main_lambda_1}
\end{align}
Here $e_\mu^a$ and $\omega_\mu^2$ are eigenvectors and eigenvalues of $\Phi_{ab}/\sqrt{M_a M_b}$, respectively, and
\allowdisplaybreaks[0]
\begin{align}
&F(0,\omega_\nu,\omega_\mu)=\phantom{\Biggl\{}\nonumber\\
&\quad\left\{
\begin{aligned}
&\frac{2}{\hbar}\left[\frac{2n_\nu+1}{2\omega_\nu}-\frac{dn_\nu}{d\omega_\nu}\right] &&\text{if}\qquad \omega_\nu=\omega_\mu\\
&\frac{2}{\hbar}\left[\frac{n_\mu+n_\nu+1}{\omega_\mu+\omega_\nu}-\frac{n_\mu-n_\nu}{\omega_\mu-\omega_\nu}\right] &&\text{if}\qquad \omega_\nu\neq\omega_\mu
\end{aligned}
\right.\,.
\label{eq:def_F0}
\end{align}
The tensor $\Theta_{abcd}$ is the solution of the
Dyson-like equation
\allowdisplaybreaks
\begin{equation}
\Theta_{abcd}=\overset{\scriptscriptstyle{(4)}}{\Phi}_{abcd}+\sum_{l_1l_2l_3l_4}\overset{\scriptscriptstyle{(4)}}{\Phi}_{abl_1l_2} \Lambda^{l_1l_2l_3l_4} \Theta_{l_3l_4cd}\,.
\label{eq:Theta_1}
\end{equation}
Notice that in all these equations the dependence of the quantities on $\boldsymbol{\Rcal}$ is understood.
We have obtained for the second derivative a relation which is different from the one found for the first derivative.
Indeed, as shown in Eq.~\eqref{eq:first_derivative_F}, the first derivative of the SCHA free energy is equal to the average
of the first derivative of the potential.
On the contrary, the second derivative of the SCHA free energy is equal to the average of the second derivative
of the potential, the SCHA matrix $\Phi_{ab}$ of Eq.~\eqref{eq:SCHA_matrix_def}, plus two terms depending on the third and fourth order SCHA tensors.
In component-free notation, we can write Eq.~\eqref{eq:sec_der_F_1} in compact form:
\begin{equation}
\frac{\partial^2F}{\partial \boldsymbol{\Rcal} \partial \boldsymbol{\Rcal}}=\boldsymbol{\Phi}+\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}
+\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\Theta}\boldsymbol{\Lambda}\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\,,
\label{eq:compact_expr_curv}
\end{equation}
where the contraction on the indices is understood.
Moreover, it is convenient to introduce a `super-index' $A=(pq)$.
In this way, for example, $\Lambda^{pqhk}=\Lambda^{AB}$, $\Theta^{pqhk}=\Theta^{AB}$
and $\overset{\scriptscriptstyle{(4)}}{\Phi}{}_{pqhk}=\overset{\scriptscriptstyle{(4)}}{\Phi}{}_{AB}$ are square symmetric `super-matrices' of order $(3N_{\text{a}})^2$,
and the contraction of indices between them can be seen as a matrix product.
As explained in the previous sections, the curvature of the free energy in a high-symmetry phase as a function of temperature
is essential in order to identify and characterize a second order phase transition. Diagonalizing the real symmetric matrix
$\partial^2F/\partial\mathcal{R}^a\partial\mathcal{R}^b$ we obtain eigenvalues and eigenvectors as a function of temperature. In the presence of a second order phase transition,
there is at least one eigenvalue that becomes negative at the transition temperature, and the corresponding eigenvector identifies
the instability distortion pattern which reduces the free energy.
By definition, the SCHA matrix $\boldsymbol{\Phi}$ is positive-definite (see comment after Eq.~\eqref{eq:def_Upsilon}). On the contrary,
as shown in Eq.~\eqref{eq:app_def_neg_lambda}, $\boldsymbol{\Lambda}$ is negative-definite thus
$\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ is negative-semidefinite. It is this term, which for reasons that will be clear later we call `bubble'
(see~Sec.~\ref{sec:Diagrammatic_representation}), that allows the second derivative of the free energy to have negative
eigenvalues. The formula obtained for $\partial^2F/\partial\mathcal{R}^a\partial\mathcal{R}^b$ also clarifies in this way
the long-standing debate about the possibility of having
second order phase transitions within the SCHA~\cite{PhysRevLett.28.895,PhysRevLett.29.369}:
the SCHA can describe a second order phase transition only if
$\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}} \ne 0$.
Using the interpretation of the 4th-rank tensors as super-matrices of order $(3N_{\text{a}})^2$,
$\boldsymbol{\Theta}$ is readily obtained by inverting Eq.~\eqref{eq:Theta_1} in matrix form:
\begin{equation}
\boldsymbol{\Theta}=\Bigl[\mathds{1}-\boldsymbol{\overset{\scriptscriptstyle{(4)}}{\Phi}}\boldsymbol{\Lambda}\Bigr]^{-1}\,\,\boldsymbol{\overset{\scriptscriptstyle{(4)}}{\Phi}}\,.
\label{eq:btheta_expr}
\end{equation}
Substituting Eq.~\eqref{eq:btheta_expr} into Eq.~\eqref{eq:compact_expr_curv} we obtain the compact expression for the
free energy Hessian:
\begin{equation}
\frac{\partial^2F}{\partial \boldsymbol{\Rcal} \partial \boldsymbol{\Rcal}}=\boldsymbol{\Phi}+
\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\,\,\boldsymbol{\Lambda}\Bigl[\mathds{1}-\boldsymbol{\overset{\scriptscriptstyle{(4)}}{\Phi}}\boldsymbol{\Lambda}\Bigr]^{-1}\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\,.
\label{eq:compact_expr_curv_daimpl}
\end{equation}
This is the equation that has been implemented .
It is also interesting to write Eq.~\eqref{eq:compact_expr_curv_daimpl} in a more symmetric fashion:
\begin{equation}
\frac{\partial^2F}{\partial \boldsymbol{\Rcal} \partial \boldsymbol{\Rcal}}=\boldsymbol{\Phi}-\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\sqrt{-\boldsymbol{\Lambda}}
\,\,
\frac{\mathds{1}}{\mathds{1}-\boldsymbol{\Xi}}
\,\,
\sqrt{-\boldsymbol{\Lambda}}\,\boldsymbol{\overset{\scriptscriptstyle{(3)}}{\Phi}}\,,
\end{equation}
where $\boldsymbol{\Xi}=-\sqrt{-\boldsymbol{\Lambda}}\,\,\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}\sqrt{-\boldsymbol{\Lambda}}$ is the adimensional real symmetric matrix that rules the convergence of the geometric series.
For example, from this formula we clearly see that
in the limiting case where the absolute values of $\boldsymbol{\Xi}$'s eigenvalues are much smaller than one and $\boldsymbol{\Xi}$ can be discarded with respect to the identity,
the curvature is given by the SCHA matrix plus the bubble only.
\section{Phonons in the SCHA}
\label{sec:phonons_in_the_SCHA}
From the results obtained, it is tempting to use the curvature of the free energy with respect to the centroids to define
a phonon-like dispersion. To this purpose, for each temperature, we consider the free energy curvature in the corresponding
equilibrium configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$, divided by the square root of the masses:
\begin{equation}
D^{\scriptscriptstyle{(F)}}_{ab}=\left.\frac{1}{\sqrt{M_aM_b}}\frac{\partial^2F}{\partial \mathcal{R}^a \partial \mathcal{R}^b}\right|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}}\,.
\label{eq:SCHA_dyn_mat}
\end{equation}
This matrix can be considered as the temperature-dependent, free energy-based, generalization of the temperature-independent harmonic dynamical matrix
\begin{equation}
D^{\scriptscriptstyle{(0)}}_{ab}=\left.\frac{1}{\sqrt{M_aM_b}}\frac{\partial^2V}{\partial R^a \partial R^b}\right|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}}\,.
\label{eq:Harmonic_dyn_mat}
\end{equation}
Here $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$ is the temperature-independent configuration for which the potential $V(\boldsymbol{R})$ has a minimum.
We associate the `free energy dynamical matrix' $D^{\scriptscriptstyle{(F)}}_{ab}$ to `free energy phonons', quasi-particles
whose energies $\hbar\Omega_{\mu}$ and polarization vectors $\epsilon^a_{\mu}$ are obtained by diagonalization as
\begin{equation}
\sum_{b}D^{\scriptscriptstyle{(F)}}_{ab}\epsilon^b_{\mu}=\Omega^2_{\mu}\epsilon^a_{\mu}\,.
\end{equation}
Since $D^{\scriptscriptstyle{(F)}}_{ab}$ is positive-definite if and only if $\partial^2F/\partial\mathcal{R}^a\partial\mathcal{R}^b|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}}$ is positive-definite,
an instability in the system corresponds to at least a frequency $\Omega_{\mu}$ becoming imaginary. It is this fact that
justifies the interpretation of $D^{\scriptscriptstyle{(F)}}_{ab}$ as a temperature-dependent generalized dynamical matrix describing
temperature-dependent anharmonic phonons.
It is worthwhile to emphasize that the theory developed so far is `static', in the sense that it is not based on time-dependent properties, but on the variation
of the free energy with respect to a static variation of the centroids position. Moreover, it is important to observe
that we cannot use
\begin{equation}
D^{\scriptscriptstyle{(S)}}_{ab}=\frac{1}{\sqrt{M_aM_b}}\Phi_{ab}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})
\label{eq:def_D}
\end{equation}
to study system instabilities and defines phonon-like particles, even if
in some cases it has given temperature-dependent anharmonic phonons in
good agreement with experiments~\cite{PhysRevB.92.140303,PhysRevLett.111.177002}.
Indeed, $D^{\scriptscriptstyle{(S)}}_{ab}$ it is not given by the second derivative of the free energy.
Moreover, by definition, $D^{\scriptscriptstyle{(S)}}_{ab}$ is positive-definite, thus it is impossible to observe any softening in its eigenvalues.
The free energy dynamical matrix $D^{\scriptscriptstyle{(F)}}_{ab}$ is a particularly important tool when we consider crystals.
Indeed, in that case we can use the same techniques that are standard for the harmonic theory~\cite{RevModPhys.73.515}.
Exploiting the translational lattice symmetry, we define the SCHA dynamical matrices $\boldsymbol{D}^{\scriptscriptstyle{(F)}}(\boldsymbol{q})$ in the unit cell as
a function of the quasi-momentum $\boldsymbol{q}$. We can explicitly calculate $\boldsymbol{D}^{\scriptscriptstyle{(F)}}(\boldsymbol{q})$ on a coarse grid
of the Brillouin zone (BZ) and later Fourier interpolate the result to obtain the matrix
on an arbitrary finer grid or a path. Thus, diagonalizing $\boldsymbol{D}^{\scriptscriptstyle{(F)}}(\boldsymbol{q})$, we obtain the spectrum
$\Omega^2_{\mu}(\boldsymbol{q})$ and the polarization vectors $\epsilon^a_{\mu}(\boldsymbol{q})$ on a path of the BZ. An imaginary phonon
in a point $\boldsymbol{q}$ indicates that the system is unstable for a distortion with
modulation $\boldsymbol{q}$ that reduces the lattice periodicity. This is, for example, what happens in charge-density wave instabilities.
Therefore, with moderate workload, it is possible to have a complete picture of the crystal instabilities.
In particular, with calculations on supercells of reasonable size it is possible, in principle, to study lattice instabilities
which are periodic on very large supercells or even incommensurate.
\section{Diagrammatic representation}
\label{sec:Diagrammatic_representation}
In this section we give a perspicuous diagrammatic description of Eq.~\eqref{eq:compact_expr_curv_daimpl}, in order to
reformulate it in a language familiar to the field theorists.
The diagrammatic description can also be useful as a basis for further developments of the theory, as we will see
later in Sec.~\ref{sec:Anstaz_for_a_dynamic_theory}.
Fixed the temperature, with the corresponding $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ we define the quadratic `SCHA Hamiltonian'
\begin{equation}
H^{\scriptscriptstyle{(S)}}=\sum_{a}\frac{{p^2_{a}}}{2M_a}+\frac{1}{2}\sum_{ab}\Phi_{ab}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})\,(R-\mathcal{R}_{\scriptscriptstyle{\text{eq}}})^a(R-\mathcal{R}_{\scriptscriptstyle{\text{eq}}})^b\,,
\label{eq:HssS_def}
\end{equation}
and we consider the corresponding SCHA thermodynamic Green function $G^{ab}_{\scriptscriptstyle{(S)}}(z)$
for the displacements normalized by masses $\sqrt{M_a}(R-\mathcal{R}_{\scriptscriptstyle{\text{eq}}})^a$.
Since $H^{\scriptscriptstyle{(S)}}$ is quadratic
\begin{equation}
\overset{\scriptscriptstyle{-1}}{G}{}^{ab}_{\scriptscriptstyle{(S)}}(z)=z^2\delta_{ab}-D_{ab}^{\scriptscriptstyle{(S)}}\,,
\label{eq:relaz_GS_DS}
\end{equation}
where $\overset{\scriptscriptstyle{-1}}{G}{}^{ab}_{\scriptscriptstyle{(S)}}$ indicates the inverse matrix of $G^{\scriptscriptstyle{(S)}}_{ab}$ (similar notation for the
inverse will be used later also in other formulas). We also consider $\chi^{abcd}_{\scriptscriptstyle{(S)}}(0)$, the SCHA `static' loop, i.e. the loop
with $G^{ab}_{\scriptscriptstyle{(S)}}$ and total frequency equal to zero:
\begin{equation}
\chi_{\scriptscriptstyle{(S)}}^{abcd}(0)=\frac{1}{\beta}\sum_l\,{G}_{\scriptscriptstyle{(S)}}^{ac}(i\Omega_l){G}_{\scriptscriptstyle{(S)}}^{bd}(i\Omega_{-l})\,,
\label{eq:static_loop}
\end{equation}
where $\Omega_l=2\pi l/\hbar\beta$ is the $l$-th Matsubara frequency.
With standard techniques for Matsubara frequency summation we obtain~\cite{mahan2000many,PhysRev.128.2589}
\begin{align}
&\frac{1}{\beta}\sum_lG_{\scriptscriptstyle{(S)}}^{ac}(i\Omega_{l})G_{\scriptscriptstyle{(S)}}^{bd}(i\Omega_{-l})=\nonumber\\
&\quad\qquad\frac{\hbar^2}{4}\sum_{\mu\nu}\frac{F(0,\omega_{\mu},\omega_{\nu})}{\omega_{\mu}\omega_{\nu}}\,e^a_{\nu}e^b_{\mu}e^c_{\nu}e^d_{\mu}\,,
\label{eq:sumMatzub_1}
\end{align}
with $F(0,\omega_{\mu},\omega_{\nu})$ defined in Eq.~\eqref{eq:def_F0}.
From Eq.~\eqref{eq:main_lambda_1}, Eq.~\eqref{eq:static_loop} and Eq.~\eqref{eq:sumMatzub_1} we obtain a relation between the tensor
$\Lambda^{abcd}$ and the static loop $\chi_{\scriptscriptstyle{(S)}}^{abcd}(0)$:
\begin{equation}
\chi_{\scriptscriptstyle{(S)}}^{abcd}(0)=-2\Lambda^{abcd}\sqrt{M_aM_bM_cM_d}\,.
\label{eq:relazione_lambda_chizero}
\end{equation}
Therefore, using Eq.~\eqref{eq:SCHA_dyn_mat} and Eq.~\eqref{eq:def_D},
formula~\eqref{eq:compact_expr_curv_daimpl} divided by the square root of the masses gives
\begin{equation}
{\boldsymbol{D}}{}^{\scriptscriptstyle{(F)}}={\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}+\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)\,,
\label{eq:diagrammatic_static_invSCHAdynmat_00}
\end{equation}
where, as usual, we have used bold symbols in component-free notation and we have defined
\begin{align}
\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)&=\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(0)\right)\nonumber\\
&\qquad\times\left[\mathds{1}-\overset{\boldsymbol{\ssf}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(0)\right)\right]^{-1}\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\,.
\label{eq:diagrammatic_static_selfenergy}
\end{align}
Here we have generalized the definition~\eqref{eq:def_D} to the $n$-th order as
\begin{equation}
\overset{\scriptscriptstyle{(n)}}{D}{}^{\scriptscriptstyle{(S)}}_{a_1\ldots a_n}=\frac{\overset{\scriptscriptstyle{(n)}}{\Phi}_{a_1\ldots a_n}(\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}})}{\sqrt{M_{a_1}\ldots M_{a_n}}}\,.
\label{eq:def_D_n}
\end{equation}
Notice that we did not use the superscript $\scriptstyle{(2)}$ for the second order tensor defined by Eq.~\eqref{eq:def_D}.
In terms of the SCHA Green function defined in Eq.~\eqref{eq:relaz_GS_DS},
Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_00} is readily written as
\begin{equation}
-{\boldsymbol{D}}{}^{\scriptscriptstyle{(F)}}=\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(S)}}(0)-\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)\,,
\label{eq:diagrammatic_static_invSCHAdynmat_0}
\end{equation}
which is equivalent to the Dyson-like equation
\begin{equation}
-\overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}{}_{\scriptscriptstyle{(F)}}=\boldsymbol{G}_{\scriptscriptstyle{(S)}}(0)+\boldsymbol{G}_{\scriptscriptstyle{(S)}}(0)\,\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)(-\overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}_{\scriptscriptstyle{(F)}})\,,
\label{eq:diagrammatic_static_invSCHAdynmat_1}
\end{equation}
where the matrix product is understood.
If the opportune diagram symmetry factors are taken into account, Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_1}
with Eq.~\eqref{eq:diagrammatic_static_selfenergy} have the Feynman diagrams representation
shown in Fig.~\ref{fig:Feyman_diagr_stat}a and Fig.~\ref{fig:Feyman_diagr_stat}b. This is the diagrammatic representation of the curvature formula~\eqref{eq:compact_expr_curv_daimpl}
(divided by the square root of the masses).
Analogous diagrammatic series has been obtained by G\"otze and Michel in Ref.~\citenum{GotzeZfurphys}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{tab_fey_static.pdf}
\caption{Figure $a)$: Diagrammatic representation of Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_1}.
Figure $b)$: Diagrammatic representation of the SCHA self-energy $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}$, Eq.~\eqref{eq:diagrammatic_static_selfenergy}.
Since in that equation only the static value $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$ is considered,
the sum over the frequencies of the internal lines is performed, but the total
frequency is kept equal to zero.
Figure $c)$: Diagrammatic representation of $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}$, the bubble part of the SCHA self-energy,
Eq.~\eqref{eq:diagrammatic_static_selfenergy_bubble}.}
\label{fig:Feyman_diagr_stat}
\end{figure}
The first term of the series giving $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$ is the SCHA `bubble' $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$.
It is given by the formula
\begin{equation}
\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)=\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(0)\right)\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}
\label{eq:diagrammatic_static_selfenergy_bubble}
\end{equation}
and corresponds to the diagram in Fig.~\ref{fig:Feyman_diagr_stat}c. The SCHA `bubble' is
the term $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$
of Eq.~\eqref{eq:compact_expr_curv}, divided by the square root of the masses.
This explains the name `bubble' given to that term.
Before concluding this section, it is worthwhile to remark that, in spite of the symbol used, at this level
the $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$ defined in Eq.~\eqref{eq:diagrammatic_static_selfenergy} is just an auxiliary quantity, without a specific physical meaning.
However, the choice of the symbol is not casual because later we will interpreted it as a self-energy.
This will give a deeper meaning to the results obtained.
\section{Stochastic implementation}
\label{sec:Stochastic_implementation}
The stochastic implementation of the SCHA (SSCHA) has demonstrated to be an efficient method
{to analyze thermal properties of solids
in situations where the harmonic approximation breaks down~\cite{PhysRevLett.111.177002,PhysRevLett.114.157004,PhysRevB.89.064302,Errea81,
PhysRevB.93.174308,PhysRevB.92.140303,0953-8984-28-49-494001}}. The SSCHA is described in Ref.~\citenum{PhysRevB.89.064302}
and consists in minimizing with {a} conjugate-gradient (CG) method
the functional $\mathcal{F}[\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}]$ with respect to $\boldsymbol{\Rcal}$ and $\boldsymbol{\varPhi}$.
The functional and its gradient
are expressed {as averages taken with $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}$
of observables $\mathds{O}(\boldsymbol{R})=\mathds{O}\bigl(V(\boldsymbol{R}),\bf(\boldsymbol{R})\bigr)$ that are functions only
of the potential $V(\boldsymbol{R})$ and
the forces $\bf(\boldsymbol{R})=-\partial V/\partial \boldsymbol{R}$}.
The method is `stochastic' because these averages are evaluated with the {importance} sampling technique.
Since the observables depend only on the position, Eqs.~\eqref{eq:avg_pos_operator}--\eqref{eq:dens_prob} apply.
The space of configurations is statistically sampled with a (large) population of finite size $N_{\mathcal{I}}$,
whose members $\boldsymbol{R}_{\scriptscriptstyle{(\Ical)}}$ are distributed according to the probability density $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$.
For each element $\boldsymbol{R}_{\scriptscriptstyle{(\Ical)}}=\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}}$, {$\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}}$} being the displacement from the
centroids $\boldsymbol{\Rcal}$,
the forces $\bf(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})$
and the potential energy $V(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})$
are calculated {by any energy-force engine,
i.e., making use of first-principles methods or empirical potentials}. In that way the average integrals can be straightforwardly
computed. However, at each step of the CG minimization algorithm the distribution probability $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$ changes. Thus, in principle,
at each minimization step a new population should be generated
and for its members the energies and the forces should be calculated. In order to reduce the number of calls to the energy-force engine,
in actual calculations a reweighting procedure is adopted~\cite{PhysRevB.89.064302}.
Energy and forces are computed only once for the population elements that are distributed according to an initially fixed probability density
$\tilde{\rho}{}_{\scriptscriptstyle{\text{in}}}(\boldsymbol{R})$.
The approximate averages for a generic distribution probability $\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{R})$ are then computed {as}
\begin{align}
\Bavg{\mathds{O}}_{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}}&\simeq
\frac{1}{N_{\mathcal{I}}}\sum_{\mathcal{I}=1}^{N_{\mathcal{I}}}
\frac{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}{\tilde{\rho}{}_{\scriptscriptstyle{\text{in}}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}\nonumber\\
&\mkern48mu\times\mathds{O}\bigl(V(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}}),\bf(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})\bigr)\vphantom{\frac{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}{\tilde{\rho}^{\scriptscriptstyle{(0)}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}}
{.}
\label{eq:average_montecarlo}
\end{align}
{Obviously, the equality holds for $N_{\mathcal{I}}\rightarrow+\infty$.}
We want to use the stochastic approach also to compute the free energy curvature through Eq.~\eqref{eq:compact_expr_curv_daimpl}.
{Considering} a configuration $\boldsymbol{\Rcal}$, after the SSCHA minimization of the functional $\mathcal{F}[\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}]$
with respect to $\boldsymbol{\varPhi}$, the SCHA matrix $\boldsymbol{\Phi}$ for that configuration is available{. T}herefore we only
need to express $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ in a form that is suited for the stochastic calculation
(here and in what follows the dependence of the matrices on $\boldsymbol{\Rcal}$ is understood).
{As demonstrated in Appendix~\ref{app:Stochastic_calculation_of_derivative_averages} (see
Eqs.~\eqref{eq:app_phi3_con_fbb},~\eqref{eq:app_SCHA3_mat_fin} and
Eqs.~\eqref{eq:app_phi4_con_fbb},~\eqref{eq:app_SCHA4_mat_fin} with Eq.~\eqref{eq:app_def_fbb_used_finale}), it can
be shown making use of integration by parts that}
\begin{subequations}\begin{align}
&\overset{\scriptscriptstyle{(3)}}{\Phi}_{abc}=-\sum_{pq}\Upsilon_{ap}\Upsilon_{bq}
\Bavg{u^pu^q\, \mathbb{f}_c}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}
\label{Eq:stoc_avg_3rd}\\
&\overset{\scriptscriptstyle{(4)}}{\Phi}_{abcd}=-\sum_{pqr}\Upsilon_{ap}\Upsilon_{bq}\Upsilon_{cr}\,
\Bavg{u^pu^qu^r\,\mathbb{f}_d}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}}\,.
\label{Eq:stoc_avg_4th}
\end{align}\label{Eq:stoc_avg_Vbb}\end{subequations}
{Here
$\Upsilon_{ab}$ is the matrix obtained from $\Phi_{ab}$ through the definition~\eqref{eq:def_Upsilon}, and
\begin{equation}
\mathbb{f}_i=\text{f}_i-\biggl[\Bavg{\text{f}_i}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}}-\sum_j\Phi_{ij}\,u^j\biggr]\,.
\label{eq:def_fbb}
\end{equation}
The equations~\eqref{Eq:stoc_avg_Vbb} express the third and {fourth} order SCHA tensors
in terms of averages of forces and displacements only (in the definition~\eqref{eq:def_fbb}
the term subtracted from the forces $\text{f}_a$ is computed analytically with negligible cost, since $\Bavg{\partial V/\partial \boldsymbol{R}}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}}$ and $\boldsymbol{\Phi}$
are known). Therefore, they can be calculated through Eq.~\eqref{eq:average_montecarlo}.
It is interesting to observe that, in the limit of an infinitely large population sampling,
adding to $\mathbb{f}_i$ a term odd in the displacements does not change the value of $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ obtained from Eq.~\eqref{Eq:stoc_avg_3rd}.
Therefore, the $\mathbb{f}_i$ used in Eq.~\eqref{Eq:stoc_avg_3rd} is actually defined only up to an additive factor that is odd in the displacements.
Analogously, if we use an infinite sampling, the $\mathbb{f}_i$ used in Eqs.~\eqref{Eq:stoc_avg_4th} is defined only up to an additive factor that is even in the displacements.
However, depending on the actual $\mathbb{f}_i$ used, we obtain different results when we use a finite sampling to
compute the averages.
The specific choice of Eq.~\eqref{eq:def_fbb}, identical for both equations~\eqref{Eq:stoc_avg_Vbb}, guarantees that if the potential $V$ is quadratic,
then the SSCHA tensors (i.e. the SCHA tensors calculated stochastically)
$\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ are correctly zero with any finite sampling used to compute the averages.
Therefore, the definition~\eqref{eq:def_fbb} reduces the stochastic error and accelerates the convergence.
Notice that if we compute the curvature of the free energy in a stationary point, since it is $\partial F/\partial\boldsymbol{\Rcal}=0$
then from Eq.~\eqref{eq:first_derivative_F} the term $\Bavg{\text{f}_i}_{{\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}}$ in Eq.~\eqref{eq:def_fbb} is zero.
In particular, this is true when we evaluate the curvature in the equilibrium configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$, which is the relevant case when
we study structural second order phase transitions.
In the limit of a fully converged stochastic calculation, the SSCHA tensors $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and
$\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ satisfy
both acoustic sum rule (ASR) and invariance with respect to permutations of indices and symmetry transformations.
Actually, in Appendix~\ref{app:Stochastic_calculation_of_derivative_averages} it is shown that
the SSCHA $n$-th order tensor satisfies the ASR with any finite population sampling,
as long as the total force acting on the system is zero for any population element (as it must be), and the ASR is satisfied by $\boldsymbol{\Phi}$.
Therefore, it is not necessary to impose any extra-condition to make $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$
satisfy the ASR.
For the invariance properties the situation is different. We can distinguish two kind of
operators acting on a tensor $\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}$:
the $n!$ operators $\mathcal{T}_{\pi}$, which permute the tensor indices according to the permutations $\pi\in\sigma_n$, and
the $N_{\scriptscriptstyle{\text{sym}}}$ operators $\mathcal{T}_{S}$, whose action corresponds to the symmetry transformations $S\in\mathcal{G}_{\scriptscriptstyle{\text{sym}}}$
(excluding lattice translations, if it is a crystal).
If we are considering a crystal, the SSCHA calculation is performed on a supercell made of $N_{\text{c}}$ unit cells, with periodic
boundary conditions. In that case we consider also the $N_{\text{c}}$ operators $\mathcal{T}_{\boldsymbol{l}}$ whose action
corresponds to the translations by lattice vectors
non commensurate with the supercell $\boldsymbol{l}\in\mathcal{G}_{\scriptscriptstyle{\text{lat}}}$. The SSCHA tensors are invariant
with respect to these operations only in the limit $N_{\mathcal{I}}\rightarrow +\infty$ (for simplicity
we consider the crystal case):
\begin{subequations}\begin{align}
\mathcal{T}_{\pi}\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&=\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&&\forall \pi\in\sigma_n \label{eq:perm_prop}\\
\mathcal{T}_{S}\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&=\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&&\forall S\in\mathcal{G}_{\scriptscriptstyle{\text{sym}}} \label{eq:sym_prop}\\
\mathcal{T}_{\boldsymbol{l}}\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&=\overset{\scriptscriptstyle{(n)}}{\boldsymbol{\Phi}}&&\forall \boldsymbol{l}\in\mathcal{G}_{\scriptscriptstyle{\text{lat}}}\,. \label{eq:lat_prop}
\end{align}\label{eq:_symmet_prop_gen}\end{subequations}
For calculations performed with finite-size populations these conditions are not satisfied.
We enforce them by applying the projectors $\mathcal{P}_{\scriptscriptstyle{\text{perm}}}$, $\mathcal{P}_{\scriptscriptstyle{\text{sym}}}$ and $\mathcal{P}_{\scriptscriptstyle{\text{lat}}}$ to the result:
\begin{subequations}\begin{align}
\mathcal{P}_{\scriptscriptstyle{\text{perm}}}&=\frac{1}{n!}\sum_{\pi\in\sigma_n}\mathcal{T}_{\pi}\label{eq:perm_proj}\\
\mathcal{P}_{\scriptscriptstyle{\text{sym}}}&=\frac{1}{N_{\scriptscriptstyle{\text{sym}}}}\sum_{S\in\mathcal{G}_{\scriptscriptstyle{\text{sym}}} }\mathcal{T}_{S} \label{eq:sym_proj}\\
\mathcal{P}_{\scriptscriptstyle{\text{lat}}}&=\frac{1}{N_{\text{c}}}\sum_{\boldsymbol{l}\in \mathcal{G}_{\scriptscriptstyle{\text{lat}}}}\mathcal{T}_{\boldsymbol{l}}\,. \label{eq:lat_proj}
\end{align}\label{eq:_symmet_proj_gen}\end{subequations}
For calculations with finite sampling the action of the projectors~\eqref{eq:_symmet_proj_gen} has two benefits:
we obtain SSCHA tensors with the correct properties and we reduce the statistical noise and improve, with negligible cost, the rapidity of the statistical convergence
with respect to $N_{\mathcal{I}}$.
Indeed, the necessity of imposing the property~\eqref{eq:perm_prop} is due to the
fact that Eqs.~\eqref{Eq:stoc_avg_Vbb} are not symmetric with respect to permutation of indices. That is caused by
the arbitrariness in the choice of the variables integrated by parts in the derivation of the formulas, shown in appendix~\ref{app:Stochastic_calculation_of_derivative_averages}.
As a consequence, an approximate evaluation of the averages causes spurious asymmetries, which are eliminated by applying the projector
$\mathcal{P}_{\scriptscriptstyle{\text{perm}}}$ to the result. The necessity of imposing the properties~\eqref{eq:sym_prop} and~\eqref{eq:lat_prop} is instead due to the fact that, in general,
the population generated to compute the averages is composed of elements whose distribution in configuration space
does not respect the symmetries of the system.
This leads to spurious fluctuations which spoil the symmetry properties of the result and which are eliminated by applying the projectors
$\mathcal{P}_{\scriptscriptstyle{\text{sym}}}$ and $\mathcal{P}_{\scriptscriptstyle{\text{lat}}}$. Applying these projectors to the result corresponds
to computing the averages through Eq.~\eqref{eq:average_montecarlo} using a larger population of $N_{\scriptscriptstyle{\text{lat}}}\times N_{\scriptscriptstyle{\text{sym}}}\times N_{\mathcal{I}}$ elements
obtained by applying the $N_{\scriptscriptstyle{\text{lat}}}\times N_{\text{sym}}$ symmetry operations on the $N_{\mathcal{I}}$ members of the original population.
In conclusion, the formulas implemented in the SSCHA are:
\allowdisplaybreaks[0]
\begin{subequations}\begin{align}
&\overset{\scriptscriptstyle{(3)}}{\Phi}_{abc}\simeq
\mathcal{P}_{\scriptscriptstyle{\text{sym}}}\mathcal{P}_{\scriptscriptstyle{\text{lat}}}\mathcal{P}_{\scriptscriptstyle{\text{perm}}}
\frac{1}{N_{\mathcal{I}}}\sum_{\mathcal{I}}\frac{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}{\tilde{\rho}{}_{\scriptscriptstyle{\text{in}}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}\nonumber\\
&\mkern80mu\times\biggl[-\sum_{pq}\Upsilon_{ap}\Upsilon_{bq}\,
u_{\scriptscriptstyle{(\Ical)}}^p\,u_{\scriptscriptstyle{(\Ical)}}^q\, \mathbb{f}_c(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})\biggr]\label{Eq:stoc_avg_Vbb_3rd_impl}\,,\\
\allowdisplaybreaks
\allowdisplaybreaks[0]
&\overset{\scriptscriptstyle{(4)}}{\Phi}_{abcd}\simeq
\mathcal{P}_{\scriptscriptstyle{\text{sym}}}\mathcal{P}_{\scriptscriptstyle{\text{lat}}}\mathcal{P}_{\scriptscriptstyle{\text{perm}}}
\frac{1}{N_{\mathcal{I}}}\sum_{\mathcal{I}}\frac{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}{\tilde{\rho}{}_{\scriptscriptstyle{\text{in}}}(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})}\nonumber\\
&\mkern10mu\times\biggl[-\sum_{pqr}\Upsilon_{ap}\Upsilon_{bq}\Upsilon_{cr}\,
u_{\scriptscriptstyle{(\Ical)}}^pu_{\scriptscriptstyle{(\Ical)}}^qu_{\scriptscriptstyle{(\Ical)}}^r\,\mathbb{f}_d(\boldsymbol{\Rcal}+\boldsymbol{u}_{\scriptscriptstyle{(\Ical)}})\biggr]
\,.\label{Eq:stoc_avg_Vbb_4th_impl}
\end{align}\label{Eq:stoc_avg_Vbb_impl}\end{subequations}
\allowdisplaybreaks
\section{Perturbative limit}
\label{sec:Perturbative_limit}
In this section we analyze the lowest perturbative order of the SCHA and of the
free energy dynamical matrix $D^{\scriptscriptstyle{(F)}}_{ab}$. First we set some definitions. Expanding the potential $V$ around
its minimum $\mathcal{R}^a_{\scriptscriptstyle{(0)}}$, the Hamiltonian $H$ is written as
\begin{equation}
H=H^{\scriptscriptstyle{(0)}}+\sum_{n\ge 3}\,\frac{1}{n!}\sum_{a_1\cdots a_n}\,\overset{\scriptscriptstyle{(n)}}{\phi}_{a_1\ldots a_n}u^{a_1}\ldots u^{a_n}\,,
\label{eq:Ham_harm+anaharm}
\end{equation}
where $u^a=R^a-\mathcal{R}^a_{\scriptscriptstyle{(0)}}$ is the displacement with respect to the potential minimum,
\begin{equation}
H^{\scriptscriptstyle{(0)}}=\sum_a\frac{p_a^2}{2M_a}+V(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})+\frac{1}{2}\sum_{ab}\phi_{ab}\,u^au^b
\label{eq:Ham_harm}
\end{equation}
is the quadratic harmonic Hamiltonian, and
\begin{equation}
\overset{\scriptscriptstyle{(n)}}{\phi}_{a_1\ldots a_n}=\left.\frac{\partial^nV}{\partial R^{a_1}\ldots\partial R^{a_n}}\right|_{\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}}
\label{eq:n-th_order_force_constant}
\end{equation}
is the $n$-th order force constant tensor. Notice that for the second order force constant matrix $\phi_{ab}$
we do not use the superscript $\scriptstyle{(2)}$. In order to avoid confusion, it is worthwhile to stress that the $n$-th
force constant $\overset{\scriptscriptstyle{(n)}}{\phi}_{a_1\ldots a_n}$ is the $n$-th derivative of the potential, evaluated at the potential minimum $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$,
whereas the $n$-th SCHA tensor $\overset{\scriptscriptstyle{(n)}}{\Phi}_{a_1\ldots a_n}(\boldsymbol{\Rcal})$, defined in Eq.~\eqref{eq:SCHA_matrix_def_nth}, is the $n$-th derivative of the potential averaged
with the distribution $\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}(\boldsymbol{\Rcal})}$.
The part of the Hamiltonian in Eq.\eqref{eq:Ham_harm+anaharm} not included in $H^{\scriptscriptstyle{(0)}}$ defines the anharmonic part of the potential,
which we treat as a (small) perturbation of $H^{\scriptscriptstyle{(0)}}$. With $G^{ab}(z)$ and $G^{ab}_{\scriptscriptstyle{(0)}}(z)$ we indicate the Green function of
$H$ and $H^{\scriptscriptstyle{(0)}}$ for the variable $\sqrt{M_a}(R^a-\mathcal{R}^a_{\scriptscriptstyle{(0)}})$, respectively. The latter is given as
\begin{equation}
\overset{\scriptscriptstyle{-1}}{G}{}^{ab}_{\scriptscriptstyle{(0)}}(z)=z^2\delta^{ab}-D^{\scriptscriptstyle{(0)}}_{ab}\,,
\label{eq:relaz_G0_D0}
\end{equation}
where $D^{\scriptscriptstyle{(0)}}_{ab}=\phi_{ab}/\sqrt{M_aM_b}$ is the harmonic dynamical matrix, already defined in Eq.~\eqref{eq:Harmonic_dyn_mat}.
The relation between the full and harmonic Green functions is given by the Dyson equation
\begin{equation}
\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}(z)=\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(0)}}(z)-\boldsymbol{\Pi}^{\scriptscriptstyle{(0)}}(z)\,,
\label{eq:Dyson_harm}
\end{equation}
which is equivalent to
\begin{equation}
{\boldsymbol{G}}(z)={\boldsymbol{G}}{}_{\scriptscriptstyle{(0)}}(z)+{\boldsymbol{G}}{}_{\scriptscriptstyle{(0)}}(z)\,\boldsymbol{\Pi}^{\scriptscriptstyle{(0)}}(z)\,{\boldsymbol{G}}(z)\,,
\label{eq:Dyson_harm_1}
\end{equation}
where, in order to use a consistent notation, we have indicated with $\boldsymbol{\Pi}^{\scriptscriptstyle{(0)}}(z)$ the harmonic self-energy, i.e.
the self-energy obtained by taking $H^{\scriptscriptstyle{(0)}}$ as non-interacting unperturbed Hamiltonian.
At the lowest perturbative order~\cite{PhysRev.128.2589}
\begin{equation}
\boldsymbol{\Pi}^{\scriptscriptstyle{(0)}}(z)\simeq\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(z)\,,
\label{eq:perturbative_harmonic_selfenergy}
\end{equation}
where $\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}$, $\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}$ and $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(z)$
are the loop, tadpole and bubble harmonic self-energies, respectively, which have the following expressions:
\begin{equation}
\overset{\scriptscriptstyle{(L)}}{\Pi}{}^{\scriptscriptstyle{(0)}}_{ab}=-\frac{1}{2}\sum_{c_1c_2}\overset{\scriptscriptstyle{(4)}}{D}{}^{\scriptscriptstyle{(0)}}_{abc_1\!c_2}
\left[\frac{1}{\beta}\sum_lG^{c_1\!c_2}_{\scriptscriptstyle{(0)}}(i\Omega_l)\right]\,,
\label{eq:pert_loop}
\end{equation}
\begin{align}
\overset{\scriptscriptstyle{(T)}}{\Pi}{}^{\scriptscriptstyle{(0)}}_{ab}=&-\frac{1}{2}\sum_{\substack{c_1c_2\\d_1d_2}}\overset{\scriptscriptstyle{(3)}}{D}{}^{\scriptscriptstyle{(0)}}_{abc_1}\overset{\scriptscriptstyle{(0)}}{G}{}_{\scriptscriptstyle{(0)}}^{c_1\!c_2}(0)
\overset{\scriptscriptstyle{(3)}}{D}{}^{\scriptscriptstyle{(0)}}_{c_2d_1\!d_2}\nonumber\\
&\qquad\qquad\qquad\qquad\times\left[\frac{1}{\beta}\sum_lG^{d_1\!d_2}_{\scriptscriptstyle{(0)}}(i\Omega_l)\right]\,,
\label{eq:pert_tadpole}
\end{align}
\begin{align}\overset{\scriptscriptstyle{(B)}}{\Pi}{}^{\scriptscriptstyle{(0)}}_{ab}(z)=&-\frac{1}{2}\sum_{\substack{c_1c_2\\d_1d_2}}\overset{\scriptscriptstyle{(3)}}{D}{}^{\scriptscriptstyle{(0)}}_{ac_1\!c_2}\overset{\scriptscriptstyle{(3)}}{D}{}^{\scriptscriptstyle{(0)}}_{bd_1\!d_2}\nonumber\\
&\quad\times\left[\frac{1}{\beta}\sum_lG_{\scriptscriptstyle{(0)}}^{c_1\!c_2}(i\Omega_l)G_{\scriptscriptstyle{(0)}}^{d_1\!d_2}(z-i\Omega_{l})\right]\,.
\label{eq:pert_bubble}
\end{align}
Here we have generalized the definition~\eqref{eq:Harmonic_dyn_mat} of the harmonic dynamical matrix to the $n$-th order:
\begin{equation}
\overset{\scriptscriptstyle{(n)}}{D}{}^{\scriptscriptstyle{(0)}}_{a_1\cdots a_n}=\frac{ \overset{\scriptscriptstyle{(n)}}{\phi}{}_{a_1\cdots a_n}}{\sqrt{M_{a_1}\cdots M_{a_n}}}\,.
\label{eq:def_dyn_harm_nth}
\end{equation}
Notice that loop and tadpole self-energies do not depend on the value of the frequency $z$. In fact they are real symmetric.
On the contrary, the bubble is a complex symmetric matrix depending on $z$.
In Fig.~\ref{fig:Feyman_diagr_pert} the diagrammatic representation of the harmonic perturbative result at the lowest order is shown.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{tab_fey_pert.pdf}
\caption{Diagrammatic description of the harmonic perturbation theory at the lowest perturbative order, see Eqs.~\eqref{eq:Dyson_harm_1}--\eqref{eq:pert_bubble}.
The dashed line corresponds to the harmonic propagator. The double solid line corresponds to the full propagator.
Notice that in Fig.~\ref{fig:Feyman_diagr_stat} we have already used a double solid line to indicate $\overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}_{\scriptscriptstyle{(F)}}$. This is not casual because later
we will interpret $\overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}_{\scriptscriptstyle{(F)}}$ as the static full propagator (see Eq.~\eqref{eq:staticG_DF}).
Third and fourth order vertices
are associated to $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(0)}}/3!$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(0)}}/4!$, respectively (see definition~\eqref{eq:def_dyn_harm_nth}).
Sum over internal degrees of freedom is performed.}
\label{fig:Feyman_diagr_pert}
\end{figure}
From the SCHA equations, retaining only the lowest order corrections to the harmonic values $\mathcal{R}^a_{\scriptscriptstyle{(0)}}$
and $\phi_{ab}$, using the SCHA matrix defined in Eq.~\eqref{eq:def_D} we obtain (see Eq.~\eqref{eq:pert_DS})
\begin{equation}
\boldsymbol{D}^{\scriptscriptstyle{(S)}}\simeq\boldsymbol{D}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}\,.
\label{eq:SCHA_Green_perturb_dyn}
\end{equation}
Equivalently, using the SCHA propagator $\boldsymbol{G}_{\scriptscriptstyle{(S)}}(z)$
defined in Eq.~\eqref{eq:relaz_GS_DS}
we can write
\begin{equation}
\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(S)}}(z)\simeq\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(0)}}(z)-\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}-\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}\,,
\label{eq:SCHA_Green_perturb}
\end{equation}
that is
\begin{equation}
{\boldsymbol{G}}{}_{\scriptscriptstyle{(S)}}(z)\simeq{\boldsymbol{G}}{}_{\scriptscriptstyle{(0)}}(z)+{\boldsymbol{G}}{}_{\scriptscriptstyle{(0)}}(z)\left[\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}\right]
{\boldsymbol{G}}{}_{\scriptscriptstyle{(S)}}(z)\,.
\label{eq:SCHA_Green_perturb_2}
\end{equation}
At the lowest perturbative order we also have (see Eq.~\eqref{eq:app_per_bubb})
\begin{equation}
\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)\simeq\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)\simeq\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(0)
\label{eq:SCHA_staticself_perturb}
\end{equation}
where $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$ and $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$ are the quantities
defined in Eq.~\eqref{eq:diagrammatic_static_selfenergy} and Eq.~\eqref{eq:diagrammatic_static_selfenergy_bubble}, respectively.
From Eqs.~\eqref{eq:SCHA_Green_perturb_dyn}--\eqref{eq:SCHA_Green_perturb_2}
we see that, at the lowest perturbative order, the SCHA and harmonic propagators
are related through the harmonic loop and tadpole self-energies only~\cite{PhysRevB.91.054304}. However, from Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_00}
and Eq.~\eqref{eq:SCHA_staticself_perturb} we see that
in order to obtain the SCHA dynamical matrix, defined in Eq.~\eqref{eq:SCHA_dyn_mat}, we need the harmonic static bubble too:
\begin{equation}
{\boldsymbol{D}}{}^{\scriptscriptstyle{(F)}}\simeq{\boldsymbol{D}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(0)\,.
\label{eq:pert_dyna_mat}
\end{equation}
Notice that, in particular, this implies that the term $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\Theta}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$
in the curvature formula, Eq.~\eqref{eq:compact_expr_curv}, can be discarded at the lowest perturbative order.
In terms of the harmonic propagator defined in Eq.~\eqref{eq:relaz_G0_D0}, the formula~\eqref{eq:pert_dyna_mat} can be written as (Cfr.~Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_1})
\begin{align}
&-\overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}{}_{\scriptscriptstyle{(F)}}\simeq \boldsymbol{G}_{\scriptscriptstyle{(0)}}(0)\nonumber\\
&\mkern40mu+\boldsymbol{G}_{\scriptscriptstyle{(0)}}(0)\left[\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}+\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(0)\right]
(- \overset{\scriptscriptstyle{-1}}{\boldsymbol{D}}{}_{\scriptscriptstyle{(F)}})\,.
\label{eq:static_perturbative}
\end{align}
Equations~\eqref{eq:SCHA_Green_perturb_2} and~\eqref{eq:static_perturbative} are the main SCHA results at the lowest harmonic
perturbative order. They are represented in diagrammatic form in Fig.~\ref{fig:pert_res}a and in Fig.~\ref{fig:pert_res}b, respectively.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{pert_res.pdf}
\caption{Diagrammatic description of the SCHA results at the lowest harmonic perturbative order. Figure a):
relation between the SCHA and the harmonic propagator,
Eq.~\eqref{eq:SCHA_Green_perturb_2}.
Figure b): relation between the free energy dynamical matrix within SCHA and the harmonic propagator, Eq.~\eqref{eq:static_perturbative}.}
\label{fig:pert_res}
\end{figure}
It is interesting to observe that, at the lowest perturbative order,
the free energy curvature takes into account only the static harmonic bubble,
whereas in the full propagator the bubble actually depends on the frequency $z$, as we can see from Eq.~\eqref{eq:perturbative_harmonic_selfenergy}.
This is consistent with the fact that we have developed only a `static' theory
(obviously, this fact does not have consequences for the tadpole and loop term, because they do not depend on the frequency).
In the next section we will investigate possible dynamic extensions of the results found thus far.
\section{Ansatz for a dynamic theory}
\label{sec:Anstaz_for_a_dynamic_theory}
In this section we propose a possible `dynamical' extension of the `static' results obtained above. This could be used to interpret
the outcomes of inelastic scattering processes between phonons and external incident particles (typically neutrons)
in the framework of the SCHA approximation. The extension that we are going to present
is reasonable because it returns the expected results in two limits. In the static limit
it gives results coherent with the ones already obtained for the free energy curvature and
at the lowest perturbative order it gives the correct results already known in literature.
Nevertheless, it is worthwhile to stress that, at variance with the `static' results, the dynamical extension that we are going
to propose is only an ansatz, reasonable but not based on a rigorous demonstration.
For that reason it can be considered as a basis for a future rigorous extension of the
static theory.
Fixed the temperature, and the relative $\mathcal{R}_{\scriptscriptstyle{\text{eq}}}^a$, we consider the full Green function
$G_{ab}(z)$ for $H$ and the Green function $G^{\scriptscriptstyle{(S)}}_{ab}(z)$ for $H^{\scriptscriptstyle{(S)}}$ in the variable $\sqrt{M_a}(R^a-\mathcal{R}_{\scriptscriptstyle{\text{eq}}}^a)$.
We consider a Dyson-type relation between them:
\begin{equation}
\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}(z)=\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(S)}}(z)-\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)\,,
\label{eq:Green_func_ansatz}
\end{equation}
which is equivalent to
\begin{equation}
{\boldsymbol{G}}(z)={\boldsymbol{G}}{}_{\scriptscriptstyle{(S)}}(z)+{\boldsymbol{G}}{}_{\scriptscriptstyle{(S)}}(z)\,\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)\,{\boldsymbol{G}}(z)\,,
\label{eq:Green_func_ansatz_1}
\end{equation}
where $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$ is the SCHA self-energy. The aim of this section is to propose an expression for $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$.
The first assumption is that its static value, i.e. its value for $z=0$, is given by Eq.~\eqref{eq:diagrammatic_static_selfenergy}.
At that level, the symbol used did not have a physical meaning.
Now we are explicitly interpreting it as the static SCHA self-energy.
Comparing Eq.~\eqref{eq:Green_func_ansatz} to Eq.~\eqref{eq:diagrammatic_static_invSCHAdynmat_0}, this is equivalent to saying that
\begin{equation}
\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}(0)=-\boldsymbol{D}^{\scriptscriptstyle{(F)}}\,.
\label{eq:staticG_DF}
\end{equation}
This is the same kind of relation that exists between the harmonic static Green function and the harmonic dynamical matrix.
Therefore, Eq.~\eqref{eq:staticG_DF} gives a deeper meaning to the consideration in Sec.\ref{sec:phonons_in_the_SCHA}
that $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ is the anharmonic generalization of the harmonic dynamical matrix. A real pole of the Green function
corresponds to the energy of a phonon with zero linewidth, i.e. with infinite lifetime. Equation~\eqref{eq:staticG_DF}
means that we observe a phonon with zero energy, i.e. we see a phonon softening and therefore an instability, when
$\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ has a null eigenvalue. This is exactly the result found in Sec.~\ref{Derivatives_of_F} and Sec.~\ref{sec:phonons_in_the_SCHA}.
Thus the interpretation of Eq.~\eqref{eq:diagrammatic_static_selfenergy} as the static SCHA self-energy is
consistent with the rigorous (static) results obtained for the free energy curvature.
The subsequent step is to give an expression for the SCHA self-energy at $z$ different from zero.
As a second part of our hypothesis, we assume for $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$
the same structure of $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$, given by Eq.~\eqref{eq:diagrammatic_static_selfenergy} and
illustrated by the diagrams in Fig.~\ref{fig:Feyman_diagr_stat}b, but readily generalized to any $z$. Therefore it is
\allowdisplaybreaks[0]
\begin{align}
\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)&=\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(z)\right)\nonumber\\
&\qquad\times\left[\mathds{1}-\overset{\boldsymbol{\ssf}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(z)\right)\right]^{-1}\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\,,
\label{eq:diagrammatic_dynamic_selfenergy}
\end{align}
\allowdisplaybreaks
with
\begin{equation}
\chi_{\scriptscriptstyle{(S)}}^{abcd}(z)=\frac{1}{\beta}\sum_l\,{G}_{\scriptscriptstyle{(S)}}^{ac}(i\Omega_l){G}_{\scriptscriptstyle{(S)}}^{bd}(z-i\Omega_{l})\,.
\label{eq:dynamic_loop}
\end{equation}
Using standard techniques for Matsubara frequencies summations~\cite{mahan2000many},
we obtain an explicit expression for this term:
\begin{align}
&\frac{1}{\beta}\sum_lG_{\scriptscriptstyle{(S)}}^{ac}(i\Omega_{l})G_{\scriptscriptstyle{(S)}}^{bd}(z-i\Omega_{l})=\nonumber\\
&\quad\qquad\frac{\hbar^2}{4}\sum_{\mu\nu}\frac{F(z,\omega_{\mu},\omega_{\nu})}{\omega_{\mu}\omega_{\nu}}\,e^a_{\nu}e^b_{\mu}e^c_{\nu}e^d_{\mu}\,,
\label{eq:sumMatzub_2}
\end{align}
where $\omega_{\mu}^2$ and $e^a_{\mu}$ are eigenvalues and corresponding eigenvectors of $D^{\scriptscriptstyle{(S)}}_{ab}$, respectively, and for $z\neq 0$
\begin{align}
F(z,\omega_\nu,\omega_\mu)&=\frac{2}{\hbar}\Biggl[\frac{(\omega_\nu+\omega_\mu)[1+n_\nu+n_\mu]}{(\omega_\nu+\omega_\mu)^2-z^2}\nonumber\\
&\qquad-\frac{(\omega_\nu-\omega_\mu)[n_\nu-n_\mu]}{(\omega_\nu-\omega_\mu)^2-z^2}\Biggr]\,.
\label{eq:def_F}
\end{align}
The assumption expressed by Eqs.~\eqref{eq:diagrammatic_dynamic_selfenergy},~\eqref{eq:dynamic_loop} is reasonable because at the lowest perturbative limit it gives the correct result. Indeed,
by using the same arguments of Sec.~\ref{sec:Perturbative_limit}, at the lowest perturbative order
we readily generalize Eq.~\eqref{eq:SCHA_staticself_perturb} to
\begin{equation}
\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)\simeq\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(z)\,.
\end{equation}
Thus, from Eq.~\eqref{eq:SCHA_Green_perturb} and Eq.~\eqref{eq:Green_func_ansatz} we obtain
\begin{equation}
\overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}(z)\simeq \overset{\scriptscriptstyle{-1}}{\boldsymbol{G}}{}^{\scriptscriptstyle{(0)}}(z)-\overset{\scriptscriptstyle{(T)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}-\overset{\scriptscriptstyle{(L)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}-\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(0)}}(z)\,,
\end{equation}
which is the correct perturbative result shown in Eqs.~\eqref{eq:Dyson_harm}
and~\eqref{eq:perturbative_harmonic_selfenergy}. In conclusion, according to our ansatz, the full Green function $\boldsymbol{G}(z)$
is (approximately) given by Eq.~\eqref{eq:Green_func_ansatz_1} with Eqs.~\eqref{eq:diagrammatic_dynamic_selfenergy},~\eqref{eq:dynamic_loop}.
In that way we obtain a minimal extension of the static theory which reproduces the correct instabilities and gives
the correct results at the lowest perturbative level. By using this formula we can study anharmonic effects in a non perturbative way also for
the dynamic case. In Fig~\ref{fig:dyn_ansatz} we give the diagrammatic expression for our ansatz, the self-energy $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$
being the one in Fig.~\ref{fig:Feyman_diagr_stat}b.
An analogous diagrammatic series has been proposed in Ref.~\citenum{PhysRevLett.24.1424}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{tab_fey_ansatz}
\caption{Diagrammatic representation of our dynamical conjecture, Eq.~\eqref{eq:Green_func_ansatz_1}. It is the generalization
to $z\neq 0$ of the static result represented in Fig.~\ref{fig:Feyman_diagr_stat}a. With $\boldsymbol{G}$ and $\boldsymbol{G}{}_{\scriptscriptstyle{(S)}}$ we indicate the full Green function and
the SCHA Green function, Eq.~\eqref{eq:relaz_GS_DS}, for the variable $\sqrt{M_a}(\mathcal{R}^a-\mathcal{R}_{\scriptscriptstyle{\text{eq}}}^a)$, respectively.
The SCHA self-energy $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}$ is represented in Fig.~\ref{fig:Feyman_diagr_stat}b.}
\label{fig:dyn_ansatz}
\end{figure}
It is interesting to observe that, inspired by the perturbative result in Eq.~\eqref{eq:perturbative_harmonic_selfenergy},
one could be tempted to naively obtain a dynamic SCHA theory simply by adding
a dynamic bubble term on top of the standard SCHA results (which, as shown in Eq.~\eqref{eq:SCHA_Green_perturb}, contain only
tadpole and loop at the lowest perturbative level).
This approach of adding a dynamic bubble has been taken, for example,
in PbTe~\cite{PhysRevLett.112.175501} and PdH~\cite{PhysRevB.87.214303},
where the strong anharmonicity induces satellite peaks in the spectral function.
Now we can see that this essentially consists in adopting our ansatz, but discarding
all the terms in $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$ described by the diagrams of Fig.~\ref{fig:Feyman_diagr_stat}b, except the non-perturbative SCHA dynamic bubble
given in Fig.~\ref{fig:Feyman_diagr_stat}c:
\begin{equation}
\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(z)=\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\left(-\frac{1}{2}\,\boldsymbol{\chi}_{\scriptscriptstyle{(S)}}(z)\right)\overset{\boldsymbol{\sst}}{\boldsymbol{D}}{}^{\scriptscriptstyle{(S)}}\,.
\label{eq:sscha_bubble}
\end{equation}
This, in general, is not justified. As long as we consider a non-perturbative situation, there is in principle
no hierarchy that allows to discard the other terms. Therefore, the term given by Eq.~\eqref{eq:sscha_bubble}
has to be considered an incomplete expression for $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$ and a better choice is to take into account the full
expression of Eq.~\eqref{eq:diagrammatic_dynamic_selfenergy}.
Of course, there can be situations in which even if the regime is not perturbative, because the third order is not smaller
than the harmonic term, nevertheless the superior orders are smaller. In that case it would be justified to use Eq.~\eqref{eq:sscha_bubble}
to evaluate $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(z)$. However, this is a further assumption that, in order to be adopted,
has to be justified case by case.
\section{Numerical test}
\label{sec:Numerical_test}
In order to give a numerical demonstration of our findings, we apply
the theory to a toy model based on the SnTe crystal (an analogous model could be used for GeTe).
SnTe crystallizes at room temperature and ambient pressure in the NaCl-structure (Fm-3m),
called $\beta$-SnTe phase, where two fcc lattices of Sn and Te interpenetrate.
At low temperature, around $100$ K, it undergoes a phase transition and stabilizes in a rhombohedral structure (R3m),
called $\alpha$-SnTe. The phase transition can be described in terms of a two-step symmetry reduction:
a fixed unit cell polar displacement, between the two fcc, along the [111] cubic direction, which eliminates the inversion center,
and a strain of the unit cell along the cube diagonal~\cite{PSSB:PSSB201248412}. We concentrate on the first distortion.
We define the interatomic potential $V(\boldsymbol{u})$ of the toy-model as a function of the displacements $u^a=R^a-\mathcal{R}_{\scriptscriptstyle{(0)}}^a$
from the equilibrium position of the rock-salt structure $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$ and we keep, beyond the quadratic part,
only the anharmonic third and fourth order terms:
\begin{align}
V(\boldsymbol{u})&=\frac{1}{2}\sum_{ab}\phi_{ab}u^a u^b+
\frac{1}{3!}\sum_{abc}\overset{\scriptscriptstyle{(3)}}{\phi}{}_{abc}\,u^a u^bu^c\nonumber\\
&\qquad+\frac{1}{4!}\sum_{abcd}\overset{\scriptscriptstyle{(4)}}{\phi}{}_{abcd}\,u^a u^bu^cu^d\,.
\end{align}
The harmonic matrix $\phi_{ab}$ has been obtained from first principle calculation
for SnTe on a 2x2x2 grid of the Brillouin zone (BZ) (details in App.~\ref{app:Toy_model_definition}).
With the experimental lattice parameter $a_{\scriptscriptstyle{\text{exp}}}=6.312\,\textup{\AA}$ we do not observe any instability in the total energy
(i.e. the harmonic matrix is positive-definite). However, a lattice instability appears and increases at $\Gamma\in$ BZ
as we increase the lattice parameter. Therefore in order to achieve, for explicative purposes, an increased instability at the harmonic level,
we calculated ab initio the harmonic matrix with a higher lattice parameter: $a_{\scriptscriptstyle{\text{toy}}}=6.562\,\textup{\AA}$.
Moreover, in order to keep the toy model as simple as possible and focus on the main purpose of the numerical test,
we ignored the LO-TO splitting at $\Gamma$, which is present in real undoped SnTe samples.
In Fig.~\ref{fig:spec_harm} we show the obtained (harmonic) phonon dispersion along a high-symmetry path of the fcc BZ.
There are imaginary phonons in several points, the optical phonon in $\Gamma$
corresponding to the highest instability.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{spec_harm.pdf}
\caption{Harmonic phonon dispersion for the toy model along an high-symmetry path of the BZ.}
\label{fig:spec_harm}
\end{figure}
For the third and fourth order contributions, we follow the model described in Refs.~\citenum{PhysRevB.90.014308, PhysRevLett.113.105501}.
We define short-range anharmonic terms by using reciprocal displacements of nearest-neighbor atoms
(in the rock-salt structure each atom has 6 nearest-neighbors). In particular, as explained in App.~\ref{app:Toy_model_definition},
in our model $\phi^{\scriptscriptstyle{(3)}}_{abc}$ is proportional to a single parameter $p_3$, and $\phi^{\scriptscriptstyle{(4)}}_{abcd}$
is a linear function of two parameters $p_4$, $p_{4\chi}$. We take
$p_4=7.63\,\text{eV}/\textup{\AA}^4$, $p_{4\chi}=4.86\,\text{eV}/\textup{\AA}^4$ and $p_3=6.70\,\text{eV}/\textup{\AA}^3$.
\subsubsection{Free energy curvature}
We consider the free energy profile obtained by displacing the atoms in the unit cell along the
[111] cubic direction. In order to describe this distortion we write the atomic position $\boldsymbol{\Rcal}$ as a function of a
scalar, adimensional parameter $Q$:
\begin{equation}
\boldsymbol{\Rcal}(Q)=\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}+Q(\boldsymbol{\Rcal}_{\scriptscriptstyle{(1)}}-\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})\,,
\label{eq:par_Q_def}
\end{equation}
where $\boldsymbol{\Rcal}_{\scriptscriptstyle{(1)}}$ is the configuration corresponding to the minimum of the potential energy
along the distortion path.
Therefore, $\boldsymbol{\Rcal}(Q)$ is linear, $Q=0$ and $Q=1$ corresponding to the high symmetry phase (Fm-3m) and
to the low-symmetry energy minimum (R3m), respectively. In Fig.~\ref{fig:F_vs_T} we show $\Delta V(\boldsymbol{\Rcal}(Q))$, the variation
of the potential (per unit cell) along this distortion path.
This curve depends on $\phi_{ab}$, $p_4$, $p_{4\chi}$. The harmonic term is responsible for the initial decrease whereas
the fourth order term gives the subsequent increase. On the contrary, due to the symmetry of the rock-salt structure,
the value of $p_3$ is not relevant for the energy pattern (as a matter of fact, the value of $p_3$ does not affect
the energy value of any unit cell configurations).
The free energy along the path has been calculated with the SSCHA on a 2x2x2 supercell.
Fixed $\boldsymbol{\Rcal}$ and the temperature, the GB functional $\mathcal{F}[\displaystyle\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\varPhi}}]$ has been minimized
with respect to $\boldsymbol{\varPhi}$ as described in Ref.~\citenum{PhysRevLett.111.177002,PhysRevB.89.064302}.
In Fig.~\ref{fig:F_vs_T} we show a complete variation path for the free energy $\Delta F(\boldsymbol{\Rcal}(Q))$ at three temperatures.
For reasons that will be clear in a while, we studied also the case without third order ($p_3=0$).
However, it is interesting to remark that at $Q=0$ the SCHA result is independent from $p_3$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{path.pdf
\caption{Variation of the free energy,
at three temperatures ($0$K, $150$K, $300$K), as a function of the atomic displacement of Eq.~\eqref{eq:par_Q_def}.
Top panel: without third order. Bottom panel: with third order, $p_3=6.70$ eV/$\textup{\AA}^3$
Vertical axis: variation of the free energy (per unit cell) with respect to the value in the undistorted position,
$\Delta F(Q)=F(Q)-F(0)$, in meV. Horizontal axis: order parameter $Q$.
In the two plots the (temperature independent) variation of the potential energy $\Delta V(Q)=V(Q)-V(0)$ is also shown.}
\label{fig:F_vs_T}
\end{figure}
A first remarkable, somewhat counterintuitive, conclusion can be deduced from the results of Fig.~\ref{fig:F_vs_T}.
While the potential energy path $V(Q)$ is independent from $p_3$, at given temperature
the two free energy paths $F(Q)$ obtained with $p_3=0$ and $p_3=6.70\,\text{eV}/\textup{\AA}^3$ are considerably different.
This has important consequences. The presence or not of a second order phase transition and, when there is such a transition,
the transition temperature $T_c$ and the low-symmetry equilibrium configuration $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ for
$T<T_c$ are properties which cannot be inferred from the potential energy profile.
\begin{figure}
\centering
\includegraphics[width=1.12\columnwidth]{d2FdQ2_finite_diff}
\caption{Curvature of the free energy in the high-symmetry phase, $d^2F/dQ^2|_{Q=0}$, as a function of temperature $T$.
The transition temperature $T_c$ is around $140$ K. Lines: curvature calculated
with Eq.~\eqref{eq:application_curvature}. Three different quantities, contracted with $d\boldsymbol{\Rcal}/dQ$, are shown (see legend).
Dots: curvature, with and without third order, estimated by finite difference from the values of the free energy
calculated for several configurations around the high-symmetry phase.}
\label{fig:d2FdQ2_finite_diff}
\end{figure}
From the values of the free energy computed near $Q=0$, the curvature in the origin $d^2F/dQ^2|_{Q=0}$
has been evaluated by finite difference. The results at four temperatures are shown with dots in Fig.~\ref{fig:d2FdQ2_finite_diff}.
We compare these values with the curvature in $Q=0$ calculated
by contracting $d\boldsymbol{\Rcal}/dQ$ with the formula for $\partial^2\mathcal{F}/\partial\boldsymbol{\Rcal}\partial\boldsymbol{\Rcal}$ of
Eq.~\eqref{eq:compact_expr_curv}:
\allowdisplaybreaks[0]
\begin{align}
\frac{d^2F}{dQ^2}&=\frac{d\boldsymbol{\Rcal}}{dQ}\frac{\partial^2 F}{\partial\boldsymbol{\Rcal}\partial\boldsymbol{\Rcal}}\frac{d\boldsymbol{\Rcal}}{dQ}\\
&=\frac{d\boldsymbol{\Rcal}}{dQ}\boldsymbol{\Phi}\frac{d\boldsymbol{\Rcal}}{dQ}
+\frac{d\boldsymbol{\Rcal}}{dQ}
\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}
\frac{d\boldsymbol{\Rcal}}{dQ}\nonumber\\
&\mkern+156mu+\frac{d\boldsymbol{\Rcal}}{dQ}
\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\Theta}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}
\frac{d\boldsymbol{\Rcal}}{dQ}\,.
\label{eq:application_curvature}
\end{align}
\allowdisplaybreaks
This formula is evaluated at $Q=0$.
In order to be consistent with the finite differences result,
all the ingredients
have been calculated by using the SSCHA on a 2x2x2 supercell.
Once the SSCHA minimization at $Q=0$ has been completed and the converged
value for $\boldsymbol{\Phi}(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})$ has been obtained, $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})$ and
$\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})$ have been computed using Eq.~\eqref{Eq:stoc_avg_Vbb_impl}.
For each temperature, we used the converged value of $\boldsymbol{\Phi}(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})$ to generate the population used to
compute the averages. Therefore, in this case it is
$\tilde{\rho}{}_{\scriptscriptstyle{\text{in}}}(\boldsymbol{R})=\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}(\boldsymbol{R})$.
Notice that, as explained in Sec.~\ref{sec:Stochastic_implementation}, since the calculation has been performed in
a stationary point of the free energy,
the term $\Bavg{\text{f}_i}_{\tilde{\rho}_{\boldsymbol{\Rcal},\boldsymbol{\Phi}}}$ on the right-hand side of Eq.~\eqref{eq:def_fbb} is zero.
For explicative purposes we have used Eq.~\eqref{eq:compact_expr_curv} to express the curvature.
Thus we have three terms in Eq.~\eqref{eq:application_curvature}
and in Fig.~\ref{fig:d2FdQ2_finite_diff} we plot three lines to show their different contributions.
The term obtained from $\boldsymbol{\Phi}$ does not depend on the value of $p_3$, whereas the other two terms depend quadratically on $p_3$.
As a consequence, $[d\boldsymbol{\Rcal}/dQ\,\boldsymbol{\Phi}\,d\boldsymbol{\Rcal}/dQ]_{Q=0}$ gives the curvature
in the high-symmetry phase when the third order is absent.
This is confirmed, within the statistical error ($\simeq 2\,\text{meV}$),
by comparing the red curve and the red dots in Fig.~\ref{fig:d2FdQ2_finite_diff}.
For $p_3=6.70\,\text{eV}/\textup{\AA}^3$ the other two terms
$[d\boldsymbol{\Rcal}/dQ\,\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\,d\boldsymbol{\Rcal}/dQ]_{Q=0}$
and
$[d\boldsymbol{\Rcal}/dQ\,\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\Theta}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\,d\boldsymbol{\Rcal}/dQ]_{Q=0}$
are necessary in order to obtain the curvature.
This is confirmed by comparing the blue curve and the blue dots in Fig.~\ref{fig:d2FdQ2_finite_diff}.
As explained in Sec.~\ref{sec:Perturbative_limit}, only at the lowest perturbative order it is possible to
neglect the term $[d\boldsymbol{\Rcal}/dQ\,\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\boldsymbol{\Lambda}\boldsymbol{\Theta}\boldsymbol{\Lambda}\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}\,d\boldsymbol{\Rcal}/dQ]_{Q=0}$.
Indeed, in Fig.~\ref{fig:d2FdQ2_finite_diff}
we show with a yellow line the curvature computed with only the SCHA matrix and the bubble. In this case the difference with respect
to the correct value increases with temperature and, even if small, it is already beyond the statistical error at $250$ K.
In this section we have numerically proved the correctness of Eq.~\eqref{eq:compact_expr_curv}.
We conclude with a consideration. As already stressed in Sec.~\ref{Derivatives_of_F}, the first term
of Eq.~\eqref{eq:application_curvature} is always positive. Therefore, with $p_3=0$ it is possible to observe
only a first-order phase transition within the SCHA approximation.
With $p_3=6.70\,\text{eV}/\textup{\AA}^3$, the plot in Fig.~\ref{fig:d2FdQ2_finite_diff} shows that the
free energy curvature in $Q=0$ changes sign for $T\simeq140$ K. However, in Fig.~\ref{fig:F_vs_T}
we see that at $T=150\,K$ the free energy has already developed a lower minimum
in $|Q|\simeq0.9$. As a consequence, the toy model studied undergoes
a first order phase transition even with $p_3$ different from zero.
\subsubsection{Phonons}
In this section we apply the concept of free energy dynamical matrix defined in Sec.~\ref{sec:phonons_in_the_SCHA}.
To be precise, fixed the temperature, we compute the second derivative of the free energy in $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$,
divided by the square root of the masses. Notice that, properly speaking, this is $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$
only for $T>T_c$, when $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ is equal to $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$, because
at temperatures below the transition temperature $\boldsymbol{\Rcal}_{\scriptscriptstyle{\text{eq}}}$ departs from $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$.
Nevertheless, for explicative purposes and having this caveat in mind, we will use the same symbol even at $T<T_c$ .
The matrix $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ is given by the matrix $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$ plus the static self-energy $\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$ which, in turn,
is made of the bubble term $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$ plus other factors, negligible at the lowest perturbative level
(see Eqs.~\eqref{eq:diagrammatic_static_invSCHAdynmat_00},~\eqref{eq:diagrammatic_static_selfenergy},~\eqref{eq:diagrammatic_static_selfenergy_bubble},~\eqref{eq:SCHA_staticself_perturb}).
Since we are considering a crystal, we exploit the lattice translational symmetry and we write the dynamical matrices
in the unit cell as a function of the quasimomentum. In Fig.~\ref{fig:spectrum} we plot the spectrum of these matrices along
a high-symmetry path of the BZ. We consider two temperatures. The matrix $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$ coincides with the free energy
dynamical matrix $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ when the third order is absent. Since $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$ is positive-definite, the spectrum is always positive.
However, with $p_3=6.70\,\text{eV}/\textup{\AA}^3$ the dynamical matrix $\boldsymbol{D}^{\scriptscriptstyle{(F)}} $ is qualitatively different from $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$.
Below the transition temperature the phonon spectrum becomes imaginary (negative eigenvalue) in $\Gamma$, and only in that point.
The other instabilities that were present in the harmonic phonon spectrum, Fig.~\ref{fig:spec_harm}, have been washed
out by the zero-point energy and anharmonicity. Notice that in this case the comparison between the harmonic and the free energy dynamical
matrix is particularly meaningful because, for symmetry reasons, both are computed in the same point $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$.
In Fig.~\ref{fig:spectrum} we also show the spectrum obtained by adding only the bubble $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$
to $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$. From the results shown in Fig.~\ref{fig:d2FdQ2_finite_diff} we expected in $\Gamma$ a very small
difference between the full formula and the one
considering only the bubble. However, here we have a more complete picture. As we can see, in other points
of the BZ the spectrum is more affected by the presence of terms beyond the bubble.
For example, at $400$ K for the 5th mode in $L$, the terms beyond the bubble change the spectrum around $13\,\text{cm}^{-1}$.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{tab_phonons.pdf}
\caption{(color online) Spectrum along a BZ high-symmetry line of the
matrix $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$, independent of $p_3$, and of the free energy dynamical matrix $\boldsymbol{D}^{\scriptscriptstyle{(F)}}=\boldsymbol{D}^{\scriptscriptstyle{(S)}}+\boldsymbol{\Pi}^{\scriptscriptstyle{(S)}}(0)$
for $p_3=6.70\,\text{eV}/\textup{\AA}^3$, at two temperatures. When the third order is equal to zero, $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ is equal to $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$.
For $p_3=6.70\,\text{eV}/\textup{\AA}^3$ the system shows phonon softening, i.e. instability, in $\Gamma$.
The spectrum obtained by adding only the bubble $\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$ to $\boldsymbol{D}^{\scriptscriptstyle{(S)}}$ is also shown.
At $400$ K it is marked the considerable difference, around $13\,\text{cm}^{-1}$, between the energies of the
5th mode in $L$ obtained with $\boldsymbol{D}^{\scriptscriptstyle{(S)}}+\overset{\scriptscriptstyle{(B)}}{\boldsymbol{\Pi}}{}^{\scriptscriptstyle{(S)}}(0)$ and with $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$.}
\label{fig:spectrum}
\end{figure*}
\subsubsection{Convergence}
Since for our test we used a toy model, i.e. an analytic potential, we could evaluate the averages
using populations of very big size at small computational cost.
However, in view of first principle applications for realistic materials,
we carefully performed convergence tests of the curvature formula with respect to the population size $N_{\mathcal{I}}$.
First, we tested the convergence of $d^2F/dQ^2|_{Q=0}$ at various temperatures.
As said, for each temperature we calculated the curvature using the converged value of $\boldsymbol{\Phi}(\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}})$ to generate the population used
to compute the averages in Eq.~\eqref{Eq:stoc_avg_Vbb_impl}.
As shown in the upper left-hand panel of Fig.~\ref{fig:conv_test}, the convergence can be considered reached with $N_{\mathcal{I}}=10^{4}$.
However, it is worthwhile to say that, in general, fitting the values of the curvature versus temperature with
a polynomial allows to wash out part of the stochastic noise and obtain good estimations for $T_c$ with smaller populations.
In this case, for example, fitting with a $4$th degree polynomial the results obtained with $N_{\mathcal{I}}=10^3$
gives a value for $T_c$ which is only $9$ K smaller than the converged one.
As we have seen in Fig.~\ref{fig:d2FdQ2_finite_diff}, for $d^2F/dQ^2|_{Q=0}$ the terms beyond the bubble, which depend on $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$,
have a limited relevance. For that reason, we performed an analogous convergence test for the frequency of the $5$th mode in $L$ of $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$. Indeed,
as shown in Fig.~\ref{fig:spectrum}, for that specific mode the terms beyond the bubble play a non negligible role
in the determination of the spectrum. Therefore, this quantity is particularly significant
to analyze the convergence of the different terms comprising the curvature formula.
Here, as in the previous paragraph, with $\boldsymbol{D}^{\scriptscriptstyle{(F)}}$ we are indicating the curvature of the free energy in $\boldsymbol{\Rcal}_{\scriptscriptstyle{(0)}}$ divided by the
square root of masses, even at temperatures below $T_c$.
As shown in the upper right-hand panel of Fig.~\ref{fig:conv_test}, also in this case the convergence can be considered reached with $N_{\mathcal{I}}=10^{4}$.
However, the absolute stochastic error is already smaller than $3\,\text{cm}^{-1}$ with $N_{\mathcal{I}}=10^3$.
It is interesting to see how the two terms $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$
and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ affect the convergence, separately.
To that end, we plot in the the other panels of Fig.~\ref{fig:conv_test} the curvature and the frequency of the chosen mode, versus temperature,
obtained once with $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ computed with different population sizes $N_{\mathcal{I}}$
but with $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ fixed to the converged value (obtained with a population of $10^5$ elements),
and in the other case the inverse. The conclusion is that the total convergence is affected
in a similar way from the two tensors $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$.
This could be surprising since the 4th order tensor $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ is obtained by averaging a quantity that depends three times on the displacements, whereas
for the 3rd order tensors $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ it is averaged a less complicated quantity which depends only two times on the displacements.
However, it has to be considered that in the curvature formula $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$ is fully contracted (at variance with $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$).
Indeed, the random fluctuations on the single components tend to cancel each other and, thus,
the convergence of a contracted tensor is expected to be faster than the convergence of a single tensor component.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{converg.pdf}
\caption{(Color online) Convergence test.
Left-hand column: curvature in the high-symmetry phase, for the distortion considered,
as a function of the temperature. Right-hand column: frequency of the 5th mode in $L$ (also highlighted
in Fig.~\ref{fig:spectrum}) as a function of temperature. For this mode the effect of the terms beyond the bubble is not negligible.
Populations of different size are used to compute the tensors $\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$ and $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$
with Eq.~\eqref{Eq:stoc_avg_Vbb_impl}.
Upper row panels: different population sizes are used to compute both tensors.
With population of $10^4$ elements the result can be considered converged within the statistical error.
Second row (third row) bottom panels: different population sizes are used to compute only
$\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$($\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$) whereas $\overset{\scriptscriptstyle{(4)}}{\boldsymbol{\Phi}}$($\overset{\scriptscriptstyle{(3)}}{\boldsymbol{\Phi}}$)
is computed with $10^5$ elements. In the two cases the convergence trend is similar.}
\label{fig:conv_test}
\end{figure*}
\section{Conclusions}
\label{sec:conclusions}
In this work we present an approach to study structural second order phase transitions in
molecules and solids within the self-consistent harmonic approximation.
The developed method allows to estimate transition temperature and instability modes.
It is based on the analytic formula giving the second derivative of the SCHA free energy
with respect to the average atomic positions.
The Hessian of the SCHA free energy is also expressed in terms of thermal averages of forces and displacements.
Therefore, the method is suitable for a stochastic implementation in conjunction with any energy-force engine.
{Considering} a configuration, it allows to calculate directly the free energy curvature
{once} the SSCHA calculation has been performed in that point. As a consequence,
it permits to avoid the very computational demanding finite difference approach of computing
the curvature through several SSCHA calculations for different configurations{~\cite{Errea81}}.
Moreover, the imposition of symmetries on the result reduces the statistical noise and speeds up
statistical convergence with respect to the population size used to compute the averages
with the {importance} sampling technique.
The efficiency of this method makes it ideal to be used
in conjunction with first principle energy-force engines to study realistic materials,
{such as ferroelectrics or CDW materials}.
With the curvature formula it is possible to find the instabilities of a general condensed matter
system. In particular, the method is especially convenient for crystals, since
by exploiting the lattice translational symmetry and the
Fourier interpolation technique it is possible
to find distortions lowering the free energy with any modulation in space (e.g.
periodic on large supercells or even incommensurate)
with SSCHA {calculations} performed on supercells of moderate size.
In order to demonstrate our findings, numerical tests have been performed on a toy model.
The results confirm both correctness of the theory and numerical efficiency of the implemented method.
In addition to its practical utility, the developed theory sheds light on several fundamental
aspects of the SCHA. In particular, the role of the auxiliary effective quadratic {Hamiltonian} is clarified.
It is shown that the SCHA matrix is only a term of the free energy Hessian and, in general,
it does not {define} an anharmonic dynamical matrix.
On the contrary, an anharmonic temperature dependent, free energy based, dynamical matrix is obtained
through the free energy curvature. It generalizes the temperature independent harmonic dynamical matrix
and defines temperature dependent anharmonic phonons.
The theory developed for the SCHA free energy curvature is static, as it does not take into account
any dynamical effects. Inspired by a perspicuous diagrammatic interpretation of the results,
we propose a tentative minimal dynamic extension of the static theory in order to associate
{spectral functions with}
anharmonic phonons and interpret the results of scattering processes in a full non-perturbative way.
{Similarly, the dynamic theory allows to calculate phonon lifetimes in the non perturbative
limit}.
At variance with the curvature formula, the suggested dynamic extension is not based on a rigorous demonstration.
Nevertheless, it is expected to give good results, because it is correct in both static limit and lowest
perturbative limit, and thus it opens the way to further theoretical developments and interesting applications.
\section*{ACKNOWLEDGMENTS}
The authors acknowledge support from the Graphene Flagship.
I.E. acknowledges financial support from the
Spanish Ministry of Economy, Industry, and Competitiveness (Grant No. FIS2016-76617-P).
M.C. acknowledges support from Agence Nationale de la Recherche under contract ANR-13-IS10-0003-01, from the Graphene Flagship,
PRACE for awarding us access to resource on Marenostrum at BSC and the computer facilities provided
by CINES, IDRIS, and CEA TGCC (Grant EDARI No.~2017091202).
|
\section{}
\begin{abstract}
Image segmentation is the process of partitioning an image into a set of meaningful regions according to some criteria. Hierarchical segmentation has emerged as a major trend in this regard as it favors the emergence of important regions at different scales. On the other hand, many methods allow us to have prior information on the position of structures of interest in the images. In this paper, we present a versatile hierarchical segmentation method that takes into account any prior spatial information and outputs a hierarchical segmentation that emphasizes the contours or regions of interest while preserving the important structures in the image. Several applications are presented that illustrate the method versatility and efficiency.
\end{abstract}
\begin{keywords}
Mathematical Morphology, Hierarchies, Segmentation, Prior-based Segmentation, Stochastic Watershed.
\end{keywords}
\section{Introduction}
In this paper, we propose a method to take advantage of any prior spatial information previously obtained on an image to get a hierarchical segmentation of this image that emphasizes its regions of interest, allowing us to get more details in the designated regions of interest of an image while still preserving its strong structural information.
Potential applications are numerous. When having a limited storage capacity (for very large images for example), this would allow us to keep details in the regions of interest as a priority.
Similarly, in situations of transmission with limited bandwidth, one could first transmit the important information of the image: the details of the face for a video-call, the pitch and the players for a soccer game and so on. One could also use such a tool as a preprocessing one, for example to focus on an individual from one camera view to the next one in video surveillance tasks. Finally, from an artistic point of view, the result is interesting and similar to a combination of focus and cartoon effects. Some of these examples are illustrated in this paper.
Image segmentation has been shown to be inherently a multi-scale problem \cite{guigues2006scale}. That is why hierarchical segmentation has become the major trend in image segmentation and most top-performance segmentation techniques \cite{arbelaez2011contour}\cite{pont2015multiscale}\cite{Ren13}\cite{meyer15} fall into this category: hierarchical segmentation does not output a single partition of the image pixels into sets but instead a single multi-scale structure that aims at capturing relevant objects at all scales. Researches on this topic are still vivid as differential area profiles \cite{ouzounis12}, robust segmentation of high-dimensional data \cite{gueguen2013local} as well as theoretical aspects regarding the concept of partition lattice \cite{Serra12}\cite{Ronse2013} and optimal partition in a hierarchy \cite{kiran2013ground}\cite{Ravi2013global}\cite{xu2016hierarchical}.
Our goal in this paper is to develop a hierarchical segmentation algorithm that focuses on certain predetermined zones of the image. The hierarchical aspect also allows us, for tasks previously described, to very simply tune the level of details wanted depending on the application.
Furthermore, our algorithm is very versatile, as the spatial prior information that it uses can be obtained by any of the numerous learning-based approaches proposed over the last decades to roughly localize objects \cite{oquab15} \cite{Lampert08} \cite{Sermanet13}. In this regard, our work joins an important research point that consists in designing approaches to incorporate prior knowledge in the segmentation, as shape prior on level sets \cite{chan2005level}, star-shape prior by graph-cut \cite{veksler2008star}, use of a shape prior hierarchical characterization obtained with deep learning \cite{chen2013deep}, or related work making use of stochastic watershed to perform targeted image segmentation \cite{Malmberg17}.
The remainder of the paper is organized as follows. Part \ref{sec:hierarchies} explains how we construct and use graph-based hierarchical segmentation. Then part \ref{sec:ourmethod} specifies how we use prior information on the image to obtain hierarchies with regionalized fineness. Several examples of applications of this method are described in part \ref{sec:examples}. Finally, conclusions and perspectives are presented in part \ref{sec:conclusion}.
\section{Hierarchies and partitions}
\label{sec:hierarchies}
\vspace{-2.5mm}
\vspace{-2.5mm}
\subsection{Graph-based hierarchical segmentation}
Obtaining a suitable segmentation directly from an image is very difficult. This is why it is often make use of hierarchies to organize and propose interesting contours by valuating them. In this section, we remind the reader how to construct and use graph-based hierarchical segmentation.
For each image, let us suppose that a fine partition is produced by an initial segmentation (for instance a set of superpixels \cite{Achanta12,machairas2015waterpixels} or the basins produced by a classical watershed algorithm \cite{meyer1990morphological}) and contains all contours making sense in the image. We define a dissimilarity measure between adjacent tiles of this fine partition.
One can then see the image as a graph, the \textit{region adjacency graph} (RAG), in which each node represents a tile of the partition; an edge links two nodes if the corresponding regions are neighbors in the image; the weight of the edge is equal to the dissimilarity between these regions. Working on the RAG is much more efficient than working on the image, as there are far less nodes in the RAG than there are pixels in the image.
Formally, we denote this graph $\mathcal{G}=(\mathcal{V},\mathcal{E},\mathbf{W})$, where $\mathcal{V}$ corresponds to the image domain or set of pixels/fine regions, $\mathcal{E} \subset \mathcal{V} \times \mathcal{V}$ is the set of edges linking neighbour regions,
$\mathbf{W}: \mathcal{E} \to \mathbb{R}^{+}$ is the dissimilarity measure usually based on local gradient information (or color or texture), for instance $\mathbf{W}(i,j) \propto |\mathbf{I}(v_i)-\mathbf{I}(v_j)|$ with $\mathbf{I}:\mathcal{V} \to \mathbb{R}$ representing the image intensity.
The edge linking the nodes $p$ and $q$ is designated by $e_{pq}$ . A path is a sequence of nodes and edges: for example the path linking the nodes $p$ and $s$ is the set $\{p, e_{pt} , t, e_{ts}, s\}$. A \textit{connected subgraph} is a subgraph where each pair of nodes is connected by a path. A \textit{cycle} is a path whose extremities coincide. A \textit{tree} is a connected graph without cycle. A \textit{spanning tree} is a tree containing all nodes. A \textit{minimum spanning tree} (MST) $\mathcal{MST}(\mathcal{G})$ of a graph $\mathcal{G}$ is a spanning tree with minimal possible weight, obtained for example using the Boruvka algorithm (the weight of a tree being equal to the sum of the weights of its edges). A \textit{forest} is a collection of trees.
A \emph{partition} $\mathsf{\pi}$ of a set $\mathcal{V}$ is a collection of subsets of $\mathcal{V}$, such that the whole set $\mathcal{V}$ is the disjoint union of the subsets in the partition, i.e., $\mathsf{\pi}=\{\mathsf{R}_1,\mathsf{R}_2,\ldots,\mathsf{R}_k\}$, such that $\forall i, \mathsf{R}_i \subseteq \mathcal{V} \ $; $ \forall i\neq j, \mathsf{R}_i \cap\mathsf{R}_j = \emptyset$ ; $\bigcup_{i}^{k}\mathsf{R}_i = \mathcal{V}$.
Cutting all edges of the $\mathcal{MST}(\mathcal{G})$ having a valuation superior to a threshold $\lambda$ leads to a minimum spanning forest (MSF) $\mathcal{F}(\mathcal{G})$, i.e. to a partition of the graph. Note that the obtained partition is the same that one would have obtain by cutting edges superior to $\lambda$ directly on $\mathcal{G}$ \cite{najman13}. Since working on the $\mathcal{MST}(\mathcal{G})$ is less costly and provides similar results regarding graph-based segmentation, we work only with the $\mathcal{MST}(\mathcal{G})$ in the sequel.
\vspace{-2.5mm}
\vspace{-2.5mm}
\begin{figure}
\begin{center}
\includegraphics[width=.55 \columnwidth]{./images/Markers-based.png}
\vspace{-2.5mm}
\caption{\textbf{A}: a partition represented by an edge-weighed graph; \textbf{B}: a minimum spanning tree of the graph, with 2 markers in blue: the highlighted edge in blue is the highest edge on the path linking the two markers ; \textbf{C}: the segmentation obtained when cutting this edge; \textbf{D} blue and orange domain are the domains of variation of the two markers generating the same segmentation.}\label{Fig:MST}
\end{center}
\vspace{-2.5mm}
\end{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
So cutting edges by decreasing valuations gives an \emph{indexed hierarchy of partitions} $(\mathcal{H},\bm{\lambda})$, with $\mathcal{H}$ a \emph{hierarchy of partitions} i.e. a chain of nested partitions $\mathcal{H}=\{\mathsf{\pi}_0, \mathsf{\pi}_1,\ldots, \mathsf{\pi}_n| \forall j,k, \quad 0 \quad \leq j\leq k\leq n \Rightarrow \mathsf{\pi}_j \sqsubseteq \mathsf{\pi}_k\}$, with $\mathsf{\pi}_n$ the single-region partition and $\mathsf{\pi}_0$ the finest partition on the image, and $\bm{\lambda}: \mathcal{H} \to \mathbb{R}^+$ being a stratification index verifying $\bm{\lambda}(\mathsf{\pi}) < \bm{\lambda}(\mathsf{\pi}')$ for two nested partitions $\mathsf{\pi} \subset \mathsf{\pi}'$. This increasing map allows us to value each contour according to the level of the hierarchy for which it disappears: this is the \emph{saliency} of the contour, and we consider that the higher the saliency, the stronger the contour.
For a given hierarchy, the image in which each contour takes as value its saliency is called \textit{Ultrametric Contour Map} (UCM)\cite{arbelaez2011contour}. Representing a hierarchy by its UCM is an easy way to get an idea of its effect because thresholding an UCM always provides a set of closed curves and so a partition. In this paper, for better visibility, we represent UCM with inverted contrast.
To get a partition for a given hierarchy, there are several possibilities:
\vspace{-1mm}
\begin{itemize}
\item simply thresholding the highest saliency values,
\item marking some nodes as important ones and then computing a partition accordingly, which is known as \textit{marker-based segmentation},
\item smartly editing the graph by finding the partition that minimizes an energetic function.
\end{itemize}
\vspace{-1mm}
In a complementary approach, we argue that the quality of the obtained partitions highly depends on the hierarchy that we use, and thus that changing the dissimilarity can lead to more suitable partitions. Indeed, if the dissimilarity reflects only a local contrast as in the hierarchy issued by the RAG, the most salient regions in the image are the small contrasted ones. So instead of departing from a simple and rough dissimilarity such as contrast and then use an sophisticated technique to get a good partition out of it, one can also try to obtain a more informative dissimilarity adapted to the content of the image such that the simplest methods are sufficient to compute interesting partitions. This way, the aforementioned techniques lead to segmentations better suited for further exploitation.
How can we construct more pertinent and informative dissimilarities?
\vspace{-2.5mm}
\vspace{-2.5mm}
\subsection{Stochastic watershed hierarchies}
\label{ssec:stochastic}
The stochastic watershed (SWS), introduced in \cite{angulo2007stochastic} on a simulation basis and extended with a graph-based approach in \cite{meyer15}, is a versatile tool to construct hierarchies. The seminal idea is to operate multiple times marker-based segmentation with random markers and valuate each edge of the $\mathcal{MST}$ by its frequency of appearance in the resulting segmentations.
Indeed, by spreading markers on the RAG $\mathcal{G}$, one can construct a segmentation as a MSF $\mathcal{F}(\mathcal{G})$ in which each tree takes root in a marked node. Marker-based segmentation directly on the $\mathcal{MST}$ is possible: one must then cut, for each pair of markers, the highest edge on the path linking them. Furthermore, there is a domain of variation in which each marker can move while still leading to the same final segmentation. More details are provided in Figure \ref{Fig:MST}.
Let us then consider on the $\mathcal{MST}$ an edge $e_{st}$ of weight $\omega_{st}$ and compute its probability to be cut. We cut all edges of the $\mathcal{MST}$ with a weight superior or equal to $\omega_{st}$, producing two trees $\tree_{s}$ and $\tree_{t}$ of roots $s$ and $t$. If at least one marker falls within the domain $\mathsf{R}_{s}$ of $\tree_{s}$ nodes and at least one marker falls within the domain $\mathsf{R}_{t}$ of $\tree_{t}$ nodes, then $e_{st}$ will be cut in the final segmentation.
Let denote $\mu(\mathsf{R})$ the number of random markers falling in a region $\mathsf{R}$. We want to attribute to $e_{st}$ the following probability value:
\begin{equation} \label{Proba}
\begin{split}
\mathbb{P}[(\mu(\mathsf{R}_{s}) \geq 1) \land (\mu(\mathsf{R}_{t}) \geq 1)] & = 1-\mathbb{P}[(\mu(\mathsf{R}_{s}) = 0) \lor (\mu(\mathsf{R}_{t}) = 0)] \\
& = 1-\mathbb{P}(\mu(\mathsf{R}_{s}) = 0)-\mathbb{P}(\mu(\mathsf{R}_{t}) = 0)+\mathbb{P}(\mu(\mathsf{R}_{s} \cup \mathsf{R}_{t}) = 0) \\
\end{split}
\end{equation}
If markers are spread following a Poisson distribution, then for a region $R$:
\begin{equation} \label{Poisson}
\mathbb{P}(\mu(R) = 0)=\exp^{-\Lambda(R)},
\end{equation}
With $\Lambda(R)$ being the expected value (mean value) of the number of markers falling in $R$. The probability thus becomes:
\begin{equation} \label{newProba}
\mathbb{P}(\mu(\mathsf{R}_{s}) \geq 1 \land \mu(\mathsf{R}_{t}) \geq 1) = 1-\exp^{-\Lambda(\mathsf{R}_{s})}-\exp^{-\Lambda(\mathsf{R}_{t})}+\exp^{-\Lambda(\mathsf{R}_{s} \cup \mathsf{R}_{t})}
\end{equation}
When the Poisson distribution has an homogeneous density $\lambda$:
\begin{equation} \label{uniform}
\Lambda(R) = \texttt{area}(R) \lambda,
\end{equation}
When the Poisson distribution has a non-uniform density $\lambda$:
\begin{equation} \label{nonuniform}
\Lambda(R)=\int_{(x,y) \in R} \lambda(x,y) \, \mathrm dx \mathrm dy
\end{equation}
The output of the SWS algorithm thus depends on the departure $\mathcal{MST}$ (structure and edges valuations) and of the probabilistic law governing the markers distribution. Furthermore, SWS hierarchies can be chained, leading to a wide exploratory space that can be used in a segmentation workflow \cite{fehri16}.
Because of its versatility and good performance, SWS represents a good departure algorithm to modify in order to inject prior information. Indeed, when having a prior information about the image, is it possible to use it in order to have more details in some parts rather than others?
\section{Hierarchies highlighthing structures of interest using prior information}
\label{sec:ourmethod}
\subsection{Hierarchy with Regionalized Fineness (HRF)}
\label{ssec:HRF}
In the original SWS, a uniform distribution of markers is used (whatever size or form they may have). In order to have stronger contours in a specific region of the image, we adapt the model so that more markers are spread in this region.
Let $E$ be an object or class of interest, for example $E = {\text{``face of a person"}}$, and $\mathbf{I}$ be the studied image. We denote by $\theta_{E}$ the probability density function (PDF) associated with $E$ obtained separately, and defined on the domain $D$ of $\mathbf{I}$, and by $\mathrm{PM}(\mathbf{I},\theta_{E})$ the probabilistic map associated, in which each pixel $p(x,y)$ of $\mathbf{I}$ takes as value $\theta_{E}(x,y)$ its probability to be part of $E$. Given such an information on the position of an event in an image, we obtain a hierarchical segmentation focused on this region by modulating the distribution of markers.
If $\lambda$ is a density defined on $D$ to distribute markers (uniform or not), we set $\theta_{E} \lambda$ as a new density, thus favoring the emergence of contours within the regions of interest.
Considering a region $\mathsf{R}$ of the image, the mean number of markers falling within $\mathsf{R}$ is then:
\begin{equation} \label{newnonuniform}
\Lambda_{E}(\mathsf{R})=\int_{(x,y) \in \mathsf{R}} \theta_{E}(x,y) \lambda(x,y) \, \mathrm dx \mathrm dy
\end{equation}
Note that if we want $N$ markers to fall in average within the domain $D$, we work with a slightly modified density:
\begin{equation} \label{eqNbMarkers}
\hat{\lambda}=\frac{N}{\mu(D)} \lambda
\end{equation}
Furthermore, this approach can be easily extended to the case where we want to take advantage of information from multiple sources. Indeed, if $\theta_{E_1}$ and $\theta_{E_2}$ are the PDF associated with two events $E_1$ and $E_2$, we can combine those two sources by using as a new density $(\theta_{E_1}+\theta_{E_2})\lambda$.
\subsection{Methodology}
We present here the steps to compute a HRF for an event $E$ given a probabilistic map $\mathrm{PM}(\mathbf{I},\theta_{E})$ providing spatial prior information on an image $\mathbf{I}$:
\begin{itemize}
\item compute a fine partition $\pi_{0}$ of the image, define a dissimilarity measure between adjacent regions and compute the RAG $\mathcal{G}$, and then the $\mathcal{MST}(\mathcal{G})$ to easily work with graphs,
\item compute a probabilistic map $\pi_{\mu}=\pi_{\mu}(\pi_{0},\mathrm{PM}(\mathbf{I},\theta_{E}))$ with each region of the fine partition $\pi_{0}$ taking as a new value the mean value of $\mathrm{PM}(\mathbf{I},\theta_{E}))$ in this region,
\item compute new values of edges by a bottom-up approach as described in section \ref{ssec:HRF}, where for each region $\mathsf{R}_{i}$ of $\pi_{0}$, $\Lambda(\mathsf{R}_{i})$ corresponds to the mean value taken by pixels of the region $\mathsf{R}_{i}$ in $\pi_{\mu}$. Note that this approach allows a highly efficient implementation using dynamic programming on graphs.
\end{itemize}
\subsection{Modulating the HRF depending on the couple of regions considered}
\label{ssec:varHRF}
If we want to favor certain contours to the detriment of others, we can modulate the density of markers in each region by taking into account the strength of the contour separating them but also the relative position of both regions.
We use the same example and notations as in section \ref{ssec:HRF}, and thus want to modulate the distribution of markers relatively to $\mathsf{R}_{s}$, $\mathsf{R}_{t}$ and their frontier.
For example, to stress the strength of the gradient separating both regions we can locally spread markers following the distribution $\chi(\mathsf{R}_{s},\mathsf{R}_{t}) \lambda$, with $\chi(\mathsf{R}_{s},\mathsf{R}_{t})=\omega_{st}$. This corresponds to the classical volume-based SWS, which allows to obtain a hierarchy that takes into account both surfaces of regions and contrast between them.
To go further, one can use any prior information in a similar way. Indeed, while using prior information to influence the output of the segmentation workflow, one might also want to choose whether the relevant information to emphasize in resulting segmentations is the foreground, the background or the transitions between them.
For example, having more details in the transition regions between background and foreground allows us to have more precision where the limit between foreground and background is actually unclear. As a matter of fact, the prior information often only provides rough positions of the foreground object with blurry contours, and such a process would allow to get precise contours of this object from the image.
Let us consider this case and define for each couple of regions $(\mathsf{R}_{s},\mathsf{R}_{t})$ a suitable $\chi(\mathsf{R}_{s},\mathsf{R}_{t})$. We then want $\chi(\mathsf{R}_{s},\mathsf{R}_{t})$ to be low if $\mathsf{R}_{s}$ and $\mathsf{R}_{t}$ both are in the background or the foreground, and high if $\mathsf{R}_{s}$ is the background and $\mathsf{R}_{t}$ in the foreground (or the opposite). We use:
\begin{equation}\label{eqNbMarkers2}
\left\{
\begin{array}{ll}
\tilde{\lambda}= \chi \lambda \\
\chi(\mathsf{R}_{s},\mathsf{R}_{t}) = \frac{\max(m(\mathsf{R}_{s}),m(\mathsf{R}_{t}))(1-\min(m(\mathsf{R}_{s}),m(\mathsf{R}_{t})))}{0.01+\sigma(\mathsf{R}_{s})\sigma(\mathsf{R}_{t})},
\end{array}
\right.
\end{equation}
$m(\mathsf{R})$ (resp. $\sigma(\mathsf{R})$) being the normalized mean (resp. normalized standard deviation) of pixels values in the region $\mathsf{R}$ of $\mathrm{PM}(\mathbf{I})$.
Thus the number of markers spread will be higher when the contrast between adjacent regions is high (numerator term) and when these regions are coherent (denominator term).
Then for each edge, its new probability to be cut is :
\begin{align} \label{newProba2}
\begin{split}
\mathbb{P}[(\mu(\mathsf{R}_{p}) \geq 1) \land (\mu(\mathsf{R}_{q}) \geq 1)] &= 1-\exp^{-\chi(\mathsf{R}_{s},\mathsf{R}_{t})\Lambda(\mathsf{R}_{p})}-\exp^{-\chi(\mathsf{R}_{s},\mathsf{R}_{t}) \Lambda(\mathsf{R}_{q})} \\
& + \exp^{-\chi(\mathsf{R}_{s},\mathsf{R}_{t}) \Lambda(\mathsf{R}_{p} \cup \mathsf{R}_{q})}
\end{split}
\end{align}
In the spirit of \cite{Chen16}, this mechanism provides us with a way to ``realign" the hierarchy with respect to the relevant prior information to get more details where the information is blurry. Similar adaptations can be thought of to emphasize details of background or foreground regions.
In the following, we illustrate the methodology exposed with some applications.
\begin{figure}[H]
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\begin{center}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imIn}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imPriorDist}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imSinvComp}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imSinv}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_100}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_1000}}
\vspace{-2.5mm}
\vspace{-2.5mm}
\caption{Hierarchical segmentation of faces. (a) Original image (b) Prior : Probabilistic map obtained thanks to face detection algorithm and (c) Volume-based SWS Hierarchy UCM (d) Volume-based HRF UCM with face position as prior (e)(f)(g) Examples of segmentations obtained with HRF - 10,100,1000 regions. }
\label{fig:FaceHRF}
\end{center}
\end{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\section{Application examples}
\label{sec:examples}
\subsection{Scalable transmission favoring regions of interest}
Let us consider a situation where one emitter wants to transmit an image through a channel with a limited bandwidth, e.g. for a videoconference call. In such a case, the more important informations to transmit are details on the face of the person on the image. Besides, we nowadays have highly efficient face detectors, using for example Haar-wavelets as features in a learning-based vision approach \cite{viola01}. Considering that for an image in entry, the face can be easily detected, we can use this information to produce a hierarchical segmentation of the image that accentuates the details around the face while giving a good sketch of the image elsewhere. Depending on the bandwidth available, we can then choose the level of the hierarchy to select and obtain the associated partition to transmit, ensuring us to convey the face with as much details as possible. Some results are presented in Figure \ref{fig:FaceHRF}, with notably a comparison between a classical volume-based SWS UCM and a volume-based HRF UCM.
\begin{figure}[H]
\vspace{-2.5mm}
\begin{center}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Flou2/blur2.jpg}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Flou2/SVFDetector2}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Flou2/imSinvComp}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Flou2/imSinv}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Flou2/Seg_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Flou2/Seg_200}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Flou2/Seg_2200}}
\vspace{-2.5mm}
\vspace{-2.5mm}
\caption{Hierarchical segmentation of non-blur objects. (a) Original image (b) Prior image obtained with non-blur zones detection algorithm (c) Volume-based SWS Hierarchy UCM (d) Volume-based HRF UCM with non-blur zones as prior (e)(f)(g) Examples of segmentations obtained with HRF - 10,200,2000 regions.}\label{Fig:blur}
\end{center}
\end{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\subsection{Artistic aspect: focus and cartoon effect}
The same method can also be used for artistic purposes. For example, when taking as prior the result of a blur detector \cite{su2011}, we can accentuate the focus effect wanted by the photograph and turn it into a cartoon effect as well - see Figure \ref{Fig:blur} for an illustration of the results.
\begin{figure}[H]
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\begin{center}
\subfigure[]{\includegraphics[width=.225\columnwidth]{./images/CNN/Bike/imIn.jpg}}
\subfigure[]{\includegraphics[width=.225\columnwidth]{./images/CNN/Bike/{imPrior_2.0}.png}}
\subfigure[]{\includegraphics[width=.225\columnwidth]{./images/CNN/Bike/imSinvComp}}
\subfigure[]{\includegraphics[width=.225\columnwidth]{./images/CNN/Bike/imSinv}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/CNN/Bike/Example_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/CNN/Bike/Example_100}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/CNN/Bike/Example_200}}
\vspace{-2.5mm}
\vspace{-2.5mm}
\caption{Hierarchical segmentation of the main class (class ''bike") in an image with heatmap issued of a CNN-based method as input. (a) Original image, (b) Heatmap issued by the CNN-based localization method, (c) Volume-based Watershed Hierarchy UCM and (d) Hierarchy with prior UCM. (e)(f)(g) Examples of segmentations obtained with HRF - 10,100,200 regions.}\label{Fig:CNN}
\end{center}
\end{figure}
\begin{figure}[H]
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\begin{center}
\subfigure[]{\includegraphics[width=.18\columnwidth]{./images/CoSeg/Heli1.jpg}}
\subfigure[]{\includegraphics[width=.18\columnwidth]{./images/CoSeg/Heli2.jpg}}
\subfigure[]{\includegraphics[width=.12\columnwidth]{./images/CoSeg/Heli3.jpg}}
\subfigure[]{\includegraphics[width=.12\columnwidth]{./images/CoSeg/Heli4.jpg}}
\subfigure[]{\includegraphics[width=.18\columnwidth]{./images/CoSeg/Heli5.jpg}}\\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli2/imPrior}}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli2/imPriorDist}}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli2/imSinvComp}}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli2/imSinv}}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli3/imSinvComp}}
\subfigure[]{\includegraphics[width=.15\columnwidth]{./images/CoSeg/Heli3/imSinv}}
\vspace{-2.5mm}
\vspace{-2.5mm}
\caption{Hierarchical co-segmentation of matched objects. (a)-(e) Images $I_{1}$ to $I_{5}$ (f) All matched key-points of $I_{3}$ with other images key-points (g) Prior for $I_{3}$: probability map generated thanks to morphological distance function (h)(i) Volume-based SWS Hierarchy UCM for $I_{3}$ and $I_{4}$ (j)(k) Volume-based HRF UCM for $I_{3}$ and $I_{4}$ with matched key-points as prior.}\label{Fig:Coseg}
\end{center}
\end{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
In the same spirit, various methods now exist to automatically roughly localize the principal object in an image. We inspire ourselves from \cite{oquab15} to do so. Using the state-of-the-art convolution neural network (CNN) classifier VGG19 \cite{simonyan14} trained on the 1000 classes ImageNet database \cite{deng09}, we first determine what is the main class in the image. Note that this CNN takes as input only images of size $224\times224$ pixels. Once it is known, we can then, by rescaling the image by a factor $s \in \{0.5,0.7,1.0,1.4,2.0,2.8\}$, compute for sub-windows of size $224\times224$ of the image the probability of appearance of the main class. By simply superimposing the results for all sub-windows, we thus obtain a probabilistic map of the main class for each rescaling factor. By max-pooling, we keep in memory the result of the scale for which the probability is the highest. The \textit{heatmap} thus produced can then be used to feed our algorithm. This way, we have at our disposal an automatized way to focus on the principal class in the scene. Some results are presented in Figure \ref{Fig:CNN}.
\begin{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\begin{center}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imIn}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imSInv_volumicAdaptative}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imSInv_volumicHRF}}
\subfigure[]{\includegraphics[width=.242\columnwidth]{./images/Face/Second/imSInv_volumicAdaptativeHRF}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_Adap_4}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_Adap_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_Adap_25}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRF_4}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRF_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRF_25}} \\
\vspace{-2.5mm}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRFAdap_4}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRFAdap_10}}
\subfigure[]{\includegraphics[width=.325\columnwidth]{./images/Face/Second/Example_HRFAdap_25}}
\vspace{-2.5mm}
\vspace{-2.5mm}
\caption{Hierarchical segmentation of faces. (a) Original image (b)(c)(d) UCM of HRF depending of the couples of regions (section \ref{ssec:varHRF}), of the regions only (section \ref{ssec:HRF}), and both combined. The rest of the images are segmentations examples with 4,10,25 regions, the hierarchies being presented in the same order.}
\label{fig:FaceHRFbis}
\end{center}
\end{figure}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\vspace{-2.5mm}
\subsection{Hierarchical co-segmentation}
Another potential application is to co-segment with the same fineness level an object appearing in several different images. For example, when given a list of images of the same object taken from different perspectives/for different conditions, we can follow the state-of-the-art matching procedure \cite{lowe04}: (i) compute all key-points in both images, (ii) compute local descriptors at these key-points, (iii) match those key-points using a spatial coherency algorithm as RANSAC. Once it is done we retain these matched key-points for both images, and generate probability maps of the appearance of the matched objects using a morphological distance function to the matched key-points.
These probability maps can then feed our algorithm, given as result a hierarchical co-segmentation that emphasizes the matched zones of the image. Some results are presented in Figure \ref{Fig:Coseg}.
\subsection{Example of the effect of the HRF highlighting transitions between foreground and background}
We illustrate here the HRF highlighting transitions between foreground and background presented in section \ref{ssec:varHRF}, by presenting its effect in the face detection example presented in figure \ref{fig:FaceHRFbis}.
\section{Conclusions and perspectives}
\label{sec:conclusion}
In this paper we have proposed a novel and efficient hierarchical segmentation algorithm that emphasizes the regions of interest in the image by using spatial exogenous information on it. The wide variety of sources for this exogenous information makes our method extremely versatile and its potential applications numerous, as shown by the examples developed in the last section. To go further, we could find a way to efficiently extend this work to videos. One could also imagine a semantic segmentation method that would go back and forth between localization algorithm and HRF to progressively refine the contours of the main objects in the image.
\FloatBarrier
\bibliographystyle{splncs03}
|
\section{Introduction}
An interesting extension of the standard model (SM) proposed by Manohar and Wise \cite{Manohar:2006ga} consists of adding a color-octet electroweak doublet scalar. The original motivation for this extension is that it consists of one of two scalar representations allowed by minimal flavor violation for scalars that do not transform under the flavor group. From this perspective the most general renormalizable potential consistent with this field content is constructed resulting in generic studies for TeV scale color octet scalars. Scalars such as these occur in specific models, for example in unification with $SU(5)$ \cite{Perez:2016qbo, Dorsner:2007fy} or with $SO(10)$ \cite{Bertolini:2013vta}.
The model has interesting phenomenology and many studies have been performed in the literature \cite{Gresham:2007ri,Gerbush:2007fe,Burgess:2009wm,He:2011ti,Dobrescu:2011aa,Bai:2011aa,Arnold:2011ra,Kribs:2012kz,Reece:2012gi,Cao:2013wqa,He:2013tla,Cheng:2015lsa,Martinez:2016fyd}. A salient feature of the MW model being that it is very difficult to observe or exclude the new scalars at the LHC. The model has a very large parameter space, but these studies have shown that it is possible to constrain it significantly by imposing custodial symmetry, partial wave unitarity and vacuum stability.
Along these lines, we have recently proposed considering the MW extension in the context of a two Higgs doublet model (2HDM). In Ref.~\cite{Cheng:2016tlc} we introduced the model and constrained its parameter space by imposing tree level theoretical constraints arising from symmetries and from perturbative unitarity. In addition we briefly discussed LHC phenomenology, concluding that the largest effects from this addition to the 2HDM would appear in corrections to one-loop couplings of the higgs boson to two gluons or photons.
Previous studies of the MW model \cite{He:2013tla} and of the 2HDM \cite{Chakrabarty:2014aya,Ferreira:2015rha} indicate that the viable parameter space is further constrained when one includes renormalization group corrections, and in this paper we apply this rationale to the MW extension of the 2HDM. We first compute the beta functions for the couplings of the model. We then consider the requirements that there be no Landau poles (LP) below a certain high mass scale $\Lambda$, that the scalar potential remains stable and that two-to-two scattering amplitudes remain perturbative at all scales below $\Lambda$. These conditions have a long history of being used as theoretical constraints from their early application to the SM Higgs boson mass \cite{Callaway:1988ya,Sher:1988mj}.
\section{The model}
The construction of the model was described in Ref.~\cite{Cheng:2016tlc}. The scalar sector of the SM is replaced by three $SU(2)$ scalar doublets: two color singlets ($\Phi_1, \Phi_2)$ and one color-octet $S$. The most general renormalizable potential for the 2HDM sector ($\Phi_1, \Phi_2)$ is well known from the literature \cite{Gunion:1989we,Branco:2011iw} and we restrict ourselves to the case of a CP conserving potential with a discrete symmetry $\Phi_1\to -\Phi_1$ that is only violated softly (although in the end we replace this with the requirement of MFV). To this known potential we add the couplings between the color octet $S$ and the two color singlets ($\Phi_1, \Phi_2)$ as well as the color octet self interactions \cite{Manohar:2006ga} resulting in
\begin{eqnarray}
V\left( \Phi_1, \Phi_2 \right) &=& m_{11}^2 \Phi_1^\dag \Phi_1 + m_{22}^2 \Phi_2^\dag \Phi_2 - m_{12}^2 \left( \Phi_1^\dag \Phi_2 + \Phi_2^\dag \Phi_1 \right) \nonumber \\
&+& \frac{\lambda_1}{2} \left( \Phi_1^\dag \Phi_1 \right)^2 + \frac{\lambda_2}{2} \left( \Phi_2^\dag \Phi_2 \right)^2
+ \lambda_3 \left( \Phi_1^\dag \Phi_1 \right) \left( \Phi_2^\dag \Phi_2 \right) \nonumber\\
&+& \lambda_4 \left( \Phi_1^\dag \Phi_2 \right) \left( \Phi_2^\dag \Phi_1 \right) + \frac{\lambda_5}{2} \left[ \left( \Phi_1^\dag \Phi_2 \right)^2 + \left( \Phi_2^\dag \Phi_1 \right)^2 \right] \nonumber\\
&+& 2m_S^2 {\rm Tr}S^{\dag i}S_i + \mu_1 {\rm Tr}S^{\dag i}S_i S^{\dag j}S_j + \mu_2 {\rm Tr}
S^{\dag i}S_j S^{\dag j}S_i + \mu_3 {\rm Tr} S^{\dag i}S_i {\rm Tr}S^{\dag j} S_j\nonumber\\
& +& \mu_4 {\rm Tr}S^{\dag i}S_j {\rm Tr}S^{\dag j}S_i + \mu_5 {\rm Tr}S_i S_j{\rm Tr}
S^{\dag i}S^{\dag j} + \mu_6 {\rm Tr}S_i S_j S^{\dag j}S^{\dag i}\nonumber \\
&+& \nu_1 \Phi_1^{\dag i}\Phi_{1i}{\rm Tr}S^{\dag j}S_j + \nu_2 \Phi_1^{\dag i}\Phi_{1j}
{\rm Tr}S^{\dag j}S_i\nonumber\\
& +& \left( \nu_3 \Phi_1^{\dag i}\Phi_1^{\dag j}{\rm Tr}S_i S_j + \nu_4 \Phi_1^{\dag i}{\rm Tr}
S^{\dag j}S_j S_i + \nu_5 \Phi_1^{\dag i}{\rm Tr}S^{\dag j}S_i S_j + {\rm h.c.} \right) \nonumber \\
&+&\omega_1 \Phi_2^{\dag i}\Phi_{2i}{\rm Tr}S^{\dag j}S_j + \omega_2 \Phi_2^{\dag i}\Phi_{2j}
{\rm Tr}S^{\dag j}S_i\nonumber\\
& +& \left( \omega_3 \Phi_2^{\dag i}\Phi_2^{\dag j}{\rm Tr}S_i S_j + \omega_4 \Phi_2^{\dag i}{\rm Tr}
S^{\dag j}S_j S_i + \omega_5 \Phi_2^{\dag i}{\rm Tr}S^{\dag j}S_i S_j + {\rm h.c.} \right) \nonumber\\
&+& \kappa_1 \Phi_1^{\dag i}\Phi_{2i}{\rm Tr}S^{\dag j}S_j +
\kappa_2 \Phi_1^{\dag i}\Phi_{2j}{\rm Tr}S^{\dag j}S_i + \kappa_3 \Phi_1^{\dag i}\Phi_2^{\dag j}{\rm Tr}S_j S_i
+ {\rm h.c.}
\label{potential}
\end{eqnarray}
In all the terms we have explicitly shown the $SU(2)$ indices $i, j$, $S_i = T^A S_i^A$, and the trace is taken over color indices.
As per our discussion in Ref.~\cite{Cheng:2016tlc}, in terms that are not part of the usual the 2HDM, we have allowed some that satisfy MFV but not the discrete symmetry mentioned above.
After symmetry breaking some of these couplings are related to scalar masses, and these relations can be readily found in the literature \cite{Manohar:2006ga,Branco:2011iw,Cheng:2016tlc}.
The Yukawa potential for this model in the flavour eigenstate basis is given by
\begin{eqnarray}
L_{Y}&=& - {\left( g_1^D \right)^\alpha}_\beta {\bar D}_{R, \alpha }\Phi_1^\dag Q_L^\beta -
{\left( g_1^U \right)^\alpha}_\beta {\bar U}_{R, \alpha}{\tilde \Phi}_1^\dag Q_L^\beta \nonumber\\
&-& {\left( g_2^D \right)^\alpha}_\beta {\bar D}_{R, \alpha } \Phi_2^\dag Q_L^\beta - {\left( g_2^U
\right)^\alpha}_\beta {\bar U}_{R, \alpha}{\tilde \Phi}_2^\dag Q_L^\beta + {\rm h.c.},\nonumber \\
&-& {\left( g_3^D \right)^\alpha}_\beta {\bar D}_{R, \alpha}S^\dag Q_L^\beta -
{\left( g_3^U \right)^\alpha}_\beta {\bar U}_{R, \alpha }{\tilde S}^\dag Q_L^\beta + {\rm h.c.}
\label{yukawas}
\end{eqnarray}
As is conventional, we use ${\tilde H}_i = \varepsilon_{ij} H_j^*$
for all three scalar doublets $H=\Phi_{1,2},S$, and $\alpha, \beta$ are flavour indices.
The large number of parameters present in Eq.~\ref{potential} and Eq.~\ref{yukawas} is reduced by the following theoretical considerations:
\begin{itemize}
\item Minimal flavour violation \cite{Chivukula:1987py,D'Ambrosio:2002ex} implies
\begin{itemize}
\item 2HDM Type I: $\eta_1^D=\eta_1^U =0$
\item 2HDM Type II: $\eta_1^U=\eta_2^D=0$
\end{itemize}
Requiring MFV instead of a discrete symmetry to define the models allows quartic terms in the scalar potential that are odd in either of the doublets, such as $\nu_{4,5}$, $\omega_{4,5}$ and $\kappa_{1,2,3}$. It also allows additional terms in the pure 2HDM sector, but we do not consider those here.
\item Custodial symmetry \cite{Sikivie:1980hm,Pomarol:1993mu,Grzadkowski:2010dj}. As discussed in \cite{Cheng:2016tlc} the least restrictive method to impose custodial symmetry results in all the $\lambda_i$'s being real and in the relations
\begin{eqnarray}
{\kappa_2} = {\kappa _3},\ 2{\nu _3} = {\nu _2},\ {\nu _4} = \nu _5^*,\ 2{\omega _3} = {\omega_2},\ {\omega _4} = \omega _5^*,\ \lambda_4 = \lambda_5.
\label{met1}
\end{eqnarray}
These conditions imply mass degeneracies $m_{H^\pm}=m_A$ and $m_{S^\pm}=m_{S_I^0}$. An alternative possibility, `twisted' custodial symmetry \cite{Gerard:2007kn,Cervero:2012cx} results instead in $m_{H^\pm}=m_H$ and $m_{S^\pm}=m_{S_R^0}$ \cite{Burgess:2009wm}.
\item Perturbative unitarity \cite{Lee:1977eg,Kanemura:1993hm, Horejsi:2005da,Ginzburg:2005dt,Grinstein:2015rtl,He:2013tla}
The two-to-two scattering matrix for this model in the neutral, color singlet channel is $18\times 18$ and in Ref.~\cite{Cheng:2016tlc} we diagonalize it numerically and illustrate the resulting constraints. The numerical results can be roughly approximated by,
\begin{subequations}
\begin{align}
&\left| \lambda_1 \right|,\left| \lambda_2 \right| \le \frac{8\pi}{3},\quad
\left| \lambda_3 \right| \le 4\pi,\quad \left| \lambda_4 \right|, \left| \lambda_5 \right| \le \frac{8\pi}{5},\\
&\left| \nu_1 \right|,\left| \nu_3 \right|, \left| \omega_1 \right|, \left| \omega_3 \right| \le 2\sqrt{2}\pi,\quad
\left| \nu_2 \right|,\left| \omega_2 \right| \le 4\sqrt{2}\pi,\\
&\left| \kappa_1 \right| \le 2\pi,\quad \left| \kappa_2 \right|, \left| \kappa_3 \right| \le 4\pi,\\
& \left| 17 \mu_3 +13 \mu_4 +13 \mu_6 \right|\leq 16 \pi,\\
& \left| \nu_4+\nu_5\right|\leq \frac{32\pi}{\sqrt{15}}, \quad
\left| \omega_4+\omega_5\right| \leq \frac{32\pi}{\sqrt{15}},\\
& \left| 12 \mu_3+10 \mu_4+7\mu_6\right|\leq 32\pi.
\label{approxlim}
\end{align}
\end{subequations}
In the present study we require the perturbative unitarity constraints, as obtained by diagonalizing the full $18\times 18$ matrix numerically, to be satisfied by the running couplings at all scales below $\Lambda$.
\item Stability, or having a positive definite Higgs potential for the 2HDM \cite{Deshpande:1977rw} implies that
\begin{equation}
\lambda_1 > 0, \quad
\lambda_2 > 0, \quad
\lambda_3 > - \sqrt{\lambda_1 \lambda_2},\quad
\lambda_3 + \lambda_4 \pm \lambda_5 > - \sqrt{\lambda_1 \lambda_2}.
\label{stabvac}
\end{equation}
We will again require that this conditions be satisfied by the running couplings at all scales below $\Lambda$. A recent paper presents a way to extend the stability conditions to additional fields \cite{Kannike:2016fmd}. Here we will only consider the effect of the color scalars through their one-loop contributions to the running of the parameters of the 2HDM, and require the conditions Eq.~\ref{stabvac} be satisfied at all scales.
\item For our numerical analysis we will use color octet scalar masses near 1~TeV as in Ref.~\cite{Cheng:2016tlc}, as masses at this scale are still allowed by LHC searches \cite{Hayreter:2017wra}.
\end{itemize}
\section{Validity of the model up to high energy scales}
\subsection{The renormalization group equations for the scalar couplings}
We now turn to the novel aspect of this paper: investigating the consequences of requiring the model to be valid up to some high energy scale. The procedure is straightforward, we first derive the corresponding renormalization group equations (RGE) for all the parameters in the model, a result we present in the Appendix. For our numerical analysis we restrict ourselves to the case with custodial symmetry and no CP violation, which reduces the RGE to the following,
\everymath{\displaystyle}
\begingroup
\allowdisplaybreaks
\begin{align*}
&16\pi^2\beta_{\lambda_1} = 12\lambda_1^2 + 4\lambda_3^2 + 4\lambda_3\lambda_4 +4\lambda_4^2
+ 8\nu_1^2 + 8\nu_1\nu_2 + 8\nu_2^2 \\
&\qquad\qquad\quad - 12 \lambda_t^4 - 3\lambda_1\left( 3g^2 - 4\lambda_t^2 + g'^2 \right) + \frac34 \left( 3g^4 + 2g^2 g'^2 + g'^4 \right),\\
&16\pi^2\beta_{\lambda_2} = 12\lambda_2^2 + 4\lambda_3^2 + 4\lambda_3\lambda_4 + 4\lambda_4^2
+ 8\omega_1^2 + 8\omega_1\nu_2 + 4\omega_2^2 + 16\omega_3^2,\\
&\qquad\qquad\quad - 3\lambda_2\left( 3g^2 + g'^2 \right) + \frac34 \left( 3g^4 + 2g^2 g'^2 + g'^4 \right),\\
&16\pi^2\beta_{\lambda_3} = 4\lambda_3^2 + 4\lambda_4^2 + 2\left(\lambda_1+\lambda_2\right)\left(3\lambda_3 + \lambda_4\right)
+ 8\nu_1\omega_1 + 4\nu_1\omega_2 + 4\nu_2\omega_1 + 8\kappa_2^2,\\
&\qquad\qquad\quad - 3\lambda_3\left( 3g^2 - 2\lambda_t^2 + g'^2 \right) + \frac34 \left( 3g^4 - 2g^2 g'^2 + g'^4 \right),\\
&16\pi^2\beta_{\lambda_4} = 12\lambda_4^2 + 8\lambda_3\lambda_4 + 2\left(\lambda_1+\lambda_2\right)\lambda_4 + 4\nu_2\omega_2
+ 8\left|\kappa_1\right|^2 + 8\kappa_1\kappa_2 + 4\kappa_2^2,\\
&\qquad\qquad\quad - 3\lambda_4\left( 3g^2 - 2\lambda_t^2 + g'^2 \right) + \frac32 g^2 g'^2,\\
&16\pi^2\beta_{\nu_1} = 6\lambda_1\nu_1 + 2\lambda_1\nu_2 + 4\lambda_3\omega_1 + 2\lambda_3\omega_2 + 2\lambda_4\omega_1 + 2\nu_1^2 + 2\nu_2^2 + 2\kappa_1^2 + 2\kappa_2^2 \\
&\qquad\qquad\quad + \nu_1\left(26\mu_1 + 17\mu_3 + 13\mu_4\right)
+\nu_2\left(\frac{32}3\mu_1 + 8\mu_3 + 2\mu_4\right) + \frac{10}3 \nu_4^2,\\
&16\pi^2\beta_{\nu_2} = 2\lambda_1\nu_2 + 2\lambda_4\omega_2 + 4\nu_1\nu_2 + 6\nu_2^2 + 4\kappa_1\kappa_2 + 6\kappa_2^2 \\
&\qquad\qquad\quad + \nu_2\left(\frac{14}3\mu_1 + \mu_3 + 9\mu_4\right) + \frac{25}3\nu_4^2,\\
&16\pi^2\beta_{\nu_4} =\left(3\kappa_1 + 9\kappa_2\right)\omega_4 + \nu_4\left(3\nu_1 + 9\nu_2 + 11\mu_1 + 3\mu_3 + 9\mu_4\right),\\
&16\pi^2\beta_{\omega_1} = 6\lambda_2\omega_1 + 2\lambda_2\omega_2 + 4\lambda_3\nu_1 + 2\lambda_3\nu_2 + 2\lambda_4\nu_1 + 2\omega_1^2 + 2\omega_2^2 + 2\kappa_1^2 + 2\kappa_2^2 \\
&\qquad\qquad\quad + \omega_1\left(26\mu_1 + 17\mu_3 + 13\mu_4\right)
+\omega_2\left(\frac{32}3\mu_1 + 8\mu_3 + 2\mu_4\right) + \frac{10}3\omega_4^2,\\
&16\pi^2\beta_{\omega_2} = 2\lambda_2\omega_2 + 2\lambda_4\nu_2 + 4\omega_1\omega_2 + 6\omega_2^2 + 4\kappa_1\kappa_2 + 6\kappa_2^2 \\
&\qquad\qquad\quad+ \omega_2\left(\frac{14}3\mu_1 + \mu_3 + 9\mu_4\right) + \frac{25}3\omega_4^2,\\
&16\pi^2\beta_{\omega_4} = \left(3\kappa_1 + 9\kappa_2\right)\nu_4 + \omega_4\left(3\omega_1 + 9\omega_2 + 11\mu_1 + 3\mu_3 + 9\mu_4\right),\\
&16\pi^2\beta_{\kappa_1} = \kappa_1\left(2\lambda_3 + 10\lambda_4 + 2\nu_1 + 2\omega_1 + 18\mu_1 + 17\mu_3 + 13\mu_4 \right)\\
&\qquad\qquad\quad +\kappa_2\left(4\lambda_4 + \frac32\nu_2 + \frac32\omega_2 + \frac{32}3\mu_1
+ 8\mu_3 + 4\mu_4\right) + \frac{10}3\nu_4\omega_4,\\
&16\pi^2\beta_{\kappa_2} = \kappa_1\left(2\nu_2 + 2\omega_2 \right) + 14\nu_4\omega_4\\
&\qquad\qquad\quad+\kappa_2\left(2\lambda_3 + 2\lambda_4 + 2\nu_1 + 4\nu_2 + 2\omega_1 + 4\omega_2 + \frac{14}3\mu_1 + \mu_3 + 9\mu_4\right) ,\\
&16\pi^2\beta_{\mu_1} = 3\nu_4^2 + 3\omega_4^2
+ 13\mu_1^2 + 6\mu_1\left(\mu_3 + \mu_4\right),\\
&16\pi^2\beta_{\mu_3} = 2\nu_1^2 + 2\nu_1\nu_2 + 2\omega_1^2 + 2\omega_1\omega_2 + 4\kappa_1^2 + 4\kappa_1\kappa_2 - \frac{10}3\left(\nu_4^2 + \omega_4^2\right)\\
&\qquad\qquad\quad + \frac{268}9\mu_1^2 + \mu_1\left(52\mu_3 + \frac{88}3\mu_4\right)
+ 20\mu_3^2 + 26\mu_3\mu_4 + 6\mu_4^2,\\
&16\pi^2\beta_{\mu_4} = \nu_2^2 + \omega_2^2 + 2\kappa_2^2 + \frac23\left(\nu_4^2 + \omega_4^2 \right) + \frac49\mu_1^2 + \frac{52}3\mu_1\mu_4 + 6\mu_3\mu_4 + 16\mu_4^2,\\
&16\pi^2\beta_{g_s} = - 6g_s^3,\\
&16\pi^2\beta_g = - \frac53 g^3,\\
&16\pi^2\beta_{g'} = \frac{25}3 g'^3,\\
&16\pi^2\beta_{\lambda_t} = \lambda_t\left(-\frac94 g^2 - 8g_s^2 + \frac{13}2\lambda_t^2
- \frac{17}{12}g'^2\right).
\end{align*}
\endgroup
In these equations $g$, $g'$ and $g_s$ are the $SU(2)_L$, $U(1)_Y$ and $SU(3)$ couplings of the SM, $\lambda_t \equiv \sqrt2 m_t / v$ is the top-quark Yukawa coupling and we use the standard definition $\beta=\frac{\rm d}{{\rm d}\ln\left(\Lambda/\Lambda_0\right)}$.
These equations have been checked against the known limits: the 2HDM \cite{Branco:2011iw}; the MW model \cite{He:2013tla}.
When we include the effect of the coupling $g'$ in the RGE we end up with high scale couplings that no longer satisfy custodial symmetry. The deviations from the symmetry limit are small as expected, proportional to $g'$, and we ignore them in our numerical analysis.
\subsection{The running of the scalar couplings}
As we run the couplings between the electroweak and high scales we find three possible behaviors: well behaved couplings at all scales; one or more of the couplings develop a LP; or even though there are no LP for the scales considered, perturbative unitarity or stability are violated at some scale in the range. We illustrate these three possibilities below.
Figure \ref{f:running1} illustrates the case of well behaved couplings up to the Planck scale. Figure \ref{f:running2} shows how it is possible to develop multiple LP at relatively low energy scales even when the couplings are small and perturbative at the electroweak scale. Finally, figure \ref{f:running3} illustrates a case where there are no LP below the Planck scale, but perturbative unitarity is violated at some point below $\Lambda_{\rm Planck}$. In this case, the violation is due to the presence of a LP just beyond $\Lambda_{\rm Planck}$.
\begin{figure}[thb]
\includegraphics[width=.45\textwidth]{fig_running1_1}
\includegraphics[width=.45\textwidth]{fig_running1_2}
\includegraphics[width=.45\textwidth]{fig_running1_3}
\includegraphics[width=.45\textwidth]{fig_running1_4}
\caption{Running couplings for a case that satisfies unitarity and stability conditions at all scales below $\Lambda_{\rm Planck}$. }
\label{f:running1}
\end{figure}
\begin{figure}[thb]
\includegraphics[width=.45\textwidth]{fig_running2_1}
\includegraphics[width=.45\textwidth]{fig_running2_2}
\includegraphics[width=.45\textwidth]{fig_running2_3}
\includegraphics[width=.45\textwidth]{fig_running2_4}
\caption{Running couplings for a case where a LP is encountered below $\Lambda_{\rm Planck}$. }
\label{f:running2}
\end{figure}
\begin{figure}[thb]
\includegraphics[width=.45\textwidth]{fig_running3_1}
\includegraphics[width=.45\textwidth]{fig_running3_2}
\includegraphics[width=.45\textwidth]{fig_running3_3}
\includegraphics[width=.45\textwidth]{fig_running3_4}
\caption{Running couplings for a case where no LP is encountered below $\Lambda_{\rm Planck}$, but the unitarity and stability conditions are not satisfied for all scales $\Lambda < \Lambda_{\rm Planck}$. }
\label{f:running3}
\end{figure}
\subsection{Constraining the allowed region of parameter space}
For this purpose we use Mathematica to solve the RGE numerically with initial conditions at the electroweak scale. The initial conditions used for the new couplings are points that satisfy perturbative unitarity and stability determined as in Ref.~\cite{Cheng:2016tlc}. For each point we evolve all the couplings up to a high scale $\Lambda$ and discard the point if a LP is detected, or if the couplings at any scale below $\Lambda$ violate the perturbative unitarity or stability conditions. This results in acceptable points satisfying a more stringent condition than the absence of LP, in the spirit of renormalization group improved unitarity bounds of Ref.~\cite{Marciano:1989ns}. The use of this condition in our numerical search makes it easier to find acceptable points than if we were to allow a LP at $\Lambda$. It would also be possible to constrain the parameters using higher order unitarity conditions, but we do not pursue this in this paper \cite{Cacchio:2016qyh,Murphy:2017ojk}.
The parameter space is too large for a completely random scan to be efficient. Instead, we follow the approach described below.
\begin{enumerate}
\item We begin our study at an intermediate energy scale $\Lambda_m$ which, for the sake of computational efficiency. We choose (by trial and error) it to be $\ln (\Lambda_m/{10^3\ {\rm GeV}}) = 10$.
\item Before running the RGE for the whole model, we generate a large sample of points within the 2HDM subspace. The sample is generated in such a way that a large portion of it is valid up to $\Lambda_m$.
\item We then use this 2HDM data set as seeds, randomly assigning values to the new couplings within a proper range, to generate a sample for the whole parameter space. From this sample we find a few hundred valid points and determine the hypercube which contains most of the solutions, a region somewhat smaller than that allowed by perturbative unitarity at the electroweak scale.
\item Starting from these few hundred points we study nearby points to expand the allowed region.
\item We finally construct the region of parameter space where the full model is valid up to the scale $\Lambda_m$ by repeating step 4 recursively for a sufficiently long time.
\item Points that are valid up to scales higher than $\Lambda_m$, are generally inside a sub-region of the allowed region up to scale $\Lambda_m$. Therefore, to find the constraints for a higher scale, we select the seed points from step 5 and repeat step 4 to construct the new allowed region.
\end{enumerate}
Our results are illustrated in Figure~\ref{f:uni-GUT} for the scales $\Lambda_m$ and $\Lambda_{\rm GUT}$ in representative two-dimensional projections. The GUT scale is chosen because of the existing $SU(5)$ \cite{Perez:2016qbo} or $SO(10)$ \cite{Bertolini:2013vta} models which can have TeV scale scalar color octets, but the figures illustrate the general trend as we require the model to be valid up to higher energy scales. In general, for the new parameters involving the color-octet scalars, the allowed parameter space is now very significantly reduced with respect to that allowed by tree-level unitarity. In addition, this procedure produces the first constraints on parameters like $\nu_4$ and $\omega_4$ which do not affect two to two processes at tree-level.
\begin{figure}[thb]
\includegraphics[width=.45\textwidth]{fig_lambda12}
\includegraphics[width=.45\textwidth]{fig_lambda34}
\includegraphics[width=.45\textwidth]{fig_nuomega1}
\includegraphics[width=.45\textwidth]{fig_nuomega2}
\includegraphics[width=.45\textwidth]{fig_kappa12}
\includegraphics[width=.45\textwidth]{fig_nuomega4}
\includegraphics[width=.45\textwidth]{fig_mu14}
\includegraphics[width=.45\textwidth]{fig_mu34}
\caption{Representative two-dimensional projections of the allowed parameter space for which the model is valid up to the intermediate scale $\Lambda_m$ (red) and $\Lambda_{\rm GUT}$ (blue). The black points are the seed points used as initial values at the electroweak scale that satisfy both perturbative unitarity and stability.
\label{f:uni-GUT}}
\end{figure}
Validity up to the GUT scale thus results in approximate one at a time constraints
\begin{eqnarray}
0\leq \lambda_{1,2} \lsim 0.5, && -0.4 \lsim \lambda_3 \lsim 0.7,\nonumber \\
-0.4 \leq \lambda_{4} \lsim 0.4, && |\omega_{1,2}| \lsim 1.6, \nonumber \\
|\nu_{1,2}| \lsim 1.6, && |\kappa_{1,2}| \lsim 1.4, \nonumber \\
|\nu_{4}| \lsim 2, && |\omega_{4}| \lsim 2, \nonumber \\
|\mu_{1,3}| \lsim 3.3, && |\mu_4| \lsim 4.2.
\end{eqnarray}
For the parameters of the 2HDM the points that produce a model valid to high energy scales are those for which $\cos(\beta-\alpha)$ is closer to zero, as in the alignment limit; and those for which $M_H$ is very close to $M_{H^\pm}$ as shown in Figure~\ref{f:2hdm-GUT}.
\begin{figure}[thb]
\includegraphics[width=.45\textwidth]{fig_angle}
\includegraphics[width=.45\textwidth]{fig_Hmass}
\caption{Points for which the 2HDM model satisfies both perturbative unitarity and stability at the electroweak scale (black) compared to those for which it is also valid up to the scale $\Lambda_m$ (red) and $\Lambda_{\rm GUT}$ (blue).
\label{f:2hdm-GUT}}
\end{figure}
\section{Conclusions}
In this paper, we provide the renormalization group equations for all the couplings in the scalar potential of a 2HDM augmented with a color-octet. As an application, we constrain the parameters of the model by requiring it to be valid up to some high scale.
The acceptable region of the parameter space that satisfies both unitarity and stability constraints without developing LP up to a high scale is determined numerically, and the resulting constraints are provided for the case $\Lambda_{\rm HIGH}=\Lambda_{\rm GUT}$. As expected, the allowed region is reduced as the scale increases; looking at the 2HDM subspace, it contracts towards the alignment limit and mass-degeneracy of heavy neutral and charged Higgses.
\begin{acknowledgments}
We thank Dianne Cook for help visualizing the multidimensional parameter space with the aid of the grand and guided tour \cite{method} with the software GGOBI \cite{visual}.
Li Cheng thanks David Atwood and Kerry Whisnant for useful discussions. The work of GV was supported in part by the Australian Government through the Australian Research Council.
\end{acknowledgments}
\pagebreak
|
\section{\bf Introduction.}
All graphs considered here are simple and finite, unless otherwise stated. Let $C_k\,(\text{respectively }P_{k}),$ denote the cycle (respectively path) on $k$ vertices. The complete graph on $n$ vertices is denoted by $K_n$ and its complement is denoted by $\overline{K}_n.$ For two graphs $G$ and $H$ their {\it tensor product}, denoted by $G\times H,$ has vertex set $V(G)\times V(H)$ in which two vertices $(g_1,h_1)\text { and }(g_2,h_2)$ are adjacent whenever $g_1g_2\in E(G)\text{ and }h_1h_2\in E(H).$ The {\it wreath product} of the graphs $G$ and $H,$ denoted by $G\circ H,$ has vertex set $V(G)\times V(H)$ in which $(g_1,\,h_1)(g_2,\,h_2)$ is an edge whenever $g_1g_2$ is an edge in $G,$ or $g_1=g_2$ and $h_1h_2$ is an edge in $H.$ Similarly, the \emph{cartesian product} of the graphs $G$ and $H,$ denoted by $G\Box H,$ has vertex set $V(G)\times V(H)$ in which $(g_1,\,h_1)(g_2,\,h_2)$ is an edge whenever $g_1=g_2$ and $h_1h_2$ is an edge in $H,$ or $h_1=h_2$ and $g_1g_2$ is an edge in $G.$ The subgraph induced by $S\subseteq V(G)$ is denoted by $\langle S\rangle.$ Similarly, the subgraph induced by $E^\prime\subseteq E(G)$ is denoted by $\langle E^\prime\rangle.$ For a graph $G,$ $G(\lambda)$ is the graph obtained from $G$ by replacing each edge of $G$ by $\lambda$ parallel edges. For a graph $G,$ $G^*$ is the symmetric digraph of $G.$
For two loopless multigraphs $G(\lambda)$ and $H(\mu),$ their tensor product, denoted by $G(\lambda)\times H(\mu),$ has the vertex set $V(G)\times V(H)$ and its edge set is described as follows: if $e=g_1g_2$ is an edge of multiplicity $\lambda$ in $G(\lambda)$ and $f=h_1h_2$ is an edge of multiplicity $\mu$ in $H(\mu),$ then corresponding to these edges there are edges $(g_1,h_1)(g_2,h_2)$ and $(g_1,h_2)(g_2,h_1)$ each of multiplicity $\lambda\mu$ in $G(\lambda)\times H(\mu)$ and $G(\lambda)\times H(\mu)$ is isomorphic to $(G\times H)(\lambda\mu).$ Hence $G(\lambda)\times H\cong G\times H(\lambda)\cong (G\times H)(\lambda).$
If $H_1,H_2,\ldots,H_k$ are edge-disjoint subgraphs of $G$ and $E(G)=E(H_1)\cup E(H_2)\cup\ldots\cup E(H_k),$ then we write $G=H_1\oplus H_2\oplus\ldots\oplus H_k.$ For a graph $G,$ if $E(G)$ can be partitioned into $E_1,$$E_2,\ldots,$$E_k$ such that $\langle E_i\rangle\cong H,$ for all $i,\,1\le i\le k,$ then we say that $H$ decomposes $G,$ or {\it $H$-decomposition} of $G$ exists. Clearly, the tensor product is commutative and distributive over edge-disjoint union of graphs, that is, if $G=H_1\oplus H_2\oplus \ldots \oplus H_k,$ then $G\times H=(H_1\times H)\oplus (H_2\times H)\oplus\ldots\oplus(H_k\times H).$
Let $G$ be a finite group and let $S$ be a symmetric subset of $G$ (that is, $s\in S \text{ implies } -s\in S).$ The vertices of the {\it Cayley graph,} $Cay(S,\, G),$ are the elements of $G$ and there is an edge between $x$ and $y$ if and only if $x-y \in S.$ Note that $Cay(S,\, G)$ is connected if and only if $S$ generates the group $G.$ A \emph{circulant} $X=Circ(n;\,L)$ is a graph with vertex set $V(X)=\{u_0,\,u_1,\,\ldots,\,u_{n-1}\}$ and edge set $E(X)=\{u_iu_{i+\ell}\,|\,i\in \mathbb{Z}_n,\,\ell \in L\},$ where $L\subseteq\left\{1,\,2,\,\ldots,\,\lfloor\frac{n}{2}\rfloor\right\}$ and $\mathbb{Z}_n$ is the set of integers modulo $n.$ The elements of $L$ are called {\it jumps.} Clearly, every circulant graph of order $n$ is a Cayley graph with the underlying group being $\mathbb{Z}_{n}.$
The problem of finding Hamilton cycle decompositions of product graphs is not new. Hamilton cycle decompositions of various product graphs, including digraphs, have been studied by many authors; see, for example, \cite {{CyclesandRays}, {Ann.DM.3.21-28}, {Bosak.Book}, {IJPAM.23.723-729}, {Comput.Mat.Appl.31.11-19}, {ARS.80.33-44}, {DM.308.3586-3606}, {DM.310.2776-2789}, {JCTB.73.119-129}, {DM.90.169-190}, GC.25.571-581}. It has been conjectured \cite{Ann.DM.3.21-28} that if both $G$ and $H$ are Hamilton cycle decomposable graphs, then $G\Box H$ is Hamilton cycle decomposable, where $\Box$ denotes the cartesian product of graphs. This conjecture has been verified to be true for a large classes of graphs \cite{DM.90.169-190}. Baranyai and Sz\'{a}sz \cite{JCTB.31.253-261} proved that if both $G$ and $H$ are even regular Hamilton cycle decomposable graphs, then $G \circ H$ is Hamilton cycle decomposable. In \cite{JCTB.73.119-129}, Ng has obtained a partial solution to the following conjecture of Alspach et al. \cite{CyclesandRays}: If $D_1$ and $D_2$ are directed Hamilton cycle decomposable digraphs, then $D_1\circ D_2$ is directed Hamilton cycle decomposable. Jha \cite{IJPAM.23.723-729} conjectured the following
\begin{con}\emph{\cite{IJPAM.23.723-729}}\label{IJPAM.23.723-729}
If both $G$ and $H$ are Hamilton cycle decomposable graphs and $G\times H$ is connected, then $G\times H$ is Hamilton cycle decomposable.
\end{con}
Conjecture \ref{IJPAM.23.723-729} was disproved, see \cite{DM.186.1-13}. Because of this, finding Hamilton cycle decompositions of the tensor products of Hamilton cycle decomposable graphs is considered to be difficult. Though Conjecture \ref{IJPAM.23.723-729} has been disproved, we believe that if the graphs $G$ and $H$ are suitably chosen, that is, with some suitable conditions imposed on them, then $G\times H$ may have Hamilton cycle decomposition. In \cite{DM.268.49-58} and \cite{ARS.80.33-44} it has been proved that $K_r \times K_s$ and $K_{r ,r} \times K_m$ are Hamilton cycle decomposable; in \cite{DM.308.3586-3606} it is shown that the tensor product of two regular complete multipartite graphs is Hamilton cycle decomposable. Hamilton cycle decompositions of the tensor products of complete bipartite graphs and complete multipartite graphs are dealt with in \cite{DM.310.2776-2789}. Also in \cite{GC.25.571-581}, Paulraja and Sivasankar proved that $(K_r\times K_s)^*,\,((K_r \circ \overline{K}_s)\times K_n)^*,\, ((K_r \times K_s)\times K_m)^*,\,((K_r\circ \overline{K}_s)\times (K_m\circ \overline{K}_n))^*$ and $(K_{r,r} \times (K_m\circ \overline{K}_n))^*$ are directed Hamilton cycle decomposable. It can be observed that $K_r,\,K_{r,\,r},\,K_r\circ\overline{K}_s$ are circulant graphs. Based on the results of \cite{{IJPAM.23.723-729}, {Comput.Mat.Appl.31.11-19}, {ARS.80.33-44}, {DM.308.3586-3606}, {DM.310.2776-2789}}, Manikandan and Paulraja conjectured the following:
\begin{con}\emph{(Manikandan and Paulraja) \cite{Thesis.RSM}.}\label{con.RSMPP}
If $G$ and $H$ are Hamilton cycle decomposable circulant graphs and at least one of them is non bipartite, then $G\times H$ is Hamilton cycle decomposable.
\end{con}
\begin{thm}\emph{\cite{Ann.DM.3.21-28}}\label{Ann.DM.3.21-28}
If both $G$ and $H$ have Hamilton cycle decompositions and at least one of them is of odd order, then $G\times H$ admits a Hamilton cycle decomposition.\hfill$\Box$
\end{thm}
One may naturally ask when is $G\times H$ Hamilton cycle decomposable, if both $G$ and $H$ are of even order. In this paper it is partially answered.
We say that an even regular circulant graph $X=Circ(n;\,L)$ has the {\it property $Q,$} if (1) the number of odd and even jumps in $L$ are equal; (2) odd jumps can be paired with even jumps so that each of the $\frac{|L|}{2}$ resulting $4$-regular graphs is connected. It is known that every $4$-regular connected circulant graph is Hamilton cycle decomposable; see \cite{JCTB.46.142-153}. Consequently, every circulant graph with property $Q$ is Hamilton cycle decomposable.
Here we prove the following main Theorem.
\begin{thm}\label{THM.CIR.1}
Let $G$ be a circulant graph with property $Q$ and let $H$ be any Hamilton cycle decomposable multigraph, then $G\times H$ is Hamilton cycle decomposable.
\end{thm}
This theorem has many interesting consequences. In particular, we have the following corollary.
\begin{cor}\label{COR.CIR.1}
If $G$ and $H$ are even regular Hamilton cycle decomposable circulant graphs and at least one of them has the property $Q,$ then $G\times H$ is Hamilton cycle decomposable.
\end{cor}
One of the consequences of Corollary \ref{COR.CIR.1} is that if $G=(K_{4n+2}-F)$ and $H=(K_{2m}-F^\prime),$ where $F$ and $F^\prime$ are $1$-factors of $K_{4n+2}$ and $K_{2m},$ respectively, then $G\times H$ is Hamilton cycle decomposable. In particular, after deleting suitable number of jumps of $K_{4n+2}$ and $K_{2m},$ the resulting graphs need not be dense but their tensor product is Hamilton cycle decomposable. This cannot be deduced from the existing results in this direction.
\section{\bf Notation and Preliminaries.}
First we present the necessary definitions here. The notation that we use here are from \cite{JCTB.46.142-153} and for the sake of completeness we give them.
The Cayley graph $\Gamma=Cay(n,\,S),$ where $S=\{a,\,b\}$ is a generating set of the finite abelian group with $2a\ne 0,\,2b\ne 0$ and $a\ne\pm b,$ is a simple connected graph. We call the edge $x(x+a)$ of $\Gamma$ an $a$-edge; the subgraph formed by the $a$-edges is a disjoint union of cycles, called $a$-cycles, each of length $k_a,$ the order of the element $a.$ Similarly, we define $b$-edges and $b$-cycles; each $b$-cycle is of length $k_b,$ the order of the element $b.$ In $\Gamma,$ let us denote the number of $a$-cycles by $\alpha$ and the number of $b$-cycles by $\beta.$ Since the length of each $a$-cycle (respectively $b$-cycle) is $k_a=\frac{n}{\alpha},$ (respectively $k_b=\frac{n}{\beta},$) $n=\alpha k_a$ (respectively $n=\beta k_b$).
As the graph $\Gamma$ is connected, the vertices $0,\,b,\,\ldots,\,\ell b,\,\ldots,\,(\alpha-1)b$ are in the $\alpha$ $a$-cycles denoted by $C_0,\,C_1,\,\ldots,C_\ell,\,\ldots,C_{\alpha-1},$ respectively, and $\alpha b$ belongs to $C_0,$ see \cite{JCTB.46.142-153}. Hence, we have $\alpha b=ca$ for some $c$ with $0\le c\le k_a-1.$ Further, every vertex $x$ of $\Gamma$ can be uniquely written as $x=ib+ja$ with $i\in \{0,\,1,\,\ldots,\, \alpha-1\}$ and $j\in \{0,\,1,\,\ldots,\, k_a-1\}.$ Using this uniqueness, label the vertices of $\Gamma$ as $(i,\,j)$ with $i\in \{0,\,1,\,\ldots,\, \alpha-1\},$ and $j\in \{0,\,1,\,\ldots,\, k_a-1\},$ where the first coordinate indicates the label of the cycle $C_i$ containing the vertex and the second coordinate indicates the position of the vertex on the cycle.
The following definition comprises the notation and defines a class of simple graphs, which we denote by $\Gamma(\alpha,\,\beta).$
\begin{defn}\emph{\cite{JCTB.46.142-153}}\label{JCTB.46.142-153}
Let $\Gamma(\alpha,\,\beta)$ denote the class of simple graphs on $\alpha k$ vertices, where $\alpha\ge 1,\,k\ge 3,\,0\le c<k,$ and $\beta=gcd(k,\,c).$ The $\alpha k$ vertices of the graph can be labeled $(i,\,j)$ with $i$ taken modulo $\alpha$ and $j$ taken modulo $k.$ The edges are \emph{(1)} First kind: $(i,\,j)(i,\,j+1)$ and \emph{(2)} Second kind: $(i,\,j)(i+1,\,j)$ for all $i\in \{0,\,1,\,\ldots,\, \alpha-2\},$ and $(\alpha-1,\,j)(0,\,j+c).$
\end{defn}
A graph $\Gamma\in\Gamma(\alpha,\,\beta)$ has $\alpha$ \lq\lq vertical" disjoint cycles $C_i,\,0\le i<\alpha,$ with a natural orientation and $\alpha-1$ horizontal {\it parallel matchings} between the cycles $C_i$ and $C_{i+1}$ for $0\le i<\alpha-1,$ and a particular {\it parallel matching} between $C_{\alpha-1}$ and $C_0$ (which depends on the value of $c).$ A graph $\Gamma\in\Gamma(2,\,\beta)$ consists of two cycles plus two perfect matchings between them. A graph $\Gamma\in\Gamma(1,\,\beta)$ consists of a cycle plus the chords joining $(0,\,j)$ to $(0,\,j+c).$ Observe that $\Gamma(\alpha,\,\beta)$ and $\Gamma(\beta,\,\alpha)$ are isomorphic classes of graphs, that is, an element $\Gamma_1\in\Gamma(\alpha,\,\beta)$ is isomorphic to an element $\Gamma_2\in\Gamma(\beta,\,\alpha)$ and vice versa, for example see Figure 1.
\begin{center}
\includegraphics{fig1.eps}
\end{center}
{\small Graphs of Figures $1(a),\,1(b)$ and $1(c)$ are isomorphic. \\$(a)$ Graph $G=Circ(12,\,\{2,\,3\}),$ where the edges of jump $2$ give $2$ disjoint cycles each of length $6$ and the edges of jump $3$ give $3$ disjoint cycles each of length $4.$ \\$(b)$ $G$ is an element of $\Gamma(2,\,3)$ with $c=3.$ \\$(c)$ $G$ in $(b)$ is drawn as an element of $\Gamma(3,\,2)$ with $c=2.$
}\begin{center}Figure 1\end{center}
\begin{defn}\emph{\cite{JCTB.46.142-153}}\label{JCTB.46.142-153-defn}
A Hamilton cycle decomposition of a graph $\Gamma\in\Gamma(\alpha,\,\beta)$ has the property $Q_1,$ if both Hamilton cycles use at least one edge of the matching between $C_i$ and $C_{i+1},$ for all $i\in\{0,\,1,\,\ldots,\,\alpha-1\},$ where $C_\alpha=C_0.$
\end{defn}
\begin{thm}\emph{\cite{JCTB.46.142-153}}\label{JCTB.46.142-153-1}
The class $\Gamma(\alpha,\,\beta)$ consists of the $4$-regular connected Cayley graphs on a finite abelian group with a generating set $\{a,\,b\},$ where $\alpha$ is the number of $a$-cycles and $\beta$ is the number of $b$-cycles. \hfill$\Box$
\end{thm}
Using the above theorem, Bermond et al. proved the following
\begin{thm}\emph{\cite{JCTB.46.142-153}}\label{JCTB.46.142-153-main}
Every $4$-regular connected Cayley graph on a finite abelian group can be decomposed into two Hamilton cycles.
\end{thm}
If $G$ and $H$ are Hamilton cycle decomposable graphs with at least one of them having odd order, then $G\times H$ is Hamilton cycle decomposable, by Theorem \ref{Ann.DM.3.21-28}; hence in what follows, {\bf we assume that both $G$ and $H$ are Hamilton cycle decomposable graphs each having an even number of vertices.} If $G$ is a connected $4$-regular circulant graph of even order with generating set $\{a,\,b\}$ where $a$ and $b$ are of different parity such that $2a\ne 0,\,2b\ne 0$ and $a\ne\pm b,$ then by the Definition \ref{JCTB.46.142-153}, $\alpha$ and $\beta$ are of different parity. For our convenience, according to the situation we consider the parity of $a$ and $b.$ {\bf In what follows, we assume that $\Gamma(\alpha,\,\beta)$ denotes the class of Cayley graphs on the abelian group $\mathbb{Z}_n$} (that is, circulants) with generating set $\{a,\,b\}$ and by Theorem \ref{JCTB.46.142-153-1}, $\Gamma(\alpha,\,\beta)$ consists of the class of $4$-regular connected circulants of order $n$ with generating set $\{a,\,b\},$ where $\alpha$ is the number of $a$-cycles in a $2$-factor and $\beta$ is the number of $b$-cycles.
In the rest of the paper, {\bf we assume that $\Gamma\in\Gamma(\alpha,\,\beta)$ is a $4$-regular circulant graph of even order with generating set $\{a,\,b\}$ where $a$ and $b$ are of different parity; consequently, $\alpha$ and $\beta$ are of different parity.}
\begin{defn}\label{defn.jha}
Let $\{H_1,\,H_2\}$ be a Hamilton cycle decomposition of a $4$-regular graph $G$ of even order with at least $6$ vertices. If $G$ contains a $4$-cycle $(a\,b\,c\,d)$ such that the edges $ab$ and $cd$ belong to one of the two Hamilton cycles and the edges $bc$ and $da$ are on the other Hamilton cycle, then the $4$-cycle is said to be an {\it alternating} $4$-cycle in $G,$ with respect to $H_1$ and $H_2.$ Further, if the vertices $a$ and $c$ \emph{(}or $b$ and $d$\emph{)} are at an odd distance along each of the two Hamilton cycles $H_i,$ then the $4$-cycle $(a\,b\,c\,d)$ is said to be an odd alternating $4$-cycle in $G$ with respect to $\{H_1,\,H_2\}.$ We say that the graph $G$ has {\it property $Q_2,$} if $G$ contains an odd alternating $4$-cycle with respect to a Hamilton cycle decomposition of it.
\end{defn}
The proof techniques we use here heavily depend on \cite{JCTB.46.142-153}.
The {\it reduced graph} $\Gamma^\prime\in\Gamma(\alpha,\,\beta)$ with respect to $\Gamma\in\Gamma(\alpha+2,\,\beta)$ is defined as follows: delete the vertices of the cycles $C_\alpha$ and $C_{\alpha+1}$ of $\Gamma,$ that is, the vertices $(\alpha,\,j)$ and $(\alpha+1,\,j),\,0\le j\le k-1,$ and add the edges $(\alpha-1,\,j)(0,\,j+c),\,0\le j\le k-1,$ joining $C_{\alpha-1}$ and $C_0,$ where $c$ is given in the definition of $\Gamma(\alpha+2,\,\beta).$ In some cases $\Gamma^\prime$ might be a multigraph and, if it is a simple graph, it is an element of $\Gamma(\alpha,\,\beta)$ ($k$ and $c$ being unchanged). Successive reductions of the resulting graphs result in a graph $\Gamma^{k}\in\Gamma(\alpha+2-2k,\,\beta)$ for some $k\ge 1.$ $\Gamma^{k}\in\Gamma(\alpha+2-2k,\,\beta)$ is also called a reduced graph of $\Gamma\in\Gamma(\alpha+2,\,\beta).$
The graph $\Gamma\in\Gamma(\nu+2,\,\delta)$ is said to be {\it lift graph} of $\Gamma^\prime\in\Gamma(\nu,\,\delta),$ if $\Gamma$ is obtained as follows: delete the edges $(\nu-1,\,j)(0,\,j+c),\,0\le j\le k-1,$ joining $C_{\nu-1}$ and $C_0$ of $\Gamma^\prime;$ then add two $a$-cycles $C_{\nu}$ and $C_{\nu+1}$ with vertices $(\nu,\,j)$ and $(\nu+1,\,j),\,0\le j\le k-1,$ to $\Gamma^\prime,$ and add the edges $(\nu-1,\,j)(\nu,\,j),\,(\nu,\,j)(\nu+1,\,j),\,(\nu+1,\,j)(0,\,j+c),\,0\le j\le k-1,$ where $c$ is as given in the definition of $\Gamma(\nu,\,\delta).$ Successive liftings of the resulting graphs result in a graph $\Gamma^{k}\in\Gamma(\nu+2k,\,\delta)$ for some $k\ge 1.$ $\Gamma^{k}\in\Gamma(\nu+2k,\,\delta)$ is also called a lift graph of $\Gamma^\prime\in\Gamma(\nu,\,\delta).$
We use $p_G(x,\,y)$ to denote the length of a path (not necessarily a shortest path) from $x$ to $y$ in the graph $G.$ By $(a,\,b)$-section, we denote a path from $a$ to $b.$
\begin{lem}\label{LEM2.directcirculant}
Every graph $\Gamma\in\Gamma(\alpha,\,\beta),$ where $1\le \alpha,\,\beta\le 2,$ has a Hamilton cycle decomposition $\{H_1,\,H_2\}$ with the properties $Q_1$ and $Q_2.$
\end{lem}
{\bf Proof.} Let $\Gamma\in\Gamma(\alpha,\,\beta).$ Since $1\le \alpha,\,\beta\le 2$ and $\alpha\ne \beta,$ without loss of generality assume that $\alpha=2$ and $\beta=1.$ Consequently, $a$ is even and $b$ is odd as $2=\alpha=gcd(n,\,a)$ and $1=\beta=gcd(n,\,b).$ Clearly, $c\ne 0$ (otherwise $\beta\ne 1).$ The existence of Hamilton cycles $H_1$ and $H_2$ in $\Gamma$ described below is in \cite{JCTB.46.142-153}.
A Hamilton cycle $H_{1}$ is obtained by deleting the edges $0a$ and $(-b)(-b+a)$ from $C_0$ and $C_1,$ respectively, and adding the edges $(-b)0$ and $(-b+a)a$ connecting these two $a$-cycles, that is, $H_1=\{C_0-\{0a\}\}\cup \{C_1-\{(-b)(-b+a)\}\}\cup \{0(-b),\,(-b+a)a\},$ see Figure 2(b); $H_2=\Gamma-E(H_1)$ is shown in the solid lines of Figure 2(c1). Clearly, these two Hamilton cycles use both $a$ and $b$-edges and hence the property $Q_1$ holds.
\begin{center}
\includegraphics[scale=0.8]{fig2.eps}
\end{center}
{\small (a). A member $\Gamma$ of $\Gamma(2,\,1)$\\ (b). Hamilton cycle $H_1$ of $\Gamma\in\Gamma(2,\,1)$\\ (c1). The solid lines are edges on $H_2.$ Other than the two $a$-edges $0a$ and $(-b)(-b+a),$ all edges of $H_2$ (and thus all edges connecting $C_0$ and $C_1)$ are $b$-edges. \\ (c2). Another drawing of the Hamilton cycle $H_2$ of $\Gamma\in\Gamma(2,\,1).$}\begin{center}Figure 2
\end{center}
\noindent{\bf Claim.} The graph $\Gamma\in\Gamma(2,\,1)$ satisfies the property $Q_2$ with respect to the above Hamilton cycle decomposition $\{H_1,\,H_2\}.$
Since $\alpha=2,$ there are two $a$-cycles $C_0$ and $C_1$ and there are two parallel matchings between $C_0$ and $C_1.$ We denote one of the matchings from $C_0$ to $C_1$ as $M_0$ and the other matching from $C_1$ to $C_0$ as $M_1;$ let the edges of $M_0$ be $(0,\,j)(1,\,j),\,0\le j\le \frac{n}{2}-1,$ and let the edges of $M_1$ be $(1,\,j)(0,\,j+c),\,0\le j\le \frac{n}{2}-1,$ where addition is taken modulo $n.$ We call the vertex $(i,\,j)\,(=ib+ja)$ of $C_0$ as the {\it corresponding vertex} of $(i+1,\,j)\,(=(i+1)b+ja)$ of $C_1$ and vice versa. We prove the existence of the property $Q_2$ with respect to the Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma$ by the parity of $c.$
\noindent{\bf Case 1.} $n\equiv 2\,(mod\,\,4).$
First, we shall obtain an even length path along $H_1,$ so that this path is a section of a suitable odd length path along $H_1$ having its ends as two \lq\lq opposite vertices" of an odd alternating $4$-cycle in $\Gamma;$ the existence of the odd alternating $4$-cycle will be proved later.
In ${\bf (A)}$ below, we obtain a required even length path along $H_1$ and in ${\bf (B)}$ we prove that the path obtained in ${\bf (A)}$ is a section of a path along $H_1$ joining two \lq\lq opposite vertices" of an odd alternating $4$-cycle.
\noindent{\bf (A).} First we consider $c$ to be even and consider the Hamilton cycle $H_1$ of $\Gamma$ obtained above. Since $M_1$ is a perfect matching of $\Gamma$ and $M_1$ matches the vertices in $C_1$ with vertices in $C_0,$ $(-b)0$ is an edge of $M_1,$ which is one of the two $M_1$ edges in $H_1,$ see Figure 3(a). Clearly, $-b\equiv (\frac{n}{2}-c)a+b\,(mod\,\,n),$ that is, $-b$ is the $(\frac{n}{2}-c+1)$th vertex of $C_1,$ starting from $b,$ see Figure 3(a). As $\frac{n}{2}$ is odd, $(\frac{n}{2}-c+1)$ is even and hence the length of the section of $C_1$ from $b$ to $-b$ (containing the vertex $b+a$) is odd.
\begin{center}
\includegraphics[scale=0.8]{fig3.eps}
{\small $H_1$ is shown in both Figures 1(a) and 1(b), wherein the bold lines are the required even length path.}\\Figure 3
\end{center}
As the length of the section of $C_1$ from $b$ to $-b$ (containing the vertex $(b+a)$) is the same as the length of the section of $C_0$ from $a$ to $(-2b+a)$ (containing the vertex $2a),$ the length of the path from $a$ to $-2b+a$ along $H_1,$ denoted by $p_{H_1}(a,\,-2b+a)$ is odd and let this path be $R.$ Now $R$ together with the edge $a(-b+a)$ is our required even length path $R_0,$ see the bold edges of Figure 3(a).
Next we suppose that $c$ is odd. As in the above paragraph, the vertex $0$ in $C_0$ is matched under $M_1$ with the vertex $-b\equiv (\frac{n}{2}-c)a+b\,(mod\,\,n);$ $\frac{n}{2}-c$ is even as $\frac{n}{2}$ is odd. Thus the length of the section of $C_1$ from $b$ to $-b$ (containing the vertex $(b+a)$) is even and hence the length of the section $R_1$ of $C_1$ from $b$ to $(-b+a)$ (containing the vertex $(-a+b)$) is even (note that $(-b)(-b+a)$ is an edge of $C_1$ which we have deleted for the construction of $H_1$). Now $R_1$ together with the edges $\{(b+a)b,\,(-b+a)a\}$ is our required even length path $R_0$ from the vertex $a$ to the vertex $b+a$ along $H_1,$ see the bold edges of Figure 3(b).
\noindent ${\bf (B).}$ $H_1$ contains exactly two edges of $M_1$ as shown in the Figure 3. Hence except the two $4$-cycles of $\Gamma,$ namely, $(0,\,a,\,a+b,\,b)$ and $(-2b,\,-2b+a,\,-b+a,\,-b),$ the other $4$-cycles, which have two consecutive vertices of $C_0$ and the corresponding two vertices of $C_1$ constitute alternating $4$-cycles (with respect to $H_1$ and $H_2$ of $\Gamma$).
If $c$ is even, it is clear that every $4$-cycle containing two consecutive vertices of the section $(-2b+a)(-2b+2a)(-2b+3a)\ldots((\frac{n}{2}-1)a)0$ of $C_0$ and the two corresponding consecutive vertices of the section $(-b+a)(-b+2a)(-b+3a)\ldots(b-a)b$ of $C_1$ is an alternating $4$-cycle of $\Gamma$ with respect to $H_1$ and $H_2,$ see the Figure 3 (note that $a=2b$ is ruled out, for otherwise, it would imply $c=1$ for the following reason: $a=2b$ implies $b=-b+a$ and since $(-b+a)a$ is an edge of $H_1$ joining $C_1$ and $C_0$ implies $c=1,$ which is not the case as we consider $c$ is even). Similarly if $c$ is odd, every $4$-cycle containing two consecutive vertices in the section $a\,(2a)\,(3a)\,\ldots\,(-2b)$ of $C_0$ and the two corresponding consecutive vertices in the section $(b+a)\,(b+2a)\,(b+3a)\,\ldots\,(-b)$ of $C_1$ is an alternating $4$-cycle of $\Gamma$ with respect to $H_1$ and $H_2;$ note if $a=-2b,$ then $a=-ca$ so $-a=ca=(\frac{n}{2}-1)a$ but then $c\equiv (\frac{n}{2}-1)\,(mod\,\,\frac{n}{2}),$ that is, $c=\frac{n}{2}-1,$ which is even, a contradiction.
Next we prove that the alternating $4$-cycles described in the above paragraph using the sections of $C_0$ and $C_1$ satisfy the property that the length of the path along $H_1$ between the opposite vertices of the $4$-cycle is of odd length and further this odd length path contains $R_0$ (described above).
Let $C=(x,\,x+a,\,x+a+b,\,x+b)$ be an alternating $4$-cycle so that $x,\,x+a\in V(C_0)$ and $x+a+b,\,x+b\in V(C_1),$ see Figure 4. The vertices $x$ and $x+b$ are the corresponding vertices in $C_0$ and $C_1.$ If $c$ is even, then the $(-2b+a,\,x)$-section of $H_1,$ contained in $C_0,$ and the corresponding $(x+b,\,(-b+a))$-section of $H_1$ contained in $C_1$ have the same length, see Figure 4(a). Hence the $(x,\,x+a+b)$-section of $H_1$ (containing the vertex $-2b$) is of odd length as it contains $R_0,$ which is of even length, see Figure 4(a); that is $p_{H_1}(x,\,x+a+b)$ is odd when $c$ is even. Similarly, if $c$ is odd then, the $(a,\,x)$-section of $H_1$ (containing the vertex $2a$), contained in $C_0,$ and the corresponding $(b+a,\,x+b)$-section of $H_1,$ contained in $C_1$ (containing the vertex $b+2a),$ have the same length. Thus the $(x,\,x+a+b)$-section of $H_1$ (containing the vertex $-b+a)$ is of odd length as it contains $R_0,$ which is of even length, see Figure 4(b). That is $p_{H_1}(x,\,x+a+b)$ is odd when $c$ is odd, see Figure 4(b).
\begin{center}
\includegraphics[scale=0.8]{fig4.eps}
\end{center}
{\small (a) The lengths of the sections $(-2b+a,\,x)$ along $H_1$ (section of $C_0$) and $(-b+a,\,x+b)$ along $H_1$ (section of $C_1$) are same and $(x+b)(x+a+b)$ is an edge \\(b) the lengths of the sections $(a,\,x)$ along $H_1$ (part of $C_0$) and $(b+a,\,x+b)$ along $H_1$ (part of $C_1$) are same and $(x+a)(x+a+b)$ is an edge.}\begin{center}Figure 4
\end{center}
Next we shall obtain an appropriate $x$ on $C_0$ so that $p_{H_2}(x,\,x+a+b)$ is also odd which gives the property $Q_2$ with respect to $H_1$ and $H_2.$ As $H_2$ contains exactly two $a$-edges, namely, $(-b)(-b+a)$ and $0a$ (see Figure 2(c2)), $H_2-\{0a,\,(-b)(-b+a)\}$ is a pair of odd length paths (since the vertices of these paths alternate between $V(C_0)$ and $V(C_1)$). Let $S_0=0\,(b)\,(2b)\,\ldots\,(-b+a)$ and $S_1=a\,(a+b)\,(a+2b)\,\ldots\,(-b)$ denote these paths, namely, the $(0,\,(-b+a))$-section and $(-b,\,a)$-section of $H_2$, respectively, see Figures 2(c1) and 2(c2).
If $S_0$ (respectively $S_1$) contains a pair of consecutive vertices $x$ and $x+a$ of $C_0$ (note that the edge $x(x+a)$ is in $H_1$ and $S_0$ is contained in $H_2$), then $p_{H_2}(x,\,x+a+b)$ is odd as the vertices of $S_0$ (respectively $S_1$) alternate between $V(C_0)$ and $V(C_1)$ and $(x+a)(x+a+b)$ is an edge in $H_2.$ We shall prove the existence of such a pair of consecutive vertices in the sequence of vertices $L_0=(-2b+a)\,(-2b+2a)\,(-2b+3a)\,\ldots\,(\frac{n}{2}-1)a\,0,$ or in $L_1=a\,(2a)\,(3a)\,\ldots\,(-2b),$ see Figure 4.
Let $c$ be even. From the construction of $H_1$ and $H_2,$ the vertices $0$ and $-2b+a$ are in $S_0,$ (see Figure 2(c1)). There are even number of vertices in $L_0$ as $R_0$ is of even length and $C_0$ has odd number of vertices (see Figure 3(a)). If there is no pair of consecutive vertices in $L_0$ of the required type in $S_0$ or $S_1,$ then the preceding vertex of $0,$ namely, $(\frac{n}{2}-1)a\,(=-a),$ and the succeeding vertex of $-2b+a,$ namely, $-2b+2a,$ are in $S_1,$ as the vertices $0$ and $-2b+a$ are in $S_0;$ consequently, there must be an odd number of vertices from $-2b+2a$ to $(\frac{n}{2}-1)a(=-a)$ in $L_0,$ which is not the case (since in $L_0,$ $-2b+2a$ to $-a$ contains even number of vertices). Therefore there must exist a pair of consecutive vertices in $L_0$ of the required type and hence $\Gamma$ satisfies property $Q_2$ with respect to $H_1$ and $H_2.$
Next we assume that $c$ is odd. As above, we shall show the existence of the required pair of consecutive vertices in $L_1=a\,(2a)\,(3a)\,\ldots\,(-2b).$ $L_1$ has even number of vertices (see Figure 3(b)). Assume that there is no pair of vertices in $L_1$ of the required type in $S_0$ or $S_1$ and hence alternate vertices of $L_1$ are in $S_0$ and $S_1.$ Then, as the vertices $a$ and $-2b$ are in $S_1,$ the vertices $2a$ and $-2b-a$ must be in $S_0$ and hence there must be an odd number of vertices from $2a$ to $-2b-a$ in $L_1,$ which is not the case (as $L_1$ contains even number of vertices from $2a$ to $-2b-a$). Thus there must exist a pair of consecutive vertices as required and hence $\Gamma$ satisfies the property $Q_2.$
This completes the proof when $n\equiv 2\,(mod\,\,4).$
\noindent Case 2: $n\equiv 0\,(mod\,\,4).$
Since $n\equiv 0\,(mod\,\,4)$ and $\Gamma\in \Gamma(2,\,1),$ $c$ is always odd, otherwise, the $b$-edges will not induce a Hamilton cycle. As $\alpha=2$ and $n\equiv 0\,(mod\,\,4),$ $a\equiv 2\,(mod\,\,4)$ as $gcd(a,\,n)=\alpha=2.$ By assumption $b$ is odd and hence $b\equiv 1\,or\,3\,(mod\,\,4);$ then $-b\equiv 3\,or \,1\,(mod\,\,4).$ For any two vertices $x$ and $y$ on $C_1,$ the path from $x$ to $y$ along $C_1$ has even length if and only if $x\equiv y\,(mod\,\,4).$ Hence the length of the section of $C_1$ from $b$ to $(-b)$ (containing the vertex $(b+a))$ is odd and the length of the section of $C_1$ from the vertex $(-b+a)$ to $b$ (containing the vertex $(b-a))$ is even and let this path be $R,$ (see Figure $4(b);$ note that the Figure $4(b)$ is for the case $n\equiv 2\,(mod\,\,4);$ the figure for the case $n\equiv 0\,(mod\,\,4)$ is similar). Now the edges of $R$ together with the edges $\{(b+a)b,\,(-b+a)a\}$ induce an even length path, say $R_1,$ from the vertex $b+a$ to $a.$ Again, the length of the section of $C_0$ from $a$ to $-2b+a$ (containing the vertex $2a$) is odd since it is of same length as the section of $C_1$ from $b$ to $-b$ (containing the vertex $(b+a))$. Now the edges of the section of $C_0$ from $a$ to $-2b+a$ together with the edge $a(-b+a)$ induce an even length path, say $R_2.$
As in Case 1, except the two $4$-cycles, namely, $(0,\,a,\,a+b,\,b)$ and $(-2b,\,-2b+a,\,-b+a,\,-b)$ in $\Gamma,$ each of the $4$-cycles formed by two consecutive vertices of $C_0$ and their corresponding vertices in $C_1$ constitute an alternating $4$-cycle with respect to $H_1$ and $H_2.$ Clearly, one of the two paths along $H_1,$ joining a pair of opposite vertices of these alternating $4$-cycles contains exactly one of the even length paths $R_1$ or $R_2$ and hence its length along $H_1$ is odd; hence if $(x,\,x+a,\,x+a+b,\,x+b)$ is an alternating $4$-cycle with $x,\,x+a\in V(C_0)$ and $x+a+b,\,x+b\in V(C_1),$ then $p_{H_1}(x,\,x+a+b)$ is odd.
We now show that in at least one of these alternating $4$-cycles, described above, $p_{H_2}(x,\,x+a+b)$ is odd. Let $S_0$ and $S_1$ be as defined in Case 1. Suppose the vertices of $C_0$ are alternately in $S_0$ and $S_1,$ then $c$ is even, but it is not the case. Therefore, there exist two consecutive vertices along $C_0$ which are in $S_0$ or $S_1.$ Thus there exists a pair of opposite vertices of an alternating $4$-cycle $(x,\,x+a,\,x+a+b,\,x+b),$ where $x\in V(C_0)$ and $x+a+b\in V(C_1)$ such that $p_{H_2}(x,\,x+a+b)$ is odd and hence $(x,\,x+a,\,x+a+b,\,x+b)$ is an odd alternating $4$-cycle. Thus $\Gamma$ satisfies property $Q_2,$ with respect to $H_1$ and $H_2.$
This completes the proof of the lemma.\hfill$\Box$
We use the following remarks in the proof of Lemma \ref{LEM1.directcirculant} given below.
\begin{lem}\label{lemma1ofbermond}\emph{\cite{JCTB.46.142-153}}
Let $\Gamma$ be a graph of $\Gamma(\alpha + 2,\beta).$ If the reduced graph $\Gamma^\prime$
admits a hamiltonian decomposition having the property $Q_1$ between $C_{\alpha-1}$ and
$C_0,$ then $\Gamma$ admits a Hamilton cycle decomposition having the property $Q_1$
between $C_{\alpha+1}$ and $C_0.$\hfill$\Box$
\end{lem}
\begin{rmk}\label{DC.remark1}
In the construction of two edge disjoint Hamilton cycles $H_1$ and $H_2$ of $\Gamma\in\Gamma(\alpha+2,\,\beta)$ from the two edge disjoint Hamilton cycles $H_1^\prime$ and $H_2^\prime,$ respectively, of $\Gamma^\prime\in\Gamma(\alpha,\,\beta),$ every $b$-edge of $H_i^\prime,\,i=1,\,2,$ connecting the vertices of $C_{\alpha-1}^\prime$ to $C_0^\prime$ is replaced by a path of odd length, see Lemma \ref{lemma1ofbermond}. Hence the parity of length of the path between any pair of vertices in $C_0,\,C_1,\,\ldots,\,C_{\alpha-1}$ along $H_i$ remains the same as the parity of length of the path between the respective vertices in $C_0^\prime,\,C_1^\prime,\,\ldots,\,C_{\alpha-1}^\prime$ of $H_i^\prime,$ where $C_i^\prime$ are the $a$-cycles of $\Gamma^\prime.$\hfill$\Box$
\end{rmk}
By $\overrightarrow{(a,\,b)}$ we denote a directed arc with tail at $a$ and head at $b.$
\begin{rmk}\label{DC.remark2}
Consider the graph $\Gamma\in\Gamma(3,\,2)$ and its reduced graph $\Gamma^\prime\in\Gamma(1,\,2).$ Let $\{H_1^\prime,\,H_2^\prime\}$ be the Hamilton cycle decomposition of $\Gamma^\prime$ guaranteed by Lemma \ref{LEM2.directcirculant}. Clearly, $H_1^\prime$ has two natural orientations one in the clockwise direction and the other in the anticlockwise direction and these two orientations induce orientations for the two $b$-edges of $H_1^\prime,$ namely, $0(-b)$ and $a(-b+a).$ We first fix the clockwise orientation of the Hamilton cycle $H_1^\prime$ of $\Gamma^\prime,$ and let the orientation of the edges $0(-b)$ and $a(-b+a)$ of $H_1^\prime$ give two arcs $\overrightarrow{(-b,\,0)}$ and $\overrightarrow{(-b+a,\,a)},$ see Figure 5. With respect to this orientation, we obtain a Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma,$ using the proof of Lemma \ref{lemma1ofbermond}, with property $Q_1,$ where the arcs $\overrightarrow{(-b,\,0)}$ and $\overrightarrow{(-b+a,\,a)}$ correpond to the $(m_1=2)$ edges $\{(\alpha-1,\,j_0)(0,\,j_0+c)\}$ and $\{(\alpha-1,\,j_1)(0,\,j_1+c)\},$ respectively, in the proof of Lemma \ref{lemma1ofbermond}.
\begin{center}
\includegraphics{fig5.eps}
\end{center}
{\small The Hamilton cycles $H_1$ and $H_2$ shown in Figure 5(a) arise out of the Hamilton cycles $H_1^\prime$ and $H_2^\prime$ with respect to the clockwise orientation of $H_1^\prime$ and, the Hamilton cycle decomposition $\{H^1,\,H^2\}$ shown in Figure 5(b) arises out of the Hamilton cycles $H_1^\prime$ and $H_2^\prime$ with respect to the anticlockwise orientation of $H_1^\prime,$ where $|\Gamma^\prime|=n=10,\,a=3,\,b=4.$}\\\vspace{-0.7cm}\begin{center}Figure 5\end{center}
If we fix the anticlockwise orientation for $H_1^\prime,$ we obtain another orientation of the edges $0(-b)$ and $a(-b+a)$ of $H_1^\prime$ and the resulting arcs are $\overrightarrow{(0,\,-b)}$ and $\overrightarrow{(a,\,-b+a)},$ see Figure $5(b).$ With respect to this orientation, we obtain a Hamilton cycle decomposition $\{H^1,\,H^2\}$ of $\Gamma,$ by proof of Lemma \ref{lemma1ofbermond}, with the property $Q_1,$ where the arcs $\overrightarrow{(0,\,-b)}$ and $\overrightarrow{(a,\,-b+a)}$ correspond to the $(m_1=2)$ edges $\{(\alpha-1,\,j_0)(0,\,j_0+c)\}$ and $\{(\alpha-1,\,j_1)(0,\,j_1+c)\},$ respectively, in the proof of Lemma \ref{lemma1ofbermond}, see Figure 5. Irrespective of the orientations of $H_1^\prime$ (that is, clockwise or anticlockwise), $H_1\cup H_2\cong H^1\cup H^2,$ that is, the two circulant graphs obtained from the union of the Hamilton cycles $H_1\cup H_2$ and $H^1\cup H^2$ in two different orientations of the edges of $H_1^\prime$ are isomorphic to each other. This isomorphism can be described by mapping $(i,\,j)$ to $(i,\,n-j),\,0\le i\ne \alpha-1,\,1\le j\le k_a-1,$ where the even integer $n$ is the number of vertices of $\Gamma^\prime,$ and the vertices $(i,\,0),0\le i\le \alpha-1$, are the fixed vertices of the isomorphism.\hfill$\Box$
\end{rmk}
\begin{lem}\label{LEM1.directcirculant}
Let $\Gamma\in\Gamma(\alpha+2,\,\beta)$ be a graph with even number of vertices and having the generating set $\{s,\,t\}$ where $s$ and $t$ are of different parity. If the reduced graph $\Gamma^\prime\in\Gamma(\alpha,\,\beta),$ of $\Gamma,$ has the properties $Q_1$ and $Q_2$ with respect to a Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime,$ then $\Gamma$ has a Hamilton cycle decomposition $\{H_1,\,H_2\}$ with the properties $Q_1$ and $Q_2.$
\end{lem}
\noindent{\bf Proof.} Throughout the proof we use $\{s,\,t\}$ and $\{a,\,b\}$ as the generating sets for $\Gamma\in\Gamma(\alpha+2,\,\beta)$ and $\Gamma^\prime\in\Gamma(\alpha,\,\beta),$ respectively. Without loss of generality assume that throughout the proof $s$ is even and $t$ is odd. As $\Gamma$ is of even order and as $s$ is even and $t$ is odd, $\alpha$ is even and $\beta$ is odd, since $\alpha+2=gcd(m,\,s)$ and $\beta=gcd(m,\,t),$ where $m$ is the number of vertices of $\Gamma.$ Let $\{H_1^\prime,\,H_2^\prime\}$ be a Hamilton cycle decomposition of $\Gamma^\prime$ and $\{H_1,\,H_2\}$ be the corresponding Hamilton cycle decomposition of $\Gamma.$
In the construction of $H_1$ and $H_2$ from $H_1^\prime$ and $H_2^\prime$ (see \cite{JCTB.46.142-153}), $H_i,\,i=1,\,2,$ is obtained from $H_i^\prime$ by replacing each of the edges of $H_i^\prime$ between $C_{\alpha-1}^\prime$ to $C_0^\prime$ by an odd length path, whose origin is in $C_{\alpha-1}$ and terminus is in $C_0$ of $\Gamma$ and the internal vertices of the paths are in $C_\alpha$ and $C_{\alpha+1}.$
As $\{H_1^\prime,\,H_2^\prime\}$ has the property $Q_1,$ there exists a Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma\in\Gamma(\alpha+2,\,\beta)$ satisfying property $Q_1$ by Lemma \ref{lemma1ofbermond}. Next we prove that $\Gamma$ satisfies the property $Q_2$ with respect to $\{H_1,\,H_2\}.$ We prove this by induction on $\alpha.$
First we explain the idea behind the proof of this Lemma. In Claim 1 below, we prove the existence of a Hamilton cycle decomposition of $\Gamma\in\Gamma(3,\,2)$ satisfying $Q_2$ from a Hamilton cycle decomposition of $\Gamma^\prime\in\Gamma(1,\,2)$ with property $Q_2$ and in Claim 2 below, we prove the existence of a Hamilton cycle decomposition of $\Gamma\in\Gamma(4,\,1)$ satisfying $Q_2$ from a Hamilton cycle decomposition of $\Gamma^\prime\in\Gamma(2,\,1)$ with property $Q_2.$ In Claim 3 below, we obtain a Hamilton cycle decomposition of $\Gamma\in\Gamma(\alpha+2,\,\beta)$ with property $Q_2$ from a Hamilton cycle decomposition of $\Gamma^\prime\in\Gamma(\alpha,\,\beta)$ with property $Q_2.$ First we assume that $\{H_1^\prime,\,H_2^\prime\}$ is the Hamilton cycle decomposition of $\Gamma^\prime\in\Gamma(\alpha,\,\beta),$ where $\{\alpha,\,\beta\}=\{1,\,2\},$ as described in Lemma \ref{LEM2.directcirculant}.
\noindent{\bf Claim 1.} For a Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime\in\Gamma(1,\,2)$ with property $Q_2,$ there is a Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma\in\Gamma(3,\,2)$ with property $Q_2.$
Let $|V(\Gamma^\prime)|=n;$ by hypothesis $\Gamma^\prime$ admits a Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ with the property $Q_2.$ We consider two cases.
\noindent{\bf Case 1.} $n\equiv 0\,(mod\,\,4).$
Let $\Gamma^\prime\in\Gamma(1,\,2)$ and let its corresponding graph in $\Gamma(3,\,2)$ be $\Gamma.$ From the definition of $\Gamma(\alpha,\,\beta),$ the labels of the vertices of the $i$th $a$-cycle are $(i,\,j),\,0\le i\le \alpha-1,\,0\le j\le k_a-1,$ where $k_a$ is the length of the $i$th $a$-cycle. We know that $\Gamma^\prime\in\Gamma(1,\,2)$ has only one $a$-cycle and hence its vertices are $(0,\,j),\,j=0,\,1,\,\ldots,\,n-1.$ Each vertex $(0,\,j)$ of $\Gamma^\prime$ gives rise to three vertices, namely, $(0,\,j),\,(1,\,j)$ and $(2,\,j),$ in $\Gamma$ and we call these three vertices of $\Gamma$ as the corresponding vertices of $(0,\,j)$ of $\Gamma^\prime$ and vice versa.
\begin{center}
\includegraphics[scale=0.9]{fig6.eps}
\end{center}
{\small $(a)$ Labeling of the vertices of the circulant graph $\Gamma^\prime\in\Gamma(1,\,2)$ on $10$ vertices with the group elements, where the generating set is $\{3,\,4\}\subset \mathbb{Z}_{10}$ $(b).$ Labeling of the vertices of the graph $\Gamma^\prime$ with ordered pairs as described in \cite{JCTB.46.142-153}. $(c).$ Labeling of the vertices of the graph $\Gamma^\prime,$ with ${\bf j}$ if the vertex has the label $(0,\,j)$ in $(b).$}\vspace{-.4cm}\begin{center}Figure 6
\end{center}
For our convenience we relabel the vertices $(0,\,j),\,0\le j\le n-1,$ of $\Gamma^\prime\in\Gamma(1,\,2)$ as ${\bf j},$ see Figure 6(c); we call it as {\it new labelling of $\Gamma^\prime$}(throughout this lemma, the new labelling of $\Gamma^\prime$ is denoted by bold face letters); in fact, in the proof of this lemma each vertex of $\Gamma^\prime$ will have three different labels, namely, $(0,\,j),\,{\bf j}$ and the other one the group element and according to our convenience and circumstances we use one of these labels. But, for the vertices of $\Gamma\in\Gamma(3,\,2)$ we use the unique label $(i,\,j)$ as in \cite{JCTB.46.142-153}. In the graph $\Gamma^\prime,$ let the vertex $-b$ be denoted by the label $(0,\,r)$ and hence in our new labeling it is denoted by ${\bf r},$ see Figure 6; for our convenience we write $-b={\bf r}.$ Clearly, ${\bf r+1}\,(=-b+a)$ is the immediate next vertex of ${\bf r}\,(=-b)$ in $C_0^\prime$ of $\Gamma^\prime$ in the new labeling. Corresponding to the vertices ${\bf r}$ and ${\bf r+1}$ of $\Gamma^\prime,$ there are two rows, each having three vertices, in $\Gamma;$ the vertices of these rows are denoted by, $(0,\,r),\,(1,\,r),\,(2,r)$ and $(0,\,r+1),\,(1,\,r+1),\,(2,r+1),$ respectively, see Figure 7.
\begin{center}
\includegraphics{fig7.eps}
\end{center}
{\small Construction of the Hamilton cycle $H_1$ of $\Gamma\in\Gamma(3,\,2)$ from the Hamilton cycle $H_1^\prime$ of $\Gamma^\prime\in\Gamma(1,\,2),$ as described in \cite{JCTB.46.142-153}.}\begin{center}
Figure 7
\end{center}
In the construction of $\Gamma\in\Gamma(3,\,2)$ from $\Gamma^\prime\in\Gamma(1,\,2),$ the two edges $(0,\,r)(1,\,r)$ and $(0,\,r+1)(1,\,r+1)$ are in $H_1$ and the two edges $(0,\,r)(0,\,r+1)$ and $(1,\,r)(1,\,r+1)$ are in $H_2,$ where $\{H_1,\,H_2\}$ is the Hamilton cycle decomposition of $\Gamma,$ corresponding to the Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime,$ as described in \cite{JCTB.46.142-153}, see Figures 7 and 9. We claim that $\Gamma$ satisfies the property $Q_2$ with respect to the Hamilton cycle decomposition $\{H_1,\,H_2\},$ where the required $4$-cycle of $\Gamma$ is $(0,\,r)\,(0,\,r+1)\,(1,\,r+1)\,(1,\,r).$
First we show that $p_{H_1}((0,\,r),\,(1,\,r+1))$ is odd. Clearly, $C_0^\prime=({\bf 0},\,{\bf 1},\,{\bf 2},\,\ldots,\,{\bf n-1},\,{\bf 0})$ is a Hamilton cycle in $\Gamma^\prime,$ with respect to our new labeling.
The Hamilton cycle $H_1$ of $\Gamma$ contains the four vertices $(0,\,r),\,(0,\,0),\,(0,\,r+1)$ and $(1,\,r+1)$ in the clockwise order as shown in the Figure 7 ( the order is guaranteed by the corresponding Hamilton cycle $H_1^\prime$ of $\Gamma^\prime$), where $(0,\,r+1)(1,\,r+1)$ is an edge of $H_1.$ As $C_0^\prime$ being an even cycle, it can be thought of as a bipartite graph with bipartition $X=\{{\bf 0},\,{\bf 2},\,{\bf 4},\,\ldots,\,{\bf n-2}\}$ and $Y=\{{\bf 1},\,{\bf 3},\,\ldots,\,{\bf n-1}\}.$ Clearly, the vertex $-b\,\big(={\bf r},$ the $({\bf r+1})$th vertex along $C_0^\prime,$ the group element $ra\,(mod\,\,n);$ we do not differentiate the labels $-b,\,{\bf r}$ and $ra$ and we denote the vertex $-b$ by $-b={\bf r}=ra$ in the three labellings of the vertices of $\Gamma^\prime\big)$ is in $X,$ because $b$ is even implies $-b$ is even; as $-b=ra$ and $a$ is odd, ${\bf r}$ is even. Further, $0\in X$ and $-b\in X$ implies $p_{C_0^\prime}({\bf 0},\,{\bf r})=p_{C_0^\prime}(0,\,-b)$ is even. Now consider the Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime$ as in Lemma \ref{LEM2.directcirculant}. The Hamilton cycle $H_1^\prime$ of $\Gamma^\prime$ is obtained from $C_0^\prime$ by deleting two edges $0a\,(={\bf 01})$ and $(-b)(-b+a)(={\bf r(r+1)})$ of $C_0^\prime$ and adding two edges $0(-b)(={\bf 0r})$ and $a(-b+a)(={\bf 1}({\bf r+1})),$ see Figure 8.
\begin{center}
\includegraphics[scale=0.9]{fig8.eps}
\end{center}
{\small Length of the $({\bf 1},\,{\bf r})$-section along $H_1^\prime$ is odd, as ${\bf r}$ (an even integer) is the length from ${\bf 0}$ to ${\bf r}$ along $C_0^\prime;$ this implies that the $({\bf 0},\,{\bf r+1})$-section along $H_1^\prime$ is of odd length.}
\begin{center}Figure 8
\end{center}
Next we prove that $p_{H_1^\prime}({\bf 0},\,{\bf r+1})\,(=p_{H_1^\prime}(0,\,-b+a))$ is odd. From the last paragraph, $p_{C_0^\prime}({\bf 0},\,{\bf r})$ is even and hence $p_{C_0^\prime}({\bf 1},\,{\bf r})$ is odd, see Figure 8; the path from ${\bf 1}$ to ${\bf r}$ along $C_0^\prime$ is also in $H_1^\prime$ and hence $p_{H_1^\prime}({\bf 1},\,{\bf r})$ is odd; consequently, as $n$ is even,\begin{eqnarray} p_{H_1^\prime}({\bf 0},\,{\bf r+1})\,is\,odd,\,\label{directcirculant.equation1}\end{eqnarray} (where we consider the $({\bf 0},\,{\bf r+1})$-section of $H_1^\prime$ not containing the vertex ${\bf r}$), see Figure 8.
Next we prove that $p_{H_1}((0,\,r),\,(1,\,r+1))$ is odd. We divide the $((0,\,r),\,(1,\,r+1))$-section of $H_1$ (containing the vertex $(1,\,r))$ into three subsections and we show that each one of them is of odd length to conclude $p_{H_1}((0,\,r),\,(1,\,r+1))$ is odd; the subsections are $((0,\,r),\,(0,\,0)),$ $((0,\,0),\,(0,\,r+1))$ and $((0,\,r+1),\,(1,\,r+1)),$ see Figure 7. From the construction of $H_1$ from $H_1^\prime,$ it is clear that as $((0,\,r),\,(0,\,0))$-section and $((0,\,r+1),\,(1,\,r+1))$-section of $H_1$ are of length 3 and length 1, respectively, it is enough to show that $((0,\,0),\,(0,\,r+1))$-subsection is of odd length. But it is easy to see that $((0,\,0),\,(0,\,r+1))$-section of $H_1$ is identical with the $({\bf 0},\,{\bf r+1})$-section of the Hamilton cycle $H_1^\prime,$ see Figure 7. The $({\bf 0},\,{\bf r+1})$-section of $H_1^\prime$ is already proved to be of odd length, by (\ref{directcirculant.equation1}). Thus $p_{H_1}((0,\,0),\,(0,\,r+1))$ is odd.
Next we prove that $p_{H_2}((0,\,r),\,(1,\,r+1))$ is odd. As above, we divide the $((0,\,r),\,(1,\,r+1))$-section of $H_2$ (not containing the vertex $(0,\,r+1)$) into four subsections, namely, $((0,\,r),\,(0,\,0)),$\,((0,\,0),\,(0,\,1)),$\,((0,\,1),\,(0,\,r-1))$ and $((0,\,r-1),\,(1,\,r+1))$-section in the cyclic order are guaranteed by $H_2^\prime,$ see \cite{JCTB.46.142-153} and Figure 9. From the construction of $H_2,$ $(0,\,0)(0,\,1)$ is an edge and the $((0,\,r-1),\,(1,\,r+1))$-section is a path of length 3, namely, $(0,\,r-1)\,(1,\,r-1)\,(1,\,r)\,(1,\,r+1),$ see Figure 9. Hence we show that the lengths of the other two sections are of different parity. This is achieved by finding the lengths of the corresponding sections in $H_2^\prime$ of $\Gamma^\prime.$ The sections in $H_2^\prime$ corresponding to the $((0,\,r),\,(0,\,0))$-section and $((0,\,1),\,(0,\,r-1))$-section of $H_2$ in $\Gamma$ are $({\bf r},\,{\bf 0})$-section and $({\bf 1},\,{\bf r-1})$-section, respectively, in the new labeling of $\Gamma^\prime.$
\begin{center}
\includegraphics[scale=0.8]{fig9.eps}
{\small $H_2^\prime$ in $\Gamma^\prime$ and $H_2$ in $\Gamma$}\end{center}
\vspace{-.5cm}
\begin{center}Figure 9
\end{center}
We shall show that $p_{H_2^\prime}({\bf r},\,{\bf 0})$ is odd and $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is even. We prove that $p_{H_2^\prime}({\bf r},\,{\bf 0})$ is odd by observing that the $({\bf 0},\,{\bf r})$-section (not containing the edge ${\bf 0r}$) of one of the $b$-cycles, namely $0\,b\,(2b)\,\ldots\,(-b)\,0\,\,\big(=({\bf 0}\,\ldots {\bf r}\,{\bf 0})\text{ in the new labeling,}$ not containing the edge ${\bf 0r}\big)$ of $\Gamma^\prime$ is a section of $H_2^\prime,$ and it is of odd length as the $({\bf 0},\,{\bf r})$-section of $H_2^\prime$ together with the edge ${\bf 0r}$ is a $b$-cycle, which is of even length in $\Gamma^\prime,$ that is, $p_{H_2^\prime}({\bf r},\,{\bf 0})$ is odd.
\begin{center}
\includegraphics{fig10.eps}
{\small A redrawing of $H_2^\prime$}\end{center}
\begin{center}Figure 10\end{center}
Next we prove that $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is even. Now consider the $({\bf 1},\,{\bf r-1})$-section of $H_2^\prime$ containing only $b$-edges. As we move along $b$-edges of the $({\bf 1},\,{\bf r-1})$-section of $H_2^\prime,$ the alternate labels of the vertices are congruent to $1\,(mod\,\,4)$ or $3\,(mod\,\,4).$ If two nonconsecutive vertices of $H_2^\prime$ are both congruent to $1\text{ or }3\,(mod\,\,4),$ then their distance along $H_2^\prime$ is even. Therefore, as $a\equiv 1\text{ or }3\,(mod\,\,4)$ implies $-b-a\equiv 1\text{ or }3\,(mod\,\,4),$ $p_{H_2^\prime}({\bf 1},{\bf r-1})$ is even.
As observed earlier, the length of the $((0,\,r),\,(1,\,r+1))$-section of $H_2$ is sum of the lengths of the sections $((0,\,r),\,(0,\,0)),$\,$((0,\,0),\,(0,\,1)),$\,$((0,\,1),\,(0,\,r-1))$ and a path of length $3$ from $(0,\,r-1)$ to $(1,\,r+1)$ in $H_2$ of $\Gamma.$ As $p_{H_2^\prime}({\bf r},\,{\bf 0})$ is odd, $p_{H_2}((0,\,r),\,(0,\,0))$ is odd by Remark \ref{DC.remark1}, the $((0,\,0),\,(0,\,1))$-section is an edge, the length of $((0,\,1),\,(0,\,r-1))$-section of $H_2$ is even, as $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is even and by Remark \ref{DC.remark1} and the last section is a $P_4=(0,\,r-1)(1,\,r-1)(1,\,r)(1,\,r+1).$ Thus in $\Gamma,$ $p_{H_2}((0,\,r),\,(1,\,r+1))$ is odd.
\noindent{\bf Case 2.} $n\equiv 2\,(mod\,\,4).$
Let $\{H_1^\prime,\,H_2^\prime\}$ be the Hamilton cycle decomposition of $\Gamma^\prime$ with property $Q_2$ obtained in Lemma \ref{LEM2.directcirculant}. As pointed out in Remark \ref{DC.remark2}, the Hamilton cycle $H_1^\prime$ has two natural orientations and with respect to each of these two orientations there is a Hamilton cycle decomposition of $\Gamma\in\Gamma(3,\,2).$ We show that in at least one of these two orientations of $H_1^\prime,$ the corresponding Hamilton cycle decomposition of $\Gamma$ has the property $Q_2$ (Note that both these two Hamilton cycle decompositions satisfy the property $Q_1$ as in Remark \ref{DC.remark2}). Consider the new labeling of $\Gamma^\prime$ described in Case 1 above.
\noindent{\bf Subcase 2.1.} $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is odd.
We know that $H_1^\prime$ has exactly two $b$-edges. In this case we consider the clockwise orientation of $H_1^\prime.$ Consider the $4$-cycle $C=((0,\,r)(0,\,r+1)(1,\,r+1)(1,\,r))$ in $\Gamma,$ see Figure 9 (the figure for the case $n\equiv 2\,(mod\,\,4)$ also resembles the same as in the case $n\equiv 0\,(mod\,\,4)$ and note that ${\bf r}$ is even).
To prove $p_{H_1}((0,\,r),\,(1,\,r+1))$ is odd, we divide the $((0,\,r),\,(1,\,r+1))$-section of $H_1,$ (containing the vertex $(1,\,r)$) into three subsections and we show that each of them is of odd length. The subsections are $((0,\,r),\,(0,\,0)),$ $((0,\,0),\,(0,\,r+1))$ and $((0,\,r+1),\,(1,\,r+1)),$ see Figure 7 (the figure for the case $n\equiv 2\,(mod\,\,4)$ also resembles the same as in the case $n\equiv 0\,(mod\,\,4)$). From the construction of $H_1$ from $H_1^\prime,$ it is clear that $((0,\,r),\,(0,\,0))$-section of $H_1$ is of length 3, see Figure 7, and $(0,\,r+1)(1,\,r+1)$ is an edge, hence it is enough to show that $((0,\,0),\,(0,\,r+1))$-section (containing the vertex $(0,\,n-1)$) is of odd length. But it is easy to see (in fact, it is the same proof as in the case $n\equiv 0\,(mod\,\,4)$) that this section of $H_1$ is the same as the $({\bf 0},\,{\bf r+1})$-section of the Hamilton cycle $H_1^\prime$ of $\Gamma^\prime,$ which is of odd length and hence $p_{H_1}((0,\,0),\,(0,\,r+1))$ is odd. This completes the proof that $p_{H_1}((0,\,r),\,(1,\,r+1))$ is odd.
Now we prove that $p_{H_2}((0,\,r)(1,\,r+1))$ is odd. For that we divide the $((0,\,r),\,(1,\,r+1))$-section of $H_2$ (not containing the vertex $(0,\,r+1)$) into four subsections, namely, $((0,\,r),\,(0,\,0)),\,((0,\,0),\,(0,\,1)),\,((0,\,1),\,(0,\,r-1))$ and $((0,\,r-1),\,(1,\,r+1)),$ see Figure 9. It is clear that the $((0,\,0),\,(0,\,1))$-section is an edge and the $((0,\,r-1),\,(1,\,r+1))$-section is a path of length $3.$ We shall show that the lengths of the subsections $((0,\,r),\,(0,\,0))$ and $((0,\,1),\,(0,\,r-1))$ are of different parity.
First we prove that the length of the subsection $((0,\,r),\,(0,\,0))$ of $H_2$ is even. Clearly, the section corresponding to $((0,\,r),\,(0,\,0))$ of $H_2$ is $({\bf r},\,{\bf 0})$-section in $H_2^\prime$ of $\Gamma^\prime,$ in the new labeling. Observe that the $({\bf 0},\,{\bf r})$-section of one of the $b$-cycles, not containing the edge ${\bf 0r},$ of $\Gamma^\prime$ is a section of $H_2^\prime,$ and it is of even length as the $({\bf 0},\,{\bf r})$-section of $H_2^\prime$ together with the edge ${\bf 0r}$ is a $b$-cycle of odd length in $\Gamma^\prime$ (as $n\equiv 2\,(mod\,\,4)$ and $\beta=2$), that is, $p_{H_2^\prime}({\bf r},\,{\bf 0})$ is even. Hence by Remark \ref{DC.remark1}, $p_{H_2}((0,r),\,(0,\,0))$ is even.
Next consider the $((0,\,1),\,(0,\,r-1))$-section (containing the vertex $(1,\,1)$) of $H_2;$ the corresponding $({\bf 1},\,{\bf r-1})$-section in $H_2^\prime$ contains only $b$-edges. As $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is odd, by assumption, and also the section $({\bf 1},\,{\bf r-1})$ of $H_2^\prime$ containing only $b$-edges, $p_{H_2}((0,\,1),\,(0,\,r-1))$ is odd, by Remark \ref{DC.remark1}. This proves $p_{H_2}((0,\,r),\,(1,\,r+1))$ is odd.
\noindent{\bf Subcase 2.2.} $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is even.
In this subcase, we consider the anticlockwise orientation of the edges of $H_1^\prime.$ Then $C=((0,\,0)(0,\,1)(1,\,1)(1,\,0))$ will be proved to be an odd alternating $4$-cycle, see Figure 11.
First we prove that $p_{H_1}((0,\,0)(1,\,1))$ is odd. For this, we divide the $((0,\,0),\,(1,\,1))$-section of $H_1$ (containing the vertex $(1,\,0)$) into three subsections, namely, $((0,\,0),\,(0,\,r)),\,((0,\,r),\,(0,\,1))$ and $((0,\,1),\,(1,\,1)).$ Clearly, $((0,\,0),\,(0,\,r))$-section is the path $(0,\,0)\,(1,\,0)\,(2,\,0)\,(0,\,r)$ of length 3 and $((0,\,1),\,(1,\,1))$-section is an edge and hence it is enough to prove that $((0,\,r),\,(0,\,1))$-section of $H_1$ is of odd length.
In $H_1^\prime$ of $\Gamma^\prime,$ $b$ is even implies $-b$ is even as $n$ is even. As $-b=ra$ and $a$ is odd implies ${\bf r}$ is even. It is an easy observation that the vertices with even numbered labels in the new labeling of $H_1^\prime$ are at odd distance from the vertex $a={\bf 1},$ the new label of the vertex $a,$ since the even numbered vertices are $2a={\bf 2},\,4a={\bf 4},$ etc. Hence $p_{H_1^\prime}({\bf 1},\,{\bf r})$ is odd as ${\bf r}$ is even. As the $((0,\,r),\,(0,\,1))$-section of $H_1$ has the same length as the $({\bf r},\,{\bf 1})$-section of $H_1^\prime,$ $p_{H_1}((0,\,r),\,(0,\,1))$ is odd.
Finally, we show that $p_{H_2}((0,\,0)(1,\,1))$ is odd. To prove this, we divide the $((0,\,0),\,(1,\,1))$-section of $H_2$ (containing the vertex $(0,\,1)$) into three subsections, namely, $((0,\,0),\,(0,\,1)),\,((0,\,1),\,(0,\,r-1))$ and $((0,\,r-1),\,(1,\,1)),$ see Figure 11. It is clear that the $((0,\,0),\,(0,\,1))$-section is an edge and the $((0,\,r-1),\,(1,\,1))$-section is the path $(0,\,r-1)\,(2,\,n-1)\,(2,\,0)\,(2,\,1)\,(1,\,1)$ on $5$ vertices, see Figure 11. Hence it is enough to show that the $((0,\,1),\,(0,\,r-1))$-section is of even length. As $p_{H_2^\prime}({\bf 1},\,{\bf r-1})$ is even, by assumption in this subcase, $({\bf 1},\,{\bf r-1})$-section of $H_2^\prime,$ containing only $b$-edges, is of even length and hence $p_{H_2}((0,\,1),\,(0,\,r-1))$ is even, by Remark \ref{DC.remark1}. Thus $p_{H_2}((0,\,0),\,(0,\,1))$ is odd. Hence the Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma$ satisfies the property $Q_2.$
\begin{center}
\includegraphics{fig11.eps}
\end{center}
{\small The Hamilton cycles $H_1$ and $H_2$ of $\Gamma\in\Gamma(3,\,2)$ corresponding to the anticlockwise orientation of $H_1^\prime$ of $\Gamma^\prime\in\Gamma(1,\,2).$}\begin{center}Figure 11
\end{center}
\noindent{\bf Claim 2.} For any Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime\in\Gamma(2,\,1)$ with the property $Q_2,$ there is a Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma\in\Gamma(4,\,1)$ with the property $Q_2.$
Consider the graph $\Gamma^\prime\in\Gamma(2,\,1).$ There are two $a$-cycles $C_0^\prime$ and $C_1^\prime;$ hence there are two parallel matchings between $C_0^\prime$ and $C_1^\prime.$ We denote the matching from $C_0^\prime$ to $C_1^\prime$ as $M_0$ and the other matching from $C_1^\prime$ to $C_0^\prime$ as $M_1;$ the edges of $M_0$ are $(0,\,j)\,(1,\,j)$ and the edges of $M_1$ are $(1,\,j)\,(0,\,j+c),$ $0\le j\le \frac{n}{2}-1,$ for some $c.
Clearly, from the construction of $H_1^\prime$ and $H_2^\prime$ of $\Gamma^\prime,$ as in \cite{JCTB.46.142-153}, every alternating $4$-cycle having two consecutive vertices along $C_0^\prime$ and two consecutive vertices along $C_1^\prime$ uses exactly two $a$-edges and two $b$-edges and both these two $b$-edges are in $M_0$ or in $M_1.$ Lemma \ref{LEM2.directcirculant} guarantees the existence of an odd alternating $4$-cycle with respect to the Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime$ and this $4$-cycle contains two $b$-edges of $M_0.$ In the construction of $H_i,\,i=1,\,2,$ of $\Gamma$ (from $H_i^\prime,\,i=1,\,2,$ of $\Gamma^\prime$) the $a$-edges and $M_0$-edges (that is, $b$-edges that belong to the matching $M_0$) of $\Gamma^\prime$ are retained as it is in the transformation of $\Gamma^\prime$ to $\Gamma.$ While obtaining $H_i$ from $H_i^\prime,$ each of the $b$-edges of $M_1$ in $H_i^\prime$ are replaced by a path of odd length whose internal vertices are in $C_2$ and $C_3.$ Hence an odd alternating $4$-cycle that exists in $\Gamma^\prime$ becomes an odd alternating $4$-cycle with respect to the Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma.$
\noindent{\bf Claim 3.} For any Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime\in\Gamma(\alpha,\,\beta)$ with property $Q_2,$ there is a Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma\in\Gamma(\alpha+2,\,\beta)$ with property $Q_2.$
In Claims 1 and 2, we proved that the existence of the property $Q_2$ in $\Gamma^\prime\in \Gamma(1,\,2)$ or $\Gamma(2,\,1)$ implies the existence of the property $Q_2$ in $\Gamma\in\Gamma(3,\,2)$ or $\Gamma(4,\,1),$ respectively. From the Hamilton cycle decomposition and construction of the odd alternating $4$-cycle of $\Gamma$ from $\Gamma^\prime,$ we have shown that the odd alternating $4$-cycle lies between $C_0$ and $C_1,$ that is having two vertices in $C_0$ and two vertices in $C_1.$ Hence, as in Claim 2, the odd alternating $4$-cycle in the Hamilton cycle decomposition $\{H_1^\prime,\,H_2^\prime\}$ of $\Gamma^\prime\in \Gamma(\alpha,\,\beta)$ is also the odd alternating $4$-cycle in the Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma\in\Gamma(\alpha+2,\,\beta);$ this completes the proof of existence of the property $Q_2$ with respect to the Hamilton cycle decomposition $\{H_1,\,H_2\}$ of $\Gamma.$
This completes the proof of the lemma.\hfill$\Box$
\begin{rmk}\label{DC.remark3}
Recall that our graphs of $\Gamma(\alpha,\,\beta)$ always have even order by assumption. As every member of $\Gamma(\alpha,\,\beta)$ is isomorphic to a member of $\Gamma(\beta,\,\alpha)$ and vice-versa, to each member of $\Gamma\in\Gamma(\alpha,\,\beta),$ there corresponds a graph $\Gamma_1\,(\cong\Gamma)\in\Gamma(\beta,\,\alpha).$ If $\alpha$ is odd, then $k_a,$ the length of each of the $a$-cycles of $\Gamma\in\Gamma(\alpha,\,2),$ must be even. Let $\Gamma\in\Gamma(\alpha,\,2)$ and let $\Gamma_1\in\Gamma(2,\,\alpha)$ be the image of $\Gamma$ under an isomorphism. Clearly, the $a$-edges and $b$-edges of $\Gamma$ are the $b$-edges and $a$-edges of $\Gamma_1,$ respectively, and vice versa. As $k_a$ is even, each $a$-cycle $C_i,\,0\le i\le \alpha-1,$ in $\Gamma$ has two $1$-factors, say $F_1^i$ and $F_2^i$ in the subgraph induced by $C_i;$ that is, for $i\in \{0,\,1,\,\ldots,\, \alpha-1\},$ $F_1^i=\{(i,\,0)(i,\,1),\,(i,\,2)(i,\,3),\,\ldots,\,(i,\,n-2)(i,\,n-1)\}$ and $F_2^i=\{(i,\,1)(i,\,2),\,(i,\,3)(i,\,4),\,\ldots,\,(i,\,n-1)(i,\,0)\}.$ The edges of the $1$-factor $F_1^i$ of $\Gamma$ become the $M_0$-edges of $\Gamma_1\in\Gamma(2,\,\alpha)$ and the edges of the $1$-factor $F_2^i$ of $\Gamma$ become the $M_1$-edges of $\Gamma_1.$ In the proof of Lemma \ref{LEM1.directcirculant}, we constructed an odd alternating $4$-cycle with respect to the Hamilton cycle decomposition of $\Gamma,$ so that the two $a$-edges use the edges of $F_1^i.$ Hence, if $\Gamma\in\Gamma(\alpha,\,2)$ has an odd alternating $4$-cycle, then the image of the $4$-cycle under the isomorphism between the corresponding graphs of $\Gamma(\alpha,\,2)$ and $\Gamma(2,\,\alpha)$ gives the odd alternating $4$-cycle in $\Gamma_1\in\Gamma(2,\,\alpha)$ using two $M_0$-edges of $\Gamma_1.$
\end{rmk}
\begin{rmk}\label{DC.remark4}
By saying, for $\Gamma\in\Gamma(\alpha,\,\beta)$ obtain the reduced graph of $\Gamma$ in $\Gamma(\alpha_1,\,\beta_1),\,\alpha_1\le\alpha,\,\beta_1\le\beta,$ we mean the following: The successive reduced graphs of $\Gamma\in\Gamma(\alpha,\,\beta)$ in $\Gamma(\alpha-2,\,\beta),\,\Gamma(\alpha-4,\,\beta),\,\ldots,\,\Gamma(\alpha_1,\,\beta),$ yields a graph $\Gamma_1\in\Gamma(\alpha_1,\,\beta).$ But the class of graphs in $\Gamma(\alpha_1,\,\beta)$ are isomorphic to the class of graphs in $\Gamma(\beta,\,\alpha_1),$ that is, to each graph of $\Gamma(\alpha_1,\,\beta)$ there is an isomorphic copy of it in $\Gamma(\beta,\,\alpha_1)$ and vice versa. Hence $\Gamma_1$ can be considered as a graph $\Gamma_2\in\Gamma(\beta,\,\alpha_1).$ Then successive reduced graphs of $\Gamma_2$ yields a graph $\Gamma_3\in\Gamma(\beta_1,\,\alpha_1).$ As the class of graphs in $\Gamma(\beta_1,\,\alpha_1)$ are isomorphic to the class of graphs in $\Gamma(\alpha_1,\,\beta_1),$ $\Gamma_3$ can be considered as a graph $\Gamma_4\in\Gamma(\alpha_1,\,\beta_1).$ We shall call $\Gamma_3\in\Gamma(\beta_1,\,\alpha_1)$ or $\Gamma_4\in\Gamma(\alpha_1,\,\beta_1)$ (note that $\Gamma_3\cong \Gamma_4$) as a reduced graph of $\Gamma\in\Gamma(\alpha,\,\beta)$ according to the circumstances.
Similarly, by saying, for $\Gamma^\prime\in\Gamma(\nu,\,\delta)$ obtain its lifted graph $\Gamma_{k+\ell}\in \Gamma(\nu+2k,\,\delta+2\ell),$ where $k$ and $\ell$ are not simultaneously zero, we mean the following: for $\Gamma^\prime\in\Gamma(\nu,\,\delta),$ by lifting, we obtain a graph $\Gamma_1\in\Gamma(\nu+2,\,\delta).$ Similarly, for $\Gamma_1\in\Gamma(\nu+2,\,\delta),$ by lifting, we obtain $\Gamma_2\in\Gamma(\nu+4,\,\delta).$ Successively, we can get the graph $\Gamma_k\in\Gamma(\nu+2k,\,\delta).$ But the class of graphs in $\Gamma(\nu+2k,\,\delta)$ are isomorphic to the class of graphs in $\Gamma(\delta,\,\nu+2k),$ that is, to each graph of $\Gamma(\nu+2k,\,\delta)$ there is an isomorphic copy of it in $\Gamma(\delta,\,\nu+2k)$ and vice versa. Hence $\Gamma_k$ can be considered as a graph of $\Gamma(\delta,\,\nu+2k).$ From this, by successive liftings, we obtain $\Gamma_{k+1}\in \Gamma(\delta+2,\,\nu+2k),\,\Gamma_{k+2}\in\Gamma(\delta+4,\,\nu+2k),\,\ldots,\,\Gamma_{k+\ell}\in\Gamma(\delta+2\ell,\,\nu+2k).$ We call $\Gamma_{k+\ell}$ as a lifted graph of $\Gamma^\prime\in\Gamma(\nu,\,\delta).$ As the two classes of graphs $\Gamma(\nu+2k,\,\delta+2\ell)$ and $\Gamma(\delta+2\ell,\,\nu+2k)$ have the same set of graphs, upto isomorphism, we can consider $\Gamma_{k+\ell}$ as an element of $\Gamma(\nu+2k,\,\delta+2\ell).$
\end{rmk}
The idea in the next theorem is based on \cite{JCTB.46.142-153}.
\begin{thm}\label{THM.directcirculant}
Every $4$-regular connected circulant graph $\Gamma\in\Gamma(\alpha,\,\beta)$ of even order with jumps of different parity has the property $Q_2$ with respect to a Hamilton cycle decomposition.
\end{thm}
\noindent{\bf Proof.} Let $\Gamma\in\Gamma(\alpha,\,\beta).$ As observed in Lemma \ref{LEM1.directcirculant}, $\alpha$ and $\beta$ are of different parity and hence we assume that $\alpha$ is even and $\beta$ is odd. Obtain the reduced graph $\Gamma^\prime\in\Gamma(1,\,2)$ of $\Gamma,$ see Remark \ref{DC.remark4}. If $\Gamma^\prime$ is simple, then it has a Hamilton cycle decomposition satisfying properties $Q_1$ and $Q_2,$ by Lemma \ref{LEM2.directcirculant}. Now we \lq\lq lift" $\Gamma^\prime$ to $\Gamma^{\prime\prime}\in\Gamma(\beta,\,2).$ By Lemma \ref{LEM1.directcirculant}, $\Gamma^{\prime\prime}$ has the properties $Q_1$ and $Q_2$ with respect to a Hamilton cycle decomposition of it. $\Gamma^{\prime\prime}$ can be considered as a graph $\Gamma^{\prime\prime\prime}\in\Gamma(2,\,\beta).$ Again lift $\Gamma^{\prime\prime\prime}$ to a graph in $\Gamma(\alpha,\,\beta),$ which is precisely $\Gamma,$ having the properties $Q_1$ and $Q_2$ with respect to a Hamilton cycle decomposition of it, by Lemma \ref{LEM1.directcirculant}. Thus the result is true if $\Gamma^\prime$ is simple.
Next we assume that $\Gamma^\prime$ is not simple. Let $\Gamma_1\in\Gamma(\alpha_1,\,\beta_1)$ be a reduced graph of $\Gamma\in\Gamma(\alpha,\,\beta)$ such that $\Gamma_1$ is simple but the reduced graph of $\Gamma_1$ in $\Gamma(\alpha_1-2,\,\beta_1)$ is not simple. The successive reduced graphs of $\Gamma_1$ results in $\Gamma^\prime\in\Gamma(2,\,1).$ Because of the above process of reductions and liftings of the classes of graphs, $\alpha_1$ and $\beta_1$ may have different parity from the parity of $\alpha$ and $\beta,$ respectively. $\Gamma^\prime$ is not simple only in three cases, see \cite{JCTB.46.142-153}:
(i) $\alpha_1=3$ and $c=0;$ in that case we obtain loops in $\Gamma^\prime.$
(ii) $\alpha_1=4$ and $c=0;$ in this case the edges in $\Gamma^\prime$ between $C_0^\prime$ and $C_1^\prime$ and, between $C_1^\prime$ and $C_0^\prime$ are the same, that is, they are multiple edges.
(iii) $\alpha_1=3$ and $c=\frac{k_a}{2};$ where $k_a$ is the order of the element $a$ in the generating set of $\Gamma_1;$ in this case we get multiple edges in $\Gamma^\prime.$
As the jumps of $\Gamma_1$ are of different parity, the first two cases correspond to the Cartesian product of cycles; in that case $c=0$ and $k_a=\beta_1.$ If $\beta_1\ge 5,$ again we reduce the graph $\Gamma_1\in\Gamma(3,\,\beta_1)$ or $\Gamma_1\in\Gamma(4,\,\beta_1)$ to a graph in $\Gamma(3,\,4)$ or $\Gamma(4,\,3).$ But $\Gamma(3,\,4)$ and $\Gamma(4,\,3)$ are isomorphic classes of graphs. Hence it is enough to consider $\Gamma(3,\,4)$ with $c=0.$ Then the class of graphs $\Gamma(3,\,4)$ with $c=0$ reduces to the single graph $C_3\Box C_4,$ where $\Box$ is the Cartesian product of graphs. In Figure 12, a Hamilton cycle decomposition is shown, where $C=(1\,2\,3\,4)$ is an odd alternating $4$-cycle.
\vspace{-.4cm}
\begin{center}
\includegraphics{fig12.eps}
\end{center}
\vspace{-.4cm}
{\small The graph $C_3\Box C_4\in\Gamma(3,\,4)$ and a Hamilton cycle decomposition $\{H_1,\,H_2\}$ with property $Q_2.$}\vspace{-.4cm}\begin{center}Figure 12
\end{center}
Next we consider the last case $\alpha_1=3,\,c=\frac{k_a}{2}.$ As $\Gamma_1$ is simple, then $k_a\ge 3$ and $\beta_1\ge 2,$ as $c=\beta_1=\frac{k_a}{2}\ge2.$ If $\beta_1\ge 4,$ we first reduce the graph $\Gamma_1$ to a graph in $\Gamma(3,\,2).$ In this case there is only one graph in $\Gamma(3,2),$ see \cite{JCTB.46.142-153},
as shown in Figure 13. In the Hamilton cycle decomposition shown in Figure 13, $C=(1\,2\,3\,4)$ is an odd alternating $4$-cycle.
\begin{center}
\includegraphics{fig13.eps}
\end{center}
\vspace{-.4cm}
{\small The only graph of $\Gamma(3,\,2)$ and a Hamilton cycle decomposition with property $Q_2.$}\vspace{-.4cm}\begin{center}Figure 13
\end{center}
\vspace{-.13cm}
To complete the proof, first we lift the graph $C_3\Box C_4\in\Gamma(3,\,4)$ of Figure 12 to the graph $G\in\Gamma(\beta,\,4)$ with properties $Q_1$ and $Q_2$ with respect to a Hamilton cycle decomposition, by Lemma \ref{LEM1.directcirculant}. Let $G_1\in\Gamma(4,\,\beta)$ be isomorphic to $G.$ As $G$ has the properties $Q_1$ and $Q_2,$ so does $G_1.$ Now we lift the graph $G_1\in\Gamma(4,\,\beta)$ to the graph $\Gamma\in\Gamma(\alpha,\,\beta).$ Clearly, $\Gamma\in\Gamma(\alpha,\,\beta)$ has a Hamilton cycle decomposition, by Lemma \ref{LEM1.directcirculant}. Similarly, we prove it for the graph in Figure 13.
This completes the proof of theorem.\hfill$\Box$
The following theorem of Jha \cite{Comput.Mat.Appl.31.11-19} is used in our proof of Theorem \ref{THM.CIR.1}.
\begin{thm}\emph{\cite{Comput.Mat.Appl.31.11-19}
\label{THM.MAIN.directcirculant}
Let $\{H_1,\,H_2\}$ be a Hamilton cycle decomposition of a $4$-regular graph $G$ of even order $m$ containing an odd alternating four cycle, that is property $Q_2,$ with respect to $\{H_1,\,H_2\},$ then the graph $C_n\times G,\,n$ even, admits a Hamilton cycle decomposition.\hfill$\Box$
\end{thm}
\noindent{\bf Proof of Theorem \ref{THM.CIR.1}.}
Let $|V(G)|=m$ and $|V(H)|=n.$ Since $H$ is Hamilton cycle decomposable, we have $H=C_n^1\oplus C_n^2\oplus \ldots\oplus C_n^\ell,$ where each $C_n^i$ is a Hamilton cycle of $H.$ If at least one of $G$ or $H$ has odd order, then $G\times H$ is Hamilton cycle decomposable, by Theorem \ref{Ann.DM.3.21-28}, and hence assume that both $G$ and $H$ are of even order. Since $G$ has the property $Q,$ $G$ can be decomposed into $4$-regular connected circulants, that is, $G=G_1\oplus G_2\oplus \ldots \oplus G_{\ell^\prime},$ where each $G_i$ is a $4$-regular circulant graph with jumps of different parity. Thus $G_i,\,1\le i\le \ell^\prime,$ can be decomposed into two Hamilton cycles with property $Q_2,$ by Theorem \ref{THM.directcirculant}. Now $G\times H\cong (G_1\oplus G_2\oplus \ldots\oplus G_{\ell^\prime})\times (C_n^1\oplus C_n^2\oplus \ldots\oplus C_n^\ell)=(G_1\times C_n^1)\oplus \ldots (G_1\times C_n^\ell)\oplus \ldots \oplus (G_{\ell^\prime}\times C_n^1)\oplus \ldots (G_{\ell^\prime}\times C_n^\ell).$ But each $G_i\times C_n^j,\,1\le i\le \ell^\prime,\,1\le j\le \ell,$ can be decomposed into Hamilton cycles, by Theorem \ref{THM.MAIN.directcirculant}.
This completes the proof of the theorem.\hfill$\Box$
\noindent {\bf Conclusion.} In \cite{{DM.268.49-58}, {IJPAM.23.723-729}, {ARS.80.33-44}, {DM.308.3586-3606}, {DM.310.2776-2789}}, existence of Hamilton cycle decompositions of the graphs $K_r\times K_s,\,K_r\times K_{s,\,s},\,K_{r(s)}\times K_{m(n)},\,K_{r,\,r}\times K_{m(n)}$ are proved; the factor graphs in the product graphs are either complete or complete multipartite graphs, which are very dense graphs. However, if we consider one of the factor graphs as circulant, in $G\times H,$ with same number of odd and even jumps, irrespective of the number of jumps, which can be paired so that the resulting set of edges induces connected $4$-regular circulants, then $G\times H$ is Hamilton cycle decomposable. This proves that with this additional condition on $G$ or $H,$ the factor graphs $G$ and $H$ need not be dense. For example, Hamilton cycle decomposition of $(K_{4r+2}-F)\times H,$ where $F$ is a $1$-factor of $K_{4r+2},$ follows from our Theorem \ref{THM.CIR.1} whenever $H$ is Hamilton cycle decomposable multigraph. Also one can conclude that tensor products of certain sparse Hamilton cycle decomposable circulant graphs are Hamilton cycle decomposable.
\noindent{\bf Acknowledgments:} The authors would like to thank the Department of Science and Technology, Government of India, New Delhi, for partial financial assistance through Grant No: SR/S4/MS:481/07.
|
\section{Introduction}
In this article, we study the minimal free resolution of the associated graded ring of the local ring $A$ of a monomial curve
$C \subset \mathbb{A}^{4}$ corresponding to an arithmetic sequence based on the standard basis theory.
The associated graded ring $G=gr_{m}(A)= \bigoplus_{i=0}^{\infty} (m^{i} / m^{i+1})$ of $A$ with maximal ideal $m$ is a standard graded $k$-algebra. Since it corresponds to the important geometric construction, it has been studied to get comprehensive information on the local ring (see \cite{rossi-valla,rossi-sharifan,molinelli-tamone1,molinelli-tamone2,molinelli-patil-tamone}).
Because the minimal finite free resolution of a finitely generated $k$-algebra is a very useful tool to extract information about the algebra, finding an explicit minimal free resolution of a standard $k$-algebra is a basic problem. This difficult problem has been extensively studied in the case of affine monomial curves \cite{sharifan-nahandi,sengupta,gimenez-sengupta-srinivasan,oneto-tamone,barucci-froberg-sahin}.
We recall that a monomial affine curve $C$ has a parametrization
\[x_0=t^{m_0}, \; x_1=t^{m_1}, \;\ldots, \; x_n=t^{m_n}\]
where $m_0,m_1,\ldots,m_n$ are positive integers with
$gcd(m_0,m_1,...,m_n)=1$. The additive semigroup, which is denoted by
$$<m_0,m_1,...,m_n>=\{\; \sum_{0 \leq i \leq n} \mathbb{N}m_i \; \mid \; \mathbb{N}=
\{0,1,2,\ldots \} \}$$
\noindent generated minimally by $m_0,m_1,...,m_n$, i.e., $m_j \notin \sum_{0 \leq i \leq n; i \neq j} \mathbb{N}m_i$ for $i \in \{0,\ldots,n\}$.
Assume that $m_0,m_1,\ldots,m_n$ be positive integers such that $0 < m_0 < m_1 < \ldots < m_n$ and $m_i=m_0 + id$ for
every $1 \leq i \leq n$, where $d$ is the common difference, i.e. the integers $m_i$'s form an arithmetic progression.
The monomial curve which is defined parametrically by
\[x_0=t^{m_0}, \; x_1=t^{m_1}, \;\ldots, \; x_n=t^{m_n}\]
\noindent such that $0 < m_0 < m_1 < \ldots < m_n$ form an arithmetic progression is called a certain monomial curve.
In order to study the associated graded ring
of a monomial curve $C$ at the origin, it is possible to consider
either the associated graded ring of
$A=k[[t^{m_0},t^{m_1},...,t^{m_n}]]$ with respect to the maximal
ideal $m=(t^{m_0},t^{m_1},...,t^{m_n})$ which is denoted by
$gr_m(k[[t^{m_0},t^{m_1},...,t^{m_n}]])$, or the ring
$k[x_0,x_1,...,x_n]/I(C)^{*}$, where $I(C)^{*}$ is the ideal generated by the polynomials $f^{*}$
for $f$ in $I(C)$, where $f^{*}$ is the homogeneous summand of $f$
of the least degree, since they are isomorphic. We recall that $I(C)^{*}$ is the defining ideal of the tangent cone of the curve $C$ at the origin.
Our main aim in this paper is to give an explicit minimal free resolution of the associated graded ring for certain monomial curves in affine 4-space. Even if one can obtain the numerical invariants of the minimal free resolution of the tangent cone of certain monomial curves in $\mathbb{A}^{4}$ by using the Theorem 4.1 and Proposition 4.6 in \cite{sharifan-nahandi}, we give the minimal free resolution of the tangent cone of certain monomial curves in affine 4-space in an explicit form by giving a new proof based on the standard basis theory.
Using the standard basis theory and knowing the minimal generating set of binomial generators of the defining ideal of certain monomial curve explicitly from \cite{patil}, we find the minimal generators of the tangent cone of a certain monomial curve in affine 4-space. By knowing the minimal generators, we show the Cohen-Macaulayness of the tangent cone of these families of curves. We obtain explicit minimal free resolution by using Schreyer's theorem but prove it using the Buchsbaum-Eisenbud theorem \cite{eisenbud-buchsbaum}.
Finally, we give the minimal graded free resolutions and as a corollary compute the Hilbert function of the tangent cones for these families. All computations have been carried out using {\footnotesize SINGULAR}\cite{singular}.
\section{Minimal generators of the associated graded ring}
In this section, we find the minimal generators of the tangent
cone of the certain monomial curve $C$ having the defining ideal as in Theorem
4.5 in \cite{patil} in affine 4-space. First, we recall the theorem which gives the construction
of the minimal set of generators for the defining ideal of certain affine monomial curve in
$\mathbb{A}^{4}$ .
Let $m_0 < m_1 < m_2 < m_3$ be positive integers with $gcd(m_0,m_1,m_2,m_3)=1$ and
assume that $m_0,m_1,m_2,m_3$ form an arithmetic progression with common difference $d$.
Let $R=k[x_0,x_1,x_2,y]$ be a polynomial ring over the field $k$. We use $y$ instead of $x_3$ by
following the same notation in \cite{sengupta,sengupta2,patil}. Let $\phi : R \rightarrow k[t^{m_0},t^{m_1},t^{m_2},t^{m_3}]$ be the $k$-algebra homomorphism defined by
\begin{center}
$\phi(x_0)=t^{m_0}$, $\phi(x_1)=t^{m_1}$, $\phi(x_2)=t^{m_2}$, $\phi(y)=t^{m_3}$
\end{center}
and $I(C)=Ker (\phi)$. Let us write $m_{0}=3a+b$ such that $a$ and $b$ are positive integers $a \geq 1$ and $b \in [1,3]$.
In \cite{sengupta}, the following theorem is given as a definition.
\begin{theorem} \label{thm1} {\rm \cite{patil}} Let \\[1mm]
\indent\hspace{1cm} $\xi_{11} := x_1^{2}-x_{0}x_{2},$\\[1mm]
\indent\hspace{1cm} $\varphi_{i} := x_{i+1}x_2-x_{i}y,\; for \; 0 \leq i \leq 1.$\\[1mm]
\indent\hspace{1cm} $\psi_{j} := x_{b+j}y^{a}-x_{0}^{a+d}x_{j},\; if \; 1 \leq b \leq 2 \;\; and \; \; 0 \leq j \leq 2-b.$\\[1mm]
\indent\hspace{1cm} $\theta := y^{a+1}-x_{0}^{a+d}x_{3-b}$. \\[1mm]
\indent\hspace{1cm} $\displaystyle
G :=\left\{
\begin{array}{lr}
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \psi_{0}, \psi_{1}\} \cup \{ \theta\}& ~ if\;\;\; b=1, \\
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \psi_{0}\} \cup \{ \theta\}& ~ if\;\;\; b=2, \\
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \theta\}& ~ if\;\;\; b=3.
\end{array}
\right.
$\\[1mm]
\noindent then, G is a minimal generating set for the defining ideal $I(C)$.
\end{theorem}
Now, we recall the definition of the negative degree reverse lexicographical
ordering among the other local orderings.
\begin{definition} {\rm \cite[p.14]{greuel-pfister}} (negative degree reverse
lexicographical ordering)
\begin{center}
$x^{\alpha} >_{ds} x^{\beta}:\Leftrightarrow$ $degx^{\alpha}<degx^{\beta}$,
where $degx^{\alpha}=\alpha_{1}+...+\alpha_{n}$, \\
or $(degx^{\alpha}=degx^{\beta}$ and $\exists 1\leq i\leq n:
\alpha_{n}=\beta_{n},...,\alpha_{i+1}=\beta_{i+1},\alpha_{i}<\beta_{i}).$
\end{center}
\end{definition}
In the following Lemma, we show that the above set G is also standard basis with respect to $>_{ds}$.
\begin{lemma}\label{lemma1}
The minimal set $G$ is a standard basis with
respect to the negative degree reverse lexicographical ordering
$>_{ds}$ with $x_0 > x_1 > x_2 > y$.
\end{lemma}
\noindent {\em Proof.}
We apply the standard basis algorithm
to the set $G$. We will prove for $b=1,2$, and $3$, respectively.
By using the notation in \cite{greuel-pfister}, we denote the leading monomial of a polynomial $f$
by $LM(f)$, the S-polynomial of the polynomials $f$ and $g$ by $spoly(f,g)$ and the
Mora's polynomial weak normal form of $f$ with respect to G by $NF(f \mid G)$.\\
\noindent {\bf Case \textit{b} $=$ 1.}\\
From the minimal generating set $G$ in Theorem~\ref{thm1}, we obtain
$$G = \big\{\xi_{11} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0} = x_{1}x_2-x_{0}y,\;\; \varphi_{1} = x_{2}^2-x_{1}y,$$
$$\psi_{0} = x_{1}y^{a}-x_{0}^{a+d+1},\;\; \psi_{1} = x_{2}y^{a}-x_{0}^{a+d}x_{1},\;\; \theta = y^{a+1}-x_{0}^{a+d}x_{2}\big\}.$$
\noindent Recalling that the ordering is the negative degree reverse lexicographical ordering, we have
$LM(\xi_{11}) = x_1^{2}$, $LM(\varphi_{0})$ = $x_{1}x_{2}$, $LM(\varphi_{1}) = x_2^{2}$,
$LM(\psi_{0}) = x_{1}y^{a}$, $LM(\psi_{1})$ = $x_{2}y^{a}$ and $LM( \theta) = y^{a+1}$.
We begin with $\xi_{11}$ and $\varphi_{0}$. $LM(\xi_{11})$ = $x_1^{2}$ and
$LM(\varphi_{0}) = x_{1}x_{2}$. We compute
${\rm spoly}(\xi_{11}, \varphi_{0})$ = $x_{0}x_{1}y-x_{0}x_{2}^2$.
$LM({\rm spoly}(\xi_{11}, \varphi_{0}))$ = $x_{0}x_{2}^2$. Among the leading monomials of the elements of $G$,
only $LM(\varphi_{1})$ divides $LM({\rm spoly}(\xi_{11}, \varphi_{0}))$.
Also ecart$(\varphi_{1})$=ecart$({\rm spoly}(\xi_{11}, \varphi_{0}))$ = 0.
${\rm spoly}(\varphi_{1}, {\rm spoly}(\xi_{11}, \varphi_{0}))$ = 0 implies
$NF({\rm spoly}(\xi_{11},\varphi_{0})\,\vert\,G)$ = 0.
Next, we choose $\xi_{11}$ and $\varphi_{1}$. Since ${\rm lcm}(LM(\xi_{11}), LM(\varphi_{1}))$
= $LM(\xi_{11}).LM(\varphi_{1})$, then $NF({\rm spoly}(\xi_{11}, \varphi_{1})\,\vert\,\{\xi_{11}, \varphi_{1}\})$ = 0.
This implies that $NF({\rm spoly}(\xi_{11}, \varphi_{1})\,\vert\,G)$ = 0.
In the same manner, $NF({\rm spoly}(\xi_{11}, \psi_{1})\,\vert\,G)$ = 0,
$NF({\rm spoly}(\xi_{11}, \theta)\,\vert\,G)$ = 0, $NF({\rm spoly}(\varphi_{0},\theta)\,\vert\,G)$ = 0,
$NF({\rm spoly}(\varphi_{1},\psi_{0})\,\vert\,G)$ = 0
and
$NF({\rm spoly}(\varphi_{1},\theta)\,\vert\,G)$ = 0.
Now, we compute S-polynomial of $\xi_{11}$ and $\psi_{0}$.
${\rm spoly}(\xi_{11}, \psi_{0}) = x_{0}^{a+d+1}x_{1}-x_{0}x_{2}y^{a}$.
Among the leading monomials of the elements of $G$,
only $LM(\psi_{1})$ divides $LM({\rm spoly}(\xi_{11}, \psi_{0}))$ = $x_{0}x_{2}y^{a}$.
Also ecart$(\psi_{1})$=ecart$({\rm spoly}(\xi_{11}, \psi_{0}))$ = $d$.
${\rm spoly}(\psi_{1}, {\rm spoly}(\xi_{11}, \psi_{0}))$ =0 implies
$NF({\rm spoly}(\xi_{11},\psi_{0})\,\vert\,G)$ = 0.
Again, we compute S-polynomial of $\varphi_{0}$ and $\varphi_{1}$.
${\rm spoly}(\varphi_{0}, \varphi_{1})$ = $x_{1}^{2}y-x_{0}x_{2}y$.
Among the leading monomials of the elements of $G$,
only $LM(\xi_{11})$ divides\\
$LM({\rm spoly}(\varphi_{0}, \varphi_{1}))$ = $x_{1}^2y$.
Also ecart$(\xi_{11}) = $ecart$({\rm spoly}(\varphi_{0}, \varphi_{1}))$ = 0.
${\rm spoly}(\xi_{11}, {\rm spoly}(\varphi_{0}, \varphi_{1}))$ = 0 implies
$NF({\rm spoly}(\varphi_{0},\varphi_{1})\,\vert\,G)$ = 0.
Now choose $\varphi_{0}$ and $\psi_{0}$. Then, S-polynomial of $\varphi_{0}$ and $\psi_{0}$ is
${\rm spoly}(\varphi_{0}, \psi_{0})=x_{0}^{a+d+1}x_{2}-x_{0}y^{a+1}$.
Once again, only $LM(\theta)$ divides $LM({\rm spoly}(\varphi_{0}, \psi_{0}))$ = $x_{0}y^{a+1}$
among the leading monomials of the elements of $G$.
Also ecart$(\theta)$ = ecart$({\rm spoly}(\varphi_{0}, \psi_{0}))$ = $d$.
${\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{0}))$ = 0 implies
$NF({\rm spoly}(\varphi_{0},\psi_{0})\,\vert\,G)$ = 0.
Similarly, ${\rm spoly}(\varphi_{0}, \psi_{1}) = x_{0}^{a+d}x_{1}^{2}-x_{0}y^{a+1}$.
Again, as in the previous case $LM(\theta)$ divides $LM({\rm spoly}(\varphi_{0}, \psi_{1}))$ = $x_{0}y^{a+1}$.
Also ecart$(\theta)$ = ecart$({\rm spoly}(\varphi_{0}, \psi_{1}))$ = $d$.
${\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{1}))$ = $x_{0}^{a+d}x_{1}^{2}-x_{0}^{a+d+1}x_2.$
Among the leading monomials of the elements of $G$,
only $LM(\xi_{11}) = x_{1}^2$ divides $LM({\rm spoly}({\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{1}))))$ = $x_{0}^{a+d}x_{1}^2$.
ecart$(\xi_{11})$ = ecart$({\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{1})))$ = 0.
${\rm spoly}(\xi_{11}, {\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{1})))$ = 0 implies
$NF({\rm spoly}(\varphi_{0},\psi_{1})\,\vert\,G)$ = 0.
Similarly, we compute ${\rm spoly}(\varphi_{1}, \psi_{1})$ = $x_{0}^{a+d}x_{1}x_{2}-x_{1}y^{a+1}$.
Among the leading monomials of the elements of $G$,
only $LM(\psi_{0})$ and $LM(\theta)$ divides $LM({\rm spoly}(\varphi_{1}, \psi_{1}))$ = $x_{1}y^{a+1}$.
Note that ecart$(\psi_{0})$ = ecart$(\theta) = d$.
Firstly, beginning with $\psi_{0}$,
${\rm spoly}(\psi_{0}, {\rm spoly}(\varphi_{1}, \psi_{1}))$ = $x_{0}^{a+d}x_{1}x_{2}-x_{0}^{a+d+1}y$.
Among the leading monomials of the elements of $G$, $LM(\varphi_{1}) = x_{1}x_{2}$ divides
$LM({\rm spoly}(\psi_{0}, {\rm spoly}(\varphi_{1}, \psi_{1})))$.
Also ecart$(\varphi_{1})$ = ecart$({\rm spoly}(\psi_{0}, {\rm spoly}(\varphi_{1}, \psi_{1})))$ = 0.
${\rm spoly}(\varphi_{1}, {\rm spoly}(\psi_{0}, {\rm spoly}(\varphi_{1}, \psi_{1})))$ = 0.
Secondly, taking $\theta$, ${\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{1}))$ = 0.
Thus, $NF({\rm spoly}(\varphi_{1},\psi_{1})\,\vert\,G)$ = 0.
We continue by computing ${\rm spoly}(\psi_{0}, \psi_{1})=x_{0}^{a+d}x_{1}^2-x_{0}^{a+d+1}x_{2}$.
$LM(\xi_{11}) = x_{1}^{2}$ divides $LM({\rm spoly}(\psi_{0}, \psi_{1}))$ = $x_{0}^{a+d}x_{1}^{2}.$
Also ecart$(\xi_{11})$ = ecart$({\rm spoly}(\psi_{0}, \psi_{1}))$ = 0.
${\rm spoly}(\xi_{11}, {\rm spoly}(\psi_{0}, \psi_{1})$ = 0 implies
$NF({\rm spoly}(\psi_{0},\psi_{1})\,\vert\,G)$ = 0.
In the same manner, ${\rm spoly}(\psi_{0}, \theta)$ = $x_{0}^{a+d}x_{1}x_{2}-x_{0}^{a+d+1}y$.
$LM(\varphi_{0}) = x_{1}x_{2}$ divides $LM({\rm spoly}(\psi_{0}, \theta))$ = $x_{0}^{a+d}x_{1}x_{2}.$
Also ecart$(\varphi_{0})$ = ecart$({\rm spoly}(\psi_{0}, \theta))$ = 0.
${\rm spoly}(\varphi_{0}, {\rm spoly}(\psi_{0}, \theta))$ = 0 implies
$NF({\rm spoly}(\psi_{0},\theta)\,\vert\,G)$ = 0.
Finally, we compute ${\rm spoly}(\psi_{1}, \theta)$ = $x_{0}^{a+d}x_{2}^{2}-x_{0}^{a+d}x_{1}y$.
$LM(\varphi_{1})$ = $x_{2}^{2}$ divides $LM({\rm spoly}(\psi_{1}, \theta))$ = $x_{0}^{a+d}x_{2}^{2}.$
Also ecart$(\varphi_{1})$ = ecart$({\rm spoly}(\psi_{1}, \theta))$ = 0.\\
${\rm spoly}(\varphi_{1}, {\rm spoly}(\psi_{1}, \theta))$ = 0 implies
$NF({\rm spoly}(\psi_{0},\theta)\,\vert\,G)$ = 0.
\vspace{0.5cm}
\noindent {\bf Case \textit{b} $=$ 2.}\\
As in the previous case, we obtain by the minimal generating set $G$ in Theorem~\ref{thm1},
$$G=\big\{\xi_{11} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0} = x_{1}x_2-x_{0}y,\;\; \varphi_{1} = x_{2}^2-x_{1}y,$$
$$\psi_{0} = x_{2}y^{a}-x_{0}^{a+d+1},\;\; \theta = y^{a+1}-x_{0}^{a+d}x_{1}\big\}.$$
$LM(\xi_{11}) = x_1^{2}$, $LM(\varphi_{0}) = x_{1}x_{2}$, $LM(\varphi_{1}) = x_{2}^{2}$,
$LM(\psi_{0}) = x_{2}y^{a}$ and $LM( \theta) = y^{a+1}$ with respect to
the negative degree reverse lexicographical ordering.
We begin with $\xi_{11}$ and $\varphi_{0}$. This case is exactly the same as in $b=1$.
Next, we choose $\xi_{11}$ and $\varphi_{1}$. As in the first case, since
${\rm lcm}(LM(\xi_{11}),LM(\varphi_{1}))$ = $LM(\xi_{11}).LM(\varphi_{1})$, then $NF({\rm spoly}(\xi_{11},\varphi_{1})\,\vert\,\{\xi_{11},\varphi_{1}\})$ = 0.
Therefore, this implies that $NF({\rm spoly}(\xi_{11},\varphi_{1})\,\vert\,G)$ = 0.
In the same manner, $NF({\rm spoly}(\xi_{11}, \psi_{0})\,\vert\,G)$ = 0,
$NF({\rm spoly}(\xi_{11}, \theta)\,\vert\,G)$ = 0, $NF({\rm spoly}(\varphi_{0},\theta)\,\vert\,G)$ = 0 and
$NF({\rm spoly}(\varphi_{1},\theta)\,\vert\,G)$ =0.
Again, we compute S-polynomial of $\varphi_{0}$ and $\varphi_{1}$.
This one is also the same as in the previous case.
Now choose $\varphi_{0}$ and $\psi_{0}$. Then, S-polynomial of $\varphi_{0}$ and $\psi_{0}$ is
${\rm spoly}(\varphi_{0}, \psi_{0})$ = $x_{0}^{a+d+1}x_{1}-x_{0}y^{a+1}$.
Once again, only $LM(\theta) = y^{a+1}$ divides $LM({\rm spoly}(\varphi_{0}, \psi_{0}))$ = $x_{0}y^{a+1}$
among the leading monomials of the elements of $G$.
Also, ecart$(\theta)$ = ecart$({\rm spoly}(\varphi_{0}, \psi_{0}))$ = $d$.
${\rm spoly}(\theta, {\rm spoly}(\varphi_{0}, \psi_{0}))$ = 0 implies
$NF({\rm spoly}(\varphi_{0},\psi_{0})\,\vert\,G)$ = 0.
Similarly, we compute ${\rm spoly}(\varphi_{1}, \psi_{0})$ = $x_{0}^{a+d+1}x_{2}-x_{1}y^{a+1}$.
Among the leading monomials of the elements of $G$,
only $LM(\theta)$ divides $LM({\rm spoly}(\varphi_{1}, \psi_{0}))$ = $x_{1}y^{a+1}$.
ecart$(\theta)$ = ecart$({\rm spoly}(\varphi_{1}, \psi_{0}))$ = $d$.
${\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{0}))$ = $x_{0}^{a+d+1}x_{2}-x_{0}^{a+d}x_{1}^{2}$.
Since ${\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{0}))$ is not zero, again among the leading
monomials of the elements of $G$, $LM(\xi_{11})$ = $x_{1}^{2}$ divides
$LM({\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{0})))$ = $x_{0}^{a+d}x_{1}^{2}$.
ecart$(\xi_{11})$ = ecart$({\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{0})))$ = 0.\\
${\rm spoly}(\xi_{11}, {\rm spoly}(\theta, {\rm spoly}(\varphi_{1}, \psi_{0})))$ = 0.
Thus, $NF({\rm spoly}(\varphi_{1},\psi_{0})\,\vert\,G)$ = 0.
Finally, we compute ${\rm spoly}(\psi_{0}, \theta)$ = $x_{0}^{a+d}x_{1}x_{2}-x_{0}^{a+d+1}y$.
$LM(\varphi_{0}) = x_{1}x_{2}$ divides $LM({\rm spoly}(\psi_{0}, \theta))$ = $x_{0}^{a+d}x_{1}x_{2}.$
Also ecart$(\varphi_{0})$ = ecart$({\rm spoly}(\psi_{0}, \theta))$ = 0.
${\rm spoly}(\varphi_{0}, {\rm spoly}(\psi_{0}, \theta))$ = 0 implies
$NF({\rm spoly}(\psi_{0},\theta)\,\vert\,G)$ = 0.
\vspace{0.5cm}
\noindent {\bf Case \textit{b} $=$ 3.}\\
Finally, by writing 3 instead of b in the minimal generating set $G$ in Theorem~\ref{thm1}, we obtain
$$G=\big\{\xi_{11} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0} = x_{1}x_2-x_{0}y,\;\; \varphi_{1} = x_{2}^2-x_{1}y,\;\;
\theta = y^{a+1}-x_{0}^{a+d+1}\big\}.$$
In the same manner, $LM(\xi_{11}) = x_1^{2}$, $LM(\varphi_{0})$ = $x_{1}x_{2}$,
$LM(\varphi_{1}) = x_{2}^{2}$ and $LM( \theta)$ = $y^{a+1}$ with respect to
the negative degree reverse lexicographical ordering $>_{ds}$.
As in the previous cases, we begin with $\xi_{11}$ and $\varphi_{0}$ and this case is exactly the same as in $b=1$.
In the same manner, $NF({\rm spoly}(\{\xi_{11},\varphi_{1})\,\vert\,G)$ = 0,
$NF({\rm spoly}(\xi_{11}, \theta)\,\vert\,G)$ = 0, $NF({\rm spoly}(\varphi_{0},\theta)\,\vert\,G)$ = 0 and
$NF({\rm spoly}(\varphi_{1},\theta)\,\vert\,G)$ = 0.
Finally, the computation of the S-polynomial of $\varphi_{0}$ and $\varphi_{1}$ also results
as in the case $b=1$.
Therefore, if $b=$1,2 and 3, we conclude that the set $G$ is a standard basis
with respect to the negative degree reverse lexicographical ordering $>_{ds}$.
\begin{flushright}
$\Box$
\end{flushright}
We can now find the minimal generating set of the tangent cone by using the above lemma.
\begin{proposition}\label{prop1} Let $C$ be a certain monomial
curve having parametrization
\[x_0=t^{m_0}, \; x_1=t^{m_1}, \; x_2=t^{m_2}, \; y=t^{m_3}\]
\noindent $m_0=3a+b$ for positive integers $a \geq 1$ and $b \in [1,3]$
and $0 < m_0 < m_1 < m_2 < m_3$ form an arithmetic progression with common difference $d$
and let the generators of the defining ideal $I(C)$ be given by the set $G$ in Theorem \ref{thm1}.
Then the defining ideal $I(C)^{*}$ of the tangent cone is generated by the set $G^{*}$ consisting of
the least homogeneous summands of the binomials in $G$.
\end{proposition}
\noindent {\em Proof}. By the
Lemma~\ref{lemma1},
\[
G :=\left\{
\begin{array}{lr}
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \psi_{0}, \psi_{1}\} \cup \{ \theta\}& ~ if\;\;\; b=1, \\
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \psi_{0}\} \cup \{ \theta\}& ~ if\;\;\; b=2, \\
\{\xi_{11} \} \cup \{ \varphi_{0}, \varphi_{1}\} \cup \{ \theta\}& ~ if\;\;\; b=3.
\end{array}
\right.
\]
as in Theorem \ref{thm1}, is a standard basis of $I(C)$ with
respect to a local degree ordering $>_{ds}$ with respect to
$x_0 > x_1 > x_2 > y$. Then, from \cite [Lemma 5.5.11]{greuel-pfister},
$I(C)^{*}$ is generated by the least homogeneous summands of the elements in the standard basis.
Thus, $I(C)^*$ is generated by\\
if $b=1$
$$G^{*}=\big\{\xi_{11}^{*} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*}= x_{1}x_2-x_{0}y,\;\; \varphi_{1}^{*}= x_{2}^2-x_{1}y,\;\; \psi_{0}^{*} = x_{1}y^{a}, $$
$$\psi_{1}^{*}= x_{2}y^{a},\;\; \theta^{*} = y^{a+1}\big\},$$
if $b=2$
$$G^{*}=\big\{\xi_{11} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*} = x_{1}x_2-x_{0}y,\;\;
\varphi_{1} ^{*}= x_{2}^2-x_{1}y,\;\; \psi_{0}^{*} = x_{2}y^{a}, \theta^{*} = y^{a+1}\big\},$$
and if $b=3$
$$G^{*}=\big\{\xi_{11}^{*} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*} = x_{1}x_2-x_{0}y,\;\;
\varphi_{1} ^{*}= x_{2}^2-x_{1}y,\;\; \theta^{*} = y^{a+1}\big\}.$$
\begin{flushright}
$\Box$
\end{flushright}
\begin{theorem}\label{thm2} Let $C$ be a certain monomial
curve having parametrization
\[x_0=t^{m_0}, \; x_1=t^{m_1}, \; x_2=t^{m_2}, \; y=t^{m_3}\]
\noindent $m_0=3a+b$ for positive integers $a \geq 1$ and $b \in [1,3]$
and $0 < m_0 < m_1 < m_2 < m_3$ form an arithmetic progression with common difference $d$.
The certain monomial curve $C$ with the defining ideal $I(C)$ as in Theorem \ref{thm1} has
Cohen-Macaulay tangent cone at the origin.
\end{theorem}
\begin{proof}
We can apply the Theorem 2.1 in \cite{arslan1} to the generators of the
tangent cone which are given by the set\\
if $b=1$
$$G^{*}=\big\{\xi_{11}^{*} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*}= x_{1}x_2-x_{0}y,\;\;
\varphi_{1}^{*}= x_{2}^2-x_{1}y,\;\; \psi_{0}^{*} = x_{1}y^{a}, $$
$$\psi_{1}^{*}= x_{2}y^{a},\;\; \theta^{*} = y^{a+1}\big\},$$
if $b=2$
$$G^{*}=\big\{\xi_{11} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*} = x_{1}x_2-x_{0}y,\;\;
\varphi_{1} ^{*}= x_{2}^2-x_{1}y,\;\; \psi_{0}^{*} = x_{2}y^{a}, \theta^{*} = y^{a+1}\big\},$$
and if $b=3$
$$G^{*}=\big\{\xi_{11}^{*} = x_1^{2}-x_{0}x_{2},\;\; \varphi_{0}^{*} = x_{1}x_2-x_{0}y,\;\;
\varphi_{1} ^{*}= x_{2}^2-x_{1}y,\;\; \theta^{*} = y^{a+1}\big\}.$$
All of these sets are
Gr\"{o}bner bases with respect to the reverse lexicographic order
with $x_0>y>x_1>x_2$. Since $x_0$ does not divide the leading
monomial of any element in $G^{*}$ in all three cases, the ring
$k[x_0,x_1,x_2,y]/I(C)^*$ is Cohen-Macaulay from Theorem 2.1
in \cite{arslan1}. Thus, $R={\rm gr}_m(k[[t^{m_0},t^{m_1},t^{m_2},t^{m_3}]])\cong
k[x_0,x_1,x_2,y]/I(C)^{*}$ is Cohen-Macaulay.
\end{proof}
\section{Minimal free resolution of the associated graded ring}
In this section, we study the minimal free resolution of
${\rm gr}_m(k[[t^{m_0},t^{m_1},t^{m_2},t^{m_3}]])$ of the certain monomial curve $C$ in affine 4-space.
\begin{theorem}\label{thm3}
Let $C$ be a certain affine monomial curve in $\mathbb{A}^{4}$
having parametrization
\[x_0=t^{m_0}, \; x_1=t^{m_1}, \; x_2=t^{m_2}, \; y=t^{m_3}\]
\noindent $m_0=3a+b$ for positive integers $a \geq 1$ and $b \in [1,3]$
and $0 < m_0 < m_1 < m_2 < m_3$ form an arithmetic progression with common difference $d$.
Then the sequence of R-modules
\begin{center}
$0 \longrightarrow R^{\beta_{3}(b)} \xrightarrow{\phi_{3}(b)} R^ {\beta_{2}(b)} \xrightarrow{\phi_{2}(b)} R^ {\beta_{1}(b)} \xrightarrow{\phi_{1}(b)} R \xrightarrow{\phi}\ G \longrightarrow 0$
\end{center}
is a minimal free resolution for the tangent cone of $C$, where
\begin{eqnarray}
\nonumber
\beta_{1}(b) =\left\{
\begin{array}{ll}
6& if\;\; b=1, \\
5& if\;\; b=2, \\
4& if\;\; b=3,
\end{array}
\right.
,&
\beta_{2}(b) =\left\{
\begin{array}{ll}
8& if\;\; b=1, \\
5& if\;\; b=2, \\
5& if\;\; b=3,
\end{array}
\right.
,&
\beta_{3}(b) =\left\{
\begin{array}{ll}
3& if\;\; b=1, \\
1& if\;\; b=2, \\
2& if\;\; b=3.
\end{array}
\right.
\end{eqnarray}
\noindent and $\phi$'s denote the canonical surjections and the maps between R-modules depend on b
{\footnotesize
$$
\phi_{1}(b=1)=\Big(
\begin{matrix}
g_1=x_{1}^{2}-x_{0}x_{2} &
g_2=x_{1}x_{2}-x_{0}y &
g_3=x_{2}^{2}-x_{1}y &
g_4=x_{1}y^{a}&
g_5=x_{2}y^{a}&
g_6=y^{a+1}
\end{matrix}\Big)
$$
$$
\phi_{2}(b=1)=
\begin{pmatrix}
x_2 & y^{a} & -y & 0 & 0 & 0 & 0 & 0 \\
-x_1& 0 & x_2 & y^{a} & 0 & 0 & 0 & 0 \\
x_0 & 0 & -x_1 & 0 & y^{a} & 0 & 0 & 0 \\
0 & -x_1 & 0 & -x_2 & 0 & x_2 & y & 0 \\
0 & x_0 & 0 & 0 & -x_2 & -x_1 & 0 & y \\
0 & 0 & 0 & x_0 & x_1 & 0 & -x_1& -x_2 \\
\end{pmatrix},
$$
$$
\phi_{3}(b=1)=
\begin{pmatrix}
y^{a} & 0 & 0 \\
-x_2& y & 0 \\
0 & y^{a} & 0 \\
x_1 & -x_2 & 0 \\
-x_0 & x_1 & 0 \\
0 & -x_2 & y \\
0 & x_1 & -x_2 \\
0 & -x_0 & x_1 \\
\end{pmatrix},
$$
}
{\footnotesize
$$
\phi_{1}(b=2)=\Big(
\begin{matrix}
g_1=x_{1}^{2}-x_{0}x_{2} &
g_2=x_{1}x_{2}-x_{0}y &
g_3=x_{2}^{2}-x_{1}y &
g_4=x_{2}y^{a}&
g_5=y^{a+1}
\end{matrix}\Big)
$$
$$
\phi_{2}(b=2)=
\begin{pmatrix}
x_2 & -y & 0 & 0 & 0 \\
-x_1& x_2 & y^{a} & 0 & 0 \\
x_0 & -x_1 & 0 & y^{a} & 0 \\
0 & 0 & -x_1 & -x_2 & y \\
0 & 0 & x_0 & x_1 & -x_2 \\
\end{pmatrix},
$$
$$
\phi_{3}(b=2)=
\begin{pmatrix}
g_{5}=y^{a+1} \\
g_{4}=x_2y^{a} \\
-g_{3}=-x_{2}^{2}+x_{1}y \\
g_{2}= x_{1} x_{2}-x_{0}y \\
g_{1}=x_{1}^{2}-x_{0}x_{2} \\
\end{pmatrix},
$$
}
{\footnotesize
$$
\phi_{1}(b=3)=\Big(
\begin{matrix}
g_1=x_{1}^{2}-x_{0}x_{2} &
g_2=x_{1}x_{2}-x_{0}y &
g_3=x_{2}^{2}-x_{1}y &
g_4=y^{a+1}
\end{matrix}\Big)
$$
$$
\phi_{2}(b=3)=
\begin{pmatrix}
x_2 & y^{a+1} & -y & 0 & 0 \\
-x_1& 0 & x_2 & y^{a+1} & 0 \\
x_0 & 0 & -x_1 & 0 & y^{a+1} \\
0 & -x_{1}^{2}+ x_{0}x_{2} & 0 & -x_{1}x_{2}+x_{0}y & - x_{2}^{2}+ x_{1}y \\
\end{pmatrix},
$$
$$
\phi_{3}(b=3)=
\begin{pmatrix}
y^{a+1} & 0 \\
-x_2 & y \\
0 & y^{a+1} \\
x_1 & -x_2 \\
-x_0 & x_1 \\
\end{pmatrix},
$$
}
\end{theorem}
\begin{proof} We will prove the theorem for the three cases, $b=$ 1, 2, and 3.\\
\noindent {\bf Case \textit{b} $=$ 1.}\\
It is easy to show that $\phi_{1}(1)\phi_{2}(1)=\phi_{2}(1)\phi_{3}(1)=0$
which proves that the above sequence is a complex.
To prove the exactness, we use Corollary 2 of Buchsbaum-Eisenbud theorem in \cite{eisenbud-buchsbaum}.
We have to show that rank $\phi_{1}(1)=1$, rank $\phi_{2}(1)=5$ and rank $\phi_{3}(1)=3$, and also that
$I(\phi_{i}(1))$ contains a regular sequence of length $i$ for all $1 \leq i \leq 3$. rank $\phi_{1}(1)=1$ is trivial.
We want to show that rank $\phi_{2}(1)=5$. Since the columns of the matrix $\phi_{2}(1)$ are related by the
generators of the defining ideal $I(C)$,
note that all $6 \times 6$ minors of $\phi_{2}(1)$ are zero. $\phi_{2}(1)$ has a non zero divisor in the kernel.
By McCoy's theorem rank $\phi_{2}(1) \leq 5$. The determinants of $5 \times 5$ minors of $\phi_{2}(1)$ are
$x_{0}g_{6}^{2}$ when the 6th row and the columns 3, 5 and 6 are deleted, and $x_{1}g_{2}^{2}$ when
the 2nd row and the columns 2, 5 and 8 are deleted. Since $\{x_{0}g_{6}^{2}, x_{1}g_{2}^{2}\}$ are relatively prime, $I(\phi_{2}(1))$ contains a regular sequence of length 2.
Also, among the $3 \times 3$ minors of $\phi_{3}(1)$, we have $\{-x_{0}g_{1}, -x_{1}g_{2},-x_{2}g_{3}\}$. They are relatively prime, so $I(\phi_{3}(1))$ contains a regular sequence of length 3.\\
\noindent {\bf Case \textit{b} $=$ 2.}\\
Clearly $\phi_{1}(2)\phi_{2}(2)=\phi_{2}(2)\phi_{3}(2)=0$ and rank $\phi_{1}(2)=1$ and rank $\phi_{3}(2)=1$.
We have to show that rank $\phi_{2}(2)=4$ and $I(\phi_{i}(1))$ contains a regular sequence of length $i$ for all $1 \leq i \leq 3$. Among the $4 \times 4$ minors of $\phi_{2}(2)$, $I(\phi_{2}(2))$ contains
$\{-g_{1}^2, -g_{2}^2\}$. These two determinants constitute a regular sequence of length 2, since they are relatively prime. \\
\noindent {\bf Case \textit{b} $=$ 3.}\\
As in the previous cases, we have to show that rank $\phi_{1}(3)=1$, rank $\phi_{2}(3)=3$ and rank $\phi_{3}(3)=2$, and also that
$I(\phi_{i}(1))$ contains a regular sequence of length $i$ for all $1 \leq i \leq 3$. rank $\phi_{1}(3)=1$ is trivial.
We have to show that rank $\phi_{2}(3)=3$. $\phi_{2}(3)$ has a non zero divisor in the kernel.
By McCoy's theorem rank $\phi_{2}(3) \leq 3$. Among the $3 \times 3$ minors of $\phi_{2}(3)$, $I(\phi_{2}(3))$ contains
$\{g_{1}^2, g_{2}^2\}$ which is a regular sequence of length 2, since they are relatively prime.
Also, among the $2 \times 2$ minors of $\phi_{3}(3)$, we have $\{g_{1}, -g_{2}, g_{3}\}$. They are
relatively prime, so $I(\phi_{3}(3))$ contains a regular sequence of length 3.
\end{proof}
\begin{corollary} Under the hypothesis of Theorem 3.1., the minimal graded free resolution of the associated graded ring $G$ is given by
if $b=$1
\begin{center}
$0 \longrightarrow R(-(a+3))^{3} \xrightarrow{\phi_{3}(b)}
R(-3)^{2} \bigoplus R(-(a+2))^{6} \xrightarrow{\phi_{2}(b)} R(-2)^{3}\bigoplus R(-(a+1))^{3}
\xrightarrow{\phi_{1}(b)} R $
\end{center}
if $b=$2
\begin{center}
$0 \longrightarrow R(-(a+4)) \xrightarrow{\phi_{3}(b)} R(-3)^2 \bigoplus R(-(a+2))^{3} \xrightarrow{\phi_{2}(b)} R(-2)^{3}\bigoplus R(-(a+1))^{2}
\xrightarrow{\phi_{1}(b)} R $
\end{center}
if $b=$3
\begin{center}
$0 \longrightarrow R(-(a+4))^{2} \xrightarrow{\phi_{3}(b)} R(-3)^{2} \bigoplus R(-(a+3))^{3} \xrightarrow{\phi_{2}(b)} R(-2)^{3}\bigoplus R(-(a+1))
\xrightarrow{\phi_{1}(b)} R $
\end{center}
\end{corollary}
\begin{flushright}
$\Box$
\end{flushright}
If $H_{G}(i)=dim_{k}(m^{i}/m^{i+1})$ is the Hilbert function of $G$, then
\begin{corollary} Under the hypothesis of Theorem 3.1., the Hilbert function of the associated graded ring $G$ is given by\\
if $b=$ 1\\
\noindent $H_{G}(i)= \left(\!\!\!\begin{array}{c} i+3 \\ 3 \end{array}\!\!\! \right)
-3\left(\! \!\! \begin{array}{c} i+1 \\ 3 \end{array}\!\!\! \right)
-3\left(\!\!\! \begin{array}{c} i-a+2 \\ 3\end{array}\!\!\! \right)
+2 \left(\!\!\! \begin{array}{c} i \\ 3 \end{array}\!\!\! \right)
+6\left(\!\!\! \begin{array}{c} i-a+1 \\ 3 \end{array}\!\!\! \right)
-3\left(\!\!\! \begin{array}{c} i-a \\ 3 \end{array}\!\!\! \right)$\\
if $b=$ 2\\
\noindent $H_{G}(i)= \left(\!\!\! \begin{array}{c} i+3 \\ 3 \end{array}\!\!\! \right)
-3\left(\!\!\! \begin{array}{c} i+1 \\ 3 \end{array}\!\!\! \right)
-2\left(\!\!\! \begin{array}{c} i-a+2 \\ 3\end{array}\!\!\! \right)
+2 \left(\!\!\! \begin{array}{c} i \\ 3 \end{array}\!\!\! \right)
+3\left(\!\!\! \begin{array}{c} i-a+1 \\ 3 \end{array}\!\!\! \right)
-\left(\!\!\! \begin{array}{c} i-a-1 \\ 3 \end{array}\!\!\! \right)$\\
if b=3\\
\noindent $H_{G}(i)= \left(\!\!\! \begin{array}{c} i+3 \\ 3 \end{array}\!\!\! \right)
-3\left(\!\!\! \begin{array}{c} i+1 \\ 3 \end{array}\!\!\! \right)
-\left(\!\!\! \begin{array}{c} i-a+2 \\ 3\end{array}\!\!\! \right)
+2\left(\!\!\! \begin{array}{c} i \\ 3 \end{array}\!\!\! \right)
+3\left(\!\!\! \begin{array}{c} i-a \\ 3 \end{array}\!\!\! \right)
-2\left(\!\!\! \begin{array}{c} i-a-1 \\ 3 \end{array}\!\!\! \right)$
\end{corollary}
\vspace{5mm}
\bibliographystyle{amsplain}
|
\section{Introduction}
Let $G=\left(V, A\right)$ be an undirected graph where~$V$ is the set of~$N< \infty$ nodes and~$A\in \left\{0,1\right\}^{N\times N}$ is the underlying adjacency matrix. Our goal in this document is to construct a preorder on the (finite) set of nodes~$V$ of~$G$ that conveys non-trivial structure: 1) it is preserved by the logistic dynamical system
\begin{equation}
\frac{d}{dt}y_{i}(t) = \left(\sum_{j\sim i}\gamma y_{j}(t)\right)\left(1-y_i(t)\right)-y_i(t)\label{eq:logistic}
\end{equation}
for $i=1,\ldots,N$, in a sense to be made precise; and 2) its underlying equivalence relation coincides with the coarsest equitable partition of a graph. The logistic dynamics consists of a standard set of coupled Ordinary Differential Equations (ODEs) that model the spread of epidemics in networks (e.g.,~\cite{porterdynamical,newman,augusto_qualitative,kurtzi,Melanie}). The parameter~$\gamma$ stands for the rate of infection of a virus in the network; the state variable~$y_i(t)$ models the likelihood of infection of an individual~$i$ at time~$t$, due to transmission from its peers. The state-variable~$y_i(t)$ can be also interpreted as the fraction of individuals infected in a community (or sub-graph click\footnote{In a click, all nodes are connected to all other nodes.})~$i$ in a network of interconnected communities (a.k.a., super-network). The sum in equation~\eqref{eq:logistic} runs over the neighbors of~$i$ and that is what accounts for the underlying network structure in the dynamics.
A preorder on the set of nodes is a binary relation $\succeq$ on $V$ that is reflexive and transitive, i.e.,
$\bullet$ $i \succeq i$ (reflexive)
$\bullet$ $i \succeq i' \succeq i'' \Rightarrow i\succeq i''$ (transitive).
Any preorder on a set $V$ naturally induces a quotient on $V$ given by the following equivalence relation
\begin{equation}
i\cong j \overset{{\sf def.}}\Longleftrightarrow i\succeq j \mbox{ and } j\succeq i.\label{eq:equivale}
\end{equation}
Since `$\succeq$' is not a partial order, the right hand side of~\eqref{eq:equivale} does not imply equality~$i=j$ (up to isomorphism), but a weaker equivalence relation. In fact, we will show that our preorder construct induces (in light of equation~\eqref{eq:equivale}) the so-called coarsest equitable partition (CEP), which is an important coloring or symmetry on a graph and that we will introduce momentarily.
It is clear that the logistic dynamical system preserves the automorphisms of a graph -- in that, if isomorphic nodes are initialized with the same initial condition, then their state evolve in synchrony -- but it is non-trivial to show that it preserves the CEP of the graph (this will follow as a corollary to the fact that it preserves the preorder~`$\succeq$' constructed in this paper).
\textbf{Outline of the paper.} Section~\ref{sec:symmetry} introduces and contrast the definitions of isomorphism and the coarsest equitable partition on graphs; Section~\ref{sec:construct} constructs the preorder; Section~\ref{sec:preserva} shows the implications of the preorder on dynamical systems such as the logistic system, in particular, the logistic dynamical system preserves the preorder and hence admits a non-trivial invariant set; Section~\ref{sec:conclusion} concludes the paper.
\vspace*{2ex}
\textbf{Preliminary Notation.}
\vspace*{2ex}
\begin{itemize}
\item $\mathbb{N}=\left\{1,2,3,\ldots\right\}$;
\item $V(G)=$ set of nodes of graph~$G$ -- sometimes denoted simply by~$V$;
\item $\mathcal{N}(i)=$ set of neighbors to the node $i$ (also named neighborhood to $i$) in the undirected graph $G=\left(V,E\right)$;
\item $d(i)=$ degree of the node $i$ in the undirected graph $G=\left(V,E\right)$;
\item $\mathbf{d}(\mathbf{p})=$ vector collecting the degrees of the nodes stacked in the vector $\mathbf{p}\in V^{\ell}$, assuming that $G=\left(V,E\right)$ is undirected;
\item $\pi_{I}^{\ell}\,:\,V^{\ell+1}\rightarrow V^{\left|I\right|}:$ is the canonical projection on the coordinates indexed by the set~$I\subset \left\{1,\ldots,\ell\right\}$.
\end{itemize}
\section{Symmetries on Graphs}\label{sec:symmetry}
In this Section, we define some notions of symmetry on graphs and present the exact notion that fits our purposes, namely, the coarsest equitable partition (CEP). We start by introducing the definition of isomorphism on graphs.
\begin{definition}[Graph Isomorphism]\label{def:isomorphgraph}
We say that two graphs~$\mathcal{G}_1=\left(V_1,E_1\right)$ and~$\mathcal{G}_2=\left(V_2,E_2\right)$ are isomorphic, and represent it by~$\mathcal{G}_1\cong \mathcal{G}_2$, whenever there exists a bijection~$f\,:\,V_1\rightarrow V_2$ with
\begin{equation}
i\sim j \Leftrightarrow f(i)\sim f(j)\nonumber
\end{equation}
where~`$i \sim j$' means that~$i$ and~$j$ are neighbors in the graph.
\end{definition}
\begin{definition}[Graph Automorphism]\label{def:automorphgraph}
We call~$f\,:V\,\rightarrow V$ an automorphism on~$G=\left(V,E\right)$ if~$f$ is an isomorphism from~$G$ onto~$G$, i.e.,~$f$ is bijective and
\begin{equation}
i\sim j \Leftrightarrow f(i)\sim f(j).\nonumber
\end{equation}
\end{definition}
We say that two nodes~$i$ and~$j$ in a graph are isomorphic, and denote it as~$i\approx j$, whenever there exists an automorphism~$f$ with~$f(i)=j$ or~$f(j)=i$. The set of automorphisms on a graph~$G$ endowed with the operation of composition conforms to a group that we refer to as~${\sf Aut}(G)$.
\begin{definition}[Orbits]\label{def:orbit}
Let~$i$ be a node in~$G$. We refer to
\begin{equation}
[i]\overset{\Delta}={\sf Aut}(G).i=\left\{f(i)\,:\,f\in {\sf Aut}(G)\right\}\nonumber
\end{equation}
as the orbit of the vertex~$i$. In words, two nodes are in the same orbit if and only if they are isomorphic.
\end{definition}
The orbits of a graph in definition~\ref{def:orbit} define an equivalence relation between nodes which allows to quotient the set of nodes into equivalence classes
\begin{equation}
\left[i\right]\overset{\Delta}=\left\{j\in V\,:\, j\approx i\right\}.\nonumber
\end{equation}
Fig.~\ref{fig:quotient} illustrates the chromatic partition of a graph according to this equivalence relation. Monochromatic graphs under this equivalence relation are referred to as vertex-transitive graphs.
\begin{figure} [hbt]
\begin{center}
\includegraphics[scale= 0.5]{network_classes.pdf}
\caption{Nodes with the same color are isomorphic nodes.}\label{fig:quotient}
\end{center}
\end{figure}
It is natural to expect that the partition illustrated in Fig.~\ref{fig:quotient} is preserved by the logistic dynamical system in the sense that if all nodes have the same initial condition, i.e.,~$\mathbf{y}(0)=\mathbf{1}_N y_0$, then, there is no reason to expect that eventually, say one of the indigo nodes will increase its state (e.g., degree of infection in an epidemics) faster than the other indigo node. Indeed, and more formally, we can show that such dynamical system preserves the partition induced by the orbits of the automorphism group of a graph. In particular, the underlying dynamical system admits a lower dimensional version quotiented by the automorphism symmetries of the graph.
Fig.~\ref{fig:frucht} illustrates a regular graph with degree~$3$, so-called Frucht graph, that entails some asymmetry. For instance, its group of automorphisms is the trivial one (for more details, refer to~\cite{implementing}), i.e., no two nodes are isomorphic in the Frucht graph. In other words, the Frucht graph is a regular graph that is not vertex-transitive. From such asymmetry, it is not clear whether nodes initialized at the same state will evolve evenly under an epidemics modeled by the logistic dynamics.
\begin{figure} [hbt]
\begin{center}
\includegraphics[scale= 0.4]{Frucht.pdf}
\caption{Frucht Graph.}\label{fig:frucht}
\end{center}
\end{figure}
Despite the asymmetry, all nodes are equivalent to each other in the CEP (or also in our preorder construct) sense. Therefore, as corollary to the results proved in this paper, nodes departing from the same state, in any regular graph (not necessarily vertex-transitive), will evolve in synchrony -- in particular, a regular network admits a $1$-dimensional dynamics version.
\textbf{Remark.} Note that the synchrony induced by the CEP on the logistic dynamics shall not be taken for granted as one can find counter-examples of dynamical systems whose qualitative properties depend on the nodes even for a regular graph (that is not vertex-transitive), refer, e.g., to~\cite{aldous1989}.
The previous discussion illustrates that even though the orbits of an automorphism are preserved by the logistic and the like dynamical systems, this is not the coarsest coloring that is preserved. Next, we introduce the coarsest equitable partition (for more details refer to~\cite{fractional}). First, let~$\mathcal{P}=\left\{\mathcal{P}_1,\ldots,\mathcal{P}_k\right\}$ be a partition on~$V$. Define
\begin{equation}
s_{\ell}(i) = \left|\mathcal{N}(i)\cap \mathcal{P}_{\ell}\right|\nonumber
\end{equation}
as the number of neighbors of~$i$ in the partition~$\ell$, and define~
\begin{equation}
\mathbf{s}(i)=\left(s_1(i),\ldots,s_{d(i)}(i)\right),\nonumber
\end{equation}
as the distribution of the~$d(i)$ neighbors of~$i$ across the~$\ell$ elements of the partition.
\begin{definition}[Equitable partition]
The partition~$\mathcal{P}$ is called equitable whenever
\begin{equation}
i\approx_{\mathcal{P}} j \Leftrightarrow \mathbf{s}(i)=\mathbf{s}(j).\nonumber
\end{equation}
\end{definition}
In words, two nodes~$i$ and $j$ lie in the same class if and only if they have the same distribution of neighbors across classes. In particular, they have the same degree. The partition induced by the orbits of the group of automorphisms in a graph is an equitable partition, but not the coarsest one in the graph. As referred to before, no two nodes in the Frucht graph in Fig.~\ref{fig:frucht}, which is a regular graph, are equivalent w.r.t. the group of automorphism whereas all nodes are equivalent w.r.t. our preorder construct equivalence, which is equitable and corresponds to the coarsest equitable partition. In certain cases, e.g., when the underlying graph~$G$ is a tree, then the orbits of the automorphism group and the coarsest equitable partition coincide. The next lemma asserts that the \textbf{coarsest} equitable partition exists in any (finite) graph.
\begin{lemma}
Let~$\mathcal{P}$ and~$\mathcal{Q}$ be two equitable partitions on~$G$. Then, there exists the finest equitable partition that is coarser than~$\mathcal{P}$ and~$\mathcal{Q}$ and it is denoted as
\begin{equation}
\mathcal{P}\vee\mathcal{Q}\nonumber
\end{equation}
\end{lemma}
We now present the characterization of the coarsest equitable partition on a graph~$G$, given by the multisets of the so-called iterated degree sequence of the nodes. First, let
\begin{eqnarray}
d^{(1)}(i) & \overset{\Delta}= & \left\{d(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber\\
d^{(2)}(i) & \overset{\Delta}= & \left\{d^{(1)}(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber\\
& \vdots & \nonumber\\
d^{(k+1)}(i) & \overset{\Delta}= & \left\{d^{(k)}(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber
\end{eqnarray}
For instance, from Fig.\ref{fig:quotient} we have
\begin{eqnarray}
d^{(1)}(2) & = & \left\{2,3\right\}\nonumber\\
d^{(2)}(2) & = & \left\{\left\{2,2\right\},\,\left\{1,2,2\right\}\right\}\nonumber\\
\end{eqnarray}
One can extend it to the infinitary sequence, so-called ultimate degree sequence
\begin{equation}
D\left(i\right)\overset{\Delta}=\left(d(i),d^{(1)}(i),d^{(2)}(i),\ldots,d^{(k)}(i),\ldots\right)\nonumber
\end{equation}
Now,~$\mathcal{P}$ is the coarsest equitable partition in a graph~$G$ if and only if
\begin{equation}
i\approx_{\mathcal{P}} j \Leftrightarrow D(i)=D(j).\nonumber
\end{equation}
Observe in particular that all nodes in a regular graph fall in the same class with respect to the coarsest equitable partition.
We rephrase a corollary to Theorem~$6.5.1$ in~\cite{fractional} as follows.
\begin{corollary}
Let~$\mathcal{P}$ be the partition induced on the set of nodes~$V$ by the iterated degree sequence on the graph~$G=\left(V,E\right)$. Let~$\mathcal{Q}$ be an equitable partition on~$V$. Then,~$\mathcal{P}$ is coarser than~$\mathcal{Q}$.
\end{corollary}
There are efficient algorithms to determine the coarsest equitable partition in~$O(N\log(N))$. Determining the finer partition induced by the orbits of the automorphism group is NP-hard. We reserve the symbol~$\approx$ as the equivalence relation associated to the coarsest equitable coloring.
\section{Preorder-construct on simple undirected graphs}\label{sec:construct}
In this Section, we construct a preorder on the set of nodes of an undirected graph~$G=\left(V,E\right)$. The idea of the preorder is to formalize the following domination concept: a node~$i$ is greater or dominates a node~$j$ whenever the degree of~$i$ is greater than the degree of~$j$; and the degrees of the neighbors of~$i$ are greater than the degrees of the neighbors of~$j$; and the degrees of the neighbors of the neighbors of~$i$ are greater than the degrees of the neighbors of the neighbors of~$j$ and so forth. This idea must be clearly formalized to make any sense.
Define
\begin{equation}
\mathcal{P}_{\ell}(i)\overset{\Delta}= \left\{\mathbf{p}^{\ell}=\left(i,i_1,\ldots,i_{\ell}\right)\in V^{\ell+1}\,:\,i_{k}\sim i_{k+1}\,\,\forall{k\leq \ell-1}\right\}\nonumber,
\end{equation}
as the set of all paths, on the graph~$G$, departing from the node~$i$ and with length~$\ell$.
Now, we introduce one of the main concepts for the preorder construct.
\begin{definition}[$\ell$-adapted function]
Let~$i,j\in V(G)$ and~$\ell \in \mathbb{N}$. We define the class of $\ell$-adapted functions on~$G$ at the pair~$\left(j,i\right)$ as the set of functions
\begin{equation}
\begin{array}{rcl} f_{ji}^{\ell}: \mathcal{P}_{\ell}(j) & \to & \mathcal{P}_{\ell}(i)\\ \left(j,j_1,\ldots,j_{\ell}\right) & \mapsto & \left(i,i_1,\ldots,i_{\ell}\right) \end{array}\nonumber
\end{equation}
fulfilling the next two properties
\begin{eqnarray}
\pi^{\ell}_{\leq m}\left(\mathbf{p}^{\ell}\right) & = & \pi^{\ell}_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\
& \Downarrow & \\
\pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & = & \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right),\nonumber
\end{eqnarray}
and
\begin{eqnarray}
\pi^{\ell}_{\leq m}\left(\mathbf{p}^{\ell}\right) & \neq & \pi^{\ell}_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\
& \Downarrow & \label{eq:notequal}\\
\pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & \neq & \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right),\nonumber
\end{eqnarray}
for all~$\mathbf{p}^{\ell},\mathbf{\widetilde{p}}^{\ell}\in \mathcal{P}_{\ell}(j)$, where we defined
\begin{equation}
\pi^{\ell}_{\leq m}\,:\,V^{\ell+1}\longrightarrow V^{m+1}\nonumber
\end{equation}
with~$\pi^{\ell}_{\leq m}\left(x_0,x_1,\ldots,x_{\ell}\right)= \left(x_0,x_1,\ldots,x_{m}\right)$, i.e.,~$\pi_{\leq m}$ returns the first~$m\leq \ell$ coordinates from the input vector.
\end{definition}
In words, if the first~$m$ coordinates of a path~$\mathbf{p}^{\ell}$ coincide with the first~$m$ coordinates of a path~$\mathbf{\widetilde{p}}^{\ell}$, then the same should hold for the images of~$\mathbf{p}^{\ell}$ and~$\mathbf{\widetilde{p}}^{\ell}$ by the map~$f^{\ell}_{ij}$. For simplicity, from now one, we will commit an abuse of notation and we will write~$\pi_{\leq m}$ instead of~$\pi_{\leq m}^{\ell}$.
We assume throughout that~$f^{\ell}_{ii}$ is the identity map for any~$i\in V(G)$, which is clearly an adapted function. We now observe two important properties.
\begin{lemma}[``Group" property]
Let~$g^{\ell}_{jk}$ and~$h^{\ell}_{ki}$ be two $\ell$-adapted functions at~$\left(j,k\right)$ and~$\left(k,i\right)$, respectively. Then,
\begin{equation}
f^{\ell}_{ji}\overset{\Delta}=h^{\ell}_{ki}\circ g^{\ell}_{jk}\nonumber
\end{equation}
is $\ell$-adapted at~$\left(j,i\right)$.
\end{lemma}
\begin{proof}
Indeed, let~$\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in \mathcal{P}_{\ell}(j)$ with
\begin{equation}
\pi_{\leq m}\left(\mathbf{p}^{\ell}\right)= \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber
\end{equation}
for some~$m\leq \ell$, then
\begin{equation}
\pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber
\end{equation}
as~$g^{\ell}_{jk}$ is $\ell$-adapted and thus,
\begin{equation}
\pi_{\leq m}\left(f^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(f^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber
\end{equation}
as~$h^{\ell}_{jk}$ is $\ell$-adapted. Similarly, we can establish the property in equation~\eqref{eq:notequal}.
\end{proof}
\begin{lemma}\label{lem:invertible}
If~$f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is an $\ell$-adapted invertible function, then~$g^{\ell}_{ij}\overset{\Delta}=\left(f^{\ell}_{ji}\right)^{-1}$ is an $\ell$-adapted function.
\end{lemma}
\begin{proof}
Assume that~$g^{\ell}_{ij}$ is not an adapted function. Then, either one of the cases below should hold true
\begin{enumerate}
\item $\pi_{\leq m}\left(\mathbf{p}^{\ell}\right)= \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)$, but $\pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)\neq \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)$ for some~$m$ and $\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in\mathcal{P}_{\ell}(i)$.
\item $\pi_{\leq m}\left(\mathbf{p}^{\ell}\right)\neq \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)$, but $\pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)$ for some~$m$ and $\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in\mathcal{P}_{\ell}(i)$.
\end{enumerate}
Now, case 1) implies that
\begin{eqnarray}
\pi_{\leq m}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\right) & = & \pi_{\leq m}\left(\mathbf{p}^{\ell}\right)\nonumber\\
& = & \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\
& = & \pi_{\leq m}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\right)\nonumber
\end{eqnarray}
which is a contradiction as
\begin{equation}
\pi_{\leq m}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\neq \pi_{\leq m}\left(g^{\ell}_{ij}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber
\end{equation}
and~$f^{\ell}_{ji}$ is an adapted function. Case 2) similarly leads to a contradiction, therefore the inverse map~$g^{\ell}_{ij}$ is an $\ell$-adapted function.
\end{proof}
We now observe that if~$f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is an adapted function, then its restriction to the $(\ell-1)$-paths~$\mathcal{P}_{\ell-1}(j)$, denoted as~$f^{\ell-1}_{ji}$, is an $(\ell-1)$-adapted function. Moreover,~$f^{\ell}_{ji}$ extends~$f^{\ell-1}_{ji}$ in that if~$\mathbf{p}^{\ell-1}=\pi_{\leq \ell-1}\left(\mathbf{p}^{\ell}\right)$, then
\begin{equation}
\pi_{\leq \ell-1}\left(f^{\ell}_{ji}(\mathbf{p}^{\ell})\right)=f^{\ell-1}_{ji}\left(\mathbf{p}^{\ell-1}\right)\label{eq:extends}
\end{equation}
or else, if~$\mathbf{p}^{\ell-1} \neq \pi_{\leq \ell-1}\left(\mathbf{p}^{\ell}\right)$, then
\begin{equation}
\pi_{\leq \ell-1}\left(f^{\ell}_{ji}(\mathbf{p}^{\ell})\right)\neq f^{\ell-1}_{ji}\left(\mathbf{p}^{\ell-1}\right)\label{eq:extends2}
\end{equation}
\begin{definition}[Adapted sequence]
We call
\begin{equation}
\left(f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)\right)_{\ell}\nonumber
\end{equation}
an adapted sequence (on~$\ell$) of adapted functions at the pair~$\left(j,i\right)$ if~$f^{\ell+1}_{ji}$ extends~$f^{\ell}_{ji}$ for any~$\ell \in \mathbb{N}$ in the sense of equations~\eqref{eq:extends}-\eqref{eq:extends2}.
\end{definition}
The following Lemma remarks that if the adapted sequence is comprised of invertible maps, then the inverse maps sequence is an adapted sequence as well.
\begin{lemma}[Inverse sequence]\label{lem:inversesequence}
Let~$\left(f^{\ell}_{ji}\right)_{\ell}$ be an adapted sequence with $f^k_{ji}$ being an invertible $k$-adapted function for all~$k$. Then, the sequence~$\left(g^{\ell}_{ij}\right)_{\ell}$, where~$g^{\ell}_{ij}\overset{\Delta}=\left(f^{\ell}_{ji}\right)^{-1}$, is an adapted sequence.
\end{lemma}
\begin{proof}
Note that each~$g^{\ell}_{ij}$ is an adapted function from Lemma~\ref{lem:invertible}. We are just left to prove that~$g^{\ell+1}_{ij}$ extends~$g^{\ell}_{ij}$ for all~$\ell$, in the sense of equations~\eqref{eq:extends}-\eqref{eq:extends2}. Suppose that~$\left(g^{\ell}_{ij}\right)_{\ell}$ is not an adapted sequence. Then, there exists an~$\ell\in \mathbb{N}$ so that~$g^{\ell+1}_{ij}$ does not extend~$g^{\ell}_{ij}$, or in other words, one of the conditions below hold true
\begin{enumerate}
\item $\mathbf{p}^{\ell}=\pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, but $\pi_{\leq \ell}\left(g^{\ell+1}_{ji}(\mathbf{p}^{\ell+1})\right) \neq g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)$;
\item $\mathbf{p}^{\ell}\neq \pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, but $\pi_{\leq \ell}\left(g^{\ell+1}_{ji}(\mathbf{p}^{\ell+1})\right) = g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)$.
\end{enumerate}
Define~$\mathbf{\widetilde{p}^{\ell}}\overset{\Delta}=g_{ij}^{\ell}\left(\mathbf{p}^{\ell}\right)$ and~$\mathbf{\widetilde{p}}^{\ell+1}\overset{\Delta}=g_{ij}^{\ell+1}\left(\mathbf{p}^{\ell+1}\right)$. Case 1) implies that
\begin{equation}
f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\neq \pi_{\leq \ell}\left(f^{\ell+1}_{ji}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)\right),\nonumber
\end{equation}
but~$f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)=\mathbf{p}^{\ell}$ and~$f^{\ell+1}_{ji}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)=\mathbf{p}^{\ell+1}$ which contradicts the assumption. Case 2) leads similarly to a contradiction. Therefore, the inverse sequence~$\left(g^{\ell}_{ij}\right)_{\ell}$ is an adapted sequence.
\end{proof}
\begin{theorem}[``Group" property for adapted sequences]\label{th:groupadaptedseq}
Let~$\left(h^{\ell}_{ki}\right)$ and~$\left(g^{\ell}_{jk}\right)$ be two adapted sequences. Then, the pointwise composition
\begin{equation}
\left(f^{\ell}_{ji}\overset{\Delta}= h^{\ell}_{ki} \circ g^{\ell}_{jk}\right)_{\ell}\nonumber
\end{equation}
is an adapted sequence.
\end{theorem}
\begin{proof}
The sequence~$\left(f^{\ell}_{ji}\right)$ must fulfill the defining conditions in equations~\eqref{eq:extends}-\eqref{eq:extends2}. Let~$\mathbf{p}^{\ell}=\pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, with~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$ and~$\mathbf{p}^{\ell+1}\in\mathcal{P}_{\ell+1}(j)$. Then,
\begin{equation}
\pi_{\leq \ell}\left(g^{\ell+1}_{jk}\left(\mathbf{p}^{\ell+1}\right)\right)=g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\label{eq:g}
\end{equation}
as~$\left(g^{\ell}_{jk}\right)_{\ell}$ is an adapted sequence. Moreover, define~$\mathbf{\widetilde{p}}^{\ell+1}\overset{\Delta}= g^{\ell+1}_{jk}\left(\mathbf{p}^{\ell+1}\right)$ and~$\mathbf{\widetilde{p}}^{\ell}\overset{\Delta}= g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)$. Thus, equation~\eqref{eq:g} can be written as
\begin{equation}
\pi_{\leq \ell}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)=\mathbf{\widetilde{p}}^{\ell}\nonumber
\end{equation}
which implies that
\begin{equation}
\pi_{\leq \ell}\left(h^{\ell+1}_{jk}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)\right)=h^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber
\end{equation}
and hence,
\begin{equation}
\pi_{\leq \ell}\left(f^{\ell+1}_{ji}\left(\mathbf{p}^{\ell+1}\right)\right)=f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right).\nonumber
\end{equation}
The second case in equation~\eqref{eq:extends2} holds similarly and~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence.
\end{proof}
Now, we introduce the main construct of this document.
\begin{definition} [Preorder]\label{def:preorder}
Let~$i,j\in V(G)$. We say that~$i\succeq j$, whenever there exists an adapted sequence
\begin{equation}
\left(f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)\right)_{\ell}\nonumber
\end{equation}
where both properties hold
\begin{itemize}
\item $f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is injective;
\item $\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\geq \mathbf{d}\left(\mathbf{p}^{\ell}\right)$
\end{itemize}
for all~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$ and all~$\ell \in\mathbb{N}$.
\end{definition}
We say that an adapted sequence~$\left(f^{\ell}_{ji}\right)$ \textbf{explains} the inequality~$i\succeq j$, whenever~$\left(f^{\ell}_{ji}\right)$ fulfills the two conditions in the definition~\ref{def:preorder}. We may also refer that $i\succeq j$ is \textbf{explained} by~$\left(f^{\ell}_{ji}\right)$ under these conditions.
\begin{theorem}
The binary relation~$\succeq$ in definition~\ref{def:preorder} is a preorder on the set of nodes~$V(G)$ of the graph~$G$, i.e., it is reflexive and transitive:
\begin{itemize}
\item \textbf{(Reflexivity)} $i\succeq i$;
\item \textbf{(Transitivity)} $i \succeq k \succeq j \Rightarrow i \succeq j$.
\end{itemize}
\end{theorem}
\begin{proof}
\underline{\textbf{Reflexivity:}} For any node~$i\in V(G)$, consider the sequence of identity maps
\begin{equation}
f^{\ell}_{ii}\,:\,\mathcal{P}_{\ell}(i) \rightarrow \mathcal{P}_{\ell}(i)\nonumber
\end{equation}
and note that this is an adapted sequence, with injective maps that preserve the degree.
\underline{\textbf{Transitivity:}} Let~$\left(h^{\ell}_{ki}\right)$ and~$\left(g^{\ell}_{jk}\right)$ explain~$i\succeq k$ and~$k\succeq j$, respectively, and define the sequence
\begin{equation}
\left(f^{\ell}_{ji}\overset{\Delta}= h^{\ell}_{ki} \circ g^{\ell}_{jk}\right)_{\ell}.\nonumber
\end{equation}
First, note that from Theorem~\ref{th:groupadaptedseq}, the sequence~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence. Since the composition of injective maps is injective, then each~$f^{\ell}_{ji}$ is injective. Now,
\begin{eqnarray}
\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & = & \mathbf{d}\left(h^{\ell}_{ki}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)\right)\nonumber\\
& \geq & \mathbf{d}\left(g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\nonumber\\
& \geq & \mathbf{d}\left(\mathbf{p}^{\ell}\right)\nonumber
\end{eqnarray}
and thus,~$i\succeq j$ (explained by~$\left(f^{\ell}_{ji}\right)$).
\end{proof}
Now, we remark that to each preorder, there is an associated equivalence relation defined as
\begin{equation}
i\cong j \Leftrightarrow \left(i\succeq j \wedge j\succeq i \right).\nonumber
\end{equation}
Checking that this is an equivalence relation is trivial. The next Theorem, provides a more explicit characterization for our preorder~$\succeq$.
\begin{theorem}\label{th:characterization}
$i \cong j$, if and only if there exists an adapted sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ such that
\begin{itemize}
\item $f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is bijective;
\item $\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)=\mathbf{d}\left(\mathbf{p}^{\ell}\right)$;
\end{itemize}
for all~$\mathbf{p}^{\ell} \in \mathcal{P}_{\ell}(j)$ and for all~$\ell$.
\end{theorem}
\begin{proof}
The \textbf{if} part is trivial: it follows as a corollary to Lemma~\ref{lem:inversesequence}.
We are left to prove the \textbf{only if} part. Let~$i\cong j$. Then, by definition,~$i\succeq j$ and~$j\succeq i$. In other words, there exist two adapted sequences~$\left(f_{ji}^{\ell}\right)$ and~$\left(g_{ij}^{\ell}\right)$ that explain~$i\succeq j$ and~$j\succeq i$, respectively. First, note that since~$f_{ji}^{\ell}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ and~$g_{ij}^{\ell}\,:\,\mathcal{P}_{\ell}(i)\rightarrow \mathcal{P}_{\ell}(j)$ are injective maps, then~$\left|\mathcal{P}_{\ell}(i)\right|=\left|\mathcal{P}_{\ell}(j)\right|$ for all~$\ell$ (Schr\"{o}der-Bernstein Theorem) and thus, since the graph is finite, we have~$\left|\mathcal{P}_{\ell}(i)\right|=\left|\mathcal{P}_{\ell}(j)\right|< \infty$, and the injective maps~$f_{ji}^{\ell}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ and~$g_{ij}^{\ell}\,:\,\mathcal{P}_{\ell}(i)\rightarrow \mathcal{P}_{\ell}(j)$ are in fact bijections.
We are to show that~$f^{\ell}_{ji}$ (or~$g^{\ell}_{ij}$) preserves the degree sequence. Define the following interlaced dynamics on~$\mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)$
\begin{itemize}
\item If $\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$: $\mathbf{p}^{\ell}\rightarrow f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\rightarrow g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\rightarrow f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\right)\rightarrow \ldots$
\item If $\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(i)$: $\mathbf{p}^{\ell}\rightarrow g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\rightarrow f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\rightarrow g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\right)\rightarrow \ldots$
\end{itemize}
We first remark that the vector degree is monotonous over the orbits of the above interlaced dynamical system. Therefore, if~$\mathbf{p}^{\ell}$ is periodic, then
\begin{itemize}
\item $\mathbf{d}\left(\mathbf{p}^{\ell}\right)=\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)$, if~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$;
\item $\mathbf{d}\left(\mathbf{p}^{\ell}\right)=\mathbf{d}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)$, if~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(i)$.
\end{itemize}
In other words, if~$\mathbf{p}^{\ell}\in \mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)$ is a periodic point (or path) w.r.t. the interlaced dynamics, then the degree is preserved by the bijections~$f^{\ell}_{ji}$ and~$g^{\ell}_{ij}$. We claim that any point (or path)~$\mathbf{p}$ is periodic. Indeed, let~$\mathbf{p}\in \mathcal{P}_{\ell}(i)$ and assume that~$\mathbf{p}$ is not periodic. This in particular implies that all the points in the backwards iterates
\begin{equation}
\left(f^{\ell}_{ji}\right)^{-1}(\mathbf{p}),\left(g^{\ell}_{ji}\right)^{-1}\left(\left(f^{\ell}_{ji}\right)^{-1}(\mathbf{p})\right),\ldots\nonumber
\end{equation}
are not periodic as well. In other words, the set of all backwards iterates has infinite cardinality, which is a contradiction as~$\left|\mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)\right|<\infty$. Therefore, we conclude that~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence of bijections preserving the vector degree.
\end{proof}
\begin{theorem}
The equivalence relation~$\cong$ is equal to the coarsest equitable partition (CEP).
\end{theorem}
\begin{proof}
We first show that~$\cong$ conforms to an equitable relation. Then, we show that any other equitable partition in the graph is finer then the coloring of~$\cong$.
\textbf{Part I: $\cong$ is equitable.}
Let~$i\cong j$. Take an arbitrary neighbour~$j_1\in\mathcal{N}(j)$ of~$j$ and let us show that there exists~$i_1\in \mathcal{N}(i)$ with~$j_1\cong i_1$. This establishes the first part.
Indeed, let~$i_1\overset{\Delta}=\pi_{2}\left(f^{1}_{ji}\left(j,j_1\right)\right)$. Our claim is that~$j_1\cong i_1$. Define~$g^{\ell}_{j_1i_1}(\mathbf{p})\overset{\Delta}=\pi_{2\leq \cdot \leq \ell+1}\left(f^{\ell+1}_{ji}(j,\mathbf{p})\right)$ for all~$\mathbf{p}\in \mathcal{P}_{\ell}(j_1)$. We observe that~$\left(g^{\ell}_{j_1i_1}\right)$ fulfills the conditions of Theorem~\ref{th:characterization} and thus,~$j_1\cong i_1$.
\textbf{Part II: $\cong$ is the coarsest equitable relation.}
We will prove that given another equitable relation~$\approx$, then
\begin{equation}
i\approx j \Rightarrow i\cong j.\nonumber
\end{equation}
For each pair~$n,m$ of $\approx$-equivalent nodes~$n\approx m$, define
\begin{equation}
f_{nm}\,:\,\mathcal{N}(n)\rightarrow \mathcal{N}(m)\nonumber
\end{equation}
to be any bijection preserving the $\approx$-classes, i.e., $f_{nm}(k)\approx k$ for all~$k\in\mathcal{N}(n)$. Note that this is possible as~$\approx$ is equitable: $n,m$ have the same degree (hence we can choose a bijection), and have the same number of neighbors per class (hence we can choose a bijection that preserves classes).
Now, assume~$i\approx j$ and define the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ by induction as
\begin{equation}
\begin{array}{rcl} f_{ji}^{1}: \mathcal{P}_{1}(j) & \to & \mathcal{P}_{1}(i)\\ \left(j,j_1\right) & \mapsto & \left(i,i_1\right) \end{array}\nonumber
\end{equation}
where~$i_1=f_{ji}(j_1)$, with~$j_1\in \mathcal{N}(j)$ and
\begin{equation}
\begin{array}{rcl} f_{ji}^{n}: \mathcal{P}_{n}(j) & \to & \mathcal{P}_{n}(i)\\ \left(j,j_1,\ldots,j_{n}\right) & \mapsto & \left(i,i_1,\ldots,i_n\right) \end{array}\nonumber
\end{equation}
where,
\begin{equation}
i_1=f_{ji}(j_1),\,i_2=f_{j_{1}i_{1}}(j_2),\,\ldots,i_n=f_{j_{n-1}i_{n-1}}(j_n)\nonumber
\end{equation}
with~$j_1\in \mathcal{N}(j)$, and~$j_k\in\mathcal{N}(j_{k-1})$ for all~$k\leq n$, i.e.,
\begin{equation}
\left(j,j_1,\ldots,j_n\right)\in\mathcal{P}_{n}(j).\nonumber
\end{equation}
Observe first that~$f_{ji}^{n}$ is well constructed as~$j_n\approx i_n$ for each $n$ above -- as one can show by induction -- and thus,~$f_{j_ni_n}$ makes sense. Moreover, the sequence~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence of bijections that preserves the vector degree. By Theorem~\ref{th:characterization}, we conclude that~$i\cong j$.
\end{proof}
The next Theorem provides an alternative practical characterization for the preorder thus constructed: it is an inductive preorder. In fact, in order to prove that the logistic dynamical system preserves the preorder, we need the result in Theorem~\ref{th:induction}.
\begin{theorem}[Induction property of $\succeq$]\label{th:induction}
\begin{equation}\label{eq:inductive}
i\succeq j \Leftrightarrow \exists{f\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i) \mbox{ injective:}} f(k)\succeq k\,\, \forall{k\in\mathcal{N}(j)}
\end{equation}
\end{theorem}
\begin{proof}
\textbf{We start by proving the implication~`$\Rightarrow$'.}
Define~$f(k)\overset{\Delta}=\pi_{2}\left(f^{1}_{ji}(j,k)\right)$. Note that~$f\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i)$ is injective. Also, define the following sequence
\begin{equation}
g^{\ell}_{kf(k)}\left(k,k_1,k_2,\ldots,k_{\ell}\right)=\pi_{2\leq \cdot \leq \ell+2}\left(f^{\ell+1}_{ji}(j,k,k_1,\ldots,k_{\ell})\right)\nonumber
\end{equation}
and note that~$\left(g^{\ell}_{kf(k)}\,:\, \mathcal{P}_{\ell}(k)\rightarrow \mathcal{P}_{\ell}(f(k))\right)_{\ell}$ is an adapted sequence. Note also that~$g^{\ell}_{kf(k)}\,:\,\mathcal{P}_{\ell}(k)\rightarrow \mathcal{P}_{\ell}(f(k))$ is injective for all~$\ell$. The monotonicity on the vector degree is clearly inherited by~$f^{\ell}_{ji}$. Therefore,~$f(k)\succeq k$.
\textbf{Now, we prove the implication~`$\Leftarrow$'.}
By definition, we have~$f(k)\succeq k$ if and only if there exists an adapted sequence~$\left(g^{\ell}_{kf(k)}\right)$ that explains the inequality. Now, define
\begin{equation}
f^{\ell}_{ji}\left(j, j_1,j_2,\ldots,j_{\ell}\right)=\left(i, g^{\ell-1}_{j_1 f(j_1)}\left(j_1,j_2,\ldots,j_{\ell}\right)\right).
\end{equation}
$\left(f^{\ell}_{ji}\right)$ thus defined is an adapted sequence that explains~$i\succeq j$.
\end{proof}
In fact, our preorder is not only inductive in the sense of equation~\eqref{eq:inductive}, but it is the maximum preorder amongst the inductive preorders as explained next. Let~$\mathcal{O}$ be the set of inductive preorders on a graph~$G$ endowed with the following partial order~$\leq$
\begin{equation}
\widetilde{\succeq}\,\, \leq \,\, \succeq \,\,\Longleftrightarrow \,\,\left(i \,\,\widetilde{\succeq}\,\, j \Rightarrow i \succeq j\right).\nonumber
\end{equation}
It is trivial to check that~$\leq$ is a partial order on~$\mathcal{O}$.
\begin{theorem}
$\succeq$ is the greatest element (a.k.a. maximum) on~$\mathcal{O}$ with respect to the partial order~$\leq$,
\begin{equation}
\widetilde{\succeq}\,\, \leq \,\,\succeq\,\,\forall{\widetilde{\succeq}\in \mathcal{O}},
\end{equation}
i.e., any other preorder~$\widetilde{\succeq}\in \mathcal{O}$ is not only comparable to~$\succeq$, but it is upperbounded by it.
\end{theorem}
\begin{proof}
Let~$\widetilde{\succeq}$ be another inductive preorder. For each pair of nodes~$\left(n,m\right)$ with~$n \,\,\widetilde{\succeq}\,\, m$, choose an injective function
\begin{equation}
f_{mn}\,:\,\mathcal{N}(m)\rightarrow \mathcal{N}(n)\nonumber
\end{equation}
that preserves~$\widetilde{\succeq}$, i.e.,~$f(k)\,\widetilde{\succeq} k$ for all~$k\in \mathcal{N}(m)$. Now, let~$i \widetilde{\succeq} j$ and define the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ by induction:
\begin{equation}
\begin{array}{rcl} f_{ji}^{1}: \mathcal{P}_{1}(j) & \to & \mathcal{P}_{1}(i)\\ \left(j,j_1\right) & \mapsto & \left(i,i_1\right) \end{array}\nonumber
\end{equation}
where~$i_1=f_{ji}(j_1)$, with~$j_1\in \mathcal{N}(j)$ and
\begin{equation}
\begin{array}{rcl} f_{ji}^{n}: \mathcal{P}_{n}(j) & \to & \mathcal{P}_{n}(i)\\ \left(j,j_1,\ldots,j_{n}\right) & \mapsto & \left(i,i_1,\ldots,i_n\right) \end{array}\nonumber
\end{equation}
where,
\begin{equation}
i_1=f_{ji}(j_1),\,i_2=f_{j_{1}i_{1}}(j_2),\,\ldots,i_n=f_{j_{n-1}i_{n-1}}(j_n)\nonumber
\end{equation}
with~$j_1\in \mathcal{N}(j)$, and~$j_k\in\mathcal{N}(j_{k-1})$ for all~$k\leq n$, i.e.,
\begin{equation}
\left(j,j_1,\ldots,j_n\right)\in\mathcal{P}_{n}(j).\nonumber
\end{equation}
Note that~$f^{\ell}_{ji}$ is well defined. Also, each~$f^{\ell}_{ji}$ is injective and preserves the degree sequence. Moreover, the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ is adapted.
\end{proof}
\section{Logistic dynamics preserves the preorder}\label{sec:preserva}
In this section, we show that the logistic dynamics~\eqref{eq:logistic} preserves the preorder~$\succeq$. As a result, this leads to the emergence of a non-trivial invariant set for this type of dynamical system over networks, whose characterization is tied to the underlying graph structure.
\begin{theorem}\label{th:preservesineq}
Let~$\mathbf{y}(0)\in\left[0,1\right]^{N}$ be so that
\begin{equation}
y_{i}(0)\geq y_{j}(0)\,\forall{i\succeq j}.\label{eq:assumption}
\end{equation}
Then,
\begin{equation}
y_{i}(t,\mathbf{y}(0))\geq y_{j}\left(t,\mathbf{y}(0)\right)\,\,\forall{t\geq 0}.
\end{equation}
\end{theorem}
Equivalently, the set
\begin{equation}
S\overset{\Delta}=\left\{\mathbf{y}\in\left[0,1\right]^{N} \,:\, y_i\geq y_k,\,\forall{i\succeq k} \right\}\label{eq:invariant}
\end{equation}
is invariant to the logistic dynamical system.
\begin{proof}
To show that~$S$ (defined in equation~\eqref{eq:invariant}) is invariant under the logistic dynamics (either continuous or discrete-time), we just need to check the qualitative behavior of the trajectory~$\left(\mathbf{y}(t)\right)$ at the border of~$S$ to establish that it does not escape from the set. We focus attention on the continuous time case.
Let~$i\succeq j$ with~$y_i(s)=y_j(s)$, for some time~$s$, i.e., the trajectory is at the border at time~$s$. Let the adaptive sequence~$\left(f_{ji}^{\ell}\right)$ explain the inequality~$i\succeq j$. For simplicity, denote~$f_{ji}:=f_{ji}^{1}$ with
\begin{equation}
f_{ji}\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i).\nonumber
\end{equation}
Assume first that
\begin{equation}
y_{k}(s) > y_{f_{ji}(k)}(s)
\end{equation}
for some neighbor~$k\in\mathcal{N}(i)$. Given the induction property of the preorder in Theorem~\ref{th:induction} and the assumption in equation~\eqref{eq:assumption}, we conclude, by inspection on the ODE equation~\eqref{eq:logistic}, that
\begin{equation}
\overset{\cdot}y_i(s)>\overset{\cdot} y_j(s)\nonumber
\end{equation}
and thus, there exists~$\delta>0$ such that~$y_i(t)>y_j(t)$ for all $t\in\left.\left[s,s+\delta\right.\right)$. In other words, the trajectory does not scape the set.
In the case where the two nodes have the same configuration in a first order neighborhood, i.e.,~$f_{ji}$ is bijective and
\begin{equation}
y_k(s)=y_{f_{ji}(k)}(s)
\end{equation}
for all~$k\in \mathcal{N}(i)$, then the previous argument does not apply since~$\overset{\cdot}y_i(s)=\overset{\cdot}y_j(s)$ and we need to consider higher order derivatives. The equations for the $(n+1)$th order derivatives of the logistic system are given by
\begin{eqnarray}
\overset{(n+1)}y_i(t) & = & \gamma \sum_{k\sim i} \left[\overset{(n)}y_k(t)\left(1-y_i(t)\right)- \sum_{m=0}^{n-1}\left(\begin{array}{c} n \\ m \end{array}\right)\overset{(m)}y_k(t) \overset{(n-m)}y_i(t)\right]-\overset{(n)}y_i(t)\label{eq:high}\\
\overset{(n+1)}y_j(t) & = & \gamma \sum_{\ell\sim j} \left[\overset{(n)}y_{\ell}(t)\left(1-y_j(t)\right)- \sum_{m=0}^{n-1}\left(\begin{array}{c} n \\ m \end{array}\right)\overset{(m)}y_{\ell}(t) \overset{(n-m)}y_j(t)\right]-\overset{(n)}y_j(t).\label{eq:high2}
\end{eqnarray}
Given a vector of indexes~$I=\left(I_1,I_2,\ldots,I_N\right)\in\mathbb{N}^N$, define~$y_I:=\left(y_{I_1},\ldots,y_{I_N}\right)$. Assume, for some~$N\geq 1$, that the adapted sequence~$\left(f_{ji}^{\ell}\right)$ is such that~$f_{ji}^{m}\,:\,\mathcal{P}_{m}(j)\rightarrow \mathcal{P}_{m}(i)$ is bijective for all~$m\leq N$ and
\begin{equation}
y_{p^{m}}=y_{f^m_{ji}\left(p^{m}\right)},
\end{equation}
for all~$p^m\in \mathcal{P}^m(j)$, and~$m\leq N$, but at~$N+1$ we have (the strict dominance)
\begin{equation}
f_{ji}^{N+1}\mbox{ injective and not bijective, or } y_{p_{N+1}^{N+1}}<y_{f_{ji}\left(p^{N+1}_{N+1}\right)}.\nonumber
\end{equation}
Then, by inspection on the higher order derivative equations~\eqref{eq:high}-\eqref{eq:high2}, and applying induction on~$N$, we conclude that
\begin{eqnarray}
\overset{(n)}y_i(s) & = & \overset{(n)}y_j(s),\,\forall{n\leq N}\\
\overset{(N+1)}y_i(s) & > & \overset{(N+1)}y_j(s)
\end{eqnarray}
and, thus, there exists~$\delta>0$ such that~$y_i(t)>y_j(t)$ for all $t\in\left.\left[s,s+\delta\right.\right)$, that is, the trajectory does not escape the set~$S$.
\end{proof}
\begin{corollary}[LD preserves CEP]\label{co:preservesCEP}
The homogeneous logistic dynamical system preserves the CEP of the graph, i.e., if~~$\mathbf{y}(0)\in\left[0,1\right]^{N}$ is so that
\begin{equation}
y_{i}(0)= y_{j}(0)\,\forall{i\cong j}.
\end{equation}
Then,
\begin{equation}
y_{i}(t,\mathbf{y}(0))= y_{j}\left(t,\mathbf{y}(0)\right)\,\,\forall{t\geq 0}\,\,\,\,\,\forall{i\cong j}.
\end{equation}
\end{corollary}
Corollary~\ref{co:preservesCEP} follows from Theorems~\ref{th:characterization} and~\ref{th:preservesineq}.
\textbf{Remark.} The proof of Theorem~\ref{th:preservesineq} naturally extends to establish the result for a broader class of dynamical systems. In particular, let~$\overline{F}\,:\,\mathbb{R}\times\mathbb{R}^{N-1}\rightarrow \mathbb{R}$ be an analytic real valued function invariant under permutations on the second vector coordinate, i.e.,
\begin{equation}
\overline{F}\left(x_1;x_2,\ldots,x_N\right)= \overline{F}\left(x_1;p\left(x_2,\ldots,x_N\right)\right)
\end{equation}
for all permutations~$p\,:\,\mathbb{R}^{N-1}\rightarrow \mathbb{R}^{N-1}$ and for all~$\mathbf{x}=\left(x_1,\ldots,x_N\right)\in\mathbb{R}^{N}$. Assume also that~$\overline{F}$ is monotonous in the second vector coordinate, i.e.,
\begin{equation}
\mathbf{x} \geq \mathbf{y}\in\mathbb{R}^{N-1}\Rightarrow \overline{F}\left(x_1;\mathbf{x}\right)\geq \overline{F}\left(x_1;\mathbf{y}\right).
\end{equation}
Given a graph~$G$, let~$F_G=\left(F_1,F_2,\ldots,F_N\right)$ be the vector field induced by~$\overline{F}$ as follows
\begin{equation}\label{eq:construct2}
F_i\left(x\right)= \overline{F}\left(x_i, x_{i_1},\ldots,x_{i_k},0,\ldots,0\right)
\end{equation}
where,~$\mathcal{N}(i)\overset{\Delta}=\left\{j\sim i\,:\, j\in V\right\}=\left\{i_1,\ldots,i_k\right\}$ is the set of neighbors to~$i$ in the graph~$G$ (i.e.,~$F$ conveys the graph structure of~$G$). Note that the ordering of~$x_{i_1},\ldots,x_{i_k}$ in equation~\eqref{eq:construct2} is not relevant as~$\overline{F}$ is invariant under permutation of the associated coordinates. By evoking Fa\`{a} di Bruno's formula (refer, e.g., to~\cite{faadi}) for higher order derivatives of~$F_G$ and resorting to a proof similar to the one in Theorem~\ref{th:preservesineq}, one can establish the result in Theorem~\ref{th:preservesineq} to the class of dynamical systems
\begin{equation}
\frac{d}{dt}\mathbf{y}(t)=F_G(\mathbf{y}(t))
\end{equation}
with vector field~$F_G$. Note that~$\mathbf{F}$ is analytic and, thus, there exists unique solution~$\left(\mathbf{y}\left(t,\mathbf{y}(0)\right)\right)$ for each initial condition~$\mathbf{y}(0)\in\mathbb{R}^N$.
\section{Concluding remarks}\label{sec:conclusion}
In this paper, we constructed a preorder on the set of nodes of a graph. Such preorder entails the classical coarsest equitable partition of the graph. Moreover, it is preserved by certain monotonous dynamical systems over networks, which leads to the existence of an invariant set whose characterization relies on the graph structure of the dynamical system. Note that in light of Corollary~\ref{co:preservesCEP}, such dynamical systems admit a lower dimensional version with underlying dimension given by the number of colors associated with the CEP of the graph, i.e., by having the same initial conditions for nodes with same color, the state of such nodes will be synchronized for all time. Also, Theorem~\ref{th:preservesineq} tells us that one can lower-bound and/or upper-bound solutions to such dynamical systems by properly lower-bounding and/or upper-bounding the corresponding initial conditions. These observations provide with an application to approximate such dynamical systems over complex networks by lower-dimensional ones: i) find the CEP of the graph (which has a cost of about~$n \log n$, with~$n$ being the number of nodes); and ii) initialize evenly the state of nodes with same color so to lower-bound and/or upper-bound the arbitrary initial conditions of interest. In this case, the solutions (with arbitrary initial conditions) can be lower-bounded and/or upper-bounded by its lower dimensional versions for all time~$t$.
\small
\bibliographystyle{IEEEtran}
|
\section{Sensitive Policies}
\label{sec:sensitivity}
Fast, sample-efficient learning is a huge challenge for deep RL. But does it have to be? The conventional wisdom is that ``shallow'' RL is sample-efficient because the hand-designed features provide a good basis for fast learning. But don't deep networks learn good features? Shouldn't deep RL therefore be \emph{faster} once good features have been learned? Indeed, one of the huge benefits of deep architectures should be the ability to learn high-sensitivity neurons that can change the policy behavior in large but structured ways. For example, if a bipedal walker policy can learn a ``direction'' neuron it should be able to toggle walking right versus walking left just by changing the sign on that neuron, which requires changing a relatively small number of connections in relatively local ways.
So the goal is to learn policy parameters $\theta$ that can change the policy in large but structured ways in response to small local changes in $\theta$. That is, we should have a \emph{sensitive} policy. We can define sensitivity in terms of the change in the reward function. For example, if our goal is to maximize $E_{\pi_\theta}[r(\tau)]$, and we take gradient steps of the form
\[
\theta' \leftarrow \theta + \alpha\nabla_\theta E_{\pi_\theta}[r(\tau)],
\]
we can define sensitivity as
\[
\lim_{\alpha \rightarrow 0} \frac{E_{\pi_{\theta'}}[r(\tau)] - E_{\pi_\theta}[r(\tau)]}{\alpha}.
\]
Equivalently, this corresponds to the derivative of the expected reward with respect to the step size. The goal is to have a large (positive) sensitivity with respect to a reward of interest. If we have that, we can learn quickly. If we have high sensitivity to \emph{many} different rewards, we can likely learn quickly even with respect to new rewards.
\section{Optimizing for Sensitivity}
Let us define \emph{locally adapted} parameters in terms of an adaptation operator $f_i(\theta)$:
\[
f_i(\theta) = \theta + \alpha\nabla_\theta E_{\pi_\theta}[r_i(\tau)],
\]
where $r_i(\tau)$ is a reward function from the set of rewards $\{r_i\}$. We can define a sensitivity optimization problem as
\[
\max_\theta E_{\pi_\theta}[r_0(\tau)] + \sum_i E_{\pi_{f_i(\theta)}}[r_i(\tau)],
\]
where $r_0(\tau)$ is a \emph{base} reward function (which may be omitted in general) that defines the main problem we are solving, while $r_i(\tau)$ are auxiliary reward functions that encourage the discovery of a policy that is sensitive to the types of rewards we care about. For example, $r_0$ for a humanoid might be to stand upright, while each $r_i$ might correspond to running in a different direction. If we solve this type of problem, we get a policy that can very quickly learn to run in different directions, and perhaps can generalize to other structured exploration behaviors.
As Chelsea pointed out, this procedure assumes that the reward expectations are smooth. This is not the same as assuming that the rewards \emph{themselves} are smooth, since the expectation of a discontinuous function might still be continuous. This also assumes that the auxiliary rewards are sufficiently diverse.
The optimization requires second derivatives, but these can likely be computed efficiently with most automatic differentiation packages.
\section{The \emph{Other} Fast Weights}
An interesting extension on the above optimization problem is to define a \emph{fast weights} mask $M$, so that only a subset of the policy parameters are adapted, and redefine $f_i$ as
\[
f_i(\theta) = \theta + \alpha\nabla_\theta E_{\pi_\theta}[r_i(\tau)] \circ M,
\]
where $\circ$ represents an elementwise product. Under the ``feature learning'' view of deep networks, this amounts to saying that some weights should be trained for sensitivity, while others are frozen during adaptation because they represent broadly useful features. $M$ might be either fixed or learned.
\section{Open Problems}
While the procedure described above may be effective for optimizing sensitive policies that adapt quickly to new situations, it leaves a number of openings for improvement. First, the adaptation procedure considered involves only one gradient step. This is fine in the ``differential'' view of sensitivity defined in Section~\ref{sec:sensitivity}, but may be limiting in general. It's also not immediately obvious how to extend this framework from policy gradient to other algorithms such as actor-critic, though the extension may be relatively straightforward.
\end{document}
\section{Introduction}
\label{sec:intro}
Learning quickly is a hallmark of human intelligence, whether it involves recognizing objects from a few examples or quickly learning new skills
after just minutes of experience. Our artificial agents should be able to do the same, learning and adapting quickly from only a few examples, and continuing to adapt as more data becomes available. This kind of fast and flexible learning is challenging, since the agent must integrate its prior experience with a small amount of new information, while avoiding overfitting to the new data. Furthermore, the form of prior experience and new data will depend on the task. As such, for the greatest applicability, the mechanism for learning to learn (or meta-learning) should be general to the task and the form of computation required to complete the task.
In this work, we propose a meta-learning algorithm that is general and model-agnostic, in the sense that it can be directly applied to any learning problem and model that is trained with a gradient descent procedure. Our focus is on deep neural network models, but we illustrate how our approach can easily handle different architectures and different problem settings, including classification, regression, and policy gradient reinforcement learning, with minimal modification.
In meta-learning, the goal of the trained model is to quickly learn a new task from a small amount of new data, and the model is
trained by the meta-learner to be able to learn on a large number of different tasks.
The key idea underlying our method is to
train the model's initial parameters such that the model has maximal performance on a new task after the parameters have been updated
through one or more gradient steps computed with a small amount of data from that new task.
Unlike prior meta-learning methods that learn an update function or learning rule~\cite{schmidhuber1987,bengiobengio1,learntolearnbygdbygd,hugo}, our algorithm does not expand the number of learned parameters nor place constraints on the model architecture (e.g. by requiring a recurrent model~\cite{mann} or a Siamese network~\cite{siameseoneshot}), and it can be readily combined with fully connected, convolutional, or recurrent neural networks. It can also be used with a variety of loss functions, including differentiable supervised losses and non-differentiable reinforcement learning objectives.
The process of training a model's parameters such that a few gradient steps, or even a single gradient step, can produce good results on a new task can be viewed from a feature learning standpoint as building an internal representation that is broadly suitable for many tasks. If the internal representation is suitable to many tasks, simply fine-tuning the parameters slightly (e.g. by primarily modifying the top layer weights in a feedforward model) can produce good results. In effect, our procedure optimizes for models that are easy and fast to fine-tune, allowing the adaptation to happen in the right space for fast learning. From a dynamical systems standpoint, our learning process can be viewed as maximizing the sensitivity of the loss functions of new tasks with respect to the parameters: when the sensitivity is high, small local changes to the parameters can lead to large improvements in the task loss.
The primary contribution of this work is a simple model- and task-agnostic algorithm for meta-learning that trains a model's parameters such that a small number of gradient updates will lead to fast learning on a new task.
We demonstrate the algorithm on different model types, including fully connected and convolutional networks, and in several distinct domains, including few-shot regression, image classification, and reinforcement learning.
Our evaluation shows that our meta-learning algorithm compares favorably to state-of-the-art one-shot learning methods designed specifically for supervised classification, while using fewer parameters, but that it can also be readily applied to regression and can accelerate reinforcement learning in the presence of task variability, substantially outperforming direct pretraining as initialization.
\section{Model-Agnostic Meta-Learning}
\label{sec:overview}
We aim to train models that can achieve rapid adaptation, a problem setting that is often formalized as few-shot learning. In this section, we will define the problem setup and present the general form of our algorithm.
\subsection{Meta-Learning Problem Set-Up}
\label{sec:problem}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathbf{x}}{\mathbf{x}}
\newcommand{f}{f}
\newcommand{\loss_{\task_i}}{\mathcal{L}_{\mathcal{T}_i}}
The goal of few-shot meta-learning
is to train a model that can quickly adapt to a new task using only a few datapoints and training iterations.
To accomplish this, the model or learner is trained during a meta-learning phase on a set of tasks, such that the trained model can quickly adapt to new tasks using only a small number of examples or trials.
In effect, the meta-learning problem treats entire tasks as training examples.
In this section, we formalize this meta-learning problem setting in a general manner, including brief examples of different learning domains.
We will discuss two different learning domains in detail in Section~\ref{sec:instantiations}.
We consider a model, denoted $f$, that maps observations $\mathbf{x}$ to outputs $\mathbf{a}$.
During meta-learning, the model
is trained to be able to adapt to a large or infinite number of tasks.
Since we would like to apply our framework to a variety of learning problems, from classification to reinforcement learning, we introduce a generic notion of a learning task below.
Formally, each task $\mathcal{T} = \{ \mathcal{L}(\mathbf{x}_1,\mathbf{a}_1,\dots,\mathbf{x}_H,\mathbf{a}_H), q(\mathbf{x}_1), q(\mathbf{x}_{t+1} | \mathbf{x}_t, \mathbf{a}_t), H \}$
consists of a loss function $\mathcal{L}$, a distribution over initial observations $q(\mathbf{x}_1)$, a transition distribution $ q(\mathbf{x}_{t+1} | \mathbf{x}_t, \mathbf{a}_t)$, and an episode length $H$. In i.i.d. supervised learning problems, the length $H\!=\!1$.
The model may generate samples of length $H$ by choosing an output $\mathbf{a}_t$ at each time $t$.
The loss $\mathcal{L}(\mathbf{x}_1,\mathbf{a}_1,\dots,\mathbf{x}_H,\mathbf{a}_H)\rightarrow \mathbb{R}$, provides task-specific feedback, which might be in the form of a misclassification loss or a cost function in a Markov decision process.
In our meta-learning scenario, we consider a distribution over tasks $p(\mathcal{T})$ that we want our model to be able to adapt to.
In the $K$-shot learning setting, the model is trained to learn a new task $\mathcal{T}_i$ drawn from $p(\mathcal{T})$ from only $K$ samples drawn from $q_i$ and feedback $\loss_{\task_i}$ generated by $\mathcal{T}_i$.
During meta-training, a task $\mathcal{T}_i$ is sampled from $p(\mathcal{T})$, the model is trained with $K$ samples and feedback from the corresponding loss $\loss_{\task_i}$ from $\mathcal{T}_i$, and then tested on new samples from $\mathcal{T}_i$.
The model $f$ is then improved by considering how the \emph{test} error on new data from $q_i$ changes with respect to the parameters. In effect, the test error on sampled tasks $\mathcal{T}_i$ serves as the training error of the meta-learning process.
At the end of meta-training, new tasks are sampled from $p(\mathcal{T})$,
and meta-performance is measured by the model's performance after learning from $K$ samples.
Generally, tasks used for meta-testing are held out during meta-training.
\subsection{A Model-Agnostic Meta-Learning Algorithm}
\label{sec:maml}
\begin{figure}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.9) \linethickness{0.5pt}
\put(0.25,-0.1){\includegraphics[width=0.75\columnwidth]{mamldiagram.pdf}}
\put(0.63,0.63){\huge$\theta$}
\put(0.58,0.3){\large$\nabla \mathcal{L}_1$}
\put(1.0,0.4){\large$\nabla \mathcal{L}_2$}
\put(0.96,0.55){\large$\nabla \mathcal{L}_3$}
\put(0.80,0.06){\large$\theta^*_1$}
\put(1.24,0.04){\large$\theta^*_2$}
\put(1.32,0.32){\large$\theta^*_3$}
\end{picture}
\vspace{-0.8cm}
\caption{
Diagram of our model-agnostic meta-learning algorithm (MAML), which optimizes for a representation $\theta$ that can quickly adapt to new tasks.
\label{fig:teaser}
\vspace{-0.5cm}
}
\end{figure}
In contrast to prior work, which has sought to train recurrent neural networks that ingest entire datasets~\cite{mann,rl2} or feature embeddings that can be combined with nonparametric methods at test time~\cite{matchingnets,siameseoneshot}, we propose a method that can learn the parameters of any standard model via meta-learning in such a way as to prepare that model for fast adaptation. The intuition behind this approach is that some internal representations are more transferrable than others. For example, a neural network might learn internal features that are broadly applicable to all tasks in $p(\mathcal{T})$, rather than a single individual task. How can we encourage the emergence of such general-purpose representations? We take an explicit approach to this problem: since the model will be fine-tuned using a gradient-based learning rule on a new task, we will aim to learn a model in such a way that this gradient-based learning rule can make rapid progress on new tasks drawn from $p(\mathcal{T})$, without overfitting. In effect, we will aim to find model parameters that are \emph{sensitive} to changes in the task, such that small changes in the parameters will produce large improvements on the loss function of any task drawn from $p(\mathcal{T})$, when altered in the direction of the gradient of that loss (see Figure~\ref{fig:teaser}). We make no assumption on the form of the model, other than to assume that it is parametrized by some parameter vector $\theta$, and that the loss function is smooth enough in $\theta$ that we can use gradient-based learning techniques.
Formally, we consider a model
represented by a parametrized function $f_\theta$
with parameters $\theta$.
When adapting to a new task $\mathcal{T}_i$, the model's parameters $\theta$ become $\theta_i'$.
In our method, the updated parameter vector $\theta_i'$ is computed using one or more gradient descent updates on task $\mathcal{T}_i$.
For example, when using one gradient update,
\vspace{-0.15cm}
$$
\vspace{-0.15cm}
\theta_i'=\theta-\alpha \nabla_\theta \loss_{\task_i}( f_\theta ).
$$
The step size $\alpha$ may be fixed as a hyperparameter or meta-learned.
For simplicity of notation, we will consider one gradient update for the rest of this section, but using multiple gradient updates is a straightforward extension.
The model parameters are trained by optimizing for the performance of $f_{\theta_i'}$ with respect to $\theta$ across tasks sampled from $p(\mathcal{T})$.
More concretely, the meta-objective is as follows:
\vspace{-0.15cm}
\begin{align*}
\vspace{-0.2cm}
\min_\theta \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \loss_{\task_i} ( f_{\theta_i'})
= \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \loss_{\task_i} ( f_{\theta - \alpha \nabla_\theta \loss_{\task_i}(f_\theta)})
\end{align*}
Note that the meta-optimization is performed over the model parameters $\theta$, whereas the objective is computed using the updated model parameters $\theta'$.
In effect, our proposed method aims to optimize the model parameters such that one or a small number of gradient steps on a new task will produce maximally effective behavior on that task.
The meta-optimization across tasks is performed via stochastic gradient descent (SGD), such that the model parameters $\theta$ are updated as follows:
\vspace{-0.15cm}
\begin{equation}
\label{eq:metaupdate}
\vspace{-0.2cm}
\theta \leftarrow \theta - \beta \nabla_\theta \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \loss_{\task_i} ( f_{\theta_i'})
\end{equation}
where $\beta$ is the meta step size. The full algorithm, in the general case, is outlined in Algorithm~\ref{alg:maml}.
\begin{algorithm}[t]
\caption{Model-Agnostic Meta-Learning}
\label{alg:maml}
\begin{algorithmic}[1]
\REQUIRE $p(\mathcal{T})$: distribution over tasks
\REQUIRE $\alpha$, $\beta$: step size hyperparameters
\STATE randomly initialize $\theta$
\WHILE{not done}
\STATE Sample batch of tasks $\mathcal{T}_i \sim p(\mathcal{T})$
\FORALL{$\mathcal{T}_i$}
\STATE Evaluate $\nabla_\theta \loss_{\task_i}(f_\theta)$ with respect to $K$ examples
\STATE Compute adapted parameters with gradient descent: $\theta_i'=\theta-\alpha \nabla_\theta \loss_{\task_i}( f_\theta )$
\ENDFOR
\STATE Update $\theta \leftarrow \theta - \beta \nabla_\theta \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \loss_{\task_i} ( f_{\theta_i'})$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
The MAML meta-gradient update involves a gradient through a gradient. Computationally, this requires an additional backward pass through $f$ to compute Hessian-vector products, which is supported by standard deep learning libraries such as TensorFlow~\cite{tensorflow}. In our experiments, we also include a comparison to dropping this backward pass and using a first-order approximation, which we discuss in Section~\ref{sec:image_results}.
\vspace{-0.15cm}
\section{Species of MAML}
\label{sec:instantiations}
\vspace{-0.15cm}
In this section, we discuss specific instantiations of our meta-learning algorithm for supervised learning and reinforcement learning. The domains differ in the form of loss function and in how data is generated by the task and presented to the model, but the same basic adaptation mechanism can be applied in both cases.
\vspace{-0.1cm}
\subsection{Supervised Regression and Classification}
\newcommand{\mathbf{y}}{\mathbf{y}}
Few-shot learning is well-studied in the domain of supervised tasks, where the goal is to learn a new function from only a few input/output pairs for that task, using prior data from similar tasks for meta-learning. For example, the goal might be to classify images of a Segway after seeing only one or a few examples of a Segway, with a model that has previously seen many other types of objects. Likewise, in few-shot regression, the goal is to predict the outputs of a continuous-valued function from only a few datapoints sampled from that function, after training on many functions with similar statistical properties.
To formalize the supervised regression and classification problems in the context of the meta-learning definitions in Section~\ref{sec:problem}, we can define the horizon
$H=1$ and drop the timestep subscript on $\mathbf{x}_t$, since the model accepts a single input and produces a single output, rather than a sequence of inputs and outputs. The task $\mathcal{T}_i$ generates $K$ i.i.d. observations $\mathbf{x}$ from $q_i$, and the task loss is represented by the error between the model's output for $\mathbf{x}$ and the corresponding target values $\mathbf{y}$ for that observation and task.
Two common loss functions used for supervised classification and regression are cross-entropy and mean-squared error (MSE), which we will describe below; though, other supervised loss functions may be used as well. For regression tasks using mean-squared error, the loss takes the form:
\vspace{-0.15cm}
\begin{align}
\vspace{-0.2cm}
\label{eq:sup1}
\loss_{\task_i}(f_\phi ) = \!\!\!\!\!\! \sum_{\mathbf{x}^{(j)}, \mathbf{y}^{(j)} \sim \mathcal{T}_i} \lVert f_\phi(\mathbf{x}^{(j)}) - \mathbf{y}^{(j)} \rVert_2^2,
\end{align}
where $\mathbf{x}^{(j)}, \mathbf{y}^{(j)}$ are an input/output pair sampled from task $\mathcal{T}_i$. In $K$-shot regression tasks, $K$ input/output pairs are provided for learning for each task.
Similarly, for discrete classification tasks with a cross-entropy loss, the loss takes the form:
\vspace{-0.15cm}
\begin{equation}
\begin{aligned}
\label{eq:sup2}
\loss_{\task_i}( f_\phi ) = \!\!\!\!\!\! \sum_{\mathbf{x}^{(j)}, \mathbf{y}^{(j)} \sim \mathcal{T}_i} \!\!\!\!\!\! &\mathbf{y}^{(j)} \log f_\phi(\mathbf{x}^{(j)}) \\
&+(1-\mathbf{y}^{(j)}) \log (1-f_\phi(\mathbf{x}^{(j)}))
\end{aligned}
\end{equation}
According to the conventional terminology, $K$-shot classification tasks use $K$ input/output pairs from each class, for a total of $NK$ data points for $N$-way classification. Given a distribution over tasks $p(\mathcal{T}_i)$, these loss functions can be directly inserted into the equations in Section~\ref{sec:maml} to perform meta-learning, as detailed in Algorithm~\ref{alg:mamlsup}.
\begin{algorithm}[t]
\caption{MAML for Few-Shot Supervised Learning}
\label{alg:mamlsup}
\begin{algorithmic}[1]
{\footnotesize
\REQUIRE $p(\mathcal{T})$: distribution over tasks
\REQUIRE $\alpha$, $\beta$: step size hyperparameters
\STATE randomly initialize $\theta$
\WHILE{not done}
\STATE Sample batch of tasks $\mathcal{T}_i \sim p(\mathcal{T})$
\FORALL{$\mathcal{T}_i$}
\STATE Sample $K$ datapoints $\mathcal{D}=\{\mathbf{x}^{(j)}, \mathbf{y}^{(j)}\}$ from $\mathcal{T}_i$
\STATE Evaluate $\nabla_\theta \loss_{\task_i}(f_\theta)$ using $\mathcal{D}$ and $\loss_{\task_i}$ in Equation~(\ref{eq:sup1}) or~(\ref{eq:sup2})
\STATE Compute adapted parameters with gradient descent: $\theta_i'=\theta-\alpha \nabla_\theta \loss_{\task_i}( f_\theta )$
\STATE Sample datapoints $\mathcal{D}_i'=\{\mathbf{x}^{(j)}, \mathbf{y}^{(j)}\}$ from $\mathcal{T}_i$ for the meta-update
\ENDFOR
\STATE Update $\theta \leftarrow \theta - \beta \nabla_\theta \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \loss_{\task_i} ( f_{\theta_i'})$ using each $\mathcal{D}_i'$ and $\loss_{\task_i}$ in Equation~\ref{eq:sup1} or~\ref{eq:sup2}
\ENDWHILE
}
\end{algorithmic}
\end{algorithm}
\subsection{Reinforcement Learning}
In reinforcement learning (RL), the goal of few-shot meta-learning is to enable an agent to quickly acquire a policy for a new test task using only a small amount of experience in the test setting. A new task might involve achieving a new goal or succeeding on a previously trained goal in a new environment. For example, an agent might learn to quickly figure out how to navigate mazes so that, when faced with a new maze, it can determine how to reliably reach the exit with only a few samples.
In this section, we will discuss how MAML can be applied to meta-learning for RL.
Each RL task $\mathcal{T}_i$ contains an initial state distribution $q_i(\mathbf{x}_1)$ and a transition distribution $q_i(\mathbf{x}_{t+1}|\mathbf{x}_t,\mathbf{a}_t)$, and the loss $\loss_{\task_i}$ corresponds to the (negative) reward function $R$. The entire task is therefore a Markov decision process (MDP) with horizon $H$, where the learner is allowed to query a limited number of sample trajectories for few-shot learning. Any aspect of the MDP may change across tasks in $p(\mathcal{T})$. The model being learned, $f_\theta$, is a policy that maps from states $\mathbf{x}_t$ to a distribution over actions $\mathbf{a}_t$ at each timestep $t \in \{1,...,H\}$. The loss for task $\mathcal{T}_i$ and model $f_\phi$ takes the form
\vspace{-0.1cm}
\begin{align}
\label{eq:rl}
\loss_{\task_i}( f_\phi) = - \mathbb{E}_{\mathbf{x}_t, \mathbf{a}_t \sim f_\phi, q_{\mathcal{T}_i}}\left[ \sum_{t=1}^H R_i(\mathbf{x}_t, \mathbf{a}_t) \right].
\end{align}
In $K$-shot reinforcement learning, $K$ rollouts from $f_\theta$ and
task $\mathcal{T}_i$, $(\mathbf{x}_1,\mathbf{a}_1,...\mathbf{x}_H)$, and the corresponding rewards $R(\mathbf{x}_t, \mathbf{a}_t)$, may be used for adaptation on a new task $\mathcal{T}_i$. Since the expected reward is generally not differentiable due to unknown dynamics, we use policy gradient methods to estimate the gradient both for the model gradient update(s) and the meta-optimization. Since policy gradients are an on-policy algorithm, each additional gradient step during the adaptation of $f_\theta$ requires new samples from the current policy $f_{\theta_{i'}}$. We detail the algorithm in Algorithm~\ref{alg:mamlrl}. This algorithm has the same structure as Algorithm~\ref{alg:mamlsup}, with the principal difference being that steps 5 and 8 require sampling trajectories from the environment corresponding to task $\mathcal{T}_i$. Practical implementations of this method may also use a variety of improvements recently proposed for policy gradient algorithms, including state or action-dependent baselines and trust regions~\cite{trpo}.
\begin{algorithm}[t]
\caption{MAML for Reinforcement Learning}
\label{alg:mamlrl}
\begin{algorithmic}[1]
{\footnotesize
\REQUIRE $p(\mathcal{T})$: distribution over tasks
\REQUIRE $\alpha$, $\beta$: step size hyperparameters
\STATE randomly initialize $\theta$
\WHILE{not done}
\STATE Sample batch of tasks $\mathcal{T}_i \sim p(\mathcal{T})$
\FORALL{$\mathcal{T}_i$}
\STATE Sample $K$ trajectories $\mathcal{D}=\{(\mathbf{x}_1,\mathbf{a}_1,...\mathbf{x}_H)\}$ using $f_\theta$ in $\mathcal{T}_i$
\STATE Evaluate $\nabla_\theta \loss_{\task_i}(f_\theta)$ using $\mathcal{D}$ and $\mathcal{L}_{\mathcal{T}_i}$ in Equation~\ref{eq:rl}
\STATE Compute adapted parameters with gradient descent: $\theta_i'=\theta-\alpha \nabla_\theta \mathcal{L}_{\mathcal{T}_i}( f_\theta )$
\STATE Sample trajectories $\mathcal{D}_i'=\{(\mathbf{x}_1,\mathbf{a}_1,...\mathbf{x}_H)\}$ using $f_{\theta_i'}$ in $\mathcal{T}_i$
\ENDFOR
\STATE Update $\theta \leftarrow \theta - \beta \nabla_\theta \sum_{\mathcal{T}_i \sim p(\mathcal{T})} \mathcal{L}_{\mathcal{T}_i} ( f_{\theta_i'})$ using each $\mathcal{D}_i'$ and $\mathcal{L}_{\mathcal{T}_i}$ in Equation~\ref{eq:rl}
\ENDWHILE
}
\end{algorithmic}
\end{algorithm}
\section{Related Work}
The method that we propose in this paper addresses the general problem of meta-learning~\cite{thrun,schmidhuber1987,naik}, which includes few-shot learning.
A popular approach for meta-learning is to train a meta-learner that learns how to update the parameters of the learner's model~\cite{bengiobengio1,schmidfastweights,bengiobengio2}. This approach has been applied to learning to optimize deep networks~\cite{hochreiter,learntolearnbygdbygd,learntooptimize}, as well as for learning dynamically changing recurrent networks~\cite{hypernets}.
One recent approach learns both the weight initialization and the optimizer, for few-shot image recognition~\cite{hugo}. Unlike these methods, the MAML learner's weights are updated using the gradient, rather than a learned update; our method does not introduce additional parameters for meta-learning nor require a particular learner architecture.
Few-shot learning methods have also been developed for specific tasks such as generative modeling~\cite{neuralstatistician,oneshotgenicml} and image recognition~\cite{matchingnets}. One successful approach for few-shot classification is to learn to compare new examples in a learned metric space using e.g. Siamese networks~\cite{siameseoneshot} or recurrence with attention mechanisms~\cite{matchingnets,comparators,prototypical}.
These approaches have generated some of the most successful results, but are difficult to directly extend to other problems, such as reinforcement learning. Our method, in contrast, is agnostic to the form of the model and to the particular learning task.
Another approach to meta-learning is to train memory-augmented models on many tasks, where
the recurrent learner is trained to adapt to new tasks as it is rolled out. Such networks have been applied to few-shot image recognition~\cite{mann,metanets} and learning ``fast'' reinforcement learning agents~\cite{rl2,learningrl}.
Our experiments show that our method outperforms the recurrent approach on few-shot classification. Furthermore, unlike these methods, our approach simply provides a good weight initialization and uses the same gradient descent update for both the learner and meta-update.
As a result, it is straightforward to finetune the learner for additional gradient steps.
Our approach is also related to methods for initialization of deep networks.
In computer vision, models pretrained on large-scale image classification have been shown to learn effective features for a range of problems~\cite{decaf}. In contrast, our method explicitly optimizes the model for fast adaptability, allowing it to adapt to new tasks with only a few examples.
Our method can also be viewed as explicitly maximizing sensitivity of new task losses to the model parameters.
A number of prior works have explored sensitivity in deep networks, often in the context of initialization~\cite{orthogonal,forgetting}. Most of these works have considered good random initializations, though a number of papers have addressed data-dependent initializers~\cite{datadependentinit,weightnorm}, including learned initializations~\cite{husken,maclaurin}. In contrast, our method explicitly trains the parameters for sensitivity on a given task distribution, allowing for extremely efficient adaptation for problems such as $K$-shot learning and rapid reinforcement learning in only one or a few gradient steps.
\section{Experimental Evaluation}
\iffalse
\begin{figure*}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.8) \linethickness{0.5pt}
\put(0.0,-0.1){\includegraphics[width=0.67\columnwidth]{mse_5shot.png}}
\put(1.33,-0.1){\includegraphics[width=0.67\columnwidth]{mse_10shot.png}}
\put(2.67,-0.1){\includegraphics[width=0.67\columnwidth]{mse_20shot.png}}
\end{picture}
\caption{Quantitative sinusoid regression results showing test-time learning curves.
Note that MAML continues to improve with additional gradient steps without overfitting to the extremely small dataset during meta-testing, and achieves a loss that is substantially lower than the baseline fine-tuning approach.
\label{fig:mamlquant}
\vspace{-0.5cm}
}
\end{figure*}
\fi
\begin{figure*}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.8) \linethickness{0.5pt}
\put(0.0,-0.0){\includegraphics[width=0.5\columnwidth]{mamlsineplot_5_7.png}}
\put(0.95,-0.0){\includegraphics[width=0.5\columnwidth]{mamlsineplot_10_0.png}}
\put(2,0.05){\line(0,1){0.7}}
\put(2.05,-0.0){\includegraphics[width=0.5\columnwidth]{pretrained001sineplot_5_7.png}}
\put(3.0,-0.0){\includegraphics[width=0.49\columnwidth]{pretrained002sineplot_10_0.png}}
\put(0.0,-0.1){\includegraphics[width=2\columnwidth]{legend_sinusoid.pdf}}
\end{picture}
\caption{Few-shot adaptation for the simple regression task. Left: Note that MAML is able to estimate parts of the curve where there are no datapoints, indicating that the model has learned about the periodic structure of sine waves. Right: Fine-tuning of a model pretrained on the same distribution of tasks without MAML, with a tuned step size. Due to the often contradictory outputs on the pre-training tasks, this model is unable to recover a suitable representation and fails to extrapolate from the small number of test-time samples.
\label{fig:mamlqual}
\vspace{-0.4cm}
}
\end{figure*}
\begin{figure}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.75) \linethickness{0.5pt}
\put(0.3,-0.15){\includegraphics[width=0.67\columnwidth]{mse_10shot.png}}
\end{picture}
\caption{Quantitative sinusoid regression results showing the learning curve at meta test-time.
Note that MAML continues to improve with additional gradient steps without overfitting to the extremely small dataset during meta-testing, achieving a loss that is substantially lower than the baseline fine-tuning approach.
\label{fig:mamlquant}
\vspace{-0.5cm}
}
\end{figure}
\iffalse
\begin{table*}[t]
\vspace{-0.5cm}
\caption{Few-shot classification test accuracy on the MiniImagenet dataset by~\citet{hugo}. MAML achieves results that outperform state-of-the-art models. The $\pm$ shows $95\%$ confidence intervals over tasks. The evaluation of the baseline methods and matching networks is from~\citet{hugo}.}
\label{tbl:miniimagenet}
\begin{center}
{\footnotesize
\begin{tabular}{|l|c|c|c|c|}
\hline
\multicolumn{1}{|c}{} & \multicolumn{2}{|c|}{5-way Accuracy} \\
& 1-shot & 5-shot \\
\hline
fine-tuning baseline & $28.86 \pm 0.54\%$ & $49.79 \pm 0.79\%$ \\
\hline
nearest neighbor baseline & $41.08 \pm 0.70\%$ & $51.04 \pm 0.65\%$ \\
\hline
matching nets~\cite{matchingnets} & $43.56 \pm 0.84\%$ & $55.31 \pm 0.73\%$ \\
\hline
meta-learner LSTM~\cite{hugo} & $43.44 \pm 0.77\%$ & $60.60 \pm 0.71\%$ \\
\hline
\textbf{MAML (ours)} & $\mathbf{47.70 \pm 1.79\%}$ & $\mathbf{62.63 \pm 0.94\%}$ \\
\hline
\end{tabular}
}
\end{center}
\vspace{-0.2cm}
\end{table*}
\fi
The goal of our experimental evaluation is to answer the following questions: (1) Can MAML enable fast learning of new tasks? (2) Can MAML be used for meta-learning in multiple different domains, including supervised regression, classification, and reinforcement learning? (3) Can a model learned with MAML continue to improve with additional gradient updates and/or examples?
All of the meta-learning problems that we consider require some amount of adaptation to new tasks at test-time. When possible, we compare our results to an oracle that receives the identity of the task (which is a problem-dependent representation) as an additional input, as an upper bound on the performance of the model. All of the experiments were performed using TensorFlow~\cite{tensorflow}, which allows for automatic differentiation through the gradient update(s) during meta-learning. The code is available online\footnote{Code for the regression and supervised experiments is at \url{github.com/cbfinn/maml} and code for the RL experiments is at \url{github.com/cbfinn/maml_rl}}.
\subsection{Regression}
We start with a simple regression problem that illustrates the basic principles of MAML. Each task involves regressing from the input to the output of a sine wave, where the amplitude and phase of the sinusoid are varied between tasks. Thus, $p(\mathcal{T})$ is continuous, where the amplitude varies within $[0.1,5.0]$ and the phase varies within $[0,\pi]$, and the input and output both have a dimensionality of $1$. During training and testing, datapoints $\mathbf{x}$ are sampled uniformly from $[-5.0,5.0]$. The loss is the mean-squared error between the prediction $f(\mathbf{x})$ and true value. The regressor is a neural network model with $2$ hidden layers of size $40$ with ReLU nonlinearities. When training with MAML, we use one gradient update with $K=10$ examples with a fixed step size $\alpha=0.01$, and use Adam as the meta-optimizer~\cite{adam}. The baselines are likewise trained with Adam. To evaluate performance, we fine-tune a single meta-learned model on varying numbers of $K$ examples, and compare performance to two baselines: (a) pretraining on all of the tasks, which entails training a network to regress to random sinusoid functions and then, at test-time, fine-tuning with gradient descent on the $K$ provided points, using an automatically tuned step size, and (b) an oracle which receives the true amplitude and phase as input. In Appendix~\ref{app:comparisons}, we show comparisons to additional multi-task and adaptation methods.
We evaluate performance by fine-tuning the model learned by MAML and the pretrained model on $K=\{5,10,20\}$ datapoints. During fine-tuning, each gradient step is computed using the same $K$ datapoints. The qualitative results, shown in Figure~\ref{fig:mamlqual} and further expanded on in Appendix~\ref{app:sinequal} show that the learned model is able to quickly adapt with only $5$ datapoints, shown as purple triangles, whereas the model that is pretrained using standard supervised learning on all tasks is unable to adequately adapt with so few datapoints without catastrophic overfitting. Crucially, when the $K$ datapoints are all in one half of the input range, the model trained with MAML can still infer the amplitude and phase in the other half of the range, demonstrating that the MAML trained model $f$ has learned to model the periodic nature of the sine wave. Furthermore, we observe both in the qualitative and quantitative results (Figure~\ref{fig:mamlquant} and Appendix~\ref{app:sinequal}) that the model learned with MAML continues to improve with additional gradient steps, despite being trained for maximal performance after one gradient step. This improvement suggests that MAML optimizes the parameters such that they lie in a region that is amenable to fast adaptation and is sensitive to loss functions from $p(\mathcal{T})$, as discussed in Section~\ref{sec:maml}, rather than overfitting to parameters $\theta$ that only improve after one step.
\subsection{Classification}
\label{sec:image_results}
\begin{table*}[t]
\vspace{-0.25cm}
\caption{Few-shot classification on held-out Omniglot characters (top) and the MiniImagenet test set (bottom). MAML achieves results that are comparable to or outperform state-of-the-art convolutional and recurrent models. Siamese nets, matching nets, and the memory module approaches are all specific to classification, and are not directly applicable to regression or RL scenarios. The $\pm$ shows $95\%$ confidence intervals over tasks. Note that the Omniglot results may not be strictly comparable since the train/test splits used in the prior work were not available. The MiniImagenet evaluation of baseline methods and matching networks is from~\citet{hugo}.}
\label{tbl:omniglot}
\begin{center}
{\footnotesize
\begin{tabular}{|l|c|c|c|c|}
\hline
\multicolumn{1}{|c}{} & \multicolumn{2}{|c|}{5-way Accuracy} & \multicolumn{2}{|c|}{20-way Accuracy}\\
{ Omniglot~\citep{omniglot}} & 1-shot & 5-shot & 1-shot & 5-shot\\
\hline
MANN, no conv ~\cite{mann} & $82.8\%$ & $94.9\%$ &-- & -- \\
\hline
\textbf{MAML, no conv (ours)} & $\mathbf{89.7 \pm 1.1}\%$ & $\mathbf{97.5 \pm 0.6}\%$ & -- & -- \\
\hline
\hline
Siamese nets~\cite{siameseoneshot} & $97.3\%$ & $98.4\%$ & $88.2\%$ & $97.0\%$ \\
\hline
matching nets~\cite{matchingnets} & $98.1\%$ & $98.9\%$ & $93.8\%$ & $98.5\%$ \\
\hline
neural statistician~\cite{neuralstatistician} & $98.1\%$ & $99.5\%$ & $93.2\%$ & $98.1\%$ \\
\hline
memory mod.~\cite{rareevents} & $98.4\%$ & $99.6\%$ & $95.0\%$ & $98.6\%$ \\
\hline
\textbf{MAML (ours)} & $\mathbf{98.7 \pm 0.4\%}$ & $\mathbf{99.9 \pm 0.1\%}$ & $\mathbf{95.8 \pm 0.3\%}$ & $\mathbf{98.9 \pm 0.2}\%$ \\
\hline
\multicolumn{5}{c}{}
\end{tabular}
\begin{tabular}{|l|c|c|c|c|}
\hline
\multicolumn{1}{|c}{} & \multicolumn{2}{|c|}{5-way Accuracy} \\
MiniImagenet~\citep{hugo} & 1-shot & 5-shot \\
\hline
fine-tuning baseline & $28.86 \pm 0.54\%$ & $49.79 \pm 0.79\%$ \\
\hline
nearest neighbor baseline & $41.08 \pm 0.70\%$ & $51.04 \pm 0.65\%$ \\
\hline
matching nets~\cite{matchingnets} & $43.56 \pm 0.84\%$ & $55.31 \pm 0.73\%$ \\
\hline
meta-learner LSTM~\cite{hugo} & $43.44 \pm 0.77\%$ & $60.60 \pm 0.71\%$ \\
\hline
\textbf{MAML, first order approx. (ours)} & $\mathbf{48.07 \pm 1.75\%}$ & $\mathbf{63.15 \pm 0.91\%}$ \\
\hline
\textbf{MAML (ours)} & $\mathbf{48.70 \pm 1.84\%}$ & $\mathbf{63.11 \pm 0.92\%}$ \\
\hline
\end{tabular}
}
\end{center}
\vspace{-0.35cm}
\end{table*}
To evaluate MAML in comparison to prior meta-learning and few-shot learning algorithms, we applied our method to few-shot image recognition on the Omniglot~\cite{omniglot} and MiniImagenet datasets. The Omniglot dataset consists of 20 instances of 1623 characters from 50 different alphabets. Each instance was drawn by a different person. The MiniImagenet dataset was proposed by~\citet{hugo}, and involves 64 training classes, 12 validation classes, and 24 test classes. The Omniglot and MiniImagenet image recognition tasks are the most common recently used few-shot learning benchmarks~\cite{matchingnets,mann,hugo}. We follow the experimental protocol proposed by~\citet{matchingnets}, which involves fast learning of $N$-way classification with 1 or 5 shots. The problem of $N$-way classification is set up as follows: select $N$ unseen classes, provide the model with $K$ different instances of each of the $N$ classes, and evaluate the model's ability to classify new instances within the $N$ classes. For Omniglot, we randomly select $1200$ characters for training, irrespective of alphabet, and use the remaining for testing. The Omniglot dataset is augmented with rotations by multiples of $90$ degrees, as proposed by~\citet{mann}.
Our model follows the same architecture as the embedding function used by~\citet{matchingnets}, which has 4 modules with a $3\times3$ convolutions and $64$ filters, followed by batch normalization~\cite{batchnorm}, a ReLU nonlinearity, and $2\times2$ max-pooling. The Omniglot images are downsampled to $28\times28$, so the dimensionality of the last hidden layer is $64$. As in the baseline classifier used by~\citet{matchingnets}, the last layer is fed into a softmax. For Omniglot, we used strided convolutions instead of max-pooling. For MiniImagenet, we used $32$ filters per layer to reduce overfitting, as done by~\cite{hugo}.
In order to also provide a fair comparison against memory-augmented neural networks~\cite{mann} and to test the flexibility of MAML, we also provide results for a non-convolutional network. For this, we use a network with $4$ hidden layers with sizes $256$, $128$, $64$, $64$, each including batch normalization and ReLU nonlinearities, followed by a linear layer and softmax. For all models, the loss function is the cross-entropy error between the predicted and true class. Additional hyperparameter details are included in Appendix~\ref{app:hyperclass}.
We present the results in Table~\ref{tbl:omniglot}. The convolutional model learned by MAML compares well to the state-of-the-art results on this task, narrowly outperforming the prior methods. Some of these existing methods, such as matching networks, Siamese networks, and memory models are designed with few-shot classification in mind, and are not readily applicable to domains such as reinforcement learning. Additionally, the model learned with MAML uses fewer overall parameters compared to matching networks and the meta-learner LSTM, since the algorithm does not introduce any additional parameters beyond the weights of the classifier itself. Compared to these prior methods, memory-augmented neural networks~\cite{mann} specifically, and recurrent meta-learning models in general, represent a more broadly applicable class of methods that, like MAML, can be used for other tasks such as reinforcement learning~\cite{rl2,learningrl}. However, as shown in the comparison, MAML significantly outperforms memory-augmented networks and the meta-learner LSTM on 5-way Omniglot and MiniImagenet classification, both in the $1$-shot and $5$-shot case.
A significant computational expense in MAML comes from the use of second derivatives when backpropagating the meta-gradient through the gradient operator in the meta-objective (see Equation~(\ref{eq:metaupdate})). On MiniImagenet, we show a comparison to a first-order approximation of MAML, where these second derivatives are omitted. Note that the resulting method still computes the meta-gradient at the post-update parameter values $\theta_i'$, which provides for effective meta-learning. Surprisingly however, the performance of this method is nearly the same as that obtained with full second derivatives, suggesting that most of the improvement in MAML comes from the gradients of the objective at the post-update parameter values, rather than the second order updates from differentiating through the gradient update. Past work has observed that ReLU neural networks are locally almost linear~\citep{linear}, which suggests that second derivatives may be close to zero in most cases, partially explaining the good performance of the first-order approximation. This approximation removes the need for computing Hessian-vector products in an additional backward pass, which we found led to roughly $33\%$ speed-up in network computation.
\subsection{Reinforcement Learning}
\iffalse
\begin{figure}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.71) \linethickness{0.5pt}
\put(-0.01,0.0){\includegraphics[width=0.51\columnwidth]{point_results.png}}
\put(1.03,-0.1){\includegraphics[width=0.48\columnwidth]{pretrain_paths_viz.png}}
\put(1.03,0.36){\includegraphics[width=0.48\columnwidth]{maml_paths_viz.png}}
\end{picture}
\caption{Left: quantitative results from 2D navigation task, Right: qualitative comparison between model learned with MAML and with fine-tuning from a pretrained network.
\label{fig:2d}
\vspace{-0.5cm}
}
\end{figure}
\fi
\begin{figure}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,1.05) \linethickness{0.5pt}
\put(0.4,0.4){\includegraphics[width=0.51\columnwidth]{point_results.png}}
\put(0.99,-0.1){\includegraphics[width=0.51\columnwidth]{pretrain_paths_viz.png}}
\put(0.0,-0.1){\includegraphics[width=0.51\columnwidth]{maml_paths_viz.png}}
\end{picture}
\caption{Top: quantitative results from 2D navigation task, Bottom: qualitative comparison between model learned with MAML and with fine-tuning from a pretrained network.
\label{fig:2d}
\vspace{-0.6cm}
}
\end{figure}
\begin{figure*}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.56) \linethickness{0.5pt}
\put(2.72,-0.1){\includegraphics[width=0.49\columnwidth]{antdirec_results.png}}
\put(1.80,-0.1){\includegraphics[width=0.49\columnwidth]{ant_results.png}}
\put(0.885,-0.1){\includegraphics[width=0.49\columnwidth]{cheetahdirec_results.png}}
\put(-0.03,-0.1){\includegraphics[width=0.49\columnwidth]{cheetah_results.png}}
\put(3.28,0.1){\includegraphics[width=0.2\columnwidth]{maml_locolegend_transparent.png}}
\put(3.69, 0.225){\includegraphics[width=0.22\columnwidth]{cheetahshot.png}}
\put(3.69,-0.07){\includegraphics[width=0.22\columnwidth]{antshot.png}}
\end{picture}
\caption{Reinforcement learning results for the half-cheetah and ant locomotion tasks, with the tasks shown on the far right. Each gradient step requires additional samples from the environment, unlike the supervised learning tasks. The results show that MAML can adapt to new goal velocities and directions substantially faster than conventional pretraining or random initialization, achieving good performs in just two or three gradient steps. We exclude the goal velocity, random baseline curves, since the returns are much worse ($<-200$ for cheetah and $<-25$ for ant).
\label{fig:loco}
\vspace{-0.5cm}
}
\end{figure*}
\iffalse
\begin{figure}
\setlength{\unitlength}{0.5\columnwidth}
\begin{picture}(1.99,0.65) \linethickness{0.5pt}
\put(0.0,-0.1){\includegraphics[height=0.35\columnwidth]{cheetahshot.png}}
\put(1,-0.1){\includegraphics[width=0.49\columnwidth]{antshot.png}}
\end{picture}
\caption{Illustration of the locomotion environments.
\label{fig:tasks}
\vspace{-0.5cm}
}
\end{figure}
\fi
To evaluate MAML on reinforcement learning problems, we constructed several sets of tasks based off of the simulated continuous control environments in the rllab benchmark suite~\cite{benchmarking}. We discuss the individual domains below. In all of the domains, the model trained by MAML is a neural network policy with two hidden layers of size $100$, with ReLU nonlinearities. The gradient updates are computed using vanilla policy gradient (REINFORCE)~\cite{reinforce}, and we use trust-region policy optimization (TRPO) as the meta-optimizer~\cite{trpo}. In order to avoid computing third derivatives, we use finite differences to compute the Hessian-vector products for TRPO. For both learning and meta-learning updates, we use the standard linear feature baseline proposed by~\citet{benchmarking}, which is fitted separately at each iteration for each sampled task in the batch. We compare to three baseline models: (a) pretraining one policy on all of the tasks and then fine-tuning, (b) training a policy from randomly initialized weights, and (c) an oracle policy which receives the parameters of the task as input, which for the tasks below corresponds to a goal position, goal direction, or goal velocity for the agent. The baseline models of (a) and (b) are fine-tuned with gradient descent with a manually tuned step size. Videos of the learned policies can be viewed at \mbox{\url{sites.google.com/view/maml}}
\vspace{-0.1cm}
\noindent {\bf 2D Navigation.}
In our first meta-RL experiment, we study a set of tasks where a point agent must move to different goal positions in 2D, randomly chosen for each task within a unit square. The observation is the current 2D position, and actions correspond to velocity commands clipped to be in the range $[-0.1,0.1]$. The reward is the negative squared distance to the goal, and episodes terminate when the agent is within $0.01$ of the goal or at the horizon of $H=100$. The policy was trained with MAML to maximize performance after $1$ policy gradient update using $20$ trajectories. Additional hyperparameter settings for this problem and the following RL problems are in Appendix~\ref{app:hyperrl}.
In our evaluation, we compare adaptation to a new task with up to 4 gradient updates, each with $40$ samples. The results in Figure~\ref{fig:2d} show the adaptation performance of models that are initialized with MAML, conventional pretraining on the same set of tasks, random initialization, and an oracle policy that receives the goal position as input. The results show that MAML can learn a model that adapts much more quickly in a single gradient update, and furthermore continues to improve with additional updates.
\vspace{-0.1cm}
\noindent {\bf Locomotion.}
To study how well MAML can scale to more complex deep RL problems, we also study adaptation on high-dimensional locomotion tasks with the MuJoCo simulator ~\cite{mujoco}. The tasks require two simulated robots -- a planar cheetah and a 3D quadruped (the ``ant'') -- to run in a particular direction or at a particular velocity. In the goal velocity experiments, the reward is the negative absolute value between the current velocity of the agent and a goal, which is chosen uniformly at random between $0.0$ and $2.0$ for the cheetah and between $0.0$ and $3.0$ for the ant. In the goal direction experiments, the reward is the magnitude of the velocity in either the forward or backward direction, chosen at random for each task in $p(\mathcal{T})$. The horizon is $H=200$, with $20$ rollouts per gradient step for all problems except the ant forward/backward task, which used $40$ rollouts per step. The results in Figure~\ref{fig:loco} show that MAML learns a model that can quickly adapt its velocity and direction with even just a single gradient update, and continues to improve with more gradient steps. The results also show that, on these challenging tasks, the MAML initialization substantially outperforms random initialization and pretraining. In fact, pretraining is in some cases worse than random initialization, a fact observed in prior RL work~\cite{actormimic}.
\iffalse
\begin{table*}[t]
\caption{MAML for meta-RL on 2D point robot - average return \todo{make this a line plot with error bars.}}
\label{tbl:pointmass}
\begin{center}
\begin{tabular}{c|c|c|c|c}
\hline
num. gradient updates & 0 & 1 & 2 & 3\\
\hline
random init. & $-82.66$ & $-57.96$ & $-53.84$ & $-46.80$ \\
\hline
init from all tasks & $-40.81$ & $-19.73$ & $-11.72$ & $-8.47$ \\
\hline
MAML (ours) & $-40.41$ & $-11.68$ & $-3.33$ & $-3.23$ \\
\hline
\hline
oracle & -2.29 & n/a & n/a & n/a \\
\hline
\end{tabular}
\end{center}
\end{table*}
\fi
\section{Discussion and Future Work}
We introduced a meta-learning method based on learning easily adaptable model parameters through gradient descent. Our approach has a number of benefits. It is simple and does not introduce any learned parameters for meta-learning. It can be combined with any model representation that is amenable to gradient-based training, and any differentiable objective, including classification, regression, and reinforcement learning. Lastly, since our method merely produces a weight initialization, adaptation can be performed with any amount of data and any number of gradient steps, though we demonstrate state-of-the-art results on classification with only one or five examples per class. We also show that our method can adapt an RL agent using policy gradients and a very modest amount of experience.
Reusing knowledge from past tasks may be a crucial ingredient in making high-capacity scalable models, such as deep neural networks, amenable to fast training with small datasets. We believe that this work is one step toward a simple and general-purpose meta-learning technique that can be applied to any problem and any model. Further research in this area can make multitask initialization a standard ingredient in deep learning and reinforcement learning.
\section*{Acknowledgements} The authors would like to thank Xi Chen and Trevor Darrell for helpful discussions, Yan Duan and Alex Lee for technical advice, Nikhil Mishra, Haoran Tang, and Greg Kahn for feedback on an early draft of the paper, and the anonymous reviewers for their comments. This work was supported in part by an ONR PECASE award and an NSF GRFP award.
|
\section{Introduction}\label{sec:intro}
Factorization theorems play an important role in understanding the all orders behavior of observables in Quantum Chromodynamics (QCD). While typically formulated at leading power, the structure of subleading power corrections is of significant theoretical and practical interest. A convenient formalism for studying factorization in QCD is the Soft Collinear Effective Theory (SCET) \cite{Bauer:2000ew, Bauer:2000yr, Bauer:2001ct, Bauer:2001yt}, an effective field theory describing the soft and collinear limits of QCD. SCET allows for a systematic power expansion in $\lambda \ll 1$ at the level of the Lagrangian, and simplifies many aspects of factorization proofs \cite{Bauer:2002nz}. SCET has been used to study power corrections at the level of the amplitude \cite{Larkoski:2014bxa} and to derive factorization theorems at subleading power for $B$ decays \cite{Lee:2004ja,Beneke:2004in,Hill:2004if,Bosch:2004cb,Beneke:2004rc,Paz:2009ut,Benzke:2010js}. More recently, progress has been made towards understanding subleading power corrections for event shape observables \cite{Freedman:2013vya,Freedman:2014uta,Moult:2016fqy,Feige:2017zci}.
In this paper, we focus on the power suppressed hard scattering operators describing the gluon initiated production (or decay) of a color singlet scalar. We present a complete operator basis to $\mathcal{O}(\lambda^2)$ in the SCET power expansion using operators of definite helicity \cite{Moult:2015aoa,Kolodrubetz:2016uim,Feige:2017zci}, and discuss how helicity selection rules simplify the structure of the basis. We also classify all operators which can contribute at the cross section level at $\mathcal{O}(\lambda^2)$, and discuss the structure of interference terms between different operators in the squared matrix element. We then perform the tree level matching onto our operators. These results can be used to study subleading power corrections either in fixed order, or resummed perturbation theory, and compliment our recent analysis for the case of $q\bar q$ initiated production \cite{Feige:2017zci}.
We will consider the production of a color singlet final state, which we take for concreteness to be the Higgs, with the underlying hard Born process
\begin{equation} \label{eq:interaction}
g_a(q_a)\, g_b(q_b) \to H(q_1)
\,,\end{equation}
where $g_{a,b}$ denote the colliding gluons, and $H$ the outgoing Higgs particle. We work in the Higgs effective theory, with an effective Higgs gluon coupling
\begin{align}\label{eq:heft_intro}
\mathcal{L}_{\text{hard}}=\frac{C_1(m_t, \alpha_s)}{12\pi v}G^{\mu \nu}G_{\mu \nu} H\,,
\end{align}
obtained from integrating out the top quark. Here $v=(\sqrt{2} G_F)^{-1/2}=246$ GeV, and the matching coefficient is known to $\mathcal{O}(\alpha_s^3)$ \cite{Chetyrkin:1997un}.
The active-parton exclusive jet cross section corresponding to \eq{interaction} can be proven to factorize for a variety of jet resolution variables. For concreteness we will take the case of beam thrust, $\tau_B$. The leading power factorized expression for the beam thrust cross section can be written schematically in the form \cite{Stewart:2010tn}
\begin{align} \label{eq:sigma}
\frac{\mathrm{d}\sigma^{(0)}}{\mathrm{d} \tau_B} &=
\int\!\mathrm{d} x_a\, \mathrm{d} x_b\, \mathrm{d} \Phi(q_a \!+ q_b; q_1)\, M(\{q_1\})\
H_g^{(0)}(\{q_i\})\:
\Bigl[ B_g^{(0)} B_g^{(0)} \Bigr]\otimes S_g^{(0)}
\,,\end{align}
where the $x_{a,b}$ denote the momentum fractions of the incoming partons, $\mathrm{d} \Phi$ denotes the Lorentz-invariant phase space for the Born process in \eq{interaction}, and $M(\{q_i\})$ denotes the measurement made on the color singlet final state.\footnote{By referring to active-parton factorization we imply that this formula ignores contributions from proton spectator interactions~\cite{Gaunt:2014ska} that occur through the Glauber Lagrangian of Ref.~\cite{Rothstein:2016bsq}. There are also perturbative corrections at ${\cal O}(\alpha_s^4)$ that are described by a single function $B_{gg}$ in place of $B_gB_g$~\cite{Zeng:2015iba,Rothstein:2016bsq}.} The dependence on the underlying hard interaction is encoded in the hard function $\widehat{H}(\{q_i\})$ and the trace is over color. The soft function $\widehat{S}$ describes soft radiation, and the beam functions $B_i$ describe energetic initial-state radiation along the beam directions \cite{Stewart:2009yx}. The factorization theorem of \Eq{eq:sigma} allows logarithms of $\tau_B$ to be resummed to all orders through the renormalization group evolution of the hard, beam and soft functions.
The factorization formula in \Eq{eq:sigma} captures all terms in the cross section scaling as $\tau_B^{-1}$, including delta function terms. More generally the cross section can be expanded in powers of $\tau_B$ as,
\begin{align}\label{eq:cross_expand}
\frac{\mathrm{d}\sigma}{\mathrm{d}\tau_B} &=\frac{\mathrm{d}\sigma^{(0)}}{\mathrm{d}\tau_B} +\frac{\mathrm{d}\sigma^{(1)}}{\mathrm{d}\tau_B} +\frac{\mathrm{d}\sigma^{(2)}}{\mathrm{d}\tau_B}+\frac{\mathrm{d}\sigma^{(3)}}{\mathrm{d}\tau_B} +{\cal O}(\tau)\,.
\end{align}
Here the superscript refers to the suppression in powers of $\sqrt{\tau_B}$ relative to the leading power cross section. This particular convention is chosen due to the power expansion in SCET, where one typically takes the SCET power counting parameter $\lambda$ to scale like $\lambda^2 \sim \tau_B$. Odd orders in \Eq{eq:cross_expand} are expected to vanish, and we will show this explicitly for $\mathrm{d}\sigma^{(1)}/\mathrm{d}\tau_B$. The first non-vanishing power correction to the cross section then arises from $\mathrm{d}\sigma^{(2)}/\mathrm{d}\tau_B$, which contains all terms that scale like $\mathcal{O}(\tau_B^0)$.
It is generally expected that the power corrections in \Eq{eq:cross_expand} obey a factorization formula similar to that of \Eq{eq:sigma}. Schematically,
\begin{align} \label{eq:sigma_sub}
&\hspace{-0.25cm}\frac{\mathrm{d}\sigma^{(n)}}{\mathrm{d}\tau_B} =
\int\!\mathrm{d} x_a\, \mathrm{d} x_b\, \mathrm{d} \Phi(q_a \!+ q_b; q_1)\,M(\{q_1\})\
\sum_{j} H^{(n_{Hj})}_{j} \otimes
\Big[ B^{(n_{Bj})}_{j} B^{(n'_{Bj})}_{j}\Big] \otimes S_j^{(n_{Sj})}
,\end{align}
where $j$ sums over the multiple contributions that appear at each order, $n_{Hj}+n_{Bj}+n_{Bj}'+n_{Sj}=n$, and $\otimes$ denotes a set of convolutions, whose detailed structure has not been specified and is known to be more complicated than typical leading power factorization theorems. We also let $\otimes$ include nontrivial color contractions. The derivation of such a formula would enable for the resummation of subleading power logarithms using the renormalization group evolution of the different functions appearing in \Eq{eq:sigma_sub}, allowing for an all orders understanding of power corrections to the soft and collinear limits.
To derive a factorization theorem in SCET, QCD is matched onto SCET, which consists of hard scattering operators in $\mathcal{L}_\mathrm{hard}$ and a Lagrangian $\mathcal{L}_\mathrm{dyn}$ describing the dynamics of soft and collinear radiation
\begin{align}
\mathcal{L}_{\text{SCET}}=\mathcal{L}_\mathrm{hard}+\mathcal{L}_\mathrm{dyn} \,.
\end{align}
The dynamical Lagrangian can be divided into two parts
\begin{align}
\mathcal{L}_\mathrm{dyn}=\mathcal{L}_{\text{fact}}+\mathcal{L}_{G}^{(0)} \,.
\end{align}
Here $\mathcal{L}_{G}^{(0)}$ is the leading power Glauber Lagrangian determined in Ref.~\cite{Rothstein:2016bsq} which couples together soft and collinear fields in an apriori non-factorizable manner, and $\mathcal{L}_{\text{fact}}$ includes both the leading interactions which can be factorized into independent soft and collinear Lagrangians, and subleading power interactions which are factorizable as products of soft and collinear fields. Our focus here is on determining the subleading power $\mathcal{L}_\mathrm{hard}$ for $gg\to H$, and $\mathcal{L}_\mathrm{dyn}$ only plays a minor role when we carry out explicit matching calculations (and $\mathcal{L}_{G}^{(0)}$ does not play a role at all since these matching calculations are tree level).
The hard scattering operators are process dependent, while the Lagrangian $\mathcal{L}_\mathrm{dyn}$ is universal and the relevant terms for our analysis are known in SCET to $\mathcal{O}(\lambda^2)$ in the power expansion \cite{Manohar:2002fd,Chay:2002vy,Beneke:2002ni,Beneke:2002ph,Pirjol:2002km,Bauer:2003mga}. A field redefinition can be performed in the effective theory \cite{Bauer:2002nz} which allows for the decoupling of leading power soft and collinear interactions in $\mathcal{L}_{\text{fact}}$. If $\mathcal{L}_{G}^{(0)}$ is proven to be irrelevant, then the Hilbert spaces for the soft and collinear dynamics are factorized, and a series of algebraic manipulations can be used to write the cross section as a product of squared matrix elements, each involving only collinear or soft fields. This provides a field theoretic definition of each of the functions appearing in \Eq{eq:sigma_sub} in terms of hard scattering operators and Lagrangian insertions in SCET. Since the Lagrangian insertions are universal, the remaining ingredient which is required to derive a subleading power factorization theorem for the $gg\to H$ process is a complete basis of subleading power hard scattering operators. The derivation of a basis, which is the goal of this paper, provides the groundwork for a systematic study of power corrections for color singlet production through gluon fusion.
An important application of the results presented in this paper is to the calculation of subleading power corrections to event shape observables for $gg\to H$, such as $0$-jettiness \cite{Stewart:2010tn}. Recently, there has been considerable interest in the use of event shape observables for performing NNLO fixed order subtractions using the $q_T$ \cite{Catani:2007vq} or $N$-jettiness \cite{Boughezal:2015aha,Gaunt:2015pea} subtraction schemes. These ideas have been applied to color singlet production \cite{Catani:2009sm,Ferrera:2011bk,Catani:2011qz,Grazzini:2013bna,Cascioli:2014yka,Ferrera:2014lca,Gehrmann:2014fva,Grazzini:2015nwa,Grazzini:2015hta,Campbell:2016yrh,Boughezal:2016wmq}, to the production of a single jet in association with a color singlet particle \cite{Boughezal:2015aha,Boughezal:2015dva,Boughezal:2016isb,Boughezal:2016dtm}, and to inclusive photon production \cite{Campbell:2016lzl}. By analytically computing the power corrections for the subtractions, their stability and numerical accuracy can be significantly improved. This was shown explicitly in \cite{Moult:2016fqy} with the SCET based analytic calculation of the leading power corrections for $0$-jettiness for $q\bar q$ initiated Drell Yan like production of a color singlet, and it would be interesting to extend this calculation to $gg\to H$. For a direct calculation of the power corrections in QCD, see \cite{Boughezal:2016zws}.
An outline of this paper is as follows. In \Sec{sec:review} we provide a brief review of SCET and of the helicity building blocks required for constructing subleading operators in SCET. In \Sec{sec:basis} we present a complete basis of operators to $\mathcal{O}(\lambda^2)$ for the gluon initiated production of a color singlet, and carefully classify which operators can contribute to the cross section at $\mathcal{O}(\lambda^2)$. In \Sec{sec:matching} we perform the tree level matching to the relevant operators. We conclude and discuss directions for future study in \Sec{sec:conclusions}.
\section{Helicity Operators in SCET}\label{sec:review}
In this section we briefly review salient features of SCET, as well as the use of helicity operators in SCET. Reviews of SCET can be found in \Refs{iain_notes,Becher:2014oda}, and more detailed discussions on the use of helicity operators can be found in \Refs{Moult:2015aoa,Kolodrubetz:2016uim,Feige:2017zci}.
\subsection{SCET}\label{sec:review_scet}
SCET is an effective field theory of QCD describing the interactions of collinear and soft particles in the presence of a hard interaction \cite{Bauer:2000ew, Bauer:2000yr, Bauer:2001ct, Bauer:2001yt, Bauer:2002nz}. Collinear particles are characterized by a large momentum along a particular light-like direction, while soft particles are characterized by having a small momentum with homogenous scaling of all its components. For each jet direction present in the problem we define two light-like reference vectors $n_i^\mu$ and $\bn_i^\mu$ such that $n_i^2 = \bn_i^2 = 0$ and $n_i\!\cdot\!\bn_i = 2$. We can then write any four-momentum $p$ as
\begin{equation} \label{eq:lightcone_dec}
p^\mu = \bn_i\!\cdot\! p\,\frac{n_i^\mu}{2} + n_i\!\cdot\! p\,\frac{\bn_i^\mu}{2} + p^\mu_{n_i\perp}\
\,.\end{equation}
A particle with momentum $p$ close to the $\vec{n}_i$ direction will be referred to as $n_i$-collinear. In lightcone coordinates its momenta scale like $(n_i\!\cdot\! p, \bn_i \!\cdot\! p, p_{n_i\perp}) \sim \bn_i\!\cdot\! p$ $\,(\lambda^2,1,\lambda)$. Here $\lambda \ll 1$ is a formal power counting parameter determined by the measurements or kinematic restrictions imposed on the QCD radiation. The choice of reference vectors is not unique, and any two reference vectors, $n_i$ and $n_i'$, with $n_i\cdot n_i' \sim \ord{\lambda^2}$ describe the same physics. The freedom in the choice of $n_i$ is represented in the effective theory as a symmetry known as reparametrization invariance (RPI) \cite{Manohar:2002fd,Chay:2002vy}. More explicitly, there are three classes of RPI transformations under which the EFT is invariant
\begin{alignat}{3}\label{eq:RPI_def}
&\text{RPI-I} &\qquad & \text{RPI-II} &\qquad & \text{RPI-III} {\nonumber} \\
&n_{i \mu} \to n_{i \mu} +\Delta_\mu^\perp &\qquad & n_{i \mu} \to n_{i \mu} &\qquad & n_{i \mu} \to e^\alpha n_{i \mu} {\nonumber} \\
&\bar n_{i \mu} \to \bar n_{i \mu} &\qquad & \bar n_{i \mu} \to \bar n_{i \mu} +\epsilon_\mu^\perp &\qquad & \bar n_{i \mu} \to e^{-\alpha} \bar n_{i \mu}\,.
\end{alignat}
The transformation parameters are assigned the power counting $\Delta^\perp \sim \lambda$, $\epsilon^\perp \sim \lambda^0$, and $\alpha\sim \lambda^0$. Additionally, while $\alpha$ can be a finite parameter, the parameters $\Delta^\perp$ and $\epsilon^\perp$ are infinitesimal, and satisfy $n_i\cdot \Delta^\perp=\bar n_i\cdot \Delta^\perp=n_i \cdot \epsilon^\perp=\bar n_i \cdot \epsilon^\perp=0$. RPI symmetries can be used to relate operators at different orders in the power expansion, and will be used in this paper to relate the Wilson coefficients of several subleading power operators to the leading power Wilson coefficients for the $gg\to H$ process. Furthermore, the RPI-III symmetry will constrain the form of the Wilson coefficients of our subleading power operators. At tree level the Wilson coefficients are simply rational functions of the large momentum components of the fields appearing in the operator, which must satisfy the rescaling symmetries of RPI-III.
SCET is constructed by decomposing momenta into label and residual components
\begin{equation} \label{eq:label_dec}
p^\mu = \lp^\mu + k^\mu = \bn_i \!\cdot\!\lp\, \frac{n_i^\mu}{2} + \lp_{n_i\perp}^\mu + k^\mu\,.
\,\end{equation}
The momenta $\bn_i \cdot\lp \sim Q$ and $\lp_{n_i\perp} \sim \lambda Q$ are referred to as the label components, where $Q$ is a typical scale of the hard interaction, while $k\sim \lambda^2 Q$ is a small residual momentum describing fluctuations about the label momentum. Fields with momenta of definite scaling are obtained by performing a multipole expansion. Explicitly, the effective theory consists of collinear quark and gluon fields for each collinear direction, as well as soft quark and gluon fields. Independent gauge symmetries are enforced for each set of fields, which have support for the corresponding momenta carried by that field \cite{Bauer:2003mga}. The leading power gauge symmetry is exact, and is not corrected at subleading powers.
In SCET, fields for $n_i$-collinear quarks and gluons, $\xi_{n_i,\lp}(x)$ and $A_{n_i,\lp}(x)$, are labeled by their collinear direction $n_i$ and their large momentum $\lp$. The collinear fields are written in a mixed representation, namely they are written in position space with respect to the residual momentum and in momentum space with respect to the large momentum components. Derivatives acting on collinear fields give the residual momentum dependence, which scales as $i \partial^\mu \sim k \sim \lambda^2 Q$, whereas the label momentum operator $\mathcal{P}^\mu$ gives the label momentum component. It acts on a collinear field as $\mathcal{P}^\mu\, \xi_{n_i,\lp} = \lp^\mu\, \xi_{n_i,\lp}$. Note that we do not need an explicit $n_i$ label on the label momentum operator, since it is implied by the field that the label momentum operator is acting on. We will use the shorthand notation $\overline {\mathcal P} = \bn_i\!\cdot\!\mathcal{P}$. We will often suppress the explicit momentum labels on the collinear fields, keeping only the label of the collinear sector, ${n_i}$. Of particular relevance for the construction of subleading power operators is the $\mathcal{P}_\perp$ operator, which identifies the $\mathcal{O}(\lambda)$ perp momenta between two collinear fields within a collinear sector.
Soft degrees of freedom are described in SCET by quark and gluon fields $q_{us}(x)$ and $A_{us}(x)$. In this paper we will restrict ourselves to the SCET$_\text{I}$ theory where the soft degrees of freedom are referred to as ultrasoft so as to distinguish them from the soft modes of SCET$_\text{II}$ \cite{Bauer:2002aj}. The operators we construct are also applicable in the SCET$_\text{II}$ theory, but additional soft operators would be required. For a more detailed discussion see \Ref{Feige:2017zci}. The ultrasoft fields carry residual momenta, $i \partial^\mu \sim \lambda^2Q$, but do not carry label momenta, since they are not associated with any collinear direction. Correspondingly, they also do not carry a collinear sector label. The ultrasoft fields are able to exchange residual momenta between distinct collinear sectors while remaining on-shell.
SCET is constructed such that manifest power counting in the expansion parameter $\lambda$ is maintained at every stage of a calculation. All fields have a definite power counting \cite{Bauer:2001ct}, shown in \Tab{tab:PC}, and the SCET Lagrangian is expanded as a power series in $\lambda$
\begin{align} \label{eq:SCETLagExpand}
\mathcal{L}_{\text{SCET}}=\mathcal{L}_\mathrm{hard}+\mathcal{L}_\mathrm{dyn}= \sum_{i\geq0} \mathcal{L}_\mathrm{hard}^{(i)}+
{\cal L}_G^{(0)} + \sum_{i\geq0} \mathcal{L}^{(i)} \,.
\end{align}
Here $(i)$ denotes objects at ${\cal O}(\lambda^i)$ in the power counting. The Lagrangians $ \mathcal{L}_\mathrm{hard}^{(i)}$ contain the hard scattering operators $O^{(i)}$, and are determined by an explicit matching calculation. The hard scattering operators encode all process dependence, while the $\mathcal{L}^{(i)}$ describe the dynamics of ultrasoft and collinear modes in the effective theory, and are universal. The terms we need are explicitly known to $\mathcal{O}(\lambda^2)$, and can be found in a summarized form in \cite{iain_notes}. Finally, ${\cal L}_G^{(0)} $ is the leading power Glauber Lagrangian \cite{Rothstein:2016bsq}, which describes the leading power coupling of soft and collinear degrees of freedom through potential operators.
\begin{table}
\begin{center}
\begin{tabular}{| l | c | c |c |c|c| r| }
\hline
Operator & $\mathcal{B}_{n_i\perp}^\mu$ & $\chi_{n_i}$& $\mathcal{P}_\perp^\mu$&$q_{us}$&$D_{us}^\mu$ \\
Power Counting & $\lambda$ & $\lambda$& $\lambda$& $\lambda^3$& $\lambda^2$ \\
\hline
\end{tabular}
\end{center}
\caption{
Power counting for building block operators in $\text{SCET}_\text{I}$.
}
\label{tab:PC}
\end{table}
In this paper we will be interested in subleading power hard scattering operators, in particular, $\mathcal{L}_\mathrm{hard}^{(1)}$ and $\mathcal{L}_\mathrm{hard}^{(2)}$. The hard effective Lagrangian at each power is given by a product of hard scattering operators and Wilson coefficients,
\begin{align} \label{eq:Leff_sub_explicit}
\mathcal{L}^{(j)}_{\text{hard}} = \sum_{\{n_i\}} \sum_{A,\cdot\cdot}
\bigg[ \prod_{i=1}^{\ell_A} \int \! \! \mathrm{d} \omega_i \bigg] \,
& \vec O^{(j)\dagger}_{A\,\lotsdots}\big(\{n_i\};
\omega_1,\ldots,\omega_{\ell_A}\big) {\nonumber}\\
& \times
\vec{C}^{(j)}_{A\,\lotsdots}\big(\{n_i\};\omega_1,\ldots,\omega_{\ell_A} \big)
\,.
\end{align}
The appropriate collinear sectors $\{n_i\}$ are determined by directions found in the collinear states of the hard process being considered. If there is a direction $n_1'$ in the state then we sum over the cases where each of $n_1$, $\ldots$, $n_4$ is set equal to this $n_1'$.\footnote{Technically the $n_i$ in $\{n_i\}$ are representatives of an equivalence class determined by demanding that distinct classes $\{n_i\}$ and $\{n_j\}$ have $n_i\cdot n_j\gg \lambda^2$.} The sum over $A,\cdot\cdot$ in \eq{Leff_sub_explicit} runs over the full basis of operators that appear at this order, which are specified by either explicit labels $A$ and/or helicity labels $\cdot\cdot$ on the operators and coefficients. The $\vec{C}^{(j)}_{A}$ are also vectors in the color subspace in which the $\mathcal{O}(\lambda^j)$ hard scattering operators $\vec O_A^{(j)\dagger}$ are decomposed. Explicitly, in terms of color indices, we follow the notation of \Ref{Moult:2015aoa} and have
\begin{align} \label{eq:Opm_color}
\vec O^\dagger_\lotsdots &= O_\lotsdots^{a_1\dotsb \alpha_n}\, \bar{T}^{\, a_1\dotsb \alpha_n}
\,, {\nonumber}\\
C_{\lotsdots}^{a_1\dotsb\alpha_n}
&= \sum_k C_{\lotsdots}^k T_k^{a_1\dotsb\alpha_n}
\equiv \bar{T}^{ a_1\dotsb\alpha_n} \vec{C}_{\lotsdots}
\,.\end{align}
Here $\bar{T}^{\, a_1\dotsb\alpha_n}$ is a row vector of color structures that spans the color conserving subspace. The $a_i$ are adjoint indices and the $\alpha_i$ are fundamental indices. The color structures do not necessarily have to be independent, but must be complete.
Hard scattering operators involving collinear fields are constructed out of products of fields and Wilson lines that are invariant under collinear gauge transformations~\cite{Bauer:2000yr,Bauer:2001ct}. The field building blocks for these operators are collinear gauge-invariant quark and gluon fields, defined as
\begin{align} \label{eq:chiB}
\chi_{{n_i},\omega}(x) &= \Bigl[\delta(\omega - \overline {\mathcal P}_{n_i})\, W_{n_i}^\dagger(x)\, \xi_{n_i}(x) \Bigr]
\,,\\
\mathcal{B}_{{n_i}\perp,\omega}^\mu(x)
&= \frac{1}{g}\Bigl[\delta(\omega + \overline {\mathcal P}_{n_i})\, W_{n_i}^\dagger(x)\,i D_{{n_i}\perp}^\mu W_{n_i}(x)\Bigr]
\,. {\nonumber}
\end{align}
For this particular definition of $\chi_{{n_i},\omega}$, we have $\omega > 0$ for an incoming quark and $\omega < 0$ for an outgoing antiquark. For $\mathcal{B}_{{n_i},\omega\perp}$, $\omega > 0$ ($\omega < 0$) corresponds to outgoing (incoming) gluons. The covariant derivative in \eq{chiB} is given by,
\begin{equation}
i D_{{n_i}\perp}^\mu = \mathcal{P}^\mu_{{n_i}\perp} + g A^\mu_{{n_i}\perp}\,,
\end{equation}
and the collinear Wilson line is defined as
\begin{equation} \label{eq:Wn}
W_{n_i}(x) = \biggl[~\sum_\text{perms} \exp\Bigl(-\frac{g}{\overline {\mathcal P}_{n_i}}\,\bn\!\cdot\! A_{n_i}(x)\Bigr)~\biggr]\,.
\end{equation}
The emissions summed in the Wilson lines are $\ord{\lambda^0}$ in the power counting. The square brackets indicate that the label momentum operators act only on the fields in the Wilson line. The collinear Wilson line, $W_{n_i}(x)$, is localized with respect to the residual position $x$, so that
$\chi_{{n_i},\omega}(x)$ and $\mathcal{B}_{{n_i},\omega}^\mu(x)$ can be treated as local quark and gluon fields from the perspective of the ultrasoft degrees of freedom.
All operators in the theory must be invariant under ultrasoft gauge transformations. Collinear fields transform under ultrasoft gauge transformations as background fields of the appropriate representation.
Dependence on the ultrasoft degrees of freedom enters the operators through the ultrasoft quark field $q_{us}$, and the ultrasoft covariant derivative $D_{us}$, defined as
\begin{equation}
i D_{us}^\mu = i \partial^\mu + g A_{us}^\mu\,.
\end{equation}
Other operators, such as the ultrasoft gluon field strength, can be constructed from the ultrasoft covariant derivative. The power counting for these operators is shown in \Tab{tab:PC}.
The complete set of collinear and ultrasoft building blocks is summarized in \Tab{tab:PC}. These can be combined, along with Lorentz and Dirac structures, to construct a basis of hard scattering operators at any order in the SCET power counting. All other field and derivative combinations can be reduced to this set by the use of equations of motion and operator relations~\cite{Marcantonini:2008qn}. As shown in \Tab{tab:PC}, both the collinear quark and collinear gluon building block fields scale as ${\cal O}(\lambda)$. Therefore, while for most jet processes only a single collinear field appears in each sector at leading power, subleading power operators can involve multiple collinear fields in the same collinear sector, as well as $\mathcal{P}_\perp$ insertions. The scaling of an operator is simply obtained by adding up the powers for the building blocks it contains. This implies that at higher powers hard scattering operators involve more and more fields, or derivative insertions, leading to any increasingly complicated structure. Furthermore, to ensure that the effective theory completely reproduces all IR limits of the full theory, as well as to guarantee that the renormalization group evolution of the operators is closed, it is essential that operator bases in SCET are complete, namely all operators consistent with the symmetries of the problem must be included. Enumerating a minimal basis of operators becomes difficult at subleading power, and it is essential to be able to efficiently identify independent operators, as well as to make manifest all symmetries of the problem.
\subsection{Helicity Operators}\label{sec:review_helicity}
An efficient approach to simplify operator bases in SCET is to use operators of definite helicity \cite{Moult:2015aoa,Kolodrubetz:2016uim,Feige:2017zci}. This general philosophy is well known from the study of on-shell scattering amplitudes, where it leads to compact expressions, removes gauge redundancies, and makes symmetries manifest. The use of helicities is also natural in SCET since the effective theory is formulated as an expansion about identified light like directions with respect to which helicities are naturally defined, and collinear fields carry these directions as labels. Furthermore, since SCET is formulated in terms of collinear gauge invariant fields, see \Eq{eq:chiB}, one can naturally project onto physical polarizations. SCET helicity operators were introduced in \cite{Moult:2015aoa} where they were used to study leading power processes with high multiplicities. This was extended to subleading power in \cite{Kolodrubetz:2016uim} where it was shown that the use of helicity operators is also convenient when multiple fields appear in the same collinear sector. In this section we briefly review SCET helicity operators, since we will use them to simplify the structure of the subleading power basis for $gg\to H$. We will follow the notation and conventions of~\cite{Moult:2015aoa,Kolodrubetz:2016uim,Feige:2017zci}. A summary of the complete set of operators that we will use is given in \Tab{tab:helicityBB}.
We define collinear gluon and quark fields of definite helicity as
\begin{subequations}
\label{eq:cBpm_quarkhel_def}
\begin{align}
\label{eq:cBpm_def}
\mathcal{B}^a_{i\pm} &= -\varepsilon_{\mp\mu}(n_i, \bn_i)\,\mathcal{B}^{a\mu}_{n_i\perp,\omega_i}
\,, \\
\label{eq:quarkhel_def}
\chi_{i \pm}^\alpha &= \frac{1\,\pm\, \gamma_5}{2} \chi_{n_i, - \omega_i}^\alpha
\,,\qquad\quad
\bar{\chi}_{i \pm}^{\bar \alpha} = \bar{\chi}_{n_i, - \omega_i}^{\bar \alpha} \frac{1\,\mp\, \gamma_5}{2}\,.
\end{align}
\end{subequations}
Here $a$, $\alpha$, and ${\bar \alpha}$ are adjoint, $3$, and $\bar 3$ color indices respectively, and the $\omega_i$ labels on both the gluon and quark building blocks are taken to be outgoing, which is also used for our helicity convention. Using the standard spinor helicity notation (see e.g. \cite{Dixon:1996wi} for an introduction)
\begin{align} \label{eq:braket_def}
|p\rangle\equiv \ket{p+} &= \frac{1 + \gamma_5}{2}\, u(p)
\,,
& |p] & \equiv \ket{p-} = \frac{1 - \gamma_5}{2}\, u(p)
\,, \\
\bra{p} \equiv \bra{p-} &= \mathrm{sgn}(p^0)\, \bar{u}(p)\,\frac{1 + \gamma_5}{2}
\,,
& [p| & \equiv \bra{p+} = \mathrm{sgn}(p^0)\, \bar{u}(p)\,\frac{1 - \gamma_5}{2}
\,, {\nonumber}
\end{align}
with $p$ lightlike, the polarization vector of an outgoing gluon with momentum $p$ can be written
\begin{equation}
\varepsilon_+^\mu(p,k) = \frac{\mae{p+}{\gamma^\mu}{k+}}{\sqrt{2} \langle kp \rangle}
\,,\qquad
\varepsilon_-^\mu(p,k) = - \frac{\mae{p-}{\gamma^\mu}{k-}}{\sqrt{2} [kp]}
\,,\end{equation}
where $k\neq p$ is an arbitrary light-like reference vector, chosen to be $\bn_i$ in \eq{cBpm_def}.
Since fermions always arise in pairs, we can define currents with definite helicities. Here we will restrict to the case of two back to back directions, $n$ and $\bar n$, as is relevant for $gg\to H$. A more general discussion can be found in \Refs{Kolodrubetz:2016uim,Feige:2017zci}. We define helicity currents where the quarks are in opposite collinear sectors,
\begin{align} \label{eq:jpm_back_to_bacjdef}
& h=\pm 1:
& J_{n \bn \pm}^{{\bar \alpha}\beta}
& = \mp\, \sqrt{\frac{2}{\omega_n\, \omega_\bn}}\, \frac{ \varepsilon_\mp^\mu(n, \bn) }{\langle \bn \mp | n \pm\rangle} \, \bar{\chi}^{\bar \alpha}_{n\pm}\, \gamma_\mu \chi^\beta_{\bn \pm}
\,, \\
& h=0:
& J_{n \bn 0}^{{\bar \alpha}\beta}
& =\frac{2}{\sqrt{\vphantom{2} \omega_n \,\omega_\bn}\, [n \bn] } \bar \chi^{\bar \alpha}_{n+}\chi^\beta_{\bn-}
\,, \qquad
(J^\dagger)_{n \bn 0}^{{\bar \alpha}\beta}=\frac{2}{\sqrt{ \vphantom{2} \omega_n \, \omega_\bn} \langle n \bn \rangle } \bar \chi^{\bar \alpha}_{n-}\chi^\beta_{\bn+}
\,, {\nonumber}
\end{align}
as well as where the quarks are in the same collinear sector,
\begin{align}\label{eq:coll_subl}
& h=0:
& J_{i0}^{{\bar \alpha} \beta}
&= \frac{1}{2 \sqrt{\vphantom{2} \omega_{\bar \chi} \, \omega_\chi}}
\: \bar \chi^{\bar \alpha}_{i+}\, \Sl{\bar n}_i\, \chi^\beta_{i+}
\,,\qquad
J_{i\bar 0}^{{\bar \alpha} \beta}
= \frac{1}{2 \sqrt{\vphantom{2} \omega_{\bar \chi} \, \omega_\chi}}
\: \bar \chi^{\bar \alpha}_{i-}\, \Sl {\bar n}_i\, \chi^\beta_{i-}
\,, \\[5pt]
& h=\pm 1:
& J_{i\pm}^{{\bar \alpha} \beta}
&= \mp \sqrt{\frac{2}{ \omega_{\bar \chi} \, \omega_\chi}} \frac{\epsilon_{\mp}^{\mu}(n_i,\bar n_i)}{ \big(\langle n_i \mp | \bar{n}_i \pm \rangle \big)^2}\:
\bar \chi_{i\pm}^{\bar \alpha}\, \gamma_\mu \Sl{\bar n}_i\, \chi_{i\mp}^\beta
\,. {\nonumber}
\end{align}
Here $i$ can be either $n$ or $\bar n$. All of these currents are manifestly invariant under the RPI-III symmetry of SCET. The Feynman rules for all currents are very simple, and are given in~\cite{Feige:2017zci}. Note that the operators $J_{n \bn \pm}^{{\bar \alpha}\beta}$, $J_{i0}^{{\bar \alpha} \beta}$, and $J_{i\bar 0}^{{\bar \alpha} \beta}$ have quarks of the same chirality, and hence are the ones that will be generated by vector gauge bosons.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|cc|ccc|c|ccc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
Field: &
$\mathcal{B}_{i\pm}^a$ & $J_{ij\pm}^{{\bar \alpha}\beta}$ & $J_{ij0}^{{\bar \alpha}\beta}$
& $J_{i\pm}^{{\bar \alpha} \beta}$
& $J_{i0}^{{\bar \alpha} \beta}$ & $J_{i\bar 0}^{{\bar \alpha} \beta}$
& $\mathcal{P}^{\perp}_{\pm}$
& $\partial_{us(i)\pm}$ & $\partial_{us(i)0}$ & $\partial_{us(i)\bar{0}}$
\\[3pt]
Power counting: &
$\lambda$ & $\lambda^2$ & $\lambda^2$
& $\lambda^2$ & $\lambda^2$& $\lambda^2$ & $\lambda$
& $\lambda^2$ & $\lambda^2$ & $\lambda^2$
\\
Equation: &
(\ref{eq:cBpm_def}) & \multicolumn{2}{c|}{(\ref{eq:jpm_back_to_bacjdef})}
& \multicolumn{3}{c|}{(\ref{eq:coll_subl})} & (\ref{eq:Pperppm})
& \multicolumn{3}{c|}{(\ref{eq:partialus})}
\\
\hline
\end{tabular}\\
\vspace{.3cm}
\begin{tabular}{|c|cc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
Field: &
$\mathcal{B}^a_{us(i)\pm}$ &
\!\!$\mathcal{B}^a_{us(i)0}$
\\[3pt]
Power counting: &
$\lambda^2$ & $\lambda^2$
\\
Equation: &
\multicolumn{2}{c|}{(\ref{eq:Bus})}
\\
\hline
\end{tabular}
\end{center}
\vspace{-0.3cm}
\caption{The helicity building blocks in $\text{SCET}_\text{I}$ that will be used to construct a basis of hard scattering operators for $gg\to H$, together with their power counting order in the $\lambda$-expansion, and the equation numbers where their definitions may be found. The building blocks also include the conjugate currents $J^\dagger$ in cases where they are distinct from the ones shown.
}
\label{tab:helicityBB}
\end{table}
At subleading power one must also consider insertions of the $\mathcal{P}_{i\perp}^\mu$ operator. Note that we can drop the explicit $i$ index on the $\mathcal{P}_\perp$ operator, as it is implied by the field that the operator is acting on. The $\mathcal{P}_{\perp}^\mu$ operator acts on the perpendicular subspace defined by the vectors $n_i, \bar n_i$, so it is naturally written as
\begin{align} \label{eq:Pperppm}
\mathcal{P}_{+}^{\perp}(n_i,\bar n_i)=-\epsilon^-(n_i,\bar n_i) \cdot \mathcal{P}_{\perp}\,, \qquad \mathcal{P}_{-}^{\perp}(n_i,\bar n_i)=-\epsilon^+(n_i,\bar n_i) \cdot \mathcal{P}_{\perp}\,.
\end{align}
The $\mathcal{P}^\perp_\pm$ operator carry helicity $h=\pm 1$.
We use square brackets to denote which fields are acted upon by the $\mathcal{P}^{\perp}_{\pm}$ operator, for example
$\mathcal{B}_{i+} \left [ \mathcal{P}^{\perp}_{+} \mathcal{B}_{i-} \right] \mathcal{B}_{i-}$,
indicates that the $\mathcal{P}^{\perp}_{+}$ operator acts only on the middle field, whereas for currents, we use a curly bracket notation
\begin{align}\label{eq:p_perp_notation}
\big\{ \mathcal{P}^{\perp}_\lambda J_{i 0 }^{{\bar \alpha} \beta} \big\}
& = \frac{1}{2 \sqrt{\vphantom{2}\omega_{\bar \chi} \, \omega_\chi }} \:
\Big[ \mathcal{P}^{\perp}_{\lambda} \bar \chi^{\bar \alpha}_{i +}\Big] \Sl {\bar n}_i \chi^\beta_{i+}
\,, \\
\big\{ J_{i0 }^{{\bar \alpha} \beta} (\mathcal{P}^{\perp}_{\lambda})^\dagger \big\}
&= \frac{1}{2\sqrt{\vphantom{2}\omega_{\bar \chi} \, \omega_\chi}} \:
\bar \chi^{\bar \alpha}_{i+} \Sl {\bar n}_i \Big[ \chi^\beta_{i+} (\mathcal{P}^{\perp}_{\lambda})^\dagger \Big]
\,, {\nonumber}
\end{align}
to indicate which of the fields within the current is acted on.
To work with gauge invariant ultrasoft gluon fields, we construct our basis post BPS field redefinition. The BPS field redefinition is defined by \cite{Bauer:2002nz}
\begin{equation} \label{eq:BPSfieldredefinition}
\mathcal{B}^{a\mu}_{n\perp}\to \mathcal{Y}_n^{ab} \mathcal{B}^{b\mu}_{n\perp} , \qquad \chi_n^\alpha \to Y_n^{\alpha {\bar \beta}} \chi_n^\beta,
\end{equation}
and is performed in each collinear sector. Here $Y_n$, $\mathcal{Y}_n$ are fundamental and adjoint ultrasoft Wilson lines. For a general representation, r, the ultrasoft Wilson line is defined by
\begin{equation}
Y^{(r)}_n(x)=\bold{P} \exp \left [ ig \int\limits_{-\infty}^0 ds\, n\cdot A^a_{us}(x+sn) T_{(r)}^{a}\right]\,,
\end{equation}
where $\bold P$ denotes path ordering. The BPS field redefinition has the effect of decoupling ultrasoft and collinear degrees of freedom at leading power \cite{Bauer:2002nz}, and it accounts for the full physical path of ultrasoft Wilson lines~\cite{Chay:2004zn,Arnesen:2005nk}.
The BPS field redefinition introduces ultrasoft Wilson lines into the hard scattering operators. These Wilson lines can be arranged with the ultrasoft fields to define ultrasoft gauge invariant building blocks. In particular, the gauge covariant derivative in an arbitrary representation, $r$, can be sandwiched by Wilson lines and decomposed as
\begin{align}\label{eq:soft_gluon}
Y^{(r)\,\dagger}_{n_i} i D^{(r)\,\mu}_{us} Y^{(r)}_{n_i }=i \partial^\mu_{us} + [Y_{n_i}^{(r)\,\dagger} i D^{(r)\,\mu}_{us} Y^{(r)}_{n_i}]=i\partial^\mu_{us}+T_{(r)}^{a} g \mathcal{B}^{a\mu}_{us(i)}\,.
\end{align}
Here we have defined the ultrasoft gauge invariant gluon field by
\begin{align} \label{eq:softgluondef}
g \mathcal{B}^{a\mu}_{us(i)}= \left [ \frac{1}{in_i\cdot \partial_{us}} n_{i\nu} i G_{us}^{b\nu \mu} \mathcal{Y}^{ba}_{n_i} \right] \,.
\end{align}
In the above equations the derivatives act only within the square brackets. Note from \eq{softgluondef}, that $n_i\cdot \mathcal{B}^{a}_{us(i)}= 0$. The Wilson lines which remain after this procedure can be absorbed into a generalized color structure, $\bar{T}_{\mathrm{BPS}}$ (see \cite{Kolodrubetz:2016uim} for more details). Determining a complete basis of color structures is straightforward, and detailed examples are given in~\cite{Feige:2017zci}.
Having defined gauge invariant ultrasoft gluon fields, we can now define ultrasoft gauge invariant gluon helicity fields and derivative operators which mimic their collinear counterparts. For the ultrasoft gluon helicity fields we define the three building blocks
\begin{equation} \label{eq:Bus}
\mathcal{B}^a_{us(i)\pm} = -\varepsilon_{\mp\mu}(n_i, \bn_i)\,\mathcal{B}^{a\mu}_{us(i)},\qquad \mathcal{B}^a_{us(i)0} =\bar n_\mu \mathcal{B}^{a \mu}_{us(i)}
\,,\end{equation}
and similarly for the ultrasoft derivative operators
\begin{equation} \label{eq:partialus}
\partial_{us(i)\pm} = -\varepsilon_{\mp\mu}(n_i, \bn_i)\,\partial^{\mu}_{us},\qquad \partial_{us(i)0} =\bar n_{i\mu} \partial^{\mu}_{us}, \qquad \partial_{us(i)\bar 0} = n_{i \mu} \partial^{\mu}_{us}
\,.\end{equation}
Unlike for the gauge invariant collinear gluon fields, for the ultrasoft gauge invariant gluon field we use three building block fields to describe the two physical degrees of freedom because the ultrasoft gluons are not fundamentally associated with any direction. Without making a further gauge choice, their polarization vectors do not lie in the perpendicular space of any fixed external reference vector. When inserting ultrasoft derivatives into operators we will use the same curly bracket notation defined for the $\mathcal{P}_\perp$ operators in \Eq{eq:p_perp_notation}.
Gauge invariant ultrasoft quark fields can also appear explicitly in operator bases at subleading powers. From \Tab{tab:PC} we see that they power count as $\mathcal{O}(\lambda^3)$, and are therefore not relevant for our construction of an $\mathcal{O}(\lambda^2)$ operator basis. Details on the structure of subleading power helicity operators involving ultrasoft quarks can be found in~\cite{Feige:2017zci}. It is important to emphasize that although ultrasoft quarks do not appear in the hard scattering operators at $\mathcal{O}(\lambda^2)$ they do appear in the calculation of cross sections or amplitudes at $\mathcal{O}(\lambda^2)$ due to subleading power Lagrangian insertions in the effective (examples where they play an important role for factorization in $B$-decays include both exclusive decays~\cite{Bauer:2002aj,Mantry:2003uz,Beneke:2003pa} and inclusive decays~\cite{Bosch:2004cb,Lee:2004ja,Beneke:2004in}). Such ultrasoft quark contributions also played an important role in the recent subleading power perturbative SCET calculation of \Ref{Moult:2016fqy}.
Finally, we note that the helicity operator basis presented in this section only provides a complete basis in $d=4$, and we have not discussed evanescent operators \cite{Buras:1989xd,Dugan:1990df,Herrlich:1994kh}. An extension of our basis to include evanescent operators would depend on the regularization scheme. However, in general additional building block fields would be required, for example an $\epsilon$ scalar gluon $\mathcal{B}^a_{\epsilon}$ to encode the $(-2\epsilon)$ transverse degrees of freedom of the gluon. As in standard loop calculations, we expect that the evanescent operators at each loop order could be straightforwardly identified and treated. Since we do not perform a one-loop matching to our operators, we leave a complete treatment of evanescent operators to future work.
\section{Operator Basis}\label{sec:basis}
In this section we enumerate a complete basis of power suppressed operators up to $\mathcal{O}(\lambda^2)$ for the process $gg\to H$. The organization of the operator basis in terms of helicity operators will make manifest a number of symmetries arising from helicity conservation, greatly reducing the operator basis. Helicity conservation is particularly powerful in this case due to the spin-$0$ nature of the Higgs. The complete basis of field structures is summarized in \Tab{tab:summary}. In \Sec{sec:discussion} we will show which operators contribute to the cross section at $\mathcal{O}(\lambda^2)$. These operators are indicated with a check mark in the table.
Examining \eq{Leff_sub_explicit} we see that the hard Lagrangian in SCET is written as a sum over label momenta of the hard operators. For the special case of two back-to-back collinear sectors this reduces to
\begin{align}\label{eq:sum_dir}
\mathcal{L}_{\text{hard}}^{(j)} = \sum_{n} \sum_{A,\cdot\cdot}
\bigg[ \prod_{i=1}^{\ell_A} \int \! \! \mathrm{d} \omega_i \bigg] \,
& \vec O^{(j)\dagger}_{A\,\lotsdots}\big(n,\bn;
\omega_1,\ldots,\omega_{\ell_A}\big) {\nonumber}\\
& \times
\vec{C}^{(j)}_{A\,\lotsdots}\big(n,\bn;\omega_1,\ldots,\omega_{\ell_A} \big)
\,.
\end{align}
When writing our basis, we therefore do not need to include operators which are identical up to the swap of $n\leftrightarrow \bar n$. This means that when writing an operator with different field structures in the two collinear sectors we are free to make an arbitrary choice for which is labeled $n$ and which $\bar n$, and this choice can be made independently for each operator. When squaring matrix elements, all possible interferences are properly incorporated by the sum over directions in \Eq{eq:sum_dir}.
As discussed in \Sec{sec:intro}, we will work in the Higgs effective theory with a Higgs gluon coupling given by the effective Lagrangian in \Eq{eq:heft_intro}. We therefore do not consider operators generated by a direct coupling of quarks to the Higgs. All quarks in the final state are produced by gluon splittings. The extension to include operators involving quarks coupling directly to Higgs, as relevant for $H\to b\bar b$, is straightforward using the helicity building blocks given in \Sec{sec:review_helicity}.
{
\renewcommand{\arraystretch}{1.4}
\begin{table}[t!]
\hspace{-0.15cm} \scalebox{0.9}{
\begin{tabular}{| c | l | l | c | c | c | }
\hline
Order & $\!$Category & Operators (equation number)
& \#$\!$ helicity & \#$\!$ of
& $\!\sigma_{2j}^{\mathcal{O}(\lambda^2)}\!\! \ne\! 0\!$
\\[-8pt]
& & & configs & \! color\! &
\\ \hline
$\mathcal{O}(\lambda^0)$
& $\! H gg$ & $O_{\mathcal{B} \lambda_1 \lambda_1}^{(0)ab}=\mathcal{B}^a_{n \lambda_1} \mathcal{B}^a_{\bar n \lambda_1} H$ \,(\ref{eq:hgg})
& 2 & 1 & $\checkmark$
\\ \hline
$\mathcal{O}(\lambda)$
& $\! H q \bar{q} g$
& $O_{\mathcal{B} n,\bar n \lambda_1 ( \lambda_i)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n,\bar n \lambda_1}^a\, J_{n\bar n\, \lambda_j}^{{\bar \alpha}\beta}\,H$ \,(\ref{eq:H1_basis},\ref{eq:H1_basis2})
& 4 & 1 & $\checkmark$
\\ \hline
$\mathcal{O}(\lambda^2)$
& $\! H q \bar q Q \bar Q$ & $O_{qQ1(\lambda_1;\lambda_2)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)n {\lambda_1}\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n {\lambda_2}\, }^{{\bar \gamma}\delta}\,H$ \,(\ref{eq:Z2_basis_qQ})
& 4 & 2 &
\\
& & $O_{qQ2(\lambda_1;\lambda_1)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q \bar{Q}) n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(Q \bar{q}) \bar{n} \, \lambda_1\, }^{{\bar \gamma}\delta}\,H $ \,(\ref{eq:Z2_basis_qQ_2})
& 2 & 2 &
\\
& & $O_{qQ3(\lambda_1;-\lambda_1)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q) n \bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n -\lambda_1\, }^{{\bar \gamma}\delta}\,H$ \,(\ref{eq:Z2_basis_qQ_3})
& 2 & 2 &
\\ \cline{2-6}
& $\! H q \bar q q \bar q$ & $O_{qq1(\lambda_1;\lambda_2)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)n {\lambda_1}\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n {\lambda_2}\, }^{{\bar \gamma}\delta}\,H$ \,(\ref{eq:Z2_basis_qq})
& 3 & 2 &
\\
& & $O_{qq3(\lambda_1;-\lambda_1)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q) n \bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(q) n\bar n -\lambda_1\, }^{{\bar \gamma}\delta}\,H$ \,(\ref{eq:Z2_basis_qq_3})
& 1 & 2 &
\\ \cline{2-6}
& $\! H q \bar q g g$ & $O_{\cB1\lambda_1 \lambda_2(\lambda_3)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n \lambda_1}^a \mathcal{B}_{\bar n \lambda_2}^b \, J_{n\,{\lambda_3} }^{{\bar \alpha}\beta} \,H $ \,(\ref{eq:Hqqgg_basis3})
& 4 & 3 & $\checkmark$
\\
& & $O_{\cB2\lambda_1 \lambda_2(\lambda_3)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n \lambda_1}^a \mathcal{B}_{\bar n \lambda_2}^b \, J_{n\,{\lambda_3} }^{{\bar \alpha}\beta} \,H $ \,(\ref{eq:Hqqgg_basis4})
& 2 & 3 &
\\ \cline{2-6}
& $\! H gggg$ & $O_{4g1\lambda_1 \lambda_2 \lambda_3 \lambda_4}^{(2)a b c d}
= S \mathcal{B}^a_{n \lambda_1} \mathcal{B}^b_{n \lambda_2} \mathcal{B}^c_{\bn \lambda_3} \mathcal{B}^d_{\bn \lambda_4} H $ \,(\ref{eq:H_basis_gggg_1})
& 3 & 9 &
\\
& & $O_{4g2\lambda_1 \lambda_2 \lambda_3 \lambda_4}^{(2)a b c d}
= S \mathcal{B}^a_{n \lambda_1} \mathcal{B}^b_{\bn \lambda_2} \mathcal{B}^c_{\bn \lambda_3} \mathcal{B}^d_{\bn \lambda_4} H$ \,(\ref{eq:H_basis_gggg_2})
& 2 & 9 & $\checkmark$
\\ \cline{2-6}
& $\mathcal{P}_\perp$ & $O_{\mathcal{P}\chi \lambda_1 (\lambda_2)[\lambda_{\mathcal{P}}]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n\lambda_1}^a \, \{J_{\bar n\, {\lambda_2} }^{{\bar \alpha}\beta}(\mathcal{P}_{\perp}^{\lambda_{\mathcal{P}}})^\dagger\}\,H$ \,(\ref{eq:Hqqgpperp_basis_same})
& 4 & 1 & $\checkmark$
\\
& & $O_{\mathcal{P}\mathcal{B} \lambda_1 \lambda_2 \lambda_3[\lambda_{\mathcal{P}}]}^{(2)abc}
= S\, \mathcal{B}_{n\lambda_1}^a\, \mathcal{B}_{\bar n \lambda_2}^b \left [\mathcal{P}_{\perp}^{\lambda_{\mathcal{P}}} \mathcal{B}_{\bar n \lambda_3}^c \right ]\! H\! $ \,(\ref{eq:Hgggpperp_basis})
& 4 & 2 & $\checkmark$
\\ \cline{2-6}
& $\!$Ultrasoft $\!\!\!$ & $O_{\chi (us(n))0:(\lambda_1)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(n)0}^a \, J_{n\bar n\,\lambda_1}^{{\bar \alpha}\beta}\,H $ \,(\ref{eq:soft_insert_basis})
& 2 & 1 &
\\
& & $O_{\chi (us(\bar n))0:(\lambda_1)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(\bar n)0}^a \, J_{n\bar n\,\lambda_1}^{{\bar \alpha}\beta}\,H $ \,(\ref{eq:soft_insert_basis2})
& 2 & 1 &
\\
& & $O_{\partial \chi (us(i))\lambda_1:(\lambda_2)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(i)\lambda_1} \, J_{n\bar n\,\lambda_2}^{{\bar \alpha}\beta}\}\,H $ \,(\ref{eq:soft_derivative_basis})
& 4 & 1 &
\\
& & $O_{\mathcal{B} (us(n))\lambda_1:\lambda_2 \lambda_3 }^{(2)abc}
=\mathcal{B}_{us(n) \lambda_1}^a \,\mathcal{B}_{n\, \lambda_2}^{b}\,\mathcal{B}_{\bn\, \lambda_3}^{c}\,H$ \,(\ref{eq:Hgggus})
& 2 & 2 & $\checkmark$
\\
& & $O_{\mathcal{B} (us(\bar n))\lambda_1:\lambda_2 \lambda_3 }^{(2)abc}
=\mathcal{B}_{us(\bar n) \lambda_1}^a \,\mathcal{B}_{n\, \lambda_2}^{b}\,\mathcal{B}_{\bn\, \lambda_3}^{c}\,H$ \,(\ref{eq:Hgggus_2})
& 2 & 2 & $\checkmark$
\\
& & $O_{\partial \mathcal{B} (us(i))\lambda_1:\lambda_2 \lambda_3 }^{(2)ab}
=\left[ \partial_{us(i) \lambda_1} \,\mathcal{B}_{n\, \lambda_2}\right] \,\mathcal{B}_{\bn\, \lambda_3}\,H $ \,(\ref{eq:Hdggus})
& 4 & 1 & $\checkmark$
\\ \hline
\end{tabular}}
\vspace{0.1cm}
\caption{Basis of hard scattering operators for $gg\to H$ up to ${\cal O}(\lambda^2)$. The $\lambda_i$ denote helicities, $S$ represents a symmetry factor present for some cases, and detailed lists of operators can be found in the indicated equation. The number of allowed helicity configurations are summarized in the fourth column. The final column indicates which operators contribute to the cross section up to $\mathcal{O}(\lambda^2)$ in the power expansion, as discussed in Sec. \ref{sec:discussion}. Counting the helicity configurations there are a total of 53 operators, of which only 28
contribute to the cross section at $\mathcal{O}(\lambda^2)$. Of those 28, only 24 have non zero Wilson coefficients at tree level since the operators in \eq{Hqqgpperp_basis_same} are absent at this order. These numbers do not include the number of distinct color configurations which are indicated in the 5th column.
}
\label{tab:summary}
\end{table}
}
\subsection{Leading Power}\label{sec:lp}
The leading power operators for $gg\to H$ in the Higgs effective theory are well known. Due to the fact that the Higgs is spin zero, the only two operators are
\begin{align}
\boldsymbol{g_n g_{\bn}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_gg_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:hgg}
&O_{\mathcal{B}++}^{(0)ab}
=\mathcal{B}_{n +}^a\, \mathcal{B}_{\bar n+}^b\, H
\,, \qquad &&O_{\mathcal{B}--}^{(0)ab}
=\mathcal{B}_{n-}^a\, \mathcal{B}_{ \bn -}^b\, H\,.
\end{alignat}
Here the purple circled denotes that this is a hard scattering operator in the effective theory, while the dashed circles indicate which fields are in each collinear sector.
Note that here we have opted not to include a symmetry factor at the level of the operator. We will include symmetry factors in the operator only when there is an exchange symmetry within a given collinear sector. We assume that overall symmetry factors which involve exchanging particles from different collinear sectors are taken into account at the phase space level. The color basis here is one-dimensional, and we take it to be
\begin{align} \label{eq:leading_color}
\bar{T}^{ab} = \delta_{ab}\,, \qquad
\bar{T}_{\mathrm{BPS}}^{ab} = \bigl( \mathcal{Y}_{n}^T \mathcal{Y}_{\bar n} \bigr)^{ab}
= \bigl( \mathcal{Y}_{\bn}^T \mathcal{Y}_{n} \bigr)^{ba}
\,.
\end{align}
\subsection{Subleading Power}\label{sec:nlp}
Due to the spin zero nature of the Higgs, the $\mathcal{O}(\lambda)$ operators are highly constrained. To simplify the operator basis we will work in the center of mass frame and we will further choose our $n$ and $\bar n$ axes so that the total label $\perp$ momentum of each collinear sector vanishes. This is possible in an SCET$_\text{I}$ theory since the ultrasoft sector does not carry label momentum, and it implies that we do not need to include operators where the $\mathcal{P}_\perp$ operator acts on a sector with a single collinear field. At $\mathcal{O}(\lambda)$ the suppression in the operator must therefore come from an explicit collinear field.
There are two possibilities for the collinear field content of the operators, either three collinear gluon fields, or two collinear quark fields and a collinear gluon field. Interestingly, the helicity selection rules immediately eliminate the possibility of $\mathcal{O}(\lambda)$ operators with three collinear gluon fields, since they cannot sum to a zero helicity state. We therefore only need to consider operators involving two collinear quark fields and a collinear gluon field. The helicity structure of these operators is also constrained. In particular, to cancel the spin of the collinear gluon field, the collinear quark current must have helicity $\pm1$. Furthermore, the quark-antiquark pair arises from a gluon splitting, since we are considering gluon fusion in the Higgs EFT, and therefore both have the same chirality. Together this implies that the quarks are described by the current $J_{n\bar n\,\pm}^{{\bar \alpha}\beta}$. The only two operators in the basis at $\mathcal{O}(\lambda)$ are
\begin{align}
\boldsymbol{q_n (\bar qg)_{\bn}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/SubleadingQG_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:H1_basis}
&O_{\mathcal{B}\bar n+(+)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta} \,H
\,,\qquad &
&O_{\mathcal{B}\bar n-(-)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{\bar n-}^a \, J_{n\bar n\,-}^{{\bar \alpha}\beta}\,H
\,,
\end{alignat}
for the case that the gluon field is in the same sector as the antiquark field, which we have taken to be $\bar n$, and
\begin{align}
\boldsymbol{(q g)_n \bar q_{\bn}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/SubleadingQG_flip_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:H1_basis2}
&O_{\mathcal{B} n-(+)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta} \,H
\,,\qquad &
&O_{\mathcal{B} n+(-)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a \, J_{n\bar n\,-}^{{\bar \alpha}\beta}\,H
\,,
\end{alignat}
for the case that the gluon field is in the same direction as the collinear quark field. In both cases the color basis is one-dimensional $\bar{T}^{a\, \alpha{\bar \beta}} = T^a_{\alpha{\bar \beta}}$. After the BPS field redefinition we have
\begin{align} \label{eq:nlp_color}
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}}
&= \left (Y_n^\dagger Y_{\bar n} T^a \right )_{\alpha \bar \beta}
\,,
& \bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}}
&= \left (T^a Y_n^\dagger Y_{\bar n} \right )_{\alpha \bar \beta}
\,,
\end{align}
for \eqs{H1_basis}{H1_basis2} respectively.
\subsection{Subsubleading Power}\label{sec:nnlp}
At $\mathcal{O}(\lambda^2)$ the allowed operators can include either additional collinear field insertions, insertions of the $\mathcal{P}_\perp$ operator, or ultrasoft field insertions. We will treat each of these cases in turn.
\subsubsection{Collinear Field Insertions}\label{sec:nnlp_collinear}
We begin by considering operators involving only collinear field insertions. At $\mathcal{O}(\lambda^2)$ the operator can have four collinear fields. These operators can be composed purely of collinear gluon fields, purely of collinear quark fields, or of two collinear gluon fields and a collinear quark current. In each of these cases helicity selection rules will restrict the possible helicity combinations of the operators.
\vspace{0.4cm}
\noindent{\bf{Two Quark-Two Gluon Operators:}}
We begin by considering operators involving two collinear quark fields and two collinear gluon fields, which are again severely constrained by the helicity selection rules. Since the two gluons fields can give either helicity $0$ or $2$, the only way to achieve a total spin zero is if the quark fields must be in a helicity zero configuration. Furthermore, since they arise from a gluon splitting they must have the same chirality. This implies that all operators must involve only the currents $J_{n\, 0}^{{\bar \alpha}\beta}$ or $J_{n\, \bar 0}^{{\bar \alpha}\beta}$, where we have taken without loss of generality that the two quarks are in the $n$-collinear sector, as per the discussion below \Eq{eq:sum_dir}. The gluons can then either be in opposite collinear sectors, or in the same collinear sector. The color basis before BPS field redefinition is identical for the two cases. It is three dimensional, and we take as a basis
\begin{equation} \label{eq:ggqqll_color}
\bar{T}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(T^a T^b)_{\alpha{\bar \beta}}\,,\, (T^b T^a)_{\alpha{\bar \beta}} \,,\, {\rm tr}[T^a T^b]\, \delta_{\alpha{\bar \beta}}
\Bigr)
\,.\end{equation}
In the case that the two collinear gluons are in opposite collinear sectors a basis of helicity operators is given by
\begin{align}
\boldsymbol{(g q\bar q)_n (g)_{\bn}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qqbarg_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Hqqgg_basis3}
&O_{\cB1++(0)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\,{0} }^{{\bar \alpha}\beta} \,H\, , \qquad
&&O_{\cB1++(\bar 0)}^{(2)ab\, {\bar \alpha}\beta}
=\mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta} \,H\, , \\
&O_{\cB1--(0)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\,{0} }^{{\bar \alpha}\beta} \,H\, , \qquad
&&O_{\cB1--(\bar 0)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{ n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta} \,H\, . {\nonumber}
\end{alignat}
The color basis after BPS field redefinition is given by
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(\mathcal{Y}_n^T \mathcal{Y}_\bn)^{cb} (T^a T^c)_{\alpha{\bar \beta}} \,,\,
(\mathcal{Y}_n^T \mathcal{Y}_\bn)^{cb} (T^c T^a)_{\alpha{\bar \beta}} \,,\,
T_F (\mathcal{Y}_n^T \mathcal{Y}_\bn)^{ab} \, \delta_{\alpha{\bar \beta}}
\Bigr)
\,,
\end{equation}
where we have used ${\rm tr}[T^a T^b] = T_F \delta^{ab}$.
In the case that the two gluons are in the same collinear sector a basis of helicity operators is given by
\begin{align}
& \boldsymbol{(q \bar q)_n (gg)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_gg_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Hqqgg_basis4}
&O_{\cB2+-(0)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a\, \mathcal{B}_{ \bar n-}^b \, J_{n\,{0} }^{{\bar \alpha}\beta}\, H\,, \qquad
&O_{\cB2+-(\bar 0)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a\, \mathcal{B}_{ \bar n-}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta}\, H\,.
\end{align}
The color basis after BPS field redefinition is
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(Y_n^\dagger Y_\bn T^a T^b Y^\dagger_{\bar n} Y_n )_{\alpha{\bar \beta}}
\,,\,
(Y_n^\dagger Y_\bn T^b T^a Y^\dagger_{\bar n} Y_n )_{\alpha{\bar \beta}}
\,,\,
{\rm tr}[T^a T^b]\, \delta_{\alpha{\bar \beta}} \Bigr)
\,.
\end{equation}
\vspace{0.4cm}
\noindent{\bf{Four Gluon Operators:}}
Operators involving four collinear gluon fields can have either two collinear gluon fields in each sector, or three collinear gluon fields in one sector. A basis of color structures before BPS field redefinition is given by
\begin{equation} \label{eq:gggg_color}
\bar{T}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
{\rm tr}[abcd] + {\rm tr}[adcb] \\ {\rm tr}[acdb] + {\rm tr}[abdc] \\ {\rm tr}[adbc] + {\rm tr}[acbd] \\
{\rm tr}[abcd] - {\rm tr}[adcb] \\ {\rm tr}[acdb] - {\rm tr}[abdc] \\ {\rm tr}[adbc] - {\rm tr}[acbd] \\ 2{\rm tr}[ab]\, {\rm tr}[cd] \\ 2{\rm tr}[ac]\, {\rm tr}[db] \\ 2{\rm tr}[ad]\, {\rm tr}[bc]
\end{pmatrix}^{\!\!\!T}
.\end{equation}
Here we have used a simplified notation, writing only the adjoint indices of the color matrices appearing in the trace. For example, ${\rm tr}[abcd] \equiv {\rm tr}[T^a T^b T^c T^d]$.
The color bases after BPS field redefinition will be given separately for each case.
For the specific case of SU($N_c$) with $N_c=3$ we could further reduce the color basis by using the relation
\begin{align}
&{\rm tr}[abcd+dcba] + {\rm tr}[acdb+bdca] + {\rm tr}[adbc+cbda]
{\nonumber}\\ & \qquad
= {\rm tr}[ab]{\rm tr}[cd] + {\rm tr}[ac]{\rm tr}[db] + {\rm tr}[ad]{\rm tr}[bc]
\,.\end{align}
We choose not to do this, as it makes the structure more complicated, and because it does not hold for $N_c>3$.
In the case that there are two collinear gluon fields in each collinear sector, a basis of helicity operators is given by
\begin{align}
& \boldsymbol{(gg)_n (gg)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4g_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:H_basis_gggg_1}
&O_{4g1++++}^{(2)a b c d}
= \,\frac{1}{4} \mathcal{B}^a_{n +} \mathcal{B}^b_{n +} \mathcal{B}^c_{\bn +} \mathcal{B}^d_{\bn +} \,H
\,,\qquad &
&O_{4g1+-+-}^{(2)a b c d}
= \, \mathcal{B}^a_{n +} \mathcal{B}^b_{n -} \mathcal{B}^c_{\bn +} \mathcal{B}^d_{\bn -} \,H
\,,\\
&O_{4g1----}^{(2)a b c d}
= \,\frac{1}{4} \mathcal{B}^a_{n -} \mathcal{B}^b_{n -} \mathcal{B}^c_{\bn -} \mathcal{B}^d_{\bn -} \,H
\,.\qquad &{\nonumber}
\end{alignat}
The spin zero nature of the Higgs implies that a number of helicity configurations do not contribute, and therefore are not included in our basis operators here.
The color basis after BPS field redefinition is given by
\begin{equation}
\bar{T}_{\text{BPS}}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] + {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] + {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] + {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] - {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] - {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] - {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
\frac{1}{2} \delta^{ab} \delta^{cd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{ac} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{bd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{ad} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{bc}
\end{pmatrix}^{\!\!\!T}
.\end{equation}
The other relevant case has three gluons in one sector, which we take to be the $\bn$ collinear sector. The basis of operators is then given by
\begin{align}
& \boldsymbol{(g)_n (ggg)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4ga_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:H_basis_gggg_2}
&O_{4g2+++-}^{(2)a b c d}
= \,\frac{1}{2} \mathcal{B}^a_{n +} \mathcal{B}^b_{\bn +} \mathcal{B}^c_{\bn +} \mathcal{B}^d_{\bn -} \,H
\,,\qquad &
&O_{4g2-+--}^{(2)a b c d}
= \,\frac{1}{2} \mathcal{B}^a_{n -} \mathcal{B}^b_{\bn +} \mathcal{B}^c_{\bn -} \mathcal{B}^d_{\bn -} \,H
\,.
\end{alignat}
In this case, the post-BPS color basis is given by
\begin{equation}
\bar{T}_{\text{BPS}}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] + {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] + {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] + {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] - {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] - {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] - {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{\bn} \mathcal{Y}^{c' c}_{\bn} \mathcal{Y}^{d' d}_{\bn}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{ab} \delta^{cd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{ac} \delta^{bd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_\bn)^{ad} \delta^{bc}
\end{pmatrix}^{\!\!\!T}
.\end{equation}
The helicity basis has made extremely simple the task of writing down a complete and minimal basis of four gluon operators, which would be much more difficult using traditional Lorentz structures. The helicity operators also make it simple to implement the constraints arising from the spin zero nature of the Higgs.
\vspace{0.4cm}
\noindent{\bf{Four Quark Operators:}}
We now consider the case of operators involving four collinear quark fields. These operators are again highly constrained by the helicity selection rules and chirality conservation, since each quark-antiquark pair was produced from a gluon splitting. In particular, these two constraints imply that there are no operators with non-vanishing Wilson coefficients with three quarks in one collinear sector. Therefore, we need only consider the cases where there are two quarks in each collinear sector.
When constructing the operator basis we must also treat separately the case of identical quark flavors $H q \bar q q\bar q$ and distinct quark flavors $H q \bar q Q\bar Q $. For the case of distinct quark flavors $H q \bar q Q\bar Q $ we will have a $q\leftrightarrow Q$ symmetry for the operators. Furthermore the two quarks of flavor $q$, and the two quarks of flavor $\bar Q$, are necessarily of the same chirality. In the case that both quarks of the same flavor appear in the same current, the current will be labeled by the flavor. Otherwise, the current will be labeled with ($q \bar{Q}$) or ($Q\bar{q}$) appropriately. For all these cases, the color basis is
\begin{equation} \label{eq:qqqq_color}
\bar{T}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\Bigl(
\delta_{\alpha{\bar \delta}}\, \delta_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\Bigr)
\,.\end{equation}
We will give results for the corresponding $\bar T_{\rm BPS}^{\, \alpha{\bar \beta}\gamma{\bar \delta}}$ basis as we consider each case below.
For the case of operators with distinct quark flavors $H q \bar q Q\bar Q $ and two collinear quarks in each of the $n$ and $\bn$ sectors there are three possibilities. There is either a quark anti-quark pair of the same flavor in each sector (e.g. $(q \bar q)_n(Q \bar Q)_\bn$), a quark and an antiquark of distinct flavors in the same sector (e.g. $(q \bar Q)_n(Q \bar q)_\bn$), or two quarks with distinct flavors in the same sector(e.g. $(q Q)_n(\bar q \bar Q)_\bn$). In the case that there is a quark anti-quark pair of the same flavor in each sector, the basis of helicity operators is
\vspace{0.3cm}\begin{align}
\boldsymbol{(q \bar q)_n(Q \bar Q)_\bn:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_low}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_qQ}
&O_{qQ1(0;0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n 0\, }^{{\bar \gamma}\delta} H
\,,\qquad &
&O_{qQ1(0;\bar 0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q) n 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n \bar 0\,}^{{\bar \gamma}\delta} H
\,,\\
&O_{qQ1(\bar 0;0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n 0\, }^{{\bar \gamma}\delta} H
\,,\qquad &
&O_{qQ1(\bar 0;\bar 0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n \bar 0\, }^{{\bar \gamma}\delta} H
\,,{\nonumber}
\end{alignat}
where we have chosen the $q$ quark to be in the $n$ sector. Since all the operators have total helicity $0$ along the $\hat n$ direction, there are only chirality constraints here and no constraints from angular momentum conservation. In the case that there is a quark anti-quark of distinct flavors in the same sector, chirality and angular momentum conservation constrains the basis to be
\begin{align}
\boldsymbol{(q \bar Q)_n(Q \bar q)_\bn:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_b_low}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align} \label{eq:Z2_basis_qQ_2}
&O_{qQ2(0;0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q \bar{Q}) n 0\, }^{{\bar \alpha}\beta}\, J_{(Q\bar{q} ) \bar{n} 0\, }^{{\bar \gamma}\delta} H
\,,\qquad
O_{qQ2(\bar 0;\bar 0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q \bar{Q}) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q\bar{q} )\bar{n} \bar 0\, }^{{\bar \gamma}\delta} H
\,.
\end{align}
For the operators in \eqs{Z2_basis_qQ}{Z2_basis_qQ_2} the color basis after BPS field redefinition is
\begin{align} \label{eq:TBPS_OqQ12}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \Big[ Y_{n}^\dagger Y_{\bar n} \Big]_{\alpha{\bar \delta}}\,\Big[ Y_{\bar n}^\dagger Y_{n} \Big]_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
When there are two quarks of distinct flavors in the same sector the basis of helicity operators is constrained by chirality and reduced further to just two operators by angular momentum conservation, giving
\begin{align}
\boldsymbol{(q Q)_n(\bar q \bar Q)_\bn:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_c_low}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Z2_basis_qQ_3}
&O_{qQ3(+;-)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar n +\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n -\, }^{{\bar \gamma}\delta} H
\,,\qquad &
O_{qQ3(-;+)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar n -\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n +\, }^{{\bar \gamma}\delta} H
\,.
\end{align}
For the operators in \eq{Z2_basis_qQ_3} the color basis after BPS field redefinition is
\begin{align} \label{eq:TBPS_OqQ3}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\alpha{\bar \delta}}\,\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \beta}}
\,,\,
\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\alpha{\bar \beta}}\,\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
In the cases considered in \eqs{Z2_basis_qQ}{Z2_basis_qQ_2} where there is a quark and antiquark field in the same collinear sector, we have chosen to work in a basis using $J_{i0}^{{\bar \alpha}\beta}$ and $J_{i\bar 0}^{{\bar \alpha}\beta}$ which contain only fields in a single collinear sector. One could also construct an alternate form for the basis, for example using the currents $J_{n\bn\lambda}^{{\bar \alpha}\beta}$. From the point of view of factorization our basis is the most convenient since the fields in the $n$ and $\bar n$-collinear sectors are only connected by color indices, which will simplify later steps of factorization proofs. In the following, we will whenever possible use this logic when deciding between equivalent choices for our basis.
For identical quark flavors the operators are similar to those in Eqs.~(\ref{eq:Z2_basis_qQ},\ref{eq:Z2_basis_qQ_3}). The distinct operators include
\begin{align}
\boldsymbol{(q \bar q)_n(q \bar q)_\bn:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4qq_low}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_qq}
&O_{qq1(0;0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \,\frac14\, J_{(q)n 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n 0\, }^{{\bar \gamma}\delta} H
\,,\qquad &
&\phantom{O_{qq1(0;\bar 0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q) n 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n \bar 0\,}^{{\bar \gamma}\delta} H
\,,}\\
&O_{qq1(\bar 0;0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n 0\, }^{{\bar \gamma}\delta} H
\,,\qquad &
&O_{qq1(\bar 0;\bar 0)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, \frac14\, J_{(q) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n \bar 0\, }^{{\bar \gamma}\delta} H
\,,{\nonumber}
\end{alignat}
\begin{align}
\boldsymbol{(q q)_n(\bar q \bar q)_\bn:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4qq_c_low}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Z2_basis_qq_3}
&O_{qq3(+;-)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar n +\, }^{{\bar \alpha}\beta}\, J_{(q) n\bar n -\, }^{{\bar \gamma}\delta} H
\,.
\end{align}
Note that in \Eq{eq:Z2_basis_qq} there are only three operators due to the equivalence between the two operators
\begin{align}
\sum_n J_{(q) n 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n \bar 0\,}^{{\bar \gamma}\delta} H
\equiv \, \sum_n J_{(q)n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n 0\, }^{{\bar \gamma}\delta} H
\,,
\end{align}
due to the fact that the $n$ label is summed over, as in \Eq{eq:sum_dir}.
We also have the same color bases as in \eqs{TBPS_OqQ12}{TBPS_OqQ3} for $O_{qq1}^{(2)}$ and $O_{qq3}^{(2)}$ respectively.
\subsubsection{$\mathcal{P}_\perp$ Insertions}\label{sec:nnlp_perp}
Since we have chosen to work in a frame where the total $\perp$ momentum of each collinear sector vanishes, operators involving explicit insertions of the $\mathcal{P}_\perp$ operator first appear at $\mathcal{O}(\lambda^2)$. The $\mathcal{P}_\perp$ operator can act only in a collinear sector composed of two or more fields. At $\mathcal{O}(\lambda^2)$, there are then only two possibilities, namely that the $\mathcal{P}_\perp$ operator is inserted into an operator involving two quark fields and a gluon field, or it is inserted into an operator involving three gluon fields.
In the case that the $\mathcal{P}_\perp$ operator is inserted into an operator involving two quark fields and a gluon field, the helicity structure of the operator is highly constrained. In particular, the quark fields must be in a helicity zero configuration. Combined with the fact that they must have the same chirality, this implies that all operators must involve only the currents $J_{\bar n\, 0}^{{\bar \alpha}\beta}$ or $J_{\bar n\, \bar 0}^{{\bar \alpha}\beta}$. Here we have again taken without loss of generality that the two quarks are in the $\bar n$-collinear sector. A basis of operators is then given by
\begin{align}
& \boldsymbol{(g)_n (q\bar q\, \mathcal{P}_\perp)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qqbarg_orderlam_perp_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:Hqqgpperp_basis_same}
&O_{\mathcal{P} \chi + (0)[+]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \big\{ \mathcal{P}_{\perp}^{+} J_{\bar n\, 0}^{{\bar \alpha}\beta} \big\}\, H
\,,\qquad &
&O_{\mathcal{P}\chi - (0)[-]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \big\{ \mathcal{P}_{\perp}^{-} J_{\bar n\, 0 }^{{\bar \alpha}\beta} \big\} \, H
\,,\\
&O_{\mathcal{P}\chi + (\bar 0)[+]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \big\{ \mathcal{P}_{\perp}^{+} J_{\bar n\, \bar0 }^{{\bar \alpha}\beta} \big\}\, H
\,,\qquad &
&O_{\mathcal{P}\chi - (\bar 0)[-]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a \, \big\{ \mathcal{P}_{\perp}^{-} J_{\bar n\, \bar0 }^{{\bar \alpha}\beta} \big\} \, H
\,.{\nonumber}
\end{alignat}
Since we have assumed that the total $\mathcal{P}_\perp$ in each collinear sector is zero, integration by parts can be used to make the $\mathcal{P}_\perp$ operator act only on either the quark, or the antiquark field, which has been used in \Eq{eq:Hqqgpperp_basis_same}. (The additional operators that are needed when we relax this assumption are discussed in \app{gen_pt}.)
The color basis is one-dimensional
\begin{equation} \label{eq:nnlp_color_quark_perp}
\bar{T}^{a\, \alpha{\bar \beta}} = T^a_{\alpha{\bar \beta}}\,.
\end{equation}
After BPS field redefinition the structure is given by
\begin{align} \label{eq:nnlp_color_quark_perpBPS}
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}}
=\left ( Y^\dagger_{\bar n} T^b \mathcal{Y}_n^{ba} Y_{\bar n} \right )_{\alpha \bar \beta}
= \bigl( \mathcal{Y}_n^T \mathcal{Y}_\bn \bigr)^{ac}\: T^c_{\alpha \bar \beta}
\,.
\end{align}
In the case that the $\mathcal{P}_\perp$ operator is inserted into an operator involving three gluon fields, the helicity selection rules simply imply that the helicities must add to zero.
A basis of operators involving three collinear gluon fields and a $\mathcal{P}_\perp^\pm$ insertion is given by
\begin{align}
& \boldsymbol{(g)_n (gg\, \mathcal{P}_\perp)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_3g_perp_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Hgggpperp_basis}
&O_{\mathcal{P}\mathcal{B} +++[-]}^{(2)abc}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n+}^c \right ] \,H
\,,\qquad && O_{\mathcal{P}\mathcal{B} ---[+]}^{(2)abc}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n-}^c \right ] \,H\,,
{\nonumber} \\
%
&O_{\mathcal{P}\mathcal{B} ++-[+]}^{(2)abc}
= \, \mathcal{B}_{n+}^a\, \mathcal{B}_{ \bar n+}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n-}^c \right ] \,H
\,,\qquad
&&O_{\mathcal{P}\mathcal{B} --+[-]}^{(2)abc}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{ \bn -}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n +}^c \right ] \,H
\,.
\end{alignat}
Note that the analogous operators with the helicities $O_{\mathcal{P}\mathcal{B} +-+[+]}^{(2)abc}$ and $O_{\mathcal{P}\mathcal{B} -+-[-]}^{(2)abc}$ are not eliminated, but instead are equivalent to those in the last row by integrating the $\mathcal{P}_\perp^\pm$ by parts onto the other $\bn$-collinear field since the total $\mathcal{P}_\perp$ in each collinear sector is zero. (The additional operators that are needed when we relax this assumption are discussed in \app{gen_pt}.)
The basis of color structures here is two dimensional,
\begin{equation} \label{eq:ggg_perp_color}
\bar{T}^{abc} =
\begin{pmatrix}
i f^{abc} \\ d^{abc}
\end{pmatrix}
\,,
\qquad
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{a'b'c'}\, {\cal Y}_n^{a'a} {\cal Y}_\bn^{b'b} {\cal Y}_{\bn}^{c'c} \\
d^{a'b'c'}\, {\cal Y}_n^{a'a} {\cal Y}_\bn^{b'b} {\cal Y}_{\bn}^{c'c}
\end{pmatrix}
=
\begin{pmatrix}
i f^{bcd} {\cal Y}_{\bn}^{a'd} {\cal Y}_n^{a'a} \\
d^{bcd}\, {\cal Y}_\bn^{a'd} {\cal Y}_n^{a'a}
\end{pmatrix}
\,.
\end{equation}
In the BPS redefined color structure we have written it both in a form that makes the structure of the Wilson lines appearing from the field redefinition clear, as well as in a simplified form.
\subsubsection{Ultrasoft Insertions}\label{sec:nnlp_soft}
At $\mathcal{O}(\lambda^2)$ we have the possibility of operators with explicit ultrasoft insertions. To have label momentum conservation these operators must have a collinear field in each collinear sector. Interestingly, despite the fact that the leading power operator has two collinear gluon fields, for the operators involving an ultrasoft insertion one can have either two collinear quark fields, or two collinear gluon fields.
The construction of an operator basis involving ultrasoft gluons is more complicated due to the fact that they are not naturally associated with a given lightcone direction. There are therefore different choices that can be made when constructing the basis. We will choose to work in a basis where all ultrasoft derivatives acting on ultrasoft Wilson lines are absorbed into $\mathcal{B}_{us}$ fields. To understand why it is always possible to make this choice, we consider two pre-BPS operators involving two collinear quark fields, and an ultrasoft derivative
\begin{align}
O^\mu_1=\bar \chi_{\bar n} (i D_{us}^\mu) \chi_n\,, \qquad O^\mu_2=\bar \chi_{\bar n} (-i \overleftarrow D_{us}^\mu) \chi_n\,,
\end{align}
where $(-i \overleftarrow D_{us}^\mu)=(i D_{us}^\mu)^\dagger$ and we have not made the contraction of the $\mu$ index explicit, as it is irrelevant to the current discussion. Performing the BPS field redefinition, we obtain
\begin{align}
O^\mu_{1\text{BPS}}= i \bar \chi_{\bar n}Y_{\bar n}^\dagger D_{us}^\mu Y_n \chi_n\,, \qquad
O^\mu_{2\text{BPS}}=-i \bar \chi_{\bar n} Y_{\bar n}^\dagger \overleftarrow D_{us}^\mu Y_n \chi_n\,.
\end{align}
If we want to absorb all derivatives acting on Wilson lines into $\mathcal{B}_{us}$ fields, we must organize the Wilson lines in the operators as
\begin{align}
O^\mu_{1\text{BPS}}= i\bar \chi_{\bar n}Y_{\bar n}^\dagger Y_n (Y_n^\dagger D_{us}^\mu Y_n) \chi_n\,, \qquad
O^\mu_{2\text{BPS}}=-i \bar \chi_{\bar n} (Y_{\bar n}^\dagger \overleftarrow D_{us}^\mu Y_{\bar n}) Y_{\bar n}^\dagger Y_n \chi_n\,.
\end{align}
Using \Eq{eq:soft_gluon} we see that this can be written entirely in terms of $\partial_{us}$ operators acting on collinear fields, and the two ultrasoft gauge invariant gluon fields $\mathcal{B}_{us(n)}$ and $\mathcal{B}_{us(\bar n)}$ for $O^\mu_{1\text{BPS}}$ and $O^\mu_{2\text{BPS}}$ respectively. Note, however, that ultrasoft gluon fields defined with respect to both lightcone directions are required. Alternatively, it is possible to work only with $\mathcal{B}_{us(n)}$, for example, but in this case we see that the ultrasoft derivative must also be allowed to act explicitly on pairs of ultrasoft Wilson lines, for example $[\partial_{us}^\mu (Y_n^\dagger Y_\bn)]$. In constructing our complete basis we will choose to avoid this so that ultrasoft derivatives acting on soft Wilson lines occur only within the explicit $\mathcal{B}_{us}$ fields. This choice also makes our basis more symmetric.
For the operators involving one ultrasoft gluon and two collinear quarks we have the basis
\begin{align}
& \boldsymbol{g_{us}(q)_n (\bar q)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_qq_soft_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:soft_insert_basis}
&O_{\chi(us(n))-:(+)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(n)-}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, H
\,,\qquad &
&O_{\chi(us(n))+:(-)}^{(2)a\,{\bar \alpha}\beta}
=\mathcal{B}_{us(n)+}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, H
\,,
\end{alignat}
with the unique color structure
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\,a\, \alpha{\bar \beta}} = \left ( T^a Y^\dagger_{n} Y_{\bn} \right )_{\alpha \bar\beta}
\,,\end{equation}
and
\begin{alignat}{2} \label{eq:soft_insert_basis2}
&O_{\chi(us(\bar n))+:(+)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(\bar n)+}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, H
\,,\qquad &
&O_{\chi(us)(\bar n))-:(-)}^{(2)a\,{\bar \alpha}\beta}
=\mathcal{B}_{us(\bar n)-}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, H
\,,
\end{alignat}
with the unique color structure
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\,a\, \alpha{\bar \beta}}=(Y_n^\dagger Y_\bn T^a)_{\alpha{\bar \beta}}
\,.\end{equation}
Note that the color structures associated with the two different projections of the $\mathcal{B}_{us}$ field are distinct. All other helicity combinations vanish due to helicity selection rules. The helicity selection rules differ between \Eq{eq:soft_insert_basis} and \Eq{eq:soft_insert_basis2} due to the different choice of reference vector for the ultrasoft field in the two cases.
We also have operators involving two collinear quark fields and a single ultrasoft derivative,
\begin{align}
& \boldsymbol{\partial_{us} (q)_n (\bar q)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_qq_deriv_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:soft_derivative_basis}
&O_{\partial \chi (us(n))-:(+)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)-}\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\}\, H
\,,\qquad &
&O_{\partial \chi (us(n))+:(-)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)+} \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\}\, H\,,
\end{alignat}
with the unique color structure given before and after BPS field redefinition by
\begin{align} \label{eq:leading_color_deriv}
\bar{T}^{\alpha{\bar \beta}} = (\delta_{\alpha{\bar \beta}})\,, \qquad
\bar{T}_{\mathrm{BPS}}^{\alpha{\bar \beta}} = \big[Y_n^\dagger Y_{\bar n} \big]_{\alpha{\bar \beta}}
\,,
\end{align}
and
\begin{alignat}{2} \label{eq:soft_derivative_basis2}
&O_{\partial^\dagger \chi (us(\bar n))+:(+)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\,+}^{{\bar \alpha}\beta}\, (i\partial_{us(\bar n)+})^\dagger\}\, H
\,,~ &
&O_{\partial^\dagger \chi (us(\bar n))-:(-)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\, -}^{{\bar \alpha}\beta}\, (i\partial_{us(\bar n)-})^\dagger\}\, H\,,
\end{alignat}
with the unique color structure given before and after BPS field redefinition by
\begin{align} \label{eq:leading_color_deriv2}
\bar{T}^{\alpha{\bar \beta}} = (\delta_{\alpha{\bar \beta}})\,, \qquad
\bar{T}_{\mathrm{BPS}}^{\alpha{\bar \beta}} = \big[Y_n^\dagger Y_{\bar n} \big]_{\alpha{\bar \beta}}
\,.
\end{align}
Although the color structure happens to be the same in both cases, we have separated them to highlight the different decompositions of the ultrasoft derivatives in the two cases. Note that the form of the ultrasoft derivatives which appear is constrained by the helicity constraints.
Similarly, we have the corresponding operators involving two collinear gluons. A basis of helicity operators involving two collinear gluons and a single ultrasoft gluon field is given by
\begin{align}
& \boldsymbol{g_{us}(g)_n (g)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_gg_soft_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Hgggus}
&O_{\mathcal{B} (us(n))0:++}^{(2)abc}
= \mathcal{B}_{us(n)0}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, H
\,, \qquad && O_{\mathcal{B} (us(n))0:--}^{(2)abc}
= \mathcal{B}_{ us(n)0}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, H
\,,
\end{align}
with the basis of color structures,\footnote{In order to see how the Wilson line structure in \Eq{eq:Z2g_colorus} arises, we look at the object $D_{us}^{ab} \mathcal{B}_{n}^c \mathcal{B}_{\bn}^d$ pre-BPS field redefinitions. This object must be contracted with a tensor to make it a singlet under ultrasoft gauge transformations. Each of these resulting forms can be mapped onto the color structures of \Eq{eq:Z2g_colorus} after performing the BPS field redefinition}
\begin{equation} \label{eq:Z2g_colorus}
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc} \\
d^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc}
\end{pmatrix}^T
\,,
\end{equation}
and
\begin{align}\label{eq:Hgggus_2}
&O_{\mathcal{B} (us(\bar n))0:++}^{(2)abc}
= \mathcal{B}_{us(\bar n)0}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, H
\,, \qquad && O_{\mathcal{B} (us(\bar n))0:--}^{(2)abc}
= \mathcal{B}_{ us(\bar n)0}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, H
\,,
\end{align}
with the basis of color structures,
\begin{equation} \label{eq:Z2g_colorus_2}
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{abd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{dc} \\
d^{abd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{dc}
\end{pmatrix}^T
\,.
\end{equation}
We have only included the $\bar{T}_{\mathrm{BPS}}^{abc}$ version of the color structure here because the $\mathcal{B}_{us(n)\lambda}^a$ are generated by BPS field redefinition.
The Wilson coefficients of the operators that include $\mathcal{B}_{us(n)0}$ can be related to the Wilson coefficients of the leading power operators using RPI symmetry (see \cite{Larkoski:2014bxa}). In particular, we have
\begin{align} \label{eq:usRPIrelation}
C^{(2)}_{\mathcal{B} n(us)0:\lambda_1, \lambda_1}&=-\frac{\partial C^{(0)}_{\lambda_1, \lambda_1} }{\partial \omega_1}
\,,
\end{align}
where $C^{(0)}_{\lambda_1, \lambda_1}$ is the Wilson coefficient for the leading power operator of \Eq{eq:hgg}. We will explicitly verify this at the level of tree level matching in \Sec{sec:matching}.
We must also consider operators with an insertion of $\partial_{us(n)}$ with two collinear gluons in different collinear sectors. The gluon equations of motion allow us to eliminate the operators $in\cdot \partial \mathcal{B}_{n\perp}$ and $i\bar n\cdot \partial \mathcal{B}_{\bar n\perp}$, which can be rewritten purely in terms of collinear objects~\cite{Marcantonini:2008qn}. Furthermore, we again choose to organize our basis of operators such that ultrasoft derivatives act on ultrasoft Wilson lines only within the $\mathcal{B}_{us}$ fields, as was done in the quark case. (We also do not include operators where the ultrasoft derivative acts on the Higgs field, since this is moved to the other fields by integration by parts.) The basis of operators involving ultrasoft derivatives is then given by
\begin{align}
& \boldsymbol{\partial_{us}(g)_n (g)_{\bn}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_gg_deriv_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:Hdggus}
&O_{\partial \mathcal{B} (us(n))\bar 0:++}^{(2)ab}
= \mathcal{B}_{n+}^a \, \left[ \partial_{us(n)\bar 0} \mathcal{B}_{\bar n+}^b\right] \,H
\,, \,~~ && O_{\partial \mathcal{B} (us(n))\bar 0:--}^{(2)ab}
= \mathcal{B}_{n-}^a \left[\partial_{us(n)\bar 0} \mathcal{B}_{\bar n-}^b\right]\, H
\,,
\end{alignat}
with the basis of color structures
\begin{equation} \label{eq:Z2gd_colorus}
\bar{T}_{\mathrm{BPS}}^{ab} =
\big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{ab}
\,.
\end{equation}
and
\begin{alignat}{2}\label{eq:Hdggus2}
&O_{\partial \mathcal{B} (us(\bar n))0:++}^{(2)ab}
= \left[ \partial_{us(\bar n)0}\, \mathcal{B}_{n+}^a\right]\, \mathcal{B}_{\bar n+}^b\, H
\,, \,~~ && O_{\partial \mathcal{B} (us(\bar n))0:--}^{(2)ab}
= \left [ \partial_{us(\bar n)0}\, \mathcal{B}_{n-}^a \right] \, \mathcal{B}_{\bar n-}^b\, H
\,,
\end{alignat}
with the basis of color structures
\begin{equation} \label{eq:Z2gd_colorus2}
\bar{T}_{\mathrm{BPS}}^{ab} =
\big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{ab}
\,.
\end{equation}
The Wilson coefficients of the operators that include $\partial_{us(n)0}$ can also be related to the Wilson coefficients of the leading power operators using RPI symmetry (see \cite{Larkoski:2014bxa}). In particular, we have
\begin{align} \label{eq:usRPIrelationb}
C^{(2)}_{\partial \mathcal{B} (us(\bar n)) 0:\lambda_1 \lambda_1}&=-\frac{\partial C^{(0)}_{\lambda_1,\lambda_1} }{\partial \omega_1}
\,,
\end{align}
where $C^{(0)}_{\lambda_1,\lambda_1}$ is the Wilson coefficient for the leading power operator of \Eq{eq:hgg}. We will explicitly show how this arises in the tree level matching in \Sec{sec:matching}.
\subsection{Cross Section Contributions and Factorization}\label{sec:discussion}
{
\renewcommand{\arraystretch}{1.6}
\begin{center}
\begin{table}[t!]
\begin{center}
\begin{tabular}{| c | l | c | c | c | c | }
\hline
& Operators & Factorization & Beam $n$ & Beam $\bar n$ & Soft \\
\hline
$\!\!\mathbf{\mathcal{O}(\lambda^0)}\!\!$
&$O_{\mathcal{B}}^{(0)} O_{\mathcal{B}}^{(0)} $
& $H_g^{(0)} B_{g}^{(0)} B_{g}^{(0)} S_g^{(0)}$
& $\mathcal{B}_n \,\hat\delta\, \mathcal{B}_n$
& $\mathcal{B}_\bn \,\hat\delta\, \mathcal{B}_\bn$
& $\mathcal{Y}_n^T \mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn^T \mathcal{Y}_n$
\\
\hline
$\!\!\mathbf{\mathcal{O}(\lambda^2)}\!\!$
&$O_{\mathcal{B}\bar n}^{(1)} O_{\mathcal{B}\bar n}^{(1)}$
& $H_{g1}^{(0)} B_q^{(0)} B_{qgg}^{(2)} S_{q}^{(0)}$
& $\bar \chi_n \,\hat\delta\,\chi_n $
& $\bar \chi_\bn \mathcal{B}_\bn \hat\delta\, \mathcal{B}_\bn\chi_\bn $
& $Y_\bn^\dagger Y_n \widehat\mathcal{M}^{(0)}\, Y_n^\dagger Y_\bn$
\\
\cline{2-6}
&$O^{(0)} O_{\cB1}^{(2)} $
& $H_{g2}^{(0)} B_{gqq}^{(2)} B_{g}^{(0)} S_{g}^{(0)}$
& $\!\!\bar \chi_n \mathcal{B}_n \chi_n \hat\delta\, \mathcal{B}_n\!\! $
& $ \mathcal{B}_\bn \,\hat\delta\, \mathcal{B}_\bn $
& $\mathcal{Y}_n^T \mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn^T \mathcal{Y}_n$
\\
\cline{2-6}
&$O^{(0)} O_{\mathcal{P}\chi}^{(2)} $
& $H_{g3}^{(0)} B_g^{(0)} B_{gq P}^{(2)} S_{g}^{(0)}$
& $\mathcal{B}_n \,\hat\delta\,\mathcal{B}_n $
& $\!\!\bar \chi_\bn [\mathcal{P}_\perp \chi_\bn] \hat\delta\, \mathcal{B}_\bn\!\! $
& $\mathcal{Y}_n^T \mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn^T \mathcal{Y}_n$
\\
\cline{2-6}
&$O^{(0)} O_{\mathcal{P} \mathcal{B}}^{(2)} $
& $H_{g4}^{(0)} B_{g}^{(0)} B_{gg P}^{(2)} S_{g}^{(0)}$
& $\bar \mathcal{B}_n \,\hat\delta\, \mathcal{B}_n $
& $\!\!\mathcal{B}_\bn [\mathcal{P}_\perp \mathcal{B}_\bn] \hat\delta\, \mathcal{B}_\bn \!\! $
& $\mathcal{Y}_n^T \mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn^T \mathcal{Y}_n$
\\
\cline{2-6}
&$O^{(0)} O_{4g2}^{(2)}$
& $H_{g5}^{(0)} B_g^{(0)} B_{gg}^{(2)} S_{g}^{(0)}$
& $\mathcal{B}_n \,\hat\delta\,\mathcal{B}_n $
& $\!\! \mathcal{B}_\bn \mathcal{B}_\bn \mathcal{B}_\bn \hat\delta\, \mathcal{B}_\bn \!\! $
& $\mathcal{Y}_n^T \mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn^T \mathcal{Y}_n$
\\
\cline{2-6}
&$\!\! O^{(0)} O_{\mathcal{B} (us)0}^{(2)} \!\!\!$
& $H_{g6}^{(0)} B_g^{(0)} B_{g}^{(0)} S_{g\mathcal{B}}^{(2)}$
& $\mathcal{B}_n \,\hat\delta\,\mathcal{B}_n $
&$\mathcal{B}_\bn \,\hat\delta\,\mathcal{B}_\bn $
& $\!\mathcal{B}_{us(n) 0}\, \mathcal{Y}_n\mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn \mathcal{Y}_n\!\!$
\\
\cline{2-6}
& $\!\!O^{(0)} O_{\partial(us)0}^{(2)} \!\!\!$
& $H_{g7}^{(0)} B_g^{(0)} B_{g}^{(0)} S_{g\partial 0}^{(2)}$
& $\mathcal{B}_n \,\hat\delta\,\mathcal{B}_n $
&$\mathcal{B}_\bn \,\hat\delta\,\mathcal{B}_\bn $
& $\!\partial_{us(n) 0}\, \mathcal{Y}_n\mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn \mathcal{Y}_n\!\!$
\\
\cline{2-6}
& $\!\!O^{(0)} O_{\partial(us)\bar 0}^{(2)} \!\!\!$
& $H_{g8}^{(0)} B_g^{(0)} B_{g}^{(0)} S_{g\partial \bar 0}^{(2)}$
& $\mathcal{B}_n \,\hat\delta\,\mathcal{B}_n $
&$\mathcal{B}_\bn \,\hat\delta\,\mathcal{B}_\bn $
& $\! \partial_{us(n) \bar 0}\, \mathcal{Y}_n\mathcal{Y}_\bn \widehat\mathcal{M}^{(0)}\, \mathcal{Y}_\bn \mathcal{Y}_n\!\!$
\\
\hline
\end{tabular}
%
\caption{Subleading beam and soft functions arising from products of hard scattering operators in the factorization of Higgs with a jet veto, and their field content. Helicity and color structures have been suppressed. We have not included products of operators whose beam and soft functions are identical to those shown by charge conjugation or $n\leftrightarrow \bn$.} \label{tab:fact_func}
\end{center}
\end{table}
\end{center}
}
While the basis of operators presented in this section is quite large, many of the operators will not contribute to a physical cross section at $\mathcal{O}(\lambda^2)$. In this section we briefly discuss the helicity operator basis, focusing in particular on understanding which operators can contribute to the cross section for an SCET$_\text{I}$ event shape observable, $\tau_B$, measured on $gg\to H$. In \Sec{sec:contribs_lam}, we show that there are no contributions to the cross section from hard scattering operators at $\mathcal{O}(\lambda)$, which would correspond to power corrections of $\sqrt{\tau_B}$. Then in \Sec{sec:contribs}, we use helicity selection rules to determine which operators can contribute at $\mathcal{O}(\lambda^2)=\mathcal{O}(\tau_B)$. The results are summarized in \Tab{tab:summary}.
Given the set of contributing operators, one can then determine the full subleading power factorization theorem for the related observables with Higgs production. Here we restrict ourselves to determining the structure of the factorization theorem terms arising purely from our subleading hard scattering operators, written in terms of hard, beam and soft functions. A summary of these results is given in \Tab{tab:fact_func}. In many cases the beam and soft functions which appear in the subleading power factorization formula are identical to those at leading power. For the case of the soft functions this simplification arises due to color coherence, allowing a simplification to the Wilson lines in the soft functions that appear. For gluon-gluon and quark-quark color channels the leading power soft functions are
\begin{align}\label{eq:soft_func_def}
S_g^{(0)}=\frac{1}{(N_c^2-1)} {\rm tr} \big\langle 0 \big| \mathcal{Y}^T_{\bar n} \mathcal{Y}_n \widehat{\mathcal{M}}^{(0)}\mathcal{Y}_n^T \mathcal{Y}_{\bar n} \big|0 \big\rangle\,, \qquad
S_q^{(0)}=\frac{1}{N_c} {\rm tr} \big\langle 0 \big| Y^\dagger_{\bar n} Y_n \widehat{\mathcal{M}}^{(0)}Y_n^\dagger Y_{\bar n} \big|0 \big\rangle\,,
\end{align}
and depend on the kinematic variables probed by the measurement operator $\widehat{\mathcal{M}}^{(0)}$.
For the beam functions, this simplification occurs since the power correction is often restricted to a single collinear sector. The other collinear sector is then described by the leading power beam functions (incoming jet functions) for gluons and quarks~\cite{Stewart:2009yx,Stewart:2010qs}
\begin{align}\label{eq:beam_func_def}
\frac{\delta^{ab}}{N_c^2-1}\, B_g^{(0)}
&=- \frac{\omega \, \theta(\omega)}{2\pi} \int \!\!\frac{dx^-}{2|\omega|}\: e^{\frac{i}{2} \ell^+ x^-} \Big\langle p \Big|\, \mathcal{B}^{\mu a}_{n\perp} \big(x^- \text{\small $\frac{n}{2}$}\big)\, \hat{\delta}\, \left[ \delta(\omega-\bar \mathcal{P}) \mathcal{B}^{b}_{n\perp \mu}(0) \right] \,\Big|p\Big\rangle
\,, \\
\frac{\delta^{\alpha {\bar \beta}}}{N_c}\, B_q^{(0)}
&= \frac{\theta(\omega)}{2\pi} \int \!\!\frac{dx^-}{2|\omega|}\: e^{\frac{i}{2} \ell^+ x^-}
\Big\langle p\Big|\, \chi_{n}^{\alpha} \big(x^- \text{\small $\frac{n}{2}$}\big) \frac{\Sl{\bar n}}{2} \,\hat{\delta}\, \left[ \delta(\omega-\bar \mathcal{P}) \bar \chi_{n}^{\bar\beta}(0) \right]
\,\Big| p \Big\rangle
\,, {\nonumber}
\end{align}
where we take $\ell^+ \gg \Lambda_{\rm QCD}^2/\omega$.
The result for the leading power measurement function $\hat{\delta}$ appearing in these beam functions depends on the factorization theorem being treated. Often the beam functions are inclusive in which case $\hat{\delta}=1$, giving functions of the momentum fraction of the struck parton $x$ and a single invariant mass momentum variable, $B_{g}^{(0)}(x,\omega \ell^+)$ and $B_{q}^{(0)}(x,\omega \ell^+)$. Here we assume an SCET$_{\text{I}}$ type measurement that does not fix the $\mathcal{P}_\perp$ of the measured particle. This assumption has been explicitly used in writing the form of the beam functions in \Eq{eq:beam_func_def}, as well as in our construction of the operator basis.
\subsubsection{Vanishing at $\mathcal{O}(\lambda)$}\label{sec:contribs_lam}
We begin by considering possible contributions to the cross section at $\mathcal{O}(\lambda)$.
While we will not discuss the factorization of the cross section in detail, the contribution of the hard scattering operators to the cross section at $\mathcal{O}(\lambda)$ can be written schematically as
{\begin{small}
\begin{align}\label{eq:xsec_lam}
&\frac{d\sigma^{(1)} }{d\tau_B} \supset N \sum_{X,i} \tilde \delta^{(4)}_q \bra{P_1 P_2} C_i^{(1)} O_i^{(1)}(0) \ket{X}\bra{X} C^{(0)} O^{(0)}(0) \ket{P_1 P_2} \delta\big( \tau_B - \tau_B^{(0)}(X) \big) +\text{h.c.}
\,.
\end{align}
\end{small}}
Here $N$ is a normalization factor, $P_1, P_2$ denote the incoming hadronic states, and we use the shorthand notation $\tilde \delta^{(4)}_q=(2\pi)^4\delta^4(q-p_X)$ for the momentum conserving delta function. This expression should merely be taken as illustrative of the operator contributions, and in particular, we have not made explicit any color or Lorentz index contractions, nor the treatment of the initial state. The summation over all final states, $X$, includes phase space integrations. The measurement of the observable is enforced by $ \delta\big( \tau_B - \tau_B^{(0)}(X) \big) $, where $\tau_B^{(0)}(X)$, returns the value of the observable $\tau_B$ as measured on the final state $X$. The explicit superscript $(0)$ indicates that the measurement function is expanded to leading power, since here we focus on the power suppression due to the hard scattering operators.
From \Eq{eq:xsec_lam} we see that hard scattering operators contribute to the $\mathcal{O}(\lambda)$ cross section through their interference with the leading power operator. The $\mathcal{O}(\lambda)$ basis of operators is given in \Eqs{eq:H1_basis}{eq:H1_basis2}, each of which involves a single collinear quark field in each collinear sector. Conservation of fermion number then immediately implies that these operators cannot have non-vanishing matrix elements with the leading power operator which consists of a single collinear gluon field in each sector. Therefore, all contributions from hard scattering operators vanish at $\mathcal{O}(\lambda)$. Although we do not consider them in this paper, using similar arguments one can show that all other sources of power corrections, such as Lagrangian insertions, also vanish at $\mathcal{O}(\lambda)$.
\subsubsection{Relevant Operators at $\mathcal{O}(\lambda^2)$}\label{sec:contribs}
Unlike the $\mathcal{O}(\lambda)$ power corrections, the power corrections at $\mathcal{O}(\lambda^2)=\mathcal{O}(\tau_B)$ will not vanish. Contributions to the cross section at $\mathcal{O}(\lambda^2)$ whose power suppression arises solely from hard scattering operators take the form either of a product of two $\mathcal{O}(\lambda)$ operators or as a product of an $\mathcal{O}(\lambda^2)$ operator and an $\mathcal{O}(\lambda^0)$ operator
{\begin{small}
\begin{align}\label{eq:xsec_lam2}
\frac{d\sigma^{(2)}}{d\tau_B} &\supset N \sum_{X,i} \tilde \delta^{(4)}_q \bra{P_1 P_2} C_i^{(2)} O_i^{(2)}(0) \ket{X}\bra{X} C^{(0)} O^{(0)}(0) \ket{P_1 P_2} \delta\big( \tau_B - \tau_B^{(0)}(X) \big) +\text{h.c.}{\nonumber} \\
&+ N \sum_{X,i,j} \tilde \delta^{(4)}_q \bra{P_1 P_2} C_i^{(1)} O_i^{(1)}(0) \ket{X}\bra{X} C_j^{(1)} O_j^{(1)}(0) \ket{P_1 P_2} \delta\big( \tau_B - \tau_B^{(0)}(X) \big)+\text{h.c.} \,.
\end{align}
\end{small}}
For $gg\to H$ the operator basis has only a single operator at $\mathcal{O}(\lambda)$ (up to helicities and $n\leftrightarrow \bar n$), which was given in \Eqs{eq:H1_basis}{eq:H1_basis2}. This operator will contribute to the cross section at $\mathcal{O}(\lambda^2)$, as indicated in \Tab{tab:fact_func}.
The contributions from $\mathcal{O}(\lambda^2)$ hard scattering operators are highly constrained since they must interfere with the leading power operator. We will discuss each possible contribution in turn, and the summary of all operators which can contribute to the $\mathcal{O}(\lambda^2)$ cross section is given in \Tab{tab:summary}. The schematic structure of the beam and soft functions arising from each of the different operator contributions is shown in \Tab{tab:fact_func}. The subleading beam and soft functions enumerated in this table are universal objects that will appear in processes initiated by different Born level amplitudes (such as $q\bar q$ annihilation), unless forbidden by symmetry. In this initial investigation, we content ourselves with only giving the field content of the beam and soft functions. In \Tab{tab:fact_func}, to save space, we do not write the external vacuum states for the soft functions, or the external proton states for the beam functions, nor do we specify the space-time positions of the fields. We do not present here the full definitions analogous to the leading power definitions given in \Eqs{eq:soft_func_def}{eq:beam_func_def}, but using the field content given in \Tab{tab:fact_func}. Deriving full definitions goes hand in hand with presenting the complete factorization theorems for these contributions, which will be given in future work.
\vspace{0.6cm}
\noindent{\bf{Two Quark-One Gluon Operators:}}
The two quark-one gluon operators, $O_{\mathcal{B}\bar n}^{(1)}$ can contribute to the cross section by interfering with themselves. These operators are interesting since they effectively have a quark like cusp, instead of a gluon like cusp as is true of the leading power operators. They contribute with a leading power quark channel soft function $S_q^{(0)}$, a quark beam function $B_q^{(0)}$ and a subleading power beam function $B_{qgg}^{(2)}$ that has fermion number crossing the cut (as indicated by its first $q$ subscript).
\vspace{0.4cm}
\noindent{\bf{Two Quark-Two Gluon Operators:}}\\
\indent In the case of the two quark-two gluon operators, the only operators that will have a non-vanishing contribution are those that have the two gluons in different collinear sectors, namely $O_{\cB1}^{(2)}$. This gives a gluon beam function $B_g^{(0)}$, soft function $S_g^{(0)}$, and a subleading power beam function with gluon quantum numbers crossing the cut $B_{gqq}^{(2)}$ (with three color contractions). The operator $O_{\cB2}^{(2)}$, which has two quarks in a helicity $0$ configuration in one collinear sector, and two gluons in a helicity $0$ configuration in the other collinear sector does not contribute, since rotational invariance implies that its interference with the leading power operator vanishes.
\vspace{0.4cm}
\noindent{\bf{Four Gluon Operators:}}
To give a non-vanishing interference with the leading power operator the four gluon operators must have an odd number of collinear gluon fields in each sector. This implies that $O_{4g1}^{(2)}$ does not contribute, while $O_{4g2}^{(2)}$ does. Once again we can prove that $O_{4g2}^{(2)}$ generates a contribution that enters with simply the leading power gluon soft function $S_g^{(0)}$ (the direct proof of this requires some fairly extensive color algebra). This happens despite the fact that the subleading power beam function $B_{gg}^{(2)}$ has six color contractions. The contribution from this four gluon operator first enters the cross section at $\mathcal{O}(\alpha_s^2)$.
\vspace{0.4cm}
\noindent{\bf{Four Quark Operators:}}
For a four quark operator to interfere with the leading power operator, it must have both zero fermion number and a helicity 1 projection in each collinear sector. This eliminates all four quark operators from contributing to the cross section at $\mathcal{O}(\lambda^2)$.
\vspace{0.4cm}
\noindent{\bf{$\mathcal{P}_\perp$ Operators:}}
Both the operators involving $\mathcal{P}_\perp$ insertions have the correct symmetry properties and therefore both $O_{\mathcal{P} \chi }^{(2)}$ and $O_{\mathcal{P} \mathcal{B} }^{(2)}$ can contribute to the $\mathcal{O}(\lambda^2)$ cross section. Both contributions have a leading power gluon beam function $B_g^{(0)}$ and soft function $S_g^{(0)}$. The operator $O_{\mathcal{P} \mathcal{B} }^{(2)}$ has a similar structure to the operator $\mathcal{O}^{(2)}_{\cP1}$ found in the quark case in \cite{Feige:2017zci}, which contributes a leading log to the thrust (beam thrust) cross section \cite{Moult:2016fqy}. It involves a subleading power beam function $B_{ggP}^{(2)}$ (with two color contractions). On the other hand, we find in \Sec{sec:match_qqg_lam2} that the operator $O_{\mathcal{P} \chi }^{(2)}$ has a vanishing Wilson coefficient at tree level, so its factorized contribution starts at least at ${\cal O}(\alpha_s^2)$ for the cross section. It has a subleading beam function $B_{gqP}^{(2)}$ with a single color contraction.
\vspace{0.4cm}
\noindent{\bf{Ultrasoft Operators:}}
The ultrasoft operators involving quark fields cannot contribute to the cross section through interference with the leading power operator due to fermion number conservation. Therefore, only the gluon operators of \Eqs{eq:Hgggus}{eq:Hdggus} contribute. They have leading power gluon beam functions $B_g^{(0)}$.
\subsubsection{Comparison with $\bar q\, \Gamma q$}\label{sec:compare}
It is interesting to briefly compare the structure of the operator basis, as well as the contributions to the $\mathcal{O}(\lambda^2)$ cross section, to the basis for a process with two collinear sectors initiated by the $\bar q \Gamma q$ current as discussed in \cite{Feige:2017zci}. The leading power factorization theorems for the two cases are essentially identical, with simply a replacement of quark and gluon jet (beam) functions, as well as the color charges of the Wilson lines in the soft functions. However, at subleading power there are interesting differences arising both from the helicity structure of the currents, as well as from the form of the leading power Wilson coefficient.
An interesting feature of $gg\to H$ is that the Wilson coefficient for the leading power operator, which is given in \Sec{sec:matching_lp}, depends explicitly on the large label momenta of the gluons at tree level. This is not the case for the $\bar q\, \Gamma q$ current, whose leading power operator has a Wilson coefficient that is independent of the large label momenta at tree level. As discussed in \Sec{sec:nnlp_soft}, the Wilson coefficients of hard scattering operators involving insertions of $n \cdot \partial$, $\bar n \cdot \partial$, or $\mathcal{B}_{us(n)0}$ are related to the derivatives of the leading power Wilson coefficients by RPI. This implies that these particular operators vanish at tree level for a $\bar q\, \Gamma q$ current, but are present at tree level for $gg\to H$. For the $\bar q \, \Gamma q$ current the power corrections from the ultrasoft sector arise instead only from subleading power Lagrangian insertions. Therefore, the nature of power corrections in the two cases is quite different in terms of the organization of the effective theory in the ultrasoft sector. However, this does not say anything about their numerical size which would require a full calculation. Furthermore, the organization of the collinear hard scattering operators is nearly identical in the two cases.
Despite this difference in the organization of the particular corrections within the ultrasoft sector of the effective theory, there is also much similarity in the way that the subleading power operators contribute to the cross section at $\mathcal{O}(\lambda^2)$. In particular, in both cases, operators involving an additional ultrasoft or collinear gluon field as compared with the leading power operator contribute as an interference of the form $\mathcal{O}(\lambda^2) \mathcal{O}(1)$, see \Tab{tab:fact_func}. This is guaranteed by the Low-Burnett-Kroll theorem \cite{Low:1958sn,Burnett:1967km}. However, the subleading hard scattering operators that have a different fermion number in each sector than the leading power operators contribute as $\mathcal{O}(\lambda) \mathcal{O}(\lambda)$. For the $gg\to H$ case, this is the $O_{\mathcal{B}\bar n}^{(1)}$ operator, while for a $q\Gamma \bar q$ current considered in \cite{Feige:2017zci}, it was a hard scattering operator involving two collinear quarks recoiling against a collinear gluon. In the NNLO calculation of power corrections for the $q\Gamma \bar q$ case \cite{Moult:2016fqy}, this operator played an important role, as it gave rise to a leading logarithmic divergence not predicted by a naive exponentiation of the one-loop result, and it is expected that the same will be true here. We plan to consider this calculation in a future work, and to understand in more detail the relation between the leading logarithmic divergences for the $q\Gamma \bar q$ current, compared with a $gg$ current.
\section{Matching}\label{sec:matching}
In this section we perform the matching to the operators relevant for the calculation of the $\mathcal{O}(\lambda^2)$ cross section, which were enumerated in \Sec{sec:contribs} and summarized in \Tab{tab:summary}. As discussed in \Sec{sec:intro}, we will work in the context of an effective Higgs gluon coupling
\begin{align}\label{eq:heft}
\mathcal{L}_{\text{hard}}=\frac{C_1(m_t, \alpha_s)}{12\pi v}G^{\mu \nu}G_{\mu \nu} H\,,
\end{align}
obtained from integrating out the top quark. Here $v=(\sqrt{2} G_F)^{-1/2}=246$ GeV, and the matching coefficient is known to $\mathcal{O}(\alpha_s^3)$ \cite{Chetyrkin:1997un}. Corrections to the infinite top mass can be included in the matching coefficient $C_1$. We use the sign convention
\begin{align}
G_{\mu \nu}^a=\partial_\mu A_\nu^a-\partial_\nu A_\mu^a+g f^{abc} A_{\mu}^b A_\nu^c\,,\qquad iD^\mu=i\partial^\mu+gA^\mu\,.
\end{align}
In the matching, we take all particles as outgoing. However, to avoid a cumbersome notation we use $\epsilon$ instead of $\epsilon^*$ for the polarization of an outgoing gluon. We also restrict to Feynman gauge although we check gauge invariance by enforcing relevant Ward identities. For operators involving collinear gluon fields gauge invariance is guaranteed through the use of the $\mathcal{B}_\perp$ fields.
The Higgs effective Lagrangian has Feynman rules for $2$, $3$, and $4$ gluons which are summarized in \App{app:expand_gluon}. Due to the non-negative powers of momenta appearing in these Feynman rules they give rise to Wilson coefficients which are less singular than those arising from power corrections to the ultrasoft and collinear dynamics of SCET. This will be seen explicitly in the subleading power matching calculations. To simplify the notation throughout this paper we will suppress the factor of $C_1(m_t, \alpha_s)/(12\pi v)$, and simply write the Feynman rules and matching relations for the operator
\begin{align}
O^{\text{hard}}=G^{\mu \nu}G_{\mu \nu} H\,.
\end{align}
The dependence on $C_1(m_t, \alpha_s)/(12\pi v)$ is trivially reinstated.
Throughout the matching, collinear gluons in the effective theory will be indicated in Feynman diagrams as a spring with a line drawn through them, collinear quarks will be indicated by dashed lines, and ultrasoft gluons will be indicated with an explicit ``us". This will distinguish them from their full theory counterparts for which standard Feynman diagram notation for quarks and gluons is used. Furthermore, for the full theory diagrams, we will use the $\otimes$ symbol to denote the vertex of the Higgs effective theory, as compared with the purple circle used to denote a hard scattering operator in the effective theory.
Due to the large number of operators present in our basis, we find it most convenient to express the results of the tree-level matching in the form of the Wilson coefficient multiplying the relevant operator. For this purpose we define a shorthand notation with a caligraphic {\cal O},
\begin{align}
\mathcal{O}_X^{(i)} = C_X^{\text{tree}} O_X^{(i)}\,,
\end{align}
where as before, the superscript indicates the power suppression, and the subscript is a label that denotes the field and helicity content.
We will write results for $\mathcal{O}_X^{(i)}$ in a form such that it is trivial to identify the tree level Wilson coefficient $C_X^{\text{tree}}$, so that higher order corrections can be added as they become available.
\subsection{Leading Power Matching}\label{sec:matching_lp}
The leading power matching is of course well known, however, we reproduce it here for completeness and to illustrate the matching procedure. The matching can be performed using a two gluon external state. Since the leading power operator is independent of any $\perp$ momenta, in performing the matching we can take the momenta
\begin{align}
p_1^\mu =\omega_1 \frac{n^\mu}{2}, \qquad p_2 =\omega_2 \frac{\bar n^\mu}{2}\,,
\end{align}
and the polarizations to be purely $\perp$, namely $\epsilon_{i}^\mu=\epsilon_{i\perp}^\mu$. All of the operators in \Sec{sec:lp} give a non-vanishing contribution to the two-gluon matrix element for this choice of polarization.
In the two gluon matrix element, this choice of polarization does not remove overlap with any of the operators in \Sec{sec:lp}.
Expanding the QCD result, we find
\begin{align}
\left.\fd{2.0cm}{figures/matching_gg_low} \right |_{\mathcal{O}(\lambda^0)} &= -2i \delta^{ab} \omega_1 \omega_2 \epsilon_{1\perp} \cdot \epsilon_{2\perp}\,.
\end{align}
This is reproduced by the leading power operator
\begin{align}
\mathcal{O}^{(0)}_{\mathcal{B}}= -2\omega_1 \omega_2 \delta^{ab} \mathcal{B}^a_{\perp \bar n, \omega_2} \cdot \mathcal{B}^b_{\perp n, \omega_1} H\,,
\end{align}
or in terms of the helicity basis of \Eq{eq:hgg}, we have
\begin{align}
\mathcal{O}^{(0)}_{\mathcal{B}++}&=2\omega_1 \omega_2 \delta^{ab} \mathcal{B}^a_{ \bar n +, \omega_2} \cdot \mathcal{B}^b_{n +, \omega_1} H\,, \qquad
\mathcal{O}^{(0)}_{\mathcal{B}--}=2\omega_1 \omega_2 \delta^{ab} \mathcal{B}^a_{ \bar n -, \omega_2} \cdot \mathcal{B}^b_{n -, \omega_1} H\,.
\end{align}
While we focus here on the case where there is zero perp momentum in each collinear sector, we also give the Feynman rule in the case that each sector has a non-zero perp momentum. This will allow us to illustrate the gauge invariance properties of the collinear gluon field $\mathcal{B}_\perp$. The expansion of the collinear gluon field with an incoming momentum $k$ is given by
\begin{align}
\mathcal{B}^\mu_{n\perp}&= A^{\mu a}_{\perp k} T^a -k^\mu_\perp \frac{\bar n \cdot A^a_{nk} T^a}{\bar n \cdot k} +\cdots \,,
\end{align}
where the dots represent terms with multiple gluon fields. The two gluon terms are given in \App{app:expand_gluon}. Gauge invariance therefore dictates the Feynman rule of our operator in the case of generic perp momenta for the two gluon fields,
\begin{align}
\fd{2.0cm}{figures/matching_gg_FeynRule_low} &= -2i \delta^{ab} \omega_1 \omega_2 \left( \epsilon_{1\perp}^{\mu} -p_{1\perp}^\mu \frac{\bar n \cdot \epsilon_1}{ \bar n \cdot p_1} \right) \left( \epsilon_{2\perp}^{\mu} -p_{2\perp}^\mu \frac{\bar n \cdot \epsilon_2}{ \bar n \cdot p_2} \right) \,.
\end{align}
We note that the additional terms are essential to enforce that the required Ward identities are satisfied, and the result is gauge invariant. While this is trivial in this simple leading power example, for the more complicated matching calculations considered in the remainder of the paper we will often perform the matching for particular kinematic configurations, and the gauge invariance of the collinear gluon fields is an important ingredient to uniquely obtain the full result.
\subsection{Subleading Power Matching}\label{sec:matching_nlp}
We now consider the matching at $\mathcal{O}(\lambda)$. In \Sec{sec:nlp} we argued that the only $\mathcal{O}(\lambda)$ operator which can contribute to the cross section at $\mathcal{O}(\lambda^2)$ has two collinear quark fields in opposite collinear sectors and a collinear gluon field. We can therefore perform the matching using this external state. For concreteness we start with the case with a quark in the $n$-collinear sector, and a gluon and antiquark in the $\bar n$-collinear sector, $(q)_n (\bar q g)_\bn$. Since the power suppression arises from the explicit fields, and all propagators are off shell, we can use the kinematics
\begin{align}
p_1^\mu =\omega_1 \frac{n^\mu}{2}, \qquad p_2 =\omega_2 \frac{\bar n^\mu}{2}, \qquad p_3 =\omega_3 \frac{\bar n^\mu}{2}\,,
\end{align}
and take the polarization of the gluon to be purely $\perp$, $\epsilon_{i}^\mu=\epsilon_{i\perp}^\mu$. This choice suffices to obtain non-zero matrix elements for all the operators we want to probe, and to distinguish them from one another.
For the matching calculations, we will use the notation
\begin{align}
u_n(i)=P_n u(p_i)\,, \qquad \text{and} \qquad v_n(i)=P_n v(i)\,, \qquad \text{with} \qquad P_n=\frac{\Sl{n} \Sl{\bar n}}{4}\,,
\end{align}
for the projected SCET spinors. Here we have taken the momentum $p_i$ to be $n$-collinear, but similar relations exist for the case that it is $\bar n$-collinear.
The spinors obey
\begin{align}
u(p_i) = \Big( 1+ \frac{\Sl{p}_{i\perp}}{\bar{n} \cdot p_i} \frac{\Sl{\bar{n}}}{2} \Big) u_n(i) \,,
\qquad
u(p_i) = \Big( 1+ \frac{\Sl{p}_{i\perp}}{n \cdot p_i} \frac{\Sl{n}}{2} \Big) u_{\bar n}(i) \,,
\end{align}
for the $n$-collinear and $\bar n$-collinear cases respectively, with direct analogs for the $v(p_i)$ spinors.
Expanding the QCD diagram to $\mathcal{O}(\lambda)$, we find
\begin{align}
\left. \fd{2.5cm}{figures/matching_lam_qqg_low}\right |_{\mathcal{O}(\lambda)} &= \frac{-2ig\omega_3}{\omega_2} \bar u_n (p_1) \Sl{\epsilon}_{3\perp} T^a v_{\bar n}(p_2)\,.
\end{align}
There are no contributions from time ordered products in the effective theory to this particular matrix element used in the matching. This is due to the fact that there are no $\mathcal{O}(\lambda^0)$ or $\mathcal{O}(\lambda^1)$ operators involving just two quark fields, and the collinear Lagrangian insertions in each section preserve the fermion number of each sector, so this particular matrix element can not be obtained from Lagrangian insertions starting from the leading power operator involving two collinear gluons. Therefore, the result must be reproduced entirely by a hard scattering operator in SCET. This operator is given by
\begin{align}
\mathcal{O}^{(1)}_{\mathcal{B} \bar n}= -2g\, \frac{\omega_3}{\omega_2}\,
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3}
\chi_{\bar n,-\omega_2} H\,.
\end{align}
or, in terms of the helicity operators of \Eq{eq:H1_basis}
\begin{align}
\mathcal{O}^{(1)}_{\mathcal{B} \bar n+(+)}&=4g\frac{\omega_3}{\omega_2} T^a_{\alpha {\bar \beta}} \sqrt{\frac{\omega_1 \omega_2}{2}} \langle \bar n n \rangle \mathcal{B}^a_{\bar n+,\omega_3} J^{{\bar \alpha} \beta}_{n\bar n+} H\,, {\nonumber} \\
\mathcal{O}^{(1)}_{\mathcal{B} \bar n-(-)}&=-4g\frac{\omega_3}{\omega_2}T^a_{\alpha {\bar \beta}} \sqrt{\frac{\omega_1 \omega_2}{2}} [ \bar n n ] \mathcal{B}^a_{\bar n-,\omega_3} J^{{\bar \alpha} \beta}_{n\bar n-} H\,.
\end{align}
The Wilson coefficient has a singularity as the energy fraction of the quark in the $\bar n$-collinear sector becomes soft. This operator will therefore contribute to the leading logarithmic divergence at the cross section level at $\mathcal{O}(\lambda^2)$. Note that this operator is explicitly RPI-III invariant, with its Wilson coefficient taking the form of a ratio of the large momentum components of the two $\bar n$ collinear fields.
For convenience, we also give the full Feynman rule for this operator
\begin{align}
\fd{3cm}{figures/matching_lam_qqg_Feyn_low}
&=-2ig T^c \frac{\omega_3 }{\omega_2} \left( \gamma^\nu_\perp-\frac{\Sl{p}_{3\perp} n^\nu}{\omega_3} \right)\,.
\end{align}
Note that this Feynman rule contains terms that were not present in the matching calculation due to the special choice of kinematics used there. These additional terms are determined by the gauge invariant gluon field, $\mathcal{B}_{\perp \bar n}$, and it is easy to see that they ensure that this operator satisfies the required Ward identities.
The matching for the operators in the case $(\bar q)_n (q g)_\bn$ can be easily obtained from the above results by exploiting charge conjugation. This gives
\begin{align}
\mathcal{O}^{(1)}_{\mathcal{B} n}= -2g \frac{\omega_3}{\omega_1} \bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp n, \omega_3} \chi_{\bn,-\omega_2} H\,,
\end{align}
and for the helicity operators in \eq{H1_basis2} we obtain
\begin{align}
\mathcal{O}^{(1)}_{\mathcal{B} n+(-)}
&=4g\frac{\omega_3}{\omega_1} T^a_{\alpha {\bar \beta}}
\sqrt{\frac{\omega_1 \omega_2}{2}} \langle \bar n n \rangle \mathcal{B}^a_{n-,\omega_3} J^{{\bar \alpha} \beta}_{n\bar n+} H
\,, {\nonumber} \\
\mathcal{O}^{(1)}_{\mathcal{B} n-(+)}
&=-4g\frac{\omega_3}{\omega_1}T^a_{\alpha {\bar \beta}}
\sqrt{\frac{\omega_1 \omega_2}{2}} [ \bar n n ]
\mathcal{B}^a_{n+,\omega_3} J^{{\bar \alpha} \beta}_{n\bar n-} H
\,.
\end{align}
This concrete matching calculation at subleading power also clearly illustrates the distinction between subleading power hard scattering operators, and the standard picture of leading power factorization in terms of splitting functions. In the leading power factorization for $H\to gq\bar q$, when the $q\bar q$ pair become collinear, the amplitude factorizes into $H\to gg$ multiplied by a universal $g\to q\bar q$ splitting function. This gives rise to a leading power contribution, due to the nearly on-shell propagator of the intermediate gluon that undergoes the splitting. For the operator considered here, the gluon which splits into the $q\bar q$ pair is far off-shell, due to the fact that the $q$ and $\bar q$ are in distinct collinear sectors. Because of this, it is represented in the effective theory by a local contribution (namely a hard scattering operator), and this operator is power suppressed. The hard scattering operators therefore describe precisely the contributions that are not captured by a splitting function type factorization. While this is particularly clear for the operator considered here, this picture remains true for the subleading power hard scattering operators for the more complicated partonic contributions considered at subsubleading power in \Sec{sec:matching_nnlp}. The hard scattering operators describe local contributions, which do not factorize in standard splitting function type picture, and therefore in general have no relation to known splitting functions which appear in the literature.
\subsection{Subsubleading Power Matching}\label{sec:matching_nnlp}
In this section we perform the tree level matching to the $\mathcal{O}(\lambda^2)$ operators, considering only those which contribute at the cross section level at $\mathcal{O}(\lambda^2)$, as discussed in \Sec{sec:contribs}. Since there are a number of operators, each with different field content, we will consider each case separately.
\subsubsection{Ultrasoft Derivative}\label{sec:us_deriv}
We begin by performing the matching to the ultrasoft derivative operators of \Sec{sec:nnlp_soft}. To perform the matching we can use a state consisting of two perpendicularly polarized collinear gluons, and we take our momenta as
\begin{align}
p_1^\mu=(\omega_1 +k_1)\frac{n^\mu}{2} +p_{1r}\frac{\bar n^\mu}{2}+p_{1\perp}^\mu\,, \qquad p_2^\mu=(\omega_2 +k_2)\frac{\bar n^\mu}{2} +p_{2r}\frac{n^\mu}{2}+p_{2\perp}^\mu\,.
\end{align}
Since we have taken non-zero label perp momentum to keep the particles on shell we will have operators contributing that involving the $\mathcal{P}_\perp$ operator. These operators were not included in our basis, since we assumed zero total perp momentum in each sector. (See \App{app:gen_pt} for the additional operators required in the case that the collinear sectors have non-vanishing perp momentum.) However, these terms are easy to identify. Dropping these terms involving the label perp momentum to identify the contributions relevant for the matching, we find
\begin{align}
\left. \fd{2.0cm}{figures/matching_gg_low}\right |_{\mathcal{O}(\lambda^2)}&=-2i \delta^{ab} \omega_1 k_2 \epsilon_{1\perp} \cdot \epsilon_{2\perp} -2i\delta^{ab} \omega_2 k_1 \epsilon_{1\perp} \cdot \epsilon_{2\perp}\,.
\end{align}
This result must be completely reproduced by hard scattering operators in the effective theory, since the relevant subleading propagator insertions are proportional to residual components of the $\perp$ momentum, which we have taken to be zero in the matching (see \App{app:expand_gluon}, and in particular \Eq{eq:lam2_gluon_prop}).
The operators given in \Sec{sec:nnlp_soft} were defined post BPS field redefinition, in which case the partial derivative operator $\partial^\mu$ acts on gauge invariant building blocks. While the distinction between pre- and post-BPS field redefinition is not relevant for the calculation of the matrix elements in this particular case, since there are no ultrasoft emissions, it of course determines the form that the operators are written in. For convenience, we give the operators both before and after BPS field redefinition. Note that the collinear gluon field transforms as an adjoint matter field under ultrasoft gauge transformations since the ultrasoft gauge field acts as a background field.
Matching onto pre-BPS field redefinition operators, we find
\begin{align}\label{eq:preBPS_softgluon}
\mathcal{O}^{(2)}_{n\cdot D}=4\omega_1 {\rm tr} \left[ \mathcal{B}^\mu_{\perp n, \omega_1} [in \cdot D_{us}, \mathcal{B}^\mu_{\perp \bar n, \omega_2} ] \right ] H\,, \quad \mathcal{O}^{(2)}_{\bar n\cdot D}=4\omega_2 {\rm tr} \left[ \mathcal{B}^\mu_{\perp \bar n, \omega_2} [i\bar n \cdot D_{us}, \mathcal{B}^\mu_{\perp n, \omega_1}] \right] H\,,
\end{align}
where the trace is over color. This color structure will be fixed by matching with an additional ultrasoft gluon in \Sec{sec:us_gluon}. To determine the operators post-BPS field redefinition, we can either directly apply the BPS field redefinition, or simply match to the operators of \Sec{sec:nnlp_soft}. We find that the operators where the ultrasoft derivative acts on the gluon fields are given by
\begin{align}
\mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us)(0)}=-2\omega_1 \mathcal{B}^{\mu a}_{\perp n, \omega_1} i n \cdot \partial \mathcal{B}^{\mu b}_{\perp \bar n, \omega_2} H\,, \quad \mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us) (\bar 0)}=-2\omega_2 \mathcal{B}^{\mu a}_{\perp \bar n, \omega_2} i\bar n \cdot \partial \mathcal{B}^{\mu b}_{\perp n, \omega_1} H\,,
\end{align}
or expanded in terms of the helicity operator basis
\begin{align}
\mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us(n)) \bar 0:++}
&=-2\omega_1 \mathcal{B}^{\mu a}_{n +, \omega_1} i \partial_{us(n)\bar 0}
\mathcal{B}^{\mu b}_{\bar n +, \omega_2} H
\,, \quad \\
\mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us(\bar n))0:++}
&=-2\omega_2 \mathcal{B}^{\mu a}_{\bar n +, \omega_2} i\partial_{us(\bar n) 0} \mathcal{B}^{\mu b}_{n +, \omega_1} H
\,, {\nonumber} \\
\mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us(n)) \bar 0:--}
&=-2\omega_1 \mathcal{B}^{\mu a}_{n -, \omega_1} i \partial_{us(n)\bar 0} \mathcal{B}^{\mu b}_{\bar n -, \omega_2} H
\,, \quad {\nonumber}\\
\mathcal{O}^{(2)ab}_{\partial \mathcal{B} (us(\bar n))0:--}
&=-2\omega_2 \mathcal{B}^{\mu a}_{\bar n -, \omega_2} i\partial_{us(\bar n) 0} \mathcal{B}^{\mu b}_{n -, \omega_1} H
\,. {\nonumber}
\end{align}
Here the color indices are contracted against the basis of color structures given in \Eq{eq:Z2gd_colorus}.
These operators also give rise, after BPS field redefinition to operators involving $\mathcal{B}_{us}$. These will be discussed in \Sec{sec:us_gluon}.
As mentioned above \Eq{eq:Hdggus}, using the gluon equations of motion we can eliminate operators involving $\bar n\cdot \partial\mathcal{B}_{\bar n}$ and $n\cdot \partial \mathcal{B}_n$ from our basis to all orders in perturbation theory. This structure of the ultrasoft derivative operators is important for the matching at $\mathcal{O}(\lambda^2)$. In particular, only the $\bar n \cdot \partial$ acts on the $n$-collinear sector, and only the $n\cdot \partial$ acts on the $\bar n$-collinear sector. These correspond to the residual components of the label momenta. In a graph consisting of only collinear particles (i.e. no ultrasoft particles) the residual components of the label momenta can be chosen to vanish, so that these operators do not contribute. In all purely collinear graphs computed in the remainder of this paper, we will always make this choice, and therefore, these operators will not contribute. However, these operators will contribute, and will play an important role, when ultrasoft particles are present in the graph.
\subsubsection{qqg}\label{sec:match_qqg_lam2}
We now consider the case of the $\mathcal{O}(\lambda^2)$ operators involving two collinear quark fields, a collinear gluon field, and a $\mathcal{P}_\perp$ insertion. In \Sec{sec:nnlp} we argued that the only such operators have both quark fields in the same collinear sector, which we will take to be the $n$-collinear sector. To perform the matching we take the kinematics
\begin{align}\label{eq:qqg_momentum}
p_1^\mu=\omega_1 \frac{n^\mu}{2}+p_\perp^\mu +p_{1r} \frac{\bar n^\mu}{2}\,, \qquad p_2^\mu=\omega_2 \frac{n^\mu}{2}-p_\perp^\mu +p_{2r} \frac{\bar n^\mu}{2}\,, \qquad p_3^\mu=\omega_3 \frac{\bar n^\mu}{2}\,.
\end{align}
With this choice all subleading Lagrangian insertions in SCET vanish. This can be seen from the explicit subleading Lagrangians and Feynman rules given in \App{app:expand_gluon} by noting that these give contributions to this matrix element the involve residual components of $\perp$ momentum, or residual components of the large label momentum, which are zero for the choice of momentum in \Eq{eq:qqg_momentum}. The result must therefore be entirely reproduced by hard scattering operators. Expanding the QCD result we find that it vanishes at $\mathcal{O}(\lambda^2)$
\begin{align}
\left. \fd{2.5cm}{figures/matching_lam2_qqg_low} \right|_{\mathcal{O}(\lambda^2)}&= 0\,.
\end{align}
This is expected since this diagram involves only collinear dynamics in a single collinear sector, and non-trivial terms will be reproduced by power suppressed Lagrangians.
Therefore, at tree level, the hard scattering operators involving two quarks in the same sector along with a $\mathcal{P}_\perp$ insertion have vanishing Wilson coefficients. We do not have an argument that the Wilson coefficients of these operators would continue to vanish at higher orders in perturbation theory, and therefore we do not expect this to be the case.
\subsubsection{ggg}\label{subsec:ggg_match_lam2}
We now consider matching to the $\mathcal{O}(\lambda^2)$ three gluon operators which have a single $\mathcal{P}_\perp$. In \Sec{sec:contribs} we have argued that the only such operators that contribute to the cross section at $\mathcal{O}(\lambda^2)$ have two gluons in the same collinear sector, which we take to be $\bar n$ for concreteness. To perform the matching, we take the kinematics as
\begin{align}
p_1=\omega_1 \frac{n^\mu}{2}\,, \qquad p_2=\omega_2 \frac{\bar n^\mu}{2}+p_\perp^\mu+p_{2r} \frac{n^\mu}{2}\,, \qquad p_3=\omega_3 \frac{\bar n^\mu}{2}-p_\perp^\mu+p_{3r} \frac{n^\mu}{2}\,.
\end{align}
As a further simplification, we can take the polarization vector of the gluon in the $n$-collinear sector to be purely $\perp$,
$\epsilon_{1}^\mu=\epsilon_{1\perp}^\mu\,.$ All of the three gluon operators in our basis give a non-vanishing contribution to the three-gluon matrix element for this choice of polarizations.
In performing the expansion of the QCD diagrams we will obtain all three projections of the polarization vectors, namely $\bar n \cdot \epsilon_{2,3}$, $n \cdot \epsilon_{2,3}$, and $p_{\perp} \cdot \epsilon_{2,3\perp}$. However, all of the operators in our basis are formed from $\mathcal{B}_{\perp}$, and therefore contain only the $n \cdot \epsilon_{2,3}$ and $p_{\perp} \cdot \epsilon_{2,3\perp}$ components. From the on-shell conditions for the gluon we have the relation
\begin{align}
\omega_2 \frac{\bar n \cdot \epsilon_2}{2}=\frac{p_\perp^2 n\cdot \epsilon_2}{2\omega_2}-p_\perp \cdot \epsilon_\perp\,,
\end{align}
and similarly for $\epsilon_3$. Note that we always use the Minkowski signature for the $\perp$ momenta, i.e. $p_\perp^2=-\vec p_\perp^{~2}$. In performing the matching one can therefore keep track of only the $\perp$ polarizations, as long as the $\bar n \cdot \epsilon$ polarizations are converted into $n \cdot \epsilon$ and $p_\perp\cdot \epsilon_\perp$ using the above equation. This allows one to simplify the structure of the matching while keeping enough terms to reconstruct operators formed from $\mathcal{B}_\perp$ gluon fields.
Expanding the QCD diagrams, and keeping only the $\perp$ terms of the polarizations we find
\begin{align}
&\left. \left(\fd{2.25cm}{figures/matching_lam2_ggg4_low} + \fd{2.25cm}{figures/matching_lam2_ggg1_low}\right) \right |_{\mathcal{O}(\lambda^2)} {\nonumber} \\
&\hspace{0cm}= -4gf^{abc} \frac{\omega_3}{\omega_2}\left( \epsilon_{1\perp} \cdot \epsilon_{2\perp} p_{\perp}\cdot \epsilon_{3\perp} -\epsilon_{2\perp} \cdot \epsilon_{3\perp} p_{\perp} \cdot \epsilon_{1\perp} \right ) -4g f^{abc} \epsilon_{1\perp}\cdot \epsilon_{2\perp} p_{\perp}\cdot \epsilon_{3\perp} +[(2,b)\leftrightarrow (3,c)]\,, {\nonumber} \\
&\left. \fd{2.25cm}{figures/matching_lam2_ggg2_low}\right|_{\mathcal{O}(\lambda^2)} =0\,, {\nonumber} \\
&\left.\fd{2.25cm}{figures/matching_lam2_ggg3_low}\right|_{\mathcal{O}(\lambda^2)} =-4g f^{abc} \left(p_\perp \cdot \epsilon_{3\perp} \epsilon_{1\perp}\cdot \epsilon_{2\perp} -p_\perp \cdot \epsilon_{1\perp} \epsilon_{2\perp}\cdot \epsilon_{3\perp} +\frac{\omega_2}{\omega_3} p_\perp \cdot \epsilon_{3\perp} \epsilon_{1\perp} \cdot \epsilon_{2\perp} \right) {\nonumber} \\
&\hspace{4cm}+[(2,b)\leftrightarrow (3,c)]\,,
\end{align}
We have shown results for the individual diagrams to emphasize the structure of the contributions, namely that only the diagrams involving an off-shell propagator or the Higgs EFT three gluon vertex contribute.
Simplifying this result, we find that the sum of the QCD diagrams is given by
\begin{align}
&\left. \left(\fd{2.25cm}{figures/matching_lam2_ggg4_low} + \fd{2.25cm}{figures/matching_lam2_ggg1_low}+\fd{2.25cm}{figures/matching_lam2_ggg2_low}+\fd{2.25cm}{figures/matching_lam2_ggg3_low}\right) \right |_{\mathcal{O}(\lambda^2)} {\nonumber} \\
= &4gf^{abc} \left (2+\frac{\omega_3}{\omega_2}+\frac{\omega_2}{\omega_3} \right) \epsilon_{2\perp}\cdot \epsilon_{3\perp} p_\perp\cdot \epsilon_{1\perp} -4gf^{abc} \left( 2+\frac{\omega_3}{\omega_2}+\frac{\omega_2}{\omega_3} \right)\epsilon_{1\perp} \cdot \epsilon_{2\perp} p_\perp \cdot \epsilon_{3\perp}{\nonumber} \\
& -4gf^{abc} \left( 2+\frac{\omega_3}{\omega_2}+\frac{\omega_2}{\omega_3} \right)\epsilon_{1\perp} \cdot \epsilon_{3\perp} p_\perp \cdot \epsilon_{2\perp} \,.
\end{align}
For the choice of kinematics and polarizations used in the matching there are no SCET subleading Lagrangian contributions at this power, for similar reasons to the case of $gq\bar q$ discussed above. Therefore, the hard scattering operators must exactly reproduce the QCD result.
We write the operators and their Wilson coefficients both in the helicity basis of \Eq{eq:Hgggpperp_basis}, as well as in a more standard Lorentz structures, as the two may prove useful for different purposes. In terms of standard Lorentz structures the tree level matching gives
\begin{align}
\mathcal{O}^{(2)}_{\mathcal{P} \cB1}&=-\left( \frac{1}{2}\right)4g \left( 2+\frac{\omega_3}{\omega_2}+ \frac{\omega_2}{\omega_3} \right)i f^{abc} \mathcal{B}^a_{n\perp,\omega_1}\cdot \left[ \mathcal{P}_\perp \mathcal{B}^b_{\bar n \perp,\omega_2}\cdot \right] \mathcal{B}_{\bar n \perp,\omega_3}^c H\,, {\nonumber} \\
\mathcal{O}^{(2)}_{\mathcal{P} \cB2}&=4g\left( 2+\frac{\omega_3}{\omega_2} + \frac{\omega_2}{\omega_3}\right) if^{abc}\left[ \mathcal{P}_\perp \cdot \mathcal{B}_{\bar n \perp,\omega_3}^a \right] \mathcal{B}^b_{n\perp,\omega_1} \cdot \mathcal{B}_{\perp \bar n, \omega_2}^c H\,.
\end{align}
We have written the first operator in this form to incorporate the symmetry factor.
In the helicity basis, we have
\begin{align}\label{eq:3g_hel_match}
&\mathcal{O}_{\mathcal{P}\mathcal{B} +++[-]}^{(2)}
= 4g if^{abc} \left(2+\frac{\omega_3}{\omega_2}+ \frac{\omega_2}{\omega_3}\right) \mathcal{B}_{n+,\omega_1}^a\, \mathcal{B}_{\bar n+,\omega_3}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n+,\omega_2}^c \right ] \,H
\,,{\nonumber} \\
&\mathcal{O}_{\mathcal{P}\mathcal{B} ---[+]}^{(2)}
=4g if^{abc}\left(2+\frac{\omega_3}{\omega_2}+ \frac{\omega_2}{\omega_3}\right) \mathcal{B}_{n-,\omega_1}^a\, \mathcal{B}_{\bar n-,\omega_3}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n-,\omega_2}^c \right ] \,H\,,
{\nonumber} \\
%
&\mathcal{O}_{\mathcal{P}\mathcal{B} ++-[+]}^{(2)}
= -2g if^{abc}\left(2+\frac{\omega_3}{\omega_2}+ \frac{\omega_2}{\omega_3}\right)\, \mathcal{B}_{n+,\omega_1}^a\, \mathcal{B}_{ \bar n-,\omega_3}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n+,\omega_2}^c \right ] \,H
\,,{\nonumber} \\
&\mathcal{O}_{\mathcal{P}\mathcal{B} -+-[-]}^{(2)}
= -2g if^{abc}\left(2+\frac{\omega_3}{\omega_2}+ \frac{\omega_2}{\omega_3}\right)\mathcal{B}_{n-,\omega_1}^a\, \mathcal{B}_{ \bn -,\omega_3}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n +,\omega_2}^c \right ] \,H
\,.
\end{align}
We therefore see explicitly that the helicity selection rules are realized in the tree level matching. Furthermore, the Wilson coefficient is formed from Bose symmetric combinations of ratios of the large momentum components of the $\bar n$ collinear fields, as required by RPI-III invariance.
For convenience, we also give the Feynman rule of the combined operator with three external gluons
\begin{align}\label{eq:feynrule_ggg3}
\fd{2.75cm}{figures/feynrule_ggg3_low}&\\%
&\hspace{-2.5cm}=4g f^{abc}\Bigl( 2 + \frac{ \omega_2}{\omega_3} + \frac{ \omega_3}{\omega_2}\Bigr)\biggl[ p_\perp^\mu g_\perp^{\nu \rho} - p_\perp^\nu g_\perp^{\mu \rho} - p_\perp^\rho g_\perp^{\mu \nu} +\frac{p^2_\perp}{\omega_2 \omega_3}\left( \omega_3 n^\nu g_\perp^{\mu \rho}- \omega_2 n^\rho g_\perp^{\mu \nu} + p_\perp^\mu n^\nu n^\rho\right) \biggr]
. {\nonumber}
\end{align}
This contains additional terms not present in the earlier matching calculation, due to the particular choice of $\perp$ polarizations used to simplify the matching. One can explicitly check that this operator satisfies the Ward identity, which is gauranteed by the fact that it is written in terms of $\mathcal{B}_\perp$ fields.
It is also interesting to note that the Wilson coefficient of this operator has a divergence as either $\omega_2$, or $\omega_3$ become soft, so that it will give rise to a leading logarithmic divergence in the cross section at $\mathcal{O}(\lambda^2)$.
\subsubsection{Ultrasoft Gluon}\label{sec:us_gluon}
The operators involving a single ultrasoft insertion were given in \Sec{sec:nnlp_soft}, and it was argued that they were related by RPI to the leading power operator involving two collinear gluons. In this section we will explicitly perform the tree level matching to verify that this relation holds. The operators in \Sec{sec:nnlp_soft} were given after BPS field redefinition, since it is more convenient when enumerating a complete basis to work with a gauge invariant ultrasoft gluon field. While it is possible to directly match to the post-BPS operators, we will first perform the matching to pre-BPS field redefinition operators involving ultrasoft covariant derivatives, and verify the color structure given in \Eq{eq:preBPS_softgluon}. We will then give the operators after BPS field redefinition.
We perform the matching to a three particle external state, with one collinear gluon in each sector, and a single ultrasoft gluon.
To simplify the matching we take the momenta of the collinear particles as
\begin{align}
p_1^\mu=\omega_1 \frac{n^\mu}{2}\,, \qquad p_2^\mu =\omega_2 \frac{\bar n^\mu}{2}\,,
\end{align}
and the momentum of the ultrasoft particle as
\begin{align}
p_3^\mu=\bar n \cdot p_3 \frac{n^\mu}{2} +n\cdot p_3 \frac{\bar n^\mu}{2} +p_{3\perp}^\mu\,,
\end{align}
where $(n\cdot p_3, \bar n \cdot p_3, p_{3\perp})\sim (\lambda^2, \lambda^2,\lambda^2)$.
The full theory QCD diagrams expanded to $\mathcal{O}(\lambda^2)$ are given by
\begin{align}
\left. \fd{2.25cm}{figures/matching_gg_soft_low}\right|_{\mathcal{O}(\lambda^2)} &= 2g \omega_2 f^{abc} \epsilon_1 \cdot \epsilon_2 \frac{\bar n\cdot p_3}{n\cdot p_3}n\cdot \epsilon_3+4g f^{abc} \omega_2 \epsilon_1 \cdot \epsilon_3 \frac{\epsilon_{2\perp} \cdot p_{3\perp}}{n \cdot p_3} {\nonumber} \\
&-4g f^{abc} \omega_2 \epsilon_2\cdot \epsilon_3 \frac{p_{3\perp} \cdot \epsilon_{1\perp}}{ n \cdot p_3}\,, {\nonumber} \\
\left.\fd{2.25cm}{figures/matching_gg_soft_2_low} \right|_{\mathcal{O}(\lambda^2)} &=-2g \omega_1 f^{abc} \epsilon_1 \cdot \epsilon_2 \frac{n\cdot p_3}{\bar n\cdot p_3}\bar n\cdot \epsilon_3-4g f^{abc} \omega_1 \epsilon_2 \cdot \epsilon_3 \frac{\epsilon_{1\perp} \cdot p_{3\perp}}{\bar n \cdot p_3} {\nonumber} \\
&+4g f^{abc} \omega_1 \epsilon_1\cdot \epsilon_3 \frac{p_{3\perp} \cdot \epsilon_{2\perp}}{\bar n \cdot p_3}\,, {\nonumber} \\
\left.\fd{2.25cm}{figures/matching_gg_soft_4_low} \right|_{\mathcal{O}(\lambda^2)} &= 0\,,{\nonumber} \\
\left.\fd{2.25cm}{figures/matching_gg_soft_3_low} \right|_{\mathcal{O}(\lambda^2)} &=2g f^{abc} \omega_1 \epsilon_1 \cdot \epsilon_2 n\cdot \epsilon_3-2g f^{abc} \omega_2 \epsilon_1 \cdot \epsilon_2 \bar n \cdot \epsilon_3 \,.
\end{align}
In this case there are also contributions from $T$ product diagrams in SCET correcting the emission of an ultrasoft gluon. Once we subtract these terms from the full theory result, the remainder will be localized at the hard scale.
The $\mathcal{O}(\lambda^2)$ Feynman rule for the emission of a ultrasoft gluon from a collinear gluon is given by (see \App{app:expand_gluon} and e.g. \cite{Larkoski:2014bxa} for the explicit Feynman rule)
\begin{align}
\fd{3cm}{figures/soft_lam2_scet_low}&=\langle | T \mathcal{B}^\nu_{n\perp}(0) \mathcal{L}^{(2)}_{A_n} | \epsilon_n, p_n; \epsilon_s,p_s \rangle=-i f^{abc} \epsilon_{n\mu} \frac{2\epsilon_{s\rho} p_{s\sigma}}{p_n^- n\cdot p_s} \left( g^{\mu \rho}_\perp g^{\sigma \nu}_\perp - g^{\mu \sigma}_\perp g^{\rho \nu}_\perp \right)\,.
\end{align}
The two SCET diagrams involving this Lagrangian insertion are given by
\begin{align}
\fd{2.25cm}{figures/matching_gg_soft_scet_low} &=\frac{4\omega_2 f^{abc} }{n\cdot p_3} (\epsilon_1 \cdot \epsilon_3 p_{3\perp} \cdot \epsilon_{2\perp} -\epsilon_{1\perp} \cdot p_{3\perp} \epsilon_2 \cdot \epsilon_3)\,, {\nonumber} \\
\fd{2.25cm}{figures/matching_gg_soft_scet_2_low} &=\frac{4\omega_1 f^{abc} }{\bar n\cdot p_3} (\epsilon_1 \cdot \epsilon_3 p_{3\perp} \cdot \epsilon_{2\perp} -\epsilon_{1\perp} \cdot p_{3\perp} \epsilon_2 \cdot \epsilon_3)\,.
\end{align}
Finally we also have contributions from the ultrasoft derivative operators of \Sec{sec:us_deriv}, with a leading power emission of a ultrasoft gluon. For these diagrams we find
\begin{align}
\fd{2.25cm}{figures/matching_gg_soft_scet_deriv_low}&= 2g \omega_2 f^{abc} \epsilon_1 \cdot \epsilon_2 \frac{\bar n\cdot p_3}{n\cdot p_3}n\cdot \epsilon_3\,, \quad
\fd{2.25cm}{figures/matching_gg_soft_scet_deriv_2_low}&= -2g \omega_1 f^{abc} \epsilon_1 \cdot \epsilon_2 \frac{n\cdot p_3}{\bar n\cdot p_3}\bar n\cdot \epsilon_3\,.
\end{align}
The SCET $T$-products therefore exactly reproduce the QCD diagrams, with the exception of the contribution from the three gluon vertex of the Higgs effective theory.
Subtracting the SCET contributions from the expansion of the QCD diagrams, we find that the hard scattering operators are given by
\begin{align}\label{eq:preBPS_softgluon2}
\mathcal{O}^{(2)}_{n\cdot D}=4\omega_1 {\rm tr} \left[ \mathcal{B}^\mu_{\perp n, \omega_1} [n \cdot D_{us}, \mathcal{B}^\mu_{\perp \bar n, \omega_2} ] \right ] H\,, \quad \mathcal{O}^{(2)}_{\bar n\cdot D}=4\omega_2 {\rm tr} \left[ \mathcal{B}^\mu_{\perp \bar n, \omega_2} [\bar n \cdot D_{us}, \mathcal{B}^\mu_{\perp n, \omega_1}] \right] H\,,
\end{align}
as stated in \Eq{eq:preBPS_softgluon}.
In terms of gauge invariant ultrasoft gluon fields we have
\begin{align}
\mathcal{O}^{(2)}_{\mathcal{B}(us(n))}&=\left(i f^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc}\right) \left (-2g \omega_2 \mathcal{B}^a_{n\perp, \omega_1} \cdot \mathcal{B}^b_{\bar n \perp, \omega_2} \mathcal{B}^c_{us(n)0} \right)\,, {\nonumber} \\
\mathcal{O}^{(2)}_{\mathcal{B}(us(\bar n))}&=\left(i f^{abd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{dc}\right) \left (-2g \omega_1 \mathcal{B}^a_{n\perp, \omega_1} \cdot \mathcal{B}^b_{\bar n \perp, \omega_2} \mathcal{B}^c_{us(\bar n)0} \right)\,,
\end{align}
where the color structures that appear at tree level are the first components of the color basis of \Eqs{eq:Z2g_colorus}{eq:Z2g_colorus_2}.
In terms of helicity operators,
\begin{align}
\hspace{-0.5cm}\mathcal{O}_{\mathcal{B}(us(n)) 0:++}^{(2)}&=-2g \left(i f^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc}\right) \omega_2 \mathcal{B}^a_{n+,\omega_1} \mathcal{B}^b_{\bar n+,\omega_2} \mathcal{B}^c_{us(n)0} H\,,{\nonumber} \\
\mathcal{O}_{\mathcal{B}(us(n)) 0:--}^{(2)}&=-2g \left(i f^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc}\right) \omega_2 \mathcal{B}^a_{n-,\omega_1} \mathcal{B}^b_{\bar n-,\omega_2} \mathcal{B}^c_{us(n) 0} H\,, {\nonumber} \\
\hspace{-0.5cm}\mathcal{O}_{\mathcal{B}(us(\bar n)) 0:++}^{(2)}&=-2g \left(i f^{abd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{dc}\right) \omega_1 \mathcal{B}^a_{n+,\omega_1} \mathcal{B}^b_{\bar n+,\omega_2} \mathcal{B}^c_{us(\bar n)0} H\,,{\nonumber} \\
\mathcal{O}_{\mathcal{B}(us(\bar n)) 0:--}^{(2)}&=-2g \left(i f^{abd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{dc}\right) \omega_1 \mathcal{B}^a_{n-,\omega_1} \mathcal{B}^b_{\bar n-,\omega_2} \mathcal{B}^c_{us(\bar n) 0} H\,.
\end{align}
This agrees with the relation derived from RPI symmetry, given in \Eq{eq:usRPIrelation}. For convenience, we also give the Feynman rule for the contribution of the hard scattering operators to a single ultrasoft emission both before BPS field redefinition
\begin{align}
\fd{2.75cm}{figures/matching_gg_soft_Feynrule_low}&=2g f^{abc} \omega_1 g^{\mu \nu}_\perp n^\rho -2g f^{abc} \omega_2 g^{\mu \nu}_\perp \bar n^\rho\,,
\end{align}
as well as after BPS field redefinition
\begin{align}
\fd{2.75cm}{figures/matching_gg_soft_Feynrule_BPS_low}&=2g f^{abc} \left[ \omega_1 \left( n^\rho - \frac{ n \cdot p_3}{\bar n \cdot p_3} \bar n^\rho \right) - \omega_2 \left(\bar n^\rho - \frac{\bar n \cdot p_3}{n \cdot p_3} n^\rho \right) \right] {\nonumber} \\
&=2g f^{abc} \left[
n^\rho \left( \omega_1 + \frac{\bar n \cdot p_3}{n \cdot p_3} \omega_2 \right)
-\bar n^\rho \left( \omega_2 + \frac{ n \cdot p_3}{\bar n \cdot p_3} \omega_1 \right) \right] \,.
\end{align}
Note that the contribution from hard scattering operators before the BPS field redefinition is local, but not gauge invariant, since before BPS field redefinition there are also SCET $T$-product diagrams involving. After BPS field redefinition, the contribution from the hard scattering operators is gauge invariant, but at the cost of locality. However, as emphasized in \cite{Feige:2017zci}, the form of the non-locality is dictated entirely by the BPS field redefinition, and is therefore not problematic. It is therefore advantageous to work in terms of the ultrasoft gauge invariant building blocks, so that the contributions from the hard scattering operators alone are gauge invariant.
Note also that here we have restricted the $\perp$ momentum of the two collinear particles to vanish for simplicity. Furthermore, because of the ultrasoft wilson lines in the color structure of \Eq{eq:Z2g_colorus}, there are also Feynman rules with multiple ultrasoft emissions. This is analogous to the familiar case of the $\mathcal{B}_\perp$ operator which has Feynman rules for the emission of multiple collinear gluons.
\subsubsection{qqgg}
A basis for the operators involving two collinear quark and two collinear gluon fields was given in \Sec{sec:nnlp_collinear}. In \Sec{sec:contribs} it was argued that the only non-vanishing contributions to the cross section at $\mathcal{O}(\lambda^2)$ arise from operators with the two collinear quarks and a collinear gluon in one sector, recoiling against a collinear gluon in the other sector.
In performing the matching to these operators there are potentially $T$-product terms from the three gluon $\mathcal{O}(\lambda^2)$ operator of \Sec{subsec:ggg_match_lam2}, where one of the gluons splits into a $q\bar q$ pair.
By choosing the momentum
\begin{align}
p_1^\mu&=\omega_1 \frac{n^\mu}{2}+p_\perp^\mu +p_{1r} \frac{\bar n^\mu}{2}\,, \quad p_2^\mu=\omega_2 \frac{n^\mu}{2}-p_\perp^\mu +p_{2r} \frac{\bar n^\mu}{2}\,, \quad
p_3^\mu=\omega_3 \frac{\bar n^\mu}{2}\,, \quad p_4^\mu=\omega_4 \frac{n^\mu}{2}\,,
\end{align}
we see from \Eq{eq:feynrule_ggg3} that all SCET $T$-product contributions vanish, so that the result must be reproduced by hard scattering operators in SCET.
Expanding the QCD diagrams to $\mathcal{O}(\lambda^2)$, we find that all the contributions from the two gluon vertex in the Higgs effective theory vanish
\begin{align}
\left. \fd{2.5cm}{figures/matching_lam2_qqgg1_low}\right|_{\mathcal{O}(\lambda^2)} &=0\,, \quad \left.\fd{2.5cm}{figures/matching_lam2_qqgg2_low}\right|_{\mathcal{O}(\lambda^2)} = 0\,, \quad \left.\fd{2.5cm}{figures/matching_lam2_qqgg3_low}\right|_{\mathcal{O}(\lambda^2)} = 0{\nonumber} \\
\left.\fd{2.25cm}{figures/matching_lam2_qqgg4_low}\right|_{\mathcal{O}(\lambda^2)} &=0\,, \quad \left.\fd{2.5cm}{figures/matching_lam2_qqgg5_low}\right|_{\mathcal{O}(\lambda^2)} =0\,.
\end{align}
This result might be anticipated from the structure of the diagrams. However, there is a non-vanishing contribution from the three-gluon vertex in the Higgs effective theory
\begin{align}
\left. \fd{2.5cm}{figures/matching_lam2_qqgg6_low} \right|_{\mathcal{O}(\lambda^2)} &=-\frac{4g^2 f^{abc} \omega_4 \epsilon_{3\perp}\cdot \epsilon_{4\perp}}{(\omega_1+\omega_2)^2} \bar u_n(p_1) T^a \frac{\Sl{\bar n}}{2} v_n(p_2)\,.
\end{align}
In terms of standard Lorentz and Dirac structures the corresponding hard scattering operator is given by
\begin{align}
\mathcal{O}^{(2)}_{\cB1}=\frac{4g^2 if^{abc} \omega_4 }{(\omega_1+\omega_2)^2}\mathcal{B}^b_{n\perp,\omega_4 }\cdot \mathcal{B}^c_{\bar n \perp,\omega_3} \bar \chi_{n,\omega_1} T^a \frac{\Sl{\bar n}}{2} \chi_{n,-\omega_2}H\,.
\end{align}
Projected onto the helicity operator basis of \Eq{eq:Hqqgg_basis3}, and using the color basis of \Eq{eq:ggqqll_color}, we find
\begin{align}
\mathcal{O}^{(2)}_{\cB1++(0)}&=-\frac{4g^2 \omega_4 }{(\omega_1+\omega_2)^2}2\sqrt{\omega_1 \omega_2} \left( (T^a T^b)_{\alpha {\bar \beta}} -(T^b T^a)_{\alpha {\bar \beta}} \right) \mathcal{B}^a_{n+,\omega_4} \mathcal{B}^b_{\bar n+,\omega_3} J_{n0}^{{\bar \alpha} \beta} H\,,{\nonumber} \\
\mathcal{O}^{(2)}_{\cB1--(0)}&=-\frac{4g^2 \omega_4 }{(\omega_1+\omega_2)^2}2\sqrt{\omega_1 \omega_2} \left( (T^a T^b)_{\alpha {\bar \beta}} -(T^b T^a)_{\alpha {\bar \beta}} \right) \mathcal{B}^a_{n-,\omega_4} \mathcal{B}^b_{\bar n-,\omega_3} J_{n0}^{{\bar \alpha} \beta} H\,, {\nonumber}\\
\mathcal{O}^{(2)}_{\cB1++(\bar0)}&=-\frac{4g^2 \omega_4 }{(\omega_1+\omega_2)^2}2\sqrt{\omega_1 \omega_2} \left( (T^a T^b)_{\alpha {\bar \beta}} -(T^b T^a)_{\alpha {\bar \beta}} \right) \mathcal{B}^a_{n+,\omega_4} \mathcal{B}^b_{\bar n+,\omega_3} J_{n\bar0}^{{\bar \alpha} \beta}H\,, {\nonumber} \\
\mathcal{O}^{(2)}_{\cB1--(\bar0)}&=-\frac{4g^2 \omega_4 }{(\omega_1+\omega_2)^2}2\sqrt{\omega_1 \omega_2} \left( (T^a T^b)_{\alpha {\bar \beta}} -(T^b T^a)_{\alpha {\bar \beta}} \right) \mathcal{B}^a_{n-,\omega_4} \mathcal{B}^b_{\bar n-,\omega_3} J_{n\bar0}^{{\bar \alpha} \beta}H\,.
\end{align}
For convenience, we also give the Feynman rule for the operator
\begin{align}
\fd{2.5cm}{figures/matching_lam2_qqgg_Feyn_low}&=-\frac{4g^2 f^{abc} T^a \omega_4}{(\omega_1+\omega_2)^2} \left( g^{\mu \nu}_\perp -\frac{p^\nu_{4\perp} \bar n^\mu}{\omega_4} \right) \frac{\Sl{\bar n}}{2}\,.
\end{align}
Again, this contains additional terms not present in the matching calculation, and it is straightforward to check that they are necessary to satisfy the required Ward identities.
\subsubsection{gggg}
Finally, we consider the matching to the operators involving four collinear gluon fields. A basis of such operators was given in \Eq{eq:H_basis_gggg_2}. In \Sec{sec:contribs} it was argued that to contribute to the cross section at $\mathcal{O}(\lambda^2)$, there must be three collinear gluons in the same sector. For concreteness, we take this to be the $\bar n$ sector. The operators with three gluons in the $n$ sector can be obtained by crossing $\bar n \leftrightarrow n$.
To perform the matching we choose the momenta as
\begin{align}\label{eq:momentagggg}
p_1^\mu&=\omega_1 \frac{ n^\mu}{2}\,, \quad p_2^\mu=\omega_2 \frac{\bar n^\mu}{2}\,, \quad p_3^\mu=\omega_3 \frac{\bar n^\mu}{2}-p_\perp^\mu +p_{3r} \frac{n^\mu}{2}\,, \quad p_4^\mu=\omega_4 \frac{\bar n^\mu}{2}+p_\perp^\mu +p_{4r} \frac{ n^\mu}{2}\,.
\end{align}
With this choice, each particle in the $\bar n$ sector is on-shell, but the sum of any two of their momenta is off-shell,
\begin{equation}
p_i^2 = 0\,, \qquad (p_1 + p_j)^2 \sim \mathcal{O}(1)\,, \qquad (p_j + p_k)^2 \sim \mathcal{O}(\lambda^2) \,, \qquad j,k=2,3,4 \,;~ j\neq k\,,
\end{equation}
which regulates all propagators. This particular choice of momenta is convenient since it simplifies $T$-product contributions from SCET. Furthermore, we take the external polarizations to be purely perpendicular, i.e. $\epsilon_i^\mu = \epsilon_{i\perp}^\mu$. All of the four gluon operators give a non-vanishing contribution to the four-gluon matrix element for this choice of polarization, allowing their Wilson coefficients to be obtained.
In computing the full theory diagrams for the matching it is convenient to separate the diagrams into those involving on-shell propagators, which will be partially reproduced by $T$-product terms in SCET, and diagrams involving only off-shell propagators. Since the four gluon operators obtain their power suppression entirely from the fields, for diagrams involving only off-shell propagators the residual momenta in \Eq{eq:momentagggg} can be ignored, as they contribute only power suppressed contributions. Diagrams with on-shell propagators are regulated by the residual momenta in \Eq{eq:momentagggg}.
We begin by considering the expansion of the full theory diagrams that don't involve any on-shell propagators. In this case, all $\perp$ momenta can be set to zero, and the result will be purely local. The relevant QCD diagrams expanded to $\mathcal{O}(\lambda^2)$ arise from the four gluon vertex in the Higgs effective theory,\\
\noindent\begin{minipage}{.3\linewidth}
\begin{equation}
\hspace{2cm}\left.\fd{2.25cm}{figures/matching_lam2_gggg1_low}\right|_{\mathcal{O}(\lambda^2)} {\nonumber}
\end{equation}
\end{minipage}%
\begin{minipage}{.7\linewidth}
\begin{align}
=&4ig^2 (f^{eab}f^{ecd}+f^{ead}f^{ecb}) \epsilon_{1\perp} \cdot \epsilon_{3\perp} \epsilon_{2\perp} \cdot \epsilon_{4\perp} {\nonumber} \\
&+4ig^2 (f^{eac}f^{ebd} +f^{ead}f^{ebc} ) \epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{2\perp} \cdot \epsilon_{3\perp} {\nonumber} \\
&+4ig^2 ( f^{eab}f^{edc}+f^{eac}f^{edb}) \epsilon_{1\perp} \cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_{4\perp} \,,
\end{align}
\end{minipage}\\
\newline
from a splitting off of the three gluon vertex,\\
\noindent\begin{minipage}{.3\linewidth}
\begin{equation}
\hspace{1cm}\left.\left( \fd{2.15cm}{figures/matching_lam2_gggg2_low}+\text{perms} \right)\right|_{\mathcal{O}(\lambda^2)} {\nonumber}
\end{equation}
\end{minipage}%
\begin{minipage}{.7\linewidth}
\begin{align}
=&2ig^2 \left( \frac{\omega_3-\omega_2}{\omega_4} \right) f^{abe}f^{cde} \epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{3\perp} \cdot \epsilon_{2\perp} {\nonumber} \\
& +[(2,d)\leftrightarrow (4,b)] +[(3,c)\leftrightarrow (4,b)] \,,
\end{align}
\end{minipage}\\
\newline
and from multiple emissions off of the two gluon vertex, either using the four gluon vertex with a single off-shell propagator\\
\noindent
\raisebox{1cm}{\begin{minipage}{.3\linewidth}
\begin{equation}
\hspace{0.25cm}\left.\left(\fd{2.15cm}{figures/matching_lam2_gggg3a_low}+\text{perms} \right) \right|_{\mathcal{O}(\lambda^2)} {\nonumber}
\end{equation}
\end{minipage}}%
\begin{minipage}{.7\linewidth}
\begin{align}
\hspace{1cm}=& 2ig^2\left( \frac{\omega_2}{\omega_3+\omega_4}\right){\nonumber} \\
& \left[ f^{bae} f^{cde} (\epsilon_{3\perp} \cdot \epsilon_{4\perp} \epsilon_{1\perp} \cdot \epsilon_{2\perp} -\epsilon_{4\perp} \cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_{1\perp}) \right.{\nonumber} \\
&+f^{bce}f^{ade} (\epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{3\perp} \cdot \epsilon_{2\perp} -\epsilon_{4\perp} \cdot \epsilon_{2\perp} \epsilon_{1\perp} \cdot \epsilon_{3\perp}){\nonumber} \\
&\left. +f^{bde}f^{ace}(\epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{3\perp} \cdot \epsilon_{2\perp}-\epsilon_{3\perp} \cdot \epsilon_{4\perp} \epsilon_{1\perp} \cdot \epsilon_{2\perp}) \right] {\nonumber} \\
&\hspace{-0cm} +[(2,d)\leftrightarrow (3,c)] +[(2,d)\leftrightarrow (4,b)] \,,
\end{align}
\end{minipage}\\
\newline
\noindent or sequential emissions with two off-shell propagators
\noindent\begin{minipage}{.3\linewidth}
\begin{equation}
\left.\left(\fd{2.15cm}{figures/matching_lam2_gggg7_low}+\text{perms}\right) \right|_{\mathcal{O}(\lambda^2)} {\nonumber}
\end{equation}
\end{minipage}%
\begin{minipage}{.7\linewidth}
\begin{align}
=&2ig^2 \frac{\omega_2 \omega_3}{\omega_4(\omega_3+\omega_4)} \epsilon_{2\perp} \cdot \epsilon_{3\perp} \epsilon_{4\perp} \cdot \epsilon_{1\perp} f^{abe}f^{ecd}{\nonumber} \\
& +[\text{perms}] \,.
\end{align}
\end{minipage}\\
In the last case we have not explicitly listed the permutations, since all possible permutations are required.
We now consider the expansion of the full theory diagrams involving on-shell propagators. These will generically involve both local and non-local pieces. The non-local pieces will be directly reproduced by $T$-products in the effective theory. The first class of diagrams involving on-shell propagators are those with all propagators on-shell. Here, at tree level, the dynamics occurs entirely within a single collinear sector. The two relevant QCD diagrams expanded to $\mathcal{O}(\lambda^2)$ are
\begin{align}
\left.\fd{2.45cm}{figures/matching_lam2_gggg3_low} \right|_{\mathcal{O}(\lambda^2)} &=0\,, \qquad
\left.\left(\fd{2.00cm}{figures/matching_lam2_gggg5_low} +\text{perms} \right)\right|_{\mathcal{O}(\lambda^2)} &=0\,,
\end{align}
both of which have vanishing subleading power contributions.
Next, we consider diagrams involving both on-shell and off-shell propagators. To simplify the results, we will often use the relation
\begin{align}
\frac{p_\perp^2}{(p_2+p_3)^2}=-\frac{\omega_3}{\omega_2}\,,
\end{align}
which will allow us to write the result in terms of a local term, which is just a rational function of the label momenta, and a non-local term, which explicitly contains the on-shell propagator. These non-local terms will be cancelled by the $T$-product diagrams in SCET. For a first class of diagrams, where we have a nearly on-shell splitting in the $\bar n$-collinear sector, we have both a local term
\begin{align}
\left.\left(\fd{2.25cm}{figures/matching_lam2_gggg4_low}\right) \right|_{\mathcal{O}(\lambda^2)}& =4ig^2 f^{aed}f^{bce}\frac{(\omega_3 - \omega_4)}{(\omega_3 + \omega_4)}\epsilon_{1\perp} \cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_{4\perp}\,,
\end{align}
when the splitting is into the particles $3$ and $4$, as well as a term that has both local and non-local pieces
\begin{align}
\left.\left(~\fd{2.25cm}{figures/matching_lam2_gggg4b_low}~ +~ \fd{2.25cm}{figures/matching_lam2_gggg4c_low}~\right) \right|_{\mathcal{O}(\lambda^2)}& \\
&\hspace{-6cm} =\frac{4ig^2 f^{aeb}f^{dce}}{\omega_4}\left[ \frac{2(\omega_2 + \omega_3)}{(p_2 + p_3)^2} p_\perp \cdot \epsilon_{1\perp} p_\perp \cdot \epsilon_{2\perp} \epsilon_{3\perp}\cdot \epsilon_{4\perp} \right.{\nonumber}\\
&\hspace{-3.2cm}\left. -(2\omega_3 + \omega_4) \epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{2\perp} \cdot \epsilon_{3\perp} \vphantom{\frac{(a)}{(b)}}\right] + [3 \leftrightarrow 4, b \leftrightarrow c , p_\perp \to -p_\perp]\,. {\nonumber}
\end{align}
As will be discussed in more detail when we consider the corresponding diagrams in the EFT, the first permutation is purely local, since there is no corresponding $T$-product term in the effective theory, and thus it must be fully reproduced by a hard scattering operator. This particular splitting allows a slight simplification in the calculation of the SCET diagrams. For a second class of diagrams, where we have an on-shell splitting emitted from an off-shell leg, we again have a purely local term
\begin{align}
\left.\left(\fd{2.05cm}{figures/matching_lam2_gggg6_low}\right) \right|_{\mathcal{O}(\lambda^2)}& =0\,,
\end{align}
as well as non-local contributions,
\begin{align}
\left.\left(~\fd{2.05cm}{figures/matching_lam2_gggg6b_low}~ +~ \fd{2.05cm}{figures/matching_lam2_gggg6c_low}~\right) \right|_{\mathcal{O}(\lambda^2)}& \\
&\hspace{-6.5cm} =2ig^2 f^{aeb}f^{dce}\left[ \left(\frac{4\omega_4}{(\omega_2 + \omega_3)(p_2 + p_3)^2}\right) p_\perp \cdot \epsilon_{1\perp} p_\perp \cdot \epsilon_{2\perp} \epsilon_{3\perp}\cdot \epsilon_{4\perp} \right.{\nonumber}\\
&\hspace{-4.5cm}\left. -\frac{\omega_3(\omega_2 - \omega_3)(\omega_2 + \omega_3 + \omega_4)^2}{\omega_2\omega_4(\omega_2 + \omega_3)^2} \epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{2\perp} \cdot \epsilon_{3\perp} \right] + [3 \leftrightarrow 4, b \leftrightarrow c , p_\perp \to -p_\perp]\,.{\nonumber}
\end{align}
Again, we see the same pattern, that the first permutation gives rise to a purely local term, while the second two permutations give rise to both local and non-local terms.
Finally, we have the diagrams involving the three gluon vertex in the Higgs effective theory. We again have a local contribution
\begin{align}
\left.\left(\fd{2.05cm}{figures/matching_lam2_gggg8_low}\right) \right|_{\mathcal{O}(\lambda^2)}& =-2ig^2 f^{ade}f^{ebc} \frac{\omega_2(\omega_3-\omega_4)}{(\omega_3 + \omega_4)^2} \epsilon_{1\perp} \cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_4\,,
\end{align}
and a non-local contribution
\begin{align}
\left.\left(~\fd{2.05cm}{figures/matching_lam2_gggg8b_low}~ +~ \fd{2.05cm}{figures/matching_lam2_gggg8c_low}~\right) \right|_{\mathcal{O}(\lambda^2)}& \\
&\hspace{-6.5cm} =2ig^2 f^{aeb}f^{dce}\left[ \frac{8}{(p_2 + p_3)^2} p_\perp \cdot \epsilon_{1\perp} p_\perp \cdot \epsilon_{2\perp} \epsilon_{3\perp}\cdot \epsilon_{4\perp} \right.{\nonumber}\\
&\hspace{-4.cm}\left. -\biggl\{\frac{ (\omega_3 + \omega_4)^2 - \omega_2\omega_3}{\omega_2\omega_4}\biggr\} \epsilon_{1\perp} \cdot \epsilon_{4\perp} \epsilon_{2\perp} \cdot \epsilon_{3\perp} \right] + [3 \leftrightarrow 4, b \leftrightarrow c , p_\perp \to -p_\perp]\,. {\nonumber}
\end{align}
The non-local terms in the above expansions must be reproduced by $T$-product terms in the effective theory. First, there are potential contributions from $\mathcal{O}^{(2)}_{\mathcal{P}\mathcal{B}}$, with the two gluon Feynman rule for $\mathcal{B}_{\bar n,\perp}$, which is given in \App{app:expand_gluon}. Such contributions give vanishing overlap for our choice of $\perp$ polarizations. There are however $T$-product contributions arising from the three gluon $\mathcal{O}^{(2)}_{\mathcal{P}\mathcal{B}}$ operator, with an $\mathcal{L}^{(0)}$ insertion.
The three gluon Feynman rule for the $\mathcal{O}^{(2)}_{\mathcal{P}\mathcal{B}}$ vertex was given in \Eq{eq:feynrule_ggg3}.
Since the $\mathcal{O}^{(2)}_{\mathcal{P}\mathcal{B}}$ operator has an explicit $\mathcal{P}_\perp$ insertion, it vanishes in the case that either of the particles in the $\bar n$ sector has no perpendicular momentum. This is why our particular choice of momenta for the matching simplifies the structure of the $T$-products. The two non-vanishing permutations are given by
\begin{align}\label{eq:ggggSCET}
& \fd{2.00cm}{figures/matching_gggg_scetb_low} + \fd{2.00cm}{figures/matching_gggg_scetc_low} \\
=&-8ig^2 f^{abe}f^{ecd} \frac{(\omega_2 + \omega_3 + \omega_4)^2 }{(\omega_3 + \omega_2)\omega_4}\left[ \frac{\omega_3}{(\omega_2 + \omega_3)} \epsilon_{1\perp}\cdot \epsilon_{4\perp} \epsilon_{3\perp}\cdot \epsilon_{2\perp} - \frac{ p_\perp\cdot\epsilon_{1\perp} p_\perp\cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_{4\perp}}{(p_2 + p_3)^2} \right] {\nonumber} \\
&+ [3 \leftrightarrow 4, b \leftrightarrow c , p_\perp \to -p_\perp]\,, {\nonumber}
\end{align}
which consists both of a local and a non-local term. The non-local terms exactly reproduce the ones obtained in the QCD expansion
\begin{align}
\hspace{-1cm}
\left(\fd{1.80cm}{figures/matching_gggg_scetb_low} + \fd{1.80cm}{figures/matching_gggg_scetc_low} \right)_\text{\!\!\!non-loc.} \hspace{-0.5cm}
& \!\! = \left(\fd{1.90cm}{figures/matching_lam2_gggg4b_low} + \fd{1.90cm}{figures/matching_lam2_gggg6b_low} + \fd{1.90cm}{figures/matching_lam2_gggg8b_low} +\! \text{ perms }\!\!\! \right)_\text{\!\!\!non-loc.} \hspace{-1cm} {\nonumber} \\[0.4cm]
&\hspace{-5cm}=8ig^2 p_\perp\cdot\epsilon_{1\perp} p_\perp\cdot \epsilon_{2\perp} \epsilon_{3\perp} \cdot \epsilon_{4\perp} \left( \frac{f^{abe}f^{ecd}} {(p_2 + p_3)^2} \frac{(\omega_2 + \omega_3 + \omega_4)^2 }{(\omega_3 + \omega_2)\omega_4} + [3 \leftrightarrow 4 , b \leftrightarrow c] \right) \,.
\end{align}
While it is of course necessary that the EFT reproduces all such non-local terms, this is also a highly non trivial cross check of both the three and four gluon matching.
The matching coefficients for the hard scattering operators are given by the remaining local terms. Before presenting the result we briefly comment on the organization of the color structure. All diagrams are proportional to $f^{abe}f^{cde}$, $f^{ace}f^{bde}$ or $f^{ade}f^{bce}$, which are related by the Jacobi identity $f^{abe}f^{cde} = f^{ace}f^{bde} - f^{ade}f^{bce}$. A basis in terms of structure constants can easily be related to the trace basis of \eqref{eq:gggg_color} using
\begin{align}\label{eq:colmatrrel}
f^{ace}f^{bde} &= {\rm tr}[abdc] + {\rm tr}[acdb] - {\rm tr}[acbd] - {\rm tr}[adbc] = e_2 - e_3\,,{\nonumber} \\
f^{ade}f^{bce} &= {\rm tr}[abcd] + {\rm tr}[adcb] - {\rm tr}[acbd] - {\rm tr}[adbc] = e_1 - e_3\,,
\end{align}
where $e_i$ is the $i$-th element of the basis in \eqref{eq:gggg_color}. We find it most convenient to write the Wilson coefficient in the $(f^{ace}f^{bde},f^{ade}f^{bce})$ basis.
After subtracting the local piece of the SCET $T-$product of \eqref{eq:ggggSCET} from the full theory graphs, and manipulating the result to bring it into a compact form, we find the following operator
\begin{align}
\mathcal{O}^{(2)}_{4g}= 16 \pi \alpha_s f^{ade}f^{bce} (\mathcal{B}^a_{n\perp,\omega_i} \cdot \mathcal{B}^b_{\bar n \perp, \omega_j})(\mathcal{B}^c_{\bar n \perp, \omega_k} \cdot \mathcal{B}_{\bar n \perp, \omega_\ell}^d) \left( 3 + \dfrac{\omega_j^3 + \omega_k^3 + \omega_\ell^3 + \omega_j\omega_k\omega_\ell}{(\omega_j + \omega_k)(\omega_j + \omega_\ell)(\omega_k + \omega_\ell)}\right) \,.
\end{align}
The Wilson coefficient is manifestly RPI-III invariant. When the matrix element of this operator is taken we are forced to sum over permutations which gives the proper Bose symmetric result, as well as inducing terms with other color structures. In terms of the helicity operators of \Eq{eq:H_basis_gggg_2}, we have
\begin{align}
\mathcal{O}^{(2)}_{4g}
&= 16 \pi \alpha_s f^{ade}f^{bce} \left( 3 + \dfrac{\omega_j^3 + \omega_k^3 + \omega_\ell^3 + \omega_j\omega_k\omega_\ell}{(\omega_j + \omega_k)(\omega_j + \omega_\ell)(\omega_k + \omega_\ell)}\right) {\nonumber} \\
&\ \ \times \Big[ \mathcal{B}^a_{n+,\omega_i} \mathcal{B}^b_{\bar n +, \omega_j} \mathcal{B}^c_{\bar n +, \omega_k} \mathcal{B}_{\bar n -, \omega_\ell}^d + \mathcal{B}^a_{n+,\omega_i} \mathcal{B}^b_{\bar n +, \omega_j} \mathcal{B}^c_{\bar n -, \omega_k} \mathcal{B}_{\bar n +, \omega_\ell}^d
{\nonumber} \\
&\qquad +\mathcal{B}^a_{n-,\omega_i} \mathcal{B}^b_{\bar n -, \omega_j} \mathcal{B}^c_{\bar n +, \omega_k} \mathcal{B}_{\bar n -, \omega_\ell}^d + \mathcal{B}^a_{n-,\omega_i} \mathcal{B}^b_{\bar n -, \omega_j} \mathcal{B}^c_{\bar n -, \omega_k} \mathcal{B}_{\bar n +, \omega_\ell}^d \Big]
{\nonumber} \\
&= 16 \pi \alpha_s
\biggl[ 3 + \dfrac{\omega_j^3 \!+\! \omega_k^3 \!+\! \omega_\ell^3 \!+\! \omega_j\omega_k\omega_\ell}{(\omega_j \!+\! \omega_k)(\omega_j \!+\! \omega_\ell)(\omega_k \!+\! \omega_\ell)}\biggr]
\Big[ (f^{ade}f^{bce}\! +\! f^{ace}f^{bde}) \mathcal{B}^a_{n+,\omega_i} \mathcal{B}^b_{\bar n +, \omega_j} \mathcal{B}^c_{\bar n +, \omega_k} \mathcal{B}_{\bar n -, \omega_\ell}^d {\nonumber} \\
& \qquad\qquad
- (f^{ade}f^{bce} + f^{abe}f^{cde} )\mathcal{B}^a_{n-,\omega_i} \mathcal{B}^b_{\bar n +, \omega_j} \mathcal{B}^c_{\bar n -, \omega_k} \mathcal{B}_{\bar n -, \omega_\ell}^d \Big]
\,.
\end{align}
We see that all the helicity selection rules are satisfied in the tree level matching, as expected. We have also checked the result using the automatic FeynArts \cite{Hahn:2000kx} and FeynRules implementation of the HiggsEffectiveTheory \cite{Alloul:2013bka}. For more complicated calculations at subleading power in SCET it would be interesting to fully automate the computation of Feynman diagrams involving power suppressed SCET operators and Lagrangians.
The four gluon operators derived in this section can be used to study $\mathcal{O}(\alpha_s^2)$ collinear contributions at $\mathcal{O}(\lambda^2)$. It would be interesting to understand in more detail the universality of collinear splittings at subleading power, as well as collinear factorization properties. For some recent work in this direction from a different perspective, see \cite{Stieberger:2015kia,Nandan:2016ohb}.
The behavior of these Wilson coefficients is also quite interesting. They exhibit a singularity as any pair of collinear particles simultaneously have their energy approach zero. This was also observed in the Wilson coefficients for operators describing the subleading collinear limits of two gluons emitted off of a $q\bar q$ vertex \cite{Feige:2017zci}.
\section{Conclusions}\label{sec:conclusions}
In this paper we have presented a complete basis of operators at $\mathcal{O}(\lambda^2)$ in the SCET expansion for color singlet production of a scalar through gluon fusion, as relevant for $gg\to H$. To derive a minimal basis we used operators of definite helicities, which allowed us to significantly reduce the number of operators in the basis. This simplification is due to helicity selection rules which are particularly constraining due to the scalar nature of the produced particle. We also classified all possible operators which could contribute to the cross section at $\mathcal{O}(\lambda^2)$. In performing this classification the use of a helicity basis again played an important role, allowing us to see from simple helicity selection rules which operators could contribute. While the total number of subleading power operators is large, the number that contribute at the cross section level is smaller. We compared the structure of the contributions to the case of a quark current, $\bar q \Gamma q$, finding interesting similarities, despite a slightly different organization in the effective theory.
A significant portion of this paper was devoted to a tree level calculation of the Wilson coefficients of the subleading power operators which can contribute to the cross section at $\mathcal{O}(\lambda^2)$. The Wilson coefficients obtained in this matching will allow for a study of the power corrections at NLO and for the study of the leading logarithmic renormalization group structure at subleading power. An initial investigation of the renormalization group properties of several subleading power operators relevant for the case of $e^+e^-\to \bar q q$ was considered in \cite{Freedman:2014uta}.
A number of directions exist for future study, with the goal of understanding factorization at subleading power. In particular, one would like to combine the hard scattering operators derived in this paper with the subleading SCET Lagrangians to derive a complete factorization theorem at subleading power for a physical event shape observable. Combined with the operators in \cite{Feige:2017zci}, all necessary ingredients are now available to construct such a subleading factorization for thrust for $\bar q q$ or $gg$ dijets in $e^+e^-$ collisions. This would also allow for a test of the universality of the structure of the subleading factorization. The operators of this paper can also be used to study threshold resummation, where power corrections of $\mathcal{O}((1-z)^0)$ have received considerable attention \cite{Dokshitzer:2005bf,Grunberg:2007nc,Laenen:2008gt,Laenen:2008ux,Grunberg:2009yi,Laenen:2010uz,Almasy:2010wn,Bonocore:2014wua,White:2014qia,deFlorian:2014vta,Bonocore:2015esa,Bonocore:2016awd}, particularly for the $q\bar q$ channel, but it would be interesting to extend this to the $gg$ case.
An interesting application of current relevance of the results presented in this paper is to the calculation of fixed order power corrections for NNLO event shape based subtractions. Gaining analytic control over power corrections can significantly improve the performance and stability of such subtraction schemes. This has been studied for $q\bar q$ initiated Drell Yan production to NNLO in \cite{Moult:2016fqy} using a subleading power operator basis in SCET (see also \cite{Boughezal:2016zws} for a direct calculation in QCD). Combined with the results for the operator basis and matching for $q\bar q$ initiated processes given in \cite{Feige:2017zci}, the operator basis presented in this paper will allow for the systematic study of power corrections for color singlet production and decay.
\begin{acknowledgments}
We thank the Erwin Schr\"odinger Institute and the organizers of the ``Challenges and Concepts for Field Theory and Applications in the Era of LHC Run-2'' workshop for hospitality and support while portions of this work were completed. This work was supported in part by the Office of Nuclear Physics of the U.S. Department of Energy under the Grant No.~DE-SC0011090, by the Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231, and the LDRD Program of LBNL. I.S. was also supported by the Simons Foundation through the Investigator grant 327942.
\end{acknowledgments}
|
\section{THE ENVELOPE-FUNCTION METHOD. INTRODUCTION TO THE HISTORY OF THE PROBLEM}
By the middle of the 20-th century, the problem arose of explaining and predicting the electronic properties of semiconductors in external fields. Various versions of the single-band effective-mass method were developed. The Luttinger--Kohn envelope-function (EF) method \onlinecite{Luttinger_1955} based on a generalization of the $\bm{kp}$-approach turned out to be very convenient. The formalism of EFs admitted a natural multiband generalization. Such a generalization was made by Keldysh in his theory of deep impurities \onlinecite{Keldysh_1964}. In [\onlinecite{Keldysh_1964}], Keldysh also noticed that, under certain conditions, EFs in a narrow-band semiconductor obey an effective Dirac equation (a system of four first-order differential equations). It is interesting that, in III--V semiconductors, this situation occurs under reversal of the sign of the strong spin--orbit splitting of the valence band. The idea of inversion of bands in crystals with strong spin--orbit interaction will be addressed below in this paper. In the same year as [\onlinecite{Keldysh_1964}], it was shown [\onlinecite{Wolff_1964}] that the spectrum of electrons and holes in bismuth near the $L$ points of the Brillouin zone should be described by an anisotropic Dirac Hamiltonian.
According to modern terminology, bismuth can be considered as a Dirac material, and the first Dirac material at that. These materials include graphene, bismuth-- antimony alloys, lead chalcogenides, 2D and 3D topological insulators, Dirac semimetals, Weyl semimetals, and many other materials. One-particle excitations in these materials are called massless (for gapless materials) or massive Dirac fermions (DFs).
The energy spectrum $E(p)$ of free DFs in the three-dimensional isotropic case is analogous to the spectrum of a relativistic electron:
\begin{equation}\label{Dirac_spectrum}
E(p)=\pm\sqrt{m^2c^4 + p^2c^2}
\end{equation}
Here $m$ is the electron effective mass, which is usually one or two orders of magnitude less than the mass of a free electron in vacuum, and $c$ is the effective velocity of light. This velocity is two orders of magnitude less than the velocity of light in vacuum; therefore, real Fermi excitations are, in fact, non-relativistic. However, relativistic corrections due to spin-orbit interaction in crystals with close bands may be rather large, which contributes to the formation of DFs. A finite mass in the spectrum of relativistic particles corresponds to a finite width of the forbidden band according to the well-known formula $E_g = 2mc^2$.
The 1960s marked the emergence of the physics of 2D electron systems. Quantum size effect was started to be used to realize the 3D $\to$ 2D transition. The conventional theory of size quantization is based on the single-band effective mass approximation with zero boundary conditions for EFs. Physically, this is justified by the impenetrability of potential barriers on the surfaces (interfaces) that confine the motion of an electron. This approximation works well in crystals with parabolic band spectrum, for example, for electrons in silicon. However, the situation is qualitatively changed when one tries to correctly describe the size quantization in a system of DFs.
The possibility of implementation of the size quantization was predicted in 1962 [\onlinecite{Sandom_1962}] for films made of narrow-band semiconductors and related semimetals with low concentration of carriers. The experimental discovery of this phenomenon [\onlinecite{Ogrin_1966}, \onlinecite{Lutskii_1966}] and its explanation [\onlinecite{Sandom_1967}] occurred a few years later when studying bismuth films.
Strictly speaking, envelope wave functions strongly differ from real wave functions of an electron in a crystal: the former functions are the envelopes of the latter. However, this fact was not of fundamental importance for qualitative considerations at the initial stage. But the development of experiment raised the question of quantitative description of the results. Unexpectedly, the problem turned out to be rather difficult. A seemingly academic question of the simplest boundary conditions for effective wave functions on the surface of a bismuth-type narrow-band semiconductor inevitably entailed the problem of description of SSs. It turned out that the simplest and physically obvious boundary conditions---zero boundary conditions for all four EFs in the Dirac equation---lead to a trivial (everywhere zero) solution due to the overdeterminacy of the problem. This is how the theoretical problem of derivation of correct boundary conditions for the envelope-function method arose, which is especially important in its multiband version. It is important that these boundary conditions should also describe the spectrum of SSs in addition to volume states.
The point is that, on the real surface of any crystal, including a Dirac crystal, there always exist electron or hole SSs, which are extrinsic and/or intrinsic. The first are attributed to defects and contamination of the surface, while the second (we will consider only these SSs) exist on the ideal surface. Intrinsic SSs were theoretically considered at the early stage of band theory in different models [\onlinecite{Tamm_1932_1,Tamm_1932_2,Maue_1935,Shockley_1939,Davison_1970}]. I.E. Tamm's works were the first ones; therefore, these SSs are called Tamm states. W. Shockley considered another popular model, and the SSs obtained by him are often called Shockley states. However, there is no essential difference between the Tamm and Shockley SSs. Moreover, in the simplest models, these states turn into each other under the variation of model parameters [\onlinecite{Davison_1970}], which may lead to confusion. Therefore, we will call intrinsic SSs Tamm--Shockley states, bearing in mind that both types of SSs are attributed to the sharp (at the atomic scale) discontinuity of the crystal potential at the surface.
The Tamm--Shockley states are not only situated (completely or partly) in the forbidden band of a bulk crystal. It is very important that these states should form a surface band of conducting states that are delocalized in the plane of the surface and are characterized by a 2D dispersion law. However, intrinsic SSs are very sensitive to surface roughness and contamination. Therefore, it is not surprising that a lot of time passed until the Tamm--Shockley states were reliably detected experimentally. As a rule, these states are investigated in ultra-high vacuum usually by local techniques (STM, ARPES, and so on). The problem of existence of an SS band under normal conditions requires special investigation. For instance, it is not quite clear how much stringent conditions should be imposed on the perfection of the surface or the interface, and what does this depend on.
In ordinary semiconductors (with wide forbidden gap and small spin--orbit interaction), a band of Tamm--Shockley states arises far from always, and the wave functions of these states are not usually mixed with the wave functions of size quantized states. The very existence conditions of these states strongly depend on the structure of the surface or the interface at the atomic scale; therefore, in spite of the 80-year history of the problem, the general principles of behaviour of these functions have been investigated rather fragmentarily, especially experimentally. The situation started to be changed in recent years, when topological materials described by a modified Dirac equation appeared. The theory of topological insulators predicts that "topological" SSs protected from backscattering should exist in these materials for topological reasons. The very fact of existence of these states does not depend on the details of the structure of a surface region.
A more general question arises: are the requirements imposed on the perfection of the surface (interface) relaxed for the manifestation of SSs on the surface of any Dirac crystal, not necessarily a topological insulator? The theoretical answer in the simplest Dirac model [\onlinecite{Volkov_Pinsker_1981_English}] is positive. Nevertheless, as expected for the Tamm--Shockley states, the spectrum of SSs for Dirac materials depends on the properties of the surface, although not in a quite obvious way.
To solve this problem, one had to derive boundary conditions of general form that are invariant under time reversal and describe DFs near an impenetrable wall. The solution of the genuine (rather than a modified, as in the theory of topological insulators later) Dirac equation in a half-space with these boundary conditions showed that SSs should appear on any surface (more precisely, for any parameters of the boundary conditions) and should be extremely strongly split with respect to spin as a result of spin--orbit interaction with the surface (this interaction is often called the Rashba interaction). These SSs are characterized by a conical dispersion law but they occurred not all the values of momenta. The "strength" of surface spin--orbit interaction is characterized by a real phenomenological parameter $a_0$ that appears in the boundary conditions. Depending on the sign of $a_0$, there exist two classes of surfaces. For the surfaces of one of the classes, the conical ("Dirac") point in the spectrum is located in the forbidden gap of the bulk material; i.e., SSs in this case have a 2D spectrum of massless spin-nondegenerate DFs, as in the case of a topological insulator. The physical meaning of the sign of $a_0$ remained unclear.
In 1985, in the famous work [\onlinecite{Volkov_Pank_1985_English}], the authors considered another Dirac model: the model of inverse heterocontact. It is described by the Dirac equation in which the mass of a DF is varied in space smoothly (at the atomic scale) and symmetrically with respect to the electron--hole (e--h) transformation. The interface is determined by the position of the plane in which the sign of the mass is reversed (the conduction and valence bands are interchanged). It is this case where the band of nondegenerate heterointerface states possessing a conical 2D spectrum of massless DFs is formed. It is remarkable that this conclusion does not depend on the details of the interface potential; one just needs the inversion of the bands. The point is that, under these conditions, the equation for DFs in the inverse contact takes a form typical for supersymmetric quantum mechanics with zero mode corresponding to interface states. This nontrivial result is actively used in modern physics of topological insulators. The Dirac point in this symmetric model is located exactly at the center of the gap.
The reason for some disagreement between the results of [\onlinecite{Volkov_Pank_1985_English}] and [\onlinecite{Volkov_Pinsker_1981_English}] and the physical meaning of the sign of $a_0$ were elucidated in the survey [\onlinecite{Volkov_review_1995_English}]. A significant e--h asymmetry was introduced into the model of an abrupt heterocontact (which, however, is still smooth at the atomic scale). As this asymmetry increases, the Dirac point was more and more shifted from the center of the gap and disappeared. For a greater discontinuity of bands at the heterojunction, the spectrum of interface states completely corresponds to the spectrum of SSs from [\onlinecite{Volkov_Pinsker_1981_English}]. The comparison of the results yielded a model expression for the boundary parameter $a_0$. The sign of this parameter generally correlated to the reversal of the sign of the gap at the interface.
Unfortunately, the incipient state of the manufacturing technology of samples did not allow one to verify these conclusions experimentally 30--20 year ago. The works [\onlinecite{Volkov_Pinsker_1981_English,Volkov_Pank_1985_English,Volkov_review_1995_English}] acquired importance in relation to the discovery of topological insulators, as well as in relation to the recently detected conducting edge states of non-topological type in nanoperfortated graphene [\onlinecite{Latyshev_2013}, \onlinecite{Latyshev_SciRep_2014}].
In the present study, we discuss the problems touched upon above and, within the formalism of EFs, formulate a minimal model that describes SSs of various, topological and non-topological, types in a number of Dirac materials. The paper is organized as follows. In Section 2, we give a brief survey of works on Tamm--Shockley-type SSs within the formalism of EFs. In Section 3, we consider topological SSs and, in Section 4, Tamm--Shockley edge states in graphene. In Section 5, we formulate conclusions.
\section{TAMM--SHOCKLEY-TYPE SURFACE STATES}
The method of effective wave-functions---envelope functions---is widely used to describe the behavior of electrons in multilayer semiconductor structures. The method of EFs can be applied to the description of smooth (at the atomic scale) fields and is inapplicable to the real case of atomically abrupt interfaces. Information on the microscopic structure of the interface can be taken into account by appropriate boundary conditions for EFs.
The problem of boundary conditions in bounded crystals has a long history. Theoretical studies on boundary conditions in semiconductor structures can be classified into two groups. The works of the first, largest, group are devoted to the derivation of "two-sided" boundary conditions relating EFs to their left and right derivatives at the interface. They contain different approaches to the solution of mathematical problems related, in particular, to possible singular behaviour of EFs on a heterointerface [\onlinecite{Ivchenko2005,Zawadzki2004,Foreman_2005,Takhtamirov_1999,Rodina_2002,
Takhtamirov_1997,Takhtamirov_2010}].
We will focus only on the works of the second group, which are devoted to the derivation of one-sided boundary conditions at the crystal--high barrier interface (in particular, at the crystal--vacuum interface). The problems of this type arise when describing the Tamm--Shockley surface (interface) states.
Under the neglect of the spin--orbit interaction, a microscopic derivation of boundary conditions for EFs at a step-like boundary between semiconductor ($z > 0$) and vacuum ($z < 0$) was apparently first presented in [\onlinecite{Volkov1976}, \onlinecite{Volkov1977}]. The boundary conditions introduced there contain boundary parameters that are analytically (but in a complicated way) expressed in terms of a complete infinite-band structure of a semiconductor analytically continued to the domain of complex quasimomenta. The numerical determination of these parameters remains an unsolved problem.
\section{SHALLOW TAMM--SHOCKLEY STATES IN A SINGLE-BAND APPROXIMATION}
In a single-band approximation, a boundary condition represents a linear relation between an EF and its normal derivative with a single boundary parameter with the dimensionality of length, which we denote below by $R$. This length characterizes the structure of the interface between semiconductor ($z > 0$) and an impenetrable barrier ($z < 0$) at the atomic scale and has the meaning of the localization depth of a shallow Tamm--Shockley state when it exists (to this end, the condition $R > 0$ should be satisfied). Moreover, the length $R$ depends on the parameters of the bulk band structure.
In [\onlinecite{Volkov_1979}], a much simpler derivation of the same boundary condition is presented from the Hermiticity of the effective Hamiltonian for EFs in a half-space bounded by an impenetrable barrier. Within such a phenomenological approach, the parameter $R$ should be determined from experiment. The high-barrier model is applicable when the interface length $R$ is large compared with the penetration depth into the barrier. The effect of the spin--orbit interaction on single-band boundary conditions and spin splitting of shallow Tamm--Shockley states in the conduction band of a semiconductor with inversion symmetry was considered in [\onlinecite{Vasko1979}]. In the model used in [\onlinecite{Vasko1979}], the spin splitting is controlled by the product of the length $R$ and the parameter of bulk spin--orbit interaction. A non-parabolic generalization of boundary conditions [\onlinecite{Vasko1979}] in an asymmetric quantum well with infinite barriers is presented in [\onlinecite{Rodina_2006}].
We begin with the analysis of a single-band boundary condition as applied to heterostructures based on III--V semiconductors with the heterointerface orientation (001). In [\onlinecite{Devizorova_2013},\onlinecite{Devizorova_2014}], the authors considered the effect of an atomically abrupt heterointerface on the effective single-band Hamiltonian of a quantum well and the spin splitting in a conduction band of symmetry $\Gamma_{6c}$ in crystals with lack of inversion symmetry. The discontinuity of bands at the heterointerface is assumed to be large, and the heterobarrier, to be impenetrable. The latter is characterized by a certain boundary condition for EFs.
The dynamics of a conduction electron for $z > 0$ within the multiband EF method is described by the system of Kohn--Luttinger $\bm{kp}$-equations
\begin{eqnarray}\label{kp_eqs}
\left\{ \left[E_n(0)+\frac{\hat{\bm{p}}^2}{2m_0}+V(z)\right]\delta_{nn'} + \frac{\hat{\bm{p}}\cdot\bm{p}_{nn'}}{m_0} \right. \nonumber \\
+ \left. \frac{\hbar}{4m_0^2c^2}\left(\bm{p}\cdot\left[\bm{\sigma}\times\bm{\nabla}V_0\right]\right)_{nn'} \right\}\Phi_{n'}=E\Phi_n
\end{eqnarray}
where $n$ is the band number, $E_{n}(0)$ is the extremum energy of the $n$-th band, $\Phi_n$ is a set of EFs, $\bm{p}_{nn'}$ is a matrix element of the momentum operator $\hat{\bm{p}}$ on Bloch functions at the center of the Brillouin zone, $m_0$ is the mass of a free electron, the last term in curly brackets is a matrix element of the operator of bulk spin--orbit interaction on Bloch functions, and $\sigma_i$, $i = x, y, z,$ are Pauli matrices.
The Hermiticity of Hamiltonian (\ref{kp_eqs}) in the half-space $z > 0$ reduces, after integration by parts, to vanishing of the surface contribution. This is equivalent to setting the matrix element of the normal component of the current operator between any pair of states at the boundary to zero:
\begin{equation}\label{normal_current}
\left.\left(\Phi^{+}_{\lambda}\hat{v}_z\Phi_{\nu}\right) \right|_{z=0}=0
\end{equation}
where $\hat{v}_z$ is a non-diagonal matrix of velocity $\left (\hat{v}_z\right)_{nn'}= \partial_{p_z} \left(H_{nn'}\right)$.
Since only the spinor corresponding to the conduction band $\Gamma_{6c}$ is large in the multicomponent function $\Phi$, we apply the unitary transformation $\Phi = e^S\phi$ [\onlinecite{Winkler}] (with regard to $\bm{kp}$ terms up to the third order inclusively), which reduces the Hamiltonian to a single-band one with effective mass $m*$ and smooth potential $V(z)$. Now, the 3D Hamiltonian of the conduction band contains contributions $\hat{H}_{BIA}$ and $\hat{H}_{SIA}$ that describe the spin splitting due to the lack of inversion symmetry of the crystal and the asymmetry of the well:
\begin{equation}\label{hamiltonians}
\hat{H} = \frac{\hat{\bm{p}}^2}{2m^*}+V(z)+\hat{H}_{BIA}+\hat{H}_{SIA},
\end{equation}
\begin{eqnarray}
\hat{H}_{BIA} =\frac{\gamma_c}{\hbar^3}\left[ \sigma_xp_x\left(\hat{p}_y^2 - \hat{p}_z^2\right)\right. \qquad\qquad\qquad\quad\nonumber \\
+\sigma_yp_y\left(\hat{p}_z^2 - \hat{p}_x^2\right) + \sigma_zp_z\left(\hat{p}_x^2 - \hat{p}_y^2\right),
\end{eqnarray}
\begin{equation}
\hat{H}_{SIA} = a_{SO} \left(\sigma_xp_y - \sigma_y p_x\right)\partial_zV(z).
\end{equation}
By the same transformation, condition (3) reduces to a certain constraint for a two-component EF. Next, we require the invariance of this constraint with respect to the operation of time reversal
\begin{equation}\label{time_inversion}
\hat{T} = i\sigma_y \hat{K},
\end{equation}
where $\hat{K}$ is the operator of complex conjugation. We obtain T-invariant boundary conditions that take into account the spin--orbit interaction in the bulk and on the symmetry interface $C_{2v}$, as well as the absence of the inversion center in the bulk crystal:
\begin{eqnarray}\label{single_band_BC}
\left[\sigma_0 - i\frac{R\hat{p}_z}{\hbar} - i \frac{2m^*\gamma_c R}{\hbar^4}\left(\sigma_yp_y - \sigma_xp_x\right)\hat{p}_z\right. \nonumber \\
- i \frac{m^*\gamma_c R}{\hbar^4}\sigma_z\left( p_x^2 - p_y^2\right) + \frac{\left(\chi + \chi^{int}\right)R}{\hbar}\left(\sigma_xp_y - \sigma_y p_x\right) \nonumber\\
\left.\left.- \frac{2m^*\gamma_c^{int}}{\hbar^3}\left(\sigma_yp_y - \sigma_x p_x\right) \right]\phi\right|_{z=0}=0.\qquad
\end{eqnarray}
Here $\sigma_0$ is the identity matrix. The real quantity $R$, which has the dimensionality of length, depends on the microscopic structure of the boundary. The constants $\gamma_c$ and $\chi$ are defined by the volume parameters; for GaAs, $|\gamma_c| = 24.4$ eV$\cdot {\rm A}^3$ and $\chi = 0.082$. The constants $\gamma_c^{int}$ and $\chi_c^{int}$ characterize the spin--orbit interaction with the interface crystal potential and are determined from a comparison with experiment [\onlinecite{Devizorova_2014}].
The Tamm problem in the half-space $z > 0$ in this case corresponds to the solution of a single-band effective mass equation with Hamiltonian (\ref{hamiltonians}) and boundary conditions (\ref{single_band_BC}). It is important that such an approach leads to the discontinuity of single-band EFs on the interface. This is attributed to the non-perturbative effect of the interface potential. Therefore, the approach allows one to describe, for example, shallow Tamm--Shockley states, if they exist, even in the single-band approximation [\onlinecite{Volkov1976,Volkov1977,Volkov_1979}]. Unfortunately, this situation is not realized for electrons of the conduction band at the (001) GaAs/AlGaAs hererojunction because $R < 0$. A quantitative comparison with precision measurements of electron spin resonance that are sensitive to the parameters of boundary conditions shows [\onlinecite{Devizorova_2014}] that $R = -2.2$ nm.
\begin{figure}
\includegraphics{fig1n.eps}
\caption{Evolution of the 2D spectrum of Tamm--Shockley SSs for increasing interface spin--orbit interaction: from (a) spin-degenerate shallow SSs through (b) weak spin--orbit splitting to (c) a conical spectrum of topological-type SSs. Bulk states are shown in gray. }
\end{figure}
On the surface or at the interface where $R > 0$, a band of shallow SSs [\onlinecite{Volkov_1979}] is formed that is described, with neglect of the spin--orbit interaction, by the first two terms in the boundary conditions (\ref{single_band_BC}) (Fig. 1a). These SSs are "suspended" over the bottom of the conduction band, and their localization depth is equal to $R$. The volume and interface contributions to the spin--orbit interaction, which are described by the remaining terms in (\ref{single_band_BC}), anisotropically split the band of SSs with respect to spin (more precisely, with respect to chirality---the projection of spin onto the direction related to momentum) (Fig. 1b). This splitting was first introduced in [\onlinecite{Vasko1979}] in the simplest isotropic model. As the interface spin--orbit interaction increases, the splitting increases and becomes comparable to the bandgap $E_g$ (Fig. 1c), and the single-band approximation fails.
\section{TAMM--SHOCKLEY STATES IN THE TWO-BAND APPROXIMATION}
The two-band model is a minimal multi-band model. For Dirac materials such as bismuth, lead chalсogenides, and many other, the Tamm problem in the simplest approximations reduces to the solution of the Dirac equation in the half-space $z > 0$,
\begin{equation}\label{Dirac_eq}
\left(
\begin{array}{cc}
mc^2 - E & c \bm{\sigma}\cdot\bm{k} \\
c \bm{\sigma}\cdot\bm{k} & -mc^2 - E
\end{array}
\right)
\left(
\begin{array}{cc}
\Psi_c \\
\Psi_v
\end{array}
\right)=0
\end{equation}
with the boundary conditions [\onlinecite{Volkov_Pinsker_1981_English}]
\begin{equation}\label{Dirac_BC}
\left[\Psi_v - ia_0\bm{\sigma}\cdot \bm{n}\Psi_c\right]_S=0.
\end{equation}
Here the two-component EFs $\Psi_c$ and $\Psi_v$ describe the conduction and valence bands, $a_0$ is a real phenomenological parameter describing the small-scale structure of the surface, $\bm{n}$ is a unit normal to the surface $S$ of the crystal, and $E_g = 2mc^2$ is the bulk band gap. The boundary conditions (\ref{Dirac_BC}) follow from the Hermiticity (self-adjointness) of the Dirac Hamiltonian in the half-space and the symmetry of the problem with respect to time reversal.
\begin{figure}
\includegraphics{fig2n.eps}
\caption{Energy spectrum of the 3D Dirac equation in a half-space. The band of non-degenerate Tamm--Shockley states is shown in red, and bulk states are shown in light gray. There are two classes of surfaces numbered by the sign of the boundary parameter $a_0$. (a) An analog of a topologically nontrivial spectrum. For a surface with positive $a_0$, SSs fill the main part of the cone with a Dirac point in the gap of a bulk material, as in a topological insulator. (b) An analog of a topologically trivial spectrum. As the sign of $a_0$ is changed, SSs are expelled from the bulk gap and fill the peripheral part of the cone. (c, d) The spectrum of a single-valley Dirac semimetal is obtained from the spectrum of the Dirac equation for zero mass. The massless limit of the Tamm problem for two classes of surfaces is shown in Figs. 2d and 2c; Figures 2a and 2b turn into Figs. 2d and 2c, respectively. }
\end{figure}
A perturbative account of remote bands gives rise to diagonal corrections to the Dirac Hamiltonian in (\ref{Dirac_BC}) that are quadratic in momentum. The corresponding band gap model, sometimes called Dimmock's model, is used in the theory of topological insulators (see below). In this section, the contribution of remote bands is neglected.
As a result of the strong spin--orbit interaction, the Tamm problem for DFs always gives rise to a non-degenerate spectrum of SSs that fills a part of the cone surface (Fig. 2). It is interesting to compare this result with the theory of topological insulators and semi-metals. The result qualitatively depends on the sign of the boundary parameter $a_0$.
For a surface with a positive value of $a_0$, a conical (''Dirac'') point in the spectrum lies in the forbidden gap of a bulk single-valley material; i.e., in this case, SSs are characterized by a 2D spectrum of massless DFs (Figs. 1c and 2a). The comparison with topological theory implies that this case is an analog of a topologically nontrivial case. Figure 2b corresponds to the topologically trivial phase, for which the parameter $a_0$ is negative. In the massless limit, we obtain SSs for Dirac's single-valley 3D semimetal (Figs. 2c, 2d).
\begin{figure}[h]
\includegraphics{fig3n.eps}
\caption{Dirac model of a symmetric heterocontact [\onlinecite{Volkov_Pank_1985_English}] in the case of reversal of the sign of interband gap is described by equations of supersymmetric quantum mechanics; the zero mode of these equations corresponds to interface states with conical spectrum (Fig. 1c) that are stable with respect to perturbations. }
\end{figure}
Analytically, the spectrum of SSs for the Dirac equation in the half-space is expressed as
\begin{equation}\label{SS_spectrum}
E_s(\bm{k}_{||})=s\frac{2a_0}{1+a_0^2}|\bm{k}_{||}|+E_0.
\end{equation}
Here the quantum number $s = \pm 1$ describes the chirality, $\bm{k}_{||} = (k_x, k_y, 0)$. The energy of the conical point is measured from the center of the gap and has the form
\begin{equation}\label{Dirac_point}
E_0=mc^2\frac{1-a_0^2}{1+a_0^2}.
\end{equation}
The branches of the spectrum (\ref{SS_spectrum}) with different chirality are implemented under the condition
\begin{equation}\label{SS_exist}
\frac{2a_0}{1+a_0^2}\frac{mc}{\hbar} - s |\bm{k}_{||}|\geq 0.
\end{equation}
Another approach to the theory of SSs is related to the formulation of models whose bandgap parameters smoothly (at the atomic scale) vary near the interface ("heterocontact"). In this case, the multi-band method of EFs can be applied in the whole space, and no boundary condition problem arises. In [\onlinecite{Volkov_Pank_1985_English}], the authors considered interface states of DFs in the model of a symmetric inverse heterocontact. These states are described by the Dirac equation in which the mass of DFs is inverted in space, the conduction and valence bands being interchanged (Fig. 3).
Near the inversion point, a 2D band of SSs arises. The equation for DFs at the inverse junction takes a form characteristic of supersymmetric (SUSY) quantum mechanics:
\begin{equation}\label{Vol_Pank_eq}
\left(
\begin{array}{cc}
-E & i\dfrac{E_g(z)}{2} + c\bm{\sigma}\cdot\bm{k} \\
-i\dfrac{E_g(z)}{2} + c\bm{\sigma}\cdot\bm{k} & -E
\end{array}
\right)
\left(
\begin{array}{cc}
\Psi_c \\
\Psi_v
\end{array}
\right)=0
\end{equation}
The zero mode of this equation describes interface states,
\begin{equation}\label{SUSY_solution}
\Psi_{\pm} =
\left(
\begin{array}{c}
0 \\
e^{i\theta/2}\\
\pm e^{i\theta/2}\\
0
\end{array}
\right) \exp\left\{\int_0^z \frac{E_g(z)}{2\hbar c}dz+ i\bm{k}_{||}\bm{r}_{||} \right\},
\end{equation}
where $e^{i\theta}=(k_x + i k_y)/k_{||}$. Solution (\ref{SUSY_solution}) is stable against to perturbations of the function $E_g(z)$ in view of SUSY. The Dirac point in this symmetric model is located precisely at the center of the gap, and the spectrum of SSs is actually described by formulas (\ref{SS_spectrum}) and (\ref{Dirac_point}) in which $a_0 = 1$.
\begin{figure*}[!]
\includegraphics{fig4n.eps}
\caption{Model of an asymmetric inverse heterocontact [\onlinecite{Volkov_review_1995_English}]. For large asymmetry, interface states cease to exist in spite of the inversion of bands. }
\end{figure*}
The model of an asymmetric contact considered in the survey [\onlinecite{Volkov_review_1995_English}] is more realistic (see Fig. 4). The asymmetry of bands is defined by the scalar potential $\varphi(z)$ that have the meaning of the bending of bands.
For a greater discontinuity of bands, the spectrum of interface states is given by
\begin{equation}\label{Volk_Pank_SS_spectra}
E=-\frac{E_{g1}\varphi_0}{E_{g0}} \pm |\bm{k}_{||}|\sqrt{1 - \left(\frac{2\varphi_0}{E_{g0}}\right)^2}
\end{equation}
where $\varphi =\varphi_0 f(z) $, $E_g(z)=E_{g0}f(z)+E_{g1}$, and the model function $f(z)$ describes a correlated variation of the band gap and the bending of bands at the heterointerface.
It makes sense to compare the results of [\onlinecite{Volkov_Pinsker_1981_English}] and [\onlinecite{Volkov_review_1995_English}]. The heterointerface spectrum (\ref{Volk_Pank_SS_spectra}) turns into the spectrum of SSs (\ref{SS_spectrum}) under the change
\begin{equation}
a_0 = -{\rm sign}\left(E_{g0}\right)\sqrt{\frac{E_{g0}-\varphi_0}{E_{g0}+\varphi_0}}
\end{equation}
This demonstrates the physical meaning of the boundary parameter $a_0$: its sign correlates with the sign of the gap, and the difference of its amplitude from 1 is the measure of the e--h asymmetry. An analog of the topologically non-trivial phase in Fig. 2a corresponds to the inversion of bands. An important consequence of (\ref{Volk_Pank_SS_spectra}) is as follows: in general, the inversion of bands is insufficient for the emergence of supersymmetric interface states. For higher e--h asymmetry, when $\varphi_0/E_{g0}>1$, these states cease to exist.
\begin{figure}[!]
\includegraphics{fig5n.eps}
\caption{A wire with inhomogeneous cross section in a longitudinal magnetic field. The amplitude of a supersymmetric SS that is implemented for half-integer values of a local magnetic flux (in units of flux quanta) passing through the wire cross section $z = z_0$. }
\end{figure}
In principle, the conditions for the emergence of SUSY states can be controlled by external fields. For example, consider [\onlinecite{Enaldiev_arxiv}] a solution to the Dirac equation (\ref{Dirac_eq}) with the boundary conditions (\ref{Dirac_BC}) for a cylindrical nanowire of radius $R_0$ in a longitudinal magnetic field $B$. The spectrum of SSs is controlled by magnetic flux $\Phi=\pi B R_0^2$ through a section of the wire. For the positive sign of the boundary parameter $a_0$ in a strong field, the quasi-classical spectrum of SSs is given by
\begin{equation}\label{SS_wire}
E=\pm\frac{2a_0c\hbar}{1+a_0^2}\sqrt{k_z^2+\frac{(j+\Phi/\Phi_0)^2}{R^2}}+E_0,
\end{equation}
where $k_z$ is a projection of quasi-momentum to the wire axis, $j$ are half-integer numbers, and $\Phi_0 = hc/e $ is the flux quantum. When the magnetic flux is a half multiple of the flux quantum, the spectrum (\ref{SS_wire}) becomes gapless. For any sign of $a_0$, the contribution of SSs to the total density of states oscillates as a function of the magnetic flux with period $\Phi_0$. This leads to Aharonov--Bohm-type oscillations of the wire resistance.
More interesting is the case of a wire with inhomogeneous cross section in a uniform longitudinal magnetic field. Magnetic flux through the section of a wire is inhomogeneous, whereby local SUSY states can be formed (Fig. 5).
In [\onlinecite{Enaldiev_arxiv}], it was found that such a SUSY solution with energy $E_0$ has the form
\begin{equation}\label{SUSY_wire}
\Psi \propto \exp\left( \frac{1}{\hbar v}\int_{z_0}^{z}\Delta(z)dz\right),
\end{equation}
where $v=2a_0c/(1+a_0^2)$ is the velocity of surface states. The effective local gap $\Delta(z)=\hbar v \left(j+ \Phi(z)/\Phi_0\right)/R(z)$ is inverted at the point $z = z_0$; therefore, solution (\ref{SUSY_wire}) is localized precisely at this point.
\section{TOPOLOGICAL SURFACE STATES}
The theory of band structure of semiconductors involving topological arguments, which were first introduced in the theory of the quantum Hall effect [\onlinecite{Thouless_1982}], gave rise to a new physical concept---a topological insulator (TI) [\onlinecite{Hasan_Kane_TI_review},\onlinecite{Qi_Zhang_review}]. In topological insulators with a strong spin--orbit interaction, the specific features of the bulk band structure are such that a band of conducting surface (or edge) states of the type shown in Figs. 1c and 2a should exist on the surface of TIs for topological reasons. This is the most fundamental prediction of topological theory. Surface DFs are non-degenerate with respect to the spin quantum number (chirality) and have a conical dispersion law, which leads to the weakening of their scattering by non-magnetic impurities. It is proved that this property is a specific feature of precisely topological surface or edge states.
To implement a TI phase, one usually should impose three qualitative conditions on the parameters of the bulk band structure: symmetry with respect to time reversal (T symmetry), strong spin--orbit interaction, and inversion of bands. As a result, topological SSs arise in the bulk gap; according to topological theory, the very fact of emergence of these states does not depend on the details of the structure of the near-surface region. The possibility of implementation of 2D TI [\onlinecite{Kane_2005}] and 3D TI [\onlinecite{Fu_Kane_2007}] phases was predicted in 2005 and 2007, respectively.
First, consider a 1D band of edge states in a 2D crystal system. The topological invariant $Z_2$ is defined by the properties of the anti-unitary matrix $w_{nm}=\langle u_n(\bm{k})|\hat{T}|u_m(-\bm{k})\rangle$, which is composed of the matrix elements of the time-reversal operator on Bloch factors of all occupied bands [\onlinecite{Hasan_Kane_TI_review}]. The T symmetry allows one to classify all 2D systems into two groups that differ by the value of the $Z_2$ invariant $\nu = \left\{0, 1\right\}$ [\onlinecite{Kane_2005}]. A crystal is said to be in the phase of a topological (trivial) insulator if the Hamiltonian characterizing it has $\nu = 1$ ($\nu = 0$). A TI phase is distinguished by the existence of an odd number of Kramers pairs of edge states at the Fermi level in the bulk gap. It is important that any T-invariant perturbations of the Hamiltonian of the crystal that do not close the gap in the spectrum cannot change the value of the $Z_2$ invariant of the system. Topological protection of SSs in a TI is understood precisely in this sense.
\begin{figure}[ht]
\includegraphics{fig6n.eps}
\caption{Scheme of evolution of the bulk spectrum of a narrow-gap crystal for increasing bulk spin--orbit interaction leading to the inversion of bands: from (a) a trivial insulator phase through (b) a gapless phase to (c) the inverted spectrum in the phase of a topological insulator. (d) Bandgap model of a semi-infinite Bi$_2$Se$_3$-type topological insulator. The curvature of the spectrum of bulk bands is contributed from the dispersion the mass term $M(\bm{k})$. The states in the gap illustrate the spectrum of topological SSs. }
\end{figure}
In the theoretical study [\onlinecite{Bernevig_2006}], the authors predicted that a 2D TI phase should be implemented in a CdTe--HgTe double heterostructure for the width of the HgCdTe quantum well greater than 6.4 nm. It is these conditions under which the electron and hole size-quantized subbands are inverted.
A two-component EF describing the states in a quantum well with two different values of the spin quantum number satisfies a modified Dirac equation (the so-called BHZ model [\onlinecite{Bernevig_2006}]):
\begin{widetext}
\begin{equation}\label{BHZ}
\left(
\begin{array}{cccc}
\varepsilon(k)+m(k) & A(k_x - ik_y) & 0 & 0 \\
A(k_x + ik_y) & \varepsilon(k)-m(k) & 0 & 0 \\
0 & 0 & \varepsilon(k)+m(k) & -A(k_x + ik_y) \\
0 & 0 & -A(k_x - ik_y) & \varepsilon(k)-m(k)
\end{array}
\right)
\left(
\begin{array}{c}
\Psi_{\uparrow} \\
\Psi_{\downarrow}
\end{array}
\right)=
E\left(
\begin{array}{c}
\Psi_{\uparrow} \\
\Psi_{\downarrow}
\end{array}
\right)
\end{equation}
\end{widetext}
Here $\varepsilon(k)= C -D(k_x^2 + k_y^2)$, $m(k)= M - B(k_x^2 + k_y^2)$,
$A, B, C, D$, and $M $are the parameters of the system, and $k_x$ and $k_y$ are the components of 2D momentum. The inversion of bands corresponds to a positive value of the product of parameters $MB > 0$.
The Tamm problem in the half-plane $x > 0$ in the model of zero boundary conditions gives rise to spin-split edge states with linear dispersion. Recently, 1D conducting edge channels associated with these states have been observed experimentally [\onlinecite{Konig_2007},\onlinecite{Nowack_2013}]. However, the spectrum of edge states has not been measured.
In [\onlinecite{Enaldiev_2015}], the authors considered the problem of sensitivity of edge states to the form of nonzero boundary conditions. They derived a general T-invariant boundary condition for the EF $\Psi=\left(\Psi_{\uparrow},\Psi_{\downarrow}\right)^T$ satisfying Eq. (\ref{BHZ}):
\begin{equation}\label{Zag_BC}
\left[F\partial_{\bm{n}}\Psi + G\Psi\right]_{\Gamma}=0
\end{equation}
where $G$, a matrix that mixes different components of the EF, is defined by six real phenomenological parameters and the matrix $F$ is defined by the bulk parameters $B, A$, and $D$; this matrix is introduced into the boundary condition (\ref{Zag_BC}) for convenience. The spatial symmetry reduces the number of independent phenomenological parameters in the matrix $G$ to four ones. The authors showed that the difference of boundary conditions from zero conditions significantly changes the spectrum of edge states and, generally speaking, can even lead to the vanishing of these states from the region of small quasi-momenta, where model (\ref{BHZ}) is applicable.
Note the following important feature of this system. Although non-zero values of the parameters $B$ and $D$ are assumed to be important for topological classification, they make a small contribution to the energy of size-quantized subbands for actual quantum wells of Hg(Cd)Te [\onlinecite{Bernevig_2006}]. Therefore, one can neglect these parameters in the Tamm problem, and Eq. (\ref{BHZ}) reduces to a 2D version of the Dirac equation. For the latter equation, the spectrum of the Tamm problem is defined by formula (\ref{SS_spectrum}) in which $\bm{k}_{||}$ is an conserving projection of quasimomentum to the 1D boundary of the 2D system considered. In this case, information on the inversion of bands is contained in the sign of the boundary parameter $a_0$ rather than in the bulk parameters of the BHZ equation (\ref{BHZ}).
The topological classification is complicated as the dimensionality increases. For a 3D TI, one usually defines four topological invariants ($\nu_0;\nu_1,\nu_2,\nu_3$), each of which can take values 0 or 1. A 3D crystal characterized by $\nu_0=1$ is called a strong TI. In the simplest case, a gapless 2D band of non-degenerate massless DFs is formed on the surface of a strong 3D TI. 3D crystals with $\nu=0$ are called weak TIs for which the appearance of SSs in the forbidden band depends on both the values of $\nu_{1,2,3}$ and the translational symmetry of the surface [\onlinecite{Fu_2007}]. In the general case, weak TIs have no gapless SSs.
It has been theoretically shown that a whole class of Bi$_2$Se$_3$-type compounds with inversion symmetry should belong to the TI phase. In this case, it is the strong spin--orbit interaction that leads to the inversion of two bands with different parity that are nearest to the Fermi level (Fig. 6). The existence of SSs in Bi$_2$Se$_3$, Bi$_2$Te$_3$, and Sb$_2$Te$_3$ compounds was experimentally established in [\onlinecite{Hsieh_PRL_2009},\onlinecite{Hsieh_Hasan_Nature_2009}]. A standard 3D TI for ARPES measurements is the crystalline compound Bi$_2$Se$_3$ in which the gap is about 0.3 eV. By this method, it was established [\onlinecite{Hsieh_PRL_2009}] that on the (111) surface there exist non-degenerate SSs with massless Dirac spectrum in the gap of the crystal.
In the approximation of EFs, these compounds are described near the center of the Brillouin zone by a modified Dirac equation, which, with neglect of the anisotropy and the e--h asymmetry, has the form
\begin{equation}\label{3D_TI}
\left(
\begin{array}{cc}
M(\bm{k}) - E & v\bm{\sigma}\cdot\bm{k} \\
v\bm{\sigma}\cdot\bm{k} & -M(\bm{k}) - E
\end{array}
\right)
\left(
\begin{array}{cc}
\Psi_c \\
\Psi_v
\end{array}
\right)=0
\end{equation}
where the significantly dispersive mass term $M(\bm{k}) = M_0 - B\bm{k}^2$ distinguishes this equation from the true Dirac equation. A topological phase is characterized by the positive value of the product of parameters $M_0B > 0$, which is attributed to the inversion of bands. The break of the e--h symmetry is described by the introduction of an additional diagonal contribution to Eq. (\ref{3D_TI}). We will neglect this asymmetry.
To describe SSs within Eq. (\ref{3D_TI}), one should specify boundary conditions that take into account the atomic-scale structure of the surface. To this end, one often uses the fact that the above-mentioned modification of the Dirac equation doubles the order of the system of differential equations. Therefore, it is simplest to use zero boundary conditions for all components of EFs [\onlinecite{Liu_Zhang_PRB_2010}, \onlinecite{Shan_2010}]: $\left[\Psi_c = \Psi_v\right]_S = 0$. Such boundary conditions, which are often called ''open'' do not take into account the details of the microscopic structure of the surface at the atomic scale, but nevertheless guarantee the existence of massless DFs on a plane surface.
However, it was demonstrated that non-zero boundary conditions may significantly affect the spectrum of SSs in the case of 3D TIs as well [\onlinecite{Enaldiev_2015}, \onlinecite{Menshov_2014}]. General boundary conditions for a four-component EF mix the values of the EF with its normal derivative on the boundary $S$:
\begin{equation}
\left[\frac{B}{M_0}\partial_{\bm{n}}\Psi+Q\Psi\right]_S=0
\end{equation}
where the $4\times 4$ matrix $Q$ should be determined from additional considerations. The T symmetry, along with the spatial symmetry, reduces the number of independent real phenomenological parameters in the matrix to three ones. For sufficiently strong surface spin--orbit interaction described by these parameters, SSs may disappear from the forbidden gap in the neighbourhood of the point $\Gamma$ at which the EF method is applicable [\onlinecite{Enaldiev_2015}].
Experiments have shown that the properties of SSs are sensitive to the perturbations of the surface potential, which can be taken into account by an appropriate choice of the parameters of the boundary conditions. For example, the doping of the surface with non-magnetic and magnetic atoms shifts the position of the Dirac point and modifies the dispersion of SSs [\onlinecite{Valla_2012, Scholz_2012, Bianchi_2011, Roy_2014}].
Another Dirac system is given by a topological crystalline insulator (TCI). In a TCI, the protection of SSs is guaranteed by two symmetries at once: the T symmetry and the spatial symmetry of the crystal and its surface [\onlinecite{Fu_2011}]. A topological classification is performed similar to the above-considered case of TIs, with the difference that the matrix $w_{nm}$ is constructed on the operator $\hat{U}\hat{T}$, where $\hat{U}$ is the operator of spatial symmetry.
Lead chalcogenides and Pb$_{1-x}$Sn$_x$Te(Se) solid solutions pass to the TCI phase at certain values of the concentration of tin when the inversion of bands of different parities occurs at the $L$ points [\onlinecite{Hsieh_NatCom_2012, Xu_2012, Dziawa_2012, Egorova_2015}]. The role of spatial symmetry in these compounds is played by mirror symmetry, which persists not for any orientations of the surface. In the Brillouin zone of this system, there are four inequivalent $L$ valleys in the forbidden band of each of which massless surface 2D DFs appear. On a (001) surface, pairs of Dirac cones of SSs are projected to a single point, leading to the degeneracy of the spectra of SSs. Atomically abrupt perturbations preserving mirror symmetry lift this degeneracy, moving apart the Dirac points in the reciprocal space without opening a gap in the spectrum of SSs [\onlinecite{Hsieh_NatCom_2012, Xu_2012, Dziawa_2012}]. However, perturbations breaking the mirror symmetry of the surface give rise to a gap in the spectrum of SSs [\onlinecite{Zeljkovic_2015}], thus breaking the topological protection. The fundamental features of the spectrum of SSs in any phase of TCIs based on lead chalcogenides can be found when one neglects the intervalley interaction in the solution of an anisotropic Dirac equation with boundary conditions (\ref{Dirac_BC}).
\section{EDGE STATES IN GRAPHENE}
The Tamm--Shockley states arise far from always. The following questions are fundamental: Under what conditions, except for those implemented in TIs, do Tamm--Shockley-type states arise? Do they exist on real interfaces under normal conditions (rather than in vacuum)? Does the conductivity through these states have a band character? The analysis carried out above shows that, in a number of Dirac materials, such states exist (in contrast to, for example, GaAs-type semiconductors).
\begin{figure}[ht]
\includegraphics{fig7n.eps}
\caption{Edge states on a graphene half-plane in each valley (on the right of the equality sign) and in the reduced valley scheme, when their centers coincide (on the left). }
\end{figure}
Quite recently, edge states of this type have been experimentally observed by resonance methods in another 2D Dirac material, graphene. It was found [\onlinecite{Latyshev_2013}] that the spectrum of Tamm--Shockley edge states in graphene is linear with respect to momentum, which is similar to a section of the conical spectrum of SSs in TIs. Moreover, the band character of the conductivity through these states in normal conditions [\onlinecite{Latyshev_SciRep_2014}] was proved by a direct transport experiment. Let us describe a theory necessary to explain these experiments.
2D DFs in graphene are described by a two-valley system of equations for two-component functions $\Psi_K$ and $\Psi_{K'}$ describing EFs in two valleys, $K$ and $K'$:
\begin{equation}\label{Graphene_eq}
\left(
\begin{array}{cc}
v \bm{\sigma}\cdot\left(\bm{k}-\bm{K}\right) & 0 \\
0 & v \bm{\sigma}\cdot\left(\bm{k}-\bm{K'}\right)
\end{array}
\right)
\left(
\begin{array}{cc}
\Psi_K \\
\Psi_{K'}
\end{array}
\right)=E
\left(
\begin{array}{cc}
\Psi_K \\
\Psi_{K'}
\end{array}
\right)
\end{equation}
where $\bm{\sigma}=(\sigma_x,\sigma_y)$ are the Pauli matrices in standard representation and $v\approx 10^6$ m/s.
General boundary conditions can be determined, as usual, from two requirements: the Hermiticity of Hamiltonian (\ref{Graphene_eq}) and the invariance of the boundary conditions with respect to time reversal. They have a rather complicated form [\onlinecite{Zagorodnev2011,McCann_2004,Zag_Dev_En_PRB,Akhmerov_2008,Basko_2009,Tkachov_2009}]
\begin{equation}\label{graphene_BC}
\left[\Psi_K \sin\beta + i e^{i\varphi}\left(e^{i\gamma\bm{\sigma}\cdot\bm{n}}+\sigma_z \cos\beta\right)\sigma_z\Psi_{K'}\right]_{\Gamma}=0
\end{equation}
Here $\beta, \varphi$, and $\gamma$ are real phenomenological parameters.
With the neglect of intervalley scattering, the system can be described by a pair of independent 2D Weyl equations with one-parameter boundary conditions [\onlinecite{Volkov_2009},\onlinecite{Ostaay_Akhmerov_PRB}] for the single-valley function $\Psi_{K(K')}=\left(\psi_{1K(K')},\psi_{2K(K')}\right)^T$:
\begin{eqnarray}\label{graphene_single_valley}
v \bm{\sigma}\cdot\bm{p}\Psi_{K(K')}=E\Psi_{K(K')} \\
\left[\psi_{1K(K')}+ia^{\tau}\psi_{2K(K')}\right]_{\Gamma}=0
\end{eqnarray}
where $a$ is a real phenomenological parameter that determines the properties of the edge of graphene at the atomic scale and the index $\tau=\pm 1$ numbers the valleys: $\tau=1$ ($\tau=-1$) in valley $K$ ($K'$).
The Tamm problem (\ref{graphene_single_valley}) is easily solved. The wave function of edge states is exponentially localized near the boundary. A typical localization length determined from experiment [\onlinecite{Latyshev_SciRep_2014}] is 2 nm. The value of $a$ should be determined from comparison with experiment; however, from a comparison with model microscopic calculations [\onlinecite{Ostaay_Akhmerov_PRB},\onlinecite{Koskinen_2008}], one should expect that it is small: $|a|\ll 1$.
For a graphene sample in the form of a half-plane, the spectrum of these states in the absence of a magnetic field has the form [\onlinecite{Basko_2009,Tkachov_2009,Volkov_2009,Ostaay_Akhmerov_PRB}] (Fig. 7):
\begin{equation}\label{ES_spectra}
E_{\tau}(k_{||})=\frac{2a}{1+a^2}\tau v k_{||},\quad \tau k_{||}>0
\end{equation}
Here $k_{||}$ is the one-dimensional momentum of an electron along the edge, measured from the center of the valley. It is important that the sign of $k_{||}$ is determined by the number of the valley. For clarity, different valleys in the figure are shown in different colors. The localization length of an edge state is equal to $1/k_{||}$.
In the case of a perpendicular magnetic field, two types of edge states exist: Tamm--Shockley states and magnetic edge states due to skipping orbits. They interact with each other, and the result of their interference depends on the geometry of a sample. For the half-plane $x > 0$, the resulting spectrum [\onlinecite{Zagorodnev2011}] is shown in Fig. 8.
\begin{figure}[ht]
\includegraphics{fig8n.eps}
\caption{Energy (in units of $\hbar v/\lambda$) of DFs as a function of the $X$ center of oscillator in a semi-infinite graphene in a magnetic field of 1 T for $a = 0.2$. The reduced valley scheme. Red solid lines show the states from valley $K$, and blue dashed lines, the states from valley $K'$. }
\end{figure}
Edge states also exist on the boundary of a circular hole (an antidot with radius $R_0$) in an infinite sheet of graphene. The finiteness of the perimeter of the antidot leads to the quantization of the momentum component parallel to the boundary. Now discrete edge states are characterized by the total angular momentum $j = \pm 1/2, \pm 3/2, \pm 5/2, \dots$, different signs of $j$ corresponding to different valleys. Within a quasiclassical approach, the spectrum of such edge states is obtained from (\ref{ES_spectra}) by replacing $k_{||}$ by a multiple of $1/R_0$. A more accurate asymptotes ($|a| \ll 1$) yields a decay-type complex spectrum:
\begin{equation}\label{ES_AD}
E=\frac{2a\hbar v}{R_0}\left(|j|-\frac{\tau}{2}\right) - i \frac{2\pi\hbar v}{R_0}\frac{|a|\left(|a|(|j|-1/2)\right)^{2|j|-1}}{\Gamma^2(|j|-1/2)}.
\end{equation}
Here $\tau j > 0$, $\Gamma(z)$ is a gamma function, and $j = \pm 3/2, \pm 5/2, \dots$. In a special case of $j = \pm 1/2$, the energy of edge states is determined by the root of the equation
\begin{equation}\label{ES_zero_mode}
ER_0\ln\left(\frac{|E|R_0}{2}\right) = - a
\end{equation}
Although these states are quasistationary in the absence of a magnetic field, their lifetime with respect to the decay into bulk states is large in the actual case of small $a$. DFs trapped in edge states perform a cyclic motion along the perimeter of the antidot (clockwise in one valley, or counterclockwise in the other), acquiring an additional Aharonov--Bohm phase in a magnetic field.
The spectrum of edge states in the antidot in a perpendicular magnetic field is controlled by a magnetic flux $\Phi=\pi R_0^2B$ through the area of the antidot:
\begin{equation}\label{ES_asymp}
E = \frac{2a\hbar v}{R_0}\left(j+\frac{\Phi}{\Phi_0}-\frac{\tau}{2}\right).
\end{equation}
The applicability conditions of this asymptotes is given by
\begin{equation}
\tau\left(j+\frac{\Phi}{\Phi_0}\right)>0, \quad |a|\ll 1.
\end{equation}
Asymptotes (\ref{ES_asymp}) is the more justified, the less the effect of the magnetic field on the orbital part of the wave function, i.e., when the localization length of an edge state is small compared with the magnetic length. In the general case, a numerically calculated spectrum is shown in Fig. 9 [\onlinecite{Enaldiev2011}].
\begin{figure}[ht]
\includegraphics{fig9n.eps}
\caption{Dimensionless energy of DFs at a graphene antidot in a magnetic field of $B = 4$ T as a function of the conserving projection of the total angular momentum $j$ in the reduced valley scheme; $a = 0.1$. The states in different valleys are shown in different colors.}
\end{figure}
These states explain the recently revealed Aharonov--Bohm effect in the resistivity of nanoperforated graphene. From comparison with experiment, we can determine the value of the boundary parameter $a = -0.05$.
Thus, we can conclude that hole-type DFs slowly rotate around every antidot in nanoperforated graphene. The sign of rotation is determined by the number of the valley, and the speed of rotation is ten times less than the volume velocity of DFs in graphene.
\section{CONCLUSIONS. MINIMAL MODEL OF SURFACE STATES FOR DIRAC ELECTRONS}
One can draw the following qualitative conclusion from the survey of literature data on Dirac materials: the Dirac property favours the emergence of a Tamm--Shockley-type band of edge and surface states. However, this band is not always situated in the bulk gap where its spectrum is analogous to the spectrum of topological states. For some values of boundary parameters, this band is expelled from the gap and its energy overlaps with the energy of bulk bands.
The analysis carried out allows us to formulate a minimal model suitable for the analytic investigation of fundamental features of the spectrum of surface or edge states in a number of Dirac systems: a generally anisotropic version of the 3D Dirac equation (\ref{Dirac_eq}) with boundary conditions (\ref{Dirac_BC}) and their 2D generalizations.
Such a model can be applied with neglect of intervalley interaction to describe SSs in the following Dirac systems.
In 3D systems: lead chalcogenides and solid solutions based on lead chalcogenides Pb$_{1-x}$Sn$_x$Se(Te)-type, bismuth- and Bi$_{1-x}$Sb$_x$-type semimetals at the $L$ point, both in the inverted and ordinary, non-inverted, modes.
In 2D systems: graphene and 2D TIs in quantum wells based on a HgTe--CdTe heteropair.
The minimal model cannot be applied to describe SSs in Bi$_2$Se$_3$, Bi$_2$Te$_3$, and Sb$_2$Te$_3$ because of a significant contribution of remote bands that forms the dispersion of the mass term in a modified Dirac equation.
\section{ACKNOWLEDGMENTS}
We are grateful to I.V. Zagorodnev for critical remarks.
This work was supported in part by the Russian Foundation for Basic Research.
|
\section{Introduction}
Let $G$ be a simple graph of {\em order} $n$ with the vertex set $V(G)$ and {\em size} $m$ with the
edge set $E(G)$. We use \cite{W} for terminology and notation,
which are not defined here. The {\em minimum} and {\em maximum degrees}
in graph $G$ are denoted by $\delta(G)$ and $\Delta(G)$, respectively. A vertex $v \in V(G)$ is called an \textit{odd} (\textit{even}) vertex if $\deg_G(v)$ is odd (even). For a graph $G=(V,E)$, let $V_{o}$ ( $V_e$) be the set of odd (even) vertices with $n_o=\mid V_o \mid$ and $n_e=\mid V_e \mid$. If $X \subseteq V(G)$, then $G[X]$ is the {\em subgraph} of $G$ {\em induced} by $X$. For disjoint subsets $X$ and $Y$ of vertices we let
$E(X,Y)$ denote the set of edges between $X$ and $Y$. The {\em
open neighborhood} $N_G(v)$ of a vertex $v\in V(G)$ is the set of
all vertices adjacent to $v$. Its {\em closed neighborhood} is
$N_G[v] = N_G(v) \cup \{v\}$. In addition, the \textit{open} and \textit{closed neighborhoods} of a subset $S \subseteq V(G)$ are $N_G(S)=\cup_{v\in S}N_G(v)$ and
$N_G[S]=N_G(S)\cup S$, respectively. The \textit{degree} of a vertex $v \in V(G)$ is $\deg_G(v)= \mid N_G(v) \mid$.
For a real-valued function $f: V(G)
\longrightarrow {\mathbb R}$ the weight of $f$ is $\omega(f)=\sum
_{v \in V(G)}f(v)$ and for a subset $S$ of $V(G)$ we define $f(S)
= \sum_{v\in S} f(v)$, so $\omega(f)=f(V(G))$.
For a positive integer $1 \leq k \leq n$, a {\em signed $k$-subdominating function} (SkSDF) of G is a function $f: V(G)\longrightarrow \{-1, 1\}$ such that $f (N_G[v]) \geq1$ for
at least $k$ vertices $v$ of $G$. The {\em signed $k$-subdomination number} for a graph $G$ is defined as $\gamma_{ks}(G) = \min\{w(f )|f \mbox{ is a SkSDF of } G\}$. The concept of the signed $k$-subdomination number was introduced and studied by Cockayne and Mynhardt \cite{CM}.
A {\em nonnegative signed dominating function} (NNSDF) of $G$ defined in
\cite{HLFZ} as a function $f : V(G) \longrightarrow \{-1, 1\}$ such
that $f(N_G[v]) \geq 0$ for all vertices $v$ of $G$. The {\em nonnegative signed
domination number} (NNSDN) of $G$ is
$\gamma^{NN}_{s}(G)= \min\{\omega(f)| f \mbox{ is an NNSDF of } G \}$.
We now introduce a
{\em nonnegative signed $k$-subdominating function} (NNSkSDF) of $G$ for a positive integer $1 \leq k \leq n$ as a function $f :
V(G) \longrightarrow \{-1, 1\}$ such that $f(N_G[v]) \geq 0$ for at
least $k$ vertices $v$ of $G$. We define the {\em nonnegative signed $k$-subdomination number}
(NNSkSDN) of $G$ by $\gamma^{NN}_{ks}(G)= \min\{\omega(f)\mid f \mbox{ is a NNS}k\mbox{SDF of }G \}$. A nonnegative signed $k$-subdominating function of weight $\gamma^{NN}_{ks}(G)$ is called a $\gamma^{NN}_{ks}(G)$-{\em function}. Note that
$\gamma^{NN}_{ns}(G)=\gamma^{NN}_{s}(G)$. Since every signed $k$-subdominating function of $G$ is a nonnegative signed $k$-subdominating function, we deduce that
$$\gamma^{NN}_{ks}(G) \leq \gamma_{ks}(G).$$
For a function $f : V(G)
\longrightarrow \{-1, 1\}$ of $G$ we define $P=\{v\in V(G)\mid
f(v)=1\}$, $M=\{v\in V(G)\mid f(v)=-1\}$, and $C_f=\{v \in V(G)| f (N_G[v])\geq 0 \}$.
In this paper, we establish some new lower bounds on $\gamma^{NN}_s(G)$ for a general
graph in terms of various different graph parameters. Some of these bounds improve several lower bounds on $\gamma^{NN}_s(G)$ presented in \cite{AS,HLFZ}. We also
initiate the study of nonnegative signed $k$-subdomination numbers in graphs, and present some sharp lower bounds for $\gamma^{NN}_{ks}(G)$ in terms of the
order and the degree sequence of a graph $G$.
\begin{observation}\label{ob}
{\rm Let $f$ be an NNSkSDF of $G$. For $v \in C_f$ if $v$ is an even vertex, then $f(N_G[v]) \geq 1$ while $f(N_G[v]) \geq 0$ if $v$ is an odd vertex. }
\end{observation}
\begin{observation}\label{Signed}
{\rm Let $1\leq k\leq n$ be a positive integer. For any even graph $G$, $$\gamma^{NN}_{ks}(G)=\gamma_{ks}(G).$$ }
\end{observation}
In this paper, we make use of the following results.
\begin{prelem}\cite{AS}\label{Ata3}
{\rm Let $G$ be a graph of order $n$ and size $m$. Then
$$\gamma^{NN}_{s}(G) \geq \frac{n}{2}-m.$$
}
\end{prelem}
\begin{prelem}\cite{AS}\label{Ata5}
{\rm Let $G$ be a graph of order $n$, size $m$ and minimum degree $\delta$. Then
$$\gamma^{NN}_{s}(G) \geq \frac{-4m+3n\lceil\frac{\delta+1}{2}\rceil-n}{3\lceil\frac{\delta+1}{2}\rceil+1}.$$
}
\end{prelem}
\begin{prelem}\cite{DHHS}\label{cycle}
{\rm For $n \geq 3$, $$\gamma_s(C_n)=\left\{
\begin{array}{lcl}
n/3 & \mbox{if} & n \equiv 0 \pmod 3, \\
\lfloor\frac{n}{3}\rfloor+1 & \mbox{if} & n \equiv 1 \pmod 3, \\
\lfloor\frac{n}{3}\rfloor+2 & \mbox{if} & n \equiv 2 \pmod 3. \\
\end{array} \right. $$ }
\end{prelem}
\begin{corollary}\cite{H}\label{Hen}
{\rm For any $r$-regular graph $G$ of order $n$,
$\gamma_{s}(G) \geq \dfrac{n}{r+1}$, for $r$ even.
Furthermore this bound is sharp.}
\end{corollary}
\begin{prelem}\cite{HLFZ}\label{Hua6}
{\rm Let $K_n$ be a complete graph. Then $\gamma^{NN}_{s}(K_n) =0$ when $n$ is even and $\gamma^{NN}_{s}(K_n) =1$ when $n$ is odd.
}
\end{prelem}
\begin{prelem}\cite{HLFZ}\label{Hua10}
{\rm For any graph $G$ with maximum degree $\Delta$ and minimum degree $\delta$, we have $$\gamma^{NN}_{s}(G) \geq \frac{\delta-\Delta}{\delta+\Delta+2}n.$$
}
\end{prelem}
\section{Lower bounds on the NNSDNs of graphs }
In this section, we present some new sharp lower bounds for $\gamma^{NN}_{s}(G)$ by using $n_e$ as the number of even vertices in a graph $G$. We begin with the following lemma.
\begin{lemma}\label{lemNN}
{\rm Let $f$ be an NNSDF of a simple connected graph $G$. Then,
\begin{enumerate}
\item $\sum_{v\in P} \deg_{G}(v) \geq n+n_e-2\mid P \mid +\sum_{v \in M} \deg_{G}(v)$.
\item $\sum_{v \in P}\deg_{G[P]}(v)\geq \sum_{v \in
P}\lceil \frac {\deg_{G}(v)-1}{2}\rceil$.
\end{enumerate}
}
\end{lemma}
\begin{proof}
For $v \in V(G)$, let $\deg_P(v)$ and $\deg_M(v)$ denote the numbers of vertices of $P$ and $M$, respectively, which are adjacent
to $v$. Clearly, $\deg_{G}(v)=\deg_{M}(v)+\deg_{P}(v)$.
Since $f(N_G[v]) \geq 0$, for every $v \in P$, $\deg_{M}(v) \leq
\deg_{P}(v)+1$, and for every $v \in M$, $\deg_{P}(v) \geq
\deg_{M}(v)+1$. Hence, if $v \in P$, then
$\deg_{M}(v) \leq \lfloor \frac{\deg_{G}(v)+1}{2}\rfloor$ and if $v
\in M$, then $\deg_{P}(v) \geq \lceil
\frac{\deg_{G}(v)+1}{2}\rceil$.
\begin{enumerate}
\item Counting the number of edges in $E(P,M)$
in two ways, we can deduce that
$$\begin{array}{ccccl}
\sum_{v \in M} \lceil \frac{\deg_{G}(v)+1}{2} \rceil & \leq & |E(P,M)| & \leq &
\sum_{v \in P} \lfloor \frac{\deg_{G}(v)+1}{2} \rfloor
\end{array}$$
\noindent It follows that
$$\begin{array}{ccl}
\sum_{v \in P}\deg_{G}(v)+ \mid P \mid & \geq & \sum_{v \in V} \lceil\frac{\deg_{G}(v)+1}{2}\rceil\\\\
&=&\sum_{v \in V_o}\frac{\deg_G(v)+1}{2}+\sum_{v \in V_e}\frac{\deg_G(v)+2}{2}\\\\
&=&\sum_{v \in V}\frac{\deg_G(v)}{2}+n_e+\frac{n_o}{2}\\\\
&=&\sum_{v \in P}\frac{\deg_G(v)}{2}+\sum_{v \in M}\frac{\deg_G(v)}{2}+n_e+\frac{n_o}{2},\\
\end{array}$$
which implies that
$$\sum_{v \in P}\frac{\deg_G(v)}{2} \geq \sum_{v \in M}\frac{\deg_G(v)}{2}+n_e+\frac{n_o}{2}-\mid P\mid.$$ Hence,
$$\begin{array}{ccl}
\sum_{v\in P} \deg_{G}(v)& \geq & \sum_{v \in M} \deg_{G}(v)+n+n_e-2\mid P \mid.\\
\end{array}$$
\item Consider the subgraph $G[P]$ induced by $P$. We have
$\deg_{G[P]}(v) = \deg_{P}(v)$ for each $v \in P$. Since
$\deg_P(v) \geq \lceil \frac {\deg_{G}(v)-1}{2}\rceil$ for each $v
\in P$, we have
$$\sum_{v \in P}\deg_{G[P]}(v)\geq \sum_{v \in P}\lceil \frac
{\deg_{G}(v)-1}{2}\rceil.$$
\end{enumerate}
\end{proof}
In the next theorem we present some lower bounds on
$ \gamma^{NN}_{s}(G)$. By using Lemma \ref{lemNN} and graph parameters such as order, size, number of even vertices, maximum and minimum degrees
we obtain some new lower bounds for $\gamma^{NN}_{s}(G)$. These new results are independent from each other.
\begin{theorem}\label{theoNN}
{\rm Let $G$ be a simple connected graph of
order $n$, minimum degree $\delta$, maximum degree $\Delta$ and the number of even vertices $n_e$.
Then
\begin{enumerate}
\item $\gamma^{NN}_{s}(G) \geq \dfrac{n\delta-n\Delta+2n_e}{\Delta+\delta+2}$,
\\
\item $\gamma^{NN}_{s} (G) \geq \dfrac{2m+n_e-n\Delta}{\Delta+1}$,
\\
\item $\gamma^{NN}_{s} (G)\geq \dfrac{n\delta+n_e-2m}{\delta+1} $,
\\
\item $\gamma^{NN}_{s} (G)\geq \lceil\dfrac{-(\delta+1)+ \sqrt{(\delta+1)^2 +8(n\delta +n+n_e)}}{2}-n\rceil$,
\\
\item $\gamma^{NN}_{s} (G) \geq \lceil\sqrt{2m+n+n_e}-n\rceil$.
\end{enumerate}
}
\end{theorem}
\begin{proof}
According to Lemma \ref{lemNN}, we have
\begin{equation}\label{eq1}
\sum_{v \in M}\deg_G(v)+n+n_e-2 \mid P \mid \leq \sum_{v \in P}\deg_G(v).
\end{equation}
\begin{enumerate}
\item \label{1} Since $\delta \leq \deg_G(v) \leq \Delta$ for each $v \in V(G)$, inequality (\ref{eq1}) follows that
$$\delta n- \mid P \mid\delta+n+n_e-2\mid P \mid\leq \sum_{v\in P}\deg_{G}(v) \leq \Delta\mid P\mid.$$
From this inequality, it is deduced that
$$\mid P \mid \geq \frac{\delta n +n+n_e}{\Delta+\delta+2}.$$
Hence, $$\gamma^{NN}_{s} (G)=2\mid P \mid-n \geq
\dfrac{n\delta-n\Delta+2n_e}{\Delta+\delta+2}.$$
\item \label{2} Since $2m=\sum_{v \in V} \deg_G(v)$ and
$\deg_G(v) \leq \Delta$ for each $v \in V(G)$, by inequality
(\ref{eq1}) it follows that
$$\begin{array}{ccl}
2m+n+n_e-2\mid P \mid & = & \sum_{v \in V}\deg_{G}(v)+n+n_e-2\mid P \mid \\\\
& \leq & 2\sum_{v \in P}\deg_{G}(v)\\\\
& \leq & 2\Delta\mid P \mid.
\end{array}$$
Therefore, $2\mid P \mid\geq \dfrac{2m+n+n_e}{\Delta+1}$, and
$\gamma^{NN}_{s} (G) \geq \dfrac{2m+n_e-n\Delta}{\Delta+1}$, as desired.
\item \label{3} Using inequality (\ref{eq1}) and the facts $2m=\sum_{v \in V}
\deg_G(v)$ and $\deg_G(v) \geq \delta$ for any $v \in V(G)$, we
have
$$\begin{array}{ccl}
2m & = & \sum_{v \in V}\deg_{G}(v)\\\\
& \geq & 2\sum_{v \in M}\deg_{G}(v)+n+n_e-2\mid P \mid\\\\
& \geq & 2n\delta-2\delta\mid P \mid+n+n_e-2\mid P \mid
\end{array}$$
It follows that
$$2\mid P \mid\geq \dfrac{2n\delta +n+n_e-2m}{\delta+1}.$$
Thus, $\gamma^{NN}_{s} (G)\geq \dfrac{n\delta+n_e-2m}{\delta+1} $, as desired.
\item \label{4} Consider $G[P]$. According to Lemma
\ref{lemNN}, we have
$$\sum_{v \in P}\deg_{G[P]}(v)\geq \sum_{v \in
P}\lceil \dfrac {\deg_{G}(v)-1}{2}\rceil \geq \sum_{v \in P} \dfrac{\deg_{G}(v)-1}{2} .$$ On the other hand,
since $G[P]$ is a simple connected graph, $$\sum_{v \in P}\deg_{G[P]}(v)\leq
\mid P \mid(\mid P \mid-1).$$ Thus,
$$\begin{array}{ccl}
2\mid P \mid(\mid P \mid-1) &\geq& \sum_{v \in P}\deg_{G}(v) -\mid P \mid\hfill\\\\
& \geq & \sum_{v \in M}\deg_{G}(v)+n+n_e-3\mid P \mid\hfill\\\\
& \geq & n\delta-\delta\mid P \mid+n+n_e-3\mid P\mid. \hfill\\\\
\end{array}$$
This implies that
$$2\mid P \mid^2+(\delta+1)\mid P \mid-(n\delta+n+n_e)\geq 0.$$
Therefore,
$$2\mid P \mid \geq \frac{-(\delta+1)+ \sqrt{(\delta+1)^2 +8(n\delta +n+n_e)}}{2},$$
and hence $\gamma^{NN}_{s} (G)\geq \lceil \dfrac{-(\delta+1)+ \sqrt{(\delta+1)^2 +8(n\delta +n+n_e)}}{2}-n\rceil$, as desired.
\item \label{5} By Parts (\ref{4}) and (\ref{2}) we have
$$2\sum_{v \in P}\deg_G(v) \leq 4\mid P \mid^2-2\mid P \mid, $$
and $$2\sum_{v \in P}\deg_G(v) \geq 2m+n+n_e-2\mid P \mid,$$
respectively. So,
$$4\mid P \mid^2\geq 2m+n+n_e,$$
which implies that
$$ \mid P \mid\geq \frac{\sqrt{2m+n+n_e}}{2}.$$
Thus, $\gamma^{NN}_{s} (G)\geq \lceil\sqrt{2m+n+n_e}-n\rceil $, as
desired.
\end{enumerate}
\end{proof}
From Theorem \ref{theoNN} (\ref{1})$-$(\ref{3}), we have the
following result. For $k=n$ by Observation \ref{Signed} when $r$ is even, we have the same bound presented in Corollary \ref{Hen} by Henning.
\begin{corollary}\label{regular}
{\rm For $r\geq 1$, if $G$ is an $r$-regular graph of order $n$,
then
$$\gamma^{NN}_{s}(G) \geq \left\{\begin{array}{ll}
\dfrac{n}{r+1} & \mbox{ if } $r$ \mbox{ is even,} \\
&\\
0 & \mbox{ if } $r$ \mbox{ is odd}. \\
\end{array}\right.$$
Furthermore, these bounds are sharp.}
\end{corollary}
\noindent In order to show that the bounds presented in Theorem \ref{theoNN} are sharp, we will give a graph $G$ and construct a $\gamma^{NN}_s(G)$-function $f$
such that $w(f)$ achieves the lower bounds, and thus the lower bounds are sharp. We also illuminate that our bounds for
some of these graphs are attainable while the corresponding bounds given in Theorems \ref{Ata3}, \ref{Ata5}, and \ref{Hua10} are not.
In fact, a trivial examples such $G \in \{K_n,C_n\}$ is sufficient for this. It is easy to
see that $\gamma^{NN}_s(K_n)$ obtains all the five bounds in Theorem \ref{theoNN} while the bound in Theorem \ref{Ata3} shows that $\gamma^{NN}_s(K_n) \geq \frac{2n-n^2}{2}$ and the bound in Theorem \ref{Ata5} is not more than $\frac{5n-n^2}{3n+5}$. As an other example, $\gamma^{NN}_s(C_n)$ attains the lower bounds in
Theorem \ref{theoNN} (\ref{1})$-$(\ref{3}), when $n\equiv 0\;\; (mod\;\;3)$ while the bounds in Theorems \ref{Ata3}, \ref{Ata5} and \ref{Hua10} are
not more than $\frac{n}{7}$. Besides, we can construct a non-complete graph with an arbitrary large order whose reaches the lower bounds in Theorem \ref{theoNN} (\ref{1})$-$(\ref{3}) as follows. Letting $t$ be a positive integer, we consider a cycle of length $2t$. For every edge, we include an additional vertex being adjacent to both endpoints
of the corresponding edge. The obtained graph is denoted by $G$. It is easy to check that the graph $G$ is a graph with $n=4t$, $m=6t$, $\delta=2$, $\Delta=4$ and $n_e=4t$.
Define a function $f : V(G) \longrightarrow \{-1, 1\}$ as follows: $f (v)=1$ for $v \in V(C_{2t})$ and $f (v)=-1$ for
$v \in V(G) \setminus V(C_{2t})$. It is easy to verify that $f$ is a $\gamma^{NN}_s(G)$-function with
$w(f ) = 0$ and bounds in Theorem \ref{theoNN} (\ref{1})$-$(\ref{3}) are also $0$, which implies that $G$ is sharp for these bounds. However, $\gamma^{NN}_s(G)$ does not attain the corresponding bounds given in Theorems \ref{Ata3}, \ref{Ata5} \ and \ref{Hua10}, which are $-4t$, $\lceil-\frac{4t}{7}\rceil$, and $-t$, respectively. Next, we show that there is also a graph $G$ different from $K_n$ such that $\gamma^{NN}_s(G)$ reaches lower bounds in
Theorem \ref{theoNN} (\ref{4})$-$(\ref{5}). Let $H$ be the Haj\'{o}s graph. We can verify that $\gamma^{NN}_s(H)=0$, and $H$ is sharp for presented bounds in Theorem \ref{theoNN} (\ref{4})$-$(\ref{5}).
\section{Lower bounds on the NNSkSDNs of graphs}
In this section, we initiate the study of the nonnegative signed $k$-subdomination number in graphs. we present some lower bounds on the nonnegative signed $k$-subdomination number of a graph in terms of the order and the degree sequence. We begin with the following lemma.
\begin{lemma}\label{lemkNN}
{\rm Let $G$ be a graph and $1\leq k \leq n$ be a positive integer. Let $f$ be a $\gamma^{NN}_{ks}(G)$-function.
Let $P_1=P \cap C_f$, $P_2=P \setminus P_1$, $M_1=M \cap C_f$ and $M_2 =M\setminus M_1$. Then,
$$\sum_{v \in P}\deg_G(v)+\mid P_1 \mid \geq \sum_{v \in P_1 \cup M_1}\lceil \dfrac{\deg_G(v)+1}{2}\rceil.$$
}
\end{lemma}
\begin{proof}
Note that if $v \in V(G)$, then $\deg_G(v)=\deg_P(v)+\deg_M(v)$. For $v \in P_1 \cup M_1$, $f (N_G[v])\geq0$. So, if $v \in P_1$, then $\deg_P(v) \geq \lceil\frac{\deg_G(v)-1}{2}\rceil$ and $\deg_M(v) \leq \lfloor\frac{\deg_G(v)+1}{2}\rfloor$. Similarly, if $v \in M_1$, then $\deg_P(v) \geq \lceil\frac{\deg_G(v)+1}{2}\rceil$ and $\deg_M(v) \leq \lfloor\frac{\deg_G(v)-1}{2}\rfloor$.
Counting the number of edges in $E(P,M)$
in two ways, we conclude that
$$\sum_{v \in M_1}\lceil\frac{\deg_G(v)+1}{2}\rceil+\sum_{v \in M_2}\deg_P(v) \leq \sum_{v \in P_1}\lfloor\frac{\deg_G(v)+1}{2}\rfloor+\sum_{v \in P_2}\deg_M(v).$$
By adding $\sum_{v \in P_1}\lceil\frac{\deg_G(v)+1}{2}\rceil$ to the both sides of the inequality we have
$$\begin{array}{ccc}
\sum_{v \in P}\deg_G(v)+\mid P_1 \mid &\geq& \sum_{v \in P_1 \cup M_1}\lceil\dfrac{\deg_G(v)+1}{2}\rceil+ \sum_{v \in M_2}\deg_G(v)\hfill\\\\
&\geq& \sum_{v \in P_1 \cup M_1}\lceil\dfrac{\deg_G(v)+1}{2}\rceil,\hfill\\
\end{array}$$
\noindent and this completes the proof.
\end{proof}
\begin{theorem}\label{theokNN}
{\rm For any graph $G$ with degree sequence $d_1\leq d_2 \leq \cdots \leq d_n$ and any positive integer $1 \leq k \leq n$,
\begin{enumerate}
\item[1.] $\gamma^{NN}_{ks}(G) \geq \dfrac{2\sum_{i=1}^{k}\lceil \frac{d_i+1}{2}\rceil}{\Delta+1}-n$.
\item[2.] $\gamma^{NN}_{ks}(G) \geq \dfrac{n\delta-4m-n+2\sum_{i=1}^{k}\lceil \frac{d_i+1}{2}\rceil}{\delta+1}$.\\
Furthermore, these bounds are sharp.
\end{enumerate}
}
\end{theorem}
\begin{proof}
Considering Lemma \ref{lemkNN} we have
\begin{equation}\label{eq31}
\sum_{v \in P}\deg_G(v)+\mid P_1 \mid \geq \sum_{v \in P_1 \cup M_1}\lceil \dfrac{\deg_G(v)+1}{2}\rceil.
\end{equation}
\begin{enumerate}
\item \label{k1} Since $\delta \leq \deg_G(v) \leq \Delta$ for each $v \in V(G)$, inequality (\ref{eq31}) follows that
$$\Delta \mid P \mid+\mid P \mid \geq \sum_{v \in P}\deg_G(v)+\mid P_1 \mid \geq \sum_{v \in P_1 \cup M_1} \lceil\dfrac{\deg_G(v)+1}{2}\rceil.$$
Note that $P_1 \cap M_1=\emptyset$ and $\mid P_1 \cup M_1 \mid=\mid P_1 \mid+\mid M_1 \mid \geq k$. So,
$$\mid P\mid \geq \dfrac{\sum_{v \in P_1 \cup M_1} \lceil\dfrac{\deg_G(v)+1}{2}\rceil}{\Delta+1} \geq \dfrac{\sum_{i=1}^k\lceil\dfrac{d_i+1}{2}\rceil}{\Delta+1}.$$
Thus,
$$\gamma^{NN}_{ks}(G)=2 \mid P \mid-n \geq \dfrac{2\sum_{i=1}^{k}\lceil \frac{d_i+1}{2}\rceil}{\Delta+1}-n.$$
\item \label{k2} Obviously, , $2m=\sum_{v \in V(G)}\deg_G(v)=\sum_{v \in P}\deg_G(v)+\sum_{v \in M}\deg_G(v)$. If we add $\mid P_1 \mid$ to the both sides of this equality, then by Lemma \ref{lemkNN} we deduce that
$$\begin{array}{ccc}
\mid P \mid \geq \mid P_1 \mid &\geq& -2m+ \sum_{v \in P_1 \cup M_1}\lceil \dfrac{\deg_G(v)+1}{2}\rceil+\sum_{v \in M}\deg_G(v)\hfill\\\\
&\geq& -2m+\sum_{i=1}^{k}\lceil\dfrac{d_i+1}{2}\rceil+\delta n - \delta \mid P\mid.\hfill\\
\end{array}$$
Therefore,
$$\mid P \mid \geq \dfrac{n\delta-2m+\sum_{i=1}^{k}\lceil \frac{d_i+1}{2}\rceil}{\delta+1},$$
and hence,
$$\gamma^{NN}_{ks}(G)=2\mid P \mid -n \geq \dfrac{n\delta-4m-n+2\sum_{i=1}^{k}\lceil \frac{d_i+1}{2}\rceil}{\delta+1}.$$
\end{enumerate}
Now suppose that $k =n$, considering that $2\sum_{i=1}^n \lceil \dfrac{d_i+1}{2}\rceil =2m+n+n_e$, we can immediately obtain those two bounds in Theorem \ref{theoNN} (\ref{2}) and (\ref{3}) from the lower bounds of Theorem
\ref{theokNN}, respectively. Since the bounds in Theorem \ref{theoNN} are sharp, so there exist graphs whose $\gamma^{NN}_{ks}(G)$ recieve the bounds in
Theorem \ref{theokNN}. Therefore, these bounds are sharp.
\end{proof}
\noindent As an immediate consequence of Theorem \ref{theokNN} we have the
following result.
\begin{corollary}\cite{HATT}\label{Hatt}
{\rm For every $r$-regular graph $G$ of order $n$, $\gamma_{ks}(G)\geq \dfrac{k(r + 2)}{r + 1}-n$ for $r$ even. }
\end{corollary}
\begin{corollary}\label{kregular}
{\rm For $r\geq 1$, if $G$ is an $r$-regular graph of order $n$,
then
$$\gamma^{NN}_{ks}(G) \geq \left\{\begin{array}{ll}
\dfrac{k(r+2)}{r+1}-n & \mbox{ if } $r$ \mbox{ is even,} \\
&\\
k-n & \mbox{ if } $r$ \mbox{ is odd}. \\
\end{array}\right.$$
Furthermore, these bounds are sharp.}
\end{corollary}
\noindent Clearly, if $r$ is even, then by Observation \ref{Signed} we have the same given bound in Corollary \ref{HATT}.
|
\section{Concluding Remarks}
The $\text{\sf FAQ}$ framework showed that many common computational tasks over a very
wide range of domains such as CSPs, machine learning, relational database,
logic, and matrix computations can be performed {\em inside} a database
using the same abstraction. The main idea is to blur the line between data
and computation, as we use the database to store, compute, and
process functions.
The glue of the framework is a simple dynamic programming algorithm called
$\functionname{InsideOut}$, which can be cast as a query-rewriting method and thus it is
readily implementable within any RDBMS.
These ideas are implemented, tested, and validated within the $\text{\sf LogicBlox}$
database system. Our theory predicts practice very well, which is a
beautiful thing to see.
\section{The {\large \sf InsideOut} Algorithm}
\label{sec:algo}
Parts of this section will be familiar to readers
who have been exposed to elementary graphical models~\cite{MR2778120}. There
are, however, a couple of ideas that are taken from new developments in
database theory~\cite{skew, NPRR12, faq}
that are likely not known in the graphical model literature.
For each factor $\psi_S$, define its {\em size} to be the number of non-zero
points under its domain:
$ |\psi_S| := \left|\bigl\{ \mv x_S \ | \
\psi_S(\mv x_S) \neq \mv 0
\bigr\}
\right|.$
{\it \bf{Basic variable elimination.}}
To describe the intuition, we first explain $\functionname{InsideOut}$ as it applies to
the special case of $\text{\sf FAQ-SS}$ (or $\problemname{SumProd}$).
The idea behind variable elimination
\cite{DBLP:journals/ai/Dechter99,MR1426261,zhangpoole94} is to `fold' common
factors, exploiting the distributive law:
\begin{eqnarray*}
&& \bigoplus_{x_{f+1}}\cdots
\bigoplus_{x_n}
\bigotimes_{S\in\mathcal E} \psi_S(\mv x_S)\\
&=&\bigoplus_{x_{f+1}}\cdots
\bigoplus_{x_{n-1}}
\bigotimes_{S\in\mathcal E-\partial(n)} \psi_S(\mv x_S)
\otimes \underbrace{
\left( \bigoplus_{x_n} \bigotimes_{S\in\partial(n)}
\psi_S(\mv x_S)\right)}_{\text{new factor } \psi_{U_n-\{n\}}},
\end{eqnarray*}
where the equality follows from the fact that $\otimes$ distributes over
$\oplus$, $\partial(n)$ denotes all edges incident to $n$ in $\mathcal H$
and $U_n=\cup_{S\in\partial(n)} S$.
Assume for now that we can somehow efficiently compute the intermediate
factor $\psi_{U_n-\{n\}}$. Then, the resulting problem is another instance of
$\text{\sf FAQ-SS}$ on a modified multi-hypergraph $\mathcal H'$,
constructed from $\mathcal H$ by removing vertex $n$ along with
all edges in $\partial(n)$, and {\em adding back} a new hyperedge $U_n-\{n\}$.
Recursively, we continue this process until all variables $X_n,\dots,X_{f+1}$
are eliminated. Textbook treewidth-based results for $\problemname{PGM}$ inference
are obtained this way \cite{MR2778120}.
In the database context (i.e.\ given an $\text{\sf FAQ}$-query over the Boolean semiring), the
intermediate result $\psi_{U_n-\{n\}}$ is essentially an intermediate
relation of a query plan, the folding technique exploiting distributive law
corresponds to ``pushing the aggregate down'' the query
plan~\cite{Cohen:2006:UAF:1142473.1142480}.
{\it\bf Introducing the indicator projections.}
While correct, basic variable elimination as described above is potentially
not very efficient for sparse input factors, i.e. factors where the number of non-zero
entries is much smaller than the product of the active domain sizes.
This is because the product that was factored out of the scope of $X_n$
might annihilate many entries of the intermediate result
$\psi_{U_n-\{n\}}$, while we have spent so much time computing $\psi_{U_n-\{n\}}$.
For example, for an $S\notin\partial(n)$ such that $S\subseteq U_n$ and tuple $\mv y_S$ such that
$\psi_S(\mv y_S)=\mv 0$, we do not need to compute the entries
$\psi_{U_n-\{n\}}(\mv x_{U_n-\{n\}})$ for which
$\mv y_{S} = \mv x_{S}$: those entries will be eliminated later
anyhow. The idea is then to only compute those
$\psi_{U_n-\{n\}}(\mv x_{U_n-\{n\}})$ values that will ``survive'' the other
factors later on.
One simple way to achieve this would be to compute, for each
$S \in \mathcal E - \partial(n)$,
an ``indicator factor'' that checks if $\psi_S(\mv x_S)$ is $\mv 0$ or not.
Formally, for any two sets $T\subseteq S$, and a given factor $\psi_S$,
the function
$\psi_{S/T} : \prod_{i \in T} \functionname{Dom}(X_i) \to \mv D$
defined by
\[
\psi_{S/T}(\mv x_T) :=
\begin{cases}
\mv 1 & \exists \mv x_{S-T} \text{ s.t. }
\psi_S(\mv x_T,\mv x_{S-T}) \neq \mv 0\\
\mv 0 & \text{otherwise}
\end{cases}
\]
is called the {\em indicator projection} of $\psi_S$ onto $T$.
Using indicator factors, $\functionname{InsideOut}$ computes the following factor
when marginalizing $X_n$ away:
\begin{multline}\label{eqn:sub:query}
\psi_{U_n-\{n\}}(\mv x_{U_n-\{n\}}) =\\
\bigoplus_{x_n} \left[ \left( \bigotimes_{S\in\partial(n)} \psi_S
\right)
\otimes \left( \bigotimes_{\substack{S\notin \partial(n),\\S\cap
U_n\neq \emptyset}}
\psi_{S/S\cap U_n}
\right)\right].
\end{multline}
Another minor tweak is the observation that, if there is a hyperedge
$S\in \mathcal E-\partial(n)$ for which $S\subset U_n$, then we do not use
the indicator projection $\psi_{S/S\cap U_n}$: we can use $\psi_S$ itself
to compute the intermediate factor $\psi_{U_n-\{n\}}$,
and then remove $\psi_S$ from $\mathcal H'$.
\begin{ex}
We explain how the ideas above are implemented in
Example~\ref{ex:db}. First, the
order in which we choose to eliminate variables might have
a huge effect on the runtime. For now, let us assume that we somehow
decided to rewrite $\varphi(b,d)$ using the following variable order, where we
trace the first couple of steps of the $\functionname{InsideOut}$ algorithm {\em without}
the indicator projection: (Example~\ref{ex:db:GYO} later explains how this order is related to the tree decomposition in Fig.~\ref{fig:example:tree}.)
\begin{eqnarray*}
\varphi(b,d) &=& \sum_c \sum_a \sum_e \sum_f \sum_g \sum_h
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_Y\\
&=& \sum_c \sum_a \sum_e \sum_f \sum_g
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\underbrace{\sum_h \psi_Y(f,h)}_{\psi_1(f)}\\
&=& \sum_c \sum_a \sum_e \sum_f \sum_g
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_1 \\
&=& \sum_c \sum_a \sum_e \sum_f
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\underbrace{\sum_g \psi_W(e,g)}_{\psi_2(e)}\\
&=& \sum_c \sum_a \sum_e \sum_f
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\end{eqnarray*}
The first two steps are straightforward, where we eliminated $g$ and $h$.
In $\text{\sf LogiQL}$, these intermediate factors are computed with the following
two rules
\begin{verbatim}
psi1[f] = s1 <- agg<<s1 = count()>> Y(f,h).
psi2[e] = s2 <- agg<<s2 = total(e)>> W(e,g).
\end{verbatim}
The mathematical abstraction corresponds to rewriting a query into a series of smaller queries.
Next, we explain how the indicator projection works when we eliminate variable
$f$. Out of the remaining factors
$\psi_R$
$\psi_S$
$\psi_T$
$\psi_U$
$\psi_V$
$\psi_1$
and
$\psi_2$
the following factors contain $f$:
$\psi_U(d,f)$
$\psi_V(e,f)$
and
$\psi_1(f)$.
If we were to multiply them together and marginalize away $f$, we would
create a new factor $\psi_3(e,d) = \sum_f \psi_U\psi_V\psi_1$ over variables $\{e, d\}$.
However, two other factors have variables that overlap with $\{e, d\}$, namely $\psi_T(b,c,d,e)$ and $\psi_2(e)$.
For $\psi_T$, we include its indicator projection $\psi_{T/\{e,d\}}$
in computing $\psi_3$.
(We will see later in Example~\ref{ex:db:runtime} how including $\psi_{T/\{e,d\}}$ can actually speed up the computation of $\psi_3$ asymptotically.)
For $\psi_2$, we can include $\psi_2$ itself.
(Recall the minor tweak we mentioned above.)
Overall,
we end up with the following definition of $\psi_3$:
\[ \psi_3(e,d) = \sum_f \psi_U \cdot \psi_V\cdot \psi_1 \cdot \psi_2 \cdot
\psi_{T/\{e,d\}}. \]
In $\text{\sf LogiQL}$, this sub-result is computed with two rules:
\begin{verbatim}
proj1(d,e) <- T(b,c,d,e). // projection rule
psi3[e,d] = s3 <- agg<<s3 = total(s1*s2)>> U(d,f),
V(e,f), psi1[f] = s1, psi2[e] = s2, proj1(d,e).
\end{verbatim}
After eliminating $f$, we are left with the following
\begin{eqnarray*}
\varphi(b,d) &=& \sum_c\sum_a\sum_e
\psi_
\psi_
\psi_
\psi_3\\
&=& \sum_c\sum_a
\psi_
\psi_
\sum_e
\psi_
\psi_T\\
&=& \sum_c\sum_a
\psi_
\psi_
\underbrace{
\sum_e
\psi_
\psi_T
\psi_{R/\{b\}} \psi_{S/\{c\}}}_{\psi_4(b,c,d)}
\end{eqnarray*}
leading to the following $\text{\sf LogiQL}$ rules
\begin{verbatim}
proj2(b) <- R(a,b).
proj3(c) <- S(a,c).
psi4[b,c,d] = s4 <- agg<<s4 = total(s3)>>
psi3[e,d] = s3, T(b,c,d,e), proj2(b), proj3(c).
\end{verbatim}
At this point, we have 3 factors left $\psi_R(a,b)$,
$\psi_S(a,c)$, and $\psi_4(b,c,d)$. We eliminate $a$ then $c$ straightforwardly:
\[
\sum_c\sum_a
\psi_
\psi_
\psi_4\\
=
\sum_c
\psi_4
\underbrace{
\sum_a
\psi_
\psi_
\psi_{4/\{b,c\}}
}_{\psi5(b,c)}\\
=
\sum_c \psi_4 \psi_5.
\]
Note that $\psi_{4/\{b, c\}}$ has values in $\{0, 1\}$ although $\psi_4$ can have any value in $\Z$.
The final $\text{\sf LogiQL}$ rules are
\begin{verbatim}
proj4(b,c) <- psi4[b,c,d] = s4. // indicator projection
psi5[b,c] = s5 <- agg<<s5 = count()>>
R(a,b), S(a,c), proj4(b,c).
output[b,d] = t <- agg<<t = total(s4*s5)>>
psi4[b,c,d] = s4, psi5[b,c] = s5.
\end{verbatim}
\label{ex:running:insideout}
\end{ex}
{\it\bf The general $\text{\sf FAQ}$ problem.}
The above strategy does not care if the variable aggregates where the same or
different: As long as $(\D, \oplus^{(n)}, \otimes)$ is a semiring, we can fold
the common factors and eliminate $X_n$. Thus, $\functionname{InsideOut}$ works almost as is
for a general $\text{\sf FAQ}$ instance (as opposed to $\text{\sf FAQ-SS}$).
Finally, when $\oplus^{(n)}=\otimes$ we simply swap the
two (identical) operators:
{\small
\begin{eqnarray*}
\varphi(\mv x_{[f]})
&=&
\cdots
\mathop{\textstyle{\bigoplus^{(n-1)}}}_{x_{n-1}}
\mathop{\textstyle{\bigoplus^{(n)}}}_{x_n}
\mathop{\textstyle{\bigotimes}}_{S\in\mathcal E} \psi_S(\mv x_S)\\
&=&
\cdots
\mathop{\textstyle{\bigoplus^{(n-1)}}}_{x_{n-1}}
\mathop{\textstyle{\bigotimes}}_{x_n \in \functionname{Dom}(X_n)}
\mathop{\textstyle{\bigotimes}}_{S\in\mathcal E} \psi_S(\mv x_S)\\
&=&
\cdots
\mathop{\textstyle{\bigoplus^{(n-1)}}}_{x_{n-1}}
\mathop{\textstyle{\bigotimes}}_{S\in\mathcal E}
\mathop{\textstyle{\bigotimes}}_{x_n \in \functionname{Dom}(X_n)}
\psi_S(\mv x_S)\\
&=&
\cdots
\mathop{\textstyle{\bigoplus^{(n-1)}}}_{x_{n-1}}
\mathop{\textstyle{\bigotimes}}_{S\notin\partial(n)}
\underbrace{\left(\psi_S(\mv x_S)\right)^{|\functionname{Dom}(X_n)|}}_{\psi'_S}
\mathop{\textstyle{\bigotimes}}_{S\in\partial(n)}
\underbrace{\mathop{\textstyle{\bigotimes}}_{x_n}
\psi_S(\mv x_S)}_{\psi_{S-\{n\}}}.
\end{eqnarray*}
}
We are left with an $\text{\sf FAQ}$-instance whose hypergraph is exactly
$\mathcal H' = \mathcal H-\{n\}$: the hypergraph obtained form $\mathcal H$ by removing
vertex $n$ from
the vertex set and all incident hyperedges.
The sub-problems are of the form of
{\em product marginalizations} of individual factors
$\psi_S$ for $S\in \partial(n)$,
each of which can be computed in linear time in $|\psi_S|$.
The product marginalization step is algorithmically
much easier because it does not create the intermediate factor
$\psi_{U_n-\{n\}}$.
As for $S\notin\partial(n)$, we replace $\psi_S$ by the power factor
$\psi'_{S}(\mv x_{S}) = \left(\psi_S(\mv x_S)\right)^{|\functionname{Dom}(X_n)|},$
which can be done in linear time with a $\log|\functionname{Dom}(X_n)|$ blowup using the
repeated squaring algorithm.
Note the key fact that this power is with respect to the product aggregate
$\otimes$. In most (if not all) applications of $\text{\sf FAQ}$, there is one additional
property: most of the time, $\otimes$ is an idempotent operator over the active domain.
For example, in the $\problemname{\#QCQ}$ problem $\otimes$ is the usual product operator and
the domain that it aggregates over is $\{0,1\}$ (before there is a sum outside).
In this case,
$\psi'_{S}(\mv x_{S}) = \left(\psi_S(\mv x_S)\right)^{|\functionname{Dom}(X_n)|} =
\psi_S(\mv x_S),$
and we do not need to spend the linear nor $\log$-blowup time.
For more details on product idempotence, see~\cite{faq-arxiv}.
{\it\bf $\text{\sf FAQ}$ sub-problems as natural joins.}
In the above we have explained how $\functionname{InsideOut}$ breaks a big problem into
smaller problems. In the product marginalization case, the sub-problems are
easy to solve: they can be solved in linear time. The most difficult problems,
however, are of the form~\eqref{eqn:sub:query}. This is exactly an $\text{\sf FAQ}$-query
where we marginalize out only one variable, with the remaining variables free.
Zooming in, problem~\eqref{eqn:sub:query} is of the form
\[ \psi_{U_n-\{n\}}(\mv x_{U_n-\{n\}}) :=
\bigoplus_{x_n} \bigotimes_{F\in\mathcal E_n}\psi_F, \]
where $\mathcal H_n=(U_n,\mathcal E_n)$ is the sub-$\text{\sf FAQ}$-query hypergraph.
The problem is solved by computing
$\psi_{U_n}(\mv x_{U_n}) := \bigotimes_{F\in\mathcal E_n}\psi_F$ first.
Once the $\psi_{U_n}$ is computed, marginalizing away $X_n$
to obtain $\psi_{U_n-\{n\}}$ is trivial.
Computing the inner product is a natural join problem in disguise. Each
input factor $\psi_S$
is represented using a table of tuples of the form
$[\mv x_S, \psi_S(\mv x_S)]$.
Essentially, $\mv x_S$ is the (compound)
key and $\psi_S(\mv x_S)$ is the value in this relation.
Again, recall that entries not in the table have $\psi_S$-value $\mv 0$.
Hence, to compute $\psi_{U_n}$ we can first join the tables $\psi_S$
using only the key space.
For each tuple $\mv x_{U_n}$ in the result of this join, we record the value
$\psi_{U_n}(\mv x) = \prod_{F\in\mathcal E_n} \psi_{F}(\mv x_F)$.
The runtime is dominated by the natural join's runtime.
{\it\bf Worst-case optimal join algorithms.}
Computing the natural join is a very well-studied problem with exciting new
developments in the past decade or so.
There are new {\em worst-case optimal} algorithms~\cite{LFTJ,NPRR12,skew,anrr}
that operate quite differently from traditional query plans, in the sense
that they no longer compute one pairwise join at a
time, but instead process the query globally.
While the vast majority of database engines today still rely on
traditional query plans, new, complex data analytics engines
are switching to worst-case optimal algorithms: $\text{\sf LogicBlox}$'s
engine~\cite{LB} is built on a worst-case optimal algorithm
called {\sf LeapFrog Triejoin}~\cite{LFTJ} ($\problemname{LFTJ}$), and the {\em Myria}
data analytics platform supports a variant of
$\problemname{LFTJ}$~\cite{DBLP:conf/sigmod/ChuBS15}.
We briefly outline these results here.
The generic form of the natural join problem can be posed in our hypergraph
language as
$Q = \ \Join_{F\in \mathcal E} R_F$, where $\mathcal H=(\mathcal V,\mathcal E)$ is the query
hypergraph.
The vertices of this hypergraph consist of all attributes.
Each hyperedge $F\in\mathcal E$ corresponds to an input relation $R_F$
whose attributes are $F$.
The natural join problem can be thought of as a constraint satisfaction
problem: each input relation $R_F$ imposes a constraint where a tuple
$\mv x_F$ satisfies the constraint if $\mv x_F \in R_F$.
A tuple $\mv x$ on all variables $\mathcal V$ is an output of the join if
the projection $\mv x_F$ satisfies $R_F$ for all $F\in\mathcal E$.
$\problemname{LFTJ}$~\cite{LFTJ} can be viewed as {\em backtracking-search} algorithm,
which was known some 50 years ago in the AI and constraint programming
world~\cite{DBLP:journals/cacm/DavisLL62,DBLP:dblp_journals/jacm/GolombB65}.
(In contrast, by saving intermediate results $\psi_{U_n-\{n\}}$ instead of re-computing them each time,
$\functionname{InsideOut}$ can be thought of as \emph{dynamic programming}.
The duality between
backtracking search and dynamic programming is
well-known~\cite{Rossi:2006:HCP:1207782}.)
$\problemname{LFTJ}$ fixes
some variable ordering $X_1,\dots,X_n$ of the query $Q$, then
performs ``leap-frogging'' to find the first binding $x_1$ that does not
yet violate any constraints $R_F$; once $x_1$ is found, it looks for the first binding $x_2$ such that the partial tuple $(x_1,x_2)$ does not violate any
constraint. The algorithm proceeds this way until either a full binding
$\mv x$ is constructed in which case $\mv x$ is an output, or no good binding
is found. For example, if no feasible binding for $x_3$ is found, then the
algorithm backtracks to the next good binding of $x_2$.
The first advantage of backtracking search is that it requires only
$O(1)$-extra space: it does not cache any computation. The second advantage,
amazingly, is that a join algorithm based on back-tracking search such as
$\problemname{LFTJ}$ or others in~\cite{NPRR12, skew} are worst-case optimal, in the sense
that the algorithm runs in time bounded by the worst-case output size.
To state the output size bound, we need the following notion. Define the
{\em fractional edge cover polytope} $\FECP(\mathcal H)$ associated with a hypergraph
$\mathcal H$ to be the set of all
vectors $\vec\lambda=(\lambda_F)_{F\in\mathcal E}$ satisfying the following linear
constraints:
\[
\vec\lambda \geq \mv 0, \text{ and }
\sum_{F \in \mathcal E, v \in F} \lambda_F \geq 1, \ \forall v \in \mathcal V.
\]
A vector $\vec\lambda \in \FECP(\mathcal H)$ is called a
{\em fractional edge cover} of $\mathcal H$.
The join output size is bounded above by $\prod_{F\in\mathcal E}|R_F|^{\lambda_F}$,
for {\em any} $\vec\lambda \in \FECP(\mathcal H)$.
The best bound $\text{\sf AGM}\xspace(\mathcal H)$, known as the $\text{\sf AGM}\xspace$-bound~\cite{AGM08,GM06},
is obtained by solving the linear program
\begin{equation}
\min \Bigl\{ \sum_{F\in\mathcal E}\lambda_F \log_2|R_F| : \vec\lambda\in
\FECP(\mathcal H)\Bigr\}.
\end{equation}
\begin{ex}
\label{ex:db:runtime}
Consider the query computing $\psi_3$ in
Example~\ref{ex:running:insideout}. The join query on the keys
has the following shape:
$Q = U(d,f) \Join V(e,f) \Join I(f) \Join J(e) \Join K(d,e).$
Then,
$\text{\sf AGM}\xspace(Q) =
|U|^{\lambda_{d,f}}
|V|^{\lambda_{e,f}}
|I|^{\lambda_f}
|J|^{\lambda_e}
|K|^{\lambda_{d,e}},
$
where $\vec\lambda$ is a fractional edge cover of the query's hypergraph.
Suppose all input relations have the same size $N$, then the optimal bound
is obtained by setting $\lambda_{d,f}=\lambda_{d,e}=\lambda_{e,f}=1/2$,
and $\lambda_d=\lambda_e=0$. Worst-case optimal algorithms run in time
$\tilde O(N^{3/2})$ for this instance. Any traditional join-tree based
plan runs in $\Omega(N^2)$-time for some input~\cite{NPRR12}.
Moreover, without the indicator projection of $T(b, c, d, e)$, there would be no $K(d, e)$ above, the best edge cover would be $\lambda_{d, f}=\lambda_{e, f}=1$, and the runtime would become $\Omega(N^2)$.
\end{ex}
{\it\bf Runtime analysis.}
Let $N$ denote the input size, $\outputsize$ the output size, and $K$ the
set of $k \in [n]$ for which $\oplus^{(k)} \neq \otimes$ (note that $[f]
\subseteq K$).
Also, let $\text{\sf AGM}\xspace(Q_k)$ denote the $\text{\sf AGM}\xspace$-bound on the $k$th sub-query's
hypergraph $\mathcal H_k$. Then, it is not hard to show~\cite{faq-arxiv} that the runtime
of $\functionname{InsideOut}$ is
\begin{equation}
\tilde O\Bigl(
N + \sum_{k \in K}\text{\sf AGM}\xspace(\mathcal H_k)+\outputsize
\Bigr).
\label{eqn:runtime1}
\end{equation}
The first term is input-preprocessing time, second
is the total subproblem solving time, and third is the unavoidable output
reporting time.
From~\eqref{eqn:runtime1},
we can write down a precise expression for the runtime of
$\functionname{InsideOut}$.
Minimizing the resulting (somewhat complicated) expression leads to the dynamic
programming algorithm for the $\problemname{MCM}$ problem
and the $\problemname{FFT}$ algorithm for
the $\problemname{DFT}$ (see~\cite{faq-arxiv} for details).
In the above discussion, we assumed that variables were eliminated in order
$X_n,X_{n-1},\dots,X_{1}$. However, there is no reason to force
$\functionname{InsideOut}$ to follow this particular order.
In particular, there might be a different variable ordering for which
expression~\eqref{eqn:runtime1} is {\em a lot} smaller and the algorithm still
works correctly on that ordering (see~\cite{faq-arxiv}).
This is where the main technical contributions of our work in~\cite{faq-arxiv} begin.
We need to answer the following two fundamental questions:
{\it Question 1.}
How do we know which variable orderings are equivalent to the original
$\text{\sf FAQ}$-query expression?
{\it Question 2.}
How do we find the ``best'' variable ordering among all
equivalent variable orderings?
In the next two sections, we sketch how we answered the above two
questions and followup questions in theory and in practice.
\section{Introduction}
The following fundamental problems from diverse domains share
a common algebraic structure involving (generalized) sums of products.
\begin{ex}($\problemname{Matrix Chain Multiplication}$ ($\problemname{MCM}$))
\label{ex:matrix mult}
Given a series of matrices $\mv A_1, \dots, \mv A_n$ over some field $\F$,
where the dimension of $\mv A_i$ is $p_i \times p_{i+1}$, $i \in [n]$,
we wish to compute the product $\mv A = \mv A_1 \cdots \mv A_n$.
The problem can be reformulated as follows.
There are $n+1$ variables $X_1,\dots,X_{n+1}$ with domains
$\functionname{Dom}(X_i) = [p_i]$, for $i\in[n+1]$.
For $i \in [n]$, matrix $\mv A_i$ can be viewed as a function of two variables
\[ \psi_{i,i+1}: \functionname{Dom}(X_i)\times \functionname{Dom}(X_{i+1}) \to \F, \]
where $\psi_{i,i+1}(x,y) = (\mv A_i)_{xy}$.
The $\problemname{MCM}$ problem is to compute the output function
\[ \varphi(x_1,x_{n+1}) =
\sum_{x_2 \in\functionname{Dom}(X_2)} \cdots \sum_{x_n \in \functionname{Dom}(X_n)} \prod_{i=1}^n
\psi_{i,i+1}(x_i,x_{i+1}).
\]
\end{ex}
\begin{ex}(Maximum A Posteriori ($\problemname{MAP}$) queries in
probabilistic graphical models ($\problemname{PGM}$s))
Consider a discrete graphical model represented by a hypergraph
$\mathcal H = (\mathcal V, \mathcal E)$.
There are $n$ discrete random variables $\mathcal V = \{X_1,\dots,X_n\}$ on finite
domains $\functionname{Dom}(X_i)$, $i\in[n]$, and $m=|\mathcal E|$ {\em factors}
\[ \psi_S : \prod_{i \in S} \functionname{Dom}(X_i) \to \R_+, \ S \in \mathcal E. \]
A typical inference task is to compute the marginal $\problemname{MAP}$ estimates,
written in the form
\[ \varphi(x_1,\dots,x_f) =
\max_{x_{f+1}\in\functionname{Dom}(X_{f+1})} \cdots
\max_{x_{n}\in\functionname{Dom}(X_{n})} \prod_{S\in\mathcal E}\psi_S(\mv x_S).
\]
\end{ex}
\begin{ex}(Conjunctive query in RDBMS)
\label{ex:db:basic}
Consider a schema with the following input relations:
$R(a,b)$,
$S(b,c)$,
$T(c,a)$,
where for simplicity let us say all attributes are integers.
Consider the following query:
\begin{verbatim}
SELECT R.a
FROM R, S, T
WHERE R.b = S.b AND S.c = T.c AND T.a = R.a;
\end{verbatim}
The above query can be reformulated as follows.
Relation $R(a,b)$ is modeled by a function $\psi_R(a,b) \to
\{\functionname{true}, \functionname{false}\}$, where $\psi_R(a,b)=\functionname{true}$ iff $(a,b) \in R$,
and relations $S(b,c)$ and $T(c, a)$ are modeled by similar functions $\psi_S(b, c), \psi_T(c, a)$. Now, computing the above query basically corresponds to computing the function $\varphi(a)\rightarrow\{\functionname{true},\functionname{false}\}$, defined as:
\[\varphi(a)=
\bigvee_b \bigvee_c \psi_R(a, b)\wedge \psi_S(b, c)\wedge \psi_T(c, a).\]
\end{ex}
\begin{ex}($\problemname{\# Quantified Conjunctive Query}$ ($\problemname{\#QCQ}$))
\label{ex:sqcq}
Let $\Phi$ be a first-order formula of the form
\[ \Phi(X_1,\dots,X_f) = \DarkBlue{Q_{f+1}} X_{f+1} \cdots
\DarkGreen{Q_n} X_n
\left( \bigwedge_{R\in\functionname{atoms}(\Phi)} R \right), \]
where $Q_i \in \{\exists, \forall\}$, for $i>f$.
The $\problemname{\#QCQ}$ problem is to {\em count} the number of tuples in relation $\Phi$
on the free variables $X_1,\dots,X_f$.
To reformulate $\problemname{\#QCQ}$,
construct a hypergraph $\mathcal H = (\mathcal V, \mathcal E)$ as follows: $\mathcal V$ is the set
of all variables $X_1,\dots,X_n$, and for each $R \in \functionname{atoms}(\Phi)$
there is a hyperedge $S = \functionname{vars}(R)$ consisting of all variables in $R$.
The atom $R$ can be viewed as a function indicating
whether an assignment $\mv x_S$ to its variables is satisfied by the atom;
namely $\psi_S(\mv x_S) = 1$ if $\mv x_S \in R$ and $0$ otherwise.
For each $i \in \{f+1,\dots,n\}$ we define an aggregate operator
\[ \textstyle{\bigoplus^{(i)} =
\begin{cases}
\max & \text{ if } Q_i = \exists, \\
\times & \text{ if } Q_i = \forall.
\end{cases}}
\]
Then, the $\problemname{\#QCQ}$ problem above is to compute the {\em constant} function
\[ \varphi = \sum_{x_1\in\functionname{Dom}(X_1)} \cdots \sum_{x_f\in\functionname{Dom}(X_f)}
\mathop{\textstyle{\DarkBlue{\bigoplus^{(f+1)}}}}_{x_{f+1}\in\{0,1\}}
\cdots
\mathop{\textstyle{\DarkGreen{\bigoplus^{(n)}}}}_{x_{n}\in\{0,1\}}
\prod_{S\in\mathcal E}\psi_S(\mv x_S).
\]
\end{ex}
It turns out that these and dozens of other fundamental problems from
constraint satisfaction ($\problemname{CSP}$),
databases,
matrix operations,
$\problemname{PGM}$ inference,
logic,
coding theory,
and complexity theory
can be viewed as special instances
of a generic problem we call the
{\bf F}unctional {\bf A}ggregate {\bf Q}uery, or
the $\text{\sf FAQ}$ problem, which we define
next. (See~\cite{faq-arxiv,AM00} for many more examples.)
Throughout the paper, we use the following convention.
Uppercase $X_i$ denotes a variable, and lowercase $x_i$ denotes a value in
the domain $\functionname{Dom}(X_i)$ of the variable.
Furthermore, for any subset $S\subseteq [n]$, define
\begin{align*}
\mv X_S &= (X_i)_{i\in S},
& \mv x_S = (x_i)_{i\in S} \in \prod_{i\in S}\functionname{Dom}(X_i).
\end{align*}
In particular, $\mv X_S$ is a tuple of variables and $\mv x_S$ is a
tuple of specific values with support $S$.
The input to $\text{\sf FAQ}$ is a set of functions and the output is a function
computed using a series of aggregates over the variables and
input functions.
More specifically, for each $i\in [n]$, let $X_i$ be a variable on some discrete
domain $\functionname{Dom}(X_i)$, where $|\functionname{Dom}(X_i)|\ge 2$.
The $\text{\sf FAQ}$ problem is to compute the following function
\begin{equation}
\label{eq:gen-faq}
\varphi(\mv x_{[f]}) =
\mathop{\DarkBlue{\textstyle{\bigoplus^{(f+1)}}}}_{x_{f+1}\in\functionname{Dom}(X_{f+1})}
\cdots
\mathop{\DarkGreen{\textstyle{\bigoplus^{(n)}}}}_{x_{n}\in\functionname{Dom}(X_{n})}
\mathop{\textstyle{\bigotimes}}_{S\in\mathcal E}\psi_S(\mv x_S),
\end{equation}
where
\begin{itemize}
\item $\mathcal H=(\mathcal V,\mathcal E)$ is a multi-hypergraph. $\mathcal V=[n]$ is the index
set of the variables $X_i$, $i\in [n]$. Overloading notation, $\mathcal V$ is
also referred to as the set of variables.
\item The set $F = [f]$ is the set of {\em free variables} for some
integer $0\le f\le n$. Variables in $\mathcal V-F$ are called
{\em bound variables}. (Free and bound are logic terminologies. Free
variables are group-by variables in database nomenclature.)
\item $\D$ is a fixed domain, such as $\{0,1\}$, $\R^+$, $\Z$.
\item For every hyperedge $S\in\mathcal E$,
$\psi_S:\prod_{i\in S} \functionname{Dom}(X_i) \to \D$ is an
{\em input} function (also called a {\em factor}).
\item For every bound variable $i>f$, $\oplus^{(i)}$ is a binary (aggregate)
operator on the domain $\D$.
\item And, for each bound variable $i>f$ either $\oplus^{(i)}=\otimes$ or
$(\D,\oplus^{(i)},\otimes)$ forms a commutative semiring \footnote{A triple
$(\D,\oplus,\otimes)$ is a {\em commutative semiring} if $\oplus$
and $\otimes$ are commutative binary operators over $\D$ satisfying the
following:
(1) $(\D, \oplus)$ is a commutative monoid with an additive
identity, denoted by $\mv 0$.
(2) $(\D, \otimes)$ is a commutative
monoid with a multiplicative
identity, denoted by $\mv 1$.
(3) $\otimes$ distributes over $\oplus$.
(4) For any element $e \in \mathbf D$, $e \otimes \mv 0
= \mv 0 \otimes e = \mv 0$.
} (with the same
$\mv 0$ and $\mv 1$). Informally, this means that we can do addition and multiplication over $\D$ and still remain in the same set.
If $\oplus^{(i)} = \otimes$, then $\oplus^{(i)}$ is called a
{\em product aggregate};
otherwise, it is a {\em semiring aggregate}. (We assume that there is at least one semiring aggregate.)
\end{itemize}
Because for $i > f$ every variable $X_i$ has its own aggregate $\bigoplus^{(i)}$
over all values $x_i \in \functionname{Dom}(X_i)$, in the rest of the paper we will
write $\bigoplus^{(i)}_{x_i}$ to mean
$\displaystyle{\mathop{\textstyle{\bigoplus^{(i)}}}_{x_{i}\in\functionname{Dom}(X_{i})}}$.
We will refer to $\varphi$ as an {\em $\text{\sf FAQ}$-query}.
We use $\text{\sf FAQ-SS}$ to denote
the special case when there is a \emph{S}ingle \emph{S}emiring aggregate,
i.e. $\oplus^{(i)}=\oplus, \forall i>f$, and $(\D, \oplus, \otimes)$ is a
semiring~\cite{AM00}.
\begin{ex}(Aggregate query in RDBMS)
\label{ex:db}
Consider the following query over relations
$R(a,b)$,
$S(a,c)$,
$T(b,c,d,e)$,
$U(d,f)$,
$V(e,f)$,
$W(e,g)$,
$Y(f,h)$, where all attributes are integers:
\begin{verbatim}
SELECT R.b, U.d, sum(W.e)
FROM R, S, T, U, V, W, Y
WHERE R.a = S.a AND R.b = T.b AND S.c = T.c
AND T.d = U.d AND T.e = V.e AND W.e = V.e
AND U.f = V.f AND Y.f = V.f GROUP BY R.b, U.d;
\end{verbatim}
We now explain how the above query can be reduced to an $\text{\sf FAQ}$
instance. Relation $R(a,b)$ is modeled with a function $\psi_R(a,b) \to
\{0,1\}$, where $\psi_R(a,b)=1$ iff $(a,b) \in R$. Similarly, we can think of
relations $S$, $T$, $U$, $V$, and $Y$ as functions
$\psi_S$,
$\psi_T$,
$\psi_U$,
$\psi_V$,
$\psi_Y$,
with $\{0,1\}$ values.
We single out one relation $W(e,g)$ where the modeling is different:
$\psi_W(e,g) = e$ if $(e,g) \in W$ and $0$ otherwise.
The corresponding $\text{\sf FAQ}$-query is
\[ \varphi(b, d) = \sum_a \sum_c \sum_e \sum_f \sum_g \sum_h
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\psi_
\]
(For readability, we did not write the argument lists of the functions $\psi_R, \psi_S$, etc. They
should be obvious from context.)
Note that a tuple in the output of the aggregate query has the schema $(b,d,\varphi(b,d))$.
The corresponding hypergraph is shown in
Fig.~\ref{fig:example:graph}. The set of free variables is $F = \{b, d\}$.
The domain is $\mv D = \Z$, the set of integers.
Note also that the above reduction to $\text{\sf FAQ}$ still works if we replace sum by another aggregate, e.g., max.
In order to explain later the connection of $\functionname{InsideOut}$ to query rewriting, we also write the above query in $\text{\sf LogiQL}$, an extension of Datalog
supported by the $\text{\sf LogicBlox}$ engine~\cite{LB}:
\begin{verbatim}
Q[b, d] = s <- agg<<s = total(e)>> R(a,b), S(a,c),
T(b,c,d,e), U(d,f), V(e,f), W(e,g), Y(f,h).
\end{verbatim}
In the above, {\tt agg} is short for aggregate, {\tt total} is equivalent to {\tt sum}
in {\tt SQL}, the notation {\tt Q[b,d]=s} means that the head predicate is
{\tt Q(b,d,s)} where {\tt (b,d)} is a key, hence the query computes
{\tt Q(b,d,sum(e))}.
\end{ex}
\input{fig_example}
The above example illustrates several important points.
First, when we defined the $\text{\sf FAQ}$ problem we did not specify how the input and
output factors are represented. The representation choice turns out to make a
huge difference in computational complexity~\cite{faq-arxiv}.
However, in practical applications the representation is usually the obvious one:
an input factor $\psi_S(\mv X_S)$ can be thought of as a table of tuples
$[\mv x_S, \psi_S(\mv x_S)]$, with the implicit assumption that if $\mv x_S$ is
not in the table then its $\psi_S$-value is $\mv 0$. (This is the additive identity
$\mv 0$ of the domain $\mv D$.)
Second, the reduction to $\text{\sf FAQ}$ is only at the syntax level. No real data conversion is
necessary. All the data we need to obtain the functions $\psi_R$,
$\psi_T$ etc. are already in the input relations.
Third, in the mathematical definition of $\varphi(b, d)$ above, the domains of all
variables are integers and so we have {\em infinite} sums. We could have
restricted all variables to their active domains; but that is not necessary
because summing over all integers or over the active domains give identical
answer: tuples not present are assumed to have values $\mv 0$.
Now that we have established the scope of $\text{\sf FAQ}$, in the remainder of
this paper we show a perhaps surprising result that an $\text{\sf FAQ}$ problem can be
solved by {\em one} simple yet efficient algorithm. The algorithm can be implemented as a set
of ordinary database queries. The runtime matches or improves upon the best known
runtimes in many application areas that the $\text{\sf FAQ}$ framework captures.
The runtime depends on the order of variable aggregates in the $\text{\sf FAQ}$ expression,
which naturally leads us to the question of how to re-order those aggregates to obtain the best runtime without changing the semantic meaning of the expression.
\section{Practical implications}
In this section we address two questions the readers might have regarding
$\functionname{InsideOut}$: (1) hurdles one might face in a practical implementation of
$\functionname{InsideOut}$, and (2) whether practical implications are as good as what the
theory says.
\paragraph*{Additional hurdles and how to solve them}
There are a couple of problems we have to solve to implement $\functionname{InsideOut}$
effectively.
The first problem is, in real-world queries, we do not just have materialized predicates as inputs,
we also have predicates such as $a<b$, $a+b=c$, negations and so on.
These predicates do not have a ``size.''
To solve this problem, one solution is to set the ``size'' of those
predicates to be $\infty$
while computing the $\text{\sf AGM}\xspace$-bound. For instance, if we have a sub-query of the form
$Q \leftarrow R(a,b),S(b,c),a+b=c$, where $R$ and $S$ are input
materialized predicates of size $N$, then by setting the size of $a+b=c$
to be infinite, $\text{\sf AGM}\xspace(Q) = N^2$. This solution does
not work for two reasons. (1) If we knew $a+b=c$, then it is easy to infer
that $|Q| \leq N$ and also to compute $Q$ in time $\tilde O(N)$:
scan over tuples in $R$, use $a+b=c$ to compute $c$, and see if $(b,c) \in S$.
In other words, the $\text{\sf AGM}\xspace$-bound is no longer tight.
(2) This solution may give an $\infty$-bound when the output size is clearly
bounded. Consider, for example, the query $Q \leftarrow R(a),S(b),a+b=c$; in
this case, $\{a,b,c\}$ is the only hyperedge covering vertex $c$ in the
fractional edge cover.
Our implementation at $\text{\sf LogicBlox}$ makes use of generalizations of $\text{\sf AGM}\xspace$ to
queries with functional dependencies and immaterialized predicates (such as
$a+b=c$). These new bounds are based on a linear program
whose variables are marginal entropies~\cite{DBLP:conf/pods/KhamisNS16,panda}.
The second problem is to select a good variable ordering to run $\functionname{InsideOut}$
on. In principle, one does not have to use the $\text{\sf AGM}\xspace$-bound or the
bounds from~\cite{DBLP:conf/pods/KhamisNS16,panda} to estimate the cost of
an $\text{\sf FAQ}$ subquery. If one were to implement $\functionname{InsideOut}$ inside any RDBMS, one
could poll that RDBMS's optimizer to figure out the cost of a
given variable ordering. However, there are $n!$ variable orderings, and
optimizer's cost estimation is time-consuming. Furthermore, some subqueries
have inputs which are intermediate results.
Hence, it is much faster to compute a variable ordering minimizing the
$\functionname{faqw}$ of the query, defined on the bounds
in~\cite{DBLP:conf/pods/KhamisNS16,panda}. As the problem is $\complexityclass{NP}$-hard,
either an approximation algorithm~\cite{faq} or a greedy heuristic suffices
in our experience.
$\functionname{InsideOut}$ is bottom-up dynamic programming. We can also solve $\text{\sf FAQ}$
queries with top-down (memoized)
dynamic programming. In hindsight, this was the approach
that Bakibayev et al.~\cite{DBLP:journals/pvldb/BakibayevKOZ13} and
Olteanu and Z\'avodn\'y~\cite{OZ15} took to solve $\text{\sf FAQ}$ over a single
semiring.
We can also limit the amount of memoization in a top-down strategy
to attain performance gain in some cases~\cite{KalinskyEK17}.
\paragraph*{Practical Impact}
It is trivial to construct classes of queries on real datasets for which
$\functionname{InsideOut}$-style of algorithms gives arbitrarily large speedups over
traditional RDBMSs. In fact, even when dynamic programming does not take
effect, the speedup of backtracking search (and thus
worst-case optimal algorithms) over traditional query plans is already
huge~\cite{DBLP:conf/sigmod/NguyenABKNRR14}.
The impact of the $\text{\sf FAQ}$-framework and the $\functionname{InsideOut}$ algorithm, however,
go much beyond these toy queries (even when run on real datasets).
$\functionname{InsideOut}$ is a component of $\text{\sf LogicBlox}$'s effort to extend $\text{\sf LogiQL}$ to be a
probabilistic programming language~\cite{DBLP:conf/icdt/BaranyCKOV16}, as
part of DARPA's PPAML and MUSE programs. The component the algorithm handles is
inference in discrete graphical models.
Learning from the beautiful work of Olteanu and
Schleich~\cite{DBLP:journals/pvldb/OlteanuS16, DBLP:conf/sigmod/SchleichOC16},
we realized~\cite{indblearn}
that $\functionname{InsideOut}$
can be used to train a large class of machine learning models {\em inside}
the database. Our implementation showed orders of magnitude speedup over the
traditional data modeler route
of exporting the data and running it through \texttt{R} or \texttt{Python}.
These models are trained with different variations of gradient descents,
whose (pre-)computation steps are $\text{\sf FAQ}$ queries. What is much more interesting
than the vanilla $\text{\sf FAQ}$ framework we presented above is that, in these
applications, we want to compute {\em many} (in the 100K-range) $\text{\sf FAQ}$ queries all at
once, making dynamic programming much more crucial to the performance.
Another related approach was considered in~\cite{Kumar:2015:LGL:2723372.2723713}.
\section{Theoretical contributions}
\input{contributions}
To answer the above questions, we start with some definitions.
A variable ordering $\sigma$ is {\em $\varphi$-equivalent} iff
permuting the variable aggregates of $\varphi$ using $\sigma$ gives an expression $\varphi'$
that is \emph{semantically-equivalent} to $\varphi$,
i.e.\ that always returns the same output as $\varphi$ {\em no matter} what the input is.
Let $\problemname{EVO}(\varphi)$ denote the set of all $\varphi$-equivalent variable
orderings.
\begin{ex}
\label{ex:EVO}
The $\text{\sf FAQ}$ query $\varphi'$ below is $\varphi$-equivalent.
\begin{eqnarray*}
\varphi &=& \sum_a\sum_d \max_b \sum_c \psi_1(a, b) \psi_2(a, c) \psi_3 (c, d),\\
\varphi' &=& \sum_a\sum_c \sum_d \max_b \psi_1(a, b) \psi_2(a, c) \psi_3 (c, d).
\end{eqnarray*}
This is because $\varphi$ can be written as
\[\varphi = \sum_a\left[\left(\sum_d \sum_c \psi_2(a, c) \psi_3 (c, d)\right)
\left(\max_b \psi_1(a, b)\right)\right].\]
\end{ex}
Now, for any $\sigma \in \problemname{EVO}(\varphi)$, let $\mathcal H^\sigma_k$ denote the $k$th
sub-query's hypergraph when we run $\functionname{InsideOut}$ on $\sigma$. Ideally, we would
like to find $\sigma$ minimizing the expression
$\sum_{k \in K} \text{\sf AGM}\xspace(\mathcal H^\sigma_k)$. However, this expression is
data-dependent and thus it is a bit difficult to handle in a
mathematically clean way. We simplify our objective by approximating the
bound~\eqref{eqn:runtime1} a little:
we upperbound $\text{\sf AGM}\xspace(\mathcal H^\sigma_k)$ by the {\em fractional edge cover number} of the
subgraph $\mathcal H^\sigma_k$,
i.e. $\text{\sf AGM}\xspace(\mathcal H^\sigma_k) \leq N^{\rho^*(\mathcal H^\sigma_k)}$, where
$\rho^*(\mathcal H^\sigma_k) := \min\{ \sum_{F\in \mathcal E}\lambda_F :
\vec\lambda\in\FECP(\mathcal H^\sigma_k)\}$.
Then, \eqref{eqn:runtime1} is upperbounded by
$\tilde O\left(
N^{\max_{k\in K} \rho^*(\mathcal H^\sigma_k)}+\outputsize
\right)$;
$\functionname{InsideOut}$ on variable ordering $\sigma$ runs in $\tilde
O(N^{\functionname{faqw}(\sigma)}+\outputsize)$-time, where
$\functionname{faqw}(\sigma) := \max_{k\in K} \rho^*(\mathcal H^\sigma_k)$.
Thus, to have the best runtime we would like to select an equivalent
ordering $\sigma$ with the smallest exponent $\functionname{faqw}(\sigma)$, called the
{\em $\text{\sf FAQ}$-width} of an $\text{\sf FAQ}$-query:
\begin{equation}
\functionname{faqw}(\varphi) :=
\min \left\{ \functionname{faqw}(\sigma) \ | \ \sigma \in \problemname{EVO}(\varphi) \right\}
\label{eqn:faqw}
\end{equation}
\begin{ex}
In Example~\ref{ex:EVO}, the original order in $\varphi$ has an $\text{\sf FAQ}$-width of 2, because eliminating $c$ first corresponds to joining $\psi_2$ and $\psi_3$ in time $\Omega(N^2)$.
In contrast, the order in $\varphi'$ has an $\functionname{faqw}$ of 1, allowing to evaluate $\varphi$ in time $O(N)$.
\end{ex}
To solve the optimization problem~\eqref{eqn:faqw}, the first problem we have to
address is to precisely characterize the set $\problemname{EVO}(\varphi)$.
Our approach, sketched in Fig.~\ref{fig:schematic}, is to construct an
{\em expression tree} of the $\text{\sf FAQ}$ query $\varphi$. The expression tree defines
a partially ordered set on the variables called the {\em precedence poset}.
Then, to complete the characterization of $\problemname{EVO}$, we show that every ordering in
$\problemname{EVO}$ is {\em component-wise equivalent} ($\problemname{CWE}$) to a linear extension of $\problemname{LinEx}(P)$. (See~\cite{faq-arxiv} for details.)
Thus, if we do not care about query complexity, we can take the orange path in
Fig~\ref{fig:schematic} and bruteforcedly compute an optimal variable ordering
$\sigma^*$, run it through $\functionname{InsideOut}$, for a total runtime of
$\tilde O(N^{\functionname{faqw}(\varphi)}+\outputsize)$.
However, in some $\text{\sf FAQ}$-framework's applications such as in graphical models, we
cannot simply sweep query complexity under the rug. Moreover, computing the
$\functionname{faqw}$ is $\complexityclass{NP}$-hard because $\functionname{faqw}$ is a strict generalization of
the fractional hypertree width ($\functionname{fhtw}$),
which is $\complexityclass{NP}$-hard~\cite{2016arXiv161101090F} to compute. Hence, we
find a good approximation for the $\functionname{faqw}$. This is the green path in
Fig.~\ref{fig:schematic}: from the expression tree, we construct a tree
decomposition for $\mathcal H$; then, from the GYO-elimination order of this tree
decomposition we obtain a variable ordering $\bar \sigma$ for which we can
show that $\functionname{faqw}(\bar \sigma) \leq \functionname{faqw}(\sigma^*) + g(\functionname{faqw}(\sigma^*))$, where
$g$ is any known approximation of $\functionname{fhtw}$
(the best of which is due to Marx~\cite{Marx:2010:AFH:1721837.1721845}).
\begin{ex}
\label{ex:db:GYO}
The variable ordering used earlier in Example~\ref{ex:running:insideout}
is a GYO-elimination order for the tree decomposition in Fig.~\ref{fig:example:tree}.
(In GYO, when we eliminate variable $X_n$, the set $U_n$ becomes a bag of the tree decomposition whose children are the bags corresponding to $\partial(n)$.)
In particular, bags $\{f, h\}$, $\{e, g\}$, and $\{d, e, f\}$ resulted from eliminating $h, g,$ and $f$ respectively.
The tree decomposition has width $\functionname{fhtw}=3/2$, same as the $\functionname{faqw}$ of the variable ordering.
\end{ex}
\input{results}
From these ideas, we obtained many corollaries, some of which are summarized
in Table~\ref{tab:results}.
The results in Table~\ref{tab:results} span three areas:
(1) $\problemname{CSP}$s and Logic;
(2) $\problemname{PGM}$s and
(3) Matrix operations.
Except for joins, problems in area (1) need the full generality of $\text{\sf FAQ}$ formulation, where
$\functionname{InsideOut}$ either improves upon existing results or yields new results.
Problems in area (2) can already be reduced to $\text{\sf FAQ-SS}$. Here, $\functionname{InsideOut}$
improves upon known results since it takes advantage of Grohe and Marx's more
recent fractional hypertree width bounds.
Finally, problems in area (3) of Table~\ref{tab:results} are classic.
$\functionname{InsideOut}$ does not yield anything new here, but it is intriguing to be able
to explain the textbook dynamic programming algorithm for
$\problemname{Matrix-Chain Multiplication}$~\cite{MR2002e:68001}
as an algorithm to find a good
variable ordering for the corresponding $\text{\sf FAQ}$-instance.
The $\problemname{DFT}$ result is a re-writing of Aji and McEliece's
observation \cite{AM00}.
|
\section{Conclusion}
|
\section{Introduction}\label{sec:intro}
Let $M$ be a compact Riemannian manifold of dimension $d$ without boundary.
For $0\leq p\leq d$, we will denote by $\Delta_p$ the Hodge-Laplace operator acting on smooth $p$-forms of $M$.
It is well known that the spectrum of $\Delta_p$, denoted by $\spec_p(M)$, is a discrete set of non-negative real numbers, repeated according to its finite multiplicity, and tending to infinity.
If $\spec_p(M)=\spec_p(M')$, $M$ and $M'$ are said to be \emph{$p$-isospectral}.
The following question appears naturally.
\begin{question}\label{question1}
For a given subset $I$ of $\{0,\dots,d\}$.
Are there $d$-dimensional non-isometric compact Riemannian manifolds $M$ and $M'$ such that they are $p$-isospectral if and only if $p$ is in $I$?
\end{question}
The main goal of this article is to study Question~\ref{question1}, for every choice $I\subset \{0,\dots,d\}$, in the class of lens spaces, and more generally in the class of odd-dimensional lens orbifolds.
A lens space is an orientable manifold with positive constant sectional curvature and cyclic fundamental group.
Its dimension is odd and it has the form $\Gamma\ba S^{2n-1}$ with $\Gamma$ a cyclic subgroup of $\SO(2n)$ acting freely on $S^{2n-1}$.
Relaxing the free action condition of $\Gamma$ we obtain a lens orbifold.
We will always assume that the lens orbifolds are odd-dimensional.
Lens spaces has been used many times as a test case for spectral questions, since their spectra can be explicitly computed.
Ikeda used generating functions to encode the $p$-spectra of a spherical space form.
This idea was very useful to construct various isospectral examples (\cite{Ikeda80_isosp-lens}, \cite{Ikeda83}, \cite{Ikeda88}, \cite{GornetMcGowan06}) and also to prove spectral rigidity results (\cite{IkedaYamamoto79}, \cite{Yamamoto80}, \cite{Ikeda80_3-dimI}, \cite{Ikeda80_3-dimII}, \cite{Ikeda97}).
Among many other results, Ikeda showed in \cite{Ikeda88} for each $p_0\geq0$ a pair of lens spaces which are $p$-isospectral for all $p$ satisfying $0\leq p\leq p_0$, and are not $(p_0+1)$-isospectral.
Subsections~\ref{subsec:Ikeda'sapproach} and \ref{subsec:Ikeda'sexamples} give a summary on some of these results.
Gornet and McGowan~\cite{GornetMcGowan06} reactivated the use of lens spaces in spectral questions by making a computational study of $p$-isospectral lens spaces (see Subsection~\ref{subsec:GornetMcGowan} for details).
They used Ikeda's generating functions to check whether two lens spaces are $p$-isospectral.
Similarly, Shams Ul Bari~\cite{Shams11} found several examples of $0$-isospectral lens orbifolds.
Also following Ikeda's approach, the Dirac operator was also considered on spin lens spaces in \cite{Bar91thesis} and \cite{Boldt15}, and in spherical space forms in \cite{Bar96}.
Very recently, Bari and Hunsicker~\cite{ShamsHunsicker17} proved the non-existence for non-isometric $0$-isospectral lens orbifolds in dimension $3$ and $4$.
By using standard representation theory on compact groups, Miatello, Rossetti and the author~\cite{LMR-onenorm} relate the $p$-spectrum of a lens space with the number of vectors of fixed one-norm in certain associated sublattice of $\Z^n$.
This was used to show the first examples of compact Riemannian manifolds that are $p$-isospectral for all $p$ but are not strongly isospectral.
The articles \cite{LMR-survey}, \cite{BoldtLauret-onenormDirac}, \cite{Lauret-spec0cyclic} and \cite{Lauret-multip} follow this approach and \cite{DeFordDoyle14} study in detail the examples of all-$p$-isospectral pairs in \cite{LMR-onenorm}.
The articles \cite{MohadesHonori16}, \cite{MohadesHonori16b} are also related to this approach.
In order to explain in detail the main goals of this article, we introduce some useful notation.
\begin{definition}\label{notation:I-isospectral}
For $I$ a subset of $\{0,\dots,d\}$, we will say that a family (two or more) of $d$-dimensional manifolds are \emph{$I$-isospectral} if they are mutually $p$-isospectral for all $p\in I$, and for any $p\not\in I$ there are at least two elements in the family that are not $p$-isospectral.
\end{definition}
In other words, $I$ is the largest subset of $\{0,\dots,d\}$ such that any two elements in the family are $p$-isospectral for all $p$ in it.
In particular, any subfamily of an $I$-isospectral family will be $I'$-isospectral for some $I'\supseteq I$.
For an orientable $d$-dimensional compact Riemannian manifold $M$, $\spec_p(M)=\spec_{d-p}(M)$ for all $p$.
Hence, when the underlying manifolds are orientable, we abbreviate $I$-isospectral for some $I\subset \{0,1,\dots, \lfloor \frac{d}{2} \rfloor\}$ to $I'$-isospectral when $I'=I\cup \{d-p:p\in I\}$.
Lens orbifolds are orientable of odd dimension, say $d=2n-1$, thus we will be considering subsets of $\{0,\dots,n-1\}$.
The aim of this article is to make a computational study of $I$-isospectral lens orbifolds.
The appendix \cite{Lauret-appendix} includes, for low values of $n$ and $q$, all families of $(2n-1)$-dimensional $I$-isospectral lens orbifolds with fundamental group of order $q$.
Section~\ref{sec:data} includes summaries of these computational results, by showing the subsets $I$ of $\{0,\dots,n-1\}$ for which there exists an $I$-isospectral family.
Section~\ref{sec:evidencedfacts} proves several facts evidenced from the data by using the tools introduced in Section~\ref{sec:preliminaries}.
All computations were made by using \cite{Sage}.
We end this section by listing the most interesting conclusions.
\begin{enumerate}
\item The most common obstruction to the existence of $I$-isospectral families is the `hole obstruction' in Proposition~\ref{prop:hole-obstruction}. If two spherical orbifolds (e.g.\ lens orbifolds) are $(p-1)$-isospectral and $(p+1)$-isospectral, then they also are $p$-isospectral. In other words, the subset $I$ cannot contain a `simple hole'.
\item It is well known that a lens space cannot be $0$-isospectral to a lens orbifold with singularities since they share a common Riemannian cover (see Subsection~\ref{subsec:Orbifoldsmanifolds}).
The computational results give strong evidences that a lens space cannot be $p$-isospectral to a lens orbifold with singularities for any $p$ (see Conjecture~\ref{conj:orbifold/manifold}).
\item We prove the non-existence of a pair of $I$-isospectral lens spaces for some choices of $I$.
Namely, when $I\subset \{0,\dots,n-1\}$ has $n-1$ elements (see Theorem~\ref{thm:[0,n-2]-isosp}), and when $I$ has $n-2$ elements and it is different to $\{0,\dots,n-3\}$ (see Theorem~\ref{thm:n-2elements}).
\item Gornet and McGowan in \cite{GornetMcGowan06} were interested in the existence of pairs of $p$-isospectral lens spaces for some $p>0$, which are not $0$-isospectral.
We found several such examples from dimension $5$ on among lens orbifolds.
We also found such examples among lens spaces from dimension $11$ on (see Remark~\ref{rem:p-isospp>0}).
\item There is evidence that the isotropy type of the singular points of $0$-isospectral lens orbifolds coincide (see Remark~\ref{rem:L(q/d;s)orbifolds}).
Furthermore, one can see that most $I$-isospectral families with $I$ non-empty and $0\not\in I$ have lens orbifolds with singular points of different isotropy types (see Remark~\ref{rem:isotropytype}).
\item If two lens spaces are $0$-isospectral, then the pair of corresponding covering spaces of the same degree are also $0$-isospectral (see Subsection~\ref{subsec:increasing-q}).
This is not true for $p>0$.
\end{enumerate}
\section{Preliminaries}\label{sec:preliminaries}
In this section we will introduce several concepts and results on lens orbifolds and their $p$-spectra.
It is based on the references \cite{IkedaTaniguchi78}, \cite{IkedaYamamoto79}, \cite{Ikeda80_isosp-lens}, \cite{Ikeda88}, \cite{LMR-repequiv}, \cite{LMR-onenorm}, \cite{Lauret-spec0cyclic}, \cite{Lauret-multip}.
These preliminaries will be used in Section~\ref{sec:evidencedfacts} to prove several facts evidenced by the computational results.
The author suggests the reader already familiar with lens spaces and their $p$-spectra to skip this section and come back to it for reference.
\subsection{$p$-Spectra of spherical orbifolds}
We assume that $M$ is a \emph{good orbifold} covered by $S^{2n-1}$, that is, $M=\Gamma\ba S^{2n-1}$ with $\Gamma$ a finite subgroup of $\SO(2n)$.
We will use the term \emph{spherical orbifold} for a space as above.
The space $\Gamma\ba S^{2n-1}$ is a manifold if and only if $\Gamma$ acts freely on $S^{2n-1}$.
In this case, $\Gamma\ba S^{2n-1}$ is usually called a \emph{spherical space form}.
For $\Gamma\subset G$ finite, let us denote by $\Delta_{\Gamma,p}$ the Hodge-Laplace operator on $p$-forms of $\Gamma\ba S^{2n-1}$, which is given by the restriction of $\Delta_p$ on $S^{2n-1}$ to $\Gamma$-invariant smooth $p$-forms on $S^{2n-1}$.
Two spherical orbifolds $\Gamma\ba S^{2n-1}$ and $\Gamma'\ba S^{2n-1}$ are said to be \emph{$p$-isospectral} if the operators $\Delta_{\Gamma,p}$ and $\Delta_{\Gamma',p}$ have the same spectra.
The space $\Gamma\ba S^{2n-1}$ is orientable, thus $\spec_p(\Gamma\ba S^{2n-1}) = \spec_{2n-1-p}(\Gamma\ba S^{2n-1})$.
In particular, $p$-isospectrality for every $0\leq p\leq n-1$ is actually equivalent to $p$-isospectrality for every $p$.
The next theorem is well known (see for instance \cite[Thm.~4.2]{IkedaTaniguchi78}, \cite[Prop.~2.1]{Ikeda88}, \cite[Thm.~1.1]{LMR-repequiv}, \cite[Prop.~2.2]{LMR-onenorm}).
It describes the $p$-spectrum of $\Gamma\ba S^{2n-1}$ in terms of the dimension of the $\Gamma$-invariant subspaces of certain irreducible representations of $\SO(2n)$.
For $\pi$ a representation of $\SO(2n)$, let $V_\pi$ denote its underlying vector space and $V_{\pi}^\Gamma$ the $\Gamma$-invariants in $V_{\pi}$.
Following the same notation as in \cite{Lauret-multip}, let $\pi_{k,p}$ be the irreducible representation of $\SO(2n)$ with highest weight $k\varepsilon_1+\sum_{j=1}^p \varepsilon_j = (k+1)\varepsilon_1+\sum_{j=2}^p\varepsilon_j$.
For $k\geq 1$, we set
\begin{equation}\label{eq1:lambda_(k,p)}
\lambda_{k,p} =
\begin{cases}
0&\quad\text{if $p=-1$},\\
(k+p)(k+2n-2-p) &\quad\text{if $0\leq p\leq n-1$.}
\end{cases}
\end{equation}
\begin{theorem}\label{thm:spectrum-general}
Fix $0\leq p\leq n-1$.
Each eigenvalue in $\spec_p(\Gamma\ba S^{2n-1})$ is of the form $\lambda_{k,p-1}$ or $\lambda_{k,p}$ for some $k\geq 1$, with multiplicity
\begin{equation*}
\mult_{\Delta_{\Gamma,p}}(\lambda_{k,p-1}) = \dim V_{\pi_{k-1,p}}^\Gamma
\quad\text{ and }\quad
\mult_{\Delta_{\Gamma,p}}(\lambda_{k,p}) = \dim V_{\pi_{k-1,p+1}}^\Gamma
\end{equation*}
respectively.
\end{theorem}
One can see from Theorem~\ref{thm:spectrum-general} that the $p$-spectrum of a spherical orbifold consists of two `strings' of multiplicities, namely $\{\mult_{\Delta_{\Gamma,p}}(\lambda_{k,p-1})\}_{k\geq1}$ and $\{\mult_{\Delta_{\Gamma,p}}(\lambda_{k,p})\}_{k\geq1}$.
Consequently, the $p$-spectrum and the $(p+1)$-spectrum of a spherical orbifold share one string.
This remark yields a very important obstruction to isospectrality (see for instance \cite[Cor.~1.2(ii)]{LMR-repequiv} or \cite[Prop.~2.1]{Ikeda88}).
\begin{proposition}\label{prop:hole-obstruction}
If two $(2n-1)$-dimensional spherical orbifolds are $(p-1)$-isospectral and $(p+1)$-isospectral, then they are $p$-isospectral.
In particular, $1$-isospectrality implies $0$-isospectrality.
\end{proposition}
\subsection{Lens spaces}\label{subsec:lensspaces}
Let $q$ be a positive integer and let $s=(s_1,\dots,s_n)\in\Z^n$ satisfying that $\gcd(q,s_1,\dots,s_n)=1$.
The cyclic group $\Gamma_{q,s}$ generated by
\begin{equation*}\label{eq:gamma}
\gamma_{q,s}:=
\diag\left(
\left[\begin{smallmatrix}\cos(2\pi{s_1}/q)&-\sin(2\pi{s_1}/q) \\ \sin(2\pi{s_1}/q)&\cos(2\pi{s_1}/q)
\end{smallmatrix}\right]
,\dots,
\left[\begin{smallmatrix}\cos(2\pi{s_m}/q)&-\sin(2\pi{s_m}/q) \\ \sin(2\pi{s_m}/q)&\cos(2\pi{s_m}/q)
\end{smallmatrix}\right]
\right)
\end{equation*}
has order $q$ and the space
\begin{equation*}
L(q;s)=L(q;s_1,\dots,s_m):=\Gamma_{q,s}\ba S^{2n-1},
\end{equation*}
is a \emph{lens orbifold}.
We will always consider the Riemannian structure induced by the round metric on $S^{2n-1}$.
The group $\Gamma_{q,s}$ acts freely on $S^{2n-1}$ if and only if $\gcd(q,s_j)=1$ for all $j$.
In this case, the Riemannian manifold $\Gamma_{q,s}\ba S^{2n-1}$ is a \emph{lens space}.
The following fact is well known (see for instance \cite[Ch.~V]{Cohen-book}).
\begin{proposition}\label{prop3:lens-isom}
Let $L=L(q;s)$ and $L'=L(q;s')$ be orbifold lens spaces.
Then, $L$ and $L'$ are isometric if and only if there exist a permutation $\sigma$ of $\{1,\dots,n\}$, $\epsilon_1,\dots,\epsilon_n\in\{\pm1\}$ and $t\in\Z$ coprime to $q$, such that
$$
s_{\sigma(j)}'\equiv t\epsilon_js_j\pmod{q}
$$
for all $1\leq j\leq n$.
\end{proposition}
\begin{remark}\label{rem:lensinC^n}
An equivalent and useful way to view the lens space $L(q;s)$ is as the quotient of $S^{2n-1}=\{(z_1,\dots,z_n)\in\C^n:|z_1|^2+\dots+|z_n|^2\}\subset \C^n$ by the action of the cyclic group of $q$-roots of unity given by
\begin{equation*}
\xi\cdot (z_1,\dots,z_n)=(\xi^{s_1}z_1,\dots,\xi^{s_n}z_n).
\end{equation*}
\end{remark}
The isotropy group of a point $x$ in a lens orbifold $L(q;s)=S^{2n-1}/\Gamma_{q,s}$, which in particular is a good orbifold, is given by the elements in $\Gamma_{q,s}$ fixing $x$.
Two points share the isotropy type if their isotropy groups are isomorphic.
The connected components of the equivalent isotropy classes of the points in $L(q;s)$ form a stratification of $L(q;s)$.
The points with a non-trivial isotropy group are called \emph{singular}, and otherwise \emph{regular}.
Since $L(q;s)$ is connected, the subset of its regular points form a single stratum, the only stratum of full dimension.
More details on orbifolds and their spectra can be found in \cite{Gordon12-orbifold}.
The isotropy classes in a lens orbifold $L(q;s)$ is determined by the multiset (set with multiplicities)
\begin{equation}\label{eq:multisetGCD}
\{\!\{\gcd(q,s_j):1\leq j\leq n\}\!\}.
\end{equation}
For example, $L(q;s)$ is a lens space if and only if \eqref{eq:multisetGCD} is equal to $\{\!\{1^n\}\!\}$ (the multiset given by $1$ repeated $n$ times).
The lens orbifold $L(4;1,0)$ contains exactly one point with non-trivial isotropy group.
Indeed, the action of a $q$-root of unity $\xi$ on a point $(z_1,z_2)$ in $S^3$ is
$$
\xi\cdot(z_1,z_2)= (\xi z_1,z_2),
$$
thus the class of the point $(0,1)$ in $L(4;1,0)$ has singular isotropy $\Gamma_{4,(1,0)}$, while $\Gamma_{4,(1,0)}$ acts freely on any other point in $L(4;1,0)$.
A more involved example is $L:=L(4;0,1,2)$, which has the three different isotropy types.
Indeed, the class in $L$ of the point $(1,0,0)$ in $S^5$ has isotropy group equal to $\Gamma_{4,(0,1,2)}$, and the class in $L$ of the points $(a,0,b)$ in $S^5$ satisfying $a^2+b^2=1$ and $b\neq 0$ have isotropy group of order $2$.
The rest of the points are regular.
\subsection{Ikeda's approach}\label{subsec:Ikeda'sapproach}
A.~Ikeda made important progress in inverse spectral geometry of spherical space forms, particularly in the study of lens spaces.
His main tool were the (two) generating functions associated to the $p$-spectrum of a lens space given.
He defined (see \cite[(2.3)]{Ikeda88})
\begin{equation}\label{eq1:F_Gamma^p}
F_\Gamma^{p}(z) =
\sum_{k\geq0} \mult_{\Delta_{\Gamma,p}} (\lambda_{k+1,p}) z^{k}
=\sum_{k\geq0} \dim V_{\pi_{k,p+1}}^\Gamma z^k,
\end{equation}
for any $0\leq p\leq n-1$ and any finite subgroup $\Gamma$ of $\SO(2n)$.
By Theorem~\ref{thm:spectrum-general}, $F_\Gamma^{p-1}(z)$ and $F_\Gamma^{p}(z)$ encode the $p$-spectrum of $\Gamma\ba S^{2n-1}$, and each generating function corresponds to each string.
From Theorem~\ref{thm:spectrum-general}, one immediately obtains the following characterization of $p$-isospectral spherical orbifolds.
\begin{proposition}\label{prop:p-isospIkeda}
Two spherical orbifolds $\Gamma\ba S^{2n-1}$ and $\Gamma'\ba S^{2n-1}$ are $p$-isospectral if and only if
$F_{\Gamma}^{p-1}(z)=F_{\Gamma'}^{p-1}(z)$ and $F_{\Gamma}^{p}(z)=F_{\Gamma'}^{p}(z).
$
\end{proposition}
We next recall the expression for $F_\Gamma^p(z)$ obtained by Ikeda.
In \cite[Thm.~3.2]{IkedaYamamoto79}, for $\Gamma$ a finite subgroup of $\SO(2n)$, it was shown that
\begin{equation}\label{eq:F_Gamma^0}
F_\Gamma^0(z) = \frac{1}{z}\left(\frac{1-z^2}{q}\sum_{\gamma\in\Gamma} \frac{1}{ \det(z-\gamma )}-1\right).
\end{equation}
Here $\det(z-\gamma)$ stands for $\det(\op{Id}_{2m}z-\gamma)=\prod_\lambda (z-\lambda)$, where $\lambda$ runs over the eigenvalues of $\gamma$.
In the case when $\Gamma\ba S^{2n-1}=L(q;s_1,\dots,s_n)$, we have that
\begin{equation}\label{eq:F_L^0}
F_\Gamma^0(z) = \frac{1}{z}\left(\frac{1-z^2}{q}\sum_{h=0}^{q-1} \prod_{j=1}^n \frac{1}{ (z-\xi_q^{h s_j})(z-\xi_q^{-h s_j})}-1\right).
\end{equation}
This equation was used in \cite{Ikeda80_isosp-lens} to give the first pair of $0$-isospectral lens spaces.
\begin{remark}\label{rem:F_L^0}
We point out that \eqref{eq:F_Gamma^0} is not identical to the one given in \cite{IkedaYamamoto79} and \cite{Ikeda80_isosp-lens}, since there $F_\Gamma^0(z)$ was defined by $\sum_{k\geq0} \dim V_{\pi_{k-1,1}}^\Gamma z^k$.
Actually, \cite[(4.7)]{LMR-survey} is missing the term $z^{-1}$.
From now on, for $L=\Gamma\ba S^{2n-1}$ a lens orbifold, we will write $F_L^p(z)$ in place of $F_\Gamma^p(z)$.
Furthermore, for a lens space $L=L(q;s)$, we write
\begin{equation}\label{eq:F_L^0abreviado}
\widetilde F_{L}^0(z) = \frac{q(zF_L^0(z)+1)}{(1-z^2) }
= \sum_{h=0}^{q-1} \prod_{j=1}^n \frac{1}{ (z-\xi_q^{h s_j})(z-\xi_q^{-h s_j})}.
\end{equation}
This new function will be very useful since $L$ and $L'$ are $0$-isospectral if and only if $\widetilde F_{L}^0(z)=\widetilde F_{L'}^0(z)$.
\end{remark}
Furthermore, Ikeda also gave an expression for $F_{\Gamma}^p(z)$ for $1\leq p\leq n-1$ and $\Gamma$ a finite subset of $\SO(2n)$, namely (see \cite[p.~394]{Ikeda88}),
\begin{equation}\label{eq:neat}
F_\Gamma^p(z) = (-1)^{p+1} z^{-p} + \frac{1}{|\Gamma|} \sum_{k=0}^p (-1)^{p-k} (z^{k-p}-z^{p-k+2})\sum_{\gamma\in\Gamma} \frac{\chi^{k}(\gamma)}{\det(z-\gamma)}.
\end{equation}
Here, $\chi^k$ denotes the character of the $k$-exterior representation $\bigwedge^k(\C^{2m})$ of $\SO(2m)$.
\subsection{Ikeda's examples}\label{subsec:Ikeda'sexamples}
We now recall briefly the examples of families of lens spaces isospectral for $p$ satisfying $0\leq p\leq p_0$.
They were given by Ikeda in \cite{Ikeda88} (see also \cite[\S4]{LMR-survey}).
These examples will be cited several times in the rest of the article.
For $q$ and $n$ positive integers, let ${\mathfrak L}_0(q;n)$ denote the classes (up to isometry) of $(2n-1)$-dimensional lens spaces $L(q;s_1,\dots,s_n)$ satisfying $s_i\not\equiv \pm s_j\pmod q$ for all $i\neq j$.
For a lens space $L=L(q;s_1,\dots,s_n)$ in ${\mathfrak L}_0(q;n)$, choose $h$ integers $\bar s_1,\dots,\bar s_h$ such that
$$
\{\pm s_1,\dots, \pm s_n, \pm \bar s_1,\dots,\pm\bar s_h\}
$$
is a set of representatives of integers mod~$q$, coprime to $q$.
Therefore $2n + 2h = \phi(q)$, where $\phi(q)$ denotes the Euler phi function.
Set $\bar L=L(q;\bar s_1,\dots,\bar s_h)$.
One can easily check that two lens spaces $L$ and $L'$ in ${\mathfrak L}_0(q;n)$ are isometric if and only if $\overline{L}$ and $\overline{L'}$ are isometric in ${\mathfrak L}_0(q;h)$ (see \cite[Prop.~3.3]{Ikeda88}).
We now assume that $q$ is a prime number and $n$ satisfies $q-1=2n+4$, in order to have $h=2$.
Set
\begin{equation}
{\mathfrak L}_p(q;n) = \left\{L(q;s)\in {\mathfrak L}_0(q;n):
\begin{array}{l}
a_1\bar s_1+a_2\bar s_2\not\equiv0\pmod q ,\\
\forall \; 1\leq |a_1|+|a_2|\leq p+2
\end{array}
\right\},
\end{equation}
thus one has the filtration
\begin{equation}\label{eq:filtration}
{\mathfrak L}_0(q;n) \supset {\mathfrak L}_1(q;n) \supset {\mathfrak L}_2(q;n)\supset\cdots\,.
\end{equation}
With this notation we can state Ikeda's result.
\begin{theorem}\label{thm:mainIkeda}
Let $q$ be an odd prime, $n=(q-5)/2$ and $0\leq p_0\leq n-1$ and let $L$ and $L'$ be lens spaces in ${\mathfrak L}_{p_0}(q;n)$.
Then $L$ and $L'$ are $p$-isospectral for all $0\leq p\leq p_0$.
If furthermore $L\in {\mathfrak L}_{p_0+1}(q;n)$ and $L'\notin {\mathfrak L}_{p_0+1}(q;n)$, then $L$ and $L'$ are not $(p_0+1)$-isospectral.
\end{theorem}
\begin{example}\label{ex:Ikeda'sexamples}
We now give the first steps of this sequence of families.
The cases satisfying $q<11$ are not interesting since ${\mathfrak L}_{0}(q;(q-5)/2)$ has less than two elements.
If $q=11$, then $n=3$ and we have that $L_1:=L(11;3,4,5)\simeq L(11;1,2,4) \in {\mathfrak L}_{0}(11;3) \smallsetminus {\mathfrak L}_{1}(11;3)$ and $L_2:=L(11;2,4,5)\simeq L(11;1,2,3) \in {\mathfrak L}_{1}(11;3)$.
Hence, $L_1$ and $L_2$ are $0$-isospectral but not $1$-isospectral, that is, they are $\{0\}$-isospectral according to Notation~\ref{notation:I-isospectral}.
For $q=13$ we have that $n=4$ and one can check that ${\mathfrak L}_{0}(13;4)$ is given by the elements
$$
\begin{array}{l}
L_1:=L(11;3,4,5,6)\simeq L(11;1,2,3,6) \in {\mathfrak L}_{0}(13;4) \smallsetminus {\mathfrak L}_{1}(13;4), \\
L_2:=L(11;2,4,5,6)\simeq L(11;1,2,3,4) \in {\mathfrak L}_{1}(13;4)\smallsetminus {\mathfrak L}_{2}(13;4),\\
L_3:=L(11;2,3,4,6)\simeq L(11;1,2,3,5) \in {\mathfrak L}_{2}(13;4).
\end{array}
$$
Therefore, these manifolds are all $0$-isospectral pairwise, $L_2$ and $L_3$ are also $1$-isospectral, $L_1$ is not $1$-isospectral to any other in the family, and there are no $2$-isospectral pairs among them.
In other words, the family $\{L_1,L_2,L_3\}$ is $\{0\}$-isospectral, $\{L_2,L_3\}$ is $\{0,1\}$-isospectral and there is no any $\{0,1,2\}$-isospectral subfamily of $\{L_1,L_2,L_3\}$.
\end{example}
\begin{remark}
Adapting this method, Shams~\cite{Shams11} found many families of $0$-isospectral lens orbifolds with singular points of dimension $\geq9$.
He worked with $q$ a power of a prime number and also a product of two different primes numbers.
\end{remark}
\subsection{Gornet and McGowan's computational study}\label{subsec:GornetMcGowan}
Gornet and McGowan made a computational study on $p$-isospectral lens spaces by following Ikeda's approach.
We next try to explain what they did in \cite{GornetMcGowan06}.
Fix $q$ a prime number and write $q-1=2n+2h$ with $n$ and $h$ positive integer numbers (they considered the cases $2\leq h\leq 4$).
For $L$ in $\mathfrak L_0(q;n)$, the generating function $F_L^p(z)$ in \eqref{eq:neat} can be written as (a common term plus) a polynomial over the $q$-th cyclotomic polynomial $\Phi_q(z)$ (see \cite[(4.6)]{Ikeda88} and \cite[(4.11)]{LMR-survey}).
Such polynomial is \cite[(3)]{GornetMcGowan06}, and its coefficients depend on certain arithmetic conditions (see \cite[(5)]{GornetMcGowan06}).
Gornet and McGowan computed these coefficients, for every prime number $q\leq 100$ and $h=2,3,4$, and then by comparing these numbers obtained the families of lens spaces in $\mathfrak L_0(q;n)$ that are $p$-isospectral, for any $0\leq p\leq n-1$.
We note that their examples are quite different from the ones given in this paper.
First at all, they work in lens spaces contained in $\mathfrak L_0(q;n)$, while we work with arbitrary lens orbifolds.
Furthermore, the dimension of their examples is subject to the choices of $q$ and $h$ ($2n-1=q-1-2h$) and consequently it grows quickly.
In our case, we fix $q$ and the dimension $2n-1$ up to $11$ ($17$ in \cite{Lauret-appendix}).
Unfortunately, the condition of $q$ being a prime number is essential in the previous argument, and it was not assumed in \cite{GornetMcGowan06}.
Sebastian Boldt communicated to the authors the following counterexample: the $3$-dimensional lens spaces $L(15;1,2)$ and $L(15;1,4)$ with fundamental group of order $15$ are not isometric and satisfy the condition for being $0$-isospectral given in \cite{GornetMcGowan06}, which is impossible by \cite{IkedaYamamoto79}.
The author thanks Sebastian Boldt for sharing this smart example.
Ruth Gornet communicated the author in the ``VI Workshop on Differential Geometry 2016'' at C\'ordoba, Argentina, that she and Jeffrey McGowan are working to extend their results to composite numbers given by the product of two primes numbers, including the case of the square of a prime number.
\subsection{One-norm approach}
In \cite{LMR-onenorm} was started a study of the spectra of lens spaces in connection with the one-norm lengths of elements in the associated \emph{congruence lattice}.
Apparently, this relation was already known by Ikeda and Yamamoto as it is indicated in \cite{Yamamoto80}.
However, they did not make any further use of this connection.
This approach works also for spherical orbifolds of the form $\Gamma\ba S^{2n-1}$ with $\Gamma$ a finite abelian subgroup of $\SO(2n)$.
We associate to a lens orbifold $L(q;s_1,\dots,s_n)$ the \emph{congruence lattice}
\begin{equation}\label{eq1:conglattice}
\mathcal L(q;s_1,\dots,s_n) :=
\left\{\sum_{j=1}^n a_j\varepsilon_j \in \Z^n: a_1s_1+\dots+a_ns_n\equiv0\pmod q\right\}.
\end{equation}
For $\mu=(a_1,\dots,a_n)\in\Z^n$, we write $\norma{\mu}=|a_1|+\dots+|a_n|$ for the \emph{one-norm} of $\mu$ and let $Z(\mu)$ be the number of zero coordinates of $\mu$.
For any subset $\mathcal L$ of $\Z^n$, set
\begin{align}\label{eq1:N_L(k)N_L(k,l)}
N_{\mathcal L} (k,\ell) &= \#\{\mu\in \mathcal L: \norma{\mu}=k,\;Z(\mu)=\ell\},\\
N_{\mathcal L} (k)& = \#\{\mu\in \mathcal L: \norma{\mu}=k\} . \notag
\end{align}
Furthermore, the \emph{one-norm} generating function of $\mathcal L$ is given by
\begin{equation}
\vartheta_{\mathcal L}(z) = \sum_{k\geq0} N_{\mathcal L}(k)z^k.
\end{equation}
Similarly, for $0\leq \ell\leq n$, we let
\begin{equation}
\vartheta_{\mathcal L}^{(\ell)}(z) = \sum_{k\geq0} N_{\mathcal L}(k,\ell)z^k.
\end{equation}
After quite some effort, the author proved in \cite[Thm.~2.2]{Lauret-multip} an explicit expression for $F_{L}^p(z)$ for any $0\leq p\leq n-1$.
\begin{theorem}\label{thm:F_L^p}
Let $L$ be a $(2n-1)$-dimensional lens orbifold with associated congruence lattice $\mathcal L$.
For each $1\leq p\leq n$,
we have that
\begin{equation}\label{eq1:F_Gamma^p-1-formula}
F_L^{p-1}(z) = \frac{1}{z^p(1-z^2)^{n-1}}\sum_{\ell=0}^{n} \vartheta_{\mathcal L}^{(\ell)}(z) \; A_{p}^{(\ell)}(z)+\frac{(-1)^p}{z^p},
\end{equation}
where \begin{align}\label{eq3:A_pl}
A_{p}^{(\ell)}(z) &= \sum_{j=1}^{p} (-1)^{j-1} \sum_{t=0}^{\lfloor\frac{p-j}{2}\rfloor} \binom{n-p+j+2t}{t}
\\ &\quad
\sum_{\beta=0}^{p-j-2t} 2^{p-j-2t-\beta} \binom{n-\ell}{\beta} \binom{\ell}{p-j-2t-\beta}
\sum_{\alpha=0}^\beta \binom{\beta}{\alpha} \sum_{i=0}^{j-1} z^{2(p-j-t-\alpha+i)}. \notag
\end{align}
\end{theorem}
\begin{remark}
It is important to note that $A_{p}^{(\ell)}(z)$ does not depend on the particular lens orbifold $L$.
Furthermore, it is a polynomial of degree $\leq 2(p-1)$ since $0\leq 2(p-j-2t-\beta) \leq 2(p-j-t-\alpha+i)\leq 2(p-j+j-1)=2(p-1)$.
\end{remark}
For example, we have that
\begin{align*}
F_L^0(z)
&= -\frac{1}{z}+\frac{1}{z(1-z^2)^{n-1}} \sum_{\ell=0}^n \theta_{\mathcal L}^{(\ell)}(z), \\
F_L^1(z)
&= \frac{1}{z^2}+ \frac{1}{z^2(1-z^2)^{n-1}} \sum_{\ell=0}^{n} \theta_{\mathcal L}^{(\ell)}(z) \Big((n-1+\ell)z^2+(n-1-\ell)\Big).
\end{align*}
As a consequence of Theorem~\ref{thm:F_L^p}, one obtains the following characterization (see \cite[Cor.~2.3]{Lauret-multip}).
\begin{theorem}\label{thm:charact[0,p]}
Let $0\leq p_0\leq n-1$ and let $L$ and $L'$ be $(2n-1)$-dimensional lens orbifolds with associated congruence lattices $\mathcal L$ and $\mathcal L'$ respectively.
Then, $L$ and $L'$ are $p$-isospectral for every $0\leq p\leq p_0$ if and only if
\begin{equation}\label{eq1:condition[0,p_0]-isosp}
\sum_{\ell=0}^n \ell^h \, \vartheta_{\mathcal L}^{(\ell)}(z) =
\sum_{\ell=0}^n \ell^h \, \vartheta_{\mathcal L_{\Gamma'}}^{(\ell)}(z)
\qquad\text{for all }0\leq h\leq p_0.
\end{equation}
In particular, $L$ and $L'$ are $0$-isospectral if and only if $\vartheta_{\mathcal L}(z)= \vartheta_{\mathcal L'}(z)$.
Moreover, $L$ and $L'$ are $p$-isospectral for all $p$ if and only if $\vartheta_{\mathcal L}^{(\ell)}(z)=\vartheta_{\mathcal L'}^{(\ell)}(z)$ for all $0\leq \ell\leq n$.
\end{theorem}
\subsection{Rational form for the one-norm generating functions}
For $q$ a positive integer and $\mathcal L\subset P(G)$, we define
\begin{align}
C(q) &=\{\mu=\textstyle\sum_{i}a_i\varepsilon_i\in P(G): |a_i|<q\quad\forall\,i\},\label{eq1:reduceterms} \\
N_{\mathcal L}^{\textrm{red}}(k,\ell) &= \#\{\mu\in C(q)\cap\mathcal L: \norma{\mu}=k,\;Z(\mu)=\ell\},\notag \\
\Phi_{\mathcal L}^{(\ell)}(z) &= \sum_{k\geq0} N_{\mathcal L}^{\textrm{red}}(k,\ell) z^k.\notag
\end{align}
Note that $\Phi_{\mathcal L}^{(\ell)}(z)$ is a polynomial of degree at most $(n-\ell)(q-1)$.
\begin{theorem}\label{thm:theta^l-rational}
Let $L$ be a $(2n-1)$-dimensional lens orbifold with fundamental group of order $q$ and associated congruence lattice $\mathcal L$.
Then, for each $0\leq \ell\leq n$,
\begin{equation*}
\vartheta_{\mathcal L}^{(\ell)} (z)
=\frac{1}{(1-z^{q})^{n-\ell}}\sum_{s=0}^{n-\ell} 2^s\binom{\ell+s}{s} z^{sq} \,\Phi_{\mathcal L}^{(\ell+s)}(z).
\end{equation*}
Moreover,
\begin{equation*}
\vartheta_{\mathcal L}(z)
=\frac{1}{(1-z^{q})^{n}} \, \displaystyle\sum_{t=0}^{n} z^{tq}\sum_{\ell=t}^n \binom{\ell}{t} \Phi_{\mathcal L}^{(\ell)}(z).
\end{equation*}
\end{theorem}
\begin{lemma}\label{lem:theta^n-1}
Let $L=L(q;s)$ be a $(2n-1)$-dimensional lens space and let $\mathcal L$ be its associated congruence lattice.
Then $\vartheta_{\mathcal L}^{(n)}(z)=1$ and
$$
\vartheta_{\mathcal L}^{(n-1)}(z)= \frac{2nz^q}{1-z^q}.
$$
\end{lemma}
\begin{proof}
Theorem~\ref{thm:theta^l-rational} ensures that $\vartheta_{\mathcal L}^{(n)}(z) = \Phi_{\mathcal L}^{(n)}(z)$ and $\vartheta_{\mathcal L}^{(n-1)}(z) = \Phi_{\mathcal L}^{(n-1)}(z) + 2nz^q\Phi_{\mathcal L}^{(n)}(z)$.
The assertions follow from $\Phi_{\mathcal L}^{(n)}(z)=1$ and $\Phi_{\mathcal L}^{(n-1)}(z)=0$.
Indeed, the first one is clear since the only element in $\Z^n$ with $n$ zero coordinates is the zero vector, which is always in $\mathcal L$.
The second one holds because a vector $\mu$ with $n-1$ zero coordinates has the form $ke_i$ for some $k\neq0$ and $1\leq i \leq n$, then $\mu$ is in $\mathcal L=\mathcal L(q;s_1,\dots,s_n)$ if and only if $ks_i\equiv 0 \pmod q$, which is equivalent to $k\equiv 0\pmod q$ because $\gcd(q,s_j)=1$ for all $j$.
\end{proof}
\section{Computational results}\label{sec:data}
In this section we include the consequences of the computational results.
This article is accompanied with the appendix \cite{Lauret-appendix}, which includes the computational results and the algorithms used.
More precisely, \cite{Lauret-appendix} shows the list of all $I$-isospectral families among $(2n-1)$-dimensional lens orbifolds for $2n-1=5$, $7$, $9$, $11$, $13$, $15$, $17$ with fundamental group of order $q\leq 200$, $120$, $75$, $59$, $44$, $32$, $27$ respectively.
Here, in each (odd) dimension, we will analyze the subsets $I$ for which there exist $I$-isospectralities among lens orbifolds with fundamental groups of low orders.
More precisely, for fixed $n$ satisfying $3\leq n\leq 6$ and each subset $I$ of $\{0,1,\dots,n-1\}$, Tables~\ref{table:summarydim5}--\ref{table:summarydim11} will indicate if there is a pair of $(2n-1)$-dimensional non-isometric $I$-isospectral lens spaces (or lens orbifolds) with fundamental group of low order.
The analogous tables in dimensions $13$, $15$ and $17$ are included in \cite{Lauret-appendix}.
When no example of $I$-isospectral pairs occurs, we will give a justification in case we are aware of it.
The most common justification for the non-existence of an $I$-isospectral pair is the `hole obstruction' in Proposition~\ref{prop:hole-obstruction}.
For instance, it is the case for every $I$ containing $1$ but not $0$, or containing $0$ and $2$ but not $1$.
The lowest dimension considered is $5$ since no examples appear in the computations for dimension $3$.
Indeed, Ikeda and Yamamoto~\cite{IkedaYamamoto79}\cite{Yamamoto80} proved the non-existence of non-isometric $0$-isospectral $3$-dimensional lens spaces.
Of couse, $1$-isospectrality is also impeded by Proposition~\ref{prop:hole-obstruction}.
Moreover, recently Shams and Hunsicker~\cite{ShamsHunsicker17} extend this result to $3$-dimensional lens orbifolds.
\subsection{Dimension 5}
The computational results in dimension 5 included in \cite{Lauret-appendix} are summarized in Table~\ref{table:summarydim5}.
In this case, the hole obstruction (Proposition~\ref{prop:hole-obstruction}) is an impediment to the existence of $I$-isospectral lens orbifolds for $I=\{0,2\},\{1\},\{1,2\}$.
\begin{table}
\caption{Dimension 5}\label{table:summarydim5}
\begin{tabular}{|c|c|c|} \hline
\raisebox{-5pt}{\rule{0mm}{18pt}}&Lens spaces & Lens orbifolds\\ \hline \hline
& \multicolumn{2}{c|}{$\emptyset$, $\{0\}$, $\{0,1,2\}$.} \\ \cline{2-3}
$\exists$
&
&$\{2\}$.\\ \hline
$\nexists$ Prop.~\ref{prop:hole-obstruction}
& \multicolumn{2}{c|}{$\{0,2\}$, $\{1\}$, $\{1,2\}$.}\\ \hline
$\nexists$ Thm.~\ref{thm:[0,n-2]-isosp}
& $\{0,1\}$.& \\ \hline
$\nexists$ Rem.~\ref{rem:[0,n-2]-isosporbifolds}
& &$\{0,1\}$. \\ \hline
$\nexists$ Thm.~\ref{thm:n-2elements}
&$\{2\}$. & \\ \hline
\end{tabular}
\end{table}
The existence of $I$-isospectral $5$-dimensional lens spaces for $I=\{0\}$ was well known.
Actually, Theorem~\ref{thm:mainIkeda} for $q=11$ (the smallest non-trivial example) gives a pair of $5$-dimensional $\{0\}$-isospectral lens spaces.
Furthermore, there are $\{0\}$-isospectral families of lens orbifolds which are not lens spaces.
An example with smallest fundamental group is given by the pair $$\{L(15;1,2,6), L(15;1,3,4)\}.$$
The classification of $5$-dimensional $\{0\}$-isospectral lens spaces or lens orbifolds seems to be a difficult task.
Recently, the existence of pairs of non-isometric $5$-dimensional $\{0,1,2\}$-isospectral lens spaces was shown.
This implies that the lens spaces are $p$-isospectral for all $p$ since the $p$-spectrum of a $5$-dimensional lens space coincides with its $(5-p)$-spectrum.
Many such examples have been found in \cite{LMR-onenorm}.
It would be interesting to have a classification of $5$-dimensional lens spaces that are $p$-isospectral for all $p$.
The computational results give evidences that such families come only in pairs.
DeFord and Doyle~\cite{DeFordDoyle14} have made an important step on this problem in arbitrary dimension.
It is very interesting that all such examples are given by lens spaces, that is, there is no any example of non-isometric $5$-dimensional `pure' lens orbifolds $p$-isospectral for all $p$.
Theorem~\ref{thm:[0,n-2]-isosp} ensures the non-existence of $5$-dimensional $\{0,1\}$-isospectral lens spaces.
The data gives evidences that the same is true for $5$-dimensional lens orbifolds (see Remark~\ref{rem:[0,n-2]-isosporbifolds}).
The non-existence of $5$-dimensional $\{2\}$-isospectral lens spaces is proved in Theorem~\ref{thm:n-2elements}.
It was surprising the existence of $\{2\}$-isospectral lens orbifolds.
Conjecture~\ref{conj:5-dim[2]-isosp} claims its classification.
A very interesting fact is that all pairs of $5$-dimensional $\{2\}$-isospectral lens orbifolds have different isotropy types.
For example, $\{L(8;0,1,3),L(8;1,3,4)\}$ is the pair with smallest fundamental group in this situation.
On the one hand, the maximal isotropy group in $L(8;1,3,4)$ is given in the class of the point $(0,0,1)\in\C^3$ (see Remark~\ref{rem:lensinC^n} for notation).
Such group is isomorphic to the cyclic group of order $4$.
On the other hand, the class of the element $(1,0,0)$ in $L(8;0,1,3)$ has isotropy group isomorphic to the cyclic group of order $8$, which of course has maximal order.
\subsection{Dimension 7}
\begin{table}
\caption{Dimension 7}\label{table:summarydim7}
\begin{tabular}{|c|c|c|} \hline
\raisebox{-5pt}{\rule{0mm}{18pt}}&Lens spaces & Lens orbifolds\\ \hline \hline
& \multicolumn{2}{c|}{$\emptyset$, $\{0\}$, $\{0,1\}$, $\{0,1,2,3\}$.} \\ \cline{2-3}
$\exists$
&
& $\{2\}$, $\{3\}$. \\ \hline
$\nexists$ Prop.~\ref{prop:hole-obstruction}
& \multicolumn{2}{c|}{$\{0,1,3\}$, $\{0,2\}$, $\{0,2,3\}$, $\{1\}$, }\\
(hole obstruction)
&\multicolumn{2}{c|}{$\{1,2\}$, $\{1,2,3\}$, $\{1,3\}$. }
\\ \hline
$\nexists$ Thm.~\ref{thm:[0,n-2]-isosp}
&$\{0,1,2\}$.& \\ \hline
$\nexists$ Rem.~\ref{rem:[0,n-2]-isosporbifolds}
&&$\{0,1,2\}$. \\ \hline
$\nexists$ Thm.~\ref{thm:n-2elements}
& $\{0,3\}$, $\{2,3\}$. &\\ \hline
$\nexists$ Rem.~\ref{rem:n-2elementsorbifolds}
&& $\{0,3\}$.\\ \hline
$\nexists$ ?
& $\{2\}$, $\{3\}$.
& $\{2,3\}$. \\ \hline
\end{tabular}
\end{table}
Table~\ref{table:summarydim7} shows a summary of the computational results in dimension 7.
Similarly as in the previous dimension, Proposition~\ref{prop:hole-obstruction} ensures the non-existence of $I$-isospectral pairs for subsets $I$ with simple holes, namely, $\{0,1,3\}$, $\{0,2\}$, $\{0,2,3\}$, $\{1\}$, $\{1,2\}$, $\{1,2,3\}$, $\{1,3\}$.
The existence of $I$-isospectral $7$-dimensional lens spaces for $I=\{0\}$ and $\{0,1\}$ were already known.
Indeed, Theorem~\ref{thm:mainIkeda} for $q=13$ shows examples to these cases.
Moreover, there exist such examples for pure lens orbifolds.
For instance, any pair of $5$-dimensional $0$-isospectral lens spaces induces a pair, via Proposition~\ref{prop:adding0s}, of $7$-dimensional $0$-isospectral lens orbifolds, which are not lens spaces.
For example, see in \cite[Table 2]{Lauret-appendix} the pair for $q=11$.
The smallest example (minimal $q$) of $\{0,1\}$-isospectral $7$-dimensional lens orbifolds which are not lens spaces is
$$
\{L(36;1,3,5,17),\ L(36;1,3,7,11)\}.
$$
An interesting fact is that $L(36;1,3,5,17)$ is isometric to $L(36;1,7,11,15)$ by multiplying its parameters by $5$ and making a reordering.
In other words, the $5$-dimensional lens spaces $L(36;1,5,17)$ and $L(36;1,7,11)$ are isometric, but to add the parameter $3$ to each of them produces a pair of $7$-dimensional non-isometric $\{0,1\}$-isospectral lens orbifolds.
As in the previous dimension, $q=49$ is the smallest $q$ with examples of pairs among $7$-dimensional lens spaces $p$-isospectral for all $p$ (i.e.\ $\{0,1,2,3\}$-isospectral).
In \cite{LMR-onenorm} and \cite{LMR-survey} are many more examples.
Differently to the previous dimension, there exist $7$-dimensional pure lens orbifolds (i.e.\ lens orbifolds with non-trivial maximal isotropy group) $p$-isospectral for all $p$.
Moreover, such examples can be constructed by adding certain non-coprime parameter to a pair of $5$-dimensional lens spaces that are $p$-isospectral for all $p$.
For instance, the pair $\{L(49;1,6,15), L(49;1,6,20)\}$ of $5$-dimensional lens spaces is $p$-isospectral for all $p$, and it induces the pairs
\begin{align*}
&
\begin{cases}
L(49;1,6,15, 0), \\ L(49;1,6,20, 0),
\end{cases}
&&
\begin{cases}
L(49;1,6,15, 7), \\ L(49;1,6,20, 7),
\end{cases}
&&
\begin{cases}
L(49;1,6,15,14), \\ L(49;1,6,20,14),
\end{cases}
&&
\begin{cases}
L(49;1,6,15,21), \\ L(49;1,6,20,21).
\end{cases}
\end{align*}
which are lens orbifolds $p$-isospectral for all $p$.
Theorem~\ref{thm:[0,n-2]-isosp} ensures the non-existence of $\{0,1,2\}$-isospectral lens spaces.
It is expected that this fact holds for lens orbifolds (see Remark~\ref{rem:[0,n-2]-isosporbifolds}).
We prove in Theorem~\ref{thm:n-2elements} the non-existence of pairs of $\{2,3\}$-isospectral lens spaces.
There are evidences to think on the non-existence of $\{2,3\}$-lens orbifolds, and $I$-isospectral lens spaces for $I=\{2\},\{3\}$.
There are many pairs of $\{2\}$ and $\{3\}$-isospectral pure lens orbifolds, beginning from small choices of $q$.
Actually, one can see a pattern for the values of $q$ satisfying this condition (see Conjecture~\ref{conj:7-dim[2][3]-isosp}).
\subsection{Dimension 9}
\begin{table}
\caption{Dimension 9}\label{table:summarydim9}
\begin{tabular}{|c|c|c|} \hline
\raisebox{-5pt}{\rule{0mm}{18pt}}\raisebox{-5pt}{\rule{0mm}{18pt}}&Lens spaces & Lens orbifolds\\ \hline \hline
& \multicolumn{2}{c|}{$\emptyset$, $\{0\}$, $\{0,1\}$, $\{0,1,2\}$, $\{0,1,2,3,4\}$.} \\ \cline{2-3}
$\exists$
&
& $\{2,3\}$, $\{2,3,4\}$, $\{3,4\}$, $\{4\}$. \\ \hline
$\nexists$ Prop.~\ref{prop:hole-obstruction}
&\multicolumn{2}{c|}{$\{0,1,2,4\}$, $\{0,1,3\}$, $\{0,1,3,4\}$, $\{0,2\}$, $\{0,2,3\}$, }\\
(hole
&\multicolumn{2}{c|}{$\{0,2,3,4\}$, $\{0,2,4\}$, $\{1\}$, $\{1,2\}$, $\{1,2,3\}$, } \\
obstruction)
&\multicolumn{2}{c|}{$\{1,2,3,4\}$, $\{1,2,4\}$, $\{1,3\}$, $\{1,3,4\}$, $\{1,4\}$, $\{2,4\}$.} \\ \hline
$\nexists$ Thm.~\ref{thm:[0,n-2]-isosp}
&$\{0,1,2,3\}$.& \\ \hline
$\nexists$ Rem.~\ref{rem:[0,n-2]-isosporbifolds}
&&$\{0,1,2,3\}$. \\ \hline
$\nexists$ Thm.~\ref{thm:n-2elements}
&$\{0,1,4\}$, $\{0,3,4\}$, $\{2,3,4\}$. & \\ \hline
$\nexists$ Rem.~\ref{rem:n-2elementsorbifolds}
&&$\{0,1,4\}$, $\{0,3,4\}$.\\ \hline
&\multicolumn{2}{c|}{$\{0,3\}$, $\{0,4\}$, $\{2\}$, $\{3\}$.}\\ \cline{2-3}
$\nexists$ ?
& $\{2,3\}$, $\{3,4\}$, $\{4\}$.
& \\ \hline
\end{tabular}
\end{table}
The computational results in dimension $9$ for $q\leq 75$ are in \cite[Table~3]{Lauret-appendix}, while their summary is in Table~\ref{table:summarydim9}.
The hole obstruction explains the non-existence of $I$-isospectral lens orbifolds for $16$ cases among $32$.
Theorem~\ref{thm:mainIkeda} does not produce any example in dimension $9$.
However, there exist examples of $I$-isospectral pair of lens spaces for $I=\{0\}$, $\{0,1\}$, and $\{0,1,2\}$.
The corresponding minimal values of $q$ are $16$, $17$ and $27$ respectively.
The pair of $\{0,1,2,3,4\}$-isospectral lens spaces with smallest $q$ appears in $q=64$.
However, $q=49$ is the smallest $q$ such that there is a pair of $\{0,1,2,3,4\}$-isospectral lens orbifolds.
Furthermore, one can see that these examples can be constructed from those in lower dimensions, in a similar way as was explained in dimension $7$.
Theorem~\ref{thm:[0,n-2]-isosp} ensures the non-existence of $\{0,1,2,3\}$-isospectral lens spaces (see also Remark~\ref{rem:[0,n-2]-isosporbifolds} for lens orbifolds).
Theorem~\ref{thm:n-2elements} guarantees the non-existence of pairs of $I$-isospectral lens spaces for $I=\{0,1,4\}, \{0,3,4\}, \{2,3,4\}$ (see Remark~\ref{rem:n-2elementsorbifolds} for $I$-isospectral lens orbifolds for $I=\{0,1,4\}, \{0,3,4\}$).
We do not know a reason for the non-existence of the cases indicated in the last two rows in Table~\ref{table:summarydim9}.
Also, it is somewhat surprising to see the existence of $\{2,3,4\}$-isospectral lens orbifolds.
\subsection{Dimension 11}
The computational results in dimension $11$ for $q\leq 59$ are in \cite[Table~4]{Lauret-appendix}, while their summary can be found in Table~\ref{table:summarydim11}.
In this case, there are $36$ non-existence cases among $64$ which are explained by the hole obstruction.
\begin{table}
\caption{Dimension 11}\label{table:summarydim11}
\begin{tabular}{|c|c|c|} \hline
\raisebox{-5pt}{\rule{0mm}{18pt}}&Lens spaces & Lens orbifolds\\ \hline \hline
& \multicolumn{2}{c|}{$\emptyset$, $\{0\}$, $\{0,1\}$, $\{0,1,2\}$, $\{0,1,2,3\}$, $\{0,1,2,3,4,5\}$, $\{4,5\}$.} \\ \cline{2-3}
$\exists$
&
& $\{0,3\}$, $\{0,5\}$, $\{2,3,4\}$,\\
&
& $\{3\}$, $\{3,4,5\}$, $\{4\}$, $\{5\}$.
\\ \hline
$\nexists$ Thm.~\ref{thm:[0,n-2]-isosp}
&$\{0,1,2,3,4\}$. &\\ \hline
$\nexists$ Rem.~\ref{rem:[0,n-2]-isosporbifolds}
&&$\{0,1,2,3,4\}$. \\ \hline
& $\{0,1,2,5\}$, $\{0,1,4,5\}$,& \\
$\nexists$ Thm.~\ref{thm:n-2elements}
& $\{0,3,4,5\}$, $\{2,3,4,5\}$.&
\\ \hline
&& $\{0,1,2,5\}$, $\{0,1,4,5\}$, \\
$\nexists$ Rem.~\ref{rem:n-2elementsorbifolds}
&&$\{0,3,4,5\}$. \\ \hline
& \multicolumn{2}{c|}{$\{0,4\}$, $\{0,1,4\}$, $\{0,1,5\}$, $\{0,3,4\}$, $\{0,4,5\}$, } \\
& \multicolumn{2}{c|}{$\{2\}$, $\{2,3\}$, $\{2,5\}$, $\{3,4\}$. } \\
\cline{2-3}
$\nexists$ ?
& $\{0,3\}$, $\{0,5\}$, $\{2,3,4\}$,
& $\{2,3,4,5\}$. \\
& $\{3\}$, $\{3,4,5\}$, $\{4\}$, $\{5\}$.
&\\ \hline
\end{tabular}
\end{table}
Theorem~\ref{thm:mainIkeda} for $q=17$ gives examples of $11$-dimensional $I$-isospectral lens spaces for $I=\{0\}$, $\{0,1\}$ and $\{0,1,2\}$.
Furthermore, there exist pairs of $11$-dimensional non-isometric $\{0,1,2,3\}$-isospectral lens spaces.
The smallest value of $q$ for such example is $52$.
Although there is no pair of $11$-dimensional lens spaces that are $p$-isospectral for all $p$ (i.e.\ $\{0,1,2,3,4,5\}$-isospectral) in \cite[Table~4]{Lauret-appendix}, such examples do exist.
By computations made by the author related to the articles \cite{LMR-onenorm} and \cite{LMR-survey}, the examples with minimum value of $q$ is when $q=100$.
They are
\begin{equation*}
\begin{cases}
L(100;1,11,21,31,41,61),\\
L(100;1,11,21,31,41,81),
\end{cases}
\qquad
\begin{cases}
L(100;1,11,21,41,71,81),\\
L(100;1,11,21,41,51,81).
\end{cases}
\end{equation*}
One can check by using \cite[Thm.~1]{DeFordDoyle14} that they are $p$-isospectral for all $p$.
As it was expected, there are pairs of $11$-dimensional lens orbifolds $p$-isospectral for all $p$ with fundamental group of order $49$.
They can be constructed by using the previous examples of lens spaces $p$-isospectral for all $p$ of lower dimension.
Similarly as in the previous dimensions, Theorem~\ref{thm:[0,n-2]-isosp} and Theorem~\ref{thm:n-2elements} explains the non-existence of a couple of cases for $11$-dimensional lens spaces, and Remark~\ref{rem:[0,n-2]-isosporbifolds} and Remark~\ref{rem:n-2elementsorbifolds} say something about the lens orbifold case.
There are $15$ (resp.\ $10$) cases among $64$ for which we do not know the existence of $I$-isospectral lens spaces (resp.\ lens orbifolds).
\subsection{Conclusions and remarks}
In the previous tables, for a fixed dimension $2n-1$, there are five possibilities for each $I$ among the $2^n$ subsets of $\{0,\dots,n-1\}$, namely,
\begin{itemize}
\item there is an $I$-isospectral family in the computational results,
\item the existence of an $I$-isospectral family is obstructed by Proposition~\ref{prop:hole-obstruction},
\item the existence of an $I$-isospectral family is obstructed by Theorem~\ref{thm:[0,n-2]-isosp} or Remark~\ref{rem:[0,n-2]-isosporbifolds},
\item the existence of an $I$-isospectral family is obstructed by Theorem~\ref{thm:n-2elements} or Remark~\ref{rem:n-2elementsorbifolds},
\item there is no any $I$-isospectral family in the computational results and this cannot be explained by the above reasons.
\end{itemize}
Table~\ref{table:numbersubsets} shows how the number of subsets $I$ in each possible situation changes according to the growth of dimension up to $17$.
We recall that in Tables~\ref{table:summarydim5}--\ref{table:summarydim11}, a subset $I$ in the row `$\nexists$ ?' means that the computational results show no $I$-isospectral family which cannot be explained by the reasons above.
Probably there exists such a family with fundamental group of higher order than the considered by the computer.
The author believes that there should exist, in dimensions $13$, $15$ and $17$, more subsets $I$ satisfying that there is an $I$-isospectral pair.
\begin{table}
\caption{It is shown the number of subsets $I$ of $\{0,\dots,n-1\}$ such that there is no $(2n-1)$-dimensional $I$-isospectral family of lens spaces and orbifolds in the computational results.
In a fixed dimension, the left (resp.\ right) column considers lens spaces (resp.\ lens orbifolds).
The third row indicates the cases explained in Section~\ref{sec:evidencedfacts} (i.e.\ Thm.~\ref{thm:[0,n-2]-isosp}, Rem.~\ref{rem:[0,n-2]-isosporbifolds}, Thm~\ref{thm:n-2elements} or Rem.~\ref{rem:n-2elementsorbifolds}).
}\label{table:numbersubsets}
\begin{center}
\begin{tabular}{c|cc|cc|cc|cc|cc|cc|cc|}
existence or reason &\multicolumn{14}{c|}{dimension}\\
for the non-existence
&\multicolumn{2}{c|}{5} &\multicolumn{2}{c|}{7} &\multicolumn{2}{c|}{9} &\multicolumn{2}{c|}{11} &\multicolumn{2}{c|}{13} &\multicolumn{2}{c|}{15} &\multicolumn{2}{c|}{17} \\ \hline\hline
$\exists$
&3&4& 4&6& 5&9& 7&14& 9&18& 6&23& 7&27\\ \hline
$\nexists$ Prop.~\ref{prop:hole-obstruction}
&\multicolumn{2}{c|}{3} & \multicolumn{2}{c|}{7}& \multicolumn{2}{c|}{16} &\multicolumn{2}{c|}{36} &\multicolumn{2}{c|}{79} &\multicolumn{2}{c|}{170} &\multicolumn{2}{c|}{361} \\ \hline
$\nexists$ Sect.~\ref{sec:evidencedfacts}
&2&1& 3&2& 4&3& 5&4& 6&5& 7&6& 8&7\\ \hline
$\nexists$ ?
&0&0& 2&1& 7&4& 16&10& 34&26& 73&57& 136&117\\ \hline \hline
Total & \multicolumn{2}{c|}{8}& \multicolumn{2}{c|}{16}& \multicolumn{2}{c|}{32} &\multicolumn{2}{c|}{64}& \multicolumn{2}{c|}{128}& \multicolumn{2}{c|}{256}& \multicolumn{2}{c|}{512}
\end{tabular}
\end{center}
\end{table}
We end this section with other observations.
\begin{remark}\label{rem:isotropytype}
At the end of the subsection corresponding to dimension $5$, we observed that the singular points in the lens orbifolds of any $\{2\}$-isospectral pair have different isotropy type.
In higher dimensions, a similar situation is repeated many times for $I$-isospectral families with $0\not\in I$.
More precisely, when $L(q;s_1,\dots,s_n)$ and $L(q;s_1',\dots,s_n')$ are $p$-isospectral for some $0<p\leq n-1$ and not $0$-isospectral, we usually have that $$
\{\!\{\gcd(q,s_j):1\leq j\leq n\}\!\} \neq \{\!\{\gcd(q,s_j'):1\leq j\leq n\}\!\}.
$$
However, this is not always the case.
For instance, the family
\begin{align*}
\{L(27;1, 3, 8, 10),\; L(27;1, 6, 8, 10),\; L(27;1, 8, 10, 12)\}
\end{align*}
is $\{2\}$-isospectral and $\{L(12;1, 1, 1, 3, 5),\; L(12;1, 1, 3, 5, 5)\}$ is $\{4\}$-isospectral.
On the other hand, it is expected that $0$-isospectral lens orbifolds have the same isotropy types (see Remark~\ref{rem:L(q/d;s)orbifolds}).
\end{remark}
\begin{remark}\label{rem:p-isospp>0}
Gornet and McGowan were interested on the existence of lens spaces $p_0$-isospec\-tral for some $p_0>0$ but not $p$-isospectral for all $0\leq p<p_0$.
We can see in Table~\ref{table:summarydim5} that such example does not exist in dimension $5$.
Indeed, in this case, $1$-isospectrality implies $0$-isospectrality by Proposition~\ref{prop:hole-obstruction} and $2$-isospectrality implies $p$-isospectrality for all $p$ by Theorem~\ref{thm:n-2elements}.
Also, Table~\ref{table:summarydim7} says that we do not know if such examples can exist in dimension $7$ since the problem of the existence of an $I$-isospectral pair is open for $I=\{2\}$ and $\{3\}$.
A similar situation happens in dimension $9$ according to Table~\ref{table:summarydim9}.
In dimension $11$ there exists a pair of $\{4,5\}$-isospectral lens spaces, in particular they are $4$-isospectral but not $p$-isospectral for all $0\leq p\leq 3$.
In \cite{Lauret-appendix} there are examples in dimension $13$ of $\{4\}$-isospectrality and $\{6\}$-isospectrality among lens spaces.
The same existence problem considering lens orbifolds is quite different since it is true from dimension $5$.
Moreover, there are many more examples of $I$-isospectral pure lens orbifolds with $0\not\in I$ than for lens spaces.
\end{remark}
\section{Evidenced facts}\label{sec:evidencedfacts}
In this section we list many facts observed in the computational results summarized in the previous section.
We will prove many of them by using the tools from Section~\ref{sec:preliminaries}.
We will also conjecture some other facts evidenced by the computational results.
\subsection{New examples increasing the dimension} \label{subsec:increasing-dimension}
We easily see from the data that one can add (or delete) $0$'s to the parameters of $0$-isospectral lens orbifolds and obtains $0$-isospectral lens orbifolds of higher (lower) dimension.
\begin{proposition}\label{prop:adding0s}
Let $L_1 =L(q;s_1,\dots,s_n)$, $L_2=L(q;s_1,\dots,s_n,0,\dots,0)$, $L_1'=L(q;s_1',\dots,s_n')$, and $L_2'=L(q;s_1',\dots,s_n',0,\dots,0)$,
where the parameter $0$ is repeated $m$ times in $L_2$ and $L_2'$.
Then, $L_1$ and $L_1'$ are $0$-isospectral if and only if $L_2$ and $L_2'$ are $0$-isospectral.
\end{proposition}
\begin{proof}
By \eqref{eq:F_L^0abreviado}, we have that
\begin{align*}
\widetilde F_{L_2}^0(z)
&= \sum_{h=0}^{q-1} \left(\prod_{j=1}^n\frac{1}{(z-\xi_q^{hs_j})(z-\xi_q^{-hs_j})}\right) \frac{1}{(z-1)^{2m}}= \frac{1}{(z-1)^{2m}} \widetilde F_{L_1}^0(z),
\end{align*}
and similarly for $L_1'$ and $L_2'$.
It follows immediately that $\widetilde F_{L_1}^0(z)=\widetilde F_{L_1'}^0(z)$ if and only if $\widetilde F_{L_2}^0(z)=\widetilde F_{L_2'}^0(z)$, and the proof is complete by Remark~\ref{rem:F_L^0}.
\end{proof}
Note that, the resulting spaces after adding $0$'s are always lens orbifolds with singular points, that is, they cannot be lens spaces.
For instance, $L(11;1,2,3)$ and $L(11;1,2,4)$ are $5$-dimensional non-isometric lens spaces and $L(11;1,2,3,0,\dots,0)$ and $L(11;1,2,4,0\dots,0)$, where $m$ is the number of zero coordinates added to each of them, are $(5+2m)$-dimensional non-isometric $0$-isospectral lens orbifolds.
\begin{remark}
We note that to add zero parameters to $p$-isospectral pairs with $p>0$ does not return $p$-isospectral pairs.
For example, $L(8;0,1,3)$ and $L(8;1,3,4)$ are $2$-isospectral, though $L(8;0,0,1,3)$ and $L(8;0,1,3,4)$ are not $2$-isospectral.
Moreover, there is no any $p$-isospectrality among $7$-dimensional lens spaces with fundamental group of order $q=8$.
\end{remark}
\subsection{New examples for increasing $q$}\label{subsec:increasing-q}
The next result gives a necessary condition to construct pairs of lens spaces covered by a pair of $0$-isospectral lens spaces.
\begin{theorem}\label{thm:L(q/d,s)}
Let $L=L(q;s)$ and $L=L(q;s')$ be $(2n-1)$-dimensional lens spaces.
If $L(q;s)$ y $L(q;s')$ are $0$-isospectral, then $L(q_1;s)$ y $L(q_1;s')$ are also $0$-isospectral for every positive divisor $q_1$ of $q$.
\end{theorem}
\begin{proof}
By \eqref{eq:F_L^0abreviado}, we have that
\begin{align}\label{eq:tildeF_L^0(z)-suma-divisores}
\widetilde F_{L}^0(z)
&= \sum_{d\mid q} \sum_{0\leq h<q\atop \gcd(h,q)=q/d} \prod_{j=1}^n\frac{1}{(z-\xi_q^{hs_j})(z-\xi_q^{-hs_j})} \\
&= \sum_{d\mid q} \sum_{h\in\Z_d^\times} \prod_{j=1}^n\frac{1}{(z-\xi_d^{hs_j})(z-\xi_d^{-hs_j})}.
\notag
\end{align}
Clearly, the poles of $\widetilde F_L^0(z)$ are primitive roots of unity of degree $d$ for some divisor $d$ of $q$.
We now assume that $L=L(q;s)$ and $L=L(q;s')$ are $0$-isospectral lens spaces, thus $\gcd(s_j,q)=\gcd(s_j',q)=1$ for all $1\leq j\leq n$.
Since $\widetilde F_{L}^0(z)= \widetilde F_{L'}^0(z)$, these functions have the same set of poles and, at each pole, they have the same order and singular part.
Hence, the sum of the singular parts of the poles at $d$-th primitive root of unity (i.e.\ $\xi_d^h$ for $h\in\Z_d^\times$) coincide, thus
$$
\sum_{h\in\Z_d^\times} \prod_{j=1}^n\frac{1}{(z-\xi_d^{hs_j})(z-\xi_d^{-hs_j})}
=
\sum_{h\in\Z_d^\times} \prod_{j=1}^n\frac{1}{(z-\xi_d^{hs_j'})(z-\xi_d^{-hs_j'})}
$$
for any divisor $d$ of $q$.
This implies that
$$
\widetilde F_{L(q_1,s)}^0(z) =
\sum_{d\mid q_1} \sum_{h\in\Z_d^\times} \prod_{j=1}^n\frac{1}{(z-\xi_d^{hs_j})(z-\xi_d^{-hs_j})}
= \widetilde F_{L(q_1,s')}^0(z) ,
$$
which completes the proof.
\end{proof}
Since lens spaces have cyclic fundamental group, their covering spaces are parameterized by the set of positive divisors of the order of the fundamental group.
More precisely, if $d$ is a positive divisor of $q$, then the $d$-covering space of $L(q;s_1,\dots,s_n)$ is $L(q/d;s_1,\dots,s_n)$.
Consequently, Theorem~\ref{thm:L(q/d,s)} says that if two lens spaces are $0$-isospectral, then their covering spaces of the same order are also $0$-isospectral.
\begin{remark}\label{rem:L(q/d;s)orbifolds}
The numerical evidences strongly support the claim that Theorem~\ref{thm:L(q/d,s)} is also valid for lens orbifolds.
However, the proof is not easy since the singular part of $\widetilde F_L^0(z)$ becomes unmanageable when $L$ is not a lens space.
Indeed, since $s_1,\dots,s_n$ are not necessarily coprime to $q$, it could be roots of unity of different orders in a same term in \eqref{eq:tildeF_L^0(z)-suma-divisores}.
If Theorem~\ref{thm:L(q/d,s)} were valid for lens orbifolds, then $0$-isospectral lens orbifolds $L(q;s_1,\dots,s_n)$ and $L(q;s_1',\dots,s_n')$ would satisfy
\begin{equation}\label{eq:conditionisotropy}
\{\!\{\gcd(q,s_j):1\leq j\leq n\}\!\} = \{\!\{\gcd(q,s_j'):1\leq j\leq n\}\!\}.
\end{equation}
These multisets (sets with multiplicities) are very important since they determine the isotropy types of the singular points of the corresponding lens orbifolds (see Subsection~\ref{subsec:lensspaces}).
\end{remark}
\subsection{$I$-isospectrality for $I$ having $n-1$ elements}\label{subsec:[0,n-2]-isosp}
One can observe from the data that there is no pair of $(2n-1)$-dimensional non-isometric lens orbifolds that are $I$-isospectral, if $I$ has $n-1$ elements. Indeed, the hole obstruction (Proposition~\ref{prop:hole-obstruction}) proves this fact for all cases except $I=\{0,\dots,n-2\}$ (i.e.\ $p$-isospectral for all $0\leq p\leq n-2$ but not $(n-1)$-isospectral).
The next result ensures the non-existence of them for lens spaces.
\begin{theorem}\label{thm:[0,n-2]-isosp}
Let $L=L(q;s)$ and $L=L(q;s')$ be $(2n-1)$-dimensional lens spaces.
If $L$ and $L'$ are $p$-isospectral for all $p$ satisfying $0\leq p\leq n-2$, then $L$ and $L'$ are also $(n-1)$-isospectral.
Consequently, there is no pair of non-isometric $(2n-1)$-dimensional $\{0,\dots,n-2\}$-isospectral lens spaces.
\end{theorem}
\begin{proof}
Write $\mathcal L=\mathcal L(q;s)$ and $\mathcal L'=\mathcal L(q;s')$ as in \eqref{eq1:conglattice}.
In order to show that $L$ and $L'$ are $p$-isospectral for all $p$, by Theorem~\ref{thm:charact[0,p]}, we will prove that $\mathcal L$ and $\mathcal L'$ are $\norma{\cdot}^*$-isospectral, that is, $\vartheta_{\mathcal L}^{(\ell)}(z) = \vartheta_{\mathcal L'}^{(\ell)}(z)$ for all $0\leq \ell\leq n$.
Abbreviating $\vartheta^{(\ell)}(z)=\vartheta_{\mathcal L}^{(\ell)}(z)-\vartheta_{\mathcal L'}^{(\ell)}(z)$, Theorem~\ref{thm:charact[0,p]} implies that our hypothesis is equivalent to
\begin{equation}\label{eq:n-2equations}
0=\sum_{\ell=0}^{n} \ell^h\vartheta^{(\ell)}(z)\qquad\text{ for all }\;0\leq h\leq n-2.
\end{equation}
Thus, we have $n-1$ equations (one for each $h$) and $n+1$ variables, namely, $\vartheta^{(0)}(z)$,$\dots$,$\vartheta^{(n)}(z)$.
This holds for arbitrary lens orbifolds.
Since $L$ and $L'$ are lens spaces, then $\vartheta^{(n)}(z)=\vartheta^{(n-1)}(z)=0$ by Lemma~\ref{lem:theta^n-1}.
We conclude that \eqref{eq:n-2equations} has $n-1$ equations and $n-1$ variables.
Moreover, the coefficients form a Vandermonde matrix, which is non-singular.
Hence, $\vartheta^{(\ell)}(z)=0$ for all $\ell$, and the proof is complete.
\end{proof}
\begin{remark}\label{rem:[0,n-2]-isosporbifolds}
We conjecture that Theorem~\ref{thm:[0,n-2]-isosp} also holds for lens orbifolds.
To complete the proof, it only remains to show that $\vartheta_{\mathcal L}^{(n-1)}(z)=\vartheta_{\mathcal L'}^{(n-1)}(z)$, which is true if $L$ and $L'$ satisfy \eqref{eq:conditionisotropy}.
Indeed, $\mu$ is a vector in $\mathcal L$ with exactly $n-1$ zero coordinates if and only if $\mu=k\varepsilon_i$ for some non-zero integer $k$ such that $ks_i\equiv0\pmod q$.
Since the last condition is equivalent to $\gcd(q,s_i)$ divides $k$, then
\begin{equation}\label{eq:theta^n-1}
\vartheta_{\mathcal L}^{(n-1)}(z) =
\sum_{i=1}^n \frac{z^{\frac{\gcd(q,s_i)}q}}{1-z^{\frac{\gcd(q,s_i)}q}}.
\end{equation}
Hence, $\vartheta_{\mathcal L}^{(n-1)}(z)$ only depends on the multiset $\{\!\{\gcd(q,s_j):1\leq j\leq n\}\!\}$, which determines the isotropy types of $L$ (see Subsection~\ref{subsec:lensspaces}).
We conclude by Remark~\ref{rem:L(q/d;s)orbifolds} that the non-existence of non-isometric $\{0,\dots,n-2\}$-isospectral lens orbifolds follows by showing that Theorem~\ref{thm:L(q/d,s)} is valid for lens orbifolds.
\end{remark}
\subsection{$I$-isospectrality for $I$ having $n-2$ elements}\label{subsec:n-2elements}
Let
\begin{equation}
\mathcal A=
\begin{pmatrix}
A_1^{(0)}(z) &\dots& A_1^{(n-2)}(z) \\
\vdots & \ddots & \vdots \\
A_{n}^{(0)}(z) &\dots& A_{n}^{(n-2)}(z)
\end{pmatrix},
\end{equation}
with $A_{p}^{(\ell)}(z)$ as in \eqref{eq3:A_pl}.
The matrix $\mathcal A$ has size $n\times(n-1)$ and coefficients in $\C(z)$.
\begin{lemma}\label{lem:matrix}
If $n\leq 9$, then the square matrix obtained by deleting any row to $\mathcal A$ is non-singular.
\end{lemma}
\begin{proof}
We have checked this claim by using \cite{Sage}.
In the appendix \cite{Lauret-appendix}, the (non-zero) determinants of such matrices are indicated.
\end{proof}
\begin{remark}
The author believes that Lemma~\ref{lem:matrix} holds for every $n$.
However, the proof does not seem easy.
\end{remark}
\begin{theorem}\label{thm:n-2elements}
Let $I\subset \{0,\dots,n-1\}$ with $n-2$ elements, excluding the case $\{0,\dots,n-3\}$.
Then, there is no any $I$-isospectral pair of non-isometric $(2n-1)$-dimensional lens spaces.
\end{theorem}
\begin{proof}
By Proposition~\ref{prop:p-isospIkeda}, two $(2n-1)$-dimensional lens spaces $L$ and $L'$ are $p$-isospectral for all $p\in I$ if $F_L^p(z)=F_{L'}^p(z)$ for all $p$ satisfying $p\in I$ or $p-1\in I$.
Convince yourself that these are at least $n-1$ different equations, as a consequence of how we choose the subset $I$.
Of course, $F_L^{-1}(z)=F_{L'}^{-1}(z)$ is not considered as an equation since they are zero by convention.
By Theorem~\ref{thm:F_L^p}, each equation $F_L^p(z)=F_{L'}^p(z)$ for some $0\leq p\leq n-1$ is equivalent to
\begin{equation}
0= \sum_{\ell=0}^n A_p^{(\ell)}(z) \vartheta^{(\ell)}(z).
\end{equation}
Here, we are abbreviating $\vartheta^{(\ell)}(z)=\vartheta_{\mathcal L'}^{(\ell)}(z)-\vartheta_{\mathcal L'}^{(\ell)}(z)$.
We conclude that we have at least $n-1$ equations with $n-1$ variables $\vartheta^{(0)}(z),\dots, \vartheta^{(n-2)}(z)$.
Indeed, $\vartheta^{(n-1)}(z)=\vartheta^{(n)}(z)=0$ by Lemma~\ref{lem:theta^n-1} since $L$ and $L'$ are lens spaces.
Hence, we obtain that $\vartheta^{(\ell)}(z)=0$ for all $\ell$ since the associated matrix has rank $n-1$ by Lemma~\ref{lem:matrix}.
Consequently, $L$ and $L'$ are $p$-isospectral for all $p$ by Theorem~\ref{thm:charact[0,p]}, and the proof is complete.
\end{proof}
\begin{remark}
Theorem~\ref{thm:n-2elements} cannot be extended to $I=\{0,\dots,n-3\}$ since there are $\{0,\dots,n-3\}$-isospectral pairs of $(2n-1)$-dimensional lens spaces, for low values of $n$ (e.g. $n=3,4,5,6,7,$).
The author thinks that such examples does not exist in every dimension.
\end{remark}
\begin{remark}\label{rem:n-2elementsorbifolds}
In case Theorem~\ref{thm:L(q/d,s)} is true for lens orbifolds (see Remark~\ref{rem:L(q/d;s)orbifolds}), we obtain the following result:
if $I\subset\{0,\dots,n-1\}$ has $n-2$ elements, $I\neq \{0,\dots,n-3\}$ and $0\in I$, then there is no any pair of $I$-isospectral non-isometric $(2n-1)$-dimensional lens orbifolds.
Indeed, following the proof of Theorem~\ref{thm:n-2elements}, it only remains to prove that $\vartheta_{\mathcal L'}^{(n-1)}(z)=\vartheta_{\mathcal L'}^{(n-1)}(z)$, which holds by Remark~\ref{rem:L(q/d;s)orbifolds} since $L$ and $L'$ are $0$-isospectral.
The assumption $0\in I$ above is essential, since Table~\ref{table:summarydim9} ensures the existence of pairs of $\{2,3,4\}$-isospectral $9$-dimensional lens orbifolds.
\end{remark}
\subsection{Orbifolds and manifolds}\label{subsec:Orbifoldsmanifolds}
We recall that a lens orbifold $L:=L(q;s_1,\dots,s_n)$ has a manifold structure if $\gcd(q,s_j)=1$ for every $j$.
In this case, $L$ is called a lens space.
The next result follows immediately from \cite[Prop.~3.4(ii)]{GordonRossetti03}, since lens orbifolds of the same dimension share the Riemannian universal cover.
\begin{theorem}\label{thm:orbifoldmanifolds}
Let $L$ and $L'$ be $0$-isospectral lens orbifolds.
If $L$ is a lens space (i.e.\ a manifold), then $L'$ is also a lens space.
\end{theorem}
For arbitrary orbifolds, the result does not hold for $p>0$ since Gordon and Rossetti~\cite{GordonRossetti03} show a counterexample by using the middle degree, that is, there is a $2d$-dimensional manifold $d$-isospectral to a $2d$-dimensional orbifold with singularities.
However, in the special case of lens orbifolds, the computational results give strong evidences of the following claim.
\begin{conjecture}\label{conj:orbifold/manifold}
If two $(2n-1)$-dimensional lens orbifolds $L$ and $L'$ are $p$-isospectral for some $0< p\leq n-1$ and $L$ is a lens space (i.e.\ a manifold), then $L'$ is a lens space.
\end{conjecture}
\subsection{$\{p\}$-isospectrality for $p>0$}\label{subsec:p-isosp}
Here, we investigate the examples appeared of $\{p\}$-isospec\-tral lens orbifolds with $p>0$ in dimension $5$ and $7$.
The computational results in \cite[Table 1]{Lauret-appendix} give evidences for the following claim.
\begin{conjecture}\label{conj:5-dim[2]-isosp}
The set of families of $5$-dimensional non-isometric $\{2\}$-isospectral lens orbifolds are given by pairs of the form
\begin{equation*}
\begin{cases}
L(8t;4,t,3t)\\
L(8t;8,t,3t)
\end{cases}
\qquad \text{for $t\geq1$ odd}.
\end{equation*}
\end{conjecture}
\begin{remark}
We sketch a possible proof.
If $L$ and $L'$ are $5$-dimensional $\{2\}$-isospectral lens orbifolds, then $F_{\Gamma}^{1}(z)=F_{\Gamma'}^{1}(z)$ and $F_{\Gamma}^{2}(z)=F_{\Gamma'}^{2}(z)$.
Theorem~\ref{thm:F_L^p} implies that these two equations are equivalent to the linear equation system
$$
\begin{array}{rrrcl}
(2z^2+2) \vartheta^{(0)}(z)
+&(3z^2+1) \vartheta^{(1)} (z)
+&4z^2\vartheta^{(2)} (z)
&=&0,\\
(z^4+4z^2+1) \vartheta^{(0)}(z)
+&(2z^4+2z^2) \vartheta^{(1)} (z)
+&(4z^4+2z^2) \vartheta^{(2)} (z)
&=&0,
\end{array}
$$
on the variables $\vartheta^{(\ell)} (z) := \vartheta_{\mathcal L}^{(\ell)}(z) - \vartheta_{\mathcal L'}^{(\ell)}(z)$ for $0\leq \ell\leq 2$.
By reducing this system we get that
$$
\begin{array}{rrcl}
(z^4+1) \vartheta^{(1)}(z)
+&4z^4\vartheta^{(2)}(z)
&=&0,\\
(2z^2-1) \vartheta^{(1)}(z)
+&2z^2\vartheta^{(0)}(z)
&=&0.
\end{array}
$$
This reduced system should be very useful to check that the sequence of pairs in Conjecture~\ref{conj:5-dim[2]-isosp} are $2$-isospectral.
A simple calculation shows that these pairs cannot be $p$-isospectral for $p=0,1$.
A more difficult task is to prove that this sequence exhausts such examples.
We could show that that $8$ divides to $q$ as follows.
We known that $\vartheta^{(\ell)}(z)$ is a rational function with denominator $(1-z^q)^{3-\ell}$ by Theorem~\ref{thm:theta^l-rational}.
We assume that $8$ does not divide to $q$, thus $\vartheta^{(1)}(z)$ converges at any primitive root of unity of order $8$, namely, $\xi_8^m$ with $m\in\Z$ odd.
Evaluating the first row of the previous system at $z=\xi_8^m$, we obtain that $0=4\xi_8^{4m}\vartheta^{(2)}(\xi_8^m)=-4\vartheta^{(2)}(\xi_8^m)$.
This equality should give a contradiction.
\end{remark}
Furthermore, \cite[Table 2]{Lauret-appendix} gives evidence on the following claim.
\begin{conjecture}\label{conj:7-dim[2][3]-isosp}
If $L(q;s)$ and $L(q;s')$ is a pair of $7$-dimensional non-isometric $\{2\}$-isospectral (resp.\ $\{3\}$-isospectral) lens orbifolds with fundamental group of order $q$, then $q$ is divisible by $3$ (resp.\ $5$).
\end{conjecture}
The higher dimensional cases are more complicated since the number of examples increase a lot.
\section{Conclusions}
Although lens spaces and lens orbifolds have provided several exotic isospectral examples, they are far away to answer in a positive way the question of R.~Miatello and J.P.~Rossetti~\cite[page 666]{MR-p-iso} (see also \cite[Problem~8.23]{CraioveanuPutaRassias-book}):
\begin{quote}
``Whether any possible combination of $p$-isospectrality is possible in a fixed dimension''.
\end{quote}
In other words, ``Is yes the answer of Question~\ref{question1} for every subset $I$?''.
According to Table~\ref{table:numbersubsets}, the major impediment is the well known hole obstruction (Proposition~\ref{prop:hole-obstruction}).
Furthermore, there are some barriers from isolated results like Theorem~\ref{thm:[0,n-2]-isosp}, Theorem~\ref{thm:n-2elements}, and probably others explaining the not known cases.
It would be desirable to have an explanation on the existence of $I$-isospectrality for each case $I$ in the rows `$\nexists$ ?' in Tables~\ref{table:summarydim5}--\ref{table:summarydim11}.
More precisely, for each of those $I$, to show an $I$-isospectral pair, or to give a proof for its non-existence.
Furthermore, it would be important to know if Theorem~\ref{thm:L(q/d,s)} holds also for lens orbifolds.
According to Remark~\ref{rem:L(q/d;s)orbifolds}, this would show that the spectral information of a lens orbifold determines the isotropy type of its singular points.
We note that this is not true for arbitrary orbifolds by \cite{ShamsStanhopeWebb06}.
The author hopes that this computational study motivates someone to consider any related problem.
\bibliographystyle{plain}
|
\section{Introduction}
Detecting and measuring quantum entanglement represent one of the goals
of quantum information science. In the last two decades, a large amount of work
has been invested in writing efficient separability criteria for both discrete- and
continuous-variable systems. Although bipartite entanglement seems to be
the simplest to detect and measure, practically we are still forced to apply
different criteria when discussing it for mixed states of composite systems.
As shown by Peres \cite{Peres}, a necessary condition of separability for an arbitrary
two-party state is the requirement to have a non-negative partially
transposed density matrix. In the case of discrete-variable systems, this requirement of
positive partial transposition (PPT) is also a sufficient condition of separability only for states
on ${\mathbb C}^2 \otimes {\mathbb C}^2$ and ${\mathbb C}^2 \otimes {\mathbb C}^3$
Hilbert spaces \cite{HHH}. For bipartite states of continuous-variable systems,
the PPT condition was first applied by Simon \cite{Simon}. Specifically, Simon proved
that preservation of the non-negativity of the density matrix under partial transposition
is not only a necessary, but also a sufficient condition for the separability of two-mode
Gaussian states (TMGSs). Moreover, the partial transposition criterion
could be expressed in an elegant symplectically invariant form valid for any TMGS \cite{Simon}.
To detect the continuous-variable entanglement for non-Gaussian states, Shchukin and Vogel
have derived an infinite series of inequalities for the moments of the state required by the PPT
condition \cite{SV}. Similar inequalities were obtained in Refs. \cite{BA2005,HZ2006}.
A somewhat parallel method to get general inseparability criteria for two-mode states
originates in a practical procedure proposed by Reid for demonstrating
the Einstein-Podolsky-Rosen (EPR) paradox \cite{EPR} in a non-degenerate parametric
amplifier \cite{Reid1}. This was done by using two non-local observables linearly built
with the canonical quadrature operators of the modes,
${\hat q}_j, {\hat p}_j, (j=1,2)$ \cite{Reid1,Reid2}:
\begin{align}
{\hat Q}(\lambda):= {\hat q}_1-\lambda {\hat q}_2,
\quad {\hat P}(\mu):= {\hat p}_1+\mu {\hat p}_2,
\label{Reid}
\end{align}
where $\lambda$ and $\mu$ are adjustable positive parameters.
As a consequence of their commutation relation,
\begin{align}
[{\hat Q}(\lambda), {\hat P}(\mu )]=i (1-\lambda \mu) {\hat I},
\label{CR-Reid}
\end{align}
we get the weak (Heisenberg) form of the uncertainty principle,
\begin{align}
\Delta Q(\lambda)\Delta P(\mu) \geqq \frac{1}{2}|1-\lambda \mu|,
\label{prodR}
\end{align}
which has to be fulfilled by any quantum state. In Eq.\ (\ref{prodR}) and in the sequel as well,
$(\Delta A)_{\hat \rho}$ denotes the standard deviation of the observable $\hat{A}$ in the
state $\hat{\rho}$, which is the square root of the variance
$$\left[ (\Delta A)_{\hat \rho} \right]^2:={\left \langle \left( {\hat A}
-{\langle {\hat A}\rangle }_{\hat \rho}{\hat I}\right)^2 \right \rangle }_{\hat \rho}=
{\langle {\hat A}^2\rangle }_{\hat \rho}-\left( {\langle {\hat A} \rangle}_{\hat \rho} \right)^2.$$
Unless $\lambda=\mu=1$, the operators\ (\ref{Reid}) are not genuine EPR observables
since they do not commute. In Refs.\cite{Reid1,Reid2}, a possible experimental
observation of the inequality
\begin{equation}
\Delta Q(\lambda)\Delta P(\mu) < \frac{1}{2}
\label{EPR}
\end{equation}
is interpreted as a sufficient condition to detect an EPR paradox. It is interesting
to remark that the EPR paradox \cite{EPR} as shown in Eq.\ (\ref{EPR}) and the concept
of steering introduced by Schr\" odinger \cite{Schr} as analyzed in Refs.\cite{W1,W2,W3}
were proven to be equivalent descriptions of non-locality. Moreover, the EPR-steering
turned out to be a different kind of non-locality stronger than quantum inseparability \cite{W1,W2}.
An important piece of progress in studying separability starting with the uncertainty principle was made
by Duan {\em et al.} \cite{Duan}. They have introduced another family of EPR-like uncertainties
in terms of the variances of the non-local one-parameter operators:
\begin{align}
& \hat Q(\alpha):=\alpha {\hat q}_1-\frac{1}{\alpha}{\hat q}_2, \qquad
\hat P_{\pm}(\alpha):=\alpha {\hat p}_1 \pm \frac{1}{\alpha}{\hat p}_2, \notag \\
& (\alpha >0).
\label{Duan}
\end{align}
The commutation relations
\begin{align}
[{\hat Q}(\alpha), {\hat P}_{\pm}(\alpha )]=i\left( {\alpha}^2 \mp \frac{1}{{\alpha}^2}\right) {\hat I}
\label{CR-Reid1}
\end{align}
lead to the product-form uncertainty relations for the EPR-like observables\ (\ref{Duan}),
\begin{align}
\Delta Q(\alpha)\Delta P_{\pm}(\alpha) \geqq \frac{1}{2}
\left| {\alpha}^2 \mp \frac{1}{{\alpha}^2}\right|,
\label{prodD}
\end{align}
which imply the sum-form inequalities:
\begin{align}
\left[ \Delta Q(\alpha)\right]^2+\left[ \Delta P_{\pm}(\alpha)\right]^2
\geqq \left| {\alpha}^2 \mp \frac{1}{{\alpha}^2}\right|.
\label{sumD}
\end{align}
In Ref.\cite{Duan}, the inequalities\ (\ref{sumD}) are strengthen for separable states:
a necessary condition of separability, consisting of one-parameter family of inequalities,
is thereby established for any two-mode state. Moreover, for the special class of TMGSs,
the strongest of these inequalities is proven to be also a sufficient condition of separability.
Some other necessary conditions of separability which employ pairs of more general
non-local observables depending on more parameters have been pointed out
\cite{Simon,MGVT,GMVT}. They are expressed in terms of the product
or the sum of the corresponding variances which include EPR-like correlations.
The aim of this paper is twofold. On the one hand, we tackle
the characterization of the separability of TMGSs based on the analysis
of EPR-like correlations. We formulate full criteria of separability for TMGSs,
using both the product and the sum of variances of two EPR-like observables.
On the other hand, we prove directly the equivalence between two approaches
of the separability property of a TMGS: an EPR-like one, initiated by Duan {\em et al.}
and developed in the present work, and the Simon PPT one.
The plan of the paper is as follows. In Sec. II, we recall the EPR-like necessary
conditions of separability for two-mode states, as discussed by Giovannetti {\em et al.}
in Ref. \cite{GMVT}. Section III is an overview of several useful properties
of the undisplaced TMGSs. In Sec. IV, the product function occurring in Eq.\ (\ref{prodR})
and one of the sum functions from Eq.\ (\ref{sumD}) are normalized according to
the formulae presented in Sec. II. We evaluate here their minimal values, which are
manifest separability indicators for any TMGS owing to the PPT separability
criterion \cite{Simon}. A two-parameter regularized sum function is examined
along the same lines in Sec. V, yielding another indicator of separability equally based
on Simon's PPT criterion. The formulae established in Secs. IV and V allow us to derive
the Peres-Simon PPT necessary condition of separability for a TMGS from each of the
three EPR-like corresponding necessary conditions we have employed. In Sec. VI,
we optimize a regularized two-mode sum-type correlation function within the framework
of the EPR-like approach developed by Duan {\em et al.} in Ref. \cite{Duan}.
We find a necessary and sufficient condition of separability related to the scaled
standard form II of the covariance matrix (CM). This EPR-like condition of separability
is explicitly proven to be fully equivalent to Simon's PPT one. Section VII summarizes
our main results insisting on the connection between the EPR-like and PPT
separability conditions for a TMGS. Appendix A presents some nontrivial identities
involving the standard-form parameters of the CM of a TMGS. In Appendix B,
we prove the positivity of the Hessian matrices of the three EPR-like correlation
functions examined in Secs. IV and V, when evaluated at their stationary points.
Appendix C points out the existence of a scaled standard form II of the CM,
which is exploited in Sec. VI.
\section{Separable two-mode states}
Let us consider a pair of EPR-like observables which are linear combinations
of the canonical quadrature operators of the two modes:
\begin{align}
& \hat Q:={\alpha}_1 {\hat q}_1- {\alpha}_2 {\hat q}_2, \qquad
\hat P:={\beta}_1 {\hat p}_1+{\beta}_2 {\hat p}_2, \notag \\
& ( {\alpha}_j >0, \;\; {\beta}_j >0: \quad j=1,2 ).
\label{GMVT}
\end{align}
The coordinates and momenta in Eq.\ (\ref{GMVT}) are defined in terms of the amplitude
operators of the modes:
\begin{align}
& \hat q_j:=\frac{1}{\sqrt{2}}\left( \hat a_j+\hat a_j^{\dag}\right), \qquad
\hat p_j:=\frac{1}{\sqrt{2}i}\left( \hat a_j-\hat a_j^{\dag}\right), \notag \\
& (j=1,2 ).
\label{quad}
\end{align}
Obviously, Reid's pair of one-parameter observables built with independent
parameters\ (\ref{Reid}), as well as the single-parameter one\ (\ref{Duan})
are particular cases of EPR-like operators\ (\ref{GMVT}). The commutation relation
\begin{align}
[{\hat Q}, {\hat P}]=i\left( {\alpha}_1{\beta}_1-{\alpha}_2{\beta}_2 \right) {\hat I}
\label{CR-GMVT}
\end{align}
shows that the operators\ (\ref{GMVT}) are genuine EPR observables if and only if
${\alpha}_1{\beta}_1= {\alpha}_2{\beta}_2.$ The corresponding Heisenberg
uncertainty relation,
\begin{align}
{\Delta Q}\, {\Delta P} \geqq \frac{1}{2}\left| {\alpha}_1{\beta}_1-{\alpha}_2{\beta}_2\right|,
\label{prodG}
\end{align}
entails the sum-type inequality
\begin{align}
\left( \Delta Q \right)^2+\left( \Delta P\right)^2
\geqq \left| {\alpha}_1{\beta}_1-{\alpha}_2{\beta}_2\right|.
\label{sumG}
\end{align}
If the two-mode state is separable, i. e., it is a convex combination of product states,
$$\hat \rho_s:=\sum_{k} w_k {\hat \rho}_1^{(k)}\otimes {\hat \rho}_2^{(k)}, \qquad
\left( w_k >0, \quad \sum_{k} w_k=1 \right),$$
then the product $(\Delta Q)_s(\Delta P)_s$ has a stronger lower bound
than in Eq.\ (\ref{prodG}) \cite{GMVT}:
\begin{align}
(\Delta Q)_s(\Delta P)_s \geqq \frac{1}{2}\left( {\alpha}_1{\beta}_1+{\alpha}_2{\beta}_2\right).
\label{prodsG}
\end{align}
Accordingly, for the sum of variances, an inequality stronger than Eq.\ (\ref{sumG}) holds:
\begin{align}
\left[ (\Delta Q)_s \right]^2+\left[ (\Delta P)_s\right]^2
\geqq {\alpha}_1{\beta}_1+{\alpha}_2{\beta}_2.
\label{sumsG}
\end{align}
It is useful to specialize the necessary conditions for the separability of a two-mode state,
Eqs.\ (\ref{prodsG}) and\ (\ref{sumsG}), to the pairs of EPR-like observables\ (\ref{Reid})
and\ (\ref{Duan}). We get the following two sets of inequalities:
\begin{align}
[\Delta Q(\lambda)]_s [\Delta P(\mu)]_s \geqq \frac{1}{2}\left( 1+\lambda \mu \right),
\label{prodsR}
\end{align}
\begin{align}
\left\{ [\Delta Q(\lambda)]_s \right\}^2+\left\{ [\Delta P(\mu)]_s \right\}^2
\geqq 1+\lambda \mu \, ;
\label{sumsR}
\end{align}
\begin{align}
[\Delta Q(\alpha)]_s [\Delta P_{\pm}(\alpha)]_s \geqq \frac{1}{2}\left( {\alpha}^2
+\frac{1}{{\alpha}^2} \right),
\label{prodsD}
\end{align}
\begin{align}
\left\{ [\Delta Q(\alpha)]_s \right\}^2+\left\{ [\Delta P_{\pm}(\alpha)]_s \right\}^2
\geqq {\alpha}^2+\frac{1}{{\alpha}^2}.
\label{sumsD}
\end{align}
The necessary conditions of separability\ (\ref{sumsD}) have first been derived
in Ref. \cite{Duan}.
\section{Two-mode Gaussian states}
For a long time, the Gaussian states (GSs) of the quantum radiation field are known to be
of central importance in various areas of quantum optics. In general, the GSs play
an important role for those quantum systems involving a quadratic bosonic Hamiltonian
that generates correlations between bosonic modes. They are achieved in condensed matter,
as well as in atomic ensembles such as trapped ions or Bose-Einstein condensates.
The GSs of light are also largely employed in quantum information processing
with continuous variables; their usefulness has been reviewed
in Refs. \cite{EP,BL,Ol,QC3}.
Especially accessible and insightful are the TMGSs. Here we recollect just the strictly
necessary notions and results \cite{PTH2001,PTH2003,ASI}
that enable us to discuss separability issues. Since these are not affected
by translations in the phase space, it is sufficient to deal with undisplaced
(zero-mean) TMGSs.
Recall that the characteristic function (CF) of an undisplaced TMGS $\hat \rho$
is a real exponential:
\begin{equation}
\chi (u)=\exp\left[ -\frac{1}{2}(Ju)^{T}{\mathcal V}(Ju) \right].
\label{CF}
\end{equation}
Its argument is a dimensionless vector $u \in {\mathbb R}^4$ whose components
are eigenvalues of the canonical quadrature operators of the modes:
\begin{equation}
u^T:=(q_1,\; p_1,\; q_2,\; p_2).
\label{uT}
\end{equation}
$J$ designates the standard matrix of the symplectic form on ${\mathbb R}^4$,
which is block-diagonal and skew-symmetric:
\begin{align}
J:=J_1 \oplus J_2: \qquad J_1=J_2:=\left(
\begin{matrix}
0 & 1\\ -1 & 0
\end{matrix}
\right).
\label{J}
\end{align}
The TMGS $\hat \rho$ is entirely specified, via the CF\ (\ref{CF}), by the real
and symmetric $4 \times 4$ CM, which is denoted ${\mathcal V}$. Its entries
are expectation values of products of the deviations from the means
of the quadratures\ (\ref{quad}). The order of the rows and columns
is indicated by the current vector\ (\ref{uT}). The CM ${\mathcal V}$ fulfills
the strong form (Robertson-Schr\"odinger) of the uncertainty relations
for the canonical quadrature observables\ (\ref{quad}):
\begin{equation}
{\mathcal V}+\frac{i}{2}J \geqslant 0.
\label{RS}
\end{equation}
The above requirement that the complex matrix ${\mathcal V}+\frac{i}{2}J$
has to be positive semidefinite is a necessary and sufficient condition
for the existence of the GS ${\hat \rho}$ \cite{Simon}. It implies the inequality
\begin{equation}
{\cal D}:=\det \left( {\mathcal V}+\frac{i}{2}J \right) \geqq 0,
\label{D}
\end{equation}
as well as the general property that the CM ${\mathcal V}$ is positive definite:
${\mathcal V}>0$. By contrast, the saturation equality ${\cal D}=0$
is a quite special feature. However, it is shared by all the pure GSs
and also by an interesting class of mixed ones. All these states are said
to be at the physicality edge.
It is often convenient to partition the CM ${\mathcal V}$ into $2\times 2$ submatrices:
\begin{align}
{\mathcal V}=\left(
\begin{matrix}
\mathcal V_{1}\; & \; {\mathcal C} \\ \\
{\mathcal C}^ T \; & \; \mathcal V_{2}
\end{matrix}
\right).
\label{part}
\end{align}
Here $\mathcal V_{j}, \; (j=1,\, 2)$, denote the CMs of the individual single-mode
reduced GSs, while $ {\mathcal C}$ displays the cross-correlations between the modes.
Let us mention that a symplectic transformation $S$ of the canonical quadrature
observables\ (\ref{quad}) in the Heisenberg picture induces a congruence
transformation of any CM:
\begin{equation}
{\mathcal V}^{\prime}=S\, {\mathcal V}\, S^T, \qquad S \in {\rm Sp}(4,\mathbb{R}).
\label{congr}
\end{equation}
At the same time, a unitary operator ${\hat U}(S)$ acting on the two-mode Fock space
${\mathcal H}_1 \otimes {\mathcal H}_2$ is associated to the symplectic matrix $S$.
Remarkably, the corresponding transformation of the TMGS in the Schr\"odinger picture,
${\hat \rho}^{\prime}={\hat U}(S)\, {\hat \rho}\, {\hat U}^{\dag}(S)$, preserves the
Gaussian nature of the state. Note that, if the symplectic transformation $S$ consists
of two separate single-mode ones,
\begin{equation}
S=S_1 \oplus S_2: \qquad S\in{\rm Sp}(2,\mathbb{R})\times {\rm Sp}(2,\mathbb{R}),
\label{local}
\end{equation}
then the above transformation of the TMGS ${\hat \rho}$ does not affect its amount
of entanglement. It has been proven in Ref. \cite{Duan} and just exploited
in Ref. \cite{Simon} the existence of a local symplectic transformation\ (\ref{local})
that leads to a standard form of the CM:
\begin{align}
{\mathcal V}(1,1):=\left(
\begin{matrix}
b_1\; & \; 0\; & \; c\; & \; 0 \\ \\
0\; & \; b_1\; & \; 0\; & \; d \\ \\
c\; & \; 0\; & \; b_2\; & \; 0 \\ \\
0\; & \; d\; & \; 0\; & \; b_2
\end{matrix}
\right),
\label{standard}
\end{align}
depending on four parameters $b_1,\, b_2,\, c,\, d$. With no loss of generality,
one can choose $b_1 \geqq b_2>0$ and $c \geqq |d|$. Following
Refs. \cite{Duan,Simon}, we apply two independent one-mode squeeze
transformations to the standard-form CM\ (\ref{standard}). Let us denote
the corresponding scaling factors by
\begin{equation}
u_1:=e^{2r_1}\geqq 1, \qquad u_2:=e^{2r_2}\geqq 1,
\label{scal}
\end{equation}
where $r_1$ and $r_2$ are the squeeze parameters in the two modes.
The transformed state has a {\em scaled standard-form} CM with the block
structure\ (\ref{part}):
\begin{align}
{\mathcal V}(u_1,\, u_2) =\left(
\begin{matrix}
{\mathcal V}_1(u_1)\; & \; {\mathcal C}\left( \sqrt{u_1u_2} \right) \\ \\
{\mathcal C}\left( \sqrt{u_1u_2} \right) \; & \; {\mathcal V}_2(u_2)
\end{matrix}
\right).
\label{ssfCM}
\end{align}
All its $2\times 2$ submatrices are diagonal. The CMs of the single-mode
reduced GSs read
\begin{align}
{\mathcal V}_j(u_j)=\left(
\begin{matrix}
b_j\,u_j\; & \; 0 \\ \\
0\; & \; \frac{b_j}{u_j}
\end{matrix}
\right), \qquad (j=1,2),
\label{V_j}
\end{align}
while the cross-correlation matrix is
\begin{align}
{\mathcal C}\left( \sqrt{u_1u_2} \right)=\left(
\begin{matrix}
c\sqrt{u_1u_2} \; & \; 0 \\ \\
0\; & \; \frac{d}{\sqrt{u_1u_2}}
\end{matrix}
\right).
\label{C}
\end{align}
The TMGSs whose CMs have a scaled standard form\ (\ref{ssfCM})-\ (\ref{C})
constitute a class characterized by the set of four standard-form parameters
$\{ \, b_1,\, b_2,\, c,\, d \} $ of the given TMGS ${\hat \rho}$, which are
${\rm Sp}(2,\mathbb{R})\times {\rm Sp}(2,\mathbb{R})$-invariant. The states
of this class are labeled by the pair of local squeeze factors $\{u_1,\, u_2\}$
\cite{Duan}. They are locally unitary similar to the given TMGS ${\hat \rho}$
and thereby possess precisely its amount of entanglement.
The Robertson-Schr\"odinger uncertainty relation\ (\ref{RS}) reduces
to the following restrictions for the standard-form parameters:
\begin{align}
b_1 \geqq \frac{1}{2}, \qquad b_2 \geqq \frac{1}{2};
\label{GS1}
\end{align}
\begin{align}
& b_1 b_2-c^2 \geqq \frac{1}{4} \max \left\{ \frac{b_1}{b_2}, \;
\frac{b_2}{b_1} \right\} \geqq \frac{1}{4}, \notag \\
& b_1 b_2-d^2 \geqq \frac{1}{4} \max \left\{ \frac{b_1}{b_2}, \;
\frac{b_2}{b_1} \right\} \geqq \frac{1}{4};
\label{GS2}
\end{align}
\begin{align}
& {\mathcal D}=\left( b_1 b_2-c^2 \right) \left( b_1 b_2-d^2 \right) \notag \\
& -\frac{1}{4} \left( b_1^2 +b_2^2+2cd \right) +\frac{1}{16} \geqq 0.
\label{GS3}
\end{align}
The inequalities\ (\ref{GS2}) give rise to another one involving the symplectic invariant
$\det({\mathcal V})=\left( b_1 b_2-c^2 \right) \left( b_1 b_2-d^2 \right) $ and thus
concerning the purity of the state:
\begin{align}
\det({\mathcal V}) \geqq \frac{1}{16} \quad \Longleftrightarrow \quad
{\rm Tr}\left( {\hat \rho}^2 \right) \leqq 1.
\label{purity}
\end{align}
According to Williamson's theorem \cite{W}, the CM ${\mathcal V}$ of any TMGS
${\hat \rho}$ is congruent to a diagonal CM via a symplectic matrix\ (\ref{congr}), which
is unique up to the sign. The corresponding diagonal entries, denoted
${\kappa}_{\pm},\,$ are positive, each one occurring twice. They are called
the symplectic eigenvalues of the CM ${\mathcal V}$ \cite{VW}. By virtue of Eq.\ (\ref{congr}),
the symplectic invariants $\det({\mathcal V})$ and ${\mathcal D}$ factor as follows:
\begin{align}
\det({\mathcal V})=\left( {\kappa}_{+} \right)^2 \left( {\kappa}_{-} \right)^2,
\label{detV1}
\end{align}
\begin{align}
{\mathcal D}=\left[ \left( {\kappa}_{+} \right)^2-\frac{1}{4} \right]
\left[ \left( {\kappa}_{-} \right)^2-\frac{1}{4} \right] \geqq 0,
\label{D1}
\end{align}
with
\begin{align}
{\kappa}_{+} \geqq {\kappa}_{-} \geqq \frac{1}{2}.
\label{kappa}
\end{align}
We employ Eqs.\ (\ref{detV1}) and\ (\ref{D1}) to get the symplectic eigenvalues in terms
of the standard-form parameters \cite{VW}:
\begin{align}
& \left( {\kappa}_{\pm} \right)^2=\frac{1}{2}\left[ \left(b_1^2+b_2^2 +2cd \right)
\pm \sqrt{\Delta} \right], \notag \\
& \Delta:=\left( b_1^2+b_2^2+2cd \right)^2 -4\, \det({\mathcal V}) \notag \\
& =\left( b_1^2-b_2^2 \right)^2+4\left( b_1 c+b_2 d \right) \left( b_2 c+b_1 d \right)
\geqq 0.
\label{kappa+-}
\end{align}
In order to formulate Simon's separability criterion for TMGSs, we have to consider
the partial transpose ${\hat \rho}^{\rm PT}$ of the TMGS ${\hat \rho}$. Basically, partial
transposition preserves the Gaussian character of the operator, does not modify
the standard-form parameters $\, b_1,\, b_2,\, c,\, $ and changes the sign of $d$.
Therefore, ${\hat \rho}^{\rm PT}$ is an undisplaced two-mode Gaussian operator whose CM
${\mathcal V}^{\rm PT}$ has the standard-form parameters $\{ \, b_1,\, b_2,\, c,\, -d \} $
\cite{Simon}. First, Simon has proven a lemma asserting that any TMGS with $d \geqq 0$
is separable. Then, he has shown that Peres' necessary condition of separability,
which claims that ${\hat \rho}^{\rm PT}$ should be a quantum state, i. e.,
\begin{equation}
{\mathcal V}^{\rm PT}+\frac{i}{2}J \geqslant 0,
\label{Peres}
\end{equation}
is also a {\em sufficient} one \cite{Simon}. The condition\ (\ref{Peres})
for the existence of a GS ${\hat \rho}^{\rm PT}$ reduces to the inequality
\begin{equation}
{\mathcal D}^{\rm PT}:=\det \left( {\mathcal V}^{\rm PT}+\frac{i}{2}J \right) \geqq 0,
\label{DPT}
\end{equation}
which is Simon's criterion of separability. It is usually written in the form\ (\ref{GS3})
with $d \to -d$. Therefore, a TMGS is separable if and only if the following inequality
is fulfilled:
\begin{align}
{\mathcal D}^{\rm PT}=\det({\mathcal V})-\frac{1}{4}\left( b_1^2+b_2^2-2 c d \right)
+\frac{1}{16} \geqq 0.
\label{DPT1}
\end{align}
Accordingly, the identity
\begin{align}
{\mathcal D}^{\rm PT}={\mathcal D}+c d
\label{d>0}
\end{align}
displays the property that all the TMGSs with $d \geqq 0$ are separable,
in agreement with Simon's lemma \cite{Simon}.
\section{Normalized separability indicators}
We write the variances of the EPR-like observables\ (\ref{GMVT}) for a TMGS
whose CM has a scaled standard form\ (\ref{ssfCM})--\ (\ref{C}):
\begin{align}
\left( \Delta Q \right)^2={\alpha}_1^2\, b_1 u_1+{\alpha}_2^2\, b_2 u_2
-2{\alpha}_1{\alpha}_2 c \sqrt{u_1u_2}\, ,
\label{varQ}
\end{align}
\begin{align}
\left( \Delta P \right)^2={\beta}_1^2\, \frac{b_1}{u_1}+{\beta}_2^2\, \frac{b_2}{u_2}
+2{\beta}_1{\beta}_2 \frac{d}{\sqrt{u_1u_2}}.
\label{varP}
\end{align}
Remark that, while all the TMGSs\ (\ref{ssfCM})--\ (\ref{C}) belonging to a class
with fixed standard-form parameters possess the same amount of entanglement,
their EPR-like variances\ (\ref{varQ}) and\ (\ref{varP}) depend, in addition,
on the local squeezings $u_1, u_2$. In what follows, we consider two functions
which are normalized as suggested by the separability
lower bounds\ (\ref{prodsG}) and\ (\ref{sumsG}):
\begin{align}
E\left( {\alpha}_1\, {\beta}_1,\, {\alpha}_2,\, {\beta}_2,\, u_1,\, u_2 \right)
:=\frac{\left( \Delta Q \right)^2 \left( \Delta P \right)^2}
{\left( {\alpha}_1{\beta}_1+{\alpha}_2{\beta}_2\right)^2},
\label{E1}
\end{align}
\begin{align}
F\left( {\alpha}_1,\, {\beta}_1,\, {\alpha}_2,\, {\beta}_2,\, u_1,\, u_2 \right)
:=\frac{\left( \Delta Q \right)^2+\left( \Delta P \right)^2}
{{\alpha}_1{\beta}_1+{\alpha}_2{\beta}_2}.
\label{f1}
\end{align}
According to Eqs.\ (\ref{prodsG}) and\ (\ref{sumsG}), for any separable
two-mode state (Gaussian or non-Gaussian), the following inequalities
are satisfied:
\begin{align}
E \geqq \frac{1}{4}, \qquad F \geqq 1.
\label{EF}
\end{align}
In the sequel, we develop a two-step program. First, in order
to handle simpler functions, we diminish the number of their independent variables
as much as possible by making suitable substitutions or choices. Second,
we find the absolute minima of the resulting simpler functions, which prove to be
manifest separability indicators. It is sufficient to restrict our search in this step
to values $d<0$ in the variance\ (\ref{varP}).
\subsection{Product form}
We start by absorbing the scaling factors in Eqs.\ (\ref{varQ}) and\ (\ref{varP})
into four new positive parameters,
\begin{align}
{\alpha}_j^{\prime}= {\alpha}_j\sqrt{u_j}, \quad
{\beta}_j^{\prime}= {\beta}_j \frac{1}{\sqrt{u_j}}, \quad (j=1,2),
\label{prime}
\end{align}
so that the variances become:
\begin{align}
\left( \Delta Q \right)^2= b_1\left( {\alpha}_1^{\prime} \right)^2
+b_2\left( {\alpha}_2^{\prime} \right)^2 -2c\,{\alpha}_1^{\prime}{\alpha}_2^{\prime},
\label{varQ1}
\end{align}
\begin{align}
\left( \Delta P \right)^2=b_1\left( {\beta}_1^{\prime} \right)^2
+b_2\left( {\beta}_2^{\prime} \right)^2 +2d\,{\beta}_1^{\prime}{\beta}_2^{\prime}.
\label{varP1}
\end{align}
Note the identity
\begin{align}
{\alpha}_1{\beta}_1+{\alpha}_2{\beta}_2={\alpha}_1^{\prime}{\beta}_1^{\prime}
+{\alpha}_2^{\prime}{\beta}_2^{\prime}.
\label{denom}
\end{align}
After replacing Eqs.\ (\ref{varQ1})-\ (\ref{denom}) into Eq.\ (\ref{E1}),
the function $E\left( {\alpha}_1^{\prime}\, {\beta}_1^{\prime},\, {\alpha}_2^{\prime},\,
{\beta}_2^{\prime} \right) $ can be further simplified. As a matter of fact, it is
a function of two positive variables,
\begin{align}
\lambda:=\frac{{\alpha}_2^{\prime}}{{\alpha}_1^{\prime}}, \qquad
\mu:=\frac{{\beta}_2^{\prime}}{{\beta}_1^{\prime}},
\label{lam-mu}
\end{align}
which reads:
\begin{align}
E\left( \lambda,\, \mu \right) :=\frac{\left[ \Delta Q (\lambda) \right]^2
\left[ \Delta P(\mu) \right]^2}{\left( 1+ \lambda \mu \right)^2}.
\label{E3}
\end{align}
The numerator of the fraction in the r. h. s. of Eq.\ (\ref{E3}) is equal to the product
of the variances of Reid's EPR-like observables\ (\ref{Reid}) in a standard-form
TMGS\ (\ref{standard}):
\begin{align}
\left[ \Delta Q (\lambda) \right]^2 =b_1 +b_2{\lambda}^2 -2c \lambda,
\label{varQ2}
\end{align}
\begin{align}
\left[ \Delta P(\mu) \right]^2 =b_1 +b_2{\mu}^2 +2d \mu.
\label{varP2}
\end{align}
We prove the following statement.
{\bf Theorem 1.} {\em A TMGS is separable if the absolute minimum
of the function $E\left( \lambda,\, \mu \right)$ is greater than or at least
equal to 1/4:}
\begin{align}
E_m:=\min_{ \{ \lambda,\, \mu \} } E\left( \lambda,\, \mu \right) \geqq \frac{1}{4}.
\label{minE}
\end{align}
{\em Proof}. The first-order derivatives
\begin{align}
& \frac{\partial \ln(E)}{\partial \lambda}=\frac{2 \left( b_2{\lambda}-c \right) }
{b_1 +b_2{\lambda}^2 -2c \lambda}-\frac{2\mu}{1+ \lambda \mu}, \notag \\
& \frac{\partial \ln(E)}{\partial \mu}=\frac{2 \left( b_2{\mu}+d \right) }
{b_1 +b_2{\mu}^2 +2d \mu}-\frac{2\lambda}{1+ \lambda \mu}
\label{der1}
\end{align}
vanish at a stationary point of the function\ (\ref{E3}). The resulting equations,
\begin{align}
& \left[ \Delta Q (\lambda) \right]^2 =\frac{1}{\mu} \left( b_2{\lambda}-c \right)
\left( 1+ \lambda \mu \right), \notag\\
& \left[ \Delta P(\mu) \right]^2 =\frac{1}{\lambda} \left( b_2{\mu}+d \right)
\left( 1+ \lambda \mu \right),
\label{stat}
\end{align}
have a unique solution:
\begin{align}
& {\lambda}_m=\frac{\left( b_1^2-b_2^2 \right)+\sqrt{{\Delta}^{\rm PT}}}
{2\left( b_1 c -b_2 d \right)}, \quad
{\mu}_m=\frac{\left( b_1^2-b_2^2 \right)+\sqrt{{\Delta}^{\rm PT}}}
{2\left( b_2 c -b_1 d \right)},
\label{minp}
\end{align}
where ${\Delta}^{\rm PT}$ is the discriminant in Eq.\ (\ref{A5}). In Appendix B
we have proven that the value\ (\ref{B4}),
\begin{align}
E_m=\frac{ \left( b_2{\lambda}_m-c \right)\left( b_2{\mu}_m+d \right)} { {\lambda}_m {\mu}_m},
\label{E_m}
\end{align}
is the absolute minimum of the function $E\left( \lambda,\, \mu \right) .$
By use of Eqs.\ (\ref{minp}) and\ (\ref{A5}), we find the formula
\begin{align}
E_m=\left( {\kappa}_{-}^{\rm PT} \right)^2.
\label{Emin}
\end{align}
When $d<0$, Simon's separability condition\ (\ref{DPT1}) in conjunction
with Eq.\ (\ref{Emin}) gives the alternative:
$$E_m \geqq \frac{1}{4} \;\;\Longleftrightarrow \;\; {\hat \rho} \;\; {\rm separable},
\qquad (d<0),$$
$$E_m < \frac{1}{4} \;\; \Longleftrightarrow \;\; {\hat \rho} \;\; {\rm entangled},
\qquad (d<0).$$
As $\left( {\kappa}_{-}^{\rm PT} \right)^2 \geqq \frac{1}{4} \Longleftrightarrow
{\mathcal D}^{\rm PT} \geqq 0,$ the first inequality holds also for $d \geqq 0$
due to the identity\ (\ref{d>0}). This result is in accordance with Simon's lemma
regarding the separability of any TMGS with $d \geqq 0$ \cite{Simon}.
We put together the above results to draw the general conclusion:
\begin{align}
& E_m \geqq \frac{1}{4} \;\;\Longleftrightarrow \;\; {\hat \rho} \;\; {\rm separable}, \notag \\
& E_m < \frac{1}{4} \;\; \Longleftrightarrow \;\; {\hat \rho} \;\; {\rm entangled}.
\label{Esep}
\end{align}
This necessary and sufficient condition of separability for TMGSs concludes the proof.
Therefore, the minimum of the normalized product\ (\ref{E3}) of two EPR-like
uncertainties is equal to the square of the smallest symplectic eigenvalue
${\kappa}_{-}^{\rm PT}$. Owing to the PPT separability criterion, it is itself
a separability indicator. A previous relationship between ${\kappa}_{-}^{\rm PT}$
and a product of EPR-like uncertainties was found by a different treatment
and from another perspective in Ref. \cite{Poon-Law}. Quite recently, the present
approach was successfully applied in Ref. \cite{He-Reid} to a special class of TMGSs,
namely, the squeezed thermal states (STSs): $c=-d>0$ \cite{PTH2003}.
An interesting related problem for TMGSs regarding steerability criteria was recently
put forward by Kogias and Adesso \cite{KA}. These authors have considered
the product of EPR-like uncertainties occurring in the Reid condition\ (\ref{EPR})
and have minimized it with respect to the pair of parameters $\{ \lambda,\mu \}$
and the local variables of a TMGS. They have thus recovered the explicit
symplectically invariant formula for the condition of steerability of GSs
by Gaussian measurements which has first been written in Refs. \cite{W1,W2}.
\subsection{Sum form}
We find it convenient to simplify the function\ (\ref{f1}) by setting
${\alpha}_1={\alpha}_2:=\alpha,\, {\beta}_1={\beta}_2=\frac{1}{\alpha}:$
\begin{align}
F\left( {\alpha}^2,\, u_1,\, u_2 \right) :=\frac{ \left[ \Delta Q(\alpha) \right]^2
+\left[ \Delta P_{+}(\alpha) \right]^2 }{ {\alpha}^2 +\frac{1}{{\alpha}^2} }.
\label{F2}
\end{align}
The correlation function\ (\ref{F2}) is built with the variances of a pair of non-local
one-parameter observables\ (\ref{Duan}) introduced
in Ref. \cite{Duan}:
\begin{align}
\left[ \Delta Q(\alpha) \right]^2=b_1 u_1 {\alpha}^2
+b_2 u_2 \frac{1}{ {\alpha}^2 } -2c\sqrt{u_1\, u_2},
\label{varQ4}
\end{align}
\begin{align}
\left[ \Delta P_{\pm}(\alpha) \right]^2 =\frac{b_1}{u_1} {\alpha}^2
+\frac{b_2}{u_2} \frac{1}{ {\alpha}^2 } \pm \frac{2d}{ \sqrt{u_1\, u_2} }.
\label{varP4}
\end{align}
Therefore, it reads:
\begin{align}
& F\left( {\alpha}^2,\, u_1,\, u_2 \right) =\frac{1}{ {\alpha}^4 +1}
\left[ b_1 \left( u_1+ \frac{1}{u_1} \right) {\alpha}^4
-2\left( c\sqrt{u_1\, u_2} \right. \right. \notag \\
& \left. \left. -d \frac{1}{ \sqrt{u_1\, u_2} }\right) {\alpha}^2
+b_2 \left( u_2+ \frac{1}{u_2} \right) \right].
\label{F3}
\end{align}
We state
{\bf Theorem 2.} {\em A TMGS is separable if the absolute minimum
of the function $F\left( {\alpha}^2,\, u_1,\, u_2 \right)$ is greater than or at least
equal to 1:}
\begin{align}
F_m:=\min_{ \{ {\alpha}^2,\, u_1,\, u_2 \} } F\left( {\alpha}^2,\, u_1,\, u_2 \right) \geqq 1.
\label{minF}
\end{align}
{\em Proof}. The first-order derivatives
\begin{align}
& \frac{\partial F}{\partial \left( {\alpha}^2 \right)} =\frac{2}{\left( {\alpha}^4 +1 \right)^2 }
\left\{ \left( c\sqrt{u_1\, u_2} -\frac{d}{ \sqrt{u_1\, u_2} } \right)
\left( {\alpha}^4 -1 \right) \right. \notag \\
& \left. +\left[ b_1 \left( u_1+ \frac{1}{u_1} \right)
-b_2 \left( u_2+ \frac{1}{u_2} \right) \right] {\alpha}^2 \right\} ,
\label{dFa}
\end{align}
\begin{align}
& \frac{\partial F}{\partial u_1} =\frac{ {\alpha}^2 }{\left( {\alpha}^4 +1 \right) u_1 }
\left[ b_1 \left( u_1 -\frac{1}{u_1} \right) {\alpha}^2 \right. \notag \\
& \left. -\left( c\sqrt{u_1\, u_2} +\frac{d}{ \sqrt{u_1\, u_2} }\right) \right] ,
\label{dFu1}
\end{align}
\begin{align}
& \frac{\partial F}{\partial u_2} =\frac{ {\alpha}^2 }{\left( {\alpha}^4 +1 \right) u_2 }
\left[ b_2 \left( u_2 -\frac{1}{u_2} \right) \frac{1}{ {\alpha}^2 } \right. \notag \\
& \left. -\left( c\sqrt{u_1\, u_2} +\frac{d}{ \sqrt{u_1\, u_2} }\right) \right]
\label{dFu2}
\end{align}
vanish at a stationary point of the function\ (\ref{F3}). We try to solve
the resulting system of stationarity equations:
\begin{align}
& \left( c\sqrt{u_1\, u_2} -\frac{d}{ \sqrt{u_1\, u_2} } \right)
\left( 1-{\alpha}^4 \right) \notag \\
& =\left[ b_1 \left( u_1+ \frac{1}{u_1} \right)
-b_2 \left( u_2+ \frac{1}{u_2} \right) \right] {\alpha}^2 ,
\label{sta}
\end{align}
\begin{align}
b_1 \left( u_1 -\frac{1}{u_1} \right) {\alpha}^2
=c\sqrt{u_1\, u_2} +\frac{d}{ \sqrt{u_1\, u_2} },
\label{stu1}
\end{align}
\begin{align}
b_2 \left( u_2 -\frac{1}{u_2} \right) \frac{1}{ {\alpha}^2 }
=c\sqrt{u_1\, u_2} +\frac{d}{ \sqrt{u_1\, u_2} }.
\label{stu2}
\end{align}
From Eqs.\ (\ref{stu1}) and\ (\ref{stu2}) it follows:
\begin{align}
{\alpha}^4 =\frac{ b_2 \left( u_2 -\frac{1}{u_2} \right) }
{ b_1 \left( u_1 -\frac{1}{u_1} \right) }\, ;
\label{alpha}
\end{align}
\begin{align}
& b_1 b_2 \left( u_1 -\frac{1}{u_1} \right) \left( u_2 -\frac{1}{u_2} \right) \notag \\
& =\left( c\sqrt{u_1 u_2} +\frac{d}{ \sqrt{u_1 u_2} } \right)^2 \,;
\label{u1u2}
\end{align}
\begin{align}
\left[ \Delta P_{+}(\alpha) \right]^2 =\left[ \Delta Q(\alpha) \right]^2.
\label{eqvar}
\end{align}
Insertion of Eqs.\ (\ref{stu1}) and\ (\ref{alpha}) into Eq.\ (\ref{sta})
gives the proportionality relation
\begin{align}
u_2={\gamma} u_1, \qquad \gamma:=\frac{b_2 c-b_1 d}{b_1 c-b_2 d} \leqq 1.
\label{gamma}
\end{align}
By substituting it into Eq.\ (\ref{u1u2}), we get a quadratic equation
in the product $p:=u_1 u_2 \geqq 1:$
\begin{align}
& \left( b_1 b_2 -c^2 \right) p^2 -\left[ b_1 b_2\left( \gamma +\frac{1}{\gamma}\right)
+2cd \right] p \notag \\
& +\left( b_1 b_2 -d^2 \right) =0
\label{eqp}
\end{align}
with
\begin{align}
\gamma +\frac{1}{\gamma}= 2+\frac{\left[ \left( b_1 -b_2 \right) \left( c +d \right) \right]^2 }
{\left( b_1 c-b_2 d \right) \left( b_2 c-b_1 d \right) }.
\label{sum}
\end{align}
Let us denote by ${\Delta}_p$ the discriminant of the quadratic trinomial
in Eq.\ (\ref{eqp}) and let $p_{\pm}$ be its roots. Making use of Eq.\ (\ref{A5}),
we find the relation
\begin{align}
{\Delta}_p=\left[ \frac{b_1 b_2 \left( c^2 -d^2 \right) }
{\left( b_1 c -b_2 d \right) \left( b_2 c -b_1 d \right) } \right]^2 {\Delta}^{\rm PT} \geqq 0.
\label{Dp}
\end{align}
Since $p_ {-} < 1$ for $c+d > 0$, the only acceptable solution of the quadratic
equation\ (\ref{eqp}) is
\begin{align}
p_{+}=\frac{c \left( b_1 b_2 -d^2 \right) -d \left( {\kappa}_{-}^{\rm PT} \right)^2 }
{-d \left( b_1 b_2 -c^2 \right) +c \left( {\kappa}_{-}^{\rm PT} \right)^2 } \geqq 1.
\label{p+}
\end{align}
With Eqs.\ (\ref{gamma}) and\ (\ref{p+}) we get the scaling factors:
\begin{align}
& u_{1m} =\left[ \frac{b_2 \left( b_1 b_2 -d^2 \right) -b_1 \left( {\kappa}_{-}^{\rm PT} \right)^2 }
{b_2 \left(b_1 b_2 -c^2 \right) -b_1 \left( {\kappa}_{-}^{\rm PT} \right)^2 } \right]^{\frac{1}{2} },
\notag \\
& u_{2m} =\left[ \frac{b_1 \left( b_1 b_2 -d^2 \right) -b_2 \left( {\kappa}_{-}^{\rm PT} \right)^2 }
{b_1 \left(b_1 b_2 -c^2 \right) -b_2 \left( {\kappa}_{-}^{\rm PT} \right)^2 } \right]^{\frac{1}{2} }.
\label{u_jm}
\end{align}
In view of Eq.\ (\ref{A5}), they have the alternative expressions:
\begin{align}
& u_{1m} =\left\{ \frac{ b_1 \left[ \sqrt{ {\Delta}^{\rm PT} }-\left( b_1^2 -b_2^2 \right) \right]
+2d \left( b_1 c -b_2 d \right) }
{ b_1 \left[ \sqrt{ {\Delta}^{\rm PT} }-\left( b_1^2 -b_2^2 \right) \right]
-2c \left( b_2 c -b_1 d \right) } \right\}^{\frac{1}{2} },
\notag \\
& u_{2m} =\left\{ \frac{ b_2 \left[ \sqrt{ {\Delta}^{\rm PT} }+\left( b_1^2 -b_2^2 \right) \right]
+2d \left( b_2 c -b_1 d \right) }
{ b_2 \left[ \sqrt{ {\Delta}^{\rm PT} }+\left( b_1^2 -b_2^2 \right) \right]
-2c \left( b_1 c -b_2 d \right) } \right\}^{\frac{1}{2} }.
\label{u_km}
\end{align}
Insertion of the scaling factors\ (\ref{u_km}) into Eq.\ (\ref{alpha})
leads to the following value of the EPR-like parameter $\alpha$:
\begin{align}
{\alpha}_m=\left[ \frac{ \sqrt{ {\Delta}^{\rm PT} }-\left( b_1^2 -b_2^2 \right) }
{ \sqrt{ {\Delta}^{\rm PT} }+\left( b_1^2 -b_2^2 \right) } \right]^{\frac{1}{4} } \leqq 1.
\label{a_m}
\end{align}
There is therefore a single stationary value of the function
$ F\left( {\alpha}^2,\, u_1,\, u_2 \right) $, Eq.\ (\ref{F3}):
\begin{align}
F_m = F\left( {\alpha}_m^2, u_{1m}, u_{2m} \right).
\label{F_m}
\end{align}
In fact, this is its absolute minimum, as proven in Appendix B.
With the solution\ (\ref{u_jm}) and\ (\ref{a_m}), we take advantage
of Eqs.\ (\ref{eqvar}), \ (\ref{A6}), and\ (\ref{A8}) to find the minimum value
\begin{align}
F_m=2 {\kappa}_{-}^{\rm PT}.
\label{Fmin}
\end{align}
By the same argument as before Eq.\ (\ref{Esep}), the PPT separability
condition\ (\ref{DPT1}) implies, via Eq.\ (\ref{Fmin}), the alternative:
\begin{align}
& F_m \geqq 1 \;\; \Longleftrightarrow \;\; {\hat \rho} \;\; {\rm separable}, \notag \\
& F_m < 1 \;\; \Longleftrightarrow \;\; {\hat \rho} \;\; {\rm entangled}.
\label{Fsep}
\end{align}
This concludes the proof.
It is worth mentioning that Eqs.\ (\ref{u_km}) and\ (\ref{a_m}), which specify
the coordinates of the minimum point, become much simpler for two classes of TMGSs,
namely, the STSs and the symmetric states. We find it convenient to write down here
the corresponding formulae.
\begin{enumerate}
\item {\em Two-mode STSs}:
\begin{align}
& c=-d>0 \;\; \Longleftrightarrow \;\; u_{1m}=u_{2m}=1: \qquad \gamma=1; \notag \\
& \sqrt{ {\Delta}^{\rm PT} }=\left( b_1 +b_2 \right) \sqrt{\delta}, \qquad
\delta:=\left( b_1 -b_2 \right)^2 +4c^2 >0: \notag \\
& \left( {\alpha}_m \right)^2= \frac{1}{2c}\left[ \sqrt{\delta}-\left( b_1 -b_2 \right) \right].
\label{STS}
\end{align}
\item {\em Symmetric TMGSs}:
\begin{align}
& b_1=b_2 =:b \;\; \Longrightarrow \;\; u_{1m}=u_{2m}=\sqrt{\frac{b+d}{b-c} }:
\quad \gamma=1; \notag \\
& \sqrt{ {\Delta}^{\rm PT} }=2b(c-d) >0, \qquad {\alpha}_m =1.
\label{sym}
\end{align}
\end{enumerate}
It is interesting to notice that in Ref.\cite{ASI} the minimal EPR-like uncertainty
in sum form was explicitly evaluated for symmetric TMGSs. The
authors have interpreted therein the smallest symplectic eigenvalue of the partially
transposed state as a quantifier of the greatest amount of EPR correlations which can be created
in a symmetric TMGS by means of local operations. Our present findings
in the general case, Eq.\ (\ref{Emin}) for the product form of the normalized
EPR-like uncertainties, and Eq.\ (\ref{Fmin})
for the sum-form ones, appear to be consistent with this interpretation.
To complete this section, we write the appropriate necessary conditions
of separability\ (\ref{EF}) for a TMGS:
\begin{align}
E\left( \lambda,\, \mu \right) \geqq \frac{1}{4}, \qquad
F\left( {\alpha}^2,\, u_1,\, u_2 \right) \geqq 1.
\label{EFGS}
\end{align}
When $d<0$, these inequalities hold for any values of the above functions,
including their minima\ (\ref{Emin}) and\ (\ref{Fmin}), respectively:
\begin{align}
E_m=\left( {\kappa}_{-}^{\rm PT} \right)^2 \geqq \frac{1}{4}, \qquad
F_m=2 {\kappa}_{-}^{\rm PT} \geqq 1.
\label{EmFm}
\end{align}
Each inequality\ (\ref{EmFm}) is equivalent to the Peres-Simon necessary
condition of separability:
\begin{align}
{\kappa}_{-}^{\rm PT} \geqq \frac{1}{2} \; \; \Longleftrightarrow \; \; {\cal D}^{\rm PT} \geqq 0.
\label{PS}
\end{align}
When $d \geqq 0$, Eq.\ (\ref{PS}) is equally valid, owing to the identity\ (\ref{d>0}).
To sum up, for any TMGS, both the EPR-like necessary conditions
of separability\ (\ref{EFGS}) imply the PPT condition\ (\ref{PS}).
On the other hand, Theorems 1 and 2 yield the above absolute minima\ (\ref{EmFm}).
Accordingly, Simon's separability criterion for a TMGS\ (\ref{DPT1}) implies that each
identity\ (\ref{EFGS}) is itself a sufficient condition of separability, as implicitly
expressed by Eqs.\ (\ref{Esep}) and\ (\ref{Fsep}).
\section{Separability indicator in regularized sum form}
We start from the Reid's pair of EPR-like observables\ (\ref{Reid}). One gets
their variances by setting
${\alpha}_1=1,\; {\alpha}_2=:\lambda,\; {\beta}_1=1,\; {\beta}_2=:\mu$
in Eqs.\ (\ref{varQ}) and\ (\ref{varP}):
\begin{align}
\left( \Delta Q \right)^2=b_1 u_1 +b_2 u_2 {\lambda}^2 -2c \sqrt{u_1 u_2}\, \lambda,
\label{varQ1a}
\end{align}
\begin{align}
\left( \Delta P \right)^2=\frac{b_1}{u_1} +\frac{b_2}{u_2}\,{\mu}^2
+2\frac{d}{\sqrt{u_1 u_2} }\,\mu.
\label{varP1a}
\end{align}
We assume that $d<0$ in Eq.\ (\ref{varP1a}) and in the subsequent ones. It is convenient
to absorb the scaling factor $u_2$ into a pair of new EPR-like parameters:
\begin{align}
\xi:=\sqrt{u_2}\, \lambda >0, \qquad \eta:=\frac{1}{\sqrt{u_2} }\, \mu >0.
\label{xi-eta}
\end{align}
This simplifies the variances\ (\ref{varQ1a}) and\ (\ref{varP1a}) as follows:
\begin{align}
\left[ {\Delta Q}(\xi) \right]^2 =b_1 u_1 +b_2 {\xi}^2 -2c \sqrt{u_1}\, \xi,
\label{varQ3}
\end{align}
\begin{align}
\left[ {\Delta P}(\eta) \right]^2 =\frac{b_1}{u_1} +b_2 {\eta}^2
+2\frac{d}{\sqrt{u_1} }\,\eta.
\label{varP3}
\end{align}
We take inspiration from the separability condition\ (\ref{sumsG}) to build
an EPR-like correlation function of the remaining independent variables
$\xi,\, \eta,\, u_1:$
\begin{align}
G\left( \xi,\, \eta,\, u_1 \right):=\left[ {\Delta Q}(\xi) \right]^2 +\left[ {\Delta P}(\eta) \right]^2
-\left( 1+\xi \eta \right).
\label{G}
\end{align}
The function\ (\ref{G}), which is non-negative for any separable two-mode state,
has the explicit form:
\begin{align}
& G\left( \xi,\, \eta,\, u_1 \right)=\left( b_1 u_1 +b_2 {\xi}^2 -2c \sqrt{u_1}\, \xi \right) \notag \\
& +\left( \frac{b_1}{u_1} +b_2 {\eta}^2 +2\frac{d}{\sqrt{u_1} }\,\eta \right) -\left( 1+\xi \eta \right).
\label{G1}
\end{align}
The following statement is true.
{\bf Theorem 3.} {\em A TMGS is separable if the absolute minimum
of the function $G\left( \xi,\, \eta,\, u_1 \right)$ is non-negative:}
\begin{align}
G_m:=\min_{ \{ \xi,\, \eta,\, u_1 \} } G\left( \xi,\, \eta,\, u_1 \right) \geqq 0.
\label{minG}
\end{align}
{\em Proof}. We write the stationarity conditions:
\begin{align}
& \frac{\partial G}{\partial \xi} =0: \qquad 2b_2 \xi -\eta = 2c\sqrt{u_1}, \notag \\
& \frac{\partial G}{\partial \eta} =0: \qquad -\xi +2b_2 \eta =-2\frac{d}{\sqrt{u_1} } \notag \\
& \frac{\partial G}{\partial u_1} =0: \qquad c u_1 \xi +d \eta
=\frac{b_1}{\sqrt{u_1} }\left( u_1^2 -1 \right).
\label{d1}
\end{align}
The system\ (\ref{d1}) is linear in the variables\ (\ref{xi-eta}). By solving
the first two equations with respect to $\xi$ and $\eta$ and then replacing
the result into the third one, we find its unique solution:
\begin{align}
& {\xi}_m =\frac{1}{b_2^2 -\frac{1}{4} }\left(b_2 c \sqrt{u_{1m} }
-\frac{1}{2} \frac{d}{ \sqrt{u_{1m} } } \right), \notag \\
& {\eta}_m =\frac{1}{b_2^2 -\frac{1}{4} } \left( \frac{1}{2} c \sqrt{u_{1m} }
-b_2 \frac{d}{ \sqrt{u_{1m} } } \right), \notag \\
& u_{1m}=\left[ \frac{b_2 \left( b_1 b_2-d^2 \right) -\frac{1}{4} b_1}
{b_2 \left( b_1 b_2-c^2 \right) -\frac{1}{4} b_1} \right]^{\frac{1}{2} }.
\label{mp}
\end{align}
In Appendix B we have proven that the only stationary value of the function\ (\ref{G1}),
\begin{align}
G_m = G\left( {\xi}_m,\, {\eta}_{m},\, u_{1m} \right),
\label{G_m}
\end{align}
is its absolute minimum. Insertion of the coordinates\ (\ref{mp}) into Eq.\ (\ref{G1})
and a subsequent straightforward calculation yields the formula
\begin{align}
& G_m=\frac{1}{b_2^2 -\frac{1}{4} }
\left( 2\left\{ \left[ b_2 \left( b_1 b_2-c^2 \right)-\frac{1}{4}b_1 \right] \right. \right. \notag \\
& \left. \left. \times \left[ b_2 \left( b_1 b_2-d^2 \right) -\frac{1}{4}b_1 \right] \right\}^{\frac{1}{2} }
-\left(b_2^2 -\frac{1}{4} -c d\right) \right).
\label{G_m1}
\end{align}
Making use of the identity\ (\ref{A10}), the minimum\ (\ref{G_m1}) can be cast into the form:
\begin{align}
&G_m =4\,{\mathcal D}^{\rm PT} \left( 2\left\{ \left[ b_2 \left( b_1 b_2-c^2 \right)-\frac{1}{4}b_1 \right]
\right. \right. \notag \\
& \left. \left. \times \left[ b_2 \left( b_1 b_2-d^2 \right) -\frac{1}{4}b_1 \right] \right\}^{\frac{1}{2} }
+\left(b_2^2 -\frac{1}{4} -c d\right) \right)^{-1}.
\label{G_m2}
\end{align}
Because $d<0$, Eq.\ (\ref{G_m2}) displays the feature that the minimum value $G_m$
and the symplectic invariant ${\mathcal D}^{\rm PT}$ have the same sign.
On the other hand, when $d \geqq 0$, the TMGS is separable owing to Simon's lemma.
Then, on account of Eqs.\ (\ref{sumsG}) and\ (\ref{d>0}), both quantities are non-negative.
Accordingly, the PPT condition of separability\ (\ref{DPT1}) imposes the alternative:
\begin{align}
& G_m \geqq 0 \;\;\Longleftrightarrow \;\; {\hat \rho} \;\; {\rm separable}, \notag \\
& G_m <0 \;\; \Longleftrightarrow \;\; {\hat \rho} \;\; {\rm entangled}.
\label{Gsep}
\end{align}
The proof is complete.
We conclude this section by writing the above-mentioned condition of separability:
if a given TMGS ${\hat \rho}$ is separable, then the function\ (\ref{G1}) is non-negative,
\begin{equation}
G\left( \xi,\, \eta,\, u_1 \right) \geqq 0.
\label{G1>0}
\end{equation}
It follows the alternative: if $d<0$, then its minimum\ (\ref{G_m2}) is also non-negative;
if $d \geqq 0$, then, owing to Eq.\ (\ref{d>0}), the determinant $ {\mathcal D}^{\rm PT} $
is non-negative:
\begin{align}
& d<0: \qquad G_m \geqq 0 \;\; \Longleftrightarrow \; \;
{\mathcal D}^{\rm PT} \geqq 0; \notag \\
& d \geqq 0: \qquad {\mathcal D}^{\rm PT} ={\mathcal D} + cd \geqq 0.
\label{DPT>0}
\end{align}
Therefore, the EPR-like necessary condition of separability\ (\ref{G1>0}),
valid for any TMGS, implies the Peres-Simon PPT condition\ (\ref{PS}).
On the other hand, the absolute minimum\ (\ref{G_m2}) given by Theorem 3,
shows that Simon's separability criterion for a TMGS, Eq.\ (\ref{PS}), entails
that the identity\ (\ref{G1>0}) is itself a sufficient condition of separability,
as implicitly expressed by Eq.\ (\ref{Gsep}).
\section{Conditions of separability: EPR-like versus PPT}
We find it appropriate to develop the line of reasoning put forward
by Duan {\em et al.} in their important work, Ref. \cite{Duan}.
We adopt their definitions and notations and employ Eq.\ (\ref{Duan})
to introduce the non-local operator
\begin{align}
& {\hat P}(\alpha):=
\begin{cases}
{\hat P}_{+}(\alpha), \qquad (d<0), \\
{\hat P}_{-}(\alpha), \qquad (d \geqq 0). \\
\end{cases}
\label{hatP}
\end{align}
Its variance,
\begin{align}
\left[ \Delta P(\alpha) \right]^2 =\frac{b_1}{u_1} {\alpha}^2
+\frac{b_2}{u_2} \frac{1}{ {\alpha}^2 }-\frac{2|d|}{ \sqrt{u_1\, u_2} },
\label{varP4a}
\end{align}
is the minimum over the pair of variances\ (\ref{varP4}) with respect to the sign of $d$:
\begin{align}
\Delta P(\alpha)=\min_{ \{ {\rm sgn}(d) \} }\{ \Delta P_{+}(\alpha), \, \Delta P_{-}(\alpha) \}.
\label{varPmin}
\end{align}
Recall the definition of the signum function of a real variable:
\begin{align}
& {\rm sgn}(x):=
\begin{cases}
-1, \qquad (x<0), \\
\;\;\; 0, \qquad (x=0), \\
\;\;\; 1, \qquad (x>0). \notag
\end{cases}
\end{align}
Guided by the necessary conditions of separability\ (\ref{sumsD}) derived
by Duan {\em et al.} in Ref. \cite{Duan}, we employ the following
EPR-like correlation function, which is a regularized sum:
\begin{align}
& K\left( {\alpha}^2,\, u_1,\, u_2 \right) := \left[ \Delta Q(\alpha) \right]^2
+\left[ \Delta P(\alpha) \right]^2 \notag \\
& -\left( {\alpha}^2 +\frac{1}{{\alpha}^2} \right).
\label{K}
\end{align}
Here the variances $\left[ \Delta Q(\alpha) \right]^2 $
and $\left[ \Delta P(\alpha) \right]^2 $ have the expressions\ (\ref{varQ4})
and\ (\ref{varP4a}), respectively. We apply the same pattern as in Sec. V,
starting from the necessary condition of separability\ (\ref{sumsD}) which reads:
\begin{align}
K\left( {\alpha}^2,\, u_1,\, u_2 \right) \geqq 0.
\label{K>0}
\end{align}
Substitution of the variances\ (\ref{varQ4}) and\ (\ref{varP4a}) into Eq.\ (\ref{K})
gives the explicit formula:
\begin{align}
& K\left( {\alpha}^2,\, u_1,\, u_2 \right)
={\alpha}^2 \left[ b_1 \left( u_1+ \frac{1}{u_1} \right) -1 \right] \notag \\
& +\frac{1}{ {\alpha}^2 } \left[ b_2 \left( u_2+ \frac{1}{u_2} \right) -1 \right] \notag \\
& -2\left( c\sqrt{u_1\, u_2} +\frac{|d|}{ \sqrt{u_1\, u_2} }\right).
\label{K1}
\end{align}
Our approach is to minimize the function\ (\ref{K1}). This reaches its minimum
with respect to the variable ${\alpha}^2$ for the value
\begin{align}
{\tilde \alpha}_m^2 =\sqrt{ \frac{ b_2 \left( u_2 +\frac{1}{u_2} \right) -1}
{ b_1 \left( u_1 +\frac{1}{u_1} \right) -1} }\, .
\label{alpha_m}
\end{align}
The obtained minimum is a function of the scaling factors:
\begin{align}
& f\left( u_1,\, u_2 \right) :=K\left( {\tilde \alpha}_m^2,\, u_1,\, u_2 \right) \notag \\
& =2\left\{ \left[ b_1 \left( u_1+ \frac{1}{u_1} \right) -1 \right]
\left[ b_2 \left( u_2+ \frac{1}{u_2} \right) -1 \right] \right\}^{ \frac{1}{2} } \notag \\
& -2\left( c\sqrt{u_1\, u_2} +\frac{|d|}{ \sqrt{u_1\, u_2} }\right).
\label{f(uu)}
\end{align}
The stationarity conditions for the function\ (\ref{f(uu)}) reduce to the following
system of equations in the unknowns $u_1$ and $u_2$:
\begin{align}
\frac{ \frac{b_1}{u_1}- \frac{1}{2} }{ b_1 u_1 -\frac{1}{2} }
=\frac{ \frac{b_2}{u_2}- \frac{1}{2} }{ b_2 u_2 -\frac{1}{2} },
\label{eq1}
\end{align}
\begin{align}
& \sqrt{\left( b_1 u_1 -\frac{1}{2} \right) \left( b_2 u_2 -\frac{1}{2} \right) }
-\sqrt{\left( \frac{b_1}{u_1}-\frac{1}{2} \right) \left( \frac{b_2}{u_2}-\frac{1}{2} \right) }
\notag \\
& = c \sqrt{ u_1 u_2} -\frac{ |d| }{ \sqrt{ u_1 u_2} }.
\label{eq2}
\end{align}
Making use of Eq.\ (\ref{eq1}), it is convenient to replace Eq.\ (\ref{eq2})
by a polynomial one:
\begin{align}
b_1 b_2 \left( u_1^2 -1 \right) \left( u_2^2 -1 \right) =\left( c u_1 u_2 -|d| \right)^2 .
\label{eq3}
\end{align}
Equations\ (\ref{eq1}) and\ (\ref{eq2}) coincide with those written by Duan {\em et al.}
in Ref. \cite{Duan} in order to define what they have called the standard form II
of the CM of a TMGS.
We state here a theorem that sharpens a result of Ref. \cite{Duan} regarding
the existence of a solution of this system in the general case.
{\bf Theorem 4.} {\em For any TMGS there exists at least a solution
$\left\{ {\tilde u}_1,\, {\tilde u}_2 \right\}$ of the algebraic system\ (\ref{eq1})
and\ (\ref{eq3}) in the classicality range of the local squeeze factors:
$${\tilde u}_1 \in [1, 2b_1], \quad {\tilde u}_2 \in [1, 2b_2]. $$}
The proof of this important theorem is given in Appendix C.
However, when trying to solve analytically the system under discussion
in the general case, one faces a non-trivial algebraic equation of degree eight.
The standard form II of the CM is important because the condition of classicality
for the corresponding TMGS,
\begin{equation}
{\mathcal V}\left( {\tilde u}_1, {\tilde u}_2 \right)-\frac{1}{2}I_4 \geqslant 0,
\label{class}
\end{equation}
is equivalent to the requirement
\begin{equation}
{\tilde f}:=f\left( {\tilde u}_1, {\tilde u}_2 \right) \geqq 0.
\label{f>0}
\end{equation}
Indeed, taking account of Eqs.\ (\ref{eq1}) and\ (\ref{eq3}), we get the twin formulae:
\begin{align}
& {\tilde f}=4\left[ \sqrt{\left( b_1{\tilde u}_1 -\frac{1}{2} \right)
\left( b_2{\tilde u}_2 -\frac{1}{2} \right) } -c\sqrt{ {\tilde u}_1{\tilde u}_2 } \right], \notag \\
& {\tilde f}=4\left[ \sqrt{ \left( \frac{b_1}{ {\tilde u}_1} -\frac{1}{2} \right)
\left( \frac{b_2}{ {\tilde u}_2 } -\frac{1}{2} \right) }
-\frac{ |d| }{ \sqrt{ {\tilde u}_1{\tilde u}_2 } } \right].
\label{twin}
\end{align}
We find it suitable to introduce the parallel notations:
\begin{align}
& {\tilde f}^{\prime}:=4\left[ \sqrt{\left( b_1{\tilde u}_1 -\frac{1}{2} \right)
\left( b_2{\tilde u}_2 -\frac{1}{2} \right) } +c\sqrt{ {\tilde u}_1{\tilde u}_2 } \right] >0, \notag \\
& {\tilde f}^{\prime \prime}:=4\left[ \sqrt{ \left( \frac{b_1}{ {\tilde u}_1} -\frac{1}{2} \right)
\left( \frac{b_2}{ {\tilde u}_2 } -\frac{1}{2} \right) }
+\frac{ |d| }{ \sqrt{ {\tilde u}_1{\tilde u}_2 } } \right] \geqq 0.
\label{twin1}
\end{align}
The matrix condition\ (\ref{class}) reduces to four inequalities:
\begin{align}
& b_1{\tilde u}_1 -\frac{1}{2} \geqq 0, \notag \\
& \left( b_1{\tilde u}_1 -\frac{1}{2} \right) \left( \frac{b_1}{ {\tilde u}_1 } -\frac{1}{2} \right)
\geqq 0, \notag \\
& \left( \frac{b_1}{ {\tilde u}_1 }
-\frac{1}{2} \right) 4^{-2}({\tilde f} {\tilde f}^{\prime}) \geqq 0, \notag \\
& \det\left[ {\mathcal V}\left( {\tilde u}_1, {\tilde u}_2 \right)-\frac{1}{2}I_4 \right]
=4^{-4} ({\tilde f} {\tilde f}^{\prime} )({\tilde f} {\tilde f}^{\prime \prime} ) \geqq 0.
\label{ineq}
\end{align}
Three of them are already satisfied, so that the only condition to be fulfilled
remains ${\tilde f} \geqq 0$, Eq.\ (\ref{f>0}). This classicality requirement
is a sufficient condition for the separability of the given TMGS ${\hat \rho}$,
whose CM is congruent with its standard form II
${\mathcal V}\left( {\tilde u}_1, {\tilde u}_2 \right) $ via a local symplectic
transformation\ (\ref{local}). Moreover, by virtue of Eq.\ (\ref{K>0}),
the inequality\ (\ref{f>0}) is also a necessary condition of separability.
Therefore, a TMGS whose CM has the standard form II is a unique state for which
separability reduces to classicality. By examining the sign
of the EPR-like correlation function
${\tilde f}:=f\left( {\tilde u}_1, {\tilde u}_2 \right)$, one can check whether
the standard-form-II TMGS is classical or not, i.e., whether it possesses or not
a well-behaved Glauber-Sudarshan $P$ representation.
To sum up, our optimization method has exploited the EPR-like correlation
function\ (\ref{K}) leading to the separability indicator ${\tilde f}$, Eq.\ (\ref{twin}).
It is worth stressing that this indicator differs in two respects
from the previous ones which are specified by Eqs.\ (\ref{Emin}), \ (\ref{Fmin}),
and\ (\ref{G_m2}). First, it has been identified without making any reference
to the partial transposition of the density matrix describing a TMGS ${\hat \rho}$.
Second, for the time being, one does not know an explicit solution
$\left\{ {\tilde u}_1,\, {\tilde u}_2 \right\}$ of Eqs.\ (\ref{eq1}) and\ (\ref{eq3}),
except for some special classes of TMGSs, such as
the thermal states (TSs), the mode-mixed thermal
states (MTSs), and the STSs \cite{PT2015,PT2016},
the symmetric ones \cite{PT2015}, as well as the states subject
to the constraint ${\tilde f}=0$. By contrast to the three indicators quoted above,
we have not established yet in generality that the existing stationary point
is unique or that it corresponds to a minimum.
In principle, the EPR-like separability criterion\ (\ref{f>0}) is as important
as Simon's condition of separability\ (\ref{DPT1}) that relies on partial transposition.
Nevertheless, it suffers from the drawback that, in the general case, it cannot be
handled analytically. This makes Simon's PPT criterion of separability\ (\ref{DPT1})
to prevail in practice. However, the next theorem explicitly connects
the EPR-like inequality\ (\ref{f>0}) to the PPT one\ (\ref{DPT1}).
{\bf Theorem 5.} {\em The EPR-like separability condition\ (\ref{f>0})
and Simon's PPT one\ (\ref{DPT1}) are fully and manifestly equivalent.}
{\em Proof}. We introduce the following function, which is symmetric
in the mode indices 1 and 2:
\begin{align}
& Z({u}_1, {u}_2):=\frac{1}{2}\left\{ \left[ \left( b_1{u}_1 -\frac{1}{2} \right)
\left( b_2{u}_2 -\frac{1}{2} \right) - c^2 { u}_1 u_2 \right] \right. \notag \\
& \left. \left[\left( \frac{b_1}{ u_1} +\frac{1}{2} \right) \left(\frac{ b_2}{u_2}
+\frac{1}{2} \right) - \frac{d^2 }{ u_1 u_2 } \right] +\left[ \left( b_1 u_1 +\frac{1}{2} \right)
\left( b_2{u}_2 +\frac{1}{2} \right)\right. \right. \notag \\ &\left.\left.- c^2 { u}_1 u_2 \right]
\times \left [\left( \frac{b_1}{ u_1} -\frac{1}{2} \right) \left(\frac{ b_2}{u_2}
-\frac{1}{2} \right) - \frac{d^2 }{ u_1 u_2 } \right] \right\}.
\label{Z}
\end{align}
The expression on the r. h. s. of Eq.\ (\ref{Z}) can be cast into a simpler form:
\begin{align}
Z({u}_1, {u}_2)=H(d){\mathcal D}+H(-d){\mathcal D}^{\rm PT} +\frac{1}{ 4u_1 u_2 } \Phi(u_1, u_2),
\label{Z1}
\end{align}
where $H(x):=\frac{1}{2}[1+{\rm sgn}(x)]$ denotes the Heaviside step function
and $\Phi(u_1, u_2)$ is the function\ (\ref{Phi}). In view
of Eqs.\ (\ref{GS3}) and\ (\ref{DPT1}) we have
\begin{align}
H(d){\mathcal D}+H(-d){\mathcal D}^{\rm PT}=\left( b_1 b_2-c^2 \right) \left( b_1 b_2-d^2 \right)
\nonumber\\
-\frac{1}{4} \left( b_1^2 +b_2^2+2c|d| \right) +\frac{1}{16}.
\label{DD}
\end{align}
Let us equate the expressions\ (\ref{Z})
and\ (\ref{Z1}) of the value $Z({\tilde u}_1, {\tilde u}_2)$ corresponding to a standard form II
${\mathcal V}\left( {\tilde u}_1, {\tilde u}_2 \right)$ of the CM. Taking account
of Eqs.\ (\ref{twin}),\ (\ref{twin1}), and\ (\ref{phi3}), we get the identity:
\begin{align}
& H(d){\mathcal D}+H(-d){\mathcal D}^{\rm PT}=\frac{1}{32}{\tilde f} \left\{ {\tilde f}^{\prime}
\left[ \left( \frac{b_1}{\tilde u_1} +\frac{1}{2} \right)
\left( \frac{ b_2}{\tilde u}_2 +\frac{1}{2} \right) \right. \right. \notag \\
& \left.\left.
- \frac{d^2 }{ {\tilde u}_1{\tilde u}_2 } \right] +{\tilde f}^{\prime\prime} \left[\left( b_1 \tilde u_1 +\frac{1}{2} \right)
\left( b_2{\tilde u}_2 +\frac{1}{2} \right) - c^2 {\tilde u}_1{\tilde u}_2 \right] \right\}.
\label{id2}
\end{align}
In Eq.\ (\ref{id2}), the expression in curly brackets is strictly positive and
${\mathcal D} \geqq 0$ for any TMGS ${\hat \rho}$. Accordingly, the separability indicator
${\tilde f}$ and the local symplectic invariant on the l. h. s. do have the same sign.
There are two cases:
\begin{itemize}
\item $d \geqq 0. \quad {\rm Then} \quad {\tilde f}\geqq 0 \;\; \text{and} \;\;
{\mathcal D}^{\rm PT}\geqq 0:
\quad {\hat \rho}$ \;\; separable;
\item $d < 0.$
\begin{align}
& {\rm Either} \;\; {\tilde f} \geqq 0\; \Longleftrightarrow \;
{\mathcal D}^{\rm PT} \geqq 0: \quad{\hat \rho} \; \; separable, \notag \\
& {\rm or} \;\; {\tilde f}<0 \; \Longleftrightarrow \; {\mathcal D}^{\rm PT} < 0:
\quad {\hat \rho}\; \;\; entangled. \notag
\end{align}
\end{itemize}
As the signs of the separability indicators $\tilde f$ and ${\mathcal D}^{\rm PT}$
coincide, we conclude they are equivalent in detecting separability of TMGSs:
\begin{align}
& {\tilde f} \geqq 0 \;\; \Longleftrightarrow \;\; {\mathcal D}^{\rm PT} \geqq 0:
\quad {\hat \rho}\; \;\; separable; \notag \\
& {\tilde f} <0 \;\; \Longleftrightarrow \;\; {\mathcal D}^{\rm PT} < 0:
\quad {\hat \rho}\; \;\; entangled.
\label{fGm}
\end{align}
The proof is complete.
It is instructive to present concisely the above-mentioned special classes
of TMGSs whose standard-form-II scaling factors are
explicitly evaluated.
\begin{enumerate}
\item {\em TMGSs with} $c=|d| $: TSs, MTSs, and STSs.
The solution
\begin{align}
{\tilde u}_1 = {\tilde u}_2 =1
\label{DuanSTS}
\end{align}
of Eqs.\ (\ref{eq1}) and\ (\ref{eq3}) is specific for TSs $(c=d=0)$,
MTSs $(c=d > 0),$ and STSs $(c=-d > 0)$. We have proven its uniqueness. It gives the minimum
value\ (\ref{alpha_m})
$${\tilde \alpha}_m^2 =\sqrt{ \frac{ b_2 -\frac{1}{2} }{ b_1 -\frac{1}{2} } }.$$
Equations\ (\ref{twin}) and\ (\ref{twin1}) give the separability indicator
\begin{align}
& {\tilde f}=4\left[ \sqrt{\left( b_1-\frac{1}{2} \right) \left( b_2 -\frac{1}{2} \right) } -c \right],
\label{fSTS}
\end{align}
and, respectively, the function
\begin{align}
& {\tilde f}^{\prime}={\tilde f}^{\prime\prime}=4\left[ \sqrt{\left( b_1-\frac{1}{2} \right)
\left( b_2 -\frac{1}{2} \right) } +c \right] \geqq 0.
\label{twin2}
\end{align}
Equation\ (\ref{fSTS}) becomes insightful when written in terms of the symplectic
eigenvalues ${\kappa}_{\pm}$ for an MTS and ${\kappa}_{\pm}^{\rm PT}$ for an STS:
\begin{align}
{\tilde f}={\tilde f}_{\rm MT}=\frac{4^2}{ {\tilde f}^{\prime} } \left( {\kappa}_{-}
-\frac{1}{2}\right) \left( {\kappa}_{+}-\frac{1}{2}\right) \geqq 0, \qquad (d>0),
\label{fMT1}
\end{align}
\begin{align}
{\tilde f}= {\tilde f}_{\rm ST}=\frac{4^2}{ {\tilde f}^{\prime} }\left( {\kappa}_{-}^{\rm PT}
-\frac{1}{2}\right) \left( {\kappa}_{+}^{\rm PT}-\frac{1}{2} \right), \qquad (d<0).
\label{fST1}
\end{align}
\item {\em Symmetric TMGSs}: $\, b_1=b_2=:b. \\$
The solution
\begin{align}
{\tilde u}_1 ={\tilde u}_2 =\sqrt{ \frac{b-|d|}{b-c} }
\label{Duansym}
\end{align}
is specific for symmetric TMGSs, provided that $c > |d|.$ The minimum
value\ (\ref{alpha_m}) of the parameter $\alpha$ is ${\tilde \alpha}_m =1$.
We have proven the uniqueness of the solution\ (\ref{Duansym}),
as well as its property of being a minimum point of the function\ (\ref{f(uu)})
written with $\, b_1=b_2=:b.$ The corresponding absolute minimum
is the EPR-like separability indicator
\begin{align}
{\tilde f}=4\left[ \sqrt{(b-c)(b-|d|)}-\frac{1}{2} \right].
\label{fsym}
\end{align}
According to Eqs.\ (\ref{kappa+-}) and\ (\ref{A5}),
\begin{align}
{\kappa}_{-}= \sqrt{(b-c)(b-d)}, \qquad {\kappa}_{-}^{\rm PT}= \sqrt{(b-c)(b+d)},
\label{kappa-}
\end{align}
so that the following relations hold:
\begin{align}
{\tilde f}=4\left( {\kappa}_{-} -\frac{1}{2} \right) \geqq 0, \qquad (d \geqq 0),
\label{fsym1}
\end{align}
\begin{align}
{\tilde f}=4\left( {\kappa}_{-}^{\rm PT} -\frac{1}{2} \right), \qquad (d<0).
\label{fsym2}
\end{align}
\item {\em TMGSs at the boundary ${\tilde f}=0.$}
According to Eq.\ (\ref{id2}),
\begin{align}
{\tilde f}= 0 \;\; \Longleftrightarrow \;\; H(d){\mathcal D}+H(-d){\mathcal D}^{\rm PT}=0.
\label{fDD}
\end{align}
The property ${\tilde f}= 0$ is therefore specific to all TMGSs with $d \geqq 0$
at the physicality edge $({\mathcal D}=0)$, as well as to all those with $d<0$
at the separability threshold $\left( {\mathcal D}^{\rm PT}=0 \right).$ For both limit
situations, we find a unique solution of the algebraic system\ (\ref{eq1}) and\ (\ref{eq3}):
\begin{align}
{\tilde u}_1=2\frac{ c\left( b_{1}b_{2}-d^2 \right) +\frac{1}{4}|d| }{b_{1}|d|+b_{2}c}, \;\;
{\tilde u}_2 =2\frac{ c\left( b_{1}b_{2}-d^2 \right) +\frac{1}{4}|d| }{b_{1}c+b_{2}|d|}.
\label{f=0}
\end{align}
\end{enumerate}
Let us focus on the expressions\ (\ref{fMT1}),\ (\ref{fST1}),
and\ (\ref{fsym1})-\ (\ref{fsym2}) of the EPR-like correlation function ${\tilde f}$
for MTSs, STSs, and symmetric TMGSs, respectively. These important special GSs
illustrate both cases: $d \geqq 0$ and $d<0$. The above explicit formulae clearly
display the equivalence between the EPR-like separability condition\ (\ref{f>0})
and the PPT one, Eq.\ (\ref{DPT1}). Such explicit expressions are not available
for an arbitrary TMGS. Anyway, in order to decide whether a given TMGS is separable
or not, then, by virtue of Eq.\ (\ref{fGm}), one is entitled to employ Simon's condition
of separability, ${\mathcal D}^{\rm PT} \geqq 0,$ instead of the less efficient formula
${\tilde f} \geqq 0$.
\section{Summary and conclusions}
This work is devoted to an explicit application of EPR-like correlations
in detecting Gaussian entanglement. Let us emphasize its most significant
achievements. First, we have tackled three correlation functions built with
variances of two EPR-like observables in a TMGS. They are a normalized product,
a normalized sum, and a non-normalized, but regularized sum. Our main results
are Theorems 1, 2, and 3, which express their absolute minima in terms of either
the smaller symplectic eigenvalue ${\kappa}_{-}^{\rm PT}$ or the Simon symplectic
invariant ${\mathcal D}^{\rm PT}$. These lower bounds are explicitly written
in Eqs.\ (\ref{Emin}), \ (\ref{Fmin}), and\ (\ref{G_m2}). All the three minima are
EPR-like separability indicators due to Simon's PPT criterion. We have exploited
these analytic results to point out that three distinct EPR-like necessary conditions of
separability for a TMGS imply the corresponding Peres-Simon PPT condition,
Eq.\ (\ref{PS}).
Second, the present development provides a fresh look at the work of Duan {\em et al.},
which is based on EPR-like observables \cite{Duan}. Among the original ideas put forward
in Ref. \cite{Duan}, the central one is the existence of a standard form II of the CM, which
is confirmed here. This CM describes a privileged TMGS whose classicality is equivalent
to the separability of the whole set of TMGSs connected to it by local
unitary transformations.
Thus, the separability properties of the whole class of TMGSs having a given set of
standard-form parameters are assigned to this standard-form-II state. The resulting
separability criterion is quite special because it is independent of the PPT condition.
However, in spite of its soundness, the original EPR-like approach \cite{Duan}
cannot decide if a TMGS is separable or not, except for the special cases discussed
in Sec. VI. This happens because it does not provide a general analytic solution.
In the present work, the EPR-like correlation function introduced in Ref. \cite{Duan}
is used in a regularized sum form. Its optimal value over the variables $\alpha, u_1,u_2,$
denoted ${\tilde f}$, turns out to be a separability indicator for TMGSs. This is obtained
as a marker of classicality for the standard-form-II TMGS.
Third, we have explicitly proven that the EPR-like indicator of separability ${\tilde f}$
is equivalent to Simon's PPT separability marker ${\mathcal D}^{\rm PT}$.
We stress that the resulting formula is the first direct connection between two distinct
approaches to the separability problem for TMGSs. It might stimulate
once more a reflection on the central role
of the uncertainty relations in quantum mechanics. Needless to say,
the importance of the EPR-like approach is enhanced by our simple proof of its manifest
consistency with Simon's PPT separability condition, whose practical usefulness
is universally acknowledged.
|
\section{Introduction}
Since the discovery of superconductivity in fluorine doped LaFeAsO \cite{Kamihara:Jacs}, the iron pnictides and chalcogenides have rapidly developed into one of the main branches of research in the field of unconventional superconductors \cite{Lumsden:Jpcm,Dai:Np,Stewart:Rmp}. Almost all previously reported iron-based superconductors have similar crystal structures, involving a slightly distorted two-dimensional Fe square lattice \cite{Johnston:Ap,Dagotto:Rmp,Dai:Rmp,Si:NRM}, where each Fe atom is caged in a tetrahedral structure.
The magnetism of the non-superconducting parent state is primarily given by the collinear stripe-like order, namely the C-type antiferromagnetic (C-AFM) order, although some exceptions exist such as in FeTe, FeSe, KFe$_2$Se$_2$,
and K$_2$Fe$_4$Se$_5$ \cite{Bishop:prl,Li:Prb09,Moon:Prl,Ma:Prl,rong:prl15,Li:Np,Lw:Prb,Zhang:rrl,Bao:Cpl,Ye:Prl,Yin:Prl}.
The structural and magnetic similarities of all these iron-based superconductors suggest common physical mechanisms leading to their magnetic and superconducting properties, despite the different ratios of the atomic elements involved.
However, recently iron chalcogenides with a different structure, the so-called 123-type $A$Fe$_2X_3$ ($A$=K, Cs, Rb, or Ba; and $X$=S, Se, or Te), have drawn considerable attention \cite{Maziopa:Jpcm,Saparov:Prb,Caron:Prb,Caron:Prb12,Lei:Prb,Arita:prb,Wang:prb16,Patel:prb}. These materials display unique quasi-one-dimensional two-leg ladder iron structures (see Fig.~\ref{Fig1}(a)) that are clearly qualitatively distinct from the other extensively studied iron pnictides/chalcogenides with iron layers. At least from the perspective of the electronic structure, the frequently mentioned Fermi surface nesting effect involving two pocket cylindrical Fermi surfaces (corresponding to the quasi-two-dimensional structure) in several iron pnictides/chalcogenides can not be relevant in these ladder systems \cite{Lumsden:Jpcm,Stewart:Rmp,Johnston:Ap,Dagotto:Rmp,Dai:Rmp}. Thus, the reduced dimensionality of the iron network makes $A$Fe$_2X_3$ a physically fascinating material that deserves further experimental and theoretical scrutiny. In fact, the two-leg ladders made of Fe atoms remind us of the previously studied two-leg ladders superconducting cuprates \cite{Dagotto:Rmp94,Dagotto:Rpp}. In the past, the study of cuprate ladders much illuminated the physics of Cu oxides, primarily because theoretical calculations involving model Hamiltonian can be carried out with good accuracy in chains and ladders and, thus, accurate theory-experiment comparisons can be done. A similar important impact of iron ladders on the field of iron superconductors is now to be expected.
In this family of ladder materials, BaFe$_2$Se$_3$ was the first compound reported to be superconducting at approximately $11$ K \cite{Maziopa:Jpcm}. However, others experiments claimed the material to be a semiconductor with the observed superconductivity probably induced by impurities \cite{Caron:Prb}. Even under these circumstances, one of our previous studies predicted an interesting property, namely that its block-type antiferromagnetic (Block-AFM) state is multiferroic due to magnetostriction effects \cite{Dong:PRL14,Dong:Ap}. The theoretically predicted polar structure has indeed been verified by a subsequent neutron study \cite{Lovesey:PS}. In addition, doping of K in the Ba site changes the ground state to the so-called CX-type antiferromagnetism (CX-AFM) (see Fig. 1(b)) \cite{Caron:Prb12}. Such CX-AFM order was also predicted to be the ground state of BaFe$_2$S$_3$, which was later also confirmed using neutron techniques \cite{Takahashi:Nm}. Under ambient conditions, BaFe$_2$S$_3$ has a semiconducting ground state with a very small gap about $0.06-0.07$ eV \cite{Gonen:Cm,Reiff:Jssc}. The most striking recent experimental discovery is that high pressure can drive BaFe$_2$S$_3$ to become superconducting when the hydrostatic pressure exceeds about $11$ GPa \cite{Takahashi:Nm} and its highest $T_{\rm c}$ can reach $24$ K at $13.5$ GPa \cite{Yamauchi:prl15}.
A subsequent density functional theory (DFT) study by Suzuki \textit{et al.} on BaFe$_2$S$_3$ confirmed
the CX-AFM nature of the magnetic state (see also our earlier work \cite{Dong:PRL14}) as well as its semiconducting
behavior under ambient conditions \cite{Suzuki:prb}. However, the pressure induced magnetic to nonmagnetic transition
(as phenomenologicall required by superconductivity) was not obvious from that DFT study, although the calculation indeed
showed a semiconductor-metal transition at about $5$ GPa. In those previous DFT calculations, the magnetism persisted with
only a small suppression upon increasing pressure. This problem might be due to the process followed to allow the relaxation
of the crystal structure under pressure, since for simplicity only the positions of sulfur atoms were optimized in the $x$-$y$ plane
in that previous study, which is sufficient for a qualitative analysis. However, in our study presented below all atomic positions were
simultaneously relaxed, allowing us to be not only qualitative but also quantitative in our analysis.
In the present publication, the magnetic properties, electronic structure, and pressure effects corresponding to BaFe$_2$S$_3$
are revisited using first-principles DFT calculations. Our results can be divided in two classes. First, similar results
as in previous DFT efforts and experiments for the ambient conditions have been obtained and confirmed. Second,
a semiconductor-metal transition accompanying the quenching of magnetism has been observed, in good agreement with experimental
observations. This phase transition is probably of first order because we observe a sudden drop of the iron's magnetic moment
and also anomalies in the crystal structure. More specifically, one of our most important results is that the underlying physics
of this transition lies in the modifications by pressure of the effective electronic density at the iron network of relevance.
Thus, the net effect of increasing pressure is equivalent to doping charge into the iron atoms, a nontrivial effect difficult to deduce from
the existing experimental information. Following these arguments, two additional materials $A$Fe$_2X$$_3$ (KFe$_2$S$_3$
and KFe$_2$Se$_3$) have been also studied theoretically by a similar procedure. The quenching of magnetic moments and metallic
phases have also been obtained both for the cases of KFe$_2$S$_3$ and KFe$_2$Se$_3$, although at higher pressures.
The theoretical observation of these pressure induced transitions suggest possible pressure-induced superconducting
states also in KFe$_2$S$_3$ and KFe$_2$Se$_3$ by analogy with what occurs in BaFe$_2$S$_3$, although we cannot explicitly
prove these predictions due to the limitations of DFT techniques to address superconductivity. We expect that our results
should motivate experimental efforts for their confirmation.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig1.eps}
\caption{(a) Schematic crystal structure of $A$Fe$_2$$X$$_3$. Purple: $A$, e.g. Ba or K; Yellow: $X$, e.g. S or Se; Brown: Fe.
(b) Sketch of the possible spin configurations that could be stabilized in two-leg iron ladders. Spin up and spin down are distinguished
by colors. (c) Schematic of the three dimensional magnetic ($\pi$, $\pi$, $0$) and ($0$, $0$, $0$) configurations between ladders.}
\label{Fig1}
\end{figure}
\section{Methods}
The DFT calculations are performed based on the generalized gradient approximation (GGA) with the
projector augmented wave (PAW) potentials, as implemented in the Vienna {\it ab initio} Simulation Package
(VASP) \cite{Kresse:Prb,Kresse:Prb96,Blochl:Prb,Perdew:Prl}. The PBE exchange function has been adopted. The plane-wave cutoff is $500$ eV.
The $k$-point mesh is $4\times$3$\times6$ for the minimum crystalline unit cell, which is accordingly adapted for the various magnetic cells
studied (e.g. $2\times$6$\times2$ for CX-AFM ($\pi$, $\pi$, $0$)). Starting from experimental values, both the lattice constants and atomic
positions are fully relaxed until the force on each atom was below $0.01$ eV/{\AA}.
To study the magnetic properties, various possible (in-ladder) magnetic structures are imposed on the Fe ladders,
as shown in Fig.~\ref{Fig1}(b). Despite the in-ladder correlation, the magnetic correlation between ladders can
also slightly affect the physical properties. Therefore, besides the simplest ($0$, $0$, $0$) order,
the ($\pi$, $\pi$, $0$) order is also studied, as indicated in Fig.~\ref{Fig1}(c).
In addition to the standard DFT calculation, the maximally-localized Wannier functions (MLWFs) have also been employed to fit five Fe's $3d$ bands and the Fermi surface, using the WANNIER90 packages \cite{Mostofi:cpc}.
Several previous DFT studies have found that even the pure GGA (or LSDA) procedures overestimates the local magnetic moments
in iron pnictides and chalcogenides \cite{Hansmann:prl10,Mazin:prb,Mazin:np,Dong:Prl,Zhang:rrl}. Thus, using GGA+$U$ with
a positive $U$ will render this inconsistency even more serious. Thus, in some studies, a negative $U$ correction was adopted
to better describe the 122-type Fe-based materials \cite{Ferber:prb,Yi:prb}. Alternatively, the exchange and correlation
kernels were rescaled by an appropriate factor \cite{Ortenzi:prl}. In the present study, both GGA+$U$ and pure GGA have been tested,
and we found that the later provides a better description of $A$Fe$_2$X$_3$ (regarding its crystalline constants, magnetic moments,
as well as band gaps). For this reason, only the pure GGA results will be presented in the rest of this publication.
\section{Results \& discussion}
\subsection{Magnetism \& electronic structure of BaFe$_2$S$_3$}
Our main DFT results with regards to the magnetic and electronic properties of the BaFe$_2$S$_3$ ladder can be summarized in four statements as described below:
(1) Without the external pressure, both the lattice constants and atomic positions were fully optimized in the
presence of magnetism. In the pure GGA calculation, the CX-AFM ($\pi$, $\pi$, $0$)
state has the lowest energy among all candidate configurations investigated, in agreement with experiments.
The CX-AFM ($0$, $0$, $0$) is only slightly higher in energy by $6.3$ meV/Fe, which is
reasonable considering the similarity between these two CX-AFM configurations.
(2) Even the pure GGA result gives a magnetic moment ($2.08$ $\mu_{\rm B}$/Fe) that is slightly higher than
the experimental one ($\sim1.2$ $\mu_{\rm B}$/Fe at low temperature, as measured by neutron scattering \cite{Takahashi:Nm}).
This overestimated local moment is quite common in DFT calculations of iron-based superconductors \cite{Hansmann:prl10,Mazin:prb,Mazin:np},
which may be due to the coexistence of localized Fe spins and itinerant electrons \cite{Tam:prl}. Note that this calculated
value also depends on how large the Wigner-Seitz radius of the iron atom is set to be. Thus, the inconsistency described above
may partially originate from the methodological difference between the VASP procedure and neutron experiments.
(3) For the experimentally relevant CX-AFM ($\pi$, $\pi$, $0$) state, the pure GGA calculation gives a small gap of $0.088$ eV, which agrees
with the experimental value ($0.06-0.07$ meV) \cite{Gonen:Cm,Reiff:Jssc} and previous DFT results \cite{Suzuki:prb}.
(4) According to the calculated density of states (DOS) (not shown), the bands near the Fermi levels are
highly hybridized between Fe's $3d$ orbitals and S's $3p$ orbitals. Such hybridization is quite common in iron pnictides/chalcogenides.
\subsection{Pressure induced transition in BaFe$_2$S$_3$}
Although previous DFT calculations correctly reproduced the magnetic ground state and electronic structure of BaFe$_2$S$_3$,
the pressure induced magnetic-nonmagnetic transition (presumably also associated to superconductivity, although beyond the DFT scope)
was not reproduced \cite{Suzuki:prb}, as explained before. Below, these pressure effects will be revisited using DFT.
By increasing the hydrostatic pressure in the calculation, our energies and Fe's magnetic moments are summarized
in Fig.~\ref{Fig3}(a-b). For all magnetic orders, the local moments decrease with pressure but the CX-AFM ($\pi$, $\pi$, $0$)
state always has the lowest energy. When the pressure reaches $10.8$ GPa (very close to the experimental critical
pressure \cite{Takahashi:Nm,Yamauchi:prl15}), the {\it static values} of the local moments drop to zero. Meanwhile,
the system becomes metallic. Such nonmagnetic metallic phase provides the conditions for superconductivity if there
are still antiferromagnetic fluctuations present (although short-range magnetic quantum fluctuations also can not be
captured in the DFT calculation). In this sense, without fine tuning parameters, our calculation correctly reproduces
the antiferromagnetic-to-nonmagnetic transition. Furthermore, we found that the quenching of magnetic moments occurs
as in a first-order transition, with a sudden drop from $0.83$ $\mu_{\rm B}$/Fe to zero at $10.8$ GPa (inset of Fig.~\ref{Fig3}(b)).
Figure~\ref{Fig3}(c) shows the compressibility, i.e. the lattice constants normalized to the ambient ones.
The $b$-axis lattice is the softest while the $c$-axis lattice is the hardest, as in experiments \cite{Takahashi:Nm}.
Such anisotropic compressibility can be intuitively understood considering the loose space between ladders and the
compact bonds along ladders. The first-order character of the transition at $10.8$ GPa can also be observed in the lattice
structure (especially for the lattice constant along the $b$-axis), as emphasized in the inset of Fig.~\ref{Fig3}(c).
This first-order characteristic was not reported previously and could be verified in future experiments.
\begin{figure}
\centering
\includegraphics[width=0.44\textwidth]{Fig2.eps}
\caption{DFT results corresponding to BaFe$_2$S$_3$, as a function of pressure. (a) Energy difference (per Fe) of various magnetic orders with respect to the ferromagnetic (FM) state. (b) Local magnetic moments of Fe. Inset: a magnified view near the transition. (c) Lattice constants normalized to the original ones for the CX-AFM ($\pi$, $\pi$, $0$) state. Inset: a magnified view near the transition. (d) Bader charge analysis.}
\label{Fig2}
\end{figure}
According to the Bader charge analysis (Fig.~\ref{Fig2}(d))~\cite{Bader:book,Tang:Jpcm,Henkelman:Cms},
there is a significant charge transfer ($\sim0.15$ electron) from S to Fe by increasing pressure from
$0$ GPa to $12$ GPa. This tendency is equivalent to the effects of electron doping by, for example, chemical substitution,
which is the standard procedure to generate superconductivity from a magnetic parent compound. In this sense,
it is reasonable to suspect that the superconductivity observed in BaFe$_2$S$_3$ probably is induced by electron doping,
in analogy with the superconductivity triggered by F doping in LaFeAsO. This is also quite similar to what happened in
previous investigations of the Cu-ladders, where a transfer of charge from chains to ladders triggers
superconductivity~\cite{Dagotto:sci96}. In summary, our calculations suggest that the superconductivity
of BaFe$_2$S$_3$ is probably caused by self-doping of electrons into the iron network. This effect can occur
in addition to the previously proposed scenario based on the broadening of the electronic bandwidth by
pressure~\cite{Takahashi:Nm}. Only further experimental work can clarify which of the two tendencies is more dominant to generate superconductivity.
A careful analysis of the DOS at the Fermi level just before
the critical pressure finds a sharp peak, i.e. a van Hove singularity, which suddenly drops
to become a valley around $11$ GPa, as shown in Fig.~\ref{Fig3}(a).
The first-order magnetic-nonmagnetic transition leads to the sudden disappearance of this van Hove singularity
since electronic bands are seriously reconstructed at the critical pressure. The Fermi surface at $11$ GPa
is shown in Fig.~\ref{Fig3}(b), with four bands crossing the Fermi level. Two of them are nesting around
the $\Gamma$ point while the other two are isolated.
\begin{figure}
\centering
\includegraphics[width=0.44\textwidth]{Fig3a.eps}
\includegraphics[width=0.48\textwidth]{Fig3b.eps}
\caption{(a) Evolution of the DOS near the Fermi level with increasing pressure. Below the critical pressure,
a van Hove singularity appears at the Fermi level, but it suddenly disappears at the critical pressure. Only
the spin-up channel is shown because the spin-down channel is exactly identical, since the system is either
antiferromagnetic or nonmagnetic. (b) Fermi surfaces at $11$ GPa. The first four (I-IV) panels are individual
Fermi surfaces. The last two (V-VI) are the total Fermi surfaces viewed from different orientations.}
\label{Fig3}
\end{figure}
\subsection{Pressure induced transition in KFe$_2$S$_3$}
Until now, BaFe$_2$S$_3$ is the only experimentally confirmed superconductor in the 123-type series of iron ladders. As a consequence, it is interesting to investigate whether there are other 123-type iron ladders that can also potentially become superconductors. KFe$_2$S$_3$ is a sister member of BaFe$_2$S$_3$, where the nominal valence of Fe is higher by $+0.5$. Experimentally, KFe$_2$S$_3$ has been synthesized \cite{Mitchell:Jssc} but their detailed physical properties have not been reported, particularly under high pressure. Structurally, this K-based 123 ladder is similar to BaFe$_2$S$_3$, with the same $Cmcm$ group.
Here, DFT calculations have been performed also on KFe$_2$S$_3$. In this study we have found that
the magnetic ground state is also of the CX-AFM type. Moreover, the local magnetic moment is about $2.3$ $\mu_{\rm B}$/Fe
in the pure GGA calculations. Compared to BaFe$_2$S$_3$, the band gap of KFe$_2$S$_3$ ($\sim0.51$~eV in pure GGA calculations)
is slightly larger. According to the atomically-resolved DOS (not shown), the states near the Fermi level are also primarily
contributed by the Fe's $3d$ orbitals, which are also highly hybridized with the S's $3p$ orbitals.
According to the Bader charge analysis, the electronic densities at the Fe and S sites in KFe$_2$S$_3$ are lower by $0.09$ and $0.16$ electrons
than those in BaFe$_2$S$_3$, respectively. Therefore, the replacement of Ba$^{2+}$ by K$^{1+}$ does not really dope the iron sites by the nominal
amount of $0.5$ holes, but those holes mostly go to the S's sites due to the partially covalent Fe-S bonds.
Since the electronic density of Fe is slightly lower in KFe$_2$S$_3$, the critical pressure should
be higher according to the Bader charge analysis. To verify this expectation, the calculated crystal constants and
magnetism (with pure GGA) are presented in Fig.~\ref{Fig4} as a function of pressure.
As shown in Fig.~\ref{Fig4}(a), the CX-AFM ($\pi$, $\pi$, $0$) state is always the ground state if it is magnetically ordered.
The suppression of the magnetic moment by pressure is shown in Fig.~\ref{Fig4}(b). The first-order character of the transition
is similar to that observed in BaFe$_2$S$_3$. The critical pressure ($\sim23$ GPa) is indeed larger, as expected. The band gap
of the CX-AFM phase reduces to zero at $17$ GPa, inducing a semiconductor to metal transition. The metallic nonmagnetic phase
above $23$ GPa may be superconducting, according to the previous experience with BaFe$_2$S$_3$. Of course, this reasoning is merely
by analogy between similar materials because DFT cannot address superconductivity directly.
The lattice constants under pressure, normalized to their ambient values, are shown in Fig.~\ref{Fig4}(c); they are also very similar
to those reported for BaFe$_2$S$_3$. The Bader charge analysis applied to KFe$_2$S$_3$ under pressure leads to the same behavior as
in BaFe$_2$S$_3$, namely pressure enhances the local electronic density at the Fe sites, as shown in Fig.~\ref{Fig4}(d). There is also
significant charge transfer ($\sim0.24$ electron per Fe) from S to Fe by increasing pressure from $0$~GPa to $25$~GPa. Interestingly,
the Bader charge densities at the critical pressures for magnetic quenching are almost identical (error bar $\delta<0.005e$ in our
calculations) for KFe$_2$S$_3$ and BaFe$_2$S$_3$, suggesting a similar physical mechanism for both compounds.
To summarize this subsection, our calculations here predict that KFe$_2$S$_3$ should be similar to BaFe$_2$S$_3$, regarding its
magnetic ground state as well as its behavior upon pressure. Thus, superconductivity is possible under higher pressure.
By increasing pressure, the transfer of electrons from S to Fe occurs, namely a self-doping process takes place that eventually
could lead to superconductivity as in the canonical layered iron superconductors.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig4.eps}
\caption{DFT results of KFe$_2$S$_3$ as a function of pressure. (a) Energies (per Fe) of various magnetic orders. (b) Local magnetic moments of Fe. (c) Lattice constants normalized to the original ones for the CX state. (d) Bader charge analysis.}
\label{Fig4}
\end{figure}
\subsection{Pressure induced transition in KFe$_2$Se$_3$ }
Since reducing the electronic density at the Fe atoms is a disadvantage to suppress magnetism with increasing pressure,
as demonstrated in KFe$_2$S$_3$ where a higher pressure than for BaFe$_2$S$_3$ was needed to induce the metallic phase,
it is natural to expect the opposite tendency in other $A$Fe$_2X_3$ compounds that naturally have higher electronic density at the Fe atoms.
According to our Bader charge analysis, at ambient conditions the electronic density at the Fe atoms
in KFe$_2$Se$_3$ (experimentally confirmed to display the CX-AFM order~\cite{Caron:Prb12}) is higher than
in the case of KFe$_2$S$_3$ by $0.12$ electrons, and even higher than that in BaFe$_2$S$_3$ by $0.03$ electrons,
due to the weak electronegativity of Se. Then, a natural speculation arises: could it be that by increasing pressure
KFe$_2$Se$_3$ is closer to metallicity, and thus perhaps also superconductivity, than BaFe$_2$S$_3$ is?
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Fig5.eps}
\caption{DFT results of KFe$_2$Se$_3$ as a function of pressure. (a) Energies (per Fe) of various magnetic orders. (b) Local magnetic
moments of Fe. (c) Lattice constants normalized to the original ones for the CX state. (d) Bader charge analysis.}
\label{Fig5}
\end{figure}
Despite the experimental studies by Caron \textit{et al.}~\cite{Caron:Prb12}, DFT calculations on KFe$_2$Se$_3$
have not been performed to our knowledge. To remedy this problem, here the pure GGA calculation has been carried
out for KFe$_2$Se$_3$. At ambient conditions, the CX-AFM state is indeed the ground state (Fig.~\ref{Fig5}(a)).
The local moment of Fe is large, reaching the value $2.65$ $\mu_{\rm B}$/Fe (slightly higher than the experimental
one $2.1$ $\mu_{\rm B}$/Fe~\cite{Caron:Prb12}) at ambient conditions (Fig.~\ref{Fig5}(b)), which is a negative signal for metallicity,
and thus for potential superconductivity.
Upon pressure, the quenching of the magnetic moment and semiconductor-metal transition indeed occurs. The gap of KFe$_2$Se$_3$ is about $0.56$ eV at ambient condition, which is gradually closed by increasing pressure to $25$~GPa. The required critical pressure for magnetic quenching reaches $29$~GPa, even higher than that of KFe$_2$S$_3$. And a different feature is that this magnetic phase transition seems to be more gradual, probably of second order or weak first order, rather than occurring by a sudden jump as observed in BaFe$_2$S$_3$ and KFe$_2$S$_3$. The Bader charge analysis is shown in Fig.~\ref{Fig5}(d) as a function of pressure. Furthermore, for the higher magnetic moment in KFe$_2$Se$_3$, the critical pressure should be higher to suppress the magnetism. Then, in this case additional charge transfer from Se to Fe may be required to suppress the magnetism.
In summary of this subsection, our DFT calculations have confirmed the CX-AFM magnetic ground state for KFe$_2$Se$_3$. Although KFe$_2$Se$_3$ owns a relative high electronic density at the Fe atoms, its large gap and large moment make it even more difficult to induce a nonmagnetic metallic phase upon pressure. And this magnetic-nonmagnetic transition may be of the second order, or weak first order. These different features may arise from the Se atoms, which are larger in size and weaker in their electronegativity.
\section{Conclusion}
In this work, the magnetic and electronic properties of BaFe$_2$S$_3$, KFe$_2$S$_3$, and KFe$_2$Se$_3$ have been
analyzed using first-principles calculations. The CX-AFM magnetic order is confirmed to be the common
magnetic ground state for all these materials. The pressure-driven semiconductor to metal transition, as well as
the antiferromagnetic-to-nonmagnetic transition, has been properly reproduced. Although the DFT technique
can not directly address a superconducting state, our study can provide helpful information to understand
the superconducting transition of BaFe$_2$S$_3$ at high pressure ($11$~GPa), which is predicted to be
a first-order transition. Our main conclusion is that the electron transfer from S to Fe, i.e. a self-doping
process, may play a key role to tune the magnetism in BaFe$_2$S$_3$ and eventually induce metallicity and potentially
superconductivity.
A similar first-order transition has also been predicted for KFe$_2$S$_3$, although the required critical pressure
is higher (about $23$~GPa). By contrast, although the magnetism can also be quenched in KFe$_2$Se$_3$, the required
pressure (about $29$~GPa) is even higher and the transition seems to be of the second order, or weak first order.
Further experiments are encouraged to verify our predictions as well as the possible existence of metallicity,
and perhaps superconductivity, in KFe$_2$S$_3$ and KFe$_2$Se$_3$.
\acknowledgments{We acknowledge valuable discussions with Rong Yu, Hongyan Lu, and Songxue Chi. This work was primarily
supported by the National Natural Science Foundation of China (Grant No. 11674055). E.D. was supported by the U.S. Department of
Energy, Office of Basic Energy Sciences, Materials Sciences and Engineering Division. Most calculations
were performed on Tianhe-2 at the National Supercomputer Center in Guangzhou (NSCC-GZ).}
\bibliographystyle{apsrev4-1}
|
\section{introduction}
Nowadays, many efforts have been devoted to producing the electric
power from renewable energy sources among which the wind power is
the most technically favorable and economically attractive \cite{Harbour:2009}.
However, volatile and uncontrollable characteristics of the wind power
generation lead to stability concerns for the secure and economic
operation of modern smart grids. As the wind penetration grows
continuously, it is imperative to investigate the impacts of wind
power generations on the system stability.
In the literature, the impacts of the wind power generation have
been studied concerning different types of stabilities
\cite{Rodriguez:2003}-\cite{Bezerra:2012}. Specifically,
\cite{Rodriguez:2003}-\cite{Loparo:2011} investigated the impacts of
different parameters (e.g., the reactive power compensation,
distance to the fault, and rotor inertia) on the transient and
frequency stabilities of a power system; \cite{Kling_pes:2002}
addressed the influence of different wind generators on the
transient stability; \cite{Harbour:2009} and \cite{Keane:2012}
studied the detrimental and beneficial influences of wind generators
on transient and small-signal stabilities by converting wind
generators to conventional synchronous generators;
\cite{Vieira:2015}\cite{Bezerra:2012} analyzed the impacts of
various control algorithms of wind generators on the long-term
stability. In those studies, the variable nature of wind power is
not considered and the wind speed is oversimplified as constant. To
address this concern, \cite{crow:2012}-\cite{Nwankpa:2000} adopted
an approach that describes the uncertainty of the wind power by
stochastic differential equations (SDEs) and investigated the
impacts of the wind generation on rotor-angle and small-signal
stabilities, in which, however, the wind power was simply modeled as
a Gaussian white noise perturbation on the power injection.
Regarding the long-term stability analysis that focuses on the time
scale when fast dynamics damp out and control devices start working,
however, a comprehensive framework is still missing in the literature to
characterize the wind power with various stochastic properties, lay
down a theoretical foundation for the stability assessment of these
stochastic systems, and develop efficient numerical tools for such
stability analysis. To address these issues, the Weibull model of
the wind speed has been incorporated into the dynamic model of the
power system to perform the long-term stability analysis
\cite{Wangxz:sde1}, {where} SDEs are applied to
describe the dynamics of the wind speed. By this SDE-based model, a
theoretical approach that approximates the stochastic model by a
deterministic model has been developed to reduce the computational
burden caused by an accurate quantification of the uncertainty. Nevertheless,
the proposed model and methodology are only applicable to continuous
power system models. On the other hand, the discrete events induced
by control and protective devices occur frequently in a long time
scale after contingencies \cite{Cutsem:book}. For instance, load tap
changers are to restore the load-side voltages; shunt compensation
switchings act to increase the transmission capability; and
OvereXcitation Limiters may be activated to protect the generators
from overheating. These discrete dynamics are generally designed to
act after the fast dynamics damp out so as to avoid unnecessary interactions with the fast dynamics
\cite{Cutsem:book}-\cite{Kundur:book}, {and they require accurate representation by discrete models in time-domain simulation \cite{Milano:2011}}. As a result, it is imperative to
integrate discrete models to perform the comprehensive
long-term stability analysis for realistic power systems.
The paper begins by showing that a power grid integrating wind
power generations can be modeled as a stochastic hybrid model (SHM),
with discrete dynamics, in a SDE-based framework in which the wind
speed model that captures various stochastic properties can be
integrated. In particular, it is analytically shown in this
framework that SHM can be approximated by a deterministic hybrid
model (DHM) which offers an accurate trajectory approximation (for
SHM) and stability assessments with high computational efficiency
if some mild sufficient conditions are satisfied. A numerical example
is presented to demonstrate the accuracy and efficiency of DHM. It
is noteworthy that SHM must be implemented whenever
{any proposed sufficient conditions are violated}.
To show this necessity, we present several numerical examples in
which DHM fails to capture the instabilities of SHM. The causes for
the failure are investigated and shown to correspond to a violation
{of} the sufficient conditions. This discussion
complements the proposed SDE-based framework, shows the application
scope of the approximation methodology, and also emphasizes the
largely-neglected necessity of the stochastic model in the long-term
stability analysis.
As the modern smart grids endeavor to incorporate high penetration of
intermittent renewable energy, integrate plug-in vehicles, and
encourage opportunistic users, the operation and control of
power grids are required to account for the resulting high
variability and uncertainty. We believe that the proposed SDE-based
framework and approximation methodology can be readily generalized
to conduct stability assessments for power systems with the
uncertainties brought about by various renewable energy sources,
plug-in vehicles, smart appliances, opportunistic users, and so
forth.
The remainder of the paper is organized as follows. Section
\ref{sectionmodel} introduces the SDE-based framework of power
system models integrating the stochastic dynamics of the wind speed.
Section \ref{sectiontheory} develops an approximation methodology
for SHM in the SDE-based framework, which provides an accurate
trajectory approximation and correct stability assessments with a
high simulation speed. In particular, a diagram is summarized at the
end of Section \ref{subsectiontheory} to illustrate the
relationships among the proposed models and theoretical results.
Furthermore, Section \ref{sectionnecessity} presents some critical
cases in which some sufficient conditions of the proposed
methodology are violated, to explain the necessity of implementing
SHM to obtain correct stability assessments.
\section{sde-based framework of hybrid models}\label{sectionmodel}
The conventional long-term stability model (i.e., the complete
dynamic model) without stochasticity for simulating the system
dynamic response to a disturbance in the $\tau$ time scale can be
described as follows {(see (22)-(25)
\cite{Wangxz:sde1} and (15) \cite{Wangxz:CAS}):}
\begin{eqnarray}
\bm{{z}_d}(k)&=&\bm{{h}_d}(\bm{z_c,x,y,z_d}(k-1))\label{slowdde}\\
\bm{{z}_{c}^\prime}&=&\bm{{h}_c({z_c,x,y,z_d})}\label{slowode}\\
\epsilon\bm{{{x}^\prime}}&=&\bm{f({z_c,x,y,z_d})}\label{fastode}\\
\bm{0}&=&\bm{g({z_c,x,y,z_d})}\label{algebraiceqn}
\end{eqnarray}
where $\tau=t\epsilon$ and $\prime$ refers to $\frac{d}{d\tau}$. Here,
(\ref{slowdde}) accounts for the long-term discrete events, such as
shunt capacitors and load tap changers (LTCs); (\ref{slowode})
depicts the slow dynamics, including self restorative loads, turbine
governors (TGs), and OvereXcitation Limiters (OXLs); (\ref{fastode})
describes the fast dynamics of components, such as synchronous
machines, doubly-fed induction generators (DFIGs), induction motors,
and exciters; and (\ref{algebraiceqn}) describes the power flow
relation and internal relationships between variables. In addition,
$\bm{h_d}$ are discrete functions; $\bm{z_d}$ are slow discrete
variables whose changing from $\bm{z_d}(k-1)$ to $\bm{z_d}(k)$
relies on (\ref{slowdde}) and occurs at times $t_k$, $1\le k\le N$.
The functions $\bm{{h}_c}$, $\bm{{f}}$, and $\bm{{g}}$ are
continuous; $\bm{z_c}$, $\bm{{x}}$, and $\bm{{y}}$ are the vectors
of slow state variables, fast state variables, and algebraic
variables, respectively; and $\epsilon$ is deemed as the reciprocal of
the maximum time constant among all components.
\subsection{Stochastic Model of Wind Speed}\label{wind model}
The impacts of the wind power on the system stability have been
addressed \cite{Rodriguez:2003}-\cite{Bezerra:2012} in which the
wind speed is termed as a constant and an entry of the vector
$\bm{y}$---algebraic variables. In this paper, we characterize the
randomness of the wind speed by a stochastic model.
Specifically, given $n_w$ independent wind energy sources that each
energy source follows a certain probability distribution, the wind
speeds of the $n_w$ sources are collectively denoted by a vector
$\bm{y_w}$ in the following model (see \cite{Wangxz:sde1} and
\cite{Milano:2013_1}):
\begin{eqnarray}
\epsilon\bm{\eta_w}^\prime &=& -A\bm{\eta_w}+\sigma\bm{\xi}={\bm{f_w}(\bm{\eta_w})+\sigma\bm{{\xi}}}, \label{wind1}\\
\bm{y_w} &=& \bm{\hat{F}_w^{-1}}\left(\bm{\hat{\Phi}}
\big(\tfrac{\bm{\eta_w}}{\sigma/\sqrt{2\bm{\alpha}}}\big)\right) =
\bm{g_w(\eta_w)}, \label{wind2}
\end{eqnarray}
where $\bm{\eta_w}, \bm{y_w} \in \mathbb{R}^{n_w}$, the matrix
$$A=\mbox{diag}(\bm{\alpha})=\mbox{diag}[\alpha_1,\ldots,\alpha_{n_w}]\in
\mathbb{R}^{{n_w}\times{n_w}}$$ determines the autocorrelation
property of $\bm{y_w}$ (see below for more details), and
$\int_0^{t}\bm{\xi}(s)ds$ is an $n_w$-dimensional Wiener process. In
addition, $\bm{\hat{F}_w}=[F_1(\eta_{w_1}),
F_2(\eta_{w_2}),...F_{n_w}(\eta_{w_{n_w}})]^T$,
$\bm{\hat{\Phi}}=[\Phi(\eta_{w_1}),\Phi(\eta_{w_2}),...\Phi(\eta_{w_{n_w}})]^T$,
and $\bm{g_w}:\mathbb{R}^{n_w}\mapsto\mathbb{R}^{n_w}$, where $F_i$
is the cumulative distribution function of the corresponding wind
speed $y_{w_i}$, and $\Phi$ is the cumulative distribution function
of a Gaussian distribution.
In model (\ref{wind1})-(\ref{wind2}) of the wind speed,
$\bm{\eta_w}$ is a vector Ornstein-Uhlenbeck process, and each
$y_{w_i}$ matches the distribution of $F_i$ by the property of the
memoryless transformation \cite{Milano:2013_1}. For example, if the
wind speed of source $n_{w_i}$ is governed by the Weibull
distribution with a shape parameter $k_i>0$ and a scale parameter
$\lambda_i>0$, then
\begin{equation}
F_{w_i}(u)={{1-e^{(u/\lambda_i)^{k_i}}}} ~\hbox{for all}~ u>0,
\end{equation}
and $y_{w_i}$ has the following statistical properties
(see (26)-(28)\cite{Milano:2013_1}):
\begin{enumerate}[(i)]
\item $\mathrm{E}[{y_{w_i}}(t)]={\lambda_i}\Gamma(1+\frac{1}{{k_i}})={\mu_{w_i}}$.
\item $\mathrm{Var}[{y_{w_i}}(t)]={\lambda_i}^2
\Gamma(1+\frac{2}{{k_i}})-\mu_{w_i}^2$.
\item $\mathrm{Aut}[{y_{w_i}}(t_k),{y_{w_i}}(t_j)]\approx
e^{-\alpha_i|t_j-t_k|}$.
\end{enumerate}
Note that $\lambda_i$, $k_i$, and $\alpha_i$ are the parameters that
determine the statistical properties of wind speed $y_{w_i}$, but
$\sigma$ does not. So $\sigma$ can be arbitrarily selected
\cite{Wangxz:sde1}\cite{Milano:2013_1}. Indeed, $\sigma$ is only
an intermediate parameter to generate the Ornstein-Uhlenbeck process
$\bm{\eta_w}$. The readers are referred to \cite{Milano:2013_1} for
more details.
\subsection{Hybrid Models}
When integrating the stochastic model (\ref{wind1})-(\ref{wind2}) of
the wind speed into the long-term stability model
(\ref{slowdde})-(\ref{algebraiceqn}), the stochastic hybrid model
(SHM) takes the following form:
\begin{eqnarray}
\bm{{z}_d}(k)&=&\bm{\bar{h}_d(\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},}\bm{z_d}(k-1)\bm{)}, \label{dde}\\
\bm{{z}_{c}}^\prime&=&\bm{\bar{h}_c({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}, \label{slow_ode}\\
\epsilon{\bm{\bar{x}}}^\prime&=&\bm{\bar{f}({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}+\sigma B \bm{\bar{\xi}}, \label{fast_ode}\\
\bm{0}&=&\bm{\bar{g}({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})},
\label{alg eqn}
\end{eqnarray}
{where $\bm{\bar{x}} \doteq
\left[\begin{smallmatrix}\bm{x}
\\ \bm{\eta_w}\end{smallmatrix}\right]$,
$\bm{\bar{y}} \doteq \left[\begin{smallmatrix}\bm{y}
\\ \bm{y_w}\end{smallmatrix}\right]$, and
$B \doteq
\left[\begin{smallmatrix}0\\I_{n_w}\end{smallmatrix}\right]$,
nonzero entries of which correspond to $n_w$ independent wind
sources. In addition, $\bm{\bar{f}}\doteq \left[\begin{smallmatrix} \bm{f} \\
\bm{f_w}
\end{smallmatrix}\right]$, $\bm{\bar{g}} \doteq \left[\begin{smallmatrix} \bm{g} \\
\bm{p}
\end{smallmatrix}\right]$ with $\bm{p} \doteq \bm{y_w}-\bm{g_w(\eta_w)}$,
and $\bar{\bm{\xi}}=\left[\begin{smallmatrix}
0\\ {\bm{\xi}}\end{smallmatrix}\right]\in \mathbb{R}^{{n_x}\times{n_w}}$.
Here, (\ref{dde}) and
(\ref{slow_ode}) are directly derived from (\ref{slowdde}) and
(\ref{slowode}), respectively, such that
$\bm{\bar{h}_d(\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},}\bm{z_d}(k-1)\bm{)}
= \bm{{h}_d}(\bm{z_c,x,y,z_d}(k-1))$ and
$\bm{\bar{h}_c({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}=
\bm{{h}_c({z_c,x,y,z_d})}$; (\ref{fast_ode}) is obtained from a
combination of (\ref{fastode}) and (\ref{wind1}), whereas (\ref{alg
eqn}) is derived by combining (\ref{algebraiceqn}) and
(\ref{wind2}).}
Recall that discrete dynamics described by (\ref{dde}) play
important roles in the long-term stability because many protective
and control devices may take effect in the long-term time scale to
restore the load-sided power, protect generators, and so on.
This study aims to show that the SHM (\ref{dde})-(\ref{alg eqn}) can
be well approximated by a deterministic hybrid model (DHM), say,
\begin{eqnarray}
\bm{{z}_d}(k) &=& \bm{\bar{h}_d(\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},}\bm{z_d}(k-1)\bm{)}, \label{det_dde}\\
\bm{{z}_{c}}^\prime &=& \bm{\bar{h}_c({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}, \label{det_slow_ode}\\
\epsilon{\bm{\bar{x}}}^\prime &=& \bm{\bar{f}({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}, \label{det_fast_ode}\\
\bm{0} &=&
\bm{\bar{g}({\bm{z_c},\bm{\bar{x}},\bm{\bar{y}},\bm{z_d}})}.
\label{det_alg eqn}
\end{eqnarray}
Note that the vector of algebraic variables $\bm{\bar{y}}$ in
(\ref{alg eqn}) and (\ref{det_alg eqn}) can be eliminated under
\textit{Assumption 1} {which is a generic property
satisfied in normal operating conditions
\cite{Cutsem:book}\cite{Chiang:book}.} \\
\textbf{\textit{Assumption 1.}} {\it The DHM
(\ref{det_dde})-(\ref{det_alg eqn}) does not encounter singularity,
i.e., $\frac{\partial{\bm{\bar{g}}}}{\partial{\bm{y}}}$ is
nonsingular along the trajectory.}\\
Under \textit{Assumption 1}, $\bm{\bar{y}}$ can be represented in
terms of $\bm{z_c}$, $\bm{\bar{x}}$, and $\bm{z_d}$ using
(\ref{det_alg eqn}), namely
$\bm{\bar{y}=m(\bm{z_c},\bm{\bar{x}},\bm{z_d})}$. Then, the SHM
(\ref{dde})-(\ref{alg eqn}) can be written as:
\begin{eqnarray}
\bm{z_d}(k)&=&\bm{H_d(\bm{z_c},\bm{\bar{x}},}\bm{z_d}(k-1)\bm{)}, \label{sde power dde}\\%\label{sde power dde-0}\\
\bm{z_c}^\prime&=&\bm{H_c(\bm{z_c},\bm{\bar{x}},\bm{z_d})}, \label{sde power slow}\\%\label{sde power slow-0}\\
\epsilon\bm{\bar{x}}^\prime&=&\bm{F(\bm{z_c},\bm{\bar{x}},\bm{z_d})}+\sigma B {\bm{\bar{\xi}}}. \label{sde power fast
\end{eqnarray}
By analogy, the DHM (\ref{det_dde})-(\ref{det_alg eqn}) is
equivalently converted to:
\begin{eqnarray}
\bm{z_d}(k)&=&\bm{H_d(\bm{z_c},\bm{\bar{x}},}\bm{z_d}(k-1)\bm{)}, \label{det power dde}\\
\bm{z_c}^\prime &=&\bm{H_c(\bm{z_c},\bm{\bar{x}},\bm{z_d})}, \label{det power slow}\\
\epsilon\bm{\bar{x}}^\prime &=&\bm{F(\bm{z_c},\bm{\bar{x}},\bm{z_d})}.
\label{det power fast}
\end{eqnarray}
In section \ref{sectiontheory}, a theoretical foundation is to be
developed to ensure the effectiveness of the approach that
approximates the SHM (\ref{sde power dde})-(\ref{sde power fast}) by
the DHM (\ref{det power dde})-(\ref{det power fast}) in the
long-term stability study. The key is to show that if some mild
conditions are satisfied, then the DHM (\ref{det power
dde})-(\ref{det power fast}) is theoretically ensured to provide an
accurate trajectory approximation and stability assessments for the
SHM (\ref{sde power dde})-(\ref{sde power fast}). Clearly, the DHM
consumes much less computational resources in the simulation
compared with the SHM and may serve as an efficient stability
assessment tool for power grids with significant wind power
generations.
\section{an approximation methodology for stochastic hybrid model}\label{sectiontheory}
The singular perturbation method for SDEs
\cite{Gentz:2006}-\cite{Freidlin:book} and sufficient conditions for
the quasi steady-state (QSS) model
\cite{Wangxz:CAS}\cite{Wangxz:wiley} are employed here to develop a
theoretical foundation for an approximation of the SHM (\ref{sde
power dde})-(\ref{sde power fast}) by the DHM (\ref{det power
dde})-(\ref{det power fast}). A numerical example using a 145-bus system is presented to demonstrate the
accuracy and efficiency of the DHM.
\subsection{Theoretical Foundation}\label{subsectiontheory}
In the SHM, when the discrete jumping is initiated, discrete
variables $\bm{z_d}$ are updated first by (\ref{sde power dde}), and
then the system acts according to (\ref{sde power slow})-(\ref{sde
power fast}) with constant $\bm{z_d}$. In this regard, one can treat
the SHM (\ref{sde power dde})-(\ref{sde power fast}) as a series of
continuous systems (\ref{sde power slow})-(\ref{sde power fast})
with constant $\bm{z_d}$ \cite{Cutsem:book}. Similarly, the DHM
(\ref{det power dde})-(\ref{det power fast}) can be considered as a
series of continuous systems (\ref{det power slow})-(\ref{det power
fast}) with constant $\bm{z_d}$. It is reasonable to assume that the
SHM and the DHM are governed by the same sequence of parameter
values $\bm{z_d}$ given the same initial condition. So, the hybrid
models (i.e., SHM and DHM) can be analyzed by comparing the
corresponding continuous systems in the series. Additionally, we
suppose that each deterministic continuous system (\ref{det power
slow})-(\ref{det power fast}) satisfies some generic
differentiability and non-degeneracy conditions (see
\textit{Assumption 2.1} \cite{Wangxz:sde1}), which are reasonable
assumptions for real-life physical systems.
If $\bm{\bar{x}}={\bm{m_1(\bm{z_c},\bm{z_d})}}$ is an asymptotically
stable
equilibrium point of the short-term stability model
$\bm{0}=\bm{F(\bm{z_c},\bm{\bar{x}},\bm{z_d})}$ for all $\bm{z_c}$ and $\bm{z_d}$,
i.e., $\bm{\bar{x}}=\bm{m_1(\bm{z_c},\bm{z_d})}$ is a stable
component of the constraint manifold, then there exists an invariant
manifold of system (\ref{det power dde})-(\ref{det power fast}):
$\bm{\bar{x}}=\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)=\bm{m_1(\bm{z_c},\bm{z_d})}+O(\epsilon)$
for sufficiently small $\epsilon$
\cite{Wangxz:sde1}\cite{Khalil:book}\cite{Alberto:article}, where
$\bm{m_1(\bm{z_c},\bm{z_d})}$ and
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$ can be not smooth. An
ellipsoidal layer ${M}(h)$ around
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$ is defined as follows:
\begin{eqnarray}
{M}(h)&\doteq&\{(\bm{z_c},\bm{\bar{x}},\bm{z_d}):\langle (\bm{\bar{x}}-\bm{m_1}^\star(\bm{z_c},\bm{z_d},\epsilon)),\nonumber\\
&&\quad
M_1^\star(\bm{z_c},\bm{z_d},\epsilon)^{-1}(\bm{\bar{x}}-\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon))\rangle<
h^2\}. \nonumber
\end{eqnarray}
Here, the matrix $M_1^\star(\bm{z_c},\bm{z_d},\epsilon)$ that represents
the cross section of $M(h)$ is properly defined (see Appendix B in
\cite{Wangxz:sde1}), and an illustration for $M(h)$ is shown in Fig.
\ref{N(h)}.
\begin{figure}[!ht]
\centering
\includegraphics[width=2.5in,keepaspectratio=true,angle=0]{mh_discrete_modify-eps-converted-to.pdf}
\caption{An illustrtion of $M(h)$ in the DHM. Here, $n_{z_c}=1$,
$n_{\bar{x}}=2$, and $M(h)$ is an ellipsoidal layer around
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$. }\label{N(h)}
\end{figure}
Given the well-defined initial condition for each continuous system
(\ref{sde power slow})-(\ref{sde power fast}) with constant
$\bm{z_d}$ (of the SHM), the following theorem shows that the
trajectories of the SHM (\ref{sde power dde})-(\ref{sde power fast})
are confined in $M(h)$ despite the changes of discrete variables,
provided that the slow manifold is stable.\\
\textbf{\textit{Theorem 1 (Sample-Path Concentration for SHM):}}\\
{\it Consider the SHM (\ref{sde power dde})-(\ref{sde power fast})
in the study region $D_{\bm{z_d}} \times D_{\bm{z_c}} \times
D_{\bm{x}}$, for some fixed $\epsilon_0>0$, $h_0>0$, there exist
$\delta_0>0$, a time $\tilde{\tau}_k$ of order $\epsilon|\mbox{log}h|$,
\textcolor{black}{and $\tau_k > \tilde{\tau}_k$} for each continuous
system (\ref{sde power slow})-(\ref{sde power fast}) with $0\le k\le
N$ such that if the following conditions (i) and (ii) are satisfied:
\begin{description}
\item[(i)] The slow manifold $\bm{\bar{x}}=\bm{m_1(\bm{z_c},\bm{z_d})}$ is a
stable component of the constraint manifold, where $\bm{z_c}\in
D_{\bm{z_c}}$ and $\bm{z_d}\in D_{\bm{z_d}}$;
\item[(ii)] The initial condition $(\bm{z_c}^k(0),\bm{\bar{x}}^k(0),\bm{z_d}(k))$ for {each} continuous
system (\ref{sde power slow})-(\ref{sde power fast}) of the SHM
satisfies that $(\bm{z_c}^k(0),\bm{\bar{x}}^k(0),\bm{z_d}(k))\in
M(\delta_0)$, where $\bm{z_c}^k(0)\in D_{\bm{z_c}}$ and
$\bm{z_d}(k)\in D_{\bm{z_d}}$ for $k\in[0,1,...N]$,
\end{description}
then, for all $\tau\in
\Pi=\cup_{i=1}^{N-1}[\tilde{\tau}_i,\tau_i)\cup[\tilde{\tau}_N,\infty]$,
the sample path $(\bm{z_c}(\tau),\bm{\bar{x}}(\tau),\bm{z_d}(\tau))$
of the SHM (\ref{sde power dde})-(\ref{sde power fast}) satisfies
the following probability property:
\begin{eqnarray}\label{theorem1}
&\mathbb{P}\{\exists\, \tau\in\Pi: &(\bm{z_c}(\tau),\bm{\bar{x}}(\tau),\bm{z_d}(\tau))\notin M(h)\}\nonumber\\
&&\leq
C_{n_{z_c},n_x}(\tau,\epsilon)e^{\frac{-h^2}{2\sigma^2}(1-O(h)-O(\epsilon))},
\end{eqnarray}
for all $\epsilon\leq\epsilon_0$, $h\leq h_0$, where the coefficient
$C_{n_{z_c},n_x}(\tau,\epsilon)=[C^{n_{z_c}}+h^{-n_x}](1+\frac{\tau}{\epsilon^2})$
is linear in $\tau$.}
\vspace{12pt}
{Proof:} See Appendix \ref{prooftheorem1}.\\
\textit{Theorem 1} shows that if conditions (i)-(ii) are satisfied,
then the probability that the sample path leaves $M(h)$ is less than
the right hand side (RHS) of (\ref{theorem1}). Specifically, if
$h\gg \sigma$, i.e., the deepness of the layer $h$ is far larger
than $\sigma$ related with wind speeds, then the RHS of
(\ref{theorem1}) becomes very small, which suggests that the sample
pathes of the SHM do not leave $M(h)$ almost surely
\cite{Wangxz:sde1}\cite{Gentz:2006}. So, there is no need to worry
about the probability when investigating the relations between the
trajectory of the SHM (\ref{sde power dde})-(\ref{sde power fast})
and that of the DHM (\ref{det power dde})-(\ref{det power fast}). On
the other hand, $\sigma$ does not influence the stochastic
properties of wind speed $\bm{y_w}$ as stated in Section \ref{wind
model} or Section III-A in \cite{Wangxz:sde1} (where $\sigma$ is
only an intermediate parameter to generate the Ornstein-Uhlenbeck
process $\bm{\eta_w}$). In this regard, $\sigma$ can be selected as
small as needed such that any adequate $h$ satisfies $h\gg\sigma$.
In other words, the requirement $h\gg\sigma$ can be readily
fulfilled in this SDE-based framework. {In addition, Theorem 2.4
\cite{Gentz:2003} has commented that for $h\gg\sigma$, the first
exit time that the solution $\bm{z_c}$ of the (continuous)
stochastic system (\ref{sde power slow}) leaves the region
$D_{\bm{z_c}}$ is very large (exponentially in $h^2/\sigma^2$), that
is, $\bm{z_c}$ still stays within $D_{\bm{z_c}}$ almost for sure
right before the (discrete) change of $\bm{z_d}$ occurs at
$t_k$. Note that, for adequately controlled systems, discrete
devices generally do not result in severe perturbations to the
system dynamics. So, these facts suggest that condition (ii) in
Theorem 1 is generally satisfied under normal operating conditions.}
Under the condition $h\gg \sigma$, we next investigate the
relationship between the trajectory of the SHM (\ref{sde power
dde})-(\ref{sde power fast}) and that of the DHM (\ref{det power
dde})-(\ref{det power fast}). If (a) the trajectory of the SHM
remains in $M(h)$ which is an $\epsilon$ neighborhood of the invariant
manifold $\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$, and (b) the
trajectory of the DHM evolves along
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$, then we show that the
distance between the trajectory of the SHM and that of the DHM can
be readily obtained. Note that \textit{Theorem 1} provides
sufficient conditions for (a). So, the remaining question is about
how to ensure (b). Incidentally, the theoretical foundation for the
quasi steady-state (QSS) model in \cite{Wangxz:CAS} has provided
sufficient conditions for (b). In particular, one of the sufficient
conditions for (b) is the condition of \textit{consistent
attraction} defined below and illustrated in Fig.
\ref{consistentattraction}.\\
\textbf{\textit{Definition 1. Consistent Attraction
\cite{Wangxz:CAS}}}: {\it By fixing $\bm{z_c}$ and $\bm{z_d}$ as the
parameters, the short-term stability model refers to (\ref{det power
fast}). We say that the DHM (\ref{det power dde})-(\ref{det power
fast}) satisfies the condition of \textit{consistent attraction} if
the initial condition is contained in the stability region of the
initial short-term stability model and whenever discrete variables
jump from $\bm{z_d}(k-1)$ to $\bm{z_d}(k)$, $k=1,2,...,N$, the point
on trajectory of the DHM immediately after $\bm{z_d}$ jump still
stays within the stability region of the corresponding short-term
stability model.}
\begin{figure}[!ht]
\centering
\includegraphics[width=3in,keepaspectratio=true,angle=0]{consistentattraction-eps-converted-to.pdf}
\caption{The situation when the DHM satisfies the condition of
consistent attraction.}\label{consistentattraction}
\end{figure}
\textit{The condition of consistent attraction} ensures that the
trajectory of the DHM is always close to the slow manifold
$\bm{m_1(\bm{z_c,z_d})}$ despite the changing of discrete variables
(if the slow manifold is also stable), then the trajectory of the
DHM always evolves along the invariant manifold
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$. Let
$(\bm{z_{c}}(\tau),\bm{\bar{x}}(\tau),\bm{z_d}(\tau))$ be the
trajectory of the SHM (\ref{sde power dde})-(\ref{sde power fast}),
and let $(\bm{z_{cD}}(\tau),\bm{\bar{x}_D}(\tau),\bm{z_d}(\tau))$ be
that of the DHM (\ref{det power dde})-(\ref{det power fast}). Then,
the following theorem reveals the relationship between the
trajectories of the two models.\\
\textbf{\textit{Theorem 2 (Trajectory Relationship for Hybrid
Models):}}\\ {\it Given $h\gg\sigma$, consider the SHM (\ref{sde
power dde})-(\ref{sde power fast}) and the DHM (\ref{det power
dde})-(\ref{det power fast}) in the study region $D_{\bm{z_d}}
\times D_{\bm{z_c}} \times D_{\bm{x}}$, for some fixed
$\epsilon_0\in(0,h)$, there exist $\delta_0>0$, a time
$\tilde{\tau}_k$ of order $\epsilon|\mbox{log}h|$ and
$\bar{\tau}_k>\tilde{\tau}_k$ for each continuous system (\ref{sde
power slow})-(\ref{sde power fast}) where $k=0,1,...,N$, such that if
the following conditions (i), (ii) and (iii) are satisfied:
\begin{description}
\item[(i)] The slow manifold $\bm{\bar{x}}=\bm{m_1(\bm{z_c},\bm{z_d})}$ is a
stable component of the constraint manifold, where $\bm{z_c}\in
D_{\bm{z_c}}$ and $\bm{z_d}\in D_{\bm{z_d}}$;
\item[(ii)] The initial condition $(\bm{z_c}^k(0),\bm{\bar{x}}^k(0),\bm{z_d}(k))$ for each continuous
system (\ref{sde power dde})-(\ref{sde power fast}) of the SHM
satisfies $(\bm{z_c}^k(0),\bm{\bar{x}}^k(0),\bm{z_d}(k))\in
M(\delta_0)$, where $\bm{z_c}^k(0)\in D_{\bm{z_c}}$ and
$\bm{z_d}(k)\in D_{\bm{z_d}}$ for $k\in[0,1,...N]$;
\item [(iii)] The DHM (\ref{det power dde})-(\ref{det power fast})
satisfies \textbf{the condition of consistent attraction},
\end{description}
then, for $\tau\in\cup_{i=1}^{N}[\tilde{\tau}_i,\bar{\tau}_i]$, the
following relations hold:
\begin{eqnarray}
|\bm{\bar{x}}(\tau)-\bm{\bar{x}_D}(\tau)|&=&O(\sigma), \label{corollary3_1}\\
|\bm{z_{c}}(\tau)-\bm{z_{cD}}(\tau)|&=&O(\sigma\sqrt{\epsilon}),
\label{corollary3_2}
\end{eqnarray}
for all $\epsilon\in(0,\epsilon_0)$.}
\vspace{12pt}
Proof: See Appendix \ref{prooftheorem2}.\\
By \textit{Theorem 2} we observe that if the proposed sufficient
conditions are satisfied, the trajectory of the SHM can be
approximated by that of the DHM as illustrated in Fig.
\ref{stochasticdeterministic}.
\begin{figure}[!ht]
\centering
\includegraphics[width=3in,keepaspectratio=true,angle=0]{stochasticdeterministic_discrete_modify-eps-converted-to.pdf}
\caption{The trajectory
$\phi_s(\tau,\bm{z_c}(0),\bm{\bar{x}}(0),\bm{z_d})$ of SHM is
bounded in $M(\sigma)$, and can be estimated by the trajectory
$\phi_l(\tau,\bm{z_c}(0),\bm{\bar{x}}_D(0),\bm{z_d})$ of DHM.
}\label{stochasticdeterministic}
\end{figure}
\begin{figure*}[!ht]
\includegraphics[width=7in]{framework.pdf}
\caption{The proposed SDE-based framework and the
approximation methodology.}\label{framework}
\end{figure*}
Generally speaking, sufficient conditions (i)-(iii) in
\textit{Theorem 2} are moderate and satisfied when the system
operates away from the stability boundary, and thus the DHM can
substitute the SHM and typically offer correct stability assessments
with less simulation time. But, as detailed in Section
\ref{sectionnecessity}, the SHM must be applied if any of the
sufficient conditions is violated.
For clarity, we summarize in Fig. \ref{framework} the proposed
SDE-based framework, relationship between different models, and
importance of derived theoretical results. In the SDE-based
framework, the stochastic model (\ref{wind1})-(\ref{wind2}) of the
wind speed is incorporated into the conventional power system model
(\ref{slowdde})-(\ref{algebraiceqn}), and the resulting hybrid model
(\ref{dde})-(\ref{alg eqn}) is equivalent to the SHM (\ref{sde power
dde})-(\ref{sde power fast}) under normal operating condition.
Specifically, \textit{Theorem 1-2} shows that the DHM (\ref{det
power dde})-(\ref{det power fast}) can well approximate the SHM
(\ref{sde power dde})-(\ref{sde power fast}). In particular,
\textit{Theorem 1} suggests that the sample paths of (\ref{sde power
dde})-(\ref{sde power fast}) are concentrated in a neighborhood
$M(h)$ of the invariant manifold
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$, while \textit{Theorem 2}
asserts that the DHM (\ref{det power dde})-(\ref{det power fast})
can provide an accurate trajectory approximation and stability
assessments for the SHM (\ref{sde power dde})-(\ref{sde power fast})
under some mild conditions. So, under normal operating conditions
and the proposed mild conditions, the DHM
(\ref{det_dde})-(\ref{det_alg eqn}) can well approximate the SHM
(\ref{dde})-(\ref{alg eqn}) in terms of the trajectory and stability
assessments.
\subsection{Numerical Illustration}
Numerical studies using a 145-bus test case \cite{testcasearchive}
are conducted in PSAT-2.1.8 \cite{Milano:PSAT} to show the accuracy
and efficiency of the derived results. The test system has 6
doubly-fed induction generators (DFIGs) driven by 6 independent
Weibull-distributed wind sources. The parameters of Weibull
distributions are referred from \cite{Milano:2013_1} which fit the
1-h wind speed data of the Cape St. James and Victoria Airport. The
readers are referred to Table 1 \cite{Milano:2013_1} for more
details. In addition, there are 50 synchronous generators (GENs)
with automatic voltage regulators (AVRs). Turbine governors (TGs)
are equipped for GEN 10-GEN 20, and OvereXcitation Limiters (OXLs)
are also equipped for GEN 1-GEN 6. The initial time delays of OXLs
are 50s. Moreover, 5 discrete load tap changers (LTCs) are installed
at Bus 79-95, Bus 1-33, Bus 79-92, Bus 1-5, and Bus 60-95,
respectively. Particularly, the discrete model of LTCs is shown
below \cite{Kundur:book}:
\begin{equation}\label{LTC_d1}
n(k+1)=\left\{\begin{array}{ll}n(k)+\triangle{n}, &\mbox{if }v>v_0+d\mbox{ and }n(k)<n^{max};\\
n(k)-\triangle{n}, &\mbox{if }v<v_0-d\mbox{ and }n(k)>n^{min};\\
n(k), &\mbox{otherwise};\end{array}\right.
\end{equation}
where $n$ is the tap changer ratio, $v$ is the controlled voltage,
$v_0$ is the reference voltage, $d$ is half of the LTC dead-band,
$n^{max}$ and $n^{min}$ are the upper and lower tap limits,
respectively. All LTCs have initial time delays of 50s and fixed
tapping delays of 10s. At 0.5s, three lines at Bus 95-138, Bus
94-138, Bus 94-95 trip.
{Note that the dynamic
models for synchronous generators and DFIGs used in this and
subsequent numerical examples are all detailed in Ch.\,17 and
Ch.\,21 \cite{Milano:PSATDOC}. Specifically, the order II and order IV
models of GENs are employed for the simulation of this 145-bus system.}
\begin{figure}[!ht]
\centering
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vmofExc10_145discrete_modify-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{tg3ofTg3_145discrete_modify-eps-converted-to.pdf}
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{VatBus95_145discrete_modify-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{QofSyn7_145discrete_modify-eps-converted-to.pdf}
\end{minipage}
\caption{A comparison of the trajectory of the SHM and that of the
DHM by simulating the 145-bus system.} \label{my145wind_discrete}
\end{figure}
Fig. \ref{my145wind_discrete} presents a comparison of the
trajectory of the SHM and that of the DHM for which the quasi
steady-state (QSS) model \cite{Cutsem:book} is implemented to obtain
the slow manifolds of the DHM. Observe that the trajectories of the
SHM always keep close to those of the DHM despite the changing of
discrete variables, and both models give the same stability
assessments that the system is stable in the long-term time scale.
Clearly, all sufficient conditions of \textit {Theorem 2} are
satisfied, then the conclusions of \textit{Theorem 2} hold.
Particularly, Fig. \ref{my145wind_discrete} shows that the DHM does
not encounter the singularity and its slow manifold is stable. In
addition, the trajectory of the DHM evolves along
$\bm{m_1^\star}(\bm{z_c},\bm{z_d},\epsilon)$ which is an
$\epsilon$-neighborhood of the slow manifold. This illustrates the
results of \textit{Theorem 2}.
Concerning the computational efficiency, the SHM takes 137.118s to
complete the simulation, whereas the DHM only consumes 57.913s. Note
that several trajectories of the SHM may be required to evaluate the
stability in critical cases. But, the time needed to simulate one
trajectory (of the SHM) can be more than twice as that required by
the DHM.
From this example, we observe that the DHM can provide an accurate
trajectory approximation and stability assessments for the SHM with
far less simulation time, provided that the proposed mild conditions
are satisfied.
\section{necessity of stochastic model}\label{sectionnecessity}
A comprehensive theoretical framework has been developed to
approximate the SHM by the DHM. Specifically, if all sufficient
conditions of \textit{Theorem 2} are satisfied, then the DHM can
provide an accurate trajectory approximation and stability
assessments for the SHM with much less simulation time. In the
section, we further present several examples in critical cases that
the DHM fails to provide a satisfactory approximation. The causes
for such failure are investigated in the nonlinear system framework
and are shown to correspond to a violation of the proposed
sufficient conditions. This discussion complements the proposed
framework and methodology and also highlights the necessity of the
stochastic model when performing the stability analysis for the
power system with significant wind power generations, especially for
the system that operates close to the stability boundary. Given the
importance of such sufficient conditions, it is imperative to
develop efficient numerical algorithms to check these conditions in
the near future.
\subsection{Numerical Example I}
This example is a modified IEEE 14-bus system. {The order V and order
VI models of GENs are employed.} A Weibull-distributed
wind source drives a DFIG at Bus 2, and 3 GENs are equipped with
AVRs and TGs. In addition, 3 exponential recovery loads (ERLs) are
at Bus 9, 10, and 14, respectively. An OXL is installed for GEN 1,
and 3 discrete LTCs are at Bus 4-9, Bus 12-13 and Bus 2-4,
respectively, the initial time delays of which are 30s and fixed
tapping delays are 10s. At 1s, three lines at Bus 6-13, Bus 7-9, and
Bus 6-11 trip. We refer the reader to Appendix \ref{appendix2-1} for
the parameter values.
A comparison between the trajectory of the DHM and that of the SHM
is shown in Fig. \ref{my14trywind1}. The slow manifold of the DHM
acquired from the QSS model is also illustrated. The DHM converges
to a long-term stable equilibrium point (SEP) with all voltages in
the nominal range, which shows that the DHM is long-term stable.
But, the sample path of the SHM suffers from a voltage collapse. So,
the DHM fails to provide a stability assessment agreeing with the
SHM.
\begin{figure}[!ht]
\centering
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vr2ofExc1_14trywind1_modify_1-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{VatBus4_14trywind1_modify_1-eps-converted-to.pdf}
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{PofSyn1_14trywind1_modify_1-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{ifofOxl1_14trywind1_modify_1-eps-converted-to.pdf}
\end{minipage}
\caption{A comparison between the trajectory of the SHM and that of
the DHM using the 14-bus system. The DHM fails to capture the
instability of the SHM.}\label{my14trywind1}
\end{figure}
The failure of the DHM is caused by a violation of condition (iii),
i.e., \textit{the condition of consistent attraction}, in
\textit{Theorem 2}. When the discrete variables (i.e., the ratios of
LTCs) change at 120s, the state of the DHM lies outside the
stability region of the corresponding short-term stability model. To
show this, the following simulations are conducted similar to the
approach in \cite{Wangxz:CAS}. When discrete variables jump at 110s,
the trajectories of two fast variables of the corresponding
short-term stability model starting from the state of the DHM are
shown in Fig. \ref{my14trywind1_fastmodel-110}. Observe that the
trajectories converge to the SEP of the corresponding short-term
stability model which shows that \textit{the condition of consistent
attraction} is satisfied at this time.
\begin{figure}[!ht]
\centering
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vr1ofExc2_14trywind1_fastmodel_110-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vr1ofExc3_14trywind1_fastmodel_110-eps-converted-to.pdf}
\end{minipage}
\caption{The trajectories of two fast variables in the short-term
stability model when $\bm{z_d}$ change at 110s. The trajectory
starting from the state of DHM converges to the SEP of the
corresponding short-term stability model which shows that the
condition of consistent attraction is
satisfied.}\label{my14trywind1_fastmodel-110}
\end{figure}
But, if the discrete variables jump at 120s, the trajectories of the
same two fast variables of the corresponding short-term stability
model are shown in Fig. \ref{my14trywind1_fastmodel-120}. Note that
trajectories starting from the state of the DHM no longer converge
to the SEP of the corresponding short-term stability model, and thus
the state of the DHM lies outside the stability region of the
short-term stability model. So, condition (iii) in \textit{Theorem
2} is violated, and the DHM can fail to provide a satisfactory
approximation for the SHM, which is true in this case. Note that the
condition of consistent attraction is a sufficient but unnecessary
condition to ensure the stability of the DHM (Theorem 5-6
\cite{Wangxz:CAS}). So, the DHM is stable in this case even though
the condition of consistent contraction is violated.
\begin{figure}[!ht]
\centering
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vr1ofExc2_14trywind1_fastmodel_modify-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{vr1ofExc3_14trywind1_fastmodel_modify-eps-converted-to.pdf}
\end{minipage}
\caption{The trajectories of two fast variables in the short-term
stability
model when $\bm{z_d}$ change at 120s. The trajectory starting from the state
of the DHM does not converge to the SEP of the corresponding short-term stability
model. So, the condition of consistent attraction is
violated.}\label{my14trywind1_fastmodel-120}
\end{figure}
From the viewpoint of physical mechanisms, the voltage collapse is
caused by an insufficient power support when LTCs try to restore the
load-side voltages in the long-term time scale. Immediately after
the contingency, the system can maintain the short-term stability by
the control of exciters. After that, LTCs start to work at 30s and
try to restore the load-side voltages and then the corresponding
load powers. At 112s, the OXL at GEN 1 is activated to protect the
generator from overheating and thus restrict the power support from
GEN 1. To make it even worse, the power output of DFIG at Bus 2
suddenly decreases as shown in Fig. \ref{windpower} because of a
sharp drop in the wind speed. The power imbalance between the loads
and the generators finally leads to the voltage collapse in the SHM.
In the DHM, however, the wind power does not change drastically as
shown in Fig. \ref{windpower}, since the wind speed is supposed to
be invariable. So, the DFIG at Bus 2 can provide enough power
required by the action of LTCs to maintain the voltage stability of
the DHM.
\begin{figure}[!ht]
\centering
\includegraphics[width=2.5in,keepaspectratio=true,angle=0]{PofDfig1_14trywind1_modify_1-eps-converted-to.pdf}
\caption{A comparison of the real power output of DFIG in the SHM
and that in the DHM. }\label{windpower}
\end{figure}
From this example, some important physical insights can be obtained.
If the wind power plays a significant role in supporting power to
maintain the stability, for example when the penetration level is
high (8.42\% in this example), then the stochastic properties of the
wind may need to be considered in the stability analysis, especially
when the system operates close to the stability boundary.
\subsection{Numerical Example II}
The second example using an IEEE 9-bus system is presented to reveal
another cause for the failure of the DHM. In the system, {the classical model of GEN is employed.} A
Weibull-distributed wind source drives a DFIG at Bus 3, and three
GENs are equipped with TGs, AVRs, and OXLs, respectively, where the
initial time delays of OXLs are 70s. In addition, three ERLs are
located at Bus 5, 6, and 8, respectively; while three discrete LTCs
are located at Bus 5-4, Bus 9-6, and Bus 2-7, respectively, the
initial time delays of which are 60s and fixed tapping delays are
10s. At 1s, a fault occurs at Bus 6 and is cleared 5 cycles later.
The parameter values are detailed in Appendix \ref{appendix2-2}.
\begin{figure}[!ht]
\centering
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{idrofDfig1_9wind_slow_1-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{VatBus4_9wind_slow_1-eps-converted-to.pdf}
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{VatBus6_9wind_slow_1-eps-converted-to.pdf}
\end{minipage}%
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[width=1.8in ,keepaspectratio=true,angle=0]{xpofExload3_9wind_slow_1-eps-converted-to.pdf}
\end{minipage}
\caption{A comparisons between the trajectories of SHM and those of
DHM. The results of \textit{Theorem 2} do not apply in this case as
condition (i) is violated.}\label{my145wind}
\end{figure}
In this case, the slow manifold $\bm{m_1(\bm{z_c},\bm{z_d})}$ is
unstable, which implies that nearby dynamics will move away from the
slow manifold. As condition (i) in \textit{Theorem 1} and
\textit{Theorem 2} is violated, neither the concentration of sample
path stated in \textit{Theorem 1} nor the trajectory relationship
described in \textit{Theorem 2} holds. The trajectory of the SHM is
not concentrated around that of the DHM, i.e., the DHM cannot
provide an accurate trajectory approximation for the SHM, but both
of them are unstable in the long-term sense.
From the perspective of physical mechanisms, the instability is
caused by the poor control of LTCs which are originally designed to
help maintain the stability. The discrete switching of LTCs makes
the slow manifold jump from the stable component of the constraint
manifold to an unstable component such that the nearby trajectories
move away. The switching events, such as LTCs and shunt
compensation, are adopted commonly as countermeasures against the
voltage instability. But, this example shows that great caution is
necessary when executing those control strategies, because
unexpected stability issues may arise, especially when more wind
power is integrated into the power grid.
\section{Concluding Remarks}\label{sectionconclusion}
{This paper proposes a comprehensive SDE-based
framework for conducting the long-term stability analysis for the
power grid with wind power generations. This framework incorporates
the discrete dynamics induced by various control devices and the
stochastic model of the wind speed with different probability
distributions. To relieve the computational burden, a DHM is
composed and can provide an accurate trajectory approximation and
correct stability assessments for the SHM under some mild sufficient
conditions. Numerical examples are further discussed to show that
the DHM can fail in some critical cases because of a violation of
the proposed sufficient conditions, which complements the proposed
SDE-based framework and also highlights the necessity of the SHM in
the stability analysis, especially if the system operates close to
the stability boundary or experiences a high variability. For the
future work, we plan to extend the present framework to the
stability analysis of power grids with various other uncertainties
and further improve the computational efficiency of the
approximation methodology using the QSS model that integrates
uncertainties.}
\appendices
\section{Proof of Theorem 1}\label{prooftheorem1}
\textit{Proof:} Conditions (i) and (ii) {ensure}
that all conditions of \textit{Theorem 1} \cite{Wangxz:sde1} are
satisfied for each fixed $\bm{z_d}(k)$, $k=0,1,...,N$. So, the
conclusions of \textit{Theorem 1} \cite{Wangxz:sde1} are valid for
each continuous system of the SHM with fixed $\bm{z_d}(k)$. Then,
there exist $\epsilon_0^k>0$, $h_0^k>0$, $\delta_0^k>0$, and a time
$\tilde{\tau}_k$ of order $\epsilon|\mbox{log}h|$ such that whenever
$\delta\leq\delta_0^k$, the following inequality
\begin{eqnarray}\label{theorem2}
&&\mathbb{P}\{\exists \tau\in[\tilde{\tau}_k,\tau_k):
(\bm{z_c}^k(\tau),\bm{\bar{x}}^k(\tau),\bm{z_d}(k))\notin M(h)\}\nonumber\\
&&\leq
C_{n_{z_c},n_x}(\tau,\epsilon)e^{\frac{-h^2}{2\sigma^2}(1-O(h)-O(\epsilon))}
\end{eqnarray}
holds for all $\epsilon\leq\epsilon_0^k$, $h\leq h_0^k$, $k\in[0,1...N]$, on
$[\tilde{\tau}_k,\tau_{k+1})$ for $k\in[0,1,...N-1]$ or on
$[\tilde{\tau}_k,\infty]$ for $k=N$. Here,
$(\bm{z_c}^k(\tau),\bm{\bar{x}}^k(\tau),\bm{z_d}(k))$ is the
solution of each continuous system (\ref{sde power slow})-(\ref{sde
power fast}) of the SHM for fixed $\bm{z_d}(k)$ with initial
condition $(\bm{z_c}^k(0),\bm{\bar{x}}^k(0),\bm{z_d}(k))$.
Let $\epsilon_0=\mbox{min}(\epsilon_0^0,\epsilon_0^1,...\epsilon_0^N)$,
$h_0=\mbox{min}(h_0^0,h_0^1,...h_0^N)$, and
$\delta_0=\mbox{min}(\delta_0^0,\delta_0^1,...\delta_0^N)$. Then,
for
$\tau\in\Pi=\cup_{i=1}^{N-1}[\tilde{\tau}_i,\tau_i)\cup[\tilde{\tau}_N,\infty]$,
the following inequality
\begin{eqnarray}\label{theorem2}
&&\mathbb{P}\{\exists \tau\in \Pi:~ (\bm{z_c}(\tau),\bm{\bar{x}}(\tau),\bm{z_d})\notin M(h)\}\nonumber\\
&&\leq
C_{n_{z_c},n_x}(\tau,\epsilon)e^{\frac{-h^2}{2\sigma^2}(1-O(h)-O(\epsilon))}
\end{eqnarray}
holds for all $\epsilon\leq\epsilon_0$, $h\leq h_0$. This completes the proof.
\section{Proof of Theorem 2}\label{prooftheorem2}
Conditions (i)-(iii) ensure that all conditions of \textit{Theorem
2} \cite{Wangxz:sde1} are satisfied for each fixed $\bm{z_d}(k)$,
$k=0,1,...,N$. So, the conclusions of \textit{Theorem 2}
\cite{Wangxz:sde1} are valid for each continuous system of the SHM
with fixed $\bm{z_d}(k)$. So, there {exist}
$\epsilon_0^k>0$, $\delta_0^k>0$, a time $\tilde{\tau}_k$ of order
$\epsilon|\mbox{log}h|$, and $\bar{\tau}_k$ such that whenever
$\delta\leq\delta_0^k$ for all
$\tau\in[\tilde{\tau}_k,\bar{\tau}_k]$, the following estimates
\begin{eqnarray}
|\bm{\bar{x}}^k(\tau)-\bm{\bar{x}_D}^k(\tau)| &=& O(\sigma),\\
|\bm{z_{c}}^k(\tau)-\bm{z_{cD}}^k(\tau)| &=& O(\sigma\sqrt{\epsilon}),
\end{eqnarray}
hold for all $\epsilon\in(0,\epsilon_0^k)$, $0\le k\le N$, Here, $(
\bm{z_c}^k(\tau), \bm{\bar{x}}^k(\tau),$ $\bm{z_d}(k) )$ is the
solution of each continuous system (\ref{sde power slow})-(\ref{sde
power fast}) of the SHM, and
$(\bm{z_{cD}}^k(\tau),\bm{\bar{x}_D}^k(\tau),\bm{z_d}(k))$ is the
solution of each continuous system (\ref{det power slow})-(\ref{det
power fast}) of the DHM for fixed $\bm{z_d}(k)$.
Let $\epsilon_0=\mbox{min}(\epsilon_0^0,\epsilon_0^1,\cdots,\epsilon_0^N)$,
$h_0=\mbox{min}(h_0^0,h_0^1,\cdots,h_0^N)$, and
$\delta_0=\mbox{min}(\delta_0^0,\delta_0^1,\cdots,\delta_0^N)$.
Similar to \textit{Theorem 1}, one can show that for all
$\tau\in\cup_{i=1}^{N}[\tilde{\tau}_i,\bar{\tau}_i]$, the following
estimates
\begin{eqnarray}
|\bm{\bar{x}}(\tau)-\bm{\bar{x}_D}(\tau)|&=&O(\sigma)\label{corollary3_1}\\
|\bm{z_{c}}(\tau)-\bm{z_{cD}}(\tau)|&=&O(\sigma\sqrt{\epsilon})\label{corollary3_2}
\end{eqnarray}
hold for all $\epsilon\in(0,\epsilon_0)$. The proof of the theorem is
completed.
\section{Parameter Values of Numerical Example I}\label{appendix2-1}
The system is modified from the 14-bus test case in PSAT-2.1.6. The
GEN at Bus 2 is replaced by DFIG. The parameter values are given in
Table \ref{appedixtable1}-\ref{appedixtable5}.
\begin{table}[!ht]
\centering \caption{Doubly-fed induction generator parameter
values}\label{appedixtable1}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
stator resistance $r_s$&$0.01$p.u.\\
\hline
stator reactance $x_s$ &$0.1$p.u.\\
\hline
rotor resistance $r_r$&$0.01$p.u.\\
\hline
rotor reactance $x_r$&$0.08$p.u.\\
\hline
magnetizing reactance $x_\mu$&$3$p.u.\\
\hline
rotor inertia $H_m$&$3$KWs/KVA\\
\hline
pitch control gain $K_p$&$10$\\
\hline
pitch control time constant $T_p$&$3$s\\
\hline
voltage control gain $K_v$&10\\
\hline
power control time constant $T_\epsilon$&$0.01$s\\
\hline
rotor radius $R$&$75$m\\
\hline
number of poles $n_p$&$4$\\
\hline
number of blades $n_b$&$3$\\
\hline
gear box ratio $\eta_{GB}$&$0.0112$\\%$0.01123596$
\hline
maximum active power $p^{max}$&$2$p.u.\\
\hline
minimum active power $p^{min}$&$-1$p.u.\\
\hline
maximum reactive power $q^{max}$&$2$p.u.\\
\hline
minimum reactive power $q^{min}$&$-1$p.u.\\
\hline
number of machines $n_g$&$1$\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Turbine governor parameter
values}\label{appedixtable2}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
reference speed $\omega^{0}_{ref}$&$1$p.u.\\
\hline
droop $R$&$0.02$p.u.\\
\hline
maximum turbine output $p^{max}$&$1.2$p.u.\\
\hline
minimum turbine output $p^{min}$&$0$p.u.\\
\hline
governor time constant $T_s$&$0.1$s\\
\hline
servo time constant $T_c$&$0.45$s\\
\hline
transient gain time constant $T_3$&$0$s\\
\hline
power fraction time constant $T_4$&$12$s\\
\hline
reheat time constant $T_5$&$50$s\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Load tap changer parameter values for the ones
at bus 4-9, bus 12-13 and bus 2-4}\label{appedixtable3}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
the reference voltage $v_0$&$1.005$, $1.01$, $0.995$ \\
\hline
half of the deadband $d$&$0.005$, $0.1$, $0.025$ p.u.\\
\hline
tap step $r$&$0.025$\\
\hline
upper tap limit $r^{max}$&$1.2$\\
\hline
lower tap limit $r^{min}$&$0.7$\\
\hline
the initial time delay $\triangle{T_0}$&$30$s\\
\hline
the sequential time delay $\triangle{T_k}$&$10$s\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Exponential recovery load parameter
values}{\label{appedixtable4}}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
active power percentage $k_p$&$100\%$\\
\hline
reactive power percentage $k_q$&$100\%$\\
\hline
active power time constant $T_p$&$10$s\\
\hline
reactive power time constant $T_q$&$10$s\\
\hline
static active power exponent $\alpha_s$&$1$\\
\hline
dynamic active power exponent $\alpha_t$&$1.5$ for the load at Bus 9\\
&$5$ for the others\\
\hline
static reactive power exponent $\beta_s$&$2$\\
\hline
dynamic reactive power exponent $\beta_t$&$2.5$ for the load at Bus 9\\
&$10$ for the others\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Over excitation limiter parameter
values}\label{appedixtable5}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
maximum field current $i_f^{lim}$&$5.1$p.u.\\
\hline
integrator time constant $T_0$&$12$s\\
\hline
maximum output signal $v_{\mbox{oxl}}$&$100$p.u.\\
\hline
\end{tabular}
\end{table}
\section{Parameter Values of Numerical Example II}\label{appendix2-2}
The system is modified from the 9-bus test system in PSAT-2.1.6.
There is a DIFG at Bus 3. The parameters of the DFIG are the same as
those for Numerical Example I in Table \ref{appedixtable1}. The
parameters of other devices are shown in Table
\ref{appedixtable6}-\ref{appedixtable9}.
\begin{table}[!ht]
\centering \caption{Turbine governor parameter
values}\label{appedixtable6}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
reference speed $\omega^{0}_{ref}$&$1$p.u.\\
\hline
droop $R$&$0.02$p.u.\\
\hline
maximum turbine output $p^{max}$&$2$p.u.\\
\hline
minimum turbine output $p^{min}$&$0.3$p.u.\\
\hline
governor time constant $T_s$&$0.1$s\\
\hline
servo time constant $T_c$&$0.45$s\\
\hline
transient gain time constant $T_3$&$0$s\\
\hline
power fraction time constant $T_4$&$12$s\\
\hline
reheat time constant $T_5$&$50$s\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Exponential recovery load parameter
values}{\label{appedixtable7}}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
active power percentage $k_p$&$40\%$\\
\hline
reactive power percentage $k_q$&$40\%$\\
\hline
active power time constant $T_p$&$10$s\\
\hline
reactive power time constant $T_q$&$10$s\\
\hline
static active power exponent $\alpha_s$&$1$\\
\hline
dynamic active power exponent $\alpha_t$&$10$ for the load at Bus 4\\
&$5$ for the others\\
\hline
static reactive power exponent $\beta_s$&$2$\\
\hline
dynamic reactive power exponent $\beta_t$&$20$ for the load at Bus 4\\
&$10$ for the others\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Load tap changer parameter values for the ones
at bus 5-4, bus 9-6, and bus 2-7}\label{appedixtable8}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
the reference voltage $v_0$&$1.005$, $1.005$, $1.02$ \\
\hline
half of the deadband $d$&$0.025$, $0.025$, $0.04$ p.u.\\
\hline
tap step $r$&$0.12$\\
\hline
upper tap limit $r^{max}$&$1.1$\\
\hline
lower tap limit $r^{min}$&$0.9$\\
\hline
the initial time delay $\triangle{T_0}$&$60$s\\
\hline
the sequential time delay $\triangle{T_k}$&$10$s\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering \caption{Over excitation limiter parameter
values}\label{appedixtable9}
\begin{tabular}{|c|c|}
\hline
Parameter&Value\\
\hline
maximum field current $i_f^{lim}$&$2.02$, $1.3$, $1.32$p.u. \\
\hline
integrator time constant $T_0$&$10$s for GEN 1-2\\
&$30$s for GEN 3\\
\hline
maximum output signal $v_{\mbox{oxl}}$&$100$p.u.\\
\hline
\end{tabular}
\end{table}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
|
\section{Introduction}
\label{s:intro}
The X-ray emission of actively star forming galaxies is dominated by
the collective signal of high-mass X-ray binaries (HMXBs),
complemented by diffuse soft X-ray emission from hot interstellar gas
(e.g. \citealt{lehetal10}). Although the HMXB population of the Milky
Way has been thoroughly studied (see \citealt{waletal15} for a recent
review), the most luminous X-ray binaries are so rare that they can be
found and counted only in other (nearby) galaxies. Such studies have
revealed that the HMXB X-ray luminosity function (XLF) has a
power-law shape: $dN/dL_{\rm{X}}\proptoL_{\rm{X}}^{-1.6}$ from $L_{\rm{X}}\lesssim 10^{36}$
to $\sim 10^{40}$~erg~s$^{-1}$ \citep{minetal12a}, with some indication
of steepening at $L_{\rm{X}}\gtrsim 10^{40}$~erg~s$^{-1}$. The bulk of the
emission from point X-ray sources in actively star forming galaxies is
thus produced by ultraluminous X-ray sources (ULXs) with $L_{\rm{X}}\gtrsim
10^{39}$~erg~s$^{-1}$, the majority of which appear to be
supercritically accreting stellar-mass black holes
(e.g. \citealt{pouetal07,fensor11,robetal16}) and neutron stars
\citep{bacetal14,musetal15,fueetal16,isretal17}.
Luminous HMXBs, including ULXs and so-called ultraluminous supersoft
sources (ULSs), exhibit a large variety of X-ray spectral shapes,
probably due to differences in the accretion rate and orientation of
the thick accretion disc with respect to the observer
(e.g. \citealt{glaetal09,sutetal13,urqsor16}) as well as in the nature
of the accretor (a neutron star vs. a black hole,
\citealt{pinetal17}). In addition, the observed X-ray spectra and
detection rates of such objects can be significantly affected by
photoabsorption in the interstellar medium (ISM) of the host galaxies
and of the Milky Way. Taking all this into account, we have
recently constructed \citep{sazkha17a} the {\sl intrinsic} HMXB XLF in
its bright end, $10^{38}<L_{{\rm X,unabs}}\lesssim 10^{40.5}$~erg~s$^{-1}$ (where
$L_{{\rm X,unabs}}$ is the absorption corrected source luminosity in the
0.25--8~keV energy band), per unit star formation rate (SFR). It can
be described by a power law, $dN/d\logL_{{\rm X,unabs}}\approx
2.0(L_{{\rm X,unabs}}/10^{39}\,{\rm
erg~s}^{-1})^{-0.6}$~$(M_\odot$~yr$^{-1})^{-1}$, which has the same
slope as the {\sl observed} HMXB XLF of \citet{minetal12a} but a factor
of $\sim 2.3$ higher normalization. We further showed that about two
thirds of the total X-ray (0.25--8~keV) emission of HMXBs is released
in the soft band (0.25--2~keV), $\sim 5\times
10^{39}$~erg~s$^{-1}$~$(M_\odot$~yr$^{-1})^{-1}$, with roughly equal
contributions from 'hard', 'soft' and 'supersoft' sources, defined
according to their intrinsic soft/total X-ray flux ratio:
\begin{eqnarray}
{\rm Hard:}& F_{0.25-2,{\rm unabs}}/F_{0.25-8,{\rm unabs}}\le 0.6,\nonumber\\
{\rm Soft:}& 0.6<F_{0.25-2,{\rm unabs}}/F_{0.25-8,{\rm unabs}}\le 0.95,\nonumber\\
{\rm Supersoft:}& F_{0.25-2,{\rm unabs}}/F_{0.25-8,{\rm unabs}}>0.95.
\label{eq:types}
\end{eqnarray}
This detailed information about the intrinsic XLF provides
interesting constraints on the population properties of HMXBs and the
physics of near- and super-critical accretion. It may also be
interesting in the context of studying the 'cosmic dawn', since HMXBs
belonging to the first generations of stars and their remnants might
have radiatively preheated the Universe before it was reionized by
UV radiation from stars and quasars (e.g. \citealt{miretal11}). In
our other recent paper \citep{sazkha17b}, we demonstrated (see also
\citealt{madfra16}) that HMXBs
could significantly heat the Universe at $z\sim 10$ if the specific
(i.e. per unit SFR) X-ray emissivity of such systems was higher by an
order of magnitude than at the present epoch and the soft X-rays
produced by HMXBs could escape from their host galaxies without strong
attenuation in their ISM. Whether or not these conditions were
fulfilled is an open question.
In \citet{sazkha17b} we used the measured ratio of the HMXB luminosity
functions in the 0.25--2 and 0.25--8~keV bands to
estimate the effective photon index of the average intrinsic X-ray
spectrum of luminous HMXBs: $\Gamma\sim 2.1$. In the present study, we
take advantage of the same sample of sources to construct the
intrinsic (i.e. corrected for observational biases), SFR-normalized
energy spectrum of the integrated emission of luminous HMXBs in the
local Universe, hereafter referred to as the {\sl intrinsic collective
spectrum of HMXBs}, and evaluate the contributions of hard, soft and
supersoft sources to it. This spectrum may find application, in
particular, in simulations of the preheating of the early Universe by
X-ray binaries.
\section{Sample}
\label{s:sample}
We make use of the 'clean sample' of X-ray sources detected by the
{\sl Chandra} X-ray Observatory \citep{wanetal16}, presumably located
in 27 nearby ($D<15$~Mpc) galaxies (mostly spirals), from
\citet{sazkha17a}. This sample had been compiled based on the following
criteria: i) the source must be located on the sky within the 25 mag
arcsec$^{-2}$ isophote of the galaxy, i.e. at radius $R<R_{25}$ in the
plane of the galaxy, but outside of its central $0.05 R_{25}$ region,
ii) there must be at least 100 photon counts from the source in some
{\sl Chandra} observation and iii) the unabsorbed
0.25--8~keV luminosity of the source must exceed
$10^{38}$~erg~s$^{-1}$. Having additionally filtered out 16 known or
suspected background active galactic nuclei (AGN) and 3 foreground
Galactic stars, we had selected 200 HMXB candidates with
luminosities ranging from $L_{{\rm X,unabs}}=10^{38}$ to $\sim 3\times
10^{40}$~erg~s$^{-1}$, and analyzed their {\sl Chandra} spectra.
\section{Analysis}
\label{s:analysis}
In \citet{sazkha17a}, we described the measured spectrum of each
source by one of the following models: i) absorbed power law, ii)
absorbed blackbody emission and iii) absorbed mutlicolour blackbody
disc emission. These best-fitting models were then used to determine
the sources' intrinsic and observed luminosities in the 0.25--8~keV
and 0.25--2~keV energy bands ($L_{{\rm X,unabs}}$, $L_{{\rm X,obs}}$, $L_{{\rm SX,unabs}}$ and
$L_{{\rm SX,obs}}$, respectively). We now use the same spectral
fits to determine the intrinsic and observed luminosities
($L_{i,{\rm unabs}}$ and $L_{i,{\rm obs}}$, respectively) in 5 narrow subbands:
0.25--0.5, 0.5--1, 1--2, 2--4, and 4--8~keV (hereafter referred to as
bands $i=1$ to 5).
Using high-quality maps of atomic (HI) and molecular (H$_2$)
gas in the sampled galaxies, we demonstrated in \citet{sazkha17a}
that the line-of-sight absorption column densities, $N_{\rm{H}}$, inferred
from the spectra of the studied X-ray sources (typically a few
$10^{21}$~cm$^{-2}$) can be attributed to the ISM of
their host galaxies. We further took advantage of the HI,
H$_2$ and SFR maps of the sampled galaxies to evaluate observational
biases associated with the detection of X-ray sources by {\sl
Chandra}, which arise due to intrinsic diversity of HMXB
spectra and X-ray absorption in the ISM. In a nutshell (see
\citealt{sazkha17a} for a detailed discussion), in the absence of
intervening absorbing gas, a soft source with a given intrinsic
luminosity $L_{{\rm X,unabs}}$ would produce more photon counts on the
detector than a hard source with the same luminosity and
location. On the other hand, observed X-ray fluxes of soft sources are more
strongly affected by photoabsorption in the ISM, so that such
sources may become hidden from {\sl Chandra} if located on the
farther side of their host galaxy. As a result, only some fraction of the
total SFR ($\sim 32$~$M_\odot$~yr$^{-1}$) in the 27 sampled galaxies
is effectively probed by {\sl Chandra} in X-rays, and this
fraction depends on the source intrinsic luminosity and spectral
type, $T$: ${\rm SFR~}(L_{{\rm X,unabs}},T)$ \citep{sazkha17a}.
In \citet{sazkha17a}, we used the ${\rm SFR~}(L_{{\rm X,unabs}},T)$
dependence to derive the HMXB XLF per unit SFR. We can now apply a
similar procedure to constuct the intrinsic collective spectrum of
HMXBs in the local Universe by summation over the sources in the sample:
\begin{equation}
S_{i,{\rm unabs}}=C_{{\rm var}}\sum_j[1-f_{{\rm LMXB}}(L_{{\rm X,unabs},j})]\frac{L_{i,{\rm unabs},j}}{{\rm
SFR~}(L_{{\rm X,unabs},j},T_j)}.
\label{eq:avspec_intr}
\end{equation}
Here, $S_{i,{\rm unabs}}$ is the HMXB emissivity [measured in units of
erg~s$^{-1}$~($M_\odot$~yr$^{-1}$)$^{-1}$] in energy band $i$ (from
1 to 5), $L_{i,{\rm unabs},j}$ is the intrinsic luminosity in band $i$ of the
$j$th source, $L_{{\rm X,unabs},j}$ is its intrinsic 0.25--8~keV luminosity and $T_j$
is its spectral type.
By means of equation~(\ref{eq:avspec_intr}), we perform a weighted
stacking of the source spectra, such that each {\sl Chandra} source
provides a contribution equal to its luminosity in a given energy band
divided by the corresponding X-ray-probed SFR. This procedure is
analogous to the $1/V_{\rm max}$ weighting method frequently used in
astronomy, with ${\rm SFR}$ playing the role of a generalized $V_{\rm
max}$. Similar procedures have been used before e.g. for estimating
the space density of stellar objects in the Galaxy taking into
account their inhomogeneous spatial distribution
(e.g. \citealt{tinetal93,sazetal06}). Our current treatment also
closely follows the calculation of the collective X-ray spectrum of
AGN in the local Universe by \citet{sazetal08}.
We have added two additional factors in
equation~(\ref{eq:avspec_intr}). The factor $1-f_{{\rm LMXB}}(L_{{\rm X,unabs},j})$ takes
into account that our sample of HMXB candidates is expected to be
contaminated by low-mass X-ray binaries (LMXBs). Their relative
contribution as a function of luminosity was estimated in
\citet{sazkha17a} based on the LMXB XLF \citep{gilfanov04} and can be
approximated as
\begin{eqnarray}
f_{{\rm LMXB}} =
\left\{
\begin{array}{ll}
0.49, & 10^{38}\leL_{{\rm X,unabs}}<10^{38.5}~{\rm erg~s}^{-1},\\
0.41, & 10^{38.5}\leL_{{\rm X,unabs}}<10^{39}~{\rm erg~s}^{-1},\\
0.06, & 10^{39}\leL_{{\rm X,unabs}}<10^{39.5}~{\rm erg~s}^{-1},\\
0, & L_{{\rm X,unabs}}\ge10^{39.5}~{\rm erg~s}^{-1}.
\end{array}
\right.
\label{eq:lmxb}
\end{eqnarray}
The LMXB contribution is thus substantial below $10^{39}$~erg~s$^{-1}$
but negligible at higher luminosities. Therefore, since most of the
emission from HMXBs is produced by ULXs (with
$L_{{\rm X,unabs}}>10^{39}$~erg~s$^{-1}$), the significant uncertainty in our
knowledge of the LMXB XLF and hence their contribution to our sample
does not translate into a significant uncertainty in the resulting
collective spectrum of HMXBs.
The additional coefficient $C_{{\rm var}}=1/1.2\approx 0.83$ in
equation~(\ref{eq:avspec_intr}) takes into account the 'variability
bias' evaluated in \citet{sazkha17a}. It results from our using
particular {\sl Chandra} observations for estimating the luminosities
of the sources (namely those with at least 100 counts from the source)
while the same sources would be weaker by $\sim 20$\% on average due
to intrinsic variability if their luminosities were measured randomly
in time.
There are two types of uncertainties associated with the
collective spectrum $S_{i,{\rm unabs}}$. One is due to uncertainties,
$\deltaL_{i,{\rm unabs},j}$, in estimation of the unabsorbed luminosities of the
inidividual sources from X-ray spectral analysis:
\begin{equation}
\delta_{i,1}=C_{{\rm var}}\sqrt{\sum_j\left([1-f_{{\rm LMXB}}(L_{{\rm X,unabs},j})]\frac{\deltaL_{i,{\rm unabs},j}}{{\rm
SFR~}(L_{{\rm X,unabs},j},T_j)}\right)^2}.
\label{eq:avspec_unc1}
\end{equation}
Another arises from the finite size of our source sample:
\begin{equation}
\delta_{i.2}=C_{{\rm var}}\sqrt{\sum_j\left([1-f_{{\rm LMXB}}(L_{{\rm X,unabs},j})]\frac{L_{i,{\rm unabs},j}}{{\rm
SFR~}(L_{{\rm X,unabs},j},T_j)}\right)^2}.
\label{eq:avspec_unc2}
\end{equation}
This uncertainty stems from the fact that in the standard $\sum
1/V_{{\rm max},j}$ estimation of space densities, the uncertainty of
each object's contribution is assumed to follow Poisson statistics
\citep{feletal76} so that the variance of the density estimate is
$\sum 1/V^2_{{\rm max},j}$ (e.g. \citealt{tinetal93}). In our case,
the contributions of individual sources to the variance of $S_{i,{\rm unabs}}$
are similarly independent of each other but must be multiplied by the
square of the corresponding coefficients in
equation~(\ref{eq:avspec_intr}).
The total uncertainty can be estimated as a combination of these
uncertainties:
\begin{equation}
\delta_i=\sqrt{\delta_{i,1}^2+\delta_{i,2}^2}.
\label{eq:avspec_unc}
\end{equation}
We can use a slightly modified stacking procedure to also compute the
{\sl observed} collective spectrum of HMXBs in the
local Universe:
\begin{equation}
S_{i,{\rm obs}}=C_{{\rm var}}\sum_j[1-f_{{\rm LMXB}}(L_{{\rm X,unabs},j})]\frac{L_{i,{\rm obs},j}}{{\rm
SFR~}(L_{{\rm X,unabs},j},T_j)}.
\label{eq:avspec_obs}
\end{equation}
This spectrum represents the integrated X-ray emission of HMXBs as
seen by the Earth's observer, i.e. uncorrected for line-of-sight
absorption. In this case, the uncertainties of the first type are
negligible due to the small $\deltaL_{i,{\rm obs},j}$ errors and those of the
second type can be computed using equation~(\ref{eq:avspec_unc2}) by
substituting $L_{i,{\rm obs},j}$ for $L_{i,{\rm unabs},j}$.
\section{Results}
\label{s:results}
\subsection{Intrinsic spectrum}
\label{s:avspec_intr}
Figure~\ref{fig:avspec_lumranges} shows the {\sl intrinsic}
collective spectrum of HMXBs with $10^{38}<L_{{\rm X,unabs}}\lesssim
10^{40.5}$~erg~s$^{-1}$ (the luminosity range spanned by our sample
of sources), obtained using
equation~(\ref{eq:avspec_intr}). It can be well fitted ($\chi^2=0.21$
for 3 degrees of freedom) by a power law:
\begin{equation}
\frac{EL_E}{\rm SFR}=(2.1\pm 0.4)\times 10^{39} \left(\frac{E}{{\rm
keV}}\right)^{-0.11\pm 0.18}\,{\rm erg~s}^{-1}\,(M_\odot~{\rm yr})^{-1}.
\label{eq:fit_l38_40.5}
\end{equation}
The quoted uncertainty for the spectral slope may be slightly
overestimated because we regard the uncertainties $\delta_i$ of
the individual spectral points as independent, although they may be
somewhat correlated due to the contribution $\delta_{i,2}$
[eq.~(\ref{eq:avspec_unc2})] from the Poisson uncertainty in the
number of sampled sources (this is only important for the three higher
energy channels, since the uncertainties in the 0.25--0.5~keV and
0.5--1~keV bands are dominated by the $\delta_{i,1}$ errors associated
with luminosity estimation for individual sources).
\begin{figure}
\centering
\includegraphics[width=\columnwidth,viewport=30 200 560
710]{avspec_l38_405_lumranges.pdf}
\caption{Intrinsic collective spectrum of HMXBs in the local
Universe (black circles) and the contributions
(these points are slightly shifted along the horizontal axis for
better visibility) of sources in three luminosity ranges:
$L_{{\rm X,unabs}}=10^{40}$--$10^{40.5}$~erg~s$^{-1}$ (red squares),
$10^{39}$--$10^{40}$~erg~s$^{-1}$ (green triangles) and
$10^{38}$--$10^{39}$~erg~s$^{-1}$ (blue stars). The solid line is
the best-fitting power law for the total spectrum
[eq.~(\ref{eq:fit_l38_40.5})].
}
\label{fig:avspec_lumranges}
\end{figure}
As could be expected from the HMXB XLF \citep{sazkha17a},
the bulk of the emission from HMXBs is provided by
ultraluminous ($L_{{\rm X,unabs}}>10^{39}$~erg~s$^{-1}$) sources (see
Fig.~\ref{fig:avspec_lumranges}). Moreover, there is an indication
that a sizeable or even dominant contribution is provided by extremely
luminous sources with $L_{{\rm X,unabs}}>10^{40}$~erg~s$^{-1}$. These results
are unlikely to be strongly affected by LMXB contamination of our
sample of sources, which we have roughy taken into account through the
$1-f_{{\rm LMXB}}$ factor in equation~(\ref{eq:avspec_intr}). Indeed, LMXBs
are only important at $L_{{\rm X,unabs}}<10^{39}$~erg~s$^{-1}$, but the overall
contribution of such relatively low-luminosity sources to the total
emission from HMXBs is small, while nearly all of our
$L_{{\rm X,unabs}}>10^{39}$~erg~s$^{-1}$ sources are expected to be HMXBs.
The large uncertainties of the collective spectrum of HMXBs in
the two softest
bands are mainly associated with the presence of two very luminous
($L_{{\rm X,unabs}}\gtrsim 10^{40}$~erg~s$^{-1}$) supersoft [per our definition,
eq.~(\ref{eq:types})] sources in our sample. As discussed in
\citet{sazkha17a}, these sources have very soft spectra (which can be
described as blackbody radiation with $kT_{\rm bb}\sim
0.06$--0.07~keV) and their inferred intrinsic luminosities are some 3
orders of magnitude higher than their observed luminosities
(apparently due to the presence of significant amounts of cold ISM in
their direction) but very uncertain (by 1--2 orders of
magnitude). Moreover, there are in total only 7 sources (2 hard, 3
soft and 2 supersoft ones) with $L_{{\rm X,unabs}}>10^{40}$~erg~s$^{-1}$ in our
sample, so that the overall contribution of such luminous sources to
the total X-ray emission produced by HMXBs is not well
constrained by the present study.
Given the large uncertainty associated with the contribution of the
most luminous sources ($L_{{\rm X,unabs}}>10^{40}$~erg~s$^{-1}$), we also
calculated the collective spectrum of HMXBs with
$10^{38}<L_{{\rm X,unabs}}<10^{40}$~erg~s$^{-1}$ (there are in total 193 such
objects in our sample), which is shown in
Fig.~\ref{fig:avspec_spectypes_fits}. This spectrum is tightly
constrained and can be well fitted ($\chi^2=0.71$ for 3 degrees of
freedom) by the following power law:
\begin{equation}
\frac{EL_E}{\rm SFR}=(1.31\pm 0.13)\times 10^{39} \left(\frac{E}{{\rm
keV}}\right)^{-0.08\pm 0.11}\,{\rm erg~s}^{-1} (M_\odot~{\rm yr})^{-1}.
\label{eq:fit_l38_40}
\end{equation}
Comparing this expression with equation~(\ref{eq:fit_l38_40.5}) we see
that the slope is unchanged, but the normalization has decreased
by $\sim 20$--50\%, which reflects the substantial contribution of the
most luminous ($L_{{\rm X,unabs}}\gtrsim 10^{40}$~erg~s$^{-1}$) sources to the
total X-ray emission from HMXBs. The derived spectral slope
(photon index) $\Gamma\approx 2.1$ confirms the conclusion of our
previous work \citep{sazkha17a,sazkha17b} that about two thirds of the
total X-ray output of HXMBs emerges in the form of soft X-rays, at
energies below 2~keV.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,viewport=30 200 560 710]{avspec_l38_40_spectypes_fits.pdf}
\caption{Intrinsic collective spectrum of HMXBs with
$10^{38}<L_{{\rm X,unabs}}<10^{40}$~erg~s$^{-1}$ per unit SFR (black circles)
and the contributions of hard, soft and supersoft sources
[according to our definition, eq.~(\ref{eq:types})]: squares, triangles
and stars, respectively. The black solid line shows the best-fitting
power law for the total spectrum [eq.~(\ref{eq:fit_l38_40})]. The
red dotted, green dashed and blue dash-dotted lines
show the corresponding fits for the hard, soft and supersoft
components [eqs.~(\ref{eq:fit_hard_l38_40}),
(\ref{eq:fit_soft_l38_40}) and (\ref{eq:fit_supersoft_l38_40}),
respectively].
}
\label{fig:avspec_spectypes_fits}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth,viewport=30 200 560 710]{avspec_l38_40_spectypes_examples_new.pdf}
\caption{Same as Fig.~\ref{fig:avspec_spectypes_fits}, but the
power-law fits for hard, soft and supersoft sources are replaced by
examples of spectra (best-fitting unabsorbed models adopted from the
literature, see main text, with ad hoc normalizations) of real
individual sources belonging to tese catagories: red dotted line --
ULX~Ho~IX~X-1, green dashed line -- ULX~NGC~5408~X-1, blue
dash-dotted lines -- NGC~247~ULS in two different states (see
references in the text). None of
these sources belong to our sample.
}
\label{fig:avspec_spectypes_examples}
\end{figure}
Although the shape of the collective spectrum of HMXBs is
consistent with a simple power law, this is merely the result of
summing over a great variety of individual spectra. In reality, as
shown in Fig.~\ref{fig:avspec_spectypes_fits}, hard, soft and
supersoft sources (again, per our definition) provide comparable
contributions to the total spectrum, and it is the soft and supersoft
sources that are largely responsible for its low-energy part. The
presented collective spectra of the hard, soft and supersoft
sources have been obtained using the same stacking procedure
[eq.~(\ref{eq:avspec_intr})] as for the total spectrum, applied to
117, 47 and 29 sources of these classes, respectively (excluding the 7
sources with $L_{{\rm X,unabs}}>10^{40}$~erg~s$^{-1}$). These spectra can be
well fitted ($\chi^2=2.79$, 1.06 and 0.63, respectively, for 3 degrees
of freedom) by the following power laws:
\begin{equation}
\left(\frac{EL_E}{\rm SFR}\right)_{{\rm hard}}=(0.47\pm 0.05)\times 10^{39} \left(\frac{E}{{\rm
keV}}\right)^{0.51\pm 0.10}\,{\rm erg~s}^{-1} (M_\odot~{\rm yr})^{-1}.
\label{eq:fit_hard_l38_40}
\end{equation}
\begin{equation}
\left(\frac{EL_E}{\rm SFR}\right)_{{\rm soft}}=(0.37\pm 0.06)\times 10^{39} \left(\frac{E}{{\rm
keV}}\right)^{-0.37\pm 0.16}\,{\rm erg~s}^{-1} (M_\odot~{\rm yr})^{-1}.
\label{eq:fit_soft_l38_40}
\end{equation}
\begin{equation}
\left(\frac{EL_E}{\rm SFR}\right)_{{\rm supersoft}}=(0.22\pm 0.09)\times 10^{39} \left(\frac{E}{{\rm
keV}}\right)^{-1.72\pm 0.21}\,{\rm erg~s}^{-1} (M_\odot~{\rm yr})^{-1}.
\label{eq:fit_supersoft_l38_40}
\end{equation}
In reality our partition of HMXBs into three classes is ad hoc, and
each of these groups exhibits significant diversity of individual
source spectra (see the best-fitting spectral parameters for our
sources in \citealt{sazkha17a}). Nevertheless, the collective
spectra for these classes allow for some generalizing
description.
First, the collective spectrum of hard sources
resembles typical, hard spectra of
ULXs in the so-called 'broadened-disc' and 'hard ultraluminous' states
introduced by \citet{sutetal13}. To demonstrate this, we show in
Fig.~\ref{fig:avspec_spectypes_examples} the best-fitting (unabsorbed) model
(\citealt{sazetal14}, their table 2) for an X-ray spectrum of the
well-known ULX Ho IX~X-1 taken by the {\sl XMM-Netwon} observatory
(observation 0657801801), which consists of i) a hard ($\Gamma=1.28$)
power-law component with a high-energy exponential cutoff at $E_{\rm
cut}=6.4$~keV and ii) a weak additional, multicolour blackbody disc
emission component with $kT_{\rm in}=0.3$~keV ({\sl cutoffpl}+{\sl diskbb}
in {\sc xspec}, \citealt{arnaud96}). We see that this spectrum nearly
matches our collective spectrum of hard sources.
Also shown in Fig.~\ref{fig:avspec_spectypes_examples} is the best-fitting
model (\citealt{sutetal13}, their table A1) for an {\sl XMM-Newton}
spectrum of NGC~5408~X-1, which according to the classification scheme
of these authors is a represenatative of the so-called 'soft
ultraluminous' ULX spectral class. In this case, the spectrum consists
of a fairly steep power-law component with $\Gamma=2.56$ and a
multicolour blackbody component with $kT_{\rm in}=0.194$~keV ({\sl
powerlaw}+{\sl diskbb}). This spectrum is fairly similar, although
not an exact match, to our collective spectrum of soft sources.
As for the spectra of our supersoft sources, most of them can be
described in terms of blackbody emission with $kT_{\rm{bb}}$ ranging from
$\sim 0.05$ to $\sim 0.25$~keV or multicolour disc emission with
$kT_{\rm{in}}$ ranging from $\sim 0.1$ to $\sim 0.3$~keV, although the
spectra of 5 supersoft sources are somewhat better described by a power law
with $\Gamma\sim 3.4$--3.8 (see \citealt{sazkha17a}). Hence, the
softer spectra in this category are similar to typical ULS spectra
\citep{diskon03,urqsor16} while the harder ones resemble the spectra
of 'normal' X-ray binaries in high/soft states associated with high
but subcritical accretion rates (see \citealt{donetal07} for a
review). Among our lowest luminosity ($10^{38}<L_{{\rm X,unabs}}\lesssim
2\times 10^{38}$~erg~s$^{-1}$) supersoft sources there may also be
present classical supersoft sources associated with accreting white
dwarfs (e.g. \citealt{soretal16}), but such objects are not expected
to provide a significant contribution to the collective spectrum of
HMXBs, according to the HMXB XLF obtained in \citet{sazkha17a}.
Some or most of the harder spectra in the supersoft
category may correspond
to intermediate states between the supersoft ultraluminous state
typical of ULSs and the soft ultraluminous state occuring in ULXs (see
above). In fact, there is growing evidence that ultraluminous
sources can make transitions between these states. One example of such
behaviour is shown in Fig.~\ref{fig:avspec_spectypes_examples} based
on the study by \citet{fenetal16}: NGC~247~ULS has
been observed by {\sl Chandra} and {\sl XMM-Newton} to switch between
i) a 'low', supersoft ultraluminous state, when its spectrum is
dominated by soft thermal emission ($kT_{\rm in}=0.11$~keV) but
exhibits an additional, weak power-law component ($\Gamma=2.1\pm
0.9$, here we use the parameters for the {\sl diskbb}+{\sl powerlaw}
model from table~3 of \citealt{fenetal16}), which dominates above
$\sim 2$~keV, and ii) a 'high' state, when the thermal component is
sowewhat harder ($kT_{\rm in}=0.15$~keV) and the power-law
($\Gamma=3.9\pm 0.4$) component provides a larger contribution to
the X-ray luminosity\footnote{We have neglected an additional, weak
component representing thermal emission from an optically thin
plasma in the best-fitting model of \citet{fenetal16} for
NGC~247~ULS in its high state.}. The latter state appears to be
intermediate between the supersoft and
soft ultraluminous states. A similar spectral transition has been
observed in the well-known ULS in M101 \citep{sorkon16}. We see from
Fig.~\ref{fig:avspec_spectypes_examples} that our collective
spectrum of supersoft sources may well be a superposition of
spectra corresponding to different states of ULSs.
We conclude that the collective spectrum of HMXBs can be described in
terms of a superposition of different (known) spectral states of near-
and super-critically accreting X-ray binaries, which probably reflect
differences in the accretion rate and inclination of the accretion
disc with respect to the observer (see a further discussion in
\S\ref{s:discuss} below).
\subsection{Observed spectrum}
\label{s:avspec_obs}
Figure~\ref{fig:avspec_obs} shows the {\sl observed} collective
spectrum of HMXBs, obtained using
equation~(\ref{eq:avspec_obs}) by stacking the weighted spectra of all
200 sources in the sample
($10^{38}<L_{{\rm X,unabs}}<10^{40.5}$~erg~s$^{-1}$). As demonstrated in the
figure, the observed spectrum is dominated by hard sources, with the
contributions of soft and especially supersoft sources being
significantly suppressed compared to the intrinsic spectrum as a
result (mainly) of attenuation of their emission in the ISM of their host
galaxies (see \citealt{sazkha17a}). This effect is only noticeable
below 2~keV. The observed spectrum can be
approximately described as the intrinsic spectrum [a power
law with $\Gamma=2.1$, see eqs.~(\ref{eq:fit_l38_40.5},
(\ref{eq:fit_l38_40})] absorbed in cold gas with column density
$N_{\rm{H}}=(1.2\pm 0.2)\times 10^{21}$~cm$^{-2}$ (see
Fig.~\ref{fig:avspec_obs}). This value is very close to the median
absorption column of $1.1\times 10^{21}$~cm$^{-2}$ for our sample of
sources, inferred from their X-ray spectra \citep{sazkha17a}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,viewport=30 200 560 710]{avspec_obs_l38_405_spectypes_galaxy.pdf}
\caption{Observed collective spectrum of HMXBs
($10^{38}<L_{{\rm X,unabs}}<10^{40.5}$~erg~s$^{-1}$) (black
circles) and the contributions of hard, soft and supersoft sources
(squares, triangles and stars, respectively). The black solid line
shows the best-fitting absorbed power-law model for the total
spectrum, with a photon index fixed at $\Gamma=2.1$ (as inferred for
the intrinsic HMXB spectrum) and $N_{\rm{H}}=1.2\times
10^{21}$~cm$^{-2}$. The magenta dotted line shows the observed
SFR-normalized X-ray spectrum of the NGC~3256 starburst galaxy
\citep{lehetal15}. The black dashed line shows our intrinsic
collective spectrum of HMXBs absorbed by $N_{\rm{H}}=5\times
10^{21}$~cm$^{-2}$ of cold gas, which is a crude estimate (based on
the atomic and melecular gas content of this galaxy, see text) of
the typical ISM column density towards the HMXBs in NGC~3256.
}
\label{fig:avspec_obs}
\end{figure}
It is interesting to compare the observed collective spectrum of
HMXBs constructed here with observed galaxy-wide
X-ray spectra of actively star forming galaxies. Several such
spectra, measured with {\sl Chandra} and {\sl NuSTAR}, have been
presented by \citet{lehetal15}, of which the most interesting is that
of the starburst galaxy NGC~3256. This galaxy has a very high total
SFR of $\sim 36$~$M_\odot$~yr$^{-1}$, which is similar to the combined
SFR of all 27 galaxies making up our sample ($\sim
32$~$M_\odot$~yr$^{-1}$). Therefore, NGC~3256 should be as
representative of the local luminous HMXB population as our sample of
galaxies.
As shown in Fig.~\ref{fig:avspec_obs}, the SFR-normalized spectrum of
NGC~3256 nearly matches our observed collective spectrum of HMXBs at
energies above $\sim 3$~keV, confirming that the total emission of actively
star forming galaxies at these energies is dominated by ULXs. It is
more difficult to compare the collective spectrum of HMXBs and
the NGC~3256 spectrum below 3~keV. First, as discussed by
\citet{lehetal15} and evident from the strong X-ray line emission
observed from NGC~3256, thermal emission from hot interstellar gas
provides a strong
contribution to the soft X-ray flux from this galaxy. Secondly, soft
X-ray emission from point sources in NGC~3256 can be significantly
suppressed by absorption in the cold component of its ISM, in fact
much stronger than for typical galaxies in our sample, which are
characterized by much lower SFRs (at most $\sim
3$~$M_\odot$~yr$^{-1}$) and smaller amounts of cold gas.
We can roughly estimate the expected absorption column density for the
X-ray sources in NGC~3256 using measurements of its total gas content
and assuming that the gas is uniformly distributed over a disc of some
characteristic radius $R$. For the atomic gas, the total mass is
estimated as $M_{\rm HI}\sim6\times 10^{9}~M_{\odot}$, with $R_{\rm
HI}\sim 30$~kpc (e.g. \citealt{casetal04}), while for the molecular
gas $M_{\rm H_2}\sim5\times 10^{9}~M_{\odot}$ with $R_{\rm H_2}\sim 6.6$~kpc
\citep{uedetal14}. This yields integrated column densities
(perpendicular to the plane of the galaxy) of $N_{\rm{H}}\sim 3 \times
10^{20}$ and $\sim 5\times10^{21}$~cm$^{-2}$ for the atomic and
molecular gas, respectively. Taking into account that NGC 3256 is
inclined at $\sim 48^\circ$ (according to
HyperLeda\footnote{http://leda.univ-lyon1.fr/}) to our line of sight
and the Galactic absorption of $\sim 7\times10^{20}$~cm$^{-2}$ in its
direction \citep{kaletal05}, we infer that X-ray sources in this
galaxy should typically be screened from us by $N_{\rm{H}}\sim 5\times
10^{21} $~cm$^{-2}$ of cold gas. Subjecting our intrinsic
($\Gamma=2.1$) collective spectrum of HMXBs to this amount of absorption
results in a spectrum shown in Fig.~\ref{fig:avspec_obs}. We see that
most of the soft X-ray emission produced by the HMXBs in NGC~3256 can
be obscured by the ISM. In reality, HMXBs are usually concentrated to
regions of active star formation and enhanced gas column density, so
the ensemble-averaged $N_{\rm{H}}$ for the HMXB population of NGC~3256 can be
even higher than in our estimate.
Comparison of Fig.~\ref{fig:avspec_lumranges} and
Fig.~\ref{fig:avspec_obs} suggests that the (mostly obscured)
population of luminous soft HMXBs in NGC~3256 probably produces a
similar amount of soft X-rays as its hot interstellar gas. This
is in agreement with the conclusion reached by
\citet{minetal12a,minetal12b}, who compared the integrated
contributions of point X-ray sources and hot ISM to the galaxy-wide
X-ray luminosity for two dosens of nearby galaxies and found both
contributions to be similar and proportional to the SFR,
albeit within a large uncertainty associated with cold-gas
absorption of soft X-rays emitted by hot gas. According to the
linear relation found by \citet{minetal12b}, the
intrinsic X-ray (0.3--10~keV) luminosity of the ISM of a galaxy with
a given SFR is expected to be $\sim 7\times 10^{39}\,({\rm
SFR}/M_\odot~\rm{yr}^{-1})$~erg~s$^{-1}$. Taking into account the
substantial intrinsic absorption in NGC~3256 (see the discussion
above) it appears from Fig.~\ref{fig:avspec_obs} that NGC~3256
is consistent with the \citet{minetal12b} correlation.
The apparent approximate parity between the total X-ray outputs of the
luminous HMXB population and hot ISM in star forming galaxies is
interesting and should be further studied in future work.
\section{Discussion}
\label{s:discuss}
Although the exact nature of the sources comprising our sample is
unknown, we have demonstrated that the intrinsic collective
spectrum of HMXBs can be described in terms of a superposition of
various known spectral states of luminous X-ray binaries: the
broadened-disc, hard ultraluminous and soft ultraluminous states known
for ULXs \citep{sutetal13}, the very soft ($T_{\rm{bb}}\sim 0.1$~keV)
blackbody-like state typical of ULSs \citep{urqsor16} and the high/soft
states of 'normal' X-ray binaries \citep{donetal07}. All these states
may be different manifestations of near- or super-critical accretion
of matter from a massive stellar companion onto a stellar-mass black
hole (or a neutron star in some systems), reflecting differences in
the accretion rate and/or in the orientation of the (thick) accretion
disc and its wind with respect to the observer
(e.g. \citealt{midetal15,fenetal16,guetal16,urqsor16}). The basic idea
discussed in these recent papers is that when the disc is observed
nearly face-on, (relatively) hard X-ray radiation from the central
funnel is directly visible. However, the central emission region can be
obscured by the wind from an observer viewing the disc at larger
inclination, so that only reprocessed softer emission will be visible.
\citet{kawetal12} have performed a detailed modelling of X-ray
spectra generated by supercritical accretion onto a stellar-mass black
hole, combining the results of hydrodynamical simulations of a
thick accretion flow with Monte-Carlo radiative transfer in this
flow. The findings of this work are in good agreement with the general
picture adopted in the aforementioned studies, namely the appearance
of a supercritical accretor should strongly depend on the viewing
angle: the source will appear more luminous and harder if observed
face-on and weaker and softer if observed at a large angle. According
to the angular dependence of the observed X-ray luminosity (for a given
accretion rate) obtained by \citet{kawetal12} (see their fig.~3),
objects viewed at intermediate angles of $i\sim 10$--$40^\circ$ should
dominate in the collective X-ray emission of the local population of
supercritical accretors (assuming, of course, that they are randomly
oriented). Therefore, the angle-integrated spectrum of such sources
should be somewhat softer than the spectra of face-on ($i=0$) objects,
namely it is expected to have an effective photon index of $\Gamma\sim
2$ according to fig.~4 in \citet{kawetal12}. This value is close to
the $\Gamma=2.1\pm 0.1$ slope of our collective
spectrum of luminous HMXBs, demonstrating that this
spectrum (as well as its composition in terms of sources of different
luminosities and spectral types) places interesting observational
constraints on supercritical accretion models.
The present study suggests that the average spectral hardness of
luminous HMXBs does not strongly depend on their luminosity (compare
the collective spectra for three different luminosity bins in
Fig.~\ref{fig:avspec_lumranges}), which seems to contradict
the general picture outlined above, according to which spectral
hardness should positively correlate with the observed
luminosity of ULXs. Part of the explanation why the collective
spectrum of our least luminous
($L_{{\rm X,unabs}}=10^{38}$--$10^{39}$~erg~s$^{-1}$) sources is as hard as
the spectra of more luminous objects may be that this low-luminosity
bin probably includes, apart from (soft) supercritical accretors,
sub- and near-critically accreting black holes and neutron stars with
relatively hard spectra. Another reason may be that the spectrum
for this luminosity bin is significantly affected by LMXB
contamination, which we have roughly taken into account [via
eq.~(\ref{eq:lmxb})] for the normalization but not for the shape
of the spectrum.
According to the collective spectrum of HMXBs, soft X-ray
emission (0.25--2~keV) dominates the total radiative output of HMXBs in the
local Universe. This fact has been frequently overlooked in previous
studies, because much of this soft X-ray emission is absorbed in the
ISM and does not reach the Earth's absorber. The lower energy boundary
of 0.25~keV in our analysis is mainly set by the sensitivity of the {\sl
Chandra} X-ray telescope. What if the collective spectrum of
HMXBs continues with nearly the same slope ($\Gamma=2.1$) to yet
lower energies? This would mean that the luminous
HMXB population produces, apart from X-rays, a comparable or even higher
luminosity, $\gtrsim 5\times
10^{39}$~erg~s$^{-1}(M_\odot$~yr$^{-1})^{-1}$, at UV and lower
frequencies. This hidden radiation, if real, may be associated with
'misaligned ULXs', i.e. supercritically accreting massive binaries
viewed at yet higher inclinations and/or having yet higher
accretion rates than ULXs and ULSs, and the famous Galactic
microquasar SS~433 may be one of such systems
(e.g. \citealt{fabrika04,pouetal07,khasaz16}). Indeed, the observed
(albeit only at $E\gtrsim 2$~keV, because of strong line-of-signt
absorption in the soft band) X-ray luminosity of SS~433 is only
$\sim 10^{36}$~erg~s$^{-1}$ (and it is associated with its baryonic jets
rather than directly with the central source), while its UV luminosity
is estimated as $\sim 10^{40}$--$10^{41}$~erg~s$^{-1}$
\citep{cheetal82,doletal97}, and this radiation is probably associated
with the photosphere of the disc wind \citep{fabrika04}.
\section{Conclusion}
\label{s:summary}
Using a sample of 200 luminous ($L_{{\rm X,unabs}}>10^{38}$~erg~s$^{-1}$) HMXB
candidates detected by {\sl Chandra} in 27 nearby galaxies, we have
constructed the collective X-ray spectrum of HMXBs in the local
Universe per unit star formation rate, corrected
for observational biases associated with intrinsic diversity of source
spectra and X-ray absorption in the ISM (of the host galaxies and
the Milky Way). This spectrum can be described by a power law with a
photon index $\Gamma=2.1\pm 0.1$ [eqs.~(\ref{eq:fit_l38_40.5}) and
(\ref{eq:fit_l38_40})] and is dominated by ultraluminous sources
($L_{{\rm X,unabs}}>10^{39}$~erg~s$^{-1}$), with comparable contributions from
hard, soft and supersoft sources [as defined in
eq.~(\ref{eq:types})]. Hard sources, whose spectra resemble those of
'classical' ULXs, dominate at energies above a few keV, while the bulk
of the soft X-ray emission (below 2~keV) is provided by soft and
supersoft sources.
If our favoured interpretation that the derived spectrum mainly
represents population- and angle-integrated emission from
supercritically accreting HMXBs is correct, then its nearly flat
shape (in $\nu F_\nu$ units) provides an interesting constraint on
theoretical models of supercritical accretion.
The strong soft X-ray emission revealed by the intrinsic collective
spectrum of HMXBs could play an important role in the early
Universe, since the ISM in the first galaxies was probably more
transparent to X-rays than in present-day galaxies, in particular due
to the lower metallicity of the former. As a result, soft and
supersoft luminous HMXBs might have been the key contributors to the
X-ray heating of the Universe prior to its reionization, as we have
discussed recently \citep{sazkha17b}. The collective spectrum of
HMXBs obtained here can thus be used as a reference spectrum for
detailed simulations of cosmic X-ray preheating.
\section*{Acknowledgments}
The authors thank the referee for useful suggestions.
|
\section{introduction}
The adatoms on graphene is an intriguing issue, which has drawn lots of research interest in the last decade\cite{chemreview2009,adatomreview,chenjh2008,solid2012, prl2011, katsnelson2010,wu2011,hujun2012,jianghua2012,hydrogenSO, zheng2015,wuxiaosong2015,indium2015, prl2008, uchoanjp, hydrogentransport,jafarijpcm, science2016,kondoreview,bulla2010,lilinnjp,zhuanghb2009,uchoakondo2011}. The adatom introduces many important physics to the graphene system, e.g. the modification of transport property\cite{chenjh2008,katsnelson2010,prl2011,solid2012}, the enhancement of spin-orbit interaction\cite{wu2011,hujun2012,jianghua2012,hydrogenSO,zheng2015,wuxiaosong2015,indium2015}, the formation of local magnetic moment\cite{prl2008,uchoanjp, hydrogentransport, jafarijpcm, science2016}, and the Kondo physics\cite{kondoreview,bulla2010,lilinnjp,zhuanghb2009,uchoakondo2011}. One most recent breakthrough is about the hydrogen atom absorbed on graphene, where the local magnetic moment of the hydrogen adatom is confirmed and manipulated in a scanning tunnelling microscopy (STM) experiment\cite{science2016}, nearly ten years after its theoretical proposal. The importance is that a non-magnetic adatom can induce a local magnetic moment on graphene in a controllable way, which can integrate the ferromagnetism into the graphene physics and has great potential in future device applications.
Theoretically, the adatom on graphene can be well described by the Anderson impurity model, where the adatom is viewed as impurity levels coupled to the conducting electrons in the metal (i.e. graphene here)\cite{anderson61}. Due to this hybridization, the discrete impurity spectral is broadened to a Lorentzian-like shape. And, considering the on-site Coulomb interaction of the impurity orbital, the energy levels of the spin up and down are separated , which can give rise to the formation of local magnetic moment. Because of the novel electronic structure of graphene, the Anderson impurity on graphene has some unique features distinct from the normal metal. For example, since the density of states (DOS) near Fermi level goes to zero, the effective hybridization between the impurity level and the graphene is very small, which prefers to form a local magnetic moment.
In comparison with the monolayer graphene, trilayer graphene (TLG) has a drastically different electronic structure\cite{rmpgraphene,guinea2006,aba2006,solidtrilayer}. Importantly, with different stacking orders, the corresponding energy dispersion of TLGs are different.
For example, the rhombohedral (ABC) stacked TLG (r-TLG) has a $\mathbf{k}^3$ dispersion near the Fermi level, and thus it has a divergent DOS at the Fermi level. Meanwhile, near the Fermi level, the bands of the Bernal (ABA) stacked TLG (b-TLG) are linear or quadratic, and the DOS is finite. Their responses to the electric field are also different. A band gap will be opened by an external electric field in r-TLG, while there is an electric field induced band overlap in the b-TLG\cite{abakoshino,wutrilayer,zhangfantrilayer,koshino2010,lujingtrilayer,heinz2011,stmtrilayergap,abctransport,helin2015}.
Stimulated by the recent experimental progress and considering the novel electronic properties of TLG, in this work, we systematically examine the behaviors of the adatoms on TLG systems by studying the Anderson impurity model via self-consistent mean field method. We carefully investigate the influences of stacking order, adsorption site, and external electric field, to which little attention has been paid in the literatures so far\cite{ding2009}. We find that the adatom on TLG has some unusual characteristics, which is distinct from the monolayer graphene as well as the normal metal.
Firstly, we find that the adatom on r-TLG can have a Fano-shaped impurity spectral density, in contrast to the common Lorentzian-like line shape. It actually results from the interference between the divergent TLG DOS peak at Fermi level and the broadened impurity level. According to our analysis, we argue that it may be general for any impurity level near the metal DOS singularity.
Secondly, we demonstrate that a perpendicular electric field can make the impurity level in r-TLG into an in-gap state, which strongly depends on the direction of the applied electric field.
The key issue is that, in addition to opening an energy gap, the electric field can also shift the energy level of the adatom if it is only coupled to the top layer.
Namely, with an electric field in one direction, an energy gap can be opened in the r-TLG and the impurity level is shifted into the gap to form an in-gap state. But, when reversing the direction of the electric field,
the impurity level is shifted away and the in-gap state can not be achieved, though the gap can still be opened. This phenomenon appears as long as the adatom is only coupled to the top layer. The in-gap state favors the formation of local magnetic moment, and greatly modifies the impurity magnetic phase diagram of the r-TLG.
Thirdly, we numerically calculate the impurity magnetic phase diagrams for TLG, where various stacking orders, adsorption sites, doping and electric field are considered. We focus on the mechanisms for tuning the adatom magnetic moment via an electric field.
The key quantity here is the hybridization between the impurity level and the conducting electrons, which is controllable by an applied electric field.
There are several typical cases. For the undoped b-TLG, the electric field induced band overlap will enhance the hybridization, so that the magnetic phase is suppressed by the electric field. Meanwhile, for the undoped r-TLG, the DOS singularity at the Fermi level results in a giant hybridization, and thus the adatom prefers to be non-magnetic. But, if an electric field is applied to open an energy gap, the hybridization is reduced and thus the adatom on the r-TLG can be tuned to be magnetic. Importantly, due to the appearance of the in-gap state, the effect of the electric field on the impurity magnetic moment in r-TLG strongly depends on the direction of electric field (i.e. the polarity of the applied bias).
We also find that the impurity magnetic moment is very sensitive to the adsorption site. The influence of doping on the impurity magnetic phase diagram on TLG has been discussed as well.
Considering the rapid experimental progress, we would like to point out that our results here can be readily tested in the STM experiment, and may have potential applications in future magnetic device on graphene systems. It should also be noted that, here we just focus on the top-site adatom, e.g. the case of hydrogen atom, and the cases for hollow and bridge sites are left to a separate work in future.
The rest of the paper is organized as follows: In Sec. \ref{model}, we give our theoretical model and present the formula in the calculation. In Sec. \ref{discussion}, we first discuss the Fano-shaped impurity spectral function on r-TLG. Then,we talk about the electric field induced in-gap state. Finally, we show the results about the impurity magnetic phase diagrams of the TLG, and related discussions are also given. A short summary is given in Sec. \ref{summary}.
\section{theoretical model and method}\label{model}
Let us first introduce the theoretical model used in this work.
We use the Anderson impurity model to describe the adatom on the TLG. The complete Hamiltonian is written as
\begin{equation}
H=H_{\textrm{TLG}} + H_{\textrm{imp}} + H_{\textrm{hyb}}.
\end{equation}
Here, $H_{\textrm{TLG}}$ is tight binding Hamiltonian of the TLG, $H_{\textrm{imp}}$ is the Hamiltonian of adatom and $H_{\textrm{hyb}}$ describes the hybridization between the impurity levels and the conducting electrons in TLG.
\begin{figure}[h]
\centering
\includegraphics[width=8cm]{fig1.eps}
\caption{(Color online) Schematic diagrams of TLG. (a) and (b): top and side views of b-TLG; (c) and (d): top and side views of r-TLG.
}
\label{tlg}
\end{figure}
For simplicity, only the nearest neighbourhood (NN) hoppings are considered, and the structures of r-TLG and b-TLG are shown in Fig. \ref{tlg}. Generally,
\begin{equation}\label{eqoftlg}
H_{\textrm{TLG}} = H_{\textrm{intra}} + H_{\textrm{inter}} + H_{\textrm{bias}}.
\end{equation}
$H_{\textrm{intra}}$ describes the intralayer hopping, and is equivalent to the Hamiltonian of monolayer graphene
\begin{equation}
H_{\textrm{intra}}=- t \sum_{l\bm{k}\sigma} [ a_{l\bm{k} \sigma}^{\dag} \phi (\bm{k}) b_{l\bm{k} \sigma} + h.c.].
\end{equation}
Here, $a_{l\bm{k}\sigma}$ ($b_{l\bm{k}\sigma}$) is the annihilation operator of the electron on A (B) sublattice, $l$, $\bm{k}$ and $\sigma$ are the layer, momentum, and spin index, respectively, $t$ is the intralayer NN hopping, $ \phi (\bm{k}) = \sum_{j=1}^{3} e^{i \bm{k} \cdot \bm{\delta}_{j} } $, and $\bm{\delta}_j$ are the three vectors pointing from one carbon atom on A sublattice to the three adjacent atoms on B sublattice in the same layer. As shown in Fig. \ref{tlg}, the interlayer hoppings for the r-TLG and b-TLG are:
\begin{equation} \label{KineInterRTLG}
H_{\textrm{inter}}^{\textrm{r}} = t_{\perp} \sum_{\bm{k} \sigma}( b_{1\bm{k} \sigma}^{\dag} a_{2\bm{k} \sigma} + b_{2\bm{k} \sigma}^{\dag} a_{3\bm{k} \sigma} + h.c.),
\end{equation}
\begin{equation} \label{KineInterBTLG}
H_{\textrm{inter}}^{b} = t_{\perp} \sum_{\bm{k} \sigma}( b_{1\bm{k} \sigma}^{\dag} a_{2 \bm{k} \sigma} + a_{2 \bm{k} \sigma}^{\dag} b_{3 \bm{k} \sigma} + h.c.).
\end{equation}
$t_\perp$ is the interlayer hopping parameter.
When a perpendicular electric field is applied, it gives
\begin{equation}
H_{\textrm{bias}}=\sum_{l \bm{k} \sigma}\frac{1}{2} ( l - 2 ) V_{g} ( a_{l \bm{k} \sigma}^{\dag} a_{l \bm{k} \sigma}+ b_{l \bm{k} \sigma}^{\dag} b_{l \bm{k} \sigma}).
\end{equation}
$V_g$ is the applied bias voltage, i.e. the potential difference between the top and bottom layers. Diagonalizing the Hamiltonian of the TLG in Eq. (\ref{eqoftlg}) , we can get the corresponding energy bands
\begin{equation}
H_{\textrm{TLG}}=\sum_{n\bm{k}\sigma} (\varepsilon_{n\bm{k}\sigma}-\mu)c^{\dag}_{n\bm{k}\sigma}c_{n\bm{k}\sigma}
\end{equation}
where $n$ is the band index and $c^{\dag}_{n\bm{k}\sigma}$ is the electron creation operator of the eigenstate. The chemical potential $\mu$ is also included in order to consider the charge doping.
In Fig. \ref{bands}, we show the energy bands for both r-TLG and b-TLG with different bias voltages as an example, which is helpful to understand the following numerical results. The electric field induced band overlap of the b-TLG, as well as the opened gap of r-TLG, are shown clearly in Fig. \ref{bands}. In this paper, we set $t=3.16$ eV, and $t_\perp = 0.39$ eV. Note that, in experiment, one way to produce a perpendicular electric field is to build a dual-gated device\cite{zhangyuanbo2009,lauaba,lau2014}, where the potential difference $V_g$ and the charge density (or chemical potential $\mu$) can be independently tuned. Another possible way is to use molecular doping\cite{dopinggap,duan2011,wuxiaosong}, which not only can induce a potential difference between layers, but also change the charge density. Here, we assume that the electric field is produced by the dual-gated device, where the impurity level is shifted in the presence of the electric field.
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm]{fig2.eps}
\caption{(Color online) Energy bands for (a) b-TLG and (b) r-TLG with different bias voltage $V_g$. Solid black lines, $V_{g} = 0 \; \textrm{eV} $; dashed blue lines, $V_{g} = 0.05 \; \textrm{eV} $; dotted red lines, $V_{g} = 0.1 \; \textrm{eV} $.
}
\label{bands}
\end{figure}
The adatom is described as a single level impurity with the on-site Coulomb interaction. Note that in this work, we assume that the adatom is on the top layer ($l=1$). Thus, in reality, the energy level of the impurity $\varepsilon_0$ can also be shifted by the external electric field\cite{levelshift}. The impurity Hamiltonian is
\begin{equation}
H_{\textrm{imp}} = \sum_{\sigma} \varepsilon_d d_{\sigma}^{\dag} d_{\sigma} + U d_{\uparrow}^{\dag} d_{\uparrow} d_{\downarrow}^{\dag} d_{\downarrow}.
\end{equation}
where $\varepsilon_d = (\varepsilon_{0}- \frac{V_g}{2})$ is the shifted energy of the impurity level by the electric field. $U$ is the on-site Coulomb interaction.
As mentioned above, we only consider the top-site adatom, i.e. the case of hydrogen atom.
So, as shown in Fig. \ref{tlg}, the adatom is on the top of the $\textrm{A}_1$ or $\textrm{B}_1$ carbon atom of the top layer. Note that in TLG, the A and B sublattice in one monolayer are no longer equivalent. The hybridization of the top site impurity are
\begin{equation}
H_{\textrm{hyb}}^{\textrm{A}_1} = \frac{V}{\sqrt{N}} \sum_{\bm{k} \sigma} ( d_{\sigma}^{\dag} a_{1 \bm{k} \sigma} + h.c. ),
\end{equation}
\begin{equation}
H_{\textrm{hyb}}^{\textrm{B}_1} = \frac{V}{\sqrt{N}} \sum_{\bm{k} \sigma} ( d_{\sigma}^{\dag} b_{1 \bm{k} \sigma} + h.c. ).
\end{equation}
Here, $N$ is the number of sites on sublattice A or B in one monolayer.
With the mean field approximation, the occupation of the impurity level is given by
\begin{equation}\label{nd}
\begin{split}
n_{\sigma} &= \int_{- \infty}^{\mu} \textrm{d} \omega \rho_{d,\sigma}(\omega) \\
&= - \frac{1}{\pi} \int_{- \infty}^{\mu} \textrm{d} \omega \textrm{Im} [ G_{dd, \sigma}^{r}(\omega) ] ,
\end{split}
\end{equation}
where $\rho_{d,\sigma}(\omega)$ is the impurity spectral density, and $G_{dd, \sigma}^r(\omega)$ is the retarded Green's function of the impurity electron. $G_{dd, \sigma}^r(\omega)$ can be expressed by the impurity self-energy
\begin{equation}\label{greenf}
G_{dd, \sigma}^r(\omega) =\frac{1}{\omega + i \eta - \varepsilon_{d} - U n_{\bar{\sigma}} - \Sigma_{dd}^{r}(\omega) }.
\end{equation}
When the adatom is on the top site, e.g. on $\textrm{A}_1$ site, the impurity self-energy is
\begin{equation}
\begin{split}\label{selfenergy}
\Sigma_{dd}^r(\omega) &=(V^2/N) \sum_{\bm{k}} g^{r}_{a_1a_1,\sigma}(\bm{k},\omega) \\
&=V^2 g^r_{a_1 a_1, \sigma}(\omega)
\end{split}
\end{equation}
where $g^{r}_{a_1a_1,\sigma}(\bm{k},t)=-i\theta(t) \langle \{a_{1\bm{k}\sigma}(t),a^\dagger_{1\bm{k}\sigma}(0)\} \rangle$ is the noninteracting Green's function of the electrons in TLG, and we define
\begin{equation}
g^r_{a_1a_1,\sigma}(\omega)\equiv (1/N)\sum_{\bm{k}} g^{r}_{a_1a_1,\sigma}(\bm{k},\omega).
\end{equation}
We see that $g^r_{a_1a_1,\sigma}(\omega)$ is actually the real space noninteracting Green's function of TLG electron on the $\textrm{A}_1$ site. Thus, the local density of states (LDOS) of pristine TLG on the $\textrm{A}_1$ site is
\begin{equation}\label{ldos}
\rho^0_{a_1}(\omega)=-\frac{1}{\pi} \textrm{Im}[g^r_{a_1 a_1, \sigma} (\omega)].
\end{equation}
Comparing with Eq. (\ref{selfenergy}), we see that $\rho^0_{a_1}(\omega)$ is equal to the imaginary part of the self-energy $\Sigma^r_{dd}(\omega)$, apart from a constant. Meanwhile, for the Anderson impurity, the hybridization function
\begin{equation}
\Gamma(\omega)=\pi V^2 \rho^0_{a_1}(\omega)
\end{equation}
is normally used to describe the the coupling between the impurity level and the conducting electrons in metal (here, we assume the adsorption site is $\textrm{A}_1$ for example). Thus, the LDOS of the metal at the adsorption site actually reflects the hybridization function. We would like to emphasize that, in normal metal, the electron distribution is uniform, so that the LDOS is uniform as well and is in some sense equivalent to the DOS. But, here, the A and B sublattices are inequivalent. Thus, the LDOS is now site dependent and is different from the DOS of the whole system. It means that only the LDOS of the adsorption site, instead of the DOS of the whole system, is meaningful for the Anderson impurity.
Numerically, Eqs.(\ref{nd}-\ref{greenf}) can be self-consistently solved, and then we can get impurity magnetic phase diagram, as well as the impurity spectral function, for all the cases.
\section{results and discussions}\label{discussion}
Here, we present our numerical results and related discussions. First, we would like to discuss the interesting impurity spectral density of the r-TLG, which may be of a Fano line shape instead of the normal Lorentzian-like one. Second, we talk about the electric field induced in-gap state in r-TLG, which can greatly influence the magnetic moment of adatom. Finally, we give the impurity magnetic phase diagrams of the TLG, where the effects of the stacking order, adsorption site, doping and electric field are discussed in detail. We interpret the mechanisms for controlling the impurity magnetic moment by electric field.
\subsection{Fano-shaped impurity spectral density in r-TLG}
To see the spectral density of the adatom, we first consider the noninteracting case, i.e. $U=0$, where the impurity levels for spin up and down are degenerate. By solving Eq. (\ref{nd}), we plot the impurity spectral $\rho_{d}(\omega)$ in Fig. \ref{fanodos} for the cases that the adatom is on the $\textrm{A}_1$ site.
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm]{fig3.eps}
\caption{(Color online) (a), (b), (c), (d): The impurity spectral density $\rho_d(\omega)$ in the noninteracting case ($U=0$). For r-TLG [(b) and (d)], the red dotted lines are the fitting curves for the Fano line shape with Eq.(\ref{rhod}). The fitting parameters $[c_{1}, \varepsilon_{1}, \Gamma_{1}, z_{2}, \varepsilon_{2} ,\Gamma_{2}, q ]$ are [0.51,$-0.032$ $\textrm{eV}$, 0.073 $\textrm{eV}$, 0.39 $\textrm{eV}$, 0.012 $\textrm{eV}$, 0.038 $\textrm{eV}$, 0.35] in (b) and
[0.49, $-0.10$ $\textrm{eV}$, 0.073 $\textrm{eV}$, 0.75 $\textrm{eV}$, 0.0043 $\textrm{eV}$, 0.020 $\textrm{eV}$, 0.76] in (d). (e) and (f): The corresponding LDOS of TLG at $A_1$ site. Other parameters are $V=4$ eV, $V_g=0$ eV, $\mu=0$ eV.
}
\label{fanodos}
\end{figure}
Obviously, the impurity spectral density of the b-TLG and r-TLG have different behaviors. For the b-TLG, the impurity spectral density coincides with the common understanding. The impurity level is broadened by the metal conducting electrons, which has a Lorentzian-like line shape. The width of the DOS peak depends on the imaginary part of the impurity self-energy, i.e. the LDOS given in Eq. (\ref{ldos}), and the corresponding LDOS is plotted in Fig. \ref{fanodos} (e). Meanwhile, the impurity spectral density of b-TLG does not depend much on the position of the impurity level $\varepsilon_0$, as shown in Fig. \ref{fanodos} (a) and (c).
The behaviours of the r-TLG are completely different. In Fig.\ref{fanodos} (b), when the impurity level is around the Fermi level ($\varepsilon_0=-0.1$ eV), the impurity spectral density is not a peak as normal, but has a dip at the Fermi level.
When $\varepsilon_0$ is shifted away ($\varepsilon_0=-0.3$ eV), it recovers to a Lorentzian-like shape, but has a small peak near the Fermi level, as shown in Fig. \ref{fanodos} (d).
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm]{fig4.eps}
\caption{(Color online) The impurity spectral density $\rho_{d,\sigma}(\omega)$ with finite on-site $U$ ($U=0.9$ eV). $\mu=0$ eV, $\varepsilon_{0} = - 0.4 \; \textrm{eV}$.
}
\label{spin}
\end{figure}
We now demonstrate that the unusual impurity spectral density of r-TLG can be interpreted as a Fano resonance, which reflects the interference between the broadened impurity level and the DOS peak of r-TLG at the Fermi level.
As we know, the shape of the $\rho_{d}(\omega)$ actually depends on the imaginary part of the self-energy, which is nearly a constant in normal metal. But, in r-TLG, this assumption is invalid. We plot the corresponding LDOS, i.e. the imaginary part of the self-energy, in Fig. \ref{fanodos} (f), which has a sharp peak at the Fermi level in contrast to the b-TLG case [see Fig. \ref{fanodos} (e)]. This is not surprising because that the DOS of r-TLG is divergent at the Fermi level. To understand the shape of $\rho_{d}(\omega)$, intuitively, we can separate $\Sigma^r_{dd}(\omega)$ into two parts by sorting the electron states in the summation in the definition of Eq. (\ref{selfenergy}),
\begin{equation}
\Sigma^r_{dd}(\omega)=\Sigma^{r(1)}_{dd}(\omega) + \Sigma^{r(2)}_{dd}(\omega).
\end{equation}
$\Sigma^{r(1)}_{dd}(\omega)$ corresponds to the constant background of the LDOS, and $\Sigma^{r(2)}_{dd}(\omega)$ is related to the peak.
Specifically, from the band structure of r-TLG (see in Fig. \ref{bands}), we see that only the states near the $E=0$ contribute to the DOS peak, and thus by summing all these states, we approximately have
\begin{equation}
\Sigma^{r(2)}_{dd}(\omega) \approx V^2 g^r_{\textrm{peak}}(\omega)
\end{equation}
where
\begin{equation}\label{grpeak}
g^r_{\textrm{peak}}(\omega) = \frac{c_{2}}{\omega - \varepsilon_2 + i\Gamma_2}.
\end{equation}
Here, $c_{2}$ is the appropriate strength of the pole, $\varepsilon_2$ is the position of DOS peak ($\varepsilon_2=0$ in this case). We just use a Lorentzian to model the DOS peak here, and in principle the peak width $\Gamma_2$ should be very narrow (several tens of meV \cite{xiao2015}) in comparison with the hybridization broadened impurity level (normally several eV). Meanwhile, the sum of all the other states gives $\Sigma^{r(1)}_{dd}(\omega)$. Note that
\begin{equation}\label{gamma1}
\Gamma_1 \equiv -\textrm{Im}[\Sigma^{r(1)}_{dd}(\omega)] \approx {\rm const}.
\end{equation}
which results from the nearly constant LDOS except for the DOS peak, as shown in Fig. \ref{fanodos} (f). Now, the impurity Green's function can be approximately expressed as
\begin{equation}\label{dyson}
G^{r}_{dd} (\omega) \approx G^{0r}_{dd}(\omega) + G^{0r}_{dd}(\omega)V g^r_{\textrm{peak}}(\omega)VG^{0r}_{dd}(\omega),
\end{equation}
where we define
\begin{equation}
G^{0r}_{dd}(\omega) \equiv \frac{c_1}{\omega-\varepsilon_d - \Sigma^{r(1)}_{dd}(\omega) + i\eta}.
\end{equation}
Clearly, the $G^{0r}_{dd}(\omega)$ gives the broadened impurity level,
\begin{equation}
\rho^0_d(\omega)=-\frac{1}{\pi}\textrm{Im}[G^{0r}_{dd}(\omega)],
\end{equation}
which should be a Lorentzian due to the relation (\ref{gamma1}), and the width of the corresponding DOS peak is given by $\Gamma_1$. And the position of the DOS peak here is $\varepsilon_{1} = \varepsilon_{d} + \mathrm{Re}[\Sigma_{dd}^{r(1)}(\omega)]$. Here, $c_1$ is also the strength of the pole. Finally, by Eq. (\ref{dyson}), we get an approximate expression of the impurity spectral density
\begin{equation}\label{rhod}
\rho_d(\omega) = \rho^{0}_d (\omega) + z_{2} [ \rho^{0}_d(\omega)]^2 \frac{q^2 - 1 + 2 q \varepsilon'}{1+ {\varepsilon'}^2},
\end{equation}
where $z_2=\frac{\pi c_2 V^2}{\Gamma_2}$, $\varepsilon' = \frac{\omega - \varepsilon_2}{\Gamma_2}$ and
\begin{equation}
q=-\frac{\textrm{Re}[G^{0r}_{dd}(\omega) ]}{\textrm{Im}[G^{0r}_{dd}(\omega) ]}
\end{equation}
The factor $(q^2 - 1 + 2 q \varepsilon')/(1+ {\varepsilon'}^2)$ indicates that the impurity spectral density has a Fano profile, which is determined by the parameter q. We want to point out that the situation here is very similar as the Kondo physics of a magnetic adatom on metal surface\cite{hgluo2004}. The Kondo resonance of a magnetic impurity will give rise to a narrow DOS peak at the Fermi level, which can interference with the broadened impurity level, and thus induce a Fano line shape of the impurity spectral density. The band structure of r-TLG has a DOS peak at the Fermi level, which actually takes the place of the Kondo peak in the Fano resonance. The distinction is that the DOS peak here is an intrinsic property of the band structure, so that the Fano-shaped impurity spectral density we predicted here does not need very low temperature, which is required in the Kondo case.
Then, we test the above picture about the Fano resonance by fitting the numerical results of the impurity spectral density of the r-TLG with the formula in Eq. (\ref{rhod}). The results of fitting are given in Fig. \ref{fanodos} (b) and (d) (see the dotted lines). We see that the formula of the Fano resonance can well describe the impurity spectral density. It means that, in Fig. \ref{fanodos} (b), the dip at the Fermi level should come from a Fano resonance between the DOS peak of r-TLG and the broadened impurity level.
When the on-site Coulomb interaction is included, a local magnetic moment can occur with proper values of the hybridization and Coulomb interaction $U$. This is well described by the Anderson impurity model. If the adatom becomes magnetic, the impurity levels for the up spin and down spin are no longer degenerate. In this case, we have two Fano-shaped impurity spectral density for both spin up and down, as shown in Fig. \ref{spin} (b) and (d). However, if the impurity is non-magnetic, the spin is degenerate, and only one impurity spectral density can be observed, see Fig. \ref{spin} (f). The corresponding cases of b-TLG are given in Fig. \ref{spin} (a), (c) and (e) as well.
The results above are all about the adatom on $A_1$ site. Actually, the Fano-shaped impurity spectral density can only be observed in the r-TLG with $A_1$ site adatom. This is because that, when adatom is on $B_1$ site, the corresponding LDOS does not have a peak at the Fermi level, even for the r-TLG.
At last, we would like to argue that the Fano-shaped impurity spectral density may be a general phenomenon when the impurity level is near a DOS singularity of the metal. Actually, in our analysis above, the only assumption is to use a Lorentzian to model the DOS peak in metal, see in Eq. (\ref{grpeak}). Thus, it is essentially suitable for any DOS singularity in metal. Namely, we predict that, when the metal has a DOS singularity and the impurity level is near the singularity, the impurity DOS should be of a Fano line shape instead of the normal Lorentzian one. We hope this prediction can be tested in future experiment.
\subsection{Electric-field-induced in-gap state in r-TLG}
The in-gap state has been intensively studied in bilayer graphene\cite{wuxiaosong,ingap2007,vacancy2010,ingap2013,ingap2015,zhujianxin}, which should also be important for the TLG.
Here, we would like to point out that, for the adatom on r-TLG, a perpendicular electric field has two important effects: shifting the impurity level and opening an energy gap. An intriguing consequence is the electric field induced in-gap state, which strongly depends on the direction of the electric field, or the polarity of the applied bias.
Assuming that the impurity level is below the Fermi level ($\varepsilon_0- \mu < 0$), a negative bias ($V_g < 0$) will not only open an energy gap but also shift the impurity level towards the gap, since $\varepsilon_d = \varepsilon_0 - V_g /2$. Thus, an in-gap state can be made in this case. However, if the direction of the electric field is reversed ($V_g>0$), a gap can still be opened but the impurity level is now moved away from the gap, so that the in-gap state can never be achieved.
\onecolumngrid
\begin{center}
\begin{figure}[h!]
\includegraphics[width=15cm]{fig5.eps}
\caption{(Color online) (a),(b),(c): The impurity spectral density $\rho_d (\omega)$ of the r-TLG-$A_1$ in the noninteracting case with different $V_g$. (d),(e),(f): The corresponding LDOS of TLG at $A_1$ site. (g),(h),(i): The corresponding energy bands of the r-TLG. $\varepsilon_{0} = -0.3 \ \textrm{eV}$, $V = 1 \ \textrm{eV}$.
}
\label{ingapfig}
\end{figure}
\end{center}
\twocolumngrid
The above picture of the electric field induced in-gap state has nothing to do with the on-site $U$ of the adatom. Thus, we first give
a concrete example in the case of $U=0$ in Fig. \ref{ingapfig} (i.e. spin degenerate case). Here, we consider an adatom with $\varepsilon_0 = -0.3$ eV on $A_1$ site, and apply bias voltages $V_g=$ $-0.2$ eV, $-0.6$ eV and $-0.8$ eV, respectively. When $V_g=-0.2$ eV, an energy gap is opened as shown in Fig. \ref{ingapfig} (g). In Fig. \ref{ingapfig} (d), we plot the corresponding LDOS at $A_1$ site. There is a peak at the edge of the conduction band but no obvious peak for the valence band.
It is known that, due to the opened gap, the DOS has singularities at the edges of both the conduction and valence bands in r-TLG. But the LDOS at $A_1$ site is asymmetric because the electrons on $A_1$ site are mainly from the conduction bands.
In Fig. \ref{ingapfig} (a), we see that the $V_g=-0.2$ eV is not large enough to move the impurity level into the gap, and we observe a broadened impurity level out of the gap. With a larger bias $V_g=-0.6$ eV, the impurity level can now be shifted into the gap, and an in-gap state appears as shown in Fig. \ref{ingapfig} (b), where the corresponding LDOS and bands are given in Fig. \ref{ingapfig} (e) and (h), respectively.
Increasing $V_g$ further, the impurity level is moved up to the conduction band, which is the case in Fig. \ref{ingapfig} (c), (f) and (i) with $V_g = -0.8$ eV. It should be noted that, in Fig. \ref{ingapfig} (c), the shifted impurity level is near the edge of conduction band. We actually get a Fano-shaped impurity spectral density, instead of the Lorentzian one. The reason is similar as the case in last section, since LDOS has a sharp DOS peak at the band edge. Reversing the direction of electric field, $V_g$ becomes positive and the impurity level will be pulled down to lower energy. So, we can not get an in-gap state with positive $V_g$. We indeed do not observe any in-gap state in our numerical results. When the impurity level is above the chemical potential ($\varepsilon_0 - \mu > 0$), the situation can be treated similarly. And if the adatom is on the $B_1$, an in-gap state can also be achieved by a proper electric field.
The in-gap state prefers to be magnetic, if the on-site Coulomb interaction is considered. This is because that , for the in-gap state, the effective coupling between the impurity level and the conducting electrons is very weak. So, a tiny on-site interaction can break the spin degeneracy, and the adatom becomes magnetic in the undoping case. We give an example in Fig. \ref{ingapmagnetic}, where a small on-site $U$ about $15$ meV can induce a spin splitting for the in-gap state, even if the hybridization constant $V=0.5$ eV is not small. In the next section, we will show that the electric-field-induced in-gap state will strongly influence the formation and control of the local magnetic moment of the adatom.
\begin{figure}[h!]
\centering
\includegraphics[width=8.5cm]{fig6.eps}
\caption{(Color online) (a) The impurity spectral density $\rho_{d,\sigma} (\omega)$ of the r-TLG-$A_1$ with $U = 0.0157 \; \textrm{eV}$. Other parameters are $\mu = 0 \; \textrm{eV}$, $V = 0.5 \; \textrm{eV}$, $V_{g} = 0.1 \; \textrm{eV}$, $\varepsilon_{0}=-7.85 \times 10^{-3} \; \textrm{eV}$. The corresponding dimensionless parameters in the magnetic phase diagram are $x=5$ and $y=0.5$. (b) The corresponding LDOS of TLG at $A_1$ site.
}
\label{ingapmagnetic}
\end{figure}
\subsection{Local magnetic moment of the adatom on TLG}
The local magnetic moment formation of an impurity in metal is well described by the Anderson impurity model. The distinct electronic structure of the TLG will make the impurity magnetic phase diagram different from the case of normal metal. We will show that both the stacking order and the external electric field can obviously modify the local magnetic moment formation on the TLG.
\begin{figure}[h!]
\centering
\includegraphics[width=8.5cm]{fig7.eps}
\caption{(Color online) (a), (c), (e), (g): The magnetic phase boundaries of an impurity on the undoped TLG ($\mu=0$ eV) with the impurity level as a variable. (b),(d),(f),(h): The corresponding LDOS of TLG on the adsorption site. $V = 0.5 \; \textrm{eV}$, $D=10$ eV.
}
\label{phase1}
\end{figure}
Let us first introduce some basic physical pictures about the local magnetic moment formation. In the Anderson impurity model, the formation of local magnetic moment depends on the occupation of the impurity level for each spin component $n_\sigma$, which is calculated self-consistently by Eq. (\ref{nd}). The local magnetic moment becomes nonzero when $n_\sigma \neq n_{\bar{\sigma}}$, otherwise the adatom is non-magnetic. Note that here we do not consider the Kondo effect, which is beyond the mean field picture and should occur at very low temperature. There are several key parameters, which govern the local magnetic moment formation. One is the on-site Coulomb interaction $U$. The Coulomb interaction favors the formation of local magnetic moment, because it tends to prohibit the double occupation of the impurity level. The hybridization between the impurity electron and the conducting electrons in metal is also a key quantity, which tends to wash out the local magnetic moment.
We use the hybridization function $\Gamma(\omega)$, or equivalently the LDOS at the adsorption site of the pristine TLG, to represent the hybridization.
In normal metal, the hybridization function can be viewed as a constant $\Delta=\pi V^2 /D$, i.e. the hybridization energy, because that the DOS in metal normally varies smoothly with the energy and can be approximately viewed as a constant. Here, $D$ is the band width and $1/D$ is a constant DOS. However, due to the unique DOS of TLG, the LDOS should not be considered as a constant any more, and will play a key role in the formation of the local magnetic moment. Another important quantity is the energy position of the impurity level relative to the Fermi level of the metal. When the impurity level is close to the Fermi level, the system is in the mixed valence region and small hybridization can destroy the local magnetic moment.
In Fig. \ref{phase1}, we show the phase boundaries of the adatom between the magnetic and non-magnetic phases for both r-TLG and b-TLG in the undoped case.
As is known, the $\textrm{A}_1$ and $\textrm{B}_1$ sites of TLG are no longer equivalent, so that two adsorption sites are considered here. Specifically, there are four concrete situations in Fig. \ref{phase1}: b-TLG with $\textrm{A}_1$ site adsorption (b-TLG-$\textrm{A}_1$), b-TLG with $\textrm{B}_1$ site adsorption (b-TLG-$\textrm{B}_1$), r-TLG with $\textrm{A}_1$ site adsorption (r-TLG-$\textrm{A}_1$), r-TLG with $\textrm{B}_1$ site adsorption (r-TLG-$\textrm{B}_1$). And in order to illustrate the influence of the external electric field, we also plot the results with different bias voltage $V_g$.
In the phase diagram, $x=\Delta/U$ and $y=(\mu-\varepsilon_d)/U$ are two dimensionless parameters, which represent the hybridization and the energy position of the impurity level relative to the chemical potential $\mu$, respectively. The two parameters $x$ and $y$ determine the magnetic phase of the adatom.
Here, we set $\mu=0$ (charge neutrality) and let $\varepsilon_d$ be variable. It corresponds to the case of the undoped TLG with different impurity energy levels $\varepsilon_0$, since $V_g$ is given.
It should be noted that, to fix $\mu$ or $\varepsilon_d$ gives different information in the case of TLG, though $y=(\mu-\varepsilon_d)/U$ may be the same. We discuss the case with fixed impurity energy level later.
In order to interpret the numerical results, we plot the corresponding LDOS in Fig. \ref{phase1} (b), (d), (f) and (h) as well, which can give the effective hybridization together with the constant hybridization energy $\Delta$.
We now discuss the results without the external electric field ($V_g=0$, see the black solid lines in Fig. \ref{phase1}). First, we see that the magnetic regions in the phase diagrams are different. The magnetic phase of r-TLG-$\textrm{A}_1$ is much smaller than other cases [see Fig. \ref{phase1}(e)], while that of the b-TLG-$B_1$ [see Fig. \ref{phase1} (c)] and r-TLG-$B_1$ [see Fig. \ref{phase1} (g)] are the largest. This is attributed to the difference of the hybridization, resulted from the influence of the LDOS. Since $\mu=0$ here, the LDOS of the TLG around the Fermi level will play the major role in the hybridization. For example, in the r-TLG-$\textrm{A}_1$ case [ Fig. \ref{phase1} (f)], the LDOS has a peak at the Fermi level, which corresponds to a giant hybridization and thus greatly suppresses the magnetic phase. In contrast, the LDOS of $\textrm{B}_1$ site of r-TLG does not have a peak at $\mu=0$ ([see Fig. \ref{phase1} (h)], so that the magnetic region is larger. The discussion above means that the adatom on the $A_1$ site of r-TLG prefers to be non-magnetic, compared with other cases.
We also notice that the phase diagram is symmetric around y=0. There are two reasons. One is that the bands for both b-TLG and r-TLG have particle-hole symmetry since we only consider the NN hopping. The other is that it is the undoped case with $\mu=0$.
Then, we discuss the influence of the external electric field. For the case of b-TLG-$\textrm{B}_1$ [see Fig. \ref{phase1} (c)], the bias voltage $V_g$ will induce a LDOS peak near the Fermi level [see Fig. \ref{phase1} (d)], because of the electric field induced band overlap as shown in Fig. \ref{bands} (b). It will considerably enhance the hybridization near the Fermi level, so that the magnetic phase is suppressed [see Fig. \ref{phase1} (c)]. Meanwhile, due to the bias voltage, the LDOS is no longer symmetric around the Fermi level, and thus the phase diagram becomes asymmetric as well, as shown in Fig. \ref{phase1} (a) and (c). As a comparison, we see that the modification of the LDOS (i.e. the hybridization) of the b-TLG-$\textrm{A}_1$ is not so obvious, so that the magnetic phase is not drastically changed, as given in Fig. \ref{phase1} (a).
\begin{figure}[t]
\centering
\includegraphics[width=8.5cm]{fig8.eps}
\caption{(Color online) Spin resolved electron's occupation of the impurity as a function of the energy level with different $V_g$. $U = 1 \; \textrm{eV}$, $\mu = 0 \; \textrm{eV}$, $V = 3.2 \; \textrm{eV}$.
}
\label{manipulation}
\end{figure}
The r-TLG has a completely different behaviour, because that the bias voltage will induce an energy gap near the Fermi level [see Fig. \ref{bands} (a)]. For the case of r-TLG-$\textrm{A}_1$, we can see that the magnetic phase is enlarged drastically by the applied bias voltage [see Fig. \ref{phase1} (e)]. This is because that, near $\mu=0$, the LDOS is changed from a peak ($V_g$=0) to nearly zero (finite $V_g$), where the hybridization has been greatly suppressed. The nearly zero LDOS implies that the impurity level is decoupled from the TLG, which can be effectively considered as a free impurity level. So, very small $U$ can magnetize the impurity and the magnetic phase is enlarged.
The in-gap state in Fig. \ref{ingapmagnetic} actually corresponds to one point in the phase diagram of Fig. \ref{phase1} (e), where $x \approx 5$ and $y \approx 0.5$.
When the adatom is on the $\textrm{B}_1$ site, the undoped case of r-TLG is special. As we mentioned above, we can still change the impurity level into an in-gap state when the adatom is on $B_1$ site. But in Fig. \ref{phase1} (g), we see that the change of the magnetic phase induced by electric field is not as obvious as the $A_1$ site case. The reason is that, for the undoped r-TLG, the effective hybridization here, i.e. LDOS at $B_1$ site, is nearly zero occasionally
[see Fig. \ref{phase1} (h)], so that the variation of the hybridization is tiny when the in-gap state appears. But as long as $\mu \neq 0$, as we will show later, the magnetic phase of the adatom on the $B_1$ site of r-TLG can be drastically changed by the electric field.
Note that $\varepsilon_d$ is the impurity level shifted by the bias voltage. Since the adatom is always on the top layer, the $\varepsilon_d$ of the positive and negative bias voltage are different. Thus, the magnetic phases for the positive and negative $V_g$ are not the same, as shown in Fig. \ref{phase1}.
The phenomena above are useful to manipulate the adatom magnetic moment on TLG. For example, when a bias voltage is applied,
it should be easier to realize the on/off transition of the magnetic moment in the case of r-TLG-$\textrm{A}_1$ than in the b-TLG-$\textrm{A}_1$ case. We give a concrete example in Fig. \ref{manipulation}, where the occupation of the impurity level is plotted for spin up and down, respectively. In Fig. \ref{manipulation} (a), we see that, for the case of b-TLG-$\textrm{A}_1$, only when the impurity level $\varepsilon_0$ is in a very small region, a bias voltage can realize the on/off transition of the local magnetic moment. Namely, when the $\varepsilon_0$ is near $-0.6$ eV, the local moment (i.e. $n_{\uparrow}-n_{\downarrow}$) is finite for $V_g=0$ but zero if a bias voltage ($V_g=0.1$ eV) is applied.
In contrast, for the r-TLG-$\textrm{A}_1$ case [see Fig. \ref{manipulation} (b)], the magnetic moment can be switched by a bias voltage within a much wider allowed energy region of the impurity level, i.e. between $-0.8$ eV and $-0.5$ eV.
All the discussions above are about the undoped TLG with $\mu=0$. In principle, the doping and bias voltage can be controlled separately in the dual-gate configuration in experiment. It should be noted that, in TLG, the effects of doping and applied bias voltage are different, where doping tunes the chemical potential but a bias voltage changes the band structure. Thus, we then give the the phase boundary of the adatom with fixed $\varepsilon_0$ in Fig. \ref{phase2}, which means that the impurity level is given and the chemical potential (doping) is tunable.
\begin{figure}[t!]
\centering
\includegraphics[width=8.5cm]{fig9.eps}
\caption{(Color online) (a),(c),(e),(g): The magnetic phase boundaries for an impurity on TLG with $\varepsilon_{0} = - 0.5 \; \textrm{eV}$. (b),(d),(f),(h): The corresponding LDOS of TLG at the adsorption site.
$V = 0.5 \; \textrm{eV}$.
}
\label{phase2}
\end{figure}
First, we see that the magnetic phase is no longer symmetric around $y=0.5$, which results from the fact that the particle-hole symmetry is absent when $\mu$ is nonzero. This is similar as the cases of monolayer and bilayer graphene in former literatures. Some interesting results appear when a bias voltage is applied. The influence of the electric field on the magnetic phase is very significant for the r-TLG, while it is tiny for b-TLG.
And, the impurity magnetic moment in r-TLG depends strongly on the sign of the applied bias $V_g$, i.e. the direction of the applied electric field.
We plot the magnetic phases for $V_g = \pm 1$ eV. As demonstrated in Fig. \ref{phase2} (e) and (g), a negative bias voltage $V_g=-1$ eV can greatly enlarge the magnetic phase, but a positive bias voltage can not. This is because that, since $\varepsilon_0 = -0.5$ eV here, as discussed in last section, a negative bias voltage can make the impurity level into an in-gap state, which prefers to be magnetic even with tiny on-site Coulomb interaction. Thus, the modification of magnetic phase is so drastic.
But for a positive bias, the in-gap state can not be achieved, so that the change of the magnetic phase is small.
Similarly, the phase diagram of b-TLG is slightly changed by the bias voltage, as given in Fig. \ref{phase2} (a) and (c), due to the absence of the in-gap state. Different from the former case, because that $\mu$ is a variable here, we do not find an intuitive relation between the LDOS and the magnetic phase, though the LDOS is also given for comparison.
\section{summary}\label{summary}
In summary, we theoretically investigate the adatom on the top-site of TLG by solving the the Anderson impurity model via self-consistent mean field method. The influences of the stacking order, nonequivalent adsorption sites, and external electric field are carefully considered.
Our main findings are: \textbf{(1)} Due to the divergent DOS at the Fermi level, the adatom on the $\textrm{A}_1$ site of r-TLG can have a Fano-shaped impurity spectral density, which may be a general property for the impurity level near the metal DOS singularity. The mechanism is similar as the case of Kondo peak induced Fano resonance, which has been intensively studied in the last two decades. \textbf{(2)} For the top-site adatom on r-TLG, the impurity level can be tuned into an in-gap state by an external electric field, which depends strongly on the polarity of the applied bias. The in-gap state greatly modifies the local moment formation of the adatom. \textbf{(3)} Except for adjusting the chemical potential, we demonstrate that controlling the hybridization between the impurity level and the conducting electrons via a bias voltage is another effective way to manipulate the magnetic moment of the adatom on TLG. We systematically calculate the magnetic phase diagram of the adatom on undoped TLG in the presence of an external electric field for various cases.
For undoped r-TLG-$\textrm{A}_1$, the divergent DOS at the Fermi level will greatly enhance the hybridization, so that the adatom is unlikely to be magnetic.
But if a bias voltage is applied, a gap is opened, hybridization is decreased and thus, the adatom can be tuned to be magnetic. Furthermore, if the impurity level is tuned into an in-gap state by a proper bias, tiny on-site Coulomb interaction can make the adatom to be magnetic, which greatly enlarges the magnetic phase in the phase diagram. The situations of the b-TLG are in contradiction to that of the r-TLG, because that electric field will induce a band overlap and increase the hybridization. The typical example is the case of b-TLG-$\textrm{B}_1$, where the electric field enhanced hybridization will obviously restrain the magnetic phase. At last, we discuss the influence of doping on the magnetic phase diagram for the TLG.
We show that the impurity magnetic phase of r-TLG strongly depends on the direction of the electric field, because of the electric-field-induced in-gap state.
We would like to compare the situations of the monolayer\cite{prl2008,jafarijpcm,uchoanjp} and bilayer graphene\cite{ding2009,bilayer2011,bilayer2014,aabilayer} with the case of TLG we studied here. The DOS of the undoped monolayer graphene is zero at the Fermi level , so that the effective hybridization is nearly zero and the adatom prefers to be magnetic. Meanwhile, in monolayer graphene, the effect of electric field is merely to change the chemical potential (doping). As for the bilayer graphene, it is in some sense similar as r-TLG, because that a bias voltage (or an electric field) can open a gap. We think that the electric-field-induced in-gap state will also appear in the bilayer case. But, for the undoped bilayer graphene, the DOS at the Fermi level is a constant, not divergent as r-TLG. Thus, it should be easier to get a magnetic adatom in bilayer than in r-TLG. And there should be no Fano shaped impurity spectral in bilayer graphene. Meanwhile, only in the b-TLG, the electric field can enhance the hybridization and suppress the local magnetic moment. We argue that the influence of the electric field on the adatom in TLG is more significant than that in the bilayer graphene.
Finally, let us discuss the related experiments. The most suitable experimental system to test our theoretical results is the hydrogen atom absorbed on TLG, which can be exactly described by the single level Anderson impurity model. In experiment, the hydrogen atom favors the top site of graphene, and the magnetic moment of hydrogen atom on graphene has been confirmed recently via STM measurement\cite{science2016}. We assume that, on the TLG, the hydrogen atom should also favor the top site, not the hollow and bridge sites. The magnetism of hydrogen adatom can be readily detected by the STM. Furthermore, the DOS peak of the undoped of r-TLG near the Fermi level has been observed recently also by the STM\cite{xiao2015}. The bias voltage induced gap has also been confirmed by various experimental techniques, including STM\cite{stmtrilayergap}, transport\cite{abctransport} and optical measurements\cite{heinz2011}. So, in principle, there is no fundamental difficulties to test our results in experiment. We believe that our work gives a comprehensive understanding about the magnetic moment of the top-site adatom on TLG, and hope that it can stimulate further theoretical and experimental interest in this intriguing issue.
\begin{acknowledgments}
J.H.G is supported by the National Natural Science Foundation of China (Grants No. 11534001, No. 11274129,). J.T.Lu is supported by the National Natural Science Foundation of China (Grant Nos. 61371015 and 11304107).
\end{acknowledgments}
|
\section{Introduction}
Recent literature addressing \textit{queueing networks} (QNs) has aimed to study inferential methods for the estimation of service requirements. These networks offer the means to describe complex stochastic systems through sets of interacting resources, and have found applications in the design of engineering and computing systems \citep{kleinrock1976queueing}, or within call centres \citep{Koole2002}, factories \citep{buzacott1993stochastic} and hospitals \citep{Osorio2009996}. Enabling the understanding of service performance is very important, since it provides quantitative input for the optimal design of interconnected service stations. Yet, drawing inference on parameters is a challenging errand, since in most applications successive network states are never fully observed. Hence, proposed approaches often rely on reduced summaries such as queue lengths, visit counts or response times, and perform inference in different ways, including regression-based estimation procedures, non-linear numerical optimization or maximum likelihood methods. For a recent review on the matter we refer the reader to \cite{Spinner201551} and references therein.
In this paper, we focus on the underlying \textit{continuous-time Markov chains} (CTMCs) induced by general-form open QNs, and we develop a flexible framework for drawing Bayesian inference on parameters that govern these models; in the presence of general patterns of missing data currently only discussed in\citep{sutton2011}. Statistical computation is very difficult within this family of models, as it involves working with often countably infinite state spaces where observations provide little indirect information. Here, we target multi-class Markovian cases with possible class switching and different service disciplines, where few or no individual job departure times are observed at specific servers. Hence, knowledge is mostly restricted to task arrival and departures times to, and from, the network. A \textit{task} is a collection of \textit{jobs} undertaken at different service stations, and high loads make it virtually impossible to determine the state of the network at any point in time, including the ordering of jobs across multiple queues. We propose an inferential framework that allows the imposition of prior system knowledge and overcomes common restrictions on service rates across popular service \textit{disciplines} in traditional modelling approaches. A key contribution is that we introduce a slice-sampling approach relying on mappings to the measurable space of task transitions across the service stations; this enables studying systems where the transition paths of tasks among the queues is unknown, and leads to an efficient sampler. The approach draws motivation from techniques aimed to explore countably infinite state spaces within \textit{Dirichlet mixture models} or \textit{infinite-state hidden Markov models} \citep{Walker2007,VanGael2008,Kalli2011}, and sits well within a uniformization oriented MCMC scheme for jump processes as presented in \cite{rao13a}.
Currently, common assumptions in inferential frameworks include the existence of complete data, product-form equilibrium distributions or unique classes with shared \textit{service rates}. However, we often encounter systems where the completion of jobs at individual stations is only occasionally registered. In addition, inference on the basis of balance may in cases be inaccurate; for instance, the existence of equilibrium in service delivery systems with human workers is a strong assumption, since workload is usually externally controlled and arrivals hardly constitute a Poisson process. In addition, there exist concerns regarding the use of steady-state metrics whenever prior knowledge and constraints are imposed on parameters \citep{armero1994assessment}; and the use of product-form solutions within popular \textit{BCMP networks} \citep{baskett1975open} restricts \textit{first come first served} (FCFS) queues to share service distributions over different task classes.
Aiming for flexible inferential methods, Bayesian procedures relying on \textit{Markov Chain Monte Carlo} techniques were first explored in \cite{sutton2011}. There, the authors discussed a latent variable model targeting networks where only subsets of transition times are observed; the method was applicable to open QNs and defined through deterministic transformations between the data and independent \textit{service times} across different disciplines. Later, \cite{Wang2016} proposed the use of a Gibbs sampler relying on product-form distributions and queue length observations, and it advanced the study of closed BCMP networks, offering an approximation method for the normalizing constant within the network's equilibrium distribution. To the best of our knowledge, no further advances exist in the study of exact Monte Carlo inferential frameworks overcoming known restrictions in the study of QNs. Yet, significant progress has been made with sampling techniques and approximate inference methods for continuous-time dynamic systems often modelled as CTMCs or \textit{continuous-time Bayesian networks} (CTBN) \citep{Nodelman2002,fan2008sampling}. However, simulating system dynamics conditioned on scarce observations remains a complex task; a review on the efficiency of various methods for this purpose (including direct sampling, rejection sampling and uniformization methods) can be found in \cite{hobolth2009}.
Recently, authors \cite{rao13a} have presented a noteworthy contribution based on the principles of uniformization \citep{Lippman:1975,jensen1953markoff}. Their work explores a class of auxiliary variable MCMC methods allowing for the efficient and exact computation of state evolutions in systems with discrete support (such as Markov jump processes). The framework relies on producing highly dependent time discretizations within subsequent blocked steps in a Gibbs sampler, and is hypothetically applicable to the study of system evolutions within QNs. However, such systems exhibit strong and characteristic temporal dependencies (cf. \cite{sutton2011}), transitions over an infinite set of states, varying specifications of service disciplines and Markovian regimes often subject to switching. Hence, we face major impediments which require elaborate implementations of slice sampling techniques \citep{neal2003}. In this work, we describe a method that controls the computational complexity within simulation procedures; for that matter, we employ families of auxiliary variables across steps in a Gibbs sampler targeting network paths. The result is a method that imposes strong restrictions within the vast space of permissible network transitions at each iteration; however, each subsequent step in the sampler allows for significant timing and routing deviations in limited numbers of tasks routed through the network, ensuring convergence to (i) the distribution of network path evolutions across its full space, given the evidence (ii) the posterior distribution of the arrival and service rates. Finally, we present results on both synthetic and real data, obtained from a service delivery tasking tool implemented in two jointly coordinated university hospitals in the United Kingdom.
The rest of the paper is organised as follows. Section \ref{QNandCTMC} describes CTMCs induced by general form QNs, introduces notions of compatibility with observations, and states the problem addressed in the work. In section \ref{UnifAuxObs} the principle of uniformization and its application to networks is briefly revised, mappings to task transitions and auxiliary variables are introduced and the proposed sampler is described. Section \ref{Examps} introduces results for three example networks of varying complexity with both synthetic and real data. Finally, Section
\ref{Discussion} offers a brief closing discussion.
\section{Queue networks and continuous-time Markov processes} \label{QNandCTMC}
Consider an open Markovian network with $M$ single service stations, a \textit{population set} $\mathcal{C}$ of different task classes and a non-deterministic \textit{network topology} defined by a family of \textit{routing probability matrices} $\mathcal{P} = \{P^c : c\in\mathcal{C}\}$, such that
\begin{itemize}
\item $P^c_{i,j}$ denotes the probability of a class $c\in\mathcal{C}$ task immediately moving to station $j$ after completing a job service in station $i$, for all $1\leq i,j \leq M$.
\item $P^c_{i,0}$ denotes the probability of a class $c\in\mathcal{C}$ task immediately exiting the network after completing a job service in station $i$, for all $1\leq i\leq M$.
\item $\sum_{j=0}^M P^c_{i,j} = 1$, for all $1\leq i\leq M, c\in\mathcal{C}$.
\end{itemize}
Furthermore, let $\lambda_c>0$ denote external arrival rates for each task class $c\in\mathcal{C}$; and $p^c_{0,i}$ the corresponding probabilities for its first job to enter station $i$, $1\leq i\leq M$. Servers in the network are assumed independent and may differ in their queueing discipline. Service times are non-negative, have constant rates, and vary over servers and classes; we denote them $\mu^c_i$ for all $1\leq i\leq M, c\in\mathcal{C}$. \textit{Switching} is allowed and thus classes are not permanent categorizations; state-dependent service rates are not considered but follow naturally.
In Figure \ref{exmplQueues} we observe two example networks further examined within Section \ref{Examps} in this paper. There, shaded circles indicate servers with exponential service rates $\mu^c_i$, all accompanied by corresponding job \textit{queueing} areas pictured as empty rectangles. Together, such server and queue pairs each represent a service station $i$, $1\leq i\leq M$. The shaded boxes are probabilistic routing junctions, where task destinations after a job service (or arrival) are determined according to $P^c$ (or $p^c$). Finally, $\lambda_c$ show rates for exponential task arrivals from outside the network.
\begin{figure}[h!]
\centering
\begin{tikzpicture}
\draw (0,0) -- ++(1.4cm,0) -- ++(0,-0.8cm) -- ++(-1.4cm,0);
\foreach \i in {1,...,4}
\draw (1.4cm-\i*8pt,0) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (1.4cm+0.41cm,-0.4cm) circle [radius=0.4cm];
\draw (0,-1.4cm) -- ++(1.4cm,0) -- ++(0,-0.8cm) -- ++(-1.4cm,0);
\foreach \i in {1,...,4}
\draw (1.4cm-\i*8pt,-1.4cm) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (1.4cm+0.41cm,-1.4cm-0.4cm) circle [radius=0.4cm];
\draw (3.2cm,-0.7cm) -- ++(1.4cm,0) -- ++(0,-0.8cm) -- ++(-1.4cm,0);
\foreach \i in {1,...,4}
\draw (4.6cm-\i*8pt,-0.7cm) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (4.6cm+0.41cm,-0.7cm-0.4cm) circle [radius=0.4cm];
\filldraw[fill=black!05!white] (-1.9cm,-0.8cm) rectangle (-1.2cm,-1.4cm);
\draw[->,line width=0.20mm] (-1.1cm,-1.1cm) -- (-0.1,-0.4cm);
\draw[->,line width=0.20mm] (-1.1cm,-1.1cm) -- (-0.1,-1.8cm);
\draw[<-,line width=0.20mm] (-1.5cm,-0.7cm) -- +(0cm,0.3cm) node[above] {$\lambda_c$};
\draw[->,line width=0.20mm] (2.3cm,-0.4cm) -- (3.1cm,-1.05cm);
\draw[->,line width=0.20mm] (2.3cm,-1.8cm) -- (3.1cm,-1.15cm);
\draw[->,line width=0.20mm] (5cm,-0.6cm) -- +(0cm,0.5cm);
\node at (1.83,-0.4cm) {$\mu^c_1$};
\node at (1.83,-1.8cm) {$\mu^c_2$};
\node at (5.03,-1.1cm) {$\mu^c_3$};
\node[align=center] at (-1.55cm,-1.15cm) {$p^c_{0,\cdot}$};
\draw (-0.1,-3) -- ++(1.4cm,0) -- ++(0,-0.8cm) -- ++(-1.4cm,0);
\foreach \i in {1,...,4}
\draw (1.3cm-\i*8pt,-3) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (1.3cm+0.41cm,-3.4cm) circle [radius=0.4cm];
\draw (3.2,-4.4cm) -- ++(-1.4cm,-0cm) -- ++(0cm,-0.8cm) -- ++(1.4cm,0);
\foreach \i in {1,...,4}
\draw (1.8cm+\i*8pt,-4.4cm) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (1.8cm-0.41cm,-1.4cm-3.4cm) circle [radius=0.4cm];
\filldraw[fill=black!05!white] (3.6cm,-3.1cm) rectangle (4.3cm,-3.7cm);
\draw[-,line width=0.20mm] (2.2cm,-3.4cm) -- ++(1.4,0.0cm);
\draw[->,line width=0.20mm] (4.3cm,-3.4cm) -- ++(1cm,0.0cm);
\draw[->,line width=0.20mm] (3.95cm,-3.7cm) -- ++(0cm,-1.1cm) -- ++(-0.65cm,0cm);
\draw[line width=0.20mm] (1cm,-4.8cm) -- ++(-1.9cm,0cm) -- ++(0cm,1.4cm);
\draw[<-,line width=0.20mm] (-0.4cm,-3.4cm) -- ++(-1.2cm,0cm) node[left] {$\lambda_c$};
\node at (1.73,-3.4cm) {$\mu^c_1$};
\node at (1.41,-4.8cm) {$\mu^c_2$};
\node[align=center] at (3.95cm,-3.45cm) {$P^c_{1,\cdot}$};
\end{tikzpicture}
\vspace{10pt}
\caption{On top, a bottleneck network with $3$ servers; bottom, $2$ networks routed in a loop with a single entry and exit server.} \label{exmplQueues}
\end{figure}
Under exponential and independence assumptions, there exists an underlying continuous-time Markov process $X=(X_t)_{t\geq 0}$ that describes the system behaviour. Formally, denoting by $\mathcal{S}$ the countably infinite set of possible states in the network, $X$ is a right-continuous stochastic process such that time-indexed variables $X_t$ are defined within a measurable space $(\mathcal{S},\Sigma_\mathcal{S})$, where $\Sigma_\mathcal{S}$ stands for the power set of $\mathcal{S}$. On a basic level, $X$ holds the ordering of jobs in each queue and server, along with their classes and task identifiers; and $\mathcal{S}$ is the multidimensional product of all possible congruent states at every station. The \textit{infinitesimal generator matrix} $Q$ of $X$ is infinite and such that
$$\mathbb{P}(X_{t+\mathrm{d}t}=x'|X_{t}=x) = \mathbb{I}(x=x') + Q_{x,x'}\mathrm{d}t + o(\mathrm{d}t)$$
for all $x,x'\in\mathcal{S}$. Elements in the generator describe rates for transitions within states in the chain, in addition $Q_{x,x'}\geq 0$ for all $x \neq x'$, and $Q_x \coloneqq Q_{x,x} = - \sum_{x'\in\mathcal{S}: x\neq x'} Q_{x,x'}$. Hence, rows in $Q$ sum to $0$, and the full rate for a state departure is given by $|Q_x|$, for all $x\in\mathcal{S}$. Note that transition rates are the product between routing probabilities and exponential rates above; for instance,
\begin{itemize}
\item $\lambda_c p^c_{0,i}$ is the transition rate among states in $\mathcal{S}$ accounting for a class-$c$ arrival to service station $i$,
\item $\mu^c_i P^c_{i,j}$ is the transition rate among states in $\mathcal{S}$ accounting for a job of class $c$ serviced at station $i$ immediately transitioning to station $j$.
\end{itemize}
\subsection{Observations} Let $\Gamma = \{0,\dots,M\}^2 \times \mathbb{N}$ define a \textit{task transition space}. A triplet $\boldsymbol{\gamma}=(i,j,k)\in \Gamma$ denotes a transition for a uniquely identifiable task $k$, with $i$ and $j$ specifying the departure and entry stations respectively. Note that it is possible to augment $\Gamma$ in order to include task classes, yet given unique identifiers this information is redundant. In this work, a transition triplet is never fully observed; instead, we define a set of partial observations $\mathcal{O} = \mathcal{O}_1 \cup \mathcal{O}_ 2 \subset \Sigma_\Gamma$, with
\begin{align*}
\ \mathcal{O}_1 = \{ \sigma \in \Sigma_\Gamma : \gamma_1=\ &\gamma_1', \ \gamma_3=\gamma_3' \text{ and } \\
& \gamma_2,\gamma_2'>0, \text{ for all } \boldsymbol{\gamma},\boldsymbol{\gamma}'\in\sigma \}, \\
\ \mathcal{O}_2 = \{ \sigma \in \Sigma_\Gamma : \gamma_2=\ &\gamma_2'=0, \ \gamma_3=\gamma_3' \text{ and } \\
& \gamma_1,\gamma_1'>0 , \text{ for all } \boldsymbol{\gamma},\boldsymbol{\gamma}'\in\sigma \},
\end{align*}
where $\Sigma_\Gamma$ stands for the power set of $\Gamma$.
\begin{definition}
An \textit{observation} in $\mathcal{O}$ is a subset of $\Gamma$ that contains all permitted task transitions in the network at some specified time $t>0$, given external information on an arrival, departure or job service.
\end{definition}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[->,line width=0.20mm] (0cm,0cm) -- (7.7cm,0cm);
\draw (0,-0.1cm) -- ++(0cm,0.2cm);
\node at (8cm,0cm) {$\boldsymbol{t}$};
\filldraw[black!60!white,draw=black] (1.2cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.1cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (6.5cm,0cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (0.75cm,0.82cm) rectangle ++(0.9,0.6);
\filldraw[black!05!white,draw=black] (3.65cm,0.82cm) rectangle ++(0.9,0.6);
\filldraw[black!05!white,draw=black] (5.5cm,0.82cm) rectangle ++(0.9,0.6);
\filldraw[black!05!white,draw=black] (6.6cm,0.82cm) rectangle ++(0.9,0.6);
\node[fill=white] at (1.23cm,-0.4cm) {\footnotesize $1_{0\rightarrow 1}$};
\node at (1.2cm,1.1cm) {\small $\boldsymbol{1_{0\rightarrow \cdot}}$};
\node[fill=white] at (4.13cm,-0.4cm) {\footnotesize $1_{1\rightarrow 3}$};
\node at (4.1cm,1.1cm) {\small $\boldsymbol{1_{1\rightarrow \cdot}}$};
\node[fill=white] at (6.53cm,-0.4cm) {\footnotesize $1_{3\rightarrow 0}$};
\node at (5.95cm,1.1cm) {\small $\boldsymbol{1_{3\rightarrow \cdot}}$};
\node at (7.05cm,1.1cm) {\small $\boldsymbol{1_{\cdot\rightarrow 0}}$};
\draw[<-,>=latex,color=black!80!white,line width=0.20mm] (1.2,0.78cm) -- ++(0cm,-0.65cm);
\draw[<-,>=latex,color=black!80!white,line width=0.20mm] (4.1,0.78cm) -- ++(0cm,-0.65cm);
\draw[<-,>=latex,color=black!80!white,line width=0.20mm] (5.95,0.78cm) -- ++(0.5cm,-0.65cm);
\draw[<-,>=latex,color=black!80!white,line width=0.20mm] (7.05,0.78cm) -- ++(-0.5cm,-0.65cm);
\end{tikzpicture}
\caption{Sample observations generated by a single task transitioning a bottleneck network with 3 servers. Observations are represented by rectangles. Dots inform us of transition times. Information below the dots specifies the actual task transitions at each step.} \label{observations}
\end{figure}
In Figure \ref{observations} we observe a bottleneck network produce four partial observations as it evolves over time. The network corresponds to that in Figure \ref{exmplQueues} (top), and observations include a single task arrival, two job services for the task, and a departure immediately after the final service. There, each task transition $(i,j,k)\in\Gamma$ is marked as $k_{i\rightarrow j}$ at its corresponding time point; note that indexes $i,j$ take the value $0$ in order to specify an external arrival or a departure. The observations take the form of elements of $\mathcal{O}$, i.e.
$$1_{i\rightarrow \cdot} = \{\boldsymbol{\gamma} \in \Gamma : \gamma_1 = i , \gamma_2 >0, \gamma_3 = 1 \}\in\mathcal{O}_1,$$ for $i\in\{1,2,3\}$, and
$$1_{\cdot\rightarrow 0} = \{\boldsymbol{\gamma} \in \Gamma : \gamma_1 > 0 , \gamma_2 =0, \gamma_3 = 1 \}\in\mathcal{O}_2.$$
In this toy example, it is possible to deduce the original path $X$ in the network when considering the available observations along with the topology in Figure \ref{exmplQueues}; including task orderings across all queues and servers at every point in time. However, in real world applications job service observations are often missing or do not exist at all. In this work, only arrivals and departures are assumed to always be available.
\subsection{Compatibility}
Let $\mathcal{T}: \mathcal{S}^2 \rightarrow \Gamma \cup \varnothing$ define a measurable function, equipped with the corresponding products of discrete algebras, which maps a pair of states $x,x'\in\mathcal{S}$ to its task transition triplet in $\Gamma$. For instance, $$\mathcal{T}(x,x')=(2,3,12)$$ should $x'$ be reachable from $x$ by servicing a job for task $12$ in server $2$ and immediately routing it to queue $3$. Note that for this to be possible, a job for task $12$ must be in server $2$ within $x$, and the remaining tasks in the system must be distributed and ordered across stations so that there will exist full agreement with $x'$. If a state $x'$ is not directly reachable from $x$, then $\mathcal{T}(x,x') = \varnothing$. We note that the pre-image of a triplet in $\mathcal{T}$ is given by a countably infinite set of pairs of network states in $\mathcal{S}$, unless bounds on the task population are imposed.
\begin{definition} \label{comp} Fix some terminal time $T>0$ and let $\{O_{t_r}\in\mathcal{O} : r=1,\dots,R\}$ be a sequence of observations at times $0\leq t_1<\dots<t_R\leq T$. Also, let $Y_{t_r} = \mathcal{T}^{-1}(O_{t_r}) \in \Sigma_\mathcal{S}^2$ for all $r=1,\dots,R$. Then, we say that a process $X$ is \textit{compatible} with an observation $O_{t_r}$, and we write $X\perp O_{t_r}$ if
$$ \lim_{s\nearrow t_r}X_s = y \quad \text{and} \quad X_{t_r} = y', $$
for some pair of network configurations $(y,y') \in Y_{t_r}$. Furthermore, we say that a process $X$ is \textit{fully compatible} with the observations if $X\perp O_{t_r}$ for all $r=1,\dots,R$.
\end{definition}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[->,line width=0.20mm] (0cm,0cm) -- (7.7cm,0cm);
\draw (0,-0.1cm) -- ++(0cm,0.2cm);
\draw[->,line width=0.20mm] (0cm,-1.5cm) -- (7.7cm,-1.5cm);
\draw (0,-1.6cm) -- ++(0cm,0.2cm);
\draw[->,line width=0.20mm] (0cm,-3cm) -- (7.7cm,-3cm);
\draw (0,-3.1cm) -- ++(0cm,0.2cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (1.2,0.68cm) -- ++(0cm,-4.2cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (6.5,0.68cm) -- ++(0cm,-4.2cm);
\node at (8cm,0cm) {$\boldsymbol{t}$};
\node at (8cm,-1.5cm) {$\boldsymbol{t}$};
\node at (8cm,-3cm) {$\boldsymbol{t}$};
\filldraw[black!60!white,draw=black] (1.2cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.1cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (6.5cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (1.2cm,-1.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (2.8cm,-1.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (6.5cm,-1.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (1.2cm,-3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.9cm,-3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (6.5cm,-3cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (0.75cm,0.8cm) rectangle ++(0.9,0.6);
\filldraw[black!05!white,draw=black] (6.05cm,0.8cm) rectangle ++(0.9,0.6);
\node[fill=white] at (1.23cm,-0.4cm) {\footnotesize $1_{0\rightarrow 1}$};
\node at (1.2cm,1.08cm) {\small $\boldsymbol{1_{0\rightarrow \cdot}}$};
\node[fill=white] at (4.13cm,-0.4cm) {\footnotesize $1_{1\rightarrow 3}$};
\node[fill=white] at (6.53cm,-0.4cm) {\footnotesize $1_{3\rightarrow 0}$};
\node at (6.5cm,1.08cm) {\small $\boldsymbol{1_{\cdot\rightarrow 0}}$};
\node[fill=white] at (1.23cm,-1.9cm) {\footnotesize $1_{0\rightarrow 2}$};
\node[fill=white] at (2.83cm,-1.9cm) {\footnotesize $1_{2\rightarrow 3}$};
\node[fill=white] at (6.53cm,-1.9cm) {\footnotesize $1_{3\rightarrow 0}$};
\node[fill=white] at (1.23cm,-3.4cm) {\footnotesize $1_{0\rightarrow 2}$};
\node[fill=white] at (4.93cm,-3.4cm) {\footnotesize $1_{2\rightarrow 3}$};
\node[fill=white] at (6.53cm,-3.4cm) {\footnotesize $1_{3\rightarrow 0}$};
\end{tikzpicture}
\caption{Example network paths, all compatible with arrival and departure observations for a single task entering and leaving a bottleneck network with three servers.} \label{compatibilityPaths}
\end{figure}
In Figure \ref{compatibilityPaths} we observe task transitions for sample paths $X$ which are compatible with the arrival and departure information as shown in Figure \ref{observations}. There, notice that the first sequence corresponds to the original path forming the observations. This time, no job services have been retained and there exist infinitely many paths $X$ that could have produced the same output, with varying transition times and task orderings across the different stations. In large networks with multiple tasks and all simultaneously transitioning the system, it is hard to picture the infinite amount of fully compatible paths $X$, unless large proportions of job services are retrieved.
\subsection{Latent network and problem statement} Denote by $x_0\in\mathcal{S}$ the initial state in $X$. In this paper, this is assumed to be an \textit{empty} state, where no jobs populate the network. It is however possible to define an initial distribution $\pi$ over states, s.t. $\pi(x)\coloneqq \mathbb{P}(X_0=x)$ for all $x\in\mathcal{S}$. Now, assume we retrieve $K\in\mathbb{N}$ observation sequences $\tilde{\boldsymbol{O}}=\{\boldsymbol{O}_k\}_{k=1,\dots,K}$ collected during different realizations $\boldsymbol{X}=\{X^k\}_{k=1,\dots,K}$ in the network; with
$$\boldsymbol{O}_k = \{O_{t_r}\in\mathcal{O} : r=1,\dots,R_k\}$$
at times $0\leq t_1<\dots<t_{R_k}\leq T_k$, for $k=1,\dots,K$.
The likelihood function is proportional to the product of network path densities fully compatible with $\tilde{\boldsymbol{O}}$, and is thus intractable. A Gibbs sampling approach centred around latent network evolutions is appropriate, iterating between paths and parameters. For that, note that every $X$ is a piecewise-constant process and may be fully characterized by a set of transition times $\boldsymbol{t}=\{t_1,\dots,t_{n}\}$ along with network states $\boldsymbol{x}=\{x_1,\dots,x_{n}\}$, so that $X\equiv (\boldsymbol{t},\boldsymbol{x})$ with $X_0 = x_0$. To ease notation, denote $\boldsymbol{\theta} \equiv (\mathcal{P},\boldsymbol{p},\boldsymbol{\lambda},\boldsymbol{\mu})$, where $\boldsymbol{p}$ is the vector of arrival routing probabilities. Now, let $\delta_{\boldsymbol{x}}$ be the number of transitions in $\boldsymbol{x}$ excluding task arrivals and departures. For each $k=1\dots,K$, the density of $(\boldsymbol{t},\boldsymbol{x})$ given $\boldsymbol{O}_k$ is (up to proportionality) such that
\begin{align}
f_{X}((\boldsymbol{t},\boldsymbol{x})|&\boldsymbol{\theta},\boldsymbol{O}_k,x_0) \nonumber \\
& \propto f_{\boldsymbol{O}}(\boldsymbol{O}_k|(\boldsymbol{t},\boldsymbol{x}),x_0) f_{X}((\boldsymbol{t},\boldsymbol{x})|\boldsymbol{\theta},x_0) \nonumber \\
& \propto (1-q)^{\delta_{\boldsymbol{x}}-\delta_{\boldsymbol{o}}} \times \mathbb{I}((\boldsymbol{t},\boldsymbol{x}) \perp O : O\in\boldsymbol{O}_k) \nonumber \\
& \times e^{Q_{x_n}(T_k-t_n)} \prod_{i=1}^n Q_{x_{i-1},x_{i}} e^{Q_{x_{i-1}}(t_i-t_{i-1})}, \label{densPath}
\end{align}
where $q$ denotes the probability that a job service in $X$ is observed, and $\delta_{\boldsymbol{o}}\leq \delta_{\boldsymbol{x}}$ is the corresponding amount of service observations in $\boldsymbol{O}_k$. This density is supported in a suitably defined space of finite network evolutions and the term on top is proportional to Bernoulli trials penalizing network paths with unobserved job services. The term below follows from the properties of the minimum of exponentially distributed random variables.
Hence, drawing parameter inference entails the complex task of simulating network configurations from \eqref{densPath}, over increasingly large state spaces and with strong conditional dependencies. In the following, we revise the notion of uniformization and sampling methods for jump processes introduced in \cite{rao13a}, and we present an auxiliary observation-variable sampler fit for inference in QN models.
\section{Uniformization and auxiliary observations} \label{UnifAuxObs}
\begin{figure*}[t]
\centering
\begin{tikzpicture}
\draw[->,line width=0.20mm] (0cm,0cm) -- (15.5cm,0cm);
\draw (0,-0.1cm) -- ++(0cm,0.2cm);
\node at (15.8cm,0cm) {$\boldsymbol{t}'$};
\filldraw[black!00!white,draw=black] (0.6cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (2.7cm,0cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (3.8cm,0cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (5.3cm,0cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (6.4cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (9.1cm,0cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (10.8cm,0cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (11.8cm,0cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (13.3cm,0cm) circle [radius=0.09cm];
\node[fill=white] at (2.73cm,-0.4cm) {\scriptsize $0\rightarrow 1$};
\node[fill=white] at (9.13cm,-0.4cm) {\scriptsize $1\rightarrow 3$};
\node[fill=white] at (13.33cm,-0.4cm) {\scriptsize $3\rightarrow 0$};
\filldraw[draw=black,fill=black!00!white] (0.0,-0.7) rectangle (0.6,-0.95);
\filldraw[draw=black,fill=black!00!white] (0.6,-0.7) rectangle (2.7,-0.95);
\filldraw[draw=black,fill=black!20!white] (2.7,-0.7) rectangle (3.8,-0.95);
\filldraw[draw=black,fill=black!20!white] (3.8,-0.7) rectangle (5.3,-0.95);
\filldraw[draw=black,fill=black!20!white] (5.3,-0.7) rectangle (6.4,-0.95);
\filldraw[draw=black,fill=black!20!white] (6.4,-0.7) rectangle (9.1,-0.95);
\filldraw[draw=black,fill=black!40!white] (9.1,-0.7) rectangle (10.8,-0.95);
\filldraw[draw=black,fill=black!40!white] (10.8,-0.7) rectangle (11.8,-0.95);
\filldraw[draw=black,fill=black!40!white] (11.8,-0.7) rectangle (13.3,-0.95);
\filldraw[draw=black,fill=black!00!white] (13.3,-0.7) rectangle (15.5,-0.95);
\node at (15.8cm,-0.8cm) {$\boldsymbol{x}'$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (0.0,-1.4) rectangle (0.9,-4cm);
\filldraw[draw=black,fill=black!00!white] (0.0,-1.4) rectangle (0.9,-1.7);
\node at (0.47cm,-1.57cm) {\scriptsize $0\rightarrow 1$};
\filldraw[draw=black,fill=black!00!white] (0.0,-1.7) rectangle (0.9,-2.1);
\node at (0.47cm,-1.92cm) {\scriptsize $0\rightarrow 2$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (0.0,-2.1) rectangle (0.9,-2.7);
\node at (0.47cm,-2.42cm) {\scriptsize $1\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (0.0,-2.7) rectangle (0.9,-3.1);
\node at (0.47cm,-2.92cm) {\scriptsize $2\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (0.0,-3.1) rectangle (0.9,-3.6);
\node at (0.47cm,-3.37cm) {\scriptsize $3\rightarrow 0$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (2.7,-1.4) rectangle (3.6,-4cm);
\filldraw[draw=black,fill=black!00!white] (2.7,-1.4) rectangle (3.6,-1.7);
\node at (2.7 + 0.47,-1.57cm) {\scriptsize $0\rightarrow 1$};
\filldraw[draw=black,fill=black!00!white] (2.7,-1.7) rectangle (3.6,-2.1);
\node at (2.7 + 0.47,-1.92cm) {\scriptsize $0\rightarrow 2$};
\filldraw[draw=black,fill=black!00!white] (2.7,-2.1) rectangle (3.6,-2.7);
\node at (2.7 + 0.47,-2.42cm) {\scriptsize $1\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (2.7,-2.7) rectangle (3.6,-3.1);
\node at (2.7 + 0.47,-2.92cm) {\scriptsize $2\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (2.7,-3.1) rectangle (3.6,-3.6);
\node at (2.7 + 0.47,-3.37cm) {\scriptsize $3\rightarrow 0$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (9.1,-1.4) rectangle (10,-4cm);
\filldraw[draw=black,fill=black!00!white] (9.1,-1.4) rectangle (10,-1.7);
\node at (9.1 + 0.47,-1.57cm) {\scriptsize $0\rightarrow 1$};
\filldraw[draw=black,fill=black!00!white] (9.1,-1.7) rectangle (10,-2.1);
\node at (9.1 + 0.47,-1.92cm) {\scriptsize $0\rightarrow 2$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (9.1,-2.1) rectangle (10,-2.7);
\node at (9.1 + 0.47,-2.42cm) {\scriptsize $1\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (9.1,-2.7) rectangle (10,-3.1);
\node at (9.1 + 0.47,-2.92cm) {\scriptsize $2\rightarrow 3$};
\filldraw[draw=black,fill=black!00!white] (9.1,-3.1) rectangle (10,-3.6);
\node at (9.1 + 0.47,-3.37cm) {\scriptsize $3\rightarrow 0$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (13.3,-1.4) rectangle (14.2,-4cm);
\filldraw[draw=black,fill=black!00!white] (13.3,-1.4) rectangle (14.2,-1.7);
\node at (13.3 + 0.47,-1.57cm) {\small $0\rightarrow 1$};
\filldraw[draw=black,fill=black!00!white] (13.3,-1.7) rectangle (14.2,-2.1);
\node at (13.3 + 0.47,-1.92cm) {\small $0\rightarrow 2$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (13.3,-2.1) rectangle (14.2,-2.7);
\node at (13.3 + 0.47,-2.42cm) {\small $1\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (13.3,-2.7) rectangle (14.2,-3.1);
\node at (13.3 + 0.47,-2.92cm) {\small $2\rightarrow 3$};
\draw[draw=black,pattern=north west lines,pattern color=black!30!white] (13.3,-3.1) rectangle (14.2,-3.6);
\node at (13.3 + 0.47,-3.37cm) {\small $3\rightarrow 0$};
\draw[-,color=black!60!white,line width=0.20mm] (-0.2cm,-1.4 cm) -- (-0.2cm,-4cm);
\draw[-,color=black!60!white,line width=0.20mm] (-0.25cm,-1.4 cm) -- (-0.15cm,-1.4cm);
\draw[-,color=black!60!white,line width=0.20mm] (-0.25cm,-4 cm) -- (-0.15cm,-4cm);
\node at (-0.53,-2.7) {$\Omega$};
\draw[-,color=black!60!white,line width=0.20mm] (10.2cm,-1.4 cm) -- (10.2cm,-3.6cm);
\draw[-,color=black!60!white,line width=0.20mm] (10.25cm,-1.4 cm) -- (10.15cm,-1.4cm);
\draw[-,color=black!60!white,line width=0.20mm] (10.25cm,-3.6 cm) -- (10.15cm,-3.6cm);
\node at (11.2,-2.53) {$\max_{x\in\mathcal{S}}|Q_x|$};
\draw[-,color=black!60!white,line width=0.20mm] (3.8cm,-1.4 cm) -- (3.8cm,-2.7cm);
\draw[-,color=black!60!white,line width=0.20mm] (3.75cm,-1.4 cm) -- (3.85cm,-1.4cm);
\draw[-,color=black!60!white,line width=0.20mm] (3.75cm,-2.7 cm) -- (3.85cm,-2.7cm);
\node at (4.3,-2.08) {$|Q_{x_2'}|$};
\draw[-,dashed,color=black!60!white,line width=0.20mm] (1cm,-4 cm) -- (2cm,-4cm);
\draw[-,dashed,color=black!60!white,line width=0.20mm] (3.7cm,-4 cm) -- (4.7cm,-4cm);
\draw[-,dashed,color=black!60!white,line width=0.20mm] (10.1cm,-4 cm) -- (11.1cm,-4cm);
\draw[-,dashed,color=black!60!white,line width=0.20mm] (14.3cm,-4 cm) -- (15.3cm,-4cm);
\end{tikzpicture}
\vspace{10pt}
\caption{Graphical representation of times $\boldsymbol{t}'$, states $\boldsymbol{x}'$ and transition probabilities for a uniformization-based simulation in a single-class bottleneck network. We observe a single task routed from entry to departure, with virtual transitions represented by empty dots. Vertical rectangles are proportionally split according the likelihood of the various possible services and arrivals.} \label{unifQueue}
\end{figure*}
A generative process for sampling $X$ requires alternating between exponentially distributed times and transitions in proportion to rates. Instead, a \textit{uniformization}-based \citep{Lippman:1975,jensen1953markoff} sampling scheme employs a \textit{dominating rate} $\Omega$ and introduces the notion of \textit{virtual transitions}, so that all times are sampled in a blocked step. In Algorithm \ref{unifProcX} we observe a uniformization procedure to produce network paths.
\begin{algorithm}[h!]
\normalsize
\caption{Uniformization procedure for process $X$}
\label{unifProcX}
\begin{algorithmic}[1] \vspace{2pt}
\State Fix a dominating rate $\Omega \geq \max_{x\in\mathcal{S}}|Q_x|$.
\State Sample transition times $0\leq t_1< \dots <t_m\leq T$ from a homogeneous Poisson process with rate $\Omega$.
\State Set initial state $X_0=x_0$.
\For {$i\in\{1,\dots,m\}$}
\State Sample $x_i$ from $\{x_{i-1}\}\cup \mathcal{X}_{x_{i-1}}$, with $$\mathcal{X}_{x_{i-1}}=\{x\in\mathcal{S}\backslash\{x_{i-1}\} : \mathcal{T}(x_{i-1},x)\neq\varnothing\},$$ \hskip\algorithmicindent and probabilities $\pi_{x_{i-1}}$ given by $$\pi_{x_{i-1}} = \{1+Q_{x_{i-1}}/\Omega\} \cup \{Q_{x_{i-1},x}/\Omega : x\in\mathcal{X}_{x_{i-1}}\}.$$
\EndFor
\State Return $(\boldsymbol{t},\boldsymbol{x})$.
\end{algorithmic}
\end{algorithm}
A proof of probabilistic equivalence between a generative and uniformized sampling scheme involves comparing the marginal distribution across states at any time $t\geq 0$, and can be found in \cite{hobolth2009}. A uniformization procedure yields an augmented set of times $\boldsymbol{t}'=\{t_1',\dots,t_{m}'\}$ and states $\boldsymbol{x}'=\{x_1',\dots,x_{m}'\}$ that accounts for both real and virtual transitions in $X$. Whenever $x_i=x_{i-1}$ we refer to transition $i$ as \textit{virtual} and note that the number of such transitions is dependent on the choice of $\Omega$. Finally, the density function in \eqref{densPath} may be rewritten to include virtual jumps, so that
\begin{align*}
f_{X}((\boldsymbol{t}',\boldsymbol{x}')|\Omega&,\boldsymbol{\theta},\boldsymbol{O}_k,x_0) \\
&\propto (1-q)^{\delta_{\boldsymbol{x}'}-\delta_{\boldsymbol{o}}} \times \mathbb{I}((\boldsymbol{t}',\boldsymbol{x}') \perp O : O\in\boldsymbol{O}_k) \nonumber \\
& \times \prod_{i=1}^m Q_{x_{i-1}',x_{i}'}^{\mathbb{I}(x_i'\neq x_{i-1}')} (\Omega+Q_{x_{i-1}'})^{\mathbb{I}(x_i'= x_{i-1}')},
\end{align*}
where terms not proportional to $(\boldsymbol{t}',\boldsymbol{x}')$ are omitted.
In practice, simulating $X$ only requires considering a limited number of candidate states in each transition; in close relation to the number of service stations. In Figure \ref{unifQueue} we observe a graphical representation of times, states and transition probabilities for a uniformization-based procedure in the single-class bottleneck network in Figure \ref{exmplQueues} (top). There, we observe only one task from entry to departure, and we notice $\boldsymbol{x}'$ is unaltered after virtual transitions. Vertical rectangles are divided in proportion to rates for services and arrivals, and infeasible services are hashed in grey (the additional hashed area in the bottom accounts for a strictly positive dominating rate $\Omega$). This determines the probabilities leading to new states at subsequent times, with virtual jumps associated to the sum of all hashed regions. Finally, removing virtual transitions within $(\boldsymbol{t}',\boldsymbol{x}')$ yields the desired realization $(\boldsymbol{t},\boldsymbol{x})$ in $X$.
\subsection{An auxiliary observation-variable sampler} A uniformization oriented approach can enable the construction of a Gibbs sampler targeting the conditional distribution $f_{X}((\boldsymbol{t},\boldsymbol{x})|\boldsymbol{\theta},\boldsymbol{O}_k,x_0)$. For such purpose, authors \cite{rao13a} show it is possible to recycle groups of real transition times within each iteration. The method applies well to many families of Markov jump processes, but it is insufficient in order to tackle complex systems such as QNs due to a quadratic cost on the number of states when producing $\boldsymbol{x}$. This is a known problem in discrete-time systems with large state spaces (such as dynamic Bayesian networks or infinite-state hidden Markov models), and proposed solutions include approximate inference methods \citep{Boyen1998,Ng2002} or the use of slice sampling techniques for exact inference \citep{VanGael2008}.
However, QNs contain strong serial dependencies, and transitions over an infinite set of states are triggered by a very reduced number of rates; hence, this can render techniques aimed at Dirichlet mixture models \citep{Walker2007,Kalli2011} or hidden Markov models unusable. A viable approach would ideally consider limited divergences in network paths $X$ over subsequent steps in a sampler; yet allowing for considerable deviations in the routing of a reduced set of tasks. Here, we describe a sampling scheme that achieves this goal by employing random auxiliary mappings to the space of task transitions $\Gamma$. Intuitively:
\begin{itemize}
\item In each iteration we will first produce additional auxiliary evidence, resulting from task transitions within the current trajectory of $X$.
\item This evidence will be used next in order to significantly restrict the explorable range of network paths in the the following sampler iteration.
\end{itemize}
This approach poses a computationally tractable technique suited for the analysis of system transitions in QNs, and will construct a Markov chain of posterior trajectories over the entire range of paths in full agreement with the original observed evidence, where reasonably distant samples in the chain are statistically unrelated.
\subsubsection{Preliminaries}
Set $\Omega > \max_{x\in\mathcal{S}}|Q_x|$ and let $\boldsymbol{t}'$ and $\boldsymbol{x}'$ define some auxiliary frames of transition times and states in $X$, including both real and virtual values. Arrival, departure and job service observations must come at transition times in $\boldsymbol{t}'$; hence, we may define an augmented set of observations $\boldsymbol{O}_k' = \{O_{t_i'}\in\mathcal{O}\cup\mathcal{O}_3 : i=1,\dots,m\}$ at times $t_1',\dots,t_m'$, with
$$\mathcal{O}_3 = \{ \{(i,j,k)\in\Gamma : i,j\neq 0 \}\cup\{\varnothing\} \}$$
and such that $\boldsymbol{O}_k\subseteq\boldsymbol{O}_k'$. This accounts for missing observations; note that since arrivals and departures are always observed, a missing observation offers evidence for either an inner transition or virtual jump in the network. For simplicity, we assume that no state is reachable from itself in a transition, so that $\mathcal{T}(x,x)=\varnothing$; however, the framework naturally extends to networks where self-transitions are a possibility. Now, denote by $\boldsymbol{u}$ an auxiliary family of subsets of $\Gamma\cup\varnothing$, such that
\begin{align}
\ \mathbb{P}(u_i=u|x_{i-1}',x_{i}') = \begin{cases}
p, & \text{if } u= \{\mathcal{T}(x_{i-1}',x_{i}')\}, \\
1-p, & \text{if }u=\Gamma\cup\varnothing,
\end{cases} \label{auxes}
\end{align}
with some fixed $p\in[0,1]$, for all $u_i\in\boldsymbol{u}$, $i=1,\dots,m$. Hence, auxiliary variables $u\in\boldsymbol{u}$ will refer to either the entire space of task transitions or sets with a single element in $\Gamma$; we note that these single element sets will be further contained within a larger observation-set $O\in\boldsymbol{O}_k'$.
Recall that in queueing networks a task transition may follow from an infinite number of network configurations; that is, there may exist an infinite amount of task orderings across the stations so that a specific job is serviced in one given server and routed to another. However, any network state can only transition to a finite space, by relocating one task in a new queue after a service or an arrival. Thus $$\{(x,x')\}\subset\mathcal{T}^{-1}\mathcal{T}(x,x')\subset \mathcal{S}^2,$$ strictly, for all $x,x'\in\mathcal{S}$. Moreover, any $u\in\boldsymbol{u}$ such that $u\neq \Gamma\cup\varnothing$ can only be produced by a limited set of uniformized paths in $X$, and compatibility definitions in Definition \ref{comp} extend naturally to these auxiliary-observation variables. Restrictions are of two types:
\begin{itemize}
\item \textit{Transition triplets} impose a transition for an identifiable task. The transition probability is identical over all pairs of compatible states $(x,x')\in\mathcal{S}$.
\item \textit{Null sets} impose virtual jumps. The transition probability (lack thereof) depends both on the network configuration and dominating rate $\Omega$.
\end{itemize}
\subsubsection{Sampler}
\begin{algorithm*}[t]
\normalsize
\caption{Forward filtering backward sampling with dynamic arrays}
\label{backForwAlgo}
\begin{algorithmic}[1] \vspace{3pt}
\State Set initial state $x_0$ and let $\alpha_0(x)=\mathbb{I}(x= x_0)$, $x\in\mathcal{S}$.
\For {$i\in\{1,\dots,m\}$} \hfill \textit{Forward Filtering}
\For {$x\in\mathcal{S}$ s.t. $\alpha_{i-1}(x)>0$}
\For {$x'\in\mathcal{S}$ s.t. $|Q_{x,x'}|>0,(x,x')\in\mathcal{T}^{-1}(O_{t_i'}) \cap \mathcal{T}^{-1}(u_i)$}
\State Update: $$\alpha_{i}(x') \gets \alpha_{i}(x') + (1-q)^{\mathbb{I}(\mathcal{T}(x,x')\neq\varnothing)} \Big (\mathbb{I}(x=x')+ \frac{Q_{x,x'}}{\Omega}\Big) \alpha_{i-1}(x)$$
\EndFor
\EndFor
\EndFor
\State Sample $x_m$ from $x\in\mathcal{S}$ with probability in proportion to $\alpha_m(x)$.
\For {$i\in\{m-1,\dots,1\}$} \hfill \textit{Backward Sampling}
\For {$x\in\mathcal{S}$ s.t. $|Q_{x,x_{i+1}}|>0,\alpha_i(x)>0,(x,x_{i+1})\in \mathcal{T}^{-1}(O_{t_{i+1}'}) \cap \mathcal{T}^{-1}(u_{i+1})$}
\State Update: $$\beta(x) \gets \alpha_i(x) (1-q)^{\mathbb{I}(\mathcal{T}(x,x_{i+1})\neq\varnothing)} \Big(\mathbb{I}(x=x_{i+1})+ \frac{Q_{x,x_{i+1}}}{\Omega}\Big)$$
\EndFor
\State Sample $x_i$ from $x\in\mathcal{S}$ in proportion to $\beta(x)$.
\EndFor
\State Remove virtual transitions in $(\boldsymbol{t}',\boldsymbol{x}')$ and return $(\boldsymbol{t},\boldsymbol{x})$.
\end{algorithmic}
\end{algorithm*}
Let $(\boldsymbol{t},\boldsymbol{x})$ denote a network path in $X$ fully compatible with $\boldsymbol{O}_k$, with $\boldsymbol{t}=\{t_1,\dots,t_n\}$ and $\boldsymbol{x}=\{x_1,\dots,x_n\}$; then, marginalizing over $(\boldsymbol{x}',\boldsymbol{u})$ the frame $\boldsymbol{t}'$ is independent of any observations and such that (cf. \cite{rao13a})
\begin{align*}
f_{\boldsymbol{t}'}(&t_1',\dots,t_m'|(\boldsymbol{t}, \boldsymbol{x}),\Omega,\boldsymbol{\theta},x_0) = \mathbb{I}(\boldsymbol{t} \subseteq \{ t_1',\dots,t_m' \} ) \times \\
& \prod_{i=0}^n(\Omega+Q_{x_i})^{\#(\{t_1',\dots,t_m'\}\cap (t_i,t_{i+1}))} e^{-(\Omega+Q_{x_i})(t_{i+1}-t_i)}
\end{align*}
with $t_0=0$ and $t_{n+1}=T$. Thus, it may be sampled in a collapsed step incorporating virtual transitions to times in $\boldsymbol{t}$, employing a succession of Poisson processes with rates $\{\Omega+Q_{x_i}$ : $x_i\in\boldsymbol{x}\}$. Note that $\boldsymbol{t}'$ along with $(\boldsymbol{t},\boldsymbol{x})$ induces preliminary sequences of missing observations in $\boldsymbol{O}_k'$ and uniformized transitions in $\boldsymbol{x}'$. Next, we target $\boldsymbol{u}|\boldsymbol{x}'$ sampling $m$ auxiliary-observation variables from \eqref{auxes}.
Finally, we obtain a new path $(\boldsymbol{t},\boldsymbol{x})$ in full agreement with both real and auxiliary observations, producing $\boldsymbol{t},\boldsymbol{x},\boldsymbol{x}'$ in a blocked step. This simplifies to sampling a sequence $\boldsymbol{x}'|\boldsymbol{t}',\boldsymbol{u},\Omega,\boldsymbol{\theta},\boldsymbol{O}_k',x_0$ and removing virtual entries; it is achieved by employing dynamic arrays within a procedure for discrete-time state-space models as shown in Algorithm \ref{backForwAlgo}. Alternatively, note it is possible to employ a particle filtering approach within a forward procedure, in order to impose further memory constraints.
\subsubsection{Properties and considerations}
Along with observations and naturally restrictive constraints on state transitions within QNs, auxiliary variables in $\boldsymbol{u}$ allow us to limit the space a sampler is allowed to explore within each iteration. These restrictions apply both within forward and backward procedures and leave the underlying filtering equations unaltered, up to proportionality. Increasing the value of $p$ will make computationally expensive iterations less likely, at the cost of a higher dependence between subsequent realizations of $X$. Also, the term $(1-q)$ enters the forward procedure penalizing network paths with unobserved transitions and is only proportionally relevant when no observation exists.
\begin{figure*}[t]
\centering
\begin{tikzpicture}
\node at (0.5cm,4cm) {\small \textbf{Iteration} $\boldsymbol{n}$};
\draw[->,line width=0.20mm] (0cm,3cm) -- (15cm,3cm);
\draw (0,2.9cm) -- ++(0cm,0.2cm);
\node at (15.3cm,3cm) {$\boldsymbol{t}$};
\draw[-,dashed,color=black!80!white,line width=0.20mm] (1.9,4.58cm) -- ++(0cm,-7.7cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (4.2,4.58cm) -- ++(0cm,-5.5cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (10.3,4.58cm) -- ++(0cm,-5.5cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (14.4,4.58cm) -- ++(0cm,-7.7cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (1.9,5.08cm) -- ++(12cm,0cm);
\filldraw[black!60!white,draw=black] (1.9cm,3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.2cm,3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (5.7cm,3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (10.3cm,3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (12.1cm,3cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (14.4cm,3cm) circle [radius=0.09cm];
\node[fill=white] at (1.93cm,2.6cm) {\footnotesize $1_{0\rightarrow 1}$};
\node[fill=white] at (4.23cm,2.6cm) {\footnotesize $2_{0\rightarrow 1}$};
\node[fill=white] at (5.73cm,2.6cm) {\footnotesize $1_{1\rightarrow 3}$};
\node[fill=white] at (10.33cm,2.6cm) {\footnotesize $1_{3\rightarrow 0}$};
\node[fill=white] at (12.13cm,2.6cm) {\footnotesize $2_{1\rightarrow 3}$};
\node[fill=white] at (14.43cm,2.6cm) {\footnotesize $2_{3\rightarrow 0}$};
\node[fill=white] at (7.2cm,5.08) {\small \textbf{Evidence}};
\filldraw[black!05!white,draw=black] (1.45cm,4.8cm) rectangle ++(0.9,0.6);
\node at (1.9cm,5.08cm) {\small $\boldsymbol{1_{0\rightarrow \cdot}}$};
\filldraw[black!05!white,draw=black] (3.75cm,4.8cm) rectangle ++(0.9,0.6);
\node at (4.2cm,5.08cm) {\small $\boldsymbol{2_{0\rightarrow \cdot}}$};
\filldraw[black!05!white,draw=black] (9.85cm,4.8cm) rectangle ++(0.9,0.6);
\node at (10.3cm,5.08cm) {\small $\boldsymbol{1_{\cdot\rightarrow 0}}$};
\filldraw[black!05!white,draw=black] (13.95cm,4.8cm) rectangle ++(0.9,0.6);
\node at (14.4cm,5.08cm) {\small $\boldsymbol{2_{\cdot\rightarrow 0}}$};
\node[fill=white] at (1.07cm,1.5cm) {\small \textbf{Add virtual jumps}};
\draw[->,line width=0.20mm] (0cm,0.5cm) -- (15cm,0.5cm);
\draw (0,0.4cm) -- ++(0cm,0.2cm);
\node at (15.3cm,0.5cm) {$\boldsymbol{t}$};
\node[fill=white] at (1.93cm,0.1cm) {\footnotesize $1_{0\rightarrow 1}$};
\node[fill=white] at (4.23cm,0.1cm) {\footnotesize $2_{0\rightarrow 1}$};
\node[fill=white] at (5.73cm,0.1cm) {\footnotesize $1_{1\rightarrow 3}$};
\node[fill=white] at (10.33cm,0.1cm) {\footnotesize $1_{3\rightarrow 0}$};
\node[fill=white] at (12.13cm,0.1cm) {\footnotesize $2_{1\rightarrow 3}$};
\node[fill=white] at (14.43cm,0.1cm) {\footnotesize $2_{3\rightarrow 0}$};
\filldraw[black!60!white,draw=black] (1.9cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (0.5cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (0.7cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (1.1cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (1.3cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.2cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (3.4cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (3.6cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (5.5cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!60!white,draw=black] (5.7cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (4.5cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (4.7cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (10.3cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (6.3cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (6.5cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (6.9cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (7.1cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (8.4cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (8.6cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (9.2cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (9.4cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (11.9cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!60!white,draw=black] (12.1cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!05!white,draw=black] (13.6cm,0.3cm) rectangle ++(0.4,0.4);
\filldraw[black!00!white,draw=black] (13.8cm,0.5cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (14.4cm,0.5cm) circle [radius=0.09cm];
\draw[-,dashed,color=black!80!white,line width=0.20mm] (0.7,0.3cm) -- ++(0cm,-1.3cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (1.3,0.3cm) -- ++(0cm,-1.3cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (3.6,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (4.7,0.3cm) -- ++(0cm,-1.3cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (5.7,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (6.5,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (7.1,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (8.6,0.3cm) -- ++(0cm,-1.3cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (9.4,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (12.1,0.3cm) -- ++(0cm,-3.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (13.8,0.3cm) -- ++(0cm,-3.4cm);
\node at (0cm,-1cm) {\large $\boldsymbol{u}$:};
\node[fill=white] at (0.7cm,-1cm) {\xmark};
\node[fill=white] at (1.3cm,-1cm) {\xmark};
\node[fill=white] at (1.9cm,-1cm) {\cmark};
\node[fill=white] at (3.6cm,-1cm) {\cmark};
\node[fill=white] at (4.2cm,-1cm) {\xmark};
\node[fill=white] at (4.7cm,-1cm) {\xmark};
\node[fill=white] at (5.7cm,-1cm) {\cmark};
\node[fill=white] at (6.5cm,-1cm) {\cmark};
\node[fill=white] at (7.1cm,-1cm) {\cmark};
\node[fill=white] at (8.6cm,-1cm) {\xmark};
\node[fill=white] at (9.4cm,-1cm) {\cmark};
\node[fill=white] at (10.3cm,-1cm) {\xmark};
\node[fill=white] at (12.1cm,-1cm) {\cmark};
\node[fill=white] at (13.8cm,-1cm) {\cmark};
\node[fill=white] at (14.4cm,-1cm) {\cmark};
\node[fill=white] at (2.01cm,-2.2cm) {\small \textbf{Empty frame, project evidence}};
\draw[->,line width=0.20mm] (0cm,-3.2cm) -- (15cm,-3.2cm);
\draw (0,-3.3cm) -- ++(0cm,0.2cm);
\node at (15.3cm,-3.2cm) {$\boldsymbol{t}$};
\filldraw[black!00!white,draw=black] (1.9cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (3.6cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.2cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (5.7cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (6.5cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (7.1cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (9.4cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (10.3cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (12.1cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (13.8cm,-3.2cm) circle [radius=0.09cm];
\filldraw[black!00!white,draw=black] (14.4cm,-3.2cm) circle [radius=0.09cm];
\draw[-,dashed,color=black!80!white,line width=0.20mm] (4.2,-3.3cm) -- ++(0cm,-2.4cm);
\draw[-,dashed,color=black!80!white,line width=0.20mm] (10.3,-3.3cm) -- ++(0cm,-2.4cm);
\node[fill=white] at (4.23cm,-3.6cm) {\footnotesize $2_{0\rightarrow 1}$};
\node[fill=white] at (10.33cm,-3.6cm) {\footnotesize $1_{3\rightarrow 0}$};
\node[fill=white] at (1.32cm,-4.7cm) {\small \textbf{FFBS: iteration $\boldsymbol{n+1}$}};
\draw[->,line width=0.20mm] (0cm,-5.7cm) -- (15cm,-5.7cm);
\draw (0,-5.8cm) -- ++(0cm,0.2cm);
\node at (15.3cm,-5.7cm) {$\boldsymbol{t}$};
\filldraw[black!60!white,draw=black] (1.9cm,-5.7cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (4.2cm,-5.7cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (6.5cm,-5.7cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (7.1cm,-5.7cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (10.3cm,-5.7cm) circle [radius=0.09cm];
\filldraw[black!60!white,draw=black] (14.4cm,-5.7cm) circle [radius=0.09cm];
\node[fill=white] at (1.93cm,-6.1cm) {\footnotesize $1_{0\rightarrow 2}$};
\node[fill=white] at (4.23cm,-6.1cm) {\footnotesize $2_{0\rightarrow 1}$};
\node[fill=white] at (6.44cm,-6.1cm) {\footnotesize $1_{2\rightarrow 3}$};
\node[fill=white] at (7.22cm,-6.1cm) {\footnotesize $2_{1\rightarrow 3}$};
\node[fill=white] at (10.33cm,-6.1cm) {\footnotesize $1_{3\rightarrow 0}$};
\node[fill=white] at (14.43cm,-6.1cm) {\footnotesize $2_{3\rightarrow 0}$};
\end{tikzpicture}
\vspace{10pt}
\caption{Task transition diagram with a single iteration in the proposed sampler, for a bottleneck network with three servers. Here, tasks 1 and 2 are observed entering and leaving the network. First, start with a path whose task transitions are fully compatible with the evidence. Then, supplement it with virtual jumps at the corresponding rates, and produce auxiliary variables across real and virtual nodes. Next, empty the uniformized frame and propagate both real and auxiliary evidence; imposing task transitions or virtual jumps within clamped nodes. Finally, repopulate the frame via forward filtering backward sampling, hence maintaining agreement with the existing evidence.} \label{samplerIter}
\end{figure*}
In Figure \ref{samplerIter} we observe a task transition diagram with a single iteration in the proposed sampler, for the bottleneck network in Figure \ref{exmplQueues} (top). In this example, two tasks (numbered 1 and 2) are observed entering and leaving the network at different times; however, there exists no information regarding job services within the network. In each iteration, the sampler begins with a network path whose task transitions are fully compatible with the existing evidence. In an initial step, the existing path is supplemented with virtual transitions at the corresponding Poisson rates. In the Figure, we observe that nodes for both virtual jumps and the unobserved job services are superimposed over shaded boxes; the boxes represent further evidence for the lack of task arrivals or departures at these times. Next, auxiliary variables are produced across real and virtual jumps, the subsets are loosely represented by ticks ($\Gamma\cup\varnothing$) and crosses ($\{\mathcal{T}(x'_{i-1},x'_i)\}$) for open and clamped nodes respectively. Then, the uniformized frame is emptied and both real and auxiliary evidence is propagated, imposing task transitions or virtual jumps within clamped nodes and resulting in a restricted frame for possible network paths. Finally, a new compatible path is sampled via forward filtering backward sampling as summarized in Algorithm \ref{backForwAlgo}; this will consider the imposed task transitions and weight successive network states over the clamped epochs. The resulting path is fully compatible with the observed evidence, however, notice that task transitions at arrival or departure times may change between iterations.
Note that by choosing $\Omega$ strictly greater than all values in the diagonal of $Q$, the resulting Markov chain over posterior network transitions is irreducible. Increasing the dominating rate will improve mixing in exchange for higher computational requirements. Finally, we note that a high value of $p$ may hinder the sampler from fully exploring the posterior range of network paths.
\subsubsection{Parameter sampling}
Finally, given a new family of network realizations $\boldsymbol{X}=\{X^k\}_{k=1,\dots,K}$ fully compatible with observation sequences $\tilde{\boldsymbol{O}}=\{\boldsymbol{O}_k\}_{k=1,\dots,K}$, we may obtain posterior samples of arrival and service rate parameters. For traditional FCFS stations this is such that
$$\lambda_c | \boldsymbol{X} \sim \text{Gamma}\big( \textstyle\delta_c, \sum_{k=1}^K T_k \big)$$ and
$$\mu^c_i | \boldsymbol{X} \sim \text{Gamma}\big( \gamma^c_i, \tau^c_i \big),$$
for $c\in\mathcal{C}, i=1,\dots,M$; assuming independent network parameters and uninformative priors. Here $\delta_c, \gamma^c_i$ and $\tau^c_i$ denote respectively the number of class $c$ arrivals, class $c$ jobs served at station $i$ and the time server $i$ has been occupied by a class $c$ job, in all realizations in $\boldsymbol{X}$. Finally, posterior probability vectors for class $c$ routings in every node $i=1,\dots,M$ are given by
$$P^c_{i,\cdot} | \boldsymbol{X} \sim \text{Dir}\big( \boldsymbol{1} + \boldsymbol{\kappa}^c_i \big)$$
where $\boldsymbol{\kappa}^c_i$ defines a vector of transition counts from server $i$ in $\boldsymbol{X}$. Arrival posteriors in $\boldsymbol{p}$ are defined the same way. We note that in order to ease identifiability in the inferential problem, it is also possible to fix parameters, incorporate conjugate priors or to impose inequality constraints and bounds across parameters; we will show examples in Section \ref{Examps} below. Also, the above expressions must be altered when stations respond to prioritization regimes other than FCFS (see Example 3 in Section \ref{Examps}).
\section{Examples} \label{Examps}
In the following, we discuss results obtained across three example networks with both synthetic and real data, in order of increasing difficulty. In all cases, results are obtained through a JAVA implementation of the proposed sampler, and starting compatible network paths have been manually assigned.
The examples demonstrate the ability of the proposed algorithm in order to handle missing data in multi-class inferential problems with varying service disciplines, class switching and imposed prior constraints. Hence, the sampler offers the means to overcome necessary assumptions linked to the common use of product form equilibrium expressions for QNs. To the best of our knowledge, there exists no alternative approach overcoming these restrictions when drawing exact inference in general open Markovian networks.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{Example1plot}
\caption{Graphical summary of output for a single-class tandem network fitted to synthetic data. On the left, we observe the posterior distribution of service rate parameters with original values $\mu_1=0.2$ and $\mu_2=0.5$. Black bars in the center correspond to autocorrelation values. On the right, a contour plot for the joint posterior density of rates (dashed lines represent real values).} \label{CorTandem}
\end{figure*}
\subsection{Tandem network} In the simplest example, we analyse simulated data on a tandem network with two $M/M/1$ stations, FCFS service disciplines and a single task class. Data is generated so that true service rates are $\mu_1=0.2$ and $\mu_2=0.5$, arrivals are given by $\lambda=0.12$ and the network topology is defined by a routing probability matrix $P$ such that $P_{1,2} = 1$ and $P_{2,0}=1$. Also, jobs enter directly into the first queue and $p_{0,1} = 1$.
For the inferential problem, job service observations (in first station) are always ignored and the only source of information are end-to-end measurements. Thus, available knowledge is limited to the times when tasks enter the queue on the first station and when they depart through the second station. Overall, we examine 5000 realizations totalling 17827 tasks during 115601 time units. For the purpose, the network topology in $P$ is fixed deterministic, since there exists a unique route from start to completion of tasks. Also, in order to ensure identifiability we impose an inequality constraint on service rates and assign fairly uninformative parameter priors, so that $$\pi(\mu_1,\mu_2)\propto \mathbb{I}(\mu_1\leq\mu_2)\times\exp(-10^{-3}(\mu_1+\mu_2)).$$
Note that the problem directly links to the inferential task with two exponentially distributed random variables when only its sum is observed, with the further complexity that unknown waiting times have to be discounted from the empirical observations.
In Figure \ref{CorTandem} (right) we observe a contour plot for the joint posterior kernel density estimation over service rates, and we notice a significant negative correlation in values (the dashed vertical and horizontal lines represent the original parameters values in the network path simulations). Results are obtained across two chains with 100000 iterations each, a 10000 burn-in stage, varying starting rates and different scales for dominating rates and probabilities producing auxiliary-observations, so that $p_1=0$ and $\Omega_1=2\max_{x\in\mathcal{S}}|Q_x|$ and $p_2=0.25, \Omega_2=1.5\max_{x\in\mathcal{S}}|Q_x|$. Note that the second chain is produced employing restrictive auxiliary-observations as opposed to the first; hence, stronger serial dependencies across subsequent latent paths in the network should be expected. Yet, the remainder plots show marginal posterior kernel density estimations for both service rates, along with an autocorrelation summary across a thinned sample in the second chain, showing a satisfactory mixing.
A discussion on the effects and computational gains resulting from employing restrictive auxiliary observations follows in the next example. In general, networks of interest are complex and $p=0$ would pose a computationally infeasible problem. Also, even in simple networks such as this example, computing times can be excessive, and considerable reductions can be traded at the cost of higher serial dependences.
\begin{table*}[t]
\caption{Summary statistics for posterior service rates along with computing times across three chains tuned differently in the bottleneck network in Figure \ref{exmplQueues} (top).} \label{summaryMCMC}
\renewcommand{\arraystretch}{1.3}
\setlength{\tabcolsep}{17pt}
\centering
\begin{tabular}{lllllllll}
\hline
& Real & \multicolumn{2}{l}{Summary} & \multicolumn{5}{l}{Quantiles} \\ \cline{3-9}
& & Mean & StDev & 2.5\% & 25\% & 52\% & 75\% & 97.5\% \\ \hline
$\mu_1^1$ & \textbf{0.3} & 0.311 & 0.019 & 0.273 & 0.297 & 0.310 & 0.323 & 0.350 \\
$\mu_1^2$ & \textbf{0.25} & 0.237 & 0.017 & 0.204 & 0.225 & 0.236 & 0.247 & 0.271 \\
$\mu_1^3$ & \textbf{0.2} & 0.185 & 0.016 & 0.154 & 0.173 & 0.184 & 0.196 & 0.218 \\
$\mu_2^1$ & \textbf{0.7} & 0.709 & 0.052 & 0.613 & 0.673 & 0.707 & 0.743 & 0.818 \\
$\mu_2^2$ & \textbf{0.5} & 0.537 & 0.040 & 0.462 & 0.509 & 0.536 & 0.564 & 0.620 \\
$\mu_2^3$ & \textbf{0.3} & 0.297 & 0.028 & 0.245 & 0.277 & 0.296 & 0.315 & 0.354 \\
$\mu_3^1$ & \textbf{1.5} & 1.626 & 0.088 & 1.458 & 1.566 & 1.625 & 1.684 & 1.802 \\
$\mu_3^2$ & \textbf{1.2} & 1.206 & 0.069 & 1.074 & 1.159 & 1.204 & 1.252 & 1.346 \\
$\mu_3^3$ & \textbf{0.8} & 0.728 & 0.052 & 0.630 & 0.692 & 0.727 & 0.762 & 0.833 \\ \hline \addlinespace[0.1cm] \hline
& $p$ & \multicolumn{2}{l}{$\Omega$} & \multicolumn{2}{l}{Run Time} & \textbf{ESS:} & Mean & Min \\ \cline{1-9}
1 & 0.7 & \multicolumn{2}{l}{$2\max_{x\in\mathcal{S}}|Q_x|$} & \multicolumn{2}{l}{6056.5s} & & 14575 & 7246 \\
2 & 0.5 & \multicolumn{2}{l}{$1.5\max_{x\in\mathcal{S}}|Q_x|$} & \multicolumn{2}{l}{24709.4s} & & 26069 & 14750 \\
3 & 0.2 & \multicolumn{2}{l}{$1.2\max_{x\in\mathcal{S}}|Q_x|$} & \multicolumn{2}{l}{73993.2s} & & 42101 & 22460 \\ \hline
\end{tabular}
\end{table*}
\subsection{Bottleneck network} We examine simulated data in the bottleneck network in Figure \ref{exmplQueues} (top), with 3 FCFS stations and 3 different task classes. The true service rates can be observed in Table \ref{summaryMCMC}, and task arrivals are given by $\lambda_1=0.08, \lambda_2=0.06$ and $\lambda_3=0.04$. In this case, along with end-to-end measurements, approximately half of all generated job service observations are retrieved so that $q=0.5$. The network topology is defined by $\{P^c,p^c:c\in\mathcal{C}\}$, where
$$P^c= \bordermatrix{ & \ 0 \ & 1 \ & 2 \ & 3 \ \cr
1 & \ 0 \ & 0 \ & 0 \ & 1 \ \cr
2 & \ 0 \ & 0 \ & 0 \ & 1 \ \cr
3 & \ 1 \ & 0 \ & 0 \ & 0 \ },$$
is identical for all three classes and assumed to be known. In addition, job entries are split evenly, i.e. $p^c_{0,1} = 0.5$ and $p^c_{0,2} = 0.5$ for all $c\in\mathcal{C}$.
In total, we analyse 500 network realizations totalling 1281 tasks during 5083 time units. In order to ease identifiability we assume the existence of a slow, medium and fast server; and assign rather uninformative parameter priors, i.e.
\begin{align*}
\pi(\mu^c_1,&\mu^c_2,\mu^c_3) \\
& \propto \mathbb{I}(\mu^c_1\leq\mu^c_2\leq\mu^c_3)\times\exp(-10^{-3}(\mu^c_1+\mu^c_2+\mu^c_3))
\end{align*}
for all $c\in\mathcal{C}$. Note that this network type may not be analysed by means of product-form representations centred around figures of queue-lengths (c.f \cite{Wang2016}). This is because traditional BCMP networks require FCFS stations to share service rates across task classes. On the other hand, an MCMC sampler as presented in \citep{sutton2011} can be extended in order to handle general service distributions and target network path transitions; however, the framework is not designed for such aim, it would require an additional Metropolis-Hastings step and it is likely to perform poorly.
\begin{table}[h!]
\caption{Correlation matrix between service rate parameters in a bottleneck network.}
\label{corMat}
\renewcommand{\arraystretch}{1.3}
\setlength{\tabcolsep}{4pt}
\centering
\begin{tabular}{crrrrrrrr}
\hline
& $\mu_1^2$ & $\mu_1^3$ & $\mu_2^1$ & $\mu_2^2$ & $\mu_2^3$ & $\mu_3^1$ & $\mu_3^2$ & $\mu_3^3$ \\
\hline
$\mu_1^1$ & -0.02 & -0.01 & -0.05 & 0.01 & 0.01 & -0.05 & 0.01 & 0.02 \\
$\mu_1^2$ & & -0.02 & 0.00 & -0.03 & 0.01 & 0.01 & -0.04 & 0.01 \\
$\mu_1^3$ & & & 0.01 & 0.01 & -0.09 & 0.00 & -0.01 & -0.06 \\
$\mu_2^1$ & & & & -0.03 & -0.05 & -0.13 & 0.01 & 0.01 \\
$\mu_2^2$ & & & & & -0.02 & 0.00 & -0.10 & 0.01 \\
$\mu_2^3$ & & & & & & 0.00 & 0.00 & -0.10 \\
$\mu_3^1$ & & & & & & & 0.01 & 0.01 \\
$\mu_3^2$ & & & & & & & & 0.01 \\ \hline
\end{tabular}
\end{table}
In Table \ref{summaryMCMC} we observe summary statistics, computing times and effective sample sizes across three chains with 100000 iterations each, a 10000 burn-in stage, varying starting rates and different scales for dominating rates $\Omega$ and probabilities $p$ producing auxiliary-observations. There, we notice a good trade-off between effective samples and the drastic decrease in computing times required when imposing strong serial relations on network paths across subsequent iterations in the sampler. This is the case even when the volume of virtual jumps produced is reduced, and emphasizes the need for such slice sampling techniques in inferential problems with QNs. In addition, Table \ref{corMat} displays the overall posterior correlation matrix between service rate parameters, and shows very mild relations in rates for each task class. There, we notice the importance of employing posterior samples from the produced chains in order to answer extrapolation-type questions in network systems. Finally, note it is possible to ease imposed restrictions on the network topology and to employ different service disciplines across servers (see next example).
\subsection{Feedback network} Finally, we show how the proposed sampling scheme may be used to analyse a real data set. For the purpose, we employ work-logs for medical clinicians. Briefly, the data set includes task requests and completions for individual doctors outside the 9:00-17:00 Monday to Friday \textit{in hours} settings. It belongs to two jointly coordinated university hospitals in the United Kingdom, together servicing a geographical region with over 2.5 million residents.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[->,line width=0.20mm] (0cm,0cm) -- (7.6cm,0cm);
\draw (0,-0.1cm) -- ++(0cm,0.2cm);
\draw (1.2,-0.1cm) -- ++(0cm,0.2cm);
\draw (2.4,-0.1cm) -- ++(0cm,0.2cm);
\draw (2.9,-0.1cm) -- ++(0cm,0.2cm);
\draw (4.2,-0.1cm) -- ++(0cm,0.2cm);
\draw (4.8,-0.1cm) -- ++(0cm,0.2cm);
\draw (6.9,-0.1cm) -- ++(0cm,0.2cm);
\node at (7.9cm,0cm) {$\boldsymbol{t}$};
\node at (0.00cm,-0.3cm) {\footnotesize $0$};
\node at (1.21cm,-0.3cm) {\footnotesize $t_1$};
\node at (2.41cm,-0.3cm) {\footnotesize $t_2$};
\node at (2.91cm,-0.3cm) {\footnotesize $t_3$};
\node at (4.21cm,-0.3cm) {\footnotesize $t_4$};
\node at (4.81cm,-0.3cm) {\footnotesize $t_5$};
\node at (6.91cm,-0.3cm) {\footnotesize $t_6$};
\draw[-,dashed,line width=0.20mm] (1.2cm,0cm) -- ++ (0cm,0.75cm);
\draw[-,dashed,line width=0.20mm] (2.4cm,0cm) -- ++ (0cm,1.5cm);
\draw[-,dashed,line width=0.20mm] (2.9cm,0cm) -- ++ (0cm,2.25cm);
\draw[-,dashed,line width=0.20mm] (4.2cm,0cm) -- ++ (0cm,0.75cm);
\draw[-,dashed,line width=0.20mm] (4.8cm,0cm) -- ++ (0cm,2.25cm);
\draw[-,dashed,line width=0.20mm] (6.9cm,0cm) -- ++ (0cm,1.5cm);
\draw[black!70!white,line width=1.3mm] (1.2cm,0.75cm) -- (4.2cm,0.75cm);
\node at (0.75cm,0.75cm) {\footnotesize \textbf{Fall}};
\draw[black!70!white,line width=1.3mm] (2.4cm,1.5cm) -- (6.9cm,1.5cm);
\node at (1.6cm,1.5cm) {\footnotesize \textbf{Clerking}};
\draw[black!70!white,line width=1.3mm] (2.9cm,2.25cm) -- (4.8cm,2.25cm);
\node at (2.15cm,2.25cm) {\footnotesize \textbf{Urgency}};
\end{tikzpicture}
\caption{Sample diagram with a subset of tasking data linked to a clinician during a shift} \label{sampleData}
\end{figure}
In Figure \ref{sampleData} we show a diagram with a small of subset of data linked to a clinician during a shift; there we observe three overlapping tasks recorded in the system (from request to completion), and each belonging to a different class. Note that it is not possible to know when the clinician was engaged with each duty; as individual jobs for tasks are not registered when queueing or being routed across teams of administrative staff, nurses and doctors. An extended description of the data set may be found in \cite{Perez201634}.
Multiple tasks are grouped across 14 categories and analysed with a \textit{feedback network} as shown in Figure \ref{feedbackQ}. There, we notice the presence of two M/M/1 servers with alternative disciplines and route switching among classes. Task observations for doctors are of roughly two kinds, based on whether they require \textit{engagement} or not. Many tasks are recorded and erased within doctor work-logs in a short time span, due to no need for action; on the other hand, the remainder of tasks exhibit long processing times indicating the need for considerable doctor activity.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw (7,0) -- ++(1.4cm,0) -- ++(0,-0.8cm) -- ++(-1.4cm,0);
\foreach \i in {1,...,4}
\draw (8.4cm-\i*8pt,0) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (8.4cm+0.41cm,-0.4cm) circle [radius=0.4cm];
\draw (9.7,-1.4cm) -- ++(-1.4cm,-0cm) -- ++(0cm,-0.8cm) -- ++(1.4cm,0);
\foreach \i in {1,...,4}
\draw (8.3cm+\i*8pt,-1.4cm) -- +(0,-0.8cm);
\filldraw[fill=black!05!white] (8.3cm-0.41cm,-1.4cm-0.4cm) circle [radius=0.4cm];
\filldraw[fill=black!05!white] (10.3cm,-0.1cm) rectangle (11cm,-0.7cm);
\draw[-,line width=0.2mm] (9.3cm,-0.4cm) -- ++(1,0.0cm);
\draw[->,line width=0.20mm] (8.805,0.1cm) -- ++(0.0,0.4cm);
\draw[->,line width=0.20mm] (11,-0.4cm) -- ++(1,0cm);
\draw[->,line width=0.20mm] (10.65cm,-0.7cm) -- ++(0cm,-1.1cm) -- ++(-0.85cm,0cm);
\draw[->,line width=0.20mm] (6.1cm,-0.5cm) -- ++(0.8,0.0cm);
\draw[<-,line width=0.20mm] (6.9cm,-0.3cm) -- ++(-1.8cm,0cm);
\draw[line width=0.20mm] (7.4cm,-1.8cm) -- ++(-1.3cm,0cm) -- ++(0cm,1.3cm);
\node[align=center] at (10.65cm,-0.4cm) {\small $P$};
\node[align=center] at (8.8cm,-0.4cm) {\tiny $FCFS$};
\node[align=center] at (7.9cm,-1.8cm) {\tiny $PS$};
\end{tikzpicture}
\vspace{10pt}
\caption{Feedback network with two M/M/1 servers and route switching among task classes.} \label{feedbackQ}
\end{figure}
In the proposed example, arrival jobs are buffered within an administrative FCFS priority type queue and depart to a transition center where they either leave the system or get routed for processing with some unknown probability. Once they are assigned to further processing, they join the doctor's processing centre and switch their routing mechanism; so they will depart the network next time they undergo administrative processing in the first queue. The service station aimed to capture strain on doctor workload is assigned a \textit{processor sharing} (PS) discipline with a single worker, aiming to accommodate doctors attending concurrent duties outside standard working hours. No job service observations are available, so that $q=0$ and only the arrival and departure times for tasks to the network are observed.
\begin{table}[b!]
\caption{Summary statistics for rate and routing parameters in a feedback network.} \label{summaryMCMC2pars}
\renewcommand{\arraystretch}{1.3}
\setlength{\tabcolsep}{12pt}
\centering
\begin{tabular}{lllll}
\hline
& & \multicolumn{2}{l}{Summary} \\ \cline{3-4}
& & Mean & StDev \\ \hline
Buffer & $\mu_1$ & 3.397 & 0.072 \\
Admission & $\mu_2^1$ & 0.587 & 0.038 \\
& $P^1_{1,0}$ & 0.250 & 0.026 \\
Certification death & $\mu_2^2$ & 0.723 & 0.159 \\
& $P^2_{1,0}$ & 0.601 & 0.061 \\
Check patient & $\mu_2^{12}$ & 0.650 & 0.051 \\
& $P^{12}_{1,0}$ & 0.399 & 0.033 \\
Clerking & $\mu_2^3$ & 0.731 & 0.018 \\
& $P^3_{1,0}$ & 0.061 & 0.007 \\
Clinical review & $\mu_2^4$ & 0.608 & 0.015 \\
& $P^4_{1,0}$ & 0.333 & 0.014 \\
Address relatives & $\mu_2^5$ & 0.361 & 0.079 \\
& $P^5_{1,0}$ & 0.470 & 0.098 \\
Drug prescribing & $\mu_2^6$ & 0.920 & 0.028 \\
& $P^6_{1,0}$ & 0.547 & 0.012 \\
Early warning & $\mu_2^7$ & 0.488 & 0.017 \\
& $P^7_{1,0}$ & 0.346 & 0.017 \\
Fall of patient & $\mu_2^8$ & 0.575 & 0.080 \\
& $P^8_{1,0}$ & 0.318 & 0.054 \\
None of above & $\mu_2^{10}$ & 0.339 & 0.013 \\
& $P^{10}_{1,0}$ & 0.393 & 0.018 \\
Other services & $\mu_2^9$ & 0.209 & 0.012 \\
& $P^9_{1,0}$ & 0.435 & 0.024 \\
Procedure request & $\mu_2^{11}$ & 0.635 & 0.027 \\
& $P^{11}_{1,0}$ & 0.387 & 0.019 \\
Test request & $\mu_2^{13}$ & 0.565 & 0.013 \\
& $P^{13}_{1,0}$ & 0.263 & 0.011 \\
Urgent response & $\mu_2^{14}$ & 0.209 & 0.028 \\
& $P^{14}_{1,0}$ & 0.306 & 0.059 \\ \hline
\end{tabular}
\end{table}
In total, we analyse a reduced subset of 10000 doctor shifts roughly distributed across 4 years of observations. The network topology is partially known; i.e. $P^c_{2,1} = p^c_{0,1} = 1$ for all task classes, and $P^c_{1,0} = 1$ after tasks have undergone processing and hence switched routing mechanism. However, $P^c_{1,0} = 1 - P^c_{1,2}$ needs to be determined for all existing task classes. Processing rates for tasks are assumed equal in the first service station and different in the PS server; we assign no constraints and we impose loosely uninformative priors such that
$$\pi(\mu^c_i) \propto \exp(-10^{-3}\mu^c_i)$$
for all $i\in\{1,2\}$ and $c\in\mathcal{C}$. Also, note that within a PS discipline posterior rates given network realizations are given by
$$\mu^c | \boldsymbol{X} \sim \text{Gamma}\big( \textstyle \gamma^c, \sum_{k=1}^K \int_{0}^{T_k} \phi_c^k(t) \mathrm{d}t \big),$$
for all $c\in\mathcal{C}$, where $\gamma^c$ denotes the number of class $c$ jobs served at the station in all realizations in $\boldsymbol{X}$; and
$$\phi_c^k(t) = \frac{\textstyle \sum_{j=0}^{J} \mathbb{I}(\text{Job $j$ is class $c$}) \cdot \mathbb{I}(a_j < t < d_j)}{\textstyle \sum_{j=0}^{J} \mathbb{I}(a_j < t < d_j)},$$
where summations are across all jobs processed in the PS station in realization $k$, and $a_j,d_j$ denote the arrival and departure times of the job to the server. In Table \ref{summaryMCMC2pars} we observe summary statistics for parameters across two chains with 100000 iterations each, a 50000 burn-in stage and varying starting rates. In one chain, we use $\Omega=2\max_{x\in\mathcal{S}}|Q_x|$ and $p = 0.75$; in the second we have $\Omega=1.5\max_{x\in\mathcal{S}}|Q_x|$ and $p = 0.5$.
\begin{table}[h!]
\caption{Point estimates and standard errors for average processing times excluding waiting times, across different tasks. These relate to times from entry to departure in network.} \label{summaryMCMC2}
\renewcommand{\arraystretch}{1.3}
\setlength{\tabcolsep}{5pt}
\centering
\begin{tabular}{llllll}
\hline
& \multicolumn{2}{l}{Completion} & & \multicolumn{2}{l}{Completion} \\ \cline{2-3} \cline{5-6}
& Mean & StErr & & Mean & StErr \\ \hline
Admission & 1.796 & 0.087 & Early w. & 1.828 & 0.046 \\
Cert. death & 0.985 & 0.141 & Fall & 1.701 & 0.176 \\
Check & 1.400 & 0.073 & None & 2.263 & 0.076 \\
Clerking & 1.856 & 0.032 & Other & 3.173 & 0.173 \\
Cl. review & 1.588 &0.027 & Procedure & 1.442 & 0.045 \\
Address rel. & 1.971 & 0.374 & Test & 1.818 & 0.032 \\
Drug pres. & 0.920 & 0.017 & Urgent & 3.877 & 0.491 \\ \hline
\end{tabular}
\end{table}
In addition, Table \ref{summaryMCMC2} shows point estimates and standard errors for average completion times in all task types, these correspond to the full processing times from entry to departure in the network (excluding queueing times) and are reported in hour units. Hence, we notice it is possible to assess workload both globally and across single components in the system, thus allowing to answer extrapolation kinds of questions on workload; i.e. in relation to means, variances and extreme values for system strain under likely alterations.
\section{Discussion} \label{Discussion}
This paper has presented a flexible approach for carrying exact Bayesian inference within known or hypothesized queueing networks. Its focus is on multi-class, open and Markovian cases and the approach is centred around the underlying continuous-time Markov chains induced by these complex stochastic systems. The proposed method relies on a slice sampling technique with mappings to the space of task transitions across servers in the network. It sits well over uniformization-oriented MCMC approaches introduced in \cite{rao13a} and can deal with missing data, imposed prior knowledge and strong serial dependencies posing a complex inferential task (cf. \cite{sutton2011}).
The need for such inferential frameworks with missing data is justified by the ability of general-form networks to allow evaluating response times in complex systems. Overall, recovering measures such as processing times is a technically difficult task when designing increasingly complex IT systems \citep{Liu200636}, or in service delivery networks (such as those in hospitals) due to ethical issues with such an intrusive process \citep{Perez201634}. Yet, QNs provide the tools to assess system alterations, diagnose poor performance or evaluate robustness to spikes in workload.
The advantage of the presented inferential method is that it permits retrospectively assessing the likely status of systems at any point in time; rather than only providing summary information on strain over individual bits. However, limitations relate to tractability restrictions with high-magnitude networks. In such cases, controlling the dimensionality of unobservable state spaces requires imposing strong serial dependencies within simulated latent network paths across steps in the sampler. This however may restrict the produced chain from exploring the posterior range of network paths efficiently. Approximate inferential frameworks relying on reduced product-form simplifications of state beliefs may improve the scalability of the method. Moreover, it is possible to explore the use of particle filtering approaches along with auxiliary variables for this purpose, since clamping explorable spaces within filtering procedures would likely ease the usual challenges regarding particle degeneracy; that is, ending with a very few particles having non-zero weights.
Also, the use of the uniformization technique will limit applications of the present framework to the study of purely Markovian processes. While it is possible to employ Markov-modulated regimes that adapt service and arrival rates to network states, this will greatly expand state spaces under consideration. Also, uniformization may deem the sampler computationally inefficient should service rates vary greatly across queues or job classes, as certain transition types will greatly dominate the underlying discrete time Markov chain.
Finally, the paper assumes that the volume of job service observations retrieved across the network is given by $\%(100\cdot q)$ of the total processing during a fixed time interval. For simplicity, $q$ is assumed fixed and known to the user. Many network structures (such as bottleneck networks) will allow for uncertainty regarding this parameter to be quantified by means of the presented sampler, as each iteration will provide a total number of network transitions complementing the observation number as a sufficient statistic. However, it is necessary to impose the knowledge of $q$ in order to ensure model identifiability whenever networks contain either global or self-loops.
\section*{Supplementary material}
Synthetic data along with a Java implementation of the algorithm can be found in \url{https://bitbucket.org/ikertxo1986/auxvarsamplerjava} or \url{https://github.com/IkerPerez/auxVarSampler}. This allows to reproduce the results within the examples above.
\section*{Acknowledgements}
We would like to thank the anonymous reviewers for their valuable remarks and suggestions that have improved the quality of this paper.
\bibliographystyle{apa}
|
\subsection{Domination and collective domination}
All graphs in this paper are assumed to be simple, namely not containing parallel edges or loops. The (open) neighborhood of a vertex $v$ in a graph $G$, denoted by $\tilde{N}(v)=\tilde{N}_G(v)$, is the set of all vertices connected to $v$. Given a set $D$ of vertices we write $\tilde{N}(D)$ for $\bigcup_{v \in D}\tilde{N}(v)$. Let $N(D)=N_G(D) = \tilde{N}(D) \cup D$.
A set $D$ is said to be {\em dominating} if $N(D)=V$ and {\em totally dominating} if
$\tilde{N}(D)=V$.
The minimal size of a dominating set is denoted by $\gamma(G)$, and the minimal size of a totally dominating set by
$\tilde{\gamma}(G)$.
There is a {\em collective} version of domination.
Given a system of graphs
$\zg=(G_1,\dots,G_k)$ on the same vertex set $V$, a system $\zd=(D_1,\dots,D_k)$ of subsets of $V$ is said to be {\em collectively dominating} if $\bigcup_{i \le k}N_{G_i}(D_i)=V$. Let $\gamma_\cup(\zg)$ be the minimum of $\sum_{i \le k}|D_i|$ over all collectively dominating systems.
\subsection{Independence and joint independence}
A set of vertices is said to be {\em independent} in $G$ if its elements are pairwise non-adjacent. The complex (closed down hypergraph) of independent sets in $G$ is denoted by $\zi(G)$. The {\em independence polytope} of $G$, denoted by
$IP(G)$, is the convex hull of the characteristic vectors of the
sets in $\zi(G)$. For a system of graphs
$\zg=(G_1,\dots,G_k)$ on $V$ the
{\em joint independence number}, $\alpha_\cap(\zg)$, is
$\max\{|I| : I \in \cap_{i \le k}\zi(G_i)\}$. The
{\em fractional joint independence number}, $\alpha_\cap^*(\zg)$, is
$\max\{\vec{x}\cdot\vec{1}:~~\vec{x}\in\bigcap_{i\le k} IP(G_i)\}$.\\
We shall mainly deal with the case $k=2$. Let us first observe that it is possible to have $\alpha_\cap^*(G_1,G_2)<\min(\alpha(G_1),\alpha(G_2))$.
\begin{example}
Let $G_1$ be obtained from the complete bipartite graph with respective sides $\{v_1, \ldots,v_6\}$ and $\{u_1,u_2\}$, by the addition of the edges $v_1v_2$, $v_3v_4$ and $u_1u_2$, and let $G_2=\bar{G_1}$. Then $\alpha(G_1)=\alpha(G_2)=4$, while $\alpha_\cap^*(G_1,G_2)=2$, the optimal vector in $IP(G_1) \cap IP(G_2)$ being the constant $\frac{1}{4}$ vector.
\end{example}
A graph $H$ is called a {\em partition graph} if it is the disjoint union of cliques. In a partition graph $\alpha=\gamma$.
The union of two systems of disjoint cliques is the line graph of a bipartite graph, having the set of cliques in one system as one side of the graph, and the set of cliques in the other system as the other side, an edge connecting two vertices (namely, cliques in different systems) if they intersect.
Thus, by K\"{o}nig's famous duality theorem \cite{konig}, we have:
\begin{theorem}\label{konig2}
If $G$ and $H$ are partition graphs on the same vertex set, then $$\alpha_\cap(G,H)=\gamma_\cup(G,H)$$.
\end{theorem}
There are graphs in which $\alpha >\gamma$, and thus
equality does not necessarily hold for general pairs $(G,H)$ of graphs, even when $G=H$. On the other hand, since a maximal independent set is dominating, we have $\gamma(G) \le \alpha(G)$ in every graph $G$. But the corresponding inequality for pairs of graphs is not necessarily true, as the following example shows.
\begin{example}\label{noteq}
Let $G=P_4$, namely the path with $3$ edges on $4$ vertices, and let $H$ be its complement. Then $\alpha_\cap(G,H)=1$
and $\gamma_\cup(G,H)=2$, so $\alpha_\cap(G,H)<\gamma_\cup(G,H)$.
\end{example}
However, as was shown in \cite{abhk}, if $\alpha_\cap$ is replaced by its
fractional version, then the non-trivial inequality in Theorem
\ref{konig2} does hold.
\begin{theorem}\label{inddom}
For any two graphs $G$ and $H$ on the same set of vertices we have $$\alpha_\cap^*(G,H)\geq\gamma_\cup(G,H)$$.
\end{theorem}
In Example \ref{noteq} $\vec{\frac{1}{2}}\in IP(G)\cap IP(H)$, and $\alpha_\cap^*(G,H)=2$, so $\alpha_\cap^*(G,H)=\gamma_\cup(G,H)$.
\begin{lemma}\label{fracit}
Let $\zv=(V_1, \ldots ,V_m)$ be a system of disjoint sets, let $\zi$ be the set of ranges of partial choice functions from $\zv$, and let $V=\bigcup_{i \le m} V_i$.
Then
$$\{f: V \to \R^+ \mid \sum_{v \in V_j}f(v)\le 1 \text{~for~ every~} j \le m \}=conv(\{\chi_I \mid I \in \zi\})$$.
\end{lemma}
\begin{proof}
Obviously, the right hand side is contained in the left hand side. For the reverse containment, let
$f: V \to \R^+$ be such that $\sum_{v \in V_j}f(v)\le 1$ for every $j \le m$, and assume for negation that it can be separated from all functions $\chi_I$, $ I \in \zi$, namely there exists a vector $\vec{u}$ such that $\sum_{v\in V} u(v)f(v)\ge 1$, and $\sum_{v\in I}u(v)=\sum_{v\in V} u(v)\chi_I(v)<1$ for all
$I \in \zi$. Since $conv(\{\chi_I \mid I \in \zi\})$ is closed down, we may assume that $\vec{u}$ is non-negative. For each $j \le m$ let $v_j$ be such that $u(v_j)$ is maximal over all $v\in V_j$, and let $I=\{v_j \mid j \le m\}$. The fact that $\sum_{v\in I}u(v)< 1$ implies then that
$$\sum_{j \le m}\sum_{v\in V_j}u(v)f(v)\le \sum_{j \le m}u(v_j)\sum_{v \in V_j} f(v)\le \sum_{j \in V_j}u(v_j)< 1, $$
a contradiction.
\end{proof}
\subsection{Independent transversals}
When one graph in the pair $(G,H)$, say $H$, is a partition graph, the parameters $\alpha_\cap(G,H)$ and $\alpha^*_\cap(G,H)$ can be described using the terminology of so-called {\em independent transversals}.
Given a graph $G$ and a partition $\zv=(V_1, \ldots ,V_m)$ of $V(G)$, an independent transversal (IT) is an independent set in $G$ consisting of the choice of one vertex from each set $V_i$. A {\em partial IT} is an independent set representing some $V_i$'s (so, it is the independent range of a partial choice function from $\zv$). A function $f : V \to \R^+$ is called a {\em partial fractional IT} if, when viewed as a vector, it belongs to $IP(G)$, and $\sum_{v\in V_j}f(v)\le 1$ for all $j \le m$. If $\sum_{v\in V_j}f(v)=1$ for all $j \le m$ then $f$ is called a {\em fractional IT}. By Lemma \ref{fracit} this means that $f \in IP(H) \cap IP(G)$, namely it is a jointly fractional independent set, where $\zv$ is the set of cliques in $H$.
For $I \subseteq [m]$ let $V_I=\bigcup_{i \in I}V_i$.
The following was proved in \cite{penny}:
\begin{theorem}\label{thm:penny}
If $\tg(G[V_I]) \ge 2|I|-1$ for every $I \subseteq [m]$ then there exists an IT.
\end{theorem}
Theorem \ref{inddom}, applied to the case in which $H$ is a partition graph, yields:
\begin{theorem}\label{fractr}
If $\gamma(G[V_I]) \ge |I|$ for every $I \subseteq [m]$ then there exists a fractional IT.
\end{theorem}
\section{Putting weights on the vertices}
In \cite{abz} a weighted version of Theorem \ref{thm:penny} was proved.
As is often the case with weighted versions, the motivation came from decompositions: weighted results give, by duality, fractional decompositions results. It is conjectured that if $|V_i| \ge 2\Delta(G)$ then there exists a partition of $V(G)$ into $\max_{i \le m}|V_i|$ IT's. The weighted version of Theorem \ref{thm:penny} yielded the existence of a fractional such decomposition.
\begin{notation}
Given a real valued function $f$ on a set $S$, and a set $A\subseteq
S$, define $f[A]=\sum_{a\in A}f(a)$. We also write $|f|=f[S]$ and we
call $|f|$ the {\em size} of $f$.
\end{notation}
\begin{definition}
Let $G=(V,E)$ be a graph, and let $w:V\to\mathbb{N}$ be a weight
function on $V$. We say that a function $f:V\to\mathbb{N} $ {\em
$w$-dominates} a set $U$ of vertices, if $f[N(u)]\geq w(u)$ for every $u\in U$. We say that $f$ is {\em $w$-dominating} if it $w$-dominates $V$.
The {\em weighted domination number}
$\gamma^w(G)$ is $\min\{|f| \mid f ~\text{is $w$-dominating}\} $
\end{definition}
This definition extends to systems of graphs:
\begin{definition}
Let $\mathcal{G}=(G_1,\dots,G_k)$ be a system of graphs on the same
vertex set $V$. Let $w:V\to\mathbb{N}$ be a non-negative weight function on $V$, and let $\zf=(f_i:~V\to\mathbb{N}, ~~i \le k)$ be a system of functions. We say that
$\zf$
$w$-dominates $\zg$ if $\sum_{i=1}^k f_i[N_{G_i}(v)]\geq w(v)$ for every $v\in V$. The {\em
weighted collective domination number } is
$$\gamma_\cup^w(\mathcal{G})=\min\{\sum_{i=1}^k
|f_i|:(f_1,\dots,f_k)\; is\; w-dominating\}.$$
The extension of the independence parameter to the weighted case is also quite natural:
$$(\alpha_\cap^w)^*(\zg)=\max\{\sum_{v \in V}x(v)w(v) \mid \vec{x}\in\bigcap_{i=1}^k
IP(G_i)\}.$$
\end{definition}
The aim of this paper is to study the following possible extension of Theorem \ref{inddom} to the weighted case.
\begin{Conjecture}\label{conj:main}
If $G$ and $H$ are graphs on the same vertex set $V$
then for any weight function $w:V\to\mathbb{N}$ we have
$$(\alpha_\cap^w)^*(G,H)\geq\gamma_\cup^w(G,H).$$
\end{Conjecture}
If $H=G$ then
the stronger $\alpha_\cap^w(G,G) \geq\gamma_\cup^w(G,G)$ is true, namely:
\begin{lemma}
$\alpha^w(G) \ge \gamma^w(G)$.
\end{lemma}
\begin{proof}
We have to exhibit a $w$-dominating function $f$ and an independent set $I$ with $|f|\leq w[I]$.
Let $V(G)=\{v_1,\dots,v_n\}$. We define a $w$-dominating function $f:V\to \mathbb{N}$ inductively.
Let $f(v_1)=w(v_1)$. Having defined $f(v_1),\dots,f(v_{i-1})$ let $$f(v_i)= [w(v_i)-\sum_{v_j\in N(v_i),\; j<i}f(v_j)]^+$$
Clearly, $f$ is $w$-dominating.
We next find an independent set $I$ such that $w[I]\geq|f|$.
Let $v_{i_1}$ be the vertex that has the maximal index over all the vertices in $V_1=V\cap supp(f)$. Since $f(v_{i_1})>0$, we have $f[N(v_{i_1})]=w(v_{i_1})$.
Suppose that we have defined the sets of vertices $V_1,V_2,\dots, V_{k-1}$ and vertices $v_{i_1},\dots,v_{i_k}$ such that $v_{i_j}$ is the vertex whose index is maximal over all the vertices in $V_j$ where $V_j=V_{j-1}\setminus N(v_{i_{j-1}})$ for every $j=1,\dots,k-1$.
Let $V_k=V_{k-1}\setminus N(v_{i_{k-1}})$ and let $v_{i_k}$ be the vertex whose index is maximal over all the vertices in $V_k$. By the definition of $f$ we have $\sum_{v_j\in N(v_{i_k}),\; j<i_k}f(v_j)=w(v_{i_k})$, so $\sum_{v_j\in V_k\cap N(v_{i_k})}f(v_j)\leq w(v_{i_k})$. We stop the process when $V_t=\emptyset$ for some $t$. In this case $I=\{v_{i_1},\dots,v_{i_{t-1}}\}$ is an independent set that satisfies $w[I]\geq |f|$ as desired.
\end{proof}
\section{The case of partition graphs}
The main result of this paper is:
\begin{theorem}\label{fractrconj}
Conjecture \ref{conj:main} is true if $H$ is a partition graph. Namely, if $H$ is a partition graph and $G$ is any graph, then
$$(\alpha_\cap^w)^*(G,H)\geq\gamma_\cup^w(G,H).$$
\end{theorem}
Let us first re-formulate the left hand side of the inequality in terms of partitions.
For a partition $\zv=(V_1, \ldots, V_m)$ of the vertex set $V$ of a graph $G$, let \begin{equation} \label{ns} \nsw(G, \zv) =\max\{ \sum_{v \in V}w(v)f(v)\; \mid \; f ~\text{is~a~fractional~ partial ~IT}\}.\end{equation}
By Lemma \ref{fracit} we have:
\begin{lemma}\label{lem:param}
$(\alpha_\cap^w)^*(G,H)=\nsw(G, \zv)$.
\end{lemma}
Let us also re-formulate the right hand side using the terminology of partitions.
Given partition $\zv=(V_1, \ldots, V_m)$ of $V(G)$, a pair of non-negative real valued functions $f$ on $V$ and $g$ on $[m]$ is said to be {\em collectively $w$-dominating} if for every vertex $v \in V_i$ we have $g(i)+f[N(v)] \ge w(v)$. Let $\gamma^w(G, \zv)$ be the minimum of $|g|+|f|$ over all collectively $w$-dominating pairs of functions. In this terminology, $\gamma_\cup^w(G,H)=\gamma^w(G,\zv)$. In addition, let $\tau^w(G, \zv)$ be the minimum of $|g|+\frac{|f|}{2}$ over all collectively $w$-dominating pairs of functions.
In \cite{abz} the following weighted version of Theorem \ref{fractr} was proved.
\begin{theorem}\label{thm:abz}
$\nu^w(G,\zv) \ge \tau^w(G, \zv)$.
\end{theorem}
\begin{remark} Note the factor $\frac{1}{2}$ difference between the definitions of $\tau^w(G, \zv)$ and $\gamma^w(G,\zv)$. It mirrors the factor $\frac{1}{2}$ difference (manifest in the factor $2$ in ``$2|I|-1$'') between the statements of
Theorems \ref{thm:penny} and \ref{fractr}. The same factor appears in the weighted case: the difference between the integral and fractional versions is the $\frac{1}{2}$ factor hidden in the right hand sides
of Theorems \ref{thm:abz} and of \ref{thm:partition} below. \end{remark}
By Lemma \ref{lem:param} the case of Conjecture \ref{conj:main} in which $H$ is a partition graph is:
\begin{theorem}\label{thm:partition}
${(\nu^w)}^*(G,\zv) \ge \gamma^w(G, \zv)$, where $\zv$ is the partition of $V$ into cliques of $H$.
\end{theorem}
\begin{proof}
Note that if $f = \sum_{I \in \zi(G)}x_I\chi_I$ then $f[V_j]=\sum_{I\in\zi(G)}x_I|I\cap
V_j|$, and thus the constraints defining the linear program for $\nsw(G, \zv)$ are $\sum_{I\in\zi(G)}x_I|I\cap
V_j|\leq 1$ and $\sum_{I\in\zi(G)}x_I=1$.
\begin{assertion}
Let
\begin{equation}\label{t}\nsw(G,\zv)=\max\{\sum_{I\in \zi(G)}x_I w[I]|\sum_{I\in\zi(G)}x_I|I\cap
V_j|\leq 1,\; \sum_{I\in\zi(G)}x_I\leq 1\}\end{equation}
\end{assertion}
\begin{proof}
Denote the right hand side by $t$.
If $f=\sum_{I\in\zi(G)}x_I\chi_I$ is an optimal solution of the linear program \eqref{ns} then $\nsw(G,\zv)=\sum_{v\in V}w(v)f(v)=\sum_{I\in \zi(G)}x_I w[I]$. Hence $\nsw(G,\zv)\leq t$.
On the other hand, suppose by negation that there exists an optimal solution of the linear program \eqref{t} that satisfies $\sum_{I\in\zi(G)}x_I=1-\epsilon$ for some $\epsilon>0$. Clearly $t>0$, and hence there exists an independent set $I_0$ such that $x_{I_0}>0$. Choose a vertex $v\in I_0$, and define a vector $\vec{x'}$ as follows. Let $x'_{I_0}=x_{I_0}-\epsilon$, $x'_{I_0\setminus\{v\}}=x'_{\{v\}}=\epsilon$ and $x'_I=x_I$ otherwise. Note that the vector $x'$ satisfies constrains of the linear program, but the weight of $\sum_{I\in\zi(G)}x'_I\chi_I$ is $\sum_{I\in\zi(G)}x_Iw[I]+\epsilon$, contradicting the maximality of the optimal solution. Hence this optimal solution is also a solution for the linear program \eqref{ns}, so, $t\leq \nsw(G,\zv)$ proving the desired equality.
\end{proof}
By LP duality
$\nsw(G, \zv)$ is the minimum of $\sum_{j=0}^m y_j$ over all vectors $\vec{y}=(y_0,y_1,\dots,y_m)$ satisfying $y_0+\sum_{j=1}^my_j |I\cap V_j| \geq w[I]$ for all $I\in\zi(G)$.
Let $\vec{y}=(y_0,y_1,\dots,y_m)$ be a vector in which the minimum is attained, meaning that $\sum_{j=0}^m y_j=\nsw(G,\zv)$, and let $g(j)=\lfloor y_j\rfloor$ for all $j \le m$. We define a new weight function $w_g$ by $w_g(v)=[w(v)-\lfloor y_{j(v)}\rfloor]^+$, where $j(v)$ is that $j$ for which $v \in V_j$. Let $V'=\{v \mid w_g(v)>0\}$ be the support of $w_g$, and let $G'=G[V']$.
For a number $s$ let $\{s\}$ be the fractional part of $s$, namely $\{s\}=s- \lfloor s \rfloor$.
\begin{assertion}
The vector $(y_0,\{y_1\},\dots,\{y_m\})$ is an optimal solution for the program dual to: $(\nu^{w_g})^*(G', \zv)$, namely
$$y_0+\sum_{j=1}^m \{y_j\}=(\nu^{w_g})^*(G',\zv):=$$ $$\max\{\sum_{I\in\zi(G')}x_I w_g[I]\mid \;\sum_{I\in\zi(G')} x_I\leq1\; \text{and} \; \forall j\;\sum_{I\in\zi(G')} x_I |I\cap V_j|\leq 1\}$$
\end{assertion}
\begin{proof}
Denote by $y$ the sum $y_0+\sum_{j=1}^m \{y_j\}$.
For every $v\in V'$, we have $w_g(v)=w(v)-\lfloor y_{j(v)}\rfloor$, hence $y_0+\sum_{j=1}^m\{y_j\} |I\cap V_j|\geq w_g[I]$ for every $I\in\zi(G')\subseteq\zi(G)$, proving that $y\geq (\nu^{w_g})^*(G',\zv)$.
For the reverse inequality, assume for negation that there exists a solution $\vec{x}=(x_0,\dots,x_m)$ such that $\sum_{j=0}^m x_j<y$.
Then the vector $\vec{x}=(x_0,x_1+\lfloor y_1\rfloor,\dots,x_m+\lfloor y_m\rfloor)$ is a solution to the original problem that satisfies $x_0+\sum_{j=1}^m x_j+\lfloor y_j\rfloor<\sum_{j=0}^m y_j=\nsw(G,\zv)$, contradicting the optimality of $\vec{x}$.
\end{proof}
Since an optimal solution of the primary problem corresponding to the weight function $w_g$ satisfies $\sum_{I\in\mathcal{I}(G')}x_I=1$, there exists a set $I$ such that $x_I>0$.
Let $I_0$ be a set of minimal weight in $supp(x)$. Then \begin{equation}\label{I_0} w_g[I_0]=w_g[I_0]\sum_{I\in\zi(G')}x_I\leq \sum_{I\in\zi(G')}x_I w_g[I]=(\nu^{w_g})^*(G',\zv)\end{equation}
Let $h:V'\to\mathbb{N}$ defined by $h(v)=w_g(v)$ if $v\in I_0$ and $h(v)=0$ otherwise.
\begin{assertion}
The function $h$ is $w_g$-dominating in $G'$.
\end{assertion}
\begin{proof}
Suppose not. Then there exists a vertex $v\in V'$ such that $w_g(v)>h(v)+h[\tilde{N}(v)]=h(v)+w_g[\tilde{N}(v)\cap I_0]$. Clearly, $v\notin I_0$, hence $h(v)=0$. The set $I'=(I_0\setminus \tilde{N}(v))\cup \{v\}$ satisfies
$$w_g[I']=w_g[I_0\setminus \tilde{N}(v)]+w_g(v)>w_g[I_0\setminus \tilde{N}(v)]+w_g[I_0\cap \tilde{N}(v)]=w_g[I_0]$$ Since $w$ is an integral function, $w_g$ is also integral, hence $w_g[I']=w_g[I_0]+k_v$ for some $k_v\geq 1$.
On the other hand, by the definition of the dual program $w_g[I']\leq y_0+\sum_{j=1}^m \{y_j\}|I'\cap V_j|$. In addition, since $x_{I_0}>0$, the complementary slackness conditions state that equality holds in the corresponding constraint in the dual problem, i.e. $w_g[I_0]= y_0+\sum_{j=1}^m \{y_j\}|I_0\cap V_j|$. Hence,
$$k_v=w_g[I']-w_g[I_0]\leq \sum_{j=1}^m \{y_j\}(|I'\cap V_j|-|I_0\cap V_j|)=\{y_{j(v)}\}-\sum_{u\in N(v)\cap I_0}\{y_{j(u)}\}<1 $$
is a contradiction.
\end{proof}
Since $g$ dominates all vertices $v\in V\setminus V'$, the pair $(g,h)$ is $w$-dominating, hence using \eqref{I_0} we have $$\gamma^w(G,\zv)=\leq |g|+|h|=|g|+w_g[I_0]\leq |g|+(\nu^{w_g})^*(G',\zv)$$ $$=\sum_{j=1}^m \lfloor y_j\rfloor+y_0+\sum_{j=1}^m \{y_j\}=\sum_{j=0}^m y_j=\nsw(G, \zv) $$
as desired.
\end{proof}
\end{section}
|
\section{Introduction}
For the numerical solution of partial differential equations, Galerkin finite
element methods are among the most popular discretization methods. In the last
decades, \textit{non-conforming} Galerkin discretizations have become very
attractive where the test and trial spaces are not subspaces of the natural
energy spaces and/or the variational formulation is modified on the discrete
level. These methods have nice properties, e.g. in different parts of the
domain different discretizations can be easily used and glued together or, for
certain classes of problems (Stokes problems, highly indefinite Helmholtz and
Maxwell problems, problems with \textquotedblleft locking\textquotedblright,
etc.), the non-conforming discretization enjoys a better stability behavior
compared to the conforming one. One of the first non-conforming finite element
space was the Crouzeix-Raviart element (\cite{Crouzeix_Raviart}, see
\cite{Brenner_Crouzeix} for a survey). It is piecewise affine with respect to
a triangulation of the domain while interelement continuity is required only
at the barycenters of the edges/facets (2D/3D).
In \cite{ccss_2012}, a family of high order non-conforming (intrinsic) finite
elements have been introduced which corresponds to a family of high-order
Crouzeix-Raviart elements in two dimensions. For Poisson's equation, this
family includes the non-conforming Crouzeix-Raviart element
\cite{Crouzeix_Raviart}, the Fortin-Soulie element \cite{Fortin_Soulie}, the
Crouzeix-Falk element \cite{Crouzeix_Falk}, and the Gauss-Legendre elements
\cite{BaranGL}, \cite{BaranCVD} as well as the standard conforming $hp$-finite elements.
In our paper we will characterize a family of high-order Crouzeix-Raviart type
finite elements in \textit{three} dimensions, first implicitly by imposing
certain jump conditions at the interelement facets. Then we derive a local
basis for these finite elements. These new finite element spaces are
non-conforming but the (broken version of the) continuous bilinear form can
still be used. Thus, our results also give insights on how far one can go in
the non-conforming direction while keeping the original forms.
The explicit construction of a basis for these new finite element spaces
require some deeper theoretical tools in the field of orthogonal polynomials
on triangles and their representations which we develop here for this purpose.
As a simple model problem for the introduction of our method, we consider
Poisson's equation but emphasize that this method is applicable also for much
more general (systems of) elliptic equations.
There is a vast literature on various conforming and non-conforming, primal,
dual, mixed formulations of elliptic differential equations and conforming as
well as non-conforming discretization. Our main focus is the\ characterization
and construction of non-conforming Crouzeix-Raviart type finite elements from
theoretical principles. For this reason, we do not provide an extensive list
of references on the analysis of specific families of finite elements spaces
but refer to the classical monographs \cite{Ciarlet}, \cite{Schwab}, and
\cite{BoffiBrezziFortin} and the references therein.
The paper is organized as follows.
In Section \ref{SecModelProblem} we introduce our model problem, Poisson's
equation, the relevant function spaces and standard conditions on its well-posedness.
In Section \ref{sectconfbasis} we briefly recall classical, conforming
$hp$-finite element spaces and their Lagrange basis.
The new non-conforming finite element spaces are introduced in Section
\ref{section_nonconforme}. We introduce an appropriate compatibility condition
at the interfaces between elements of the mesh so that the non-conforming
perturbation of the original bilinear form is consistent with the local error
estimates. We will see that this compatibility condition can be inferred from
the proof of the second Strang lemma applied to our setting. The weak
compatibility condition allows to characterize the non-conforming family of
high-order Crouzeix-Raviart type elements in an \textit{implicit} way. In this
section, we will also present explicit representations of non-conforming basis
functions of general degree $p$ while their derivation and analysis is the
topic of the following sections.
Section \ref{SecExplConstr} is devoted to the explicit construction of a basis
for these new non-conforming finite elements. It requires deeper theoretical
tools from orthogonal polynomials on triangles and their representation which
we will derive for this purpose in this section.
It is by no means obvious whether the constructed set of functions is linearly
independent and span the non-conforming space which was defined implicitly in
Section \ref{section_nonconforme}. These questions will be treated in Section
\ref{SecPropNC}.
Finally, in Section \ref{SecConclusion} we summarize the main results and give
some comparison with the two-dimensional case which was developed in
\cite{ccss_2012}.
\section{Model Problem\label{SecModelProblem}}
As a model problem we consider the Poisson equation in a bounded Lipschitz
domain $\Omega\subset\mathbb{R}^{d}$ with boundary $\Gamma:=\partial\Omega$.
First, we introduce some spaces and sets of functions for the coefficient
functions and solution spaces.
The Euclidean scalar product in $\mathbb{R}^{d}$ is denoted for $\mathbf{a
,\mathbf{b}\in\mathbb{R}^{d}$ by $\mathbf{a}\cdot\mathbf{b}$. For $s\geq0$,
$1\leq p\leq\infty$, let $W^{s,p}\left( \Omega\right) $ denote the classical
(real-valued) Sobolev spaces with norm $\left\Vert \cdot\right\Vert
_{W^{s,p}\left( \Omega\right) }$. The space $W_{0}^{s,p}\left(
\Omega\right) $ is the closure with respect to the $\left\Vert \cdot
\right\Vert _{W^{s,p}\left( \Omega\right) }$ of all $C^{\infty}\left(
\Omega\right) $ functions with compact support. As usual we write
$L^{p}\left( \Omega\right) $ short for $W^{0,p}\left( \Omega\right) $. The
scalar product and norm in $L^{2}\left( \Omega\right) $ are denoted by
$\left( u,v\right) :=\int_{\Omega}uv$ and $\left\Vert \cdot\right\Vert
:=\left( \cdot,\cdot\right) ^{1/2}$. For $p=2$, we use $H^{s}\left(
\Omega\right) $, $H_{0}^{s}\left( \Omega\right) $ as shorthands for
$W^{s,2}\left( \Omega\right) $, $W_{0}^{s,2}\left( \Omega\right) $. The
dual space of $H_{0}^{s}\left( \Omega\right) $ is denoted by $H^{-s}\left(
\Omega\right) $. We recall that, for positive integers $s$, the seminorm
$\left\vert \cdot\right\vert _{H^{s}\left( \Omega\right) }$ in $H^{s}\left(
\Omega\right) $ which contains only the derivatives of order $s$ is a norm in
$H_{0}^{s}\left( \Omega\right) $.
We consider the Poisson problem in weak form
\begin{equation}
\text{Given }f\in L^{2}\left( \Omega\right) \text{ find }u\in H_{0
^{1}\left( \Omega\right) \text{\quad}a\left( u,v\right) :=\left(
\mathbb{A}\nabla u,\nabla v\right) =\left( f,v\right) \text{\quad}\forall
v\in H_{0}^{1}\left( \Omega\right) . \label{varform
\end{equation}
Throughout the paper we assume that the diffusion matrix $\mathbb{A}\in
L^{\infty}\left( \Omega,\mathbb{R}_{\operatorname*{sym}}^{d\times d}\right)
$ is symmetric and satisfie
\begin{equation}
0<a_{\min}:=\underset{\mathbf{x}\in\Omega}{\operatorname*{ess}\inf
\inf_{\mathbf{v\in}\mathbb{R}^{d}\backslash\left\{ 0\right\} }\frac{\left(
\mathbb{A}\left( \mathbf{x}\right) \mathbf{v}\right) \cdot\mathbf{v
}{\mathbf{v}\cdot\mathbf{v}}\,\leq\underset{\mathbf{x}\in\Omega
{\operatorname*{ess}\sup}\sup_{\mathbf{v}\in\mathbb{R}^{d}\backslash\left\{
0\right\} }\frac{\left( \mathbb{A}\left( \mathbf{x}\right) \mathbf{v
\right) \cdot\mathbf{v}}{\mathbf{v}\cdot\mathbf{v}}=:a_{\max}<\infty
\label{aeps
\end{equation}
and{ that there exists a partition $\mathcal{P}:=(\Omega_{j})_{j=1}^{J}$ of
$\Omega$ into $J$ (possibly curved) polygons (polyhedra for }$d=3$) {such
that, for some appropriate }$r\in\mathbb{N}$, it hold
\begin{equation}
\left\Vert \mathbb{A}\right\Vert _{{PW^{{r},\infty}\left( \Omega\right)
}:=\max_{1\leq j\leq J}\left\Vert \left. \mathbb{A}\right\vert _{\Omega_{j
}\right\Vert _{W^{r,\infty}\left( \Omega_{j}\right) }<\infty. \label{aeps2
\end{equation}
Assumption (\ref{aeps}) implies the well-posedness of problem (\ref{varform})
via the Lax-Milgram lemma.
\section{Conforming hp-Finite Element Galerkin
Discretization\label{sectconfbasis}}
In this paper we restrict our studies to bounded, polygonal ($d=2$) or
polyhedral ($d=3$) Lipschitz domains $\Omega\subset\mathbb{R}^{d}$ and regular
finite element meshes $\mathcal{G}$ (in the sense of \cite{Ciarlet})
consisting of (closed) simplices $K$, where hanging nodes are not allowed. The
local and global mesh width is denoted by $h_{K}:=\operatorname*{diam}K$ and
$h:=\max_{K\in\mathcal{G}}h_{K}$. The boundary of a simplex $K$ can be split
into $\left( d-1\right) $-dimensional simplices (facets for $d=3$ and
triangle edges for $d=2$) which are denoted by $T$. The set of all facets in
$\mathcal{G}$ is called $\mathcal{F}$; the set of facets lying on
$\partial\Omega$ is denoted by $\mathcal{F}_{\partial\Omega}$ and defines a
triangulation of the surface $\partial\Omega$. The set of facets in $\Omega$
is denoted by $\mathcal{F}_{\Omega}$. As a convention we assume that simplices
and facets are closed sets. The interior of a simplex $K$ is denoted by
$\overset{\circ}{K}$ and we write $\overset{\circ}{T}$ to denote the
(relative) interior of a facet $T$. The set of all simplex vertices in the
mesh $\mathcal{G}$ is denoted by $\mathcal{V}$, those lying on $\partial
\Omega$ by $\mathcal{V}_{\partial\Omega}$, and those lying in $\Omega$ by
$\mathcal{V}_{\Omega}$. Similar the set of simplex edges in $\mathcal{G}$ is
denoted by $\mathcal{E}$, those lying on $\partial\Omega$ by $\mathcal{E
_{\partial\Omega}$, and those lying in $\Omega$ by $\mathcal{E}_{\Omega}$.
We recall the definition of conforming $hp$-finite element spaces (see, e.g.,
\cite{Schwab}). For $p\in\mathbb{N}_{0}:=\left\{ 0,1,\ldots\right\} $, let
$\mathbb{P}_{p}^{d}$ denote the space of $d$-variate polynomials of total
degree $\leq p$. For a connected subset $\omega\subset\Omega$, we write
$\mathbb{P}_{d}^{p}\left( \omega\right) $ for polynomials of degree $\leq p$
defined on $\omega$. For a connected $m$-dimensional manifold $\omega
\subset\mathbb{R}^{d}$, for which there exists a subset $\hat{\omega
\in\mathbb{R}^{m}$ along an affine bijection $\chi_{\omega}:\hat{\omega
}\rightarrow\omega$, we set $\mathbb{P}_{p}^{m}\left( \omega\right)
:=\left\{ v\circ\chi_{\omega}^{-1}:v\in\mathbb{P}_{p}^{m}\left( \hat{\omega
}\right) \right\} $. If the dimension $m$ is clear from the context, we
write $\mathbb{P}_{p}\left( \omega\right) $ short for $\mathbb{P}_{p
^{m}\left( \omega\right) $.
The conforming $hp$-finite element space is given b
\begin{equation}
S_{\mathcal{G},\operatorname*{c}}^{p}:=\left\{ u\in C^{0}\left(
\overline{\Omega}\right) \mid\forall K\in\mathcal{G}\quad\left. u\right\vert
_{K}\in\mathbb{P}_{p}\left( K\right) \right\} \cap H_{0}^{1}\left(
\Omega\right) . \label{hpfinele
\end{equation}
A Lagrange basis for $S_{\mathcal{G},\operatorname*{c}}^{p}$ can be defined as
follows. Let
\begin{equation}
\widehat{\mathcal{N}}^{p}:=\left\{ \frac{\mathbf{i}}{p}:\mathbf{i
\in\mathbb{N}_{0}^{d}\text{ with }i_{1}+\ldots+i_{d}\leq p\right\}
\label{defnodalpointsref
\end{equation}
denote the equispaced unisolvent set of nodal points on the $d$-dimensional
unit simplex
\begin{equation}
\widehat{K}:=\left\{ \mathbf{x}\in\mathbb{R}_{\geq0}^{d}\mid x_{1
+\ldots+x_{d}\leq1\right\} . \label{defrefelement
\end{equation}
For a simplex $K\in\mathcal{G}$, let $\chi_{K}:\widehat{K}\rightarrow K$
denote an affine mapping. The set of nodal points is given b
\begin{equation
\begin{array}
[c]{l
\mathcal{N}^{p}:=\left\{ \chi_{K}\left( \mathbf{\hat{N}}\right)
\mid\mathbf{\hat{N}}\in\widehat{\mathcal{N}}^{p},K\in\mathcal{G}\right\} ,\\
\mathcal{N}_{\Omega}^{p}:=\mathcal{N}^{p}\cap\Omega,\qquad\mathcal{N
_{\partial\Omega}^{p}:=\mathcal{N}^{p}\cap\partial\Omega.
\end{array}
\label{defNodalpoints
\end{equation}
The Lagrange basis for $S_{\mathcal{G},\operatorname*{c}}^{p}$ can be indexed
by the nodal points $\mathbf{N}\in\mathcal{N}_{\Omega}^{p}$ and is
characterized by
\begin{equation}
B_{p,\mathbf{N}}^{\mathcal{G}}\in S_{\mathcal{G},\operatorname*{c}}^{p
\quad\text{and\quad}\forall\mathbf{N}^{\prime}\in\mathcal{N}_{\Omega}^{p}\quad
B_{p,\mathbf{N}}^{\mathcal{G}}\left( \mathbf{N}^{\prime}\right)
=\delta_{\mathbf{N},\mathbf{N}^{\prime}}, \label{basisfunctions
\end{equation}
where $\delta_{\mathbf{N},\mathbf{N}^{\prime}}$ is the Kronecker delta.
\begin{definition}
\label{Deflocbasis}For all $K\in\mathcal{G}$, $T\in\mathcal{F}_{\Omega}$,
$E\in\mathcal{E}_{\Omega}$, $\mathbf{V}\in\mathcal{V}_{\Omega}$, the
conforming spaces $S_{K,\operatorname*{c}}^{p}$, $S_{T,\operatorname*{c}}^{p
$, $S_{E,\operatorname*{c}}^{p}$, $S_{\mathbf{V},\operatorname*{c}}^{p}$ are
given as the spans of the following basis function
\begin{align*}
S_{K,\operatorname*{c}}^{p} & :=\operatorname*{span}\left\{ B_{p,\mathbf{N
}^{\mathcal{G}}\mid\mathbf{N}\in\overset{\circ}{K}\cap\mathcal{N}_{\Omega
^{p}\right\} ,\\
S_{T,\operatorname*{c}}^{p} & :=\operatorname*{span}\left\{ B_{p,\mathbf{N
}^{\mathcal{G}}\mid\mathbf{N}\in\overset{\circ}{T}\cap\mathcal{N}_{\Omega
^{p}\right\} ,\\
S_{E,\operatorname*{c}}^{p} & :=\operatorname*{span}\left\{ B_{p,\mathbf{N
}^{\mathcal{G}}\mid\mathbf{N}\in\overset{\circ}{E}\cap\mathcal{N}_{\Omega
^{p}\right\} ,\\
S_{\mathbf{V},\operatorname*{c}}^{p} & :=\operatorname*{span}\left\{
B_{p,\mathbf{V}}^{\mathcal{G}}\right\} \text{.
\end{align*}
\end{definition}
The following proposition shows that these spaces give rise to a direct sum
decomposition and that these spaces are locally defined. To be more specific
we first have to introduce some notation.
For any facet $T\in\mathcal{F}_{\Omega}$, vertex $\mathbf{V}\in\mathcal{V
_{\Omega}$, and $E\in\mathcal{E}_{\Omega}$ we define the set
\begin{equation
\begin{array}
[c]{ll
\mathcal{G}_{T}:=\left\{ K\in\mathcal{G}:T\subset\partial K\right\} , &
\omega_{T}:
{\displaystyle\bigcup\limits_{K\in\mathcal{G}_{T}}}
K,\\
\mathcal{G}_{\mathbf{V}}:=\left\{ K\in\mathcal{G}:\mathbf{V}\in\partial
K\right\} , & \omega_{\mathbf{V}}:
{\displaystyle\bigcup\limits_{K\in\mathcal{G}_{\mathbf{V}}}}
K,\\
\mathcal{G}_{E}:=\left\{ K\in\mathcal{G}:E\subset\partial K\right\} , &
\omega_{E}:
{\displaystyle\bigcup\limits_{K\in\mathcal{G}_{E}}}
K.
\end{array}
\label{defoftrianglesubsets
\end{equation}
\begin{proposition}
\label{PropSpaceDecomp}Let $S_{K,\operatorname*{c}}^{p}$,
$S_{T,\operatorname*{c}}^{p}$, $S_{E,\operatorname*{c}}^{p}$, $S_{\mathbf{V
,\operatorname*{c}}^{p}$ be as in Definition \ref{Deflocbasis}. Then the
direct sum decomposition hold
\begin{equation}
S_{\mathcal{G},\operatorname*{c}}^{p}=\left(
{\displaystyle\bigoplus\limits_{\mathbf{V}\in\mathcal{V}_{\Omega}}}
S_{\mathbf{V},\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{E\in\mathcal{E}_{\Omega}}}
S_{E,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
S_{T,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{K\in\mathcal{G}}}
S_{K,\operatorname*{c}}^{p}\right) . \label{spacedecomposition
\end{equation}
\end{proposition}
\section{Galerkin Discretization with Non-Conforming Crouzeix-Raviart Finite
Elements\label{section_nonconforme}}
\subsection{Non-Conforming Finite Elements with Weak Compatibility Conditions}
In this section, we will characterize a class of non-conforming finite element
spaces implicitly by a weak compatibility condition across the facets. For
each facet $T\in\mathcal{F}$, we fix a unit vector $\mathbf{n}_{T}$ which is
orthogonal to $T$. The orientation for the inner facets is arbitrary but fixed
while the orientation for the boundary facets is such that $\mathbf{n}_{T}$
points toward the exterior of $\Omega$. Our non-conforming finite element
spaces will be a subspace o
\[
C_{\mathcal{G}}^{0}\left( \Omega\right) :=\left\{ u\in L^{\infty}\left(
\Omega\right) \mid\forall K\in\mathcal{G\quad}\left. u\right\vert
_{\overset{\circ}{K}}\in C^{0}\left( \overset{\circ}{K}\right) \right\}
\]
and we consider the skeleton
{\displaystyle\bigcup\limits_{T\in\mathcal{F}}}
T$ as a set of measure zero.
For $K\in\mathcal{G}$, we define the restriction operator $\gamma
_{K}:C_{\mathcal{G}}^{0}\left( \Omega\right) \rightarrow C^{0}\left(
K\right) $ b
\[
\left( \gamma_{K}w\right) \left( \mathbf{x}\right) =w\left(
\mathbf{x}\right) \quad\forall\mathbf{x}\in\overset{\circ}{K
\]
and on the boundary $\partial K$ by continuous extension. For the inner facets
$T\in\mathcal{F}$, let $K_{T}^{1},K_{T}^{2}$ be the two simplices which share
$T$ as a common facet with the convention that $\mathbf{n}_{T}$ points into
$K_{2}$. We set $\omega_{T}:=K_{T}^{1}\cup K_{T}^{2}$. The jump $\left[
\cdot\right] _{T}:C_{\mathcal{G}}^{0}\left( \Omega\right) \rightarrow
C^{0}\left( T\right) $ across $T$ is defined b
\begin{equation}
\left[ w\right] _{T}=\left. \left( \gamma_{K_{2}}w\right) \right\vert
_{T}-\left. \left( \gamma_{K_{1}}w\right) \right\vert _{T}.
\label{defjumps
\end{equation}
For vector-valued functions, the jump is defined component-wise. The
definition of the non-conforming finite elements involves orthogonal
polynomials on triangles which we introduce first.
Let $\widehat{T}$ denote the (closed) unit simplex in $\mathbb{R}^{d-1}$, with
vertices $\mathbf{0}$, $\left( 1,0,\ldots,0\right) ^{\intercal}$, $\left(
0,1,0,\ldots,0\right) ^{\intercal}$, $\left( 0,\ldots,0,1\right)
^{\intercal}$. For $n\in\mathbb{N}_{0}$, the set of orthogonal polynomials on
$\widehat{T}$ is given b
\begin{equation}
\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) :=\left\{
\begin{array}
[c]{ll
\mathbb{P}_{0}\left( \widehat{T}\right) & n=0,\\
\left\{ u\in\mathbb{P}_{n}\left( \widehat{T}\right) \mid\int_{\widehat{T
}uv=0\quad\forall v\in\mathbb{P}_{n-1}\left( \widehat{T}\right) \right\} &
n\geq1.
\end{array}
\right. \label{defPnn-1senk
\end{equation}
We lift this space to a facet $T\in\mathcal{F}$ by employing an affine
transform $\chi_{T}:\widehat{T}\rightarrow T$
\[
\mathbb{P}_{n,n-1}^{\perp}\left( T\right) :=\left\{ v\circ\chi_{T
^{-1}:v\in\mathbb{P}_{n,n-1}^{\perp}\left( T\right) \right\} .
\]
The orthogonal polynomials on triangles allows us to formulate the
\textit{weak compatibility condition} which is employed for the definition of
non-conforming finite element spaces
\begin{equation}
\left[ u\right] _{T}\in\mathbb{P}_{p,p-1}^{\perp}\left( T\right)
,\ \forall T\in\mathcal{F}_{\Omega}\quad\text{and\quad}\left. u\right\vert
_{T}\in\mathbb{P}_{p,p-1}^{\perp}\left( T\right) ,\ \forall T\in
\mathcal{F}_{{\partial\Omega}}. \label{poinprep
\end{equation}
We have collected all ingredients for the (implicit) characterization of the
non-conforming Crouzeix-Raviart finite element space.
\begin{definition}
\label{DefinitionE}The non-conforming finite element space $S_{\mathcal{G
}^{p}$ with weak compatibility conditions across facets is given b
\begin{equation}
S_{\mathcal{G}}^{p}:=\left\{ u\in L^{\infty}\left( \Omega\right)
\mid\forall K\in\mathcal{G}\quad\gamma_{K}u\in\mathbb{P}_{p}\left( K\right)
\text{ and }u\text{ satisfies\ (\ref{poinprep})}\right\} . \label{DefSGp
\end{equation}
\end{definition}
The non-conforming Galerkin discretization of (\ref{varform}) for a given
finite element space $S$ which satisfies $S_{\mathcal{G},\operatorname*{nc
}^{p}\subset S\subset S_{\mathcal{G}}^{p}$ reads
\begin{equation}
\text{Given }f\in L^{2}\left( \Omega\right) \text{ find }u_{S}\in
S\text{\quad}a_{\mathcal{G}}\left( u_{S},v\right) :=\left( \mathbb{A
\nabla_{\mathcal{G}}u_{S},\nabla_{\mathcal{G}}v\right) =\left( f,v\right)
\mathbf{\qquad\forall}v\in S \label{ncGalDisc
\end{equation}
where
\[
\nabla_{\mathcal{G}}u\left( \mathbf{x}\right) :=\nabla u\left(
\mathbf{x}\right) \qquad\forall\mathbf{x}\in\Omega\backslash\left(
{\displaystyle\bigcup\limits_{T\in\mathcal{F}}}
\partial T\right) .
\]
\subsection{Non-Conforming Finite Elements of Crouzeix-Raviart Type in
3D\label{SecNCFECR}}
The definition of the non-conforming space $S_{\mathcal{G}}^{p}$ in
(\ref{DefSGp}) is implicit via the weak compatibility condition. In this
section, we will present explicit representations of non-conforming basis
functions of Crouzeix-Raviart type for general polynomial order $p$. These
functions together with the conforming basis functions span a space
$S_{\mathcal{G},\operatorname*{nc}}^{p}$ which satisfies the inclusions
$S_{\mathcal{G},\operatorname*{c}}^{p}\subsetneq S_{\mathcal{G
,\operatorname*{nc}}^{p}\subseteq S_{\mathcal{G}}^{p}$ (cf. Theorem
\ref{thm-convergence-piecewise}). The derivation of the formula and their
algebraic properties will be the topic of the following sections.
We will introduce two types of non-conforming basis functions: those whose
support is one tetrahedron and those whose support consists of two adjacent
tetrahedrons, that is tetrahedrons which have a common facet. For details and
their derivation we refer to Section \ref{SecExplConstr} while here we focus
on the representation formulae.
\subsubsection{Non-Conforming Basis Functions Supported on One Tetrahedron}
The construction starts by defining \textit{symmetric orthogonal polynomials}
$b_{p,k}^{\operatorname*{sym}}$, $0\leq k\leq d_{\operatorname*{triv}}\left(
p\right) -1$ on the reference triangle $\widehat{T}$ with vertices $\left(
0,0\right) ^{\intercal}$, $\left( 1,0\right) ^{\intercal}$, $\left(
0,1\right) ^{\intercal}$, wher
\begin{equation}
d_{\operatorname*{triv}}\left( p\right) :=\left\lfloor \frac{p
{2}\right\rfloor -\left\lfloor \frac{p-1}{3}\right\rfloor . \label{dtrivfirst
\end{equation}
We define the coefficient
\[
M_{i,j}^{\left( p\right) }=\left( -1\right) ^{p}~_{4}F_{3}\left(
\genfrac{}{}{0pt}{}{-j,j+1,-i,i+1}{-p,p+2,1
;1\right) \frac{2i+1}{p+1}\qquad0\leq i,j\leq p,
\]
where $_{p}F_{q}$ denotes the generalized hypergeometric function (cf.
\cite[Chap. 16]{NIST:DLMF}). The $_{4}F_{3}$-sum is understood to terminate at
$i$ to avoid the $0/0$ ambiguities in the formal $_{4}F_{3}$-series. These
coefficients allow to define the polynomial
\[
r_{p,2k}\left( x_{1},x_{2}\right) :=2\sum\limits_{0\leq j\leq p/2
M_{2j,2k}^{\left( n\right) }b_{p,2j}+b_{p,2k}\qquad0\leq k\leq p/2,
\]
where $b_{p,k}$, $0\leq k\leq p$, are the basis for the orthogonal polynomials
of degree $p$ on $\widehat{T}$ as defined afterwards in (\ref{defbnk}). Then,
a basis for the symmetric orthogonal polynomials is given b
\begin{equation}
b_{p,k}^{\operatorname*{sym}}:=\left\{
\begin{array}
[c]{ll
r_{p,p-2k} & \text{if }p\text{ is even,}\\
r_{p,p-1-2k} & \text{if }p\text{ is odd,
\end{array}
\right. \qquad k=0,1,\ldots,d_{\operatorname*{triv}}\left( p\right) -1.
\label{defbpksym
\end{equation}
The non-conforming Crouzeix-Raviart basis function $B_{p,k}^{\widehat
{K},\operatorname*{nc}}\in\mathbb{P}_{p}\left( \widehat{K}\right) $ on the
unit tetrahedron $\widehat{K}$ is characterized\ by its values at the nodal
points in $\widehat{\mathcal{N}}^{p}$ (cf. (\ref{defnodalpointsref})). For a
facet $T\subset\partial\widehat{K}$, let $\chi_{T}:\widehat{T}\rightarrow T$
denote an affine pullback to the reference triangle. Then $B_{p,k
^{\widehat{K},\operatorname*{nc}}\in\mathbb{P}_{p}\left( \widehat{K}\right)
$ is uniquely defined b
\begin{equation}
B_{p,k}^{\widehat{K},\operatorname*{nc}}\left( \mathbf{N}\right) :=\left\{
\begin{array}
[c]{ll
b_{p,k}^{\operatorname*{sym}}\circ\chi_{T}^{-1}\left( \mathbf{N}\right) &
\forall\mathbf{N}\in\widehat{\mathcal{N}}^{p}\text{ s.t. }\mathbf{N}\in
T\text{ for some facet }T\subset\partial\widehat{K},\\
0 & \forall\mathbf{N}\in\widehat{\mathcal{N}}^{p}\backslash\partial\widehat{K
\end{array}
\right. \qquad k=0,1,\ldots,d_{\operatorname*{triv}}\left( p\right) -1.
\label{DefBpkKhut
\end{equation}
\begin{remark}
\label{RemDef}In Sec. \ref{SecFullSymPoly}, we will prove that the polynomials
$b_{p,k}^{\operatorname*{sym}}$ are totally symmetric, i.e., invariant under
affine bijections $\chi:\widehat{K}\rightarrow\widehat{K}$. Thus, any of these
functions can be lifted to the facets of a tetrahedron via affine pullbacks
and the resulting function on the surface is continuous. As a consequence, the
value $B_{p,k}^{\widehat{K},\operatorname*{nc}}\left( \mathbf{N}\right) $ in
definition (\ref{DefBpkKhut}) is independent of the choice of $T$ also for
nodal points $\mathbf{N}$ which belong to different facets.
It will turn out that the value $0$ at the inner nodes could be replaced by
other values without changing the arising non-conforming space. Other choices
could be preferable in the context of inverse inequalities and the condition
number of the stiffness matrix. However, we recommend to choose these values
such that the symmetries of $B_{p,k}^{\widehat{K},\operatorname*{nc}}$ are preserved.
\end{remark}
\begin{definition}
\label{DefSymSpaceK}The non-conforming tetrahedron-supported basis functions
on the reference element are given b
\begin{equation}
B_{p,k}^{\widehat{K},\operatorname*{nc}}=\sum_{\mathbf{N}\in\widehat
{\mathcal{N}}^{p}\cap\partial\widehat{K}}B_{p,k}^{\widehat{K
,\operatorname*{nc}}\left( \mathbf{N}\right) B_{p,\mathbf{N}}^{\mathcal{G
}\qquad k=0,1,\ldots,d_{\operatorname*{triv}}\left( p\right) -1 \label{18b
\end{equation}
with values $B_{p,k}^{\widehat{K},\operatorname*{nc}}\left( \mathbf{N
\right) $ as in (\ref{DefBpkKhut}). For a simplex $K\in\mathcal{G}$ the
corresponding non-conforming basis functions $B_{p,k}^{K,\operatorname*{nc}}$
are given by lifting $B_{p,k}^{\widehat{K},\operatorname*{nc}}$ via an affine
pullback $\chi_{K}$ from $\widehat{K}$ to $K\in\mathcal{G}$
\[
\left. B_{p,k}^{K,\operatorname*{nc}}\right\vert _{\overset{\circ}{K^{\prime
}}}:=\left\{
\begin{array}
[c]{ll
B_{p,k}^{\widehat{K},\operatorname*{nc}}\circ\chi_{K}^{-1} & K=K^{\prime},\\
0 & K\neq K^{\prime}.
\end{array}
\right.
\]
and span the spac
\begin{equation}
S_{K,\operatorname*{nc}}^{p}:=\operatorname*{span}\left\{ B_{p,k
^{K,\operatorname*{nc}}:k=0,1,\ldots,d_{\operatorname*{triv}}\left( p\right)
-1\right\} . \label{DefSkpsym
\end{equation}
\end{definition}
\begin{example}
The lowest order of $p$ such that $d_{\operatorname*{triv}}\left( p\right)
\geq1$ is $p=2$. In this case, we get $d_{\operatorname*{triv}}\left(
p\right) =1$. In Figure \ref{FigSym} the function $b_{p,k
^{\operatorname*{sym}}$ and corresponding basis functions $B_{p,k
^{K,\operatorname*{nc}}$ are depicted for $\left( p,k\right) \in\left\{
\left( 2,0\right) ,\left( 3,0\right) ,\left( 6,0\right) ,\left(
6,1\right) \right\} $
\begin{table}[tbp] \centering
\begin{tabular}
[c]{|l|l|l|l|}\hlin
{\includegraphics[
height=1.1779in,
width=1.4503in
{bsym_2_0.swp.eps
&
{\includegraphics[
height=1.177in,
width=1.4434in
{bsym_3_0.swp.eps
&
{\includegraphics[
height=1.1718in,
width=1.4434in
{bsym_6_0.swp.eps
&
{\includegraphics[
height=1.1666in,
width=1.4364in
{bsym_6_1.swp.eps
\\%
{\includegraphics[
height=1.1018in,
width=1.4676in
{bsymm3D_2_0.eps
&
{\includegraphics[
height=1.0914in,
width=1.4538in
{bsymm3D_3_0.eps
&
{\includegraphics[
height=1.0862in,
width=1.4468in
{bsymm3D_6_0.eps
&
{\includegraphics[
height=1.081in,
width=1.4399in
{bsymm3D_6_1.eps
\\
\multicolumn{1}{|c|}{$b_{2,0}^{\operatorname*{sym}},B_{2,0
^{K,\operatorname*{nc}}$} & \multicolumn{1}{|c|}{$b_{3,0}^{\operatorname*{sym
},B_{3,0}^{K,\operatorname*{nc}}$} & \multicolumn{1}{|c|}{$b_{6,0
^{\operatorname*{sym}},B_{6,0}^{K,\operatorname*{nc}}$} &
\multicolumn{1}{|c|}{$b_{6,1}^{\operatorname*{sym}},B_{6,1
^{K,\operatorname*{nc}}$}\\\hline
\end{tabular}
\caption{Symmetric orthogonal polynomials on the reference triangle and corresponding tetrahedron-supported non-conforming basis functions.\label{FigSym}
\end{table
\end{example}
\subsubsection{Non-Conforming Basis Functions Supported on Two Adjacent
Tetrahedrons}
The starting point is to define orthogonal polynomials $b_{p,k
^{\operatorname*{refl}}$ on the reference triangle $\widehat{T}$ which are
mirror symmetric\footnote{The superscript \textquotedbllef
$\operatorname*{refl}$\textquotedblright\ is a shorthand for \textquotedblleft
reflection\textquotedblright\ and explained in Section \ref{SubSecIrr}.} with
respect to the angular bisector in $\widehat{T}$ through $\mathbf{0}$ and
linear independent from the fully symmetric functions $b_{p,k
^{\operatorname*{sym}}$. We se
\begin{equation}
b_{p,k}^{\operatorname*{refl}}:=\frac{1}{3}\left( 2b_{p,2k}\left(
x_{1},x_{2}\right) -b_{p,2k}\left( x_{2},1-x_{1}-x_{2}\right)
-b_{p,2k}\left( 1-x_{1}-x_{2},x_{1}\right) \right) \qquad0\leq k\leq
d_{\operatorname*{refl}}\left( p\right) -1, \label{brefl1stdef
\end{equation}
wher
\begin{equation}
d_{\operatorname*{refl}}\left( p\right) :=\left\lfloor \frac{p+2
{3}\right\rfloor . \label{defdreflp
\end{equation}
Let $K_{1}$, $K_{2}$ denote two tetrahedrons which share a common facet, say
$T$. The vertex of $K_{i}$ which is opposite to $T$ is denoted by
$\mathbf{V}_{i}$. The procedure of lifting the nodal values to the facets of
$\omega_{T}:=K_{1}\cup K_{2}$ is analogous as for the basis functions
$B_{n,k}^{K,\operatorname*{nc}}$. However, it is necessary to choose the
pullback $\chi_{i,\tilde{T}}:\widehat{T}\rightarrow\tilde{T}$ of a facet
$\tilde{T}\subset\partial K_{i}\backslash\overset{\circ}{T}$ such that the
origin is mapped to $\mathbf{V}_{i}$.
\begin{equation}
B_{p,k}^{T,\operatorname*{nc}}\left( \mathbf{N}\right) :=\left\{
\begin{array}
[c]{ll
b_{p,k}^{\operatorname*{refl}}\circ\chi_{i,\tilde{T}}^{-1}\left(
\mathbf{N}\right) & \forall\mathbf{N}\in\mathcal{N}^{p}\text{ s.t.
}\mathbf{N}\in\tilde{T}\text{ for some facet }\tilde{T}\subset\partial
K\backslash\overset{\circ}{T}_{i},\\
0 & \forall\mathbf{N}\in\mathcal{N}^{p}\cap\overset{\circ}{\omega_{T}
\end{array}
\right. \qquad k=0,1,\ldots,d_{\operatorname*{refl}}\left( p\right) -1.
\label{defvaluesreflt
\end{equation}
Again, the value $0$ at the inner nodes of $\omega_{T}$ could be replaced by
other values without changing the arising non-conforming space.
\begin{definition}
\label{PropBpmTncbasis}The non-conforming facet-oriented basis functions are
given b
\begin{equation}
B_{p,k}^{T,\operatorname*{nc}}=\sum_{\mathbf{N}\in\mathcal{N}^{p}\cap
\partial\omega_{T}}B_{p,k}^{T,\operatorname*{nc}}\left( \mathbf{N}\right)
\left. B_{p,\mathbf{N}}^{\mathcal{G}}\right\vert _{\omega_{T}}\qquad\forall
T\in\mathcal{F}_{\Omega},\quad k=0,1,\ldots,d_{\operatorname*{refl}}\left(
p\right) -1 \label{defedgesupp
\end{equation}
with values $B_{p,k}^{T,\operatorname*{nc}}\left( \mathbf{N}\right) $ as in
(\ref{defvaluesreflt}) and span the spac
\begin{equation}
S_{T,\operatorname*{nc}}^{p}:=\operatorname*{span}\left\{ B_{p,k
^{T,\operatorname*{nc}}:k=0,1,\ldots,d_{\operatorname*{refl}}\left( p\right)
-1\right\} . \label{STpncdef
\end{equation}
The non-conforming finite element space of Crouzeix-Raviart type is given b
\begin{equation}
S_{\mathcal{G},\operatorname*{nc}}^{p}:=\left(
{\displaystyle\bigoplus\limits_{E\in\mathcal{E}_{\Omega}}}
S_{E,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
S_{T,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{K\in\mathcal{G}}}
S_{K,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{K\in\mathcal{G}}}
S_{K,\operatorname*{nc}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
\operatorname*{span}\left\{ B_{p,0}^{T,\operatorname*{nc}}\right\} \right)
. \label{DefSCR
\end{equation}
\end{definition}
\begin{remark}
In Sec. \ref{Sectaureflcomp}, we will show that the polynomials $b_{p,k
^{\operatorname*{refl}}$ are mirror symmetric with respect to the angular
bisector in $\widehat{T}$ through $\mathbf{0}$. Thus, any of these functions
can be lifted to the outer facets of two adjacent tetrahedrons via (oriented)
affine pullbacks as employed in (\ref{defvaluesreflt}) and the resulting
function on the surface is continuous. As a consequence, the value
$B_{p,k}^{T,\operatorname*{nc}}\left( \mathbf{N}\right) $ in definition
(\ref{defvaluesreflt}) is independent of the choice of $T$ also for nodal
points $\mathbf{N}$ which belong to different facets.
In Theorem \ref{Theorem33}, we will prove that (\ref{DefSCR}), in fact, is a
direct sum and a basis is given by the function
\[
B_{p,\mathbf{N}}^{\mathcal{G}}\quad\forall\mathbf{N}\in\mathcal{N}_{\Omega
}\backslash\mathcal{V},\quad B_{p,k}^{K,\operatorname*{nc}}\quad\forall
K\in\mathcal{G},0\leq k\leq d_{\operatorname*{triv}}\left( p\right) -1,\quad
B_{p,0}^{T,\operatorname*{nc}}\quad\forall T\in\mathcal{F}_{\Omega}.
\]
Also we will prove that $S_{\mathcal{G},\operatorname*{c}}^{p}\subsetneq
S_{\mathcal{G},\operatorname*{nc}}^{p}\subseteq S_{\mathcal{G}}^{p}$. This
condition implies that the convergence estimates as in Theorem
\ref{thm-convergence-piecewise} are valid for this space. We restricted the
reflection-type non-conforming basis functions to the lowest order $k=0$ in
order to keep the functions linearly independent.
\end{remark}
\begin{example}
The lowest order of $p$ such that $d_{\operatorname*{refl}}\left( p\right)
\geq1$ is $p=1$. In this case, we get $d_{\operatorname*{refl}}\left(
p\right) =1$. In Figure \ref{FigRefl} the function $b_{p,k
^{\operatorname*{refl}}$ and corresponding basis functions $B_{p,k
^{T,\operatorname*{nc}}$ are depicted for $\left( p,k\right) \in\left\{
\left( 1,0\right) ,\left( 2,0\right) ,\left( 4,0\right) ,\left(
4,1\right) \right\} $
\begin{table}[tbp] \centering
\begin{tabular}
[c]{|c|c|c|c|}\hlin
{\includegraphics[
height=1.222in,
width=1.5065in
{prefl_1_0.swp.eps
&
{\includegraphics[
height=1.2168in,
width=1.5273in
{prefl_2_0.swp.eps
&
{\includegraphics[
height=1.222in,
width=1.4987in
{prefl_4_0.swp.eps
&
{\includegraphics[
height=1.2168in,
width=1.5273in
{prefl_4_1.swp.eps
\\%
{\includegraphics[
height=1.2721in,
width=1.51in
{brefl3D_1_0_full_Feb17opti.eps
&
{\includegraphics[
height=1.2695in,
width=1.5866in
{brefl3D_2_0_full_Feb17opti.eps
&
{\includegraphics[
height=1.2721in,
width=1.5004in
{brefl3D_4_0_full_Feb17opti.eps
&
{\includegraphics[
height=1.2687in,
width=1.4183in
{brefl3D_4_1_full_Feb17opti.eps
\\
\multicolumn{1}{|c|}{$b_{1,0}^{\operatorname*{refl}},B_{1,0
^{T,\operatorname*{nc}}$} & \multicolumn{1}{|c|}{$b_{2,0
^{\operatorname*{refl}},B_{2,0}^{T,\operatorname*{nc}}$} &
\multicolumn{1}{|c|}{$b_{4,0}^{\operatorname*{refl}},B_{4,0
^{T,\operatorname*{nc}}$} & \multicolumn{1}{|c|}{$b_{4,1
^{\operatorname*{refl}},B_{4,1}^{T,\operatorname*{nc}}$}\\\hline
\end{tabular}
\caption{Orthogonal polynomials of reflection type and corresponding non-conforming basis functions which are supported on two adjacent tetrahedrons.
The common facet is horizontal and the two tetrahedrons are on top of each other.\label{FigRefl}
\end{table
\end{example}
\subsection{Error Analysis}
In this subsection we present the error analysis for the Galerkin
discretization (\ref{ncGalDisc}) with the non-conforming finite element space
$S_{\mathcal{G}}^{p}$ and subspaces thereof. The analysis is based on the
second Strang lemma and has been presented for an intrinsic version of
$S_{\mathcal{G}}^{p}$ in \cite{ccss_2012}.
For any inner facet $T\in\mathcal{F}$ and any $v\in S_{\mathcal{G}}^{p}$,
condition (\ref{poinprep}) implies $\int_{T}\left[ v\right] _{T}=0$ : hence,
the jump $\left[ v\right] _{T}$ is always zero-mean valued. Let $h_{T}$
denote the diameter of $T$. The combination of a Poincar\'{e} inequality with
a trace inequality then yield
\begin{equation}
\left\Vert \left[ u\right] _{T}\right\Vert _{L^{2}\left( T\right) }\leq
Ch_{T}\left\vert \left[ u\right] _{T}\right\vert _{H^{1}\left( T\right)
}\leq\tilde{C}h_{T}^{1/2}\left\vert u\right\vert _{H_{\operatorname*{pw}
^{1}\left( \omega_{T}\right) }, \label{poincarecondspa
\end{equation}
wher
\[
\left\vert u\right\vert _{H_{\operatorname*{pw}}^{p}\left( \omega_{T}\right)
}:=\left( \sum_{K\subset\omega_{T}}\left\vert u\right\vert _{H^{p}\left(
K\right) }^{2}\right) ^{1/2}\text{.
\]
In a similar fashion we obtain for all boundary facets $T\in\mathcal{F
_{\partial\Omega}$ and all $u\in S_{\mathcal{G}}^{p}$ the estimat
\begin{equation}
\left\Vert u\right\Vert _{L^{2}\left( T\right) }\leq\tilde{C}h_{T
^{1/2}\left\vert u\right\vert _{H_{\operatorname*{pw}}^{1}\left( \omega
_{T}\right) }. \label{poincarecondspb
\end{equation}
{We say that the exact solution }${u}\in H${$_{0}^{1}\left( \Omega\right) $
is piecewise smooth over the partition $\mathcal{P=}\left( \Omega_{j}\right)
_{j=1}^{J}$},{ if there exists some positive integer $s$ such that
\[
u_{|\Omega_{j}}\in H^{1+s}\left( \Omega_{j}\right) \quad\text{for
}j=1,2,\ldots,J.
\]
We write }$u\in${$PH^{1+s}(\Omega)$ and refer for further properties and
generalizations to non-integer values of }$s${, e.g., to \cite[Sec.
4.1.9]{SauterSchwab2010}.}
{For the approximation results, the finite element meshes $\mathcal{G}$ are
assumed to be \textit{compatible} with the partition $\mathcal{P}$ in the
following sense: for all $K\in\mathcal{G}$, there exists a single index $j$
such that }$\overset{\circ}{{K}}${$\cap\Omega_{j}\neq\emptyset$. }
The proof that $\left\vert \cdot\right\vert _{H_{\operatorname*{pw}
^{1}\left( \Omega\right) }$ is a norm on $S_{\mathcal{G}}^{p}$ is similar as
in \cite[Sect. 10.3]{scottbrenner3}: For $w\in H_{0}^{1}\left( \Omega\right)
$ this follows from $\left\vert w\right\vert _{H_{\operatorname*{pw}
^{1}\left( \Omega\right) }=\left\Vert \nabla w\right\Vert $ and a Friedrichs
inequality; for $w\in S_{\mathcal{G}}^{p}$ the condition $\left\Vert
\nabla_{\mathcal{G}}w\right\Vert =0$ implies that $\left. w\right\vert _{K}$
is constant on all simplices $K\in\mathcal{G}$. The combination with $\int
_{T}w=0$ for all $T\in\mathcal{F}_{\partial\Omega}$ leads to $\left.
w\right\vert _{K}=0$ for the outmost simplex layer via a Poincar\'{e}
inequality, i.e., $\left. w\right\vert _{K}=0$ for all $K\in\mathcal{G}$
having at least one facet on $\partial\Omega$. This argument can be iterated
step by step over simplex layers towards the interior of $\Omega$ to finally
obtain $w=0$.
\begin{theorem}
\label{thm-convergence-piecewise}Let $\Omega\subset\mathbb{R}^{d}$ be a
bounded, polygonal ($d=2$) or polyhedral ($d=3$) Lipschitz domain and let
$\mathcal{G}$ be a regular simplicial finite element mesh for $\Omega$. Let
the diffusion matrix $\mathbb{A}\in L^{\infty}\left( \Omega,\mathbb{R
_{\operatorname*{sym}}^{d\times d}\right) $ satisfy assumption (\ref{aeps})
and let $f\in L^{2}\left( \Omega\right) $. As an additional assumption on
the regularity, we require that the exact solution of (\ref{varform})
satisfies $u\in PH{^{1+s}\left( \Omega\right) }$ for some positive integer
$s$ and $\left\Vert \mathbb{A}\right\Vert _{{PW^{{r},\infty}\left(
\Omega\right) }}<\infty$ holds with $r:=\min\left\{ p,s\right\} $. Let the
continuous problem (\ref{varform}) be discretized by the non-conforming
Galerkin method (\ref{ncGalDisc}) with a finite dimensional space $S$ which
satisfies $S_{\mathcal{G},\operatorname*{c}}^{p}\subset S\subset
S_{\mathcal{G}}^{p}$ on{ a compatible mesh $\mathcal{G}$}. Then,
(\ref{ncGalDisc}) has a unique solution which satisfie
\[
\left\vert u-u_{S}\right\vert _{H_{\operatorname*{pw}}^{1}\left(
\Omega\right) }\leq Ch^{r}\left\Vert u\right\Vert _{PH{^{1+r}\left(
\Omega\right) }}.
\]
The constant $C$ only depends on $a_{\min}$, $a_{\max}$, $\left\Vert
\mathbb{A}\right\Vert _{{PW^{{r},\infty}\left( \Omega\right) }}$, $p$, $r$,
and the shape regularity of the mesh.
\end{theorem}
\proof
The second Strang lemma (cf. \cite[Theo. 4.2.2]{Ciarlet}) applied to the
non-conforming Galerkin discretization (\ref{ncGalDisc}) implies the existence
of a unique solution which satisfies the error estimat
\[
\left\vert u-u_{S}\right\vert _{H_{\operatorname*{pw}}^{1}\left(
\Omega\right) }\leq\left( 1+\frac{a_{\max}}{a_{\min}}\right) \inf_{v\in
S}\left\vert u-v\right\vert _{H_{\operatorname*{pw}}^{1}\left( \Omega\right)
}+\frac{1}{a_{\min}}\sup_{v\in S}\frac{\left\vert \mathcal{L}_{u}\left(
v\right) \right\vert }{\left\vert v\right\vert _{H_{\operatorname*{pw}
^{1}\left( \Omega\right) }},
\]
wher
\[
\mathcal{L}_{u}\left( v\right) :=a_{\mathcal{G}}\left( u,v\right) -\left(
f,v\right) .
\]
The approximation properties of $S$ are inherited from the approximation
properties of $S_{\mathcal{G},\operatorname*{c}}^{p}$ in the first infimum
because of the inclusion $S_{\mathcal{G},\operatorname*{c}}^{p}\subset S$. For
the second term we obtai
\begin{equation}
\mathcal{L}_{u}\left( v\right) =\left( \mathbb{A}\nabla u,\nabla
_{\mathcal{G}}v\right) -\left( f,v\right) . \label{defL-piecewise
\end{equation}
Note that $f\in L^{2}\left( \Omega\right) $ implies that $\operatorname{div
\left( \mathbb{A}\nabla u\right) \in L^{2}\left( \Omega\right) $ and, in
turn, that the normal jump $\left[ \mathbb{A}\nabla u\cdot\mathbf{n
_{T}\right] _{T}$ equals zero and the restriction $\left. \left(
\mathbb{A}\nabla u\cdot\mathbf{n}_{T}\right) \right\vert _{T}$ is well
defined for all $T\in\mathcal{F}$. We may apply simplexwise integration by
parts to (\ref{defL-piecewise}) to obtai
\[
\mathcal{L}_{u}\left( v\right) =-\sum_{T\in\mathcal{F}_{\Omega}}\,\int
_{T}\left( \mathbb{A}\nabla u\cdot\mathbf{n}_{T}\right) \left[ v\right]
_{T}+\sum_{T\in\mathcal{F}_{\partial\Omega}}\int_{T}\left( \mathbb{A}\nabla
u\cdot\mathbf{n}_{T}\right) v.
\]
Let $K_{T}${ be one simplex in $\omega_{T}$. For }$1\leq i\leq d$,{ let}
$q_{i}\in\mathbb{P}_{d}^{p-1}\left( K_{T}\right) $ denote the best
approximation of $w_{i}:=\left. \left( \sum_{j=1}^{d}A_{i,j}\partial
_{j}u\right) \right\vert _{K_{T}}$ with respect to the $H^{1}\left(
K_{T}\right) $ norm. Then, $\left. q_{i}\right\vert _{T}n_{T,i}\in
\mathbb{P}_{d-1}^{p-1}\left( T\right) $ for $1\leq i\leq d$, and the
inclusion $S\subset S_{\mathcal{G}}^{p}$ implie
\begin{align}
\left\vert \mathcal{L}_{u}\left( v\right) \right\vert \leq & \left\vert
-\sum_{T\in\mathcal{F}_{\Omega}}\int_{T}\left( \sum_{i=1}^{d}\left(
w_{i}-q_{i}\right) \cdot n_{T,i}\right) \left[ v\right] _{T}\right\vert
\label{Luest1}\\
& +\left\vert \sum_{T\in\mathcal{F}_{\partial\Omega}}\int_{T}\left(
\sum_{i=1}^{d}\left( w_{i}-q_{i}\right) \cdot n_{T,i}\right) v\right\vert
\nonumber\\
\leq & \sum_{T\in\mathcal{F}_{\Omega}}\left\Vert \left[ v\right]
_{T}\right\Vert _{L^{2}\left( T\right) }\sum_{i=1}^{d}\left\Vert w_{i
-q_{i}\right\Vert _{L^{2}\left( T\right) }\nonumber\\
& +\sum_{T\in\mathcal{F}_{\partial\Omega}}\left\Vert v\right\Vert
_{L^{2}\left( T\right) }\sum_{i=1}^{d}\left\Vert w_{i}-q_{i}\right\Vert
_{L^{2}\left( T\right) }.\nonumber
\end{align}
Standard trace estimates and approximation properties lead t
\begin{align}
\left\Vert w_{i}-q_{i}\right\Vert _{L^{2}\left( T\right) } & \leq C\left(
h_{T}^{-1/2}\left\Vert w_{i}-q_{i}\right\Vert _{L^{2}\left( K_{T}\right)
}+h_{T}^{1/2}\left\vert w_{i}-q_{i}\right\vert _{H^{1}\left( K_{T}\right)
}\right) \label{Luest2}\\
& \leq Ch_{T}^{r-1/2}\left\vert w_{i}\right\vert _{H^{r}\left( K_{T}\right)
}\leq Ch_{T}^{r-1/2}\left\Vert u\right\Vert _{H^{1+r}\left( K_{T}\right)
},\nonumber
\end{align}
where $C$ depends only on $p$, $r$, $\left\Vert \mathbb{A}\right\Vert
_{{W^{r}\left( K_{T}\right) }}$, and the shape regularity of the mesh{.}The
combination of (\ref{Luest1}), (\ref{Luest2}) and (\ref{poincarecondspa
),(\ref{poincarecondspb}) along with the shape regularity of the mesh leads to
the consistency estimat
\begin{align*}
\left\vert \mathcal{L}_{u}\left( v\right) \right\vert & \leq C\left(
\sum_{T\in\mathcal{F}_{\Omega}}h_{T}^{r}\left\Vert u\right\Vert _{H^{1+r
\left( K_{T}\right) }\left\vert v\right\vert _{H_{\operatorname*{pw}
^{1}\left( \omega_{T}\right) }+\sum_{T\in\mathcal{F}_{\partial\Omega}
h_{T}^{r}\left\Vert u\right\Vert _{H^{1+r}\left( K_{T}\right) }\left\vert
v\right\vert _{H_{\operatorname*{pw}}^{1}\left( \omega_{T}\right) }\right)
\\
& \leq\tilde{C}h^{r}\left\Vert u\right\Vert _{PH{^{1+r}\left( \Omega\right)
}}\left\vert v\right\vert _{H_{\operatorname*{pw}}^{1}\left( \Omega\right)
},
\end{align*}
which completes the proof.\hfil
\endproof
\begin{remark}
If one chooses in (\ref{poinprep}) a degree $p^{\prime}<p$ for the
orthogonality relations in (\ref{poinprep}), then the order of convergence
behaves like $h^{r^{\prime}}\left\Vert e\right\Vert _{H^{1+r^{\prime}}\left(
\Omega\right) }$, with $r^{\prime}:=\min\left\{ p^{\prime},s\right\} $,
because the best approximations $q_{i}$ now belong to $P_{d-1}^{p^{\prime
-1}\left( T\right) $.
\end{remark}
\section{Explicit Construction of Non-Conforming Crouzeix-Raviart Finite
Elements\label{SecExplConstr}}
\subsection{Jacobi Polynomials}
Let $\alpha,\beta>-1$. The \emph{Jacobi polynomial} $P_{n}^{\left(
\alpha,\beta\right) }$ is a polynomial of degree $n$ such that
\[
\int_{-1}^{1}P_{n}^{\left( \alpha,\beta\right) }\left( x\right) \,q\left(
x\right) \left( 1-x\right) ^{\alpha}\left( 1+x\right) ^{\beta}\,dx=0
\]
for all polynomials $q$ of degree less than $n$, and (cf. \cite[Table
18.6.1]{NIST:DLMF}
\begin{equation}
P_{n}^{\left( \alpha,\beta\right) }\left( 1\right) =\frac{\left(
\alpha+1\right) _{n}}{n!},\qquad P_{n}^{\left( \alpha,\beta\right) }\left(
-1\right) =\left( -1\right) ^{n}\frac{\left( \beta+1\right) _{n}}{n!}.
\label{Pnormalization
\end{equation}
Here the \emph{shifted factorial} is defined by $\left( a\right)
_{n}:=a\left( a+1\right) \ldots\left( a+n-1\right) $ for $n>0$ and
$\left( a\right) _{0}:=1$. The Jacobi polynomial has an explicit expression
in terms of a \emph{terminating Gauss hypergeometric series} (see (cf.
\cite[18.5.7]{NIST:DLMF})
\begin{equation}
\,\mbox{}_{2}F_{1}\!\left(
\genfrac{}{}{0pt}{}{-n,b}{c
;z\right) :=\sum_{k=0}^{n}\frac{\left( -n\right) _{k}\left( b\right)
_{k}}{\left( c\right) _{k}\,k!}\,z^{k} \label{F21rep
\end{equation}
as follow
\begin{equation}
P_{n}^{\left( \alpha,\beta\right) }\left( x\right) =\frac{\left(
\alpha+1\right) _{n}}{n!}\,\,\mbox{}_{2}F_{1}\!\left(
\genfrac{}{}{0pt}{}{-n,n+\alpha+\beta+1}{\alpha+1
;\frac{1-x}{2}\right) . \label{defJP
\end{equation}
\subsection{Orthogonal Polynomials on Triangles\label{SecOrthoPolyT}}
Recall that $\widehat{T}$ is the (closed) unit triangle in $\mathbb{R}^{2}$
with vertices $\widehat{\mathbf{A}}_{0}=\left( 0,0\right) ^{\intercal}$,
$\widehat{\mathbf{A}}_{1}=\left( 1,0\right) ^{\intercal}$, and
$\widehat{\mathbf{A}}_{3}=\left( 0,1\right) ^{\intercal}$. An orthogonal
basis for the space $\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $
was introduced in \cite{Proriol_ortho} and is given by the functions $b_{n,k
$, $0\leq k\leq n$,
\begin{equation}
b_{n,k}(\mathbf{x}):=\left( x_{1}+x_{2}\right) ^{k}\,P_{n-k}^{(0,2k+1)
\left( 2\left( x_{1}+x_{2}\right) -1\right) \,P_{k}^{\left( 0,0\right)
}\left( \frac{x_{1}-x_{2}}{x_{1}+x_{2}}\right) , \label{defbnk
\end{equation}
where $P_{k}^{\left( 0,0\right) }$ are the Legendre polynomials (see
\cite[18.7.9]{NIST:DLMF})\footnote{The Legendre polynomials with normalization
$P_{k}^{\left( 0,0\right) }\left( 1\right) =1$ for all $k=0,1,\ldots$ can
be defined \cite[Table 18.9.1]{NIST:DLMF} via the three-term recursio
\begin{equation}
P_{0}^{\left( 0,0\right) }\left( x\right) =1;\quad P_{1}^{\left(
0,0\right) }\left( x\right) =x;\quad\text{and\quad}\left( k+1\right)
P_{k+1}^{\left( 0,0\right) }\left( x\right) =\left( 2k+1\right)
xP_{k}^{\left( 0,0\right) }\left( x\right) -kP_{k-1}^{\left( 0,0\right)
}\left( x\right) \quad\text{for }k=1,2,\ldots, \label{rekLk
\end{equation}
from which the well-known relation $P_{k}^{\left( 0,0\right) }\left(
x\right) =\left( -1\right) ^{k}P_{k}^{\left( 0,0\right) }\left(
x\right) $ for all $k\in\mathbb{N}_{0}$ follows.}. From (\ref{rekLk})
(footnote) it follows that these polynomials satisfy the following symmetry
relatio
\begin{equation}
b_{n,k}\left( x_{1},x_{2}\right) =\left( -1\right) ^{k}b_{n,k}\left(
x_{2},x_{1}\right) \quad\forall n\geq0,\forall\left( x_{1},x_{2}\right) .
\label{symmbnk
\end{equation}
By combining (\ref{F21rep}) - (\ref{defbnk}), an elementary calculation leads
to\footnote{Further special values ar
\
\begin{array}
[c]{ll
b_{n,0}\left( 0,0\right) =P_{n}^{\left( 0,1\right) }\left( -1\right)
=\left( -1\right) ^{n}\frac{\left( 2\right) _{n}}{n!}=\left( -1\right)
^{n}\left( n+1\right) , & b_{n,k}\left( 0,0\right) =0,~1\leq k\leq n,\\
b_{n,k}\left( 1,0\right) =P_{n-k}^{\left( 0,2k+1\right) }\left( 1\right)
P_{k}^{\left( 0,0\right) }\left( 1\right) =1,0\leq k\leq n, &
b_{n,k}\left( 0,1\right) =P_{n-k}^{\left( 0,2k+1\right) }\left( 1\right)
P_{k}^{\left( 0,0\right) }\left( -1\right) =\left( -1\right) ^{k},~0\leq
k\leq n.
\end{array}
\]
} $b_{n,0}\left( 0,0\right) =\left( -1\right) ^{n}\left( n+1\right) $.
Let
\begin{equation}
E^{\operatorname{I}}:=\overline{\widehat{\mathbf{A}}_{0}\widehat{\mathbf{A
}_{1}}\text{,\quad}E^{\operatorname{II}}:=\overline{\widehat{\mathbf{A}
_{0}\widehat{\mathbf{A}}_{2}}\text{,\quad and\quad}E^{\operatorname{III
}:=\overline{\widehat{\mathbf{A}}_{1}\widehat{\mathbf{A}}_{2}}
\label{defedges
\end{equation}
denote the edges of $\widehat{T}$. For $\operatorname{Z}\in\left\{
\operatorname{I},\operatorname{II},\operatorname{III}\right\} $, we introduce
the linear restriction operator for the edge $E^{\operatorname{Z}}$ by
$\gamma^{\operatorname{Z}}:C^{0}\left( \widehat{T}\right) \rightarrow
C^{0}\left( \left[ 0,1\right] \right) $ b
\begin{equation}
\gamma^{\operatorname{I}}u:=u\left( \cdot,0\right) ,\quad\gamma
^{\operatorname{II}}u:=u\left( 0,\cdot\right) ,\quad\gamma
^{\operatorname{III}}u=u\left( 1-\cdot,\cdot\right) \label{defgamma
\end{equation}
which allows to defin
\[
b_{n,k}^{\operatorname{I}}:=\gamma^{\operatorname{I}}b_{n,k},\quad
b_{n,k}^{\operatorname{II}}:=\gamma^{\operatorname{II}}b_{n,k},\quad
b_{n,k}^{\operatorname{III}}:=\gamma^{\operatorname{III}}b_{n,k
,\quad\text{for }k=0,1,\ldots,n.
\]
\begin{lemma}
\label{LemRestLinIndep}For any $\operatorname{Z}\in\left\{ \operatorname{I
,\operatorname{II},\operatorname{III}\right\} $, each of the systems $\left(
b_{n,k}^{\operatorname{Z}}\right) _{k=0}^{n}$, form a basis of $\mathbb{P
_{n}\left( \left[ 0,1\right] \right) $.
\end{lemma}
\proof
First note that $\left\{ x^{j}\left( x-1\right) ^{n-j}:0\leq j\leq
n\right\} $ is a basis for $\mathbb{P}_{n}\left( \left[ 0,1\right]
\right) $; this follows from expanding the right-hand side of $x^{m
=x^{m}\left( x-\left( x-1\right) \right) ^{n-m}$. Specialize the formula
\cite[18.5.8]{NIST:DLMF
\[
P_{m}^{\left( \alpha,\beta\right) }\left( s\right) =\frac{\left(
\alpha+1\right) _{m}}{m!}\left( \frac{1+s}{2}\right) ^{m}~_{2}F_{1}\left(
\genfrac{}{}{0pt}{}{-m,-m-\beta}{\alpha+1
;\frac{s-1}{s+1}\right)
\]
to $m=n-k$, $\alpha=0,\beta=2k+1,s=2x-1$ to obtai
\begin{align}
b_{n,k}^{\operatorname{I}}\left( x\right) & =x^{n}~_{2}F_{1}\left(
\genfrac{}{}{0pt}{}{k-n,-n-k-1}{1
;\frac{x-1}{x}\right) \label{defbtilde}\\
& \overset{\text{(\ref{F21rep})}}{=}\sum_{i=0}^{n-k}\frac{\left( k-n\right)
_{i}\left( -n-k-1\right) _{i}}{i!i!}x^{n-i}\left( x-1\right) ^{i}.
\label{clnk
\end{align}
The highest index $i$ of $x^{n-i}\left( x-1\right) ^{i}$ in $b_{n,k
^{\operatorname{I}}\left( x\right) $ is $n-k$ with coefficient
$\dfrac{\left( 2k+2\right) _{n-k}}{\left( n-k\right) !}\neq0$. Thus the
matrix expressing $\left[ b_{n,0}^{\operatorname{I}},\ldots,b_{n,n
^{\operatorname{I}}\right] $ in terms of $\left[ \left( x-1\right)
^{n},x\left( x-1\right) ^{n-1},\ldots,x^{n}\right] $ is triangular and
nonsingular; hence $\left\{ b_{n,k}^{\operatorname{I}}:0\leq k\leq n\right\}
$ is a basis of $\mathbb{P}_{n}\left( \left[ 0,1\right] \right) $. The
symmetry relation $b_{n,k}^{\operatorname{II}}=\left( -1\right) ^{k
b_{n,k}^{\operatorname{I}}$ for $0\leq k\leq n$ (cf. (\ref{symmbnk})) shows
that $\left\{ b_{n,k}^{\operatorname{II}}:0\leq k\leq n\right\} $ is also a
basis of $\mathbb{P}_{n}\left( \left[ 0,1\right] \right) $. Finally
substituting $x_{1}=1-x,x_{2}=x$ in $b_{n,k}$ results i
\begin{equation}
b_{n,k}^{\operatorname{III}}\left( x\right) =P_{n-k}^{\left( 0,2k+1\right)
}\left( 1\right) P_{k}^{\left( 0,0\right) }\left( 1-2x\right) ,
\label{bnkIII
\end{equation}
and $P_{n-k}^{\left( 0,2k+1\right) }\left( 1\right) =1$ (from
(\ref{Pnormalization})). Clearly $\left\{ P_{k}^{\left( 0,0\right) }\left(
1-2x\right) :0\leq k\leq n\right\} $ is a basis for $\mathbb{P}_{n}\left(
\left[ 0,1\right] \right) $
\endproof
\begin{lemma}
\label{TraceLemma}Let $v\in\mathbb{P}_{n}\left( \left[ 0,1\right] \right)
$. Then, there exist unique orthogonal polynomials $u^{\operatorname{Z}
\in\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $, $\operatorname{Z
\in\left\{ \operatorname{I},\operatorname{II},\operatorname{III}\right\} $
with $v=\gamma^{\operatorname{Z}}u^{\operatorname{Z}}$. Thus, the linear
extension operator $\mathcal{E}^{\operatorname{Z}}:\mathbb{P}_{n}\left(
\left[ 0,1\right] \right) \rightarrow\mathbb{P}_{n,n-1}^{\perp}\left(
\widehat{T}\right) $ is well defined by $\mathcal{E}^{\operatorname{Z
}v:=u^{\operatorname{Z}}$.
\end{lemma}
\proof
From Lemma \ref{LemRestLinIndep} we conclude that $\gamma^{\operatorname{Z}}$
is surjective. Since the polynomial spaces are finite dimensional the
assertion follows fro
\[
\dim\mathbb{P}_{n}\left( \left[ 0,1\right] \right) =n+1=\dim
\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) .
\
\endproof
The orthogonal polynomials can be lifted to a general triangle $T$.
\begin{definition}
Let $T$ denote a triangle and $\chi_{T}$ an affine pullback to the reference
triangle $\widehat{T}$. Then, the space of orthogonal polynomials of degree
$n$ on $T$ i
\[
\mathbb{P}_{n,n-1}^{\perp}\left( T\right) :=\left\{ v\circ\chi_{T
^{-1}:v\in\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) \right\} .
\]
\end{definition}
From the transformation rule for integrals one concludes that for any
$u=v\circ\chi_{T}^{-1}\in\mathbb{P}_{n,n-1}^{\perp}\left( T\right) $ and all
$q\in\mathbb{P}_{n-1}\left( T\right) $ it hold
\begin{equation}
\int_{T}uq=\int_{T}\left( v\circ\chi_{T}^{-1}\right) q=2\left\vert
T\right\vert \int_{\widehat{T}}v\left( q\circ\chi_{T}\right) =0
\label{orthogonalityT
\end{equation}
since $q\circ\chi_{T}\in\mathbb{P}_{n-1}\left( \widehat{T}\right) $. Here
$\left\vert T\right\vert $ denotes the area of the triangle $T$.
\subsection{Totally Symmetric Orthogonal Polynomials\label{SecFullSymPoly}}
In this section, we will decompose the space of orthogonal polynomials
$\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $ into three
irreducible modules (see \S \ref{SubSecIrr}) and thus, obtain a direct sum
decomposition $\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right)
=\mathbb{P}_{n,n-1}^{\perp,\operatorname*{sym}}\left( \widehat{T}\right)
\oplus\mathbb{P}_{n,n-1}^{\perp,\operatorname*{refl}}\left( \widehat
{T}\right) \oplus\mathbb{P}_{n,n-1}^{\perp,\operatorname{sign}}\left(
\widehat{T}\right) $. We will derive an explicit representation for a basis
of the space of totally symmetric polynomials $\mathbb{P}_{n,n-1
^{\perp,\operatorname*{sym}}\left( \widehat{T}\right) $ in
\S \ref{ConSymmBas} and of the space of reflection symmetric polynomials
$\mathbb{P}_{n,n-1}^{\perp,\operatorname*{refl}}\left( \widehat{T}\right) $
in \S \ref{Sectaureflcomp}.
We start by introducing, for functions on triangles, the notation of total
symmetry. For an arbitrary triangle $T$ with vertices $\mathbf{A}_{0}$,
$\mathbf{A}_{1}$, $\mathbf{A}_{2}$, we introduce the set of permutations
$\Pi=\left\{ \left( i,j,k\right) :i,j,k\in\left\{ 0,1,2\right\} \text{
pairwise disjoint}\right\} $. For $\pi=\left( i,j,k\right) \in\Pi$, define
the affine mapping $\chi_{\pi}:T\rightarrow T$ by
\begin{equation}
\chi_{\pi}\left( \mathbf{x}\right) =\mathbf{A}_{i}+x_{1}\left(
\mathbf{A}_{j}-\mathbf{A}_{i}\right) +x_{2}\left( \mathbf{A}_{k
-\mathbf{A}_{i}\right) . \label{defchipi
\end{equation}
We say a function $u$, defined on $T$, has \textit{total symmetry} i
\[
u=u\circ\chi_{\pi}\quad\forall\pi\in\Pi\text{.
\]
The space of totally symmetric orthogonal polynomials i
\begin{equation}
\mathbb{P}_{n,n-1}^{\perp,\operatorname*{sym}}\left( \widehat{T}\right)
:=\left\{ u\in\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) :u\text{
has total symmetry}\right\} . \label{defpsymhat
\end{equation}
The construction of a basis of $\mathbb{P}_{n,n-1}^{\perp,\operatorname*{sym
}\left( \widehat{T}\right) $ requires some algebraic tools which we develop
in the following.
\subsubsection{The decomposition of $\mathbb{P}_{n,n-1}^{\perp}\left(
\widehat{T}\right) $ or $\mathbb{P}_{n}\left( \left[ 0,1\right] \right) $
into irreducible $\mathcal{S}_{3}$ modules\label{SubSecIrr}}
We use the operator $\mathcal{\gamma}^{\operatorname{I}}$ (cf. (\ref{defgamma
)) to set up an action of the symmetric group $\mathcal{S}_{3}$ on
$\mathbb{P}_{n}\left( \left[ 0,1\right] \right) $ by transferring its
action on $\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $ on the
basis $\left\{ b_{n,k}\right\} $. It suffices to work with two generating
reflections. On the triangle $\chi_{\left\{ 0,2,1\right\} }\left(
x_{1},x_{2}\right) =\left( x_{2},x_{1}\right) $ and thus $b_{n,k}\circ
\chi_{\left\{ 0,2,1\right\} }=\left( -1\right) ^{k}b_{n,k}$ (this follows
from (\ref{symmbnk})). The action of $\chi_{\left\{ 0,2,1\right\} }$ is
mapped to $\sum_{k=0}^{n}\alpha_{k}b_{n,k}^{\operatorname{I}}\mapsto\sum
_{k=0}^{n}\left( -1\right) ^{k}\alpha_{k}b_{n,k}^{\operatorname{I}}$, and
denoted by $R$. For the other generator we use $\chi_{\left\{ 1,0,2\right\}
}\left( x_{1},x_{2}\right) =\left( 1-x_{1}-x_{2},x_{2}\right) $. Under
$\gamma^{\operatorname{I}}$ this corresponds to the map $\sum_{k=0}^{n
\alpha_{k}b_{n,k}^{\operatorname{I}}\left( x\right) \mapsto\sum_{k=0
^{n}\alpha_{k}b_{n,k}^{\operatorname{I}}\left( 1-x\right) $ which is denoted
by $M$. We will return later to transformation formulae expressin
\[
b_{n,k}\circ\chi_{\left\{ 1,0,2\right\} }\left( x_{1},x_{2}\right)
=\left( 1-x_{1}\right) ^{k}P_{n-k}^{\left( 0,2k+1\right) }\left(
1-2x_{1}\right) P_{k}^{\left( 0,0\right) }\left( \frac{1-x_{1}-2x_{2
}{1-x_{1}}\right)
\]
in the $\left\{ b_{n,k}\right\} $-basis. Observe that $\left( MR\right)
^{3}=I$ because $\chi_{\left\{ 1,0,2\right\} }\circ\chi_{\left\{
0,2,1\right\} }\left( x_{1},x_{2}\right) =\left( 1-x_{1}-x_{2
,x_{1}\right) $ and this mapping is of period 3. It follows that each of
$\left\{ M,R\right\} $ and $\left\{ \chi_{\left\{ 1,0,2\right\}
,\chi_{\left\{ 0,2,1\right\} }\right\} $ generates (an isomorphic copy of)
$\mathcal{S}_{3}$. It is a basic fact that the relations $M^{2}=I,R^{2}=I$ and
$\left( MR\right) ^{3}=I$ define $\mathcal{S}_{3}$. The representation
theory of $\mathcal{S}_{3}$ informs us that there are three nonisomorphic
irreducible representations
\begin{align*}
\tau_{\operatorname*{triv}} & :\chi_{\left\{ 0,2,1\right\}
\rightarrow1,\chi_{\left\{ 1,0,2\right\} }\rightarrow1;\\
\tau_{\operatorname{sign}} & :\chi_{\left\{ 0,2,1\right\} }\rightarrow
-1,\chi_{\left\{ 1,0,2\right\} }\rightarrow-1;\\
\tau_{\operatorname*{refl}} & :\chi_{\left\{ 0,2,1\right\}
\rightarrow\sigma_{1}:
\begin{bmatrix}
-1 & 0\\
0 & 1
\end{bmatrix}
,\chi_{\left\{ 1,0,2\right\} }\rightarrow\sigma_{2}:
\begin{bmatrix}
\frac{1}{2} & 1\\
\frac{3}{4} & -\frac{1}{2
\end{bmatrix}
.
\end{align*}
(The subscript \textquotedblleft\textit{$\operatorname*{refl}$
\textquotedblright\textit{ }designates the \textit{reflection
representation).\textit{ }Then the eigenvectors of $\sigma_{1},\sigma_{2}$
with $-1$ as eigenvalue are $\left( -1,0\right) ^{\intercal}$ and $\left(
2,-3\right) ^{\intercal}$ respectively; these two vectors are a basis for
$\mathbb{R}^{2}$. Similarly the eigenvectors of $\sigma_{1}$ and $\sigma_{2}$
with eigenvalue $+1$, namely $\left( 0,1\right) ^{\intercal}$, $\left(
2,1\right) ^{\intercal}$, form a basis. Form a direct sum
\[
\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) :=\left(
{\displaystyle\bigoplus\limits_{j\geq0}}
E_{j}^{\left( \operatorname*{triv}\right) }\right) \oplus\left(
{\displaystyle\bigoplus\limits_{j\geq0}}
E_{j}^{\left( \operatorname{sign}\right) }\right) \oplus\left(
{\displaystyle\bigoplus\limits_{j\geq0}}
E_{j}^{\left( \operatorname*{refl}\right) }\right) ,
\]
where the $E_{j}^{\left( \operatorname*{triv}\right) },E_{j}^{\left(
\operatorname{sign}\right) },E_{j}^{\left( \operatorname*{refl}\right) }$
are $\mathcal{S}_{3}$-irreducible and realizations of the representations
$\tau_{\operatorname*{triv}},\tau_{\operatorname{sign}},\tau
_{\operatorname*{refl}}$ respectively. Let $d_{\operatorname*{triv}}\left(
n\right) ,d_{\operatorname{sign}}\left( n\right) ,d_{\operatorname*{refl
}\left( n\right) $ denote the respective multiplicities, so that
$d_{\operatorname*{triv}}\left( n\right) +d_{\operatorname{sign}}\left(
n\right) +2d_{\operatorname*{refl}}\left( n\right) =n+1$. The case $n$ even
or odd are handled separately. If $n=2m$ is even then the number of
eigenvectors of $R$ having $-1$ as eigenvalue equals $m$ (the cardinality of
$\left\{ 1,3,5,\ldots,2m-1\right\} $). The same property holds for $M$ since
the eigenvectors of $M$ in the basis $\left\{ x^{2m}\left( x-1\right)
^{2m-j}\right\} $ are explicitly given by $\left\{ x^{2m-2\ell}\left(
x-1\right) ^{2\ell}-x^{2\ell}\left( x-1\right) ^{2m-2\ell}:0\leq\ell\leq
m\right\} $. Each $E_{j}^{\left( \operatorname*{refl}\right) }$ contains
one $\left( -1\right) $-eigenvector of $\chi_{\left\{ 1,0,2\right\} }$ and
one of $\chi_{\left\{ 0,2,1\right\} }$ and each $E_{j}^{\left(
\operatorname{sign}\right) }$ consists of one $\left( -1\right)
$-eigenvector of $\chi_{\left\{ 0,2,1\right\} }$. This gives the equation
$d_{\operatorname*{refl}}\left( n\right) +d_{\operatorname{sign}}\left(
n\right) =m$. Each $E_{j}^{\left( \operatorname*{refl}\right) }$ contains
one $\left( +1\right) $-eigenvector of $\chi_{\left\{ 1,0,2\right\} }$ and
one of $\chi_{\left\{ 0,2,1\right\} }$ and each $E_{j}^{\left(
\operatorname*{triv}\right) }$ consists of one $\left( +1\right)
$-eigenvector of $\chi_{\left\{ 0,2,1\right\} }$. There are $m+1$
eigenvectors with eigenvalue $1$ of each of $\chi_{\left\{ 1,0,2\right\} }$
and $\chi_{\left\{ 0,2,1\right\} }$ thus $d_{\operatorname*{refl}}\left(
n\right) +d_{\operatorname*{triv}}\left( n\right) =m+1$.
If $n=2m+1$ is odd then the eigenvector multiplicities are $m+1$ for both
eigenvalues $+1,-1$. By similar arguments we obtain the equations
$d_{\operatorname*{refl}}\left( n\right) +d_{\operatorname{sign}}\left(
n\right) =m+1$, $d_{\operatorname*{refl}}\left( n\right)
+d_{\operatorname*{triv}}\left( n\right) =m+1$. It remains to find one last
relation for both, even and odd cases.
To finish the determination of the multiplicities $d_{\operatorname*{triv
}\left( n\right) ,d_{\operatorname{sign}}\left( n\right)
,d_{\operatorname*{refl}}\left( n\right) $ it suffices to find
$d_{\operatorname*{triv}}\left( n\right) $. This is the dimension of the
space of polynomials in $\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right)
$ which are invariant under both $\chi_{\left\{ 0,2,1\right\} }$ and
$\chi_{\left\{ 1,0,2\right\} }.$ Since these two group elements generate
$\mathcal{S}_{3}$ this is equivalent to being invariant under each element of
$\mathcal{S}_{3}$ .This property is called \textit{totally symmetric}. Under
the action of $\gamma^{\operatorname{I}}$ this corresponds to the space of
polynomials in $\mathbb{P}_{n}\left( \left[ 0,1\right] \right) $ which are
invariant under both $R$ and $M$. We appeal to the classical theory of
symmetric polynomials: suppose $\mathcal{S}_{3}$ acts on polynomials in
$\left( y_{1},y_{2},y_{3}\right) $ by permutation of coordinates then the
space of symmetric (invariant under the group) polynomials is exactly the
space of polynomials in $\left\{ e_{1},e_{2},e_{3}\right\} $ the elementary
symmetric polynomials, namely $e_{1}=y_{1}+y_{2}+y_{3}$, $e_{2}=y_{1
y_{2}+y_{1}y_{3}+y_{2}y_{3}$, $e_{3}=y_{1}y_{2}y_{3}$. To apply this we set up
an affine map from $\widehat{T}$ to the triangle in $\mathbb{R}^{3}$ with
vertices $\left( 2,-1,-1\right) $, $\left( -1,2,-1\right) $, $\left(
-1,-1,2\right) $. The formula for the map i
\[
y\left( x\right) =\left( 2-3x_{1}-3x_{2},3x_{1}-1,3x_{2}-1\right) .
\]
The map takes $\left( 0,0\right) ,\left( 1,0\right) ,\left( 0,1\right) $
to the three vertices respectively. The result i
\begin{align*}
e_{1}\left( y\left( x\right) \right) & =0,\\
e_{2}\left( y\left( x\right) \right) & =-9\left( x_{1}^{2}+x_{1
x_{2}+x_{2}^{2}-x_{1}-x_{2}\right) -3,\\
e_{3}\left( y\left( x\right) \right) & =\left( 3x_{1}-1\right) \left(
3x_{2}-1\right) \left( 2-3x_{1}-3x_{2}\right) .
\end{align*}
Thus any totally symmetric polynomial on $\widehat{T}$ is a linear combination
of $e_{2}^{a}e_{3}^{b}$ with uniquely determined coefficients. The number of
linearly independent totally symmetric polynomials in $\left(
{\displaystyle\bigoplus\limits_{j=1}^{n}}
\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) \right) \oplus
\mathbb{P}_{0}\left( \widehat{T}\right) $ equals the number of solutions of
$0\leq2a+3b\leq n$ with $a,b=0,1,2,\ldots$. As a consequence
$d_{\operatorname*{triv}}\left( n\right) =\operatorname*{card}\left\{
\left( a,b\right) :2a+3b=n\right\} $. This number is the coefficient of
$t^{n}$ in the power series expansion of
\[
\frac{1}{\left( 1-t^{2}\right) \left( 1-t^{3}\right) }=\left(
1+t^{2}+t^{3}+t^{4}+t^{5}+t^{7}\right) \left( 1+2t^{6}+3t^{12
+\ldots\right) .
\]
From $d_{\operatorname*{triv}}\left( n\right) =\operatorname*{card}\left(
\left\{ 0,2,4,\ldots\right\} \cap\left\{ n,n-3,n-6,\ldots\right\} \right)
$ we deduce the formula (cf. (\ref{dtrivfirst})
\[
d_{\operatorname*{triv}}\left( n\right) =\left\lfloor \frac{n
{2}\right\rfloor -\left\lfloor \frac{n-1}{3}\right\rfloor .
\]
As a consequence: if $n=2m$ then $d_{\operatorname{sign}}\left( n\right)
=d_{\operatorname*{triv}}\left( n\right) -1$ and $d_{\operatorname*{refl
}\left( n\right) =m+1-d_{\operatorname*{triv}}\left( n\right) $; if
$n=2m+1$ then $d_{\operatorname{sign}}\left( n\right)
=d_{\operatorname*{triv}}\left( n\right) $ and $d_{\operatorname*{refl
}\left( n\right) =m+1-d_{\operatorname*{triv}}\left( n\right) $. From this
the following can be derived: $d_{\operatorname{sign}}\left( n\right)
=\left\lfloor \frac{n-1}{2}\right\rfloor -\left\lfloor \frac{n-1
{3}\right\rfloor $ and $d_{\operatorname*{refl}}\left( n\right)
=\left\lfloor \frac{n+2}{3}\right\rfloor $. Here is a table of values in terms
of $n\operatorname{mod}6$
\
\begin{vmatrix}
n & d_{\operatorname*{triv}}\left( n\right) & d_{\operatorname{sign}}\left(
n\right) & d_{\operatorname*{refl}}\left( n\right) \\
6m & m+1 & m & 2m\\
6m+1 & m & m & 2m+1\\
6m+2 & m+1 & m & 2m+1\\
6m+3 & m+1 & m+1 & 2m+1\\
6m+4 & m+1 & m & 2m+2\\
6m+5 & m+1 & m+1 & 2m+2
\end{vmatrix}
.
\]
\subsubsection{Construction of totally symmetric polynomials\label{ConSymmBas
}
Let $M$ and $R$ denote the linear maps $Mp\left( x_{1},x_{2}\right)
:=p\left( 1-x_{1}-x_{2},x_{2}\right) $ and $Rp\left( x_{1},x_{2}\right)
:=p\left( x_{2},x_{1}\right) $ respectively. Both are automorphisms of
$\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $. Note $Mp=p\circ
\chi_{\left\{ 1,0,2\right\} }$ and $Rp=p\circ\chi_{\left\{ 0,2,1\right\}
}$ (cf.\ Section \ref{SubSecIrr}).
\begin{proposition}
Suppose $0\leq k\leq n$ the
\begin{align}
Rb_{n,k} & =\left( -1\right) ^{k}b_{n,k};\label{Rbnk}\\
Mb_{n,k} & =\left( -1\right) ^{n}\sum_{j=0}^{n}~_{4}F_{3}\left(
\genfrac{}{}{0pt}{}{-j,j+1,-k,k+1}{-n,n+2,1
;1\right) \frac{2j+1}{n+1}b_{n,j}. \label{Mbnkeq
\end{align}
\end{proposition
\proof
The $_{4}F_{3}$-sum is understood to terminate at $k$ to avoid the $0/0$
ambiguities in the formal $_{4}F_{3}$-series. The first formula was shown in
Section \ref{SubSecIrr}. The second formula is a specialization of
transformations in \cite[Theorem 1.7(iii)]{Dunkl_3symm}: this paper used the
shifted Jacobi polynomial $R_{m}^{\left( \alpha,\beta\right) }\left(
s\right) =\frac{m!}{\left( \alpha+1\right) _{m}}P_{m}^{\left( \alpha
,\beta\right) }\left( 1-2s\right) $. Setting $\alpha=\beta=\gamma=0$ in the
formulas in \cite[Theorem 1.7(iii)]{Dunkl_3symm} results in $b_{n,k}=\left(
-1\right) ^{k}\dfrac{\theta_{n,k}}{k!\left( n-k\right) !}$ and
$Mb_{n,k}=\dfrac{\phi_{n,k}}{k!\left( n-k\right) !}$, where $\theta_{n,k}$,
$\phi_{n,k}$ are the polynomials introduced in \cite[p.690]{Dunkl_3symm}. More
precisely, the arguments $v_{1},v_{2},v_{3}$ in $\theta_{n,k}$ and $\phi
_{n,k}$ are specialized to $v_{1}=x_{1},v_{2}=x_{2}$ and $v_{3}=1-x_{1}-x_{2
$
\endproof
\begin{proposition}
The range of $I+RM+MR$ is exactly the subspace $\left\{ p\in\mathbb{P
_{n,n-1}^{\perp}\left( \widehat{T}\right) :RMp=p\right\} $.
\end{proposition}
\proof
By direct computation $\left( MR\right) ^{3}=I$ (cf. Section \ref{SubSecIrr
). This implies $\left( RM\right) ^{2}=MR$. If $p$ satisfies $RMp=p$ then
$Mp=Rp$ and $p=MRp$. Now suppose $RMp=p$ then $\left( I+RM+MR\right)
\frac{1}{3}p=p$; hence $p$ is in the range of $I+RM+MR$. Conversely suppose
$p=\left( I+RM+MR\right) p^{\prime}$ for some polynomial $p^{\prime}$, then,
$RM\left( I+RM+MR\right) p^{\prime}=\left( RM+\left( RM\right)
^{2}+I\right) p^{\prime}=p$
\endproof
Let $M_{i,j}^{\left( n\right) },R_{i,j}^{\left( n\right) }$ denote the
matrix entries of $M,R$ with respect to the basis $\left\{ b_{n,k}:0\leq
k\leq n\right\} $, respectively (that is $Mb_{n,k}=\sum_{j=0}^{n
b_{n,j}M_{j,k}^{\left( n\right) }$) . Let $S_{i,j}^{\left( n\right) }$
denote the matrix entries of $MR+RM+I$. The
\begin{align*}
R_{i,j}^{\left( n\right) } & =\left( -1\right) ^{i}\delta_{i,j
;~M_{i,j}^{\left( n\right) }=\left( -1\right) ^{n}~_{4}F_{3}\left(
\genfrac{}{}{0pt}{}{-i,i+1,-j,j+1}{-n,n+2,1
;1\right) \frac{2i+1}{n+1};\\
S_{i,j}^{\left( n\right) } & =\left( \left( -1\right) ^{j}+\left(
-1\right) ^{i}\right) M_{i,j}^{\left( n\right) }+\delta_{i,j}.
\end{align*}
Thus $S_{i,j}^{\left( n\right) }=2M_{i,j}^{\left( n\right) }+\delta_{i,j}$
if both $i,j$ are even, $S_{i,j}^{\left( n\right) }=-2M_{i,j}^{\left(
n\right) }+\delta_{i,j}$ if both $i,j$ are odd , and $S_{i,j}^{\left(
n\right) }=0$ if $i-j\equiv1\operatorname{mod}2$
\endproof
\begin{corollary}
For $0\leq k\leq\frac{n}{2}$ each polynomial $r_{n,2k}:=2\sum\limits_{0\leq
j\leq n/2}M_{2j,2k}^{\left( n\right) }b_{n,2j}+b_{n,2k}$ is totally
symmetric and for $0\leq k\leq\frac{n-1}{2}$ each polynomial $r_{n,2k+1
=-2\sum\limits_{0\leq j\leq\left( n-1\right) /2}M_{2j+1,2k+1}^{\left(
n\right) }b_{n,2j+1}+b_{n,2k+1}$ satisfies $Mp=-p=Rp$ (the sign representation).
\end{corollary}
\proof
The pattern of zeroes in $\left[ M_{i,j}^{\left( n\right) }\right] $ shows
that $r_{n,2k}=\left( MR+RM+I\right) b_{n,2k}\in\operatorname*{span}\left\{
b_{n,2j}\right\} $ and thus satisfies $Rr_{n,2k}=r_{n,2k}$; combined with
$RMr_{n,2k}=r_{n,2k}$ this shows $r_{n,2k}$ is totally symmetric. A similar
argument applies to $\left( MR+RM+I\right) b_{n,2k+1}$
\endproof
\begin{theorem}
\label{TheoCD14}The functions $b_{n,k}^{\operatorname*{sym}}$, $0\leq k\leq
d_{\operatorname*{triv}}\left( n\right) -1$, as in (\ref{defbpksym}) form a
basis for the totally symmetric polynomials in $\mathbb{P}_{n,n-1}^{\perp
}\left( \widehat{T}\right) $.
\end{theorem}
\proof
We use the homogeneous form of the $b_{n,m}$ as in \cite{Dunkl_3symm}, that
is, se
\begin{gather*}
b_{n,2m}^{\prime}\left( v\right) =\left( v_{1}+v_{2}+v_{3}\right)
^{n}b_{n,2m}\left( \frac{v_{1}}{v_{1}+v_{2}+v_{3}},\frac{v_{2}}{v_{1
+v_{2}+v_{3}}\right) \\
=\left( v_{1}+v_{2}+v_{3}\right) ^{n-2m}P_{n-2m}^{\left( 0,4m+1\right)
}\left( \frac{v_{1}+v_{2}-v_{3}}{v_{1}+v_{2}+v_{3}}\right) \left(
v_{1}+v_{2}\right) ^{2m}P_{2m}^{\left( 0,0\right) }\left( \frac
{v_{1}-v_{2}}{v_{1}+v_{2}}\right) .
\end{gather*}
Formally $b_{n,j}^{\prime}\left( v\right) =\left( -1\right) ^{j}\left(
j!\left( n-j\right) !\right) ^{-1}\theta_{n,j}\left( v\right) $ with
$\theta_{n,j}$ as in \cite[p.690]{Dunkl_3symm}. The expansion of such a
polynomial is a sum of monomials $v_{1}^{n_{1}}v_{2}^{n_{2}}v_{3}^{n_{3}}$
with $\sum_{i=1}^{3}n_{i}=n$. Symmetrizing the monomial results in the sum of
$v_{1}^{m_{1}}v_{2}^{m_{2}}v_{3}^{m_{3}}$ where $\left( m_{1},m_{2
,m_{3}\right) $ ranges over all permutations of $\left( n_{1},n_{2
,n_{3}\right) $. The argument is based on the occurrence of certain indices
in $b_{n,m}$. For a more straightforward approach to the coefficients we use
the following expansions (with $\ell=n-2k,\beta=2k+1$)
\begin{gather}
\left( v_{1}+v_{2}+v_{3}\right) ^{\ell}P_{\ell}^{\left( 0,\beta\right)
}\left( \frac{v_{1}+v_{2}-v_{3}}{v_{1}+v_{2}+v_{3}}\right) =\left(
-1\right) ^{\ell}\left( v_{1}+v_{2}+v_{3}\right) ^{\ell}P_{\ell}^{\left(
\beta,0\right) }\left( \frac{-v_{1}-v_{2}+v_{3}}{v_{1}+v_{2}+v_{3}}\right)
\label{Pn(0,beta)}\\
=\left( -1\right) ^{\ell}\frac{\left( \beta+1\right) _{\ell}}{\ell!
\sum_{i=0}^{\ell}\frac{\left( -\ell\right) _{i}\left( \ell+\beta+1\right)
_{i}}{i!\left( \beta+1\right) _{i}}\left( v_{1}+v_{2}\right) ^{i}\left(
v_{1}+v_{2}+v_{3}\right) ^{\ell-i};\nonumber
\end{gather}
an
\[
\left( v_{1}+v_{2}\right) ^{2k}P_{2k}^{\left( 0,0\right) }\left(
\frac{v_{1}-v_{2}}{v_{1}+v_{2}}\right) =\frac{1}{\left( 2k\right) !
\sum_{j=0}^{2k}\frac{\left( -2k\right) _{j}\left( -2k\right) _{j}\left(
-2k\right) _{2k-j}}{j!}v_{2}^{j}v_{1}^{2k-j}.
\]
First let $n=2m$. The highest power of $v_{3}$ that can occur in
$b_{2m,2m-2k}^{\prime}$ is $2k$, with corresponding coefficient $\frac{\left(
4m-4k+1\right) _{2k}}{\left( 2k\right) !}\sum_{j=0}^{2m-2k}c_{j}v_{2
^{j}v_{1}^{2m-j}$ for certain coefficients $\left\{ c_{j}\right\} $. Recall
that $d_{\operatorname*{triv}}\left( n\right) $ is the number of solutions
$\left( i,j\right) $ of the equation $3j+2i=2m$ (with $i,j=0,1,2,\ldots$).
The solutions can be listed as $\left( m,0\right) ,\left( m-3,2\right)
,\left( m-6,4\right) \ldots\left( m-3\ell,2\ell\right) $ where
$\ell=d_{\operatorname*{triv}}\left( n\right) -1$. By hypothesis $\left(
m-3k,2k\right) $ occurs in the list and thus $m-3k\geq0$ and $m-k\geq2k$.
There is only one possible permutation of $v_{1}^{m-k}v_{2}^{m-k}v_{3}^{2k}$
that occurs in $b_{2m,2m-2k}^{\prime}$ and its coefficient is $\frac{\left(
2k-2m\right) _{m-k}^{3}}{\left( 2m-2k\right) !}\neq0$. Hence there is a
triangular pattern for the occurrence of $v_{1}^{m}v_{2}^{m}$, $v_{1
^{m-1}v_{2}^{m-1}v_{3}^{2}$, $v_{1}^{m-2}v_{2}^{m-2}v_{3}^{4},\ldots$in the
symmetrizations of $b_{2m,2m}^{\prime}$, $b_{2m,2m-2}^{\prime}$, \ldots\ with
nonzero numbers on the diagonal and this proves the basis property when $n=2m$.
Now let $n=2m+1$. The highest power of $v_{3}$ that can occur in
$b_{2m+1,2m-2k}^{\prime}$ is $2k+1$, with coefficient $\frac{\left(
4m-4k+1\right) _{2k+1}}{\left( 2k+1\right) !}\sum_{j=0}^{2m-2k}c_{j
v_{2}^{j}v_{1}^{2m-j}$ for certain coefficients $\left\{ c_{j}\right\} $.
The solutions of $3j+2i=2m+1$ can be listed as $\left( m-1,1\right) ,\left(
m-4,3\right) ,\left( m-7,5\right) \ldots\left( m-1-3\ell,2\ell+1\right) $
where $\ell=d_{\operatorname*{triv}}\left( n\right) -1$. By hypothesis
$\left( m-1-3k,2k+1\right) $ occurs in this list, thus $m-k\geq2k+1$. There
is only one possible permutation of $v_{1}^{m-k}v_{2}^{m-k}v_{3}^{2k+1}$ that
occurs in $b_{2m+1,2m-2k}^{\prime}$ and its coefficient is $\frac{\left(
2k-2m\right) _{m-k}^{3}}{\left( 2m-2k\right) !}\neq0$. As above, there is a
triangular pattern for the occurrence of $v_{1}^{m}v_{2}^{m}v_{3}$,
$v_{1}^{m-1}v_{2}^{m-1}v_{3}^{3}$, $v_{1}^{m-2}v_{2}^{m-2}v_{3}^{5},\ldots$ in
the symmetrizations of $b_{2m+1,2m}^{\prime}$, $b_{2m+1,2m-2}^{\prime}$,
\ldots\ with nonzero numbers on the diagonal and this proves the basis
property when $n=2m+1$
\endproof
The totally symmetric orthogonal polynomials can be lifted to a general
triangle $T$.
\begin{definition}
\label{DefBasissymT}Let $T$ denote a triangle. The space of totally symmetric,
orthogonal polynomials of degree $n$ i
\begin{align}
\mathbb{P}_{n,n-1}^{\perp,\operatorname*{sym}}\left( T\right) & :=\left\{
u\in\mathbb{P}_{n,n-1}^{\perp}\left( T\right) :u\text{ has total
symmetry}\right\} \label{defbnkt}\\
& =\operatorname*{span}\left\{ b_{n,m}^{T,\operatorname*{sym}}:0\leq m\leq
d_{\operatorname*{triv}}\left( n\right) -1\right\} , \label{defBnmT
\end{align}
where the lifted symmetric basis functions are given by $b_{n,m
^{T,\operatorname*{sym}}:=b_{n,m}^{\operatorname*{sym}}\circ\chi_{T}^{-1}$ for
$b_{n,m}^{\operatorname*{sym}}$ as in Theorem \ref{TheoCD14} and an affine
pullback $\chi_{T}:\widehat{T}\rightarrow T$.
\end{definition}
\subsubsection{A Basis for the $\tau_{\operatorname*{refl}}$ component of
$\mathbb{P}_{n,n-1}^{\perp}\left( T\right) $\label{Sectaureflcomp}}
As explained in Section \ref{SubSecIrr} the space $\mathbb{P}_{n,n-1}^{\perp
}\left( \widehat{T}\right) $ can be decomposed into the $\tau
_{\operatorname*{triv}}$-, the $\tau_{\operatorname{sign}}$- and the
$\tau_{\operatorname*{refl}}$-component. A basis for the $\tau
_{\operatorname*{triv}}$ component are the fully symmetric basis functions
(cf. Section \ref{ConSymmBas}).
Next, we will construct a basis for all of $\mathbb{P}_{n,n-1}^{\perp}\left(
\widehat{T}\right) $ by extending the totally symmetric one. It is
straightforward to adjoin the $d_{\operatorname{sign}}\left( n\right) $
basis, using the same technique as for the fully symmetric ones: the monomials
which appear in $p$ with $Rp=-p=Mp$ must be permutations of $v_{1}^{n_{1
}v_{2}^{n_{2}}v_{3}^{n_{3}}$ with $n_{1}>n_{2}>n_{3}$. As in Theorem
\ref{TheoCD14} for $n=2m$ argue on monomials $v_{1}^{m-k}v_{2}^{m-1-k
v_{3}^{2k+1}$ and the polynomials $b_{2m,2m-2k-1}^{\prime}$ with $0\leq k\leq
d_{\operatorname{sign}}\left( n\right) -1=d_{\operatorname*{triv}}\left(
n\right) -2$, and for $n=2m+1$ use the monomials $v_{1}^{m+1-k}v_{2
^{m-k}v_{3}^{2k}$ and $b_{2m+1,2m-2k}$ with $0\leq k\leq
d_{\operatorname*{triv}}\left( n\right) -1=d_{\operatorname{sign}}\left(
n\right) -1.$
As we will see when constructing a basis for the non-conforming finite element
space, the $\tau_{\operatorname{sign}}$ component of $\mathbb{P
_{n,n-1}^{\perp}\left( \widehat{T}\right) $ is not relevant, in contrast to
the $\tau_{\operatorname*{refl}}$ component. In this section, we will
construct a basis for the $\tau_{\operatorname*{refl}}$ polynomials in
$\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right) $. Each such polynomial
is an eigenvector of $RM+MR$ with eigenvalue $-1$. We will show that the
polynomial
\begin{equation}
b_{n,k}^{\operatorname*{refl}}=\frac{1}{3}\left( 2I-RM-MR\right)
b_{n,2k},~0\leq k\leq\frac{n-1}{3}, \label{defqnk
\end{equation}
are linearly independent (and the same as introduced in (\ref{brefl1stdef}))
and, subsequently, that the se
\begin{equation}
\left\{ RMb_{n,k}^{\operatorname*{refl}},MRb_{n,k}^{\operatorname*{refl
}:0\leq k\leq\frac{n-1}{3}\right\} \label{basisreflrmmr
\end{equation}
is a basis for the $\tau_{\operatorname*{refl}}$ subspace of $\mathbb{P
_{n,n-1}^{\perp}\left( \widehat{T}\right) $. (The upper limit of $k$ is as
in (\ref{basisreflrmmr}) $d_{\operatorname*{refl}}\left( n\right) -1$ (cf.
(\ref{defdreflp})).) Note tha
\begin{equation
\begin{array}
[c]{c
RMb_{n,k}^{\operatorname*{refl}}=\frac{1}{3}\left( 2RM-MR-I\right)
b_{n,2k},\\
MRb_{n,k}^{\operatorname*{refl}}=\frac{1}{3}\left( 2MR-I-RM\right) b_{n,2k},
\end{array}
\label{defqnkplus
\end{equation}
because $\left( RM\right) ^{2}=MR$. Thus the calculation of these
polynomials follows directly from the formulae for $\left[ M_{ij}\right] $
and $\left[ R_{ij}\right] $. The method of proof relies on complex
coordinates for the triangle.
\begin{lemma}
\label{P00(v1v2)}For $k=0,1,2,\ldots
\begin{align*}
P_{2k}^{\left( 0,0\right) }\left( s\right) & =\left( -1\right)
^{k}\frac{\left( k+\frac{1}{2}\right) _{k}}{k!}\sum_{j=0}^{k}\frac{\left(
-k\right) _{j}^{2}}{j!\left( \frac{1}{2}-2k\right) _{j}}\left(
1-s^{2}\right) ^{k-j},\\
\left( v_{1}+v_{2}\right) ^{2k}P_{2k}^{\left( 0,0\right) }\left(
\frac{v_{1}-v_{2}}{v_{1}+v_{2}}\right) & =\left( -1\right) ^{k
\frac{\left( k+\frac{1}{2}\right) _{k}}{k!}\sum_{j=0}^{k}\frac{\left(
-k\right) _{j}^{2}}{j!\left( \frac{1}{2}-2k\right) _{j}}4^{k-j}\left(
v_{1}v_{2}\right) ^{k-j}\left( v_{1}+v_{2}\right) ^{2j}.
\end{align*}
\end{lemma
\proof
Start with the formula (specialized from a formula for Gegenbauer polynomials
\cite[18.5.10]{NIST:DLMF}
\[
P_{2k}^{\left( 0,0\right) }\left( s\right) =\left( 2s\right) ^{2k
\frac{\left( \frac{1}{2}\right) _{2k}}{\left( 2k\right) !}~_{2
F_{1}\left(
\genfrac{}{}{0pt}{}{-k,\frac{1}{2}-k}{\frac{1}{2}-2k
;\frac{1}{s^{2}}\right) .
\]
Apply the transformation (cf. \cite[15.8.1]{NIST:DLMF}
\[
_{2}F_{1}\left(
\genfrac{}{}{0pt}{}{-k,b}{c
;t\right) =\left( 1-t\right) ^{k}~_{2}F_{1}\left(
\genfrac{}{}{0pt}{}{-k,c-b}{c
;\frac{t}{t-1}\right)
\]
with $t=1/s^{2}$; then $\dfrac{t}{t-1}=\dfrac{1}{1-s^{2}}$ and $s^{2k}\left(
1-\frac{1}{s^{2}}\right) ^{k}=\left( -1\right) ^{k}\left( 1-s^{2}\right)
^{k}$. Also $2^{2k}\frac{\left( \frac{1}{2}\right) _{2k}}{\left( 2k\right)
!}=\frac{\left( \frac{1}{2}\right) _{2k}}{k!\left( \frac{1}{2}\right)
_{k}}=\frac{\left( k+\frac{1}{2}\right) _{k}}{k!}$. This proves the first
formula. Set $s=\dfrac{v_{1}-v_{2}}{v_{1}+v_{2}}$ then $1-s^{2}=\dfrac
{4v_{1}v_{2}}{\left( v_{1}+v_{2}\right) ^{2}}$ to obtain the second one
\endproof
Introduce complex homogeneous coordinates
\begin{align*}
z & =\omega v_{1}+\omega^{2}v_{2}+v_{3}\\
\overline{z} & =\omega^{2}v_{1}+\omega v_{2}+v_{3}\\
t & =v_{1}+v_{2}+v_{3}.
\end{align*}
Recall $\omega=e^{2\pi\mathrm{\operatorname*{i}}/3}=-\frac{1}{2
+\frac{\mathrm{\operatorname*{i}}}{2}\sqrt{3}$ and $\omega^{2}=\overline
{\omega}$. The inverse relations ar
\begin{align*}
v_{1} & =\frac{1}{3}\left( -\left( \omega+1\right) z+\omega\overline
{z}+t\right) \\
v_{2} & =\frac{1}{3}\left( \omega z-\left( \omega+1\right) \overline
{z}+t\right) \\
v_{3} & =\frac{1}{3}\left( z+\overline{z}+t\right) .
\end{align*}
Suppose $f\left( z,\overline{z},t\right) $ is a polynomial in $z$ and
$\bar{z}$ then $Rf\left( z,\overline{z},t\right) =f\left( \overline
{z},z,t\right) $ and $Mf\left( z,\overline{z},t\right) =f\left(
\omega\overline{z},\omega^{2}z,t\right) $. Thus $RMf\left( z,\overline
{z},t\right) =f\left( \omega^{2}z,\omega\overline{z},t\right) $ and
$MRf\left( z,\overline{z},t\right) =f\left( \omega z,\omega^{2}\overline
{z},t\right) $. The idea is to write $b_{n,2k}$ in terms of $z,\overline
{z},t$ and apply the projection $\Pi:=\frac{1}{3}\left( 2I-MR-RM\right) $.
To determine linear independence it suffices to consider the terms of highest
degree in $z,\overline{z}$ thus we set $t=v_{1}+v_{2}+v_{3}=0$ in the formula
for $b_{n,2k}$ (previously denoted $b_{n,2k}^{\prime}$ using the homogeneous
coordinates, see proof of Theorem \ref{TheoCD14}). From formula
(\ref{Pn(0,beta)}) and Lemma \ref{P00(v1v2)
\begin{align*}
b_{n,2k}^{\prime}\left( v_{1},v_{2},0\right) & =\left( n-2k+2\right)
_{n-2k}\left( v_{1}+v_{2}\right) ^{n-2k}\left( -1\right) ^{k}\frac{\left(
k+\frac{1}{2}\right) _{k}}{k!}\\
& \times\sum_{j=0}^{k}\frac{\left( -k\right) _{j}^{2}}{j!\left( \frac
{1}{2}-2k\right) _{j}}4^{k-j}\left( v_{1}v_{2}\right) ^{k-j}\left(
v_{1}+v_{2}\right) ^{2j}.
\end{align*}
The coefficient of $\left( v_{1}v_{2}\right) ^{k}\left( v_{1}+v_{2}\right)
^{n-2k}$ in $b_{n,2k}^{\prime}\left( v_{1},v_{2},0\right) $ is nonzero, and
this is the term with highest power of $v_{1}v_{2}$. Thus $\left\{
b_{n,2k}^{\prime}\left( v_{1},v_{2},0\right) :0\leq k\leq\frac{n-2
{3}\right\} $ is a basis for $\mathrm{\operatorname*{span}}\left\{ \left(
v_{1}v_{2}\right) ^{k}\left( v_{1}+v_{2}\right) ^{n-2k}:0\leq k\leq
\frac{n-2}{3}\right\} $. The next step is to show that the projection $\Pi$
has trivial kernel. In the complex coordinates $v_{1}+v_{2}=-\frac{1
{3}\left( z+z-t\right) =-\frac{1}{3}\left( z+z\right) $ and $v_{1
v_{2}=\frac{1}{9}\left( z^{2}-z\overline{z}+\overline{z}^{2}\right) $
(discarding terms of lower order in $z,\overline{z}$, that is, set $t=0$).
\begin{proposition}
\label{Prop17}If $\Pi\sum_{k=0}^{\left\lfloor \left( n-1\right)
/3\right\rfloor }c_{k}\left( z+\overline{z}\right) ^{n-2k}\left(
z^{2}-z\overline{z}+\overline{z}^{2}\right) ^{k}=0$ then $c_{k}=0$ for all
$k$.
\end{proposition}
\proof
For any polynomial $f\left( z,\overline{z}\right) $ we have $\Pi f\left(
z,\overline{z}\right) =\frac{1}{3}\left( 2f\left( z,\overline{z}\right)
-f\left( \omega^{2}z,\omega\overline{z}\right) -f\left( \omega z,\omega
^{2}\overline{z}\right) \right) $. In particula
\begin{gather*}
\Pi\left( z+\overline{z}\right) ^{n-2k}\left( z^{2}-z\overline{z
+\overline{z}^{2}\right) ^{k}=\Pi\left( z+\overline{z}\right)
^{n-3k}\left( z^{3}+\overline{z}^{3}\right) ^{k}\\
=\frac{1}{3}\left\{ 2\left( z+\overline{z}\right) ^{n-3k}-\left(
\omega^{2}z+\omega\overline{z}\right) ^{n-3k}-\left( \omega z+\omega
^{2}\overline{z}\right) ^{n-3k}\right\} \left( z^{3}+\overline{z
^{3}\right) ^{k}.
\end{gather*}
By hypothesis $n-3k\geq1$. Evaluate the expression at $z=e^{\pi
\mathrm{\operatorname*{i}/}6}+\varepsilon$ where $\varepsilon$ is real and
near $0$. Note $e^{\pi\mathrm{\operatorname*{i}/}6}=\frac{1}{2}\left(
\sqrt{3}+\mathrm{\operatorname*{i}}\right) $. The
\begin{align*}
z+\overline{z} & =\sqrt{3}+2\varepsilon,\\
\omega^{2}z+\omega\overline{z} & =-\varepsilon,\\
\omega z+\omega^{2}\overline{z} & =-\sqrt{3}-\varepsilon,\\
z^{3}+\overline{z}^{3} & =3\varepsilon+3\sqrt{3}\varepsilon^{2
+2\varepsilon^{3},
\end{align*}
an
\begin{align*}
& \frac{1}{3}\left\{ 2\left( z+\overline{z}\right) ^{n-3k}-\left(
\omega^{2}z+\omega\overline{z}\right) ^{n-3k}-\left( \omega z+\omega
^{2}\overline{z}\right) ^{n-3k}\right\} \left( z^{3}+\overline{z
^{3}\right) ^{k}\\
& =\frac{1}{3}\left\{ \left( 2-\left( -1\right) ^{n-3k}\right)
\times3^{\left( n-3k\right) /2}-\left( -\varepsilon\right) ^{n-3k
+C\varepsilon+O\left( \varepsilon^{2}\right) \right\} \varepsilon
^{k}\left( 3+3\sqrt{3}\varepsilon+2\varepsilon^{2}\right) ^{k},
\end{align*}
where $C=3^{\left( n--3k-1\right) /2}\left( n-3k\right) \left( 4-2\left(
-1\right) ^{n-3k}\right) $ (binomial theorem). The dominant term in the
right-hand side is $\left( 2-\left( -1\right) ^{n-3k}\right) 3^{\left(
n-k\right) /2-1}\varepsilon^{k}$. Now suppose $\Pi\sum\limits_{k=0
^{\left\lfloor \left( n-1\right) /3\right\rfloor }c_{k}\left(
z+\overline{z}\right) ^{n-2k}\left( z^{2}-z\overline{z}+\overline{z
^{2}\right) ^{k}=0$. Evaluate the polynomial at $z=e^{\pi
\mathrm{\operatorname*{i}/}6}+\varepsilon$. Let $\varepsilon\rightarrow0$
implying $c_{0}=0$. Indeed write the expression a
\[
\sum_{k=0}^{\left\lfloor \left( n-1\right) /3\right\rfloor }c_{k}\left(
2-\left( -1\right) ^{n-3k}\right) 3^{\left( n-k\right) /2-1
\varepsilon^{k}\left( 1+O\left( \varepsilon\right) \right) =0.
\]
Since $2-\left( -1\right) ^{n-3k}\geq1$ this shows $c_{k}=0$ for all $k$
\endproof
We have shown:
\begin{proposition}
Suppose $\Pi\sum\limits_{k=0}^{\left\lfloor \left( n-1\right)
/3\right\rfloor }c_{k}b_{n,2k}=0$ then $c_{k}=0$ for all $k$; the cardinality
of the set (\ref{basisreflrmmr}) is $d_{\operatorname*{refl}}\left( n\right)
$.
\end{proposition}
\begin{theorem}
\label{TheoCDBasis}\
\begin{enumerate}
\item[a.] The polynomials $\left\{ \Pi b_{n,2k}:0\leq k\leq\frac{n-1
{3}\right\} $ are linearly independent.
\item[b.] The set $\left\{ RM\Pi b_{n,2k},MR\Pi b_{n,2k}:0\leq k\leq
\frac{n-1}{3}\right\} $ is linearly independent and defines a basis for the
$\tau_{\operatorname*{refl}}$ component of $\mathbb{P}_{n,n-1}^{\perp}\left(
\widehat{T}\right) $.
\end{enumerate}
\end{theorem}
\proof
In general $\Pi z^{a}\overline{z}^{b}=z^{a}\overline{z}^{b}$ if $a-b\equiv
1,2\operatorname{mod}3$ and $\Pi z^{a}\overline{z}^{b}=0$ if $a-b\equiv
0\operatorname{mod}3$. Expand the polynomials $w_{k}\left( z,\overline
{z}\right) :=\Pi\left( z+\overline{z}\right) ^{n-3k}\left( z^{3
+\overline{z}^{3}\right) ^{k}$ by the binomial theorem to obtai
\[
\Pi\left( z+\overline{z}\right) ^{n-3k}\left( z^{3}+\overline{z
^{3}\right) ^{k}=\sum_{\substack{j=0\\n-2j\equiv1,2\operatorname{mod
3}}^{n-3k}\binom{n-3k}{j}z^{n-3k-j}\overline{z}^{j}\left( z^{3}+\overline
{z}^{3}\right) ^{k}.
\]
The
\begin{align*}
RMw_{k}\left( z,\overline{z}\right) & =\sum_{j=0,n-2j\equiv
1,2\operatorname{mod}3}^{n-3k}\binom{n-3k}{j}\omega^{2j-n}z^{n-3k-j
\overline{z}^{j}\left( z^{3}+\overline{z}^{3}\right) ^{k},\\
MRw_{k}\left( z,\overline{z}\right) & =\sum_{j=0,n-2j\equiv
1,2\operatorname{mod}3}^{n-3k}\binom{n-3k}{j}\omega^{n-2j}z^{n-3k-j
\overline{z}^{j}\left( z^{3}+\overline{z}^{3}\right) ^{k}.
\end{align*}
Firstly we show that $\left\{ RMw_{k},MRw_{k}\right\} $ is linearly
independent for $0\leq k\leq\frac{n-1}{3}$. For each value of
$n\operatorname{mod}3$ we select the highest degree terms from $RMw_{k}$ and
$MRw_{k}$: (i) $n=3m+1$, $\omega^{2}z^{3m+1}+\omega\overline{z}^{3m+1}$ and
$\omega z^{3m+1}+\omega^{2}\overline{z}^{3m+1}$, (ii) $n=3m+2,$ $\omega
z^{3m+2}+\omega^{2}\overline{z}^{3m+2}$ and $\omega^{2}z^{3m+2}+\omega
\overline{z}^{3m+2}$, (iii) $n=3m$, $\left( n-3k\right) \left( \omega
^{2}z^{3m}\overline{z}+\omega z\overline{z}^{3m}\right) $ and $\left(
n-3k\right) \left( \omega z^{3m}\overline{z}+\omega^{2}z\overline{z
^{3m}\right) $ (by hypothesis $n-3k\geq1$). In each case the two terms are
linearly independent (the determinant of the coefficients is $\pm\left(
\omega-\omega^{2}\right) =\mp\mathrm{\operatorname*{i}}\sqrt{3}$). Secondly
the same argument as in the previous theorem shows that $\sum_{k=0
^{\left\lfloor \left( n-1\right) /3\right\rfloor }\left\{ c_{k
RMw_{k}+d_{k}MRw_{k}\right\} =0$ implies $c_{k}RMw_{k}+d_{k}MRw_{k}=0$ for
all $k$. By the first part it follows that $c_{k}=0=d_{k}$. This completes the
proof
\endproof
\begin{remark}
The basis $b_{n,k}$ for $\mathbb{P}_{n,n-1}^{\perp}\left( \widehat{T}\right)
$ in (\ref{defbnk}) is mirror symmetric with respect to the angular bisector
in $\widehat{T}$ through the origin for even $k$ and is mirror skew-symmetric
for odd $k$. This fact makes the point $\mathbf{0}$ in $\widehat{T}$ special
compared to the other vertices. As a consequence the functions defined in
Theorem \ref{TheoCDBasis}.a reflects the special role of $\mathbf{0}$. Part b
shows that it is possible to define a basis with functions which are either
symmetric with respect to the angle bisector in $\widehat{T}$ through $\left(
1,0\right) ^{\intercal}$ or through $\left( 0,1\right) ^{\intercal}$ by
\textquotedblleft rotating\textquotedblright\ the functions $\Pi b_{n,2k}$ to
these vertices
\[
RM\left( \Pi b_{n,2k}\right) \left( x_{1},x_{2}\right) =\left( \Pi
b_{n,2k}\right) \left( x_{2},1-x_{1}-x_{2}\right) \quad\text{and\quad
}MR\left( \Pi b_{n,2k}\right) \left( x_{1},x_{2}\right) =\left( \Pi
b_{n,2k}\right) \left( 1-x_{1}-x_{2},x_{1}\right) .
\]
Since the dimension of $E^{\left( \operatorname*{refl}\right) }$ is
$2d_{\operatorname*{refl}}\left( n\right) =2\left\lfloor \frac{n+2
{3}\right\rfloor $ is not (always) a multiple of $3$, it is, in general, not
possible to define a basis where all three vertices of the triangle are
treated in a symmetric way.
\end{remark}
\begin{definition}
\label{RemMirror}Le
\begin{equation}
\mathbb{P}_{n,n-1}^{\perp,\operatorname*{refl}}\left( \widehat{T}\right)
:=\operatorname*{span}\left\{ RM\Pi b_{n,2k},MR\Pi b_{n,2k}:0\leq k\leq
\frac{n-1}{3}\right\} . \label{defPnn-1refl
\end{equation}
This space is lifted to a general triangle $T$ by fixing a vertex $\mathbf{P}$
of $T$ and settin
\begin{equation}
\mathbb{P}_{n,n-1}^{\perp,\operatorname*{refl}}\left( T\right) :=\left\{
u\circ\chi_{\mathbf{P},T}^{-1}:u\in\mathbb{P}_{n,n-1}^{\perp
,\operatorname*{refl}}\left( \widehat{T}\right) \right\} ,
\label{defReflspace
\end{equation}
where the lifting $\chi_{\mathbf{P},T}$ is an affine pullback $\chi
_{\mathbf{P},T}:\widehat{T}\rightarrow T$ which maps $\mathbf{0}$ to
$\mathbf{P}$.
The basis $b_{n,k}^{\operatorname*{refl}}$ to describe the restrictions of
facet-oriented, non-conforming finite element functions to the facets is
related to a reduced space and defined as in (\ref{defqnk}) with lifted
version
\begin{equation}
b_{n,k}^{\mathbf{P},T}:=b_{n,k}^{\operatorname*{refl}}\circ\chi_{\mathbf{P
,T}^{-1},\quad0\leq k\leq\frac{n-1}{3}. \label{defbmitttriangle
\end{equation}
\end{definition}
\begin{remark}
The construction of the spaces $\mathbb{P}_{p,p-1}^{\perp,\operatorname*{sym
}\left( T\right) \ $and $\mathbb{P}_{p,p-1}^{\perp,\operatorname*{refl
}\left( T\right) $ (cf. Definitions \ref{DefBasissymT} and \ref{RemMirror})
implies the direct sum decompositio
\begin{equation}
\operatorname*{span}\left\{ b_{p,2k}\circ\chi_{\mathbf{P},T}^{-1}:0\leq
k\leq\left\lfloor p/2\right\rfloor \right\} =\mathbb{P}_{p,p-1
^{\perp,\operatorname*{sym}}\left( T\right) \oplus\mathbb{P}_{p,p-1
^{\perp,\operatorname*{refl}}\left( T\right) . \label{defmirrorP
\end{equation}
It is easy to verify that the basis functions $b_{p,k}^{\mathbf{P},T}$ are
mirror symmetric with respect to the angle bisector in $T$ through
$\mathbf{P}$. However, the space $\mathbb{P}_{n,n-1}^{\perp
,\operatorname*{refl}}\left( T\right) $ is independent of the choice of the
vertex $\mathbf{P}$.
In Appendix \ref{AA} we will define further sets of basis functions for the
$\tau_{\operatorname*{refl}}$ component of $\mathbb{P}_{n,n-1}^{\perp}\left(
\widehat{T}\right) $ -- different choices might be preferable for different
kinds of applications.
\end{remark}
\subsection{Simplex-Supported and Facet-Oriented Non-Conforming Basis
Functions}
In this section, we will define non-conforming Crouzeix-Raviart type functions
which are supported either on one single tetrahedron or on two tetrahedrons
which share a common facet. As a prerequisite, we study in
\S \ref{SecOrthPolyTriStar} piecewise orthogonal polynomials on triangle
stars, i.e., on a collection of triangles which share a common vertex and
cover a neighborhood of this vertex (see Notation \ref{Notstar}). We will
derive conditions such that these functions are continuous across common edges
and determine the dimension of the resulting space. This allows us to
determine the non-conforming Courzeix-Raviart basis functions which are either
supported on a single tetrahedron (see \S \ref{SecBasSKnc}) or on two adjacent
tetrahedrons (see \S \ref{SecBasSTnc}) by \textquotedblleft
closing\textquotedblright\ triangle stars either by a single triangle or
another triangle star.
\subsubsection{Orthogonal Polynomials on Triangle
Stars\label{SecOrthPolyTriStar}}
The construction of the functions $B_{p,k}^{K,\operatorname*{nc}}$ and
$B_{p,k}^{T,\operatorname*{nc}}$ as in (\ref{DefSkpsym}) and
(\ref{defedgesupp}) requires some results of continuous, piecewise orthogonal
polynomials on \textit{triangle stars} which we provide in this section.
\begin{notation}
\label{Notstar}A subset $C\subset\Omega$ is a \emph{triangle star} if $C$ is
the union of some, say $m_{C}\geq3$, triangles $T\in\mathcal{F}_{C
\subset\mathcal{F}$, i.e., $C
{\displaystyle\bigcup\limits_{T\in\mathcal{F}_{C}}}
T$ and there exists some vertex $\mathbf{V}_{C}\in\mathcal{V}$ such tha
\begin{equation
\begin{array}
[c]{ll
\mathbf{V}_{C}\text{ is a vertex of }T & \forall T\in\mathcal{F}_{C},\\
\exists\text{ a continuous, piecewise affine mapping }\chi:D_{m_{C
}\rightarrow C & \text{such that }\chi\left( 0\right) =\mathbf{V}_{C}.
\end{array}
\label{deftrianglestar
\end{equation}
Here, $D_{k}$ denotes the regular closed $k$-gon (in $\mathbb{R}^{2}$).
\end{notation}
For a triangle star $C$, we defin
\[
\mathbb{P}_{p,p-1}^{\perp}\left( C\right) :=\left\{ u\in C^{0}\left(
C\right) \mid\forall T\in\mathcal{F}_{C}:\left. u\right\vert _{T
\in\mathbb{P}_{p,p-1}^{\perp}\left( T\right) \right\} .
\]
In the next step, we will explicitly characterize the space $\mathbb{P
_{p,p-1}^{\perp}\left( C\right) $ by defining a set of basis functions. Set
$\mathbf{A}:=\mathbf{V}_{C}$ (cf. (\ref{deftrianglestar})) and pick an outer
vertex in $\mathcal{F}_{C}$, denote it by $\mathbf{A}_{1}$, and number the
remaining vertices $\mathbf{A}_{2},\ldots,\mathbf{A}_{m_{C}}$ in
$\mathcal{F}_{C}$ counterclockwise. We use the cyclic numbering convention
$\mathbf{A}_{m_{C}+1}:=\mathbf{A}_{1}$ and also for similar quantities.
For $1\leq\ell\leq m_{C}$, let $\mathbf{e}_{\ell}:=\left[ \mathbf{A
,\mathbf{A}_{\ell}\right] $ be the straight line (convex hull) between and
including $\mathbf{A}$, $\mathbf{A}_{\ell}$. Let $T_{\ell}\in\mathcal{F}_{C}$
be the triangle with vertices $\mathbf{A}$, $\mathbf{A}_{\ell}$,
$\mathbf{A}_{\ell+1}$. Then we choose the affine pullbacks to the reference
element $\widehat{T}$ b
\[
\chi_{\ell}\left( x_{1},x_{2}\right) :=\left\{
\begin{array}
[c]{ll
\mathbf{A}+x_{1}\left( \mathbf{A}_{\ell}-\mathbf{A}\right) +x_{2}\left(
\mathbf{A}_{\ell+1}-\mathbf{A}\right) & \text{if }\ell\text{ is odd,}\\
\mathbf{A}+x_{1}\left( \mathbf{A}_{\ell+1}-\mathbf{A}\right) +x_{2}\left(
\mathbf{A}_{\ell}-\mathbf{A}\right) & \text{if }\ell\text{ is even.
\end{array}
\right.
\]
In this way, the common edges $\mathbf{e}_{\ell}$ are parametrized by
$\chi_{\ell-1}\left( t,0\right) =\chi_{\ell}\left( t,0\right) $ if
$3\leq\ell\leq m_{C}$ is odd and by $\chi_{\ell-1}\left( 0,t\right)
=\chi_{\ell}\left( 0,t\right) $ if $2\leq\ell\leq m_{C}$ is even. The final
edge $\mathbf{e}_{1}$ is parametrized by $\chi_{1}\left( t,0\right)
=\chi_{m_{C}}\left( t,0\right) $ if $m_{C}$ is even and by $\chi_{1}\left(
t,0\right) =\chi_{m_{C}}\left( 0,t\right) $ (with interchanged arguments!)
otherwise. We introduce the se
\[
R_{p,C}:=\left\{
\begin{array}
[c]{ll
\left\{ 0,\ldots,p\right\} & \text{if }m_{C}\text{ is even,}\\
\left\{ 2\ell:0\leq\ell\leq\left\lfloor \frac{p}{2}\right\rfloor \right\} &
\text{if }m_{C}\text{ is odd
\end{array}
\right.
\]
and define the functions (cf. (\ref{defbnkt}), (\ref{defReflspace}),
(\ref{defmirrorP})
\begin{equation}
\left. b_{p,k}^{C}\right\vert _{T_{\ell}}:=b_{p,k}\circ\chi_{\ell
^{-1},\qquad\forall k\in R_{p,C}. \label{basispatch
\end{equation}
\begin{lemma}
\label{Lemcont}For a triangle star $C$, a basis for $\mathbb{P}_{p,p-1
^{\perp}\left( C\right) $ is given by $b_{p,k}^{C}$, $k\in R_{p,C}$. Furthe
\begin{equation}
\dim\mathbb{P}_{p,p-1}^{\perp}\left( C\right) =\left\{
\begin{array}
[c]{cc
p+1 & \text{if }m_{C}\text{ is even,}\\
\left\lfloor \frac{p}{2}\right\rfloor +1 & \text{if }m_{C}\text{ is odd.
\end{array}
\right. \label{dimformula
\end{equation}
\end{lemma
\proof
We show that $\left( b_{p,k}^{C}\right) _{k\in R_{p,C}}$ is a basis of
$\mathbb{P}_{p,p-1}^{\perp}\left( C\right) $ and the dimension formula.
\textbf{Continuity across }$\mathbf{e}_{\ell}$ \textbf{for odd }$3\leq\ell\leq
m_{C}$.
The definition of the lifted orthogonal polynomials (see (\ref{defbnkt}),
(\ref{defReflspace}), (\ref{defmirrorP})) implies that the continuity across
$\mathbf{e}_{\ell}$ for odd $3\leq\ell\leq m_{C}$ is equivalent t
\[
\sum_{k=0}^{p}\alpha_{p,k}^{\left( \ell-1\right) }b_{p,k}^{\operatorname{I
}=\sum_{k=0}^{p}\alpha_{p,k}^{\left( \ell\right) }b_{p,k}^{\operatorname{I
}.
\]
From Lemma \ref{LemRestLinIndep} we conclude that the continuity across such
edges is equivalent to
\begin{equation}
\alpha_{p,k}^{\left( \ell-1\right) }=\alpha_{p,k}^{\left( \ell\right)
}\qquad\forall0\leq k\leq p.
\end{equation}
\textbf{Continuity across }$\mathbf{e}_{\ell}$ \textbf{for even }$2\leq
\ell\leq m_{C}$.
Note that $\chi_{2}\left( 0,t\right) =\chi_{3}\left( 0,t\right) $. Taking
into account (\ref{defbnkt}), (\ref{defReflspace}), (\ref{defmirrorP}) we see
that the continuity across $\mathbf{e}_{\ell}$ is equivalent t
\[
\sum_{k=0}^{p}\alpha_{p,k}^{\left( 2\right) }b_{p,k}^{\operatorname{II
}=\sum_{k=0}^{p}\alpha_{p,k}^{\left( 3\right) }b_{p,k}^{\operatorname{II}}.
\]
From Lemma \ref{LemRestLinIndep} we conclude that the continuity across
$\mathbf{e}_{\ell}$ for even $2\leq\ell\leq m_{C}$ is again equivalent to
\begin{equation}
\alpha_{p,k}^{\left( \ell-1\right) }=\alpha_{p,k}^{\left( \ell\right)
}\qquad\forall0\leq k\leq p.
\end{equation}
\textbf{Continuity across }$\mathbf{e}_{1}$
For even $m_{C}$ the previous argument also applies for the edge
$\mathbf{e}_{1}$ and the functions $b_{p,k}^{C}$, $0\leq k\leq p$, are
continuous across $\mathbf{e}_{1}$. For odd $m_{C}$, note that $\chi
_{1}\left( t,0\right) =\chi_{m_{C}}\left( 0,t\right) $. Taking into
account (\ref{defbnkt}), (\ref{defReflspace}), (\ref{defmirrorP}) we see that
the continuity across $\mathbf{e}_{1}$ is equivalent to
\[
\sum_{k=0}^{p}\alpha_{p,k}^{\left( 1\right) }b_{p,k}^{\operatorname{I}
=\sum_{k=0}^{p}\alpha_{p,k}^{\left( m_{C}\right) }b_{p,k}^{\operatorname{II
}.
\]
Using the symmetry relation (\ref{symmbnk}) we conclude that this is
equivalent t
\[
\sum_{k=0}^{p}\alpha_{p,k}^{\left( 1\right) }b_{p,k}^{\operatorname{I}
=\sum_{k=0}^{p}\alpha_{p,k}^{\left( m_{C}\right) }\left( -1\right)
^{k}b_{p,k}^{\operatorname{I}}.
\]
From Lemma \ref{LemRestLinIndep} we conclude that this, in turn, is equivalent
t
\begin{equation
\begin{array}
[c]{cl
\alpha_{p,k}^{\left( 1\right) }=\alpha_{p,k}^{\left( m_{C}\right) } &
k\text{ is even,}\\
\alpha_{p,k}^{\left( 1\right) }=-\alpha_{p,k}^{\left( m_{C}\right) } &
k\text{ is odd.
\end{array}
\end{equation}
From the above reasoning, the continuity of $b_{p,k}^{C}$ across
$\mathbf{e}_{1}$ follows if $\alpha_{n,k}^{\left( \ell\right) }=0$ for odd
$k$ and all $1\leq\ell\leq m_{C}$.
The proof of the dimension formula (\ref{dimformula}) is trivial
\endproof
\subsubsection{A Basis for the Symmetric Non-Conforming Space
$S_{K,\operatorname*{nc}}^{p}$\label{SecBasSKnc}}
In this section, we will prove that $S_{K,\operatorname*{nc}}^{p}$ (cf.
(\ref{DefSkpsym})) satisfie
\[
S_{K,\operatorname*{nc}}^{p}\oplus S_{K,\operatorname*{c}}^{p}=S_{K
^{p}:=\left\{ u\in S_{\mathcal{G}}^{p}:\operatorname*{supp}u\subset
K\right\} ,
\]
where $S_{\mathcal{G}}^{p}$ is defined in (\ref{hpfinele}) and, moreover, that
the functions $B_{p,k}^{K,\operatorname*{nc}}$, $k=0,1,\ldots
,d_{\operatorname*{triv}}\left( p\right) -1$, as in (\ref{DefBpkKhut}),
(\ref{DefSkpsym}) form a basis of $S_{K,\operatorname*{nc}}^{p}$.
Let $T$ denote one facet of $K$ and let $C:=\partial K\backslash\overset
{\circ}{T}$. Since $C$ is a triangle star with $m_{C}=3$, we can apply Lemma
\ref{Lemcont} to obtain that
\[
\left. S_{K}^{p}\right\vert _{C}:=\left\{ \left. u\right\vert _{C}:u\in
S_{K}^{p}\right\} \subset\operatorname*{span}\left\{ b_{p,2k}^{C}:0\leq
k\leq\left\lfloor \frac{p}{2}\right\rfloor \right\} .
\]
The continuity of $b_{p,2k}^{C}$ implies that the restriction $b_{p,2k
^{\partial T}:=\left. b_{p,2k}^{C}\right\vert _{\partial T}$ is continuous.
From (\ref{bnkIII}) we conclude that
\begin{equation}
\left. b_{p,2k}^{\partial T}\right\vert _{E}=P_{2k}^{E}\qquad\forall
E\subset\partial T, \label{defbn2kdT
\end{equation}
where $P_{2k}^{E}$ is the Legendre polynomial of even degree $2k$ scaled to
the edge $E$ with endpoint values $+1$ and symmetry with respect to the
midpoint of $E$. Hence, we are looking for orthogonal polynomials
$\mathbb{P}_{p,p-1}^{\perp}\left( T\right) $ whose traces on $\partial T$
are linear combination of $b_{p,2k}^{\partial T}$, $0\leq k\leq\left\lfloor
\frac{p}{2}\right\rfloor $. From (\ref{symmbnk}) we deduce that they have
total symmetry, i.e., belong to the space $\mathbb{P}_{p,p-1}^{\perp
,\operatorname*{sym}}\left( T\right) $ (cf. Definition \ref{DefBasissymT}).
For $0\leq m\leq d_{\operatorname*{triv}}\left( p\right) -1$, let
$b_{p,m}^{\partial K,\operatorname*{sym}}:\partial K\rightarrow\mathbb{R}$ be
defined facet-wise for any $T\subset\partial K$ b
\begin{equation}
\left. b_{p,m}^{\partial K,\operatorname*{sym}}\right\vert _{T
:=b_{p,m}^{T,\operatorname*{sym}}\qquad0\leq m\leq d_{\operatorname*{triv
}\left( p\right) -1. \label{defsurfacesym
\end{equation}
Finally, we extend the function $b_{p,m}^{\partial K,\operatorname*{sym}}$ to
the total simplex $K$ by polynomial extension (cf. (\ref{DefBpkKhut}),
(\ref{18b})
\begin{equation}
B_{p,m}^{K,\operatorname*{nc}}=\sum_{\mathbf{N}\in\mathcal{N}^{p}\cap\partial
K}b_{p,m}^{\partial K,\operatorname*{sym}}\left( \mathbf{N}\right) \left.
B_{p,\mathbf{N}}^{\mathcal{G}}\right\vert _{K}\qquad0\leq m\leq
d_{\operatorname*{triv}}\left( p\right) -1. \label{defsimplsupp
\end{equation}
These functions are the same as those introduced in Definition
\ref{DefSymSpaceK}. The above reasoning leads to the following Proposition.
\begin{proposition}
For a simplex $K$, the space of non-conforming, simplex-supported
Crouzeix-Raviart finite elements can be chosen as in (\ref{DefSkpsym}) and the
functions $B_{p,k}^{K,\operatorname*{nc}}$, $0\leq k\leq
d_{\operatorname*{triv}}\left( p\right) -1$ are linearly independent.
\end{proposition}
\subsubsection{A Basis for $S_{T,\operatorname*{nc}}^{p}$\label{SecBasSTnc}}
Let $T\in\mathcal{F}_{\Omega}$ be an inner facet and $K_{1},K_{2
\in\mathcal{G}$ such that $T=K_{1}\cap K_{2}$ and $\omega_{T}=K_{1}\cup K_{2}$
(cf. (\ref{defoftrianglesubsets})) with the convention that the unit normal
$\mathbf{n}_{T}$ points into $K_{2}$. In this section, we will prove that a
space $\tilde{S}_{T,\operatorname*{nc}}^{p}$ which satisfie
\begin{equation}
\tilde{S}_{T,\operatorname*{nc}}^{p}\oplus\left(
{\displaystyle\bigoplus\limits_{i=1}^{2}}
S_{K_{i},\operatorname*{nc}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{i=1}^{2}}
S_{K_{i},\operatorname*{c}}^{p}\right) \oplus S_{T,\operatorname*{c}
^{p}=S_{T}^{p}:=\left\{ u\in S_{\mathcal{G}}^{p}:\operatorname*{supp
u\subset\omega_{T}\right\} \label{directsplitfacetspace
\end{equation}
can be chosen as $\tilde{S}_{T,\operatorname*{nc}}^{p
:=S_{T,\operatorname*{nc}}^{p}$ (cf. (\ref{STpncdef})) and, moreover, that the
functions $B_{p,k}^{T,\operatorname*{nc}}$, $k=0,1,\ldots
,d_{\operatorname*{refl}}\left( p\right) -1$, as in (\ref{defedgesupp}) form
a basis of $S_{T,\operatorname*{nc}}^{p}$.
Let $C_{i}:=\left( \partial K_{i}\right) \backslash\overset{\circ}{T}$,
$i=1,2$, denote the triangle star (cf. Notation \ref{Notstar}) formed by the
three remaining triangles of $\partial K_{i}$. We conclude from Lemma
\ref{Lemcont} that a basis for $\mathbb{P}_{p,p-1}^{\perp}\left(
C_{i}\right) $ is given by $b_{p,2k}^{C_{i}}$, $0\leq k\leq\left\lfloor
\frac{p}{2}\right\rfloor $ (cf. (\ref{basispatch})). Any function $u$ in
$S_{T}^{p}$ satisfie
\begin{equation
\begin{array}
[c]{cl
\gamma_{K_{i}}u\in\mathbb{P}_{p}\left( K_{i}\right) & i=1,2,\\
\left. \left( \gamma_{K_{i}}u\right) \right\vert _{T^{\prime}}\in
\mathbb{P}_{p,p-1}^{\perp}\left( T^{\prime}\right) & \forall T^{\prime
}\subset C_{i},\quad i=1,2,\\
\left[ u\right] _{T}\in\mathbb{P}_{p,p-1}^{\perp}\left( T\right) . &
\end{array}
\label{jumpfacet
\end{equation}
Since any function in $S_{T}^{p}$ is continuous on $C_{i}$, we conclude from
Lemma \ref{Lemcont} (with $m_{C_{i}}=3$) tha
\begin{equation}
\left. u\right\vert _{C_{i}}\in\mathbb{P}_{p,p-1}^{\perp}\left(
C_{i}\right) \text{\quad and\quad}\left. \gamma_{K_{i}}u\right\vert
_{\partial T}\in\operatorname*{span}\left\{ b_{p,2k}^{\partial T}:0\leq
k\leq\left\lfloor \frac{p}{2}\right\rfloor \right\} \qquad i=1,2
\label{twosidedlinearcombis
\end{equation}
with $b_{p,2k}^{\partial T}$ as in (\ref{defbn2kdT}).
To identify a space $\tilde{S}_{T,\operatorname*{nc}}^{p}$ which satisfies
(\ref{directsplitfacetspace}) we consider the jump condition in
(\ref{jumpfacet}) restricted to the boundary $\partial T$. The symmetry of the
functions $b_{p,2k}^{\partial T}$ implies that $\left[ u\right] _{T
\in\mathbb{P}_{p,p-1}^{\perp,\operatorname*{sym}}\left( T\right) $, i.e.,
there is a function $q_{1}\in S_{K_{1},\operatorname*{nc}}^{p}$(see
(\ref{DefSkpsym})) such that $\left[ u\right] _{T}=\left. q_{1}\right\vert
_{T}$ and $\tilde{u}$, defined by $\left. \tilde{u}\right\vert _{K_{1}
=u_{1}+q_{1}$ and $\left. \tilde{u}\right\vert _{K_{2}}=u_{2}$, is continuous
across $T$. On the other hand, all functions $u\in S_{T}^{p}$ whose
restrictions $\left. u\right\vert _{\omega_{T}}$ are discontinuous can be
found in $S_{K_{1},\operatorname*{nc}}^{p}\oplus S_{K_{2},\operatorname*{nc
}^{p}$. In view of the direct sum in (\ref{directsplitfacetspace}) we may thus
assume that the functions in $\tilde{S}_{T,\operatorname*{nc}}^{p}$ are
continuous in $\omega_{T}$.
To finally arrive at a direct decomposition of the space in the right-hand
side of (\ref{directsplitfacetspace}) we have to split the spaces
$\mathbb{P}_{p,p-1}^{\perp}\left( C_{i}\right) $ into a direct sum of the
spaces of totally symmetric orthogonal polynomials and the spaces introduced
in Definition \ref{RemMirror} and glue them together in a continuous way. We
introduce the functions $b_{p,k}^{C_{i},\operatorname*{sym}}:=\left.
b_{p,k}^{\partial K_{i},\operatorname*{sym}}\right\vert _{C_{i}}$, $0\leq
k\leq d_{\operatorname*{triv}}\left( p\right) -1$, with $b_{p,k}^{\partial
K_{i},\operatorname*{sym}}$ as in (\ref{defsurfacesym}) and define
$b_{p,k}^{C_{i},\operatorname*{refl}}$, $0\leq k\leq d_{\operatorname*{refl
}\left( p\right) -1$, piecewise by $\left. b_{p,k}^{C_{i
,\operatorname*{refl}}\right\vert _{T^{\prime}}:=b_{p,k}^{\mathbf{A
_{i},T^{\prime}}$ for $T^{\prime}\subset C_{i}$ with $b_{p,k}^{\mathbf{A
_{i},T^{\prime}}$ as in (\ref{defbmitttriangle}). The mirror symmetry of
$b_{p,k}^{\mathbf{A}_{i},T^{\prime}}$ with respect to the angular bisector in
$T^{\prime}$ through $\mathbf{A}_{i}$ implies the continuity of $b_{p,k
^{C_{i},\operatorname*{refl}}$. Hence
\begin{equation}
\mathbb{P}_{p,p-1}^{\perp}\left( C_{i}\right) =\operatorname*{span}\left\{
\left. b_{p,k}^{C_{i},\operatorname*{sym}}\right\vert _{C_{i}}:0\leq k\leq
d_{\operatorname*{triv}}\left( p\right) -1\right\} \oplus
\operatorname*{span}\left\{ b_{p,k}^{C_{i},\operatorname*{refl}}:0\leq k\leq
d_{\operatorname*{refl}}\left( p\right) -1\right\} . \label{paux
\end{equation}
Since the traces of $b_{p,k}^{C_{i},\operatorname*{sym}}$ and $b_{p,k
^{C_{i},\operatorname*{refl}}$ at $\partial T$ are continuous and are, from
both sides, the same linear combinations of edge-wise Legendre polynomials of
even degree, the gluing $\left. b_{p,k}^{\partial\omega_{T
,\operatorname*{sym}}\right\vert _{C_{i}}:=b_{p,k}^{C_{i},\operatorname*{sym
}$ and $\left. b_{p,k}^{\partial\omega_{T},\operatorname*{refl}}\right\vert
_{\dot{C}_{i}}:=b_{p,k}^{C_{i},\operatorname*{refl}}$, $i=1,2$, defines
continuous functions on $\partial\omega_{T}$. Since the space
$S_{T,\operatorname*{nc}}^{p}$ must satisfy a direct sum decomposition (cf.
(\ref{directsplitfacetspace})), it suffices to consider the functions
$b_{p,k}^{\partial\omega_{T},\operatorname*{refl}}$ for the definition of
$S_{T,\operatorname*{nc}}^{p}$. The resulting non-conforming facet-oriented
space $S_{T,\operatorname*{nc}}^{p}$ was introduced in Definition
\ref{PropBpmTncbasis} and $\tilde{S}_{T,\operatorname*{nc}}^{p}$ can be chosen
to be $S_{T,\operatorname*{nc}}^{p}$.
\begin{proposition}
\label{CornonorthoT}For any $u\in S_{T,\operatorname*{nc}}^{p}$, the following
implication hold
\[
\left. u\right\vert _{T}\in\left. S_{T,\operatorname*{nc}}^{p}\right\vert
_{T}\cap\mathbb{P}_{p,p-1}^{\perp}\left( T\right) \implies u=0.
\]
\end{proposition
\proof
Assume there exists $u\in S_{T,\operatorname*{nc}}^{p}$ with $\left.
u\right\vert _{T}\in\left. S_{T,\operatorname*{nc}}^{p}\right\vert _{T
\cap\mathbb{P}_{p,p-1}^{\perp}\left( T\right) $. Let $K$ be a simplex
adjacent to $T$. Then $u_{K}=\left. u\right\vert _{K}$ satisfies $\left.
u_{K}\right\vert _{T^{\prime}}\in\mathbb{P}_{p,p-1}^{\perp}\left( T^{\prime
}\right) $ for all $T^{\prime}\subset\partial K$ and, thus, $u_{K}\in
S_{K,\operatorname*{nc}}^{p}$. Since $\left. S_{K,\operatorname*{nc}
^{p}\right\vert _{T^{\prime}}\cap\left. S_{T,\operatorname*{nc}
^{p}\right\vert _{T^{\prime}}=\left\{ 0\right\} $ for $T^{\prime}\in\partial
K\backslash\overset{\circ}{T}$ we conclude that $u_{K}=0$
\endproof
Note that Definition \ref{PropBpmTncbasis} and Proposition \ref{CornonorthoT}
neither imply a priori that the functions
\[
\left. B_{p,k}^{T,\operatorname*{nc}}\right\vert _{K},\quad\forall
T\subset\partial K,\quad k=0,\ldots,d_{\operatorname*{refl}}\left( p\right)
-1
\]
are linearly independent nor tha
\begin{equation}
\forall T\subset\partial K\quad\text{it holds\quad}\sum_{T^{\prime}\subset
C}\left. B_{p,m}^{T^{\prime},\operatorname*{nc}}\right\vert _{T
=\mathbb{P}_{p,p-1}^{\perp,\operatorname*{refl}}\left( T\right)
\quad\text{for the triangle star }C=\partial K\backslash\overset{\circ}{T}
\label{reflspacecontained
\end{equation}
holds. These properties will be proved next. Recall the projection $\Pi
=\frac{1}{3}\left( 2I-MR-RM\right) $ from Proposition \ref{Prop17}. We
showed (Theorem \ref{TheoCDBasis}.a) that$\left\{ b_{p,k
^{\operatorname*{refl}}:0\leq k\leq\frac{p-1}{3}\right\} $ is linearly
independent, where $b_{p,k}^{\operatorname*{refl}}:=\Pi b_{p,2k}$.
Additionally $Rb_{p,k}^{\operatorname*{refl}}=b_{p,k}^{\operatorname*{refl}}$
which implies $b_{p,k}^{\operatorname*{refl}}\left( 0,x_{1}\right)
=b_{p,k}^{\operatorname*{refl}}\left( x_{1},0\right) $, and the restriction
$x_{1}\longmapsto b_{p,k}^{\operatorname*{refl}}\left( x_{1},1-x_{1}\right)
$ is invariant under $x_{1}\mapsto1-x_{1}$. For four non-coplanar points
$A_{0},A_{1},A_{2},A_{3}$ let $K$ denote the tetrahedron with these vertices.
For any $k$ such that $0\leq k\leq\frac{p-1}{3}$ define a piecewise polynomial
on the faces of $K$ as follows: choose a local $\left( x_{1},x_{2}\right)
$-coordinate system for $A_{0}A_{1}A_{2}$ so that the respective coordinates
are $\left( 0,0\right) ,\left( 1,0\right) ,\left( 0,1\right) $, and
define $Q_{k}^{\left( 0\right) }$ on the facet equal to $b_{p,k
^{\operatorname*{refl}}$. Similarly define $Q_{k}^{\left( 0\right) }$ on
$A_{0}A_{2}A_{3}$ and $A_{0}A_{3}A_{1}$ (with analogously chosen local
$\left( x_{1},x_{2}\right) $-coordinate systems), by the property
$b_{p,k}^{\operatorname*{refl}}\left( 0,x_{1}\right) =b_{p,k
^{\operatorname*{refl}}\left( x_{1},0\right) $. $Q_{k}^{\left( 0\right) }$
is continuous at the edges $A_{0}A_{1}$, $A_{0}A_{2}$, and $A_{0}A_{3}$. The
values at the boundary of the triangle star equal $b_{p,k
^{\operatorname*{refl}}\left( x_{1},1-x_{1}\right) $; note the symmetry and
thus the orientation of the coordinates on the edges $A_{1}A_{2}$, $A_{2
A_{3}$, $A_{3}A_{1}$ is immaterial. The value of $Q_{k}^{\left( 0\right) }$
on the triangle $A_{1}A_{2}A_{3}$ is taken to be a degree $p$ polynomial,
totally symmetric, with values agreeing with $b_{p,k}^{\operatorname*{refl
}\left( x_{1},1-x_{1}\right) $ on each edge.
Similarly $Q_{k}^{\left( 1\right) },Q_{k}^{\left( 2\right)
,Q_{k}^{\left( 3\right) }$ are defined by taking $A_{1},A_{2},A_{3}$ as the
center of the construction, respectively.
\begin{theorem}
\label{THeoQkilinindep}a) The functions $Q_{k}^{\left( i\right) }$, $0\leq
k\leq d_{\operatorname*{refl}}\left( p\right) -1$, $i=0,1,2,3$ are linearly independent.
b) Property (\ref{reflspacecontained}) holds.
\end{theorem}
The proof involves a series of steps. The argument will depend on the values
of the functions on the three rays $A_{0}A_{1}$, $A_{0}A_{2}$, $A_{0}A_{3}$,
each one of them is given coordinates $t$ so that $t=0$ at $A_{0}$ and $t=1$
at the other end-point. For a fixed $k$ let $q\left( t\right) =b_{p,k
^{\operatorname*{refl}}\left( t,0\right) $, $\widehat{q}\left( t\right)
=b_{p,k}^{\operatorname*{refl}}\left( 1-t,0\right) $ and $\widetilde
{q}\left( t\right) =b_{p,k}^{\operatorname*{refl}}\left( t,1-t\right) $.
\begin{lemma}
\label{Lemsump=0}Suppose $0\leq k\leq\frac{p-1}{3}$ and $0\leq t\leq1$ then
$q\left( t\right) +\widehat{q}\left( t\right) +\widetilde{q}\left(
t\right) =0$.
\end{lemma}
\proof
The actions of $RM$ and $MR$ on polynomials $f\left( x_{1},x_{2}\right) $
are given by $MRf\left( x_{1},x_{2}\right) =f\left( 1-x_{1}-x_{2
,x_{1}\right) $ and $RMf\left( x_{1},x_{2}\right) =f\left( x_{2
,1-x_{1}-x_{2}\right) $. Polynomials of $\tau_{\operatorname*{refl}}$-type
satisfy $f+RMf+MRf=0$. Apply this relation to $b_{p,k}^{\operatorname*{refl}}$
with $x_{1}=t$ and $x_{2}=0$ with the resul
\[
b_{p,k}^{\operatorname*{refl}}\left( t,0\right) +b_{p,k
^{\operatorname*{refl}}\left( 1-t,t\right) +b_{p,k}^{\operatorname*{refl
}\left( 0,1-t\right) =0.
\]
The fact that $b_{p,k}^{\operatorname*{refl}}\left( x_{1},x_{2}\right)
=b_{p,k}^{\operatorname*{refl}}\left( x_{2},x_{1}\right) $ finishes the
proof
\endproof
\textbf{Proof of Theorem \ref{THeoQkilinindep}. }Consider the contribution of
$Q_{k}^{\left( 1\right) }$ to the values on the ray $A_{0}A_{1}$: because
$Q_{k}^{\left( 1\right) }$ is constructed taking the origin at $A_{1}$ and
because of the reverse orientation of the ray we see that the value of
$Q_{k}^{\left( 1\right) }$ is given by $\widehat{q}$. The value of
$Q_{k}^{\left( 1\right) }$ on the ray $A_{0}A_{2}$ is $\widetilde{q}$ (by
the symmetry of $\widetilde{q}$ the orientation of the ray does not matter).
The other functions are handled similarly, and the contributions to the three
rays are given in this table
\
\begin{vmatrix}
& Q_{k}^{\left( 0\right) } & Q_{k}^{\left( 1\right) } & Q_{k}^{\left(
2\right) } & Q_{k}^{\left( 3\right) }\\
A_{0}A_{1} & q & \widehat{q} & \widetilde{q} & \widetilde{q}\\
A_{0}A_{2} & q & \widetilde{q} & \widehat{q} & \widetilde{q}\\
A_{0}A_{3} & q & \widetilde{q} & \widetilde{q} & \widehat{q
\end{vmatrix}
.
\]
We use $q_{k},\widetilde{q}_{k},\widehat{q}_{k}$ to denote the polynomials
corresponding to $b_{p,k}^{\operatorname*{refl}}$. Suppose that the linear
combination $\sum_{k=0}^{\left\lfloor \left( p-1\right) /3\right\rfloor
}\sum_{i=0}^{3}c_{k,i}Q_{k}^{\left( i\right) }=0$. Evaluate the sum on the
three rays to obtain the equations
\begin{align*}
0 & =\sum_{k=0}^{\left\lfloor \left( p-1\right) /3\right\rfloor }\left\{
c_{k,0}q_{k}+c_{k,1}\widehat{q}_{k}+\left( c_{k,2}+c_{k,3}\right)
\widetilde{q}_{k}\right\} =\sum_{k=0}^{\left\lfloor \left( p-1\right)
/3\right\rfloor }\left\{ \left( c_{k,1}-c_{k,0}\right) \widehat{q
_{k}+\left( c_{k,2}+c_{k,3}-c_{k,0}\right) \widetilde{q}_{k}\right\} \\
0 & =\sum_{k=0}^{\left\lfloor \left( p-1\right) /3\right\rfloor }\left\{
c_{k,0}q_{k}+c_{k,2}\widehat{q}_{k}+\left( c_{k,1}+c_{k,3}\right)
\widetilde{q}_{k}\right\} =\sum_{k=0}^{\left\lfloor \left( p-1\right)
/3\right\rfloor }\left\{ \left( c_{k,2}-c_{k,0}\right) \widehat{q
_{k}+\left( c_{k,1}+c_{k,3}-c_{k,0}\right) \widetilde{q}_{k}\right\} ,\\
0 & =\sum_{k=0}^{\left\lfloor \left( p-1\right) /3\right\rfloor }\left\{
c_{k,0}q_{k}+c_{k,3}\widehat{q}_{k}+\left( c_{k,1}+c_{k,2}\right)
\widetilde{q}_{k}\right\} =\sum_{k=0}^{\left\lfloor \left( p-1\right)
/3\right\rfloor }\left\{ \left( c_{k,3}-c_{k,0}\right) \widehat{q
_{k}+\left( c_{k,1}+c_{k,2}-c_{k,0}\right) \widetilde{q}_{k}\right\} .
\end{align*}
We used Lemma \ref{Lemsump=0} to eliminate $q_{k}$ from the equations. In
Theorem \ref{TheoCDBasis}.b we showed the linear independence of $\left\{
RMb_{p,k}^{\operatorname*{refl}},MRb_{p,k}^{\operatorname*{refl}}:0\leq
k\leq\frac{p-1}{3}\right\} $, and in Lemma \ref{LemRestLinIndep} that the
restriction map $f\mapsto f\left( x_{1},0\right) $ is an isomorphism from
the orthogonal polynomials $\mathbb{P}_{p,p-1}^{\bot}$ to $\mathbb{P
_{p}\left( \left[ 0,1\right] \right) $. Thus the projection of the set is
also linearly independent, that is, $\left\{ \widetilde{q}_{k},\widehat
{q}_{k}:0\leq k\leq\frac{p-1}{3}\right\} $ is a linearly independent set of
polynomials on $0\leq t\leq1$. This implies all the coefficients in the above
equations vanish: the $\widehat{q}_{k}$ terms show $c_{k,0}=c_{k,1
=c_{k,2}=c_{k,3}$ and then the $\widetilde{q}_{k}$-terms show $2c_{k,0
-c_{k,0}=c_{k,0}=0$.
To prove (\ref{reflspacecontained}) it suffices to transfer the statement to
the reference element $\widehat{T}$. The pullbacks of the restrictions
$\left. B_{p,m}^{T^{\prime},\operatorname*{nc}}\right\vert _{T}$, $T^{\prime
}\subset C$, are given b
\begin{equation}
b_{n,k}^{\operatorname*{refl}}=\Pi b_{n,2k}\text{,}\quad\widetilde{b
_{n,k}^{\operatorname*{refl}}:=RM\Pi b_{n,2k}\text{,\quad}\widehat{b
_{n,k}^{\operatorname*{refl}}:=MR\Pi b_{n,2k}\text{, }k=0,\ldots
d_{\operatorname*{refl}}\left( n\right) -1\text{.} \label{bbb
\end{equation}
Since $b_{n,k}^{\operatorname*{refl}}\in\mathbb{P}_{n,n-1}^{\perp
,\operatorname*{refl}}\left( \widehat{T}\right) $ (cf. (\ref{T1+T2})) it
follow
\[
\mathbb{P}_{n,n-1}^{\perp,\operatorname*{refl}}\left( \widehat{T}\right)
\overset{\text{(\ref{defPnn-1refl})}}{=}\operatorname*{span}\left\{
\widetilde{b}_{n,k}^{\operatorname*{refl}},\widehat{b}_{n,k
^{\operatorname*{refl}}:0\leq k\leq d_{\operatorname*{refl}}\left( n\right)
-1\right\} =\operatorname*{span}\left\{ b_{n,k}^{\operatorname*{refl
},\widetilde{b}_{n,k}^{\operatorname*{refl}},\widehat{b}_{n,k
^{\operatorname*{refl}}:0\leq k\leq d_{\operatorname*{refl}}\left( n\right)
-1\right\} .
\
\endproof
\section{Properties of Non-Conforming Crouzeix-Raviart Finite
Elements\label{SecPropNC}}
The non-conforming Crouzeix-Raviart finite element space $S_{\mathcal{G
,\operatorname*{nc}}^{p}$ satisfies $S_{\mathcal{G},\operatorname*{c}
^{p}\subsetneq S_{\mathcal{G},\operatorname*{nc}}^{p}\subset S_{\mathcal{G
}^{p}$ (cf. Section \ref{SecNCFECR}). In this section, we will present a basis
for $S_{\mathcal{G},\operatorname*{nc}}^{p}$ and discuss whether the inclusion
$S_{\mathcal{G},\operatorname*{nc}}^{p}\subset S_{\mathcal{G}}^{p}$, in fact,
is an equality.
\subsection{A Basis for Non-Conforming Crouzeix-Raviart Finite Elements}
We have defined conforming and non-conforming sets of functions which are
spanned by functions with local support. In this section, we will investigate
the linear independence of these functions. We introduce the following space
\[
S_{\operatorname*{sym},\operatorname*{nc}}^{p}:
{\displaystyle\bigoplus\limits_{K\in\mathcal{G}}}
S_{K,\operatorname*{nc}}^{p},\quad S_{\operatorname*{refl},\operatorname*{nc
}^{p}:
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
S_{T,\operatorname*{nc}}^{p},
\]
where $S_{K,\operatorname*{nc}}^{p}$ and $S_{T,\operatorname*{nc}}^{p}$ are as
in Definitions \ref{DefSymSpaceK} and \ref{PropBpmTncbasis}. For some $0\leq
k\leq d_{\operatorname*{refl}}\left( p\right) -1$, we introduce the subspace
$S_{\operatorname*{refl},\operatorname*{nc}}^{p,k}\subset
S_{\operatorname*{refl},\operatorname*{nc}}^{p}$ b
\[
S_{\operatorname*{refl},\operatorname*{nc}}^{p,k}:
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
\left\{ B_{p,m}^{T,\operatorname*{nc}}:0\leq m\leq k\right\} .
\]
Further we will need the conforming finite element space (cf. (\ref{hpfinele
), Def. \ref{Deflocbasis}), where the vertex-oriented functions are removed,
i.e.
\[
\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}:=\left(
{\displaystyle\bigoplus\limits_{E\in\mathcal{E}_{\Omega}}}
S_{E,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{T\in\mathcal{F}_{\Omega}}}
S_{T,\operatorname*{c}}^{p}\right) \oplus\left(
{\displaystyle\bigoplus\limits_{K\in\mathcal{G}}}
S_{K,\operatorname*{c}}^{p}\right) .
\]
\begin{theorem}
\label{Theorem33}The sum
\begin{equation}
S_{\mathcal{G},\operatorname*{c}}^{p}\oplus S_{\operatorname*{sym
,\operatorname*{nc}}^{p},\qquad S_{\operatorname*{sym},\operatorname*{nc}
^{p}\oplus S_{\operatorname*{refl},\operatorname*{nc}}^{p} \label{directsums
\end{equation}
are direct. The su
\begin{equation}
S_{\mathcal{G},\operatorname*{c}}^{p}+S_{\operatorname*{refl
,\operatorname*{nc}}^{p} \label{nondirectsum
\end{equation}
is \textbf{not} direct. The su
\begin{equation}
\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}\oplus S_{\operatorname*{sym
,\operatorname*{nc}}^{p}\oplus S_{\operatorname*{refl},\operatorname*{nc
}^{p,0} \label{finalspace
\end{equation}
is direct.
\end{theorem}
\proof
\textbf{Part 1.} We prove that the sum $S_{\operatorname*{sym
,\operatorname*{nc}}^{p}\oplus S_{\operatorname*{refl},\operatorname*{nc}
^{p}$ is direct.
From Proposition \ref{CornonorthoT} we know that the sum $\left.
S_{T,\operatorname*{nc}}^{p}\right\vert _{T}\oplus\mathbb{P}_{p,p-1}^{\perp
}\left( T\right) $ is direct. Let $\Pi_{T}:L^{2}\left( T\right)
\rightarrow\mathbb{P}_{p-1}\left( T\right) $ denote the $L^{2}\left(
T\right) $ orthogonal projection. Since $\mathbb{P}_{p-1}\left( T\right) $
is the orthogonal complement of $\mathbb{P}_{p,p-1}^{\perp}\left( T\right) $
in $\mathbb{P}_{p}\left( T\right) $ and since $\mathbb{P}_{p,p-1}^{\perp
}\left( T\right) \cap\left. S_{T,\operatorname*{nc}}^{p}\right\vert
_{T}=\left\{ 0\right\} $, the restricted mapping $\Pi_{T}:\left.
S_{T,\operatorname*{nc}}^{p}\right\vert _{T}\rightarrow\mathbb{P}_{p-1}\left(
T\right) $ is injective and the functions $q_{p,k}^{T}:=\Pi_{T}\left(
\left. B_{p,k}^{T,\operatorname*{nc}}\right\vert _{T}\right) $, $0\leq k\leq
d_{\operatorname*{refl}}\left( p\right) -1$, are linearly independent and
belong to $\mathbb{P}_{p-1}\left( T\right) $. We define the functional
\[
J_{p,k}^{T}\left( w\right) :=\int_{T}wq_{p,k}^{T}\quad0\leq k\leq
d_{\operatorname*{refl}}\left( p\right) -1\text{.
\]
Next we consider a general linear combination and show that the conditio
\begin{equation}
\sum_{K\subset\mathcal{G}}\sum_{i=0}^{d_{\operatorname*{triv}}\left(
p\right) -1}\alpha_{i}^{K}B_{p,i}^{K,\operatorname*{nc}}+\sum_{K\subset
\mathcal{G}}\sum_{T^{\prime}\subset\partial K}\sum_{j=0
^{d_{\operatorname*{refl}}\left( p\right) -1}\beta_{j}^{T^{\prime}
B_{p,j}^{T^{\prime},\operatorname*{nc}}\overset{!}{=}0 \label{ureplinindep1
\end{equation}
implies that all coefficients are zero.
We apply the functionals $J_{p,k}^{T}$ to (\ref{ureplinindep1}) and use the
orthogonality between $\mathbb{P}_{p,p-1}^{\perp}\left( T\right) $ and
$q_{p,k}^{T}$ to obtai
\begin{equation}
\sum_{K\subset\mathcal{G}}\sum_{T^{\prime}\subset\partial K}\sum
_{j=0}^{d_{\operatorname*{refl}}\left( p\right) -1}\beta_{j}^{T^{\prime
}J_{p,k}^{T}\left( B_{p,j}^{T^{\prime},\operatorname*{nc}}\right)
\overset{!}{=}0. \label{ureplinindep2
\end{equation}
For $T^{\prime}\neq T$ it holds $J_{p,k}^{T}\left( B_{p,i}^{T^{\prime
},\operatorname*{nc}}\right) =0$ since $\left. \left. B_{p,i}^{T^{\prime
},\operatorname*{nc}}\right\vert _{K}\right\vert _{T}$ is an orthogonal
polynomial. Thus, equation (\ref{ureplinindep2}) is equivalent t
\begin{equation}
\sum_{j=0}^{d_{\operatorname*{refl}}\left( p\right) -1}\beta_{j}^{T
J_{p,k}^{T}\left( B_{p,j}^{T,\operatorname*{nc}}\right) \overset{!}{=}0.
\label{ureplinindep3
\end{equation}
The matrix $\left( J_{p,k}^{T}\left( B_{p,j}^{T,\operatorname*{nc}}\right)
\right) _{k,j=0}^{d_{\operatorname*{refl}}\left( p\right) -1}$ is regular
becaus
\[
J_{p,k}^{T}\left( B_{p,j}^{T,\operatorname*{nc}}\right) =\int_{T
B_{p,j}^{T,\operatorname*{nc}}q_{p,k}^{T}=\int_{T}B_{p,j
^{T,\operatorname*{nc}}\Pi_{T}\left( \left. B_{p,k}^{T,\operatorname*{nc
}\right\vert _{T}\right) =\int_{T}B_{p,j}^{T,\operatorname*{nc}
B_{p,k}^{T,\operatorname*{nc}
\]
and $\left( \left. B_{p,k}^{T,\operatorname*{nc}}\right\vert _{T}\right)
_{k}$ are linearly independent. Hence we conclude from (\ref{ureplinindep3})
that all coefficients $\beta_{j}^{T}$ are zero and the condition
(\ref{ureplinindep1}) reduces t
\[
\sum_{K\subset\mathcal{G}}\sum_{i=0}^{d_{\operatorname*{triv}}\left(
p\right) -1}\alpha_{i}^{K}B_{p,i}^{K,\operatorname*{nc}}\overset{!}{=}0.
\]
The left-hand side is a piecewise continuous function so that the condition is
equivalent to $\sum_{i=0}^{d_{\operatorname*{triv}}\left( p\right) -1
\alpha_{i}^{K}B_{p,i}^{K,\operatorname*{nc}}\overset{!}{=}0$ for all
$K\in\mathcal{G}$. Since $B_{p,i}^{K,\operatorname*{nc}}$ is a basis for
$\left. S_{K,\operatorname*{nc}}^{p}\right\vert _{K}$ we conclude that all
$\alpha_{i}^{K}$ are zero.
\textbf{Part 2. }Next we prove that $\left( S_{\operatorname*{sym
,\operatorname*{nc}}^{p}\oplus S_{\operatorname*{refl},\operatorname*{nc
}^{p,0}\right) \cap\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}=\left\{
0\right\} $ and we show this by contradiction. Let $u\in\left(
S_{\operatorname*{sym},\operatorname*{nc}}^{p}\oplus S_{\operatorname*{refl
,\operatorname*{nc}}^{p,0}\right) \cap\tilde{S}_{\mathcal{G
,\operatorname*{c}}^{p}$ which satisfies $u\neq0$. We decompose
$u=u_{\operatorname*{sym}}+u_{\operatorname*{refl}}$ with
$u_{\operatorname*{sym}}\in S_{\operatorname*{sym},\operatorname*{nc}}^{p}$
and $u_{\operatorname*{refl}}\in S_{\operatorname*{refl},\operatorname*{nc
}^{p}$. We prove by contradiction that $u_{\operatorname*{sym}}\in
C^{0}\left( \Omega\right) $. Assume that $u_{\operatorname*{sym}}\notin
C^{0}\left( \Omega\right) $. Then, there exists a facet $T\subset
\mathcal{F}_{\Omega}$ such that $\left[ u_{\operatorname*{sym}}\right]
_{T}\neq0$. Then, $\left[ u_{\operatorname*{refl}}\right] _{T}=-\left[
u_{\operatorname*{sym}}\right] _{T}$ is a necessary condition for the
continuity of $u$. However, $\left[ u_{\operatorname*{sym}}\right] _{T
\in\mathbb{P}_{p,p-1}^{\perp,\operatorname*{sym}}\left( T\right) $ while
$\left[ u_{\operatorname*{refl}}\right] _{T}\in\mathbb{P}_{p,p-1
^{\perp,\operatorname*{refl}}\left( T\right) $ and there is a contradiction
because $\mathbb{P}_{p,p-1}^{\perp,\operatorname*{sym}}\left( T\right)
\cap\mathbb{P}_{p,p-1}^{\perp,\operatorname*{refl}}\left( T\right) =\left\{
0\right\} $. Hence, $u_{\operatorname*{sym}}\in C^{0}\left( \Omega\right) $
and, in turn, $u_{\operatorname*{refl}}\in C^{0}\left( \Omega\right) $.
\medskip
Since $u\neq0$, at least, one of the functions $u_{\operatorname*{sym}}$ and
$u_{\operatorname*{refl}}$ must be different from the zero function.
\textbf{Case a. }We show $u_{\operatorname*{sym}}=0$ by contradiction: Assume
$u_{\operatorname*{sym}}\neq0$. Then, $\left. u_{\operatorname*{sym
}\right\vert _{T}\neq0$ for all facets $T\in\mathcal{F}$. (Proof by
contradiction: If $\left. u_{\operatorname*{sym}}\right\vert _{T}=0$ for some
$T\in\mathcal{F}$, we pick some $K\in\mathcal{F}$ which has $T$ as a facet.
Since $\left. u_{\operatorname*{sym}}\right\vert _{K}\in\left.
S_{K,\operatorname*{nc}}^{p}\right\vert _{K}$ we have $\left.
u_{\operatorname*{sym}}\right\vert _{T^{\prime}}=0$ for all facets $T^{\prime
}$ of $K$ and $\left. u_{\operatorname*{sym}}\right\vert _{K}=0$. Since
$u_{\operatorname*{sym}}$ is continuous in $\Omega$, the restriction $\left.
u_{\operatorname*{sym}}\right\vert _{K^{\prime}}$ is zero for any $K^{\prime
}\in\mathcal{G}$ which shares a facet with $K$. This argument can be applied
inductively to show that $u_{\operatorname*{sym}}=0$ in $\Omega$. This is a
contradiction.) We pick a boundary facet $T\in\mathcal{F}_{\partial\Omega}$.
The condition $u\in\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}$ implies
$u=0$ on $\partial\Omega$ and, in particular, $\left. u\right\vert
_{T}=\left. u_{\operatorname*{sym}}\right\vert _{T}+\left.
u_{\operatorname*{refl}}\right\vert _{T}=0$. We use again the argument
$\mathbb{P}_{p,p-1}^{\perp,\operatorname*{sym}}\left( T\right)
\cap\mathbb{P}_{p,p-1}^{\perp,\operatorname*{refl}}\left( T\right) =\left\{
0\right\} $ which implies $u_{\operatorname*{sym}}=0$ and this is a
contradiction to the assumption $u_{\operatorname*{sym}}\neq0$.
\textbf{Case b. }From Case a we know that $u_{\operatorname*{sym}}=0$, i.e.,
$u_{\operatorname*{refl}}=u$, and it remains to show $u_{\operatorname*{refl
}=0$. The condition $u_{\operatorname*{refl}}\in\tilde{S}_{\mathcal{G
,\operatorname*{c}}^{p}$ implies $\left. u_{\operatorname*{refl}}\right\vert
_{\partial\Omega}=0$ and $u_{\operatorname*{refl}}\left( \mathbf{V}\right)
=0$ for all vertices $\mathbf{V}\in\mathcal{V}$.
The proof of Case b is similar than the proof of Case a and we start by
showing for a tetrahedron, say $K$, with a facet on the boundary that $\left.
u_{\operatorname*{refl}}\right\vert _{K}=0$ and employ an induction over
adjacent tetrahedrons to prove that $u_{\operatorname*{refl}}=0$ on every
tetrahedron in $\mathcal{G}$.
We consider a boundary facet $T_{0}\in\mathcal{F}_{\partial\Omega}$ with
adjacent tetrahedron $K\subset\mathcal{G}$. We denote the three other facets
of $K$ by $T_{i}$, $1\leq i\leq3$, and for $0\leq i\leq3$, the vertex of $K$
which is opposite to $T_{i}$ by $\mathbf{A}_{i}$.
\textbf{Case b.1. }First we consider the case that there is one and only one
\textit{other} facet, say, $T_{1}$ which lies in $\partial\Omega$. Then
$\left. u_{\operatorname*{refl}}\right\vert _{T}=\left. u_{2}\right\vert
_{T}+\left. u_{3}\right\vert _{T}$ for some $u_{i}\in S_{T_{i
,\operatorname*{nc}}^{p,0}:=\operatorname*{span}\left\{ B_{p,0
^{T_{i},\operatorname*{nc}}\right\} $, $i=2,3$. From Theorem
\ref{TheoCDBasis}.b we conclude that the sum $\left. S_{T_{2
,\operatorname*{nc}}^{p,0}\right\vert _{T}\oplus\left. S_{T_{3
,\operatorname*{nc}}^{p,0}\right\vert _{T}$ is direct. The condition $\left.
u_{\operatorname*{refl}}\right\vert _{T}=0$ then implies $u_{2}=u_{3}=0$.
Thus, we have proved $\left. u_{\operatorname*{refl}}\right\vert _{K}=0$.
\textbf{Case b.2. }The case that there are exactly two \textit{other} facets
which are lying in $\partial\Omega$ can be treated in a similar way.
\textbf{Case b.3. }Next, we consider the case that $T_{i}\in\mathcal{F
_{\Omega}$ for $i=1,2,3$. Note that $\left. u_{\operatorname*{refl
}\right\vert _{T}=\sum_{i=1}^{3}\left. u_{i}\right\vert _{T}$ for some
$u_{i}\in S_{T_{i},\operatorname*{nc}}^{p,0}$. On $T$ we choose a local
$\left( x_{1},x_{2}\right) $-coordinate system such that $\mathbf{A
_{1}=\mathbf{0}$, $\mathbf{A}_{2}=\left( 1,0\right) ^{\intercal}$,
$\mathbf{A}_{3}=\left( 0,1\right) ^{\intercal}$. From (\ref{defqnk}) and
(\ref{defqnkplus}) we conclude that
\[
b_{n,k}^{\operatorname*{refl}}+RMb_{n,k}^{\operatorname*{refl}}+MRb_{n,k
^{\operatorname*{refl}}=0.
\]
This implies $\left. u_{2}\right\vert _{T}=RM\left( \left. u_{1}\right\vert
_{T}\right) =\left. u_{1}\right\vert _{T}\circ\chi_{\left\{ 3,2,1\right\}
}$ and $\left. u_{3}\right\vert _{T}=MR\left( \left. u_{1}\right\vert
_{T}\right) =\left. u_{1}\right\vert _{T}\circ\chi_{\left\{ 2,3,1\right\}
}$ (cf. (\ref{defchipi})) and, in turn, that the restrictions $u_{i}^{E}$ of
$u_{i}$ to the edge $E_{i}=T_{i}\cap T_{0}$, $1\leq i\leq3$, are the
\textquotedblleft same\textquotedblright, more precisely, the affine pullbacks
of $u_{i}^{E}$ to the interval $\left[ 0,1\right] $ are the same. From Lemma
\ref{TraceLemma}, we obtain tha
\begin{equation}
\left. u_{1}\right\vert _{T_{1}}\circ\chi_{1}=\left. u_{2}\right\vert
_{T_{2}}\circ\chi_{2}=\left. u_{3}\right\vert _{T_{3}}\circ\chi_{3},
\label{affequi
\end{equation}
where $\chi_{i}:\widehat{T}\rightarrow T_{i}$ are affine pullbacks to the
reference triangle such that $\chi_{i}\left( \mathbf{0}\right)
=\mathbf{A}_{0}$.
This implies that the functions $u_{i}$ at $\mathbf{A}_{0}$ have the same
value (say $w_{0}$) and, from the condition $u_{\operatorname*{refl}}\left(
\mathbf{A}_{0}\right) =3w_{0}=0$, we conclude that $u_{i}\left(
\mathbf{A}_{0}\right) =0$. The values of $u_{i}$ at the vertex $\mathbf{A
_{i}$ of $K$ (which is opposite to $T_{i}$) also coincide and we denote this
value by $v_{0}$. Since $\left. u_{\operatorname*{refl}}\right\vert _{T}=0$
it holds $u_{\operatorname*{refl}}\left( \mathbf{A}_{i}\right) =2w_{0
+v_{0}=0$. From $w_{0}=0$ we conclude that also $v_{0}=0$. Let $\chi_{i,T_{0
}:\widehat{T}\rightarrow T_{0}$ denote an affine pullback with the property
$\chi_{i,T_{0}}\left( \mathbf{0}\right) =\mathbf{A}_{i}$. Hence,
\begin{equation}
\widehat{u_{i}}:=\left. u_{i}\right\vert _{T_{0}}\circ\chi_{i,T_{0}}^{-1
\in\operatorname*{span}\left\{ b_{p,0}^{\operatorname*{refl}}\right\}
\label{defuihut
\end{equation}
with values zero at the vertices of $\hat{T}$. Note tha
\begin{equation}
b_{p,0}\left( 0,0\right) =\left( -1\right) ^{p}\left( p+1\right)
\quad\text{and\quad}b_{p,0}\left( 1,0\right) =b_{p,0}\left( 0,1\right) =1.
\label{bp000
\end{equation}
The vertex properties (\ref{bp000}) along the definition of $b_{p,k
^{\operatorname*{refl}}$ (cf. (\ref{defqnk})) imply tha
\begin{align}
b_{p,0}^{\operatorname*{refl}}\left( 1,0\right) & =b_{p,0
^{\operatorname*{refl}}\left( 0,1\right) =\frac{1}{3}\left( 1-\left(
-1\right) ^{p}\left( p+1\right) \right) =c_{p},\label{defcp}\\
b_{p,0}^{\operatorname*{refl}}\left( 0,0\right) & =-2b_{p,0
^{\operatorname*{refl}}\left( 1,0\right) .\nonumber
\end{align}
Since $c_{p}\neq0$ for $p\geq1$ we conclude that $\widehat{u}_{i}=0$ holds.
Relation (\ref{defuihut}) implies $\left. u_{i}\right\vert _{T_{0}}=0$ and
thus $u_{i}=0$. From $\left. u_{\operatorname*{refl}}\right\vert _{T
=\sum_{i=1}^{3}\left. u_{i}\right\vert _{T}$ we deduce that $\left.
u_{\operatorname*{refl}}\right\vert _{K}=0$.
The Cases b.1-.3 allow to proceed with the same induction argument as for Case
a and $u_{\operatorname*{refl}}=0$ follows by induction.
\textbf{Part 3. }An inspection of Part 2 shows that, for the proof of Case a,
it was never used that the vertex-oriented basis functions have been removed
from $S_{\mathcal{G},\operatorname*{c}}^{p}$ and Case a holds verbatim for
$S_{\mathcal{G},\operatorname*{c}}^{p}$. This implies that the first sum in
(\ref{directsums}) is direct.
\textbf{Part 4.} The fact that the sum $S_{\mathcal{G},\operatorname*{c}
^{p}+S_{\operatorname*{refl},\operatorname*{nc}}^{p}$ is \textit{not} direct
is postponed to Proposition \ref{PropVertremoval}
\endproof
\begin{proposition}
\label{PropVertremoval}For any vertex $\mathbf{V\in}\mathcal{V}_{\Omega}$ it
holds $B_{p,\mathbf{V}}^{\mathcal{G}}\in S_{\operatorname*{sym
,\operatorname*{nc}}^{p}\oplus S_{\operatorname*{refl},\operatorname*{nc
}^{p,0}\oplus\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}$.
\end{proposition}
\proof
We will show the stronger statement $B_{p,\mathbf{V}}^{\mathcal{G}}\in
S_{\operatorname*{refl},\operatorname*{nc}}^{p,0}\oplus\tilde{S
_{\mathcal{G},\operatorname*{c}}^{p}$. It suffices to construct a continuous
function $u_{\mathbf{V}}\in S_{\operatorname*{refl},\operatorname*{nc}}^{p}$
which coincides with $B_{p,\mathbf{V}}^{\mathcal{G}}$ at all vertices
$\mathbf{V}^{\prime}\in\mathcal{V}$ and vanishes at $\partial\Omega$; then,
$B_{p,\mathbf{V}}^{\mathcal{G}}-u_{\mathbf{V}}\in\tilde{S}_{\mathcal{G
,\operatorname*{c}}^{p}$ and the assertion follows. Recall the known values of
$b_{p,0}^{\operatorname*{refl}}$ at the vertices of the reference triangle and
the definition of $c_{p}$ as in (\ref{defcp}). Let $K\in\mathcal{G}$ be a
tetrahedron with $\mathbf{V}$ as a vertex. The facets of $K$ are denoted by
$T_{i}$, $0\leq i\leq3$, and the vertex which is opposite to $T_{i}$ is
denoted by $\mathbf{A}_{i}$. As a convention we assume that $\mathbf{A
_{0}=\mathbf{V}$. For every $T_{i}$, $1\leq i\leq3$, we define the function
$u_{T_{i}}\in S_{T_{i},\operatorname*{nc}}^{p}$ by setting (cf.
(\ref{defbmitttriangle})
\[
\left. u_{T_{i}}\right\vert _{T_{0}}=b_{p,0}^{\operatorname*{refl}}\circ
\chi_{\mathbf{A}_{i},T_{0}}^{-1},
\]
where $\chi_{\mathbf{A}_{i},T_{0}}:\widehat{T}\rightarrow T_{0}$ is an affine
pullback which satisfies $\chi_{\mathbf{A}_{i},T_{0}}\left( \mathbf{0
\right) =\mathbf{A}_{i}$. (It is easy to see that the definition of
$u_{T_{i}}$ is independent of the side of $T_{i}$, where the tetrahedron $K$
is located.) From (\ref{defqnk}) and (\ref{defqnkplus}) we conclude that
$\left. \sum_{i=1}^{3}u_{T_{i}}\right\vert _{T_{0}}=0$ holds. We proceed in
the same way for all tetrahedrons $K\in\mathcal{G}_{\mathbf{V}}$ (cf.
(\ref{defoftrianglesubsets})). This implies that
\begin{equation}
\tilde{u}_{\mathbf{V}}:=\sum_{\substack{T\in\mathcal{F}_{\Omega
\\\mathbf{V}\in T}}u_{T} \label{sumutildev
\end{equation}
vanishes at $\Omega\backslash\overset{\circ}{\omega_{\mathbf{V}}}$ (cf.
(\ref{defoftrianglesubsets})). By construction the function $\tilde
{u}_{\mathbf{V}}$ is continuous. At $\mathbf{V}$, the function $u_{T_{i}}$ has
the value (cf. (\ref{defcp})
\[
u_{T_{i}}\left( \mathbf{V}\right) =c_{p
\]
so that $\tilde{u}_{\mathbf{V}}\left( \mathbf{V}\right) =Cc_{p}$, where $C$
is the number of terms in the sum (\ref{sumutildev}). Since $c_{p}>0$ for all
$p\geq1$, the function $u_{\mathbf{V}}:=\frac{1}{Cc_{p}}\tilde{u}_{V}$ is well
defined and has the desired properties
\endproof
\begin{remark}
We have seen that the extension of the basis functions of $S_{\mathcal{G
,\operatorname*{c}}^{p}$ by the basis functions of $S_{\operatorname*{refl
,\operatorname*{nc}}^{p}$ leads to linearly depending functions. On the other
hand, if the basis functions of the subspace $S_{\operatorname*{refl
,\operatorname*{nc}}^{p,0}$ are added and the vertex-oriented basis functions
in $S_{\mathcal{G},\operatorname*{c}}^{p}$ are simply removed, one arrives at
a set a linear independent functions which span a larger space than
$S_{\mathcal{G},\operatorname*{c}}^{p}$. Note that $S_{\operatorname*{refl
,\operatorname*{nc}}^{p,0}=S_{\operatorname*{refl},\operatorname*{nc}}^{p}$
for $p=1,2,3$.
One could add more basis functions from $S_{\operatorname*{refl
,\operatorname*{nc}}^{p}$ but then has to remove further basis functions from
$\tilde{S}_{\mathcal{G},\operatorname*{c}}^{p}$ or formulate side constraints
in order to obtain a set of linearly independent functions.
\end{remark}
We finish this section by an example which shows that there exist meshes with
fairly special topology, where the inclusio
\begin{equation}
S_{\mathcal{G},\operatorname*{c}}^{p}+S_{\operatorname*{sym
,\operatorname*{nc}}^{p}+S_{\operatorname*{refl},\operatorname*{nc}
^{p}\subset S_{\mathcal{G}}^{p} \label{maxspace
\end{equation}
is strict. We emphasize that the left-hand side in (\ref{maxspace}), for
$p\geq4$, defines a larger space than the space in (\ref{finalspace}) since it
contains \textit{all} non-conforming functions of reflection type.
\begin{example}
Let us consider the octahedron $\Omega$ with vertices $\mathbf{A}^{\pm
}:=\left( 0,0,\pm1\right) ^{\intercal}$ and $\mathbf{A}_{1}:=\left(
1,0,0\right) ^{\intercal}$, $\mathbf{A}_{2}:=\left( 0,1,0\right)
^{\intercal}$, $\mathbf{A}_{3}:=\left( -1,0,0\right) ^{\intercal}$,
$\mathbf{A}_{4}:=\left( 0,-1,0\right) ^{\intercal}$. $\Omega$ is subdivided
into a mesh $\mathcal{G}:=\left\{ K_{i}:1\leq i\leq8\right\} $ consisting of
eight congruent tetrahedrons sharing the origin $\mathbf{0}$ as a common
vertex. The six vertices at $\partial\Omega$ have the special topological
property that each one belongs to exactly four surface facets.
Note that the space defined by the left-hand side of (\ref{maxspace}) does not
contain functions whose restriction to a surface facet, say $T$, belongs to
the $\tau_{\operatorname{sign}}$ component of $\mathbb{P}_{n,n-1}^{\perp
}\left( T\right) $. Hence, the inclusion in (\ref{maxspace}) is strict if we
identify a function in $S_{\mathcal{G}}^{p}$ whose restriction to some surface
facet is an orthogonal polynomial of \textquotedblleft$\operatorname{sign}$
type\textquotedblright. Let $\widehat{q}\neq0$ be a polynomial which belongs
to the $\tau_{\operatorname{sign}}$ component of $\mathbb{P}_{n,n-1}^{\perp
}\left( T\right) $ on the reference element. Denote the (eight) facet on
$\partial\Omega$ with the vertices $\mathbf{A}^{\pm}$, $\mathbf{A}_{i}$,
$\mathbf{A}_{i+1}$ by $T_{i}^{\pm}$ for $1\leq i\leq4$ (with cyclic numbering
convention) and choose affine pullbacks $\chi_{\pm,i}:\widehat{T}\rightarrow
T_{i}^{\pm}$ as $\chi_{\pm,i}\left( \mathbf{x}\right) :=\mathbf{A}^{\pm
}+x_{1}\left( \mathbf{A}_{i}-\mathbf{A}^{\pm}\right) +x_{2}\left(
\mathbf{A}_{i+1}-\mathbf{A}^{\pm}\right) $. Then, it is easy to verify (use
Lemma \ref{Lemcont} with even $m_{C}$) that the function $q:\partial
\Omega\rightarrow\mathbb{R}$, defined by $\left. q\right\vert _{T_{i}^{\pm
}:=\widehat{q}\circ\chi_{\pm,i}^{-1}$ is continuous on $\partial\Omega$. Hence
the \textquotedblleft finite element extension\textquotedblright\ to the
interior of $\Omega$ via
\[
Q:=\sum_{\mathbf{N}\in\mathcal{N}^{p}\cap\partial\Omega}q\left(
\mathbf{N}\right) B_{p,\mathbf{N}}^{\mathcal{G}
\]
defines a function in $S_{\mathcal{G}}^{p}$ which is not in the space defined
by the left-hand side of (\ref{maxspace}).
We state in passing that the space $S_{\mathcal{G}}^{p}$ does not contain any
function whose restriction to a boundary facet, say $T$, belongs to the
$\tau_{\operatorname{sign}}$ component of $\mathbb{P}_{p,p-1}^{\perp}\left(
T\right) $ if there exists at least one surface vertex which belongs to an
\emph{odd} number of surface facets. In this sense, the topological situation
considered in this example is fairly special.
\end{example}
\section{Conclusion\label{SecConclusion}}
In this article we developed explicit representation of a local basis for
non-conforming finite elements of the Crouzeix-Raviart type. As a model
problem we have considered Poisson-type equations in three-dimensional
domains; however, this approach is by no means limited to this model problem.
Using theoretical conditions in the spirit of the second Strang lemma, we have
derived conforming and non-conforming finite element spaces of arbitrary
order. For these spaces, we also derived sets of local basis functions. To the
best of our knowledge, such explicit representation for general polynomial
order $p$ are not available in the existing literature. The derivation
requires some deeper tools from orthogonal polynomials of triangles, in
particular, the splitting of these polynomials into three irreducible
irreducible $\mathcal{S}_{3}$ modules.
Based on these orthogonal polynomials, simplex- and facet-oriented
non-conforming basis functions are defined. There are two types of
non-conforming basis functions: those whose supports consist of one
tetrahedron and those whose supports consist of two adjacent tetrahedrons. The
first type can be simply added to the conforming $hp$ basis functions. It is
important to note that the span of the functions of the second type contains
also conforming functions and one has to remove some conforming functions in
order to obtain a linearly independent set of functions. We have proposed a
non-conforming space which consists of a) all basis functions of the first
type and b) a reduced set of basis functions of the second type and c) of the
conforming basis functions without the vertex-oriented ones. This leads to a
set of linearly independent functions and is in analogy to the well known
lowest order Crouzeix-Raviart element.
It is interesting to compare these results with high-order Crouzeix-Raviart
finite elements for the two-dimensional case which have been presented in
\cite{ccss_2012}. Facets $T$ of tetrahedrons in 3D correspond to edges $E$ of
triangles in 2D. As a consequence the dimension of the space of orthogonal
polynomials $\mathbb{P}_{p,p-1}^{\perp}\left( E\right) $ equals one. For
even degree $p$, one has only non-conforming basis functions of
\textquotedblleft symmetric\textquotedblright\ type (which are supported on a
single triangle) and for odd degree $p$, one has only non-conforming basis
functions of \textquotedblleft reflection\textquotedblright\ type (which are
supported on two adjacent triangles). It turns out that adding the non
conforming symmetric basis function to the conforming $hp$ finite element
space leads to a set of linearly independent functions which is the analogue
of the first sum in (\ref{directsums}). If the non-conforming basis functions
of reflection type are added, the set of vertex-oriented conforming basis
functions have to be removed from the conforming space. This is in analogy to
the properties (\ref{nondirectsum}) and (\ref{finalspace}).
Future research is devoted on numerical experiments and the application of
these functions to system of equations as, e.g., Stokes equation and the
Lam\'{e} system.\medskip
\noindent\textbf{Acknowledgement} This work was supported in part by ENSTA,
Paris, through a visit of S.A. Sauter during his sabbatical. This support is
gratefully acknowledged.
|
\section{Supplementary material}
\section{Laser scheme for the Creutz model}
\setcounter{equation}{0}
We give a description of the cold atomic proposal. Consider atoms with two internal long-lived states $\{|g\rangle,|e\rangle\}$ trapped in an optical zigzag lattices \cite{Zhang15a} which are spatially displaced horizontally by the lattice constant $a$ between the $g$ and $e$ states. The spatial overlap between nearest neighbour orbital wavefunctions (within the single band approximation) gives rise to the following tight-binding model:
\begin{eqnarray}
H_{zz}=-J_D \sum_{m\in \mathbb{Z}} \biggl( |m+1,1\rangle_g\, _g\langle m,2| +|m+1,2\rangle_e\, _e\langle m,1|+\tr{h.c.} \biggr),\nonumber
\end{eqnarray}
where $|m,1\rangle_g$ denotes an atom in the internal $g$ state at the position $m$ in the lower chain (e.g., $|m,2\rangle_e$ denotes internal state $e$, position $m$ and in the upper chain.) and $J_D$ is the hopping strength set by the optical lattice depth, see Fig. 5a. In addition, we impose a superlattice potential \cite{Gerbier10a} to modify spatially the onsite energies of $\{|g\rangle,|e\rangle\}$, with three characteristic energy parameters $V_1,V_2,V_1+V_2$ and they are staggered in between the upper and lower chains, resulting in the energy level diagram shown in the main text Fig. 4b and Fig. 5b. The various energy differences are given by
\begin{eqnarray}
&&\omega_1=E_0+V_1+V_2,\tr{\ \ }\omega_2=E_0+V_1+2 V_2,\tr{\ \ }\omega_3=E_0+V_1,\tr{\ \ }\nonumber\\
&&\omega_4=E_0+2 V_1,\tr{\ \ }\omega_5=E_0+2 V_2,\tr{\ \ }\omega_6=E_0+2 V_1+2 V_2,\tr{\ \ }\omega_7=E_0,
\end{eqnarray}
and denoting the eight energy levels as $\{|g_i\rangle,|e_i\rangle\}$ for $i=1-4$. By turning on laser couplings $|g\rangle \leftrightarrow |e\rangle$ (with the Rabi frequency set to unity, for simplicity in notation), the atom-light coupling Hamiltonian \cite{Dalibard11a}, including the onsite energy shifts, in the rotating-wave-approximation is
\begin{eqnarray}
H_{a-l}&=&e^{-i\omega_1 t-i\phi_1}|e_1\rangle\, \langle g_1|+e^{-i\omega_2 t-i\phi_2}|e_1\rangle\, \langle g_2|+e^{-i\omega_1 t-i\phi_1}|e_2\rangle\, \langle g_2|+e^{-i\omega_3 t-i\phi_3}|e_2\rangle\, \langle g_1 |
\nonumber\\
&+&e^{-i\omega_1 t-i\phi_1}|e_3\rangle\, \langle g_3|+e^{-i\omega_2 t-i\phi_2}|e_4 \rangle\, \langle g_3|
+e^{-i\omega_1 t-i\phi_1}|e_4\rangle\, \langle g_4|+e^{-i\omega_3 t-i\phi_3}|e_3\rangle\, \langle g_4|\nonumber\\
&+&e^{-i\omega_4 t-i\phi_4}|e_3\rangle\, \langle g_1|+e^{-i\omega_5 t-i\phi_5}|e_1\rangle\, \langle g_3|+e^{-i\omega_6 t-i\phi_6}|e_4\rangle\, \langle g_2|+e^{-i\omega_7 t-i\phi_7}|e_2\rangle\, \langle g_4|+\tr{h.c.}\nonumber\\
&+&(-V_1) |g_1\rangle\, \langle g_1|+(E_0+V_2) |e_1\rangle\, \langle e_1|+(-V_1-V_2) |g_2\rangle\, \langle g_2|+(E_0) |e_2\rangle\, \langle e_2|\nonumber\\
&+&(E_0+V_1) |e_3\rangle\, \langle e_3|+(-V_2) |g_3\rangle\, \langle g_3|+(V_1+V_2) |e_4\rangle\, \langle e_4|+(0) |g_4\rangle\, \langle g_4|.
\end{eqnarray}
\begin{figure}[!b]
\begin{center}
\includegraphics[width=15cm]{fa2.pdf}
\end{center}
\caption{(a) Two zigzag chains trapping $g$ and $e$ states, respectively. (b) The energy level diagram in the presence of the superlattice potential.}
\end{figure}
The phases $\phi_j$ for $j=1-7$ are associated with the initial seven resonance lasers. We now perform the following unitary transformations on the internal states:
\begin{eqnarray}
&&|g_1\rangle \rightarrow e^{-i V_1 t}|g_1\rangle;\tr{\ \ } |g_2\rangle \rightarrow e^{-i (V_1+V_2) t}|g_2\rangle;\tr{\ \ }
|e_1\rangle \rightarrow e^{i (E_0+V_2) t}|e_1\rangle;\tr{\ \ } |e_2\rangle \rightarrow e^{i E_0 t}|e_2\rangle;\nonumber\\
&&|g_3\rangle \rightarrow e^{-i V_2 t}|g_3\rangle;\tr{\ \ } |g_4\rangle \rightarrow e^{-i 0 t}|g_4\rangle;\tr{\ \ \ \ \ \ \ \ \ \ }
|e_3\rangle \rightarrow e^{i (E_0+V_1) t}|e_3\rangle;\tr{\ \ } |e_4\rangle \rightarrow e^{i (E_0+V_1+V_2) t}|e_4\rangle;\nonumber
\end{eqnarray}
and arrive at a time-independent atom-light coupling Hamiltonian
\begin{eqnarray}
H_{a-l}&=&e^{-i\phi_1}|e_1\rangle\, \langle g_1|+e^{-i\phi_2}|e_1\rangle\, \langle g_2|+e^{-i\phi_1}|e_2\rangle\, \langle g_2|+e^{-i\phi_3}|e_2\rangle\, \langle g_1 |
\nonumber\\
&+&e^{-i\phi_1}|e_3\rangle\, \langle g_3|+e^{-i\phi_2}|e_4 \rangle\, \langle g_3|
+e^{-i\phi_1}|e_4\rangle\, \langle g_4|+e^{-i\phi_3}|e_3\rangle\, \langle g_4|\nonumber\\
&+&e^{-i\phi_4}|e_3\rangle\, \langle g_1|+e^{-i\phi_5}|e_1\rangle\, \langle g_3|+e^{-i\phi_6}|e_4\rangle\, \langle g_2|+e^{-i\phi_7}|e_2\rangle\, \langle g_4|+\tr{h.c.}\nonumber
\end{eqnarray}
We demand the laser phases to be locked at $\phi_1=-\phi_2=-\phi_3=-\theta$, $\phi_4=-\phi_5=\phi_6=-\phi_7=-\phi$. To take advantage of the gauge freedom, we note that there are actually four observable gauge-invariant phases given
\begin{eqnarray}
\Phi_1=2\phi_1-\phi_4-\phi_5,\tr{\ \ }\Phi_2=2\phi_2-\phi_5-\phi_6,\tr{\ \ }\Phi_3=2\phi_1-\phi_6-\phi_7,\tr{\ \ }\Phi_4=2\phi_3-\phi_4-\phi_7.
\end{eqnarray}
The phase locking condition amounts to two further conditions on the gauge-invariant phases $\Phi_1+\Phi_2=\Phi_3+\Phi_4=0$. The latter can facilitate in the design of the coupling of the various laser phases.
\\
\\
By controlling the Rabi frequencies and taking into account the finite spatial overlap between the single-band wavefunctions we finally obtain the Hamiltonian for the generalized Creutz scheme
\begin{eqnarray}
H_{zz}+H_{a-l}&=&H_{zz}+ \hbar \Omega_1 \sum_{m\in \mathbb{Z}} \biggl(e^{i \phi} |m,1\rangle_e\, _g\langle m,2 |+ \tr{h.c.}\biggr)\nonumber\\
&&+ \hbar \Omega_2 \sum_{m\in \mathbb{Z}} \biggl( e^{i\theta} |m+1,2\rangle_e \, _g\langle m,2 |+ e^{-i\theta} |m+1,1\rangle_g \, _e\langle m,1 |+\tr{h.c.}\biggr)\nonumber
\end{eqnarray}
|
\section{Introduction}
Changepoint detection is a central problem in fields such as finance
or genomics, where $n$ data are gathered in a sequence over time or
space. Many models define the optimal changepoints using maximum
likelihood, resulting in a discrete optimization problem. Multiple
changepoint detection models seek the optimal $K$ segments
($K-1$ changes), which amounts to optimizing likelihood
parameters over a space that contains $O(n^{K-1})$ discrete
arrangements of changepoints. In general this problem can be solved
in $O(Kn^2)$ time using the original dynamic programming algorithm of
\citet{segment-neighborhood}. Recently proposed pruning techniques
reduce the number of changepoints considered by the algorithm, thus
reducing time complexity to $O(K n\log n)$ while maintaining
optimality \citep{pruned-dp, johnson, fpop}.
In ``unconstrained'' changepoint models, there are no contraints
between model parameters on separate segments. To regularize and obtain a more
interpretable model, it is often desirable to introduce constraints
between model parameters before and after changepoints.
For example, the main problem that motivates this paper is peak
detection in ChIP-seq data, which provide noisy measurements of
protein binding or modification throughout a genome \citep{practical}. An
up-down constrained changepoint detection model has been shown to
achieve state-of-the-art peak detection accuracy in ChIP-seq data
\citep{HOCKING-PeakSeg}. The constraints of this model force an up
change in the segment mean parameter after each down change, and vice
versa.
The fastest existing solver for this problem is the Constrained
Dynamic Programming Algorithm (CDPA), which has two issues. First, it
is a heuristic algorithm that is not guaranteed to recover the optimal
solution. Second, its $O(Kn^2)$ quadratic time complexity is too slow
for use on large data sets. In this
paper we propose a new algorithm that fixes both of these issues.
\subsection{Contributions and organization}
We begin by discussing previous research into pruning techniques for
solving unconstrained changepoint detection problems
(Section~\ref{sec:related}), then state the constrained optimization
problems (Section~\ref{sec:models}). Our main contribution is
Section~\ref{sec:algorithms}, which generalizes the functional pruning
technique of \citet{pruned-dp}, thus providing a new Generalized
Pruned Dynamic Progamming Algorithm (GPDPA) for solving a class of
constrained changepoint detection problems. We show that the GPDPA
achieves state-of-the-art speed and accuracy in genomic data with
several different labeled patterns (Section~\ref{sec:results}), then
conclude by discussing the significance of our contributions
(Section~\ref{sec:discussion}).
\begin{table*}[b!]
\centering
\begin{tabular}{r|c|c}
& No pruning & Functional pruning \\
\hline
Unconstrained & Dynamic Programming Algorithm (DPA) & Pruned DPA (PDPA) \\
& Optimal solution, $O(Kn^2)$ time & Optimal solution, $O(Kn\log n)$ time\\
& \citet{segment-neighborhood} & \citet{pruned-dp, johnson} \\
\hline
Up-down constrained & Constrained DPA (CDPA) & Generalized Pruned DPA (GPDPA) \\
& Sub-optimal solution, $O(Kn^2)$ time & Optimal solution, $O(Kn\log n)$ time\\
& \citet{HOCKING-PeakSeg} & \textbf{This paper} \\
\hline
\end{tabular}
\caption{Our contribution is
the Generalized Pruned Dynamic Programming Algorithm (GPDPA),
which uses a functional pruning technique
to compute the constrained optimal $K-1$ changepoints
in a sequence of $n$ data, in $O(K n\log n)$ time on average.}
\label{tab:contribution}
\end{table*}
\section{Related work}
\label{sec:related}
There are many efficient algorithms available for computing the
optimal $K-1$ changepoints in $n$ data
points. \citet{segment-neighborhood} proposed an $O(K n^2)$ algorithm
for computing the sequence of models with $1,\dots,K$ segments.
\citet{optimal-partitioning} consider a related approach, which
introduces a penalty for each changepoint, rather than fixing the
number of changepoints. Their $O(n^2)$ algorithm computes the single
model for a given penalty constant $\lambda$. Both of these algorithms
recover the optimal solution, and follow from using dynamic
programming updates \citep{bellman} to recursively compute the maximum
likelihood from 1 to $n$ data points. Alternatively there are methods
which are computationally faster but are not guaranteed to find the
optimal segmentation. The most popular of these is the binary
segmentation algorithm which has $O(Kn)$ worst-case time complexity
\citep{binary-segmentation}. An L1 relaxation of this problem is known
as the fused lasso signal approximator, for which efficient solvers
also exist \citep{flsa}.
Several pruning methods have been recently proposed in order to reduce
time complexity, while maintaining optimality. \citet{pruned-dp} and
\citet{phd-johnson} independently discovered a functional pruning
technique, which results in algorithms with $O(n\log n)$ average time
complexity. \citet{pelt} proposed an inequality pruning technique,
which results in an algorithm with average time complexity from $O(n)$
to $O(n^2)$, depending on the number of changes. \citet{fpop} provides
a clear discussion on the differences between the two pruning
techniques.
All algorithms discussed thus far are for solving problems with no
constraints between adjacent segment mean parameters, but there are
many examples of constrained changepoint detection models. Rather
than searching all possible changepoints and likelihood parameters,
the idea is to use a constraint in order to search a smaller, more
interpretable model space. For example, \citet{haiminen2008algorithms}
propose an $O(Kn^2)$ algorithm for unimodal regression, which enforces
no up changes after the first down change. \citet{HOCKING-PeakSeg}
proposed an $O(Kn^2)$ algorithm for peak detection, which enforces a
down change after each up change, and vice versa.
Isotonic regression is another example of a constrained changepoint
detection model. There is no limit on the number of segments $K$, but
the segment means are constrained to be non-decreasing. This problem
can be solved in $O(n)$ time using the pool-adjacent-violators
algorithm \citep{mair2009isotone}, or in $O(n\log n)$ time using a
dynamic programming algorithm \citep{isotonic-dp}. An L1 relaxation of
this problem is known as nearly-isotonic regression
\citep{tibshirani2011nearly}. A problem known as reduced isotonic
regression occurs by imposing an additional constraint of $K$ segments
\citep{reduced-monotonic-regression}. The techniques for solving this
problem lead to sub-quadratic time algorithms
\citep{hardwick2014optimal}, but do not generalize to other kinds of
constraints (such as unimodal regression or peak detection).
Our contribution in this paper is proving that the functional pruning
technique can be generalized to constrained changepoint models
(Table~\ref{tab:contribution}). Our resulting Generalized Pruned
Dynamic Programming Algorithm (GPDPA) enjoys $O(Kn\log n)$ time
complexity, and works for any changepoint model with affine
constraints between adjacent segment means (including isotonic
regression, unimodal regression, and peak detection).
\section{Isotonic regression and changepoint models}
\label{sec:models}
Although our proposed algorithm can solve many constrained changepoint
detection problems (Section~\ref{sec:general}), we will simplify our
discussion by emphasizing the isotonic regression model.
\subsection{Classical isotonic regression}
The classical isotonic regression model is defined as the most likely
sequence of non-decreasing segment means. More
precisely, assume that the data $\mathbf y\in\mathbb R^n$ are a realization
of a probability distribution with mean parameter $\mathbf m\in\mathbb R^n$. For
example, assuming $y_t \sim \mathcal N(m_t, \sigma^2)$ and performing
maximum likelihood inference results in a convex minimization problem
with affine constraints,
\begin{align}
\label{eq:isotonic}
\minimize_{\mathbf m\in\mathbb R^n} &\ \
\sum_{t=1}^n \ell(y_t, m_t)\\
\text{subject to} & \ \ m_t \leq m_{t+1},\, \forall t<n.
\nonumber
\end{align}
The convex loss function $\ell:\mathbb R\times \mathbb R\rightarrow\mathbb R$ in the
case of the Gaussian likelihood is the square loss
$\ell(y, m) = (y-m)^2$. This optimization problem (\ref{eq:isotonic})
is referred to as isotonic regression, and can be efficiently solved
in $O(n)$ time using the Pool-Adjacent-Violators Algorithm (PAVA)
\citep{isotonic-unifying}.
Since isotonic regression imposes no limit on the number of
changepoints ($m_t < m_{t+1}$), it tends to overfit. For example,
consider the toy data set $\mathbf y= \left[
\begin{array}{cccccc}
2 & 5 & 30 & 34 & 600 & 621
\end{array}
\right] \in\mathbb R^6$. Because these data are strictly increasing, the
isotonic regression (\ref{eq:isotonic}) solution is the trivial model
$m_t=y_t$. However, these data contain only two large changes. To
recover these changes, we could instead use the segment neighborhood
model, which we discuss in the next section.
\subsection{Segment neighborhood changepoint model}
The segment neighborhood model of \citet{segment-neighborhood} uses
the same cost function as isotonic regression, but a different
constraint set. There is no constraint on the direction of changes,
but there must be exactly $K\leq n$ distinct segments ($K-1$ changes).
\begin{align}
\label{eq:optimal_segment_neighborhood}
\minimize_{\mathbf m\in\mathbb R^n} &\ \
\sum_{t=1}^n \ell(y_t, m_t)\\
\text{subject to} &\ \ \sum_{t=1}^{n-1} I(m_t \neq m_{t+1}) = K-1.
\nonumber
\end{align}
This optimization problem is non-convex since the model complexity is
the number of changepoints, measured via the non-convex indicator
function $I$. Nonetheless, the optimal solution can be computed in
$O(K n^2)$ time using the standard dynamic programming algorithm
\citep{segment-neighborhood}. By exploiting the structure of the
convex loss function $\ell$, the pruned dynamic programming algorithm
of \citet{pruned-dp} computes the same optimal solution in faster
$O(K n \log n)$ time.
Unlike isotonic regression, the segment neighborhood model does not
constrain the direction of the changes. Thus, for some data sets
$\mathbf y$, the segment neighborhood model may recover a change down
($m_t > m_{t+1}$). For applications where isotonic regression is used,
it would be desirable to compute a model with $K$ non-decreasing
segment means. This results in the reduced isotonic regression
problem, which we introduce in the next section.
\subsection{Reduced isotonic regression}
The idea of fitting a non-decreasing function with a limited number of
changepoints has been previously described as reduced isotonic
regression \citep{reduced-monotonic-regression}. Combining the
constraints of the isotonic regression (\ref{eq:isotonic}) and segment
neighborhood (\ref{eq:optimal_segment_neighborhood}) problems gives
\begin{align}
\label{eq:reduced}
\minimize_{\mathbf m\in\mathbb R^n} &\ \
\sum_{t=1}^n \ell(y_t, m_t)\\
\text{subject to} &\ \ \sum_{t=1}^{n-1} I(m_t \neq m_{t+1}) = K-1,
\nonumber\\
&\ \ m_t \leq m_{t+1},\, \forall t<n.
\nonumber
\end{align}
In the next section, we explain how functional pruning
can be used for solving this and related changepoint problems.
\newcommand{C}{C}
\newcommand{\mathcal{M}}{\mathcal{M}}
\section{Functional
pruning algorithms for constrained
changepoint models}
\label{sec:algorithms}
We begin by discussing an algorithm for solving the reduced isotonic
regression problem, then explain how the algorithm
generalizes to other constrained changepoint problems.
\subsection{Equivalent optimization space}
The reduced isotonic regression problem (\ref{eq:reduced}) has $n$
segment mean variables $m_t$, one for each data point $t$. To derive
our algorithm, we re-write the problem in terms of the mean
$u_k\in\mathbb R$ and endpoint $t_k\in\{1,\dots,n\}$ for each
segment $k\in\{1,\dots, K\}$.
\begin{definition}[Reduced isotonic regression optimization space]
\label{def:Ibar}
Let $(\mathbf u, \mathbf t)\in{\mathcal I}^n_K$ be the set of
non-decreasing segment means $u_1\leq\cdots\leq u_K$ and
increasing changepoint indices $0=t_0<t_1<\cdots<t_{K-1}<t_K=n$.
\end{definition}
Each segment mean $u_k$ is assigned to data points
$\tau\in(t_{k-1},t_k]\subset\{1,\dots,n\}$, resulting in the following
cost for each segment $k\in\{1, \dots, K\}$,
\begin{equation}
\label{eq:h}
h_{t_{k-1}, t_k}(u_k) = \sum_{\tau=t_{k-1}+1}^{t_k} \ell(y_\tau, u_k).
\end{equation}
The reduced isotonic regression problem can be equivalently written as
\begin{equation}
\label{eq:isotonic_ut}
\minimize_{(\mathbf u, \mathbf t)\in{\mathcal I}^n_K}
\sum_{k=1}^K
h_{t_{k-1}, t_k}(u_k)
\end{equation}
Rather than explicitly summing over data points $i$ as in problem
(\ref{eq:reduced}), this problem uses the equivalent sum over segments $k$.
\begin{figure*}[t!]
\centering
\input{figure-compare-unconstrained}
\input{figure-compare-cost}
\vskip -0.5cm
\caption{Comparison of previous unconstrained algorithm
(\textcolor{Min}{grey}) with new algorithm that constrains segment
means to be non-decreasing (\textcolor{Ckt}{red}), for the toy data
set $\mathbf y= [ 2, 1, 0, 4 ] \in\mathbb R^4$ and the square
loss. \textbf{Left:} rather than computing the unconstrained
minimum (constant grey function), the new algorithm computes the
min-less operator (red), resulting in a larger cost when the
segment mean is less than the first data point ($\mu <
2$). \textbf{Right:} adding the cost of the second data point
$(\mu-1)^2$ and minimizing yields equal means $u_1=u_2=1.5$ for
the constrained model and decreasing means $u_1=2,\, u_2=1$ for
the unconstrained model.}
\label{fig:compare-unconstrained}
\end{figure*}
\subsection{Dynamic programming update rules}
\label{sec:dyn-prog}
Optimization problem (\ref{eq:isotonic_ut}) has $K$ segment mean
variables $u_k$ and $K-1$ changepoint index variables $t_k$. Minimizing over all
variables except the last segment mean $u_K$ results in the following
definition of the optimal cost.
\begin{definition}[Optimal cost with last segment mean $\mu$]
\label{def:fcc}
Let $C_{K,n}(\mu)$ be the optimal cost of the segmentation
with $K$ segments, up to data point $n$, with last segment mean
$\mu$:
\begin{equation}
C_{K,n}(\mu) = \min_{(\mathbf u, \mathbf t)\in{\mathcal I}^n_K \ | \ u_K = \mu} \
\left\{ \sum_{k=1}^K
h_{t_{k-1}, t_k}(u_k) \right\}.
\end{equation}
\end{definition}
As in the PDPA of \citet{pruned-dp}, our proposed dynamic programming
algorithm uses an exact representation of the
$C_{k,t}:\mathbb R\rightarrow\mathbb R$ cost functions. Each $C_{k,t}(\mu)$ is
represented as a piecewise function on intervals of $\mu$. This is
implemented as a linked list of FunctionPiece objects in C++ (for
details see Section~\ref{sec:pseudocode}). Each element of the linked list
represents a convex function piece, and implementation details depend
on the choice of the loss function $\ell$ (for an example using the
square loss see Section~\ref{sec:example-comparison}).
In the original unconstrained PDPA, computing the $C_{k,t}(u_k)$
function requires taking the minimum of $C_{k,t-1}(u_k)$ (a function
of the last segment mean $u_k$) and
$\hat C_{k-1,t-1} = \min_{u_{k-1}} C_{k-1,t-1}(u_{k-1})$ (the constant
loss resulting from an unconstrained minimization with respect to the
previous segment mean $u_{k-1}$). The main novelty of our paper is the
discovery that this update can also be computed efficiently for
constrained problems. For example in reduced isotonic regression the second
term is no longer a constant, but instead a function of $u_k$,
$C_{k-1,t-1}^{\leq}(u_k) = \min_{u_{k-1}\leq u_k}
C_{k-1,t-1}(u_{k-1})$, which we refer to as the min-less operator
(Figure~\ref{fig:compare-unconstrained}, left).
\begin{definition}[Min-less operator]
\label{def:min-less}
Given any real-valued function $f:\mathbb R\rightarrow\mathbb R$, we define the min-less
operator of that function as $f^\leq(\mu)=\min_{x\leq \mu} f(x)$.
\end{definition}
The min-less operator is used in the following Theorem, which states
the update rules used in our proposed algorithm.
\begin{theorem}[Generalized Pruned Dynamic Programming Algorithm
for reduced isotonic regression]
\label{thm:gpdpa}
The optimal cost functions $C_{k,t}$ can be recursively computed
using the following update rules.
\begin{enumerate}
\item For $k=1$ we have
$C_{1,1}(\mu)=\ell(y_1,\mu)$, and for the other data
points $t>1$ we have
\begin{equation}
C_{1,t}(\mu)=C_{1,t-1}(\mu)+\ell(y_t,\mu)
\end{equation}
\item For $k>1$ and $t=k$ we have
\begin{equation}
C_{k,k}(\mu)=\ell(y_k, \mu)+C_{k-1,k-1}^\leq(\mu)
\end{equation}
\item In all other cases we have
\begin{equation}
C_{k,t}(\mu)=\ell(y_t,\mu)+
\min\{
C_{k-1,t-1}^\leq(\mu),\,
C_{k,t-1}(\mu)
\}.
\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof}
Case 1 and 2 follow from Definition~\ref{def:fcc}, and there is a proof for case 3 in Section~\ref{sec:proof}.
\end{proof}
The dynamic programming algorithm requires computing $O(Kn)$ cost
functions $C_{k,t}$. As in the original pruned dynamic programming
algorithm, the time complexity of the algorithm is $O(K n I)$ where
$I$ is the number of intervals (convex function pieces; candidate
changepoints) that are used to represent the cost functions. The
theoretical maximum number of intervals is $I=O(n)$, implying a time
complexity of $O(K n^2)$ \citep{pruned-dp-new}. However, this maximum
is only achieved in pathological synthetic data sets, such as a
monotonic increasing data sequence. The average number of intervals in
real data sets is empirically $I=O(\log n)$, as we will show in
Section~\ref{sec:results_time}. Thus the average time complexity of
the algorithm is $O(K n \log n)$.
\subsection{Example and comparison with unconstrained case}
\label{sec:example-comparison}
To clarify the discussion, consider the
toy data set $\mathbf y= \left[
\begin{array}{cccccc}
2 & 1 & 0 & 4
\end{array}
\right] \in\mathbb R^4$ and the square loss $\ell(y,\mu)=(y-\mu)^2$. The first
step of the algorithm is to compute the minimum and the maximum of the
data (0,4) in order to bound the possible values of the segment
mean $\mu$. Then the algorithm computes the optimal cost in $k=1$ segment up
to data point $t=1$:
\begin{equation}
C_{1,1}(\mu) = (2-\mu)^2=4 - 4\mu + \mu^2\text{ (for $\mu\in[0,4]$)}
\end{equation}
This function can be stored for all values of $\mu$ via the three
real-valued coefficients ($\text{constant}=4$, $\text{linear}=-4$,
$\text{quadratic}=1$). To compute the optimal cost in $K=2$ segments,
we first compute the min-less operator (red curve on left of
Figure~\ref{fig:compare-unconstrained}),
\begin{equation}
C_{1,1}^\leq(\mu) =
\begin{cases}
4 - 4\mu + \mu^2 &\text{ if }\mu\in[0,2],\, \mu'=\mu,\\
0 + 0\mu + 0\mu^2 & \text{ if }\mu\in[2,4],\, \mu'=2.
\end{cases}
\end{equation}
This function can be stored as a list of two
intervals of $\mu$ values, each with associated real-valued
coefficients. In addition, to facilitate recovery of the optimal
parameters, we store the previous segment mean $\mu'$ and endpoint
(not shown). Note that $\mu'=\mu$ means that the equality constraint
is active ($u_1=u_2$).
By adding the first min-less function $C_{1,1}^\leq(\mu)$ to the
cost of the second data point $(\mu-2)^2$ we obtain the optimal cost in $K=2$
segments up to data point $t=2$,
\begin{equation}
C_{2,2}(\mu) =
\begin{cases}
5 - 6\mu + 2\mu^2 &\text{ if }\mu\in[0,2],\, \mu'=\mu,\\
1 - 2\mu + 1\mu^2 &\text{ if }\mu\in[2,4],\, \mu'=2.
\end{cases}
\end{equation}
Note that the minimum of this function is achieved at $\mu=1.5$ which
occurs in the first of the two function pieces (red curve on right of
Figure~\ref{fig:compare-unconstrained}), with an equality constraint
active. This implies the optimal model up to data point $t=2$ with
$k=2$ non-decreasing segment means actually has no change
($u_1=u_2=1.5$). In contrast, the minimum of the cost computed by the
unconstrained algorithm is at $u_2=1$ (grey curve on right of
Figure~\ref{fig:compare-unconstrained}), resulting in a change down
from $u_1=2$.
\subsection{The PeakSeg up-down constraint}
\label{sec:PeakSeg}
The PeakSeg model described by \citet{HOCKING-PeakSeg} is the most
likely segmentation where the first change is up, all up changes are
followed by down changes, and all down changes are followed by up
changes. More precisely, the constrained optimization problem can be
stated as
\begin{align}
\label{eq:PeakSeg}
\minimize_{
\substack{\mathbf u\in\mathbb R^K \\
0=t_0<t_1<\cdots<t_{K-1}<t_K=n
}
} &\ \
\sum_{k=1}^K h_{t_{k-1}, t_k}(u_k)\\
\text{subject to \hskip 0.9cm} &\ \ u_{k-1} \leq u_k\ \forall k\in\{2,4,\dots\},
\nonumber\\
&\ \ u_{k-1} \geq u_k\ \forall k\in\{3,5,\dots\}.
\nonumber
\nonumber
\end{align}
\begin{figure*}[t!]
\centering
\input{figure-2-min-envelope}
\vskip -0.5cm
\caption{
Demonstration of GPDPA for the PeakSeg model~(\ref{eq:PeakSeg})
with $k=3$ segments. Cost functions are stored as piecewise
functions on intervals (black dots show limits between function
pieces). \textbf{Left:} the min \textcolor{Min}{$M_{3,34}$} is the
minimum of two functions: \textcolor{MinMore}{$C^{\geq}_{2,34}$}
is the cost if the second segment ends at data point $t=34$ (the
min-more operator forces a non-increasing change after), and
\textcolor{Ckt}{$C_{3,34}$} is the cost if the second segment ends
before that. \textbf{Middle:} the cost \textcolor{Ckt}{$C_{3,35}$}
is the sum of the min \textcolor{Min}{$M_{3,34}$} and the cost of
the next data point \textcolor{Data}{$\ell_{35}$}. \textbf{Right:}
in the next step, all previously considered changepoints are
pruned (cost \textcolor{Ckt}{$C_{3,35}$}), since the model with a the second
segment ending at data point $t=35$ is always less costly
(\textcolor{MinMore}{$C^{\geq}_{2,35}$}). }
\label{fig:min-envelope}
\end{figure*}
Our proposed Generalized Pruned Dynamic Programming Algorithm (GPDPA)
can be used to solve the PeakSeg problem. The initialization $k=1$ is
the same as in the reduced isotonic regression solver
(Section~\ref{sec:dyn-prog}). The dynamic programming updates for even
$k\in\{2, 4, \dots\}$ are also the same. However, to constrain non-increasing
changes, the updates for odd $k\in\{3, 5, \dots\}$ are
\begin{equation}
C_{k,t}(\mu) = \ell(y_t, \mu) + \min\{
C_{k-1,t-1}^\geq(\mu),\, C_{k,t-1}(\mu)
\},
\end{equation}
where the min-more operator is defined for any function $f:\mathbb R\rightarrow\mathbb R$ as
$f^\geq(\mu) = \min_{x\geq \mu} f(x)$. Figure~\ref{fig:min-envelope}
shows the geometric interpretation of the min-more operator, along
with an example of how the $\min\{\}$ operation performs pruning.
We implemented this algorithm using the Poisson loss
$\ell(y, \mu) = \mu - y\log \mu$, since our application in
Section~\ref{sec:results-chip-seq} is on count data
$y\in\mathbb Z_+ = \{0, 1, 2, \dots\}$.
We implemented this algorithm in C++, and our free/open-source code is
available as the PeakSegPDPA function in the coseg R package for
constrained optimal segmentation
(\url{https://github.com/tdhock/coseg}). Implementation details can be
found in Section~\ref{sec:pseudocode}.
\subsection{General affine inequality constraints
between adjacent segment means}
\label{sec:general}
In this section we briefly discuss how our proposed Generalized Pruned
Dynamic Programming Algorithm (GPDPA) can be used to solve any
optimization problem with affine inequality constraints
between adjacent segment means. For each change $k\in\{1,\dots,K-1\}$,
let $a_k,b_k,c_k\in\mathbb R$ be arbitrary coefficients that define affine
functions $g_k(u_k, u_{k+1})=a_k u_k + b_k u_{k+1} + c_k$. The
changepoint detection problem with general affine constraints is
\begin{align}
\label{eq:min_general_affine_inequality}
\minimize_{
\substack{
\mathbf u\in\mathbb R^K\\
0=t_0<t_1<\cdots<t_{K-1}<t_K=n
}
} &\ \
\sum_{k=1}^K h_{t_{k-1}, t_k}(u_k)\\
\text{subject to \hskip 0.9cm} &\ \ \forall k\in\{1,\dots,K-1\},
\nonumber\\
&\ \ g_k(u_k, u_{k+1})\leq 0.\nonumber
\nonumber
\end{align}
Some examples of models that are special cases:
\begin{enumerate}
\item If we take all $a_k,b_k,c_k=0$ then the constraints are
trivially satisfied, we
recover the unconstrained segment neighborhood problem
(\ref{eq:optimal_segment_neighborhood}).
\item If we take all $a_{k} =1$, $b_{k}=-1$ and $c_{k} = 0$ we recover
the reduced isotonic regression problem
(\ref{eq:isotonic_ut}).
\item For the PeakSeg problem (\ref{eq:PeakSeg}),
we take all $c_{k} = 0$. For odd $k\in\{1,3,\dots\}$ we take
$a_{k} =1$, $b_{k}=-1$ and for even $k\in\{2,4,\dots\}$ we take
$a_{k} =-1$, $b_{k}=1$.
\end{enumerate}
To solve these problems, we need to compute the analog of the
min-less/more operator, which we call the constrained minimization
operator. For any cost function $f:\mathbb R\rightarrow\mathbb R$ and constraint
function $g:\mathbb R\times\mathbb R\rightarrow\mathbb R$, we define the constrained
minimization operator $f^g:\mathbb R\rightarrow\mathbb R$ as
\begin{equation}
\label{eq:constrained-min-operator}
f^g(u_{k}) = \min_{u_{k-1} : g(u_{k-1}, u_{k})\leq 0} f(u_{k-1}).
\end{equation}
When $g$ is affine, the constrained minimization operator is either
non-decreasing or non-increasing. In this case it can be computed
using a simple algorithm that scans the piecewise function $f$ either
from left to right or right to left. When a local minimum is found,
its value is recorded, and a constant function piece is added (for
details see pseudocode for MinLess algorithm in
Section~\ref{sec:MinLess}). The constrained minimization operator is
used in the following general dynamic programming update rule which
can be used to compute the solution to
(\ref{eq:min_general_affine_inequality})
\begin{equation}
\label{eq:general_dp}
C_{k,t}(\mu) = \ell(y_t,\mu) + \min\{
C_{k,t-1}(\mu),\,
C_{k-1,t-1}^{g_{k-1}}(\mu)
\}.
\end{equation}
We note that this update rule is valid for constraint functions $g$
more general than affine functions. However, the
closed-form computation of the constrained minimization operator
(\ref{eq:constrained-min-operator}) would possibly be much more
difficult for these more general constraint functions (e.g. quadratic
constraint functions).
\section{Results on peak detection in ChIP-seq data}
\label{sec:results-chip-seq}
\label{sec:results}
The real data analysis problem that motivates this work is the
detection of peaks in ChIP-seq data \citep{practical}, which are
typically represented as a vector of non-negative counts
$\mathbf y\in\mathbb Z_+^n$ of aligned sequence reads for $n$ continguous
bases in a genome. Data sizes are between $n=10^5$ (maximum of the
benchmark we consider) and $n=10^8$ (largest region with no gaps in
the human genome hg19). A peak detector can be represented
as a function $c(\mathbf y)\in\{0,1\}^n$ for binary classification at
every base position. The positive class is peaks (genomic regions with
large values, representing protein binding or modification) and the
negative class is background noise (small values).
In the supervised learning framework of \citet{HOCKING2016-chipseq}, a
data set consists of $m$ count data vectors
$\mathbf y_1,\dots,\mathbf y_m$ along with labels $L_1,\dots, L_m$
that identify regions with and without peaks. Briefly, the number of
errors $E[c(\mathbf y_i), L_i]$ is the total of false positives
(negative labels with a predicted peak) plus false negatives (positive
labels with no predicted peak). The benchmark consists of seven
histone ChIP-seq data sets, each with a different peak pattern
(experiment type, labeler, cell types). The goal in each data set is
to learn the pattern encoded in the labels, and find a classifier $c$
that minimizes the total number of incorrectly predicted labels in a
held-out test set:
\begin{equation}
\label{eq:learn}
\minimize_c
\sum_{i=1}^m
E\left[
c(\mathbf y_i), L_i
\right].
\end{equation}
\citet{HOCKING-PeakSeg} proposed a constrained dynamic programming
algorithm (CDPA) to approximately compute the optimal changepoints,
subject to the PeakSeg up-down constraint
(Section~\ref{sec:PeakSeg}). The CDPA has been shown to achieve
state-of-the-art peak detection accuracy, by classifying even-numbered
segments $k$ as peaks, and odd-numbered segments $k$ as background
noise. However, its quadratic $O(Kn^2)$ time complexity makes it too
slow to run on large ChIP-seq data sets.
\begin{figure*}[t!]
\centering
\parbox{0.49\textwidth}{
\input{figure-PDPA-intervals-log-log}
}
\parbox{0.49\textwidth}{
\input{figure-PDPA-timings-log-log}
}
\vskip -0.8cm
\caption{Empirical speed analysis on 2752 count data vectors from
the histone mark ChIP-seq benchmark. For each vector we ran the
GPDPA with the up-down constraint and a max of $K=19$
segments. The expected time complexity is $O(KnI)$ where $I$ is
the average number of intervals (function pieces; candidate
changepoints) stored in the $C_{k,t}$ cost
functions. \textbf{Left}: number of intervals stored is
$I=O(\log n)$ (median, inter-quartile range, and maximum over all
data points $t$ and segments $k$). \textbf{Right}: time
complexity of the GPDPA is $O(n\log n)$ (median line and min/max
band).}
\label{fig:timings}
\end{figure*}
In this section, we show that our proposed GPDPA can be used to
overcome this speed drawback, while maintaining state-of-the-art
accuracy. To show the importance of enforcing the up-down constraint,
we consider the unconstrained Pruned Dynamic Programming Algorithm
(PDPA) of \citet{pruned-dp} as a baseline
(Table~\ref{tab:contribution}). We also compare against two popular
heuristics from the bioinformatics literature, in order to demonstrate
that constrained optimization algorithms such as the CDPA and GPDPA
are more accurate.
\subsection{Empirical time complexity in ChIP-seq data}
\label{sec:results_time}
The ChIP-seq benchmark consists of seven labeled histone data
sets.
Overall there are 2752 count data vectors $\mathbf y_i$ to segment,
varying in size from $n=87$ to $n=263169$ data. For each count data
vector $\mathbf y_i$, we ran each algorithm (CDPA, PDPA, GDPDA) with a
maximum of $K=19$ segments. This implies a maximum of 9 peaks (one for
each even-numbered segment), which is more than enough in these
relatively small data sets. To analyze the empirical time complexity,
we recorded the number of intervals stored in the $C_{k,t}$ cost
functions (Section~\ref{sec:algorithms}), as well as the computation
time in seconds.
As in the PDPA, the time complexity of our proposed GPDPA is
$O(K n I)$, which depends on the number of intervals $I$ (candidate
changepoints) stored in the $C_{k,t}$ cost functions
\citep{pruned-dp-new}. We observed that the number of intervals stored
by the GPDPA increases as a sub-linear function of the number of data
points $n$ (left of Figure~\ref{fig:timings}). For the largest data
set ($n=263169$), the algorithm only stored median=16 and maximum=43
intervals. The most intervals stored was 253 for one data set with
$n=7776$. These results suggest that our proposed GPDPA only stores on
average $O(\log n)$ intervals (possible changepoints), as in the
original PDPA. The overall empirical time complexity is thus
$O(K n \log n)$ for $K$ segments and $n$ data points.
We recorded the timings of each algorithm for computing models with up
to $K=19$ segments (a total of 10 peak models $k\in\{1,3,\dots,19\}$,
from 0 to 9 peaks). Since $K$ is constant, the expected time
complexity was $O(n^2)$ for the CDPA and $O(n \log n)$ for the PDPA
and GPDPA. In agreement with these expectations, our proposed GPDPA
shows $O(n\log n)$ asymptotic timings similar to the PDPA (right of
Figure~\ref{fig:timings}).
It is clear that the $O(n^2)$ CDPA algorithm is slower than the other
two algorithms, especially for larger data sets. For the largest count
data vector ($n=263169$), the CDPA took over two hours, but the GPDPA
took only
about two minutes. Our proposed GPDPA is nearly as fast as MACS
\citep{MACS}, a heuristic from the bioinformatics literature which
took about 1 minute to compute 10 peak models for this data set.
The total computation time to process all 2752 count data vectors was
156 hours for the CDPA, and only 6 hours for the GPDPA (26 times
faster). Overall, these results suggest that our proposed GPDPA enjoys
$O(n\log n)$ time complexity in ChIP-seq data, which makes it possible
to use for very large data sets.
\subsection{Test accuracy in ChIP-seq data}
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{figure-test-error-dots}
\vskip -0.5cm
\caption{Four-fold cross-validation
was used to estimate peak detection accuracy.
Each panel shows one of seven ChIP-seq data sets,
labeled by experiment (Broad H3K36me3),
labeler (AM), and cell types (immune).
Each black circle shows test AUC in one of four
cross-validation folds, the shaded grey circle is the mean, and
the vertical line is the maximum mean in each data set. It is
clear that the proposed GPDPA is
just as accurate as the previous state-of-the-art CDPA, and both are
more accurate than the other baseline methods.
}
\label{fig:test-error-dots}
\end{figure*}
For the optimal changepoint detection algorithms (CDPA, PDPA, GPDPA),
the prediction problem simplifies to selecting the number of segments
$K_i\in \{1, 3,\dots, 19\}$ for each data vector $i$, resulting in a
predicted peak vector $c^{K_i}(\mathbf y_i)\in\{0,1\}^n$. We select the
number of segments using an oracle penalty
$K_i^\lambda=\argmin_k l_{ik} + \lambda o_{ik}$
\citep{cleynen2013segmentation}, where $l_{ik}$ is the Poisson loss and
$o_{ik}$ is the oracle model complexity for the model with $k$
segments for data vector $i$.
The problem thus simplifies to learning a scalar
penalty constant $\lambda$,
\begin{equation}
\label{eq:learn-lambda}
\minimize_{\lambda}
\sum_{i=1}^m E\left[
c^{K_i^\lambda}(\mathbf y_i),
L_i\right].
\end{equation}
To demonstrate that changepoint detection algorithms are more accurate
than typical heuristics from the bioinformatics literature, we also
compared with the MACS and HMCanBroad methods \citep{MACS,
HMCan}. MACS is a popular heuristic for data with a sharp peak
pattern such as H3K4me3, and \mbox{HMCanBroad} is a popular heuristic
for data with a broad peak pattern such as H3K36me3. Although they are
not designed for supervised learning, we trained them by performing
grid search over a single significance threshold parameter (qvalue for
MACS and finalThreshold for HMCanBroad).
In each of the seven data sets in the histone benchmark,
we performed four-fold cross-validation and computed test AUC (area
under the Receiver Operating Characteristic curve) to estimate the
accuracy of each algorithm. The previous algorithm with
state-of-the-art accuracy on this benchmark was the CDPA, which
enforces the up-down constraint on segment means. We expected our
proposed GPDPA to perform just as well, since it also enforces that
constraint. In agreement with our expectation, we observed that the
CDPA and GPDPA yield comparable test AUC in all seven data sets
(Figure~\ref{fig:test-error-dots}). In contrast, the unconstrained
PDPA had much lower test AUC in several data sets, because of lower
true positive rates. These results provide convincing evidence that
the constraint is necessary for optimal peak detection accuracy.
Since the baseline HMCanBroad algorithm was designed for data with a
broad peak pattern, we expected it to perform well in the H3K36me3
data. In agreement with this expectation, HMCanBroad showed
state-of-the-art test AUC in two H3K36me3 data sets (broad peak
pattern), but was very inaccurate in four H3K4me3 data sets (sharp
peak pattern). We expected the baseline MACS algorithm to perform well
in the H3K4me3 data sets, since it was designed for data with a sharp
peak pattern. In contrast to this expectation, MACS had test AUC
values much lower than the optimization-based algorithms in all seven
data sets (Figure~\ref{fig:test-error-dots}). These results suggest
that constrained optimal changepoint detection algorithms are more
accurate than the heuristics from the bioinformatics literature.
\section{Discussion and conclusions}
\label{sec:discussion}
Algorithms for changepoint detection can be classified in terms of
time complexity, optimality, constraints, and pruning techniques
(Table~1). In this paper, we investigated generalizing the functional
pruning technique originally discovered by \citet{pruned-dp} and
\citet{phd-johnson}. We showed that the functional pruning technique can
be used to compute optimal changepoints subject to affine constraints
on adjacent segment mean parameters.
We showed that our proposed Generalized Pruned Dynamic Programming
Algorithm (GPDPA) enjoys the same log-linear $O(Kn\log n)$ time
complexity as the original unconstrained PDPA, when applied to peak
detection in ChIP-seq data sets (Figure~\ref{fig:timings}). However,
we observed that the up-down constrained GPDPA is much more accurate
than the unconstrained PDPA (Figure~\ref{fig:test-error-dots}). These
results suggest that the up-down constraint is necessary for computing
a changepoint model with optimal peak detection accuracy. Indeed, we
observed that the GPDPA enjoys the same state-of-the-art accuracy as
the previous best, the relatively slow quadratic $O(Kn^2)$ time
CDPA.
We observed that the heuristic algorithms which are popular in the
bioinformatics literature (MACS, HMCanBroad) are much less accurate
than the optimal changepoint detection algorithms (CDPA, PDPA,
GPDPA). In the past these sub-optimal heuristics have been preferred
because of their speed. For example, the CDPA took 2 hours to compute
10 peak models in the largest data set in the ChIP-seq benchmark,
whereas the GPDPA took 2 minutes, and the MACS heuristic took 1
minute. Using our proposed GPDPA, it is now possible to compute highly
accurate models in an amount of time that is comparable to heuristic
algorithms. Our proposed GPDPA can now be used as an optimal
alternative to heuristic algorithms, even for large data sets.
For future work we will be interested in exploring pruning techniques
for other constrained changepoint models. When the number of expected
changepoints grows with the number of data points, then $K=O(n)$ and
our proposed GPDPA has $O(n^2 \log n)$ average time complexity (since
it computes all models with $1,\dots,K$ segments). We have already
started modifying the GPDPA for optimal partitioning
\citep{optimal-partitioning}, which results in the Generalized
Functional Prunining Optimal Partitioning (GFPOP) algorithm
(Section~\ref{sec:GFPOP}). It computes the $K$-segment model for a
single penalty constant $\lambda$ (without computing models with
$1,\dots,K-1$ segments) in $O(n\log n)$ time.
\section{Reproducible Research Statement}
The source code and data used to create this manuscript (including all
figures) is available at
\url{https://github.com/tdhock/PeakSegFPOP-paper}
\section{Acknowledgements}
This work was supported by a Discovery Frontiers project grant, ``The
Cancer Genome Collaboratory,'' jointly sponsored by the Natural
Sciences and Engineering Research Council (NSERC), Genome Canada (GC),
the Canadian Institutes of Health Research (CIHR) and the Canada
Foundation for Innovation (CFI).
\newpage
|
\section{Introduction}
Type~IIP supernovae (SNe) are the most common type of core-collapse SNe \citep[e.g.,][]{li2011lossii} and they are known to be explosions of red supergiants (RSGs, \citealt{smartt2015review} for a review). Recent large-field and high-cadence transient surveys are starting to catch SNe~IIP within a day after their explosions \citep[e.g.,][]{yaron2017iipcsm,garnavich2016keplerbreakout,quimby2007earlyiip}. These early photometric and spectroscopic properties of SNe~IIP often do not agree with those predicted theoretically. For example, rise times of SNe~IIP are often faster than predicted \citep[e.g.,][]{gonzalez2015iiprise,gall2015earlyiip,rubin2016manyearlyiip} and early SN~IIP spectra also often show narrow lines \citep[e.g.,][]{khazov2016earlyiispectra,smith2015iiniip} as observed in SNe~IIn \citep{schlegel1990iin}. One way to resolve this discrepancy is to assume that there are dense circumstellar media (CSM) surrounding RSG SN progenitors \citep[e.g.,][]{moriya2011iipcsm,nagy2016iipbump,morozova2016iil}.
Recently, SN~2013fs provided a clear case of a RSG SN progenitor surrounded by dense CSM at the immediate vicinity of the progenitor \citep{yaron2017iipcsm}. SN~2013fs was caught within a few hours after the explosion and its first spectrum was taken in about 6 hours after the explosion. It had a flash spectroscopic feature with narrow lines indicating the existence of dense CSM \citep[e.g.,][]{gal-yam2014flash,groh2014flash,grafener2016flashmodel}. However, the narrow lines disappeared in several days and the spectra changed to those of normal SNe~IIP. Light-curve (LC) modeling by \citet{morozova2016iil} also suggested that a dense CSM located at the immediate vicinity of the progenitor is required to explain the early LC properties of SN~2013fs.
The high CSM density in the immediate vicinity of the SN progenitors has been suggested to be caused by an increase of the progenitors' mass-loss rates several years before their explosions.\footnote{Alternatively, the matter ejected by a wind can be confined by a strong external radiation field to make dense CSM \citep{mackey2014pico}.} In the canonical stellar evolution theory, no significant increase in the mass-loss rates in several years before the SN explosions is expected \citep{langer2012review} and several mechanisms to explain the possible mass-loss enhancement has been suggested \citep[e.g.,][]{quataert2012wavemassloss,moriya2014neutrino,heger1997superwind,yoon2010cantiello,moriya2015langer}. A precise estimate of the mass-loss rate shortly before the explosion is essential in understanding the unknown mechanisms of the pre-SN mass loss. In this Letter, we show that the effect of wind acceleration on the CSM density structure has a great impact on early LCs of SNe~IIP. A wind launched at the stellar surface is gradually accelerated to reach the terminal velocity. Therefore, the wind velocity just above the stellar surface is lower than the terminal velocity. This leads to a higher density of the CSM close to the progenitor than in the case where a constant wind velocity is assumed for a given mass-loss rate. Here, we show that SN progenitors' mass-loss rates shortly before their explosions can be significantly overestimated if the effect of wind acceleration is not properly taken into account by taking SN~2013fs as an example.
\section{Progenitor system}\label{sec:progenitor}
\subsection{Progenitor model}
As our focus in this Letter is on SNe~IIP, we adopt a RSG progenitor model obtained with the public stellar evolution code \texttt{MESA} \citep{paxton2011mesa1,paxton2013mesa2,paxton2015mesa3}. The progenitor has a zero-age main-sequence mass of 12~\ensuremath{\mathrm{M}_\odot}\ with the solar metallicity. The Ledoux criterion for convection is used with a mixing-length parameter of 2.0 and a semi-convection parameter of 0.01. Overshooting is applied on top of the hydrogen burning convective core with a step function. The adopted overshoot parameter is $0.3~H_P$, where $H_P$ is the pressure scale height at the outer boundary of the convective core. The standard `Dutch' wind scheme is used for both hot and cold winds, with a scaling factor of 1.0. The model has been evolved to the stage of oxygen burning in the core, from which the hydrogen enveolpe structure hardly changes until the core collapse.
The RSG progenitor has 10.3~\ensuremath{\mathrm{M}_\odot}\ with the hydrogen-rich envelope of 6.1~\ensuremath{\mathrm{M}_\odot}. Its radius is $R_\star=607~\ensuremath{\mathrm{R}_\odot}$ and the effective temperature is 3500~K. The final mass-loss rate of the progenitor according to the `Dutch' wind prescription is $1.4\times 10^{-6}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$. However, as explained below, we assume that the wind mass-loss rate is enhanced shortly before the SN explosion such that a dense CSM is formed while the progenitor structure is not significantly affected.
\begin{table}
\centering
\caption{Wind properties.}
\label{tab:wind}
\begin{tabular}{ccccc}
\hline
\ensuremath{\dot{M}} & $v_\infty$ & $v_0$ & $\beta$ & CSM radius \\
$\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ & \ensuremath{\mathrm{km~s^{-1}}} & $\mathrm{m~s^{-1}}$ & & cm \\
\hline
$10^{-3}$ & 10 & 10 & 5 & $10^{15}$\\
$10^{-4}$ & 10 & 1 & 5 & $10^{15}$ \\
0.15$^a$ & 10 & - & 0 & $1.3\times 10^{14}$ \\
\hline
\multicolumn{5}{l}{$^a$ A constant $\ensuremath{v_\mathrm{wind}}$ model by \citet{morozova2016iil}.}\\
\end{tabular}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth]{density.eps}
\caption{
Density structure of CSM. Our CSM model has a constant mass-loss rate but the $\beta$-law wind velocity ($\beta=5$). We also show a dashed line where a constant mass-loss rate $(10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}})$ and a constant wind velocity $(10~\ensuremath{\mathrm{km~s^{-1}}})$ are assumed.
The CSM density of SN~2013fs estimated by \citet{yaron2017iipcsm} from spectral modeling is shown with a square.
The CSM density of SN~2013fs estimated by \citet{morozova2016iil} from early LC modeling is also presented.
}
\label{fig:csm}
\end{figure}
\subsection{CSM}
CSM density (\ensuremath{\rho_\mathrm{CSM}}) is determined by the progenitor's mass-loss rate (\ensuremath{\dot{M}}) and wind velocity (\ensuremath{v_\mathrm{wind}}) as
\begin{equation}
\ensuremath{\rho_\mathrm{CSM}} (r) = \frac{\ensuremath{\dot{M}}}{4\pi \ensuremath{v_\mathrm{wind}} (r)}r^{-2}.
\end{equation}
In this work, we adopt a $\beta$-law velocity profile as
\begin{equation}
\ensuremath{v_\mathrm{wind}} (r) = v_0 + (v_\infty - v_0) \left( 1 - \frac{R_0}{r} \right)^\beta, \label{eq:betavel}
\end{equation}
where $v_0$ is the initial wind velocity at the stellar surface, $v_\infty$ is the final wind velocity, and $R_0$ is where the wind is launched. We find that $\beta \simeq 5$ provides the best fit to the SN~2013fs LC and we focus on this $\beta$ in this Letter.
Observations \citep[e.g.,][]{bennett2010rsgwind,marshall2004agbwindacc} indicate that RSG winds have larger values of $\beta$, i.e., slower wind acceleration, than OB supergiants that have $\beta\simeq 0.5-1$ \citep[e.g.,][]{groenewegen1989ostarmassloss,haser1995ostarmassloss,puls1996ostarmassloss}. For example, the $\beta$-law velocity profile with $\beta\simeq 3.5$ is found to match a RSG $\zeta$ Aurigae \citep{baade1996wind}. We also set $R_0=R_\star$ in this work assuming that the wind is launched at the stellar surface. $R_0$ can vary depending on the wind acceleration mechanisms \citep[e.g.,][]{bennett2010rsgwind,lamers1999windbook}. For example, $R_0$ can be $2-3~R_\star$ in the dust-driven winds. We found that models with $R_0 = R_\star$ match SN~2013fs and we focus on this case in this Letter. We investigate the effects of different $\beta$ and $R_0$ in our future work.
We present LCs for two different mass-loss rates: $10^{-3}$ and $10^{-4}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ (Table~\ref{tab:wind} and Fig.~\ref{fig:csm}). We set the terminal wind velocity $v_\infty=10~\ensuremath{\mathrm{km~s^{-1}}}$. The initial velocity $v_0$ is chosen so that the density structure at the stellar surface is smoothly connected. We cut our dense CSM at an arbitrary radius of $10^{15}~\mathrm{cm}$.
Fig.~\ref{fig:csm} shows the density structure of our models. We also present the density structure that gives the best fit to the early LCs of SN~2013fs assuming a constant \ensuremath{\dot{M}}\ and \ensuremath{v_\mathrm{wind}}\ \citep{morozova2016iil}. Our CSM have similar density to those found by \citet{morozova2016iil} near the progenitor, but our mass-loss rates are only $10^{-3}-10^{-4}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ while $0.15~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ is required when the constant wind velocity of 10~\ensuremath{\mathrm{km~s^{-1}}}\ is assumed. The CSM mass in our $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ model is $0.5~\ensuremath{\mathrm{M}_\odot}$ while that in the model of \citet{morozova2016iil} is $0.4~\ensuremath{\mathrm{M}_\odot}$.
\begin{figure*}
\includegraphics[width=\columnwidth]{y1258e10.eps}
\includegraphics[width=\columnwidth]{y1264e10.eps}
\caption{
Synthetic LCs from our RSG progenitor with an explosion energy of $10^{51}~\mathrm{erg}$ and LCs of SN~2013fs. The left panel shows our model with $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ and the right shows that with $10^{-4}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$. The LC models with dashed lines are those without CSM.
}
\label{fig:lcs_wcsm}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{y1258e13.eps}
\caption{Synthetic LCs from our $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ model with the explosion energy of $1.3\times 10^{51}~\mathrm{erg}$.}
\label{fig:lcs_highe}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{y1202e10.eps}
\caption{Synthetic LCs from a $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ model with a constant wind velocity of $10~\ensuremath{\mathrm{km~s^{-1}}}$ without the wind acceleration. An explosion energy of $10^{51}~\mathrm{erg}$ is adopted.}
\label{fig:1e-3const}
\end{figure}
\section{Light curves}\label{sec:lcs}
\subsection{Methods}
We use a one-dimensional multi-group radiation hydrodynamics code \texttt{STELLA} to investigate LCs numerically \citep[e.g.,][]{blinnikov1998sn1993j,blinnikov2000sn1987a,blinnikov2006sniadeflg,chugai2002windacc}. Starting from the progenitor models presented in the previous section, we initiate explosions by putting thermal energy just below the mass cut that we set at $1.4~\ensuremath{\mathrm{M}_\odot}$. We put $0.1~\ensuremath{\mathrm{M}_\odot}$ of \ensuremath{^{56}\mathrm{Ni}}\ at the center, but it does not affect the early LCs we present here.
\texttt{STELLA} evaluates spectral energy distributions (SEDs) at each time step. We obtain multicolor LCs by convolving filter transmission functions to the SEDs. Here, we adopt the $U$ band filter of Swift/UVOT \citep{poole2008swiftuvot} and the $g$, $r$, and $i$ filters of the Sloan Digital Sky Survey (SDSS, \citealt{doi2010sdss}). To compare with SN~2013fs, we apply a redshift correction with $z=0.0119$ and a Galactic extinction of $E(B-V)=0.0346$ assuming the extinction curve of \citet{cardelli1989exti} with $R_V =3.1$ to our theoretical SEDs.
The multicolor LCs of SN~2013fs, with which we compare our models, are presented by \citet{valenti2016typeii} and \citet{yaron2017iipcsm}. We acquire machine-readable data through the Open Supernova Catalog \citep{guillochon2017osc}. We take the $U$, $g$, $R$, and $i$-band data for comparison to cover a wide spectral range. Although we use the SDSS $r$-band filter to obtain multicolor LCs from our theoretical SEDs and compare them with the $R$-band observations, the differences between the observed $R$-band and the $r$-band magnitudes are negligibly small \citep{yaron2017iipcsm}.
\subsection{SN~2013fs}
Fig.~\ref{fig:lcs_wcsm} shows LCs obtained from our CSM configuration. We adopt the explosion energy of $10^{51}~\mathrm{erg}$ to match the luminosity during the later plateau phase. The dashed lines in the figure are LC models without CSM. The existence of the dense CSM can explain the early LCs of SN~2013fs as presented by \citet{morozova2016iil}.
We find that our LC model with the CSM having $\ensuremath{\dot{M}}=10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ provides a LC that matches SN~2013fs reasonably well. The LC model with $\ensuremath{\dot{M}}=10^{-4}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ also shows a fast rise but it does not become as bright as SN~2013fs. Although the $U$ and $g$-band LCs of SN~2013fs are well reproduced by the $\ensuremath{\dot{M}}=10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ model, it is slightly fainter than SN~2013fs in the $r$ and $i$ bands by $\sim 0.4~\mathrm{mag}$. The $r$ and $i$-band brightness can be matched to SN~2013fs by increasing the explosion energy to $1.3\times 10^{51}~\mathrm{erg}$, but then the $U$ and $g$-band magnitudes becomes brighter by $\sim 0.4~\mathrm{mag}$. This slight difference may be caused by, e.g., incompleteness of opacity information we adopt or higher metallicity of the progenitor of SN~2013fs. LCs from a constant mass-loss rate of $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ and a constant wind velocity of 10~\ensuremath{\mathrm{km~s^{-1}}}\ without the wind acceleration do not become bright enough to explain the early LCs (Fig.~\ref{fig:1e-3const}).
A fundamental difference between the model of \citet{morozova2016iil} and ours is in the mass-loss rates. \citet{morozova2016iil} find that an extremely high mass-loss rate ($\gtrsim 0.1~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$) is required to explain the early LC of SN~2013fs by assuming a constant wind velocity of 10~\ensuremath{\mathrm{km~s^{-1}}}. However, we find that a CSM from $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ with the terminal velocity of 10~\ensuremath{\mathrm{km~s^{-1}}}\ can provide a reasonable fit to the early LC if we take the velocity change due to the wind acceleration into account. Although the mass-loss rate and wind velocity are different in the two models, the final CSM mass is comparable in them ($0.5~\ensuremath{\mathrm{M}_\odot}$ in our $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ model and $0.4~\ensuremath{\mathrm{M}_\odot}$ in \citealt{morozova2016iil}). Therefore, the CSM mass is likely better constrained by the early LCs.
\citet{yaron2017iipcsm} estimate the CSM density at the immediate vicinity of the progenitor of SN~2013fs by modeling their flash spectra. We show their constraint on the CSM density with a square in Fig.~\ref{fig:csm}. Our model with $10^{-4}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ is consistent with their estimates but it does not result in a LC that is bright enough to explain the early LC. Our best LC model with $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ has higher density by a factor of 10 and the model obtained by \citet{morozova2016iil} has even higher density. \citet{yaron2017iipcsm} may have underestimated $\dot{M}$ by several factors because of the neglect of light-travel-time effects in their study \citep{grafener2016flashmodel}. Together with the altered density structure, this may bring our results in better agreement with their density estimates.
\section{Discussion and conclusions}
\citet{yaron2017iipcsm} interpret that the high CSM density needed to explain the early SN properties of SN~2013fs is caused by the increase of the progenitor's mass-loss rate shortly before the explosion. The early phase LC modeling of \citet{morozova2016iil} indicated that the progenitor's mass-loss rate may have been as high as $0.15~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ assuming a constant wind velocity of 10~\ensuremath{\mathrm{km~s^{-1}}}. However, we have shown that only $\sim 10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ is sufficient to explain the early LC if we consider wind acceleration. The inferred mass-loss rate of $\sim 10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ is still very high compared to those of observed RSGs \citep[e.g.,][]{mauron2011rsgmassloss,goldman2017agbrsgwind}. This high mass-loss rate of $\sim 10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ could not have been maintained for more than $\sim 10^3$~years before the SN explosion, as otherwise most of the hydrogen envelope would have been stripped off. The plateau duration of SN~2013fs ($\simeq 80~\mathrm{days}$) implies that its progenitor had a massive hydrogen-rich envelope.
In addition, it takes 500~years to reach $10^{14}~\mathrm{cm}$ and then only 50 years to reach $10^{15}~\mathrm{cm}$ in our wind acceleration model with $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$. \citet{yaron2017iipcsm} found that the dense CSM only extends up to $\simeq 10^{15}~\mathrm{cm}$ in SN~2013fs. Therefore, the mass-loss enhancement should not last more than about 550 years in order not to have too extended dense CSM. On the contrary, if the mass-loss enhancement lasts less than $\sim 100~\mathrm{years}$, the dense CSM does not extend enough to affect the early SN properties. Therefore, the progenitor must have undergone an abrupt increase of mass loss starting at around 500~years before the explosion in our model.
As the estimated CSM mass in SN~2013fs is similar ($\simeq 0.5~\ensuremath{\mathrm{M}_\odot}$) in the constant wind velocity model and the wind acceleration model, the difference in the estimated mass-loss rates makes a significant difference in the estimates for the period of the mass-loss enhancement before the explosion. In our model with $\sim 10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$, the mass-loss enhancement is estimated to occur in the final 500~years to the explosion.
If we assume $\sim 0.1~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$, the mass-loss enhancement should only occur in the final several years to the explosion. These estimates are important in constraining possible mass-loss mechanisms like wave-driven mass loss \citep[e.g.,][]{shiode2014wavemassloss}.
It is possible that the dense CSM formed by the mass-loss enhancement affects not only the early SN properties but also observational properties of underlying RSG SN progenitors. The effects on the observational properties of the progenitors strongly depend on the opacity in the dense CSM. If the CSM has a temperature similar to the RSG photosphere, the opacity is expected to be $\sim 10^{-3}~\mathrm{cm^2~g^{-1}}$ at most \citep{ferguson2005lowtempopacity} and it does not affect the RSG properties significantly. However, the unknown mechanisms of the mass-loss enhancement can strongly affect the opacity in the CSM. For example, if the opacity in the dense CSM with $10^{-3}~\ensuremath{\mathrm{M}_\odot}~\mathrm{yr^{-1}}$ becomes $0.01~\mathrm{cm^2~g^{-1}}$, the photosphere locates at $4~R_\star$ and the effective temperature is altered to about 1800~K.
Moreover, if the mass-loss enhancement leads to a formation of a large amount of dusts, the progenitor would be enshrouded by dusts and significantly reddened. Then, if the mass-loss enhancement starts $\sim 100$~years before the explosion in some RSG SN progenitors as we suggest, some of detected RSG SN progenitors so far may have been significantly affected by absorption in the CSM and their progenitor masses may have been estimated lower than they are, for example \citep[e.g.,][]{walmswell2012dustrsgprobsol,beasor2016dustyrsg}.
The observational data of early LCs of SNe~IIP have dramatically increased during the last several years. They give evidence that not a small fraction of SNe~IIP progenitors may be surrounded by immediate dense CSM \citep[e.g.,][]{gonzalez2015iiprise}, like in the case of SN~2013fs. The present study indicates that the effect of wind acceleration should not be ignored in the analysis of early SN~IIP LCs. Addressing this issue with more details including spectral properties is required to better understand the mass-loss history of their progenitors. We also note that the wind acceleration itself is not known well. For example, the shape of the early SN LCs may be affected by the CSM structure and early SN observations may also be useful to constrain wind acceleration such as $\beta$.
\section*{Acknowledgements}
We thank the referee, James Fuller, for constructive comments that improved this work.
TJM thanks the Yukawa Institute for Theoretical Physics at Kyoto University, where this work was initiated during the YITP-T-16-05 on "Transient Universe in the Big Survey Era: Understanding the Nature of Astrophysical Explosive Phenomena". TJM is supported by the Grant-in-Aid for Research Activity Start-up of the Japan Society for the Promotion of Science (16H07413). SCY is supported by the Korea Astronomy and Space Science Institute under the R\&D program (Project No. 3348- 20160002) supervised by the Ministry of Science, ICT and Future Planning. GG is supported by the Deutsche Forschungsgemeinschaft, Grant No. GR 1717/5. Grant no. IZ73Z0 152485 SCOPES Swiss National Science Foundation supports work of SIB. Numerical computations were partially carried out on PC cluster at Center for Computational Astrophysics, National Astronomical Observatory of Japan.
\bibliographystyle{mnras}
|
\section{Introduction}
Climate change is causing detrimental effects to society's well being as temperatures increase, extreme events become more intense\cite{pachauri2014climate}, and sea levels rise\cite{nicholls2010sea}. Natural resources that society depends on, such as agriculture, freshwater, and coastal systems, are vulnerable to increasing temperatures and more extreme weather events. Similarly transportation systems, energy systems, and urban infrastructure allowing society to function efficiently continue to degrade due to the changing climate. Furthermore, the health and security of human beings, particularly those living in poverty, are vulnerable to extreme weather events with increasing intensity, duration, and frequency~\cite{trenberth2012framing}. Scientists and stakeholders across areas such as ecology, water, and infrastructures, require access to credible and relevant climate data for risk assessment and adaptation planning.
Earth System Models (ESMs) are physics-based numerical models which run on massive supercomputers to project the Earth’s response to changes in atmospheric greenhouse gas (GHG) emissions scenarios. Archived ESM outputs are some of the principal data products used across many disciplines to characterize the likely impacts and uncertainties of climate change~\cite{taylor2012overview}. These models encode physics into dynamical systems coupling atmospheric, land, and ocean effects. ESMs provide a large number of climate variables, such as temperature, precipitation, wind, humidity, and pressure, for scientists to study and evaluate impacts. The computationally demanding nature of ESMs limits spatial resolution between 1 and 3 degrees. These resolutions are too course to resolve critical physical processes, such as convection which generates heavy rainfall, or to assess the stakeholder-relevant local impacts of significant changes in the attributes of these processes~\cite{schiermeier2010real}.
Downscaling techniques are used to mitigate the low spatial resolution of ESMs through dynamical and statistical modeling. Dynamical downscaling, also referred to as regional climate models (RCMs), account for local physical processes, such as convective and vegetation schemes, with sub-grid parameters within ESM boundary conditions for high-resolution projections. Like ESMs, RCMs are computationally demanding and are not transferable across regions. In contrast, the statistical downscaling (SD) technique learns a functional form to map ESMs to high resolution projections by incorporating observational data. A variety of statistical and machine learning models, including linear models~\cite{hessami2008automated}, neural networks~\cite{cannon2011quantile}, and support vector machines~\cite{Ghosh2010}, have been applied to SD and is discussed further in section 4 (Related Work). Despite the availability of many techniques, we are not aware of any SD method which explicitly captures spatial dependencies in both low-resolution and high-resolution climate data. Furthermore, traditional methods require observational data at the high-resolution target, meaning that regions with little observation data, often the poorest regions which are most effected by climate change, are unable to receive downscaled climate data needed for adaptation.
The lack of explicit spatial models in SD of ESMs motivated us to study the applicability of computer vision approaches, most often applied to images, to this problem. More specifically, advances in single image super-resolution (SR) correspond well to SD, which learn a mapping between low- and high-resolution images. Moreover, as SR methods attempt to generalize across images, we aim to provide downscaled climate projections to areas without high-resolution observations through what may be thought of as transfer learning. Though we will discuss this topic further in section 4 (Related Work), we found that super-resolution convolutional neural networks were able to capture spatial information in climate data to improve beyond existing methods.
Lastly, we present a framework using our super-resolution approach to downscale ensemble ESMs over the Continental United States (CONUS) at a daily temporal scale for four emission scenarios by using NASA's Earth Exchange (NEX) platform. NEX provides a platform for scientific collaboration, knowledge sharing and research for the Earth science community. As part of NEX, along with many other earth science data products, NASA scientists have already made monthly downscaled ESMs for CONUS up to the year 2100 at 30 arc seconds (NEX-DCP30) that are openly available to the public. However, the downscaling methodology, bias correction spatial disaggregation, has limitations and the monthly scale reduces the applicability to studying extreme events. The improvement of such data products is vital for scientists to study local impacts of climate change to resources society depends on.
\subsection{Key Contributions}
The key contributions are as follows:
\vspace{-1mm}
\begin{itemize}
\item We present DeepSD, an augmented stacked super-resolution convolutional neural network for statistical downscaling of climate and earth system model simulations based on observational and topographical data.
\item DeepSD outperforms a state-of-the-art statistical downscaling method used by the climate and earth science communities as well as a suite of off-the-shelf data mining and machine learning methods, in terms of both predictive performance and scalability.
\item The ability of DeepSD to outperform and generalize beyond grid-by-grid predictions suggests the ability to leverage cross-grid information content in terms of similarity of learning patterns in space, while the ability to model extremes points to the possibility of improved ability beyond matching of quantiles. Taken together, this leads to the new hypothesis that methods may be able to use spatial neighborhood information to predict in regions where data may be sparse or low in quality.
\item For the first time, DeepSD presents an ability to generate, in a scalable manner, downscaled products from model ensembles, specifically, simulations from different climate modeling groups across the world run with different emissions trajectories and initial conditions.
\item DeepSD provides NASA Earth Exchange (NEX) a method of choice to process massive climate and earth system model ensembles to generate downscaled products at high resolutions which can then be disseminated to researchers and stakeholders.
\end{itemize}
\subsection{Organization of the Paper}
The remainder of the paper is organized as follows. Section 2 (Earth Science Data) presents necessary data used for SD along with their associated challenges. Section 3 (Statistical Downscaling) discusses the problem of SD. Section 4 (Related Work) discusses techniques previously applied to SD along with an overview of super-resolution methods. Furthermore, we discuss the relationships between images and climate data. Section 5 (Methodology) presents DeepSD, the augmented stacked super-resolution convolutional neural network formulation. In section 6 (Experiments) we compare our method to another SD technique and three off-the-shelf machine learning approaches and outline the process by which we will scale our method to many climate model simulations. In section 7 (conclusion) we briefly discuss results, limitations, and future work
\section{Earth Science Data}
Earth science data stems from a variety of areas, including climate simulations, remote sensing through satellite observations, and station observations. The spatio-temporal nature of such data causes heavy computational and storage challenges. For instance, a single climate variable at the daily temporal and 4km spatial scales over only the United States requires 1.2GB of storage. Multiplying this effect over a large number of variables, including precipitation, temperature, and wind, globally creates high storage and processing requirements. Furthermore, analysis of these complex datasets require both technical and domain expertise.
ESM outputs, as discussed previously, are one form of earth science data which is crucial to the understanding of our changing climate. The most recent ESMs are a product of the fifth phase of the Coupled Model Intercomparision Project which simulate the climate through a dynamical system coupling effects from the atmosphere, land, and ocean~\cite{taylor2012overview}. However, it is well understood that holes in these models exist, including low-resolution and lack of model agreement, particularly for precipitation~\cite{schiermeier2010real}.
We can harness information in observational datasets in order to learn statistical models mapping ESM outputs to a higher resolutions. Observational datasets are available through a variety of sources, including satellite observations, station observations, and a mixture of both, namely reanalysis datasets. Often, SD models are built to downscale ESMs directly to a observational station while others aim to downscale to a grid based dataset. Gridded observational datasets are often built by aggregating station observations to a defined grid. For example, in our application, we obtain precipitation through the PRISM dataset at a 4km daily spatial resolution which aggregates station observations to a grid with physical and topographical information~\cite{daly2008physiographically}. We then upscale the precipitation data to $1/8^{\circ}$ (~$12.5$ km) as our high-resolution observations. Following, we upscale further to $1^{\circ}$ corresponding to a low-resolution precipitation, as applied in~\cite{pierce2014statistical}. The goal is then to learn a mapping between our low-resolution and high-resolution datasets.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{lr-hr-copy.pdf}
\caption{Prism Observed Precipitation: A) Low resolution at $1.0^{\circ}$ ($\sim 100$km). B) High resolution at $1/8^{\circ}$ ($\sim 12.5$km).}
\vspace{-0.5cm}
\label{fig:lr-hr}
\end{figure}
Furthermore, topography has large effects on weather and climate patterns with lower temperatures, more precipitation, and higher winds~\cite{daly2008physiographically,daly1994statistical}. Taking advantage of the valuable topographical information at different scales, where $1/2^{\circ}$ may capture large scale weather patterns while $1/8^{\circ}$ spatial resolution can capture high-resolution precipitation biases.
Each of the earth science data products discussed inherently possess rich spatial dependencies, much like images. However, traditionally statistical downscaling methods, particularly regression based models, vectorize spatial data, removing this spatial structure. While colored images contain channels consisting of, for example, red, green, and blue, climate data may be represented analogously such that the channels correspond to climate variables and topographical data. Similar approaches have been applied to satellite datasets for image classification~\cite{basu2015deepsat} and resolution enhancement~\cite{zhang2016deep}. Though climate data is more complex than images due to it's dynamics and chaotic nature, we propose that this representation allows scientists to approach the data in an unconventional manner and apply augmented models developed for image processing.
\section{Statistical Downscaling}
SD is the problem of mapping a low resolution climate variable to a high resolution projection. This mapping, which must transform a single grid point to multiple points is an ill-posed problem, one with many possible solutions (see Figure~\ref{fig:lr-hr}). However, we can mitigate the ill-posed problem by including static high-resolution topography data in conjunction with other low-resolution climate variables. We learn the SD model using observed climate datasets and then infer downscaled ESM projections. Spatial and temporal non-stationarity of the changing climate system challenges traditional statistical techniques. Downscaling precipitation further challenges these methods with sparse occurrences and skewed distributions. The combination of an ill-posed problem, uncertainty in the climate system, and data sparsity propagates uncertainty in downscaled climate projections further.
\section{Related Work}
As mentioned previously, SD has a rich and expansive history in the climate community. SD consists of three fundamental categories: regression models and weather classification schemes which improve spatial resolution while weather generators increase temporal resolution (ie. monthly to daily)~\cite{wilby2004guidelines}. As our interest is in increasing spatial resolution we will review regression methods and weather classification.
Regression methods applied to SD are wide in scope, both linear and non-linear, and vary based on the specific climate variable and temporal scale. For instance, downscaling daily precipitation, which we will focus on, relies on a sparse observational dataset where few days contain rainfall while the amount of rainfall in those days follow a skewed distribution. Automated Statistical Downscaling (ASD) presents a traditional framework for this problem where a classification model is first used to classify days with precipitation followed by a regression to estimate the amount~\cite{hessami2008automated}. Similar approaches, among others, include quantile regression neural networks~\cite{cannon2011quantile}, bayesian model averaging\cite{Zhang2015}, and expanded downscaling\cite{burger1996expanded}. Each of these regression models learns a statistical relationship between observed low- and high-resolution pairs and is then applied to ESMs. Another widely used approached is Bias Corrected Spatial Disaggregation (BCSD), which begins by bias correcting a ESM to match the distribution of the high-resolution observed dataset followed by interpolation and spatial scaling to correct for local biases\cite{wood2004hydrologic,thrasher2012technical}. Though BCSD is a simple approach, it has been shown to perform well compared to more complex methods~\cite{Burger2012,maurer2008utility}. Furthermore, we have shown that BCSD performs similarly, or better, when compared to off-the-shelf ASD approaches~\cite{vandal2017intercomparison}.
Weather classification methods take a different approach to statistical downscaling through nearest neighbor estimates, grouping weather events into similar types. Given a set of observed low- and high-resolution pairs, one can compute a distance measure between an ESM and the low-resolution observations to select the nearest high-resolution estimate. Constructed analogues furthers the method by performing a regression on a group of the nearest neighbor estimates\cite{hidalgo2008}. More advances, but similar approaches, have recently been presented, including Hierarchical Bayesian inference models\cite{manor2015bayesian}.
While the approaches discussed above are often sufficient in downscaling means, they tend to fail at downscaling extreme events. For instance, ASD approaches perform reasonably well at downscaling average precipitation~\cite{hessami2008automated} but performs poorly at the extremes~\cite{Burger2012}. As discussed by B\"{u}rger et al.~\cite{Burger2012} and Mannshardt-Shamseldin et al.~\cite{mannshardt2010downscaling}, as well as others, specific approaches to downscaling extremes are often required. These specialized approaches, such as those using Generalized Extreme Value theory, have been developed for this purpose~\cite{mannshardt2010downscaling,hashmi2011comparison}. Ideally, a single approach to downscaling leveraging all available information would capture both averages and extremes, giving the user a more credible dataset.
To our knowledge little work has been attempted to explicitly capture spatial properties for improving downscaled projections. As computer vision approaches are built to exploit the spatial structure of images, we are motivated to understand the applicability of such methods to climate datasets. As introduced previously, we represent climate variables as channels, analogous to images, and model them similarly. However, we note that this presents an analogy and not a direct correspondence.
Using the analogy between climate datasets and images, we can relate statistical downscaling to image super-resolution, where one aims to learn a mapping from low- to high-resolution image pairs. Specifically, single image super-resolution (SR), as the name suggests, increases the resolution of a single image, rather than multiple images, from a scene.
The most successful approaches to SR have been shown to be patch based (or example-based) techniques, achieving state-of-the-art performance~\cite{timofte2014a,dong2014learning,wang2015deep}. Originally proposed by Glasner et al.~\cite{glasner2009super}, patch based methods exploit self-similarity between images to produce exemplar patches. This approach has evolved into different variations of nearest neighbor techniques between low- and high-resolution patches through what is known as dictionary learning~\cite{freeman2002example,chang2004super,yang2010image}. Dictionary learning approaches to SR are analogous to those presented by weather classification SD schemes. Furthermore, approaches including kernel regression~\cite{yang2013fast}, random forests\cite{schulter2015fast}, and anchored neighborhood regression~\cite{timofte2014a}, have been proposed for SR to improve accuracy and speed, all related to methods presented in SD literature~\cite{Ghosh2010,burrows1995cart,stoner2013asynchronous}. Sparse-coding techniques, a form of dictionary learning, have recently shown state-of-the-art results in both speed and accuracy~\cite{timofte2014a}.
Convolutional neural networks were recently presented as a generalization of sparse coding, improving upon past state-of-the-art performances\cite{dong2014learning,wang2015deep}. The sparse coding generalization, non-linearity, network flexibility, and scalability to large datasets presents an opportunity to apply Super Resolution Convolutional Neural Networks to SD\cite{dong2014learning}.
\section{Methodology}
\newcommand{\operatornamewithlimits{argmin}}{\operatornamewithlimits{argmin}}
This section begins by describing and formulating Super-Resolution Convolutional Neural Networks (SRCNN), as presented by~\cite{dong2014learning}. We then introduce a stacked SRCNN architecture such that the output of one SRCNN is the input to the following SRCNN. DeepSD, the adaptation of a stacked SRCNNs to SD, is then introduced.
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{activation-w-filters-basic.eps}
\caption{Augmented SRCNN Architecture. From the left to right: Precipitation and Elevation sub-image pair, filters learned in layer 1, layer 1 activations, layer 2 filters, layer 2 activations, layer 3 filters, and HR precipitation label.}
\vspace{-0.2cm}
\label{fig:filters-activations}
\end{figure*}
\subsection{Super-resolution CNN}
SR methods, given a low-resolution (LR) image, aim to accurately estimate a high-resolution image (HR). As presented by Dong et al.~\cite{dong2014learning}, a CNN architecture can be designed to learn a functional mapping between LR and HR using three operations, patch extraction, non-linear mappings, and reconstruction. The LR input is denoted as $\mathbf{X}$ while the HR label is denoted as $\mathbf{Y}$.
A three layer CNN is then constructed as follows to produce a high resolution estimate and presented in Figure~\ref{fig:filters-activations}. Layer 1 is formulated as $$F_1(\mathbf{X}) = max(0, W_1 * \mathbf{X} + B_1),$$ where `$*$' is the convolution operation and the $max$ operation applies a Rectified Linear Unit~\cite{nair2010rectified} while $W_1$ and $B_1$ are the filters and biases, respectively. $W_1$ consists of $n_1$ filters of size $c \times f_1 \times f_1$. The filter size, $f_1 \times f_1$, operates as an overlapping patch extraction layer where each patch is represented as a high-dimensional vector.
Correspondingly, layer 2 is a non-linear operation such that $$F_2(\mathbf{X}) = max(0, W_2 * F_1(\mathbf{X}) + B_2)$$ where $W_2$ consists of $n_2$ filters of size $n_1 \times f_2 \times f_2$ and $B_2$ is a bias vector. This non-linear operation maps high-dimensional patch-wise vectors to another high-dimensional vector.
A third convolution layer is used to reconstruct a HR estimate such that $$F(\mathbf{X}) = W_3 * F_2(\mathbf{X}) + B_3.$$ Here, $W_3$ contains $1$ filter of size $n_2 \times f_3 \times f_3$. The reconstructed image $F(\mathbf{X})$ is expected to be similar to the HR image, $\mathbf{Y}$
This end-to-end mapping then requires us to learn the parameters $\Theta = \{W_1, W_2, W_3, B_1, B_2, B_3 \}$. A Euclidean loss function with inputs $\{\mathbf{X}_i\}$ and labels $\{\mathbf{Y}_i\}$ is used where the optimization objective is defined as:
\begin{equation} \label{eq:srcnn-loss}
\operatornamewithlimits{argmin}_{\Theta} \sum_{i=1}^n \Vert F(\mathbf{X}_i;\Theta) - \mathbf{Y}_i \Vert_2^2
\end{equation}
such that $n$ is the number of training samples (batch size).
We note that convolutions in layers 1, 2, and 3 decrease the image size depending on the chosen filter sizes, $f_1$,$f_2$, and $f_3$. At test time, padding using the replication method is applied before the convolution operation to ensure the size of the prediction and ground truth correspond. During training, labels are cropped such that $\mathbf{Y}$ and $F(\mathbf{X}_i ; \Theta)$, without padding, are of equal size.
\subsection{Stacked SRCNN}
Traditional SR methods are built for resolution enhancements of factors from 2 to 4 while statistical downscaling conservatively requires resolution increases of factors from 8 to 12. Rather than enhancing resolution directly to 8-12x, as SR applications typically do, we take an alternative approach. To achieve such a large resolution improvement, we present stacked SRCNNs such that each SRCNN increases the resolution by a factor of $s$. This approach allows the model to learn spatial patterns at multiple scales, requiring less complexity in the spatial representations. The approach of stacking networks has been widely used in deep learning architectures, including stacked denoising autoencoders~\cite{vincent2010stacked} and stacked RBMs for deep belief networks~\cite{hinton2006reducing}. However, contrary to the above networks where stacking is applied in an unsupervised manner, each SRCNN is trained independently using their respective input/output resolutions and stacked at test time.
A similar approach using cascading super-resolution networks showed positive results for upscaling factors below 4~\cite{wang2015deep}, however through experimentation we found that cascading SRCNNs performed worse than stacked SRCNNs. The ability of arbitrarily upscaling ground truth images to lower resolution allows for input/output pairs to be produced at multiple scales to train stacked SRCNNs. However, while training a cascading model, the output of each SRCNN is the input to the following SRCNN, which may be leading to unnecessary error propagation through the network.
\subsection{DeepSD}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{deepsdflow.png}
\caption{Layer by layer resolution enhancement from DeepSD using stacked SRCNNs. Top Row: Elevation, Bottow Row: Precipitation. Columns: 1.0$^{\circ}$, 1/2$^{\circ}$, 1/4$^{\circ}$ and 1/8$^{\circ}$ spatial resolutions.}
\vspace{-0.5cm}
\label{fig:deepsd}
\end{figure}
We now present DeepSD, an augmented and specific architecture of stacked SRCNNs, as a novel SD technique. When applying SR to images we generally only have a LR image to estimate a HR image. However, during SD, we may have underlying high-resolution data coinciding with this LR image to estimate the HR images. For instance, when downscaling precipitation we have two types on inputs including LR precipitation and static topographical features such as HR elevation and land/water masks to estimate HR precipitation. As topographical features are known beforehand at very high resolutions and generally do not change over the period of interest they can be leveraged at each scaling factor. As done when training stacked SRCNNs, each SRCNN is trained independently with it's associated input/output pairs. As presented in figure~\ref{fig:deepsd}, inference is executed by starting with the lowest resolution image with it's associated HR elevation to predict the first resolution enhancement. The next resolution enhancement is estimated from the previous layer's estimate and it's associated HR elevation. This process is repeated for each trained SRCNN. Figure~\ref{fig:deepsd} illustrates this process with a precipitation event and it's various resolution improvements. We see that this stacked process allows the model to capture both regional and local patterns.
\section{Application of DeepSD}
Though high resolution precipitation is crucial to climate adaptation, it makes up two of the four major holes in climate science~\cite{schiermeier2010real}. Furthermore, both statistical and dynamical downscaling approaches have been shown to add little information beyond coarse ESMs when applied to precipitation~\cite{kerr2013forecasting,Burger2012}. This motivates our application to downscale daily precipitation over the CONUS, a region where data is credible and abundant at high resolutions. As presented above, we use daily precipitation from the PRISM dataset~\cite{daly2008physiographically} and elevation from Global 30 Arc-Second Elevation Data Set (GTOPO30) provided by the USGS. These datasets are used to train and test DeepSD, which we compare to BCSD, a widely used statistical downscaling technique, as well as three off-the-shelf machine learning regression approaches. The years 1980 to 2005 were used for training (9496 days) while the years 2006 and 2014 (3287 days) were used for testing. Lastly, we present a scalable framework on the NASA Earth Exchange (NEX) platform to downscale 20 GCMs for multiple emission scenarios.
\subsection{Training DeepSD}
Our experiments downscale daily precipitation from $1.0^{\circ}$ to $1/8^{\circ}$, an 8x resolution enhancement, using three SRCNN networks each providing a 2x resolution increase ($1.0^{\circ} \rightarrow 1/2^{\circ} \rightarrow 1/4^{\circ} \rightarrow 1/8^{\circ}$).
\noindent
\subsubsection*{Data preprocessing} Data for a single day at the highest resolution, $1/8^{\circ}$, covering CONUS is an ``image'' of size 208x464. Precipitation and elevation are used as input channels while precipitation is the sole output. Images are obtained at each resolution through up-sampling using a bicubic interpolation. For instance up-sampling to $1.0^{\circ}$ decreases the image size from 208x464 to 26x58. Precipitation features for the first SRCNN, downscaling from $1.0^{\circ}$ to $1/2^{\circ}$, are first up-sampled to $1.0^{\circ}$ and then interpolated for a second time to $1/2^{\circ}$ in order to correspond to the output size of 52x116. This process is subsequently applied to each SRCNN depending on it's corresponding resolution. During the training phase, 51x51 sub-images are extracted at a stride of 20 to provide heterogeneity in the training set. The number of sub-images per year (1095, 9125, and 45,625) increase with resolution. Features and labels are normalized to zero mean and unit variance.
\noindent
\subsubsection*{Training Parameters} All SRCNNs are trained with the same set of parameters, selected using those found to work well by Dong et al.~\cite{dong2014learning}. Layer 1 consists of 64 filters of 9x9 kernels, layer 2 consists of 32 fully connected neurons (1x1 filters), and the output layer uses a 5x5 kernel (see Figure~\ref{fig:filters-activations}). Higher resolution models which have a greater number of sub-images may gain from larger kernel sizes and an increased number of filters. Each network is trained using Adam optimization~\cite{kingma2014adam} with a learning rate of $10^{-4}$ for the first two layers and $10^{-5}$ for the last layers. Each model was trained for $10^{7}$ iterations with a batch size of 200. Tensorflow~\cite{abadi2016tensorflow} was utilized to build and train DeepSD. Training harnessed three Titan X GPUs on an NVIDIA DIGITS DevBox by independently training each SRCNN. Inference was then executed sequentially on a single Titan X GPU on the same machine.
\subsection{Comparison}
\begin{table}
\centering
\begin{tabular}{lrrrrr}
\toprule
{} & Bias & Corr & RMSE & Skill$^{1}$ & Runtime\\
Model & {\tiny (mm/day)} & & {\tiny (mm/day)} & & (secs) \\
\midrule
Lasso & 0.053 & 0.892 & 2.653 & 0.925 & 1297 \\
ANN & 0.049 & 0.862 & 3.002 & 0.907 & 2015 \\
SVM & -1.489 & 0.886 & 3.205 & 0.342 & 27800\\
BCSD & -0.037 & 0.849 & 4.414 & \textbf{0.955} & \textbf{13} \\
DeepSD & \textbf{0.022} & \textbf{0.914} & \textbf{2.529} & 0.947 & 71 \\
\bottomrule
\end{tabular}
\caption{Comparison of predictive ability between all five methods for 1000 randomly selected locations in CONUS. Runtime is computed as the amount of time to downscale 1 year of CONUS.}
\vspace{-0.5cm}
\label{tab:asd-compare}
\end{table}
\footnote[1]{See section 6.2.2 for details.}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{rmse_map.pdf}
\caption{Daily Root Mean Square Error (RMSE) computed at each location for years 2006 to 2014 (test set) in CONUS for A) DeepSD and B) BCSD. Red corresponds to high RMSE while blue corresponds to low RMSE.}
\vspace{-0.2cm}
\label{fig:rmse-map}
\end{figure*}
\subsubsection{State-of-the-Art Methods}
Bias Correction Spatial Disaggregation (BCSD)~\cite{wood2004hydrologic} is a simple but effective method for statistical downscaling. Many studies have compared BCSD to a variety of other downscaling methods and have found good results in estimating the underlying distribution of precipitation~\cite{Burger2012}. In our experiments we apply the daily BCSD technique~\cite{thrasher2012technical} to precipitation over CONUS. First, the high-resolution precipitation is linearly interpolated to the low-resolution grid. Contrary to applying BCSD to ESMs, we are not required to perform quantile mapping as the distributions are identical. Next, the low-resolution precipitation is interpolated back to high-resolution such that the fine-grained information is lost. Then, scaling factors are computed by dividing high-resolution observations with the interpolated data over the training set (1980-2005). Lastly, the interpolated data is multiplied by the scaling factors to provide downscaled projections. For a more detailed description of this implementation of BCSD see~\cite{thrasher2012technical}. The projections over the test set (2006-2014) are then used for comparison to BCSD.
A second set of methods, Automated-Statistical Downscaling (ASD)~\cite{hessami2008automated}, is applied to compare a variety of regression techniques to DeepSD. ASD consists of two steps for downscaling precipitation: 1. Classifying rainy/non-rainy days ($\>$mm), 2. Estimating total precipitation on rainy days. Hence, this approach requires both classification and regression methods. We compare three ASD approaches using logistic and lasso regression, support vector machine (SVM) classifier and regression, and artificial neural network (ANN) classifier and regression. The Lasso penalty parameter at each location was chosen using 3-fold cross-validation. The SVMs were trained with a linear kernel and a penalty parameter of $1.0$. Each ANN consists of a single layer of 100 units connected with a sigmoid function. A 9 by 9 box for the LR precipitation surrounding the downscaled location is selected as features. Each downscaled location requires individually optimized parameters making the process computationally intensive and complex. All features and labels are normalized to zero mean and unit variance. Hence, we randomly selected 1000 locations to downscale as a trade-off between complexity and statistical certainty around our results.
\subsubsection{Daily Predictability}
\begin{table}
\centering
\begin{tabular}{llrrrr}
\toprule
& & Bias & Corr & RMSE & Skill$^{1}$ \\
Season & Model & {\tiny (mm/day)} & & {\tiny (mm/day)} & \\
\midrule
DJF & BCSD & \textbf{0.02} & 0.89 & 2.36 & 0.95 \\
& DeepSD & -0.03 & \textbf{0.95} & \textbf{1.53} & \textbf{0.94} \\
JJA & BCSD & \textbf{0.01} & 0.78 & 4.15 & \textbf{0.92} \\
& DeepSD & -0.05 & \textbf{0.86} & \textbf{3.29} & 0.91 \\
MAM & BCSD & \textbf{0.01} & 0.87 & 3.02 & \textbf{0.94} \\
& DeepSD & -0.03 & \textbf{0.93} & \textbf{2.29} & 0.93 \\
SON & BCSD & \textbf{0.01} & 0.87 & 3.27 & \textbf{0.94} \\
& DeepSD & -0.04 & \textbf{0.93} & \textbf{2.31} & \textbf{0.94} \\
\bottomrule
\end{tabular}
\caption{Comparison of Predictive Ability between DeepSD and BCSD for each season, Winter, Summer, Spring, and Fall. Values are computed at each location in CONUS and averaged.}
\vspace{-0.5cm}
\label{tab:seasonal}
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{extreme-stats.pdf}
\caption{Comparison of DeepSD and BCSD for increasingly extreme precipitation. At each location in CONUS, all precipitation events above a percentile threshold (x-axis) are selected. Percentile thresholds between 90 and 99.9 are used. A) RMSE, B) Correlation, C) Bias, and D) Skill are computed at each location and averages over CONUS. The confidence bounds of each metric are taken from the 25th and 75th quantiles.}
\label{fig:extremes}
\end{figure*}
DeepSD's ability to provide credible projections is crucial to all stakeholders. While there are many facets to statistical downscaling, we use a few key metrics to show DeepSD's applicability. Root mean square error (RMSE) and Pearson's correlation are used to capture the predictive capabilities of the methods. Figure~\ref{fig:rmse-map} maps this RMSE (mm/day) for each location. Bias, the average error, presents the ability to estimate the mean while a skill score metric, as presented in~\cite{perkins2007evaluation}, is used to measure distribution similarity. Skill is computed as
\begin{equation}
skill = \sum_{i=1}^n minimum(Z_o^{(i)}, Z_m^{(i)})
\end{equation}
such that $Z_o$ and $Z_m$ are the observed and DeepSD's empirical probability density function while n is the number of bins. Hence, the skill score is between 0 and 1 where 1 is the best.
Our first experiment compares five approaches, DeepSD, BCSD, Lasso, SVM, and ANN, on their ability to capture daily predictability, presented in Table~\ref{tab:asd-compare}. The four metrics discussed above are computed and averaged over the 1000 randomly selected locations in CONUS where ASD methods were trained. We find that DeepSD outperforms the other approaches in terms of bias, correlation, and RMSE and closely behind BCSD in terms of skill. Furthermore, we find that SVM performs poorly in testing while having the longest runtime. Similarly, the least complex ASD method, Lasso, outperforms the more complex ANN. As expected, BCSD, a method built around estimating the underlying distribution, does well in minimizing bias and estimating the underlying distribution. For these reasons, in conjunction with our previous findings~\cite{vandal2017intercomparison}, the remaining experiments will limit the methods to DeepSD and BCSD.
In the next experiment compare DeepSD and BCSD, the two scalable and top performing methods from the previous experiment, with each metric over CONUS. Each metric is computed per location and season using the daily observations and downscaled estimates then averaged over CONUS, as presented in Table
~\ref{tab:seasonal}. We find that DeepSD has high predictive capabilities for all seasons, higher correlation and lower RMSE, when compared to BCSD. Similar results are shown in Figure~\ref{fig:rmse-map} where DeepSD has a lower RMSE than BCSD for 79\% of CONUS. Furthermore, we find each method's ability to estimate the underlying distribution well with low bias, $<0.5\text{ mm/day}$, and a high skill score of $\sim0.98$. As BCSD is built specifically to minimize bias and match the underlying distribution, DeepSD's performance is strong. Overall, DeepSD outperforms BCSD for the chosen set metrics.
\subsubsection{Predicting Extremes}
As discussed in section 4, downscaling both averages and extreme events with a single method is challenging. Our last experiment tests this challenge by comparing DeepSD's ability to estimate extreme precipitation events when compared to BCSD, an approach shown to perform well~\cite{Burger2012}. A varying quantile threshold approach is used to test each methods ability to capture extreme events. For instance, given a downscaled location we compute RMSE, correlation, bias, and skill for all precipitation events greater the 90th percentile. This is done for a range of percentiles between 90 and 99.9 and averaged over all locations in CONUS. Along with the mean, we select the 25th and 75th quantiles of each metric over CONUS and plot them as confidence bounds in Figure~\ref{fig:extremes}. Figure~\ref{fig:extremes} presents BCSD's loss of predictive capability when compared to DeepSD. We find that BCSD over-estimates extremes at upper quantiles while DeepSD is relatively stable. Though RMSE, Corr, and Skill becomes worse at these extremes, DeepSD consistently outperforms BCSD, most often with thinner confidence bounds. These results show DeepSD's ability to perform well for increasingly extreme precipitation events. DeepSD's performance is impressive given that literature has shown that traditional techniques tend to fail when downscaling averages and extremes simultaneously. We hypothesize that capturing nearby spatial information allows DeepSD to isolate areas where extreme precipitation events are more likely than others.
\subsection{Scalability on NASA's NEX}
Comprehensive studies of climate change requires much more than a single ESM simulation but rather multiple projections from different models, emission scenarios, and initial conditions in order to capture uncertainty. In total, CMIP5 contains more than 20 models at 4 emission scenarios (Representative Concentration Pathways (RCPs) 2.6,4.5,6.5,8.6), a variable number of initial conditions, and multiple climate variables at a daily temporal resolution. Generally, each prospective projection is available from 2006 to 2100 while retrospective projections are available from 1850 to 2005. Limiting the downscaled projections to encompass CONUS at $1/8^{\circ} \times 1/8^{\circ}$, a single simulation requires 134MB. Following the current timeframes of downscaled projections on NASA's NEX platform, downscaling from the year 1950 to 2005 requires 7.4GB of storage while each prospective run needs 13GB. Hence, the final dataset size is 1.2TB. When downscaled further to PRISM's native resolution, $1/16^{\circ} \times 1/16^{\circ}$, the dataset size increases to approximately 5TB. Furthermore, the dataset scales linearly as more variables are added, including temperature minimum and maximum.
We test computational scalability by computing the amount of time taken to downscale 1 year of CONUS, presented in Table~\ref{tab:asd-compare}. For DeepSD, this includes the 3 feed-forward processes and their corresponding interpolations computed on a single GPU in NVIDIA DIGITS DevBox. Runtime for each of the ASD methods is estimated and scaled from the length of time to downscale the 1000 selected locations using 40 CPUs in parallel (Intel Xeon CPU E5-2680 2.8GHz). BCSD's runtime is computed as the amount of time taken to quantile map, interpolate, and scale a years worth of data. We find that BCSD and DeepSD widely outperform the ASD approaches. While BCSD provides the quickest runtime, DeepSD is still scalable.
Though DeepSD is a highly scalable method, due to a single feed-forward neural network architecture, generating such large datasets still requires heavy computational power. However, storage and compute resources are satisfied by dedicated access to the Pleiades supercomputer housed in NASA's Advanced Supercomputer Division (HECC) at NASA Ames. High resolution projections can be quickly computed using GPU's, which are available on each node, in coordination with the Message Passing Interface (MPI). High-resolution projections are then stored on NEX's filesystem which has currently 2.3PB of rapid-access storage in addition to large scale tape storage accessible on the HECC platform. In this paper we present a methodology for statistical downscaling, DeepSD, that leverages recent advances in image super-resolution and convolutional neural networks. DeepSD differs from previous SD methods by explicitly capturing spatial structure while improving scalability. A brief comparison with baseline SD techniques, BCSD and ASD, shows promising results in predictive capabilities when downscaling precipitation over the continental United States. Lastly, we describe how DeepSD can be scaled using NASA's Earth Exchange platform to provide an ensemble of downscaled climate projections from more than 20 ESMs.
\section{Conclusion}
Though DeepSD shows promise for SD, there are still some limitations in our experimentation regarding spatial and temporal generalization. An advantage of DeepSD is that a single trained model is able to downscale spatial heterogeneous regions. However, we do not test predictability in regions where the model was not trained. Future work will examine this hypothesis to understand DeepSD's credibility in regions with few observations. Second, we do not test temporal non-stationarity, a longstanding problem in statistical downscaling. Evaluation under non-stationarity can be tested using approaches presented by Salvi et al.~\cite{Salvi2015}, such that training and testing data is split between cold/warm years. As there is a single model for all locations, including cold and warm climates, we hypothesize that DeepSD is capable of capturing non-stationarity.
Furthermore, future work can improve multiple facets of DeepSD. For instance, the inclusion of more variables such as temperature, wind, and humidity at different pressure levels of the atmosphere may capture more climate patterns. Also, downscaling multiple climate variables simultaneously could be explored to find similar spatial patterns in the high-resolution datasets, such as high temperatures and increased precipitation. Most importantly, DeepSD fails to capture uncertainty around its downscaled projections, a key factor in adapting to climate change. Recent advances in Bayesian Deep Learning concepts~\cite{gal2016uncertainty} may aid in quantifying uncertainty. Though these limitations exist, DeepSD is a scalable architecture with high predictive capabilities which provides a novel framework for statistical downscaling.
\begin{acks}
This work was supported by NASA Earth Exchange (NEX), \grantsponsor{}{National Science Foundation CISE Expeditions in Computing}{https://www.nsf.gov/awardsearch/showAward?AWD_ID=1029711} under grant number:~\grantnum{}{1029711}, \grantsponsor{}{National Science Foundation CyberSEES}{https://www.nsf.gov/awardsearch/showAward?AWD_ID=1442728} under grand number:~\grantnum{}{1442728}, and \grantsponsor{}{National Science Foundation BIGDATA}{https://www.nsf.gov/awardsearch/showAward?AWD_ID=1447587} under grant number:~\grantnum{}{1447587}. The GTOPO30 dataset was distributed by the Land Processes Distributed Active Archive Center (LP DAAC), located at USGS/EROS, Sioux Falls, SD. \url{http://lpdaac.usgs.gov}. We thank Arindam Banerjee for valuable comments.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\label{introduction}
The nature of dark matter remains one of the most puzzling questions
in modern cosmology. Its assumed that elementary particles could be
the constituents of dark matter for which there have been several experimental
attempts to perform a direct detection of such dark matter particle, but still
a clear evidence for such particles remains absent \cite{Undagoitia:2015gya}.
Examples of these experiments are the Large Underground Xenon (LUX) experiment
\cite{Akerib:2012ys}, the DAMA/Libra experiment \cite{Bernabei:2008yh}, ZEPLIN
experiment \cite{Akimov:2006qw}, XENON10 experiment \cite{Aprile:679A} or DAMIC
experiment at SNOLAB \cite{Barreto:2011zu}, \cite{Chavarria:2014ika, Aguilar-Arevalo:2016zop, Aguilar-Arevalo:2016ndq}. Another
approach to determine the nature of dark matter is to find specific
features in the dynamics of visible matter in the presence of the
gravitational potential of a dark matter halo.
Dark matter is usually assumed to be a stable particle with no
electric charge, and with a negligible cross-section with baryonic
particles. However, it does interact gravitationally with the rest of
the matter (in general relativistic terms, dark matter does curve
space-time). It is also assumed that dark matter particles only
interact with themselves gravitationally, or that at least any
non-gravitational self-interaction is very weak. This means that a
fluid description of dark matter is not adequate (even though this
description may lead to some intuitive properties
\cite{Barranco:2013wy}). Indeed, the fluid approximation requires that
the mean-free path of particles is much smaller than the
characteristic size of the system, which clearly does not apply to
dark matter particles. One should mention that there are serious
proposals that consider that dark matter can be described as an
ultra-light scalar field, and such proposals seem to ameliorate some
of the problems related with the usual dark matter paradigm (see
e.g.~\cite{Matos:1999et,Matos:2000xi,Matos:2003pe}).
If one considers dark matter as a collection of non-interactive
particles there are at least two different ways to study their
dynamics. The standard approach has been to study them using
increasingly sophisticated direct gravitational N-body simulations,
with today can use billions of particles \cite{Navarro:1995iw, Harker:2005um}.
In N-body simulations, each particle represents a very large number of dark matter
particles and the interaction between two particles is computed by assuming that
they have a finite size and density profile, which leads to an effective softening
force at small scales \cite{Athanassoula:10.1046},\cite{Dehnen:10.1046}. In this way,
a given density velocity field is represented by
a set of particles \cite{Hockney:1981csup}. For more details on the techniques and
status of these simulations one can refer to \cite{Bagla:2004au}.
Another approach is to use a continuum
approximation based on kinetic theory and the Boltzmann
equation~\cite{BT2}. In this second approach a collection of
particles is described in a collective way by means of a distribution
function $f$ defined as a probability density in the so-called phase
space, namely the space defined by generalized coordinates and
momenta. When the particles are non-interactive, one obtains the
homogeneous Boltzmann equation, also known as the Vlasov
equation~\cite{BT2}. Such description is nearly equivalent to the
numerical description of dark matter by means of the N-body
simulations, although there are conceptual and quantitative
differences in the two approaches~\cite{Colombi:2015eia}.
Solutions of the Vlasov equation are frequently not easy to
interpret. It is quite easy to solve in general, since by just by
expressing the distribution function as a function of conserved
quantities we automatically obtain an exact, and stationary,
solution~\cite{Andreasson:2005qy}. However, such general solutions
often do not have a clear physical interpretation, and can lead to
misunderstandings~\cite{Shapiro:1983du}. We have found it
productive then, to solve the Vlasov equation dynamically in order to study
the evolution in phase space of an initial distribution function,
which allows for a more clear physical interpretation. Using this
approach, we study the evolution of a spherically symmetric
distribution function in different gravitational potentials, which
describe several models of dark matter halos. The distribution
function is assumed to describe a dark matter inhomogeneity within the
dark matter halo. The aim of this work is first to develop a robust
numerical integration code for the Vlasov equation, and then use it to test
the stability of inhomogeneities in dark mater halos, as well as to obtain
non-trivial final stationary states for the initial inhomogeneity.
The paper is organized as follows: In Section~\ref{Sec1} we present a
brief review of kinetic theory, introducing the Vlasov equation. In
Section~\ref{Sec2} we present four different halo gravitational
potentials, and derive a dimensionless Vlasov equation for the
corresponding models. In Section~\ref{Sec3} we describe the code used
to perform the numerical simulations of the evolution of the
distribution function by means of the Vlasov equation. After that, we
present our results for each halo model in Section~\ref{Sec4}.
Finally, we discuss our findings in Section~\ref{Sec5}.
\section{Kinetic theory and Vlasov equation}
\label{Sec1}
\subsection{Distribution function in spherical symmetry}
The particle distribution function $f$ lives in a higher dimensional
phase-space: six dimensional in the case of classical mechanics, and
seven dimensional in relativistic physics (down from eight due to the
constraint on the magnitude of the momenta, the so-called mass-shell
condition $p_\mu p^\mu=-m^2 c^2$, where $m$ is the particle's mass and
$c$ denotes the speed of light). The general explicit invariant volume
element in momentum space in the general relativistic description is
\begin{equation} \label{eq:dw}
d \omega = \frac{d^3 \, p_{*}}{p^0\,\sqrt{-g}} \; ,
\end{equation}
where $p_{*}$ denotes the covariant momenta, $p_i=\partial L /
\partial \dot{x}^i$, with $L$ the Lagrangian, and where $g$ is the
determinant of the spacetime metric. In the Newtonian limit this
volume element takes the form
\begin{equation}
d \omega = \frac{d^3\,p_*}{\sqrt{h}} \; ,
\end{equation}
where now $h$ is the determinant of the flat metric in the coordinates
under consideration, and the momenta $p_{*}$ corresponds to the
conjugate momenta. Notice that since in general the volume element in
physical space is given by \mbox{$dV = \sqrt{h} \; d^3 x$}, one finds
that the phase space volume element $dV d \omega$ is in fact
invariant.
Keeping in the Newtonian limit, when using Cartesian coordinates the
invariant volume element in momentum space takes the form $d \omega =
dp_x dp_y dp_z$, and when using spherical coordinates it becomes
\begin{equation}
d \omega=\frac{dp_r \; dp_\theta \; dp_\varphi}{r^2 \sin \theta} \; .
\label{eq:dw0}
\end{equation}
There is in fact another expression that is commonly used for
the invariant volume element in momentum space in spherical
coordinates, namely \mbox{$d \omega = p^2 \sin \theta_p \; dp \;
d\theta_p \; d\varphi_p$}, where $p$ is the magnitude of the
momentum, and $(\theta_p,\varphi_p$) are angles defined in momentum
space in the same way as ($\theta$,$\varphi$) are defined in the space
sector, so that for example $p_x = p \sin \theta_p \cos \varphi_p, p_y
= p \sin \theta_p \sin \varphi_p, p_z = p \cos \theta_p$. This volume
element can be related to the former expression, Eq.~(\ref{eq:dw0}),
by the Jacobian of the corresponding coordinate
transformation~\cite{Dominguez:2015,Jimenez:2016}.
In the following we will consider a stationary background
gravitational potential with spherical symmetry, and a distribution
function $f(t,x^i,p_i)$ that is also spherically symmetric. Moreover,
instead of the angular conjugated momenta, $p_\theta, p_\varphi$, we
will use the conserved total angular momentum squared, $L^2$, and an
auxiliary angle, $\psi$:
\begin{eqnarray}
L^2 &=& p_\theta^2 + \frac{p_\varphi^2}{\sin^2 \theta}, \label{eq:L2} \\
\psi &=& \arctan\left(\frac{p_\theta \sin\theta}{p_\varphi}\right),
\end{eqnarray}
so that the invariant volume element, Eq.~(\ref{eq:dw}), takes the form
\begin{equation}
d \omega = \frac{1}{2 r^2} \; dp_r \, dL^2\, d\psi
= \frac{L}{r^2} \; dp_r \, dL \, d\psi \; .
\label{eq:dw1}
\end{equation}
Finally, since we will be working with spherical symmetry, the
distribution function will have no dependence on $\psi$.
We will therefore integrate over $\psi$ so the volume element used for this
work will have the final form
\begin{equation}
d \bar{\omega} = \int \left( d \omega \right) \; d \psi
= \frac{\pi}{r^2} \; dp_r\,dL^2
= \frac{2 \pi L}{r^2} \; dp_r\,dL \; .
\label{eq:dw2}
\end{equation}
In spherical symmetry the distribution function will have the form $f
= f(t,r,p_r,L)$. This implies that in the case of spherical symmetry,
the phase space is effectively tri-dimensional. At this point it is
important to mention a couple of details about this distribution
function. First, even though we are in spherical symmetry, the
individual particles can have non-trivial angular momentum. The
spherical symmetry will be maintained as long as at any given point in
space with a certain radial distance $r$ from the origin, the
tangential velocities of the particles are distributed uniformly in
every possible direction. Second, notice that since $f$ depends on
the angular momentum $L$, and this is defined
through~\eqref{eq:L2}, the distribution $f$ when rewritten in terms of
$(p_r,p_\theta,p_\varphi)$ will in fact depend on the angle $\theta$.
This might seem odd since after all we are in spherical symmetry, but
as we will see below this dependence in fact cancels out in the Vlasov
equation.
We can now define the particle density $\rho_f$ in physical space, the
momentum density $j_f$, and the mean value over momentum space of an
arbitrary function $g$, will take the form
\begin{eqnarray}
\rho_f(t,r) &\equiv& \int f(t,r,p_r,L) \; d \bar{\omega}
= \frac{2 \pi}{r^2} \int f(t,r,p_r,L) \; L \; dp_r \; dL \; , \label{eq:rho1} \\
j_f(t,r) &\equiv& \int p_r f(t,r,p_r,L) \; d \bar{\omega}
= \frac{2 \pi}{r^2} \int p_r f(t,r,p_r,L) \; L \; dp_r \; dL \; ,
\label{eq:curr1} \\
\bar{g}(t,r) &\equiv& \int g f(t,r,p_r,L) \; d \bar{\omega}
= \frac{2 \pi}{r^2} \int g f(t,r,p_r,L) \; L\,dp_r\,dL \; . \label{eq:mean}
\end{eqnarray}
The total number of particles can then be defined as
\begin{equation}
N = \int f \; dV d \bar{\omega} = 8 \pi^2 \int f L \; dr dp_r dL
= 4 \pi \int \rho_f r^2 dr \; .
\label{eq:totalN}
\end{equation}
One can also define of the average value of a given physical quantity
over the complete phase space as
\begin{equation}
\left<g\right> (t) \equiv 8 \pi^2 \int g f(t,r,p_r,L) \; L \; dp_r \; dL \; dr
= 4 \pi \int \bar{g}(t,r) \; r^2 dr \; .
\label{eq:avg}
\end{equation}
As a further simplifying assumption, we will also assume that all the
particles have the same conserved angular momentum $L_0$, so that,
using a delta function, we can write $f(t,r,p_r,L) =
F(t,r,p_r)\delta(L-L_0)$, obtaining
\begin{eqnarray}
\rho_f(t,r) &=& \frac{2 \pi L_0}{r^2} \int F(t,r,p_r; L_0) \; dp_r \; ,
\label{eq:rho2} \\
j_f(t,r) &=& \frac{2 \pi L_0}{r^2} \int p_r F(t,r,p_r; L_0) \; dp_r \; ,
\label{eq:curr2} \\
\bar{g}(t,r) &=& \frac{2 \pi L_0}{r^2} \int g F(t,r,p_r; L_0) \; dp_r \; ,
\label{eq:mean2} \\
\left<g\right> (t) &=& 8 \pi^2 L_0 \int g F(t,r,p_r; L_0) \; dp_r \; dr \; ,
\label{eq:avg2} \\
N &=& 8 \pi^2 L_0 \int F(t,r,p_r; L_0) \; dr dp_r \; .
\label{eq:totalN2}
\end{eqnarray}
One should also mention the fact that for the particular case when the
motion is purely radial and the angular momentum is zero, $L_0=0$, the
above equations are in fact not correct and the derivation has to be
made again from the top (a naive application of these expressions
would seem to indicate that all the integrals above vanish). That
is, we have to begin with a Dirac delta dependence of the distribution function
in the angular momenta
\begin{equation}
f(t,r,p_r,p_\theta,p_\varphi)= F(t,r,p_r) \; \delta(p_\theta) \; \delta(p_\varphi/\sin
\theta)
\end{equation}
Transforming the Dirac deltas from coordinates
($p_\theta$,$p_\varphi$) to the new coordinates ($L$,$\psi$) yields
\begin{equation*}
\delta(p_\theta) \; \delta(p_\varphi / \sin \theta)
= \frac{1}{L} \; \delta(L) \; \delta(\psi) \; .
\end{equation*}
So that the expressions for the density, current, mean value and
average in the case of zero angular momentum become
\begin{eqnarray}
\rho_f(t,r) &=& \frac{1}{r^2} \int F(t,r,p_r) \; dp_r \; , \label{eq:rho3} \\
j_f(t,r) &=& \frac{1}{r^2} \int p_r \; F(t,r,p_r) \; dp_r \; , \label{eq:curr3} \\
\bar{g}(t,r) &=& \frac{1}{r^2} \int g \; F(t,r,p_r) \; dp_r \; , \label{eq:mean3} \\
\left<g\right> (t) &=& 4 \pi \int g \; F(t,r,p_r) \; dp_r \; dr \; , \label{eq:avg3} \\
N &=& 4 \pi \int F(t,r,p_r) \; dp_r \; dr \; . \label{eq:totalN3}
\end{eqnarray}
\subsection{Vlasov equation}
Let us now consider the Vlasov equation. In the general case, the
collisionless Boltzmann equation, also known as the Vlasov equation,
has the form
\begin{equation}
\frac{\partial f}{\partial t} + \frac{dx^i}{dt} \; \frac{\partial f}{\partial x^i}
+ \frac{dp^i}{dt} \; \frac{\partial f}{\partial p^i} = 0 \; .
\end{equation}
In spherical symmetry, and with the definitions given above, this
reduces to
\begin{equation}
\frac{\partial f}{\partial t} = - \frac{p_r}{m} \; \frac{\partial f}{\partial r}
- \frac{p_\theta}{m r^2} \; \frac{\partial f}{\partial \theta}
- \dot{p}_r \; \frac{\partial f}{\partial p_r}
- \dot{p}_\theta \; \frac{\partial f}{\partial p_\theta} \; . \label{eq:Vla0}
\end{equation}
Notice that, as mentioned before, even in spherical symmetry the
distribution function has a dependency on $\theta$ when written in
terms of the conjugate momenta.
Now, for a central potential $\Phi(r)$, the equations of motion are
\begin{eqnarray}
\dot{p}_r &=& - \frac{\partial {\cal H}}{\partial r}
= - \frac{\partial \Phi}{\partial r} + \frac{L^2}{m r^3} \; , \\
\dot{p}_\theta &=& - \frac{\partial {\cal H}}{\partial \theta}
= \frac{{p_\varphi}^2 \cos \theta}{m r^2 \sin^3 \theta} \; ,
\end{eqnarray}
where ${\cal H}$ is the Hamiltonian of the system. This implies that
\begin{eqnarray}
\frac{p_\theta}{m r^2} \; \frac{\partial f}{\partial \theta}
+ \dot{p}_\theta \frac{\partial f}{\partial p_\theta}
&=& \frac{p_\theta}{m r^2} \; \frac{\partial f}{\partial \theta}
+ \frac{p_\varphi^2 \cos \theta}{m r^2\sin^3 \theta} \; \frac{\partial f}{\partial p_\theta}
\nonumber \\
&=& \frac{p_\theta}{m r^2} \left( -\frac{2 p_\varphi^2 \cos\theta}{\sin^3 \theta} \right)
\frac{\partial f}{\partial L^2}
+ \frac{p_\varphi^2\cos\theta} {m r^2 \sin^3 \theta} \;
\left( 2 p_\theta \right) \frac{\partial f}{\partial L^2} \nonumber \\
&=& 0 \; ,
\end{eqnarray}
where we have used the chain rule to calculate the derivatives of $f$
with respect to $\theta$ and $p_\theta$ in terms of its derivative
with respect to $L^2$. This implies that in spherical symmetry the
Vlasov equation takes the final form:
\begin{equation}
\frac{\partial f}{\partial t}= - \frac{p_r}{m} \; \frac{\partial f}{\partial r}
+ \left( \frac{\partial \Phi}{\partial r} - \frac{L^2}{m r^3} \right) \frac{\partial f}{\partial p_r}
\; .
\label{eq:Vla1}
\end{equation}
Since there are no terms with derivatives with respect to $L$ in this
equation, we see that different values of the angular momenta evolve
independently of each other. The Vlasov equation in spherical symmetry
is then effectively two-dimensional. In the corresponding temporal evolution shown below,
for simplicity we will in fact assume that all particles have exactly
the same value of the angular momentum $L_0$, and leave the case of a
distribution of different angular momentum for a follow up paper.
\subsection{Continuity equation}
\label{sec:continuity}
Using the Vlasov equation we can derive the continuity equation for
the particles. For that we start from the Vlasov equation written as
\begin{equation}\label{eq:apx_b_vlasov}
\frac{\partial f}{\partial t} + \frac{p_r}{m}\frac{\partial f}{\partial r}
- \frac{\partial \Phi_{\rm eff}(r)}{\partial r} \;
\frac{\partial f}{\partial p_r} = 0 \; ,
\end{equation}
where $\Phi_{\rm eff}(r) \equiv
\Phi(r)+ L^2/2 m r^2$ is the effective
potential. Integrating over momentum space we find
\begin{equation}\label{eq:vlasovint1}
\int \frac{\partial f}{\partial t} \; d \bar{\omega}
+ \int \frac{p_r}{m} \; \frac{\partial f}{\partial r} \; d \bar{\omega}
- \int \frac{\partial \Phi_{\rm eff}(r)}{\partial r} \;
\frac{\partial f}{\partial p_r} \; d \bar{\omega} = 0 \; ,
\end{equation}
with $d \bar{\omega}$ the volume element given in equation~\eqref{eq:dw1}.
For now on we will assume that the distribution function is either of
compact support in phase space. The first term of the above equation
is easy to simplify:
\begin{equation}
\int \frac{\partial f}{\partial t} \; d \bar{\omega}
= \frac{\partial} {\partial t} \int f \; d \bar{\omega}
= \frac{\partial \rho_f} {\partial t} \; ,
\end{equation}
where have used the definition of the particle density $\rho_f$,
equation~\eqref{eq:rho1}. In a similar way we find for the second
term:
\begin{equation}
\int \frac{p_r}{m} \; \frac{\partial f}{\partial r} \; d \bar{\omega}
= \frac{2 \pi}{m r^2} \; \int p_r \; \frac{\partial f} {\partial r} \; L \; dp_r dL
= \frac{2 \pi}{m r^2} \; \frac{\partial} {\partial r} \int p_r f L \; dp_r dL
= \frac{1}{m r^2} \; \frac{\partial} {\partial r} \left( r^2 j_f \right) \; ,
\end{equation}
where we have now used the definition of the momentum density $j_f$,
equation~\eqref{eq:curr1}. Finally, the last term can be easily
shown to vanish for a distribution function of compact support:
\begin{equation}
\int \frac{\partial \Phi_{\rm eff}(r)}{\partial r} \;
\frac{\partial f}{\partial p_r} \; d \bar{\omega}
= \frac{2 \pi}{r^2} \; \frac{\partial \Phi_{\rm eff}(r)}{\partial r}
\int \frac{\partial f}{\partial p_r} \; L \; dp_r dL
= \frac{2 \pi}{r^2} \; \frac{\partial \Phi_{\rm eff}(r)}{\partial r}
\int \left( \int \frac{\partial f}{\partial p_r} \; d p_r \right) L \; dL
= 0 \; ,
\end{equation}
where we assumed that $f$ is zero for large values of $|p_r|$. If we
now define the mass density as $\rho_m := m \rho_f$, the Vlasov
equation integrated over momentum space reduces to
\begin{equation}
\frac{\partial \rho_m}{\partial t} + \frac{1}{r^2} \;
\frac{\partial} {\partial r} \left( r^2 j_f \right) = 0 \; ,
\end{equation}
which is nothing more than the standard continuity equation in
spherical symmetry. By integrating the continuity equation in physical
space it can now be easily shown, by using the divergence theorem, that
the number of particles $N$ defined above in
equation~\eqref{eq:totalN} is conserved in the sense that $\partial N
/ \partial t = 0$.
\subsection{Virial theorem}
\label{sec:virial}
When the distribution function depends only on the conserved energy
and angular momentum, the system is automatically in a steady (or
equilibrium) state. However, for arbitrary initial conditions, the
system can reach equilibrium whenever
$\rho_{,\,t}=0$.
The Virial theorem allows us to determine if the system has reached a
steady state. To derive the virial equation, we start from the Vlasov
equation in spherical symmetry, multiply it by $r p_r$, and
integrate over phase space:
\begin{equation}\label{eq:int_vlasov}
\int r p_r \; \frac{\partial f}{\partial t} \; dV d\bar{\omega}
+ \frac{1}{m} \int r p_r^ 2 \; \frac{\partial f}{\partial r} \; dV d\bar{\omega}
- \int r p_r \; \frac{\partial \Phi_{\rm eff}(r)}{\partial r} \;
\frac{\partial f}{\partial p_r} \; dV d\bar{\omega}
= 0 \; .
\end{equation}
The first term above can be rewritten as:
\begin{equation}
\int r p_r \; \frac{\partial f}{\partial t} \; dV d\bar{\omega}
= \frac{\partial}{\partial t} \int r p_r f \; dV d \bar{\omega}
= \frac{\partial}{\partial t} \left< r p_r \right> \; .
\end{equation}
For the second term in equation~\eqref{eq:int_vlasov} we find:
\begin{eqnarray}
\frac{1}{m} \int r p_r^ 2 \; \frac{\partial f}{\partial r} \; dV d\bar{\omega}
&=& \frac{8 \pi^2}{m} \int p_r^2 \left( r \frac{\partial f}{\partial r} \right) L
\; dr dL dp_r
= - \frac{8 \pi^2}{m} \int p_r^2 f L \; dr dL dp_r \nonumber \\
&=& - \frac{1}{m} \; \left< p_r^2 \right> = - 2 \left< K_r \right> \; ,
\end{eqnarray}
where in the second term above we integrated by parts over $r$, using
the fact that $f$ has compact support,
and where $K_r = p_r^2/2m$ is
the radial kinetic energy.
For the third term of~\eqref{eq:int_vlasov} we find:
\begin{eqnarray}
\int r p_r \; \frac{\partial \Phi_{\rm eff}(r)}{\partial r} \;
\frac{\partial f}{\partial p_r} \; dV d\bar{\omega}
&=& 8 \pi^2 \int r \frac{\partial \Phi_{\rm eff}(r)}{\partial r}
\left( p_r \; \frac{\partial f}{\partial p_r} \right) L \; dr dL dp_r \nonumber \\
&=& - 8 \pi^2 \int r \frac{\partial \Phi_{\rm eff}(r)}{\partial r} f L \;
dr dL dp_r \nonumber \\
&=& \left< r F_{\rm eff}(r) \right> \; ,
\end{eqnarray}
where again we have integrated by parts,
but now over $p_r$, and where
we have defined the effective force as \mbox{$F_{\rm eff}(r) = -
\partial \Phi_{\rm eff}(r) / \partial r$}. The quantity $\left< r F_{\rm eff}(r)
\right>$ in known as the \textit{virial}.
Collecting our results we find that the integrated Vlasov equation
becomes
\begin{equation}
\frac{\partial}{\partial t} \left< r p_r \right> =
2\left< K_r \right> + \left< r F_{\rm eff}(r) \right> \; .
\label{eq:virial1}
\end{equation}
The last result is known as the virial theorem. In a steady state the
distribution function is independent of time, so the virial theorem
implies
\begin{equation}
2 \left< K_r \right> + \left< r F_{\rm eff}(r) \right> = 0 \; .
\label{eq:virial2}
\end{equation}
Notice that if we rewrite the effective force as
\begin{equation}
F_{\rm eff}(r) = - \frac{\partial \Phi_{\rm eff}(r)}{\partial r}
= - \frac{\partial \Phi}{\partial r} + \frac{L^2}{mr^3} \; ,
\end{equation}
then we can rewrite equation~\eqref{eq:virial2} above as
\begin{equation}
2 \left< K_r \right> + \left< r F_g(r) \right>
+ \left< L^2 / m r^2\right> = 0 \; ,
\end{equation}
with $F_g(r)=-\partial \Phi / \partial r$ the purely gravitational
force. If we now recognize that the last term above is nothing more
than twice the angular kinetic energy, the virial theorem becomes
\begin{equation}
2 \left< K_{\rm tot} \right> + \left< r F_g(r) \right> = 0 \; ,
\label{eq:virial3}
\end{equation}
where now $K_{\rm tot}$ is the total kinetic energy. When a system
has reached an equilibrium state and the above equation is satisfied,
we say that the system has \textit{virialized}.
For the special case when the external potential $\Phi(r)$ behaves as
$r^n$, the virial theorem can be rewritten as
\begin{equation}
2 \left< K_{\rm tot} \right> - n \left< \Phi(r) \right> = 0 \; .
\label{eq:virial4}
\end{equation}
This is the form frequently found in textbooks (often for the special
case $n=-1$ corresponding to a simple point-particle gravitational
potential). However, it is not the form we are interested in since
the external potentials we will consider for this work are not as simple
as a power law.~\footnote{When instead of particles moving in an external
potential we consider the case of self-gravitating particles, then
the virial theorem can also be derived, and in that case we do find
that equation~\eqref{eq:virial4} is satisfied with $n=-1$ since the
mutual forces between the particles are purely gravitational.}
\subsection{Entropy}
The entropy is constant for a system that obeys the Vlasov
equation. Indeed, starting from the definition of entropy
\begin{equation}
S = - \int f \ln f \; dV d \bar{\omega} \; ,
\label{def:S}
\end{equation}
it can be seen that this is a conserved quantity for any system
obeying the Vlasov equation:
\begin{eqnarray}
\frac{d S}{dt} &=& -\int \left(1 + \ln f \right) \; \frac{\partial f}{\partial t}
\; dV d \bar{\omega} \nonumber \\
&=& \int \left( 1 + \ln f \right) \left[ \frac{p_r}{m} \frac{\partial f}{\partial r}
- \frac{\partial \Phi(r)}{\partial r} \frac{\partial f}{\partial p_r} \right]
\; dV d \bar{\omega} \nonumber \\
&=& \int \left[ \frac{p_r}{m} \frac{\partial(f\ln f)}{\partial r}
- \frac{\partial \Phi(r)}{\partial r} \frac{\partial(f\ln f)}{\partial p_r} \right]
\; dV d \bar{\omega} \; , \nonumber
\end{eqnarray}
the terms in square brackets can be integrated over phase space and
can easily be seen to vanish if $f$ has compact support, and $f$ and $\Phi$
smooth and finite. Hence
\begin{equation}
\frac{d S}{d t} = 0 \; ,
\end{equation}
so that the entropy is a constant as claimed. This result, though
straightforward, is somewhat counter intuitive since one would expect
the entropy to increase. However, entropy will only increase for
interacting particles for which the collision term of the Boltzmann
equation is non-zero (this is the content of the well-known $H$
theorem). In fact, one can also easily see that the integral over
phase space of any arbitrary, smooth and finite function of $f$ is also conserved,
with the total particle number $N$ and the total entropy $S$ special cases.
In the case of the Vlasov equation, one does find
that if we start with a well defined and localized
distribution function $f$, during evolution the system will move to a
new state where the distribution function is dispersed in a large
section of phase space, but the entropy still remains constant. We
then see that entropy is not a very useful quantity if one wants to
study the way in which the distribution function evolves to fill in a
large region of phase space in this case. Perhaps it would be more
useful to apply ergodic theory to describe the evolution in space
space of systems that obey the Vlasov equation, but this is a subject
that goes beyond the scope of this work.~\footnote{Notice that
calculating the entropy is also problematic numerically, as the
value of $\ln f$ becomes minus infinity in regions where $f$
vanishes, and takes extremely large negative values in regions where
$f$ is very small, which can cause serious problems of numerical
accuracy even if one assumes that $f$ does not quite vanish
anywhere.}
\section{Halo models}
\label{Sec2}
Having set the definitions and discussed the basic properties of the
Vlasov equation, we will now apply the formalism to describe the
stability of perturbations in galactic halo models. In order to do so, we will use the
Vlasov equation, Eq.~(\ref{eq:Vla1}), in a background gravitational
potential of several halo models, and study the evolution of a given
initial localized distribution function $f$ in that potential. We
imagine such a distribution to represent a small perturbation in the
matter distribution of the halo. We perform several numerical tests
which will allow us to determine the fate of the such initial
perturbation.
We will consider four different dark matter halo models which are
commonly discussed in the literature, namely the isothermal, the
truncated isothermal, the Burkert and the Navarro-Frenk-White
models. All of them will be described by a dark matter density:
\begin{eqnarray}
\rho_{\rm iso} = \frac{{\sigma_0}_{iso}^2}{2 \pi G r^2} \; , &\hspace{1cm}&
\rho_{\rm iso-tr} = \frac{{\sigma_0}_{iso-tr}^2}{2 \pi G \left(r^2 + {r_s}^2 \right)}
\; , \nonumber \\
\rho_{\rm Burkert} = \frac{{\rho_0}_{\rm Burkert}}{\left(1 + \frac{r}{r_s} \right) \,
\left( 1 + \frac{r^2}{{r_s}^2} \right)} \; , &\hspace{1cm}&
\rho_{\rm NFW} = \frac{{\rho_0}_{\rm NFW}}{\frac{r}{r_s} \, \left(1 + \frac{r}{r_s}
\right)^2}
\; , \label{rhos}
\end{eqnarray}
where $\sigma$ is a characteristic velocity, $r_s$ is a characteristic
radius, and $\rho_0$ a characteristic density for each case. In order
to be able to compare the behavior of an inhomogeneity in each halo,
we will start by writing them in similar terms. The characteristic
speed $\sigma_0$ can be related to a characteristic density by means
of a characteristic mass $M_0$ and a pattern radius $R_0$, as
\begin{equation}
{\sigma_0}^2 = \frac{G M_0}{R_0} = \frac{G \left( \frac{4 \pi}{3} \rho_0 {R_0}^3
\right)}{R_0} = \frac{4 \pi G \rho_0 {R_0}^2}{3} \; ,
\label{eq:sigma_rho}
\end{equation}
where $G$ is Newton's gravitational constant. Writing now the radial
coordinate as multiple of the pattern radius, $r=R_0\,d$, with $d$ a
dimensionless number, it is possible to rewrite the different halo
densities as
\begin{eqnarray}
\rho_{\rm iso} = \frac{2 {\rho_0}_{\rm iso}}{3 d^2} \; , &\hspace{1cm}&
\rho_{\rm iso-tr} = \frac{2 {\rho_0}_{\rm iso-tr}}{3 d^2 \left(1 + \frac{{d_s}^2}
{d^2}\right)} \; , \nonumber \\
\rho_{\rm Burkert} = \frac{{\rho_0}_{\rm Burkert}}{\left( 1 + \frac{d}{d_s} \right)
\left(1 + \frac{d^2}{{d_s}^2}\right)} \; , &\hspace{1cm}&
\rho_{\rm NFW} = \frac{{\rho_0}_{\rm NFW}}{\frac{d}{d_s}
\left(1 + \frac{d}{d_s}\right)^2} \; , \label{rhos1}
\end{eqnarray}
where $r_s$ has been rewritten as $r_s=R_0\,d_s$. Choosing now
$d_s=1$, and demanding that we obtain the same value of the density
$\rho$ for each halo at $d=3$ for the four models (see Fig.~(\ref{Fig:rhos})),
we find that the four density profiles can be
rewritten in terms of just one characteristic density $\rho_0$:
\begin{eqnarray}
\rho_{\rm iso} = \frac{{\rho_0}_{\rm iso}}{d^2} \; , &\hspace{1cm}&
\rho_{\rm iso-tr} = \frac{10 {\rho_0}_{\rm iso}}{9 \left(1 + d^2\right)}
\; , \nonumber \\
\rho_{\rm Burkert} = \frac{40 {\rho_0}_{\rm iso}}{9 \left(1 + d\right)¡
\left(1 + d^2\right)} \; , &\hspace{1cm}&
\rho_{\rm NFW} = \frac{16 {\rho_0}_{\rm iso}}{3 d \left(1 + d\right)^2} \; .
\label{rhos2}
\end{eqnarray}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.60]{fig01.pdf}}
\caption{Plots of the density profile $\rho(d)$ of the four halos
considered. Notice that our chosen calibration is such that the
densities are equal at $d=3$ in the four cases.}
\label{Fig:rhos}
\end{figure}
Given the density profile $\rho$, we can find that the halo mass
distribution $m(r)$ as
\begin{equation}
m(r) = 4 \pi \int \rho r^2 dr \; ,
\end{equation}
so that the gravitational potential is given by $\Phi=-G m(r)/r$.
Where we can also introduce a dimensionless potential $V(d)$ with the
help of the particle mass $m$ and a characteristic speed $\sigma_0$ as
\begin{equation}
\Phi(r)=m\,{\sigma_0}^2 V(d) \; ,
\end{equation}
In this way, the gradient of the gravitational potential can be
written as
\begin{equation}
\frac{\partial \Phi(r)}{\partial r}
= \frac{m {\sigma_0}^2}{R_0} \; \frac{\partial V(d)}{\partial d} \; .
\end{equation}
Making now use of the relation between $\sigma_0$ and $\rho_0$ given
by Eq.~(\ref{eq:sigma_rho}), the dimensionless gradient of the
gravitational potential for each halo becomes:
\begin{eqnarray}
\frac{\partial V_{\rm iso}(d)}{\partial d} = \frac{3}{d} \; , &\hspace{1cm}&
\frac{\partial V_{\rm iso-tr}(d)}{\partial d} = \frac{10 \left(
1 - \frac{\arctan\left(d\right)}{d}\right)}{3 d} \; , \nonumber \\
&& \label{eqs:phisd} \\
\frac{\partial V_{\rm Burkert}(d)}{\partial d}
= \frac{10 \left(\ln\left(\left(1+d^2\right) \left(1+d\right)^2\right)
- 2 \arctan\left(d\right)\right)}{3 d^2} \; , &\hspace{1cm}&
\frac{\partial V_{\rm NFW}(d)}{\partial d}
= \frac{16 \left(\ln\left(1+d\right) - \frac{d}{1+d}\right)}{d^2}
\; . \nonumber
\end{eqnarray}
At this point it is also convenient to rewrite the Vlasov equation,
Eq.~(\ref{eq:Vla1}), in a dimensionless form. As already mentioned,
defining a pattern mass $M_0$ and a pattern distance $R_0$ allows us
to define a pattern density $\rho_0$ and from it a pattern velocity
$\sigma_0$. Considering that we are dealing with a single type of
particles characterized by a mass $m$, we can continue this idea and
construct all the needed pattern quantities, {\it i.e.} a pattern time
$T_0=R_0/\sigma_0$, a pattern radial momentum $p_0=m \sigma_0$, and a
pattern angular momentum $l_0=m\,\sigma_0\,R_0$. Rewriting then the
time as $t=T_0\,{\cal T}$, the radial momentum as $p_r=p_0\,{\cal
P}_r$, and the angular momentum as $L=l_0\,{\cal L}$, we obtain the
following dimensionless Vlasov equation:
\begin{equation}
\frac{\partial \cal{F}}{\partial {\cal T}} = - {\cal P}_r \,
\frac{\partial \cal{F}}{\partial d} + \left(\frac{\partial
V(d)}{\partial d} - \frac{{\cal L}^2}{d^3}\right) \frac{\partial
\cal{F}}{\partial {\cal P}_r} \; ,
\label{eq:Vla_dimless}
\end{equation}
where we have defined $f=f_0\,\cal{F}$, with $\cal{F}$ a dimensionless
distribution function, and $f_0$ a normalization constant with units of probability density in phase
space.
For each halo model, we use the dimensionless gradient of the
gravitational potential given by Eqs.~(\ref{eqs:phisd}), in order to
evolve the distribution function. The corresponding dimensionless
potentials are given by:
\begin{eqnarray}
&& V_{\rm iso}(d) = 3 \ln(d) \; , \hspace{7cm}
V_{\rm iso-tr}(d) = \frac53 \left( \ln(1+d^2)
+ 2 \frac{\arctan(d)}{d} \right)\,, \nonumber \\
&& \label{eqs:phis} \\
&& V_{\rm Burkert}(d) = \frac{10}{3 d} \left(2 \left(1+d\right)
\left( \arctan\left(d\right) - \ln \left(1+d\right) \right)
- \left(1-d\right) \ln\left(1+d^2\right) \right) \; ,
\hspace{1cm}
V_{\rm NFW}(d) = -16 \frac{\ln(1+d)}{d} \; . \nonumber
\end{eqnarray}
The dimensionless effective potential is obtained by adding the
centrifugal term. Fig (\ref{Fig:Phis_ef}) shows the plots for the
effective potential in each case.
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.60]{fig02.pdf}}
\caption{Plots of the effective dimensionless potential of the four
halos considered with ${\cal L}=3.5$.}
\label{Fig:Phis_ef}
\end{figure}
The profiles for the potential gradients are shown in
Figure~(\ref{Fig:dPhis}), and the corresponding graphs for the
gradient of the effective potential (including the centrifugal term)
are shown in Figure~(\ref{Fig:dPhis_ef}).
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.60]{fig03.pdf}}
\caption{Plots for minus the derivative of the dimensionless
gravitational potential $V(d)$ for each of the four halos
considered.}
\label{Fig:dPhis}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.60]{fig04.pdf}}
\caption{Plots for the derivative of the effective dimensionless
gravitational potential for the four halos considered with ${\cal
L}=3.5$.}
\label{Fig:dPhis_ef}
\end{figure}
\section{Numerical code}
\label{Sec3}
In this section, we describe our \textit{Vlaso-ollin} code, which evolves the
distribution function by means of the Vlasov equation, Eq.(\ref{eq:Vla_dimless})
for several background gravitational potentials.
\subsection{Flux conservative methods}
The Vlasov equation~\eqref{eq:Vla_dimless} can be trivially rewritten
in conservation form as:
\begin{equation}
\label{eq:Vla_cons}
\frac{\partial \cal{F}}{\partial {\cal T}} + \frac{\partial}{\partial d}
\left[ {\cal P}_r {\cal F} \right] + \frac{\partial}{\partial {\cal P}_r}
\left[ \left(\frac{\partial V(d)}{\partial d}
- \frac{{\cal L}^2}{d^3} \right) {\cal F} \right] = 0 \; .
\end{equation}
We will now define the fluxes in the $d$ and ${\cal P}_r$ directions,
$F_d$ and $F_{{\cal P}_r}$, as
\begin{equation}
F_d \equiv {\cal P}_r {\cal F} \; , \quad
F_{{\cal P}_r} \equiv \frac{\partial V_{\rm eff}(d)}{\partial d} {\cal F} \; ,
\end{equation}
where as before
\begin{equation}
V_{\rm eff} \equiv V(d) + \frac{1}{2}\frac{{\cal L}^2}{d^2} \; .
\end{equation}
The Vlasov equation then becomes
\begin{equation}
\frac{\partial \cal{F}}{\partial \cal{T}}
+ \frac{\partial F_d}{\partial d}
+ \frac{\partial F_{{\cal P}_r}}{\partial {\cal P}_r} = 0 \; .
\end{equation}
We have written a numerical code to solve the last equation using a
flux conservative finite difference method~\cite{Leveque92}. A
crucial consequence of a using conservative method for the Vlasov
equation is that the numerical integration will conserve the total
number of particles $N$ exactly (up to machine round-off error), so
that any change in the number of particles will be the result of
particles leaving (or entering) the computational domain through the
boundaries.
In our code phase-space is discretized in a rectangular grid of size
$N_r \times N_p$, with $0 \leq d \leq d_{max}$ and \mbox{$-{\cal
P}_{max} \leq {\cal P} \leq {\cal P}_{max}$}, and with grid sizes
$\Delta d$ and $\Delta {\cal P}$. The code uses a method of lines,
with a three-step iterative Crank-Nicholson time integrator. We use a
flux-limiter method for the calculation of the fluxes $F_d$ and
$F_{{\cal P}_r}$ at the cell interfaces. The limiter methods minmod,
van Leer, superbee and monotonized-centered (MC)~\cite{Leveque92} were
tested for the special case of an external potential corresponding to
a constant density star, and particles with zero angular momentum, and
the MC limiter was found to show the best results in both preserving
the number of particles and preserving the form of the density profile
for stationary solutions: minmod is very diffusive in the sense that
stationary profiles tend to diminish in amplitude and become wider
(while conserving the integral), while the van Leer and superbee
limiters, though less diffusive, in general deform the stationary
profiles. The reason why we found it necessary to use sophisticated
high resolution flux-limiter methods is because such methods can be
shown to be total-variation-diminishing (TVD), and this guarantees
that the solution has no spurious oscillations. This ensures that in
regions where the distribution function $f$ is close to zero it
remains well behaved and does not become negative (which would be
unphysical, but can easily happen with numerical methods that are not
TVD). The limiter methods we use are second order accurate everywhere
except at extrema
of the distribution function, where they become only
first order.
\subsection{Boundary conditions}
There are two different types of boundaries to consider in our
simulations: the external boundaries of the computational domain at
$|{\cal P}|={\cal P}_{max}$ and $d=d_{max}$, and the internal boundary
at $d=r=0$.
For the external boundaries the boundary condition we use depends on
the sign of the coefficient of the corresponding derivative. That is,
in the radial direction we consider the sign of the momentum ${\cal
P}_r$: For negative values of ${\cal P}_r$, corresponding to
particles that would enter the computational domain, we just set the
distribution function $f$ to zero, while for positive values of ${\cal
P}_r$, corresponding to particles leaving the domain, we calculate
the fluxes all the way to the boundary using one-sided differences.
In the momentum directions we do the same thing, but in this case we
consider the sign of the force term $\partial V_{\rm eff}(d)/\partial
d$.
The boundary at $d=r=0$ is of a different type and corresponds to the
origin of the radial coordinate. Notice that at $r=0$ the centrifugal
term becomes singular, and the gravitational force can also become
singular for some of the halo models we are considering (the
isothermal and NFW models). To avoid dealing with singular quantities
we use a finite differencing grid that staggers the origin, so that
the first grid point is located at $d = \Delta d/2$. We add a
fictitious point at $d = -\Delta d/2 $ in order to be able to
calculate derivatives at $d=0$. This fictitious point also allows us
to impose adequate parity conditions on the different quantities.
Now, in the case of non-zero angular momentum the particles are never
allowed to reach the origin so that the distribution function $f$
should be zero there. But in the case of zero angular momentum there
is nothing to stop them from reaching $r=0$. If moreover, the
gravitational potential is regular there (as in the case of the
truncated-isothermal and Burkert models), the particles that reach the
origin can in principle just come back on the opposite side and
oscillate around the origin. This is in fact an interesting case, as
particles that reach the origin with a given negative radial momentum
will move back out of the origin with a positive radial momentum. In
order to capture this behavior we use the fictitious points at $d =
-\Delta d/2 $ to impose the boundary condition $f(-d,-{\cal
P}_r)=f(d,{\cal P}_r)$. We find that when doing so, particles
with zero angular momentum in a regular potential just describe orbits
around the origin of phase space, in a very similar way as one would
expect for the case of a simple harmonic oscillator.
When the angular momentum is non-zero, particles coming from infinity
should reach a finite minimum radius, so in principle one would not
need a special boundary condition at $d=0$. In practice, however, we
have found that in this case the distribution function $f$ might still
attain small non-zero values at $d=r=0$ due to numerical errors, and
using the boundary condition described above results in robust
evolutions.
There is a final important point that should be made about the
behavior of the distribution function $f$ at the origin for the case
of zero angular momentum. From the definition of the particle density
$\rho_f$ and the momentum density $j_f$, equations~\eqref{eq:rho1}
and~\eqref{eq:curr1}, one can easily see that if the distribution
function $f$ does not vanish at the origin then $\rho_f$ and $j_f$ will
be singular there (in fact, $f$ must vanish at least as $r^2$). This
actually makes perfect physical sense: if an infalling spherical shell
of matter reaches the origin at the same time, then the matter density
will clearly become infinite there as all the particles are now at a
single point. For an external (regular) gravitational potential this
represents no problem since we can just interpret the distribution
function statistically: every individual particle behaves
independently of the others and will just oscillate around the origin.
But if we were to consider the case of self-gravitating particles,
then as they reach the origin the mass density will become infinite
there, so that the gravitational potential will become singular, and
our description in terms of the Vlasov equation will break down. All
the simulations shown below will be for the case of an external
gravitational potential (the particles have no self-gravity) and
non-zero angular momentum, so that this problem will never arise, but
it is something that should be kept in mind for the study of systems
of self-gravitating particles with the Vlasov equation.
\subsection{CFL condition}
Since we are using an explicit scheme, in order to get a stable
evolution we need to satisfy the Courant-Friedrichs-Lewy (CFL)
condition. During the evolution we then choose the
time step as $\Delta t=\min(\Delta t_d,\Delta t_p)$, with
\begin{equation}
\Delta t_d = \frac{C\Delta d}{{\cal P}_{max}} \; , \qquad
\Delta t_p = \frac{C\Delta {\cal P}}{|F|_{max}} \; .
\end{equation}
where $F \equiv \partial V_{\rm eff}(d)/\partial d$ and $C$ a number
of order 1 which in two dimensions must be such that $C\leq
1/\sqrt{2}$. In all the simulations shown below we choose $C=0.7$ and
$P_{max}=6$. The maximum value of the force $F_{max}$ is calculated a
$t=0$ and does not change in time since the background potential is
fixed.
The fact that the centrifugal term ${\cal L}/d^3$ in the effective
force diverges at $d=0$ is a serious problem and forces us to use very
small values for $\Delta t$ in order to satisfy the CFL condition. In
order to avoid this problem we introduce a small parameter $\epsilon$
in the denominator that modifies the centrifugal potential for small
values of $d$, and the modified centrifugal force is obtained by deriving the new potential.
\begin{eqnarray}
\frac{{\cal L}^2}{d^2} &\longrightarrow& \frac{{\cal
L}^2}{d^2+\epsilon^2} \; , \nonumber \\
\frac{{\cal L}^2}{d^3} &\longrightarrow& \frac{{\cal
L}^2\,d}{(d^2+\epsilon^2)^2} \; .
\end{eqnarray}
For a given value of the angular momentum ${\cal L}$ we choose $\epsilon$ in
such a way as to guarantee that the centrifugal force is minimally
modified at the smallest radius that can be reached by a particle with
that angular momentum. Assuming that the particle starts far away
with initial momentum ${\cal P}_i$, and that for small values of $d$ the
centrifugal force dominates over the gravitational force, then
conservation of energy implies that the minimum radius the particle
can reach is such that ${\cal L}^2/d_{\rm min}^2 \sim {{\cal P}_i}^2 / 2 m$, which
implies $d_{\rm min} \sim \sqrt{2 m} \; {\cal L} / | {\cal P}_i |$. In our
simulations we typically take $\epsilon=d_{min}/10$.
This is clearly not an ideal solution as in involves modifying the
effective force so that a small error is introduced for small radii,
and in practice still results in values of $\Delta t$ that are much
smaller than $\Delta d$, so that our code is rather slow. We are
exploring ways to improve our code, perhaps by using an operator
splitting method with an implicit scheme for the centrifugal term. We
will report on this elsewhere.
\subsection{Initial data}
For our simulations we consider an initial localized distribution of
particles in phase space, in the background gravitational potential of
the different halo models. We interpret this as a inhomogeneity in
the dark matter halo, and use the Vlasov equation to determine the
final state of such perturbation.
We consider an initial phase space distribution $f$ given as
\begin{equation}
f(t=0) = \frac{N_0}{8 \pi^3 L_0 \sigma_d \sigma_{\cal P}} \left(
e^{-\frac{(d-d_0)^2}{{\sigma_d}^2}}
e^{-\frac{({\cal P}_d-{{\cal P}_r}_0)^2}{{\sigma_{\cal P}}^2}} +
e^{-\frac{(d+d_0)^2}{{\sigma_d}^2}}
e^{-\frac{({\cal P}_d+{{\cal P}_d}_0)^2}{{\sigma_{\cal P}}^2}} \right) \; .
\label{eq:f0}
\end{equation}
This is normalized so that the total number of particles is $N_0$, and
in all our simulations we take $N_0=1$. Notice that, when the angular
momentum is zero, the normalization factor in the initial distribution
is instead $1/(4 \pi^2 \sigma_r \sigma_{\cal P})$.
The above initial data represents a spherical shell of particles, all
of which have the same angular momentum $L_0$, with a Gaussian
distribution centered at a radius $d_0$ and radial momentum ${{\cal P}_d}_0$,
with widths $\sigma_d$ and $\sigma_{\cal P}$ respectively. The initial data
is constructed so that the symmetry $(d,{\cal P}) \to (-d,-{\cal P})$ discussed
above when we described the boundary conditions is preserved
(incidentally, this form of the initial data also guarantees that the
normalization is exact when integrating over the radial coordinate
from 0 to infinity). We remind the reader at this point that even if
all the particles have the same angular momentum $L_0$, by
construction their individual motions are uniformly distributed in all
angular directions, so that the overall spherical symmetry is
preserved. For the simulations shown below we take the following
parameters for the initial data: $d_0=3.0$, ${{\cal P}_d}_0=-0.5$,
$\sigma_d=0.5$, and $\sigma_{\cal P}=0.25$. The corresponding initial data is
shown in Figure~(\ref{Fig:f0}).
\begin{figure}[H]
\centering
\includegraphics[scale=0.45,angle=0]{fig05.pdf}
\caption{Initial distribution function in phase space for $N_0=1$,
$d_0=3.0$, ${{\cal P}_d}_0=-0.5$, $\sigma_d=0.5$, and $\sigma_{\cal P}=0.25$.}
\label{Fig:f0}
\end{figure}
\section{Results}
\label{Sec4}
The time evolution of the initial distribution function $\mathcal{F}$
is given by solving equation~\eqref{eq:Vla_cons} in the numerical
mesh. We perform a series of simulations for different values of the
angular momentum, $L=2.0,2.5,3.0,3.5$. We have not considered smaller
values of the angular momentum because in those cases we have found
that the particles fall very rapidly to the center, while larger vales
of the angular momentum result in the particles escaping the
computational domain. In all our evolutions we also fix the mass of
the individual particles to $m=1$.
For the simulations shown below we considered two different
resolutions $\Delta d = \Delta {\cal P} = 0.05$ and $\Delta d = \Delta
{\cal P} = 0.1$. The boundaries of the computational domain
are located at $d_{max}=20$ and ${\cal P}_{max}=6$. This range is
chosen so that at the higher resolution we use only a very small
fraction of the particles has escaped through the boundaries at the
end of the simulation. As we will see below, that particles that
escape do so mostly because of numerical dissipation.
\subsection{Time evolution for $\mathcal{L}=3.5$}
As an example of the dynamics of the Vlasov equation, in this section
we will concentrate on presenting the results of the time evolution of
the distribution function for the case of our largest value of the
angular momentum, namely $\mathcal{L}=3.5$.
Figures~\ref{pics:PS_iso}-\ref{pics:PS_NFW} show the evolution of the
distribution function for our four different halo models. In each
figure we show six snapshots at different times during the evolution.
The first five snapshots correspond to the same times in all cases,
namely $\mathcal{T}=0,9.33, 18.84, 30.79,45.71$. The time
corresponding to the last snapshot differs for each halo model because
in each case we show an epoch when the system has finally reached a
stationary state. Figure~\ref{Fig:FD_all_ff} shows again the final
stationary distribution function in phase space for the four different
models. We can see that all four halo models the distribution
function reaches a stationary state that corresponds to orbits in
phase space around a central point. This is not surprising, since
individual particles with a given angular momentum are expected to
have orbits around the center of gravity with some minimum and maximum
radii (the individual orbits will not be precisely elliptical since
the potentials for the different halos are not simply $1/d$). The
stationary states look similar for each halo model, even if they
are not reached at the same time; for the isothermal, Burkert and
NFW models the stationary state is reached at time $\mathcal{T} \sim 120$
and for the truncated isothermal this state is reached at time $\mathcal{T} \sim 260$.
Next, in Figures~\ref{pics:D_iso}-\ref{pics:D_NFW} we show the time
evolution of the integrated particle density $\rho_f(d)$
for the four
halo models and the same times as before. Again we see how even
though the initial stages of the evolution are different for each halo
model, the final stationary particle density is quite similar, with a
characteristic two-hump shape corresponding to particles that describe
an orbit in phase space and accumulate mostly at the extrema of those
orbits as seen in physical space.
Even though we have only shown the case with angular momentum
$\mathcal{L}=3.5$, simulations with different values of the angular
momentum behave in a very similar way.
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.34]{fig06.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig07.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig08.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig09.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig10.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig11.pdf}}
\caption{Time evolution of the distribution function $\mathcal{F}$ for
the isothermal model with angular momentum $\mathcal{L}=3.5$. The
different panels correspond to times $\mathcal{T}=0, 9.33, 18.84,
30.79, 45.71, 200.39$.}
\label{pics:PS_iso}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.34]{fig06.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig12.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig13.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig14.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig15.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig16.pdf}}
\caption{Time evolution of the distribution function $\mathcal{F}$ for
the truncated isothermal model with angular momentum
$\mathcal{L}=3.5$. The different panels correspond to times
$\mathcal{T}=0, 9.33, 18.84, 30.79, 45.71, 259.38$.}
\label{pics:PS_trun}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.34]{fig06.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig17.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig18.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig19.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig20.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig21.pdf}}
\caption{Time evolution of the distribution function $\mathcal{F}$ for
the Burkert model with angular momentum $\mathcal{L}=3.5$. The
different panels correspond to times $\mathcal{T}=0, 9.33, 18.84,
30.79, 45.71, 196.85$.}
\label{pics:PS_bur}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.34]{fig06.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig22.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig23.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig24.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig25.pdf}}
\subfigure{\includegraphics[scale=0.34]{fig26.pdf}}
\caption{Time evolution of the distribution function $\mathcal{F}$ for
the NFW model with angular momentum $\mathcal{L}=3.5$. The different
panels correspond to times $\mathcal{T}=0, 9.33, 18.84, 30.79, 45.71, 189.20$.}
\label{pics:PS_NFW}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.35]{fig27.pdf}}
\subfigure{\includegraphics[scale=0.35]{fig28.pdf}}
\subfigure{\includegraphics[scale=0.35]{fig29.pdf}}
\subfigure{\includegraphics[scale=0.35]{fig30.pdf}}
\caption{Final virialized distribution function for the four different
halo models with angular momentum $\mathcal{L}=3.5$. Starting in the
top left panel: isothermal, truncated isothermal, Burkert and
NFW.}
\label{Fig:FD_all_ff}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.2, angle=0]{fig31.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig32.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig33.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig34.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig35.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig36.pdf}}
\caption{
Time evolution of the integrated particle density $\rho_f(d)$ multiplied for the
$d^2$ factor for the isothermal model with angular momentum
$\mathcal{L}=3.5$.
The different panels correspond to times $\mathcal{T}=0, 9.33, 18.84,
30.79, 45.71, 200.39$, same that those in
Figure~\ref{pics:PS_iso}.}
\label{pics:D_iso}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.2, angle=0]{fig37.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig38.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig39.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig40.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig41.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig42.pdf}}
\caption{Time evolution of the integrated particle density $\rho_f(d)$
multiplied for the $d^2$ factor
for the isothermal truncated model with angular momentum
$\mathcal{L}=3.5$.
The different panels correspond to times
$\mathcal{T}=0, 9.33, 18.84, 30.79, 45.71, 259.38$, same that those in
Figure~\ref{pics:PS_trun}.}
\label{pics:D_trun}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.2, angle=0]{fig43.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig44.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig45.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig46.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig47.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig48.pdf}}
\caption{Time evolution of the integrated particle density $\rho_f(d)$
multiplied for the $d^2$ factor
for the Burkert model with angular momentum $\mathcal{L}=3.5$.
The different panels correspond to times $\mathcal{T}=0, 9.33, 18.84,
30.79, 45.71, 196.85$, same that
those in Figure~\ref{pics:PS_bur}.}
\label{pics:D_bur}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.2, angle=0]{fig49.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig50.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig51.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig52.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig53.pdf}}
\subfigure{\includegraphics[scale=0.2, angle=0]{fig54.pdf}}
\caption{Time evolution of the integrated particle density $\rho_f(d)$
multiplied for the $d^2$ factor
for the NFW model with angular momentum $\mathcal{L}=3.5$.
The different
panels correspond to times $\mathcal{T}=0, 9.33, 18.84, 30.79, 45.71, 189.20$,
same that those in Figure~\ref{pics:PS_NFW}.}
\label{pics:D_NFW}
\end{figure}
\subsection{One particle motion}
In order to understand the results from the previous section, it is
perhaps instructive to consider the motion of a single particle in a
gravitational potential determined by the different halo models
considered.
For a particle starting at an initial radius $d_0$, with initial
radial momentum ${{\cal P}_d}_0$, and with a given angular momentum $L$,
conservation of energy implies that
\begin{equation}
\frac{{{\cal P}_d}^2}{2 m} + \frac{{\cal L}^2}{2 m d^2} + \Phi(d) =
\frac{{{{\cal P}_d}_0}^2}{2 m} + \frac{{\cal L}^2}{2 m {d_0}^2} + \Phi(d_0) \; .
\end{equation}
From this one can determine the turning points of the orbit, or the
radius of a circular orbit for a given value of the angular
momentum.
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{2}{*}{$\cal L$} & \multirow{2}{*}{$V$} & \multicolumn{2}{c|}{Turning points} & $\rho_{max}$ \\ \cline{3-5}
& & $d_{min}$ & $d_{max}$ & $d$ \\ \hline
2 & Isothermal & 1.03 & 3.18 & 0.8 \\ \hline
2 & Iso. Trun. & 1.59 & 3.33 & 1.25 \\ \hline
2 & Burkert & 1.16 & 3.19 & 0.92 \\ \hline
2 & NFW & 0.95 & 3.15 & 0.75 \\ \hline
2.5 & Isothermal & 1.44 & 3.23 & 1.02 \\ \hline
2.5 & Iso. Trun. & 2.06 & 3.49 & 1.67 \\ \hline
2.5 & Burkert & 1.56 & 3.25 & 1.17 \\ \hline
2.5 & NFW & 1.31 & 3.19 & 0.97 \\ \hline
3 & Isothermal & 1.88 & 3.33 & 1.32 \\ \hline
3 & Iso. Trun. & 2.46 & 3.85 & 2.02 \\ \hline
3 & Burkert & 2.00 & 3.37 & 1.47 \\ \hline
3 & NFW & 1.72 & 3.26 & 1.23 \\ \hline
3.5 & Isothermal & 2.35 & 3.58 & 1.65 \\ \hline
3.5 & Iso. Trun. & 2.69 & 4.54 & 2.32 \\ \hline
3.5 & Burkert & 2.41 & 3.67 & 1.77 \\ \hline
3.5 & NFW & 2.17 & 3.41 & 1.52 \\ \hline
\end{tabular}
\caption{Positions of the turning points for each configuration of angular moment and halo potential and position of the maximum
density value at the final virialized distribution function.}
\label{tbl:turning points}
\end{table}
It is interesting to note
that the turning points for a single
particle with initial position and momentum corresponding to the
maximum of the initial phase space distribution does not coincide with
the position of the maxima of the integrated energy density in the
equilibrium state (Table \ref{tbl:turning points}). The reason for this, may be the so-called
statistical pressure: even though the particles are non interacting,
it seems that the collective motion is different from the individual
motion of particles initially at the center of the distribution.
\subsection{Virialization}
As mentioned above, in all our simulations the evolution of the
distribution function eventually reaches a stationary state. Such a
state should satisfy the virial theorem, equation~\eqref{eq:virial2}.
In order to verify this we compute the average of the radial kinetic
energy $\left <{{\cal P}_d}^2/2m \right>$, and the average of the virial
$\left< F_{\rm eff}(d)\,d \right>$, and compare them during the
evolution.
We find that in each case, as the initial distribution evolves toward
a stationary state, the average of the radial kinetic energy and the
average of the virial both reach constant values that satisfy the
virial theorem~\eqref{eq:virial2}.
Figures~\ref{Fig:virialIso}-\ref{Fig:virialNFW} show plots of $\left<
{{\cal P}_d}^2/2m \right>$ and $\left< - F_{\rm eff}(d)\,d/2 \right>$ for each
of the different halo models, and for four values of the angular
momentum in each case. We find that in each case, after strong
initial oscillations, the average kinetic energy and virial reach a
constant value satisfying the virial theorem. We also find that this
virialization process takes longer for larger values of the angular
momentum, while that the final constant value of the average kinetic
energy becomes smaller. Notice also that while $\left< {{\cal P}_r}^2/2m
\right>$ is always positive, $\left< - F_{\rm eff}(d)\,d/2 \right>$ can
become negative during the initial portions of the evolution.
\begin{figure}[H]
\centering
\includegraphics[scale=0.65]{fig55.pdf}
\caption{
Virialization for the case of the isothermal model for
different values of the angular momentum \mbox{$\mathcal{L}=2.0,2.5,3.0,3.5$}.}
\label{Fig:virialIso}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.65]{fig56.pdf}
\caption{
Virialization for the case of the truncated isothermal model
for different values of the angular momentum \mbox{$\mathcal{L}=2.0,2.5,3.0,3.5$}.}
\label{Fig:virialTrun}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.65]{fig57.pdf}
\caption{
Virialization for the case of the Burkert model for different
values of the angular momentum \mbox{$\mathcal{L}=2.0,2.5,3.0,3.5$}.}
\label{Fig:virialBurk}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.65]{fig58.pdf}
\caption{
Virialization for the case of the NFW model for different
values of the angular momentum \mbox{$\mathcal{L}=2.0,2.5,3.0,3.5$}.}
\label{Fig:virialNFW}
\end{figure}
Figure~\ref{Fig:virialcompare} shows a comparison of the virialization
process for the four different halo models in the specific case with
angular momentum ${\cal L}=3.5$. We can see that the NFW model
virializes very rapidly, while the truncated isothermal model takes
much longer than all other models to virialize. Again we see that
models that take longer to virialize reach smaller values of the final
average kinetic energy.
\begin{figure}[H]
\centering
\includegraphics[scale=0.65]{fig59.pdf}
\caption{Comparison of the virialization process for the different
halo models in the case with $\mathcal{L}=3.5$.}
\label{Fig:virialcompare}
\end{figure}
\subsection{Conservation of particles}
As mentioned above, since we are using a conservative numerical method,
the total number of particles defined in~\eqref{eq:totalN} should be
conserved up to machine round off error in our simulations. Changes
in the total number of particles should reflect only the particles
that leave the computational domain through the external boundaries.
We find, however, that numerical diffusion can cause particles to
reach the boundaries that would not do so physically.
In order to illustrate this behavior, in Figures~\ref{pics:Npart_low}
and~\ref{pics:Npart_high}
we show the time evolution of the total
number of particles integrated over the whole computational domain,
for the different halo models and different values of the angular
momentum (we remind the reader that this number has been normalized to
$1$ in the initial data). Figure~\ref{pics:Npart_low} shows the results
for low resolution runs with $\Delta d = \Delta {\cal P} = 0.1$,
while Figure~\ref{pics:Npart_high} shows the results for the higher
resolution runs with $\Delta d = \Delta {\cal P} = 0.05$.
There are several things to notice from the figures. First, at large
times we are indeed loosing particles through the boundaries. A more
detailed analysis of the fluxes through the different boundaries shows
that we are loosing them mostly through the boundary at
negative momenta. Second, even though at the
lower resolution the loss of particles is very significant by the end
of the simulations (in one case we loose as much as $70\%$ of the
particles), at the higher resolution this loss is much lower (in the
worst case only about $8\%$), which indicates that this loss of
particles is mostly due to numerical diffusion which becomes smaller
at higher resolutions.
Considering only the results from the higher resolution,
Figure~\ref{pics:Npart_high}, we notice that more particles are lost
for low values of the angular momentum (for ${\cal L}=2.0$ the NWF
halo looses $8\%$ of the particles), while for larger angular momentum
the loss of particles becomes extremely small (for ${\cal L}=3.5$ the
NFW halo looses only about $0.0007\%$). We also notice that the NFW
halo is the one which has more trouble keeping the particles inside
the computational domain whereas the truncated isothermal is the one
which best preserves the total number of particles.
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.4]{fig60.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig61.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig62.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig63.pdf}}
\caption{Total number of particles integrated over the complete
computational domain for the different halo models, and different
values of the angular momentum ${\cal L}$, for a resolution of $\Delta
d = \Delta {\cal P} = 0.1$.}
\label{pics:Npart_low}
\end{figure}
\begin{figure}[H]
\centering
\subfigure{\includegraphics[scale=0.4]{fig64.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig65.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig66.pdf}}
\subfigure{\includegraphics[scale=0.4]{fig67.pdf}}
\caption{Total number of particles integrated over the complete
computational domain for the different halo models, and different
values of the angular momentum ${\cal L}$, for a resolution of $\Delta
d = \Delta {\cal P} = 0.05$.}
\label{pics:Npart_high}
\end{figure}
\section{Discussion}
\label{Sec5}
We have studied the (Newtonian) Vlasov equation in spherical symmetry
for the case of particles moving in a background gravitational
potential. We have shown that in this case the problem effectively
reduces to that of a three-dimensional phase-space corresponding to
the radial coordinate $r$, the radial momentum $p_r$, and the total
angular momentum $L$. Moreover, since the angular momentum is
conserved, it can effectively
be decoupled from the other phase space
coordinates, so that one can consider the motion of particles all of
which have the same value of the angular momentum $L_0$,
reducing the problem to a two-dimensional case. We have
constructed a numerical code to solve directly the Vlasov
equation given several values of angular momenta $L_0$
using a conservative TVD (flux limiter) scheme.
We have used our code to study the evolution of an initial localized
distribution of particles in phase space in the background
gravitational potential of four different models for a dark matter
halos, namely the isothermal, truncated isothermal, Burkert and NFW
models. We interpret this initial distribution as a inhomogeneity in
the dark matter halo. Knowing that the
resulting Vlasov equation implies the continuity equation and also
the standard virial theorem for stationary solutions, we have used
these results for testing the corresponding temporal evolution.
We find that, for all the cases considered,
during evolution the initial distribution leads to non-trivial
stationary final distributions that satisfy the virial theorem. The
detailed properties of such final distributions change from one type of
halo to another, but the overall features are very similar. We believe
the detailed study of the evolution of such inhomogeneities could be
relevant to characterize the observed halos.
Even though the virial theorem has to be satisfied for equilibrium
states, by itself it does not tell us under what conditions an arbitrary
initial configuration will evolve toward a stationary configuration,
nor how such an evolution will develop. It is well known that an
arbitrary distribution that depends only on conserved quantities like
the total energy $E$ and angular momentum $L$, $f=f(E,L)$, is
automatically a stationary solution of the Vlasov equation. However,
for arbitrary initial configurations such as those studied here, we
can't predict a priori what stationary solution will be reached, or
even if such a stationary situation will be reached. As we have seen,
the increase of entropy is not a useful concept in this case, since
the Vlasov equation preserves entropy. We therefore believe that the
use of ergodic theory might be useful in this case.
It is important to mention that the dynamical description of dark
matter when seen from the point of view of kinetic theory, as has been
done here, differs from the standard description in terms of N-body
simulations \cite{Harker:2005um}, though of course both descriptions should coincide in the
limit of a very large number of bodies~\cite{Colombi:2015eia}
Still, a statistical description in terms of kinetic theory allows us to
define a continuous distribution of matter, as opposed to a discrete
number of point particles, which could have important advantages and
in particular would be easier to generalize to the case of general
relativity (where point particles are conceptually problematic).
Also, our description here is very different from other approaches
that consider dark matter as a type of fluid as
in~\cite{Barranco:2013wy}, or as a ultra-light scalar field as
in~\cite{Burt:2011pv,Barranco:2012qs,Barranco:2013rua}.
The results obtained and presented in this manuscript are encouraging,
showing that dark matter inhomogeneities can be well described by
kinetic theory, leading to non-trivial equilibrium configurations.
With these results, there are two clear avenues for further research.
The first is to consider the case of particles with a distribution of
different values of the angular momentum, which would require a
three-dimensional code for the phase space coordinates ($r,p_r,L$).
On the other hand, one can consider the case of a system of
self-gravitating particles, which would require us to solve the
Poisson equation to obtain the gravitational potential. And of
course, one can move away from the simple Newtonian description to a
general relativistic one (see
e.g.~\cite{MartinGarcia:2001nh,Akbarian:2014gna}). All these
problems are currently being investigated, and we will report their
progress in the near future.
\bibliographystyle{bibtex/prsty}
|
\section{Introduction}\label{sec:intro}
One of the primary goals of the study of Quantum Chromodynamics (QCD) is an understanding of the all orders behavior of observables, traditionally formalized through either an operator product expansion (OPE)~\cite{Wilson:1969zs} or factorization theorems~\cite{Collins:1985ue,Collins:1988ig,Collins:1989gx}. For observables that can be handled with an OPE, a lot is known about the form of power corrections. Examples include deep inelastic scattering where the OPE has been carried out to twist-4~\cite{Jaffe:1981td,Jaffe:1982pm,Ellis:1982wd,Ellis:1982cd}, inclusive B-decays where the OPE is known to ${\cal O}(1/m_b^4)$~\cite{Mannel:1993su}, and Quarkonia production and decay, see~\cite{Brambilla:2010cs} for a review. The description of observables with more complicated dynamics typically relies on factorization theorems and much less is known about the structure of power corrections in these cases. Power corrections have been considered for Drell-Yan~\cite{Qiu:1990xxa,Qiu:1990xy,Korchemsky:1994is,Beneke:1995pq,Korchemsky:1996iq} at ${\cal O}(\Lambda_{\rm QCD}^2/Q^2)$, for inclusive $B$ decays in the endpoint region at $\mathcal{O}((1-z)^0,(\Lambda_{\rm QCD}/m_b) ^{1,2})$~\cite{Bauer:2001mh,Leibovich:2002ys,Bauer:2002yu,Lee:2004ja,Mannel:2004as,Bosch:2004cb,Beneke:2004in,Tackmann:2005ub,Benzke:2010js}, for exclusive $B$ decays at $\mathcal{O}(\Lambda_{\rm QCD}/m_b)$~\cite{Mantry:2003uz,Blechman:2004vc,Beneke:1999br,Beneke:2000ry,Keum:2000ph,Keum:2000wi,Bauer:2004tj,Arnesen:2006vb,Arnesen:2006dc}, for event shapes $\tau$ in $e^+e^-$, $ep$, and $pp$ collisions at ${\cal O}(\Lambda_{\rm QCD}^k/(Q\tau)^k)$ \cite{Korchemsky:1994is,Dokshitzer:1995qm,Korchemsky:1995zm,Dokshitzer:1995zt,Dokshitzer:1997ew,Dokshitzer:1997iz,Korchemsky:1997sy,Korchemsky:1998ev,Korchemsky:1999kt,Korchemsky:2000kp,Belitsky:2001ij,Berger:2003pk,Lee:2006fn,Lee:2006nr,Hoang:2007vb,Gehrmann:2012sc}, and at $\mathcal{O}((1-z)^0)$ for threshold resummation \cite{Dokshitzer:2005bf,Grunberg:2007nc,Laenen:2008gt,Laenen:2008ux,Grunberg:2009yi,Laenen:2010uz,Almasy:2010wn,Bonocore:2014wua,White:2014qia,deFlorian:2014vta,Bonocore:2015esa,Bonocore:2016awd}.
A convenient formalism for studying factorization in QCD is the Soft Collinear Effective Theory (SCET) \cite{Bauer:2000ew, Bauer:2000yr, Bauer:2001ct, Bauer:2001yt}, an effective field theory describing the soft and collinear limits of QCD. SCET allows for a systematic power expansion at the level of the Lagrangian, and simplifies many aspects of factorization proofs \cite{Bauer:2002nz}. SCET has been used to study power corrections at the level of the amplitude \cite{Larkoski:2014bxa} and to derive factorization theorems for $B$ decays using subleading power operators (eg.~\cite{Bauer:2002aj,Beneke:2003pa,Mantry:2003uz,Blechman:2004vc,Bauer:2004tj,Lee:2004ja,Hill:2004if,Bosch:2004cb,Beneke:2004in,Benzke:2010js}), where many interesting processes only start at subleading power. More recently, progress has been made towards understanding the subleading factorization and resummation of the event shape thrust in $e^+e^-$ \cite{Freedman:2013vya,Freedman:2014uta}. Such subleading factorization theorems are technically cumbersome, and significant work is still required to gain a simplified and more complete understanding.
In this paper we consider the formalism required to study subleading factorization theorems in SCET, focusing in particular on subleading hard scattering operators. Using the results of \Ref{Kolodrubetz:2016uim}, we further develop and explore a set of SCET helicity operator building blocks which is valid for constructing operator bases at any order in the power expansion. These operators extend the leading power basis of \Ref{Moult:2015aoa}, where it was shown that the use of helicity operators greatly simplifies the construction of operator bases, as well as matching calculations, for processes with many final state jets (see also \cite{Moult:2014pja} for an application of helicity operators to Higgs processes, where they were used to simplify the matching to fixed order helicity amplitudes). As we will see, helicity operators can also be used to simplify the construction of subleading power bases of hard scattering operators, where multiple fields can appear in the same collinear sector. After reviewing the helicity operator building blocks, we will focus on the case of hard scattering operators involving two back-to-back collinear sectors, which is relevant for proving subleading factorization theorems for a number of phenomenologically important process, namely $e^+e^-\to$ dijets, $e^- p\to e^-$ jet, and threshold Drell-Yan or Drell-Yan with an inclusive jet veto (which we refer to as constrained Drell-Yan for short).
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.23\columnwidth]{figures/Subleading1_long}
\hspace{0.1cm}
\includegraphics[width=0.23\columnwidth]{figures/Subleading2_long}
\hspace{0.1cm}
\includegraphics[width=0.23\columnwidth]{figures/Subleading3_long}
\hspace{0.1cm}
\includegraphics[width=0.23\columnwidth]{figures/Subleading4_long}
\raisebox{0cm}{ \hspace{-0.2cm}
$a$)\hspace{3.4cm}
$b$)\hspace{3.4cm}
$c$)\hspace{3.4cm}
$d$)\hspace{4cm} }
\\[-25pt]
\end{center}
\vspace{-0.4cm}
\caption{
Examples of the contributions to thrust, $\tau$, in the dijet limit, at leading power in a) and b), and subleading power in c) and d). There is an extra collinear gluon in a) from splitting, and in b) there is an extra gluon from soft emission. In c) the extra energetic gluon is collinear with the quark, but occurs without a nearly onshell parent propagator. Likewise in d) the extra soft emission amplitude is subleading. }
\label{fig:subleadingamp}
\end{figure}
As an example to illustrate the expected form of subleading factorization theorems for collider observables, consider the $e^+e^-$ event shape thrust $T=1-\tau$~\cite{Farhi:1977sg} at center-of-mass energy $Q$. In the $Q\gg\Lambda_{\rm QCD}$ and $\tau\ll 1$ limit \cite{Korchemsky:1999kt,Fleming:2007qr,Schwartz:2007ib,Abbate:2010xh} where the events are dominated by back-to-back jets, one can derive a leading power factorization formula for this process, given by
\begin{align} \label{eq:fact0}
\frac{\mathrm{d}\sigma}{\mathrm{d}\tau}
&= Q\sigma_0\, H^{(0)}(Q,\mu) \int\!\! \mathrm{d} s\: J^{(0)}_\tau(s,\mu)
\ S^{(0)}_\tau \Big(Q\tau -\frac{s}{Q},\mu \Big)
\ +\ {\cal O}\Big(\tau^0 , \frac{\Lambda_{\rm QCD}}{Q\tau }\Big) \,.
\end{align}
This factorization formula involves hard contributions from the scale $Q$ in a leading power hard function $H^{(0)}$, collinear contributions from the scale $Q\tau^{1/2}$ in the leading power thrust jet function obtained by combining two standard jet functions, $J_\tau^{(0)} = [ J^{(0)}\otimes J^{(0)}]$, and ultrasoft contributions from the scales $Q\tau$ and $\Lambda_{\rm QCD}$ in the leading power thrust soft function $S_\tau^{(0)}$. Here $\sigma_0$ is the $e^+e^-\to q\bar q$ Born cross section and $\mu$ is the factorization scale. For simplicity, we take $Q\tau \sim \Lambda_{\rm QCD}$ so that we do not have to elaborate further on the factorization of the soft function into perturbative and non-perturbative parts. The factorized product $H^{(0)} \times J_\tau^{(0)} \otimes S_\tau^{(0)}$ in \eq{fact0} includes contributions at all orders in $\alpha_s$ and powers of $\ln\tau$, which have the power law scaling of ${\cal O}(\tau^{-1})$, including $\delta(\tau)$ terms. As indicated, there are perturbative and nonperturbative power corrections to this formula starting at ${\cal O}(\tau^0)$. The soft function $S_\tau^{(0)}$ contains all nonperturbative corrections of ${\cal O}\Big( \frac{\Lambda_{\rm QCD}^k}{\tau (Q\tau)^k} \Big)$, so the first missing power corrections are ${\cal O}(\frac{\Lambda_{\rm QCD}}{Q\tau})$, which is also ${\cal O}(\tau^0)$ in our simplified counting.
In general the cross section can be expanded in powers of $\tau$,
\begin{align}
\frac{\mathrm{d}\sigma}{\mathrm{d}\tau} &=\frac{\mathrm{d}\sigma^{(0)}}{\mathrm{d}\tau} +\frac{\mathrm{d}\sigma^{(1)}}{\mathrm{d}\tau} +\frac{\mathrm{d}\sigma^{(2)}}{\mathrm{d}\tau}+\frac{\mathrm{d}\sigma^{(3)}}{\mathrm{d}\tau} +{\cal O}(\tau)\,,
\end{align}
where $\mathrm{d}\sigma^{(n)}/\mathrm{d}\tau\sim \tau^{-1+n/2}$ denotes the suppression relative to the leading term in powers of $\sqrt{\tau}$. Although for most observables the odd terms vanish, this convention gives a simple correspondence with the amplitude level power expansion. The explicit expression for $\mathrm{d}\sigma^{(0)}/\mathrm{d}\tau$ is given by the first term, $H^{(0)} \times J_\tau^{(0)} \otimes S_\tau^{(0)}$, shown in \eq{fact0}. In SCET the hard scattering operators and Lagrangians governing the soft and collinear dynamics can be factorized from each other at any order in the power expansion.\footnote{This assumes that leading power Lagrangian interactions that can couple soft and collinear modes through Glauber exchange operators~\cite{Rothstein:2016bsq} that involve $1/\mathcal{P}_\perp^2$ potential can be ignored at the active parton level. It is known that this is the case for the full $e^+e^-\to $2-jet event shapes at leading power and for inclusive Drell-Yan at leading power \cite{Collins:1988ig}, and that this is not the case for spectator effects and ${\cal O}(\alpha_s^4)$ perturbative corrections in certain Drell-Yan event shapes~\cite{Gaunt:2014ska,Zeng:2015iba}. } Therefore factorization theorems can also be derived for the power suppressed contributions $\mathrm{d}\sigma^{(i)}/\mathrm{d}\tau$, for $i>0$, corresponding to the power corrections in \eq{fact0}. A prototypical example of this is the factorization theorem for the decay rate of $b\to s\gamma$ at large $E_\gamma$, which has the same form as \eq{fact0} at leading power, and where a factorization theorem for the ${\cal O}(\tau^0)$ terms has been derived~\cite{Lee:2004ja}. It involves subleading hard, jet, and soft functions, since power corrections can not change the relevant degrees of freedom. Based on this we expect that the higher order power corrections to the thrust cross section will obey schematic factorization theorems of the form
\begin{align} \label{eq:sigma}
&\hspace{-0.25cm}\frac{\mathrm{d}\sigma^{(n)}}{\mathrm{d}\tau} =
Q\sigma_0
\sum_{j} H^{(n_{Hj})}_{j} \otimes \Big[ J^{(n_{Jj})}_{j} J^{(n'_{Jj})}_{j}\Big] \otimes S_j^{(n_{Sj})}
,\end{align}
where $j$ sums over the multiple contributions that appear at each order, $n_{Hj}+n_{Jj}+n_{Jj}'+n_{Sj}=n$, and $\otimes$ denotes a set of convolutions. In this formula, the cross section at each power has been factorized into hard functions $H_j^{(n_{Hj})}$, jet functions $J_j^{(n_{Jj})}$, and soft functions $S_j^{(n_{Sj})}$ which may be leading or subleading power depending on the value of the $n_{Xj}$ indices. The hard function contains the dependence on the underlying hard partonic process, but can be chosen to be independent of the particular event shape that is being measured. The jet functions (which describe the collinear radiation along the jet directions) as well as the soft function (which describes the soft radiation in the event) depend on the particular measurement function.
Deriving a subleading power factorization theorem using SCET, like that in \eq{sigma}, consists of several steps. First, one must demonstrate the existence of a finite basis of hard scattering operators in SCET at the appropriate order in the power expansion, and determine an explicit basis of such operators. Matching calculations from QCD to SCET are required to determine the Wilson coefficients of these operators and the structure of their collinear Wilson lines. The soft and collinear dynamics are entirely described by the Lagrangian of the effective theory, and the subleading power SCET Lagrangian is also required at the same order in the power expansion. At leading power, the BPS field redefinition \cite{Bauer:2002nz} can be used to factorize the Lagrangian into non-interacting pieces describing each collinear sector, as well as the soft sector. The only exception to this is the leading power Glauber Lagrangian~\cite{Rothstein:2016bsq} which couples together soft and collinear fields, and which will violate factorization if it can not be shown to give canceling contributions or that it is irrelevant. Power suppressed Lagrangians have been analyzed in the literature~\cite{Beneke:2002ni,Chay:2002vy,Manohar:2002fd,Pirjol:2002km,Beneke:2002ph,Bauer:2003mga}, and the \mbox{${\rm SCET}_{\rm I}$}\xspace Lagrangian is currently known to ${\cal O}(\lambda^2)$ \cite{Bauer:2003mga} (excluding power suppressed Glauber exchange operators). Beyond leading power and after the BPS field redefinition, the subleading Lagrangians (including power suppressed Glauber operators) will involve factorized products of soft and collinear fields. Here factorization at the Lagrangian level again only requires showing that the leading power Glauber Lagrangian is not required. Next, one must consider factorization of the observable, and demonstrate that one can define suitable measurements that are separately made in the soft and collinear matrix elements. Finally, starting with the full QCD expression for the appropriate observable, one must go through a number of expansions and algebraic manipulations to factorize the cross section into a product of squared matrix elements, each involving only collinear or soft fields. This step leads to field theoretic definitions of the subleading jet, soft and hard functions appearing in the factorization theorem of \eq{sigma}. The degree to which these steps require lengthy and tedious calculations is determined by the complexity of the operator basis.
Traditionally, an operator basis is constructed by enumerating all possible operators consistent with symmetry constraints. These operators are formed from the SCET fields, along with Lorentz, Dirac and color structures. Beyond leading power, the determination of a minimal operator basis becomes complicated, even for processes with a limited number of collinear sectors, such as $pp\to \mu^+\mu^-$ (constrained Drell-Yan) or $e^+e^- \to$ dijets. The algebraic manipulations in SCET required to achieve factorization are similarly complicated, making subleading factorization laborious. In this paper, we show that by working with operators of definite helicity, the operator basis is easy to construct, and does not involve complicated Lorentz or Dirac structures, simplifying the algebraic manipulations required for factorization. Many symmetry properties are also made manifest in the helicity basis. We demonstrate how these provide simplifications both at the level of the hard scattering operator basis as well as in factorized matrix elements.
An outline of this paper is as follows. In \sec{scet} we provide a brief review of SCET with an emphasis on the field content, power counting, and construction of SCET operators. In \sec{helops} we describe our basis of helicity operators and discuss their symmetry properties. We focus in particular on the treatment of operators involving ultrasoft fields, and use the BPS field redefinition to define collinear and ultrasoft gauge invariant operator building blocks. Additionally, selection rules on the hard scattering operators due to angular momentum conservation are described \cite{Kolodrubetz:2016uim}. Extensions of the formalism to SCET$_\text{II}$ as well as SCET with massive collinear quarks are also discussed, as are complications associated with evanescent operators. We then demonstrate the utility of our helicity basis in \sec{eeJets} by constructing an $\mathcal{O}(\lambda^2)$ basis of hard scattering operators with two back-to-back collinear sectors, as relevant for $e^+e^-\to$ dijets, $e^- p\to e^-$+jet, and constrained Drell-Yan. Using the symmetry properties of the operators, we enumerate those which can contribute to the factorized cross section at $\mathcal{O}(\lambda^2)$, and in \Sec{sec:matching}, we perform the tree level matching to these operators. We conclude in \sec{conclusions}.
\section{Review of SCET}\label{sec:scet}
SCET is an effective field theory of QCD describing the interactions of collinear and soft particles in the presence of a hard interaction \cite{Bauer:2000ew, Bauer:2000yr, Bauer:2001ct, Bauer:2001yt, Bauer:2002nz}. Since SCET describes collinear particles (which are characterized by a large momentum along a particular light-like direction), as well as soft particles, it is natural to use light-cone coordinates. For each jet direction we define two light-like reference vectors $n_i^\mu$ and ${\bf n}_i^\mu$ such that $n_i^2 = {\bf n}_i^2 = 0$ and $n_i\!\cdot\!{\bf n}_i = 2$. One typical choice for these quantities is
\begin{equation}
n_i^\mu = (1, \vec{n}_i)
\,,\qquad
{\bf n}_i^\mu = (1, -\vec{n}_i)
\,,\end{equation}
where $\vec{n}_i$ is a unit three-vector. Given a choice for $n_i^\mu$ and ${\bf n}_i^\mu$, any four-momentum $p$ can then be written as
\begin{equation} \label{eq:lightcone_dec}
p^\mu = {\bf n}_i\!\cdot\! p\,\frac{n_i^\mu}{2} + n_i\!\cdot\! p\,\frac{{\bf n}_i^\mu}{2} + p^\mu_{n_i\perp}\
\,.\end{equation}
A particle with momentum $p$ close to the $\vec{n}_i$ direction, such that the components of $p$ scale as $(n_i\!\cdot\! p, {\bf n}_i \!\cdot\! p, p_{n_i\perp}) \sim {\bf n}_i\!\cdot\! p$ $\,(\lambda^2,1,\lambda)$, where $\lambda \ll 1$ is a small formal power counting parameter, are referred to as $n_i$ collinear. The formal scaling of $\lambda$ is determined by the form of measurements or kinematic restrictions on the QCD radiation. To ensure that $n_i$ and $n_j$ refer to distinct collinear directions, we must have
\begin{equation} \label{eq:nijsep}
n_i\!\cdot\! n_j \gg \lambda^2 \qquad\text{for}\qquad i\neq j
\,.\end{equation}
Since distinct reference vectors, $n_i$ and $n_i'$, with $n_i\cdot n_i' \sim \ord{\lambda^2}$ both describe the same collinear physics, one can label a collinear sector by any member of a set of equivalent vectors, $\{n_i\}$. This freedom is manifest as a symmetry of the effective theory known as reparametrization invariance (RPI) \cite{Manohar:2002fd,Chay:2002vy}. Specifically, the three classes of RPI transformations are
\begin{alignat}{3}\label{eq:RPI_def}
&\text{RPI-I} &\qquad & \text{RPI-II} &\qquad & \text{RPI-III} {\nonumber} \\
&n_{i \mu} \to n_{i \mu} +\Delta_\mu^\perp &\qquad & n_{i \mu} \to n_{i \mu} &\qquad & n_{i \mu} \to e^\alpha n_{i \mu} {\nonumber} \\
&\bar n_{i \mu} \to \bar n_{i \mu} &\qquad & \bar n_{i \mu} \to \bar n_{i \mu} +\epsilon_\mu^\perp &\qquad & \bar n_{i \mu} \to e^{-\alpha} \bar n_{i \mu}\,.
\end{alignat}
Here, we have $\Delta^\perp \sim \lambda$, $\epsilon^\perp \sim \lambda^0$, and $\alpha\sim \lambda^0$. The parameters $\Delta^\perp$ and $\epsilon^\perp$ are infinitesimal, and satisfy $n_i\cdot \Delta^\perp=\bar n_i\cdot \Delta^\perp=n_i \cdot \epsilon^\perp=\bar n_i \cdot \epsilon^\perp=0$. RPI will be exploited to simplify the structure of the subleading power operator basis in \sec{eeJets}.
The effective theory is constructed by expanding momenta into label and residual components
\begin{equation} \label{eq:label_dec}
p^\mu = \lp^\mu + k^\mu = {\bf n}_i \!\cdot\!\lp\, \frac{n_i^\mu}{2} + \lp_{n_i\perp}^\mu + k^\mu\,.
\,\end{equation}
Here, ${\bf n}_i \cdot\lp \sim Q$ and $\lp_{n_i\perp} \sim \lambda Q$ are the large label momentum components, where $Q$ is the scale of the hard interaction, while $k\sim \lambda^2 Q$ is a small residual momentum describing fluctuations about the label momentum. A multipole expansion is then performed to obtain fields with momenta of definite scaling, namely collinear quark and gluon fields for each collinear direction, as well as soft quark and gluon fields. Independent gauge symmetries are enforced for each set of fields.
Due to the multipole expansion, the SCET fields for $n_i$-collinear quarks and gluons, $\xi_{n_i,\lp}(x)$ and $A_{n_i,\lp}(x)$, are written in position space with respect to the residual momentum and in momentum space with respect to the large momentum components. They are labeled by their collinear direction $n_i$ and their large momentum $\lp$. The label momentum operator $\mathcal{P}_{n_i}^\mu$, gives the large label component of the momentum, $\mathcal{P}_{n_i}^\mu\, \xi_{n_i,\lp} = \lp^\mu\, \xi_{n_i,\lp}$, while derivatives give the residual momentum dependence, $i \partial^\mu \sim k \sim \lambda^2 Q$. The label momentum operator is defined such that when acting on a product of fields, $\mathcal{P}_{n_i}$ gives the sum of the label momenta of all $n_i$-collinear fields. We will often use the shorthand notation $\overline {\mathcal P}_{n_i} = {\bf n}\!\cdot\!\mathcal{P}_{n_i}$ for the large label momentum component.
Soft degrees of freedom are described in the effective theory by separate quark and gluon fields. We will assume that we are working in the SCET$_\text{I}$ theory where these soft degrees of freedom are referred to as ultrasoft so as to distinguish them from the soft modes of SCET$_\text{II}$ \cite{Bauer:2002aj}. Extensions of our formalism to treat SCET$_\text{II}$ problems will be discussed in \Sec{sec:scet_II}. In SCET$_\text{I}$, the ultrasoft modes do not carry label momenta, but have residual momentum dependence with $i \partial^\mu \sim \lambda^2Q$, and are able to exchange residual momenta between different collinear sectors. They are therefore described by fields $q_{us}(x)$ and $A_{us}(x)$ without label momenta, and without a collinear sector label.
SCET is formulated as an expansion in powers of $\lambda$, constructed so that manifest power counting is maintained at all stages of a calculation. As a consequence of the multipole expansion, all fields and derivatives acquire a definite power counting \cite{Bauer:2001ct}, shown in \Tab{tab:PC}. The SCET Lagrangian is also expanded as a power series in $\lambda$
\begin{align} \label{eq:SCETLagExpand}
\mathcal{L}_{\text{SCET}}=\mathcal{L}_\mathrm{hard}+\mathcal{L}_\mathrm{dyn}= \sum_{i\geq0} \mathcal{L}_\mathrm{hard}^{(i)}+\mathcal{L}_G^{(0)}+\sum_{i\geq0} \mathcal{L}^{(i)} \,,
\end{align}
where $(i)$ denotes objects at ${\cal O}(\lambda^i)$ in the power counting. The Lagrangians $ \mathcal{L}_\mathrm{hard}^{(i)}$ contain the hard scattering operators $O^{(i)}$, whose structure is determined by the matching process, as described in \Sec{sec:matching}. The leading power Glauber Lagrangian~\cite{Rothstein:2016bsq}, $\mathcal{L}_G^{(0)}$, describes leading power interactions between soft and collinear modes in the form of potentials. It breaks factorization unless it can be shown to cancel out or absorbed into other interactions such as Wilson line directions. The $\mathcal{L}^{(i)}$ describe the dynamics of ultrasoft and collinear modes in the effective theory, including subleading power corrections to the Glauber Lagrangian. The subleading Lagrangians (excluding subleading power corrections to the Glauber Lagrangian) are explicitly known to $\mathcal{O}(\lambda^2)$, and the relevant ones for our analysis can be found, along with their Feynman rules in \cite{iain_notes}.
Factorization theorems used in jet physics are typically derived at leading power in $\lambda$. In this case, interactions involving hard processes in QCD are matched to a basis of leading power SCET hard scattering operators $O^{(0)}$, the dynamics in the effective theory are described by the leading power Lagrangian, $\mathcal{L}^{(0)}$, and the measurement function, which defines the action of the observable, is expanded to leading power. Higher power terms in the $\lambda$ expansion, known as power corrections, arise from three sources: subleading power hard scattering operators $O^{(i)}$, subleading Lagrangian insertions, and subleading terms in the expansion of the measurement functions which act on soft and collinear radiation. The first two sources are independent of the details of the particular measurement, only requiring that it is an SCET$_\text{I}$ dijet observable, while the third depends on its precise definition. Although we will not discuss subleading measurement functions in this paper, an example, for the case of thrust, is given in \App{app:meas}.
\begin{table}
\begin{center}
\begin{tabular}{| l | c | c |c |c|c| r| }
\hline
Operator & $\mathcal{B}_{n_i\perp}^\mu$ & $\chi_{n_i}$& $\mathcal{P}_\perp^\mu$&$q_{us}$&$D_{us}^\mu$ \\
Power Counting & $\lambda$ & $\lambda$& $\lambda$& $\lambda^3$& $\lambda^2$ \\
\hline
\end{tabular}
\end{center}
\caption{
Power counting for building block operators in $\text{SCET}_\text{I}$.
}
\label{tab:PC}
\end{table}
Gauge invariant collinear operators in the effective theory are constructed out of products of gauge invariant building blocks. These building blocks are formed from gauge invariant combinations of fields and Wilson lines~\cite{Bauer:2000yr,Bauer:2001ct}. The collinearly gauge-invariant quark and gluon fields are defined as
\begin{align} \label{eq:chiB}
\chi_{{n_i},\omega}(x) &= \Bigl[\delta(\omega - \overline {\mathcal P}_{n_i})\, W_{n_i}^\dagger(x)\, \xi_{n_i}(x) \Bigr]
\,,\\
\mathcal{B}_{{n_i}\perp,\omega}^\mu(x)
&= \frac{1}{g}\Bigl[\delta(\omega + \overline {\mathcal P}_{n_i})\, W_{n_i}^\dagger(x)\,i D_{{n_i}\perp}^\mu W_{n_i}(x)\Bigr]
\,. {\nonumber}
\end{align}
Here we have chosen a convention such that for $\chi_{{n_i},\omega}$, we have $\omega > 0$ for an incoming quark and $\omega < 0$ for an outgoing antiquark. For $\mathcal{B}_{{n_i},\omega\perp}$, $\omega > 0$ ($\omega < 0$) corresponds to outgoing (incoming) gluons. The covariant derivative which appears in \eq{chiB} is defined as,
\begin{equation}
i D_{{n_i}\perp}^\mu = \mathcal{P}^\mu_{{n_i}\perp} + g A^\mu_{{n_i}\perp}\,,
\end{equation}
and $W_{n_i}$ are Wilson lines of ${n_i}$-collinear gluons in label momentum space defined as
\begin{equation} \label{eq:Wn}
W_{n_i}(x) = \biggl[~\sum_\text{perms} \exp\Bigl(-\frac{g}{\overline {\mathcal P}_{n_i}}\,{\bf n}\!\cdot\! A_{n_i}(x)\Bigr)~\biggr]\,,
\end{equation}
In general the structure of Wilson lines must be derived by a matching calculation from QCD. These Wilson lines sum up arbitrary emissions of ${n_i}$-collinear gluons off of particles from other sectors, which due to the power expansion always appear in the ${\bar{n}_i}$ direction. The gluon emissions summed in the Wilson lines are $\ord{\lambda^0}$ in the power counting. In \eqs{chiB}{Wn} the label momentum operators act only on the fields inside the square brackets. The Wilson line $W_{n_i}(x)$ is localized with respect to the residual position $x$, and we can therefore treat
$\chi_{{n_i},\omega}(x)$ and $\mathcal{B}_{{n_i},\omega}^\mu(x)$ as local quark and gluon fields from the perspective of ultrasoft derivatives $\partial^\mu$ that act on $x$.
The complete set of collinear and ultrasoft building blocks for constructing hard scattering operators or subleading Lagrangians at any order in the power counting is given in \Tab{tab:PC}. All other field and derivative combinations can be reduced to this set by the use of equations of motion and operator relations~\cite{Marcantonini:2008qn}. Since these building blocks carry vector or spinor Lorentz indices they must be contracted to form scalar operators, which also involves the use of objects like $\{n_i^\mu, {\bf n}_i^\mu, \gamma^\mu, g^{\mu\nu}, \epsilon^{\mu\nu\sigma\tau}\}$. A key advantage of the helicity operator approach discussed below is that this is no longer the case; all the building blocks will be scalars.
As shown in \Tab{tab:PC}, both the collinear quark and collinear gluon building block fields scale as ${\cal O}(\lambda)$. For the majority of jet processes there is a single collinear field operator for each collinear sector at leading power. (Although for fully exclusive processes that directly produce hadrons there will be multiple building blocks from the same sector in the leading power operators since they form color singlets in each sector.) Also, since $\mathcal{P}_\perp\sim \lambda$, this operator will not typically be present at leading power (exceptions could occur, for example, in processes picking out P-wave quantum numbers). At subleading power, operators for all processes can involve multiple collinear fields in the same collinear sector, as well as $\mathcal{P}_\perp$ operator insertions. The power counting for an operator is obtained by simply adding up the powers for the building blocks it contains. To ensure consistency under renormalization group evolution the operator basis in SCET must be complete, namely all operators consistent with the symmetries of the problem must be included.
Dependence on the ultrasoft degrees of freedom enters the operators through the ultrasoft quark field $q_{us}$, and the ultrasoft covariant derivative $D_{us}$, defined as
\begin{equation}
i D_{us}^\mu = i \partial^\mu + g A_{us}^\mu\,,
\end{equation}
from which we can construct other operators including the ultrasoft gluon field strength. All operators in the theory must be invariant under ultrasoft gauge transformations. Collinear fields transform under ultrasoft gauge transformations as background fields of the appropriate representation. The power counting for these operators is shown in \Tab{tab:PC}. Since they are suppressed relative to collinear fields, ultrasoft fields typically do not enter factorization theorems in jet physics at leading power. An example where ultrasoft fields enter at leading power is $B \to X_s \gamma$ in the photon endpoint region, which is described at leading power by a single collinear sector, and an ultrasoft quark field for the b quark~\cite{Bauer:2000ew}.
\section{Helicity Operators}\label{sec:helops}
The use of on-shell helicity amplitudes has been fruitful for the study of scattering amplitudes in gauge theories and gravity (see e.g. \cite{Dixon:1996wi,Elvang:2013cua,Dixon:2013uaa,Henn:2014yza} for pedagogical reviews). By focusing on amplitudes for external states with definite helicity and color configurations many simplifications arise. The helicity approach to SCET operators of \Ref{Moult:2015aoa} takes advantage of the fact that collinear SCET fields are themselves gauge invariant, and are each associated with a fixed external label direction with respect to which helicities can naturally be defined. Instead of considering operators formed from Lorentz and Dirac structures (each of which contributes to multiple states with different helicity combinations) helicity operators can be associated with external states of definite helicity. This approach greatly simplifies the construction of a minimal operator bases for processes with many active partons, and facilitates the matching to fixed order calculations which are often performed using spinor helicity techniques.
We now briefly summarize our spinor helicity conventions. Further identities, as well as our phase conventions, can be found in \app{helicity}. To simplify our discussion we take all momenta and polarization vectors as outgoing, and label all fields and operators by their outgoing helicity and momenta. We use the standard spinor helicity notation
\begin{align} \label{eq:braket_def}
|p\rangle\equiv \ket{p+} &= \frac{1 + \gamma_5}{2}\, u(p)
\,,
& |p] & \equiv \ket{p-} = \frac{1 - \gamma_5}{2}\, u(p)
\,, \\
\bra{p} \equiv \bra{p-} &= \mathrm{sgn}(p^0)\, \bar{u}(p)\,\frac{1 + \gamma_5}{2}
\,,
& [p| & \equiv \bra{p+} = \mathrm{sgn}(p^0)\, \bar{u}(p)\,\frac{1 - \gamma_5}{2}
\,, {\nonumber}
\end{align}
with $p$ lightlike. The polarization vector of an outgoing gluon with momentum $p$ can be written
\begin{equation}
\varepsilon_+^\mu(p,k) = \frac{\mae{p+}{\gamma^\mu}{k+}}{\sqrt{2} \langle kp \rangle}
\,,\qquad
\varepsilon_-^\mu(p,k) = - \frac{\mae{p-}{\gamma^\mu}{k-}}{\sqrt{2} [kp]}
\,,\end{equation}
where $k\neq p$ is an arbitrary lightlike reference vector.
The polarization vectors and spinors satisfy the standard identities.
\begin{align}
&p\cdot \epsilon_{\pm}(p,k)=k \cdot\epsilon_{\pm}(p,k) =0\,, \qquad \epsilon_{\pm}(p,k)\cdot \epsilon_{\pm}(p,k)=0\,, \qquad \epsilon_{\pm}(p,k)\cdot \epsilon_{\mp}(p,k)=-1\,, {\nonumber}\\
&\Sl{\epsilon}_+(p,k)|k\rangle=\Sl{\epsilon}_-(p,k)|k]=0\,, \qquad
[k|\Sl{\epsilon}_-(p,k)=\langle k|\Sl{\epsilon}_+(p,k)=0\,.
\end{align}
Additional identities can be found in \App{app:helicity}.
In this section we discuss the extension of the helicity operator approach of \cite{Moult:2015aoa} to subleading powers. We review the full set of subleading power building block operators introduced in \cite{Kolodrubetz:2016uim}, and provide more details about them. In \sec{coll} we describe operators involving only collinear fields, and in \sec{pperp} we describe operators involving insertions of the $\mathcal{P}_\perp$ operator. We also give all the Feynman rules for these operators. The organization of color bases is discussed in \sec{color}. The inclusion of ultrasoft fields in the hard scattering operators is more involved, since the usual SCET building blocks are gauge covariant rather than gauge invariant. In \sec{BPS} we discuss the BPS field redefinition, and how it can be used to define ultrasoft gauge invariant helicity operators to be used as basis elements. In \sec{constr_subops}, we give the complete list of scalar building blocks needed to construct an operator basis at arbitrary power in $\lambda$. Next, in \sec{extensions} we examine the extension of this formalism to \mbox{${\rm SCET}_{\rm II}$}\xspace, SCET with massive quarks and evanescent operators. We briefly discuss how to carry out matching calculations at subleading power in \sec{matching}, and the C and P properties of the operators in \sec{PandC}. In \Sec{sec:ang_cons} we discuss interesting constraints from angular momentum conservation which first appear at subleading power when there are multiple fields in the same collinear sector.
\subsection{Collinear Gauge Invariant Helicity Building Blocks}\label{sec:coll}
We define a collinear gluon field of definite helicity as
\begin{equation} \label{eq:cBpm_def}
\mathcal{B}^a_{i\pm} = -\varepsilon_{\mp\mu}(n_i, {\bf n}_i)\,\mathcal{B}^{a\mu}_{n_i\perp,\omega_i}
\,,\end{equation}
where $a$ is an adjoint color index. This is sufficient for the treatment of collinear gluons even at subleading power. With this definition, for an outgoing gluon with polarization $\pm$, momentum $p$, $p^0>0$ (or an incoming gluon with polarization $\mp$, momentum $-p$, $p_0<0$), and color $a$, the nonzero tree-level Feynman rules are\footnote{The precise definition of this delta function and measure are
\begin{align} \label{eq:labelsums}
\ldel(\lp_i - p) &\equiv \delta_{\{n_i\},p}\,\delta(\omega_i - {\bf n}_i\cdot p)\,,{\nonumber}\\
\int\!\mathrm{d}\lp &\equiv \sum_{\{n_i\}} \int\!\mathrm{d}\omega_i {\nonumber}
\,,\end{align}
where
\begin{equation}
\delta_{\{n_i\},p} =
\begin{cases}
1 &\quad n_i\cdot p = \ord{\lambda^2}
\,,\\
0 &\quad \text{otherwise}{\nonumber}
\,.\end{cases}
\end{equation}
The Kronecker delta is nonzero if the collinear momentum $p$ is in the $\{ n_i\}$
equivalence class, i.e. $p$ should be considered as collinear with
the $i$th jet. The sum in the second line runs over the different
equivalence classes. }
\begin{align} \label{eq:gluonbaseFR}
\langle g_\pm^a(p) | \mathcal{B}_{i \pm}^b |0 \rangle & = \delta^{ab} \ldel (\lp_i -p)\,, \\
\langle 0 | \mathcal{B}_{i \pm}^b |g_\mp^a(-p) \rangle & = \delta^{ab} \ldel (\lp_i -p), {\nonumber}
\end{align}
where we have followed \Ref{Moult:2015aoa} in using $k = {\bf n}$ as our reference vector. Despite the fact that $\mathcal{B}_{i \pm}^a=\mathcal{B}_{i \pm}^a(x)$, our external gluon state has zero residual momentum, so we do not get an additional phase.
We also define quark fields with definite helicity, given by
\begin{align} \label{eq:definitehelicityquarkdef}
\chi_{i \pm}^\alpha &= \frac{1\,\pm\, \gamma_5}{2} \chi_{n_i, - \omega_i}^\alpha\,,\qquad \bar{\chi}_{i \pm}^{\bar \alpha} = \bar{\chi}_{n_i, - \omega_i}^{\bar \alpha} \frac{1\,\mp\, \gamma_5}{2}\,.
\end{align}
For external quarks of definite helicity, with $n_i$-collinear momentum $p$, the spinor appearing in SCET Feynman rules is,
\begin{align}
\frac{1\,\pm\, \gamma_5}{2} \frac{\slashed{n}_i \slashed{\bar{n}}_i}{4} u(p) = \frac{\slashed{n}_i \slashed{\bar{n}}_i}{4} | p \pm \rangle \equiv |p \pm \rangle_{n_i},
\end{align}
where $|p \pm \rangle_{n_i}$ is a convenient short-hand notation for the projected spinor, and is proportional to $|n_i \pm \rangle$ (see \ref{eq:ketn}). Using this, we get the nonzero tree-level Feynman rules for incoming ($p^0<0$) and outgoing ($p^0>0$) quarks with definite helicity $\pm$ and color $\alpha$ (or ${\bar \alpha}$),
\begin{align} \label{eq:quarkbaseFR}
\Mae{0}{\chi^\beta_{i\pm}}{q_\pm^{\bar \alpha}(-p)}
&= \delta^{\beta{\bar \alpha}}\,\ldel(\lp_i - p)\, \ket{(-p_i)\pm}_{n_i} \,,\\
\Mae{q_\pm^\alpha(p)}{\bar\chi^{\bar \beta}_{i\pm}}{0}
&= \delta^{\alpha{\bar \beta}}\,\ldel(\lp_i - p)\, {}_{n_i\!}\bra{p_i\pm}
\,, {\nonumber} \\
\Mae{0}{\bar\chi^{\bar \beta}_{i\pm}}{\bar{q}_\mp^\alpha(-p)}
&= \delta^{\alpha{\bar \beta}}\,\ldel(\lp_i - p)\,{}_{n_i\!}\bra{(-p_i)\pm}
\,, {\nonumber} \\
\Mae{\bar{q}_\mp^{\bar \alpha}(p)}{\chi^\beta_{i\pm}}{0}
&= \delta^{\beta{\bar \alpha}}\,\ldel(\lp_i - p)\, \ket{p_i\pm}_{n_i} \,.{\nonumber}
\end{align}
We wish to take advantage of the fact that fermions come in pairs, to simplify the treatment of Dirac structures when constructing an operator basis. \Ref{Moult:2015aoa} therefore defined the currents
\begin{align} \label{eq:jpm_def}
J_{ij\pm}^{{\bar \alpha}\beta}
& = \mp\, \sqrt{\frac{2}{\omega_i\, \omega_j}}\, \frac{ \varepsilon_\mp^\mu(n_i, n_j) }{\langle n_j\mp | n_i\pm\rangle} \, \bar{\chi}^{\bar \alpha}_{i\pm}\, \gamma_\mu \chi^\beta_{j\pm}
\,, \\
J_{ij0}^{{\bar \alpha}\beta}
& =\frac{2}{\sqrt{\vphantom{2} \omega_i \,\omega_j}\, [n_i n_j] } \bar \chi^{\bar \alpha}_{i+}\chi^\beta_{j-}
\,, \qquad
(J^\dagger)_{ij0}^{{\bar \alpha}\beta}=\frac{2}{\sqrt{ \vphantom{2} \omega_i \, \omega_j} \langle n_i n_j \rangle } \bar \chi^{\bar \alpha}_{i-}\chi^\beta_{j+}
. {\nonumber}
\end{align}
These currents have been defined such that they are invariant under an RPI-III transformation, which can be easily from the fact that $\omega_i$ scales as ${\bf n}_i$ and the $| n_i \rangle$ scale as $\sqrt{n_i}$. In \secs{eeJets}{matching}, we will focus on the case of two back-to-back jet directions, so it is worth writing down the currents in that case. The tree-level Feynman rules for the currents with general sectors are given by
\begin{align}\label{eq:tree_feyn}
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | J_{12 +}^{{\bar \beta}_1 \beta_2} | 0 \rangle
&= e^{i\Phi(J_{12+})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, \\
\langle q_{-}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{12 -}^{{\bar \beta}_1 \beta_2} | 0 \rangle
&= e^{i\Phi(J_{12-})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber} \\
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{12\, 0}^{{\bar \beta}_1 \beta_2} | 0 \rangle
&= e^{i\Phi(J_{12\,0})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber} \\
\langle q_{-}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | (J^\dagger)_{12 \,0}^{{\bar \beta}_1 \beta_2} | 0 \rangle
&= e^{i\Phi(J^\dagger_{12\,0})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber}
\end{align}
where the phases appearing here are given by
\begin{align} \label{eq:phases1}
e^{i\Phi(J_{12\pm})} &= \langle n_1 \mp | \bar{n}_1 \pm \rangle \langle n_2 \pm | \bar{n}_2 \mp \rangle
\frac{\langle \bar{n}_1 \pm | p_1 \mp \rangle \langle \bar{n}_2 \mp | p_2 \pm \rangle}
{8\,\sqrt{\vphantom{2}\omega_1 \, \omega_2}}
\,,\\
e^{i\Phi(J_{12\, 0})} &= \langle \bar{n}_1 n_1 \rangle \langle n_2 \bar{n}_2 \rangle \frac{[p_1 \bar{n}_1] [\bar{n}_2 p_2] }{8\,\sqrt{\vphantom{2}\omega_1\,\omega_2}}
\,,{\nonumber} \\
e^{i\Phi(J^\dagger_{12\, 0})} &= [ \bar{n}_1 n_1 ] [ n_2 \bar{n}_2 ] \frac{\langle p_1 \bar{n}_1 \rangle \langle \bar{n}_2 p_2 \rangle }{8\,\sqrt{\vphantom{2}\omega_1\, \omega_2}}
\,.{\nonumber}
\end{align}
Note that the spinor products $\langle \bar{n}_1 n_1 \rangle$, $\langle n_2 \bar{n}_2 \rangle $, etc., depend on the choice of quantization axis for the spinors, and hence are not all trivial even though $n_1\cdot {\bf n}_1=n_2\cdot{\bf n}_2=2$.
If we consider two back-to-back collinear directions given by $n$ and ${\bf n}$, our currents have definite helicity, given by
\begin{align} \label{eq:jpm_back_to_bacjdef}
& \text{h=$\pm$ 1:}
& J_{n {\bf n} \pm}^{{\bar \alpha}\beta}
& = \mp\, \sqrt{\frac{2}{\omega_n\, \omega_{\bf n}}}\, \frac{ \varepsilon_\mp^\mu(n, {\bf n}) }{\langle {\bf n} \mp | n \pm\rangle} \, \bar{\chi}^{\bar \alpha}_{n\pm}\, \gamma_\mu \chi^\beta_{{\bf n} \pm}
\,, \\
& \text{h=0:}
& J_{n {\bf n} 0}^{{\bar \alpha}\beta}
& =\frac{2}{\sqrt{\vphantom{2} \omega_n \,\omega_{\bf n}}\, [n {\bf n}] } \bar \chi^{\bar \alpha}_{n+}\chi^\beta_{{\bf n}-}
\,, \qquad
(J^\dagger)_{n {\bf n} 0}^{{\bar \alpha}\beta}=\frac{2}{\sqrt{ \vphantom{2} \omega_n \, \omega_{\bf n}} \langle n {\bf n} \rangle } \bar \chi^{\bar \alpha}_{n-}\chi^\beta_{{\bf n}+}
. {\nonumber}
\end{align}
The current $J_{n {\bf n} 0}^{{\bar \alpha} \beta}$ transforms as a scalar under rotations about the $n$ axis, i.e. has helicity zero.\footnote{In Ref.~\cite{Moult:2015aoa} the $J_{ij0}^{{\bar \alpha}\beta}$ current was denoted as $J_{ijS}^{{\bar \alpha}\alpha}$. We choose to use the $0$ subscript to emphasize the helicity in the back-to-back case and conform with our notation for subleading currents below.} Similarly, the currents $J_{n {\bf n} \pm}^{{\bar \alpha} \beta}$ have helicity $h=\pm1$.
Our notation above for the $\mathcal{B}^a_{i\pm}$ and $\chi_{i \pm}^\alpha$ fields and the currents $J_{ij\pm}^{\bar\alpha\beta}$, $J_{ij\,0}^{\bar\alpha\beta}$, and $(J^\dagger)_{ij\, 0}^{\bar\alpha\beta}$ follows Ref.~\cite{Moult:2015aoa}, and these objects suffice for the construction of leading power operators. The $\Phi$ phases in \eq{phases1} were set to zero in Ref.~\cite{Moult:2015aoa}, since with only one particle in each collinear sector we are free to choose the $\tilde p_i$ to have zero label $\perp$-momenta, and with this choice all the phases vanish. However, at subleading power, multiple collinear fields can be present in the same collinear sector and the phases can not a priori be set to zero. Note that the phases for each current are given by a product of phases, one from each collinear sector.
We now look at how to treat the sectors at subleading power that contain multiple collinear fields, as was discussed in \cite{Kolodrubetz:2016uim}. Collinear gluons appear in gauge invariant building blocks of definite helicity, and therefore operators with multiple collinear gluons in the same sector can simply be obtained by multiplying copies of the gluon building blocks, such as $\mathcal{B}^a_{i +}\mathcal{B}^b_{i+}$. However, for quarks we must introduce new helicity currents. For $i=j$ the products of quark building blocks in \eq{jpm_def} all vanish ($\bar\chi_{i\pm}^{\bar\alpha} \gamma_\mu \chi_{i\pm}^\beta=0$, $\bar \chi_{i+}^{\bar\alpha} \chi_{i-}^\beta=0$, and $\bar\chi_{i-}^{\bar\alpha} \chi_{i+}^\beta=0$) and hence are not suitable for handling quarks in the same collinear sector. This follows from the SCET projection relations
\begin{align} \label{eq:proj}
\frac{\Sl n_i \Sl {\bar n}_i}{4} \chi_{n_i}=\chi_{n_i}, \qquad \Sl n_i \chi_{n_i}=0
\,,
\end{align}
which enforce that a quark anti-quark pair of the same chirality, in the same sector, must have zero helicity, while a quark anti-quark pair of opposite chirality must have helicity $\pm 1$.
Indeed, the scalar current $\bar \chi_{n_i} \chi_{n_i}=0$, vanishes, as do the plus and minus helicity components of the vector current $\bar \chi_{n_i} \gamma_\perp^{\pm} \chi_{n_i}=0$.
We therefore define the helicity operators involving two collinear quarks in the same sector as \cite{Kolodrubetz:2016uim}
\begin{align}\label{eq:coll_subl}
& \text{h=0:}
& J_{i0}^{{\bar \alpha} \beta}
&= \frac{1}{2 \sqrt{\vphantom{2} \omega_{\bar \chi} \, \omega_\chi}}
\: \bar \chi^{\bar \alpha}_{i+}\, \Sl{\bar n}_i\, \chi^\beta_{i+}
\,,\qquad
J_{i\bar 0}^{{\bar \alpha} \beta}
= \frac{1}{2 \sqrt{\vphantom{2} \omega_{\bar \chi} \, \omega_\chi}}
\: \bar \chi^{\bar \alpha}_{i-}\, \Sl {\bar n}_i\, \chi^\beta_{i-}
\,, \\[5pt]
& \text{h=$\pm 1$:}
& J_{i\pm}^{{\bar \alpha} \beta}
&= \mp \sqrt{\frac{2}{ \omega_{\bar \chi} \, \omega_\chi}} \frac{\epsilon_{\mp}^{\mu}(n_i,\bar n_i)}{ \big(\langle n_i \mp | \bar{n}_i \pm \rangle \big)^2}\:
\bar \chi_{i\pm}^{\bar \alpha}\, \gamma_\mu \Sl{\bar n}_i\, \chi_{i\mp}^\beta
\,. {\nonumber}
\end{align}
Note that these currents are only labeled by a single collinear sector $i$. Once again, we can easily see the RPI-III invariance, as the scaling power of $n_i$ is the same as the scaling power of ${\bf n}_i$ in each of these currents. The $J_{i0}^{{\bar \alpha} \beta}$ and $J_{i\bar 0}^{{\bar \alpha} \beta}$ transform as a scalar under rotations about the $n_i$ axis, i.e. have helicity zero. Similarly, the operators $J_{i\pm}^{{\bar \alpha} \beta}$ have helicity $h=\pm1$. These currents use only the reference vector associated with the particular jet in question for their construction. The Feynman rules for these currents with external quark states are
\begin{align}\label{eq:same_sector_feynrules}
& \langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | J_{i0}^{{\bar \beta}_1 \beta_2} | 0 \rangle = e^{i\Phi(J_{i0})}\: \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)\,, \\
& \langle q_{-}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{i\bar 0}^{{\bar \beta}_1 \beta_2} | 0 \rangle = e^{i\Phi(J_{i\bar 0})}\: \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)\,, {\nonumber} \\
& \langle q_{+}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{i +}^{{\bar \beta}_1 \beta_2} | 0 \rangle = e^{i\Phi(J_{i+})}\: \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)\,,{\nonumber} \\
& \langle q_{-}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | J_{i -}^{{\bar \beta}_1 \beta_2} | 0 \rangle = e^{i\Phi(J_{i-})}\: \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)\,, {\nonumber}
\end{align}
where the phases here are given by
\begin{align}
e^{i\Phi(J_{i0})} &= \frac{1}{2} \frac{ [p_1 \bar{n}_i ] \langle \bar{n}_i p_2 \rangle}
{\sqrt{\vphantom{2} \omega_1 \, \omega_2} }\,, \qquad
%
e^{i\Phi(J_{i\bar0})} = \frac{1}{2} \frac{\langle p_1 \bar{n}_i \rangle [ \bar{n}_i p_2 ]}
{\sqrt{\vphantom{2} \omega_1 \, \omega_2} }\,,\\
%
e^{i\Phi(J_{i\pm})} &= \frac{1}{2} \frac{\langle p_1 \pm | \bar{n}_i \mp \rangle \langle \bar{n}_i \pm | p_2 \mp \rangle}{\sqrt{\vphantom{2} \omega_1 \, \omega_2}}
\,.{\nonumber}
\end{align}
Together, the currents in \Eq{eq:jpm_def}, as well as \Eq{eq:coll_subl} are sufficient to describe collinear sectors with multiple quark fields. The complete set of quark currents will also include those with ultrasoft quark building blocks, which we will consider below in \Sec{sec:BPS}.
\subsection{$\mathcal{P}_\perp$ Operators}\label{sec:pperp}
Along with multiple collinear fields in the same sector, subleading power operators can involve explicit insertions of the $\mathcal{P}_{i\perp}^\mu$ operator, where $i$ denotes a particular collinear sector. $\mathcal{P}_{i\perp}^\mu$ is included as part of the operator to ensure that the Wilson coefficient includes only the dependence on the hard kinematics and has a uniform power counting. Since the $\mathcal{P}_{i\perp}^\mu$ operator acts on the perpendicular subspace defined by the vectors $n_i, \bar n_i$, which is spanned by the polarization vectors $\epsilon(n_i, \bar n_i)$, it naturally decomposes as
\begin{align} \label{eq:Pperppm}
\mathcal{P}_{i\perp}^{+}(n_i,\bar n_i)=-\epsilon^-(n_i,\bar n_i) \cdot \mathcal{P}_{i\perp}\,, \qquad \mathcal{P}_{i\perp}^{-}(n_i,\bar n_i)=-\epsilon^+(n_i,\bar n_i) \cdot \mathcal{P}_{i\perp}\,.
\end{align}
It is important to emphasize that the subscript $\pm$ refers to the helicity about the $n_i$ axis, and not the lightcone components of the momenta. This decomposition is performed for the $\mathcal{P}_{i\perp}$ operator in each sector. Note that it suffices to allow the operator $\mathcal{P}_{i\perp}$ to act only on fields with collinear label $i$.\footnote{The $i$-sector operator $\mathcal{P}_{i\perp}^\pm$ does not in general have a well defined power counting when acting on the field $\mathcal{B}^\lambda_{j\perp}$ for $j \ne i$. Instead, when we carry out the multipole expansion we decompose any derivative acting on $\mathcal{B}^\lambda_j$ using vectors for the $j$-sector, as $i\partial_{\rm tot}^\mu = (n_j^\mu/2) {\bf n}_j\cdot\mathcal{P}_j + \mathcal{P}_{j\perp}^\mu + i\partial_{\rm us}^\mu$. Therefore only the ${\cal O}(\lambda)$ $\perp$-operator $\mathcal{P}_{j\perp}^\mu$ acts on $\mathcal{B}^\lambda_{j\perp}$.} Therefore, when acting on a field, we will drop the collinear sector label on $\mathcal{P}_{i\perp}$, as it is determined by the label of the field. For example, we will simply write
\begin{align}
\mathcal{P}_{i \perp} \cdot \mathcal{B}_{i \perp} = \mathcal{P}_{\perp} \cdot \mathcal{B}_{i \perp} \,.
\end{align}
To see how this decomposition applies to operators written in more familiar notation, we consider the example operator $ \mathcal{P}_\perp \cdot \mathcal{B}_{i\perp}$. Using the completeness relation
\begin{align}\label{eq:completeness}
\sum\limits_{\lambda=\pm} \epsilon^\lambda_\mu(n_i,\bar n_i) \left (\epsilon^\lambda_\nu(n_i,\bar n_i) \right )^*=-g_{\mu \nu}+\frac{n_{i\mu} \bar n_{i\nu}+n_{i\nu} \bar n_{i\mu}}{n_i\cdot \bar n_i} = - g^\perp_{\mu \nu} ( n_i, {\bf n}_i)\,,
\end{align}
the decomposition into our basis is given by\footnote{The sign convention in \Eq{eq:Pperppm} is made so that dot products, as in \Eq{eq:pdotbexample}, agree with using a $(+,-,-,-)$ metric for the contraction.}
\begin{align}\label{eq:pdotbexample}
\mathcal{P}_\perp \cdot \mathcal{B}_{i\perp}=-\mathcal{P}_{\perp}^{+} \mathcal{B}_{i-}-\mathcal{P}_{\perp}^{-} \mathcal{B}_{i+}\,.
\end{align}
When acting within an operator containing multiple fields, square brackets are used to denote which fields are acted upon by the $\mathcal{P}_{\perp}^{\pm}$ operator. For example
\begin{align}
\mathcal{B}_{i+} \left [ \mathcal{P}_{\perp}^{\pm} \mathcal{B}_{i-} \right] \mathcal{B}_{i+}\,,
\end{align}
indicates that the $\mathcal{P}_{\perp}^{+}$ or $\mathcal{P}_{\perp}^{-}$ operator acts only on the middle field.
In general, we can decompose the action of the $\mathcal{P}_{\perp}^{\pm}$ operators into a superposition of terms where $\mathcal{P}_{\perp}^{\pm}$ will act only on a single field within a quark current. To define a general notation for these currents, we will follow \cite{Kolodrubetz:2016uim} and use curly braces and write $\mathcal{P}_{\perp}^{\pm}$ to the left of the current if it acts on only the first field in the current and write $(\mathcal{P}_{\perp}^{\pm})^\dagger$ to the right of the current if it acts on only the second field. As an example, we can look at the helicity currents with two quarks in the same sector with a single $\mathcal{P}_{\perp}^{\pm}$ insertion acting either on the first or second field, we write
\begin{align}\label{eq:p_perp_notation}
\big\{ \mathcal{P}_{\perp}^\lambda J_{i 0 }^{{\bar \alpha} \beta} \big\}
& = \frac{1}{2 \sqrt{\vphantom{2}\omega_{\bar \chi} \, \omega_\chi }} \:
\Big[ \mathcal{P}_{\perp}^{\lambda} \bar \chi^{\bar \alpha}_{i +}\Big] \Sl {\bar n}_i \chi^\beta_{i+}
\,, \\
\big\{ J_{i0 }^{{\bar \alpha} \beta} (\mathcal{P}_{\perp}^{\lambda})^\dagger \big\}
&= \frac{1}{2\sqrt{\vphantom{2}\omega_{\bar \chi} \, \omega_\chi}} \:
\bar \chi^{\bar \alpha}_{i+} \Sl {\bar n}_i \Big[ \chi^\beta_{i+} (\mathcal{P}_{\perp}^{\lambda})^\dagger \Big]
\,. {\nonumber}
\end{align}
Following the notation defined above, the use of the use of curly brackets $\{ \mathcal{P}_\perp^\lambda \cdots \}$ and $\{ \cdots (\mathcal{P}_\perp^\lambda)^\dagger \}$ indicate that the $\mathcal{P}_\perp$ operators act on only one of quark fields in the current, in the manner shown.
Also note for \eq{p_perp_notation} that the choice of $\lambda=+$ or $-$ for $\mathcal{P}_\perp^\lambda$ is independent of the $\pm$ choice for the quark building block fields.
The same notation will be used for $\mathcal{P}_{\perp}^{\pm}$ insertions into other currents. If we wish to instead indicate a $\mathcal{P}_\perp$ operator that acts on both building blocks in a current then we use the standard square bracket notation, for example, $\big[\mathcal{P}_\perp^\lambda J_{i \pm }^{{\bar \alpha} \beta}\big]$.
The Feynman rules for collinear operators involving insertions of the $\mathcal{P}_\perp$ operator follow straightforwardly from the corresponding Feynman rules without the $\mathcal{P}_\perp$ insertion, as given in Eqs. (\ref{eq:cBpm_def}), (\ref{eq:tree_feyn}), and (\ref{eq:same_sector_feynrules}). For example, using the $h=+1$ current of \Eq{eq:tree_feyn} as an example, we have
\begin{align}\label{eq:tree_feyn_perp}
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | \big\{\mathcal{P}_\perp^+ J_{12 +}^{{\bar \beta}_1 \beta_2} \big\} | 0 \rangle
&=-\epsilon^-(n_i,\bar n_i)\cdot \tilde p_{1\perp}\, e^{i\Phi(J_{12+})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber} \\
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | \big\{ \mathcal{P}_\perp^- J_{12 +}^{{\bar \beta}_1 \beta_2} \big\}| 0 \rangle
&=-\epsilon^+(n_i,\bar n_i)\cdot \tilde p_{1\perp}\, e^{i\Phi(J_{12+})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber} \\
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | \big\{ J_{12 +}^{{\bar \beta}_1 \beta_2} (\mathcal{P}_\perp^+)^\dagger \big\} | 0 \rangle
&=-\epsilon^-(n_i,\bar n_i)\cdot \tilde p_{2\perp}\, e^{i\Phi(J_{12+})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,, {\nonumber} \\
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | \big\{ J_{12 +}^{{\bar \beta}_1 \beta_2} (\mathcal{P}_\perp^-)^\dagger \big\} | 0 \rangle
&=-\epsilon^+(n_i,\bar n_i)\cdot \tilde p_{2\perp}\, e^{i\Phi(J_{12+})}\ \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}
\tilde{\delta} (\lp_1 - p_1) \tilde{\delta} (\lp_2 - p_2)
\,,
\end{align}
where the phase $\Phi(J_{12+})$ was defined in \Eq{eq:phases1}.
\subsection{Color Bases}\label{sec:color}
To this point we have focused on the helicity structure of the operators, with color indices left free. Consider an operator formed from a product of the currents of collinear helicity fields, $O^{a_1\dotsb \alpha_n}$. An important feature of each of the collinear helicity fields is that they are collinear gauge invariant. Furthermore, all the operators, including $\mathcal{P}_\perp$, behave as local operators with respect to ultrasoft gauge transformations, transforming like background fields. This implies that for the collinear operators, the constraints of gauge invariance are equivalent to that of a global color. It is therefore straightforward to write down a color basis for these operators. Following and generalizing the notation of \Ref{Moult:2015aoa}, we write
\begin{equation} \label{eq:Opm_color}
\vec O^\dagger_\lotsdots = O_\lotsdots^{a_1\dotsb \alpha_n}\, \bar{T}^{\, a_1\dotsb \alpha_n}
\,.
\end{equation}
Here $\bar{T}^{\, a_1\dotsb\alpha_n}$ is a row vector of color structures that spans the color conserving subspace. The $a_i$ are adjoint indices and the $\alpha_i$ are fundamental indices. The color structures do not necessarily have to be independent, but must be complete. Color structures which do not appear in the matching at a particular order will be generated by renormalization group evolution. Subtleties associated with the use of non-orthogonal color bases were discussed in detail in \Ref{Moult:2015aoa}, but they will not play a role here because we do not explicitly carry out the full factorization. Note that a decomposition as in \Eq{eq:Opm_color} is not possible in a gauge invariant manner in the full theory due to the covariant derivative, $D^\mu=\partial^\mu +ig A^\mu$, which does not transform uniformly under color. In other words, the full theory gauge invariance relates different possible color structures.
The goal of the subscripts on the $O$ in \eq{Opm_color} is to enumerate the helicities of the gluon, quark, and derivative building blocks in the operator. We have introduced it in a manner that is general enough to account for the presence of the ultrasoft building blocks that will be discussed in later sections. Outside of all parentheses the ultrasoft gluon helicities are listed first, followed by a colon and the enumeration of the collinear gluon helicities (note that in the absence of any ultrasoft gluons we omit the colon entirely). The helicities of the various types of quark currents are listed inside the round parentheses and are separated by colons. (In addition we use a semicolon to distinguish quark currents involving different flavors, though this notation is not made explicit in \eq{Opm_color}.) Finally in the square brackets we first list the $\mathcal{P}_\perp$ helicities, followed by a colon and then entries $\pm$ or $0$ to indicate the presence of ultrasoft derivatives $\partial_{us(i)\pm}$ and $\partial_{us(i)0}$ to be discussed below. Explicit examples that fully exploit this notation will be given in \sec{constr_subops}. When working at leading power, it is usually true that all possible combinations of helicity labels need to be included in the basis. As will be reviewed in \Sec{sec:ang_cons}, angular momentum conservation of the hard scattering process places many constraints on subleading power operators \cite{Kolodrubetz:2016uim}, so that various helicity combinations can be eliminated. Operators with $\mathcal{P}_\perp$s acting on collinear gauge invariant objects can also often be eliminated by RPI and momentum conservation considerations.
We can demonstrate how the helicity and color decomposition works with two leading power examples, which only involve operators formed from collinear fields. We consider $pp \to 2$ jets, and for simplicity, we restrict to the $q \bar q\, q' \bar q'$ channel with distinct quark flavors to demonstrate the use of the collinear quark fields, and the $gggg$ channel to demonstrate the use of the collinear gluon fields. For the quark channel we take $q$ to be $n_1$ collinear, $\bar{q}$ to be $n_2$ collinear, $q'$ to be $n_3$ collinear, and $\bar{q}'$ to be $n_4$ collinear. For the gluon channel we take the collinear gluon fields to lie in four distinct collinear sectors, labelled $n_1$ through $n_4$.
Using the notation of the traditional SCET building blocks in \Tab{tab:PC}, such operators are given by
\begin{align}
O^{\, {\bar \alpha} \beta{\bar \gamma}\delta} = \bar \chi_{n_1}^{\bar \alpha} \Gamma_1 \chi^\beta_{n_2} \bar \chi^{\bar \gamma}_{n_3} \Gamma_2 \chi^\delta_{n_4}\,,
\end{align}
for the four quark case, and
\begin{align}\label{eq:oldgggg}
O^{a b c d} = \mathcal{B}_{n_1 \perp}^{ \mu a} \mathcal{B}_{n_2 \perp }^{\nu b} \mathcal{B}_{n_3 \perp}^{ \sigma c} \mathcal{B}_{n_4 \perp}^{ \delta d} \Gamma_{\mu \nu \sigma \delta},
\end{align}
for the four gluon case. Here $\Gamma_{\mu \nu \sigma \delta}$ is a shorthand for all allowed Lorentz structures, while $\Gamma_1$ and $\Gamma_2$ are shorthand for all possible Lorentz and Dirac structures, including contractions. Actually enumerating a minimal basis of these structures is a nontrivial task. On the other hand, using the helicity basis described in this section, we find \cite{Moult:2015aoa} that there are four independent helicity operators for the quark process,
\begin{alignat}{2} \label{eq:qqQQ_basis}
&O_{(+;+)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{12+}^{{\bar \alpha}\beta}\, J_{34+}^{{\bar \gamma}\delta}
\,, \qquad &
&O_{(+;-)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{12+}^{{\bar \alpha}\beta}\, J_{34-}^{{\bar \gamma}\delta}
\,,\\
&O_{(-;+)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{12-}^{{\bar \alpha}\beta}\, J_{34+}^{{\bar \gamma}\delta}
\,, \qquad&
&O_{(-;-)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{12-}^{{\bar \alpha}\beta}\, J_{34-}^{{\bar \gamma}\delta}
\,,{\nonumber}
\end{alignat}
with a color basis given by
\begin{equation} \label{eq:qqqq_color}
\bar{T}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\Bigl(
\delta_{\alpha{\bar \delta}}\, \delta_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\Bigr)
\,.\end{equation}
Similarly, we can immediately write down a basis of helicity operators for the gluon process
\begin{align} \label{eq:helicitygggg}
\mathcal{O}^{a b c d}_{++++}
&= \frac{1}{4!} \mathcal{B}^a_{1 +} \mathcal{B}^b_{2 +} \mathcal{B}^c_{3 +} \mathcal{B}^d_{4 +} \,\,, &\mathcal{O}^{a b c d}_{+++-} &= \frac{1}{3!} \mathcal{B}^a_{1 +} \mathcal{B}^b_{2 +} \mathcal{B}^c_{3 +} \mathcal{B}^d_{4 -}\,,
\\
\mathcal{O}^{a b c d}_{++--}
&= \frac{1}{4} \mathcal{B}^a_{1 +} \mathcal{B}^b_{2 +} \mathcal{B}^c_{3 -} \mathcal{B}^d_{4 -} \,\,, &\mathcal{O}^{a b c d}_{+---} &= \frac{1}{3!} \mathcal{B}^a_{1 -} \mathcal{B}^b_{2 -} \mathcal{B}^c_{3 -} \mathcal{B}^d_{4 +}\,,
{\nonumber} \\
\mathcal{O}^{a b c d}_{----}
&= \frac{1}{4!} \mathcal{B}^a_{1 -} \mathcal{B}^b_{2 -} \mathcal{B}^c_{3 -} \mathcal{B}^d_{4 -}\,, {\nonumber}
\end{align}
with a color basis
\begin{equation} \label{eq:color_gggg}
\bar{T}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
{\rm tr}[abcd] + {\rm tr}[dcba] \\ {\rm tr}[acdb] + {\rm tr}[bdca] \\ {\rm tr}[adbc] + {\rm tr}[cbda] \\
2{\rm tr}[ab]\, {\rm tr}[cd] \\ 2{\rm tr}[ac]\, {\rm tr}[db] \\ 2{\rm tr}[ad]\, {\rm tr}[bc]
\end{pmatrix}^{\!\!\!T}
.\end{equation}
(See \Ref{Moult:2015aoa} for a more detailed discussion. For a pedagogical review of color bases for QCD amplitudes see \Refs{Dixon:1996wi,Dixon:2013uaa}.) Note that there is no complication of dealing with Dirac structures, or using equations of motion to determine the minimal operator basis. The operators only encode relevant information on the helicities and color configuration of the particles. By using the helicity building blocks, which behave like scalar fields, we reduce the process of constructing an operator basis to simply enumerating unique combinations of the scalar objects. The hard kinematics is then described by the Wilson coefficients of these operators.
The simplicity of the color bases for the collinear operators does not, however, naively extend to operators involving ultrasoft fields. Indeed, the ultrasoft derivatives are local at the ultrasoft scale, requiring the use of the ultrasoft covariant derivative $D_{us}$, and reintroducing the problem of the color decomposition that is present in the full theory, namely that the constraints of gauge invariance must be implemented. In the next section we show how this issue can be overcome by using objects that account for the action of the BPS field redefinition. By working in terms of the resulting more non-local operators, we can reduce the constraints of gauge invariance to global color, enabling us to extend the simple color decomposition discussed in this section for the collinear building blocks to ultrasoft building blocks.
\subsection{Ultrasoft Gauge Invariant Helicity Building Blocks}\label{sec:BPS}
The BPS field redefinition is defined by \cite{Bauer:2002nz}
\begin{equation} \label{eq:BPSfieldredefinition}
\mathcal{B}^{a\mu}_{n\perp}\to \mathcal{Y}_n^{ab} \mathcal{B}^{b\mu}_{n\perp} , \qquad \chi_n^\alpha \to Y_n^{\alpha \bar \beta} \chi_n^\beta,
\end{equation}
and is performed in each collinear sector. Here $Y_n$, $\mathcal{Y}_n$ are fundamental and adjoint ultrasoft Wilson lines, respectively, and we note that
\begin{equation} \label{eq:adjointtofundamental}
Y_n T^a Y_n^\dagger = T^b {\cal Y}_n^{ba}\,.
\end{equation}
For a general representation, r, the ultrasoft Wilson line is defined by
\begin{equation}
Y^{(r)}_n(x)=\bold{P} \exp \left [ ig \int\limits_{-\infty}^0 ds\, n\cdot A^a_{us}(x+sn) T_{(r)}^{a}\right]\,,
\end{equation}
where $\bold P$ denotes path ordering. The BPS field redefinition has the effect of decoupling the ultrasoft degrees of freedom from the leading power collinear Lagrangian \cite{Bauer:2002nz}. When this is done consistently for S-matrix elements it accounts for the full physical path of ultrasoft Wilson lines~\cite{Chay:2004zn,Arnesen:2005nk}, so $Y_n^{(r)}$ may also occur with a path from $(0,\infty)$. Indeed for $e^+e^-\to$ dijets all the ultrasoft Wilson lines occur with paths from $(0,\infty)$ (see eg.~\cite{Bauer:2002ie}).
After the BPS field redefinition, the fields $\mathcal{B}_{n\perp}$, and $\chi_n$ are ultrasoft gauge singlets, but still carry a global color index. We can use the BPS field redefinition to define ultrasoft quark and gluon fields that are ultrasoft gauge invariant. These operators are non-local at the ultrasoft scale, and involve the ultrasoft Wilson lines. For their construction, it is essential that the non-locality is dictated by the form of the BPS decoupling. In particular, the matching is first done onto the SCET Lagrangian pre-BPS field redefinition, which is local at the hard scale, and then the BPS decoupling is performed.
We begin by defining an ultrasoft gauge invariant quark building block field
\begin{align} \label{eq:usgaugeinvdef}
\psi_{us(i)}=Y^\dagger_{n_i} q_{us}\,,
\end{align}
where the direction of the Wilson line $n_i$ is a label for a collinear sector. Since the ultrasoft quarks are not naturally associated with an external label direction, $n_i$ can be chosen arbitrarily. However, there is often a convenient or obvious choice. The definition in \Eq{eq:usgaugeinvdef} straightforwardly generalizes to matter in an arbitrary representation. We also perform the following decomposition of the gauge covariant derivative in an arbitrary representation, $r$,
\begin{align}\label{eq:soft_gluon}
Y^{(r)\,\dagger}_{n_i} i D^{(r)\,\mu}_{us} Y^{(r)}_{n_i }=i \partial^\mu_{us} + [Y_{n_i}^{(r)\,\dagger} i D^{(r)\,\mu}_{us} Y^{(r)}_{n_i}]=i\partial^\mu_{us}+T_{(r)}^{a} g \mathcal{B}^{a\mu}_{us(i)}\,,
\end{align}
where we have defined the ultrasoft gauge invariant gluon building block field by
\begin{align} \label{eq:softgluondef}
g \mathcal{B}^{a\mu}_{us(i)}= \left [ \frac{1}{in_i\cdot \partial_{us}} n_{i\nu} i G_{us}^{b\nu \mu} \mathcal{Y}^{ba}_{n_i} \right] \,.
\end{align}
In the above equations the derivatives act only within the square brackets. Again, the choice of collinear sector label $n_i$ here is arbitrary. This is the ultrasoft analogue of the gauge invariant collinear gluon field, which can be written
\begin{align}
g\mathcal{B}_{n_i\perp}^{A\mu} =\left [ \frac{1}{\bar \mathcal{P}} \bar n_{i\nu} i G_{n_i}^{B\nu \mu \perp} \mathcal{W}^{BA}_{n_i} \right]\,.
\end{align}
From the expression for the gauge invariant ultrasoft gluon field of \Eq{eq:softgluondef} we see the price we have paid for working with ultrasoft gauge invariant operators. Unlike the ultrasoft fields, the building block field $\mathcal{B}^{A\mu}_{us(i)}$ is non-local at the scale $\lambda^2$, and depends on the choice of a collinear direction $n_i$. Note that in the case that the derivative operator acts in the opposite direction, we have
\begin{align}\label{eq:soft_gluon_reverse}
Y^{(r)\,\dagger}_{n_i} i \overleftarrow{D}^{(r)\,\mu}_{us} Y^{(r)}_{n_i }=i \overleftarrow \partial^\mu_{us} + [Y_{n_i}^{(r)\,\dagger} i \overleftarrow{D}^{(r)\,\mu}_{us} Y^{(r)}_{n_i}]=i \overleftarrow\partial^\mu_{us}-T_{(r)}^{a} g \mathcal{B}^{a\mu}_{us(i)}\,,
\end{align}
The subleading Lagrangians can also be written after BPS field redefinition in terms of the gauge invariant ultrasoft gluon field, as was done in Ref.~\cite{Larkoski:2014bxa},
\begin{align} \label{eq:LKB}
{\cal L}_{n_i}^{(1)\text{BPS}} &= \hat K_{n_i}^{(1)} + \hat K_{\mathcal{B} n_i\mu}^{(1) a} g \mathcal{B}_{us(i)}^{a\mu} +\mathcal{L}_{\xi_n \psi_{us}}^{(1)\text{BPS}}
\,, \\
{\cal L}_{n_i}^{(2)\text{BPS}} &= \hat K_{n_i}^{(2)} + \hat K_{\mathcal{B} n_i\mu}^{(2) a} g \mathcal{B}_{us(i)}^{a\mu}
+ \hat K_{\mathcal{B} \mathcal{B} n_i\mu\nu}^{(2) ab} g \mathcal{B}_{us(i)}^{a\mu} g \mathcal{B}_{us(i)}^{b\nu} +\mathcal{L}_{\chi_n \psi_{us}}^{(2)\text{BPS}}
\,. {\nonumber}
\end{align}
Here $\hat K_{n_i}^{(1)}$ and $\hat K_{n_i}^{(2)}$ contain only collinear fields and $i\partial_{us}^\mu$ derivatives. Their particular form is not relevant for the current discussion, we merely want to emphasize that the decomposition into gauge invariant building blocks is also convenient at the level of the Lagrangian, allowing it to be written in a factorized form. The terms $\mathcal{L}_{\xi_n \psi_{us}}^{(1)\text{BPS}}$ and $\mathcal{L}_{\xi_n \psi_{us}}^{(2)\text{BPS}}$ involve ultrasoft quark fields. Note that the superscript $i$ for $\hat K^{(i)}$ indicates the Lagrangian that these terms contribute to and not their individual power counting.
With the ultrasoft gauge invariant operators defined, we can now introduce ultrasoft fields and currents of definite helicity, following closely the collinear operators, but with some important differences. Throughout this section we implicitly work post BPS field redefinition. In \Sec{sec:constr_subops} we will show how an operator basis can be constructed prior and post BPS field redefinition, and how to easily treat the corresponding color bases.
We begin by defining ultrasoft gluon helicity fields which are ultrasoft gauge invariant
\begin{equation} \label{eq:Bus}
\mathcal{B}^a_{us(i)\pm} = -\varepsilon_{\mp\mu}(n_i, {\bf n}_i)\,\mathcal{B}^{a\mu}_{us(i)},\qquad \mathcal{B}^a_{us(i)0} =\bar n_\mu \mathcal{B}^{a \mu}_{us(i)}
\,.\end{equation}
From \eq{softgluondef}, we can see that the ultrasoft gluon field satisfies the relation
\begin{align}\label{eq:ndotBiszero}
n_i\cdot \mathcal{B}^{a}_{us(i)}= 0\,.
\end{align}
For the collinear gauge invariant gluon field there are only two building block fields, which correspond to the two physical helicities of the gluon. On the other hand, for the ultrasoft gauge invariant gluon field we use three building block fields to describe the two physical degrees of freedom. This occurs because the ultrasoft gluons are homogeneous and not fundamentally associated with any direction. Therefore, without making a further gauge choice their polarization vectors do not lie in the perpendicular space of any fixed external reference vector. Note that if we use the ultrasoft gauge freedom to choose $\mathcal{B}^a_{us(j)0}=0$, then we will still have $\mathcal{B}^a_{us(i)\pm}\ne 0$ and $\mathcal{B}^a_{us(i)0} \ne 0$ for $i\ne j$. We could remove $\mathcal{B}^a_{us(i)0}$ using the ultrasoft gluon equation of motion, in a manner analogous to how $[W_n^\dagger i n\cdot D_n W_n]$ is removed for the collinear building blocks. However this would come at the expense of needing to allow inverse ultrasoft derivatives, $1/(in\cdot\partial_{us})$, to appear when building operators. In the collinear case these $1/\overline {\mathcal P}$ factors are ${\cal O}(\lambda^0)$ and can be absorbed into the Wilson coefficients, but this is not possible for the ultrasoft case. Therefore we choose to forbid inverse ultrasoft derivatives and allow $\mathcal{B}^a_{us(i)0}$ to appear.
When writing down the Feynman rules for external ultrasoft gluons, we have the freedom to choose the reference vector for their polarizations. Choosing a general reference vector $k$, the tree level Feynman rules for the ultrsoft gluon field are
\begin{align} \label{eq:usgluonbaseFR}
\langle g_{us}^a(p) | \mathcal{B}_{us(i) \pm}^b(x) |0 \rangle & = - \varepsilon_{\mp \mu}(n_i {\bf n}_i) \big[ \varepsilon^\mu(p,k) - \frac{p^\mu}{n \cdot p} n \cdot \varepsilon(p,k)\big] \delta^{ab}e^{ip\cdot x}\,,
\\
\langle g_{us}^a(p) | \mathcal{B}_{us(i) 0}^b(x) |0 \rangle & = \big[ {\bf n} \cdot \varepsilon^\mu(p,k) - \frac{{\bf n} \cdot p}{n \cdot p} n \cdot \varepsilon(p,k)\big] \delta^{ab}e^{ip\cdot x}\,,
{\nonumber} \\
\langle 0 | \mathcal{B}_{us(i) \pm}^b(x) |g_{us}^a(-p) \rangle & = - \varepsilon_{\mp \mu}(n_i {\bf n}_i) \big[ \varepsilon^{* \mu}(p,k) - \frac{p^\mu}{n \cdot p} n \cdot \varepsilon^*(p,k)\big] \delta^{ab}e^{ip\cdot x}\,,
{\nonumber} \\
\langle 0 | \mathcal{B}_{us(i) 0}^b(x) |g_{us}^a(-p) \rangle & = \big[ {\bf n} \cdot \varepsilon^{*\mu}(p,k) - \frac{{\bf n} \cdot p}{n \cdot p} n \cdot \varepsilon^*(p,k)\big] \delta^{ab}e^{ip\cdot x}\,. {\nonumber}
\end{align}
We also decompose the ultrasoft partial derivative operator $\partial_{us}^\mu$ into lightcone components,
\begin{equation} \label{eq:partialus}
\partial_{us(i)\pm} = -\varepsilon_{\mp\mu}(n_i, {\bf n}_i)\,\partial^{\mu}_{us},\qquad \partial_{us(i)0} =\bar n_{i\mu} \partial^{\mu}_{us}, \qquad \partial_{us(i)\bar 0} = n_{i \mu} \partial^{\mu}_{us}
\,.\end{equation}
If the $\partial_{us(i)\bar 0}$ operator acts on an $n$-collinear field, then it can be eliminated using the equations of motion \cite{Marcantonini:2008qn}, and therefore such a combination does not need to be included in our basis.
In contrast with the collinear case, we cannot eliminate the ${\bf n}_i \cdot \partial_{us}$ using the equations of motion without introducing inverse ultrasoft derivative (e.g. $1/({\bf n}_i \cdot \partial_{us})$), so we instead keep these operators explicitly in our basis. When inserting ultrasoft derivatives into operators we will use the same curly bracket notation defined for the $\mathcal{P}_\perp$ operators in \Eq{eq:p_perp_notation}. In other words, $\{i \partial_{us(i)\lambda} J\}$ indicates that the ultrasoft derivative acts from the left on the first field in $J$ and $\{ J (i\partial_{us(i) \lambda} )^\dagger \}$ indicates that it acts from the right on the second field in $J$. Note that the appearance of $\partial_{us(i) 0}$ and $\partial_{us(i)\bar 0}$ is also constrained by RPI-III invariance.
Gauge invariant ultrasoft quark fields also appear explicitly in the operator basis at subleading powers. Due to fermion number conservation they are conveniently organized into scalar currents. From \Tab{tab:PC}, we see that ultrasoft quark fields power count like $\lambda^3$. However, for factorization theorems involving a single collinear sector, as appear in factorization theorems describing a variety of both inclusive, and exclusive $B$ decays,
operators involving ultrasoft quarks appear at leading power. The currents involving ultrasoft quarks that are necessary to define subleading power operators at any desired order are
\begin{align} \label{eq:Jus}
J_{i(us)\pm}^{{\bar \alpha}\beta}
&= \mp
\frac{\varepsilon_\mp^\mu(n_i, \bar n_i)}{ \langle {\bf n}_i \mp | n_i \pm \rangle}\:
\bar{\chi}^{\bar \alpha}_{i\pm}\, \gamma_\mu \psi^\beta_{us(i)\pm}
\,, \\
J_{i(\overline{us})\pm}^{{\bar \alpha}\beta}
&=\mp \frac{\varepsilon_\mp^\mu(\bar{n}_i, n_i)}{ \langle n_i \mp | \bar{n}_i \pm \rangle}\: \bar{\psi}_{us(i) \pm}^{\bar \alpha}\, \gamma_\mu \chi^\beta_{i\pm}
\, , {\nonumber} \\
J_{i(us)0}^{{\bar \alpha}\beta}
&= \bar \chi^{\bar \alpha}_{i+}\psi^\beta_{us(i)-}
\,, \qquad\qquad\qquad
(J^\dagger)_{i(us)0}^{{\bar \alpha}\beta}
= \bar \psi^{\bar \alpha}_{us(i)-} \chi^\beta_{i+}
\,, {\nonumber}\\
J_{i(\overline{us})0}^{{\bar \alpha}\beta}
&= \bar\psi^{\bar \alpha}_{us(i)+} \chi^\beta_{i-}
\,, \qquad\qquad\qquad
(J^\dagger)_{i(\overline{us})0}^{{\bar \alpha}\beta}
= \bar \chi^{\bar \alpha}_{i-}\psi^\beta_{us(i)+}
\,, {\nonumber}\\
J_{(us)^2 ij\pm}^{{\bar \alpha}\beta}
&= \mp\, \frac{\varepsilon_\mp^\mu(n_i, n_j)}{\langle n_j\mp | n_i\pm\rangle}\:
\bar{\psi}^{\bar \alpha}_{us(i)\pm} \gamma_\mu \psi^\beta_{us(j)\pm}
\,,{\nonumber} \\
J_{(us)^2 ij 0}^{{\bar \alpha}\beta}
&=
\bar \psi^{\bar \alpha}_{us(i)+}\psi^\beta_{us(j)-}
\,,\qquad\qquad\quad
(J^\dagger)_{(us)^2 ij 0}^{{\bar \alpha}\beta}
=
\bar \psi^{\bar \alpha}_{us(i)-}\psi^\beta_{us(j)+}
\,. {\nonumber}
\end{align}
For the mixed collinear-ultrasoft currents we choose the collinear sector label $i$ in order to specify the ultrasoft quark building block field. For the ultrasoft-ultrasoft currents, there is freedom to choose two sectors to use to construct the collinear gauge invariant ultrasoft fields, and we leave the choices arbitrary, $i$ and $j$, which appear as subscripts after the $(us)^2$ on the currents. Although the ultrasoft quark carries a label, this label is only associated with the ultrasoft Wilson line structure used to define the building block field. In particular, it is important to realize that the ultrasoft quark field does not satisfy the projection relations of \eq{proj}, that is satisfied by the collinear quarks. Note that all the ultrasoft quark currents are already RPI-III invariant except for $J_{(us)^2 ij 0}^{{\bar \alpha}\beta}$ for generic $i$ and $j$.
The Feynman rules for the collinear-ultrasoft currents to external states are
\begin{align}\label{eq:feyn_col_us}
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{us-}^{{\bar \alpha}_2}(p_2) | J_{1(us)+}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &= \frac{\sqrt{\omega_1}}{2}e^{i\Phi(J_{1(us)+})} \langle n_1 p_2 \rangle\, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) e^{ip_2\cdot x}\,, \\
\langle q_{-}^{\alpha_1}(p_1) \bar{q}_{us+}^{{\bar \alpha}_2}(p_2) | J_{1(us)-}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &=\frac{\sqrt{\omega_1}}{2} e^{i\Phi(J_{1(us)-})} [ n_1 p_2 ] \, \, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) e^{ip_2\cdot x}\,, {\nonumber} \\
\langle q_{us+}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | J_{2(\overline{us})+}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &=\sqrt{\frac{\omega_2}{2}} e^{i\Phi(J_{2(\overline{us})+})} [n_2 p_1] \, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_2 - p_2) e^{ip_1\cdot x}\,, {\nonumber} \\
\langle q_{us-}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{2(\overline{us})-}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &=\sqrt{\frac{\omega_2}{2}} e^{i\Phi(J_{2(\overline{us})+})} \langle n_2 p_1\rangle \, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_2 - p_2) e^{ip_1\cdot x}\,, {\nonumber} \\
\langle q_{+}^{\alpha_1}(p_1) \bar{q}_{us+}^{{\bar \alpha}_2}(p_2) | J_{1(us)0}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &= \sqrt{\frac{\omega_1}{2}} e^{i\Phi(J_{1(us)0})} [n_1 p_2] \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) e^{ip_2\cdot x} \,, {\nonumber} \\
\langle q_{us-}^{\alpha_1}(p_1) \bar{q}_{-}^{{\bar \alpha}_2}(p_2) | (J^\dagger)_{2(us)0}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &= \sqrt{\frac{\omega_2}{2}} e^{i\Phi((J^\dagger)_{2(us)0})} \langle n_2 p_1\rangle \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_2 - p_2) e^{ip_1\cdot x} \,, {\nonumber} \\
\langle q_{us+}^{\alpha_1}(p_1) \bar{q}_{+}^{{\bar \alpha}_2}(p_2) | J_{2(\overline{us})0}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &= \sqrt{\frac{\omega_2}{2}} e^{i\Phi(J_{2(\overline{us})0})} [n_2 p_1] \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_2 - p_2) e^{ip_1\cdot x}\,, {\nonumber} \\
\langle q_{-}^{\alpha_1}(p_1) \bar{q}_{us-}^{{\bar \alpha}_2}(p_2) | (J^\dagger)_{1(\overline{us}) 0}^{{\bar \beta}_1 \beta_2}(x) | 0 \rangle &= \sqrt{\frac{\omega_1}{2}} e^{i\Phi((J^\dagger)_{1(\overline{us})0})} \langle n_1 p_2\rangle \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2} \tilde{\delta} (\lp_1 - p_1) e^{ip_2\cdot x} \,. {\nonumber}
\end{align}
The phases appearing in the Feynman rules of \Eq{eq:feyn_col_us} are defined as
\begin{align}
& e^{i\Phi(J_{1(us)+})}=\frac{ [\bar n_1 p_1 ]}{\sqrt{2 \omega_1}} \,,
& e^{i\Phi(J_{1(us)-})}=\frac{ \langle\bar n_1 p_1 \rangle}{\sqrt{2 \omega_1} } \,, \\
& e^{i\Phi(J_{2(\overline{us})+})}=\frac{ \langle \bar n_2 p_2 \rangle}{\sqrt{2 \omega_2} } \,,
& e^{i\Phi(J_{2(\overline{us})+})}=\frac{[ \bar n_2 p_2 ]}{\sqrt{2 \omega_2} } \,, {\nonumber} \\
& e^{i\Phi(J_{1(us)0})}=\frac{ \langle n_1 {\bf n}_1 \rangle [ {\bf n}_1 p_1]}{2\sqrt{2 \omega_1} } \,,
& e^{i\Phi((J^\dagger)_{2(us)0})}=\frac{- [ n_2 {\bf n}_2 ] \langle {\bf n}_2 p_2\rangle}{2\sqrt{2 \omega_2} } \,, {\nonumber}\\
& e^{i\Phi((J)_{2(\overline{us})0})}=\frac{- \langle n_2 {\bf n}_2 \rangle [ {\bf n}_2 p_2]}{2\sqrt{2 \omega_2} } \,,
& e^{i\Phi((J^\dagger)_{1(\overline{us})0})}=\frac{ [ n_1 {\bf n}_1 ] \langle {\bf n}_1 p_1\rangle}{2\sqrt{2 \omega_1} } \,. {\nonumber}
\end{align}
The phases involve only the momentum of the collinear field.
Additional $\mathcal{P}_\perp^\pm$ insertions into the mixed ultrasoft-collinear currents are defined following the notation of \eq{p_perp_notation}, where we note that in \mbox{${\rm SCET}_{\rm I}$}\xspace only the collinear fields carry label $\perp$ momentum. The Feynman rules for the currents involving two ultrasoft quark fields are
\begin{align}
\langle q_{us+}^{\alpha_1}(p_1) \bar{q}_{us-}^{{\bar \alpha}_2}(p_2) | J_{(us)^2 12+}^{{\bar \beta}_1 \beta_2} | 0 \rangle &=\frac{[n_2 p_1] \langle n_1 p_2 \rangle}{\sqrt{2}\, n_1 \cdot n_2} \, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}\,,\\
\langle q_{us-}^{\alpha_1}(p_1) \bar{q}_{us+}^{{\bar \alpha}_2}(p_2) | J_{(us)^2 12-}^{{\bar \beta}_1 \beta_2} | 0 \rangle &=\frac{ \langle n_2 p_1\rangle [ n_1 p_2] }{\sqrt{2}\, n_1 \cdot n_2}\, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}\,, {\nonumber} \\
\langle q_{us+}^{\alpha_1}(p_1) \bar{q}_{us+}^{{\bar \alpha}_2}(p_2) | J_{(us)^2 12\, 0 }^{{\bar \beta}_1 \beta_2} | 0 \rangle &= [p_1 p_2] \, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}\,, {\nonumber} \\
\langle q_{us-}^{\alpha_1}(p_1) \bar{q}_{us-}^{{\bar \alpha}_2}(p_2) | (J^\dagger)_{(us)^2 12\, 0}^{ {\bar \beta}_1 \beta_2} | 0 \rangle &=\langle p_1 p_2 \rangle\, \delta^{\alpha_1 {\bar \beta}_1} \delta^{\beta_2 {\bar \alpha}_2}\,. {\nonumber}
\end{align}
Due to the dependence on two ultrasoft momenta, there is no natural way to separate phases from these Feynman rules as was done for previous currents.
\subsection{Constructing Operator Bases}\label{sec:constr_subops}
In this section we describe how the building blocks of the previous sections can be combined to define bases of hard scattering operators. The hard scattering component of the SCET Lagrangian, \Eq{eq:SCETLagExpand}, has an explicit expansion in powers of $\lambda$,
\begin{equation} \label{eq:Leff_sub}
\mathcal{L}_{\text{hard}} = \sum_{j\geq0} \mathcal{L}^{(j)}_{\text{hard}}
\,.\end{equation}
Here $j$ denotes suppression by ${\cal O}(\lambda^j)$ with respect to the leading power hard scattering operators. The effective Lagrangian for hard scattering operators at each power is given by,
\begin{align} \label{eq:Leff_sub_explicit}
\mathcal{L}^{(j)}_{\text{hard}} = \sum_{\{n_i\}} \sum_{A,\cdot\cdot}
\bigg[ \prod_{i=1}^{\ell_A} \int \! \! \mathrm{d} \omega_i \bigg] \,
& \vec O^{(j)\dagger}_{A\,\lotsdots}\big(\{n_i\};
\omega_1,\ldots,\omega_{\ell_A}\big) {\nonumber}\\
& \times
\vec{C}^{(j)}_{A\,\lotsdots}\big(\{n_i\};\omega_1,\ldots,\omega_{\ell_A} \big)
\,.
\end{align}
The appropriate collinear sectors $\{n_i\}$ are determined by the hard process being considered, and the sum over $A,\cdot\cdot$ runs over the full basis of operators that appear at this order, which are specified by either explicit labels $A$ and/or helicity labels $\cdot\cdot$ on the operators and coefficients. The operators also satisfy momentum conservation for various momenta, including the ${\cal O}(\lambda^0)$ $\omega_i$'s. Here the $\vec{C}^{(j)}_{A}$ are ${\cal O}(\lambda^0)$ Wilson coefficients, and are also vectors in the color subspace in which the $\mathcal{O}(\lambda^j)$ hard scattering operators $\vec O_A^{(j)\dagger}$ are decomposed. Explicitly, in terms of color indices, we have
\begin{equation} \label{eq:Cpm_color}
C_{\lotsdots}^{a_1\dotsb\alpha_n}
= \sum_k C_{\lotsdots}^k T_k^{a_1\dotsb\alpha_n}
\equiv \bar{T}^{ a_1\dotsb\alpha_n} \vec{C}_{\lotsdots}
\,.\end{equation}
Note that the color bases used to decompose the hard scattering operators at each order in $\lambda$ are in general distinct, since at higher powers more building blocks appear which carry additional color indices. The Wilson coefficients depend only on the jet directions, $\{n_i\}$, and the large label components of the operators $\omega_i$. The number of $\omega_i$'s depends on the specific operator we are considering since at subleading power multiple collinear fields can appear in the same collinear sector and we must consider the inclusion of ultrasoft building blocks with no $\omega_i$ labels. For a given operator $A\cdot\cdot$, we label the number of $\omega_i$'s as $\ell_{A\cdot\cdot}$ in \eq{Leff_sub_explicit} and do the integral over each $\omega_i$. The operators $\vec O^\dagger_{\lotsdots}$ are given by products of the quark and gluon helicity operators of \Tab{tab:helicityBB}.
At subleading power, this is complicated by the fact that the ultrasoft Wilson lines which appear in the ultrasoft gauge invariant building blocks only appear after the BPS field redefinition. There are two possible approaches to dealing with this issue, both of which give the same answer. We take the attitude that one can use whichever is most convenient. A priori, we might like to first know how the ultrasoft fields are color contracted in the operator in order to choose the appropriate collinear vector $n_i$ for defining each ultrasoft building block. In one approach we first determine the full color basis involving contractions for the ultrasoft fields' color indices, and then choose the building blocks. Alternatively, we can simply pick fixed $n_i$ vectors for the ultrasoft building blocks, so that we have common ultrasoft objects for all operators. In this case there will be products of ultrasoft Wilson lines in the operators to compensate for a choice that does not correspond with the color contraction.
In either approach a basis of operators can be constructed in the form
\begin{equation} \label{eq:Opm_BPScolor}
\vec O^\dagger_{\lotsdots}
= O_{\lotsdots}^{a_1\dotsb \alpha_n}\, \bar{T}_{\mathrm{BPS}}^{\, a_1\dotsb \bar \alpha_n}
\,.
\end{equation}
Here $O_{\lotsdots}^{a_1\dotsb \alpha_n}$ is formed from products of the collinear and ultrasoft gauge invariant helicity building blocks constructed in the previous sections, for which the complete list is given in \Tab{tab:helicityBB}. The meaning of $ \bar{T}_{\mathrm{BPS}}^{\, a_1\dotsb \alpha_n}$ will be discussed below. When utilizing $\mathcal{P}_{i\perp}^{\pm}$, $\partial_{us(i)\pm}$, and $\partial_{us(i)0}$ to construct operators, these derivatives can either act on a single building block object like $\mathcal{B}^a_{i\pm}$ or $\mathcal{B}^a_{us(i)\pm}$, or on a bilinear object like one of the currents $J$. For cases where they act on a bilinear object we use the curly bracket notation introduced in \eq{p_perp_notation} to indicated which of the two objects in the bilinear the derivative acts on. Note that $\mathcal{P}_{i\perp}^{\pm}$ can only act on $n_i$-collinear building blocks, whereas $\partial_{us(i)\pm}$ and $\partial_{us(i)0}$ can act on any building block field. The convention for the subscripts used on $O^{a_1 \cdots \alpha_n}$ was defined below \eq{Opm_color}. A couple of more complicated examples of the use of this notation are
\begin{align}
O_{A\:-0: +-(+\bar 0:-)}^{ a_1 a_2 a_3 a_4 {\bar \alpha} \beta {\bar \gamma} \delta \bar\sigma \tau}
&= \mathcal{B}_{us(1)-}^{a_1} \mathcal{B}_{us(1)0}^{a_2} \mathcal{B}_{1+}^{a_3} \mathcal{B}_{2-}^{a_4}
J_{3+}^{{\bar \alpha}\beta} J_{2\,\bar 0}^{{\bar \gamma}\delta} J_{45-}^{\bar\sigma\tau}
\,, \\
O_{B\: -:+-(\bar 0:-)[+:-]}^{ a_1 a_2 a_3 {\bar \alpha} \beta {\bar \gamma} \delta}
&= \mathcal{B}_{us(1)-}^{a_1} \big[ i\partial_{us(1)}^- \mathcal{B}_{2+}^{a_2} \big] \mathcal{B}_{2-}^{a_3} J_{3\,\bar 0}^{{\bar \alpha} \beta} \big\{ P_{\perp}^+ J_{14-}^{{\bar \gamma}\delta} \big\} \,.{\nonumber}
\end{align}
Besides illustrating the correspondence between the operator subscripts and the helicity building blocks, these examples also highlight some of the limits of the notation. In particular, specifying the collinear gluon building block helicities does not determine whether the corresponding building blocks are in the same or different collinear sectors, specifying the helicities for mixed collinear-ultrasoft currents does not fix the sector of the collinear field in this building block, and the notation for the derivatives does not fix which object they act on. To distinguish these type of differences we adopt additional explicit labels on the operators, which here are $A$ and $B$.
\begin{table}
\begin{center}
\begin{tabular}{|ccc|cccc|cccc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
$\mathcal{B}_{i\pm}^a$ & $J_{ij\pm}^{{\bar \alpha}\beta}$ & $J_{ij0}^{{\bar \alpha}\beta}$ & $\mathcal{P}_{\perp}^{\pm}$ & $J_{i\pm}^{{\bar \alpha} \beta}$ &
$J_{i0}^{{\bar \alpha} \beta}$ & $J_{i\bar 0}^{{\bar \alpha} \beta}$ &
$\mathcal{B}^a_{us(i)\pm}$ & $\mathcal{B}^a_{us(i)0}$ &
$\partial_{us(i)\pm}$ & $\partial_{us(i)0}$
\\[3pt]
$\lambda$ & $\lambda^2$ & $\lambda^2$
& $\lambda$ & $\lambda^2$& $\lambda^2$ & $\lambda^2$
& $\lambda^2$ & $\lambda^2$ & $\lambda^2$ & $\lambda^2$
\\
\hline
\end{tabular}\\
\vspace{0.3cm}
\begin{tabular}{|cccc|cc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
$J_{i(us)\pm}^{{\bar \alpha}\beta}$ &
$J_{i(\overline{us})\pm}^{{\bar \alpha}\beta}$ &
$J_{i(us)0}^{{\bar \alpha}\beta}$ &
$J_{i(\overline{us})0}^{{\bar \alpha}\beta}$ &
$J_{(us)^2ij\pm}$ & $J_{(us)^2ij0}$
\\[3pt]
$\lambda^4$ & $\lambda^4$ & $\lambda^4$
& $\lambda^4$ & $\lambda^6$& $\lambda^6$
\\
\hline
\end{tabular}
\end{center}
\vspace{-0.3cm}
\caption{
Power counting for the complete set of helicity building block operators in $\text{SCET}_\text{I}$, where the definitions for these objects are given in Eqs.~(\ref{eq:cBpm_def},~\ref{eq:definitehelicityquarkdef},~\ref{eq:jpm_def},~\ref{eq:coll_subl},~\ref{eq:Pperppm},~\ref{eq:Bus},~\ref{eq:partialus},~\ref{eq:Jus}). The building blocks also include the conjugate currents $J^\dagger$ in cases where they are distinct from the ones shown.
}
\label{tab:helicityBB}
\end{table}
The $\bar{T}_{\mathrm{BPS}}^{\, a_1\dotsb \alpha_n}$ in \eq{Opm_BPScolor} generalizes the color structure decomposition of \Sec{sec:color}, because it is defined to include ultrasoft Wilson lines that arise from the BPS field redefinition. It is important to emphasize that products of ultrasoft Wilson lines, like $Y_{n_1}^\dagger Y_{n_2}$, should not be viewed as independent building blocks. The structure of ultrasoft Wilson lines is entirely determined by the form of the BPS field decoupling. Several examples will be given below to further clarify this point. The form in \eq{Opm_BPScolor} is convenient for factorization, since for operators involving only collinear field insertions it is already written in a factorized form.
It is also useful to have an equivalent form of the operator basis where we leave the ultrasoft Wilson lines in the operator itself, which would maintain the exact color decompositions of \Sec{sec:color}. In order to do this, we can define an operator $\widetilde{O}$ such that after BPS field redefinitions we have
\begin{align}\label{eq:Otildedef}
\vec O^\dagger_{\lotsdots}
= \widetilde{O}_{\lotsdots}^{a_1\dotsb \alpha_n}\, \bar{T}^{\, a_1\dotsb\bar \alpha_n}.
\end{align}
As we will demonstrate below, converting between $O$ and $\widetilde{O}$ is a simple exercise in reorganizing where we place the ultrasoft Wilson lines, and either form is equally valid as a basis. The decomposition in \eq{Opm_BPScolor} provides a compact way to track both the ultrasoft Wilson lines and color structure in one object, and hence will be used for many of our later examples.
We can compare the complete basis of helicity operator building blocks given in \Tab{tab:helicityBB} with the traditional form of the building block basis given in Table~\ref{tab:PC}. While there are more building blocks with the helicity operators, there is a great benefit from the fact that each of them is a scalar. So, while constructing operators with \Tab{tab:PC} is a complex exercise in deducing all possible Lorentz structures, with the helicity operator approach we simply have to write down all possible products of the helicity building blocks at a given power.
In order to demonstrate how \eq{Opm_BPScolor} works in practice, we now give some simple examples. We begin by discussing cases involving only collinear fields. Here the BPS field redefinition is not necessary to define color decomposed helicity operators, but is essential in the proof of factorization. Note that insertions of the $\mathcal{P}_\perp$ operator have no effect, since they do not carry color and commute with the ultrasoft Wilson lines $[\mathcal{P}_\perp, Y_n]=0$, so we will only consider examples without $\mathcal{P}_\perp$ insertions.
We begin with a simple leading power example with gluon and quark current building blocks,
\begin{equation} \label{eq:349}
O^{a\bar \alpha \beta}_{+(\pm)}= \mathcal{B}_{1 +}^a\,J_{23\pm}^{{\bar \alpha}\beta}\,, \qquad O^{a\bar \alpha \beta}_{-(\pm)}= \mathcal{B}_{1 -}^a\,J_{23\pm}^{{\bar \alpha}\beta} \,.
\end{equation}
In this case there is a unique color structure before BPS field redefinition, namely
\begin{equation} \label{eq:gqqcolor}
\bar{T}^{ a \alpha{\bar \beta}} = \left ( T^a \right )_{\alpha {\bar \beta}}\,.
\end{equation}
After BPS field redefinition, we find the Wilson line structure,
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}} = \left ( Y^\dagger _{n_2} T^b \mathcal{Y}_{n_1}^{ba} Y_{n_3} \right )_{\alpha \bar\beta}\,.
\end{equation}
The key point is that this structure is entirely determined by the form of the operator in \Eq{eq:349}, as well as the structure of the BPS field redefinition. The ultrasoft Wilson lines arise only from the BPS field redefinition of the collinear fields in the building blocks, and not from hard matching. We can also reorganize the ultrasoft Wilson lines to group them into the operator, with the form in \eq{Otildedef}, which gives
\begin{align}
\widetilde{O}^{a\bar \alpha \beta}_{+(\pm)}
&= (\mathcal{Y}_{n_1} \mathcal{B}_{1 +})^a (Y^\dagger_{n_2}\,J_{23 \pm} Y_{n_3})^{{\bar \alpha}\beta}\,, \qquad \widetilde{O}^{a\bar \alpha \beta}_{-(\pm)}= (\mathcal{Y}_{n_1} \mathcal{B}_{1 -})^a\,(Y^\dagger_{n_2} J_{23 \pm} Y_{n_3})^{{\bar \alpha}\beta} \,.
\end{align}
with the color structure as given in \Eq{eq:gqqcolor}.
Another illustrative example is the four quark operator discussed in \sec{color}
\begin{alignat}{2} \label{eq:qqQQ_basis}
&O_{(+;+)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{q12+}^{{\bar \alpha}\beta}\, J_{q'34+}^{{\bar \gamma}\delta}
\, \qquad &
&O_{(+;-)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{q12+}^{{\bar \alpha}\beta}\, J_{q'34-}^{{\bar \gamma}\delta}
\,,\\
&O_{(-;+)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{q12-}^{{\bar \alpha}\beta}\, J_{q'34+}^{{\bar \gamma}\delta}
\,, \qquad&
&O_{(-;-)}^{{\bar \alpha}\beta{\bar \gamma}\delta}
= J_{q12-}^{{\bar \alpha}\beta}\, J_{q'34-}^{{\bar \gamma}\delta}
\,,{\nonumber}
\end{alignat}
with color basis
\begin{equation} \label{eq:qqqq_color_second}
\bar{T}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\Bigl(
\delta_{\alpha{\bar \delta}}\, \delta_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\Bigr)
\,.\end{equation}
After BPS field redefinition, the color basis becomes
\begin{equation}\label{eq:353}
\bar{T}_{\mathrm{BPS}}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\left( \left[ Y_{n_1}^\dagger Y_{n_4} \right ]_{\alpha{\bar \delta}}\,\left[ Y_{n_3}^\dagger Y_{n_2} \right ]_{\gamma{\bar \beta}}\,,\, \left[ Y_{n_1}^\dagger Y_{n_2} \right ]_{\alpha{\bar \beta}}\, \left[ Y_{n_3}^\dagger Y_{n_4} \right ]_{\gamma{\bar \delta}}
\right)
\,.\end{equation}
This demonstrates how the decomposition in \Eq{eq:Opm_BPScolor} provides a convenient way of organizing the ultrasoft Wilson lines for the collinear operators. Once again, we could choose to simply organize the ultrasoft Wilson lines in the operators $\widetilde{O}^{{\bar \alpha} \beta \bar{\gamma} \delta}$, which would then be multiplied by the color structure given in \Eq{eq:qqqq_color_second}.
Next we consider an example involving an ultrasoft building block. We again consider four quarks, but now using currents built from two $n_1$-collinear building blocks $\chi_{1\pm}^\beta$ and $\bar\chi_{1\pm}^{\bar \alpha}$, one $n_2$-collinear building block $\bar\chi_{2\pm}^{\bar \gamma}$, and one ultrasoft building block $\psi_{us(i)\pm}^\delta$. We choose to pair up these fields into the currents $J_{1\lambda}^{{\bar \alpha}\beta}$ with $\lambda=\pm,0,\bar 0$, and $J_{2(us)\lambda'}^{{\bar \gamma}\delta}$ with $\lambda'=\pm,0$ or $(J^\dagger)_{2(\overline{us})0}^{{\bar \gamma}\delta}$ . This notation implies that we have made the choice of the $2$-sector for the ultrasoft building block field, $\psi_{us(2)\pm}^\delta$. The basis of operators is then
\begin{align} \label{eq:usBBeg}
O_{A(\lambda:\lambda')}^{{\bar \alpha}\beta{\bar \gamma}\delta} &= J_{1\lambda}^{{\bar \alpha}\beta}\: J_{2(us)\lambda'}^{{\bar \gamma}\delta}
\,,\qquad O_{B(\lambda:0)}^{{\bar \alpha}\beta{\bar \gamma}\delta} = J_{1\lambda}^{{\bar \alpha}\beta}\: (J^\dagger)_{2(\overline{us})0}^{{\bar \gamma}\delta}
\,,
\end{align}
where there are 16 distinct operators once we take into account the allowed helicity choices for $\lambda$, $\lambda'$. Again we have the color basis $\bar T^{{\bar \alpha}\beta{\bar \gamma}\delta}$ in \eq{qqqq_color_second}. Due to the different structure of fields the color basis after BPS field redefinition for this example is
\begin{align}
\bar{T}_{\mathrm{BPS}}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\left( \left[ Y_{n_1}^\dagger Y_{n_2} \right ]_{\alpha{\bar \delta}}\,\left[ Y_{n_2}^\dagger Y_{n_1} \right ]_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
In the second color structure we have no $Y_{n_1}$ Wilson lines because the $n_1$-collinear fields are color contracted, and the correct $Y_{n_2}^\dagger$ Wilson line is already contained in the $\psi_{us(2)\pm}^\delta$ building block inside $J_{2(us)\lambda'}^{{\bar \gamma}\delta}$. In the first color structure we have $[Y_{n_2}^\dagger Y_{n_1}]$ determined by the collinear building blocks, and then need $[Y_{n_1}^\dagger Y_{n_2}]$ in order to swap the ultrasoft reference vector to the $1$-sector when this factor multiplies $J_{2(us)\lambda'}^{{\bar \gamma}\delta}$. Writing out the $\widetilde O^{{\bar \alpha}\beta{\bar \gamma}\delta} \bar{T}^{\, \alpha{\bar \beta}\gamma{\bar \delta}}$ form of the operators for this case, the two color structures give
\begin{align}
&\widetilde O_{A(\lambda:\lambda')}^{{\bar \alpha}\beta{\bar \beta}\alpha}
= J_{1\lambda}^{{\bar \alpha}\beta}\:
\big( Y_{n_1} Y_{n_2}^\dagger J_{2(us)\lambda'} Y_{n_2} Y_{n_1}^\dagger \big)^{{\bar \beta}\alpha}
\,, \qquad
\widetilde O_{A(\lambda:\lambda')}^{{\bar \alpha}\alpha{\bar \gamma}\gamma}
= J_{1\lambda}^{{\bar \alpha}\alpha}\: J_{2(us)\lambda'}^{{\bar \gamma}\gamma}
\,, \\
& \widetilde O_{B(\lambda:0)}^{{\bar \alpha}\beta{\bar \beta}\alpha}
= J_{1\lambda}^{{\bar \alpha}\beta}\:
\big( Y_{n_1} Y_{n_2}^\dagger (J^\dagger)_{2(\overline{us})0} Y_{n_2} Y_{n_1}^\dagger \big)^{{\bar \beta}\alpha}
\,, \qquad \widetilde O_{B(\lambda:0)}^{{\bar \alpha}\alpha{\bar \gamma}\gamma}
= J_{1\lambda}^{{\bar \alpha}\alpha}\: (J^\dagger)_{2(\overline{us})0}^{{\bar \gamma}\gamma}
\,. {\nonumber}
\end{align}
From this example we see that the advantage of using $\bar{T}_{\mathrm{BPS}}^{\, \alpha{\bar \beta}\gamma{\bar \delta}}$ is that we do not need to be concerned with the color contractions when specifying $O_{(\lambda:\lambda')}^{{\bar \alpha}\beta{\bar \gamma}\delta}$.
Note that an equivalent way of specifying the basis in \eq{usBBeg} would be to have started with the other possible grouping of the fermion building blocks, using $J_{12+}^{{\bar \gamma}\beta}$, $J_{12-}^{{\bar \gamma}\beta}$, $J_{120}^{{\bar \gamma}\beta}$ or $(J^\dagger)_{12\,0}^{{\bar \gamma}\beta}$, together with $J_{1(us)+}$, $J_{1(us)-}$, $J_{1(us)0}$, or $(J^\dagger)_{1(\overline{us})0}$. This again gives 16 choices for the operator helicities, but now we are using the ultrasoft building block with the reference direction as the $1$-sector. The final result is the same with relations between elements of the two bases.
As another example involving ultrasoft building blocks, we consider a dijet operator that in a traditional approach has an ultrasoft derivative insertion, $\bar \chi_{n_1+}^{\bar \alpha} \big( \bar{n}_2 \cdot {D}_{us} \big) \chi_{n_2 -}^\beta$. After BPS field redefinition, we can rewrite this operator as
\begin{equation}
\bar \chi_{n_1+} \big( \bar{n}_2 \cdot {D}_{us} \big) \chi_{n_2 -}\to \bar \chi_{n_1+} \left ( Y^\dagger _{n_1} Y_{n_2} \right ) i\bar{n}_2 \cdot {\partial}_{us} \chi_{n_2 -}+ \chi_{n_1+} \left ( Y^\dagger _{n_1} Y_{n_2} \right ) g {\cal B}^a_{us(2)0} T^a \chi_{n_2 -}\,.
\end{equation}
Here we have chosen the $2$ sector to define the ultrasoft building block gluon field, ${\cal B}_{us(2)0}^a = {\bf n}_2\cdot B_{us(2)}^a$. In terms of our basis of ultrasoft gauge invariant helicity operators, we can write the two operators as
\begin{equation}
O^{\bar \alpha \beta}=\frac{-2}{\sqrt{\vphantom{2} \omega_1 \,\omega_2}\, [n_1 n_2] } \bar \chi^{\bar \alpha}_{1 +} i\partial_{us(2) 0} \chi^{\beta}_{2 -}=\big\{J_{120}^{{\bar \alpha} \beta} (i\partial_{us(2) 0})^\dagger\big\}\,, \qquad
\bar{T}_{\mathrm{BPS}}^{\alpha \bar \beta}=\left ( Y^\dagger _{n_1} Y_{n_2} \right )_{\alpha \bar\beta}\,,
\end{equation}
and
\begin{equation}
O^{\bar \alpha \beta a}= \frac{2}{\sqrt{\vphantom{2} \omega_1 \,\omega_2}\, [n_1 n_2] } \bar \chi^{\bar \alpha}_{1 +} g {\cal B}^a_{us(2) 0} \chi^{\beta}_{2 -}=J_{120}^{{\bar \alpha} \beta} g {\cal B}^a_{us(2) 0} \,, \qquad \bar{T}_{\mathrm{BPS}}^{\alpha \bar \beta a}=\left ( Y^\dagger _{n_1} Y_{n_2} T^a \right )^{\alpha \bar\beta}.
\end{equation}
The decomposition can always be done in the form of \eq{Opm_BPScolor}. Using ultrasoft Wilson lines arising from the BPS field redifinition, gauge covariant derivatives can be converted into the ultrasoft partial derivative and the ultrasoft building block gluon field. Products of the building blocks will then be linked in color space by ultrasoft Wilson lines. In particular, the remaining ultrasoft derivatives do not act on the linking Wilson lines, they only act on the ultrasoft Wilson lines and other fields that appear in the definitions of the ultrasoft gauge invariant building blocks. In general we see that given a form of an operator $O$ with color structures $\bar T$, it is straightforward to determine $\bar T_{\rm BPS}$, and thus also $\widetilde O$.
\subsection{Extensions}\label{sec:extensions}
Throughout this paper we have discussed our operator basis in the language of the \mbox{${\rm SCET}_{\rm I}$}\xspace theory with massless quarks and in $d=4$ dimensions. SCET is also used in its \mbox{${\rm SCET}_{\rm II}$}\xspace (with soft rather than ultrasoft fields) and SCET$_+$ \cite{Bauer:2011uc,Procura:2014cba,Larkoski:2015zka,Pietrulewicz:2016nwo} incarnations, with massive collinear particles, and with dimensional regularization in $d=4-2\epsilon$ dimensions, so we discuss the necessary extensions for each of these here.
\subsubsection{SCET$_{\text{II}}$}\label{sec:scet_II}
For a certain class of observables, including $p_T$ dependent measurements and exclusive decays, the theory \mbox{${\rm SCET}_{\rm II}$}\xspace provides the appropriate effective field theory description~\cite{Bauer:2002aj}. In \mbox{${\rm SCET}_{\rm II}$}\xspace the soft and collinear modes live on the same invariant mass hyperbola, and therefore modes mediating interactions between the soft and collinear modes are off-shell and can be integrated out of the theory, generating both collinear and soft Wilson lines. A convenient way of matching to the \mbox{${\rm SCET}_{\rm II}$}\xspace theory is to first match QCD onto an \mbox{${\rm SCET}_{\rm I}$}\xspace theory with a larger offshellness for the collinear fields \cite{Bauer:2002aj}. The BPS field redefinition can then be used to decouple the ultrasoft and collinear modes, giving rise to Wilson lines in the operators, as discussed in \Sec{sec:BPS}. One can then match this decoupled theory to \mbox{${\rm SCET}_{\rm II}$}\xspace by lowering the virtuality of the external collinear modes to the soft scale, and relabeling ultrasoft modes as soft. This matching calculation will be trivial (1-to-1) in cases where there are not time-ordered products of two or more subleading operators or Lagrangians in the \mbox{${\rm SCET}_{\rm I}$}\xspace theory \cite{Bauer:2003mga}. Furthermore, in the matching procedure terms of a given order in $\lambda$ in the \mbox{${\rm SCET}_{\rm I}$}\xspace theory will only contribute to terms at that same order or higher in the \mbox{${\rm SCET}_{\rm II}$}\xspace theory. The resulting Wilson coefficients in the \mbox{${\rm SCET}_{\rm II}$}\xspace theory now also involve $n\cdot k_s$ momenta of soft building blocks, from integrating out hard-collinear momenta with offshellness of order ${\bf n}\cdot p_n n\cdot k_s\sim Q^2\lambda$.
\begin{table}
\begin{center}
\begin{tabular}{|ccc|cccc|ccccc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
$\mathcal{B}_{i\pm}^a$ & $J_{ij\pm}^{{\bar \alpha}\beta}$ & $J_{ij0}^{{\bar \alpha}\beta}$ & $\mathcal{P}_{\perp}^{\pm}$ & $J_{i\pm}^{{\bar \alpha} \beta}$ &
$J_{i0}^{{\bar \alpha} \beta}$ & $J_{i\bar 0}^{{\bar \alpha} \beta}$ &
$\mathcal{B}^a_{s(i)\pm}$ & $\mathcal{B}^a_{s(i)0}$ &
$\partial_{s(i)\pm}$ & $\partial_{s(i)0}$ & $\partial_{s(i)\bar 0}$
\\[3pt]
$\lambda$ & $\lambda^2$ & $\lambda^2$
& $\lambda$ & $\lambda^2$& $\lambda^2$ & $\lambda^2$
& $\lambda$ & $\lambda$ & $\lambda$ & $\lambda$ & $\lambda$
\\
\hline
\end{tabular}\\
\vspace{0.3cm}
\begin{tabular}{|cccc|cc|}
\hline \phantom{x} & \phantom{x} & \phantom{x}
& \phantom{x} & \phantom{x} & \phantom{x}
\\[-13pt]
$J_{i(s)\pm}^{{\bar \alpha}\beta}$ &
$J_{i(\overline{s})\pm}^{{\bar \alpha}\beta}$ &
$J_{i(s)0}^{{\bar \alpha}\beta}$ &
$J_{i(\overline{s})0}^{{\bar \alpha}\beta}$ &
$J_{(s)^2ij\pm}$ & $J_{(s)^2ij0}$
\\[3pt]
$\lambda^{5/2}$ & $\lambda^{5/2}$ & $\lambda^{5/2}$
& $\lambda^{5/2}$ & $\lambda^3$& $\lambda^3$
\\
\hline
\end{tabular}
\end{center}
\vspace{-0.3cm}
\caption{
Power counting for the full set of helicity building block operators in $\text{SCET}_\text{II}$. Again we must add the conjugate currents $J^\dagger$ in cases where they are distinct from the ones shown.
}
\label{tab:helicityBB2}
\end{table}
Since the helicity operator building blocks listed in \Tab{tab:helicityBB} are defined after BPS field redefinition, they also directly apply to the description of the hard scattering operators in the \mbox{${\rm SCET}_{\rm II}$}\xspace theory. We simply need to replace ultrasoft fields and Wilson lines by those involving soft fields, $q_{us}\to q_s$, $Y_{n_1}\to S_{n_1}$, etc. This matching should be done at the level of the $\widetilde O$ operators in the \mbox{${\rm SCET}_{\rm I}$}\xspace theory, so that we do not have Wilson lines grouped with the color structures in the resulting \mbox{${\rm SCET}_{\rm II}$}\xspace operators. There is a 1-to-1 correspondence between the appropriate building blocks in the two theories. We have the same building blocks for collinear fields, and operators are now built from the soft building blocks for gluons
\begin{align}
\mathcal{B}^a_{s(i)\pm} = -\varepsilon_{\mp\mu}(n_i, {\bf n}_i)\,\mathcal{B}^{a\mu}_{s(i)},\qquad \mathcal{B}^a_{s(i)0} =\bar n_\mu \mathcal{B}^{a \mu}_{s(i)}
\,,
\end{align}
where $\mathcal{B}_{s(i)}^{a\mu}$ is defined as in \eq{softgluondef} but with soft fields. The soft quark building block $\psi_{s(i)}= S_{n_i}^\dagger q_s$ appears in currents that are directly analogs of those containing the ultrasoft quark building block, namely\footnote{The notation here is chosen to make the \mbox{${\rm SCET}_{\rm I}$}\xspace to \mbox{${\rm SCET}_{\rm II}$}\xspace matching simpler, which in some cases comes at the expense of using a different normalization for the soft and collinear currents in the \mbox{${\rm SCET}_{\rm II}$}\xspace theory.}
\begin{align} \label{eq:Js_scetII}
J_{i(s)\pm}^{{\bar \alpha}\beta}
&= \mp \frac{2}{\sqrt{\vphantom{2} \omega_{i}}} \:
\frac{\varepsilon_\mp^\mu(n_i, \bar n_i)}{ \langle {\bf n}_i \mp | n_i \pm \rangle}\:
\bar{\chi}^{\bar \alpha}_{i\pm}\, \gamma_\mu \psi^\beta_{s(i)\pm}
\,, \\
J_{i(\overline{s})\pm}^{{\bar \alpha}\beta}
&=\mp \frac{2}{\sqrt{\vphantom{2} \omega_{i}}}
\frac{\varepsilon_\mp^\mu(\bar{n}_i, n_i)}{ \langle n_i \mp | \bar{n}_i \pm \rangle}\: \bar{\psi}_{s(i) \pm}^{\bar \alpha}\, \gamma_\mu \chi^\beta_{i\pm}
\, , {\nonumber} \\
J_{i(s)0}^{{\bar \alpha}\beta}
&= \sqrt{\frac{2}{\omega_{i}}}\:
\bar \chi^{\bar \alpha}_{i+}\psi^\beta_{s(i)-}
\,, \qquad\qquad\qquad
(J^\dagger)_{i(s)0}^{{\bar \alpha}\beta}
= \sqrt{\frac{2}{\omega_{i}}}\:
\bar \psi^{\bar \alpha}_{s(i)-} \chi^\beta_{i+}
\,, {\nonumber}\\
J_{i(\overline{s})0}^{{\bar \alpha}\beta}
&= \sqrt{\frac{2}{\omega_{i}}}\:
\bar\psi^{\bar \alpha}_{s(i)+} \chi^\beta_{i-}
\,, \qquad\qquad\qquad
(J^\dagger)_{i(\overline{s})0}^{{\bar \alpha}\beta}
= \sqrt{\frac{2}{\omega_{i}}}\:
\bar \chi^{\bar \alpha}_{i-}\psi^\beta_{s(i)+}
\,, {\nonumber}\\
J_{(s)^2 ij\pm}^{{\bar \alpha}\beta}
&= \mp\, \frac{\varepsilon_\mp^\mu(n_i, n_j)}{\langle n_j\mp | n_i\pm\rangle}\:
\bar{\psi}^{\bar \alpha}_{s(i)\pm} \gamma_\mu \psi^\beta_{s(j)\pm}
\,,{\nonumber} \\
J_{(s)^2 ij 0}^{{\bar \alpha}\beta}
&=
\bar \psi^{\bar \alpha}_{s(i)+}\psi^\beta_{s(j)-}
\,,\qquad\qquad\quad
(J^\dagger)_{(s)^2 ij 0}^{{\bar \alpha}\beta}
=
\bar \psi^{\bar \alpha}_{s(i)-}\psi^\beta_{s(j)+}
\,. {\nonumber}
\end{align}
The full set of \mbox{${\rm SCET}_{\rm II}$}\xspace building blocks is listed in \Tab{tab:helicityBB2}. Here the soft derivatives $\partial_{s(i)\pm}$ and $\partial_{s(i)0}$ act only on soft building block fields. From these building blocks we see that the helicity formalism can also be used to greatly simplify the construction of operator bases for \mbox{${\rm SCET}_{\rm II}$}\xspace processes.
\subsubsection{SCET$_+$}\label{sec:scet_plus}
In cases where multiple measurements are made on the same jet additional degrees of freedom must be added to SCET. A general class of effective theories to describe such situations is the class of SCET$_+$ theories \cite{Bauer:2011uc,Procura:2014cba,Larkoski:2015zka,Pietrulewicz:2016nwo}. In addition to collinear and soft modes, these effective field theories typically contain (multiple) collinear-soft modes, which exhibit both a collinear, and a soft scaling. Such effective field theories have been used, for example, for the calculation \cite{Larkoski:2015kga} of the $D_2$ \cite{Larkoski:2014gra} jet substructure observable.
While subleading power corrections to SCET$_+$ theories have not been studied, we wish to emphasize that our helicity operator approach extends also straightforwardly to such theories. In SCET$_+$ theories, subleading power hard scattering operators will involve not only collinear and (ultra)soft fields, but also collinear soft fields. Although we will not do it in this paper, it is then a simple exercise to write a basis of helicity operator building blocks, including also such collinear soft fields. Indeed, the helicity operator formalism has already been used to simplify matching calculations in SCET$_+$ at leading power in \cite{Pietrulewicz:2016nwo}.
\subsubsection{SCET with Massive Collinear Quarks}\label{sec:scet_massive}
The effective field theory description of the dynamics of the ultrasoft and collinear modes, as discussed thus far, is appropriate for massless quarks. For certain cases of phenomenological relevance, including boosted top production, the quark mass is an IR scale with the same parametric scaling as the $\perp$ momenta of collinear particles. In this case, the quark mass must be included in $\mathcal{L}_\mathrm{dyn}$ \cite{Leibovich:2003jd,Rothstein:2003wh} for collinear quarks, soft or ultrasoft quarks, or both. For example, the leading power collinear quark Lagrangian for massive quarks is given by
\begin{align}\label{eq:collinear_massive}
\mathcal{L}^{m(0)}_{n \xi} &= \bar{\xi}_n\left[ i n \cdot D_{ns} +\left( i \slashed{D}_{n \perp}-m \right) W_n \frac{1}{\overline{\mathcal{P}}_n} W_n^\dagger \left( i \slashed{D}_{n \perp} +m \right) \right] \frac{\slashed{\bar{n}}}{2} \xi_n\,.
\end{align}
where $i D_{ns} = in\cdot\partial_{us} + g n\cdot A_{us} + gn\cdot A_n$.
Since the mass appears as an IR scale in the effective theory, the hard scattering operators for the case of massive SCET are the same as for massless SCET and the helicity operator basis presented in this paper also applies. However, as compared to the leading power Lagrangian for massless collinear quarks, the mass terms in \Eq{eq:collinear_massive} imply that quark chirality is not conserved by the soft and collinear dynamics of the effective field theory. This means that symmetry arguments relying on the conservation of helicity no longer apply. (We will use such symmetry arguments to reduce the number of hard scattering operators that can contribute to the $e^+e^- \to $ dijets or constrained Drell-Yan cross-section at $\mathcal{O}(\lambda^2)$ for massless SCET.) Nevertheless, the helicity operator basis still provides a convenient way of organizing hard scattering SCET operators involving boosted massive quarks.
\subsubsection{Evanescent Operators}
One subtlety of the helicity operator basis is that it relies on massless quarks and gluons having two helicities, a feature which is specific to $4$ dimensions. In dimensional regularization, divergences are regularized by analytically continuing the particle momenta to $d=4-2\epsilon$ dimensions. In a general scheme, the helicities of quarks and gluons live in $d^g_s$, and $d^q_s$ dimensional spaces respectively, although in most commonly used schemes, only $d^g_s$ is analytically continued. Different schemes within dimensional regularization differ in their treatment of $d^g_s$ for internal (unobserved) and external (observed) particles. Evanescent operators \cite{Buras:1989xd,Dugan:1990df,Herrlich:1994kh} are defined as those whose tree level matrix elements vanish as $\epsilon\to 0$. However in loop calculations these matrix elements can multiply $1/\epsilon$ poles and lead to contributions that must be included in matching and higher order anomalous dimension calculations. For explicit discussions within the context of SCET calculations, see \Refs{Hill:2004if,Becher:2004kk,Beneke:2005gs,Ali:2007sj}. Such evanescent operators can not be specified using only our helicity building block fields.
In \Ref{Moult:2015aoa} a discussion of scheme dependence was given for leading power helicity operators, and it was shown that evanescent operators do not appear when using SCET helicity operators for leading power matching calculations in exclusive jet processes. However, evanescent operators could appear at loop level when working to subleading power. The required extension of our helicity operator basis depends in detail on the regularization scheme, but in general requires the inclusion of additional fields, for example an $\epsilon$ scalar gluon $\mathcal{B}^a_{\epsilon}$ to encode the $(-2\epsilon)$ transverse degrees of freedom, and quark currents $J_\epsilon$ which involve Dirac structures that would vanish if $\epsilon=0$. Since we do not consider the explicit one-loop matching and evolution of the helicity operators in this paper, we postpone a detailed discussion of evanescent operators to future work. However, we expect that at each loop order, the possible evanescent operators can be easily identified and treated. We note that a calculation of the leading power inclusive jet and soft functions in different regularization schemes, including the treatment of $\epsilon$ scalar gluons was presented in \Ref{Broggio:2015dga}.
\subsection{Parity and Charge Conjugation Properties}\label{sec:PandC}
It may initially seem that having distinct operators for each external helicity configuration greatly increases the number of operators. However, as is known from the study of helicity amplitudes, this is not the case. Parity and charge conjugation relations allow one to relate operators with distinct helicity configurations. An understanding of these relations is therefore essential for minimizing the number of matching calculations. The use of parity and charge conjugation relations is not limited to theories which exhibit C or P symmetry. Amplitudes and Wilson coefficients can be decomposed into pieces each of which have definite properties under C or P.
The C/P properties for the helicity building blocks involving collinear fields are as follows. Under parity, we have
\begin{align} \label{eq:subPproperties}
\P\, \mathcal{B}^a_{i\pm}(n_i; \omega_i; x)\, \P
&= \mathcal{B}^a_{i \mp}(n_i^\P; \omega_i; x^\P)
\,, \\
\P\, J^{{\bar \alpha}\beta}_{i j\pm}(n_i,n_j; \omega_i,\omega_j; x)\, \P
&= J^{{\bar \alpha}\beta}_{i j\mp}(n_i^\P,n_j^\P; \omega_i,\omega_j; x^\P)
{\nonumber} \,, \\
\P J_{i 0}^{{\bar \alpha} \beta} (n_i; \omega_1,\omega_2;x) \P
&= J_{i \bar 0 }^{{\bar \alpha} \beta}(n_i^\P; \omega_1,\omega_2;x^\P)
\,, {\nonumber} \\
\P J_{i \bar 0}^{{\bar \alpha} \beta} (n_i; \omega_1,\omega_2;x) \P
&= J_{i 0 }^{{\bar \alpha} \beta}(n_i^\P; \omega_1,\omega_2;x^\P)
{\nonumber} \,, \\
\P J_{i \pm}^{{\bar \alpha} \beta} (n_i; \omega_1,\omega_2;x) \P
&= J_{i \mp }^{{\bar \alpha} \beta}(n_i^\P; \omega_1,\omega_2;x^\P)
{\nonumber}\,,
\end{align}
where we have made the dependence on $n_i$, $\omega_i$, and $x$ explicit, and the parity-transformed vectors are $x^\P_\mu = x^\mu$.
Under charge conjugation we have,
\begin{align} \label{eq:subCproperties}
\C\, \mathcal{B}^a_{i\pm}(n_i;\omega_i)\, T^a_{\alpha{\bar \beta}}\,\C
&= - \mathcal{B}^a_{i \pm}(n_i;\omega_i) T^a_{\beta{\bar \alpha}}
\,,\\
\C\, J^{{\bar \alpha}\beta}_{i j\pm}(n_i,n_j;\omega_i, \omega_j)\,\C
&= -J^{{\bar \beta}\alpha}_{j i\mp} (n_j,n_i;\omega_j, \omega_i)
\,,{\nonumber} \\
\C J_{i 0}^{{\bar \alpha} \beta}(n_i; \omega_1,\omega_2) \C
&= -J_{i \bar 0 }^{{\bar \beta} \alpha}(n_i; \omega_2,\omega_1)
\,, {\nonumber} \\
\C J_{i \bar 0}^{{\bar \alpha} \beta}(n_i; \omega_1,\omega_2) \C
&= -J_{i 0}^{{\bar \beta} \alpha}(n_i; \omega_2,\omega_1)
{\nonumber} \,, \\
\C J_{i\pm}^{{\bar \alpha} \beta} (n_i; \omega_1,\omega_2) \C
&= -J_{i \pm }^{{\bar \beta} \alpha} (n_i; \omega_2,\omega_1)
{\nonumber} \,.
\end{align}
Under parity, the $\mathcal{P}_{\perp}^{\pm}$ operators transform as
\begin{equation} \label{eq:Pperpparity}
\P \mathcal{P}_{\perp}^{\pm} \P = -\mathcal{P}_{\perp}^{\mp} \, .
\end{equation}
Since charge conjugation exchanges the order of fields within a quark current, we have
\begin{equation} \label{eq:Pperpcharge}
C \{ J_{i 0}^{{\bar \alpha} \beta} (\mathcal{P}_{\perp}^{\pm})^\dagger\} C = -\{(\mathcal{P}_{\perp}^{\pm}) J_{i \, \,\bar 0}^{{\bar \beta} \alpha}\} \,,
\end{equation}
along with similar relations for the other operators that involve $\mathcal{P}^{\pm}_\perp$ insertions.
Although for our main example in \sec{eeJets} we will not use the operators of \Eq{eq:Jus}, which involve ultrasoft quarks, for completeness we give their C/P properties. Under parity, the mixed ultrasoft collinear operators transform as
\begin{align}
\P J_{i(us) \pm}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= J_{i(us) \mp}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, \\
\P J_{i(\overline{us}) \pm}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= J_{i(\overline{us}) \mp}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, {\nonumber} \\
\P J_{i(us) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= (J^\dagger)_{i(us) 0}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, {\nonumber} \\
\P J_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= (J^\dagger)_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, {\nonumber} \\
\P (J^\dagger)_{i(us) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= J_{i(us) 0}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, {\nonumber} \\
\P (J^\dagger)_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\P&= J_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i^\P;\omega_i)
\,, {\nonumber}
\end{align}
while under charge conjugation, we have
\begin{align}
\C J_{i(us) \pm}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&=-J_{i(\overline{us}) \mp}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,, \\
\C J_{i(\overline{us}) \pm}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&=-J_{i(us) \mp}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,, {\nonumber} \\
\C J_{i(us) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&= -J_{i(\overline{us}) 0}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,, {\nonumber} \\
\C J_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&= -J_{i(us) 0}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,, {\nonumber} \\
\C (J^\dagger)_{i(us) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&= -(J^\dagger)_{i(\overline{us}) 0}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,, {\nonumber} \\
\C (J^\dagger)_{i(\overline{us}) 0}^{{\bar \alpha} \beta}(n_i;\omega_i)\C&= -(J^\dagger)_{i(us) 0}^{{\bar \beta} \alpha}(n_i;\omega_i)
\,. {\nonumber}
\end{align}
The C/P properties of the currents involving two ultrasoft quarks are identical to those of standard quark bilinears, so we do not give them here. Finally, the C/P properties of the SCET$_\text{II}$ operators of \Eq{eq:Js_scetII}, which involve soft quarks, are easily obtained from the SCET$_\text{I}$ results above.
As a simple example to demonstrate the use of C/P relations, we consider $e^+e^- \to q\bar q$ through an off-shell photon at leading power in SCET, which we will consider in more depth in \sec{eeJets}. We will label the quark and antiquark by $1,2$ and the electron and positron by $3,4$. It is well known that at leading power there is a single current using traditional SCET operators,
\begin{align}\label{eq:trad_operator}
J^{\mu {\bar \alpha} \beta }= \bar{\chi}^{\bar \alpha}_{1} \gamma^\mu \chi^\beta_{2}.
\end{align}
The free Lorentz index is contracted with the leptonic tensor to form a scalar. On the other hand, the helicity basis consists of four scalar operators,
\begin{alignat}{2}\label{eq:ee_helop}
O^{(0){\bar \alpha} \beta}_{(+;+)}
&= J^{\bar \alpha \beta}_{12+}\,J_{e+}
\,,& \qquad
O^{(0){\bar \alpha} \beta}_{(+;-)}
&= J^{\bar \alpha \beta}_{12+ }\,J_{e-}
\,, \\
O^{(0){\bar \alpha} \beta}_{(-;+)}
&= J^{\bar \alpha \beta}_{12- }\,J_{e+ }
\,, &\qquad
O^{(0){\bar \alpha} \beta}_{(-;-)}
&= J^{\bar \alpha \beta}_{12- }\,J_{e- }
\,,{\nonumber}
\end{alignat}
which already include the leptons through the lepton helicity current $J_{e\pm}$. The leptonic helicity currents are defined in an identical manner to the leading power QCD current of \Eq{eq:jpm_def}, but without the corresponding Wilson lines or color indices
\begin{equation}
J_{e\pm}\equiv J_{34\pm}
= \mp\, \sqrt{\frac{2}{\omega_3\, \omega_4}}\, \frac{ \varepsilon_\mp^\mu(n_3, n_4) }{\langle n_4\mp | n_3\pm\rangle} \, \bar{e}_{3\pm}\, \gamma_\mu e_{4\pm}
\,.\end{equation}
For the $e^+e^- \to q\bar q$ process, there is a unique color structure for either \eq{trad_operator} or \eq{ee_helop},
\begin{align}
\bar{T}^{ \alpha \bar \beta}=(\delta_{\alpha \bar \beta})\,.
\end{align}
Invariance under parity implies that the Wilson coefficients for the helicity operators are related by
\begin{align} \label{eq:Cdijetparity}
C_{(+;+)}(n_1,n_2;\omega_1,\omega_2;\omega_3,\omega_4)
&=C_{(-;-)}(n_1^\P,n_2^\P;\omega_1,\omega_2;\omega_3,\omega_4)
\,, \\
C_{(+;-)}(n_1,n_2;\omega_1,\omega_2;\omega_3,\omega_4)
&=C_{(-;+)}(n_1^\P,n_2^\P;\omega_1,\omega_2;\omega_3,\omega_4)
\,. {\nonumber}
\end{align}
When doing the matching, we sum over the $n$'s as in \eq{Leff_sub_explicit}, so we are free to rewrite $n_i^\P \rightarrow n_i$. Since we consider the process to all orders in the strong interaction, but only leading order in the electromagnetic interaction, the leptons couple through the current $\langle 1\pm |\gamma^\mu | 2\pm \rangle=\langle 2\mp |\gamma^\mu | 1\mp \rangle$. This implies the further relation
\begin{align} \label{eq:CdijetC}
C_{(+;+)}(n_1,n_2;\omega_1,\omega_2;\omega_3,\omega_4)
&=C_{(+;-)}(n_1,n_2;\omega_1,\omega_2;\omega_4,\omega_3) \,.
\end{align}
These relations can be easily checked by considering the tree level matching. At tree level, the Wilson coefficients are given by
\begin{align}\label{eq:ee_wilson}
C_{(+;+)}&=- e^2 q^2
\frac{2[13]\ang{24}}{s_{34}} \,,~~
C_{(+;-)}=-e^2 q^2
\frac{2[14]\ang{23}}{s_{34}} \,,~~ \\
C_{(-;+)}&=-e^2 q^2
\frac{2[23]\ang{14}}{s_{34}} \,,~~
C_{(-;-)}=-e^2 q^2
\frac{2[24]\ang{13}}{s_{34}} \,. {\nonumber}
\end{align}
These satisfy the above relations, by noting from \app{helicity} that parity simply interchanges $[]\leftrightarrow \langle \rangle$. Together the three relations in \eqs{Cdijetparity}{CdijetC} provide the necessary information to indicate that the matching onto the helicity operator basis came from a vector current. Therefore, only one coefficient of the helicity operators needs to be computed in a matching calculation at any order in $\alpha_s$. Note that the basis constructed in \eq{ee_helop} also works for mediation through a Z-boson, where axial coupling also needs to be considered.
Further examples of the use of C and P to simplify helicity operator bases can be found in Ref.~\cite{Moult:2015aoa}, and below in \sec{eeJets}.
\subsection{Constraints from Angular Momentum Conservation}\label{sec:ang_cons}
The use of operators with definite helicities makes manifest symmetries related to rotational invariance. As discussed in detail in \cite{Kolodrubetz:2016uim}, constraints from conservation of angular momentum can greatly reduce the basis of hard scattering operators appearing at subleading powers, when multiple collinear fields can appear in each collinear sector. Conservation of angular momentum implies the general constraint \cite{Kolodrubetz:2016uim}
\begin{align} \label{eq:Jmin}
J^{(i)}_\text{min}\leq \sum_{j \text{ with } \hat n_j\neq \hat n_i} J_{\rm min}^{(j)} \,,
\end{align}
where $J^{(i)}_\text{min}$ is the minimum angular momentum carried by the $n_i$-collinear sector. This can be related to the helicities of the building blocks in a given sector by $J^{(i)}_\text{min}=|h_{n_i}^{\rm tot}|$, where the helicities in the $n_i$-collinear sector of some operator add up to $h_{n_i}^{\rm tot}$. From this, we immediately get the constraint,
\begin{align} \label{eq:hmin}
|h_{n_i}^\text{tot}| \leq \sum_{j \text{ with } \hat n_j\neq \hat n_i} | h_{n_j}^\text{tot}| \,,
\end{align}
where it is important to count back-to-back collinear directions only once in this sum, considering the helicity about their common axis.
These selection rules are particularly simple for the case of $e^+e^- \to$ dijets (or constrained Drell-Yan) which we study here. In this case, there are two axes along which particles move, namely the axis of the colliding $e^+e^-$ pairs, and the $n$ axis of the jets. In the helicity operator approach, the helicities of all operators are defined with respect to these axes. For the case of $e^+e^-\to$ dijets proceeding through an off-shell photon or $Z$ boson, the coupling to vector bosons guarantees that the electron pair has a combined helicity along the collision axis of $|h_{e^+e^-}|=1$, as shown in \Fig{fig:helicity_constraint}. To have a non-zero amplitude, the helicity state of the outgoing jets defined along the $n$ axis must have an overlap with this helicity $1$ state. In particular, we must have $h_{n}^\text{tot} = -1,0,1$. At subleading power, when there are multiple collinear fields in the $n$ and $\bar n$ sectors, this means that the helicities of the fields must be arranged in particular combinations, and considerably simplifies the basis.
As an example, consider a subleading power operator involving an additional collinear gluon field in the $n$ collinear sector. Without imposing constraints from angular momentum conservation, a basis of allowed helicity operators is
\begin{alignat}{2} \label{eq:Z1_basis_cons}
&O_{+(+;\pm)}^{(1)a\,{\bar \alpha}\beta}
=\mathcal{B}_{n+}^a \, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{+(-;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, J_{n\bar n\,-}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, \\
&O_{-(+;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{-(-;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\,J_{e\pm }
\,. {\nonumber}
\end{alignat}
However, the first and fourth operators have $|h_{n}^\text{tot}| = 2$ and thus are not allowed, and can be eliminated from the basis. This configuration is shown schematically in \Fig{fig:helicity_constraint} a). Only the operators where the helicity of the $n$ collinear quark and gluon fields are opposite are allowed in the basis. The actual basis of allowed operators is therefore simpler, and is just given by
\begin{alignat}{2} \label{eq:Z1_basis_cons_reduced}
&O_{-(+;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{+(-;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, J_{n\bar n\,-}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,
\end{alignat}
A schematic illustration of these configurations is shown in \Fig{fig:helicity_constraint} b). Using these restrictions, we have therefore eliminated half of the potential operators. Similar constraints will play an important role in simplifying the complete basis of subleading power operators given in \Sec{sec:eeJets}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.35\columnwidth]{figures/helicity_constraint_b}
\hspace{1.4cm}
\includegraphics[width=0.35\columnwidth]{figures/helicity_constraint_c}
\raisebox{0cm}{ \hspace{-3.0cm}
$a$)\hspace{6.6cm}
$b$)\hspace{2.4cm} }
\\[-25pt]
\end{center}
\vspace{-0.4cm}
\caption{
A illustration of the helicity selection rule for $e^+e^-\to $ dijets. In a) the collinear particles along the $n$ axis carry $|h=2|$, and have a vanishing projection onto the $J_{e\pm}$ current. In b), the collinear particles carry $|h=0|$ and therefore have a non-vanishing projection onto the $J_{e\pm}$ current. }
\label{fig:helicity_constraint}
\end{figure}
\section{Hard Scattering Operators with Two Collinear Directions}\label{sec:eeJets}
To demonstrate the simplicity of the helicity-operator approach, we construct the ${\cal O}(\lambda)$ and ${\cal O}(\lambda^2)$ basis of power suppressed hard scattering operators with two collinear directions. A summary of the complete set of operators is given in \Tab{tab:summary}. For concreteness, we take the color singlet particles involved in the hard scattering to be an $e^+e^-$ pair, with the interaction proceeding through an off-shell $\gamma$ or $Z$. The basis is valid to all orders in $\alpha_s$, and leading order in the electroweak couplings. Since the helicity operators from which our basis of hard scattering operators are composed are manifestly crossing symmetric, our operators can be used as a basis of hard scattering operators in factorization proofs for $e^+e^- \to$ dijet event shapes, constrained Drell-Yan, or DIS producing a single jet.
The extension to $\mathcal{O}(\lambda^2)$ is necessary, since when we square power suppressed jet amplitudes to calculate a cross section, the $\mathcal{O}(\lambda)$ power corrections to the cross section tend to vanish \cite{Beneke:2003pa,Lee:2004ja,Freedman:2013vya}. We will explicitly show the vanishing of the contributions to the $\mathcal{O}(\lambda)$ cross section from hard scattering contributions in \Sec{sec:ee_lambda}. We then present the hard scattering operators that contribute to the factorized cross section at $\mathcal{O}(\lambda^2)$ in \Sec{sec:ee_lambda2}. A set of ${\cal O}(\lambda)$ and ${\cal O}(\lambda^2)$ operators for $e^+e^-\to$ dijets has been presented in \Ref{Freedman:2013vya}, although no claim of completeness was made, and a different formulation of SCET was used. In \Sec{sec:compare_earlier} we will briefly compare our all orders basis with the operators of \Ref{Freedman:2013vya}.
{
\renewcommand{\arraystretch}{1.4}
\begin{table}[t!]
\hspace{-0.4cm}
\scalebox{0.842}{
\begin{tabular}{| c | l | l | c |c| c |}
\hline
$\!$Order$\!$ & $\!$Category$\!$ & Operators (equation number)
& \#$\!$ helicity & \#$\!$ of
& $\!\sigma_{2j}^{\mathcal{O}(\lambda^2)}\!\! \ne\! 0\!$
\\[-8pt]
& & & configs & $\!$color$\!$ &
\\ \hline
$\!\mathcal{O}(\lambda^0)\!$
& $\! e \bar e q \bar q$ & $O_{(\lambda_1;\pm)}^{(0){\bar \alpha}\beta}
=J^{{\bar \alpha}\beta}_{n \bar n \lambda_1}J_{e \pm }$ \,(\ref{eq:LP_basis})
& 4 & 1 & $\checkmark$ \\
\hline
$\!\mathcal{O}(\lambda)\!$
& $\! e \bar{e} q \bar{q} g$
& $O_{n\bar n1,2 \lambda_1 ( \lambda_2;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n,\bar n \lambda_1}^a\, J_{n\bar n\, -\lambda_1}^{{\bar \alpha}\beta}\,J_{e\pm }$ \,(\ref{eq:Z1_basis},\ref{eq:Z1_basis_flip})
&8 & 1 & $\checkmark$
\\
& & $O_{\bar{n}\lambda_1 ( \lambda_2:\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n \lambda_1}^a\, J_{\bar n \lambda_2\, }^{{\bar \alpha}\beta}\, J_{e\pm }$ \,(\ref{eq:Z1_basis_diff})
&8 & 1 & $\checkmark$
\\ \cline{2-6}
& $\! e \bar{e} ggg$ & $O_{\mathcal{B}\lambda_1 \lambda_2 \lambda_3 (\pm)}^{(1)abc}
= S\, \, \mathcal{B}_{n \lambda_1}^a\, \mathcal{B}_{\bar n \lambda_2}^b\, \mathcal{B}_{\bar n \lambda_3}^c J_{e\pm }$ \,(\ref{eq:eeggg})
&8 & 2 & $\checkmark$
\\ \hline
$\!\mathcal{O}(\lambda^2)\!$
& $\! e \bar{e} q \bar q Q \bar Q$ &
$O_{qQ1(\lambda_1;\lambda_2:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)n {\lambda_1}\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z2_basis_qQ})
& 8 & 2 &
\\
& & $O_{qQ2(\lambda_1;\lambda_1:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q \bar{Q}) n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(Q \bar{q}) \bar{n} \, \lambda_1\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z2_basis_qQ_2})
&4 & 2 &
\\
& & $O_{qQ3(\lambda_1;-\lambda_1;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q) n \bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n -\lambda_1\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z2_basis_qQ_3})
&4 & 2 & \\
& & $O_{qQ4(\lambda_1:\lambda_2;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \lambda_1}^{{\bar \alpha}\beta}\, J_{(Q) n \bar n\, {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z3_basis_qQ})
&8 & 2 & $\checkmark$
\\
& & $O_{qQ5(\lambda_1:\lambda_2;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(Q)\bar n n {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z3_basis_qQ})
&8 & 2 & $\checkmark$
\\ \cline{2-6}
& $\! e \bar{e} q \bar q q \bar q$ &
$O_{qq1(\lambda_1;\lambda_2:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)n {\lambda_1}\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z2_basis_qq})
& 8 & 2 &
\\
& & $O_{qq3(\lambda_1;-\lambda_1;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q) n \bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(q) n\bar n -\lambda_1\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z2_basis_qq_3})
&2 & 2 & \\
& & $O_{qq4(\lambda_1:\lambda_2;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \lambda_1}^{{\bar \alpha}\beta}\, J_{(q) n \bar n\, {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z3_basis_qq})
&8 & 2 & $\checkmark$
\\
& & $O_{qq5(\lambda_1:\lambda_2;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \lambda_1\, }^{{\bar \alpha}\beta}\, J_{(q)\bar n n {\lambda_2}\, }^{{\bar \gamma}\delta}\,J_{e\pm } $ \,(\ref{eq:Z3_basis_qq})
&8 & 2 & $\checkmark$
\\ \cline{2-6}
& $\! e \bar{e} q \bar q g g$
& $O_{\cB1\lambda_1 \lambda_2(\lambda_3;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= S \mathcal{B}_{n\lambda_1}^a \mathcal{B}_{n \lambda_2}^b \, J_{n\bar n\, \lambda_3 }^{{\bar \alpha}\beta} \,J_{e\pm }$ \,(\ref{eq:eeqqgg_basis1})
&8 &3& $\checkmark$
\\
& & $O_{\cB2\lambda_1 \lambda_2(\lambda_3;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= S \mathcal{B}_{n\lambda_1}^a \mathcal{B}_{n \lambda_2}^b \, J_{\bar n n\, \lambda_3 }^{{\bar \alpha}\beta} \,J_{e\pm }$ \,(\ref{eq:eeqqgg_basis1a})
&8 &3& $\checkmark$
\\
& & $O_{\cB3\lambda_1 \lambda_2(\lambda_3;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n\lambda_1}^a \mathcal{B}_{\bar n \lambda_2}^b \, J_{n\bar n\, \lambda_3 }^{{\bar \alpha}\beta} \,J_{e\pm } $ \,(\ref{eq:eeqqgg_basis2})
&12 &3& $\checkmark$
\\
& & $O_{\cB4\lambda_1 \lambda_2(\lambda_3:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n \lambda_1}^a \mathcal{B}_{\bar n \lambda_2}^b \, J_{n\,{\lambda_3} }^{{\bar \alpha}\beta} \,J_{e\pm } $ \,(\ref{eq:eeqqgg_basis3})
& 8 &3&
\\
& & $O_{\cB5\lambda_1 \lambda_2(\lambda_3:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n \lambda_1}^a \mathcal{B}_{\bar n \lambda_2}^b \, J_{n\,{\lambda_3} }^{{\bar \alpha}\beta} \,J_{e\pm } $ \,(\ref{eq:eeqqgg_basis4})
& 4 &3&
\\ \cline{2-6}
& $\! e \bar{e} gggg$ & $O_{4g1\lambda_1 \lambda_2 \lambda_3 \lambda_4(\pm)}^{(2)a b c d}
= S \mathcal{B}^a_{n \lambda_1} \mathcal{B}^b_{n \lambda_2} \mathcal{B}^c_{{\bf n} \lambda_3} \mathcal{B}^d_{{\bf n} \lambda_4} J_{e\pm } $ \,(\ref{eq:Z2_basis_gggg_1})
& 6 &9&
\\
& & $O_{4g2\lambda_1 \lambda_2 \lambda_3 \lambda_4(\pm)}^{(2)a b c d}
= S \mathcal{B}^a_{n \lambda_1} \mathcal{B}^b_{{\bf n} \lambda_2} \mathcal{B}^c_{{\bf n} \lambda_3} \mathcal{B}^d_{{\bf n} \lambda_4} J_{e\pm } $ \,(\ref{eq:Z2_basis_gggg_2})
& 4 &9&
\\ \cline{2-6}
& $\! \mathcal{P}_\perp$ & $O_{\cP2 \lambda_1 (\lambda_2:\pm)[\lambda_{\mathcal{P}}]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n\lambda_1}^a \, \{J_{\bar n\, {\lambda_2} }^{{\bar \alpha}\beta}(\mathcal{P}_{\perp}^{\lambda_{\mathcal{P}}})^\dagger\}\,J_{e\pm }$ \,(\ref{eq:eeqqgpperp_basis_same})
& 8 & 1 &
\\
& & $O_{\cP1n,\bar n \lambda_1 (\lambda_2;\pm)[\lambda_{\mathcal{P}}]}^{(2)a\,{\bar \alpha}\beta}
=\left [ \mathcal{P}_{\perp}^{\lambda_{\mathcal{P}}} \mathcal{B}_{n,\bar n \lambda_1}^a \right ] \, J_{n\bar n\, \lambda_2}^{{\bar \alpha}\beta}\,J_{e\pm } $ \,(\ref{eq:eeqqgpperp_basis},\ref{eq:eeqqgpperp_basis_flip})
&24 & 1 & $\checkmark$
\\
& & $O_{\mathcal{P}\mathcal{B} \lambda_1 \lambda_2 \lambda_3(\pm)[\lambda_{\mathcal{P}}]}^{(2)abc}
= S\, \mathcal{B}_{n\lambda_1}^a\, \mathcal{B}_{\bar n \lambda_2}^b\, \left [\mathcal{P}_{\perp}^{\lambda_{\mathcal{P}}} \mathcal{B}_{\bar n \lambda_3}^c \right ]\, J_{e\pm } $ \,(\ref{eq:eegggpperp_basis})
&8 & 2 &
\\ \cline{2-6}
& $\!$Ultrasoft$\!\!\!$
& $O_{\mathcal{B}(us(i))\lambda_1:(\lambda_2;\pm) }^{(2)a\,{\bar \alpha}\beta}
=\mathcal{B}_{us(i) \lambda_1}^a \, J_{n\bar n\, \lambda_2}^{{\bar \alpha}\beta}\,J_{e\pm } $ \,(\ref{eq:soft_insert_basis},\ref{eq:soft_insert_basis2})
&8 &1&
\\
& & $O_{\mathcal{B}(us(i))0:(\lambda_1;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(i)0}^a \, J_{n\bar n\,\lambda_1}^{{\bar \alpha}\beta}\,J_{e\pm } $ \,(\ref{eq:soft_insert_basis},\ref{eq:soft_insert_basis2})
& 8 &1& $\checkmark$
\\
& & $O_{\partial(us(i))\lambda_1:(\lambda_2;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(i)\lambda_1} \, J_{n\bar n\,\lambda_2}^{{\bar \alpha}\beta}\}\,J_{e\pm } $ \,(\ref{eq:soft_derivative_basis})
& 8 &1&
\\
& & $O_{\partial(us(i))0,\bar 0:(\lambda_1;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(i)0,\bar 0} \, J_{n\bar n\,\lambda_1}^{{\bar \alpha}\beta}\}\,J_{e\pm } $ \,(\ref{eq:soft_derivative_basis})
& 8 &1& $\checkmark$
\\
& & $O_{(us(i))\lambda_1:\lambda_2 \lambda_3(\pm) }^{(2)abc}
=\mathcal{B}_{us(i) \lambda_1}^a \,\mathcal{B}_{n\, \lambda_2}^{b}\,\mathcal{B}_{{\bf n}\, \lambda_3}^{c}\,J_{e\pm } $ \,(\ref{eq:eegggus},\ref{eq:eegggus2})
& 24 &2&
\\
& & $O_{\partial \mathcal{B} (us(i))\lambda_1:\lambda_2 \lambda_3(\pm) }^{(2)ab}
=\left[ \partial_{us(i) \lambda_1} \mathcal{B}_{n\, \lambda_2}\right] \mathcal{B}_{{\bf n}\, \lambda_3} J_{e\pm } $ (\ref{eq:eedggus},\ref{eq:eedggus2})\!
& 24 &2&
\\
\hline
\end{tabular}}
\vspace{-0.1cm}
\caption{Basis of hard scattering operators to ${\cal O}(\lambda^2)$ with an electron current $J_{e\pm}$ and two back-to-back collinear sectors. Here the $\lambda_i$ denote helicities, $S$ represents a symmetry factor, and $\mathcal{B}_{n,\bar n \lambda_1}^a$ indicates $\mathcal{B}_{n \lambda_1}^a$ or $\mathcal{B}_{\bar n \lambda_1}^a$. The allowed values for the $\lambda_i$ helicities are given in the indicated equation, and the total count is given in the indicated column. The last column indicates which operators can contribute to $e^+ e^- \to$ dijet event shapes and other two-direction processes up to $\mathcal{O}(\lambda^2)$ in the power expansion and at any order in $\alpha_s$, as discussed in detail in Sec. \ref{sec:ee_lambda2}. }
\label{tab:summary}
\vspace{-2.5cm}
\end{table}
}
One simplification that we make when constructing our basis is that we work in the center of mass frame, and only consider operators that are non-vanishing in this frame. This is natural for Drell-Yan (CM frame), $e^+e^- \to$ dijet event shapes (the CM frame of the jets), and is also a convenient frame for theoretical studies of DIS (the Breit frame). Because of the conservation of label momentum in SCET, this choice of frame allows us to take the strongly interacting collinear sectors to be back-to-back. We therefore describe them by the back-to-back light-like vectors $n_1=n=(1,\hat n)$ and $n_2={\bf n}=(1,-\hat n)$. Due to this choice we will label the helicity currents with $n$ and $\bar n$, as in $J_{n{\bf n}\pm}^{{\bar \alpha}\beta}$, instead of with collinear sector numbers, $J_{12\pm}^{{\bar \alpha}\beta}$. In \eq{Leff_sub_explicit} the hard Lagrangian in SCET is written as a sum over label momenta of the hard operators. For the special case of two back-to-back collinear sectors this reduces to
\begin{align}\label{eq:sum_dir}
\mathcal{L}_{\text{hard}}^{(j)} = \sum_{n} \sum_{A,\cdot\cdot}
\bigg[ \prod_{i=1}^{\ell_A} \int \! \! \mathrm{d} \omega_i \bigg] \,
& \vec O^{(j)\dagger}_{A\,\lotsdots}\big(n{\bf n};
\omega_1,\ldots,\omega_{\ell_A}\big) {\nonumber}\\
& \times
\vec{C}^{(j)}_{A\,\lotsdots}\big(n{\bf n};\omega_1,\ldots,\omega_{\ell_A} \big)
\,.
\end{align}
When constructing a complete basis, we therefore do not need to include operators which are identical up to the swap of $n\leftrightarrow \bar n$. For a given operator, we can therefore choose the $n$ and $\bar n$ labels arbitrarily, and this can be done independently for each operator. When squaring matrix elements, all possible interferences are properly incorporated by accounting for the sum over directions in \Eq{eq:sum_dir}.
Furthermore, we choose to align the $n$ and ${\bar n}$ axes with the jets or protons, such that the overall label $\perp$ momentum of each collinear sector is zero. As a consequence, operators involving $\mathcal{P}_\perp$ acting on the complete set of fields in a collinear sector vanish. Insertions of $\mathcal{P}_\perp$ can still be non-vanishing when two or more collinear field building blocks appear in the same sector, and hence will first appear in our analysis at ${\cal O}(\lambda^2)$. The presence of ultrasoft degrees of freedom carrying residual momentum implies that, unlike the label perp momentum, the residual perp momentum of a collinear sector does not necessarily vanish, and can be exchanged with the ultrasoft sector.
All outgoing quark fields are taken to be massless. Chiral symmetry is violated in QCD by heavy quark masses, such as the top quark mass, and by non-perturbative effects. When matching QCD to SCET with only massless external fields we assume we are working below the scale where a top-quark can be produced, so top quarks appear only in closed loops and chirality is conserved at each order in $\alpha_s$ by the matching procedure (though not by the low energy non-perturbative dynamics, such as the chiral condensate). This remains true when considering QCD corrections to the Z exchange for any of the processes governed by the back-to-back collinear operators. All operators appearing in our basis must therefore preserve chirality.
Throughout this section we will use $P_Z$ to denote the ratio of the $Z$ and photon propagators,
\begin{equation}
P_{Z}(s) = \frac{s}{s-M_{Z}^2 + i \Gamma_{Z} M_{Z}}
\,,\end{equation}
and we will use $v^i_{L,R}$ for the coupling of particle $i$ to the Z boson, whose explicit expressions are given by
\begin{align}
v_L^i = \frac{2 T_3^i - 2Q^i \sin^2 \theta_W}{\sin(2\theta_W)}
\,, \quad
v_R^i = - \frac{2Q^i \sin^2 \theta_W}{\sin(2\theta_W)}
\,,\end{align}
where $T_3^i$ is the third component of weak isospin. Since we use helicities to label the operators and Wilson coefficients, it is convenient to define the weak couplings in terms of helicities for both the quark and lepton currents,
\begin{equation}
v_+^l=v_R^l, \qquad v_-^l=v_L^l, \qquad v_+^q=v_R^q,\qquad v_-^q=v_L^q\,.
\end{equation}
For color, we use the normalization ${\rm tr}[T^a\,T^b] = 1/2\,\delta^{ab}$, i.e. $T_F=1/2$, and write the antisymmetric and symmetric structure constants of $SU(3)$, as $f^{abc}, \,d^{abc}$ respectively. In the case of the collinear operators, we present the color structure both before and after BPS field redefinition.
When labeling particles, the highest two subscripts will be used to refer to the electron and positron respectively, which will always appear in the current
\begin{equation}
J_{e \pm } \equiv J_{e ij\pm }
= \mp \sqrt{\frac{2}{\omega_i\, \omega_j}}\, \varepsilon_\mp^\mu(n_i, n_j)\, \frac{\bar{e}_{i\pm} \gamma_\mu e_{j \pm}} {\langle n_j\mp | n_i\pm\rangle}
\,.
\end{equation}
Since this current appears in every operator, for notational convenience we will drop the explicit $ij$ label on the current, denoting it simply by $J_{e\pm}$.
Due to the relatively large number of operators present up to $\mathcal{O}(\lambda^2)$ we provide a summary of the complete set of operators in \Tab{tab:summary}, along with the number of helicity configurations for each operator. There are a total of $256$ helicity configurations in our operator basis, $128$ of which can contribute at $\mathcal{O}(\lambda^2)$, of which $112$ have tree level contributions. This number does not include the different color configurations, which are also indicated in the table. The leading power basis can be found in \Sec{sec:LP}, the $\mathcal{O}(\lambda)$ subleading basis in \Sec{sec:sub} and the $\mathcal{O}(\lambda^2)$ subleading basis in \Sec{sec:subsub}.
\subsection{Leading Power Operators}\label{sec:LP}
To begin this analysis, we review the leading power back-to-back operators and Wilson coefficients, which are combined according to \eq{Leff_sub_explicit} to give ${\cal L}_{\rm hard}^{(0)}$. By power counting, the leading power operators consist of either two collinear quark building blocks or two collinear gluon building blocks, one in the $n$-collinear and one in the ${\bf n}$-collinear sector. For the $eegg$ channel we have a process with one offshell spin-1 ($\gamma/Z$) particle and two onshell spin-1 particles ($gg$). The Wilson coefficients then all vanish by Yang's theorem~\cite{Landau:1948kw,PhysRev.77.242}, so we omit these operators. Therefore only the $ee q \bar q$ channel contributes at leading power. While one should sum over the flavor of the outgoing quarks, this is trivial to implement, and therefore we do not make the flavor in the quark current explicit. The leading power helicity operators are given by
\begin{align}
& \boldsymbol{q \bar{q}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_qq}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{flalign}\label{eq:LP_basis}
O_{(+;\pm)}^{(0){\bar \alpha}\beta}
=J^{{\bar \alpha}\beta}_{n \bar n+}J_{e \pm }\,, \qquad
O_{(-;\pm)}^{(0){\bar \alpha}\beta}
=J^{{\bar \alpha}\beta}_{n \bar n-} J_{e\pm } \,.
\end{flalign}
Here, and throughout this section, the bracketed superscript indicates the suppression in powers of $\lambda$ of the operator relative to the leading power operators. Since these are the leading power operators, it is zero in this case.
The color basis is one dimensional, and is given before and after BPS field redefinition by
\begin{align} \label{eq:leading_color}
\bar{T}^{\alpha{\bar \beta}} = (\delta_{\alpha{\bar \beta}})\,, \qquad
\bar{T}_{\mathrm{BPS}}^{\alpha{\bar \beta}} = \big[Y_n^\dagger Y_{\bar n} \big]_{\alpha{\bar \beta}}
\,.
\end{align}
Since the weak interactions break $C$ and $P$ symmetry, it is convenient to expand the Wilson coefficients into components with well defined $C/P$ properties. We use the decomposition
\begin{align} \label{eq:Z0_expand_Wilson}
\vec{C}^{(0)}_{(\lambda_q;\lambda_l)}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4)
= e^2 \,
\bigg\{ &\big[Q^\ell Q^q + v_{\lambda_l}^\ell v_{\lambda_q}^q P_Z(s_{34})\big] \vec{C}^{(0)}_{q(\lambda_q;\lambda_l)}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4)
{\nonumber}\\ &
\hspace{-2cm}+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{34}) \bigg] \vec{C}^{(0)}_{v(\lambda_q;\lambda_l)}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) {\nonumber} \\
&\hspace{-2cm}+ \frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{34}) \vec{C}^{(0)}_{a(\lambda_q;\lambda_l)}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) \bigg\}
\,,
\end{align}
where $\lambda_q$ and $\lambda_l$ are the quark and lepton current helicities. Our notation follows that of \cite{Moult:2015aoa}. Here we have extracted only electroweak couplings from the Wilson coefficients, so that each of $\vec{C}_{q,v,a}$ is a power series in $\alpha_s$. In \eq{Z0_expand_Wilson} we have split the amplitude into a contribution, $\vec{C}_q$, arising from the matching contributions where the vector boson couples directly to the final-state quark line, and contributions $\vec{C}_v$, $\vec{C}_a$ where the vector boson couples to a quark loop through a vector or axial coupling, respectively. This decomposition is valid since we work only to leading order in the electroweak couplings. We have also made the assumption that all quarks, except for the top, are massless. This implies that only the $b,t$ isodoublet contributes to $\mathcal{A}_a$. This assumption can trivially be lifted, but many helicity amplitudes are calculated assuming this approximation.
Charge and parity conjugation can be used to derive relations between the Wilson coefficients. Parity relates the Wilson coefficients by
\begin{align} \label{eq:leadingCP}
C^{(0)\alpha {\bar \beta}}_{q\,(\lambda_{12};\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= C^{(0)\alpha {\bar \beta}}_{q\,(-\lambda_{12};-\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4)
\,,\\
C^{(0)\alpha {\bar \beta}}_{v\,(\lambda_{12};\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= C^{(0)\alpha {\bar \beta}}_{v\,(-\lambda_{12};-\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4)
\,,{\nonumber}\\
C^{(0)\alpha {\bar \beta}}_{a\,(\lambda_{12};\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= - C^{(0)\alpha {\bar \beta}}_{a\,(-\lambda_{12};-\lambda_{34})}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4)
\,. {\nonumber}
\end{align}
Here $\lambda_{12}=\pm$ denotes the helicity label of the helicity building block with momentum labels $\omega_1$ and $\omega_2$ describing the two collinear quark fields, and similarly for $\lambda_{34}$. This notation will be used throughout this section, with the additional allowance for $\lambda_i = 0$ or $\bar{0}$ when appropriate. For later applications, we also introduce the notation
\begin{align}
\text{if}\,\, \lambda = 0 \,\, \text{then}\, -\lambda = \bar{0}\,.
\end{align}
Note that $n$ and ${\bf n}$ are not swapped here since after applying parity we always make an additional swap $n\leftrightarrow {\bf n}$ so that we get back the same form of operators. Since we always work to leading order in the weak and electromagnetic couplings, the leptons couple only through the vector and axial-vector currents which satisfy
\begin{align}\label{eq:lo_weak}
\mae{3\pm}{\gamma^\mu}{4\pm} = \mae{4\mp}{\gamma^\mu}{3\mp}, \qquad
\mae{3\pm}{\gamma^\mu\gamma_5}{4\pm} = - \mae{4\mp}{\gamma^\mu\gamma_5}{3\mp} \,.
\end{align}
These relations will be used throughout our analysis. Here they imply
\begin{align} \label{eq:lep_reduce}
C_{q\,(\lambda_{12};\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= C_{q\,(\lambda_{12};-\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_4, \omega_3)
\,,\\
C_{v\,(\lambda_{12};\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= C_{v\,(\lambda_{12};-\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_4, \omega_3)
\,,{\nonumber}\\
C_{a\,(\lambda_{12};\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_3, \omega_4) &= - C_{a\,(\lambda_{12};-\lambda_{34})}^{(0)\alpha{\bar \beta}}(n,{\bf n}; \omega_1, \omega_2; \omega_4, \omega_3)
\,.{\nonumber}
\end{align}
The relations in \eqs{lo_weak}{lep_reduce} imply that only the three Wilson coefficients with the helicity label $(+;+)$ need to be calculated to get all twelve coefficients.
\subsection{Subleading Power Operators} \label{sec:sub}
From the power counting of the operators in \Tab{tab:PC}, we see that the $\mathcal{O}(\lambda)$ suppressed operators have three ${\cal O}(\lambda)$ collinear building block fields, or two collinear building block fields and a single $\mathcal{P}_\perp$ insertion. Our choice for $n$ and ${\bf n}$ eliminates operators that have a $\mathcal{P}_\perp$ acting on a complete collinear sector. Therefore, we only need to consider operators consisting of three collinear field building blocks at $\mathcal{O}(\lambda)$. There are two possibilities for the field content of the operators: two collinear quarks and a collinear gluon, or three collinear gluons. We shall discuss each in turn.
The helicity operators involving two collinear quarks and a single collinear gluon consist of a single leptonic current, a quark current, and a collinear gluon building block field. For each helicity configuration, we must consider the cases where both collinear quarks are in the same sector, or in different sectors. The quarks are necessarily in a same chirality pair, which simplifies the operator basis. The basis of helicity operators was already constructed in \sec{ang_cons} and is given by
\begin{align}
& \boldsymbol{(gq)_n (\bar q)_{\bf n} :}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qg_orderlam_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z1_basis}
&O_{n\bar n1+(-;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, J_{n\bar n\,-}^{{\bar \alpha}\beta} \,J_{e\pm }
\,,\qquad &
&O_{n\bar n1-(+;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a \, J_{n\bar n\,+}^{{\bar \alpha}\beta}\,J_{e\pm }
\,,
\end{alignat}
and
\begin{align}
& \boldsymbol{(q)_n (g\bar q)_{\bf n} :}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qg_orderlam_flip_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z1_basis_flip}
&O_{n\bar n2-(-;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{\bar n-}^a\, J_{n\bar n\,-}^{{\bar \alpha}\beta} \,J_{e\pm }
\,,\qquad &
&O_{n\bar n2+(+;\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a \, J_{n\bar n\,+}^{{\bar \alpha}\beta}\,J_{e\pm }
\,,
\end{alignat}
for the case that the quarks are in different collinear sectors, and
\vspace{0.3cm}
\begin{align}
& \boldsymbol{(g)_n (q \bar q)_{\bf n} :}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qqbarg_orderlam}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z1_basis_diff}
&O_{\bar{n}+(0:\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, J_{\bar n0\, }^{{\bar \alpha}\beta}\,J_{e\pm }
\,,\qquad &
&O_{\bar{n}+(\bar 0:\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, J_{\bar n\bar 0\, }^{{\bar \alpha}\beta}\,J_{e\pm }
\,,\\
&O_{\bar{n}-(0:\pm)}^{(1)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, J_{\bar n0\, }^{{\bar \alpha}\beta}\,J_{e\pm }
\,,\qquad &
&O_{\bar{n}-(\bar 0:\pm)}^{(1)a\,{\bar \alpha}\beta}
=\mathcal{B}_{n-}^a\, J_{\bar n\bar 0\, }^{{\bar \alpha}\beta} \,J_{e\pm }
\,,{\nonumber}
\end{alignat}
in the case that they are in the same sector. Note that the operators in \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} are distinct, and therefore both need to be included in the basis, while in the case that both quarks are in the same sector, it is sufficient to chose both quark fields to be in the same sector, since the direction is summed over. In \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} we have eliminated two of the possible helicity combinations, as was discussed in detail in \Sec{sec:ang_cons}.
The color basis is one dimensional, and is given by
\begin{equation} \label{eq:Z1q_color}
\bar{T}^{a\, \alpha{\bar \beta}} = T^a_{\alpha{\bar \beta}}\,.
\end{equation}
After BPS field redefinition the structure of the ultrasoft Wilson lines is different depending on whether the quarks are in different or the same collinear sectors. We find
\begin{equation} \label{eq:BPSgqqcolor}
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}} = \left (T^a Y_n^\dagger Y_{\bar n} \right )_{\alpha \bar \beta}
\,,\qquad
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}} = \left (Y_n^\dagger Y_{\bar n} T^a \right )_{\alpha \bar \beta}
\,,\qquad
\bar{T}_{\mathrm{BPS}}^{ a \alpha{\bar \beta}} = \mathcal{Y}_n^{ba} \mathcal{Y}_{\bf n}^{bc} T^c_{\alpha \bar \beta}
\,,
\end{equation}
for the operators in \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} and \eq{Z1_basis_diff}, respectively. We have used \eq{adjointtofundamental} to simplify the equations in \Eq{eq:BPSgqqcolor} and we will continue to do so throughout this section when it simplifies the relevant Wilson line structures.
The Wilson coefficients of the operators in both \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} and \eq{Z1_basis_diff} can be expanded as
\begin{align} \label{eq:Z1_qqg_expand}
\vec{C}^{(1)}(n,{\bf n};\omega_1; \omega_2, \omega_3; \omega_4,\omega_5)
= & e^2\, \,
\bigg\{ \big[Q^\ell Q^q + v_{\lambda_l}^\ell v_{\lambda_q}^q P_Z(s_{45})\big] \vec{C}^{(1)}_q(n,{\bf n};\omega_1; \omega_2, \omega_3; \omega_4,\omega_5) {\nonumber}
\\ &
+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{45}) \bigg] \vec{C}^{(1)}_v(n,{\bf n};\omega_1; \omega_2, \omega_3; \omega_4,\omega_5) {\nonumber} \\
& + \frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{45})\vec{C}^{(1)}_a(n,{\bf n};\omega_1; \omega_2, \omega_3; \omega_4,\omega_5) \bigg\}
\,,
\end{align}
where the components of the Wilson coefficient have the same meaning as in \eq{Z0_expand_Wilson}. C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:parityrel1}
\vec{C}^{(1)}_{v\,\lambda_1(\lambda_{23}:\lambda_{45})}(n,{\bf n};\{\omega_i\}) &= \vec{C}^{(1)}_{v\,-\lambda_1(-\lambda_{23}:-\lambda_{45})}(n,{\bf n};\{\omega_i\})\,, \\
\vec{C}^{(1)}_{v\,\lambda_1(\lambda_{23}:\lambda_{45})}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &= -\vec{C}^{(1)}_{v\,\lambda_1(-\lambda_{23}:-\lambda_{45})}(n,{\bf n};\omega_1;\omega_3,\omega_2;\omega_5,\omega_4)
\,,{\nonumber} \\
\vec{C}^{(1)}_{v\,\lambda_1(\lambda_{23}:\lambda_{45})}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &= \vec{C}^{(1)}_{v\,\lambda_1(\lambda_{23}:-\lambda_{45})}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_5,\omega_4)\,. {\nonumber}
\end{align}
Here the $\lambda_i$ denote generic helicity labels for the corresponding helicity building blocks, subject to the constraints of angular momentum conservation, as in \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} and \eq{Z1_basis_diff}. $\vec C^{(1)}_q$ and $\vec C^{(1)}_a$ satisfy the same relations, but with an additional negative sign for each of the equations in the case of $\vec C^{(1)}_a$. We are using the notation where $\lambda = \pm 1, 0$ or $\bar{0}$. Additionally, if $\lambda = 0$, then we take $-\lambda = \bar{0}$. Combined, these relations imply that for each of the $\vec{C}^{(1)}_q$, $\vec{C}^{(1)}_v$, and $\vec{C}^{(1)}_a$, only a single Wilson coefficient needs to be calculated for the operators in \Eq{eq:Z1_basis} (for example, $\vec C_{+(- ; +)}$) and one for those in \Eq{eq:Z1_basis_diff} (for example, $\vec C_{+(0:+)}$). In particular, restricting to a mediating photon, and ignoring processes which proceed through a fermion loop, only a single Wilson coefficient is needed in each of the two cases.
Operators involving three collinear gluon fields do not appear in the matching until one-loop, and are therefore not of immediate phenomenological interest. However, we include them here both for completeness and to demonstrate the simplicity of the helicity operator approach. The basis of three gluon operators is
\begin{align}
& \boldsymbol{(g)_n (gg)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_3g}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:eeggg}
&O_{\mathcal{B}+++(\pm)}^{(1)abc}
= \frac{1}{2}\, \mathcal{B}_{n +}^a\, \mathcal{B}_{\bar n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad &&O_{\mathcal{B}---(\pm)}^{(1)abc}
= \frac{1}{2}\, \mathcal{B}_{n-}^a\, \mathcal{B}_{ {\bf n} -}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,, {\nonumber}\\
&O_{\mathcal{B}++-(\pm)}^{(1)abc}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{ \bar n+}^b\, \mathcal{B}_{ \bar n-}^c\, J_{e\pm }
\,, \qquad &&O_{\mathcal{B}-+-(\pm)}^{(1)abc}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{ \bar n+}^b\, \mathcal{B}_{ \bar n-}^c\, J_{e\pm }
\,,
\end{alignat}
where we have taken the two gluon fields to be in the $\bar n$ collinear sector.
Here the factors of $1/2$ are included for convenience as symmetry factors. Note that when writing this basis we have used the angular momentum constraints discussed in \Sec{sec:ang_cons} to eliminate the other two helicity combinations. These missing combinations have $h=\pm2$ about the $\hat n$ axes, and therefore vanish.
The basis of color structures here is two dimensional,
\begin{equation} \label{eq:Z1g_color}
\bar{T}^{abc} =
\begin{pmatrix}
i f^{abc} \\ d^{abc}
\end{pmatrix}
\,,
\qquad
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{a'b'c'}\, {\cal Y}_n^{a'a} {\cal Y}_{\bf n}^{b'b} {\cal Y}_{{\bf n}}^{c'c} \\
d^{a'b'c'}\, {\cal Y}_n^{a'a} {\cal Y}_{\bf n}^{b'b} {\cal Y}_{{\bf n}}^{c'c}
\end{pmatrix}
=
\begin{pmatrix}
i f^{bcd} {\cal Y}_{{\bf n}}^{a'd} {\cal Y}_n^{a'a} \\
d^{bcd}\, {\cal Y}_{\bf n}^{a'd} {\cal Y}_n^{a'a}
\end{pmatrix}
\,.
\end{equation}
Once again we have simplified the Wilson line structures after the BPS field redefinition.
For $e \bar e ggg $, the intermediate boson must couple to a fermion loop, so the Wilson coefficient can be expanded as
\begin{align} \label{eq:Z1_ggg_expand}
\vec{C}^{(1)}(n,{\bf n};\omega_1, \omega_2,\omega_3; \omega_4,\omega_5)
&= e^2 \,
\bigg\{\frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{45}) \vec{C}^{(1)}_a(n,{\bf n};\omega_1, \omega_2,\omega_3; \omega_4,\omega_5) \\
&+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{45}) \bigg] \vec{C}^{(1)}_v(n,{\bf n};\omega_1, \omega_2,\omega_3; \omega_4,\omega_5) \bigg\} {\nonumber}
\,,
\end{align}
where, as in \eq{Z1_qqg_expand}, $\vec{C}^{(1)}_v$, $\vec{C}^{(1)}_a$ correspond to the contributions from the intermediate boson coupling through either the vector or axial couplings respectively, and we have suppressed the helicity labels on all coefficients. C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:chargerel1}
\vec{C}^{(1)}_{v\,\lambda_1\lambda_2\lambda_3(\lambda_{45})}(n,{\bf n};\{\omega_i\}) &= \vec{C}^{(1)}_{ v\,-\lambda_1\,-\lambda_2\,-\lambda_3(-\lambda_{45})}(n,{\bf n};\{\omega_i\}) \,, \\
\vec{C}^{(1)}_{v\,\lambda_1\lambda_2\lambda_3(\lambda_{45})}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_4,\omega_5) &= \begin{pmatrix*}[r] -1 & 0 \\ 0 & 1\end{pmatrix*} \vec{C}^{(1)}_{v\,\lambda_1\lambda_2\lambda_3(-\lambda_{45})}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_5,\omega_4)
{\nonumber} \,, \\
\vec{C}^{(1)}_{v\,\lambda_1\lambda_2\lambda_3(\lambda_{45})}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_4,\omega_5) &= \vec{C}^{(1)}_{v\,\lambda_1\lambda_2\lambda_3(-\lambda_{45})}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_5,\omega_4) \,, {\nonumber}
\end{align}
where the $\lambda_i=\pm$ denote generic helicity labels of the corresponding helicity building blocks, subject to the constraint of angular momentum conservation, as in \Eq{eq:eeggg}, and we have expressed the color structures in the bases of \Eq{eq:Z1g_color}. $\vec C^{(1)}_q$ and $\vec C^{(1)}_a$ satisfy the same relations but with an additional overall negative sign in both equations for $\vec C^{(1)}_a$.
The charge conjugation relations of \eq{chargerel1} imply that to all orders in $\alpha_s$ only the Wilson coefficients for the color structure $d^{abc}$ are non-zero for the vector current, whereas for the axial current, only the Wilson coefficients corresponding to the color structure $if^{abc}$ are non-zero. These statements remain true under renormalization group evolution. The helicity relations of \Eq{eq:chargerel1} then imply that only a single Wilson coefficient for a chosen helicity needs to be calculated for each of the color structures.
\subsection{Sub-subleading Power Operators} \label{sec:subsub}
The construction of the $\mathcal{O}(\lambda^2)$ power suppressed operator basis is slightly more involved, so we divide the discussion into several subsections. The full list of operators can be found in Table \ref{tab:summary}. We separately discuss operators involving only collinear building block fields, operators involving $\mathcal{P}_\perp$ insertions, and operators involving insertions of ultrasoft building blocks.
\subsubsection{Collinear Field Insertions}
Operators involving four collinear fields, corresponding to the partonic processes $e \bar e gggg$, $e \bar eq \bar q q\bar q$, $e \bar e q \bar q Q\bar Q $, $e \bar eggq\bar q$, appear at sub-subleading power. We will consider each in turn.
\vspace{0.4cm}
\noindent{\bf{Four Quark Operators:}}
We begin by considering the case of operators involving four collinear quark fields. When constructing the operator basis, we must consider the case that there are two collinear quarks in each collinear sector, or three collinear quarks in one sector, and one in the other. We must also treat separately the case of identical quark flavors $e \bar eq \bar q q\bar q$ and distinct massless quark flavors $e \bar e q \bar q Q\bar Q $. For the case of distinct quark flavors $e \bar e q \bar q Q\bar Q $ we will have a $q\leftrightarrow Q$ symmetry for the operators. Furthermore the two quarks of flavor $q$, and the two quarks of flavor $\bar Q$, are necessarily of the same chirality. In the case that both quarks of the same flavor appear in the same current, the current will be labeled by the flavor. Otherwise, the current will be labeled with ($q \bar{Q}$) or ($Q\bar{q}$) appropriately. For all these cases, the color basis is
\begin{equation} \label{eq:qqqq_color}
\bar{T}^{\, \alpha{\bar \beta}\gamma{\bar \delta}} =
\Bigl(
\delta_{\alpha{\bar \delta}}\, \delta_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\Bigr)
\,.\end{equation}
We will give results for the corresponding $\bar T_{\rm BPS}^{\, \alpha{\bar \beta}\gamma{\bar \delta}}$ basis as we consider each case below.
For the case of operators with distinct quark flavors $e \bar e q \bar q Q\bar Q $ and two collinear quarks in each of the $n$ and ${\bf n}$ sectors there are three possibilities. There is either a quark anti-quark pair of the same flavor in each sector (e.g. $(q \bar q)_n(Q \bar Q)_{\bf n}$), a quark and an antiquark of distinct flavors in the same sector (e.g. $(q \bar Q)_n(Q \bar q)_{\bf n}$), or two quarks with distinct flavors in the same sector(e.g. $(q Q)_n(\bar q \bar Q)_{\bf n}$). In the case that there is a quark anti-quark pair of the same flavor in each sector, the basis of helicity operators is
\begin{align}
\boldsymbol{(q \bar q)_n(Q \bar Q)_{\bf n}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_qQ}
&O_{qQ1(0;0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
&O_{qQ1(0;\bar 0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q) n 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n \bar 0\,}^{{\bar \gamma}\delta} J_{e\pm }
\,,\\
&O_{qQ1(\bar 0;0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
&O_{qQ1(\bar 0;\bar 0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q) \bar n \bar 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,{\nonumber}
\end{alignat}
where we have chosen the $q$ quark to be in the $n$ sector. Since all the operators have total helicity $0$ along the $\hat n$ direction, there are only chirality constraints here and no constraints from angular momentum conservation. In the case that there is a quark anti-quark of distinct flavors in the same sector, chirality and angular momentum conservation constrains the basis to be
\begin{align}
\boldsymbol{(q \bar Q)_n(Q \bar q)_{\bf n}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_b}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align} \label{eq:Z2_basis_qQ_2}
&O_{qQ2(0;0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q \bar{Q}) n 0\, }^{{\bar \alpha}\beta}\, J_{(Q\bar{q} ) \bar{n} 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad
O_{qQ2(\bar 0;\bar 0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q \bar{Q}) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q\bar{q} )\bar{n} \bar 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,.
\end{align}
For the operators in \eqs{Z2_basis_qQ}{Z2_basis_qQ_2} the color basis after BPS field redefinition is
\begin{align} \label{eq:TBPS_OqQ12}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \Big[ Y_{n}^\dagger Y_{\bar n} \Big]_{\alpha{\bar \delta}}\,\Big[ Y_{\bar n}^\dagger Y_{n} \Big]_{\gamma{\bar \beta}}\,,\, \delta_{\alpha{\bar \beta}}\, \delta_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
When there are two quarks of distinct flavors in the same sector the basis of helicity operators is constrained by chirality and reduced further to just two operators by angular momentum conservation, giving
\begin{align}
\boldsymbol{(q Q)_n(\bar q \bar Q)_{\bf n}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_c}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Z2_basis_qQ_3}
&O_{qQ3(+;-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar n +\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n -\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
O_{qQ3(-;+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar n -\, }^{{\bar \alpha}\beta}\, J_{(Q) n\bar n +\, }^{{\bar \gamma}\delta} J_{e\pm }
\,.
\end{align}
For the operators in \eq{Z2_basis_qQ_3} the color basis after BPS field redefinition is
\begin{align} \label{eq:TBPS_OqQ3}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\alpha{\bar \delta}}\,\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \beta}}
\,,\,
\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\alpha{\bar \beta}}\,\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
In the cases in \eqs{Z2_basis_qQ}{Z2_basis_qQ_2} where there is a quark and antiquark field in the same collinear sector, we have chosen to work in a basis using $J_{i0}^{{\bar \alpha}\beta}$ and $J_{i\bar 0}^{{\bar \alpha}\beta}$ which contain only fields in a single collinear sector. One could also construct an alternate form for the basis, for example using the currents $J_{n{\bf n}\lambda}^{{\bar \alpha}\beta}$. However, from the point of view of factorization, our basis is more convenient. The fields in the $n$ and $\bar n$ collinear sectors are only connected by color indices, which will simplify later steps of factorization proofs. In the following, we will make this choice for our basis whenever possible.
For identical quark flavors the operators have the same structure as in Eqs.~(\ref{eq:Z2_basis_qQ},\ref{eq:Z2_basis_qQ_2},\ref{eq:Z2_basis_qQ_3}), except the operators $O^{(2)}_{qQ1}$ and $O^{(2)}_{qQ2}$ are no longer distinct. A basis of operators is then given by
\begin{align}
\boldsymbol{(q \bar q)_n (q \bar{q})_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_d}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_qq}
&O_{qq1(0;0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
&O_{qq1(0;\bar 0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q) n 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n \bar 0\,}^{{\bar \gamma}\delta} J_{e\pm }
\,,\\
&O_{qq1(\bar 0;0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
&O_{qq1(\bar 0;\bar 0:\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) \bar n \bar 0\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,{\nonumber}
\end{alignat}
and
\begin{align}
\boldsymbol{(q q)_n(\bar q \bar q)_{\bf n}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_c_same}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:Z2_basis_qq_3}
&O_{qq3(+;-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=\, J_{(q) n \bar n +\, }^{{\bar \alpha}\beta}\, J_{(q) n\bar n -\, }^{{\bar \gamma}\delta} J_{e\pm }
\,,\qquad &
O_{qq3(-;+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \, J_{(q)n \bar n -\, }^{{\bar \alpha}\beta}\, J_{(q) n\bar n +\, }^{{\bar \gamma}\delta} J_{e\pm }
\,.
\end{align}
We also have the same color bases as in \eqs{TBPS_OqQ12}{TBPS_OqQ3} for $ O_{qq1}^{(2)}$ and $O_{qq3}^{(2)}$ respectively.
We must also consider the operators with three collinear quarks in one sector, and one quark in the other. To minimize the number of operators to display, we exploit the $q\leftrightarrow Q$ and $n\leftrightarrow \bar n$ symmetry to restrict ourselves to the case where the single quark (or antiquark) has flavor $Q$ and is in the $n$ collinear sector. The basis for the distinct flavor case with three quarks in the same collinear sector is then
\begin{align}
\boldsymbol{(Q)_n (\bar{Q} q \bar q)_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_31_a}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align} \label{eq:Z3_basis_qQ}
& O_{qQ4(0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n 0\, }^{{\bar \alpha}\beta}\, J_{(Q)n \bar n+\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qQ4(\bar 0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q) n \bar n +\,}^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\\
& O_{qQ4(0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n 0 \, }^{{\bar \alpha}\beta}\, J_{(Q)n \bar n -\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qQ4(\bar 0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \bar 0\,}^{{\bar \alpha}\beta}\, J_{(Q)n \bar n -\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,, {\nonumber} \\
& O_{qQ5(0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n 0\,}^{{\bar \alpha}\beta}\, J_{(Q)\bar n n +\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qQ5(\bar 0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(Q)\bar n n +\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,{\nonumber}\\
& O_{qQ5(0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= J_{(q)\bar n 0\, }^{{\bar \alpha}\beta}\, J_{(Q)\bar n n -\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qQ5(\bar 0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
=J_{(q)\bar n \bar 0\,}^{{\bar \alpha}\beta}\, J_{(Q)\bar n n -\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,.{\nonumber}
\end{align}
Note that unlike the case with two quarks in each collinear sector in \Eq{eq:Z2_basis_qQ}, here angular momentum conservation does not impose constraints beyond those from chirality, and the flavor diagonal nature of QCD and tree level electroweak interactions.
For the $O_{qQ4}^{(2)}$ operators the color basis after BPS field redefinition is
\begin{align} \label{eq:TBPS_OqQ4}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \,\delta_{\alpha{\bar \delta}}\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \beta}}
\,,\,
\delta_{\alpha{\bar \beta}}\left[ Y_{n}^\dagger Y_{\bar n} \right ]_{\gamma{\bar \delta}}
\right)
\,,
\end{align}
while for the $O_{qQ5}^{(2)}$ operators the corresponding basis is
\begin{align} \label{eq:TBPS_OqQ5}
\bar{T}_{\mathrm{BPS}}^{ \alpha{\bar \beta}\gamma{\bar \delta}} &=
\left( \left[ Y_{\bar n}^\dagger Y_{n} \right ]_{\alpha{\bar \delta}}\,\delta_{\gamma{\bar \beta}}
\, \,,\,
\delta_{\alpha{\bar \beta}} \left[ Y_{\bar n}^\dagger Y_{n} \right ]_{\gamma{\bar \delta}}
\right)
\,.
\end{align}
In the case of identical quark flavors, the same basis of eight terms as in \eq{Z3_basis_qQ} define $O_{qq4}^{(2)}$ and $O_{qq5}^{(2)}$, and the BPS color basis is as in \eq{TBPS_OqQ4} for $O_{qq4}^{(2)}$, and as in \eq{TBPS_OqQ5} for $O_{qq5}^{(2)}$. For convenience we add additional symmetry factors to the following operators,
\begin{align}
\boldsymbol{(q)_n (\bar q q \bar{q})_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4q_31_b}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align} \label{eq:Z3_basis_qq}
& O_{qq4(0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \frac{1}{2} J_{(q)\bar n 0\, }^{{\bar \alpha}\beta}\, J_{(q)n \bar n+\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qq4(\bar 0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \frac{1}{2} J_{(q)\bar n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q) n \bar n +\,}^{{\bar \gamma}\delta} \,J_{e\pm }
\,, \\
& O_{qq5(0:+;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \frac{1}{2} J_{(q)\bar n 0\,}^{{\bar \alpha}\beta}\, J_{(q)\bar n n +\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,,\qquad
O_{qq5(\bar 0:-;\pm)}^{(2){\bar \alpha}\beta{\bar \gamma}\delta}
= \frac{1}{2} J_{(q)\bar n \bar 0\, }^{{\bar \alpha}\beta}\, J_{(q)\bar n n +\, }^{{\bar \gamma}\delta} \,J_{e\pm }
\,.{\nonumber}
\end{align}
Having enumerated a complete basis for all types of four quark operators at ${\cal O}(\lambda^2)$, we now consider relations that follow from $C$ and $P$.
To make these relations explicit we expand the Wilson coefficients for $e\bar e q\bar qQ\bar Q$ as
\begin{align} \label{eq:Z2_expand}
\vec{C}^{(2)}_{qQi} (n,{\bf n};\omega_1, \omega_2; & \omega_3, \omega_4;\omega_5,\omega_6)
= e^2 \,
\bigg\{\big[Q^\ell Q^q + v_{\lambda_l}^\ell v_{\lambda_q}^q P_Z(s_{56})\big] \vec{C}^{(2)}_{q,qQi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6)
{\nonumber} \\ &
\hspace{-0.85cm}+ \big[Q^\ell Q^Q + v_{\lambda_l}^\ell v_{\lambda_Q}^Q P_Z(s_{56}) \big] \vec{C}^{(2)}_{Q,qQi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4; \omega_5,\omega_6) {\nonumber} \\
& \hspace{-0.85cm} +\frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{56}) \vec{C}^{(2)}_{a,qQi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6) \bigg\}
\ {\nonumber} \\ &
\hspace{-0.85cm}+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{56}) \bigg] \vec{C}^{(2)}_{v,qQi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6)
\,.
\end{align}
Since we have accounted for symmetry factors explicitly in the the operators, for the case of identical quark flavors, $e\bar e q\bar qq\bar q$, we have the relation
\begin{align}
& \vec{C}^{(2)}_{qqi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6)={\nonumber} \\
&\hspace{3cm}\vec{C}^{(2)}_{qQi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6)- \vec{C}^{(2)}_{qQi}(n,{\bf n};\omega_1, \omega_4; \omega_3, \omega_2;\omega_5,\omega_6)
\,.
\end{align}
We now discuss relations between different helicity operators due to symmetry constraints. C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:4quarkPrelation}
\vec{C}^{(2)}_{v,\,qQi(\lambda_{12}:\lambda_{34};\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6&) \hspace{-0.05cm}=\hspace{-0.05cm} \vec{C}^{(2)}_{v,\,qQi(-\lambda_{12}:-\lambda_{34};-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6), {\nonumber} \\
\vec{C}^{(2)}_{v,\,qQj(\lambda_{12}:\lambda_{34};\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6&) \hspace{-0.05cm}=\hspace{-0.05cm} - \vec{C}^{(2)}_{v,\,qQj(-\lambda_{12}:-\lambda_{34};-\lambda_{56})}(n,{\bf n};\omega_2,\omega_1;\omega_4,\omega_3;\omega_6,\omega_5),
{\nonumber} \\
\vec{C}^{(2)}_{v,\,qQi(\lambda_{12}:\lambda_{34};\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6&)\hspace{-0.05cm}=\hspace{-0.05cm} \vec{C}^{(2)}_{v,\,qQi(\lambda_{12}:\lambda_{34};-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_6,\omega_5).
\end{align}
Here the $\lambda_i$ denote generic helicity labels of the corresponding helicity building blocks, subject to the constraints from angular momentum conservation discussed in this section. For the scalar currents, we are again using the convention that for $\lambda = 0$, $-\lambda = \bar{0}$.
The same relations hold for $\vec{C}^{(2)}_{Q,qQi}$ and $\vec{C}^{(2)}_{q,qQi}$, and there is an additional overall minus sign in all these relations for $\vec{C}^{(2)}_{a,qQi}$.
\vspace{0.4cm}
\noindent{\bf{Two Quark-Two Gluon Operators:}}
We now consider the operators involving two collinear quark and two collinear gluon building blocks, corresponding to the partonic process $e \bar eq\bar qgg$. The quark and antiquark have the same chirality, but are not necessarily in the same collinear sector, as is also the case for the collinear gluons. The color basis for these channels is three dimensional, and we take our color basis to be
\begin{equation} \label{eq:ggqqll_color}
\bar{T}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(T^a T^b)_{\alpha{\bar \beta}}\,,\, (T^b T^a)_{\alpha{\bar \beta}} \,,\, {\rm tr}[T^a T^b]\, \delta_{\alpha{\bar \beta}}
\Bigr)
\,.\end{equation}
We begin with operators that have the quarks in opposite collinear sectors, and two gluons in the same collinear sector. A basis for these operators is
\begin{align}
\boldsymbol{(gg q)_n (\bar q)_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qgg}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eeqqgg_basis1}
&O_{\cB1++(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \frac{1}{2} \mathcal{B}_{n+}^a\, \mathcal{B}_{n+}^b\, J_{n\bar n\, - }^{{\bar \alpha}\beta} J_{e\pm }
\, , \qquad
&&O_{\cB1--(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \frac{1}{2} \mathcal{B}_{ n-}^a\, \mathcal{B}_{ n-}^b \, J_{n\bar n\, + }^{{\bar \alpha}\beta} J_{e\pm }
\, , {\nonumber}\\
&O_{\cB1+-(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{n-}^b \, J_{n\bar n\, +}^{{\bar \alpha}\beta} J_{e\pm }
\, , \qquad
&&O_{\cB1+-(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \, \mathcal{B}_{n+}^a\, \mathcal{B}_{ n-}^b J_{n\bar n\, -}^{{\bar \alpha}\beta} J_{e\pm }
\, ,
\end{alignat}
\begin{align}
\boldsymbol{(gg \bar q)_n (q)_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qbargg}}}
{\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eeqqgg_basis1a}
&O_{\cB2++(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \frac{1}{2} \mathcal{B}_{n+}^a\, \mathcal{B}_{n+}^b\, J_{\bar n n\, + }^{{\bar \alpha}\beta} J_{e\pm }
\, , \qquad
&&O_{\cB2--(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \frac{1}{2} \mathcal{B}_{ n-}^a\, \mathcal{B}_{ n-}^b \, J_{\bar n n\, - }^{{\bar \alpha}\beta} J_{e\pm }
\, , {\nonumber} \\
&O_{\cB2+-(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{n-}^b \, J_{\bar n n\, +}^{{\bar \alpha}\beta} J_{e\pm }
\, , \qquad
&&O_{\cB2+-(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \, \mathcal{B}_{n+}^a\, \mathcal{B}_{ n-}^b J_{\bar n n\, -}^{{\bar \alpha}\beta} J_{e\pm }
\, .
\end{alignat}
Here we have used constraints from angular momentum conservation to eliminate operators which do not have $h=0,\pm1$ along the $\hat n$ axis, and we have taken the two gluon fields to be in the $n$ collinear sector. For the operators in \eq{eeqqgg_basis1} the color basis after BPS field redefinition is
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(T^a T^bY^\dagger_n Y_{\bar n})_{\alpha{\bar \beta}}\,,\, (T^b T^aY^\dagger_n Y_{\bar n})_{\alpha{\bar \beta}} \,,\, {\rm tr}[T^a T^b]\, [Y^\dagger_n Y_{\bar n}]_{\alpha{\bar \beta}}
\Bigr)
\,,
\end{equation}
while for the operators in \eq{eeqqgg_basis1a} it is
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(Y^\dagger_{\bf n} Y_{n} T^a T^b)_{\alpha{\bar \beta}}\,,\,
(Y^\dagger_{\bf n} Y_{n}T^b T^a)_{\alpha{\bar \beta}} \,,\,
{\rm tr}[T^a T^b]\, [Y^\dagger_{\bf n} Y_{n}]_{\alpha{\bar \beta}}
\Bigr)
\,.
\end{equation}
Next we consider the operators with two gluon building blocks in distinct collinear sectors. When the quarks and gluons are both in distinct collinear sectors the basis of operators is
\begin{align}
\boldsymbol{(g q)_n (g \bar q)_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qg}}} {\nonumber}
\end{align}
\vspace{-0.45cm}
\begin{alignat}{2} \label{eq:eeqqgg_basis2}
&O_{\cB3++(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\bar n\, + }^{{\bar \alpha}\beta} J_{e\pm } \, , \qquad
&&O_{\cB3--(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\bar n\, -}^{{\bar \alpha}\beta} J_{e\pm }\, ,
{\nonumber} \\
&O_{\cB3++(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\bar n\, - }^{{\bar \alpha}\beta} J_{e\pm } \, , \qquad
&&O_{\cB3--(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\bar n\, +}^{{\bar \alpha}\beta} J_{e\pm }\, ,
\\
&O_{\cB3+-(-;\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\bar n\, - }^{{\bar \alpha}\beta} J_{e\pm }\, , \qquad
&&O_{\cB3-+(+;\pm)}^{(2)ab\, {\bar \alpha}\beta}
=\mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\bar n\, + }^{{\bar \alpha}\beta} J_{e\pm } \, , {\nonumber}
\end{alignat}
and the color basis after BPS field redefinition is
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(T^a Y^\dagger_n Y_{\bar n} T^b)_{\alpha{\bar \beta}}\,,\, (Y^\dagger_n T^d \mathcal{Y}_{{\bf n}}^{db} T^c \mathcal{Y}^{ca}_n Y_{\bar n})_{\alpha{\bar \beta}} \,,\, {\rm tr}[ T^c \mathcal{Y}^{ca}_n T^d \mathcal{Y}_{{\bf n}}^{db} ]\, [Y^\dagger_n Y_{\bar n}]_{\alpha{\bar \beta}}
\Bigr)
.
\end{equation}
Here operators with $J_{{\bf n} n\lambda}^{{\bar \alpha}\beta}$ are obtained from those in \eq{eeqqgg_basis2} by $n\leftrightarrow {\bf n}$. When the two quarks are in the same collinear sector the basis is given by
\begin{align}
\boldsymbol{(g q\bar q)_n (g)_{{\bf n}}:} {\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qqbarg}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eeqqgg_basis3}
&O_{\cB4++(0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\,{0} }^{{\bar \alpha}\beta} J_{e\pm }\, , \qquad
&&O_{\cB4++(\bar 0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
=\mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta} J_{e\pm }\, , \\
&O_{\cB4--(0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\,{0} }^{{\bar \alpha}\beta} J_{e\pm }\, , \qquad
&&O_{\cB4--(\bar 0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{ n-}^a\, \mathcal{B}_{\bar n-}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta} J_{e\pm }\, . {\nonumber}
\end{alignat}
The color basis after BPS field redefinition is given by
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(\mathcal{Y}_n^T \mathcal{Y}_{\bf n})^{cb} (T^a T^c)_{\alpha{\bar \beta}} \,,\,
(\mathcal{Y}_n^T \mathcal{Y}_{\bf n})^{cb} (T^c T^a)_{\alpha{\bar \beta}} \,,\,
T_F (\mathcal{Y}_n^T \mathcal{Y}_{\bf n})^{ab} \, \delta_{\alpha{\bar \beta}}
\Bigr)
\,.
\end{equation}
In writing \eq{eeqqgg_basis3} we have again used constraints of angular momentum conservation to restrict the allowed operators in the basis.
Finally we consider the basis of operators with both quarks in the same collinear sector, and both gluons in the other collinear sector. Imposing angular momentum conservation reduces the basis to two distinct operators
\begin{align}
& \boldsymbol{(q \bar q)_n (gg)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_gg}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{align}\label{eq:eeqqgg_basis4}
&O_{\cB5+-(0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a\, \mathcal{B}_{ \bar n-}^b \, J_{n\,{0} }^{{\bar \alpha}\beta} J_{e\pm }\,, \qquad
&O_{\cB5+-(\bar 0:\pm)}^{(2)ab\, {\bar \alpha}\beta}
= \mathcal{B}_{\bar n+}^a\, \mathcal{B}_{ \bar n-}^b \, J_{n\,{\bar 0} }^{{\bar \alpha}\beta} J_{e\pm }\,.
\end{align}
The color basis after BPS field redefinition is
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\, ab \alpha{\bar \beta}}
= \Bigl(
(Y_n^\dagger Y_{\bf n} T^a T^b Y^\dagger_{\bar n} Y_n )_{\alpha{\bar \beta}}
\,,\,
(Y_n^\dagger Y_{\bf n} T^b T^a Y^\dagger_{\bar n} Y_n )_{\alpha{\bar \beta}}
\,,\,
{\rm tr}[T^a T^b]\, \delta_{\alpha{\bar \beta}} \Bigr)
\,.
\end{equation}
We have chosen to write the operators with both quarks in the $n$ sector.
For the $e\bar e q\bar q g g$ operators in Eqs.~(\ref{eq:eeqqgg_basis1},\ref{eq:eeqqgg_basis2},\ref{eq:eeqqgg_basis3},\ref{eq:eeqqgg_basis4}) we expand the Wilson coefficients as
\begin{align} \label{eq:Z2_qqgg_expand}
& \vec{C}^{(2)}_{Bi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6)
= e^2\,
\bigg\{ \big[Q^\ell Q^q + v_{\lambda_l}^\ell v_{\lambda_q}^q P_Z(s_{56})\big] \vec{C}^{(2)}_{q,Bi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6) {\nonumber}
\\ &
\hspace{4cm}+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{56}) \bigg] \vec{C}^{(2)}_{v,Bi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6) {\nonumber}
\\ &
\hspace{4cm} + \frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{56}) \vec{C}^{(2)}_{a,Bi}(n,{\bf n};\omega_1, \omega_2; \omega_3, \omega_4;\omega_5,\omega_6) \bigg\}
\,.
\end{align}
C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:2q2gCPrelation}
&\vec{C}^{(2)}_{v,\,\mathcal{B}\, i\lambda_1 \lambda_2(\lambda_{34}:\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6) {\nonumber} \\
& \hspace{4cm}= \vec{C}^{(2)}_{v,\,\mathcal{B}\, i-\lambda_1 -\lambda_2(-\lambda_{34}:-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6)\,, \\
&\vec{C}^{(2)}_{v,\,\mathcal{B}\, i\lambda_1 \lambda_2(\lambda_{34}:\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6) {\nonumber} \\
&\hspace{4cm}= \begin{pmatrix*}[r] 0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 1\end{pmatrix*} \vec{C}^{(2)}_{v,\,\mathcal{B}\, i\lambda_1 \lambda_2(-\lambda_{34}:-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_4,\omega_3;\omega_5,\omega_6)\,,
{\nonumber} \\
&\vec{C}^{(2)}_{v,\,\mathcal{B}\, i\lambda_1 \lambda_2(\lambda_{34}:\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_5,\omega_6) {\nonumber} \\
& \hspace{4cm}= \vec{C}^{(2)}_{v,\,\mathcal{B}\, i\lambda_1 \lambda_2(\lambda_{34}:-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2;\omega_3,\omega_4;\omega_6,\omega_5)\,, {\nonumber}
\end{align}
where the index $i$ runs from 1 to 4. Here the $\lambda_i$ denote generic helicity labels of the corresponding helicity building blocks, subject to the constraints from angular momentum conservation discussed in this section. The same two relations hold for $\vec{C}^{(2)}_{q}$, and hold with the addition of an overall minus sign for $\vec{C}^{(2)}_{a}$.
\vspace{0.4cm}
\noindent{\bf{Four Gluon Operators:}}
Finally, we consider $\mathcal{O}(\lambda^2)$ hard scattering operators involving four collinear gluons. The four gluon channel gives a highly suppressed contribution for $e^+e^- \to$ dijets and Drell-Yan, but we nevertheless present it here for completeness. It also provides a nice demonstration of the helicity basis approach, as the construction of a minimal basis of four gluon operators is quite difficult otherwise. The helicity operators that include four gluons were presented in the example of in \Eq{eq:helicitygggg} for the case of four well separated collinear sectors. To adapt these operators to the case of two collinear sectors, we need to restrict the sector labels to $n$ and ${\bf n}$ and impose the angular momentum constraints of \sec{ang_cons}. The color basis for the four gluon operators will include more structures than were used in \Eq{eq:color_gggg}, as we now have to allow axial couplings and CP violation. Our choice for this basis is
\begin{equation} \label{eq:gggg_color}
\bar{T}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
{\rm tr}[abcd] + {\rm tr}[dcba] \\ {\rm tr}[acdb] + {\rm tr}[bdca] \\ {\rm tr}[adbc] + {\rm tr}[cbda] \\
{\rm tr}[abcd] - {\rm tr}[dcba] \\ {\rm tr}[acdb] - {\rm tr}[bdca] \\ {\rm tr}[adbc] - {\rm tr}[cbda] \\ 2{\rm tr}[ab]\, {\rm tr}[cd] \\ 2{\rm tr}[ac]\, {\rm tr}[db] \\ 2{\rm tr}[ad]\, {\rm tr}[bc]
\end{pmatrix}^{\!\!\!T}
\,.\end{equation}
For the specific case of SU($N_c$) with $N_c=3$ it is possible to further reduce the color basis by using relations of the form
\begin{align}
&{\rm tr}[abcd+dcba] + {\rm tr}[acdb+bdca] + {\rm tr}[adbc+cbda]
{\nonumber}\\ & \qquad
= {\rm tr}[ab]{\rm tr}[cd] + {\rm tr}[ac]{\rm tr}[db] + {\rm tr}[ad]{\rm tr}[bc]
\,.\end{align}
We prefer not to use this relation since it makes the structure more complicated, and does not hold for $N_c>3$, and hence one can not look at the large $N_c$ scaling of results if one uses such relations.
To construct a complete basis of four gluon operators with two collinear sectors, we need to consider two cases. First, when we have two gluons in each sector a basis of operators is
\begin{align}
& \boldsymbol{(gg)_n (gg)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4g}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_gggg_1}
&O_{4g1++++(\pm)}^{(2)a b c d}
= \,\frac{1}{4} \mathcal{B}^a_{n +} \mathcal{B}^b_{n +} \mathcal{B}^c_{{\bf n} +} \mathcal{B}^d_{{\bf n} +} J_{e\pm }
\,,\qquad &
&O_{4g1+-+-(\pm)}^{(2)a b c d}
= \, \mathcal{B}^a_{n +} \mathcal{B}^b_{n -} \mathcal{B}^c_{{\bf n} +} \mathcal{B}^d_{{\bf n} -} J_{e\pm }
\,,\\
&O_{4g1----(\pm)}^{(2)a b c d}
= \,\frac{1}{4} \mathcal{B}^a_{n -} \mathcal{B}^b_{n -} \mathcal{B}^c_{{\bf n} -} \mathcal{B}^d_{{\bf n} -} J_{e\pm }
\,,\qquad &{\nonumber}
\end{alignat}
where we have used angular momentum constraints to eliminate operators that contain only one $+$ or one $-$ helicity. The basis of color structures after BPS field redefinition is given by
\begin{equation}
\bar{T}_{\text{BPS}}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] + {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] + {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] + {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] - {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] - {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] - {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{n} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
\frac{1}{2} \delta^{ab} \delta^{cd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{ac} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{bd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{ad} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{bc}
\end{pmatrix}^{\!\!\!T}
.\end{equation}
The other relevant case has three gluons in one sector, and we can take advantage of the $n\leftrightarrow {\bf n}$ symmetry to choose the three gluons to be in the ${\bf n}$ collinear sector. The basis of operators is then given by
\begin{align}
& \boldsymbol{(g)_n (ggg)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_4ga}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:Z2_basis_gggg_2}
&O_{4g2+++-(\pm)}^{(2)a b c d}
= \,\frac{1}{2} \mathcal{B}^a_{n +} \mathcal{B}^b_{{\bf n} +} \mathcal{B}^c_{{\bf n} +} \mathcal{B}^d_{{\bf n} -} J_{e\pm }
\,,\qquad &
&O_{4g2-+--(\pm)}^{(2)a b c d}
= \,\frac{1}{2} \mathcal{B}^a_{n -} \mathcal{B}^b_{{\bf n} +} \mathcal{B}^c_{{\bf n} -} \mathcal{B}^d_{{\bf n} -} J_{e\pm }
\,,
\end{alignat}
where we have once again used conservation of angular momentum to restrict to these particular helicity choices. In this case, we have
\begin{equation}
\bar{T}_{\text{BPS}}^{ abcd} =
\frac{1}{2}\begin{pmatrix}
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] + {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] + {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] + {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{b'} T^{c'} T^{d'} ] - {\rm tr}[ T^{d'} T^{c'} T^{b'} T^{a'} ]) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{c'} T^{d'} T^{b'} ] - {\rm tr}[ T^{b'} T^{d'} T^{c'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
({\rm tr}[ T^{a'} T^{d'} T^{b'} T^{c'} ] - {\rm tr}[ T^{c'} T^{b'} T^{d'} T^{a'} ] ) \mathcal{Y}^{a' a}_{n} \mathcal{Y}^{b' b}_{{\bf n}} \mathcal{Y}^{c' c}_{{\bf n}} \mathcal{Y}^{d' d}_{{\bf n}}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{ab} \delta^{cd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{ac} \delta^{bd}
\\
\frac{1}{2} (\mathcal{Y}^T_n \mathcal{Y}_{\bf n})^{ad} \delta^{bc}
\end{pmatrix}^{\!\!\!T}
.\end{equation}
Once again for simplicity we have not used identities to simplify some of these Wilson line structures. Also we note that
Just as in the case of three gluons, the $e \bar{e} gggg$ channel must proceed through a fermion loop, so we can decompose the Wilson coefficient as
\begin{align} \label{eq:Z2_gggg_expand}
\vec{C}^{(2)}(n,{\bf n};\omega_1, \omega_2,&\omega_3,\omega_4; \omega_5,\omega_6)
= e^2 \,
\bigg\{\frac{v_{\lambda_l}^\ell}{\sin(2\theta_W)} P_Z(s_{56}) \vec{C}^{(2)}_a(n,{\bf n};\omega_1, \omega_2,\omega_3, \omega_4;\omega_5,\omega_6) {\nonumber} \\
&\hspace{-1cm}+\sum_{j=1}^{n_f} \bigg[Q^\ell Q^j + \frac{v_{\lambda_l}^\ell}{2} (v_L^j +v_R^j) P_Zs_{56}) \bigg] \vec{C}^{(2)}_v(n,{\bf n};\omega_1, \omega_2,\omega_3, \omega_4;\omega_5,\omega_6) \bigg\} \,,
\end{align}
where we have suppressed all of the helicity labels and $\vec{C}^{(2)}_a$ and $\vec{C}^{(2)}_v$ correspond to the axial or vector coupling contributions respectively. With this Wilson coefficient expansion and the color basis from \Eq{eq:gggg_color}, C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:chargerelgggg}
\vec{C}^{(2)}_{v\,\lambda_1\lambda_2\lambda_3 \lambda_4(\lambda_{56})}(n,{\bf n};\{\omega_i\}) &= \vec{C}^{(2)}_{ v\,-\lambda_1\,-\lambda_2\,-\lambda_3\,-\lambda_4(-\lambda_{56})}(n,{\bf n};\{\omega_i\}) \,, \\
\vec{C}^{(2)}_{v\,\lambda_1\lambda_2\lambda_3 \lambda_4(\lambda_{56})}(n,{\bf n};\omega_1,\omega_2,\omega_3,\omega_4;\omega_5,\omega_6) &= \hat{V}_{4g} \vec{C}^{(2)}_{v\,\lambda_1\lambda_2\lambda_3 \lambda_4(-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2,\omega_3,\omega_4;\omega_6,\omega_5) \,, {\nonumber} \\
\vec{C}^{(2)}_{v\,\lambda_1\lambda_2\lambda_3 \lambda_4(\lambda_{56})}(n,{\bf n};\omega_1,\omega_2,\omega_3,\omega_4;\omega_5,\omega_6) &= \vec{C}^{(2)}_{v\,\lambda_1\lambda_2\lambda_3 \lambda_4(-\lambda_{56})}(n,{\bf n};\omega_1,\omega_2,\omega_3,\omega_4;\omega_6,\omega_5)
{\nonumber} \,,
\end{align}
where $\hat{V}_{4g}$ is diagonal in the space defined by \Eq{eq:gggg_color} with +1 for the first three entries, -1 for the middle three entries and +1 for the final three entries. The $\lambda_i$ are generic helicity labels, but are restricted by the constraints from angular momentum conservation discussed earlier. $\vec{C}^{(2)}_a$ satisfies the same relations with an additional negative sign.
\subsubsection{$\mathcal{P}_\perp^\pm$ Insertions}
For our choice of kinematics, hard scattering operators with explicit $\mathcal{P}_\perp^\pm $ insertions first arise at $\mathcal{O}(\lambda^2)$. Since operators involving a $\mathcal{P}_\perp^\pm$ insertion that acts on an entire collinear sector vanish, the only non-vanishing $\mathcal{O}(\lambda^2)$ operators involve single insertions of $\mathcal{P}_\perp^\pm$ into subleading power operators with two collinear fields in the same sector. The contributing partonic processes are identical to those considered at subleading power, so we can decompose the Wilson coefficients following \eqs{Z1_qqg_expand}{Z1_ggg_expand}, and use the color bases of \Eqss{eq:Z1q_color}{eq:BPSgqqcolor}{eq:Z1g_color}.
For the insertions of $\mathcal{P}_\perp^\pm$ into operators involving two quarks and a gluon, a basis of operators for the case that the quarks are in distinct collinear sectors is
\begin{align}
& \boldsymbol{(gq \mathcal{P}_\perp)_n (\bar q)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qg_orderlam_perp_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eeqqgpperp_basis}
&O_{\cP1n + (+;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{n+}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1n - (-;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{n-}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\\
&O_{\cP1n - (+;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{n-}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1n - (+;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{n-}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,{\nonumber}\\
&O_{\cP1n + (-;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{n+}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1n + (-;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{n+}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, {\nonumber}
\end{alignat}
and
\begin{align}
& \boldsymbol{(q)_n (g\bar q \mathcal{P}_\perp)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qg_orderlam_perp_flip_low}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eeqqgpperp_basis_flip}
&O_{\cP1\bar n + (+;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{{\bf n}+}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1\bar n - (-;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{{\bf n}-}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\\
&O_{\cP1\bar n - (+;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{{\bf n}-}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1\bar n - (+;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{{\bf n}-}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,{\nonumber}\\
&O_{\cP1\bar n + (-;\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{+} \mathcal{B}_{{\bf n}+}^a \right ]\, J_{n\bar n\, +}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\cP1\bar n + (-;\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \left [ \mathcal{P}_{\perp}^{-} \mathcal{B}_{{\bf n}+}^a \right ]\, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, {\nonumber}
\end{alignat}
which we refer to as $\mathcal{P} 1$ operators. In the case that they are in the same collinear sector the basis is,
\begin{align}
& \boldsymbol{(g)_n (q\bar q\, \mathcal{P}_\perp)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_qqbarg_orderlam_perp}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:eeqqgpperp_basis_same}
&O_{\cP2 + (0:\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \big\{ \mathcal{P}_{\perp}^{+} J_{\bar n\, 0}^{{\bar \alpha}\beta} \big\}\, J_{e\pm }
\,,\qquad &
&O_{\cP2 - (0:\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a\, \big\{ \mathcal{P}_{\perp}^{-} J_{\bar n\, 0 }^{{\bar \alpha}\beta} \big\} \, J_{e\pm }
\,,\\
&O_{\cP2 + (\bar 0:\pm)[+]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n+}^a\, \big\{ \mathcal{P}_{\perp}^{+} J_{\bar n\, \bar0 }^{{\bar \alpha}\beta} \big\}\, J_{e\pm }
\,,\qquad &
&O_{\cP2 - (\bar 0:\pm)[-]}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{n-}^a \, \big\{ \mathcal{P}_{\perp}^{-} J_{\bar n\, \bar0 }^{{\bar \alpha}\beta} \big\} \, J_{e\pm }
\,,{\nonumber}
\end{alignat}
and we refer to these as $\cP2$ operators. If we integrate the $\mathcal{P}_\perp^\pm$ by parts in \eq{eeqqgpperp_basis} then it gives the operators involving $[\mathcal{P}_\perp^\pm J_{n{\bf n}\lambda}^{{\bar \alpha}\beta}]$, and doing this in \eq{eeqqgpperp_basis_same} gives the terms $\{ J_{{\bf n} 0}^{{\bar \alpha}\beta} (\mathcal{P}_\perp^\pm)^\dagger\}$ and $\{ J_{{\bf n} \bar 0}^{{\bar \alpha}\beta} (\mathcal{P}_\perp^\pm)^\dagger\}$, thus explaining why these structures do not appear as separate terms in the basis. In \eq{eeqqgpperp_basis} there is only only field in the ${\bf n}$ direction, so any operators that contain $\{J_{n {\bf n} \pm}^{{\bar \alpha} \beta} (\mathcal{P}_\perp^\lambda)^\dagger\}$ vanish. Similarly, in \eq{eeqqgpperp_basis_same}, any operators that contain $[\mathcal{P}^\lambda_\perp \mathcal{B}_{n}^\pm]$ are zero for our choice of kinematics. In both \eq{eeqqgpperp_basis} and \eq{eeqqgpperp_basis_same} we have used angular momentum conservation of the hard scattering process to reduce the helicity combinations allowed in the basis.
C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients for the $\mathcal{P} 1$ operators
\begin{align} \label{eq:Pparityrel1}
\vec{C}^{(2)}_{v,\,\mathcal{P} 1 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{P}]}(n,{\bf n};\{\omega_i\}) &= - \vec{C}^{(2)}_{v,\,\mathcal{P} 1 -\lambda_1 (-\lambda_{23};-\lambda_{45})[-\lambda_{P}]}(n^\P,{\bf n}^\P;\{\omega_i\})\,, \\
\vec{C}^{(2)}_{v,\,\mathcal{P} 1 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &=\!-\vec{C}^{(2)}_{v,\,\mathcal{P} 1 -\lambda_1 (-\lambda_{23}:-\lambda_{45})[-\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_3,\omega_2;\omega_5,\omega_4)\,, {\nonumber} \\
\vec{C}^{(2)}_{v,\,\mathcal{P} 1 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &= \vec{C}^{(2)}_{v,\,\mathcal{P} 1 \lambda_1 (\lambda_{23}:-\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_5,\omega_4)\,,{\nonumber}
\end{align}
which hold for both the $n$ and $\bar n$ versions of the operators,
and similarly for the $\mathcal{P} 2$ operators
\begin{align}
\vec{C}^{(2)}_{v,\,\mathcal{P} 2 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\{\omega_i\}) &= - \vec{C}^{(2)}_{v,\,\mathcal{P} 2 -\lambda_1 (-\lambda_{23}:-\lambda_{45})[-\lambda_{\mathcal{P}}]}(n^\P,{\bf n}^\P;\{\omega_i\})\,, \\
\vec{C}^{(2)}_{v,\,\mathcal{P} 2 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &= \vec{C}^{(2)}_{v,\,\mathcal{P} 2 \lambda_1 (-\lambda_{23}:-\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_3,\omega_2;\omega_5,\omega_4)\,,{\nonumber}\\
\vec{C}^{(2)}_{v,\,\mathcal{P} 2 \lambda_1 (\lambda_{23}:\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_4,\omega_5) &= \vec{C}^{(2)}_{v,\,\mathcal{P} 2 \lambda_1 (\lambda_{23}:-\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1;\omega_2,\omega_3;\omega_5,\omega_4)\,.{\nonumber}
\end{align}
A basis of operators involving three collinear gluon fields and a $\mathcal{P}_\perp^\pm$ insertion is given by
\begin{align}
& \boldsymbol{(g)_n (gg\, \mathcal{P}_\perp)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Subleading_3g_perp}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:eegggpperp_basis}
&O_{\mathcal{P}\mathcal{B} +++(\pm)[-]}^{(2)abc}
= \mathcal{B}_{n+}^a\, \mathcal{B}_{\bar n+}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n+}^c \right ] J_{e\pm }
\,,\qquad && O_{\mathcal{P}\mathcal{B} ---(\pm)[+]}^{(2)abc}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{\bar n-}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n-}^c \right ] J_{e\pm }\,,
{\nonumber} \\
%
&O_{\mathcal{P}\mathcal{B} ++-(\pm)[+]}^{(2)abc}
= \, \mathcal{B}_{n+}^a\, \mathcal{B}_{ \bar n+}^b\, \left [\mathcal{P}_{\perp}^{+} \mathcal{B}_{\bar n-}^c \right ] J_{e\pm }
\,,\qquad
&&O_{\mathcal{P}\mathcal{B} --+(\pm)[-]}^{(2)abc}
= \mathcal{B}_{n-}^a\, \mathcal{B}_{ {\bf n} -}^b\, \left [\mathcal{P}_{\perp}^{-} \mathcal{B}_{\bar n +}^c \right ] J_{e \pm }
\,.
\end{alignat}
We have used angular momentum conservation to eliminate certain helicity combinations. Note that the analogous operators with the helicities $O_{\mathcal{P}\mathcal{B} +-+(\pm)[+]}^{(2)abc}$ and $O_{\mathcal{P}\mathcal{B} -+-(\pm)[-]}^{(2)abc}$ are not eliminated, but instead are equivalent to those in the last row by integrating the $\mathcal{P}_\perp^\pm$ by parts onto the other ${\bf n}$-collinear field.
C/P relations combined with the ability to flip the helicity of the electron current, as described in \Eq{eq:lo_weak}, give the following relations between Wilson coefficients
\begin{align} \label{eq:Pchargerel1}
\vec{C}^{(2)}_{v,\,\mathcal{P} \mathcal{B} \lambda_1 \lambda_2 \lambda_3 (\lambda_{45})[\lambda_{\mathcal{P}}]}(n,{\bf n};\{\omega_i\}) &= -\vec{C}^{(2)}_{ v,\,\mathcal{P} \mathcal{B} -\lambda_1 -\lambda_2 -\lambda_3 (-\lambda_{45}) [-\lambda_{\mathcal{P}}]}(n^\P,{\bf n}^\P;\{\omega_i\}) \,, \\
&\hspace{-5.05cm}\vec{C}^{(2)}_{v,\,\mathcal{P} \mathcal{B} \lambda_1 \lambda_2 \lambda_3 (\lambda_{45}) [\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_4,\omega_5)
{\nonumber} \\
&= \begin{pmatrix*}[r] -1 & 0 \\ 0 & 1\end{pmatrix*} \vec{C}^{(2)}_{v\,\mathcal{P} \mathcal{B} \lambda_1 \lambda_2 \lambda_3 (-\lambda_{45}) [\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_5,\omega_4)
{\nonumber} \,, \\
&\hspace{-5.05cm}\vec{C}^{(2)}_{v,\,\mathcal{P} \mathcal{B} \lambda_1 \lambda_2 \lambda_3 (\lambda_{45}) [\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_4,\omega_5) =\vec{C}^{(2)}_{v\,\mathcal{P} \mathcal{B} \lambda_1 \lambda_2 \lambda_3 (-\lambda_{45}) [\lambda_{\mathcal{P}}]}(n,{\bf n};\omega_1,\omega_2,\omega_3;\omega_5,\omega_4)
{\nonumber} \,.
\end{align}
As was the case for the operators involving three collinear gluon fields discussed in \Sec{sec:sub}, the charge conjugation relations of \Eq{eq:Pchargerel1} imply that to all orders in $\alpha_s$ only the Wilson coefficients for the color structure $d^{abc}$ are non-zero for the vector current, whereas for the axial current, only the Wilson coefficients corresponding to the color structure $if^{abc}$ are non-zero. These statements remain true under renormalization group evolution.
\subsubsection{Ultrasoft Insertions}\label{sec:soft_basis}
At $\mathcal{O}(\lambda^2)$ operators first appear which involve a single $B_{us(i)\lambda}^{a}$ with $\lambda=\pm,0$ (for example $B_{us(i)\lambda}^a J_{n{\bf n}\lambda'}^{{\bar \alpha}\beta}$) or an insertion of a ultrasoft derivative (for example, $\{ \partial_{us(i)0} J_{n{\bf n}\lambda'}^{{\bar \alpha}\beta} \}$). There are no contributions involving the ultrasoft quark current building blocks, like $J_{i(us)\lambda}^{{\bar \alpha}\beta}$. Even though these mixed ultrasoft-collinear currents have the correct power counting, they do not involve the collinear fields that are needed to conserve the large momentum flow in the hard scattering processes being considered.
Before listing the basis of operators, it is worth emphasizing the distinction between the treatment of label and residual $\perp$ momentum. In SCET$_\text{I}$ ultrasoft fields do not carry label momenta. Because only the collinear sectors carry label momentum, we are able to choose the collinear sectors back-to-back, with zero total $\perp$ momentum in each collinear sector. However, for the residual components of the momentum, it is inconsistent to simultaneously choose $\bar n=(1,-\vec n)$, and to set the $\perp$ component of the residual momentum in both sectors to zero. This is because the ultrasoft fields also carry $\mathcal{O}(\lambda^2)$ residual momentum, which can cause the jet direction to recoil by this small amount. Furthermore, because ultrasoft fields carry residual momentum, we cannot, for example, say that the two collinear sectors carry equal and opposite residual momenta, and therefore we cannot in general relate ultrasoft derivatives acting on one sector to ultrasoft derivatives acting on another sector to reduce the basis. Ultrasoft derivatives acting on both sectors must therefore be included in the basis.
When constructing a basis of operators involving ultrasoft gluons, different choices can be made due to the fact that the ultrasoft gluons are not naturally associated with a given lightcone direction. This corresponds to a choice of which light like vector is used to define the $\mathcal{B}_{us(n_i)}$ field of \Eq{eq:soft_gluon}. To guide our choice, we will always choose to work in a basis where ultrasoft derivatives acting on ultrasoft Wilson lines are absorbed into $\mathcal{B}_{us}$ fields, and do not appear explicitly in the operator. As an example, consider the pre-BPS operators
\begin{align}
O^\mu_1=\bar \chi_{\bar n} \overrightarrow D_{us} \chi_n\,, \qquad O^\mu_2=\bar \chi_{\bar n} \overleftarrow D_{us} \chi_n\,,
\end{align}
where we have not made the contraction of the $\mu$ index explicit.
Performing the BPS field redefinition, we obtain
\begin{align}
O^\mu_{1\text{BPS}}=\bar \chi_{\bar n}Y_{\bar n}^\dagger \overrightarrow D_{us} Y_n \chi_n\,, \qquad O^\mu_{2\text{BPS}}=\bar \chi_{\bar n} Y_{\bar n}^\dagger \overleftarrow D_{us} Y_n \chi_n
\end{align}
To absorb all ultrasoft derivatives acting on Wilson lines into $\mathcal{B}_{us}$ fields, we can rearrange the Wilson lines in the operators as
\begin{align}
O^\mu_{1\text{BPS}}=\bar \chi_{\bar n}Y_{\bar n}^\dagger Y_n (Y_n^\dagger \overrightarrow D_{us} Y_n) \chi_n\,, \qquad O^\mu_{2\text{BPS}}=\bar \chi_{\bar n} (Y_{\bar n}^\dagger \overleftarrow D_{us}Y_{\bar n}) Y_{\bar n}^\dagger Y_n \chi_n
\end{align}
Using the definition of the ultrasoft gluon field, \Eq{eq:soft_gluon}, we see that this can be written entirely in terms of $\partial_{us}$ operators acting on collinear fields, as well as the gauge invariant ultrasoft gluon fields $\mathcal{B}_{us(n)}$ and $\mathcal{B}_{us(\bar n)}$. In this way of organizing the basis, ultrasoft gluon fields defined using both $n$ and $\bar n$ are required. It should be clear from this example that it is also possible to work entirely with only $\mathcal{B}_{us(n)}$ or $\mathcal{B}_{us(\bar n)}$. However, in this case we see that we would have ultrasoft derivatives in the operators acting on dangling ultrasoft Wilson lines. To avoid this, and to make our basis more symmetric, we choose to work with both $\mathcal{B}_{us(n)}$ or $\mathcal{B}_{us(\bar n)}$.
For the operators involving one ultrasoft gluon and two collinear quarks, we have the basis
\begin{align}
& \boldsymbol{g_{us}(q)_n (\bar q)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_qq_soft}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:soft_insert_basis}
&O_{\mathcal{B}(us(n))-:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(n)-}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\mathcal{B}(us(n))+:(-;\pm)}^{(2)a\,{\bar \alpha}\beta}
=\mathcal{B}_{us(n)+}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, \\
&O_{\mathcal{B}(us(n))0:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(n)0}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\mathcal{B}(us(n))0:(-;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(n)0}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, {\nonumber}
\end{alignat}
with the unique color structure
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\,a\, \alpha{\bar \beta}} = \left ( T^a Y^\dagger_{n} Y_{{\bf n}} \right )_{\alpha \bar\beta}
\,,\end{equation}
and
\begin{alignat}{2} \label{eq:soft_insert_basis2}
&O_{\mathcal{B}(us(\bar n))+:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(\bar n)+}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\mathcal{B}(us(\bar n))-:(-;\pm)}^{(2)a\,{\bar \alpha}\beta}
=\mathcal{B}_{us(\bar n)-}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, \\
&O_{\mathcal{B}(us(\bar n))0:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(\bar n)0}^a\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\, J_{e\pm }
\,,\qquad &
&O_{\mathcal{B}(us(\bar n))0:(-;\pm)}^{(2)a\,{\bar \alpha}\beta}
= \mathcal{B}_{us(\bar n)0}^a \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\, J_{e\pm }
\,, {\nonumber}
\end{alignat}
with the unique color structure
\begin{equation}
\bar{T}_{\mathrm{BPS}}^{\,a\, \alpha{\bar \beta}} = \left ( Y^\dagger_{n} Y_{{\bf n}} T^a \right )_{\alpha \bar\beta}
\,.\end{equation}
The helicity selection rules act different for the two projections of the $\mathcal{B}_{us}$ fields due to the different definition of helicity in the two cases.
The Wilson coefficients of the operators that include $\mathcal{B}_{us(n)0}$ can be related to the Wilson coefficients of the leading power operators using RPI symmetry (see \cite{Larkoski:2014bxa}). In particular, we have
\begin{align} \label{eq:usRPIrelation}
C^{(2)}_{\mathcal{B}(us)0:(\lambda_1,\pm)}&=-\frac{\partial C^{(0)}_{(\lambda_1;\pm)} }{\partial \omega_1}
\,,
\end{align}
where $C^{(0)}_{(\lambda_1;\pm)}$ is the Wilson coefficient for the leading power dijet operator of \sec{LP}. As we will show in \Sec{sec:matching}, the leading power Wilson coefficients for the case of back to back jets are independent of $\omega_1$ and $\omega_2$, at tree level, so that this Wilson coefficient vanishes at tree level. We will also show explicitly that they do not arise in the tree level matching calculation in \Sec{sec:match_soft}. However, we will also show that the Wilson coefficient is non-vanishing at $\mathcal{O}(\alpha_s)$.
The operators involving $\mathcal{B}_{us(n)\pm}^a$ do not seem to be related to the leading power Wilson coefficient. In \cite{Larkoski:2014bxa} it was shown that in the general case of $N$ jets, certain subleading operators involving $\mathcal{B}_{us(n)\pm}^a$ are generated by the RPI expansion of the leading power operator. However, these particular operators vanish for the case of back to back jets that we consider here. Interestingly, as we will show below, only the operators $O_{\mathcal{B}(us)0:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}$ and $O_{\mathcal{B}(us)\bar 0:(+;\pm)}^{(2)a\,{\bar \alpha}\beta}$ can contribute to the dijet cross section at $\mathcal{O}(\lambda^2)$, and therefore, the new Wilson coefficients (which are not related by RPI) do not contribute at this order.
We also have operators involving two collinear quark fields and a single ultrasoft derivative. In writing the basis, we can use the fact that quark equations of motion can be used to rewrite $in\cdot \partial \chi_n$ and $i\bar n \cdot \partial \chi_{\bar n}$ in terms of purely collinear operators. Therefore, these combinations of derivatives do not need to be included in our basis. A basis of derivative operators is then given by
\begin{align}
& \boldsymbol{\partial_{us} (q)_n (\bar q)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_qq_deriv}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2} \label{eq:soft_derivative_basis}
&O_{\partial(us(n))-:(+;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)-}\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\}\, J_{e\pm }
\,,\qquad &
&O_{\partial(us(n))+:(-;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)+} \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\}\, J_{e\pm }\,,
{\nonumber} \\
&O_{\partial(us(n))0:(+;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)0}\, J_{n\bar n\,+}^{{\bar \alpha}\beta}\}\, J_{e\pm }
\,,\qquad &
&O_{\partial(us(n))0:(-;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{\partial_{us(n)0} \, J_{n\bar n\, -}^{{\bar \alpha}\beta}\}\, J_{e\pm }\,,
{\nonumber} \\
&O_{\partial^\dagger(us(\bar n))+:(+;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\,+}^{{\bar \alpha}\beta}\, (i\partial_{us(\bar n)+})^\dagger\}\, J_{e\pm }
\,,\qquad &
&O_{\partial^\dagger(us(\bar n))-:(-;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\, -}^{{\bar \alpha}\beta}\, (i\partial_{us(\bar n)-})^\dagger\}\, J_{e\pm }\,,
{\nonumber} \\
&O_{\partial^\dagger(us(\bar n))\bar 0:(+;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\,+}^{{\bar \alpha}\beta}\, (i\partial_{us(\bar n)\bar 0})^\dagger\}\, J_{e\pm }
\,,\qquad &
&O_{\partial^\dagger(us(\bar n))\bar 0:(-;\pm)}^{(2)\,{\bar \alpha}\beta}
= \{ J_{n\bar n\, -}^{{\bar \alpha}\beta}\,(i\partial_{us(\bar n)\bar 0})^\dagger\}\, J_{e\pm }
\,.
\end{alignat}
The helicity decomposition for these ultrasoft operators is more cumbersome due to the fact that not only the $\pm$ helicities appear, and for this particular case, it is perhaps simpler to use the more traditional operator basis, in contrast to the case with multiple collinear fields.
The color structure of these operators is exactly the same as for the leading power operator given in \eq{leading_color}. The Wilson coefficients of the operators that include a $\partial_{us(n)0}$ or $\partial_{us(n)\bar 0}$ are related via RPI to the Wilson coefficients of the leading power operator by
\begin{align} \label{eq:uspartialRPIrelation}
C^{(2)}_{\partial(us)0:(\lambda_1,\pm)}=-\frac{\partial C^{(0)}_{(\lambda_1;\pm)} }{\partial \omega_1} \,, \qquad C^{(2)}_{\partial(us)\bar 0:(\lambda_1,\pm)}=-\frac{\partial C^{(0)}_{(\lambda_1;\pm)} }{\partial \omega_2}
\, .
\end{align}
This is true also of the operators where the derivatives act on the Wilson lines, as these arise only through the BPS field redefinition of the same operator.
As we will show in \Sec{sec:matching}, the leading power Wilson coefficients for the case of back to back jets are independent of $\omega_1$ and $\omega_2$ at tree level, so that these Wilson coefficients in fact vanish at the lowest order in the matching. We will also show this explicitly in the matching calculation in \Sec{sec:match_soft}. It is also interesting to mention the physical interpretation of the vanishing of these contributions. As was discussed in \cite{Larkoski:2014bxa} these derivative terms can be interpreted as the orbital angular momentum contribution to the tree level Low-Burnett-Kroll (LBK) theorem \cite{Low:1958sn,Burnett:1967km}, which vanishes for back to back jets.
In the case of ultrasoft derivative insertions, the operators that include a $\partial_{us(n)\pm}$ are related by RPI to $\mathcal{O}(\lambda)$ operators that involve the insertion of $\mathcal{P}_{\pm}$ into the leading power operators of \Eq{eq:LP_basis}. Indeed, RPI implies that the label momentum and derivative operator must always appear in the combination $\mathcal{P}^\mu+i\partial^\mu$. Therefore the Wilson coefficients of these operators are equal. However, we have chosen to work in the center of mass frame, where insertions of $\mathcal{P}_{\pm}$ into the leading power operator vanish, and so such operators do not appear explicitly in our basis. Therefore we include the Wilson coefficients of these operators in our basis.
We also have operators involving two collinear gluons and a single ultrasoft gluon field. Since their Wilson coefficients start at one-loop order these are of limited phenomenological relevance, but are included as a further example of our approach. The basis of such operators is given by
\begin{align}
& \boldsymbol{g_{us}(g)_n (g)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_gg_soft}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:eegggus}
&O_{(us(n))+:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)+}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(n))+:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)+}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,, \\
&O_{(us(n))-:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)-}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us)-:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)-}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,, {\nonumber} \\
&O_{(us(n))+:-+(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)+}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(n))-:+-(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)-}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,,{\nonumber}
\end{alignat}
with the two dimensional basis of color structures,\footnote{In order to see how the Wilson line structure in \Eq{eq:Z2g_colorus} arises, we look at the object $D_{us}^{ab} \mathcal{B}_{n}^c \mathcal{B}_{{\bf n}}^d$ pre-BPS field redefinitions. This object must be contracted with a tensor to make it a singlet under ultrasoft gauge transformations. Each of these resulting forms can be mapped onto the color structures of \Eq{eq:Z2g_colorus} after performing the BPS field redefinition}
\begin{equation} \label{eq:Z2g_colorus}
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc} \\
d^{abd}\, \big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{dc}
\end{pmatrix}^T
\,,
\end{equation}
and
\begin{alignat}{2}\label{eq:eegggus2}
&O_{(us(\bar n))-:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)-}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(\bar n))-:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)-}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,, \\
&O_{(us(\bar n))+:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)+}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(\bar n))+:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)+}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,, {\nonumber} \\
&O_{(us(\bar n))-:-+(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)-}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us)+:+-(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)+}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,,{\nonumber}
\end{alignat}
with the basis of color structures
\begin{equation} \label{eq:Z2g_colorus2}
\bar{T}_{\mathrm{BPS}}^{abc} =
\begin{pmatrix}
i f^{acd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{db} \\
d^{acd}\, \big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{db}
\end{pmatrix}^T
\,.
\end{equation}
Here we have only included the $\bar{T}_{\mathrm{BPS}}^{abc}$ version of the color structure here because the $\mathcal{B}_{us(n)\lambda}^a$ are generated by BPS field redefiniton. When constructing this basis we have used the angular momentum constraints discussed in \Sec{sec:ang_cons} to eliminate the other two helicity combinations in each case.
In \eq{eegggus} we have not included the operators
\begin{align}
&O_{(us(n))0:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)0}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(n))0:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(n)0}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }\,,
{\nonumber} \\
&O_{(us(\bar n))0:++(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)0}^a\, \mathcal{B}_{n+}^b\, \mathcal{B}_{\bar n+}^c\, J_{e\pm }
\,, \qquad && O_{(us(\bar n))0:--(\pm)}^{(2)abc}
= \mathcal{B}_{us(\bar n)0}^a\, \mathcal{B}_{n-}^b\, \mathcal{B}_{\bar n-}^c\, J_{e\pm }
\,.
\end{align}
The coefficients of these operators are related by RPI to the derivative of the leading power operators for $e\bar e gg$, and therefore also vanish by Yang's theorem. The Wilson coefficients of the operators in \eq{eegggus} are not constrained by RPI considerations.
We can also consider operators with an insertion of $\partial_{us(n)}$ with two collinear gluons in different collinear sectors. As for the case of ultrasoft derivative insertions into the quark operators, the gluon equations of motion allow us to eliminate the operators $in\cdot \partial \mathcal{B}_{n\perp}$ and $i\bar n\cdot \partial \mathcal{B}_{\bar n\perp}$. However, these operators already vanish by Yang's theorem, as they are related by RPI to the leading power operators for $e\bar e gg$. A basis of helicity operators is then given by
\begin{align}
& \boldsymbol{\partial_{us}(g)_n (g)_{{\bf n}}:}{\vcenter{\includegraphics[width=0.18\columnwidth]{figures/Leading_gg_deriv}}} {\nonumber}
\end{align}
\vspace{-0.4cm}
\begin{alignat}{2}\label{eq:eedggus}
&O_{\partial \mathcal{B} (us(n))+:++(\pm)}^{(2)ab}
= \mathcal{B}_{n+}^a\,\left[ \partial_{us(n)+} \mathcal{B}_{\bar n+}^b \right]\, J_{e\pm }
\,, ~\, && O_{\partial \mathcal{B} (us(n))+:--(\pm)}^{(2)ab}
= \mathcal{B}_{n-}^a \, \left[ \partial_{us(n)+} \mathcal{B}_{\bar n-}^b\right]\, J_{e\pm }
\,,
{\nonumber} \\
&O_{\partial \mathcal{B} (us(n))-:++(\pm)}^{(2)ab}
= \mathcal{B}_{n+}^a \, \left[ \partial_{us(n)-} \mathcal{B}_{\bar n+}^b\right] \, J_{e\pm }
\,, ~\, && O_{\partial \mathcal{B} (us(n))-:--(\pm)}^{(2)ab}
= \mathcal{B}_{n-}^a \left[\partial_{us(n)-} \mathcal{B}_{\bar n-}^b\right]\, J_{e\pm }
\,,
{\nonumber} \\
&O_{\partial \mathcal{B} (us(n))+:-+(\pm)}^{(2)ab}
= \mathcal{B}_{n-}^a \, \left[ \partial_{us(n)+} \mathcal{B}_{\bar n+}^b \right]\, J_{e\pm }
\,, \, ~&& O_{\partial \mathcal{B} (us(n))-:+-(\pm)}^{(2)ab}
= \mathcal{B}_{n+}^a \, \left [\partial_{us(n)-}\, \mathcal{B}_{\bar n-}^b\right]\, J_{e\pm }
\,,
\end{alignat}
with the basis of color structures
\begin{equation} \label{eq:Z2gd_colorus}
\bar{T}_{\mathrm{BPS}}^{ab} =
\big({\cal Y}_n^T {\cal Y}_{\bar n}\big)^{ab}
\,,
\end{equation}
and
\begin{alignat}{2}\label{eq:eedggus2}
&O_{\partial \mathcal{B} (us(\bar n))-:++(\pm)}^{(2)ab}
= \left[ \partial_{us(\bar n)-}\, \mathcal{B}_{n+}^a\right]\, \mathcal{B}_{\bar n+}^b\, J_{e\pm }
\,, \,~ && O_{\partial \mathcal{B} (us(\bar n))-:--(\pm)}^{(2)ab}
= \left [ \partial_{us(\bar n)-}\, \mathcal{B}_{n-}^a \right] \, \mathcal{B}_{\bar n-}^b\, J_{e\pm }
\,,
{\nonumber} \\
&O_{\partial \mathcal{B} (us(\bar n))+:++(\pm)}^{(2)ab}
= \left [ \partial_{us(\bar n)+}\, \mathcal{B}_{n+}^a \right] \, \mathcal{B}_{\bar n+}^b\, J_{e\pm }
\,, ~\, && O_{\partial \mathcal{B} (us(\bar n))+:--(\pm)}^{(2)ab}
= \left [ \partial_{us(\bar n)+}\, \mathcal{B}_{n-}^a \right] \, \mathcal{B}_{\bar n-}^b\, J_{e\pm }
\,,
{\nonumber} \\
&O_{\partial \mathcal{B} (us(\bar n))-:-+(\pm)}^{(2)ab}
= \left [\partial_{us(\bar n)-}\, \mathcal{B}_{n-}^a \right] \, \mathcal{B}_{\bar n+}^b\, J_{e\pm }
\,, ~\, && O_{\partial \mathcal{B} (us(\bar n))+:+-(\pm)}^{(2)ab}
= \left [ \partial_{us(\bar n)+}\, \mathcal{B}_{n+}^a \right] \, \mathcal{B}_{\bar n-}^b\, J_{e\pm }
\,,
\end{alignat}
with the basis of color structures
\begin{equation} \label{eq:Z2gd_colorus2}
\bar{T}_{\mathrm{BPS}}^{ab} =
\big({\cal Y}_{\bar n}^T {\cal Y}_{n}\big)^{ab}
\,.
\end{equation}
We have not included any operator with $\partial_{us(n)0}$ or $\partial_{us(n)\bar 0}$ acting on two collinear gluons, as these will all have Wilson coefficients that are related to the coefficient of the two gluon operator, which vanishes due to Yang's theorem.
\subsection{Cross Section Contributions}\label{sec:ee_discuss}
While the basis of hard scattering operators presented in this section is quite large, and we have focused on providing a complete basis to allow for an understanding of all possible contributions, many of these operators will not contribute to a calculation of a particular cross section. In this section we will consider the case of event shapes in $e^+e^-\to$ dijets, and discuss how symmetry arguments can be used to show which operators can contribute to the cross section up to $\mathcal{O}(\lambda^2)$. In \Sec{sec:ee_lambda}, we begin by proving that the hard scattering operators do not generate a contribution to the cross section at $\mathcal{O}(\lambda)$, and then in \Sec{sec:ee_lambda2}, we discuss which operators contribute to the cross section at $\mathcal{O}(\lambda^2)$, and the particular form of their contribution, by listing the operator content of the resulting jet and soft functions. A summary of which operators contribute is given in \Tab{tab:summary}.
\subsubsection{Vanishing at $\mathcal{O}(\lambda)$}\label{sec:ee_lambda}
For $e^+e^-\to$ dijets event shapes described by SCET$_\text{I}$, the leading $\mathcal{O}(\lambda)$ power corrections vanish \cite{Beneke:2003pa,Lee:2004ja,Freedman:2013vya}. This is expected because fixed order calculations indicate the leading correction should scale as $e$, while an $\mathcal{O}(\lambda)$ power correction would scale as $\sqrt{e}$ for our power counting. In this section, we use our formalism to show explicitly that this is the case for contributions from the hard scattering operators. Similar arguments can also be used to show that Lagrangian contributions vanish. The $\mathcal{O}(\lambda)$ observable expansion terms also vanish, as discussed in \App{app:meas}.
While we will not discuss the factorization of the cross section in detail, the contribution of the hard scattering operators to the cross section at $\mathcal{O}(\lambda)$ can be written
{\begin{small}
\begin{align}\label{eq:xsec_lam}
&\frac{d\sigma}{d\tau}^{(1)} \supset N \sum_{X,i} \tilde \delta^{(4)}_q \bra{0} C_i^{(1)} \tilde{O}_i^{(1)}(0) \ket{X}\bra{X} C^{(0)} \tilde{O}^{(0)}(0) \ket{0} \delta\big( \tau - \tau^{(0)}(X) \big) +\text{h.c.}
\,.
\end{align}
\end{small}}
Here $N$ is a normalization factor. We use the shorthand notation $\tilde \delta^{(4)}_q=(2\pi)^4\delta^4(q-p_X)$ for the momentum conserving delta function. The summation over all final states, $X$, includes phase space integrations. Here $L$ denotes the $e^+e^-$ leptonic initial state. The measurement of the observable is enforced by $ \delta\big( e - e(X) \big) $, where $e(X)$, returns the value of the observable $e$ as measured on the final state $X$.
From the expression for the $\mathcal{O}(\lambda)$ power correction to the cross section in \Eq{eq:xsec_lam}, we see that the only contributions from hard scattering operators arise from matrix elements of an $\mathcal{O}(\lambda)$ operator with an $\mathcal{O}(\lambda^0)$ operator. When determining whether or not the insertion of a given operator vanishes, we can make arguments based on fermion number conservation or angular momentum conservation either before or after factorization. Before factorization, the matrix elements in the cross section given in \Eq{eq:Z1_basis} that contribute at $\mathcal{O}(\lambda)$ with the insertion of an $\mathcal{O}(\lambda)$ operator can be written as
\begin{align} \label{eq:subleadingmatrix_general}
\langle 0| O^{(0)}(x) \widehat \mathcal{M}^{(0)} O^{(1)}(0) |0 \rangle\,.
\end{align}
Here the sum over the complete set of states $|X\rangle$ has been performed using the measurement operator $\widehat \mathcal{M}^{(0)}$, where $\widehat \mathcal{M}^{(0)} |X\rangle = \delta(e - e^{(0)}(X)) |X\rangle$. As we are taking a vacuum matrix element, we must have that $O^{(0)}(x) O^{(1)}(0)$ conserves fermion number and angular momentum. If this matrix element does not vanish, we can move to the factorized state, where we split the operators into components in the $n$, ${\bf n}$ and ultrasoft sectors. This will give us factorized matrix elements that become our jet and soft functions,
\begin{align} \label{eq:subleadingmatrix_factorized}
\langle 0| O^{(0)}_n(x) \widehat \mathcal{M}_n^{(0)} O_n^{(1)}(0) |0 \rangle \langle 0| O^{(0)}_{{\bf n}}(x) \widehat \mathcal{M}_n^{(0)} O_{{\bf n}}^{(1)}(0) |0 \rangle \langle 0| O^{(0)}_{us}(x) \widehat \mathcal{M}_{us}^{(0)} O_{us}^{(1)}(0) |0 \rangle\,.
\end{align}
In this form, it is clear that each sector must exhibit both fermion number and angular momentum conservation, so we can make these arguments at the level of the factorized matrix elements, providing an even stronger constraint. In other words, if we examine the field content in each sector, we must have the same number of quarks and anti-quarks and must conserve angular momentum. As shown in \Tab{tab:summary}, there are only two operator structures appearing at $\mathcal{O}(\lambda)$, and a single operator at $\mathcal{O}(\lambda^0)$ , so we can consider each of the possible contributions in turn.
$ \boldsymbol{O^{(1)}_{ggg} O^{(0)}_{qq}}$ \textbf{ vanish:}
We first consider the $\mathcal{O}(\lambda)$ hard scattering operators involving three collinear gluons, given in \Eq{eq:eeggg}. After factorization of this matrix element one obtains a vacuum matrix element involving a single quark field in each collinear sector, coming from the leading power operator. The leading order Lagrangian separately conserves fermion number in each collinear sector, and therefore this contribution vanishes.
$ \boldsymbol{O^{(1)}_{qqg} O^{(0)}_{qq}}$ \textbf{ vanish:}
Next we consider the contribution from the $\mathcal{O}(\lambda)$ hard scattering operators involving two collinear quarks, and one collinear gluon. As for the three gluon operator, fermion number conservation immediately eliminates any possible contribution from the operators of \Eq{eq:Z1_basis_diff}, where the collinear quarks are both in the same sector. To eliminate the $\mathcal{O}(\lambda)$ contribution from the operators of \Eq{eq:Z1_basis}, we can use symmetry arguments, similar to those in \Sec{sec:ang_cons}. For the operator of \Eq{eq:Z1_basis}, the matrix elements entering the factorized expression for the cross section are of the form
\begin{align} \label{eq:subleadingmatrix}
\langle 0| (J^{\bar \alpha \beta}_{n\bar n \lambda_1}(x))^\dagger \widehat \mathcal{M}^{(0)} J^{\bar \delta \gamma}_{n\bar n \lambda_2}(0) \mathcal{B}^a_{(n,\bar n) \lambda_3}(0) |0 \rangle\,.
\end{align}
The $\lambda_i=\pm$ denote arbitrary helicities, and the $(n,\bar n)$ subscript on the gluon denotes that it can be associated with either collinear sector. Since RPI has been used to choose the axes of the collinear sectors as back-to-back, all helicities are defined with respect to a common $\hat{n}$ axis. The SCET Lagrangian preserves rotational invariance about the $\hat{n}$ axis, which implies that this matrix element vanishes since with an odd number of $\pm$ helicities it can not transform as a scalar.
An identical argument follows for the operator of \Eq{eq:Z1_basis_flip}.
\subsubsection{Relevant Hard Scattering Operators at $\mathcal{O}(\lambda^2)$}\label{sec:ee_lambda2}
Having shown that there are no contributions to the dijet cross section at $\mathcal{O}(\lambda)$, we now discuss contributions at $\mathcal{O}(\lambda^2)$. To see which can contribute, we focus on contributions arising from our basis of hard scattering operators, although we will also briefly mention contributions from subleading Lagrangian insertions. Due to the power counting $\lambda \sim \sqrt{e}$, the ${\cal O}(\lambda^2)$ power corrections correspond to ${\cal O}(e)$ power corrections, and will not in general vanish. While there are large number of $\mathcal{O}(\lambda^2)$ hard scattering operators given in \Tab{tab:summary}, we can use similar arguments to those of \Sec{sec:ee_lambda} to severely restrict the number of operators that contribute to the dijet cross section at $\mathcal{O}(\lambda^2)$. The resulting set of operators are indicated in \Tab{tab:summary}, and \Tab{tab:fact_func} shows which products of hard scattering operators contribute to the factorized cross section and the schematic form of the corresponding hard, jet and soft functions.
Despite the fact that we are working at subleading power, in many cases the jet and soft functions which appear in the factorization are identical to those at leading power, with only several new power suppressed functions appearing, as can be seen in \Tab{tab:fact_func}. For the case of the soft functions this simplification arises due to color coherence, allowing a simplification to the Wilson lines in the soft functions that appear. For quark-quark and gluon-gluon color channels the leading power soft functions are
\begin{align}
S_q^{(0)}=\frac{1}{N_c} {\rm tr} \langle 0 | Y^\dagger_{\bar n} Y_n \widehat{\mathcal{M}}^{(0)}Y_n^\dagger Y_{\bar n}|0\rangle\,, \qquad S_g^{(0)}=\frac{1}{(N_c^2-1)} {\rm tr} \langle 0 | \mathcal{Y}^T_{\bar n} \mathcal{Y}_n \widehat{\mathcal{M}}^{(0)}\mathcal{Y}_n^T \mathcal{Y}_{\bar n}|0\rangle\,,
\end{align}
and depend on the kinematic variables probed by the measurement operator $\widehat{\mathcal{M}}^{(0)}$.
For the jet functions, this simplification occurs since the power correction is often restricted to a single collinear sector. The other collinear sector is then described by the leading power jet functions for quarks and gluons
\begin{align}
\delta^{\alpha\bar \beta} \Big( \frac{\Sl{n}}{2} \Big)^{\!ss'\!} J_q^{(0)}
&= \int \!\!\frac{dx^-}{|\omega|}\: e^{\frac{i}{2} \ell^+ x^-}
\Big\langle 0\Big|\, \chi_{n}^{s\alpha} \big(x^- \text{\small $\frac{n}{2}$}\big) \,\hat{\delta}\, \bar \chi_{n,\omega}^{s'\bar\beta}(0)
\,\Big| 0 \Big\rangle
\,, \\
\delta^{ab} g^{\mu\nu}_\perp J_g^{(0)}
&=-\omega\! \int \!\!\frac{dx^-}{|\omega|}\: e^{\frac{i}{2} \ell^+ x^-} \Big\langle 0 \Big|\, \mathcal{B}^{\mu a}_{\perp} \big(x^- \text{\small $\frac{n}{2}$}\big)\, \hat{\delta}\, \mathcal{B}^{\nu b}_{\perp,\omega}(0) \,\Big|0\Big\rangle
\,. {\nonumber}
\end{align}
The form of the leading power measurement function $\hat{\delta}$ appearing in these jet functions will depend on the precise factorization theorem being treated. Here we assume an SCET$_{\text{I}}$ type measurement that does not fix the $\mathcal{P}_\perp$ of the measured particle. Often the jet functions are inclusive in which case $\hat{\delta}=(2\pi)^2 \delta^2(\omega_\perp+\mathcal{P}_\perp^2)$, giving functions of a single invariant mass momentum variable, $J_{q}^{(0)}(\omega \ell^+-\vec \omega_\perp^{\,2})$ and $J_{g}^{(0)}(\omega \ell^+-\vec \omega_\perp^{\,2})$.
Contributions to the cross section at $\mathcal{O}(\lambda^2)$ whose power suppression arises solely from hard scattering operators, take the form either of a product of two $\mathcal{O}(\lambda)$ operators or as a product of an $\mathcal{O}(\lambda^2)$ operator and an $\mathcal{O}(\lambda^0)$ operator:
{\begin{small}
\begin{align}\label{eq:xsec_lam2}
&\frac{d\sigma}{d\tau}^{(2)} \supset N \sum_{X,i} \tilde \delta^{(4)}_q \bra{0} C_i^{(2)} \tilde{O}_i^{(2)}(0) \ket{X}\bra{X} C^{(0)} \tilde{O}^{(0)}(0) \ket{0} \delta\big( \tau - \tau^{(0)}(X) \big) +\text{h.c.}{\nonumber} \\
&+ N \sum_{X,i,j} \tilde \delta^{(4)}_q \bra{0} C_i^{(1)} \tilde{O}_i^{(1)}(0) \ket{X}\bra{X} C_j^{(1)} \tilde{O}_j^{(1)}(0) \ket{0} \delta\big( \tau - \tau^{(0)}(X) \big)+\text{h.c.} \,.
\end{align}
\end{small}}
We first consider the contributions from products of $\mathcal{O}(\lambda)$ hard scattering operators, where we have several categories that could possibly contribute.
\noindent \vbox{$ \boldsymbol{O^{(1)}_{ggg} O^{(1)}_{ggg}}$ \textbf{ contribute:}
After factorization, the contribution to the $\mathcal{O}(\lambda^2)$ cross section from the product of two $\mathcal{O}(\lambda)$ three gluon operators of \Eq{eq:eeggg} gives rise to jet functions involving either two or four collinear gluon fields. The schematic factorization is given by $H_{q1}^{(0)} J_g^{(0)} J_{gg}^{(2)} S_g^{(0)}$, shown in \Tab{tab:fact_func}, where the jet functions $J_{gg}^{(2)}$ involving four $\mathcal{B}$ fields have two different color contractions. Rotational invariance arguments, similar to those presented in \Sec{sec:ee_lambda}, but applied after factorization into separate matrix elements for the $n$ and $\bar n$ sectors, imply that jet functions involving three collinear gluon building block fields vanish.
}
\noindent \vbox{
$ \boldsymbol{O^{(1)}_{qqg} O^{(1)}_{qqg}}$ \textbf{ contribute:}
Next consider the contribution to the $\mathcal{O}(\lambda^2)$ cross section from the square of the operators involving two collinear quarks, and a collinear gluon, a basis of which were given in \Eqs{eq:Z1_basis}{eq:Z1_basis_flip} and \Eq{eq:Z1_basis_diff}. These give rise to the factorized contributions $H_{q2}^{(0)} J_g^{(0)} J_{qq}^{(2)} S_g^{(0)}$, and $H_{q3}^{(0)} J_q^{(0)} J_{qgg}^{(2)} S_{q}^{(0)}$ in \Tab{tab:fact_func}. Here the factorization theorems involve subleading jet functions $J_{qq}^{(2)}$ with four collinear quarks (one color contraction), or $J_{qgg}^{(2)}$ with two collinear quarks and two collinear gluons (one color contraction). The exact color structure is not displayed, but is simple to obtain from the color structure of the hard scattering operators after BPS field redefinition. Rotational invariance arguments can be used to restrict the particular helicity configurations which give non-vanishing contributions. In both these cases the ultrasoft functions are leading power, and are actually given by the same product of Wilson lines that appear in leading power factorization theorems, as indicated in \Tab{tab:fact_func} with the notation $S_g^{(0)}$ and $S_{q}^{(0)}$.
}
\noindent $ \boldsymbol{O^{(1)}_{ggg} O^{(1)}_{qqg}}$ \textbf{ vanish:}
The contribution from the product of an operator containing three collinear gluons and an operator containing a collinear gluon and two collinear quarks must vanish. If the two quarks are in different sectors, then each jet function will contain only one fermion and will vanish by fermion number conservation. With 2 quarks in the same sector, we have the product of $O_{\bar{n}\lambda_1 ( \lambda_2:\pm)}^{(1)a\,{\bar \alpha}\beta}$ and $O_{\mathcal{B}\lambda_3 \lambda_4 \lambda_5 (\pm)}^{(1)bcd}$, and we can see that one sector will contain three objects of helicity $\pm1$ and thus the jet function in that sector will vanish by the angular momentum arguments considered earlier. So, there is no nonvanishing contribution from an $\mathcal{O}(\lambda)$ three gluon operator and an $\mathcal{O}(\lambda)$ two quark, one gluon operator.
We now look at the contribution of $\mathcal{O}(\lambda^2)$ hard scattering operators to the cross section. Since these operators are at the desired order, they must be combined with our leading operator which has a collinear quark building block in each of the $n$ and ${\bf n}$ directions. Once again, there are several cases to consider.
\noindent $ \boldsymbol{O^{(2)}_{qqqq} O^{(0)}_{qq}}$ \textbf{ contribute:}
First, we consider the inclusion of the four quark operators. Conservation of fermion number within each collinear sector ensures the the contributions to the cross section involving the four quark operators with two quarks within each collinear sector vanish. Therefore, only the four quark operators of \Eq{eq:Z3_basis_qQ}, involving three quarks in one collinear sector, and a single quark in the other collinear sector can give a nonzero result for the cross section at this order. They contribute to the factorized contribution $H_{q7}^{(0)} J_q^{(0)} J_{qq'}^{(2)} S_q^{(0)}$ shown in \Tab{tab:fact_func}, which involves a jet function $J_{qq'}^{(2)}$ with $4$ $\chi_{\bf n}$ (or 4 $\chi_n$) fields (with two independent color contractions). Again, the corresponding soft function is simply the leading power $S_q^{(0)}$ in all cases.
\noindent $ \boldsymbol{O^{(2)}_{gggg} O^{(0)}_{qq}}$ \textbf{ vanish:}
The operators of \Eqs{eq:Z2_basis_gggg_1}{eq:Z2_basis_gggg_2} involving four collinear gluons do not contribute to the cross section at $\mathcal{O}(\lambda^2)$, since when multiplied with the leading power operator, the factorized matrix element would violate fermion number in the $n$ and $\bar n$ sectors.
\noindent $ \boldsymbol{O^{(2)}_{ggg\mathcal{P}} O^{(0)}_{qq}}$ \textbf{ vanish:}
An identical argument also applies to the operators of \Eq{eq:eegggpperp_basis} involving three collinear gluons and a single $\mathcal{P}_\perp$ insertion, which therefore do not contribute to the $\mathcal{O}(\lambda^2)$ cross-section.
{
\renewcommand{\arraystretch}{1.6}
\begin{table}[t!]
\begin{center}
\begin{tabular}{| c | l | c | c | c | c | }
\hline
& Operators & Factorization & Jet $n$ & Jet $\bar n$ & Soft \\
\hline
$\!\!\mathbf{\mathcal{O}(\lambda^0)}\!\!$
&$\! O^{(0)} O^{(0)} $
& $H_q^{(0)} J_q^{(0)} J_q^{(0)} S_q^{(0)}$
& $\bar \chi_n \,\hat\delta\, \chi_n$
& $\bar \chi_{\bf n} \,\hat\delta\, \chi_{\bf n}$
& $Y_{\bf n}^\dagger Y_n \widehat\mathcal{M}^{(0)} Y_n^\dagger Y_{\bf n}$
\\
\hline
$\!\!\mathbf{\mathcal{O}(\lambda^2)}\!\!$
&$\! O_\mathcal{B}^{(1)} O_\mathcal{B}^{(1)} $
& $H_{q1}^{(0)} J_g^{(0)} J_{gg}^{(2)} S_g^{(0)}$
& $\mathcal{B}_n \,\hat\delta\, \mathcal{B}_n$
& $\!\!\mathcal{B}_{\bf n} \mathcal{B}_{\bf n} \hat\delta\, \mathcal{B}_{\bf n} \mathcal{B}_{\bf n}\!\!$
& ${\cal Y}_n^T {\cal Y}_{\bf n} \widehat\mathcal{M}^{(0)} {\cal Y}_{\bf n}^T {\cal Y}_n$
\\
\cline{2-6}
& $\! O_{\bar n}^{(1)} O_{{\bf n}}^{(1)} $
& $H_{q2}^{(0)} J_g^{(0)} J_{qq}^{(2)} S_g^{(0)}$
& $\mathcal{B}_n \,\hat\delta\, \mathcal{B}_n$
& $\bar \chi_{\bf n} \chi_{\bf n} \hat\delta\, \bar \chi_{\bf n} \chi_{\bf n} $
& $\mathcal{Y}_{\bf n}^T \mathcal{Y}_n \widehat\mathcal{M}^{(0)} \mathcal{Y}_n^T \mathcal{Y}_{\bf n} $
\\
\cline{2-6}
&$\! O_{n\bar n1}^{(1)} O_{n\bar n1}^{(1)}$
& $H_{q3}^{(0)} J_q^{(0)} J_{qgg}^{(2)}\, S_q^{(0)}$
& $\bar \chi_n \,\hat\delta\, \chi_n $
& $\!\!\bar \chi_{\bf n} \mathcal{B}_{\bf n} \,\hat\delta\, \mathcal{B}_{\bf n} \chi_{\bf n}\!\! $
& $Y_{\bf n}^\dagger Y_n \widehat\mathcal{M}^{(0)} Y_n^\dagger Y_{\bf n}$
\\
\cline{2-6}
&$\! O^{(0)} O_{\cB1,\cB2}^{(2)} $
& $H_{q4}^{(0)} J_q^{(0)} J_{qgg'}^{(2)} S_{q}^{(0)}$
& $\bar \chi_n \,\hat\delta\, \chi_n $
& $\!\!\bar \chi_{\bf n} \mathcal{B}_{\bf n} \mathcal{B}_{\bf n} \,\hat\delta\, \chi_{\bf n}\!\! $
& $Y_n^\dagger Y_{\bf n} \widehat\mathcal{M}^{(0)} Y_{\bf n}^\dagger Y_n$
\\
\cline{2-6}
&$\! O^{(0)} O_{\cP1\bar n}^{(2)} $
& $H_{q5}^{(0)} J_q^{(0)} J_{qgP}^{(2)} S_q^{(0)}$
& $\bar \chi_n \,\hat\delta\, \chi_n $
& $\!\!\bar \chi_{\bf n} [\mathcal{P}_{\!\perp} \mathcal{B}_{\bf n}] \hat\delta\, \chi_{\bf n}\!\!$
& $Y_n^\dagger Y_{\bf n} \widehat\mathcal{M}^{(0)} Y_{\bf n}^\dagger Y_n$
\\
\cline{2-6}
&$\! O^{(0)} O_{\cB3}^{(2)} $
& $\!\!H_{q6}^{(0)}\! J_{qg}^{(1)} J_{qg}^{(1)} \{ S_{qg}^{(0)}, S_q^{(0)} \}\!\!$
& $\!\!\bar \chi_n \mathcal{B}_n \,\hat\delta\, \chi_n \!\! $
& $\bar \chi_{\bf n} \,\hat\delta\, \mathcal{B}_{\bf n} \chi_{\bf n} $
& $Y_{\bf n}^\dagger Y_n \widehat\mathcal{M}^{(0)} Y_n^\dagger \mathcal{Y}_{\bf n} \mathcal{Y}_n Y_{\bf n}$
\\
\cline{2-6}
&$\! O^{(0)} O_{qQ4,5}^{(2)}$
& $H_{q7}^{(0)} J_q^{(0)} J_{qq'}^{(2)} S_q^{(0)}$
& $\bar \chi_n \,\hat\delta\, \chi_n $
& $\!\!\bar \chi_{\bf n} \chi_{\bf n} \bar \chi_{\bf n} \,\hat\delta\, \chi_{\bf n}\!\!$
& $Y_{\bf n}^\dagger Y_n \widehat\mathcal{M}^{(0)} Y_n^\dagger Y_{\bf n}$
\\
\cline{2-6}
&$\! O^{(0)} O_{\mathcal{B} (us)0}^{(2)} \!\!$
& $H_{q8}^{(0)} J_q^{(0)} J_{q}^{(0)} S_{q\mathcal{B}}^{(2)}$
& $\bar \chi_n\,\hat\delta\, \chi_n $
&$\bar \chi_{\bf n}\,\hat\delta\, \chi_{\bf n} $
& $\!\!\mathcal{B}_{us(n) 0}\, Y_n^\dagger Y_{\bf n} \widehat\mathcal{M}^{(0)} Y_{\bf n}^\dagger Y_n\!\!$
\\
\cline{2-6}
& $\! O^{(0)} O_{\partial(us)0}^{(2)} \!\! $
& $H_{q9}^{(0)} J_q^{(0)} J_{q}^{(0)} S_{q\partial 0}^{(2)}$
& $\bar \chi_n\,\hat\delta\, \chi_n $
&$\bar \chi_{\bf n}\,\hat\delta\, \chi_{\bf n} $
& $\!\!\partial_{us(n) 0}\, Y_n^\dagger Y_{\bf n} \widehat\mathcal{M}^{(0)} Y_{\bf n}^\dagger Y_n\!\!$
\\
\cline{2-6}
& $\! O^{(0)} O_{\partial(us)\bar 0}^{(2)}\!\!$
& $H_{q10}^{(0)} J_q^{(0)} J_{q}^{(0)} S_{q\partial \bar 0}^{(2)}$
& $\bar \chi_n\,\hat\delta\, \chi_n $
&$\bar \chi_{\bf n}\,\hat\delta\, \chi_{\bf n} $
& $\!\!\partial_{us(n) \bar 0}\, Y_n^\dagger Y_{\bf n} \widehat\mathcal{M}^{(0)} Y_{\bf n}^\dagger Y_n\!\!$
\\
\hline
\end{tabular}
\caption{Subleading jet and soft functions arising from products of hard scattering operators in the factorization of dijet event shapes and their field content. Helicity and color structures have been suppressed. We have not included products of operators whose jet and soft functions are identical to those given in the table by charge conjugation or $n\leftrightarrow \bar n$.}
\label{tab:fact_func}
\end{center}
\end{table}
}
\noindent $ \boldsymbol{O^{(2)}_{ggqq} O^{(0)}_{qq}}$ \textbf{ contribute:}
Another source of non-trivial contributions to the cross section at $\mathcal{O}(\lambda^2)$ comes from the hard scattering operators involving two collinear quarks, and two collinear gluons. To have a non-vanishing contribution, fermion number conservation within each collinear sector guarantees that the hard scattering operators must have a single quark building block field in each sector. This restricts us to the operators of \Eqs{eq:eeqqgg_basis1}{eq:eeqqgg_basis2}, as indicated by the check marks in \Tab{tab:summary}. These operators give rise to subleading jet functions involving two $\chi$ fields, and either one or two $\mathcal{B}$ fields ($J_{qg}^{(1)}$ or $J_{qgg}^{(2)}$ respectively), as shown by the factorized contributions $H_{q6}^{(0)} J_{qg}^{(1)} J_{qg}^{(1)} \{S_{qg}^{(0)}, S_{q}^{(0)} \}$ and $H_{q4}^{(0)} J_q^{(0)} J_{qgg'}^{(2)} S_{q}^{(0)}$ in \Tab{tab:fact_func}. There is a unique color structure present in the definition of $J_{qg}^{(1)}$. The $J_{qgg'}^{(2)}$ here appear for two different color contractions and are distinct from the $J_{qgg}^{(2)}$ function appearing in the $H_{q3}^{(0)} J_q^{(0)} J_{qgg}^{(2)} S_{q}^{(0)}$ case due to the location of their measurement function. The new soft function $S_{qg}^{(0)}$ has the Wilson line structure shown by the Soft entry in \Tab{tab:fact_func}, and is generated with a unique color contraction. It only appears in the subleading power cross section even though it has leading power scaling.
\noindent $ \boldsymbol{O^{(2)}_{qqg\mathcal{P}} O^{(0)}_{qq}}$ \textbf{ contribute:}
Similar arguments also apply to the subleading operators involving two collinear quarks, a collinear gluon, and a $\mathcal{P}_\perp$ insertion. In particular, the operators of \Eq{eq:eeqqgpperp_basis_same} do not contribute due to fermion number conservation, while the operators of \Eqs{eq:eeqqgpperp_basis}{eq:eeqqgpperp_basis_flip} do contribute. These give rise to the factorized contribution $H_{q5}^{(0)} J_q^{(0)} J_{qgP}^{(2)} S_q^{(0)}$ in \Tab{tab:fact_func}, as well as an identical function with $n\leftrightarrow \bar n$. There is a unique color contraction for the jet function $J_{qgP}^{(2)}$. Rotational invariance arguments can be used to restrict the particular helicity configurations which give non-vanishing contributions.
\noindent $ \boldsymbol{O^{(2)}_{us} O^{(0)}_{qq}}$ \textbf{ contribute:}
As shown in \Tab{tab:fact_func}, the only operators involving an explicit ultrasoft field insertion which contribute to the dijet cross section at $\mathcal{O}(\lambda^2)$ are those involving the $0$ helicity component of the ultrasoft gluon, $\mathcal{B}_{us(n) 0}$, or the ultrasoft partial operators $\partial_{us(n) 0}$, $\partial_{us(n) \bar 0}$. These operators give rise to the only subleading power soft functions arising from hard scattering operators, as seen in \Tab{tab:fact_func}. It is also interesting to note that for both these contributions, the hard function is fully determined by the RPI relations, as was argued in \Sec{sec:soft_basis}. The operators of \Eq{eq:soft_insert_basis} involving the $h=\pm$ components of the ultrasoft gluon field, along with the collinear quark current, vanish when multiplied with the leading power operator, due to angular momentum conservation. The contributions of the operators of \Eq{eq:eegggus} involving an ultrasoft gluon and two collinear gluons vanish when multiplied with the leading power operator due to fermion number conservation.
A complete analysis of all contributions in \Tab{tab:fact_func}, in particular of their detailed helicity and color structures, as well as their fixed order cross section contributions is beyond the scope of this paper, and is left for presentation in future work.
\subsection{Comparison with Earlier Literature}\label{sec:compare_earlier}
In this section we perform a brief comparison of our operator basis with the operators considered in \Ref{Freedman:2013vya}. These operators were used to study power corrections suppressed by ${\cal O}(\lambda^2)$ for the thrust event shape in $e^+e^-\to$ dijets. The goals of our two works are different. While we have focused on constructing a complete basis of operators valid at any order in $\alpha_s$, \Ref{Freedman:2013vya} instead derives only the set of operators that arise from tree level matching, and then uses them to explicitly calculate and confirm the ${\cal O}(\alpha_s\tau^0)$ terms in $d\sigma/d\tau$.
Despite the difference in goals, we believe it is still interesting to perform a comparison between the forms of the operators in each case. Some care must be taken, since a different formulation of SCET (first presented in \Ref{Freedman:2011kj}) is used in \Ref{Freedman:2013vya}, as compared to this paper. In the formulation of SCET used in \Ref{Freedman:2013vya}, the dynamics of each collinear sector, as well as the ultrasoft sector, is described by a copy of the QCD Lagrangian, which does not have a power expansion. This implies that the operator basis of \Ref{Freedman:2013vya} must also include terms which incorporate what would be termed subleading Lagrangian corrections in the standard formulation of SCET. Additionally, while the standard formulation of SCET used in this paper implements separate momentum conservation of residual and label momentum using the multipole expansion with labels, in the formulation of \Ref{Freedman:2013vya}, momentum is not conserved, and additional operators must be included to compensate for this. These terms must be distinguished to perform the comparison. At leading power there are no subleading Lagrangian insertions, and therefore the organization of the two SCET frameworks is equivalent.
At $\mathcal{O}(\lambda)$ in the power expansion, the basis of \Ref{Freedman:2013vya} does not consider operators involving three collinear gluons, as given in \Eq{eq:eeggg}. These operators have vanishing Wilson coefficient at tree level, with the first non-zero correction appearing through diagrams involving a quark loop, but can contribute to event shape cross sections at $\mathcal{O}(\lambda^2)$ and higher orders in the $\alpha_s$ expansion, as shown in \Tab{tab:summary}. Operators with two collinear quarks in different sectors, and a collinear gluon are also given in \Ref{Freedman:2013vya} in their operators $O_2^{(1_{a_n})}$, $O_2^{(1_{b_n})}$ and $O_2^{(1_{a_{\bar n}})}$, $O_2^{(1_{b_{\bar n}})}$. Operators with two collinear quarks in the same sector, and a collinear gluon in the other sector, corresponding to \Eq{eq:Z1_basis_diff} are given in \Ref{Freedman:2013vya} by $O_2^{(1_{e_n})}$, $O_2^{(1_{f_n})}$. It can be shown using the conservation of the QCD current that to all orders in $\alpha_s$ in the matching, these operators appear only in the combinations $O_2^{(1_{a_n})}-O_2^{(1_{b_n})}$, $O_2^{(1_{a_{\bar n}})}-O_2^{(1_{b_{\bar n}})}$, and $O_2^{(1_{e_n})}-O_2^{(1_{f_n})}$, as discussed in \Sec{sec:ang_cons}. This was first shown in \cite{Freedman:2014uta}, and used to simplify the basis of \Ref{Freedman:2013vya}. In this case, upon setting the total $\mathcal{P}_\perp$ in each sector to be zero, we find agreement with the operators in
\Eqs{eq:Z1_basis}{eq:Z1_basis_diff}. In \Ref{Freedman:2013vya}, they also include $\mathcal{O}(\lambda)$ corrections to the ultrasoft sector in (A7). In the SCET framework used in this paper, these do not appear in our operator basis due to power counting. Instead, corrections to the ultrasoft dynamics are incorporated through subleading Lagrangian insertions.
At $\mathcal{O}(\lambda^2)$ in the power expansion, the basis of \Ref{Freedman:2013vya} neglects operators involving four collinear quark fields, as included in our basis in Eqs. (\ref{eq:Z2_basis_qQ})-(\ref{eq:Z3_basis_qq}). These operators appear in the tree level matching, and satisfy all necessary symmetry relations to contribute to event shape cross sections at $\mathcal{O}(\lambda^2)$ and $\alpha_s^2$. It would be interesting to compute their explicit contribution to the cross-section, or to present an argument showing that they do not contribute. The operators involving four collinear gluons of \Eqs{eq:Z2_basis_gggg_1}{eq:Z2_basis_gggg_2} are also not considered in \Ref{Freedman:2013vya}. These operators are of limited phenomenological relevance, as they first appear at loop level in the matching, and furthermore, do not contribute to event shape cross sections at $\mathcal{O}(\lambda^2)$, as explained. In \Ref{Freedman:2013vya} they also neglect operators involving two quarks in the same sector combined with two gluons or one gluon and one $\mathcal{P}_\perp$ insertion, as in \Eqs{eq:eeqqgg_basis3}{eq:eeqqgpperp_basis_same}. Again, we have shown for the particular case of dijet event shapes, that such operators do not contribute at $\mathcal{O}(\lambda^2)$. Operators involving two quarks in opposite collinear sectors with two collinear gluons are included in \Ref{Freedman:2013vya} in operators $O_2^{(2 {b_n})}$ and $O_2^{(2 {b_{\bar n }})}$. In this case, our basis is quite different, as we have used the gluon equations of motion in the effective theory to eliminate the $n\cdot \mathcal{B}_n$ field, as is commonly done in the SCET literature \cite{Marcantonini:2008qn}. The operators $O_2^{(2 {a_n})}$, $O_2^{(2 {A_n})}$, $O_2^{(2 {\delta_n})}$, $O_2^{(2 {a_{\bar n}})}$, $O_2^{(2 {A_{\bar n }})}$, and $O_2^{(2 {\delta_{\bar n}})}$ from \Ref{Freedman:2013vya} will contribute to this case after simplification with the equations of motion. This elimination is important for the construction of our helicity basis, as it allows us to work only in terms of the physical polarizations of the $\mathcal{B}_n$ field, namely $\mathcal{B}_\perp^\pm$, and thus simplifies our basis. For the case of two gluons and two quarks, \Ref{Freedman:2013vya} does not include operators where the two quarks are in the same sector or the two gluons are in separate sectors, which can first contribute at $\alpha_s^2$. As was the case at $\mathcal{O}(\lambda)$, \Ref{Freedman:2013vya} also includes operators which in our language arise from subleading Lagrangian insertions. Counting only operators that arise from two quarks in separate sectors and two gluons in the same sector, we see that we have 4 operators while \Ref{Freedman:2013vya} has 8, which implies that the simplifications from working in the center of mass frame and using the equations of motion are useful for reducing the number of operators.
Finally, \Ref{Freedman:2013vya} also has operators in (A11) involving the expansion of momentum conserving delta functions, denoted $\mathcal{O}^{(\delta)}$, which are required in their formulation of SCET. In the standard formulation of SCET, both label, and residual momenta are conserved, so that these operators are not required in our basis. This distinction also modifies the measurement functions in the two approaches, and the subleading operators $\mathcal{O}^{(\delta)}$ are required to maintain consistency between the approaches. Indeed, a particularly convenient feature of our construction is that \Tab{tab:helicityBB} gives the full list of building blocks required to construct a complete basis of hard scattering operators for an arbitrary number of collinear directions and to arbitrary power in the expansion parameter $\lambda$. The completeness of our operator basis ensures that it is closed under renormalization group evolution.
\section{Matching at Subleading Power}
\label{sec:matching}
Having identified the relevant operators which contribute to the cross section in \Tab{tab:summary}, in this section we carry out the matching to determine the lowest order Wilson coefficients for these operators. At ${\cal O}(\lambda^2)$ operators with up to four collinear fields are present in the basis. The amplitudes are known for $e^+e^- \to 3$ partons (and related crossings) at $2$ loops \cite{Garland:2001tf,Garland:2002ak}, and $e^+e^- \to 4$ partons (and related crossings) at $1$ loop \cite{Bern:1996ka,Bern:1997sc}. As the focus of the present paper is on the structure of the operators, we will content ourselves with performing the tree level matching, leaving the higher loop matching for future work. This implies in particular that we do not match to operators involving only gluon fields, which necessarily first appear at loop level. Although we have emphasized the utility of the helicity basis for counting operators, in this section we will perform the calculation using free Lorentz indices, and then projecting onto definite helicities to obtain the Wilson coefficients for the helicity operators. This allows us to carry out the matching for a starting current with a general Dirac structure $\Gamma$. We will give the results for the operators both in terms of standard Lorentz and Dirac structures, and then projected onto our helicity operator basis.
We begin in \Sec{sec:matching_general} with a general discussion of matching at subleading power. We then consider explicit matching calculations.
From \Tab{tab:summary}, we see that we must therefore match only to operators involving one additional collinear gluon, which are considered in \Sec{sec:matching_NLP}, two additional collinear gluons, which are considered in \Sec{sec:matching_NNLP_qqgg}, or an additional $q\bar q$ pair, which are considered in \Sec{sec:matching_NNLP_4q}. Although the operators involving insertions of ultrasoft operators which contribute to the $\mathcal{O}(\lambda^2)$ cross section have their Wilson coefficients determined by RPI, we also perform the matching calculation with a single ultrasoft gluon in \Sec{sec:match_soft} to explicitly verify this.
\subsection{General Formalism}\label{sec:matching_general}
In this section we briefly describe matching to the subleading helicity operators in \mbox{${\rm SCET}_{\rm I}$}\xspace.
As in the leading power case, QCD is matched to an effective SCET hard scattering Lagrangian, which governs the interactions at the hard scale. As has been discussed, this hard scattering Lagrangian has an explicit expansion in powers of $\lambda$,
\begin{equation} \label{eq:Leff_sub}
\mathcal{L}_{\text{hard}} = \sum_{j\geq0} \mathcal{L}^{(j)}_{\text{hard}}
\,,\end{equation}
where $j$ denotes suppression by ${\cal O}(\lambda^j)$ with respect to the leading power hard scattering operators. The effective Lagrangian for hard scattering operators at each power is given by,
\begin{align} \label{eq:Leff_sub_explicit_later}
\mathcal{L}^{(j)}_{\text{hard}} = \sum_{\{n_i\}} \sum_{A,\cdot\cdot}
\bigg[ \prod_{i=1}^{\ell_A} \int \! \! \mathrm{d} \omega_i \bigg] \,
& \vec O^{(j)\dagger}_{A\,\lotsdots}\big(\{n_i\};
\omega_1,\ldots,\omega_{\ell_A}\big) {\nonumber}\\
& \times
\vec{C}^{(j)}_{A\,\lotsdots}\big(\{n_i\};\omega_1,\ldots,\omega_{\ell_A} \big)
\,.
\end{align}
The Wilson coefficients, $\vec{C}^{(j)}_{A}$, are determined by performing a matching calculation from QCD to SCET. When matching at subleading powers, one must take into account subleading Lagrangian insertions with lower power hard scattering operators, arising from the fact that the Lagrangian describing the ultrasoft and collinear dynamics is also a power expansion in $\lambda$,
\begin{align}
\mathcal{L}_\mathrm{dyn}=\mathcal{L}^{(0)}+\mathcal{L}^{(1)}+\cdots \,.
\end{align}
To any given power we will only need to consider a finite number of insertions of $\mathcal{L}^{(i)}$ for $i \geq1$, as these are constrained by the power counting. However, we must consider arbitrary $\mathcal{L}^{(0)}$ insertions, constrained only by the order in $\alpha_s$ to which we are working.
Consider the tree level matching at order $\lambda^p$. We assume that the Wilson coefficients at all lower powers, $\vec C^{(q)}_A$ with $q<p$, have already been determined. The matching can be performed with an arbitrary external state $\bra{X}$, as long as it is chosen to have non-zero matrix elements for the specific color and helicity structure of the operator. (If the basis is over complete with regards to color, then the organization of the matching results will also depend on the convention adopted, but we do not encounter this issue in this paper.) In general these specifications will not pick out a particular operator from among those with the same color and helicity structures, so there would still be a sum on $A$. The remaining distinguishing feature used is the dependence on momenta (for example, looking at the $p_\perp$-momenta that appear to determine where a $\mathcal{P}_\perp^\pm$ is acting in the operator). This then enables us to write down a matching equation for a fixed $A$ and fixed helicities, and color channel. For notational simplicity we assume all these specifications to be made with the state $\bra{X_A}$.
Such an external state can arise at tree level only from $\vec O^{(p)\dagger}_A$, or from subleading Lagrangian insertions into lower power operators. One then has the following matching equation
\begin{align} \label{eq:lamp_match}
-i(\mathcal{A}^{\mathrm{tree}})^{(\lambda^p)}
=C^{(p)} \big\langle X_A\big| \vec O^{(p)\dagger}_A \big|0 \big\rangle^{\mathrm{tree}}_{\mathcal{L}^{(0)}}
+ \sum_{n=1} \big\langle X_A\big| [\mbox{\small$\prod$} \mathcal{L}^{(k)}]^{(n)} \mathcal{L}^{(p-n)}_{\text{hard}} \big|0 \big\rangle^{\mathrm{tree}}_{\mathcal{L}^{(0)}}\,,
\end{align}
where $[\mbox{\small$\prod$} \mathcal{L}^{(k)}]^{(n)}$ represents all combinations of subleading SCET Lagrangian insertions $\mathcal{L}^{(k\ge 1)}$ whose powers sum to $n$. The ${\cal L}^{(0)}$ subscript on the matrix elements indicates that they are evaluated with any number of insertions of the leading power SCET Lagrangian. Here $(\mathcal{A}^{\mathrm{tree}})^{(\lambda^p)}$ is the corresponding amplitude in full QCD expanded to pick out the order $\lambda^p$ component. The momentum of external particles in this amplitude are assigned a power counting corresponding to the building blocks in $\vec O^{(p)\dagger}_A$. This means that if two external particles are in the same collinear sector then we take their collinear limit when expanding the matrix element, and that we take the ultrasoft limit for momenta of ultrasoft particles. The matrix element, $\bra{X_A} \vec O^{(p)\dagger}_A |0 \rangle^{\mathrm{tree}}_{\mathcal{L}^{(0)}}$, is evaluated using the Feynman rules for the helicity operators. For the case with $N$ distinct collinear particles at leading power, the sum on the right side of \eq{lamp_match} does not contribute, as there is nothing to expand when each of the particles are separated to distinct sectors. However, at subleading power either collinear particles in the same sector or ultrasoft particles will give contributions in this expansion. Most modern fixed order computations are performed using spinor helicity and color decomposition techniques, which give compact results. Using our basis of helicity operators, the helicity amplitudes can be directly used, since we match to external states with definite helicities. For the particularly simple cases in this paper, we find it is easier to simply project the different helicities from a tensor, however, the helicity formalism still provides a powerful way of organizing the operators.
Beyond tree level, the matching is still conceptually straightforward, but technically more demanding due to the need to evaluate loop diagrams in both the full and effective theories. \eq{lamp_match} still holds, but now we must consider each of the Wilson coefficients and matrix elements as an expansion in $\alpha_s$, and go beyond the leading term. At leading power it is often possible to arrange the choice of IR and UV regulators such that the SCET loop diagrams are scaleless in the matching calculation. At subleading power we in general need more than one particle in a given sector in order to have nonzero overlap with the operator, and this can introduce non-trivial momentum scales in the SCET loop integrals such that loop calculations in SCET can not be avoided in this manner.
\subsection{Setup and Leading Power Matching}
\label{sec:matching_LP}
In the following sections we carry out the matching starting from a full QCD quark current with an arbitrary spin structure $\Gamma$,
\begin{align}
J^{\Gamma}=\bar \psi \Gamma \psi\,.
\end{align}
For vector and axial-vector $\Gamma$ this current is (partially)-conserved in QCD, while for the scalar, pseudo-scalar and tensor cases $J^{\Gamma}$ has an anomalous dimension in QCD. We will denote the full theory vertex with an $\otimes$ symbol, and hard scattering operators in the effective theory will be denoted with a purple circle.
The leading power Wilson coefficient is of course well known, however, we reproduce it here for completeness. The unique leading power operator, written in the form of a current is
\begin{align}
\mathcal{O}^{(0)\Gamma}=\bar \chi_{n} \Gamma \chi_{\bar n}\,,
\end{align}
It's Wilson coefficient for any spin structure $\Gamma$ is given to $\mathcal{O}(\alpha_s)$ by
\begin{align}\label{eq:LP_wilson}
C^{(0)}=1+\frac{\alpha_s(\mu) C_F}{4\pi} \left( -\log^2 \left[ \frac{-\omega_1 \omega_2-i0}{\mu^2} \right] +3\log \left[ \frac{-\omega_1 \omega_2-i0}{\mu^2} \right] -8 +\frac{\pi^2}{6} \right) \,.
\end{align}
Throughout this section, we will restrict ourselves to the tree level matching, however, we have given the Wilson coefficient of \Eq{eq:LP_wilson} to one loop, since it will be used to demonstrate the RPI relations of \Sec{sec:soft_basis} for the operators involving ultrasoft insertions, which are first non-trivial at this order.
\subsection{Subleading Matching with a Single Collinear Gluon}
\label{sec:matching_NLP}
We begin by considering the matching to operators involving two collinear quark fields and a collinear gluon field, along with a possible $\mathcal{P}_\perp$ insertion. A basis of such operators was given in \Eqs{eq:Z1_basis}{eq:Z1_basis_diff}, as well as \Eq{eq:eeqqgpperp_basis}.
The matching coefficients for these operators can be derived by considering matrix elements with a $q\bar q g$ final state. At $\mathcal{O}(\alpha_s)$, the QCD diagrams for the production of a single gluon from a quark current with spin structure $\Gamma$ are
\begin{align}
\fd{3cm}{figures/matching_subleadingvertex_2_low} &= \bar{u}(p_1) (i g T^a \Sl{\epsilon}^*) \frac{ i (\Sl{p}_1 + \Sl{p}_3)}{ (p_1+p_3)^2} \Gamma v(p_2) \,, {\nonumber}\\
\fd{3cm}{figures/matching_subleadingvertex_1_low}&= \bar{u}(p_1) \Gamma \frac{- i (\Sl{p}_2 + \Sl{p}_3)}{ (p_2+p_3)^2} (i g T^a \Sl{\epsilon}^*)v(p_2) \,.
\end{align}
The required matching coefficients is obtained by expanding these amplitudes in the required kinematic limits, namely as the gluon becomes collinear with either the quark or antiquark, or when the quark and antiquark become collinear. We will consider each of these cases in turn.
We note that while we will restrict our attention to tree level matching in this section, it would be particularly interesting to extend the matching to one-loop. Indeed, the one-loop matching to the operators involving a single additional collinear gluon field is the only ingredient related to the hard scattering operators, beyond the matching coefficients presented in this paper, that would be required to perform a full analytic $\mathcal{O}(\alpha_s^2)$ fixed order calculation of the subleading cross section for an $e^+e^-\to$ dijets observable. We leave this to future work.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(q)_{ n} (\bar q g)_{\bar n}:}$
}
We first consider the case where the gluon and antiquark are $\bar n$ collinear, and the quark is $n$ collinear. In this case we have hard scattering operators at both $\mathcal{O}(\lambda)$, which are independent of $\mathcal{P}_\perp$ and operators at $\mathcal{O}(\lambda^2)$, which depend on $\mathcal{P}_\perp$. To extract the matching coefficients for both sets of operators in a single calculation, we expand the QCD amplitudes with a non-zero $\mathcal{P}_\perp$ between the gluon and antiquark. We take the momenta of the particles as
\begin{align}
p_1^\mu=\frac{\omega_1}{2}n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2} \bar n^\mu+p_\perp^\mu +\frac{ p_{2r}}{2} n^\mu\,, \qquad
p_3^\mu=\frac{\omega_3}{2}\bar n^\mu -p_\perp^\mu +\frac{ p_{3r}}{2} n^\mu\,,
\end{align}
where the light cone components are assigned a collinear scaling. The $\omega_i$ and $p_\perp$ are taken to be purely label momenta. Following the notation of \Sec{sec:coll}, throughout this section, we will use the notation
\begin{align}
u_n(i)=P_n u(p_i)\,, \qquad \text{and} \qquad v_n(i)=P_n v(i)\,,
\end{align}
for the projected SCET spinors, where the momentum $p_i$ is $n$ collinear, and similarly for the case that it is $\bar n$ collinear. They obey
\begin{align}
u(p_i) = \Big( 1+ \frac{\Sl{p}_{i\perp}}{\bar{n} \cdot p_i} \frac{\Sl{\bar{n}}}{2} \Big) u_n(i) \,,
\qquad
u(p_i) = \Big( 1+ \frac{\Sl{p}_{i\perp}}{n \cdot p_i} \frac{\Sl{n}}{2} \Big) u_{\bar n}(i) \,,
\end{align}
for the $n$ collinear and $\bar n$ collinear cases respectively, with direct analogs for the $v(p_i)$ spinors.
Expanding the first diagram in $\lambda$, we find
\begin{align}\label{eq:ggqn_match1}
\fd{3cm}{figures/matching_subleadingvertex_casea1_low}
&=- g \frac{n\cdot \epsilon^*}{ \omega_3} \bar{u}_{ n}(1) T^a \Gamma v_{\bar n}(2) \\
&\hspace{-2cm}
- \frac{g}{\omega_1 \omega_3}\bar{u}_{n}(1) T^a \left( \omega_3 \Sl{\epsilon}_\perp^*\frac{\Sl{\bar n}}{2}\Gamma- n\cdot\epsilon^* \frac{\Sl{\bar n}}{2} \Sl{p}_\perp \Gamma +\frac{ n\cdot \epsilon^* \, \omega_1}{\omega_2} \Gamma \Sl{p}_\perp \frac{\Sl{n}}{2} \right) v_{\bar n}(2) {\nonumber} \\
&\hspace{-2cm} - \frac{g}{\omega_1 \omega_3}\bar{u}_{n}(1) T^a
\biggl( n\cdot \epsilon^*\, p_{3r} \Gamma - \Sl{\epsilon}_\perp^* \Sl{p}_\perp \Gamma + \frac{\omega_3}{\omega_2}\Sl{\epsilon}_\perp^*\frac{\Sl{\bar n}}{2} \Gamma \Sl{p}_\perp \frac{\Sl{n}}{2}-\frac{n\cdot\epsilon^*}{\omega_2} \frac{\Sl{\bar n}}{2} \Sl{p}_\perp \Gamma \Sl{p}_\perp \frac{ \Sl{n}}{2} \biggr) v_{\bar n}(2)
{\nonumber} \\
&\hspace{-2cm}+\mathcal{O}(\lambda^3)
\,, {\nonumber}
\end{align}
where terms in the first line scale like $\mathcal{O}(\lambda^0)$, terms in the second line scale like $\mathcal{O}(\lambda^1)$, and so on. Similarly, we find
\begin{align}\label{eq:ggqn_match2}
\fd{3cm}{figures/matching_subleadingvertex_casea2_low}
&=\frac{g}{(p_{2r}+p_{3r})}\bar u_{n}(1) T^a \Gamma
\Bigl( \epsilon^*\cdot \bar n +\frac{ \Sl{\epsilon}_\perp^* \Sl{p}_\perp }{\omega_2} \Bigr)v_{\bar n}(2){\nonumber} \\
& +\frac{g}{2(\omega_{2}+\omega_{3})}\bar u_{n}(1) T^a \Gamma \Bigl( \frac{\epsilon^*\cdot n }{\omega_2} \Sl{p}_\perp \Sl{n} + \Sl{n} \Sl{\epsilon}^*_\perp \Bigr) v_{\bar n}(2)\,.
\end{align}
The first line is $\mathcal{O}(\lambda^0)$ and second line is $\mathcal{O}(\lambda^1)$. Interestingly, this result is exact with no corrections beyond $\mathcal{O}(\lambda)$.
The $\mathcal{O}(\lambda^0)$ terms are reproduced by $T$-products of the leading SCET Lagrangian.
Since we have performed the matching with no residual momenta for the large or $\perp$ components, all subleading Lagrangian contributions from the $T$-products in the second term on the right hand side of \Eq{eq:lamp_match} vanish. This includes those with insertions of the ultrasoft derivative operators of \Eqs{eq:soft_derivative_basis}{eq:eedggus} and the leading order SCET Lagrangian. Since we have shown by combining the RPI relation of \Eq{eq:uspartialRPIrelation} with the matching to the leading power operator in \Eq{eq:LP_wilson} that these operators first appear at $\mathcal{O}(\alpha_s)$, such terms will not contribute to the matching. With this setup the QCD result at $\mathcal{O}(\lambda^1)$ and $\mathcal{O}(\lambda^2)$ must therefore be exactly reproduced by our basis of hard scattering operators.
Starting at $\mathcal{O}(\lambda)$, summing the results of \Eqs{eq:ggqn_match1}{eq:ggqn_match2} and rearranging, we find
\begin{align}
\left( \fd{3cm}{figures/matching_subleadingvertex_casea1_low}+\fd{3cm}{figures/matching_subleadingvertex_casea2_low}\right)\left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_casea2_low}} \right|_{\mathcal{O}(\lambda)}& \\
&\hspace{-8cm}
= - \frac{g}{\omega_1} \bar u_{n} (1) T^a \Bigl( \Sl{\epsilon}^*_\perp+\frac{\epsilon^*\cdot n \Sl{p}_\perp}{\omega_3} \Bigr) \frac{\Sl{\bar n}}{2} \Gamma v_{\bar n}(2)
+ \frac{g}{(\omega_2 +\omega_3)} \bar u_{n} (1) T^a \Gamma \frac{\Sl{n}}{2} \Bigl( \Sl{\epsilon}^*_\perp+\frac{\epsilon^*\cdot n \Sl{p}_\perp}{\omega_3} \Bigr) v_{\bar n}(2)\,.{\nonumber}
\end{align}
This can be recognized as the one gluon matrix element of the operator
\begin{align}\label{eq:matched_1gluon_barn_general}
\mathcal{O}^{(1)\Gamma}_{\mathcal{B} \bar n}&=\frac{g}{\omega_2+\omega_3} \bar \chi_{n,\omega_1} \Gamma \frac{\Sl{n}}{2}\Sl{\mathcal{B}}_{\perp \bar n,\omega_3} \chi_{\bar n,-\omega_2}-\frac{g}{\omega_1}\bar \chi_{n,\omega_1}\Sl{\mathcal{B}}_{\perp \bar n,\omega_3} \frac{\Sl{\bar n}}{2} \Gamma \chi_{\bar n,-\omega_2}\,.
\end{align}
In the particular case that $ \Gamma=\gamma^\mu$, this simplifies to
\begin{align}\label{eq:matched_1gluon_barn}
\mathcal{O}^{(1)\gamma^\mu}_{\mathcal{B} \bar n}&= g\left( \frac{\omega_1 n^\mu-(\omega_2 +\omega_3) \bar n^\mu }{\omega_1 (\omega_2 +\omega_3)} \right) \bar \chi_{n, \omega_1} \Sl{\mathcal{B}}_{\perp \bar n,\omega_3} \chi_{\bar n,-\omega_2}\,,
\end{align}
which we note for later use.
In the center-of-momentum frame conservation of momentum implies that $\omega_1=\omega_2+\omega_3=Q$, further simplifying the structure of the operator. However, throughout this section, we will not perform such a simplification, and will write the result for the Wilson coefficient for generic values of the label momenta.
At $\mathcal{O}(\lambda^2)$, only the diagram in \Eq{eq:ggqn_match1} contributes. The result can be simplified using the on-shell condition for the collinear gluon,
\begin{align}
p_{3r}\omega_3+p_\perp^2=0\,,
\end{align}
after which one finds
\begin{align}
&\left( \fd{3cm}{figures/matching_subleadingvertex_casea1_low}+\fd{3cm}{figures/matching_subleadingvertex_casea2_low}\right)\left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_casea2_low}}\right|_{\mathcal{O}(\lambda^2)}\\
&=\frac{g}{\omega_1 \omega_2} \bar u_{n} (1) T^a\Bigl( \Sl{\epsilon}^*_\perp +\frac{\epsilon^*\cdot n \Sl{p}_\perp}{\omega_3} \Bigr) \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{p}_\perp v_{\bar n}(2)
+\frac{g}{\omega_1 \omega_3} \bar u_{n} (1)T^a \Bigl( \Sl{\epsilon}^*_\perp +\frac{\epsilon^*\cdot n \Sl{p}_\perp}{\omega_3} \Bigr) \Sl{p}_\perp \Gamma v_{\bar n}(2)\,.{\nonumber}
\end{align}
This can be recognized as the one gluon matrix element of the two SCET hard scattering operators
\begin{align}\label{eq:matched_nbarP}
\mathcal{O}^{(2)\mu}_{\mathcal{P} \bar n1}
&=- \frac{g}{\omega_1 \omega_3} \bar \chi_{ n,\omega_1}
\big[\Sl\mathcal{B}_{\perp \bar n, \omega_3} \Sl{\mathcal{P}}^\dagger_{\perp}\big]
\Gamma \chi_{\bar n,-\omega_2} \,, {\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{P} \bar n2}
&= - \frac{g}{\omega_1 \omega_2} \bar \chi_{n,\omega_1}
\Big[\Sl\mathcal{B}_{\perp \bar n, \omega_3} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{P}}^\dagger_{\perp}\Big] \chi_{\bar n,-\omega_2}
\,.
\end{align}
Of particular interest is the fact that the Wilson coefficients exhibit singularities as the momentum fraction of the gluon or quark vanishes, which will be associated with logarithms in the subleading power cross section.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(gq)_{n} (\bar q )_{\bar n}:}$
}
In the case that the gluon is collinear with the quark, we can immediately obtain the operators by charge conjugation, and a relabeling. To be clear on the particle labeling, we consider
\begin{align}
p_1^\mu=\frac{\omega_1}{2}n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2} \bar n^\mu+p_\perp^\mu +\frac{ p_{2r}}{2} n^\mu\,, \qquad
p_3^\mu=\frac{\omega_3}{2}n^\mu -p_\perp^\mu +\frac{ p_{3r}}{2} \bar n^\mu\,.
\end{align}
Analogously to the case that the gluon is in the same collinear sector as the antiquark, we find the $\mathcal{O}(\lambda)$ operator
\begin{align}\label{eq:matched_1gluon_n_general}
\mathcal{O}^{(1)\mu}_{\mathcal{B} n}&=-\frac{g}{\omega_1+\omega_3} \bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp n,\omega_3} \frac{\Sl{\bar n}}{2} \Gamma \chi_{\bar n,-\omega_2}+\frac{g}{\omega_2}\bar \chi_{n,\omega_1} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{\perp n,\omega_3} \chi_{\bar n,-\omega_2}\,.
\end{align}
Again, in the case that we take $ \Gamma=\gamma^\mu$, this simplifies to
\begin{align}\label{eq:matched_1gluon_n}
\mathcal{O}^{(1)\mu}_{\mathcal{B} n}&=g\left( \frac{(\omega_1+\omega_3) n^\mu-\omega_2 \bar n^\mu }{\omega_2 (\omega_1 +\omega_3)} \right) \bar \chi_{n, \omega_1} \Sl{\mathcal{B}}_{\perp n,\omega_3} \chi_{\bar n,-\omega_2}\,.
\end{align}
The two $\mathcal{O}(\lambda^2)$ operators are
\begin{align}\label{eq:matched_nP}
\mathcal{O}^{(2)\mu}_{\mathcal{P} n1}
&=-\frac{g}{\omega_1 \omega_2} \bar \chi_{n,\omega_1}
\Bigl[ \Sl{\mathcal{P}}_\perp \frac{\Sl{\bar n}}{2}
\Gamma \frac{\Sl{n}}{2} \Sl\mathcal{B}_{\perp n, \omega_3} \Bigr]
\chi_{\bar n,-\omega_2}
\,, {\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{P} n2}
&= - \frac{g}{\omega_2 \omega_3}
\bar \chi_{n,\omega_1} \Gamma
\bigl[ \Sl{\mathcal{P}}_\perp \Sl\mathcal{B}_{\perp n, \omega_3} \bigr]
\chi_{\bar n,-\omega_2}
\,.
\end{align}
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(q \bar q)_{\bar n} ( g)_{n}:}$
}
We now consider the case where both the quark and anti-quark are in the same ($\bar n$) collinear sector and the gluon is in the $\bar n$ collinear sector. As discussed in \Sec{sec:eeJets}, in this case, the operators of this form only contribute to the cross section in the form of a matrix element with themselves. We therefore only need the matching to $\mathcal{O}(\lambda)$, if we are interested in $\mathcal{O}(\lambda^2)$ contributions to the cross section. We can therefore set the $\perp$ momentum to zero when performing the matching. We take the kinematics as
\begin{align}
p_1^\mu=\frac{\omega_1}{2}\bar n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2}\bar n^\mu\,, \qquad
p_3^\mu=\frac{\omega_3}{2}n^\mu\,.
\end{align}
Expanding the QCD results the ${\cal O}(\lambda^0)$ terms cancel between diagrams, and at ${\cal O}(\lambda)$ we find
\begin{align}\label{eq:matching_qqsame1}
\left( \fd{3cm}{figures/matching_subleadingvertex_casec1_low} \right) \left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_casec1_low}} \right|_{\mathcal{O}(\lambda)}
&= - \frac{g}{\omega_1}\bar u_{\bar n}(1) T^a \Sl{\epsilon}^*_\perp \frac{\Sl{n}}{2} \Gamma v_{\bar n}(2)\,,
\\
\left( \fd{3cm}{figures/matching_subleadingvertex_casec2_low} \right) \left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_casec2_low}} \right|_{\mathcal{O}(\lambda)}
&=
+ \frac{g}{\omega_2 } \bar u_{\bar n}(1) T^a \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_\perp v_{\bar n}(2)
\,. {\nonumber}
\end{align}
We therefore see that for this configuration, both QCD diagrams contribute.
These contributions can be recognized as the one gluon matrix element of the two SCET operators
\begin{align}\label{eq:match_2qsame}
\mathcal{O}^{(1)\mu}_{\chi\chi\bar n1}&=-\frac{g}{\omega_1} \bar \chi_{\bar n, \omega_1} \Sl{\mathcal{B}}_{\perp n,\omega_3} \frac{\Sl{n}}{2} \Gamma \chi_{\bar n,-\omega_2}\,, {\nonumber} \\
\mathcal{O}^{(1)\mu}_{\chi\chi\bar n2}&=\frac{g}{\omega_2} \bar \chi_{\bar n, \omega_1} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{\perp n,\omega_3} \chi_{\bar n,-\omega_2}\,.
\end{align}
Both Wilson coefficients exhibit a singularity as the energy fraction as one or the other quark goes to zero. It is interesting to note that this structure is dictated by the RPI-III symmetry of the theory. In particular, due to the presence of the $\Sl{n}$ projector which necessarily appears between the two $\bar n$ collinear quark fields, the Wilson coefficient must behave like $1/\omega_i$, where $i=2$, or $3$. This hints at the possible universality of this structure in subleading power collinear limits.
\subsection{Subleading Matching with Two Collinear Gluons}
\label{sec:matching_NNLP_qqgg}
We now consider the matching to operators involving two collinear gluons. As discussed in \Sec{sec:eeJets}, there are two relevant classes of such operators that contribute to the cross section at $\mathcal{O}(\lambda^2)$: those involving two collinear gluons in the same collinear sector, and those involving two gluons in opposite sectors, both of which have have the quark anti-quark pair in opposite sectors. We consider each case in turn.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q)_{\bar n} ( qgg)_{n}:}$
}
A basis of $\mathcal{O}(\lambda^2)$ operators with two collinear gluons in the same sector was given in \Eq{eq:eeqqgg_basis1}. Since these operators get their power suppression from the explicit collinear gluon fields, we can simplify the matching by taking the particle momenta as
\begin{align}
p_1^\mu&=\frac{\omega_1}{2}n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2}\bar n^\mu\,, {\nonumber} \\
p_3^\mu&=\frac{\omega_3}{2}n^\mu+p_\perp^\mu +\frac{p_{3r}}{2} \bar n^\mu\,, \qquad
p_4^\mu=\frac{\omega_4}{2}n^\mu-p_\perp^\mu +\frac{p_{4r}}{2} \bar n^\mu\,,
\end{align}
where the particle labeling is the same as before, but with $p_4$ labeling the additional gluon. This choice of momenta also removes any contributions from subleading Lagrangian insertions. We do however, get contributions from the subleading hard scattering operators, $\mathcal{O}^{(2)\mu}_{\mathcal{P} n1,2}$, which must be disentangled from the operator coefficients we want to determine. For these operators we find
\begin{align}\label{eq:2q2gsame_SCET}
\left( \fd{3cm}{figures/matching_2q2g_same_SCET_contribution_3_b}+\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_3}\right) \left. \vphantom{\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_3_b}} \right |_{\mathcal{O}(\lambda^2)}& \\
&\hspace{-4.5cm} = \frac{-g^2\omega_4}{\omega_1 \omega_2 (\omega_1+\omega_4)} \bar u_n (1) T^b T^a \Sl{\epsilon}^*_{4\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{3\perp} v_{\bar n} (2) +\left[ (3,a)\leftrightarrow (4,b) \right] , {\nonumber} \\
\left( \fd{3cm}{figures/matching_2q2g_same_SCET_contribution_4_b}+\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_4} \right)\left. \vphantom{\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_4_b}} \right |_{\mathcal{O}(\lambda^2)}&{\nonumber} \\
&\hspace{-4.5cm} = \frac{g^2}{\omega_1 \omega_2 \omega_3 p_{4r}} \bar u_n (1) T^b T^a \Sl{\epsilon}^*_{4\perp} \Sl{p}_\perp \Gamma \Sl{p}_\perp \Sl{\epsilon}^*_{3\perp} v_{\bar n} (2) +\left[ (3,a)\leftrightarrow (4,b) \right]\,. {\nonumber}
\end{align}
Interestingly, we see that with our choice of momentum, this gives rise to one term which localizes, with an $\mathcal{O}(\lambda^0)$ denominator, and a term which does not, with a $1/p_{4r}$. The corresponding non-abelian graphs vanish with our choice of momentum,
\begin{align}
\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_3na}=0\,, \qquad
\fd{3cm}{figures/matching_2q2g_same_SCET_contribution_4na}=0\,.
\end{align}
Note that we do not have to consider the operators involving ultrasoft derivatives from \Eqs{eq:soft_derivative_basis}{eq:eedggus} in the matching, since we have chosen to only use a residual momentum component for the small momentum component, and we have shown by RPI that this component of the derivative first appears at $\mathcal{O}(\alpha_s)$.
We now consider the expansion of the full theory diagrams.
We begin by expanding each of the QCD diagrams for the production of two gluons off of the $q\bar q$ pair to this order. For the independent emission diagrams where both gluons with $n$-collinear scaling are emitted from the quark with $\bar n$-collinear scaling, we have
\begin{align}\label{eq:qqggmatch_1}
&\fd{3cm}{figures/matching_2q2g_diagram1_low}+\fd{3cm}{figures/matching_2q2g_diagram2_low} \\
&= \bar{u}(p_1) \Gamma \frac{-i (\Sl{p}_2 + \Sl{p}_3 + \Sl{p}_4)}{(p_2 + p_3 + p_4)^2} (i g T^b) \Sl{\epsilon}^*_{4} \frac{ -i (\Sl{p}_2 + \Sl{p}_3)}{(p_2 + p_3)^2} (i g T^a) \Sl{\epsilon}_{3}^* v(p_2) +\left[ (3,a)\leftrightarrow (4,b) \right]\,. {\nonumber}
\end{align}
Expanding, we can pick out the $\mathcal{O}(\lambda^2)$ contribution of these diagrams, which is given by
\begin{align}
\left( \fd{3cm}{figures/matching_2q2g_diagram1_low}+\fd{3cm}{figures/matching_2q2g_diagram2_low} \right)\left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram2_low}} \right|_{\mathcal{O}(\lambda^2)}&\\
&\hspace{-3cm}= -\frac{g^2}{\omega_2(\omega_3+\omega_4)}\bar u_{n}(1) T^b T^a \Gamma \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} v_{\bar n}(2)+\left[ (3,a)\leftrightarrow (4,b) \right]\,,{\nonumber}
\end{align}
which is local, having only $\mathcal{O}(\lambda^0)$ momenta in the denominator.
Similarly, for the independent emission diagrams where a single gluon is emitted from the $n$ collinear quark, we have
\begin{align}\label{eq:qqggmatch_2}
&\fd{3cm}{figures/matching_2q2g_diagram3_low}+\fd{3cm}{figures/matching_2q2g_diagram4_low} \\
&= \bar{u}(p_1) (i g T^b) \Sl{\epsilon}^*_{4} \frac{i (\Sl{p}_1 + \Sl{p}_4)}{(p_1 + p_4)^2} \Gamma \frac{ -i (\Sl{p}_2 + \Sl{p}_3)}{(p_2 + p_3)^2} (i g T^a) \Sl{\epsilon}_{3}^* v(p_2) +\left[ (3,a)\leftrightarrow (4,b) \right]
\,. {\nonumber}
\end{align}
The $\mathcal{O}(\lambda^2)$ contribution of these diagrams is given by
\begin{align}
\left( \fd{3cm}{figures/matching_2q2g_diagram3_low}+\fd{3cm}{figures/matching_2q2g_diagram4_low} \right)\left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram4_low}}\right |_{\mathcal{O}(\lambda^2)}& \\
&\hspace{-6cm}= -g^2 \frac{1}{\omega_1 \omega_2} \bar u_n (1)T^b T^a \Sl{\epsilon}^*_{4\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{3\perp} v_{\bar n}(2) +\left[ (3,a)\leftrightarrow (4,b) \right]\, {\nonumber}\\
&\hspace{-6cm} +g^2 \frac{1}{\omega_1 \omega_2 \omega_3 p_{4r}} \bar u_n (1)T^b T^a \Sl{\epsilon}^*_{4\perp} \Sl{p}_\perp \Gamma \Sl{p}_\perp \Sl{\epsilon}^*_{3\perp} v_{\bar n} (2) +\left[ (3,a)\leftrightarrow (4,b) \right]\,. {\nonumber}
\end{align}
Here we recognize both a local term, as well as a $T$-product like term, which is reproduced by the SCET diagrams in \Eq{eq:2q2gsame_SCET}. This is expected from the topology of the diagram.
There are also non-abelian diagrams. In the case that the gluon is emitted from the $\bar n$ collinear quark, we find
\begin{align}\label{eq:qqggmatch_3}
\fd{3cm}{figures/matching_2q2g_diagram7_low} &= -i g^2 f^{abc} \bar u(p_1) \Gamma T^c \frac{(\Sl{p}_2+\Sl{p}_3+\Sl{p}_4 )}{(p_2+p_3+p_4)^2(p_3+p_4)^2}\gamma^\rho v(p_2){\nonumber} \\
&\cdot \left[ g^{\mu \nu}(p_4-p_3)^\rho-g^{\nu \rho}(2p_4+p_3)^\mu +g^{\rho \mu}(p_4+2p_3)^\nu \right]\epsilon^*_{4\nu} \epsilon^*_{3\mu} \,.
\end{align}
The $\mathcal{O}(\lambda^2)$ contribution of this diagram vanishes
\begin{align}
\fd{3cm}{figures/matching_2q2g_diagram7_low}\left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram7_low}}\right |_{\mathcal{O}(\lambda^2)}&=0 \,.
\end{align}
Finally, there are three diagrams in which both gluons are emitted from the $n$ collinear quark
\begin{align}
&\left( \fd{3cm}{figures/matching_2q2g_diagram5_low}+\fd{3cm}{figures/matching_2q2g_diagram6_low}+\fd{3cm}{figures/matching_2q2g_diagram8_low}\right) \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram8_low}} \right |_{\mathcal{O}(\lambda^2)}=0\,.
\end{align}
Each of these diagrams individually gives a vanishing contribution at $\mathcal{O}(\lambda^2)$, as might naively be expected due to the presence of on-shell propagators.
Subtracting the SCET matrix elements from the full theory results, we find that the result is given by the tree level matrix element of the two operators
\begin{align}\label{eq:matched_twogluon_samesector}
\mathcal{O}_{\mathcal{B}\cB1}^{(2)\mu}&=\frac{-g^2 }{ \omega_2 (\omega_1+\omega_3)} \bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{n\perp,\omega_3} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{n\perp,\omega_4} \chi_{\bar n,-\omega_2}\,, {\nonumber} \\
\mathcal{O}_{\mathcal{B}\cB2}^{(2)\mu}&=\frac{-g^2}{\omega_2(\omega_3+\omega_4)} \bar \chi_{n,\omega_1} \Gamma \Sl{\mathcal{B}}_{n\perp, \omega_3} \Sl{\mathcal{B}}_{n\perp, \omega_4}\chi_{\bar n,-\omega_2}\,.
\end{align}
The behavior of these Wilson coefficients is interesting, in that they exhibit a singularities as a pair of collinear particles in the $n$ direction simultaneously have their energy approach zero.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(q)_{\bar n} (\bar qgg)_{n}:}$
}
To obtain the matching to the operators involving two collinear gluons and an antiquark in the same collinear sector, we can simply apply charge conjugation to the operators in \Eq{eq:matched_twogluon_samesector} and then relabel to obtain
\begin{align}\label{eq:matched_twogluon_samesector_conj}
\mathcal{O}_{\mathcal{B}\cB1'}^{(2)\mu}&=\frac{-g^2 }{ \omega_1 (\omega_2+\omega_4)} \bar \chi_{{\bf n},\omega_1} \Sl{\mathcal{B}}_{n\perp,\omega_3} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{n\perp,\omega_4} \chi_{n,-\omega_2}\,, {\nonumber} \\
\mathcal{O}_{\mathcal{B}\cB2'}^{(2)\mu}&=\frac{-g^2}{\omega_1(\omega_3+\omega_4)} \bar \chi_{{\bf n},\omega_1} \Sl{\mathcal{B}}_{n\perp, \omega_3} \Sl{\mathcal{B}}_{n\perp, \omega_4} \Gamma\chi_{n,-\omega_2}\,.
\end{align}
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar qg)_{\bar n} ( qg)_{n}:}$
}
A basis of $\mathcal{O}(\lambda^2)$ operators with two collinear gluons in the opposite sectors was given in \Eq{eq:eeqqgg_basis2}. These operators get their power suppression from the explicit collinear gluon fields, and to simplify the matching we decompose the particle momenta as
\begin{align}
p_1^\mu&=\frac{\omega_1}{2}n^\mu+p_{1\perp}+\frac{p_{1r}}{2}\bar n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2}\bar n^\mu+p_{2\perp} +\frac{p_{2r}}{2}n^\mu\,, {\nonumber} \\
p_3^\mu&=\frac{\omega_3}{2}\bar n^\mu-p_{2\perp} +\frac{p_{3r}}{2}n^\mu\,, \qquad
p_4^\mu=\frac{\omega_4}{2} n^\mu-p_{1\perp}+\frac{p_{4r}}{2}\bar n^\mu\,,
\end{align}
where the particle labeling is the same as before, but with $p_4$ labeling the additional gluon. Furthermore, we choose the gluon polarizations to be perp in the matching. This choice of momenta also removes any contributions from subleading Lagrangian insertions. However, there are also SCET contributions coming from the $\mathcal{O}(\lambda^2)$ operators $\mathcal{O}^{(2)\mu}_{\mathcal{P} \bar n1,2}$, and $\mathcal{O}^{(2)\mu}_{\mathcal{P} n1,2}$. The four contributions from these operators are given by
\begin{align}
\fd{3cm}{figures/matching_2q2g_SCET_contribution_1}
\hspace{0.4cm}
&=- \frac{g^2\, \bar u_n (1) T^b T^a \Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} \Sl{p}_{2\perp} \Gamma v_{\bar n}(2)}{\omega_1 \omega_3(\omega_1+\omega_4)(p_{1r}+p_{4r})}
\,, {\nonumber} \\
\fd{3cm}{figures/matching_2q2g_SCET_contribution_2}
\hspace{0.4cm}
&= - \frac{g^2\, \bar u_n (1) T^b T^a \Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{p}_{2\perp} v_{\bar n}(2)}{\omega_1 \omega_2(\omega_1+\omega_4)(p_{1r}+p_{4r})}
\,, {\nonumber} \\
\fd{3cm}{figures/matching_2q2g_SCET_contribution_3}
\hspace{0.4cm}
&= - \frac{g^2\, \bar u_n (1) T^b T^a \Sl{p}_{1\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{4\perp}\Sl{\epsilon}^*_{3\perp}\Sl{p}_{2\perp} v_{\bar n}(2)}{(\omega_2+\omega_3)\omega_1 \omega_2 (p_{2r}+p_{3r})}
\,, {\nonumber} \\
\fd{3cm}{figures/matching_2q2g_SCET_contribution_4}
\hspace{0.4cm}
&=- \frac{g^2\, \bar u_n (1) T^b T^a \Gamma \Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp}\Sl{\epsilon}^*_{3\perp} \Sl{p}_{2\perp} v_{\bar n}(2)}{(\omega_2+\omega_3)\omega_2 \omega_4 (p_{2r}+p_{3r})}
\,.
\end{align}
Note that we do not have to consider the operators involving ultrasoft derivatives from \Eqs{eq:soft_derivative_basis}{eq:eedggus} in the matching, since we have chosen to only have a residual momentum for the small momentum component, and we have shown by the RPI relation of \Eq{eq:uspartialRPIrelation} combined with the matching to the leading power operator in \Eq{eq:LP_wilson} that this component of the derivative first appears at $\mathcal{O}(\alpha_s)$.
Since the matrix elements in full QCD are identical to the case when the gluons are in different collinear sectors, as was just considered, here we just give the results of the diagrams evaluated to $\mathcal{O}(\lambda^2)$. Unlike the previous case where both gluons were in the same collinear sector, in this case all possible diagrams contribute at $\mathcal{O}(\lambda^2)$. Using also the direction of the gluon to indicate whether it is taken to have $n$-collinear or $\bar n$-collinear momentum, we find
\begin{align}\label{eq:qqggmatch_alldiagrams}
&\fd{3cm}{figures/matching_2q2g_diagram1_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram1_opp_low} } \right |_{\mathcal{O}(\lambda^2)}
= -\frac{g^2\, \bar u_n (1) T^b T^a \Gamma \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} v_{\bar n}(2) }{(\omega_2+\omega_3)\omega_4 }
{\nonumber} \\
&\hspace{4.5cm}
- \frac{g^2\,\bar u_n (1) T^b T^a \Gamma \Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp}\Sl{\epsilon}^*_{3\perp} \Sl{p}_{2\perp} v_{\bar n}(2) }{(\omega_2+\omega_3)\omega_2 \omega_4 (p_{2r}+p_{3r})} {\nonumber} \\
&\hspace{4.5cm}
- \frac{g^2\,\bar u_n (1) T^b T^a \Sl{p}_{1\perp}\frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{4\perp}\Sl{\epsilon}^*_{3\perp}\Sl{p}_{2\perp} v_{\bar n}(2) }{(\omega_2+\omega_3)\omega_1 \omega_2 (p_{2r}+p_{3r})}
\,,{\nonumber} \\
&\fd{3cm}{figures/matching_2q2g_diagram2_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram2_opp_low}} \right |_{\mathcal{O}(\lambda^2)}
= -\frac{g^2\, \bar u_n (1) T^a T^b \Gamma \Sl{\epsilon}^*_{3\perp} \Sl{\epsilon}^*_{4\perp} v_{\bar n}(2)}{\omega_2\omega_4 }
\,,
{\nonumber} \\
&\fd{3cm}{figures/matching_2q2g_diagram5_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram5_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
=-\frac{g^2\,\bar u_n (1) T^b T^a \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} \Gamma v_{\bar n}(2) }{(\omega_1+\omega_4)\omega_3 }
{\nonumber}\\
&\hspace{4.5cm}
-\frac{g^2\, \bar u_n (1) T^b T^a\Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{p}_{2\perp} v_{\bar n}(2)}{\omega_1 \omega_2(\omega_1+\omega_4)(p_{1r}+p_{4r})}
{\nonumber} \\
&\hspace{4.5cm}
- \frac{g^2\, \bar u_n (1) T^b T^a\Sl{p}_{1\perp} \Sl{\epsilon}^*_{4\perp} \Sl{\epsilon}^*_{3\perp} \Sl{p}_{2\perp} \Gamma v_{\bar n}(2)}{\omega_1 \omega_3(\omega_1+\omega_4)(p_{1r}+p_{4r})}
\,, {\nonumber} \\
&\fd{3cm}{figures/matching_2q2g_diagram6_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram6_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
=-\frac{g^2\,\bar u_n (1) T^a T^b \Sl{\epsilon}^*_{3\perp} \Sl{\epsilon}^*_{4\perp} \Gamma v_{\bar n}(2) }{\omega_1\omega_3 } \,, {\nonumber}\\
&\fd{3cm}{figures/matching_2q2g_diagram3_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram3_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
= -\frac{g^2\,\bar u_n (1) T^b T^a \Sl{\epsilon}^*_{4\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{3\perp} v_{\bar n}(2) }{(\omega_2+\omega_3)(\omega_1+\omega_4) }
\,, {\nonumber}\\
&\fd{3cm}{figures/matching_2q2g_diagram4_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram4_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
= \frac{-g^2\,\bar u_n (1) T^a T^b\Sl{\epsilon}^*_{3\perp} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\epsilon}^*_{4\perp} v_{\bar n}(2) }{\omega_1 \omega_2}
\,, {\nonumber} \\
&\fd{3cm}{figures/matching_2q2g_diagram7_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram7_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
=-\frac{ig^2 f^{abc} \epsilon^*_{3\perp} \cdot \epsilon^*_{4\perp}\,
\bar u_n (1) T^c \Gamma v_{\bar n}(2) }{\omega_3 \omega_4} \,, {\nonumber}\\
&\fd{3cm}{figures/matching_2q2g_diagram8_opp_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2g_diagram8_opp_low}} \right|_{\mathcal{O}(\lambda^2)}
=-\frac{ig^2 f^{abc}\epsilon^*_{3\perp} \cdot \epsilon^*_{4\perp} \,
\bar u_n (1) T^c \Gamma v_{\bar n}(2) }{\omega_3 \omega_4} \,.
\end{align}
The sum of all these contributions must be reproduced by hard scatting operators in SCET, after the non-local terms, which are easily recognizable, have been subtracted. The simplest approach is just to associate to each of the above diagrams an operator with a different Lorentz and color structures. This makes the structure the most transparent, and simplifies the projection to helicity operators. It also allows us to easily treat a completely general Dirac structure, $ \Gamma$, inserted at the vertex. We can therefore write the operators generated by the tree level matching as
\begin{align}\label{eq:matched_twogluon_diffsector}
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB1}
&= \frac{-g^2}{(\omega_2+\omega_3) \omega_4} \,
\bar \chi_{n,\omega_1} \Gamma \Sl{\mathcal{B}}_{\perp n, \omega_4} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3} \chi_{\bar n, -\omega_2}
\,, \quad
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB2}
= \frac{-g^2}{\omega_2 \omega_4} \,
\bar \chi_{n,\omega_1} \Gamma \Sl{\mathcal{B}}_{\perp \bar n, \omega_3} \Sl{\mathcal{B}}_{\perp n, \omega_4} \chi_{\bar n, -\omega_2}
\,, {\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB3}
&= \frac{-g^2}{(\omega_1+\omega_4) \omega_3} \,
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp n, \omega_4} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3} \Gamma \chi_{\bar n, -\omega_2}
\,, \quad
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB4}
= \frac{-g^2}{\omega_1 \omega_3} \,
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3}
\Sl{\mathcal{B}}_{\perp n, \omega_4} \Gamma \chi_{\bar n, -\omega_2}
\,, {\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB5}
&= \frac{-g^2}{(\omega_2+\omega_3)(\omega_1+\omega_4)} \,
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp n, \omega_4} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3} \chi_{\bar n, -\omega_2}
\,, {\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB6}
&=\frac{-g^2}{\omega_1 \omega_2} \,
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp \bar n, \omega_3} \frac{\Sl{\bar n}}{2} \Gamma \frac{\Sl{n}}{2} \Sl{\mathcal{B}}_{\perp n, \omega_4} \chi_{\bar n, -\omega_2}
\,,{\nonumber} \\
\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB7}
&= \frac{-2g^2}{\omega_3 \omega_4} \,
\bar \chi_{n,\omega_1} \bigl( \mathcal{B}_{\perp \bar n, \omega_3} \cdot \mathcal{B}_{\perp n, \omega_4} - \mathcal{B}_{\perp n, \omega_4} \cdot \mathcal{B}_{\perp \bar n, \omega_3} \bigr) \Gamma \chi_{\bar n, -\omega_2}
\,.
\end{align}
A number of these operators exhibit singularities in the Wilson coefficients as a single one of the gluons or quarks becomes soft. This behavior is distinct from that of the Wilson coefficients of the operators involving two gluons in a single collinear sector, given in \Eq{eq:matched_twogluon_samesector}, which only have a singularity when a pair of particles becomes soft. Of particular interest, are the operators $\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB2}$, $\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB4}$, $\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB6}$, $\mathcal{O}^{(2)\mu}_{\mathcal{B}\bar\cB7}$, which have a separate divergence in each collinear sector.
In \Sec{sec:proj_hel_nnlp} these operators will be projected to a helicity basis, which simplifies their structure.
\subsection{Subleading Matching with Four Collinear Quarks}\label{sec:matching_NNLP_4q}
Finally, we consider the matching to the operators involving four collinear quark fields. As discussed in \Sec{sec:eeJets}, the only relevant such operators for the cross section at $\mathcal{O}(\lambda^2)$ involve a single collinear quark operator in one sector, and three collinear quark operators in the other sector. A basis of such operators was given in \Eq{eq:Z3_basis_qQ}.
Since the operators get their power suppression from the explicit collinear quark fields, we can simplify the matching by taking the particle momenta as
\begin{align}
p_1^\mu&=\frac{\omega_1}{2} n^\mu \,, \qquad
p_2^\mu=\frac{\omega_2}{2}\bar n^\mu\,, {\nonumber} \\
p_3^\mu&=\frac{\omega_3}{2}\bar n^\mu+p_\perp^\mu+\frac{p_{3r}}{2} n^\mu\,, \qquad
p_4^\mu=\frac{\omega_4}{2}\bar n^\mu -p_\perp^\mu+\frac{p_{4r}}{2}n^\mu\,.
\end{align}
The momentum labeling corresponds to that shown in the figure in \Eq{eq:qqqqmatch_1}. This choice of momentum also removes contributions from subleading Lagrangian insertions, as well as from the $\mathcal{O}(\lambda^2)$ single gluon operators
\begin{align}
\fd{3cm}{figures/matching_2q2Q_SCET_contribution}=0\,.
\end{align}
The QCD diagrams at $\mathcal{O}(\lambda^2)$ must therefore be exactly reproduced by the hard scattering operators. As before, the matching coefficients can be calculated by expanding the QCD diagrams. In the case where a gluon splits into a quark anti-quark pair, with one in each sector, we find
\begin{align}\label{eq:qqqqmatch_1}
\fd{3cm}{figures/matching_2q2Q_diagram1_low} &= \bar{u}(p_1) (i g T^a \gamma_\nu) v(p_2) \frac{-i}{(p_1+p_2)^2} \bar{u}(p_3) \Gamma \frac{-i (\Sl{p}_1 + \Sl{p}_2 + \Sl{p}_4)}{(p_1 + p_2 + p_4)^2} (ig T^a \gamma^\nu) v(p_2) \,, {\nonumber} \\
\fd{3cm}{figures/matching_2q2Q_diagram1a_low} &= \bar{u}(p_1) (i g T^a \gamma_\nu) v(p_2) \frac{-i}{(p_1+p_2)^2} \bar{u}(p_3) (ig T^a \gamma^\nu) \frac{i (\Sl{p}_1 + \Sl{p}_2 + \Sl{p}_3)}{(p_1 + p_2 + p_3)^2} \Gamma v(p_2) \,.
\end{align}
The $\mathcal{O}(\lambda^2)$ contribution of these diagrams is given by
\begin{align}
\fd{3cm}{figures/matching_2q2Q_diagram1_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2Q_diagram1_low}}\right |_{\mathcal{O}(\lambda^2)}&= \frac{g^2}{\omega_1\omega_2 (\omega_2 + \omega_4) } \bar{u}_{n} (1) T^a \gamma_\nu v_{\bar n} (2) \bar{u}_{\bar n} (3) \Gamma \frac{\Sl {n}}{2} \gamma^\nu T^a v_{\bar n}(4) \,, {\nonumber}\\
\fd{3cm}{figures/matching_2q2Q_diagram1a_low} \left. \vphantom{\fd{3cm}{figures/matching_2q2Q_diagram1a_low} }\right |_{\mathcal{O}(\lambda^2)}&=-\frac{g^2}{\omega_1\omega_2 (\omega_2 + \omega_3) } \bar{u}_{n} (1) T^a \gamma_\nu v_{\bar n} (2) \bar{u}_{\bar n} (3) \gamma^\nu T^a \frac{\Sl {n}}{2} \Gamma v_{\bar n}(4) \,.
\end{align}
We could also have contributions from a gluon emitted from the $n$ collinear quark, which splits into two quarks, both in the $\bar n$ collinear sector. The full theory diagram is given by
\begin{align}
\fd{3cm}{figures/matching_2q2Q_diagram2_low}=\bar u (p_3) (igT^a) \gamma^\rho v(p_4) \bar u(p_1) (igT^a) \gamma^\rho i \frac{\Sl{p}_1+\Sl{p}_3+\Sl{p}_4}{(p_1+p_3+p_4)^2} \Gamma v(p_2) \frac{(-i)}{(p_3+p_4)^2}\,.
\end{align}
Expanding this, we find that its $\mathcal{O}(\lambda^2)$ term vanishes
\begin{align}
\left.\fd{3cm}{figures/matching_2q2Q_diagram2_low}\right |_{\mathcal{O}(\lambda^2)}&= 0\,.
\end{align}
Finally, there is a contribution from a gluon emitted from the $\bar n$ collinear quark, which splits into two quarks. The full theory diagram is given by
\begin{align}
\fd{3cm}{figures/matching_2q2Q_diagram3_low}=\bar u (p_3) (igT^a) \gamma^\rho v(p_4) \bar u(p_1) \Gamma (-i) \frac{\Sl{p}_2+\Sl{p}_3+\Sl{p}_4}{(p_2+p_3+p_4)^2} (igT^a) \gamma^\rho v(p_2) \frac{(-i)}{(p_3+p_4)^2}\,.
\end{align}
Expanding this to $\mathcal{O}(\lambda^2)$, we find that its contribution at this order also vanishes
\begin{align}
\left.\fd{3cm}{figures/matching_2q2Q_diagram3_low}\right |_{\mathcal{O}(\lambda^2)}&=0\,.
\end{align}
We therefore find that the QCD result is reproduced by two SCET operators
\begin{align}\label{eq:matched_4qn}
\mathcal{O}^{(2)\Gamma}_{4\chi1a}
&=\frac{g^2}{\omega_1 \omega_2 (\omega_2 + \omega_4)} \,
\Big[ \bar \chi_{n, \omega_1} T^a \gamma_\nu \chi_{\bar n,-\omega_2}\Big]
\Big[ \bar \chi_{\bar n, \omega_3} \Gamma \frac{\Sl n}{2} \gamma^\nu T^a \chi_{\bar n,-\omega_4} \Big]
\,, {\nonumber} \\
\mathcal{O}^{(2)\Gamma}_{4\chi1b}
&=-\frac{g^2}{\omega_1 \omega_2 (\omega_2 + \omega_3)}\, \Big[ \bar \chi_{n, \omega_1} T^a \gamma_\nu \chi_{\bar n,-\omega_2}\Big]
\Big[ \bar \chi_{\bar n, \omega_3} \gamma^\nu T^a \frac{\Sl n}{2} \Gamma \chi_{\bar n,-\omega_4} \Big]
\,.
\end{align}
Similarly, for the case where there is an antiquark in the $n$ collinear sector instead of a quark, the SCET operators are given by
\begin{align}\label{eq:matched_4qbarn}
\mathcal{O}^{(2)\Gamma}_{4\bar \chi1a}
&=\frac{g^2}{\omega_1 \omega_2 (\omega_2 + \omega_4) } \,
\Big[ \bar \chi_{\bar n, \omega_2} T^a \gamma_\nu \chi_{n,-\omega_1}\Big]
\Big[ \bar \chi_{\bar n, \omega_3} \Gamma \frac{\Sl n}{2} \gamma^\nu T^a \chi_{\bar n,-\omega_4} \Big]
\,, {\nonumber} \\
\mathcal{O}^{(2)\Gamma}_{4\bar \chi1b}
&=-\frac{g^2}{\omega_1 \omega_2 (\omega_2 + \omega_3) } \,
\Big[ \bar \chi_{\bar n, \omega_2} T^a \gamma_\nu \chi_{n,-\omega_1} \Big] \Big[ \bar \chi_{\bar n, \omega_3} \gamma^\nu T^a \frac{\Sl n}{2} \Gamma \chi_{\bar n,-\omega_4} \Big]
\,.
\end{align}
Both operators exhibit singularities as certain quarks go soft.
\subsection{Subleading Matching with a Single Ultrasoft Gluon}\label{sec:match_soft}
In this section we consider the matching with a the emission of a single ultrasoft gluon, which allows us to probe the operators involving ultrasoft insertions. These operators are determined by RPI, so the determination of their Wilson coefficients does not require a new calculation, nevertheless it is instructive to see how this matching works.\footnote{Note that we have also shown that they do not contribute to the dijet cross section at $\mathcal{O}(\alpha_s \lambda^2)$.} For the corresponding tree level matching computation, the necessary SCET diagram calculations were carried out explicitly in Ref.~\cite{Larkoski:2014bxa} for an arbitrary current. For completeness we give the full theory and SCET results needed for the matching calculation in the case studied here. This illustrates how contributions from the SCET Lagrangian insertions are treated in the matching, as discussed in general in \Sec{sec:matching_general}, which can not be avoided in this case. We carry out this matching prior to making the BPS field redefinition.
To perform the matching calculation, we route the residual momenta such that it is only carried by the external ultrasoft particle, any intermediate collinear fields, and then out through the hard current. We take
\begin{align}
p_1 =\frac{\omega_1}{2} n^\mu\,, \qquad p_2 =\frac{\omega_2}{2} \bar n^\mu\,, \qquad p_s=\frac{n \cdot p_s}{2}\bar n^\mu+p_{s\perp} +\frac{\bar n \cdot p_s}{2}n^\mu\,,
\end{align}
where the momentum $p_s$ is ultrasoft and hence purely residual.
Expanding the QCD diagrams, we find
\begin{align}
\fd{3cm}{figures/matching_subleadingvertex_2_soft_low}
&=\frac{-g \epsilon_\mu^* \bar u_n(1) T^a
\Bigl[ (\omega_1 \!+\! {\bf n}\cdot p_s) n^\mu + \gamma_\perp^\mu \Sl{p}_{s\perp}
\! + (\gamma_\perp^\mu n\cdot p_s \!-\! \Sl{p}_{s\perp}\! n^\mu)
\frac{\slashed{\bar{n}}}{2}
\Bigr] \Gamma v_{\bar n}^*(2)
}{\omega_1\, n\cdot p_s}
,{\nonumber}\\
\fd{3cm}{figures/matching_subleadingvertex_1_soft_low}
&= \frac{g \epsilon_\mu^* \bar u_n(1) T^a \Gamma
\Bigl[ (\omega_2 \!+\! n\cdot p_s){\bf n}^\mu + \Sl{p}_{s\perp}\gamma_\perp^\mu
\!+ \frac{\slashed{n}}{2} (\gamma_\perp^\mu {\bf n}\cdot p_s
\!-\! \Sl{p}_{s\perp}\! {\bf n}^\mu)
\Bigr] v_{\bar n}^*(2)
}{\omega_2\, {\bf n}\cdot p_s} .
\end{align}
Next we drop the standard eikonal ${\cal O}(\lambda^0)$ terms, and using the equations of motion $n\cdot p_s {\bf n}\cdot p_s = -p_{s\perp}^2$ to write the result in a form which is clearly gauge invariant graph by graph for the ${\cal O}(\lambda^2)$ terms. This gives
\begin{align} \label{eq:fullsoft1gluonmatch}
\fd{3cm}{figures/matching_subleadingvertex_2_soft_low} \left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_2_soft_low}}\right |_{\mathcal{O}(\lambda^2)}
& \!\! = \frac{-g\epsilon_\mu^* \bar u_n(1) T^a
}{\omega_1\, n\cdot p_s}
\biggl\{
\frac{p_{s\perp}^\mu n\!\cdot\! p_s \!-\! p_{s\perp}^2\! n^\mu}{n\!\cdot\! p_s}
+ \frac{[\gamma_\perp^\mu,\Sl{p}_{s\perp}]}{2}
+ \frac{(\gamma_\perp^\mu n\!\cdot\! p_s \!-\! \Sl{p}_{s\perp}\! n^\mu)\slashed{\bar{n}}}{2}
\biggr\}
{\nonumber}\\
& \quad \times \Gamma v_{\bar n}^*(2)
\,, {\nonumber}\\[5pt]
\fd{3cm}{figures/matching_subleadingvertex_1_soft_low} \left. \vphantom{\fd{3cm}{figures/matching_subleadingvertex_1_soft_low}}\right |_{\mathcal{O}(\lambda^2)}
& \!\! = \frac{g\epsilon_\mu^* \bar u_n(1) \Gamma
}{\omega_2\, {\bf n}\cdot p_s}
\biggl\{
\frac{p_{s\perp}^\mu {\bf n}\!\cdot\! p_s \!-\! p_{s\perp}^2\! {\bf n}^\mu}
{{\bf n}\!\cdot\! p_s}
- \frac{[\gamma_\perp^\mu,\Sl{p}_{s\perp}]}{2}
+ \frac{\slashed{n}(\gamma_\perp^\mu {\bf n}\!\cdot\! p_s \!-\! \Sl{p}_{s\perp}\! {\bf n}^\mu)}{2}
\biggr\}
{\nonumber}\\
& \quad \times T^a v_{\bar n}^*(2)
\,.
\end{align}
These results are written in the same form given in Ref.~\cite{Larkoski:2014bxa}, which is also the form predicted by the LBK relation~\cite{Low:1958sn,Burnett:1967km}.
Unlike our earlier analysis of the purely collinear graphs, with a ultrasoft particle present there are residual momenta in collinear propagators. Therefore, in the matching with \Eq{eq:lamp_match} there are non-vanishing contributions from insertions of the subleading power SCET Lagrangians. Insertions of $\mathcal{L}^{(1)}$ vanish as they involve label $\perp$ momentum, which we chose to be zero. The Feynman rules for the corrections from $\mathcal{L}^{(2)}$ to an $n$-collinear quark propagator, and to the emission of a ultrasoft gluon are given by
\begin{align}
&\fd{3cm}{figures/feynman_rule_subsubleading_prop}
\: =i \frac{\Sl{\bar n}}{2}\frac{p_{r\perp}^2}{\bar n\cdot p}
\,,
&
\raisebox{0.3cm}{\fd{2.7cm}{figures/feynman_rule_soft_emission_subsub}}
& \: =igT^a \biggl( \frac{\gamma^\mu_\perp\Sl{p}_{1r\perp}}{\bar n \cdot p} + \frac{\Sl{p}_{2r\perp}\gamma^\mu_\perp}{\bar n \cdot p} \biggr) \frac{\slashed{\bar{n}}}{2}.
\end{align}
Computing the SCET diagrams that contribute to the matrix element, we find that the propagator insertions of $\mathcal{L}^{(2)}$ and corrections to the gluon emission from $\mathcal{L}^{(2)}$ give
\begin{align}
& \fd{3cm}{figures/matching_scet1g_nnlp_2_soft_prop} +\fd{3cm}{figures/matching_scet1g_nnlp_1_soft_prop}
=g\biggl(
-\frac{n \cdot \epsilon^*\, \bar n \cdot p_s}{\omega_1\,n\cdot p_s}
+\frac{\bar n \cdot \epsilon^*\, n \cdot p_s}{\omega_2\,\bar n\cdot p_s}
\biggr) \bar u_n (1) T^a \Gamma v_{\bar n}^* (2)
\,, {\nonumber}\\
& \fd{3cm}{figures/matching_scet1g_nnlp_2_soft} +\fd{3cm}{figures/matching_scet1g_nnlp_1_soft}
=g\, \bar u_n (1) T^a \biggl(
- \frac{\Sl{\epsilon}^*_\perp \Sl{p}_{s\perp} \Gamma}
{\omega_1\,n\cdot p_s}
+ \frac{\Gamma \Sl{p}_{s\perp} \Sl{\epsilon}^*_\perp}
{\omega_2\, {\bf n} \cdot p_s}
\biggr) v_{\bar n}^* (2)
\,,{\nonumber}\\
& \text{Sum}
= \frac{ g \bar u_n(1) T^a \Gamma}{\omega_2\, {\bf n}\cdot p_s}
\biggl\{ \frac{p_{s\perp}^\mu {\bf n}\!\cdot\! p_s \!-\! p_{s\perp}^2\! {\bf n}^\mu}
{{\bf n}\!\cdot\! p_s} - \frac{[\gamma_\perp^\mu,\Sl{p}_{s\perp}]}{2}
\biggr\} v_{\bf n}^*(2)
{\nonumber}\\
&\qquad\
- \frac{ g \bar u_n(1) T^a}{\omega_1\, n\cdot p_s}
\biggl\{ \frac{p_{s\perp}^\mu n\!\cdot\! p_s \!-\! p_{s\perp}^2\! n^\mu}{n\!\cdot\! p_s} + \frac{[\gamma_\perp^\mu,\Sl{p}_{s\perp}]}{2}
\biggr\} \Gamma v_{\bf n}^*(2)
\,.
\end{align}
Note that the sum is gauge invariant for each leg. Subtracting off these SCET matrix elements leaves only the third term in each line of the full theory results in \eq{fullsoft1gluonmatch}. These terms are reproduced by the SCET operators
\begin{align}\label{eq:soft_current_perp}
\mathcal{O}^{(2)\Gamma}_{s\perp1}&
= \bar \chi_{n,\omega_1} \Gamma\, \frac{i\Sl{D}_{\perp us}}{(-\omega_2)}
\frac{\Sl{n}}{2} \chi_{\bar n, -\omega_2}
\,,
& \mathcal{O}^{(2)\Gamma}_{s\perp2}&
= \bar \chi_{n,\omega_1} \frac{\Sl{\bar n}}{2}
\frac{(-i \overleftarrow{\Sl{D}}_{\perp us})}{\omega_1}
\Gamma \chi_{\bar n, -\omega_2}\,.
\end{align}
Each of these operators gives two contributions, one where the $D_\perp$ produces a ultrasoft gluon at the vertex, and one where the $D_\perp$ acts as a derivative, and the ultrasoft gluon is produced by an insertion of $\mathcal{L}^{(0)}$:
\begin{align}
\Bigl\langle \mathcal{O}^{(2)\Gamma}_{s\perp1} \Bigr\rangle
&= g\, \bar u_n(1) \Gamma T^a \frac{\slashed{n}}{2} \biggl( \frac{\Sl{\epsilon}_\perp}{\omega_2}
-\frac{\Sl{p}_{s\perp}{\bf n}\cdot\epsilon^*}{\omega_2\,{\bf n}\cdot p_s} \biggr)
v_{\bf n}^*(2)
\,,{\nonumber}\\[5pt]
\Bigl\langle \mathcal{O}^{(2)\Gamma}_{s\perp2} \Bigr\rangle
&= - g\, \bar u_n(1) T^a \biggl(
\frac{\Sl{\epsilon}_\perp}{\omega_1}
-\frac{\Sl{p}_{s\perp} n\cdot\epsilon^*}{\omega_1\, n\cdot p_s} \biggr)
\frac{\slashed{\bar{n}}}{2} \Gamma v_{\bf n}^*(2)
\,.
\end{align}
The Wilson coefficients of these operators are both fixed by RPI to all orders in perturbation theory, and our tree level matching result agrees with this relation.
We see that only the $D_\perp$, or correspondingly in terms of helicity operators, the $\partial_{\pm}$ and $\mathcal{B}_{\pm}$, are required to reproduce the tree level QCD result. Operators of the form $\bar \chi_{\bar n, \omega_1} i n\cdot D_{us} \Gamma \chi_{n, \omega_2}$ or $\bar \chi_{\bar n, \omega_1} i \bar n\cdot \overleftarrow{D}_{us} \Gamma \chi_{n, \omega_2}$ are turned into purely collinear operators by the leading power SCET quark equations of motion, and therefore are not included in the basis, as was discussed for the helicity operators in \Sec{sec:soft_basis}.
Prior to the BPS field redefinition, the operators involving the other components of $D^\mu_{us}$, which could appear at the one gluon level are
\begin{align}
\mathcal{O}_{s{\bf n}}^{(2)\Gamma} &= C_{s{\bf n}}^{(2)}\:
\bar \chi_{n, \omega_1} (-i {\bf n}\cdot \overleftarrow{D}_{us})
\Gamma \chi_{{\bf n}, -\omega_2}
\,,
&\mathcal{O}_{sn}^{(2)\Gamma} &= C_{s n}^{(2)}\:
\bar \chi_{n, \omega_1} \Gamma i n\cdot D_{us}
\chi_{{\bf n}, -\omega_2}
\,,
\end{align}
but these two operators do not show up in the tree level matching since their coefficients $C_{sn}^{(2)}$ and $C_{s{\bf n}}^{(2)}$ are ${\cal O}(\alpha_s)$. As noted in \Sec{sec:soft_basis}, the Wilson coefficients of these two operators are related to derivatives of the leading power operator. Since the Wilson coefficient of the leading power operator is $1$ at tree level, the derivative RPI relation implies that the Wilson coefficient of these operators vanish, which is reproduced by our explicit tree level matching. Using the RPI relations of \Sec{sec:soft_basis}, we can determine the Wilson coefficients of these operators at one-loop. Taking derivatives of the leading power Wilson coefficient of \Eq{eq:LP_wilson} gives
\begin{align}\label{eq:wilson_coeff_soft_mu}
C_{sn}^{(2)}=\frac{\partial C^{(0)}}{d\omega_1}
&=-\frac{\alpha_s C_F}{4\pi} \frac{1}{\omega_1}
\biggl[ -2 \ln \Bigl( \frac{-\omega_1\omega_2-i0}{\mu^2} \Bigr) + 3 \biggr]
+ {\cal O}(\alpha_s^2)
\,, \\
C_{s\bar n}^{(2)}=\frac{\partial C^{(0)}}{d\omega_2}
&=-\frac{\alpha_s C_F}{4\pi} \frac{1}{\omega_2}
\biggl[ -2 \ln \Bigl( \frac{-\omega_1\omega_2-i0}{\mu^2} \Bigr) +3 \biggr]
+ {\cal O}(\alpha_s^2)
\,. {\nonumber}
\end{align}
In the calculation of the cross section, conservation of momentum will enforce $\omega_1=\omega_2$, so that the Wilson coefficients of these two operators are identical. This will hold to all orders in perturbation theory. Combined with the arguments of \Sec{sec:ee_lambda2}, showing that the $\perp$ ultrasoft operators don't contribute to the $e^+e^-\to$ dijets cross section at $\mathcal{O}(\lambda^2)$, this shows that ultrasoft operators will first contribute at $\mathcal{O}(\alpha_s^2)$. The $\mathcal{O}(\alpha_s)$ ultrasoft contributions therefore arise only from the subleading SCET Lagrangian.
\subsection{Projection onto Helicities}\label{sec:proj_hel}
Given the general result for the Wilson coefficients in SCET, in this section we project the operators onto helicities to show how they appear in the helicity operator basis. For simplicity, we will do this projection for the specific case of a vector current, namely $ \Gamma=\gamma^\mu$, as relevant for $e^+e^-\to $ dijets proceeding through an off shell photon. We then compare with the operators of the helicity basis given in \Sec{sec:eeJets}.
In setting up the decomposition of the Wilson coefficients, as given, for example, for the leading power operator in \Eq{eq:Z0_expand_Wilson}, the couplings of the photon to both the electrons and the quarks has been extracted. To extract the Wilson coefficient and helicity operators from one of the currents, $\mathcal{O}_j^{(i)\nu}$ we can simply expand the contraction
\begin{align}\label{eq:contraction_master}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}_j^{(i)\nu}\,,
\end{align}
using the completeness relation for polarization vectors
\begin{align}\label{eq:completeness2}
\sum\limits_{\lambda=\pm} \epsilon^\lambda_\mu(n_i,\bar n_i) \left (\epsilon^\lambda_\nu(n_i,\bar n_i) \right )^*=-g_{\mu \nu}+\frac{n_{i\mu} \bar n_{i\nu}+n_{i\nu} \bar n_{i\mu}}{n_i\cdot \bar n_i} = - g^\perp_{\mu \nu} ( n_i, {\bf n}_i)\,.
\end{align}
Here we have taken $s=Q^2$, so that a large light cone momentum $Q$ is deposited in each hemisphere.
The helicity operator basis of \Sec{sec:eeJets} involves helicity operators defined with respect to the axes of the jets, namely $n$ and $\bar n$, as well as with respect to the axes of the incoming electrons, namely $n_e$ and $\bar n_e$. In performing the projection in \Eq{eq:contraction_master}, the non-trivial angular dependence of the Wilson coefficient arises from inner products of the polarization vectors with respect to the axis of the electrons and those defined with respect to the axis of the jets. While we will not in general expand these inner products of polarization vectors, as they allow for an intuitive interpretation of each term, we note that they can straightforwardly be written in terms of spinor products as
\begin{align}\label{eq:polar_innerprod}
\epsilon_+(n,\bar n)\cdot \epsilon_+ (n_e, \bar n_e) &=\frac{[n n_e]\langle \bar n_e \bar n\rangle}{\langle \bar n n \rangle \langle \bar n_e n_e\rangle}\,, \qquad
\epsilon_+(n,\bar n)\cdot \epsilon_- (n_e, \bar n_e) =-\frac{[n \bar n_e]\langle n_e \bar n\rangle}{\langle \bar n n \rangle [ \bar n_e n_e ]}\,, {\nonumber}\\
\epsilon_-(n,\bar n)\cdot \epsilon_+ (n_e, \bar n_e) &=-\frac{[\bar n n_e]\langle \bar n_e n \rangle}{[\bar n n] \langle\bar n_e n_e \rangle}\,, \qquad
\epsilon_-(n,\bar n)\cdot \epsilon_- (n_e, \bar n_e) =\frac{[\bar n \bar n_e]\langle n_e n \rangle}{[ \bar n n ] [\bar n_e n_e ]}\,,
\end{align}
Since these factors will appear whenever two spin $1$ currents are projected onto each other, we will define a shorthand notation
\begin{align}
\epsilon_+(n,\bar n)\cdot \epsilon_+ (n_e, \bar n_e) =d_{++}\,, \qquad \epsilon_+(n,\bar n)\cdot \epsilon_- (n_e, \bar n_e) = d_{+-}\,, {\nonumber} \\
\epsilon_-(n,\bar n)\cdot \epsilon_+ (n_e, \bar n_e) =d_{-+}\,, \qquad \epsilon_-(n,\bar n)\cdot \epsilon_- (n_e, \bar n_e) = d_{--}\,.
\end{align}
Similarly, inner products of the $n$ and $\bar n$ vectors with the electron polarization vectors can be expanded as
\begin{align}
n\cdot \epsilon_+ (n_e, \bar n_e) &= \frac{[n_e n] \langle n \bar n_e \rangle}{\sqrt{2} \langle \bar n_e n_e \rangle }\,, \qquad \bar n\cdot \epsilon_+ (n_e, \bar n_e) = \frac{[n_e \bar n] \langle \bar n \bar n_e \rangle}{\sqrt{2} \langle \bar n_e n_e \rangle }\,, {\nonumber} \\
n\cdot \epsilon_- (n_e, \bar n_e) &= -\frac{\langle n_e n\rangle [ n \bar n_e ]}{\sqrt{2} [ \bar n_e n_e ] }\,, \qquad \bar n\cdot \epsilon_- (n_e, \bar n_e) = -\frac{\langle n_e \bar n\rangle [ \bar n \bar n_e ]}{\sqrt{2} [ \bar n_e n_e ] }\,.
\end{align}
We will occasionally perform this expansion if it simplifies the result.
The spinor products in general give a phase
\begin{align}
\langle i j \rangle =\sqrt{|s_{ij}|}e^{i\phi_{ij}}\,, \qquad [ij]=\sqrt{|s_{ij}|}e^{-i(\phi_{ij}+\pi)}\,,
\end{align}
see \App{app:helicity} for details. To slightly simplify the expressions, we will choose to define our spinors with respect to to the jet axis $n$, which we take to be in the $z$ direction, $n^\mu=(1,0,0,1)$. In this case we simplify the spinor products between $n$ and $\bar n$
\begin{align}
[n \bar n]=-\sqrt{2}\,, \qquad \langle n \bar n \rangle=\sqrt{2} \,.
\end{align}
We also note that depending on the collinear sector of the operator, the helicities are defined either with respect to the $n$ or $\bar n$ axes. The relation
\begin{align}
\epsilon^\mu_\pm(n, \bar n)=\epsilon^\mu_\mp(\bar n, n)\,,
\end{align}
allows for the trivial exchange of the corresponding decompositions.
Finally, we comment on the organization of this section. Instead of providing tables of the Wilson coefficients, we find that it is more transparent to show the decomposition of each of the currents generated in the tree level matching onto the helicity operators. The way we have performed the matching, we have typically associated a single Feynman diagram in the matching with an SCET current. Therefore, each of these currents typically projects onto a single helicity operator. The Wilson coefficients for a particular helicity operator can then trivially be read off from the corresponding projection.
\subsubsection{Leading Power}\label{sec:proj_hel_lp}
We begin by considering the projection onto the leading power helicity basis of \Sec{sec:LP}. Since this is the first operator we are considering, we go through the projection onto the helicity basis in slightly more detail. Using the completeness relation for the polarization vectors of \Eq{eq:completeness2},
as well as the projection relations for the SCET spinors, we find
\begin{align}
&\hspace{-1cm}\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(0)\nu}={\nonumber} \\
&\left( \frac{-i}{Q^2} \right) \epsilon_-(n_e, \bar n_e)\cdot \epsilon_+(n, \bar n) \left( \epsilon_+^\mu(n_e, \bar n_e) \bar v_e \gamma^\mu u_e \right) \left( \epsilon_-^\rho (n,\bar n) \mathcal{O}^{(0)}_\rho \right){\nonumber} \\
+&\left( \frac{-i}{Q^2} \right) \epsilon_+(n_e, \bar n_e)\cdot \epsilon_+(n, \bar n) \left( \epsilon_-^\mu(n_e, \bar n_e) \bar v_e \gamma^\mu u_e \right) \left( \epsilon_-^\rho (n,\bar n) \mathcal{O}^{(0)}_\rho \right){\nonumber} \\
+&\left( \frac{-i}{Q^2} \right) \epsilon_-(n_e, \bar n_e)\cdot \epsilon_-(n, \bar n) \left( \epsilon_+^\mu (n_e, \bar n_e) \bar v_e \gamma^\mu u_e \right) \left( \epsilon_+^\rho (n,\bar n) \mathcal{O}^{(0)}_\rho \right){\nonumber} \\
+&\left( \frac{-i}{Q^2} \right) \epsilon_+(n_e, \bar n_e)\cdot \epsilon_-(n, \bar n) \left( \epsilon_-^\mu (n_e, \bar n_e) \bar v_e \gamma^\mu u_e \right) \left( \epsilon_+^\rho (n,\bar n) \mathcal{O}^{(0)}_\rho \right)\,.
\end{align}
Here we see explicitly the four different helicity combinations which arise from the single operator, and that the Wilson coefficient arises as the inner product of the helicity vectors defined with the different axes, as expected form the point of view of spin projection.
Written in terms of the helicity currents, we have
\begin{align}\label{eq:LP_helicity_matched}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(0)\nu}&= \delta_{\alpha {\bar \beta}}\left( \frac{-i}{\sqrt{2}} \right) \left (d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J_{\bar q q +}^{{\bar \alpha} \beta} \right. - d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J_{\bar q q +}^{{\bar \alpha} \beta} \right. \right. \left. \right. {\nonumber} \\
&\hspace{1cm}\left. +d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J_{\bar q q -}^{{\bar \alpha} \beta} \right. -d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J_{\bar q q -}^{{\bar \alpha} \beta} \right. \right) \,.
\end{align}
We therefore see that each of the helicity combinations in the leading power operator basis of \Eq{eq:LP_basis} is reproduced, as required from the $C$ and $P$ relations. The Wilson coefficients can then immediately be read off of \Eq{eq:LP_helicity_matched}. Note that the spinor products appearing in this expression can be simplified given a particular choice of reference axes, but we have chosen to leave them in this general form.
\subsubsection{Subleading Power}\label{sec:proj_hel_nlp}
We now consider the projection onto the $\mathcal{O}(\lambda)$ subleading power helicity operator basis of \Sec{sec:sub}.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q )_{\bar n} ( qg)_{n}}$ and $\boldsymbol{(\bar qg )_{\bar n} ( q)_{n}:}$
}
We begin by considering the case where the quark and antiquark are in different collinear sectors. The operators found in \Eqs{eq:matched_1gluon_barn}{eq:matched_1gluon_n} can be expanded in terms of the helicity currents. We have
\begin{align}
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp n} \chi_{\bar n,-\omega_2} &=\sqrt{\omega_1 \omega_2} J_{n\bar n -}\mathcal{B}_{n+} + \sqrt{\omega_1 \omega_2} J_{n\bar n +}\mathcal{B}_{n-}\,, {\nonumber} \\
\bar \chi_{n,\omega_1} \Sl{\mathcal{B}}_{\perp \bar n,\omega_3} \chi_{\bar n,-\omega_2} &=\sqrt{\omega_1 \omega_2} J_{n\bar n -}\mathcal{B}_{\bar n-} + \sqrt{\omega_1 \omega_2} J_{n\bar n +}\mathcal{B}_{\bar n +}\,,
\end{align}
where the helicity of the gluon fields is opposite in the two cases due to the fact that it is defined with respect to different axes.
From this, we immediately obtain the helicity expansions of the two operator. In the case that the gluon is in the $n$ collinear sector, we have
\begin{align}\label{eq:proj_1Bn}
&\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(1)\nu}_{\mathcal{B} n}=gT^a_{\alpha {\bar \beta}}\left(\frac{-i \sqrt{\omega_1 \omega_2}}{Q}\right)\\
&\left [ [\bar n_e n_e] \left( \frac{(\omega_1+\omega_3) n\cdot \epsilon^-(n_e, \bar n_e) -\omega_2 \bar n\cdot \epsilon^-(n_e, \bar n_e) }{\omega_2 (\omega_1+\omega_3)} \right) \left(J_{\bar e e -} J^{{\bar \alpha} \beta}_{n\bar n -}\mathcal{B}^a_{n+} + J_{\bar e e -} J^{{\bar \alpha} \beta}_{n\bar n +}\mathcal{B}^a_{n-} \right) \right. {\nonumber} \\
&\left .+ \langle \bar n_e n_e \rangle \left( \frac{(\omega_1+\omega_3) \cdot \epsilon^+(n_e, \bar n_e) -\omega_2 \bar n\cdot \epsilon^+(n_e, \bar n_e) }{\omega_2 (\omega_1+\omega_3)} \right) \left(J_{\bar e e +} J^{{\bar \alpha} \beta}_{n\bar n -}\mathcal{B}^a_{n+} + J_{\bar e e +} J^{{\bar \alpha} \beta}_{n\bar n +}\mathcal{B}^a_{n-} \right) \right ]{\nonumber}\,,
\end{align}
and in the case that the gluon is in the $\bar n$ collinear sector, we find
\begin{align}\label{eq:proj_1Bnbar}
&\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(1)\nu}_{\mathcal{B} \bar n}=gT^a_{\alpha {\bar \beta}}\left(\frac{-i \sqrt{\omega_1 \omega_2}}{Q}\right) \\
&\left [ [\bar n_e n_e] \left( \frac{\omega_1 n\cdot \epsilon^-(n_e, \bar n_e) -(\omega_2+\omega_3) \bar n\cdot \epsilon^-(n_e, \bar n_e) }{\omega_1 (\omega_2+\omega_3)} \right) \left(J_{\bar e e -} J^{{\bar \alpha} \beta}_{n\bar n -}\mathcal{B}^a_{\bar n-} + J_{\bar e e -} J^{{\bar \alpha} \beta}_{n\bar n +}\mathcal{B}^a_{\bar n+} \right) \right . {\nonumber} \\
&\left . + \langle \bar n_e n_e \rangle \left( \frac{\omega_1 \cdot \epsilon^+(n_e, \bar n_e) -(\omega_2+\omega_3) \bar n\cdot \epsilon^+(n_e, \bar n_e) }{\omega_1 (\omega_2+\omega_3)} \right) \left( J_{\bar e e +} J^{{\bar \alpha} \beta}_{n\bar n -}\mathcal{B}^a_{\bar n-} + J_{\bar e e +} J^{{\bar \alpha} \beta}_{n\bar n +}\mathcal{B}^a_{\bar n+} \right) \right ] \,.{\nonumber}
\end{align}
Note that while the Wilson coefficients of the individual helicity operators are more complicated, and explicitly involve spinor products which incorporate the angular dependence of the scattering process, the helicity structure is very simple. Indeed, comparing the operators generated, with those in the basis of
\Eq{eq:Z1_basis}, we explicitly see that the helicity selection rules are respected by the tree level matching. This will continue to hold to all orders in perturbation theory.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q q )_{\bar n} ( g)_{n}:}$
}
We now consider the operator with a single collinear gluon, where both the quarks are in the same sector. In performing the matching, we found the two operators of \Eq{eq:match_2qsame}.
Projecting these onto helicities, we find
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(1)\nu}_{\chi\chi\bar n1}&= T^a_{\alpha {\bar \beta}}\frac{-g}{\omega_1} \left( \frac{-i\sqrt{2 \omega_1 \omega_2}}{Q} \right)\, \\
&\left [ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} \mathcal{B}^a_{n+} J^{{\bar \alpha} \beta}_{\bar n 0} \right. \right .- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} \mathcal{B}^a_{n+} J^{{\bar \alpha} \beta}_{\bar n 0} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} \mathcal{B}^a_{n-} J^{{\bar \alpha} \beta}_{\bar n \bar 0} \right. \left .- d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} \mathcal{B}^a_{n-} J^{{\bar \alpha} \beta}_{\bar n \bar 0} \right. \right ] \,,{\nonumber}
\end{align}
and
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(1)\nu}_{\chi\chi\bar n2}&=T^a_{\alpha {\bar \beta}} \frac{g}{\omega_2} \left( \frac{-i\sqrt{2 \omega_1 \omega_2}}{Q} \right) \\
&\left [ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} \mathcal{B}^a_{n+} J^{{\bar \alpha} \beta}_{\bar n \bar 0} \right. \right. - d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} \mathcal{B}^a_{n+} J^{{\bar \alpha} \beta}_{\bar n \bar 0} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} \mathcal{B}^a_{n-} J^{{\bar \alpha} \beta}_{\bar n 0} \right. \left.- d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} \mathcal{B}^a_{n-} J^{{\bar \alpha} \beta}_{\bar n 0} \right. \right ]\,. {\nonumber}
\end{align}
Note that all the operators expected in the basis of \Eq{eq:Z1_basis_diff} are generated, although half come from the first Lorentz structure generated in the tree level matching, while half come from the second Lorentz structure. Again, the expected helicity selection rules, namely that the two quarks have net helicity zero is respected.
\subsubsection{Subsubleading Power}\label{sec:proj_hel_nnlp}
We now consider the projection of the operators found in the tree level matching onto the $\mathcal{O}(\lambda^2)$ helicity operator basis of \Sec{sec:subsub}.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q g \mathcal{P}_\perp)_{\bar n} ( q)_{n}:}$
}
We first consider the operators involving an insertion of the $\mathcal{P}_\perp$ operator, which contribute to the cross section at $\mathcal{O}(\lambda^2)$. The basis of helicity operators was given in \Eq{eq:eeqqgpperp_basis}, and they involve a single collinear gluon field, and a single insertion of the $\mathcal{P}_\perp$ operator.
The two currents found in the tree level matching were given in \Eq{eq:matched_nbarP}. We find
\begin{align}
\hspace{0cm}\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{P}\bar n1}&=-T^a_{\alpha {\bar \beta}} \frac{g}{\omega_1 \omega_3}\left( \frac{-i\sqrt{2 \omega_1 \omega_2} }{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n \bar n+} [ \mathcal{B}^a_{\bar n-} \mathcal{P}_{+}^\dagger] \right. \right. - d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n \bar n+} [ \mathcal{B}^a_{\bar n-} \mathcal{P}_{+}^\dagger] \right.{\nonumber} \\
&+d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n \bar n -} [\mathcal{B}^a_{\bar n+} \mathcal{P}_{-}^\dagger] \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n \bar n -} [ \mathcal{B}^a_{\bar n+} \mathcal{P}_{-}^\dagger] \right. \right] \,,{\nonumber}
\end{align}
and
\begin{align}
\hspace{0cm}\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{P}\bar n2}&= T^a_{\alpha {\bar \beta}}\frac{g}{\omega_1 \omega_2}\left( \frac{-i\sqrt{2 \omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n \bar n +} [ \mathcal{B}^a_{\bar n-} \mathcal{P}^\dagger_{-} ] \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n \bar n +}[ \mathcal{B}^a_{\bar n-} \mathcal{P}^\dagger_{-}] \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n \bar n -} [\mathcal{B}^a_{\bar n+} \mathcal{P}^\dagger_{+} ] \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n \bar n -} [\mathcal{B}^a_{\bar n+} \mathcal{P}^\dagger_{+}] \right. \right] \,. {\nonumber}
\end{align}
We see that these project down into two different classes of helicity operators. Of the operators in the helicity basis of \Eq{eq:eeqqgpperp_basis} the first four are absent, while the last eight obtain non-zero matching coefficients at tree level. As expected all expected helicity selection rules are respected.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q )_{\bar n} (gg q)_{n}:}$
}
We now consider the projection onto the helicity operators for the case of two collinear gluons in the same collinear sector. A basis of helicity operators for this case was given in \Eq{eq:eeqqgg_basis1}.
The operators arising in the tree level matching were given in \Eq{eq:matched_twogluon_samesector}. Projecting onto helicities, we find
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B} \cB1}&=\frac{1}{2} \left( (T^a T^b)_{\alpha {\bar \beta}} +(T^b T^a)_{\alpha {\bar \beta}} \right) \frac{g^2 \sqrt{ \omega_1 \omega_2}}{\omega_2(\omega_1+\omega_3)} \left( \frac{-i\sqrt{2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n+} \mathcal{B}^a_{n-} \mathcal{B}^b_{n-} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{n-} \mathcal{B}^b_{n-} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{n+} \mathcal{B}^b_{+} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{n+} \mathcal{B}^b_{n+} \right. \right] \,, {\nonumber}
\end{align}
where we have included the appropriate symmetry factor. And
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B}\cB2}&=\frac{1}{2} \left( (T^a T^b)_{\alpha {\bar \beta}} +(T^b T^a)_{\alpha {\bar \beta}} \right) \frac{-g^2 \sqrt{ \omega_1 \omega_2}}{\omega_2(\omega_3+\omega_4)} \left( \frac{-i\sqrt{2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n-} \mathcal{B}^a_{n-} \mathcal{B}^b_{n+} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{n-} \mathcal{B}^b_{n+} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{n+} \mathcal{B}^b_{-} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{n+} \mathcal{B}^b_{n-} \right. \right] \,. {\nonumber}
\end{align}
Again we see that of the helicity combinations present in the basis of \Eq{eq:eeqqgg_basis1}, four are not present, while the other eight are generated in the tree level matching.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q g )_{\bar n} ( qg)_{n}:}$
}
Similar projections apply to the case when the two gluons are in different collinear sectors. A basis of helicity operators for this case was given in \Eq{eq:eeqqgg_basis2}. The operators arising in the tree level matching were given in \Eq{eq:matched_twogluon_diffsector}. Because of the way we have organized the operators, namely each corresponding to a particular Feynman diagram, the helicity structure of the Feynman diagram is reflected in particular helicity correlations in the projected helicity operator. However, summing over all the Feynman diagrams, we will generate all the different helicity combinations in our helicity basis.
We will write the color structure of the operators in terms of the color basis of \Eq{eq:ggqqll_color}. The organization of the operators in \Eq{eq:matched_twogluon_diffsector} corresponds quite closely to this color decomposition, and therefore the projection of each operator will typically contribute to only one of the elements of the color basis.
We begin by projecting onto helicity operators the operators $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB1}$ and $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB2}$ of \Eq{eq:matched_twogluon_diffsector}. We find
\begin{align}\label{eq:2B_proj_1}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB1}&= \frac{-g^2}{(\omega_2+\omega_3) \omega_4} (T^b T^a)_{\alpha {\bar \beta}} \left( \frac{-i\sqrt{2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n+} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n+} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n-} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n-} \right. \right] \,,{\nonumber}
\end{align}
and
\begin{align}\label{eq:2B_proj_2}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB2}&= \frac{-g^2}{\omega_2 \omega_4} (T^a T^b)_{\alpha {\bar \beta}} \left( \frac{-i\sqrt{ 2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n-} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n+} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n-} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{n+} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{n+} \right. \right] \,, {\nonumber}
\end{align}
which is identical up to a relabeling of $n \leftrightarrow \bar n$, and flipping the helicities correspondingly for the gluon fields. Note that because the two gluon helicity fields are in the $\bar n$ and $n$ directions, the helicities of the $\mathcal{B}$ fields are defined with respect to different axes.
We can similarly project the operators $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB3}$ and $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB4}$ onto the helicity operator basis. We find
\begin{align}\label{eq:2B_proj_3}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB3}&= \frac{-g^2}{(\omega_1+\omega_4) \omega_3} (T^b T^a)_{\alpha {\bar \beta}} \left( \frac{-i\sqrt{ 2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n-} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n-} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n+} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n+} \right. \right] \,,{\nonumber}
\end{align}
and
\begin{align}\label{eq:2B_proj_4}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB4}&= \frac{-g^2}{\omega_1 \omega_3} (T^a T^b)_{\alpha {\bar \beta}} \left( \frac{-i\sqrt{ 2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{n+} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{n+} \right.{\nonumber} \\
&+d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n-} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n-} \right. \right] \,. {\nonumber}
\end{align}
In this case both the $\mathcal{B}$ fields are again in the same helicity, but their correlation with the helicity of the quark current is opposite to that of \Eqs{eq:2B_proj_1}{eq:2B_proj_2}.
The operators $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB5}$ and $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB6}$ give a different helicity contribution, where the gluon fields have opposite helicity. We have
\begin{align}\label{eq:2B_proj_5}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B} \bar \cB5}&= \frac{g^2}{(\omega_2+\omega_3)(\omega_1+\omega_4)} (T^b T^a)_{\alpha {\bar \beta}}\left( \frac{-i\sqrt{ 2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n+} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^b_{n-} \mathcal{B}^a_{\bar n+} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n-} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^b_{n+} \mathcal{B}^a_{\bar n-} \right. \right] \,, {\nonumber}
\end{align}
and
\begin{align}\label{eq:2B_proj_6}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B} \bar \cB6}&= \frac{g^2 }{\omega_1 \omega_2} (T^a T^b)_{\alpha {\bar \beta}} \left( \frac{-i \sqrt{2\omega_1 \omega_2}}{Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{ n-} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n+} \mathcal{B}^a_{\bar n+} \mathcal{B}^b_{ n-} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n+} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \mathcal{B}^a_{\bar n-} \mathcal{B}^b_{n+} \right. \right] \,. {\nonumber}
\end{align}
Finally, the operator $\mathcal{O}^{(2)\nu}_{\mathcal{B}\bar \cB7}$ gives
\begin{align}\label{eq:2B_proj_7}
&\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{\mathcal{B} \bar \cB7}=-\frac{2g^2}{\omega_3 \omega_4} \left( \frac{-i\sqrt{ \omega_1 \omega_2}}{\sqrt{2}Q} \right) \\
& \left( \mathcal{B}^a_{\omega_3+}\mathcal{B}^b_{\omega_4+}(T^aT^b)_{\alpha {\bar \beta}}+ \mathcal{B}^a_{\omega_3-}\mathcal{B}^b_{\omega_4-} (T^aT^b)_{\alpha {\bar \beta}} - \mathcal{B}^b_{\omega_4+}\mathcal{B}^a_{\omega_3+}(T^bT^a)_{\alpha {\bar \beta}}- \mathcal{B}^b_{\omega_4-}\mathcal{B}^a_{\omega_3-} (T^bT^a)_{\alpha {\bar \beta}} \right) {\nonumber} \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n +} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n +} \right.+d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \alpha} \beta}_{n\bar n -} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \alpha} \beta}_{n\bar n -} \right. \right] \,. {\nonumber}
\end{align}
Therefore we see that all the different helicity combinations present in the basis of \Eq{eq:eeqqgg_basis2} are generated in the tree level matching.
\vspace{0.2cm} \noindent
\underline{
$\boldsymbol{(\bar q Q\bar Q)_{\bar n} ( q)_{n}:}$
}
The four quark operators generated in the matching were given in \Eq{eq:matched_4qn}. Projecting onto the basis of helicity operators, and using the color basis of \Eq{eq:qqqq_color}, we find
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{4\chi1a}&= \frac{g^2 \sqrt{ \omega_1 \omega_2 \omega_3 \omega_4} }{\omega_1 \omega_2 (\omega_2 + \omega_4)} \frac{1}{2}(\delta_{\gamma {\bar \beta}} \delta_{{\bar \delta} \alpha} -\frac{1}{N_c} \delta_{\gamma {\bar \delta}} \delta_{\alpha {\bar \beta}}) \left( \frac{-i}{\sqrt{2}Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \gamma} \delta}_{(q) n\bar n+} J^{{\bar \alpha} \beta}_{(Q)\bar n \bar 0} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle \ J_{\bar e e+} J^{{\bar \gamma} \delta}_{(q) n\bar n+} J^{(Q){\bar \alpha} \beta}_{\bar n \bar 0} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \gamma} \delta}_{(q) n\bar n-} J^{{\bar \alpha} \beta}_{(Q)\bar n 0} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \gamma} \delta}_{(q) n\bar n-} J^{{\bar \alpha} \beta}_{(Q)\bar n 0} \right. \right] \,,{\nonumber}
\end{align}
and
\begin{align}
\bar v_e \gamma^\mu u_e \left( \frac{-ig_{\mu \nu}}{Q^2} \right) \mathcal{O}^{(2)\nu}_{4\chi1b}&=\frac{g^2 \sqrt{ \omega_1 \omega_2 \omega_3 \omega_4} }{\omega_1 \omega_2 (\omega_2 + \omega_3) } \frac{1}{2}(\delta_{\gamma {\bar \beta}} \delta_{{\bar \delta} \alpha} -\frac{1}{N_c} \delta_{\gamma {\bar \delta}} \delta_{\alpha {\bar \beta}}) \left( \frac{-i}{\sqrt{2}Q} \right) \\
&\left[ d_{+-} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \gamma} \delta}_{(q) n\bar n+} J^{{\bar \alpha} \beta}_{(Q)\bar n 0} \right. \right.- d_{++} \left. \langle \bar n_e n_e\rangle \ J_{\bar e e+} J^{{\bar \gamma} \delta}_{(q) n\bar n+} J^{(Q){\bar \alpha} \beta}_{\bar n 0} \right.{\nonumber} \\
&+ d_{--} \left. [\bar n_e n_e] J_{\bar e e-} J^{{\bar \gamma} \delta}_{(q) n\bar n-} J^{{\bar \alpha} \beta}_{(Q)\bar n \bar 0} \right. \left. - d_{-+} \left. \langle \bar n_e n_e\rangle J_{\bar e e+} J^{{\bar \gamma} \delta}_{(q) n\bar n-} J^{{\bar \alpha} \beta}_{(Q)\bar n \bar 0} \right. \right] \,.{\nonumber}
\end{align}
We see that all the helicity combinations of \Eq{eq:Z3_basis_qQ} are generated in the tree level matching, as are both possible color structures. Furthermore, all the expected helicity selection rules are respected.
The projection of the operator \Eq{eq:matched_4qn} with an antiquark instead of a quark in the $n$ collinear sector proceeds identically, so we will not discuss it explicitly.
\subsection{Discussion}\label{sec:discuss}
Having performed the tree level matching both onto operators formed of standard Dirac and Lorentz structures, as well as projecting these onto operators in the helicity basis of \Sec{sec:eeJets}, here we briefly comment on the two approaches, and their advantages and disadvantages. The major advantage of the helicity operator approach is in enumerating a complete basis, and in making symmetry arguments about which operators will contribute to the factorization. This is greatly simplified by the use of helicity operators. Enumerating the basis is straightforward, and non-trivial relationships between different operators, for example using spin Fierz relations are absent. On the other hand, enumerating a complete and minimal basis of operators in the standard approach (which we did not do in this paper) is significantly more complicated. Furthermore, many of the symmetry arguments that were used to show that particular matrix elements of operators do not contribute to the cross section are obscured. Therefore for performing the formal factorization to all orders in perturbation theory, where a complete basis is essential, the helicity operator approach offers a clear advantage. This will become even more essential if one were to consider the case of three jets, for example. As noted in the case of dijets, a large number of the operators don't contribute. This can be seen easily in the helicity operator approach where many symmetries are manifest.
However, as has been seen in the matching, since each of the helicity building blocks is a scalar object, in the helicity approach all angular correlations are contained in the Wilson coefficients. This means that the Wilson coefficients are slightly more complicated objects. For example, in the case of the $e^+e^-\to$ dijets that we have considered here, the Wilson coefficients contain the inner products of polarization vectors, which carry all angular dependence. Furthermore, for relatively simple final states, such as those considered here, it is arguably more efficient to compute first a general current, and then project to helicity amplitudes. This may no longer be true if more jet directions are involved, since for higher multiplicity states, particularly involving gluons, it is well known that the calculation of individual helicity amplitudes can be significantly simpler.
Regardless of the exact techniques used to perform the matching, the helicity approach offers significant formal advantages for understanding the all orders structure of the operators, and for generating complete operator basis. The exact procedure which is then used to match to these operators can then be whatever is most convenient.
In a forthcoming paper we have performed a similar analysis, using the helicity operators introduced in this paper to study a complete subleading operator basis for $gg\to H$, as well as performing the tree level matching to those operators which contribute to the cross section at $\mathcal{O}(\lambda^2)$ \cite{usggh:2016}. In that case we also found the helicity operator approach to be convenient. In particular, the spin zero nature of the Higgs leads to even more stringent helicity selection rules than for the $\bar q \Gamma q$ current considered here. The structure of the Wilson coefficients in the case of $gg\to H$ is also simpler due to the fact that there are no angular correlations between the initial and final state, and therefore the Wilson coefficients for the helicity operators remain simple. We therefore believe that the use of helicity operators is a particularly powerful approach for simplifying subleading power operator bases in SCET, much more generally than the $\bar q \Gamma q$ current considered here.
\section{Conclusions}\label{sec:conclusions}
In this paper we have presented a basis of SCET helicity operator building blocks valid to all orders in the power expansion. This involved the use of helicity operator building blocks with multiple collinear fields in the same collinear sector, as well as ultrasoft gauge invariant helicity fields describing ultrasoft degrees of freedom, as summarized in \Tab{tab:helicityBB}. These operators allow for efficient organization of both helicity and color information. At subleading power interesting selection rules arise from the conservation of angular momentum \cite{Kolodrubetz:2016uim}, which constrain the allowed hard scattering operators in a basis. The use of helicity operators, color organization, and ultrasoft gauge invariant building blocks greatly simplifies the construction of subleading power operator bases in SCET.
To demonstrate the efficiency of the helicity-operator approach, we explicitly constructed a complete basis of hard scattering operators from a spin-1 current with two back-to-back collinear sectors up to $\mathcal{O}(\lambda^2)$. Due to the manifest crossing symmetry of our operator basis, this basis is applicable to studying power corrections for a number of phenomenologically relevant processes, including $e^+e^-\to$ dijets, $e^- p\to e^-$ jet, and constrained Drell-Yan. As an example, we discussed in some detail the structure of the factorization in SCET for $e^+e^-\to$ dijet event shapes at subleading power, and detailed the different sources of power corrections. Symmetry relations, which are manifest in the helicity operator basis, were used to show the vanishing of hard scattering contributions to the dijet cross section at $\mathcal{O}(\lambda)$. Using our basis of hard scattering operators we enumerated and studied the field content of the subleading jet and soft functions which arise from the subleading hard scattering operators at $\mathcal{O}(\lambda^2)$ in the expansion of the cross section. We then performed a tree level matching calculation, showing the operators which arise at tree level, both in a more standard notation in terms of Dirac and Lorentz structures, as well as projected into the helicity basis. We contrasted the different forms of the operators and their utility for different purposes. The explicit results for the matching of the subleading power operators will be useful for further studies of power corrections both in fixed order and resummed perturbation theory.
Since relatively little is known about the structure of factorization theorems at subleading power a number of directions exist for future study. It would be interesting to study in more detail RPI relations between operators for the subleading dijet operators to understand if relations beyond those given in this paper could be derived. The renormalization group evolution of the subleading power helicity operators in SCET is of considerable interest. The anomalous dimensions of leading power operators are well understood, and exhibit many universal features due to their connections with the soft limits of gauge theories, and it would interesting to determine to what extent such features persist to subleading power, and what new properties emerge. A study of the RG structure of $\mathcal{O}(\lambda)$ operators was considered in \Ref{Freedman:2014uta}. The RG evolution of higher twist operators has also been well studied \cite{Balitsky:1987bk,Ratcliffe:1985mp,Ji:1990br,Ali:1991em,Kodaira:1996md,Balitsky:1996uh,Mueller:1997yk,Belitsky:1997zw,Belitsky:1999ru,Belitsky:1999bf,Vogelsang:2009pj,Braun:2009mi}, and should exhibit similar structures. Finally, an understanding of the numerical impact of the subleading corrections for $e^+e^-\to $ dijets event shapes would be of considerable interest, for example, for improving extractions of $\alpha_s$ performed in \cite{Hoang:2015hka, Abbate:2010xh, Becher:2008cf, Chien:2010kc}.
Another potentially interesting application of subleading factorization theorems is to improving subtraction schemes for higher order perturbative calculations. Subtraction schemes based on factorization theorems include the recently introduced $N$-jettiness subtraction scheme \cite{Boughezal:2015aha,Boughezal:2015dva,Gaunt:2015pea}, based on the $N$-jettiness event shape \cite{Stewart:2010tn}, as well as the SCET based subtraction scheme for NNLO semileptonic top quark decays of \Ref{Gao:2012ja}. Subleading factorization theorems would allow for the subtraction of the next-to-singular terms, potentially improving the numerical accuracy and speed of the techniques. This was emphasized in \cite{Gaunt:2015pea}, and was first studied numerically in \cite{Boughezal:2016wmq}. The feasibility of extending these schemes to subleading power will rely on a convenient organization of the subleading factorization theorem, which should be aided by the simplicity of the helicity operator approach. A detailed study of power corrections for $N$-jettiness subtractions using the operators in this paper was presented in \cite{Moult:2016fqy}, where the terms of $\mathcal{O}(\alpha_s^2 \log^3 \tau)$ were computed explicitly. See also \cite{Boughezal:2016zws} for a calculation of the power corrections using alternative methods.
More broadly, we also envision that the helicity operator approach could be useful for constructing subleading operators in other processes, including $B$ physics and higher twist DIS where power corrections have been more thoroughly studied. Although power corrections have yet to begin to play an important phenomenological role in jet physics, we have demonstrated that a particular part of the factorization at subleading power, namely the construction of a basis of hard scattering operators, can be greatly simplified by the use of helicity operators, which we hope will prove useful in the future study of factorization theorems at subleading power.
\begin{acknowledgments}
We thank Matthew Schwartz, Simon Freedman, Raymond Goerke, Andrew Larkoski, Duff Neill, Gherardo Vita and Hua Xing Zhu for helpful discussions. We thank the Erwin Schroedinger Institute and the organizers of the ``Jets and Quantum Fields for LHC and Future Colliders'' and ``Challenges and Concepts for Field Theory and Applications in the Era of LHC Run-2'' workshops for hospitality and support while portions of this work were completed. This work was supported in part by the Office of Nuclear Physics of the U.S. Department of Energy under the Grant No. DE-SCD011090, by the Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231, NSERC of Canada, the Center for the Fundamental Laws of Nature at Harvard University, and the LDRD Program of LBNL. I.S. was also supported by the Simons Foundation through the Investigator grant 327942.
\end{acknowledgments}
|
\section{Introduction}
A new property of particle masses was recently shown when studying the mass variation versus the mass increase for adjacent meson and baryon masses of given families \cite{boris}.
The investigated function is:
\begin{equation}
m_{(n+1)} - m_{n} = f [(m_{(n+1)} + m_{n})/2]\\
\end{equation}
where $m_{(n+1)}$ corresponds to the (n+1) hadron mass value. The difference of two successive masses was plotted versus the mean value of the two nearby masses. Such studies were restricted to hadron families holding at least five masses. Regular oscillations were observed
giving rise to a new symmetry, the symmetry of oscillation. It was noticed in \cite{boris} that "the existence of composite hadrons, results from the addition of several forces, related to strong interaction, that combine in, at least, one attractive and one repulsive force. The equilibrium among these forces allows the hadron to exist, otherwise the composite mass will either disintegrate, or mix into a totally new object with loss of the individual components". As in classical physics, these opposite forces may generate oscillating behaviour.
The obtained data are fitted using a cosine function:
\begin{equation}
\Delta M = \alpha_{0} + \alpha_{1} cos( (M - M_{0}) / M_{1}) \\
\end{equation}
where M$_{0}$ /M$_{1}$ is defined within 2$\pi$. All coefficients, and masses used to draw the figures are in MeV units. The quantitative information is given in Tables I and II presented below. The oscillation periods are P = 2 $\pi M_{1}$. Both $\alpha_{0}$ and $\alpha_{1}$ are adjusted on the extreme values on all figures.
Whereas smaller periods than those given in the Tables may also reproduce the data, we show in the following figures, the largest possible values.
The discussion concerned the oscillatory periods, and not the oscillation amplitudes which need theoretical study outside the scope of previous and present papers.
For the same reason, the existence of substructures in hadrons, we expect to observe oscillations in nuclei made with nucleons. Such study will be considered after the hadronic masses.
\section{Application to hadronic masses}
The masses and widths are read from the Review of Particle Physics \cite{olive}, taking into account all the data reported, even if, in some cases, they are omitted from the summary table.
In our previous paper \cite{boris} data were shown for the following meson families with fixed quantum numbers: $f_0$ in fig. 1(a), $f_2$ in fig. 1(b),
charmonium (${\it c}{\bar c}$) $0^{-}(1^{--})$ in fig. 4(a) and bottomonium (${\it b}{\bar b})$
$0^{-}(1^{--})$ in fig. 4(b).
Several other data were also shown in the same paper without restriction to given quantum numbers for charmed in fig. 2(a), charmed strange in fig. 2(b), (${\it c}{\bar c}$) in fig. 3(a), (${\it b}{\bar b}$) in fig. 3(b) mesons. The corresponding figures displayed several data outside the fitting curves, mainly for (${\it c}{\bar c}$) but also for charmed strange mesons.
Resulting data were also shown for several baryon families without selection of given quantum numbers, contrary to the indication reported in column $J^{PC}$ of \cite{boris}. They are: $N^{*}$ in fig. 5(a), $\Delta$ in fig. 6(a), $\Lambda$ in fig. 6(b), $\Lambda_C$ in fig. 5(b), $\Xi$ in fig. 7(a) and $\Xi_C$ in fig. 7(b).
All previous data were fitted with oscillations with use of a few first masses. The fit gets often spoiled over a few MeV. The comparison between the selection of charmonium and bottomonium data with fixed quantum numbers, and without spin selection, suggests the relevance to restrict the study to particle families with given spins. Of course the necessity to have still at least five known masses remains. This condition will reduce the possibilities of application.
This study is done below, where the figures shown in \cite{boris} with such criteria of given spin, are repeated here in purpose of consistency.
The masses reported in \cite{olive} are used independently of the number of
attributed stars. When the name is different from the mass, the mass is used.
\begin{figure}[ht]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=19 136 537 546,clip,scale=0.8] {fh1.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for
$N^{*}$ (1/2$^{+}$),
$N^{*}$ (1/2$^{+}$ + 1/2$^{-}$), $N^{*}$ (3/2$^{+}$ + 3/2$^{-}$), and
$N^{*}$ (1/2$^{+}$ +1/2$^{-}$ + 3/2$^{+}$ + 3/2$^{-}$) baryons.}
\end{figure}
Fig. 1 shows in inserts (a), (b), (c), and (d) the data for $N^{*}$ (1/2$^{+}$),
$N^{*}$ (1/2$^{+}$ + 1/2$^{-}$), $N^{*}$ (3/2$^{+}$ + 3/2$^{-}$), and
$N^{*}$ (1/2$^{+}$ +1/2$^{-}$ + 3/2$^{+}$ + 3/2$^{-}$). Although two data in insert (b) lie outside the curve, both fits in inserts (a) and (b) are obtained with the same period P = 390~MeV. We will use that to add subsequently the masses having the same spin but different parities, and allow therefore to get more data with five, or more masses analysed simultaneously.
Fig. 1(c) shows nice fit for
$N^{*} (3/2^{+}$) + $N^{*} (3/2^{-})$. The period here is P =201~MeV. Fig. 1(d) shows the result for $N^{*}$ baryon masses, having both parities and both spins (1/2) and (3/2). These data are fitted with P = 201~MeV.
Therefore we will later on add the data having different parities and study them separately for different spins.
\begin{figure}[ht]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=21 331 520 546,clip,scale=0.8] {fh2.eps}}
\caption{Color on line. Inserts (a) and (b) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for
$N^{*}$ (5/2$^{+}$ + 5/2$^{-}$) baryons and $\eta (0^{-})$ mesons.}
\end{figure}
\begin{figure}[b]
\centering
\scalebox{0.58}[0.68]{
\includegraphics[bb=17 336 524 551,clip,scale=0.8] {fh3.eps}}
\caption{Color on line. Inserts (a) and (b) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for
$\Delta$ (1/2), $\Delta$ (3/2) baryons.}
\end{figure}
\begin{figure}[t]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=20 137 518 551,clip,scale=0.8] {fh4.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $\Lambda~(1/2^{+})$ and ($1/2^{-}$), $\Lambda~(3/2^{+})$ and ($3/2^{-})$, $\Sigma~(1/2^{+})$ and ($1/2^{-}$), $\Sigma~(3/2^{+})$ and $(3/2^{-})$ baryons.}
\end{figure}
\begin{figure}[ht]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=13 131 522 549,clip,scale=0.8] {fh5.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $f_0$ ($0^{++}$), $f_2$ ($2^{++}$), (${\it c}{\bar c}$) $0^{-}(1^{--})$, and (${\it b}{\bar b})$ $0^{-}(1^{--})$ mesons.}
\end{figure}
\begin{figure}[ht]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=24 136 525 552,clip,scale=0.8] {fh6.eps}}
\caption{Color on line. Inserts (a), (b), and (c), show successively the mass difference between successive masses, plotted versus both corresponding mean masses for K ($0^{+}$) and K ($0^{-}$), K ($1^{+}$) and K ($1^{-}$), and K ($2^{+}$) and K ($2^{-}$) mesons.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=18 231 518 547,clip,scale=0.8] {fh7.eps}}
\caption{Color on line. Variation of the oscillating periods of the hadronic mass families. Full red circles (blue squares) show the results for mesons (baryons).}
\end{figure}
Fig. 2 shows in insert (a) the mass difference between successive masses, plotted versus both corresponding mean masses for $N^{*}$ (5/2$^{+}$) + $N^{*}$ (5/2$^{-}$). The corresponding period
P~=~390~MeV is the same as the one obtained to describe the variation of the data for $N^{*}$~(J=1/2) fig.~1(b) and $N^{*}$~(J=3/2) fig.~1(c) baryons.
The data and fit for the $\eta$ meson are shown in fig.~2(b). The very precised masses are well fitted with the period P = 622~MeV. A mass at M = 762~MeV is introduced, following the recent suggestion \cite{BTETG}.
Fig. 3 shows the data for $\Delta$ baryons J = (1/2) in insert (a) and J = (3/2) in insert (b), without distinguishing the parities. Both data are fitted with the same period
P~=~201~MeV.
Fig. 4 shows the the data for strange baryons $\Lambda~(1/2^{+})$ and ($1/2^{-}$) in insert (a), $\Lambda~(3/2^{+})$ and ($3/2^{-}$) in insert (b),
$\Sigma (1/2^{+})$ and $(1/2^{-})$ in insert (c), $\Sigma (3/2^{+})$ and $(3/2^{-})$ in insert (d). The data in fig. 4(a) are fitted by an oscillating curve, although here more simple functions are possible.
Fig. 5 shows the results for $f_{0}$ ($0^{++}$) and $f_{2}$ ($2^{++}$) light unflavoured mesons in inserts (a) and (b).
The $\sigma$ or $f_{0}$(500) meson is broad and its mass (which unprecision is taken to $\Delta$M = 125~MeV), is badly determined. The reasonable fit allows to extrapolate the masses of the next $f_{0}$ not extracted experimentally up to now. They are M $\approx$ 2670 and 2760~ MeV. In the same way, the masses of the next $f_{2}$ mesons can be tentatively predicted to be: M $\approx$ 2380, 2450, and 2625 MeV.
Inserts (c) and (d) show the results for
(${\it c}{\bar c}$) $0^{-}(1^{--})$, and (${\it b}{\bar b})$ $0^{-}(1^{--})$ mesons.
The mass of the last quoted meson used in fig. 5(c), X(4660) $?^{?}(1^{--})$ fits perfectly in this distribution, and is therefore kept, assigning tentatively the same quantum numbers. The extrapolation allows to predict tentatively the next corresponding $\Psi$ masses: M $\approx$ 4805 and 5080~MeV. In the same way, the tentatively extrapolated $\Upsilon$ masses are: M $\approx$ 11330 and 11560~MeV.
Fig. 6 shows the data for strange kaons, K ($0^{+ -}$) in insert (a),
K ($1^{+ -}$) in insert (b), and K ($2^{+ -}$) in insert (c). Oscillatory shapes must be used for fits. The corresponding periods decrease regularly from P = 691~MeV, to P=408~MeV, and finally to P=201~MeV for increasing spins in inserts: (a), (b), and (c).
\begin{table}[t]
\caption{Quantitative information concerning the oscillation behavior of some mesons and baryons analysed previously. P is the period (in MeV).}
\label{Table I}
\vspace{5.mm}
\begin{tabular}{c c c c c c}
\hline\underline{}
name&q.c.&fig.&J&mass&P\\
\hline
$K_{0}$&$q{\bar s}$&6(a)&0&493.7&691\\
$\eta$&$q{\bar q}$&2(b)&0$^{-+}$&547.9&622\\
$f_{0}$&$q{\bar q}$&5(a)&$0^{++}$&475&647\\
$K_{1}$&$q{\bar s}$&6(b)&1&892&408\\
$f_{2}$&$q{\bar q}$&5(b)&$2^{++}$&1275&415\\
$K_{2}$&$q{\bar s}$&6(c)&2&1425&201\\
charm.&$c{\bar c}$&5(c)&$1^{--}$&2981.5&358\\
botto.&$b{\bar b}$&5(d)&$1^{--}$&9391&452\\
$N^{*}$&qqq&1(b)&1/2&939&390\\
$N^{*}$&qqq&1(c)&3/2&1520&201\\
$N^{*}$&qqq&2(a)&5/2&1675&390\\
$\Lambda$&qqs&4(a)&1/2&1115.7&220\\
$\Lambda$&qqs&4(b)&3/2&1519.5&396\\
$\Delta$&qqq&3(a)&1/2&1620&201\\
$\Delta$&qqq&3(b)&3/2&1232&201\\
$\Sigma$&qqs&4(c)&1/2&1189.4&396\\
$\Sigma$&qqs&4(d)&3/2&1385&176\\
\hline
\end{tabular}
\end{table}
All obtained periods fitting the previous data are reported in table I and fig.~7. Meson (baryon) periods are plotted with full red circles (full blue squares) versus the lower mass of each family. They all are located in three well defined ranges, the same for mesons and baryons. Three general properties are observed:
The periods of the three (meson) families with the lowest spin (J = 0):
$f_{0}$, $K^{0}$, and $\eta$ are located in the highest range close to P$\approx$655~MeV.
The periods corresponding to the other families distributed in the two other ranges,
favor the intermediate range for lower spin. This is true for periods of K ~(J=1) compared to K~(J=2), N~$(J=1/2)$ compared to N~$(J=3/2)$, $\Sigma$~$(J=1/2)$ compared to $\Sigma~$(J=3/2). However the opposite is observed for periods corresponding to
$\Lambda~(J=1/2)$ and $\Lambda~(J=3/2)$. And also the period of oscillation corresponding to the masses of N (J=5/2) is larger than that of N (J=3/2) and is equal to that of N (J=1/2), suggesting here again an oscillatory behaviour. Such behaviour is indeed observed in fig.~7 with P~=~357~MeV, better adjusted to meson than to baryon results. So the periods of $\Delta~(J=1/2)$ and $\Delta~(J=3/2)$) lie outside the distribution.
We notice that the distribution reported in fig.~7, fits the period correponding to (${\it c}{\bar c}$) $0^{-}(1^{--})$ mesons and also the period corresponding to (${\it b}{\bar b})$ $0^{-}(1^{--})$ mesons not plotted on fig.~7 since the very large gap between masses. The distribution reported in fig.~7 differs from that reported in \cite{Tatischeff:2016psc}. New data are analysed in the present paper, when several data, without spin selection, were used in \cite{Tatischeff:2016psc}.
The mean values of the three ranges shown in fig.~7, are pointed out by dashed lines (in green on line) at M(0) $\approx$ 655~MeV, M(1) $\approx$ 385~MeV, and M(2) $\approx$ 195~MeV. The meson oscillating periods, shown in red full circles, are gathered together following the relation: J - $\left |{S}\right |$ + 2*I which value equals 0, 1, and 2 for the three ranges from up to down. This relation does not apply for ${\it f_{2}}$ mesons.
\section{Application to nuclei masses}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=39 136 518 546,clip,scale=0.8] {fh8.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for respectively $^{7}$Li, $^{8}$Li, $^{9}$Be, and $^{10}$B nuclei.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.90]{
\includegraphics[bb=28 229 532 546,clip,scale=0.8] {fh9.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for respectively $^{17}$N, $^{17}$O, $^{17}$F, and $^{12}$B nuclei.}
\end{figure}
It is reasonable to expect oscillations in the nuclei mass levels for the same reason as before for hadrons. The nucleons in nuclei are bound by opposite forces. This property is studied below using data from \cite{Lederer} \cite{Ajzenberg-Selove:1979uqf,Ajzenberg-Selove:1968ztr,Ajzenberg-Selove:1976uid} when not specified. We start the analysis without spin selection, considering all level masses.
Fig.~8 shows the mass difference between successive masses plotted versus the corresponding massses for $^{7}$Li (red) P=2.45~MeV, $^{8}$Li (blue) P=3.02~MeV, $^{9}$Be (green) P=2.76~MeV, and $^{10}$B (purple) P=1.88~MeV respectively in inserts (a), (b), (c), and (d). We observe an increase of the level number with increasing mass. We observe also that the fit between data and calculated curves spoils after the five-six first MeV in the case of $^{10}$B nucleus.
For heavier nuclei, these properties are amplified as seen in fig.~9.
Fig.~9 shows the mass difference between successive masses plotted versus the corresponding masses for $^{17}$N P=1.70~MeV, $^{17}$O P=1.88~MeV, $^{17}$F
P=2.39~MeV, and $^{12}$B P=1.76~MeV nuclei. In spite of the large mass differences for all data, emphasized by the log scale, the first data are rather well fitted, then followed by a large number of spread data.
So the situation is comparable to the one observed for hadrons, and therefore brings us to separate the nuclei level masses by their spins.
The next figures will study the oscillation properties of nuclei level masses having the same spin.
Although a large number of level masses are known for the majority of nuclei, rather few have a number of known quantum numbers allowing the same studies as previously done (five or more level masses with the same spin).
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=20 136 515 554,clip,scale=0.8] {fh10.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{10}$B nucleus: all spins, J=1, J=2, and J=3 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=20 136 515 554,clip,scale=0.8] {fh11.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{14}$N nucleus: all spins, J=1, J=2, and J=3 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=40 138 520 547,clip,scale=0.8] {fh12.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{15}$N nucleus: all spins, J=1/2, J=3/2, and J=5/2 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=22 238 517 546,clip,scale=0.8] {fh13.eps}}
\caption{Color on line. Inserts (a), (b), and (c) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{16}$N nucleus: J=1, J=2, and J=3 respectively. Insert (d) shows the result for $^{4}$He J=2 levels, P=5.03~MeV.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=21 83 518 546,clip,scale=0.8] {fh14.eps}}
\caption{Color on line. Inserts (a), (b), (c), (d), and (e) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{16}$O nucleus: J=0, J=1, J=2, J=3, and J=4 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.78]{
\includegraphics[bb=24 130 516 545,clip,scale=0.8] {fh15.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{20}$Ne nucleus: all spins, J=0, J=2, and J=3 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.68]{
\includegraphics[bb=29 229 520 543,clip,scale=0.8] {fh16.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{56}$Fe nucleus: J=0, J=2, J=3, and J=4 respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.74]{
\includegraphics[bb=31 128 518 544,clip,scale=0.8] {fh17.eps}}
\caption{Color on line. Inserts (a), (b), and (c) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for J=5/2 $^{25}$Al and $^{27}$Al, $^{155}$Tb and $^{159}$Tb, and $^{165}$Dy and $^{165}$Er nuclei respectively.(See text).}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.74]{
\includegraphics[bb=28 84 516 552,clip,scale=0.8] {fh18.eps}}
\caption{Color on line. Inserts (a), (b), (c), (d), and (e) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for J=2 $^{26}$Mg, $^{194}$Pt, $^{214}$Po, $^{154}$Gd, and $^{230}$Th nuclei respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.74]{
\includegraphics[bb=20 82 516 546,clip,scale=0.8] {fh19.eps}}
\caption{Color on line. Inserts (a), (b), (c), (d), (e), and (f) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for J=2 $^{62}$Zn, $^{80}$Se, $^{100}$Ru, $^{92}$Nb, $^{146}$La, and $^{132}$Ce nuclei respectively.}
\end{figure}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{0.58}[0.74]{
\includegraphics[bb=17 239 520 545,clip,scale=0.8] {fh20.eps}}
\caption{Color on line. Inserts (a), (b), (c), and (d) show successively the mass difference between successive masses, plotted versus both corresponding mean masses for $^{208}$Pb J=3, J=4, J=5, and J=6 respectively.}
\end{figure}
Fig.10 shows the mass difference between successive masses plotted versus the corresponding masses for $^{10}$B nucleus. The four inserts (a), (b), (c), and (d) show data with: all spins P = 1.885~MeV, J=1 P=2.2~MeV, J=2 P=2.2~MeV, and J=3 P=2.2~MeV. The data for separated spins (inserts (b), (c), and (d)) are well fitted with the same period of oscillation. However the small number of levels with spin J=3 (insert (d)) involves an arbitrary fit.
Fig.~11 shows the mass difference between successive masses plotted versus the corresponding masses for $^{14}$N nucleus. The four inserts (a), (b), (c), and (d) show data with: all spins P = 1.82~MeV, J=1 P=2.01~MeV, J=2 P=2.01~MeV, and J=3 P=2.04~MeV. The data for separated spins (inserts (b), (c), and (d)) are well fitted with almost the same period of oscillation.
Fig.~12 shows the mass difference between successive masses plotted versus the corresponding masses for $^{15}$N nucleus. The four inserts (a), (b), (c), and (d) show data with: all spins P = 1.885~MeV, J=1/2 P=1.885~MeV, J=3/2 P=1.885~MeV, and J=5/2 P=1.63~MeV. The data for separated spins are better fitted in inserts (b) and (d) than in (c).
Fig.~13 shows the mass difference between successive masses plotted versus the corresponding masses for $^{16}$N nucleus \cite{Tilley:1993zz}. The three inserts (a), (b), and (c) show data with: J=1 P = 1.82~MeV, J=2 P=2.07~MeV, and J=3 P=2.39~MeV. Insert (d) shows the result for the levels J=2 P=5.03~MeV of the $^{4}$He nucleus \cite{Tilley:1992zz}. The data are well fitted.
Fig.~14 shows the mass difference between successive masses plotted versus the corresponding masses for $^{16}$O nucleus \cite{Tilley:1993zz}. The five inserts (a), (b), (c), (d), and (e) show data with: J=0 P = 5.03~MeV, J=1 P=3.33~MeV, J=2 P=3.08~MeV, J=3 P=2.51~MeV, and J=4 P=3.96~MeV. One data with J=2, close to 16~MeV, is outside the fit. For all nuclei, at large excitation energy, the spins of some levels are unknown, therefore these levels are ignored.
Fig.~15 shows the mass difference between successive masses plotted in log scale versus the corresponding masses for $^{20}$Ne nucleus. The four inserts (a), (b), (c), and (d) show data with: all spins P = 1.885~MeV, J=0 P=1.885~MeV, J=2 P=1.885~MeV, and J=3 P=1.885~MeV. The data for separated spins (inserts (b), (c), and (d)) are rather well fitted with the same period of oscillation, the same as obtained for $^{15}$N (J=3/2).
Fig.~16 shows the mass difference between successive masses plotted versus the corresponding masses for $^{56}$Fe nucleus \cite{56Fe}. The four inserts (a), (b), (c), and (d) show data with: J=0 P =2.07~MeV, J=2 P=1.76~MeV, J=3 P=1.63~MeV, and J=4 P=1.00~MeV. The data are well fitted; here the oscillatory periods decrease with increasing spins. The fit in insert (a) is undetermined due to small number of data. There is two data outside the fit in insert (b) and one in insert (c).
Fig.17 shows the mass difference between successive masses plotted versus the corresponding masses for several nuclei with spin J=5/2. The data in insert (a) correspond to $^{25}$Al (full blue squares) and $^{27}$Al (full red circles) fitted with the period P=2.45~MeV.
The data in insert (b) correspond to $^{155}$Tb (full blue squares) and $^{159}$Tb (full red circles) fitted with the period P=0.575~MeV. The fit is obtained using $^{155}$Tb data. Three red data over four, corresponding to $^{159}$Tb, lie close to the same curve. The data in insert (c) correspond to $^{165}$Er (full blue squares) fit with the period P=0.314~MeV, and $^{165}$Dy (full red circles) \cite{165Dy} fit with the period P=0.547~MeV. Both nuclei $^{165}$Er and $^{165}$Dy differ by only by one proton (and one neutron), therefore the large difference between their oscillating periods is unclear.
Fig.~18 shows the mass difference between successive masses plotted versus the corresponding masses for several nuclei with spin J=2. The data in insert (a), (b), (c), (d), and (e) correspond to $^{26}$Mg P=1.885~MeV, $^{194}$Pt P=0.49~MeV, $^{214}$Po P=0.446~MeV, $^{154}$Gd P=0.49~MeV, and $^{230}$Th P=0.427~MeV nuclei respectively. Here again the periods decrease with increasing nuclei masses.
Fig.~19 shows the mass difference between successive masses plotted versus the corresponding masses for several other nuclei with spin J=2. The data in insert (a), (b), (c), (d), (e), and (f) correspond to $^{62}$Zn \cite{62Zn} P=0.817, $^{80}$Se \cite{80Se} P=0.452~MeV, $^{100}$Ru \cite{100Ru} P=0.352~MeV, $^{92}$Nb P=0.377~MeV, $^{146}$La P=0.817~MeV, and $^{132}$Ce P=2.14~MeV nuclei respectively.
Fig.~20 shows the mass difference between successive masses plotted versus the corresponding masses for $^{208}$Pb \cite{208Pb} in log scale. Inserts (a), (b), (c), and (d) correspond respectively to data having the following spins: J=3 P=0.942~MeV, J=4 P=0.754~MeV, J=5 P=0.88~MeV, and J=6 P=0.503~MeV.
The extracted periods do not fullfil the trend observed previously, namely to decrease with increasing spins. However only J=4 or preferably J=5 data are concerned with this comment. We observe that both corresponding inserts (b) and (c) exhibit one data outside the fit which remains eventually doubtful, asking eventually for more data.
\begin{table}[t]
\caption{Quantitative information concerning the oscillation behavior of some J=2 nuclei levels analysed previously. P is the period (in MeV).}
\label{Table 2}
\vspace{5.mm}
\begin{tabular}{c c c}
\hline\underline{}
nucleus&fig.&P(MeV)\\
\hline
$^{4}$He&13(d)&5.03\\
$^{10}$B&10(c)&2.2\\
$^{14}$N&11(c)&2.0\\
$^{16}$N&13(b)&2.07\\
$^{16}$O&14(c)&3.08\\
$^{20}$Ne&15(c)&1.885\\
$^{26}$Mg&18(a)&1.885\\
$^{56}$Fe&16(b)&1.76\\
$^{62}$Zn&19(a)&0.817\\
$^{80}$Se&19(b)&0.452\\
$^{92}$Nb&19(d)&0.377\\
$^{100}$Ru&19(c)&0.352\\
$^{132}$Ce&19(f)&2.14\\
$^{146}$La&19(e)&0.817\\
$^{154}$Gd&18(d)&0.490\\
$^{194}$Pt&18(b)&0.490\\
$^{214}$Po&18(c)&0.446\\
$^{230}$Th&18(e)&0.427\\
\hline
\end{tabular}
\end{table}
\begin{figure}[h]
\centering
\hspace*{-0.2cm}
\scalebox{1.2}[1]{
\includegraphics[bb=36 342 295 543,clip,scale=0.8] {fh21.eps}}
\caption{Color on line. Period variation versus the mass number A for J=2 levels.}
\end{figure}
\begin{table}[t]
\caption{Quantitative information concerning the oscillation behavior of some nuclei levels with spin different from 2, analysed previously. P(m) is the period (in MeV).}
\label{Table 2}
\vspace{5.mm}
\begin{tabular}{c c c c}
\hline\underline{}
Spin&nucleus&fig.&P(MeV)\\
\hline
0&$^{16}$O&14(a)&5.03\\
&$^{20}$Ne&15(b)&1.885\\
&$^{56}$Fe&16(a)&2.07\\
\hline
1/2&$^{15}$N&12(b)&1.885\\
\hline
1&$^{10}$B&10(b)&2.2\\
&$^{14}$N&11(b)&2.01\\
& $^{16}$N&13(a)&1.82\\
& $^{16}$O&14(b)&3.33\\
\hline
3/2&$^{15}$N&12(c)&1.885\\
\hline
5/2&$^{15}$N&12(d)&1.633\\
&$^{25}$Al&17(a)&2.45\\
&$^{27}$Al&17(a)&2.45\\
&$^{155}$Tb&17(b)&0.575\\
&$^{159}$Tb&17(b)&0.575\\
&$^{165}$Dy&17(c)&0.547\\
&$^{165}$Er&17(c)&0.314\\
\hline
3&$^{10}$B&10(d)&2.2\\
&$^{14}$N&11(d)&2.04\\
&$^{16}$N&13(c)&2.39\\
&$^{16}$O&14(d)&2.51\\
&$^{20}$Ne&15(d)&1.885\\
&$^{56}$Fe&16(c)&1.63\\
&$^{208}$Pb&20(a)&0.94\\
\hline
4&$^{16}$O&14(e)&3.96\\
&$^{56}$Fe&16(d)&1.005\\
&$^{208}$Pb&20(b)&0.754\\
\hline
5&$^{208}$Pb&20(c)&0.88\\
\hline
6&$^{208}$Pb&20(d)&0.50\\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\scalebox{1.2}[1]{
\includegraphics[bb=41 331 282 546,clip,scale=0.8] {fh22.eps}}
\caption{Color on line. Period variation versus the mass number A for J $\not=$ 2. J=1 data are shown with red full circles, J=3 with blue full squares, and J=4 with green full stars.}
\end{figure}
\section{Discussion and Conclusion}
Table II shows the periods of oscillation of the nuclei levels J=2 studied previously. Their variation versus the mass number A is displayed in fig.~21. Fast increases are observed for $^{4}$He, $^{16}$O, and $^{132}$Ce, with an abrupt fall between $^{56}$Fe and $^{62}$Zn. The two first high data are related to doubly magic number nuclei. We observe larger periods for closed shells or subshells, followed by smaller periods. The fall between $^{56}$Fe and $^{62}$Zn
should then be attributed to the passing through the magic number Z=28. The $^{132}$Ce has 58 protons which close the ${\it 1f_{7/2}}$ subshell.
The missing of enough known spins for the nuclei levels with neutron (or proton) numbers close to other magic numbers, prevents to study these mass regions.
Table III and fig.~22 show the periods of oscillation of the other nuclei level periods studied previously. J=0 data are shown with black empty squares, J=1 with red full circles,
J=3 with blue full squares, J=4 with green full stars, and J=5/2 with purple full triangles. The periods concentrate between both dashed lines and decrease with increasing masses, with a fast jump for all $^{16}$O periods J=0, 1, 3, and 4. Table III shows that the period of variation of different spin levels remain constant for light nuclei like $^{10}$B, $^{14}$N, and $^{20}$Ne. This is also the case for $^{16}$O except for the J = 4 levels. For heavier nuclei, the periods decrease for increasing spins.
For $^{56}$Fe nucleus for increasing spins J=0, 2, 3, and 4, the periods are respectively: P=2.07, 1.8, 1.63, and 1.005~MeV. For $^{208}$Pb, the period for J=5 is somewhat larger than expected for a regular decrease.
When these studies considered all spins \cite{Tatischeff:2016psc}, the figures of period variation versus the masses exhibited nice shapes in agreement with a clear oscillation for the first several masses only. In the present study done for given spins, the agreement of data versus calculations is good in all ranges where almost all levels have a known spin.
This study should be extended for higher mass hadrons, not known presently. It was already mentioned that a minimum of five masses of given quantum numbers must exist.
The same remark holds for nuclear levels. Whereas a lot of nuclear levels is known, the spin of many of them is ignored. Moreover, there are few levels with unknown spin, which masses are located between the masses of known spin levels. This may then alter the data of higher mass levels.
In conclusion the paper shows that the oscillating periods of mesons and baryons follow the same variation. This symmetry of oscillation is observed for the masses of hadrons and masses of nuclei levels which display oscillatory behaviours well observed using the relation (1) and well fitted with the cosine function (2). Such behaviour requires the need for a theoretical study to describe the oscillating distributions and particularly the oscillation amplitudes.
|
\section{Introduction}
\IEEEPARstart{A}{} traditional approach for transmitting digital
information over a communication link with high spectral efficiency
($\eta\geq2$ bit/s/Hz) is to combine good error-correcting code with
high-order modulation. Over last two decades, several capacity-approaching
coded modulation techniques have been proposed, including low-density
parity check (LDPC) coded modulation {[}1, 2{]}, turbo-coded modulation
and turbo trellis-coded modulation {[}3{]}. In this paper, we propose
an alternative bandwidth-efficient modulation technique, which does
not rely on any error correction coding in the traditional sense,
while still achieving error rates close to the best known coded modulation
schemes. Our method is based on serial concatenation of an orthogonal
linear transform, such as DFT or WHT, with memoryless nonlinearity.
We demonstrate that such a simple signal construction may exhibit
properties of random code ensemble, as a result approaching channel
capacity if decoded using a maximum-likelihood (ML) decoder.
When the proposed modulation scheme relies on DFT transform, it can
be viewed as an OFDM signal distorted by very strong memoryless nonlinearity.
In the context of OFDM, any nonlinearity is usually considered a highly
undesirable phenomenon. In fact, the sensitivity of OFDM technology
to nonlinear distortions is usually cited as a major drawback of OFDM.
Nonetheless, in {[}4{]}, the authors theoretically proved that the
bit error rate (BER) performance of OFDM transmission subjected to
strong memoryless nonlinearity can outperform linear transmission,
provided that an optimal ML-receiver is used. Moreover, in {[}5{]},
we proposed a practical message-passing receiver for nonlinearly distorted
OFDM signals, showing that the BER performance of hard-clipped OFDM
signals can be up to 2 dB better than linear OFDM transmission. In
this paper, we go one step further to demonstrate that if the uncoded
OFDM signal is distorted by memoryless nonlinearity with certain properties,
the resultant signal waveform may achieve BER performance close to
the best known capacity\textendash approaching coded modulation techniques.
The same performance can also be achieved by replacing DFT in the
transmitter with WHT, resulting in simpler hardware/software implementation
of the encoder and the decoder.
\section{Proposed modulation technique}
\subsection{Achieving capacity by serial concatenation of orthogonal transform
and memoryless nonlinearity}
Consider the modulation scheme illustrated in Figure \ref{fig:direct_modulator}.
In the transmitter, $N$ uncoded real or complex baseband modulation
symbols $\ensuremath{\boldsymbol{\mathbf{x}}=\left\{ x_{k}\right\} }$
(e.g. $M$-QAM, or $M$-PAM) are first transformed by means of orthogonal
block transform and then passed through memoryless nonlinearity block.
The signal at the modulator output can be expressed as:
\begin{equation}
\ensuremath{s_{n}=f\left(z_{n}\right),\quad n=0,1,...,N-1}\label{eq: main_modulator}
\end{equation}
where $f(z)$ is a memoryless nonlinear function, and
\begin{equation}
\ensuremath{{\bf z}={\bf Fx},}
\end{equation}
where $\mathbf{z}=\left\{ z_{n}\right\} $, and $\mathbf{F}$ is a
$N\times N$ (unitary) orthogonal transform matrix. In this paper,
we consider two types of transforms: Walsh-Hadamard transform, and
discrete Fourier transform (real-valued or complex).
Let us now assume that the memoryless nonlinear function $f(z)$ can
be represented as $\ensuremath{f\left(z\right)=\underbrace{g\left(...g\left(g\left(z\right)\right)\right)}_{l-times}}$,
where $g(z)$ is a deterministic chaotic map. Thus, $f(z)$ has certain
properties of chaotic iterated functions, namely, \emph{sensitive
dependence on initial conditions} {[}6{]}. Under such assumptions,
and if $N\rightarrow\infty$ and $l\rightarrow\infty$, the ensemble
of waveforms generated by (\ref{eq: main_modulator}) possesses all
major properties of a random code ensemble, because even small modifications
in the message sequence $\ensuremath{\left\{ x_{k}\right\} }$ (e.g.,
in a single bit) lead to:
\begin{description}
\item [{a)}] small modifications, at least, for all samples of the intermediate
signal $z_{n}$ due to the spreading properties of the orthogonal
transform operation, and
\item [{b)}] large (and pseudo-random) modifications for all samples of
waveform $s_{n}$ due to the properties of the nonlinear function
$f(z)$.
\end{description}
Therefore, we conjecture that where $f(z)$ has the above-mentioned
properties if the waveforms generated by (\ref{eq: main_modulator}) are
demodulated using a ML decoder, system performance may approach channel
capacity. Unfortunately, brute-force, maximum-likelihood decoding
of (\ref{eq: main_modulator}) is prohibitively complex, generally
of the order $O(M^{N})$, where $M$ is the modulation order. However,
as in many practical applications, we can rely on belief-propagation
techniques to approximate ML-decoding. A perfect candidate for decoding
the signals generated by (\ref{eq: main_modulator}) is the generalized
approximate message passing (GAMP) algorithm {[}7{]}. Having already
been applied to the decoding of clipped OFDM signals in {[}5{]}, it
has demonstrated good performance and reasonably good convergence
behavior.
\begin{figure}[tbh]
\includegraphics[scale=1.1]{modulator}\caption{\label{fig:direct_modulator}Proposed modulation scheme}
\end{figure}
\subsection{GAMP algorithm and its modifications}
In the additive white Gaussian noise (AWGN) channel, the model of
the received signal can be expressed as
\begin{equation}
\ensuremath{y_{n}=f\left(z_{n}\right)+w_{n},}\quad n=0,1,...,N-1\label{eq:gamp_model}
\end{equation}
where $w_{n}$ is the AWGN term with zero mean and variance $\mu^{w}$.
Model (\ref{eq:gamp_model}) is equivalent to a general problem formulation
for the GAMP algorithm {[}7{]}, which belongs to a class of Gaussian
approximations of loopy belief propagation for dense graphs. In our
decoder implementation, we use the sum-product variant of the GAMP
algorithm, which approximates minimum mean-squared error estimates
of $\mathbf{x}$ and $\mathbf{z}$.
During simulation study, we discovered that the convergence behavior
and bit and frame error rate (FER) performance of the conventional
GAMP algorithm applied to the decoding of (\ref{eq:gamp_model}) may
be improved by several simple algorithm modifications. In particular,
we used damped version of the GAMP algorithm, which is known to improve
convergence for certain mixing matrices {[}8{]}, and introduced scaling
factor for AWGN term variance $\mu^{w}\rightarrow\alpha\mu^{w}$.
In addition, we implemented the final decision selection and post-processing
scheme as will be explained in the next section that helped us overcome
certain deficiencies of a conventional GAMP algorithm. To reduce computational
complexity we opted to use a version of GAMP algorithm with scalar
step-sizes {[}7{]}. Note that the scalar step-size version of the
GAMP algorithm is optimal for WHT and complex DFT since in both cases
$\left|F_{n,k}\right|^{2}=1$. Moreover, our experimental study revealed
that the performance loss for real DFT case ($\left|F_{n,k}\right|^{2}\neq1$)
was also marginal.
The GAMP algorithm adapted to our problem is summarized below. The
algorithm generates a sequence of estimates ${\bf \hat{x}}\left(t\right)$,
${\bf \hat{z}}\left(t\right)$, for $t=1,2,...$ through the following
recursions:
\emph{Parameters:}
$t_{max}$ the maximum number of iterations,
$\alpha$ the noise scaling factor,
$\beta$ the damping factor.
\emph{Step 1) Initialization:}
$t=1$, ${\bf \hat{x}}\left(1\right)=\boldsymbol{0}$, $\tilde{\mathbf{x}}\left(1\right)=\boldsymbol{0}$,
${\bf \boldsymbol{\mu}}^{x}\left(1\right)=\boldsymbol{1}$, ${\bf \hat{s}}\left(0\right)=\boldsymbol{0}$
\emph{Step 2) Estimation of output nodes:}
\begin{equation}
\mu_{n}^{p}\left(t\right)=\frac{1}{N}\sum\limits _{k=0}^{N-1}\mu_{k}^{x}\left(t\right),\forall n\label{eq:gamp_algo_start}
\end{equation}
\begin{equation}
\hat{p}_{n}\left(t\right)=\sum\limits _{k=0}^{N-1}F_{n,k}\hat{x}_{k}\left(t\right)-\mu_{n}^{p}\left(t\right)\hat{s}_{n}\left(t-1\right),\forall n
\end{equation}
\begin{equation}
\hat{z}_{n}\left(t\right)=\frac{1}{C}\int\limits _{-\infty}^{\infty}ze^{-\frac{\left(y_{n}-f\left(z\right)\right)^{2}}{2\alpha\mu^{w}}-\frac{\left(\hat{p}_{n}\left(t\right)-z\right)^{2}}{2\mu_{n}^{p}\left(t\right)}}dz,\forall n\label{eq:Integral_start}
\end{equation}
\begin{equation}
\mu_{n}^{z}\left(t\right)=\frac{1}{C}\int\limits _{-\infty}^{\infty}z^{2}e^{-\frac{\left(y_{n}-f\left(z\right)\right)^{2}}{2\alpha\mu^{w}}-\frac{\left(\hat{p}_{n}\left(t\right)-z\right)^{2}}{2\mu_{n}^{p}\left(t\right)}}dz-\left(\hat{z}_{n}\left(t\right)\right)^{2},\forall n
\end{equation}
where
\begin{equation}
C=\int\limits _{-\infty}^{\infty}e^{-\frac{\left(y_{n}-f\left(z\right)\right)^{2}}{2\alpha\mu^{w}}-\frac{\left(\hat{p}_{n}\left(t\right)-z\right)^{2}}{2\mu_{n}^{p}\left(t\right)}}dz,\forall n\label{eq:Integral_end}
\end{equation}
\begin{equation}
\hat{s}_{n}\left(t\right)=\left(1-\beta\right)\hat{s}_{n}\left(t-1\right)+\beta\dfrac{\hat{z}_{n}\left(t\right)-\hat{p}_{n}\left(t\right)}{\mu_{n}^{p}\left(t\right)},\forall n
\end{equation}
\begin{equation}
\mu_{n}^{s}\left(t\right)=\left(1-\beta\right)\mu_{n}^{s}\left(t-1\right)+\beta\frac{1-\frac{\mu_{n}^{z}\left(t\right)}{\mu_{n}^{p}\left(t\right)}}{\mu_{n}^{p}\left(t\right)},\forall n
\end{equation}
\emph{Step 3) Estimation of input nodes}:
\begin{equation}
\tilde{x}_{k}\left(t\right)=\left(1-\beta\right)\tilde{x}_{k}\left(t-1\right)-\beta\hat{x}_{k}\left(t\right)
\end{equation}
\begin{equation}
\mu_{k}^{r}\left(t\right)=\left(\frac{1}{N}\sum\limits _{n=0}^{N-1}\mu_{n}^{s}\left(t\right)\right)^{-1},\forall k
\end{equation}
\begin{equation}
\hat{r}_{k}\left(t\right)=\tilde{x}_{k}\left(t\right)+\mu_{k}^{r}\left(t\right)\sum\limits _{n=0}^{N-1}F_{n,k}^{*}\hat{s}_{k}\left(t\right),\forall k
\end{equation}
\begin{equation}
\hat{x}_{k}\left(t+1\right)=\sum\limits _{m=1}^{M}d_{m}P_{m,k},\forall k
\end{equation}
\begin{equation}
\mu_{k}^{x}\left(t+1\right)=\sum\limits _{m=1}^{M}\left(d_{m}-\hat{x}_{k}\left(t+1\right)\right)^{2}P_{m,k},\forall k\label{eq:gamp_algo_end}
\end{equation}
where
\begin{equation}
P_{m,k}=\frac{e^{-\frac{\left(d_{m}-\hat{r}_{k}\left(t\right)\right)^{2}}{2\mu_{k}^{r}\left(t\right)}}}{\sum\limits _{l=1}^{M}e^{-\frac{\left(d_{l}-\hat{r}_{k}\left(t\right)\right)^{2}}{2\mu_{k}^{r}\left(t\right)}}},
\end{equation}
$M$ is the number of points in the signal constellation, and $\left\{ d_{m}\right\} $
is the vector of constellation points. For example, for 2-PAM modulation,
$\left\{ d_{m}\right\} =\left[-1{\rm \quad+}1\right]$, and for 4-PAM
modulation $\left\{ d_{m}\right\} =\left[-3\mathord{\left/\vphantom{-3{\sqrt{5}}}\right.\kern -\nulldelimiterspace}\sqrt{5}\quad-1\mathord{\left/\vphantom{-1{\sqrt{5}}}\right.\kern -\nulldelimiterspace}\sqrt{5}{\rm \quad+}1\mathord{\left/\vphantom{1{\sqrt{5}}}\right.\kern -\nulldelimiterspace}\sqrt{5}{\rm \quad+}{\rm 3}\mathord{\left/\vphantom{{\rm 3}{\sqrt{5}}}\right.\kern -\nulldelimiterspace}\sqrt{5}\right]$.
Steps (\ref{eq:gamp_algo_start})\textendash (\ref{eq:gamp_algo_end})
are repeated with $t\rightarrow t+1$ until $t_{max}$ iterations
have been performed.
We presented here the GAMP algorithm version for the real-valued modulation
and real orthogonal transform (e.g. $M$-PAM and WHT or real-DFT).
Nonetheless, the extension to the complex case is straightforward.
Moreover, in case of Cartesian-type nonlinearity $f(z)$ the complexity
of the decoding algorithm with $N$ complex input symbols is essentially
the same as the complexity of the real-valued algorithm with $2N$
input symbols.
More details on the GAMP algorithm and its thorough analysis can be
found in {[}7{]}.
\subsection{Choosing optimal nonlinearity $f(z)$}
Choosing the optimal shape of nonlinearity $f(z)$ is not easy, requiring
a balance between two conflicting requirements. Firstly, the nonlinear
function should be reasonably \textquotedbl{}chaotic\textquotedbl{}
to guarantee sensitivity to initial conditions and, therefore, random-like
properties of the coded waveforms. On the other hand, our experimental
study implies that a \textquotedbl{}truly chaotic\textquotedbl{} shape
of nonlinearity $f(z)$ precludes GAMP algorithm from converging to
the ML-solution. Therefore, we adopted an \emph{ad hoc} procedure
to select and optimize the shape of the memoryless nonlinearity $f(z)$.
Our choice of nonlinearity $f(z)$ relies on two key ideas:
\begin{itemize}
\item $f(z)$ should contain a linear or almost linear region around $f(0)$
to allow the message passing decoder to converge.
\item $f(z)$ should \emph{resemble} a chaotic iterated function in other
regions to guarantee sensitivity to initial conditions.
\end{itemize}
For a general representation of $f(z)$, we suggest using the flexible,
piece-wise linear model:
\begin{equation}
\ensuremath{f\left(z\right)=\left\{ \begin{array}{l}
{\mathop{\rm sgn}}\left(z\right)\left(a_{0}\left|z\right|+b_{0}\right),{\rm \quad if}\;0\le\left|z\right|<T_{1}\\
{\mathop{\rm sgn}}\left(z\right)\left(a_{1}\left|z\right|+b_{1}\right),{\rm \quad if\;}T_{1}\le\left|z\right|<T_{2}\\
...\\
{\mathop{\rm sgn}}\left(z\right)\left(a_{i-1}\left|z\right|+b_{i-1}\right),{\rm \quad if\;}T_{i-1}\le\left|z\right|<T_{i}\\
{\mathop{\rm sgn}}\left(z\right)\left(a_{i}\left|z\right|+b_{i}\right),{\rm \quad if\;}\left|z\right|\ge T_{i}
\end{array}\right.}\label{eq:Nonlinearity equation}
\end{equation}
with some parameters $\boldsymbol{a}=G_{0}\left\{ a_{0},a_{1},...,a_{i}\right\} $,
$\boldsymbol{b}=\left\{ b_{0},b_{1},...,b_{i}\right\} $, and $\boldsymbol{T}=G_{0}^{-1}\left\{ 0,T_{1},T_{2},...,T_{i}\right\} $,
where $G_{0}$ is a scaling factor that does not affect the ``shape''
of nonlinearity $f(z)$. For complex modulation (i.e. complex DFT),
we apply function (\ref{eq:Nonlinearity equation}) separately to
the real and imaginary components of $z$ (Cartesian-type nonlinearity).
To select the parameters of (\ref{eq:Nonlinearity equation}), we
devised dozens of different functions $f(z)$ that satisfy the above
mentioned heuristic requirements, and then optimized parameter $G_{0}$
at our target BER and FER (namely, BER$\leq10^{-5}$, FER$\leq$1\%).
This non-exhaustive search procedure yielded several ``good'' sets
of parameters $\boldsymbol{a}$, $\boldsymbol{b}$, and $\boldsymbol{T}$,
some of which are summarized in Table 1 and illustrated in Figure
\ref{fig:Nonlinearities}.
\begin{table}[tbh]
\caption{\label{tab:Nonlinearities}Example parameters of nonlinearity $f(z)$}
\centering{
\begin{tabular}{|c|>{\raggedright}p{0.85\columnwidth}|}
\hline
\emph{No.} & \emph{Parameters}\tabularnewline
\hline
1 & $\boldsymbol{a}G_{0}^{-1}$= \{1 2 2 -2 -2 2 2 -2 -2 -0.5\}\newline$\boldsymbol{b}$
= \{0 -2 -2.5 4 4.5 -4 -4.5 6 6.5 2.5\}\newline$\boldsymbol{T}G_{0}$=
\{0 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3\}; $G_{0}=0.53$\tabularnewline
\hline
2 & $\boldsymbol{a}G_{0}^{-1}$= \{1 2 2 -2 -2 2 2 -2 -2 -0.5\}\newline$\boldsymbol{b}$
= \{0 -2 -3.5 4 3.5 -4 -4.5 6 6.5 2.5\}\newline$\boldsymbol{T}G_{0}$=
\{0 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3\}; $G_{0}=0.5125$\tabularnewline
\hline
3 & $\boldsymbol{a}G_{0}^{-1}$= \{1.25 2 2 -2 -2 2 2 -2 -2 -0.5\}\newline$\boldsymbol{b}$
= \{0 -1.6 -3.1 3.6 3.1 -3.6 -4.1 5.6 6.1 2.4\}\newline$\boldsymbol{T}G_{0}$=
\{0 0.8 1.05 1.3 1.55 1.8 2.05 2.3 2.55 2.8\}; $G_{0}=0.415$\tabularnewline
\hline
\end{tabular}
\end{table}
\begin{figure}[tbh]
\includegraphics[scale=0.61]{nonlinearity2}
\caption{\label{fig:Nonlinearities}Three examples of nonlinear functions $f(z)$:
nonlinearity 1 (top), nonlinearity 2 (middle), and nonlinearity 3
(bottom).}
\end{figure}
\section{Simulation results}
The performance of the proposed modulation scheme with the GAMP-based
decoder was studied by means of Monte-Carlo simulation. Our initial
simulation-based study revealed that the conventional GAMP algorithm
(with $\alpha=1$, and $\beta=1$) exhibits several deficiencies,
in particular, relatively high error-floor, slow convergence and occasional
instability problems (divergence with the increase of number of iterations).
It turns out that the error-floor issue was, at least, in part due
to the decoder algorithm. To alleviate these deficiencies we implemented
several enhancement techniques. Firstly, we appended the cyclic-redundancy
check (CRC) block to each data frame and used it for the early stopping
criterion. Secondly, after each decoding iteration we calculated the
Euclidean distance $E(t)$ between the received vector $\{y_{n}\}$
and the reconstructed waveform $f\left(\sum\limits _{k=0}^{N-1}F_{n,k}\hat{x}_{k}\left(t\right)\right)$,
and if a CRC error was detected after $t_{max}$ iterations the final
decision was based on the vector $\left\{ \hat{x}_{k}\left(t\right)\right\} $
that corresponded to the minimum Euclidean distance $E(t)$. Although
this procedure did not improve FER, it minimized the BER for incorrectly
decoded frames. Thirdly, we experimentally discovered that the decoder
convergence speed can be significantly improved by using noise scaling
factor $\alpha<1$ simultaneously with damping ($\beta<1$). However,
such a technique could result in occasional algorithm divergence,
therefore to achieve better performance we implemented the following
procedure: $t_{max}/2$ iterations were performed using the modified
GAMP algorithm ($\alpha<1$, $\beta<1$), and if the decoder could
not converge within the first $t_{max}/2$ iterations, all internal
variables were reset and the subsequent $t_{max}/2$ iterations were
performed with the conventional GAMP settings ($\alpha=1$, $\beta=1$).
The overall decoding algorithm flowchart is depicted in Figure \ref{fig:Decoding-algorithm-flowchart}.
\begin{figure}[tbh]
\includegraphics[scale=0.53]{algorithm}\caption{\label{fig:Decoding-algorithm-flowchart}Decoding algorithm flowchart}
\end{figure}
Our simulation results indicate that the performance of the proposed
scheme with WHT, real DFT or complex DFT with Cartesian-type nonlinearity
is almost identical. Therefore, from complexity point of view it seems
preferable to use WHT, and, hence, most of our simulation results
reported here were obtained for WHT-based scheme.
Figure \ref{fig:BER-vs-EbNo-capacity} illustrates the BER vs $E_{b}/N_{0}$
curves for 2-PAM input modulation with three nonlinearities (Table
\ref{tab:Nonlinearities}/Figure \ref{fig:Nonlinearities}) and different
values of $N$. In all simulations, the variance of $\{z_{n}\}$ was
normalized to 1, the maximum number of iterations ($t_{max}$) was
set to 100, and integrals (\ref{eq:Integral_start})\textendash (\ref{eq:Integral_end})
were approximated using numerical summation. At the initialization
step the parameters $\alpha$, $\beta$ were set to $\alpha=0.71$,
$\beta=0.875$.
Remarkably, for the selected nonlinearity parameters $G_{0}$, $\boldsymbol{a}$,
$\boldsymbol{b}$, $\mathbf{T}$ the proposed modulation scheme exhibits
behavior similar to random-like codes with the presence of a waterfall
region and error-floor, and, as expected, performance improves for
larger frame sizes. Moreover, the proposed modulation scheme with
nonlinearity 3 and $N$=16384 achieves target BER=$10^{-5}$ at $E_{b}/N_{0}=3.3$
dB, which represents 6.3 dB gain over uncoded 2-PAM or 4-QAM modulation
and is only about 1.5 dB away from the \emph{unconstrained} AWGN channel
capacity. These results are better or within $0.1\div0.3$ dB of the
performance of the capacity-approaching, bandwidth-efficient modulation
schemes with $\eta=2$ bit/s/Hz reported in {[}1-3, 9, 10{]}. BER
curves for some of these advanced coded modulation schemes with comparable
frame sizes are reproduced in Figure \ref{fig:BER-vs-EbNo} for comparison.
\begin{figure}[tbh]
\includegraphics[scale=0.57]{n1K_16K_capacity}\caption{\label{fig:BER-vs-EbNo-capacity}BER vs $E_{b}/N_{0}$ for the proposed
modulation scheme with different nonlinearities and frame sizes in
AWGN channel ($\eta=2$ bit/s/Hz)}
\end{figure}
Although we set $t_{max}=100$, the average number of decoder iterations
was significantly lower. For example, at $BER=10^{-5}$ the average
number of decoder iterations was about 26 for the frame size $N$=16384,
and about 17 for the frame size $N$=4096, and only 7 iterations for
the frame size $N$=1024.
A considerable performance improvement over uncoded modulation was
also observed for 4-PAM or 16-QAM input modulation ($\eta=4$ bit/s/Hz).
However, in case of 4-PAM/16-QAM, the GAMP algorithm is sub-optimal
in terms of BER performance, and the example nonlinearities that we
use for 4-QAM/2-PAM (Table \ref{tab:Nonlinearities}) are not optimal
either. Therefore, application of the proposed technique to high-order
modulation formats ($\eta\geq4$ bit/s/Hz) is an open research topic.
\begin{figure}[tbh]
\includegraphics[scale=0.46]{7a_n16384_compare_w_papers_light}
\caption{\label{fig:BER-vs-EbNo}Performance comparison of the proposed technique
(WHT, Nonlinearity 3, $N$=16384) with state-of-the-art coded modulation
schemes ($2$ bit/s/Hz):\protect \\
1) ARJ4A LDPC, $N$=16384, 16-APSK {[}9{]}, \protect \\
2) ARJ4A LDPC, $N$=16384, 8-PSK {[}9{]}, \protect \\
3) Regular LDPC, BICM, $d_{v}$=3, $n$=20000, 4-PAM {[}1{]}\protect \\
4) Regular LDPC, BICM, $d_{v}$=3, $n$=10000, 4-PAM {[}1{]}\protect \\
5) Irregular LDPC, BICM, $d_{v}$=15, $n$=20000, 4-PAM {[}1{]}\protect \\
6) Irregular LDPC, BICM, $d_{v}$=15, $n$=10000, 4-PAM {[}1{]}\protect \\
7) eIRA LDPC, BICM, $n$=10000, 4-PAM {[}10{]}\protect \\
8) eIRA LDPC, BICM, $n$=9000, 8-PSK {[}10{]}\protect \\
9) Turbo TCM, $N$=10000, 8-PSK {[}3{]}}
\end{figure}
\section{Discussion}
We believe that the proposed joint coding-modulation technique might
be useful in some wired and wireless communication applications, since
it has several interesting properties:
\begin{itemize}
\item Good BER performance: As illustrated in the previous section the BER
performance is on par with that of the best known coded modulation
schemes.
\item Relatively low decoder complexity: The GAMP-based decoder complexity
per iteration is dominated by two orthogonal transform operations,
which, in case of fast WHT, require only $N\log_{2}(N)$ real additions
or subtractions. The input/output nonlinear step is generally a scalar
operation with complexity of order $O(N)$. Although, in our simulation
model, we used numerical integration to approximate integrals (\ref{eq:Integral_start})\textendash (\ref{eq:Integral_end}),
these can be expressed in closed-form using tabulated Gauss error
function, and, consequently, can be simplified or implemented via
lookup tables. It is also possible to use simpler max-sum version
of the GAMP algorithm to trade-off performance and complexity {[}7{]}.
\item The choice of nonlinearity $f(z)$ may be tailored to the requirements
of the transmission system in order to improve overall system efficiency.
For example, $f(z)$ may be selected to improve power conversion efficiency,
or illumination-to-communication conversion efficiency in visible
light communication applications {[}11{]}.
\item An interesting application of the proposed modulation scheme is to
use it in combination with a conventional, linear OFDM transmitter,
as illustrated in Figure \ref{fig:precoder_for_ofdm}. Such a system
arrangement may result in the OFDM signal with reduced PAPR. Due to
the presence of nonlinearity $f(z)$, the distribution of signal samples
at the output of the modulator illustrated in Figure \ref{fig:precoder_for_ofdm}
resembles the distribution of $M$-QAM/$M$-PAM signal affected by
additive Gaussian noise. It can be shown that if the nonlinearity
$f(z)$ has a relatively large linear region the PAPR of such a signal
will be much lower than that of a conventional OFDM signal. Figure
\ref{fig:CCDF} compares the complementary cumulative distribution
functions (CCDF) of PAPR for signals generated by the transmission
system illustrated in Figure \ref{fig:precoder_for_ofdm} with nonlinearity
1 (complex-valued OFDM with Cartesian-type nonlinearity), and a conventional
OFDM system. We analyzed CCDF for Nyquist sampled waveforms and four-times
oversampled waveforms, since four-times oversampling provides a good
approximation of the continuous-time PAPR {[}12{]}. As may be seen,
at probability $10^{-4}$ the PAPR of the continuous signal generated
by the transmission system illustrated in Figure \ref{fig:precoder_for_ofdm}
is approximately 2.3 dB lower than the PAPR of a conventional OFDM
signal. The difference is increased to 3.9dB for Nyquist-sampled waveforms.
\end{itemize}
It should be noted that the proposed method has several limitations
and open issues. Firstly, in this paper, we focused primarily on the
case $\eta=2$ bit/s/Hz. One possible way to apply this technique
to systems with higher spectral efficiency is to use the output waveform
puncturing. This approach (i.e. random puncturing) seems to work well
up to $\eta=2.5\div3$ bit/s/Hz. It is also possible to use higher-order
input modulations (16-QAM/4-PAM), however, as we mentioned earlier,
extension of this technique to spectral efficiency $\eta\geq4$ bit/s/Hz
is still an open research topic. Secondly, the GAMP algorithm is based
on Gaussian and quadratic approximations, and, therefore, it is apparently
sub-optimal, especially, for small frame sizes. We were able to improve
performance of the GAMP-based decoder using several \emph{ad hoc}
techniques. However, it is still unclear, how close the GAMP decoder
performance is to the theoretical ML performance. Finally, our choice
of nonlinearity $f(z)$ is generally based on a heuristic and experimental
approach. The problem here is that the performance of the proposed
scheme is limited not only by distance distribution of encoded waveforms
but also by non-idealities of the decoding algorithm, and this fact
greatly complicates optimization strategy.
\begin{figure}[tbh]
\includegraphics[scale=0.8]{ofdm_precoder}
\caption{\label{fig:precoder_for_ofdm}Proposed modulation scheme as a pre-coder
for a conventional OFDM transmitter}
\end{figure}
\begin{figure}[tbh]
\includegraphics[scale=0.5]{ccdf_papr}
\caption{\label{fig:CCDF}CCDF of PAPR for the proposed pre-coded OFDM (Figure
\ref{fig:precoder_for_ofdm}) with $N$=1024 and nonlinearity \#1:\protect \\
1) Conventional OFDM (Nyquist sampling)\protect \\
2) Conventional OFDM (four-times oversampling)\protect \\
3) OFDM with the proposed modulation (Nyquist sampling)\protect \\
4) OFDM with the proposed modulation (four-times oversampling)}
\end{figure}
\balance
\section{Conclusions}
In this paper, we proposed a novel joint coding-modulation technique
based on serial concatenation of orthogonal linear transformation
(e.g., WHT or DFT) with memoryless nonlinearity. We demonstrated that
such a simple signal construction may exhibit properties of a random
code ensemble, as a result approaching channel capacity. Our computer
simulations confirmed that if the decoder relies on the approximate
message passing algorithm, the proposed modulation technique exhibits
performance on par with state-of-the-art coded modulation schemes
that use capacity-approaching component codes. The proposed technique
could be extended to modulation formats with higher spectral efficiency
($\eta\geq4$ bit/s/Hz) and other types of orthogonal transformations,
offering one possible direction for our future research.
|
\section{Introduction}
Granular matter is ubiquitous in nature. However, due to the complexity of the real particle--particle interactions, the standard approaches of continuum mechanics and thermodynamics are still limited in providing meaningful descriptions of the states in which these systems can be. Edwards and Oakeshot introduced a tentative approach inspired by the ideas of equilibrium statistical mechanics to formally describe the global properties of a static granular pack. Since the introduction of this theory ---where the entropy of the systems is governed by the spatial disorder of the grains \cite{edwards}---, a number of studies have used it to frame the interpretation of the results of specific experiments. The most relevant case is the so-called ``Chicago experiment'', where a column of grains was repeatedly tapped following an annealing-type protocol \cite{chicago,chicago2}. The main outcome of this experiment is that a stationary state can be reached, where the mean volume fraction, $\phi$, is a well defined
function of the tap amplitude, $\Gamma$. Others have also obtained seemingly reproducible states without the need of annealing \cite{ribiere}. However, it has been shown recently, by simulation of frictionless grains, that these stationary states are not necessarily ergodic \cite{frenkel}. At low $\Gamma$, different members of an ensemble of steady-states prepared with a well defined protocol may sample a different region of the phase space, as the fluctuations of $\phi$ indicate.
In this paper, we demonstrate that not only the volume but also the force moment tensor, $\Sigma$, are sampled in a non-ergodic fashion and that ergodicity seems to be recovered if all particle--particle contacts are lost during each tap. This sets a clear limit to the range of driving forces able to generate a sequence of configurations for which the Edwards framework can be applied.
\section{Numerical protocol}
We simulated using the LAMMPS package \cite{plimpton2007} a quasi-two-dimensional cell containing $N=1000$ spherical particles of diameter $d$. The cell is $1.1\;d$ thick and $27.8\;d$ wide (the granular column is about 35 layers deep) to have a one to one representation of a previously introduced experimental device \cite{pugnaloni2010,ardanza2014topological}. We use a model for soft frictional spheres described in Refs. \cite{brilliantov1996, silbert2001}. The normal component, $F_\mathrm{n}$, of the contact interaction is given by an elastic repulsive force proportional to the overlap of the interacting spheres and a dissipative term proportional to the normal component of the relative velocity. The tangential term, $F_\mathrm{t}$, implements an elastic shear force and a damping force. The shear force takes into account the cumulated tangential displacement between the particles while they remain in contact. Whenever $F_\mathrm{t} > \mu F_\mathrm{n}$ ($\mu$ is the
friction coefficient), this lower dynamic friction force
is used. In this work we use the same interaction parameters as in Ref. \cite{pugnaloni2008,pugnaloni2011}. The wall--particle interaction is defined with the same parameters as the particle--particle force. Tapping is simulated by imposing an external vertical motion to the cell. This pulse is a single sinusoidal cycle $A \sin(\omega t) $. We fix $\omega=2\pi\times33$ Hz and use the tap amplitude $A$ as control parameter. The tap intensity is characterized by $\Gamma=A\omega^2/g$. The mechanical equilibrium after each tap is deemed achieved if the kinetic energy of each particle has fallen (in average) below $10^{-6} mgd$. Where $m$ is the mass of one particle and $g$ the acceleration of gravity.
We study 20 independent realizations of a decreasing ramp of the tap amplitude. We initially fill the cell by placing the spheres at random positions before letting them deposit under the action of gravity. In each realization, we decrease $\Gamma$ in small steps, from $\Gamma=20.0$ down to $\Gamma=0.8$, and apply $200$ taps for each $\Gamma$. Note that for $\Gamma<1.0$, the column of grains does not detach from the base during a tap. The $200$ taps at each value of $\Gamma$ are enough to reach a steady-state. We do not observe any drift of the mean values of $\phi$ or $\Sigma$ after the initial $100$ taps, which we will discard later in our analysis. Finally, we also study a cyclic annealing protocol: starting from the final configuration at $\Gamma=0.8$ for each of the former realizations; the tap amplitude is cyclically increased and decreased every 200 taps in the range $0.8<\Gamma<5.5$ in order to compare the steady-states reached with an alternative method.
\section{Data analysis}
To measure the packing fraction we use the 2D Voronoi tessellation (implemented in \cite{voro++}) of the $x$--$z$ plane projection of the particle positions, disregarding the third coordinate on the thin direction of the cell. Then, we associate to each particle a ``local volume'' fraction by dividing the particle area by the corresponding Voronoi area. In order to avoid boundary effects, we disregard particles closer than $2d$ to the lateral walls. Following the recommendations in Ref. \cite{gago2015}, we analyze horizontal slices of the granular column $15\;d$ thick measured at approximately the same depth with respect to the free surface in order to retrieve unbiased results for the force moment tensor due to the uneven free surface. Averaging over the $N$ particles contained in the slice of interest, we obtain the volume fraction of each static configuration. To obtain the steady-state $\phi$ corresponding to a given $\Gamma$, we averaged this quantity over the last
$100$ configurations obtained for each tap intensity. We
also obtain the force moment tensor $\sigma_{i}^{\alpha\beta}= \sum \limits_{c} r_c^{\alpha} f_c^{\beta}$ of each particle in the slice of interest. Here, the sum runs over all the contacts $c$ of the particle, $\overrightarrow{r}_c$ is the vector from the center of the grain to the contact $c$ and $\overrightarrow{f}_c$ is the corresponding contact force. We apply the same averaging protocol used for $\phi$ to obtain the force moment tensor for the configuration and $\Sigma$ for the steady-state of a given realization and $\Gamma$.
\begin{figure}[t!]
\includegraphics[width=1\columnwidth,trim= 0cm 0cm 0 0cm,clip]{fig1.eps}
\caption {Ensemble average of the steady-state packing fraction $\phi$ (a) and trace of the force moment tensor $\rm{Tr}(\Sigma)$ (b) as a function of $\Gamma$. The error bars correspond to the standard deviation over the $20$ averaged realizations. The insets show the same results and two of the $20$ independent realizations (dashed lines) in the low-tap-intensity region. The error bars on the single realization data correspond to the estimated standard error of the mean.}
\label{Fig1}
\end{figure}
\begin{figure}[t]
\includegraphics[width=1\columnwidth,trim= 0cm 0cm 0 0cm,clip]{fig2.eps}
\caption {$\phi$ (a) and $\rm{Tr}(\Sigma)$ (b) as a function of the tap number for two of the $20$ independent realizations at $\Gamma=0.8$. The notched boxes and ``violin'' diagrams shown suggest that both realizations can be hardly considered as representing the same steady-state.}
\label{Fig2}
\end{figure}
\section{Results}
In Fig.~\ref{Fig1}(a) the ensemble average of the steady-state $\phi$ (i.e., averaged over the 20 independent realizations) is displayed as a function of $\Gamma$. The error bars correspond to the standard deviation over the 20 realizations. As observed by a number of authors, the curve seems to be very well defined with independent realizations falling within a very narrow range of $\phi$ values for any given $\Gamma$. In the past, this led to the conclusion that this was a truly reversible process, where lowering or raising $\Gamma$ would lead to the same steady-state $\phi$. In the inset of Fig.~\ref{Fig1}(a) two of the independent realizations are shown for the low-tap-intensity region. From this picture, it is clear that steady-states corresponding to a given $\Gamma$ can differ from one realization to another. Notice that in the inset the error bars for the two isolated realizations correspond to the standard error of the mean (SEM), which gives an estimate of the uncertainty of the mean
value reported rather than the size of the $\phi$ fluctuations. For these two realizations, although the mean $\phi$ seems to agree within the estimated error for intensities $\Gamma>1.5$, it is clear that they are different for low $\Gamma$. This is consistent with the findings of Paillusson and Frenkel \cite{frenkel} for frictionless spheres under event-driven simulations. However, in our simulations we are able to extract the stress state of the system as well as the history of the contacts. These reveal valuable information, as we discuss below.
In Fig.~\ref{Fig1}(b), we show the trace $\rm{Tr}(\Sigma)$ of $\Sigma$ averaged over all 20 realizations as a function of $\Gamma$. As before, the error bars indicate the standard deviation over the 20 realizations. As we can see, the variability of the mean stress is significantly large at low $\Gamma$. This is not due to the large fluctuations during a given series of taps but to the variations observed from one realization of the protocol to the other. Indeed, the inset in Fig.~\ref{Fig1}(b) shows that, for low $\Gamma$, the mean values of $\rm{Tr}(\Sigma)$ for two of the 20 realizations have a relatively small SEM (i.e., the fluctuations in each realization are small). However, realizations differ from each other. The difference between results corresponding to different realizations become much more evident here than in the case of the $\phi$--$\Gamma$ plot. We suggest that the stress tensor may be more sensitive and then more suitable to sense if ergodicity is fulfilled in experimental data. Overall, as
$\Gamma$ is decreased, different realizations explore non-overlapping ranges of volume/stress. Therefore, temporal averages (on a single time series) do not match with ensemble averages (over the realizations).
\begin{figure}[t!]
\includegraphics[width=1\columnwidth,trim=0mm 0mm 0mm 0mm, clip]{fig3.eps}
\caption {Steady-state $\phi$ (a) and $\rm{Tr}(\Sigma)$ (b) as a function of $\Gamma$ corresponding to a full annealing protocol. Starting from the filled circles (red) increasing ramp, and following by filled squares (blue), open circles (green) and open squares (magenta). The error bars correspond to the estimated SEM.}
\label{Fig3}
\end{figure}
Since the number of taps we have explored for each $\Gamma$ may be small to assume that the steady-state has been properly sampled, we carried out $2000$ additional taps at $\Gamma=0.8$ for each realization. Since some of the signals are not normally distributed, we confirmed the stationarity of these states by using a non-parametric test at a level of significance of $5 \%$ \cite{fractal,R}. Two of the normally distributed realizations are displayed in Fig.~\ref{Fig2} as a function of the tap number. Comparing the corresponding notched boxes and ``violin'' diagrams of both signals, it is clear that the states do not match. Hence, the \emph{reversible branch} found in Ref. \cite{chicago2} is not so at low $\Gamma$ in our case since truly stationary states with distinguishable $\phi$ and $\rm{Tr}(\Sigma)$ may be obtained on different realizations of the annealing protocol. Of course, this is harder to detect in $\phi$ since the dispersion between realizations is much smaller compared with the
range of $\phi$ values obtained at different $\Gamma$.
The former results also confirm the hypothesis that non-ergodicity is present in a typical tapping protocol beyond the special case reported in Ref. \cite{frenkel}, where the steady-states obtained did not follow any annealing-type protocol. Hence, we observe this non-ergodic behavior even after annealing the system from high tapping strengths. To stress this point and assess if the speed of the annealing may prevent the system from reaching a unique steady-state on each realization, we apply a slower cyclic annealing protocol (similar to the one introduced in \cite{chicago}) to each of the 20 final states at low $\Gamma$ in order to reproduce the ``reversible branch''. In Fig.~\ref{Fig3} we display a sequence of two successive up and down ramps applied to one of the 20 initial realizations using $\Gamma$-steps about one half of those used in Fig. \ref{Fig1}. Although in the scale used for $\phi$ the steady-state packing fraction seems reversible, a close inspection shows that the states have distinguishable
$\phi$ at low $\Gamma$ [see Fig. \ref{Fig3}(a) and the corresponding inset]. This is much more evident when the stress is analyzed [see Fig.~\ref{Fig3}(b)].
\begin{figure}[t!]
\includegraphics[width=1\columnwidth,trim= 0cm 0cm 0 0cm,clip]{fig4.eps}
\caption {P-value (up-triangles, right axis) as a function of $\Gamma$ for the Kruskal--Wallis \cite{kruskal} one-way analysis of variance for $\phi$ (a) and $\rm{Tr}(\Sigma)$ (b). The horizontal dotted line corresponds to the significance level used (1\%). The black circles correspond to the $\phi$ and $\rm{Tr}(\Sigma)$ data from Fig.~\ref{Fig1}.}
\label{Fig4}
\end{figure}
In order to set a criterion to decide if the steady-states are not ergodic for a given $\Gamma$, we show in Fig. \ref{Fig4} the $p$-values for the Kruskal--Wallis \cite{kruskal} one-way analysis of variance performed on the 20 realizations at each $\Gamma$. This simple non-parametric test allows for the rejection of the null hypothesis that all 20 data series are drawn from a unique distribution (which does not need to be normal), hence that they correspond to a unique steady-state. If the $p$-value is significant (in our case $p>0.01$), then we cannot rule out the possibility that the 20 series come from the same steady-state. As we can see from Fig. \ref{Fig4}(a), the test run on the data for $\phi$ indicates that for $\Gamma<5.0$, the null hypothesis must be rejected and therefore there exist at least two out of the 20 steady-states that are not the same. However, for higher $\Gamma$, the test is significant and then the 20 realizations may correspond to the same steady-state. Interestingly, when the test
is
run on $\Sigma$ [see Fig. \ref{Fig4}(b)], the steady-state seems to be unique for all 20 realizations if $\Gamma>3.75$. Although differences between realizations are simpler to detect on visual inspection of $\rm{Tr}(\Sigma)$, it is actually $\phi$ that sets a higher threshold for the $\Gamma$ values needed to ensure an ergodic steady-state (i.e., $\Gamma>5.0$).
\begin{figure}[ht!]
\includegraphics[width=1\columnwidth,trim= 0cm 0cm 0 0cm,clip]{fig5.eps}
\caption {(a) Percentage $C/C_0 \times 100$ of persistent contacts (black circles) as a function of $\Gamma$, averaged over $5$ taps on $6$ independent realizations. The error bars correspond to the standard deviation over the $6$ realizations. Up-triangles (green) correspond to the p-values in Fig.~\ref{Fig4} for $\phi$ and down-triangles (magenta) to $\rm{Tr}(\Sigma)$. (b) Same as (a) for frictionless grains.}
\label{Fig5}
\end{figure}
The previous results indicate that the steady-states sampled at low tap intensities do not only depend on $\Gamma$ but on the particular history of each realization. Notice that this goes beyond the history dependent out-of-equilibrium trajectories already reported in tapped systems \cite{josserand} since here we are focusing on the steady-states. One may hypothesize that the constraints imposed by the contacts is one of the reasons for the non-ergodic behavior at low tap. If a contact persists from one tap to the next, the contact force after coming back to rest will depend on the history of all contacts of that particular grain. In order to test this idea, we analyze the evolution of all contacts during each tap and identify those that persist (i.e., contacts that did not break at any time during the pulse of energy). Figure ~\ref{Fig5} shows the average ratio $C/C_0$ of persistent contacts, $C$, to the total number of contacts, $C_0$, as a function of $\Gamma$. For this calculation, each contact was
tracked during the final $5$ taps for each $\Gamma$ on $6$ of the independent realizations and only grains that fall within the layer of interest, as discussed above, where included in the analysis. The percentage of persistent contacts is very small but non-zero up to $\Gamma \approx 5.0$. As it is expected, when $\Gamma$ is increased sufficiently all the contacts are broken and new ones are made during each tap (resulting in $C/C_0=0$). This transition coincides with the value of $\Gamma$ where the realizations seem to sample the same steady-state (see the $p$-values included in Fig.~\ref{Fig5}). Therefore, when small taps are applied, the aging of some of the contacts seem to lead the system to sample different regions of the phase space during independent realizations. However, if all contacts are made anew at each tap, the sampling becomes compatible with the idea of ergodicity introduced in Fig.~\ref{Fig4}. In order to generalize this result, we also simulated
frictionless grains. Interestingly, the same conclusion drawn for frictional grains is true for frictionless ones: different realizations seem to sample the same steady-state only if all contacts are made anew upon each tap [see Fig.~\ref{Fig5} (b)].
\section{Conclusions}
Our analysis of the steady-states of tapped granular systems indicate that these states are history-dependent for tap intensities below a certain threshold. This is in contradiction with the general assumption that macroscopic time averages ---such as the volume fraction--- can be recovered when the amplitude of the perturbation applied to the system is tuned back and forth. The differences between independent realizations become particularly noticeable in the stress distribution. These findings show that the postulates of the equilibrium statistical thermodynamics may not be always fulfilled to describe the steady state of static granular systems (see also Ref. \cite{irastorza} for a discussion on the Boltzmann distribution failure for an analytically solvable model). Focusing on tap intensities that warrant that all contacts are made anew after each tap may allow exploring the available phase space in agreement with the ergodic hypothesis. However, gentle perturbations deserve an approach that includes memory effects to suitably describe the states.
In that sense, non-equilibrium thermodynamic approaches may be a suitable alternative \cite{lebon}. Further research on such alternative formalisms, the effect of other types of forcing mechanisms (e.g., shear), and possible extensions to other complex systems (e.g., active matter) become necessary.
\begin{acknowledgements}
This work has been partially supported by projects PICT-2012-2155 ANPCyT (Argentina), FIS2011-26675 MINECO (Spain) and PIUNA (Universidad de Navarra).
\end{acknowledgements}
|
\section{Introduction}
Rogue (freak) waves are generally described as high amplitude waves with a height bigger than $2-2.2$ times the significant waveheight in a stochastic wavefield \cite{Kharif}. They have been extensively studied in recent years in the fields including but are not limited to hydrodynamics, optics, quantum mechanics, Bose-Einstein condensation, acoustics and finance, just to name a few \cite{Akhmediev2009b, bayindir2016, Akhmediev2009a, Akhmediev2011, FirstOpticalRW, Bay_Zeno, Bay_arxNoisyTun, Bay_arxChaotCurNLS}. The research has started with the investigation of the nonlinear Schr\"{o}dinger equation (NLSE). Discovery of the unexpected rational rogue wave solutions of the NLSE resulted in seminal studies of rogue waves, such as \cite{Akhmediev2009b}. Rogue wave dynamics of some of the extensions of the NLSE, such as the Sasa-Satsuma and the Kundu-Eckhaus equations, are also studied recently \cite{Soto2014RwSSchaotic, BayPRE1, BayPRE2, Bay_arxNoisyTunKEE}. It is natural to expect that in a medium whose dynamics are governed by nonlinear equations such as the NLSE and NLSE, rogue waves can also emerge, therefore investigation of the dynamics of different models needs further attention.
Development of the rogue wave early warning systems and technology is an active area of research and is crucially important for the marine environment to safeguard the ocean travel, oceanic structures and machinery such as wave energy harvesters \cite{Akhmediev2011, bayindir2016, Bay_arxEarlyDetectCS}. Two of the few early detection methods proposed in 1D are to use the emerging triangular Fourier spectra (i.e. triangular supercontinnum generation) to detect if a rogue wave is going to emerge and to use the emerging wavelet spectra to locate its emergence location \cite{Akhmediev2011, bayindir2016, Bay_arxEarlyDetectCS}. These methods work well for the single rogue waves observed in fiber optics and hydrodynamic wave flumes and lead to early warning time scales on the order of the temporal width of the rogue wave. However enhancement of the early warning times for stochastic wavefields requires further attention and development of realistic solutions such as the development of the electronic equipment to capture rogue wave emergence may take many long efforts.
To our best knowledge the early detection rogue waves are only studied in 1D and no studies exist about the early detection mechanisms of 2D rogue waves. With this motivation, we analyze the spectral properties of 2D rogue waves. Since the correct form of the 2D NLSE is not integrable, we use a radially symmetric version of the 1D NLSE and its Peregrine and Akhmediev-Peregrine solitons solutions. Although this form of the 2D NLSE does not rely upon an analytical basis, it can exhibit the characteristics of the localized peaked structures of 2D rogue wave profiles and their conical spectral forms, very similar to the 1D case. We propose to use the emerging conical spectra of the 2D rogue waves before they become evident in time as an early detection technique and thus we discuss their dynamics. With this aim, we propose an efficient method for the acquisition of the emerging conical 2D rogue wave spectra. We first construct the 1D Radon transforms of the emerging conical 2D spectra of the wavefield using CS. Then we construct 2D spectra from those projections using FBP. Since emerging 2D conical spectra can be treated as a sparse signal, the method can successfully capture the emerging conical spectra. We numerically show that this approach can produce indistinguishable results from the classical sampling approach, but it supersedes classical sampling approach due to greatly reduced sampling requirement.
\section{Methodology}
\subsection{Review of the Nonlinear Schr\"{o}dinger Equation}
The 2D dynamics of nonlinear ocean waves, optical waves and quantum vibrations can be modeled by the 2D NLSE \cite{Akhmediev2011, Bay_Zeno, Zakharov1968}. Since 2D NLSE is not integrable, some integrable extensions are proposed in the literature which admits ration soliton solutions \cite{KunduArxiv}. However, whether they can model the realistic dynamics of 2D rogue waves or not is a question which needs further attention. In order to analyze the early detection mechanism of the 2D rogue waves, we consider radially symmetric version of the 1D NLSE in this study. Although this 2D model does not have an analytical basis it can exhibit the localized peak structures of the rogue waves. The radially symmetric rational soliton solutions of the NLSE can be used to understand the dynamics of 2D rogue waves, which are accepted as accurate rogue wave models in 1D \cite{Akhmediev2009b}. Thus we consider the radially symmetric NLSE given as
\begin{equation}
i\psi_t + \frac{1}{2} \psi_{rr} + \left|\psi \right|^2 \psi =0
\label{eq01}
\end{equation}
where $r=\sqrt{x^2+y^2},t$ are the spatial and temporal variables, $i$ is the imaginary number and $\psi$ is the complex amplitude known as the wavefunction in optics and quantum mechanics but the wavefield envelope in hydrodynamics. This notation is mainly used in hydrodynamics and quantum mechanics whereas $t$ and $r$ axes are switched in fiber optics studies, where NLSE is used to describe the dynamics of light pulses in nonlinear fiber optical media. It is known that the NLSE given by Eq.(\ref{eq01}) admits many different types of analytical solutions among which the first and higher order rational soliton solution are considered as accurate rogue wave models \cite{Akhmediev2009b}. For stochastic wavefields where the analytical solution is unknown, the NLSE can be numerically solved by some numerical techniques such as the spectral method \cite{bay2009, demiray, BayTWMS2016, bay_cssfm, Karjadi2010, Karjadi2012, Bay_cssfmarx, bayindir2016nature, canuto, trefethen}. However in this study we limit ourselves with the analytical solutions of the NLSE. The radially symmetric 2D Peregrine soliton can be written as
\begin{equation}
\psi_1=\left[1-4\frac{1+2it}{1+4r^2+4t^2} \right] \exp{[it]}
\label{eq02}
\end{equation}
where $t$ and $r$ denotes the time and space, respectively \cite{Akhmediev2009b, Peregrine}. The Peregrine soliton is only a first order rational soliton solution in the Darboux hierarchy of the NLSE and higher order rational soliton solutions do exist \cite{Akhmediev2009b}. Throughout many simulations \cite{Akhmediev2009b, Akhmediev2009a, Akhmediev2011} and some experiments \cite{Kibler}, it has been confirmed that rogue waves can be in the form of the first (Peregrine) and higher order rational soliton solutions of the NLSE.
Second order rational soliton solution of the NLSE is Akhmediev-Peregrine soliton \cite{Akhmediev2009b}, which is considered to be a model for rogue waves with higher amplitude than the Peregrine soliton. The formula of Akhmediev-Peregrine soliton is given as
\begin{equation}
\psi_2=\left[1+\frac{G_2+it H_2}{D_2} \right] \exp{[it]}
\label{eq03}
\end{equation}
where
\begin{equation}
G_2=\frac{3}{8}-3r^2-2r^4-9t^2-10t^4-12r^2t^2
\label{eq04}
\end{equation}
\begin{equation}
H_2=\frac{15}{4}+6r^2-4r^4-2t^2-4t^4-8r^2t^2
\label{eq05}
\end{equation}
and
\begin{equation}
\begin{split}
D_2=\frac{1}{8} & [ \frac{3}{4}+9r^2+4r^4+\frac{16}{3}r^6+33t^2+36t^4 \\
& +\frac{16}{3}t^6-24r^2t^2+16r^4t^2+16r^2t^4 ]
\label{eq06}
\end{split}
\end{equation}
where $t$ is the time and $r$ is the space parameter \cite{Akhmediev2009b}. Using Darboux transformation formalism this soliton can be obtained using the Peregrine soliton as the seed solution \cite{Akhmediev2009b}. Many numerical simulations also confirm that rogue waves in the NLSE framework can also be in the form of Akhmediev-Peregrine soliton \cite{Akhmediev2011, Akhmediev2009b, Akhmediev2009a} however to our best knowledge an experimental verification of this soliton do not exist yet. We use 2D radially symmetric versions of the Peregrine and Akhmediev-Peregrine solitons as 2D rogue wave models.
\subsection{Review of the Compressive Sampling}
\noindent Compressive sampling (CS) is an efficient sampling technique which exploits the sparsity of the signal for its reconstruction by using far fewer samples than the requirements of the classical Shannon-Nyquist sampling theorem states \cite{Candes, Candes2006}. CS has been intensively studied as a mathematical tool in applied sciences and engineering and currently some engineering devices such as the single pixel video cameras and efficient A-D converters relies on CS algorithm. We try to give a very brief summary of the CS in this section and refer the reader to \cite{Candes, Candes2006} for a comprehensive discussion and derivation.
Let $\psi$ be a $K$-sparse signal with $N$ elements, that is only $K$ of the $N$ elements of $\psi$ are nonzero. Using orthonormal basis transformations with transformation a matrix of ${\bf \Psi}$, $\psi$ can be represented in any transformed domain in terms of the basis functions. Most common orthogonal transformations used in the literature are the Fourier, wavelet or discrete cosine transforms. Using the orthogonal transformation it is possible to rewrite the signal as $\psi= {\bf \Psi} \widehat{ \psi}$ where $\widehat{ \psi}$ is the coefficient vector. Keeping the non-zero coefficients and discarding the zero coefficients of $\psi$, it is possible to get $\psi_s= {\bf \Psi}\widehat{ \psi}_s$ where $\psi_s$ denotes the signal with non-zero entries only.
CS algorithm guarantees that a $K$-sparse signal $\psi$ which has $N$ elements can exactly be reconstructed from $M \geq C \mu^2 ({\bf \Phi},{\bf \Psi}) K \textnormal{ log (N)}$ measurements with a very high probability. In here $C$ is a positive constant and $\mu^2 ({\bf \Phi},{\bf \Psi})$ is the mutual coherence between the sensing ${\bf \Phi}$ and transform bases ${\bf \Psi}$ \cite{Candes, Candes2006}. Taking $M$ projections randomly and using the sensing matrix ${\bf \Phi}$ the sampled signal can be written as $g={\bf \Phi} \psi$. Therefore the CS problem can be rewritten as
\begin{equation}
\textnormal{ min} \left\| \widehat{ \psi} \right\|_{l_1} \ \ \ \ \textnormal{under constraint} \ \ \ \ g={\bf \Phi} {\bf \Psi} \widehat{ \psi}
\label{eq07}
\end{equation}
where $\left\| \widehat{ \psi} \right\|_{l_1}=\sum_i \left| \widehat{ \psi}_i\right|$. So that, among all signals that satisfy the given constraints mentioned above, the ${l_1}$ minimization solution of the problem is $\psi_{{}_{CS}} ={\bf \Psi} \widehat{ \psi}$. $l_1$ minimization is only one of the techniques that can be used for finding the solution of this optimization problem and other methods exist \cite{Candes, Candes2006}. Details of the CS can be seen in \cite{Candes, Candes2006}. In the current study we use the sparsity property of the 1D Radon transforms of the emerging conical 2D rogue wave spectra.
\subsection{Review of the Filtered Back Projection Algorithm}
In this section we sketch a very brief review of the FBP algorithm. The projections of a 2D function $\psi(x,y)$, which refers to the envelope of the wavefield or probability of finding an atomic particle at a specific $(x,y)$ at a given time in our study, can be computed using the Radon transform as
\begin{equation}
\widetilde{\psi}_R (r, \theta)=\int \int \psi(x,y) \delta(r-x \cos \theta-y \sin \theta)dxdy
\label{eq08}
\end{equation}
where $\theta$ is the projection angle defined from the $x$ axis. In a typical computerized tomography approach first these projections are obtained, then the full image is backprojected from these projections. However it is known that unfiltered tomographic data results in a high intensity blurring at the center of the image. In order to remove such an artifact, generally a filter is applied. Here we use a ramp filter applied in the Fourier domain as
\begin{equation}
\overline{\psi}(\rho,\theta)= F_r^{-1} \left| k_r \right| F_r \widetilde{\psi}_R (r, \theta)
\label{eq09}
\end{equation}
where $F_r$ and $F_r^{-1}$ show the forward and inverse Fourier transform operations, respectively and $\rho$ is the radial wavenumber parameter. However, other choices of the filter also exist. Then the image can be reconstructed from these projections by means of the back projection operation given as
\begin{equation}
\psi(x,y) = B \overline{\psi}(\rho,\theta)= \int_0^\pi \overline{\psi} (x \cos \theta+y \sin \theta) d\theta
\label{eq10}
\end{equation}
In a typical computed tomography approach, this integral is evaluated in a discrete fashion. The process summarized here is known as the FBP algorithm of the computed tomography. The reader is referred to \cite{Dudgeon}, for a comprehensive discussion of the FBP algorithm.
\subsection{Proposed Method}
In this paper we propose using the conical spectral features before 2D rogue waves becomes evident in time as an early detection mechanism. To efficiently measure such emerging spectra we propose a tomographic approach. We first construct the 1D Radon transforms of the emerging conical 2D spectra of the wavefield using CS. This principle works because such projections are sparse signals, with nonzero entries are located around central wavenumber. Then, we construct 2D spectra from those projections using FBP. For the radially symmetric versions of the Peregrine and Akhmediev-Peregrine solitons we show that emerging conical spectral features of 2D rogue waves before they become evident in time can be acquired efficiently by the proposed method.
The tomographic method proposed in here does not necessarily have to be used with the same reconstruction techniques. For example, the CS can be utilized by random selection of the projection angles rather than equally spaced projection angles. Instead of using FBP, it is possible to use reconstruction techniques such as inverse Radon transform, Fourier domain reconstruction algorithm and ordered subsets expectation maximization techniques, just to name a few. All would have some advantages and disadvantages, but the underlying tomographic approach for the early detection of 2D rogue waves would be same in principle for all such techniques.
\section{Results and Discussion}
\subsection{Early Detection of the 2D Peregrine Soliton by the Proposed Method}
In this section we numerically test the proposed algorithm for radially symmetric 2D Peregrine soliton. In the first step we take random samples along a slice in the physical domain to obtain the emerging triangular 1D spectra at various times. Then by applying the $l_1$ minimization of the CS algorithm to those random samples acquired in the physical domain, we obtain the sparse triangular spectra. A result obtained this way is depicted in Fig.~\ref{fig1}. In Fig.~\ref{fig1}a, we show the 1D Peregrine soliton at times $t=0$ and $t=2$. Fig.~\ref{fig1}b we compare the triangular spectra of the Peregrine soliton at $t=0$ obtained by classical and compressive sampling. The normalized root-mean-square (nrms) difference between these two spectra depicted in Fig.~\ref{fig1}b are $1.56 \times 10^{-10}$. We repeat the same procedure at $t=2$ and compare the triangular spectra of the Peregrine soliton at $t=2$ obtained by classical and compressive sampling in Fig.~\ref{fig1}c, where the nrms difference between these two spectra is $7.91\times 10^{-04}$. Both of these results are obtained using $N=1024$ classical and $M=64$ compressive samples. Due to time reversal property of the phenomena studied in the frame of the NLSE, the results for $t=2$ is no different than results for the $t=-2$, thus they may be used for early detection purposes. Additionally, the detection of emerging triangular spectra can be performed starting around $t=-5$ and may be longer early detection times in the Kundu-Eckhaus equation regime \cite{BayPRE2}. We also observe that the CS is capable of constructing the triangular spectra with far fewer samples than $M=64$ when the rogue wave is at its peak at $t=0$. The use of the CS for the early detection of the 1D rogue waves is introduced and studied in \cite{Bay_arxEarlyDetectCS}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig1}
\end{center}
\caption{\small a) Peregrine soliton at $t=0$ and $t=2$ b) the Fourier spectrum of the Peregrine soliton at $t=0$ obtained by $N=1024$ classical and $M=64$ compressive samples c) the Fourier spectrum of the Peregrine soliton at $t=2$ obtained by $N=1024$ classical and $M=64$ compressive samples.}
\label{fig1}
\end{figure}
For the 2D tomographic approach proposed above the 1D Radon transforms, i.e. the projections, of the 2D wave surface should be obtained. We obtain those projections using the perpendiculars to the slices shown above where the necessary summations are done discretely. However this is not a must, 1D Radon transforms can directly be measured using compressive samples.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig2}
\end{center}
\caption{\small Radially symmetric Peregrine soliton in 2D domain at $t=0$.}
\label{fig2}
\end{figure}
In Fig.~\ref{fig2} the radially symmetric 2D version of the Peregrine soliton at $t=0$ and in Fig.~\ref{fig3} its conical spectra obtained by $N_x=N_y=1024$ classical samples are depicted. This conical spectra begins to develop around $t=-5$, thus it can be used for the early detection of the 2D radially symmetric Peregrine soliton.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig3}
\end{center}
\caption{\small The Fourier spectrum of the radially symmetric Peregrine soliton at $t=0$ obtained using $N_x=N_y=1024$ classical samples.}
\label{fig3}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig4}
\end{center}
\caption{\small The Fourier spectrum of the radially symmetric Peregrine soliton at $t=0$ obtained using $M=64$ compressive samples and FBP algorithm with angles of $0:1:179$ degrees}
\label{fig4}
\end{figure}
In Fig.~\ref{fig4}, we present the same rogue wave spectrum obtained by the tomographic approach proposed above, where 1D Radon transforms are computed using $M=64$ compressive samples along each lines equally spaced with angles of $0:1:179$ degrees and then the FBP algorithm is used for the reconstruction of the 2D spectrum from those projections. A comparison of the results depicted in Fig.~\ref{fig3} and in Fig.~\ref{fig4} indicate that the proposed tomographic approach can successfully capture the spectral features of the 2D rogue waves, thus enables their early detection.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig5}
\end{center}
\caption{\small 3D plot of the Fourier spectrum of the radially symmetric Peregrine soliton at $t=0$ obtained using $M=64$ compressive samples and filtered backprojection algorithm with angles of $0:20:160$ degrees.}
\label{fig5}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig6}
\end{center}
\caption{\small Contour plot of the Fourier spectrum of the radially symmetric Peregrine soliton at $t=0$ obtained using $M=64$ compressive samples on slicing lines and filtered backprojection algorithm with angles of $0:20:160$ degrees and the projections.}
\label{fig6}
\end{figure}
In order to discuss the effects of using less projections in the tomographic approach for the early detection of the Peregrine soliton, we depict the spectra in 3D and in contour map format obtained using $9$ projection at angles of $0:20:160$ degrees in Fig.~\ref{fig5} and in Fig.~\ref{fig6}, respectively. As expected, as the number of projections decrease the capture of the conical spectral shape of the emerging rogue wave becomes harder. At central wavenumbers, the conical peak still appears and may be useful for early detection purposes, but it is surrounded by other spectral components which makes it harder to recognize if the emerging wave is a rogue wave. One possible technique to reduce the defects of small number of projections is to select projection angles randomly, which may lead to more accurate results since CS would perform better for a sparse signal when selections are random.
\subsection{Early Detection of the 2D Akhmediev-Peregrine Soliton by the Proposed Method}
Next we turn our attention to the radially symmetric Akhmediev-Peregrine soliton and assess the applicability of the proposed approach for its early detection.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=3.4in]{fig7}
\end{center}
\caption{\small a) Akhmediev-Peregrine soliton at $t=0$ and $t=2$ b) the Fourier spectrum of the Akhmediev-Peregrine soliton at $t=0$ obtained using $N=1024$ classical and $M=64$ compressive samples c) the Fourier spectrum of the Akhmediev-Peregrine soliton at $t=2$ obtained using $N=1024$ classical and $M=64$ compressive samples.}
\label{fig7}
\end{figure}
In Fig.~\ref{fig7}a, we show the 1D Akhmediev-Peregrine soliton at times $t=0$ and $t=2$. In Fig.~\ref{fig7}b we compare the triangular spectra of the Akhmediev-Peregrine soliton at $t=0$ obtained by classical and compressive sampling. The normalized root-mean-square (nrms) difference between these two spectra depicted in Fig.~\ref{fig7}b are $0.0016$. We again repeat the same procedure at $t=2$ and compare the triangular spectra of the Akhmediev-Peregrine soliton at $t=2$ obtained by classical and compressive sampling in Fig.~\ref{fig7}c, where the nrms difference between these two spectra is $0.0027$. Similar to the Peregrine soliton case, both of these results are obtained using $N=1024$ classical and $M=64$ compressive samples. We also observe that, similar to the Peregrine soliton case, the CS is capable of constructing the triangular spectra with far fewer samples than $M=64$ when the Akhmediev-Peregrine soliton is at its peak at $t=0$ \cite{Bay_arxEarlyDetectCS}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=3.4in]{fig8}
\end{center}
\caption{\small Radially symmetric Akhmediev-Peregrine soliton in 2D domain at $t=0$.}
\label{fig8}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=3.4in]{fig9}
\end{center}
\caption{\small The Fourier spectrum of the radially symmetric Akhmediev-Peregrine soliton at $t=0$ obtained using $N_x=N_y=1024$ classical samples.}
\label{fig9}
\end{figure}
In Fig.~\ref{fig8} the radially symmetric 2D version of the Akhmediev-Peregrine soliton at $t=0$ and in Fig.~\ref{fig9} its conical spectra obtained by $N_x=N_y=1024$ classical samples are depicted. This conical spectra begins to develop around $t=-5$, thus it can be used for the early detection of the 2D radially symmetric Akhmediev-Peregrine soliton, as in the case of the Peregrine soliton discussed above.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=3.4in]{fig10}
\end{center}
\caption{\small The Fourier spectrum of the radially symmetric Akhmediev-Peregrine soliton at $t=0$ obtained using $M=64$ compressive samples and FBP with angles of $0:1:179$ degrees}
\label{fig10}
\end{figure}
In Fig.~\ref{fig10}, we present the same Akhmediev-Peregrine rogue wave spectrum obtained by the tomographic approach proposed above, where 1D Radon transforms are computed using $M=64$ compressive samples along each lines equally spaced with angles of $0:1:179$ degrees and then the FBP algorithm is used for the reconstruction of the 2D spectrum from those projections. Again, a comparison of the results depicted in Fig.~\ref{fig9} and in Fig.~\ref{fig10} indicates that the proposed tomographic approach can successfully capture the spectral features of the 2D Akhmediev-Peregrine soliton, thus enables their early detection before they become evident in time using spectral data.
\section{Conclusion}
In this paper we have proposed an efficient method for the early detection of 2D rogue waves. We have showed that as for the early detection of the 1D rogue waves their emerging triangular spectra can be used; so the emerging 2D conical rogue wave spectra of the 2D rogue waves can be used for their early warning. We have proposed and numerically tested a method which can efficiently be used to detect 2D rogue wave emergence. In the proposed method we have constructed the 1D Radon transforms of the emerging conical 2D spectra of the wavefield using CS and then constructed 2D spectra from those projections using FBP. We have showed that the proposed approach can successfully and efficiently detect the single rogue wave emergence in 2D, with early warning times around the temporal width of the rogue wave peak, similar to 1D case. As a future work experimental verification of the proposed method would be necessary. It should also be tested for the analytical rogue waves solutions of NLSE type equations which are physically significant, as well as for the stochastic wavefields that are triggered by the modulation instability. Additionally, other options for the tomographic acquisition technique do exist. These include but are not limited to using CS with random projections instead of equally spaced projections and using other reconstruction algorithms such as the inverse Radon transform, Fourier domain reconstruction algorithm and the ordered subsets expectation maximization techniques instead of the FBP.
|
\section{Introduction} \label{sec:Intro}
Languages evolve as they pass from one mind to another. Immersed in a world of infinite variation, our cognitive architecture constrains what we can perceive, process, and produce. Cognitive constraints, such as learning biases, shape languages as they evolve and can help to explain the structure of language \citep{bever1970cognitive,slobin1973cognitive,newport1988constraints,newport2016statistical,Chater2008,christiansen2016now,Culbertson2012,Kirby2014iterated}. Early on, debate over the nature of these biases was polarized: Chomsky's nativism program explained linguistic structure as the product of a language-specific acquisition device \citep{Chomsky1957} while behaviorists claimed general-purpose learning mechanisms, such as reinforcement learning, could explain language acquisition \citep{skinner1957verbal}. Recent experimental research has found domain-general learning mechanisms underpin many aspects of language learning \citep{saffran2007domain}, such as the statistical learning involved in word segmentation by infants \citep{Saffran1996} and how memory constraints modulate learners' productions of probabilistic variation in language \citep{Chang2009}.
However, it is likely that a mixture of domain-general and domain-specific mechanisms are involved in language learning \citep[e.g.][]{pearl2009domain,culbertson2016simplicity}.
This paper offers a first attempt to quantify the relative contribution of domain-general and domain-specific learning mechanisms to linguistic regularization behavior.
Regularization is a well-documented process by which learners impose structure on data by reducing the amount of variation in that data. When language learners encounter linguistic elements in free variation, such as two realizations of a particular phoneme, two synonyms for one meaning, or two possible word orders for constructing a clause, they tend to reduce that free variation by either eliminating one of the variants, or conditioning their variant use on some aspect of the context (e.g. on the adjacent linguistic context).
Natural languages rarely exhibit free (i.e. unconditioned) variation \citep{givon1985function} and the regularization behavior of language learners is likely to be the cause.
Regularization has been documented extensively in natural language use and in the laboratory. In natural language, regularization occurs
in children's acquisition of language
\citep{Berko1958,Marcus1992,Singleton2004,Smith2007mam},
during the formation of creole languages from highly variable pidgin languages \citep{Bickerton1981,Sankoff1979,Degraff1999,Lumsden1999,Meyerhoff2000,Becker2003},
during the formation of new signed languages
\citep{Senghas1997,Senghas2000,Senghas2001},
and in historical trends of language change
\citep{Schilling1994,Lieberman2007,vanTrijp2013}.
In the laboratory, regularization has been studied in depth through artificial language learning experiments with children
\citep{HudsonKam2005,HudsonKam2009,Wonnacott2011a,culbertson2015harmonic}
and adults \citep{Wonnacott2005,HudsonKam2005,HudsonKam2009,Reali2009,Smith2010a,Perfors2012,Culbertson2012,feher2016structural,smith2017language}. Here we focus on regularization of lexical variation by adult learners in an artificial language learning paradigm. Future research should explore whether our results generalize to regularization by child learners.
Behavioral experiments offer special insight into the regularization process, because they allow researchers to present participants with controlled linguistic variation, precisely measure the way participants transform that variation, and test hypotheses about what causes participants to alter patterns of variation. For example, \citet{HudsonKam2009} investigated the regularization of pseudo-determiners in an artificial language learning experiment. In Experiment 1, adult participants were trained on a language that consisted of several verbs, several nouns (divided into 2 noun classes), 2 \emph{main} determiners (one for each noun class), and zero to 16 \emph{noise} determiners (which could occur with any noun). In the training language, each noun occurred with its main determiner on 60\% of exposures; the remaining exposures were equally divided across the noise determiners. In the testing phase, participants described scenes using the language they had learned. When participants encountered only two noise determiners during training, they regularized slightly by producing the main determiners with 70\% of the nouns, rather than the 60\% they observed in the training language. Regularization increased with the number of noise determiners, reaching its highest level with 16 noise determiners, where the main determiners were produced with nearly 90\% of the nouns. In Experiment 2, Hudson Kam \& Newport showed that adult participants regularize the same artificial language less when the noise determiners are conditioned on particular nouns in a more predictable and consistent way.
These results are consistent with Newport's Less-is-More hypothesis. Originally conceived as an explanation for why children regularize more than adults \citep{Newport1990}, it states that learners with limited memory capacity may regularize inconsistent input because they have difficulty storing and retrieving forms that are lower in frequency or used less consistently. Regularization behavior varies considerably between children and adults (see e.g. \citealp{HudsonKam2009}, Experiment 3). However, regularization due to memory limitations may also apply to adults, albeit to a lesser degree \citep{Chang2009}.
Overall, the Less-is-More hypothesis constitutes a domain-general account of linguistic regularization in terms of cognitive constraints on memory encoding and retrieval. If this hypothesis describes a truly domain-general effect, we should expect to see the same kind of regularization behavior in non-linguistic domains.
\citet{Gardner1957} conducted a frequency prediction experiment in which adult participants had to predict which of several lights would flash in any given trial. When participants observed two lights flashing at random in a 60:40 ratio (light A flashed 60\% of the time and light B flashed 40\% of the time), they \emph{probability matched} this ratio in their predictions, meaning that about 60\% of their guesses were that light A would flash next and about 40\% of their guesses were on light B. They also probability matched when observing a 70:30 ratio. However, when participants were trained on three lights (four ratios were tested: 70:15:15, 70:20:10, 60:20:20, and 60:30:10), they regularized by over-predicting the most frequent light and under-predicting the less frequent lights, which is similar to the behavior of Hudson Kam \& Newport's (2009) participants. In another experiment, \citet{Kareev1997} report an effect of individual differences in working memory capacity (as determined by a digit-span test) on participants' perception of the correlation of two probabilistic variables. Participants with lower memory capacity overestimated the most common variant, whereas participants with higher capacity did not. Similarly, \citet{Dougherty2003} show that participants with lower working memory were less likely to consider alternative choices in an eight-item prediction task and were also less likely to consider the low-frequency alternatives than participants with higher working memory. Each of these cases can be identified as regularization where the higher-frequency variants are over-represented in participants' behavior.
There is therefore strong evidence for the existence of domain-general drivers of regularization, but the extent to which they account for the level of regularity that we observe in language is not clear. This is because domain-specific learning mechanisms may play a role on their own, or interact with general mechanisms. For example, \citet{Perfors2012} presented seven carefully controlled manipulations of cognitive load during the encoding stage of an artificial language learning task and found no effect on regularization behavior. This suggests that the Less-is-More Hypothesis may apply more to retrieval than to storage, and that the effects of working memory found in the non-linguistic experiments of \citet{Kareev1997} and \citet{Dougherty2003} may not operate as strongly in language learning. Furthermore, \citet{Reali2009} show an effect of domain on regularization behavior: participants reduce variation when learning about words but increase variation when learning about coin flips. However, cognitive load was lower in the coin flipping condition (one coin was flipped, whereas 6 objects were named), so it is unclear whether the higher cognitive load or linguistic domain caused participants to regularize in the word learning task.
In the following, we present a two-by-two experimental design that manipulates cognitive load (following \citealp{HudsonKam2009} and \citealp{Gardner1957}) and task domain (directly comparing regularization in linguistic and non-linguistic domains). To manipulate cognitive load we vary the number of stimuli a learner must track concurrently. We manipulate task domain by manipulating the type of stimuli the learner must track: objects being named with words (linguistic domain) or marbles being drawn from containers (non-linguistic domain). Our method is closely based on the artificial language learning experiment in \citet{Reali2009} and our high load linguistic condition replicates their Experiment 1.
Although we compare regularization behavior in one particular linguistic task (word learning) to one particular non-linguistic task (marble drawing), any differences in regularization behavior revealed by this comparison constitute an existence proof for general and language-specific drivers of regularization behavior.
Little is known about how regularization behavior compares across different levels of language and the only systematic study of this to date (comparing morphology to word order) reports no global difference in regularization behavior across these two levels \citep{Saldana2017}.
Our two-by-two design can easily be extended to various linguistic tasks at different levels of language (e.g. phonology, morphology, and word order variation) and appropriately matched non-linguistic tasks (coin flipping, flashing light prediction, etc.) to determine the generalizability of the present results.
Based on the work reviewed above, we predict that regularization behavior will increase when cognitive load is raised. We also predict that regularization behavior will increase when the task is presented with linguistic stimuli. However, we have no clear prediction about the existence of an interaction between domain and cognitive load, or the relative amount of variation that will be removed from the data due to load or domain.
Knowing the relative contribution of domain-general and domain-specific biases to structure in language is important because it tells us how much we can ground our theories of language learning in general mechanisms of memory and statistical learning.
In order to address these questions, we need a principled measure of regularization that is comparable across different distributions of variation and stimuli domains. In Section \ref{sec:entropy}, we provide this measure by formalizing the definition of regularization as the reduction of entropy in a data set. Readers may skip this section if they are willing to accept the following statement: the amount of variation a participant regularizes is equal to the drop in entropy of their productions relative to their observations.
In Section 3, we present the experimental method and design.
In Section 4, we present the main result (both cognitive load and linguistic stimuli elicit regularization), followed by three supporting analyses that explore regularization behavior in greater depth.
In Section 5, we use our empirical data to investigate the evolution of regularity as learners' biases are repeatedly applied under a model of cultural transmission. This gives us a sense of how predictive known regularization biases can be for the level of regularity found in culturally-transmitted behaviors, such as languages.
\section{Defining and quantifying regularization} \label{sec:entropy}
In the existing literature, regularization is described as the elimination or reduction of free variation. Therefore, we will define regularization in terms of this lost variation and quantify it as the amount of variation that was lost from learners' productions when compared to the data the learners observed. The amount of variation in any data set can be quantified by the information-theoretic notion of entropy \citep[e.g.][]{Cover1991} and a growing number of studies are using entropy measures to analyze regularization behavior \citep[e.g.][]{Smith2010a, Perfors2012, fedzechkina2014communicative, Ferdinand2015, cuskley2015adoption, perfors2016adult, feher2016structural, smith2017language,Saldana2017,samara2017acquiring}.
The variation in a distribution of items, such as linguistic variants, can be quantified by \emph{Shannon entropy} \citep{Shannon1948}:
\begin{equation}
H(V) = -\sum_{v_i \in V} p(v_i) \log_2 p(v_i)
\end{equation}
\label{eq:ent}
\noindent
where $V$ is the set of linguistic variants in question,
$p(V)$ is the probability distribution over those variants,
and $p(v_i)$ is the probability of $i$th variant in that set.
For example, take the probability distribution over the 4 determiners used in the ``2 noise determiner" condition of \citet{HudsonKam2009}'s artificial language learning experiment: $p(V) = \{0.3, 0.3, 0.2, 0.2\}$. The Shannon entropy of this distribution is 1.97 bits.
Imagine a participant who was trained on this language and on testing produced the distribution $p(V') = \{0.7, 0.1, 0.1, 0.1\}$. The Shannon entropy of $p(V')$ is 1.36 bits and the change in variation is -0.61 bits. This means that 0.61 bits of variation among determiners was regularized (i.e. removed) by the participant. Or, more intuitively, $\frac{0.61}{1.97} \cdot 100 = 31\%$ of the variation in determiners was regularized by the participant.
Variation can also be lost when variants become conditioned on other linguistic variables or contexts.
For example, each determiner may have a conditional probability $p(v_i|c_j)$ of being produced with a particular noun class $c_j$, such that if one knows the class of the noun, one is better able to predict which determiner a speaker of that language will use with that noun. The variation in a distribution of items, after a conditioning variable is taken into account, is quantified by \emph{conditional entropy} \citep{Shannon1948}:
\begin{equation}
H(V|C) = -\sum_{c_j \in C} p(c_j) \sum_{v_i \in V} p(v_i|c_j) \log_2 p(v_i|c_j)
\label{eq:condent}
\end{equation}
\noindent
where $V$ is the set of linguistic variants and $C$ is the set of conditioning contexts.
Again, $p(V)$ is the probability distribution over variants,
$p(C)$ is the probability distribution over contexts,
$p(v_i|c_j)$ is the conditional probability of observing the $i$th variant in the $j$th context,
and $p(c_j)$ is the probability that the $j$th context occurs.
Given the format of this equation, we can see that the conditional entropy is the sum of the entropy of variants per context, weighted by the probability of each context.
Assume for a moment that the $p(V)$ distribution over determiners is not conditioned on each noun class, meaning that all determiners have the same probabilities regardless of the noun class they are used with, for example: $p(v_i|c_1)$ = \{0.3, 0.3, 0.2, 0.2\} and $p(v_i|c_2)$ = \{0.3, 0.3, 0.2, 0.2\}. Assume also that any noun has the following probabilities of being in noun class 1 or 2: $p(C) = \{0.6, 0.4\}$. Let us call this mapping A. The conditional entropy of mapping A is 1.97 bits, identical to the entropy of the determiners themselves, because the noun class carries no information about which determiner is used.
We can contrast this with another mapping, mapping B, where determiner use is conditioned on noun class such that $p(v_i|c_1)$ = \{0.5, 0.5, 0.0, 0.0\} and $p(v_i|c_2)$ = \{0.0, 0.0, 0.5, 0.5\}. Here, the first two determiners in the set are exclusively produced with noun class 1 and the third and fourth determiners are exclusively produced with noun class 2.
The conditional entropy of mapping B is 1.00 bit, while its entropy over determiners remains at 1.97 bits.
If a participant had been trained on a language with mapping A and produced a language with mapping B, then they would have regularized 0.97 bits, or $\frac{0.97}{1.97} \cdot 100 = 49\%$ of the variation in mapping A.
\begin{figure} [t]
\centering
\begin{tikzpicture}[fill=white]
\draw
[black,thick] (0,0) circle [radius=2.1]
[black,thick] (2.3,0) circle [radius=1.7]
(-1,2.3) node [text=black] {$H(V)$}
(3,1.9) node [text=black] {$H(C)$}
(-0.7,0) node [text=black] {$H(V|C)$}
(3,0) node [text=black] {$H(C|V)$}
(1.4,0) node [text=black] {$I(V;C)$}
(1,-3.5) node [text=black] {$H(V,C)$}
(-2,-3)--(4,-3)
(-2,-3)--(-2,-2.5)
(4,-3)--(4,-2.5);
\end{tikzpicture}
\caption{The relationship between entropy quantities in a mapping between linguistic variants ($V$) and their conditioning contexts ($C$).}
\label{fig:ent_venn}
\end{figure}
Based on these examples, it should be clear that $H(V)$ and $H(V|C)$ are two different kinds of variation that a language can have. $H(V)$ is about the total number of different linguistic forms there are and how often each one is used. $H(V|C)$ is about how often different linguistic elements occur together. It is important to note that $H(V)$ and $H(V|C)$, by themselves, do not fully describe the variation in a mapping between linguistic variants and contexts. Figure \ref{fig:ent_venn} shows the six quantities that are relevant to a complete description of the variation in a linguistic mapping. The largest quantity, $H(V,C)$, is the total amount of variation in the mapping and is equal to the area covered by the two overlapping circles. $H(V,C)$ is the joint entropy of the mapping:
\begin{equation}
H(V,C) = -\sum_{v_i \in V} \sum_{c_j \in C} p(v_i,c_j) \log_2 p(v_i,c_j)
\label{eq:joint}
\end{equation}
\noindent
where $p(v_i,c_j)$ is the joint probability of observing the $i$th variant and the $j$th context together. Looking at Figure \ref{fig:ent_venn}, it is possible to imagine how the joint entropy of the system can increase by moving the two circles away from one another. As the circles move apart, $V$ and $C$ carry less information about one another. This has the effect of increasing the two conditional entropy values and reducing the mutual information, $I(C;V)$, between $V$ and $C$. Mutual information is not a measure of variation, but one of structure: it measures how much uncertainty is reduced in V when C is known, $I(V;C) = H(V) - H(V|C)$, and how much uncertainty is reduced in C when V is known, $I(C;V) = H(C) - H(C|V)$. Note that $I(V;C) = I(C;V)$. The five entropy values, on the other hand, are all measures of variation. Although they each refer to different types of variation, they are related and do constrain one another. For example, $H(V|C)$ can never be larger than $H(V)$ and $H(C|V)$ can never be larger than $H(C)$. It is also important to note that $H(V|C) \neq H(C|V)$ (because $H(V)$ and $H(C)$ can take different values). Regularization can be quantified in terms of these five entropy values and be said to occur when one or more of these values decreases. \\
\indent {\bf Regularization} is the reduction or elimination \\
\indent of entropy in a data set. \\
We define regularization as any reduction to the space in Figure \ref{fig:ent_venn}. Regularization can occur by eliminating linguistic variants (reducing $H(V)$), eliminating conditioning contexts (reducing $H(C)$), or increasing the degree to which variants and contexts are conditioned on one another (reducing $H(V|C)$ and/or $H(C|V)$). Joint entropy always decreases when there is a net loss of variation. Mutual information, on the other hand, does not necessarily change when regularization occurs. In the following experiment, we construct a stimuli set in which lexical items are variants and the objects they refer to are contexts. In a matched non-linguistic stimuli set, marbles are variants and the containers they are drawn from are contexts. The experiment is designed such that $H(V)$, $H(V|C)$, and $H(V,C)$ will always change by the same number of bits when participants regularize and $H(C)$, $H(C|V)$, and $I(V;C)$ can not be changed by participants.
\section{Frequency learning experiment} \label{sec:experiment}
In this experiment we manipulate cognitive load and task domain, allowing us to quantify the amount of variation participants regularize due to each source. Participants observe an input mapping among stimuli and then produce behavior from which an output mapping is extracted. Finally, they estimate the frequencies of the input stimuli and these estimates are compared to their output behavior.
\subsection{Participants}
573 participants were recruited via Amazon's Mechanical Turk crowdsourcing platform and completed our experiment online. Informed consent was obtained for experimentation. Participant location was restricted to the USA and verified by a post-hoc check of participant IP address location. 61 participants were excluded on the basis of the following criteria: failing an Ishihara color vision test (15), self-reporting the use of a pen or pencil during the task in an exit questionnaire (10), not reporting their sex or age (6), self-reporting an age below 18 (1), or having previously participated in this or any of our related experiments, as determined by their user ID with MTurk (26). More participants were recruited than necessary with the expectation that some would be excluded by these criteria. Once the predetermined number of participants per condition was met, the last participants were excluded (3). All participants (included and excluded) received the full monetary reward for participating in this task, which was 0.10 USD in the one-item conditions (\emph{marbles1} and \emph{words1}) and 0.60 USD in the six-item conditions (\emph{marbles6} and \emph{words6}).\footnote{Data collection began in 2012, when this was standard reimbursement for participants recruited through MTurk. In current practice, standard reimbursement is US federal minimum wage.}
The average time taken to complete the one-item conditions was 3 minutes and 50 seconds, with a standard deviation of 1 minute and 27 seconds. Average time to complete the six-item conditions was 11 minutes and 32 seconds, with a standard deviation of 2 minutes and 6 seconds.
Of the final 512 participants, 274 reported female, 238 reported male, and the mean age was 33.7 years (min = 18, max = 72) with a standard deviation of 11.3 years.
\subsection{Materials and Stimuli}
The experiment was coded up as a Java applet that ran in the participant's web browser in a 600x800-pixel field. Photographs of 6 different containers (a bucket, bowl, jar, basket, box, and pouch) and computer-generated images of marbles in 12 different colors (blue, orange, red, teal, pink, olive, lime, purple, black, yellow, grey, and brown) served as non-linguistic stimuli. Photographs of 6 different novel objects (resembling mechanical gadgets) and 12 different nonsense words (\emph{buv}, \emph{kal}, \emph{dap}, \emph{mig}, \emph{pon}, \emph{fud}, \emph{vit}, \emph{lem}, \emph{seb}, \emph{nuk}, \emph{gos}, \emph{tef}) served as linguistic stimuli. Stimuli were chosen to have similar visual complexity across domain (determined by gzip complexity and area of stimuli).
Marbles and words were organized into fixed pairs that maximized distinctiveness between the stimuli in the pair. The stimuli lists above appear in order of these pairings (blue and orange were paired, \emph{buv} and \emph{kal} were paired, etc.). Marble colors were paired to differ in hue and brightness. Within-pair hue differences were greater than $120^{\circ}$ (i.e. chosen from approximately opposite sides of the color wheel) and within-pair brightness differences were greater than 20\%. Words were paired to be contrastive. Within-pair words utilized different letters and vowels and within-pair consonants differed by place of articulation. These stimuli are closely based on the word stimuli used in \citet{Reali2009} and selected to not look or sound like existing words when pronounced by an American English speaker. Words were presented visually and were not accompanied by auditory stimuli.
\begin{figure*} [t] \centering
\centering
\includegraphics[width=180mm]{new_schema2.pdf}
\put(-512,246){Low load, observation phase}
\put(-248,246){Low load, production phase}
\put(-512,113){High load, observation phase}
\put(-248,113){High load, production phase}
\footnotesize
\put(-492,140){trial 1}
\put(-386,154){trial 2}
\put(-277,181){10} \put(-282,172){trials}
\put(-228,140){trial 1}
\put(-70,162){trial 2}
\put(-12,181){10} \put(-17,172){trials}
\put(-492,6){trial 1}
\put(-386,20){trial 2}
\put(-277,47){60} \put(-282,38){trials}
\put(-228,6){trial 1}
\put(-70,27){trial 2}
\put(-12,47){60} \put(-17,38){trials}
\footnotesize
\put(-491,152){1 sec}
\put(-438,159){2 sec}
\put(-385,166){1 sec}
\put(-333,173){2 sec}
\put(-228,152){select}
\put(-178,159){confirm}
\put(-121,167){2 sec}
\put(-73,174){confirm}
\put(-491,18){1 sec}
\put(-438,25){2 sec}
\put(-385,32){1 sec}
\put(-333,39){2 sec}
\put(-228,18){select}
\put(-178,25){confirm}
\put(-121,32){2 sec}
\put(-73,39){confirm}
\caption{Schema of the experiment's observation and production phases. \emph{Top:} Low cognitive load condition. \emph{Bottom:} High cognitive load condition.
Examples shown are the linguistic condition. In the non-linguistic condition, containers are shown in place of the object, and marbles are shown in place of the words.}
\label{fig:schema}
\end{figure*}
\subsection{Conditions and Design} \label{Exp2_conditions}
We use a two-by-two design to investigate the effects of domain and cognitive load in four experimental conditions: \\
\noindent
1) \emph{Non-linguistic single frequency learning (marbles1)} \\
Participants observed two marble colors being drawn from one container at a particular ratio (for example, 5 blue marbles and 5 orange marbles displayed in random order). Participants were then asked to demonstrate what another several draws from the same container are likely to look like. They were not asked to predict specific future draws and thus no feedback was given. Participants observed 10 marble draws and produced 10 marble draws.
Each participant observed a set of draws in one of six possible ratios: 5:5, 6:4, 7:3, 8:2, 9:1, and 10:0. These constitute six \emph{input ratio conditions}. We will refer to the ratio that a participant observed as the \emph{input ratio} and the ratio that the participant produced as the \emph{output ratio}. There were 32 participants in each input ratio condition, totaling 192 participants in \emph{marbles1}. Container stimuli were randomized across participants: each participant saw one of the six containers. Equal numbers of participants saw each container. Marble pairs were also randomized across participants: each participant saw one of the six marble pairs. Equal numbers of participants saw each marble pair. One variant in each pair was randomly assigned to be the majority variant (i.e. have the frequency of 6, 7, 8, 9, or 10). The full details of the observation and production regimes can be found in Section \ref{sec:procedure} and Figure \ref{fig:schema}. \\
\noindent
2) \emph{Non-linguistic multiple frequency learning (marbles6)} \\
This condition is similar to the \emph{marbles1} condition, with the difference that participants observed and produced 10 draws each from 6 different containers, where each container differed in the ratio of the two marble colors. Containers, marble pairs, and input ratios were randomly assigned to one another, without replacement, and these assignments were randomized between participants. Each participant saw all six of the containers, all six of the marble pairs, and all six of the input ratios (the same input ratios as were used in the \emph{marbles1} condition: 5:5, 6:4, 7:3, 8:2, 9:1 and 10:0). There were 64 participants in this condition, yielding data for 384 (64x6) input ratios. \\
\noindent
3) \emph{Linguistic single frequency learning (words1)} \\
This condition is similar to the \emph{marbles1} condition, differing only by the use of linguistic stimuli (objects and words) instead of the non-linguistic stimuli (containers and marbles) and minimal adaptation of the instructions to the linguistic domain. Participants observed one object being named with two words at a particular ratio (for example, \emph{buv} 5 times and \emph{kal} 5 times, in random order) and were then asked to name the object like they had observed it being named.
They were not asked to predict specific future namings and thus no feedback was given. Participants observed 10 namings and produced 10 namings. Each of the 6 possible input ratios (same ratios as used in \emph{marbles1}) was observed by 32 participants, totaling 192 participants. \\
\noindent
4) \emph{Linguistic multiple frequency learning (words6)} \\
This condition is similar to the \emph{marbles6} condition, again differing only by the use of linguistic stimuli and minimal adaptation of the instructions to the linguistic domain. This condition constitutes a replication of the word learning experiment in \citet{Reali2009}, but with different object stimuli, modified word stimuli, and participants who completed the experiment online rather than in the laboratory. There were 64 participants in this condition, yielding data for 384 (64x6) input ratios.
\subsection{Procedure} \label{sec:procedure}
\begin{figure} [h]
\includegraphics[width=89mm]{sliders.pdf}
\caption{Screen shot of the sliders page in the high cognitive load linguistic condition, showing three answers selected. Participants could change their answers up until ``Save Answers" was clicked. ``Back" took participants back to the question and instruction about the sliders. In the low load condition, only one slider was shown.}
\label{fig:sliders}
\end{figure}
The experiment consisted of an observation phase and a production phase. Figure \ref{fig:schema} shows the structure and timing of the trials. Participants were not told how many observation or production trials there would be. In the observation phase, marble/word stimuli were presented in random order. In the high load conditions, the containers/objects were presented in random order. In each production trial, the left-right location of the two marbles/words was randomized. When the participant moused over an answer, a 100x100 pixel box was displayed around the choice. When clicked, the box remained and an OK button appeared equidistant between the two choices. Participants could change their answer and clicked OK to confirm their final response. Their choice was shown over the container/object and the next trial began. The OK button served to re-center the participant's cursor between trials.
After the production phase, participants were asked to estimate the generating ratio that underlies the input ratio they saw. This was accomplished by asking them how many marbles of each color were in each container, or how often each word is said for each object in the artificial language.
Participants provided their response with a discrete slider over 11 options of relative percentages (Figure \ref{fig:sliders}).
\subsection{Entropy of the training stimuli set} \label{sec:stim_entropy} \label{sec:stimset_ent}
\begin{table}[t]
\centering
\scalebox{0.8}{
\begin{tabular}{ | l | c | c | c | c | c | c | c | c | c | c | c | c |}
\hline
& $v_1$ & $v_2$ & $v_3$ & $v_4$ & $v_5$ & $v_6$ & $v_7$ & $v_8$ & $v_9$ & $v_{10}$ & $v_{11}$ & $v_{12}$ \\
\hline
$c_1$ & 5 & 5 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\hline
$c_2$ & 0 & 0 & 6 & 4 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\hline
$c_3$ & 0 & 0 & 0 & 0 & 7 & 3 & 0 & 0 & 0 & 0 & 0 & 0 \\
\hline
$c_4$ & 0 & 0 & 0 & 0 & 0 & 0 & 8 & 2 & 0 & 0 & 0 & 0 \\
\hline
$c_5$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 9 & 1 & 0 & 0 \\
\hline
$c_6$ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 10 & 0 \\
\hline
\end{tabular}}
\caption{Co-occurance frequencies among the twelve variants and six contexts in the experimental stimuli set. Each cell gives the number of times that the participant observed variant{$_i$} along with context{$_j$}.}
\label{tab:freqs}
\end{table}
Each participant observes a stimuli set that is composed of co-occurrances between marbles and containers or words and objects. For the purpose of quantifying the variation in the stimuli sets, we consider the marbles and words to be variants and consider the containers and objects to be contexts. Table \ref{tab:freqs} shows the co-occurrance frequencies between contexts and variants. In the high cognitive load conditions, this table describes the complete stimuli set that each participant was trained on in the observation phase. In the low cognitive load conditions, each participant was trained on only one row from the this table. Figure \ref{fig:stim_venn} shows the entropy values associated with Table \ref{tab:freqs} and describes the population-level variation in stimuli. These values are the same across conditions, allowing the direct comparison of mean change in entropy between conditions.
It is important to note that the experimental design prevents participants from changing $H(C)$ because contexts are presented the same number of times in the observation and production phases. $H(C|V)$ cannot be changed either because the only production options are the two variants that were shown per context in the observation phase. If participants regularize, $H(V)$, $H(V|C)$, and $H(V,C)$ will drop by the same number of bits.
The entropy of the stimuli that one participant observes in the high cognitive load condition is identical to Figure \ref{fig:stim_venn}. However, the entropy of stimuli in the low cognitive load condition is lower and varies by the input ratio observed: in condition 5:5, 6:4, 7:3, 8:2, 9:1, and 10:0, $H(V) = H(V|C) = H(V,C) = I(V;C) = $ 1 bit, 0.97 bits, 0.88 bits, 0.72 bits, 0.47 bits, and 0 bits, respectively, and $H(C) = H(C|V) = 0$.
\begin{figure} [t]
\centering
\begin{tikzpicture}[fill=white]
\draw
[black,thick] (0,0) circle [radius=2.1]
[black,thick] (0.5,-.4) circle [radius=1.3]
(-1,2.4) node [text=black] {$H(V) = 3.26$}
(3,1.8) node [text=black] {$H(C) = 2.58$}
(0,1.3) node [text=black] {$H(V|C) = 0.67$}
(3,-1.5) node [text=black] {$H(C|V) = 0$}
(0.5,-0.5) node [text=black] {$I(V;C) = 2.58$}
(0,-3.5) node [text=black] {$H(V,C) = 3.26$}
(-2,-3)--(2,-3)
(-2,-3)--(-2,-2.5)
(2,-3)--(2,-2.5);
\draw [<-,thick] (1.5,0.5)--(2.3,1.5);
\end{tikzpicture}
\caption{Entropy of the training stimuli (in bits). In the linguistic condition, $V$ is the distribution over words and $C$ is the distribution over objects. In the non-linguistic condition, $V$ is the distribution over marbles and $C$ is the distribution over containers. Refer back to Section \ref{sec:entropy} for the definition of each quantity. The experiment is designed so participants can change the size of the outer circle only.}
\label{fig:stim_venn}
\end{figure}
\begin{figure*} [t] \centering
\includegraphics[width=140mm]{6panes.pdf}
\caption{Each row shows the results of one experimental condition. Each column corresponds to one of the six input ratios, ranging from 5:5 (left) to 10:0 (right).
Each pane contains the distribution of output ratios that participants produced in response to one input ratio.
Output ratios are displayed on the x-axis as the number of times a participant produced variant \emph{x} from the input ratio \emph{x:y}, where variant \emph{x} corresponds to whatever marble/word was in the majority during the observation phase. (In the 5:5 input ratio a random marble/word was coded as variant \emph{x}.) All input ratios are indicated by a dashed line.}
\label{fig:2x2_6panels}
\end{figure*}
\section{Results} \label{sec:results}
First, we describe participant behavior and present the main result: cognitive load and linguistic stimuli both elicit regularization behavior. Next, three supporting analyses explore regularization behavior in greater depth: Section \ref{sec:encoding} shows that participants' regularization behavior was due to production biases rather than an encoding bias, Section \ref{sec:individual} analyzes individual differences in regularization during this experiment, and Section \ref{sec:primacy} shows primacy effects help explain why some individuals regularized with the minority variant, rather than the majority variant.
\subsection{Regularization behavior profiles}
Before analyzing the data in terms of its entropy, we first visually inspect how participants changed each input ratio. In Figure \ref{fig:2x2_6panels}, each panel shows the distribution of ratios that participants produced in response to each input ratio they observed, per experimental condition.
The first row (\emph{marbles1}) shows clear probability matching behavior, where both the mean and mode of participant responses are near the input ratio. Participants in this condition tended to successfully reproduce their input ratio, with a small amount of error. The second row (\emph{marbles6}) shows clear regularization behavior. Participants in this condition have moved distributional mass away from the input ratio and toward the maximally regular ratios, 0:10 and 10:0. Responses to the 5:5 input ratio seem to be a combination of probability matching behavior (13 participants also produced a 5:5 ratio) and regularization behavior (15 participants produced maximally regular ratios).
The third row (\emph{words1}) shows a mixture of probability matching and regularization behavior for all input ratios. Roughly half of the participants appear to have probability matched with error rates similar to \emph{marbles1}, and roughly half of the participants appear to have regularized at levels comparable to \emph{marbles6}. In the 10:0 input condition, none of the participants choose the unseen word on any production trial. The fourth row (\emph{words6}) shows a similar regularization profile to \emph{marbles6}, but with a more extreme movement of distributional mass to the edges, such that the majority of participants produced maximally regular ratios. This condition constitutes a successful replication of the first experiment reported in \citet{Reali2009}.
\subsection{Regularization per condition} \label{sec:netreg}
In this section, we report the differences in regularization behavior within and between the four experimental conditions. We do this by calculating the change in Shannon entropy for each pair of input-output ratios obtained from participants. For example, if a participant observes a 5:5 ratio of orange and blue marbles for the jar, and then produces a 6:4 ratio of orange and blue marbles for the jar, the Shannon entropy for that pair of input-output ratios changes by $-0.12$ bits.\footnote{From here onward, whenever we refer to the ``entropy of a ratio" we mean the Shannon entropy of the two variants in ratio $x$:$y$, where the probability distribution over the variants is $p(V) = \{\frac{x}{10},\frac{y}{10}\} $.} Figure \ref{fig:4bars} shows the mean change in entropy for all input-output ratio pairs per condition. Negative values mean participants made ratios more regular on average.
To assess the significance of differences in regularization within and between conditions, a linear mixed effects regression analysis was performed using R \citep{R2013} and \emph{lme4} \citep{Bates2013}. The dependent variable was the change in entropy of the input-output ratios. Experimental condition was the independent variable. Participant was entered as a random effect (with random intercepts). No obvious deviations from normality or homoscedasticity were apparent in the residual plots.
Within-condition changes were assessed by re-leveling the model to obtain the intercept value for each condition. The intercept equals the condition's mean change in entropy and the regression analysis provides a t-statistic to evaluate whether this mean is significantly different from zero. Three of the four experimental conditions elicited a significant amount of regularization behavior (Figure \ref{fig:4bars}). Participants regularized an average of 0.17 bits in \emph{marbles6} ($S.E. = 0.03, t(1152) = -5.53, p < .001$), 0.19 bits in \emph{words1} ($S.E. = 0.03, t(1152) = -6.52, p < .001$), and 0.36 bits in \emph{words6} ($S.E. = 0.03, t(1152) = -11.34, p < .001$). In \emph{marbles1}, the mean loss of 0.01 bits was not significantly different from zero, which indicates that participants are probability matching in this condition ($S.E. = 0.03, t(1152) = -0.35, p = 0.73$). Overall, participants regularized 26\%, 28\%, and 53\% of the conditional entropy in \emph{marbles6}, \emph{words1}, and \emph{words6}, respectively.
Pairwise comparison of regularization between conditions is also obtained from this re-leveled model. All pairwise comparisons showed a significant difference in regularization behavior at the $p < .001$ level, except for that between \emph{words1} and \emph{marbles6} ($S.E. = 0.04, t(1152) = 0.34, p = 0.73$).
\begin{figure} [t]
\includegraphics[width=90mm]{4bars_entropy.pdf}
\put(-121.5,166){***}
\put(-77,166){***}
\put(-32.5,166){***}
\caption{Entropy drops when learners regularize.
Each bar shows the average change in Shannon entropy over all pairs of input-output ratios, per condition. Stars indicate significant difference from zero. Error bars indicate the 95\% confidence intervals computed with the bootstrap percentile method \citep{Efron1979}. A significant drop in entropy means that participants regularized in that condition. Non-significant differences from zero are obtained when participants probability match. The lower and upper bounds on mean entropy change for this experiment are $-0.67$ and $+0.33$ bits.}
\label{fig:4bars}
\end{figure}
\subsection{Domain vs. cognitive load} \label{sec:manips_lmer}
Effects of the experimental manipulations were assessed by constructing a full linear mixed effects model with three independent variables (i.e. fixed effects) and their interaction: domain, cognitive load, and entropy of the input ratio. The dependent variable was the change in entropy of the input-output ratios. Participant was entered as a random effect (with random intercepts). The significance of each fixed effect was determined by likelihood ratio tests, on the full model (described above) against a reduced model which omits the effect in question. There was a significant effect of domain ($\chi^2(4) = 46.048, p < .001$), cognitive load ($\chi^2(4) = 105.07, p < .001$), and input ratio ($\chi^2(4) = 520.23, p < .001$). Interactions between fixed effects were also determined by likelihood ratio tests by comparing a reduced model (which omits all interactions) to one which includes the interaction of interest. Two interactions were found to be significant: cognitive load and input ratio ($\chi^2(1) = 74.695, p < .001$) and domain and input ratio ($\chi^2(1) = 4.4462, p = 0.03$). The interaction between domain and cognitive load was not significant ($\chi^2(1) = 0.0059, p = 0.94$).
\begin{figure} [t]
\includegraphics[width=90mm]{4bars_freq.pdf}
\put(-166,166){***}
\put(-121.5,166){***}
\caption{Raw changes in frequency fail to capture regularization behavior. Each bar shows the average difference between the number of times participants observed the majority variant in the training set and the number of times they produced that variant in the testing phase. Error bars indicate the 95\% confidence intervals computed with the bootstrap percentile method \citep{Efron1979}. Values significantly higher than zero indicate a population-level trend in over-producing the majority variant. Values significantly lower than zero indicate a population-level trend in over-producing the minority.
}
\label{fig:maj_4bars}
\end{figure}
Therefore, the best-fit model contained an interaction between domain and input ratio, an interaction between cognitive load and input ratio, but only an additive relationship between domain and cognitive load (loglikelihood $= -278.71$). A summary of the best-fit model is included in the Appendix, Table \ref{tab:bestfit1}. The effect of input ratio on entropy change is due to different amounts of regularization being possible under each input ratio (the maximum drop in entropy achievable under the 5:5 through 0:10 ratios are 1, 0.97, 0.88, 0.72, 0.47, and 0 bits, respectively). As input entropy increases from 0 to 1 bits, output entropy changes by $-0.14$ bits.
This means that participants regularize more when the entropy of the input ratio increases from 0 bits (the 10:0 ratio) to 1 bit (the 5:5 ratio). The interactions mean that the effect of input entropy on output entropy is greater by $-0.1$ bits when linguistic stimuli are used and greater by $-0.5$ bits when cognitive load is high. The additive relationship suggests that domain and cognitive load are independent drivers of regularization behavior.
\subsection{Frequency-based analysis of regularization} \label{sec:freq_analysis}
In much of the linguistic regularization literature to date, regularization is measured in terms of stimulus frequency, rather than entropy. In this section, we repeat the analyses from Section \ref{sec:netreg} and \ref{sec:manips_lmer} with a different dependent variable, \emph{change in frequency of the majority variant} \citep[as in, e.g.][]{HudsonKam2005, Reali2009}, to illustrate the difference between these two approaches.
Figure \ref{fig:maj_4bars} shows the mean change in frequency of the majority variant ($x$ from input ratio $x$:$y$). For example, if a participant produces a 7:3 ratio in response to a 9:1 input ratio, there is a -0.2 change in majority variant frequency for that pair of input-output ratios. In the 5:5 input condition, a random variant was encoded as the ``majority" variant. Positive changes mean participants over-produced the majority variant and negative changes mean participants over-produced the minority variant. In Figure \ref{fig:maj_4bars} we see that none of the conditions elicit over-production of the majority variant on average, despite the fact that participants in \emph{marbles6}, \emph{words1}, and \emph{words6} are clearly regularizing input ratios (compare to Figure \ref{fig:4bars}).
Applying the analysis in Section \ref{sec:netreg} to the change in majority variant frequency, we find that neither linguistic condition shows a significant change in majority variant frequency (\emph{words1}: $S.E. = 0.02, t(1152) = 0.22, p = 0.82$; \emph{words6}: $S.E. = 0.01, t(1152) = -0.65, p = 0.51$). However, the frequency-based analysis does reveal something that the entropy-based analysis was unable to capture: a significant over-production of the \emph{minority} variant in the marble-drawing domain, \emph{marbles1} ($S.E. = 0.02, t(1152) = -2.882$, $p=.004$) and \emph{marbles6} ($S.E. = 0.01, t(1152) = -3.269$, $p=.001$).
To determine the effects of the experimental manipulations, we apply the analysis in Section \ref{sec:manips_lmer} to the change in majority variant frequency (and we change the fixed effect \emph{entropy of the input ratio} to \emph{input frequency of the majority variant} in order to match the dependent variable). We find a significant effect of domain ($\chi^2(4) = 16.391, p = 0.003$) and input frequency ($\chi^2(4) = 14.634, p = 0.006$) on change in majority variant frequency, but no significant effect of cognitive load ($\chi^2(4) = 3.0755, p = 0.55$). We also find a significant interaction between domain and input frequency ($\chi^2(4) = 6.7741, p = 0.009$).
Therefore, the best-fit model contains an effect of domain, input frequency, and an interaction between domain and input frequency (loglikelihood $= -77.0$, see Appendix Table \ref{tab:bestfit_freq}).
In summary, the frequency analysis fails to capture the effect of cognitive load on regularization behavior and fails to capture the fact that participants are eliminating variation in the linguistic domain. The reason mean change in frequency is not different than zero in the linguistic domain is because participants sometimes regularized with the majority variant and other times regularized with the minority variant, in a way that tends to cause frequency changes to average out to zero. However, as is clear from the raw data, it would be incorrect to conclude that participants are probability matching in the linguistic domain.
\begin{figure} [t]
\includegraphics[width=86mm]{estimates.pdf}
\put(-164,176){*}
\put(-149.5,116){*}
\put(-110,82){***}
\put(-65.5,81){***}
\put(-21.5,55){***}
\caption{Production bias, not encoding bias, drives regularization.
\emph{Dark grey:} Average difference in regularity between the input ratios participants actually observed and their estimates of the underlying ratio that generated the input ratio. A significant increase in entropy means that participants estimated the underlying ratio to be more variable than the input ratio, and a significant decrease means they estimated it to be more regular. \emph{Light grey:} Average difference between production ratio regularity and estimated ratio regularity. Error bars indicate the 95\% confidence intervals computed with the bootstrap percentile method \citep{Efron1979}.}
\label{fig:4bars_Q1}
\end{figure}
\subsection{Regularization during encoding} \label{sec:encoding}
As discussed in the Introduction, regularization behavior is often explained as a result of general cognitive limitations on memory encoding and/or retrieval.
The high cognitive load manipulation in this experiment affected both the observation and production phases because both phases consisted of 60 interleaved trials.
Therefore, the regularization behavior we observed could be due to encoding multiple frequencies under load (during the observation phase) and/or retrieving frequencies under load (during the production phase). Furthermore, it is possible that linguistic domain may have a specific effect on the encoding of frequency information.
To determine whether encoding errors contribute to participants' regularization behavior in this experiment, we asked participants to estimate (using a slider) the underlying ratio that generated the marble draws or naming events they observed, per container or object (see Section \ref{sec:procedure}, last paragraph).
If participants' estimates are not significantly different from the ratios they observed, then we can assume frequency encoding was unbiased. This result would point to a production-side driver of regularization.
Figure \ref{fig:4bars_Q1} (dark grey bars) shows the average change in entropy between participants' estimates and the actual input ratios they observed. The same linear mixed effects regression analysis described in Section \ref{sec:netreg} was applied to this data, using the change between input and estimate entropy as the dependent variable. Only one condition, \emph{marbles1}, elicited a significant difference (of 0.05 bits) between the input ratios and estimates ($S.E. = 0.02, t(1152) = 2.29, p = 0.02$). In this condition, participants estimated the generating ratio to be significantly \emph{more} variable than the ratio they had observed, indicating a slight encoding bias toward variability. None of the conditions show any bias toward regularity in participants' estimates.
Effects of the experimental manipulations were assessed by the same procedure described in Section \ref{sec:manips_lmer}, using change between input and estimate entropy as the dependent variable. The best-fit model contained a significant effect of domain ($\chi^2(4) = 11.735, p = 0.02$), cognitive load ($\chi^2(4) = 34.916, p < .001$), and input ratio ($\chi^2(4) = 562.04, p < .001$) (loglikelihood = $-72.558$, see Appendix Table \ref{tab:bestfit_encode}). One interaction was found to be a significant predictor of participants' estimates: cognitive load and input ratio ($\chi^2(1) = 27.916, p < .001$). Interactions between domain and input ratio ($\chi^2(1) = 0.7554, p = 0.38$) and domain and cognitive load ($\chi^2(1) = 0.6741, p = 0.41$) were not significant.
Although the estimate data shows no bias toward regularity, the same factors that affected regularization behavior (cognitive load, domain, and input ratio) also affect participants' estimates.
Additionally, we find that the cognitive load manipulation resulted in noisier estimates ($F = 56.487, p < .001$, with Levene's test for homogeneity of variance), whereas the domain manipulation did not ($F = 0.4416, p = 0.51$). This suggests that the high load condition was indeed more difficult than the low load condition and that the two domains were well-matched in terms of difficulty and stimuli complexity.
Figure \ref{fig:4bars_Q1} (light grey bars) shows the difference in entropy between the ratio participants produced and their estimate of that ratio, i.e. the extent to which their productions were more regular than their own estimate of their input data. The same linear mixed effects regression analysis described in Section \ref{sec:netreg} was applied to this data, using the difference in entropy between the produced and estimated ratios as the dependent variable. In all conditions, production ratios are significantly more regular than the estimates participants made (\emph{marbles1}: $S.E. = 0.03, t(1152) = -2.55, p = 0.01$; \emph{marbles6}: $S.E. = 0.03$,
\noindent
$t(1152) = -6.97, p < .001$; \emph{words1}: $S.E. = 0.03, t(1152) = -8.20, p < .001$; \emph{words6}: $S.E. = 0.03, t(1152) = -10.69$,
\noindent\
$p < .001$).
This means that regularization occurs during the production phase and is likely to be involved in the retrieval and use of frequency information. Interestingly, production-side regularization occurs in all four conditions, even in \emph{marbles1} where participants probability matched their productions to their inputs (effectively ``correcting" the variability bias in their estimates). This suggests that regularity is broadly associated with frequency production behavior, even in cases that do not lead to overt regularization behavior.
In summary, raising cognitive load resulted in noisier encoding, however the noise was not biased in the direction of regularity. Estimates in the linguistic domain were not biased toward regularity either. It appears that the bulk of regularization occurs during the production-side of the experiment and is likely to involve processes of frequency retrieval and use.
\subsection{Individual differences in frequency learning strategy} \label{sec:individual}
The bimodal distributions over output ratios (refer back to Figure \ref{fig:2x2_6panels}) suggest individual differences in frequency learning strategies. We break frequency learning behavior into three categories: \emph{regularizing}, \emph{probability matching}, and \emph{variabilizing}. How many participants fall into each category? And in the high load conditions, where participants respond to more than one item, how consistent are their responses with one strategy?
We define \emph{probability matching} as sampling from the input ratio, with replacement. This leads to output ratios that are binomially distributed\footnote{Humans can probability match with variance that is significantly lower than binomial variance \citep[][pp.45-57]{Ferdinand2015}. Therefore, the definition of probability matching used in this paper is a conservative one.}
about the mean (where the mean equals the input ratio). Although the single most likely output ratio a participant could sample is the set of input ratios itself, most probability matchers will sample a ratio that has higher or lower entropy than the input ratio. We will classify participants who produced ratios within the 95\% confidence interval of sampling with replacement behavior as probability matchers. We classify participants as \emph{variabilizers} if they produced ratios with significantly higher entropy than likely under probability matching behavior. These could be participants who were attempting to produce a maximally variable set (all 5:5 ratios) or randomly selecting among the two choices on each production trial. Likewise, we classify participants as \emph{regularizers} if they produced ratios with significantly lower entropy than likely under probability matching behavior. It is important to note that a participant with a very weak bias for regularity or variability may consistently produce data that falls within the 95\% confidence range of probability matching. However, we take a conservative approach by grouping individuals as regularizers or variabilizers only when probability matching has low probability.
\begin{figure*} [t] \centering
\includegraphics[width=180mm]{individual.pdf}
\caption{Linguistic and non-linguistic stimuli evoke different frequency learning strategies. Data are from the high cognitive load conditions \emph{marbles6} (top) and \emph{words6} (bottom). The x-axis shows participant number, sorted by their conditional entropy (low to high). The y-axis shows the frequency of the majority variant in the participant's output; each point represents performance on a single container/object, and there are therefore 6 points per participant. The shaded region contains all participants classified as probability matchers. Participants to the left of the shaded region are classified as regularizers and participants to the right are classified as variabilizers.}
\label{fig:individual}
\end{figure*}
In the low load conditions, where participants only sample one ratio, the 95\% confidence intervals on output ratios were determined with the Clopper-Pearson exact method.\footnote{95\% confidence interval on probability matching per input ratio: \\ 5:5, $0.19 \leq x \leq 0.81$; 6:4, $0.26 \leq x \leq 0.88$; 7:3, $0.35 \leq x \leq 0.93$; 8:2, $0.44 \leq x \leq 0.97$; 9:1, $0.55 \leq x \leq0.99$; 10:0, $0.69 \leq x \leq 1$, where $x$ is the frequency of the majority variant.}
In the high cognitive load conditions, where participants sample a set of six ratios, we classify the set of ratios according to their conditional entropy $H(V|C)$ (refer back to Section \ref{sec:entropy}). The 95\% confidence interval on conditional entropy for probability matching in this experimental setup is 0.43 to 0.75 bits (determined by $10^{5}$ runs of simulated probability matching behavior). Participants who produced data with entropy in the range $0.43 \leq x \leq 0.75$ were classified as probability matchers, those who produced data in the range $0 \leq x < 0.43$ were classified as regularizers, and those who produced data in the range $0.75 < x \leq 1$ were classified as variabilizers.
\begin{table}[h]
\scalebox{0.88}{
\begin{tabular}{ | l | c | c | c | }
\hline
& regularizers & probability matchers & variabilizers \\
\hline
\emph{marbles1} & 10 (5\%) & 173 (90\%) & 9 (5\%) \\
\hline
\emph{words1} & 50 (26\%) & 139 (72\%) & 3 (2\%) \\
\hline
\emph{marbles6} & 30 (47\%) & 15 (23\%) & 19 (10\%) \\
\hline
\emph{words6} & 42 (66\%) & 14 (22\%) & 8 (12\%) \\
\hline
\end{tabular}}
\caption{Participants classified by frequency learning strategy. Percentages show how the strategies break down within each condition.}
\label{tab:individual}
\end{table}
Table \ref{tab:individual} shows the number of participants that fell into each frequency learning category, per condition. All strategies are represented within each experimental condition. There is a significant effect of cognitive load ($\chi^2(2) = 151.63, p < .001$) and domain ($\chi^2(2) = 31.49, p < .001$) on the distribution of frequency learning strategies, meaning that the experimental manipulations elicit different frequency learning strategies from participants. Because fewer data points were collected from participants in the low load condition, probability matching behavior is not easily ruled out, hence the high number of participants classified as probability matchers in \emph{marbles1} and \emph{words1}. It is possible that the difference in dataset size between the low and high conditions is responsible for the significant effect of load. The effect of domain, however, is reliably due to the experimental manipulation. Therefore, the remainder of this section focuses on the high load data.
Figure \ref{fig:individual} shows the set of six output ratios that each participant produced in the high cognitive load conditions. The sets are sorted by their entropy and the shaded box shows the sets that fell into the $0.43 \leq x \leq 0.75$ bit range (classified as probability matchers). Participants to the left of the box are classified as regularizers and participants to the right are classified as variabilizers. More regularizers were found in the linguistic domain, more variabilizers were found in the non-linguistic domain, and probability matchers seem equally likely to be found in either domain.
At the extreme left of the x-axis, we see the subset of regularizers, numbering 6 participants in \emph{marbles6} and 22 in \emph{words6}, who produced a maximally regular set (all 10:0 or 0:10, conditional entropy = 0 bits). No participants produced a maximally variable set (all 5:5 ratios, conditional entropy = 1 bit). Participants are more likely to maximally regularize in the linguistic condition ($\chi^2(1) = 10.2857, p = .001$). Although some participants regularized with the majority variant exclusively, \emph{no} participants regularized with the minority variant exclusively. Points in the 0-4 range on the the y-axis correspond to output ratios that contained a large number of minority variant productions (i.e. the majority variant had frequency of between 0 and 4). Most participants regularized with 1-2 minority variants and 4-5 majority variants.
In summary, we found that all frequency learning strategies, \emph{regularizing}, \emph{probability matching}, and \emph{variabilizing}, are present in each condition and the use of linguistic stimuli causes more participants to consistently regularize.
\subsection{Primacy and recency effects on regularization} \label{sec:primacy}
Studies on regularization often find that participants regularize by over-producing or over-predicting the majority variant, and this serves as the standard definition of regularization \citep[e.g.][]{HudsonKam2005}. However, many studies report some participants who regularize with the minority variant \citep[e.g.][]{HudsonKam2009,Reali2009, Smith2010a, Culbertson2012, Perfors2012, perfors2016adult}. What causes some participants to regularize with the majority variant, and others to regularize with the minority variant? In the previous section, we saw minority regularization is not due to individual differences in frequency learning behavior. If minority regularization is not a feature of individuals, it may be a feature of the training data they received.
One possible data-driven explanation for minority regularization lies in the effects of a stimulus's primacy and recency on participant behavior. In the observation phase, participants were presented with a randomly-ordered sequence of variants, such that the probability of any particular variant occurring at the beginning or end of the input sequence is proportional to its frequency in the sequence. Therefore, some participants would have received minority variants toward the beginning and/or end of the sequence, whereas others would have not.
Many experiments on the serial recall of lexical items show that participants are better at recalling the first and last few items in a list of words \citep[e.g.][]{Deese1957,Murdock1962}. This effect also extends to the learning of mappings between words and referents: \citet{Poepsel2014} found that when participants in a cross-situational learning task were confronted with several possible synonyms for an object, their confidence in a correct mapping was positively correlated with the primacy of that mapping in the observation phase. Therefore, we investigated the effect of the minority variant's position in the input sequence on participants' tendency to regularize with the minority variant.
Unlike most research on primacy and recency (which present participants with a long list of unique stimuli), our input sequences only consist of two variants, presented several times each. Therefore, we can quantify the strength of minority primacy as the imbalance of the variants across the input sequence. To do this, we will use the notion of net torque. In this analogy, we consider the input sequence to be a weightless lever of length 10 (the number of observation trials), we consider each minority variant to be a weight of one unit which is placed on the lever according to its observation trial number, and we assume the lever is balanced on a fulcrum at its center. The sum of the distance of the weights located right of center minus the sum of the distance of the weights left of center is the net torque. We will use the following standardization of net torque\footnote{Thanks to Andrew Berdahl for providing this solution. The score was standardized in order to de-correlate net torque with input ratio.}, and refer to it as the primacy score:
\begin{equation}
primacy\ score = - \Bigg(\frac{ \sum\limits_{d=1}^N w_d(d - \frac{N+1}{2})}
{m(N-m)/2} \Bigg)
\end{equation}
\label{position_score}
\noindent
where $w$ is the sequence of weights and $d$ is the distance of that weight from the start of the sequence. In the 5:5 input sequences, a random variant is coded as the ``minority" variant. $N$ is the length of $w$ and $m$ is the total number of minority variants in the sequence. Positive values mean that the minority variants occur more toward the beginning of the sequence and negative values mean they occur more toward the end of the sequence. The maximum primacy score is 1 and the minimum is -1. The average primacy score is 0 and is obtained when the sequence is balanced (i.e. minority variants are equally distributed early and late in the input sequence). For example (where 1 indicates an occurrence of the minority variant in the input sequence), the primacy score of sequence 1110000000 is 1, 0000000001 is $-1$, 0101001000 is 0.33, 1000000001 is 0, and 0000110000 is 0.
\begin{table}[h]
\centering
\raggedright
\scalebox{0.78}{
\begin{tabular}{ | l | c | c | c | c | }
\hline
production sequences & \emph{marbles1} & \emph{words1} & \emph{marbles6} & \emph{words6} \\
\hline
total & 192 & 192 & 384 & 384 \\
\hline
regularized & 43 & 85 & 201 & 241 \\
\hline
regularized w/minority & 16 (37\%) & 18 (21\%) & 53 (26\%) & 63 (26\%) \\
\hline
\end{tabular}}
\caption{Number of regularized production sequences per condition. Parentheses show the number of minority-regularized sequences as a percentage of all regularized sequences.}
\label{tab:reggers1}
\end{table}
Primacy analyses were restricted to the input sequences that participants regularized. Table \ref{tab:reggers1} shows a breakdown of the number of regularized production sequences per experimental condition (i.e. all output sequences that had lower entropy than their corresponding input sequence). Participants regularized a total of 570 input sequences.
Figure \ref{fig:primacy} plots the primacy scores of the 420 sequences that were regularized with the majority variant (grey) and the 150 sequences that were regularized with the minority variant (black). We constructed a logit mixed effects model of regularization type (majority or minority regularization) as a function of primacy score. Participant was entered as a random effect (with random intercepts). A likelihood ratio test was performed on this model and a reduced model which omits primacy score as a predictor. We found a significant effect of primacy score on regularization type ($\chi^2(1)=6.4082, p = 0.01$). On average, primacy score is 0.11 points higher ($\pm$ 0.04 standard errors) in sequences that were regularized with the minority. This means that participants are more likely to regularize with the minority variant when they saw it toward the beginning of their input sequence (i.e. when minority variant primacy is high).
However, minority regularization is not entirely explained by minority primacy. As can be seen in Figure \ref{fig:primacy}, minority regularization was obtained across all primacy scores and even when the minority was maximally recent (left-most black bar).
In summary, we found that participants who saw the minority variant toward the beginning of the observation phase were more likely to regularize with the minority variant. This helps explain some of the individual differences in regularization behavior, by grounding those differences in the properties of the data each participant observed.
\begin{figure} [t]
\includegraphics[width=90mm]{primacy.pdf}
\caption{Participants are more likely to regularize with the minority variant when they observe it toward the beginning of the input sequence.
The $x$-axis is the primacy of the minority variant in the input sequence, ranging from $-1$ (maximal recency) to 1 (maximal primacy). Bars show the number of input sequences that were regularized by over-producing the minority variant (black) and by over-producing majority variant (grey).}
\label{fig:primacy}
\end{figure}
\section{Predicting the evolution of regularity}
\begin{figure*} [t]
\includegraphics[width=183mm]{Q_matrices.pdf}
\caption{The data from the experiment is used to predict the cultural evolution of regularization.
\emph{Top}: Estimated transition matrices for each experimental condition contain the probabilities that a learner produces any given output ratio from any given input ratio (presented in terms of the frequency of variant $x$ in each input ratio \emph{x:y}). The shading of the cells denote the transition probabilities between states. Each row in the matrix corresponds to the distribution of output ratios produced in response to one input ratio (rows are the same distributions in Figure \ref{fig:2x2_6panels}, only smoothed). For example, row 5 in the \emph{marbles1} transition matrix corresponds to the upper left panel of Figure \ref{fig:2x2_6panels}, and the probability of transitioning from $s_{t-1=5}$ to $s_{t=6}$ is equivalent to the (smoothed) proportion of participants that produced a 6:4 ratio when trained on a 5:5 ratio. Likewise, rows 4 and 6 correspond to the 6:4 panel in Figure \ref{fig:2x2_6panels}, but this distribution is flipped in row 4 to display the results in terms of the minority variant. \emph{Bottom}: The stationary distribution shows the percentage of learners who will produce each output ratio, after the ratios have evolved for an arbitrarily large number of generations. Each stationary distribution is the solution to the matrix above it.}
\label{fig:Qmatrices}
\end{figure*}
In the previous sections, we showed that learners regularize novel word frequencies due to domain-general and domain-specific constraints. This was accomplished by analyzing \emph{one cycle of learning}, which spans the perception, processing, and production of a set of variants. Although this informs us about the relevant constraints that may underpin regularity in word learning, and even \emph{how much} regularity each constraint imposes on a given data set, it does not necessarily tell us how much regularity we will expect to see in a set of linguistic variants over time. This is because languages are transmitted between generations of learners and are therefore subject to multiple learning cycles, where each individual has an opportunity to impose some amount of regularity on the language.
In this section, we address the complex relationship between regularization biases and the level of regularity found in culturally transmitted data. In particular, we will focus on the evolution of regularity in the \emph{marbles6} and \emph{words1} conditions, because these two conditions elicited similar amounts of regularization behavior from two very different causes: domain-general and domain-specific constraints on frequency learning. Would a data set which is culturally transmitted under conditions of only high cognitive load (as in \emph{marbles6}) or only linguistic framing (as in \emph{words1}) ultimately acquire the same amount of regularity?
To answer this question, we will explore the dynamics of change in our existing data using an iterated learning model of cultural transmission \citep{Kirby2014iterated} in which the output of one learner serves as the input to another \citep[e.g.][]{Kirby2001,Brighton2002,Smith2003,Kirby2008,Reali2009,Smith2010a}. Several cycles of iterated learning result in a walk over the complex landscape of constraints that shape the transmitted behavior, and several walks can be used to estimate this landscape and its likely evolutionary trajectories. \citet{Griffiths2007} have shown that iterated learning is equivalent to a Markov process, which is a discrete-time random process over a sequence of values of a random variable, $v_{t=1}, v_{t=2}, ..., v_{t=n}$, such that the random variable is determined only by its most recent value \citep[p.535]{Papoulis1984}:
\begin{equation}
P(v_{t} | v_{t=1}, v_{t=2}, ..., v_{t-1}) = P(v_{t} | v_{t-1})
\label{eq:Q}
\end{equation}
\noindent
This describes a memoryless, time-invariant, process in which only the previous value ($v_{t-1}$) has an influence on the current value ($v_t$). This is the case for iterated learning chains when learners only observe the behaviors of the previous generation. All of the possible values of the random variable constitute the state space of this system. A Markov process is fully specified by the probabilities with which each state will lead to every other state and these probabilities between states can be represented as a transition matrix, \textbf{Q} \citep[p.3]{Norris2008}. The probabilities in \textbf{Q} are the landscape over which a culturally transmitted dataset evolves.
In our experimental data, each state $s$ corresponds to one of the eleven possible ratios: $s_0, s_1, ..., s_{10} =$ \{0:10, 1:9, 2:8, 3:7, 4:6, 5:5, 6:4, 7:3, 8:2, 9:1, and 10:0\}, where $s_{t-1}$ is the input ratio and $s_t$ is the output ratio. Our experiment was designed so that \textbf{Q} could be estimated for each of the four experimental conditions, by collecting data from participants in each of the eleven possible states. Figure \ref{fig:Qmatrices} (top row) shows the estimated transition matrix from each experimental condition. Each estimation consists of the raw data in that condition, smoothed with a small value $\epsilon = \textstyle\frac{1}{length(row)^2}$. Each cell in the matrix, \textbf{Q}$_{ij}$, gives the transition probability from state $s_{i=t-1}$ to state $s_{j=t}$.
The transition matrices can be used to estimate the regularity of the data after an arbitrarily large number of learning cycles. No matter what start state is used to initialize an iterated learning chain, an arbitrarily large number of iterations will converge to a stationary distribution, $\vec{s}$. The stationary distribution is defined as $\vec{s}$\textbf{Q}$ = \vec{s}$, meaning that once the data take the form of the stationary distribution and serve as the input to $\textbf{Q}$, the output will be the same distribution and the subsequent generations of data will not change anymore. The stationary distribution is a probability distribution over all states in the system, where each probability corresponds to the proportion of time the system will spend in each state, and can be solved for any matrix by decomposing the matrix into its eigenvalues and eigenvectors: $\vec{s}$ is proportional to the first eigenvector. Figure \ref{fig:Qmatrices} (bottom row) shows the stationary distribution for each transition matrix. From these distributions, we see that an arbitrarily long iterated learning chain will produce maximally regular (0:10 and 10:0) ratios approximately 40\% of the time when participants are learning about 12 marbles and six containers (\emph{marbles6}) and approximately 80\% of the time when participants are learning about two words and one object (\emph{words1}). The difference between stationary distributions here means that the evolutionary dynamics of these two experimental conditions differ.
\begin{figure} [t] \centering
\includegraphics[width=90mm]{8bars_stat.pdf}
\caption{Same learning biases lead to different degrees of regularization after many generations of cultural transmission.
\emph{Dark grey}: Average change in entropy after one learning cycle (same data in Figure \ref{fig:4bars}, reprinted here for comparison).
\emph{Light grey}: Average change in entropy of variants after convergence to the stationary distribution (i.e. after an infinite number of learning cycles).
Error bars indicate 95\% confidence intervals, computed by the bootstrap percentile method \citep{Efron1979} on 10,000 resamples of the transition matrix, where each matrix was solved for its stationary distribution and mean change in entropy.}
\label{fig:stat_reg}
\end{figure}
We calculate the level of regularity in the stationary distribution by multiplying the Shannon entropy of the ratio (defined by each state, $\vec{s}_i$) by the probability of observing that state, $p(\vec{s}_i)$. The results are 0.61 bits of conditional entropy $H(V|C)$ in \emph{marbles1}, 0.43 bits in \emph{marbles6}, 0.16 bits in \emph{words1}, and 0.24 bits in \emph{words6}. We compare these values to the results of the experiment (the average conditional entropy achieved after one learning cycle), which was 0.66 bits in \emph{marbles1}, 0.50 bits in \emph{marbles6}, 0.48 bits in \emph{words1}, and 0.32 bits in \emph{words6}.
Figure \ref{fig:stat_reg} plots these values in terms of entropy change: as the difference between the mean input entropy and the mean output entropy after one learning cycle (in dark grey) and after convergence to the stationary distribution (in light grey). Here we see that, despite showing similar mean entropy change in the experiment, the regularization biases involved in \emph{marbles6} and \emph{words1} ultimately produce different levels of regularity via cultural transmission (inferred by the non-overlapping 95\% confidence intervals in stationary regularity between \emph{marbles6} and \emph{words1}). This is due to the different distribution of probabilities within the transition matrices. These probabilities constitute different landscapes that attract iterated learning chains into different regions of the state space. One reason why the \emph{words1} data regularizes more than the other data sets, is that it has a markedly lower probability of transitioning out of the 10:0 and 0:10 states, trapping generations of learners in this highly regular region for longer amounts of time.
In summary, we have shown that the regularity elicited by two different constraints on frequency learning (the domain-general regularization biases involved in \emph{marbles6} and the domain-specific regularization biases involved in \emph{words1}) is similar in one generation of learners, but displays different evolutionary dynamics under simulated cultural transmission. This finding has important implications for the relationship between learning biases and structure in language: it means that culturally transmitted systems, such as language, do not necessarily mirror the biases of its learners \citep[see][]{Kirby1999, kirby2004ug, smith2017language}. Previously, we showed that cognitive load and linguistic domain are independent sources of regularization in individual learners. Looking at the data from individual learners, we may even infer that cognitive load and linguistic domain inject similar amounts of regularity into language. However, the fact that \emph{words1} has higher stationary regularity than \emph{marbles6} means, at least in terms of the present data, that the amount of regularity we ultimately expect to find in a language is not simply predicted from a learner's biases. Instead, the process of cultural transmission is an indispensable piece of the puzzle in explaining how learning biases shape languages.
\section{Discussion}
Regularity in language is rooted in the cognitive apparatus of its learners. In this paper, we have shown that linguistic regularization behavior results from at least two, independent sources in cognition. The first is domain-general and involves constraints on frequency learning when cognitive load is high. The second is domain-specific and is triggered when the frequency learning task is framed with linguistic stimuli.
Cognitive load was manipulated by varying the number of stimuli in a frequency learning task. When participants observed and produced for more stimuli, they regularized stimuli frequencies \emph{more} on average than when they were observing and producing for fewer stimuli. This result held when stimuli were non-linguistic (marbles and containers) and when stimuli were linguistic (words and objects) and has previously been observed in separate non-linguistic and linguistic experiments \citep{Gardner1957,HudsonKam2009}. We have shown, within the same experimental setting and for identical distributions of variation, that increasing cognitive load causes participants to regularize both non-linguistic and linguistic stimuli. Furthermore, we have shown that participants regularize a similar amount of variation in both cases, eliminating 24.6\% of the variation in marbles conditioned on containers and 25.5\% of the variation in words conditioned on objects. This similarity suggests that learners have general limits on the amount of variation they can process and reproduce, which are independent of the learning domain. It is quite possible that cognitive load makes a fixed contribution to regularization behavior, however it remains to be seen whether this result holds over different learning domains and cognitive load manipulations.
One possible alternative explanation for the cognitive load effect on regularization behavior is the differing length of the two tasks. Our design kept duration per stimulus constant across conditions, rather than total task duration, because it is unknown how stimuli presentation length affects regularization behavior. However, it is possible that participants' attention was lower at the end of the high cognitive load tasks, causing them to over-produce the stimuli they saw early on in the training phase. Given our finding that primacy affects minority regularization more than recency, this could be the case, although that effect was quite small. Future research should address the effect of stimulus duration and presentation order on the degree to which participants regularize.
Domain was manipulated by varying the type of stimuli used in the frequency learning task. When participants observed and produced mappings between words and objects, they regularized more than participants who observed and produced mappings between marbles and containers. Participants appear to have a higher baseline regularization behavior when learning about linguistic stimuli: an additional 27\% of variation was regularized due to linguistic domain in each cognitive load condition (26.7\% in the low condition, 27.4\% in the high condition).
The use of linguistic stimuli may trigger any number of domain-specific learning mechanisms or production strategies. One possibility is that the stimuli manipulation changed participants' pragmatic inferences about the frequency learning task. In an artificial language learning task, \cite{perfors2016adult} showed that participants regularize more when they believe that the variation in labels for objects can be the result of typos, suggesting that participants are more likely to maintain variation when they think it is meaningful. It is possible that participants make different assumptions about the importance of variation in marbles versus words when they are required to demonstrate what they have learned to an experimenter. However, it is not clear what these assumptions may be. Another possibility is that the use of linguistic stimuli encourages participants to consider the communicative consequences of variation. Participants in artificial language learning tasks regularize more when they are allowed to communicate with one another \citep{feher2016structural,smith2014eliminating} and even when they erroneously believe they are communicating with another participant \citep{feher2016structural}. This suggests that participants strategically regularize variation in situations that are potentially communicative and may be the reason that regularization is observed in a wide range of language learning tasks, including the present study.
The use of non-linguistic stimuli may not fully put participants into a non-linguistic task framing: participants may be saying ``orange", ``blue", etc as they observe the marbles or verbalizing a rule such as ``there are more blue marbles in the jar". If humans rely on language for solving complex problems, they may trigger linguistic representations when stimuli become more complex, regardless of the learning domain. Following this logic, any increase in cognitive load could make any task more linguistic. This would change the interpretation of our results, such that degree-differences in regularization behavior map on to degree-differences in the amount of linguistic representation involved. Adopting this interpretation of our experiment, however, would require an overhaul of the definition of ``domain-general learning mechanisms" in statistical learning.
We also investigated the role of encoding errors on regularization behavior. After the production phase, participants estimated the ratio of the variants associated with each container or object they had observed.
We found that the same factors that affected production data also affected estimates (domain, cognitive load, and input frequency).
However, the estimates themselves were not significantly more regular than the input ratios that participants observed. This suggests that participants had access to somewhat accurately encoded frequency information when making their estimates. Because participants regularized their productions without showing a corresponding bias in estimates, this implies that the bulk of their regularization occurred during the production phase of the task. This production-side interpretation is in line with the results of \citet{Chang2009}, who showed that adult participants regularize more when stimuli retrieval is made harder; \citet{Perfors2012}, who found adult participants do not regularize when encoding is made harder; and \citet{schwab2018regularization}, who show that children regularize during production despite demonstrated awareness of all word forms used during training. This result also suggests that the Less-is-More hypothesis \citep{Newport1990}, which states that learners regularize because they fail to encode, store, or retrieve lower-frequency linguistic forms, applies more to retrieval and less to encoding. However, it is possible that biased encoding could result from more complex mappings than those used in this experiment.\footnote{\citet{Vouloumanos2008} found that learners are able to encode and retrieve fine-grained differences in the statistics of low-frequency mappings between words and objects (which we calculate had a joint entropy of 4.41 bits), but failed to encode and retrieve fine-grained differences for a more complex stimuli set (with a joint entropy of 5.15 bits). The joint entropy of our high cognitive load mappings, at 3.26 bits, is within \citet{Vouloumanos2008}'s demonstrated threshold for accurate frequency representation.}
An alternative explanation for the difference between the estimates and productions could be due to how these two types of data were elicited from participants. It is likely the estimation question elicited more explicit knowledge about stimuli frequencies and the production task elicited more implicit knowledge (see \citealp{cleeremans1998implicit} for review). If this is the case, it would mean that participants' explicit knowledge of observed frequencies is more accurate than their implicit knowledge and imply that regularization behavior is more closely associated with implicit knowledge retrieval.
This paper also explored the topic of minority regularization in depth. We found that minority regularization is not the result of individual differences. Although participants did differ in frequency learning strategies (we found regularizers, probability matchers, and variabilizers in all four conditions), most participants regularized with only one or two minority variants. Therefore, we investigated differences in the randomized stimuli that each participant saw and found that participants are significantly more likely to regularize with the minority variant when it occurs toward the beginning of the observation sequence. This primacy effect is in line with the results of \citet{Poepsel2014}, which showed that participants in a cross-situational learning task had higher confidence in the correctness of a mapping between words and referents when those items co-occurred early in the observation phase. We also demonstrated how minority regularizers can confound regularization analyses which are based on the majority variant's change in frequency (Section \ref{sec:freq_analysis}) and argue that regularization should not be defined exclusively as ``overproduction of the highest-frequency or dominant form". Alternative analyses that overcome this issue are \citet{Perfors2012,perfors2016adult}'s \emph{regularization index} and entropy-based analyses, as we use here (see Section \ref{sec:entropy}).
There are several pros and a few cons to using entropy-based analyses. Regularization occurs whenever learners increase the predictability of a linguistic system and therefore directly equates to a system's decrease in entropy. Entropy measures allow us to quantify linguistic variation directly, in a mathematically-principled way based on predictability. They also allow us to quantify all of the variation in a set of linguistic variants at once. Analyses based on majority-variant frequency only tell us about changes to one or a subset of the variants in a language. Overproduction of majority forms certainly can cause a language's entropy to drop, but regularity also can increase when minority forms are overproduced or when forms are maintained but conditioned on other linguistic contexts or meanings. Entropy measures also force us to be explicit about what type of linguistic variation we are analyzing (i.e. the number of variants or their predictability in conditioning contexts) and allow direct comparison between different experiments (see footnote 5 as an example). However, entropy and frequency analyses are sensitive to different aspects of linguistic data. Entropy is better for quantifying regularization and positively identifying it, whereas frequency is better for detecting a population-level trend in over- or under-producing a particular variant. For example, the frequency method used in Section \ref{sec:freq_analysis} did not capture the effect of cognitive load on frequency learning behavior, but it did capture an interesting domain difference that the entropy analysis missed: marble drawers overproduced the minority variant on average, whereas word learners did not. These two methods also show differences in the classification of probability matching behavior: the entropy method identified \emph{marbles1} as consistent with probability matching behavior and the frequency method did not (because there is a significant bias toward the minority variant). This raises important questions about the nature of probability matching: should it be defined as reproducing the same amount of variation (as the entropy measure captures) or reproducing the same amount of variation \emph{along with} the correct mapping of variation to stimuli (as the frequency measure captures)?
Overall, this paper explored how various cognitive constraints on frequency learning give rise to regularization behavior. But what can detailed knowledge of these constraints tell us about the regularity of languages?
One possibility is that the relationship between constraints on learning and structure of languages is straightforward, such that learning biases can be directly read off the typology of languages in the world \citep[e.g.][]{baker2002atoms} or the probability that a learner ends up speaking any given language can be read off the probability of the language in its prior \citep{Griffiths2007}. Under other conditions, however, cultural transmission distorts the effects of learners' biases on the data they transmit, making it impossible to simply read learning biases off of language universals \citep{Kirby1999, kirby2004ug}. Often, this distortion increases the effects of the bias over time, such that weak biases have strong effects on the structure of culturally transmitted data \citep[e.g.][]{Kalish2007,Kirby2007, Griffiths2008,Reali2009,Smith2010a, thompson2015transmission}. However, the opposite can also occur: biases can have weaker effects or no effects at all \citep{smith2017language}. This suggests that cultural transmission increases the complexity of the relationship between individual learning biases and the structure of language.
By plugging the data obtained from our population of participants into a model of cultural transmission, we also found a complex relationship between regularization biases and regularity in culturally transmitted datasets.
Although participants produced similar amounts of regularity in response to the domain and load manipulations, the domain manipulation resulted in significantly higher regularity once the data was culturally transmitted.
Future research should explore whether similar effects of domain and demand on regularization behavior will be seen with child, rather than adult, learners. It is possible that child learners would show different relative contributions of domain general and domain specific biases than adults do. Therefore, to the extent that language change is shaped specifically by biases in child acquisition \citep[see][for opposing views]{HudsonKam2005,slobin2014ontogenesis}, this may have implications for the role of language-specific and domain-general biases in shaping the evolution of linguistic regularity.
\section{Conclusion}
When learners observe and reproduce probabilistic variation, we find they regularize (reduce variation) when cognitive load is high and when stimuli are linguistic. We conclude that linguistic regularization behavior is a co-product of domain-general and domain-specific biases on frequency learning and production. Furthermore, we find that load and domain affect how participants encode frequency information. However, encoded frequencies are not more regular than the data participants observed: the bulk of regularization occurs when participants produce data. Finally, we show that the relative contributions of load and domain to the regularity in a set of linguistic variants can change when data are transmitted culturally. In order to understand how various regularity biases create regularity in language, experiments that quantify learning biases need to be coupled with cultural transmission studies.
\section*{Acknowledgements}
This research was supported by the University of Edinburgh's College Studentship, the SORSAS award, and the Engineering and Physical Sciences Research Council.
The writeup was supported by the Omidyar Fellowship.
The reported experiment was conducted with the approval of the Linguistics and English Language Ethics Committee at the University of Edinburgh.
We thank Bill Thompson, Tom Griffiths, Florencia Reali, Simon DeDeo, Luke Maurits, and the members of the Centre for Language Evolution for their feedback on this project. We thank Daniel Richardson for providing experimental stimuli and thank our reviewers, Amy Perfors and two anonymous, for their excellent comments.
\section*{Appendix}
\begin{table}[!htb]
\centering
\scalebox{0.87}{
\begin{tabular}{ | l | c | c | c | }
\hline
& estimate & S.E. & t statistic \\
\hline
Intercept & 0.08567 & 0.04321 & $t(1152) = 1.983$ \\
\hline
Load (high) & 0.16715 & 0.04740 & $t(1152) = 3.527$ \\
\hline
Domain (words) & $-0.11373$ & 0.04149 & $t(1152) = -2.742$ \\
\hline
Input entropy & $-0.14094$ & 0.05494 & $t(1152) = -2.565$ \\
\hline
Load * Input entropy & $-0.49568$ & 0.05622 & $t(1152) = -8.816$ \\
\hline
Domain * Input ent. & $-0.09968$ & 0.04596 & $t(1152) = -2.169$ \\
\hline
\end{tabular}}
\caption{The best-fit linear mixed effects model for change in entropy, see Section \ref{sec:manips_lmer}.}
\label{tab:bestfit1}
\end{table}
\begin{table}[!htb]
\centering
\scalebox{0.87}{
\begin{tabular}{ | l | c | c | c | }
\hline
& estimate & S.E. & t statistic \\
\hline
Intercept & 0.11093 & 0.04804 & $t(1152) = 2.309$ \\
\hline
Domain (words) & -0.12580 & 0.06794 & $t(1152) = -1.852$ \\
\hline
Input frequency & -0.21521 & 0.06217 & $t(1152) = -3.462$ \\
\hline
Domain * Input freq. & 0.22925 & 0.08792 & $t(1152) = 2.608$ \\
\hline
\end{tabular}}
\caption{The best-fit linear mixed effects model for change in frequency, see Section \ref{sec:freq_analysis}.}
\label{tab:bestfit_freq}
\end{table}
\begin{table}[!htb]
\centering
\scalebox{0.87}{
\begin{tabular}{ | l | c | c | c | }
\hline
& estimate & S.E. & t statistic \\
\hline
Intercept & $0.31193$ & 0.03127 & $t(1152) = 9.976$ \\
\hline
Load (high) & $0.13083$ & 0.03672 & $t(1152) = 3.563$ \\
\hline
Domain (words) & $-0.05145$ & 0.02001 & $t(1152) = -2.571$ \\
\hline
Input entropy & $-0.37040$ & 0.03901 & $t(1152) = -9.496$ \\
\hline
Load * Input ent. & $-0.24258$ & 0.04563 & $t(1152) = -5.316$ \\
\hline
\end{tabular}}
\caption{The best-fit linear mixed effects model for estimates, see Section \ref{sec:encoding}.}
\label{tab:bestfit_encode}
\end{table}
|
\section{Introduction}
XTE J1709-267 is a recurrent soft X-ray transient and Type 1 burst source that has a recurrence time of 2-3 years (\citealt{atel255}; \citealt{atel1302}; \citealt{atel2729}; \citealt{atel4304}; \citealt{atel5319}). The source was first discovered to be in outburst in 1997 \citep{marshall97} and is associated with the globular cluster NGC 6293 \citep{jonker04b}. It is located a distance of 8.5 kpc away \citep{Lee06}. The typical 2-10 keV flux during outburst is $\sim2\times10^{-9}$ ergs cm$^{-2}$ s$^{-1}$ (\citealt{marshall97}; \citealt{cocchi98}; \citealt{jonker03}, \citeyear{jonker04a}, \citeyear{jonker04b}; \citealt{atel255}; \citealt{atel1302}; \citealt{degenaar13}).
Broad iron line profiles have been seen in low-mass X-ray binaries (LMXBs) that contain a neutron star (NS) as the primary accreting compact object (e.g. \citealt{BS07}; \citealt{papitto08}; \citealt{cackett08}, \citeyear{cackett09}, \citeyear{cackett10}; \citealt{disalvo09}; \citealt{Egron13}; \citealt{miller13}). The effects of gravitational redshift and Doppler shift/boosting are imprinted on these emission lines. These relativistic effects become stronger closer to the compact object \citep{Fabian89}. Hence, the profile of the Fe K$_{\alpha}$ line gives a direct measure of the position of inner disk. Furthermore, since the disk must truncate at or before the surface of the star, the Fe K$_{\alpha}$ line can be used to set an upper limit for the radius of the NS (\citealt{cackett08}, \citeyear{cackett10}; \citealt{reis09}; \citealt{miller13}; \citealt{degenaar15}).
Two likely scenarios for disk truncation above $\sim1\%$ L$_{\mathrm{Edd}}$ are: (1) pressure balance between the accreting material and NS's magnetic field or (2) the boundary layer extending from the surface of the NS, impeding the disk. Thus, studies of disk reflection can also be used to set an upper limit on the magnetic field strength (\citealt{cackett09}; \citealt{Pap09}; \citealt{miller11}; \citealt{degenaar14}, \citeyear{degenaar16}; \citealt{king16}; \citealt{ludlam16}) or the extent of the boundary layer (\citealt{IP09}; \citealt{king16}; \citealt{ludlam16}, \citealt{chiang16b}).
MAXI/GSC registered that XTE J1709-267 had renewed activity in 2016 May 31 \citep{atel9108}.
We obtained two $\sim20$ ks observations with $\emph{NuSTAR}$ \citep{nustar} during this outburst while the source was in the soft state. We detect a broad Fe K$_{\alpha}$ line that we model as relativistic reflection to determine the extent of the inner accretion disk and measure the inclination. There are no previous Fe K detections for this source, making this the first detailed reflection analysis.
\section{Observations and Data Reduction}
$\emph{NuSTAR}$ observations of XTE J1709-267 were taken on 2016 June 8 (Obsids 90201025002 and 90201025003). There are two detectors aboard $\emph{NuSTAR}$ that collect data: focal plane module A (FPMA) and focal plane module B (FPMB). Lightcurves and spectra were created using a 120$^{\prime \prime}$ circular extraction region centered around the source using the {\sc{nuproducts}} tool from {\sc nustardas} v1.5.1 with {\sc caldb} 20160421. A background was generated and subtracted using another region of the same dimension away from the source. There were no Type 1 X-ray bursts present in the lightcurves, but there was an increase in count rate near the end of the second observation (see Figure 1). We create gti files in order to separate the observation by low and high count rate. Preliminary modeling of the spectra with a simple continuum multiplied by a cross normalization constant is performed to determine how well the detectors agree with one another. We fixed the constant to unity for the FPMA and allowed it to float for the FPMB. The floating constant was found to be within 0.95-1.05 in each case. We proceeded to combine the two source spectra, background spectra, and ancillary response matrices via {\sc addascaspec}. We use {\sc addrmf} to create a single redistribution matrix file. We then combined the two observations of the same count rate as per \citet{king16}, resulting in a total combined exposure time for the spectrum of $\sim62$ ks for the lower flux regime and $\sim15$ ks for the spectrum generated from the higher flux. We will refer to these spectra as the low and high hereafter. The spectra were grouped using {\sc grppha} to have a minimum of 25 counts per bin.
\begin{figure}
\centering
\includegraphics[width=8.2cm]{1709lc.eps}
\caption{Lightcurve of the $\emph{NuSTAR}$ observation of J1709-267 using 500 s time bins. The vertical dashed line indicates the change in count rate in which we divided the observation.
}
\label{fig:contour}
\end{figure}
\section{Spectral Analysis and Results}
We use XSPEC version 12.9.0 \citep{arnaud96} in this work with all errors quoted at 90$\%$ confidence level. Errors were calculated from Monte Carlo Markov Chain (MCMC) of length 100,000. We perform fits over the 3.0-30.0 keV energy range. Above 30 keV, the spectra are background dominated. We account for the equivalent neutral hydrogen column density along the line of sight via {\sc tbnewer}\footnote{Wilms, Juett, Schulz, Nowak, in prep, http://pulsar.sternwarte.uni-erlangen.de/wilms/research/tbabs/index.html}. We fix the absorption column to the \citet{dl90} value of $0.237\times10^{22}$ cm$^{-2}$ (consistent with previous spectral studies of this source; e.g. \citealt{jonker03}, \citealt{degenaar13}), since $\emph{NuSTAR}$ lacks the lower energy bandpass to constrain this on its own.
Initial fits were performed with an absorbed single temperature blackbody component ({\sc bbodyrad}) to model the corona or boundary layer, and a multi-temperature blackbody ({\sc diskbb}) to account for the accretion disk emission. This combination of models gave a particularly poor fit in each case ($\chi^{2}_{\mathrm{low}}/dof=1873.69/672$ \& $\chi^{2}_{\mathrm{high}}/dof=954.38/670$), partly owing to the presence of strong disk reflection features in the spectrum. We added a power-law component with the photon index bound at an upper limit of 4.0, which has been observed in astrophysical sources such as black hole X-ray novae (\citealt{sobczak00}; \citealt{park04}). The addition of a power-law component improved the the overall fit for the low flux case by $\Delta \chi^{2}_{\mathrm{low}}=350$ for 4 d.o.f. ($11\sigma$ improvement). This continuum model is in agreement with the framework laid out in \citet{lin07} for NS transients in the soft state, though they use a broken power-law component instead. The additional power-law component was not statistically necessary in the high flux state, therefore, we do not use it in that case. The reflection is still unaccounted for by these models. Figure 2 shows an asymmetric Fe K emission line, commonly associated with relativistic disk reflection. The red wing extends down to $\sim 5$ keV while the blue wing drops around $\sim7$ keV.
\begin{figure}
\centering
\includegraphics[angle=270,width=8.2cm]{lowFeprofile.eps}
\includegraphics[angle=270,width=8.2cm]{highFe3.eps}
\caption{Ratio of the data to the continuum model for $\emph{NuSTAR}$ observation of XTE J1709-267 during the period of lower (top) and higher (bottom) flux. The iron line region from 5-8 keV was ignored to prevent the feature from skewing the fit. A simple disk blackbody, single temperature blackbody, and power-law were fit to model the continuum over the energies of 3.0-5.0 keV and 8.0-10.0 keV for the low flux case. For the high flux state, a disk blackbody and single temperature blackbody were fit to model the continuum over the energies of 3.0-5.0 keV and 8.0-10.0 keV. Fitting up to 10 keV models both the continuum and some reflection continuum, but gives an unhindered view of the Fe K$_{\alpha}$ line.
}
\label{fig:feline}
\end{figure}
The {\sc reflionx}\footnote{http://www-xray.ast.cam.ac.uk/$\sim$mlparker/reflionx$\_$models/reflionx$\_$bb.mod} \citep{reflionx} model describes reflection from an ionized disk. We included a modified version that assumes the disk is illuminated by a blackbody, rather than a power law. To account for relativistic Doppler shifts and gravitational redshifts, we convolved {\sc reflionx} with {\sc relconv} \citep{relconv}. The parameters of {\sc reflionx} are as follows: ionization parameter ($\xi$), temperature of the incident blackbody in keV ($kT$), iron abundance ($A_{Fe}$), redshift ($z$), and normalization. The parameters of {\sc relconv} are as follows: inner emissivity index ($q_{in}$), outer emissivity index ($q_{out}$), dimensionless spin parameter ($a_{*}$), inner disk radius in units of inner most stable circular orbit (ISCO; $R_{in}$), and outer disk radius in units of gravitational radii ($R_{out}$).
A few reasonable conditions were enforced when making fits with {\sc relconv} and {\sc reflionx}. First, we fixed the spin parameter, $a_{*}$ (where $a_{*}=cJ/GM^{2}$), in the model {\sc relconv} to 0 in the subsequent fits since NS in LMXBs have $a_{*} \leq 0.3$ (\citealt{miller11}; \citealt{Galloway08}). Corrections for frame-dragging for $a_{*}<0.3$ give errors $\ll10\%$ \citep{Miller98} since the position of the ISCO is nearly constant for values of low spin. Therefore, this does not hinder our estimate of the inner disk radius. Further, the outer disk radius has been fixed to 400 $R_{g}$ (where $R_{g} = GM/c^{2}$).
Last, we tied the outer emissivity index, $q_{out}$ to the inner emissivity index, $q_{in}$, to create a constant emissivity index. We are unable to constrain the inner disk radius vs. inclination space when allowing the emissivity index to be a free parameter. We fix $q=3$ as would be expected for a disk in flat, Euclidean geometry illuminated by a point source (see \citealt{wilkins12} for review). Additionally, different plausible geometries for illuminating the disk around neutron stars, such as boundary layers or hot spots, appear to produce the same $r^{-3}$ emissivity profile (D. Wilkins, priv. comm.). Since the accretion disks surrounding NSs do not undergo extreme relativistic effects such as those around maximally spinning black holes, we do not expect steeper emissivity profiles. A shallower profile like $r^{-2}$ has been postulated based upon self-consistent MHD simulations for extended coronal emission surrounding a black hole (BH) and relies solely on mass, spin, and mass accretion rate \citep{kinch16}. However, the shallower profile may intimately depend on the specific set of parameters that were input into the simulation ($M_{BH}=10\ M_{\odot}$, $a_{*}=0$, and mass accretion rate at $1\%$ of Eddington) and, thus, may not be directly translatable to NSs.
The overall model we used for the low flux spectrum was {\sc tbnewer}*({\sc diskbb}+{\sc bbody}+{\sc pow}+{\sc relconv}*{\sc reflionx}). This model provides a better fit with $\chi^{2}_{\mathrm{low}}/dof=622.6/663$. This is a $>$21$\sigma$ improvement over the model that does not take into account disk reflection for each case. The overall model we used for the high flux spectrum was {\sc tbnewer}*({\sc diskbb}+{\sc bbody}+{\sc relconv}*{\sc reflionx}). This model provides an improvement of $>$15$\sigma$ ($\chi^{2}_{\mathrm{high}}/dof=652.51/665$). Parameters and values can be seen in Table 1. Figure 3 shows the best fit spectra.
For the low flux case, the {\sc diskbb} component has a temperature of $kT=1.64_{-0.04}^{+0.03}$ keV and norm$=4.9\pm0.4$ km$^{2}$/100\ kpc$^{2}$. The {\sc bbodyrad} component has a temperature of $kT=2.44_{-0.03}^{+0.02}$ keV and normalization of $0.36_{-0.01}^{+0.07}$ km$^{2}$/100\ kpc$^{2}$. The blackbody and disk blackbody normalizations are implausibly small but this is understood to be the result of spectral hardening in atmospheres above pure blackbody emission (\citealt{london86}; \citealt{shimura95}; \citealt{merloni00}). The power law may or may not be physical but is still needed at the 8$\sigma$ level. It has a steep photon index of $\Gamma=3.99_{-0.30}^{+0.01}$ with a normalization of $0.32_{-0.10}^{+0.05}$. The inner disk radius is truncated at $R_{in}=13.8_{-1.8}^{+3.0}\ R_{g}$ and the inclination was found to be $25.2_{-1.1}^{+2.6} \ ^{\circ}$.
For the high flux case, the {\sc diskbb} component has a temperature of $kT=1.76_{-0.04}^{+0.05}$ keV and norm$=9\pm1$ km$^{2}$/100\ kpc$^{2}$. The {\sc bbodyrad} component has a temperature of $kT=2.44\pm0.04$ keV and normalization of $0.7_{-0.1}^{+0.5}$ km$^{2}$/100\ kpc$^{2}$. The inner disk radius is truncated further out at $R_{in}=23.4_{-5.4}^{+15.6}\ R_{g}$, though it is consistent with the value found in the low flux state at the 3$\sigma$ level. The inclination is $29_{-7}^{+10}\ ^{\circ}$, which also agrees with what is found from the low flux spectrum.
\begin{table}
\caption{J1709-267 Reflection Modeling}
\label{tab:refl}
\begin{center}
\begin{tabular}{llcc}
\hline
Component & Parameter & Low & High \\
\hline
{\sc tnewer}
&$N_\mathit{H} (10^{22}) ^{\dagger}$
&$0.237$&$0.237$
\\
{\sc diskbb}
&$kT$
&$1.64_{-0.04}^{+0.03}$
&$1.76_{-0.04}^{+0.05}$
\\
&norm
&$4.9\pm0.4$
&$9\pm1$
\\
{\sc bbodyrad}
&$kT$
&$2.44_{-0.03}^{+0.02}$
&$2.44\pm0.04$
\\
&norm
&$0.36_{-0.01}^{+0.07}$
&$0.7_{-0.1}^{+0.5}$
\\
{\sc power law}
&$\Gamma$
&$3.99_{-0.30}^{+0.01}$
&---
\\
&norm
&$0.32_{-0.10}^{+0.05}$
&---
\\
{\sc relconv}
&$q ^{\dagger}$
&3.0
&3.0
\\
&$a_{*} ^{\dagger}$
&0&0
\\
&$\mathit{i} (^{\circ})$
&$25.2_{-1.1}^{+2.6}$
&$29_{-7}^{+10}$
\\
&$R_\mathit{in} (ISCO) $
&$2.3_{-0.3}^{+0.5}$
&$3.9_{-0.9}^{+2.6}$
\\
&$R_\mathit{in} (R_{g}) $
&$13.8_{-1.8}^{+3.0}$
&$23.4_{-5.4}^{+15.6}$
\\
&$R_\mathit{out} (R_\mathit{g}) ^{\dagger}$
&400&400
\\
{\sc reflionx}
&$\xi$
&$200_{-30}^{+80}$
&$130_{-20}^{+10}$
\\
&$A_\mathit{Fe} $
&$0.57_{-0.04}^{+0.22}$
&$0.51_{-0.1}^{+1.1}$
\\
&$\mathit{z} ^{\dagger}$
&0&0
\\
&norm
&$0.19_{-0.07}^{+0.02}$
&$1.0_{-0.4}^{+0.2}$
\\
&$F_{\mathrm{unabs},\ 0.5-50.0\ keV}$
&$2.0_{-1.0}^{+0.6}$
&$2.6^{+1.9}_{-1.1}$
\\
&$L_{\mathrm{unabs},\ 0.5-50.0\ keV}$
&$1.7^{+0.5}_{-0.8}$
&$2.2^{+1.6}_{-1.0}$
\\
\hline
&$\chi^{2}$(dof)
&622.6 (663)
&652.5 (665)
\\
\hline
$^{\dagger}$ = fixed\\
\end{tabular}
\medskip
Note.--- Errors are quoted at $90 \%$ confidence level. The absorption column density was fixed to the \citet{dl90} value and given in units of cm$^{-2}$. The power law index was bound at an upper limit of 4.0. The {\sc reflionx} model used has been modified for a blackbody illuminating the accretion disk. The blackbody temperature was tied to the temperature of the blackbody used to model the continuum emission. The iron abundance, $A_{Fe}$, has a hard lower limit of 0.5. Flux is given in units of $10^{-9}$ ergs cm$^{-2}$ s$^{-1}$. Luminosity is calculated at a distance of 8.5 kpc and given in units of $10^{37}$ ergs s$^{-1}$. For reference, 1 ISCO $= 6\ R_{g}$ for $a_{*}=0$.
\end{center}
\end{table}
We find the iron abundance to be $A_{Fe}=0.5-1.6$. The low abundance for this source is likely due to its association with a globular cluster. Globular clusters tend to host older populations of stars and therefore have a lower metallicity. \citet{Lee06} find the metallicity in NGC 6293 is $\sim 1/100$ of solar abundance. We were unable to explore a lower iron abundance due to the hard lower limit of the model ($A_{Fe}=0.5$), however, anomalously high iron abundances have been seen in many reflection studies (e.g., \citealt{parker15}, \citeyear{parker16}; \citealt{fuerst16}; \citealt{garcia15}; \citealt{walton14}, \citeyear{walton16}). It is possible that this high $A_{Fe}$ measurement correctly describes the atmosphere of the accretion disk and not the overall abundances within the accretion flow due to the ionization structure skewing the relative abundances there. Furthermore, the overabundance found in our fits may be the result of effects from dense gas in the disk that is not accounted for by current models. This would cause the abundance to increase to replicate the continuum for a lower density disk that is allowed by the atomic data set within current reflection models (see \citealt{garcia16} for more detail).
Figure 4 shows a contour plot for the inner disk radius versus the inclination for the fits in Table 1. The unabsorbed 0.5-50.0 keV flux changes from $2.0\times10^{-9}$ ergs cm$^{-2}$ s$^{-1}$ to $2.6\times10^{-9}$ ergs cm$^{-2}$ s$^{-1}$. At a distance of 8.5 kpc, this gives a luminosity change of $1.7\times10^{37}$ ergs s$^{-1}$ to $2.2\times10^{37}$ ergs s$^{-1}$. In other words, this is a change from 0.04 L$_{\mathrm{Edd}}$ to 0.06 L$_{\mathrm{Edd}}$ ($L_{\mathrm{Edd}}=3.8\times10^{38}$ ergs s$^{-1}$; \citealt{kuulkers03}). We check that our results are not dependent on the our choice of $q=3$ by fixing the index to $q=2$ \& $q=2.5$ and find they are consistent with the inclination and $R_{in}$ at the $3\sigma$ level in each case.
\begin{figure}
\centering
\includegraphics[angle=270,width=8.2cm]{lowspeccomps.eps}
\includegraphics[angle=270,width=8.2cm]{spechigh.eps}
\caption{XTE J1709-267 low (top) and high (bottom) spectrum fit when from 3.0-30.0 keV with a {\sc diskbb} (red dash line), {\sc blackbody} (purple dot dot dot dash line), and the modified version of {\sc reflionx} (blue dot dash line) that assumes an input blackbody spectrum. For the low flux spectrum, an additional power-law (orange dot line) component was needed. The panel below shows the ratio of the data to the model. Table 1 lists parameter values. The data were rebinned for plotting purposes.
}
\label{fig:feline}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8.2cm]{contours_lowhigh_rg_2.eps}
\caption{Contour plot for inner disk radius versus inclination at the 68\%, 90\%, and 99\% confidence level for the low (black) and high (red) flux portion of the observations. The best fit values are denoted with a cross for the low flux and circle for the high flux.
}
\label{fig:contour}
\end{figure}
\section{Discussion}
We have performed spectral analysis of the soft X-ray transient XTE J1709-267 during its 2016 outburst. This is the first report and analysis of a broad Fe-K$_{\alpha}$ line in this source. Our observations capture a change in flux emitted from the source which we separate into low and high flux spectra. We find the disk to be truncated at $13.8_{-1.8}^{+3.0}\ R_{g}$ (90\% confidence level) in the lower flux spectrum. The disk appears to move outwards to $23.4_{-5.4}^{+15.6}\ R_{g}$ during the higher flux state, but is consistent with the low flux results at the 3$\sigma$ level. The unabsorbed $0.5-50.0$ keV flux at the time of the observation started at $2.0\times10^{-9}$ ergs cm$^{-2}$ s$^{-1}$ and increased to $2.6\times10^{-9}$ ergs cm$^{-2}$ s$^{-1}$, which is consistent with the typical flux observed during outburst (\citealt{marshall97}; \citealt{cocchi98}; \citealt{jonker03}, \citeyear{jonker04a}, \citeyear{jonker04b}; \citealt{atel255}; \citealt{atel1302}; \citealt{degenaar13}). Additionally, we find a low inclination of $22^{\circ}-39^{\circ}$. There are no previous estimates of the inclination for this system. The disk is likely truncated by a boundary layer surrounding the NS or the magnetosphere.
\citet{PS01} lay out the Newtonian framework for boundary layer behavior for different mass accretion rates. We estimate the mass accretion rate for XTE J1709-267 to be $1.5\times10^{-9}$ M$_{\odot}$ yr$^{-1}$ for the lower flux portion and $1.9\times10^{-9}$ M$_{\odot}$ yr$^{-1}$ for the higher flux from the 0.5-50.0 keV unabsorbed luminosity and using an efficiency of $\eta=0.2$ \citep{sibsun00}. Using Equation (25) in \citet{PS01}, we estimate that the boundary layer extends from the surface of the NS out to $\sim0.9-1.1\ R_{g}$ (assuming 1.4 M$_{\odot}$). Additional factors, such as spin and viscosity in the layer, can extend this region to be consistent with the inner edge of the accretion disk that we measured from our reflection analysis in the low flux. However, this would not be consistent with the larger inner disk radius from the high flux spectrum.
XTE J1709-267 was at a relatively low Eddington fraction ($\sim0.04-0.06$) during the time of the observation. The truncation of the disk could be due to a pressure balance between the magnetic field and accretion in the disk. We can place an upper limit on the strength of the field using the upper limit of $R_{in}=16.8\ R_{g}$ from the low flux spectrum. Assuming a mass of 1.4 M$_{\odot}$, taking the distance to be 8.5 kpc, and using the unabsorbed flux from 0.5-50.0 keV of $2.0\times10^{-9}$ erg cm$^{-2}$ s$^{-1}$ as the bolometric flux, we can determine the magnetic dipole moment, $\mu$, from Equation (1) taken from \citet{cackett09}, which was adapted from \citet{IP09}. If we assume an angular anisotropy, $f_{ang}$, and conversion factor, $k_{A}$, of unity, as well as an accretion efficiency of $\eta=0.2$, then $\mu\simeq3.7\times10^{26}$ G cm$^{3}$. This corresponds to a magnetic field strength of $B\simeq7.5\times10^{8}$ G at the magnetic poles for a NS of 10 km. The magnetic field strength at the pole is twice as strong as at the equator. This is within the range for magnetic field strength for accreting millisecond pulsars ($10^{7}-10^{9}$ G; \citealt{mukherjee15}). The high flux solution gives a maximum magnetic field strength in excess of $B>10^{9}$ G ($3.7\times10^{9}$ G). However, no X-ray pulsations have been detected from this source, so there are no indications that the magnetic field has truncated the disk and channeled material to the magnetic poles.
Recent analyses of a similar nature have been done for $\emph{NuSTAR}$ studies of the transient NS LMXBs 1RXS J180408.9-34205 (RXS J1804) and Aquila X-1 (Aql X-1). \citet{ludlam16} found the inner disk of RXS J1804 to be $R_{in}\leq11.1\ R_{g}$ in the hard state. They find that a magnetic field strength of $B\leq0.3-1.0\times10^{9}$ G at the poles or a boundary layer that is roughly the stellar radius in size is needed to truncate the disk at $11.1\ R_{g}$. \citet{degenaar16} found similar estimates of the magnetic field strength ($B\leq2\times10^{8}$ G) and inner disk radius ($R_{in}\leq1.5$ ISCO) while RXS J1804 was in the soft state. \citet{king16} found a truncated disk at $R_{in}=15\pm3\ R_{g}$ around Aql X-1. They estimate a boundary layer of $R_{B}=7.8\ R_{g}$ \citep{king16} would be surrounding Aql X-1 given the efficiency and mass accretion rate. But if the disk was not truncated by a boundary layer and instead by the magnetosphere, they obtain an upper limit on the magnetic field of $B<5\pm2\times10^{8}$ G \citep{king16}. Both RXS J1804 and Aql X-1 had at least one Type 1 X-ray burst during their observation, suggesting that material was still reaching the surface.
\section{Summary}
Using $\emph{NuSTAR}$, we perform the first reflection study of the soft X-ray transient XTE J1709-267 during its 2016 June outburst. Our observations catch the source during a change in luminosity from $0.04-0.06$ L$_{\mathrm{Edd}}$. We find the disk is truncated prior to the NS surface at a distance of $13.8_{-1.8}^{+3.0}\ R_{g}$ at 0.04 L$_{\mathrm{Edd}}$ and increases out to $23.4_{-5.4}^{+15.6}\ R_{g}$ at 0.06 L$_{\mathrm{Edd}}$. The disk is likely truncated by a boundary layer surrounding the NS. We estimate that the boundary layer extends from the surface out to $\sim0.9-1.1\ R_{g}$ for the mass accretion rate and efficiency of the disk. However, though viscosity and spin effects can increase the extent of the boundary layer, at low Eddington fraction the boundary layer is not likely to halt the accretion disk at a large radius (i.e., at the large radius implied by the high flux solution). An alternative explanation is that the disk is truncated by the magnetosphere. Conservative estimates place an upper limit on the magnetic field strength to be $B\leq0.75-3.70\times10^{9}$ G at the magnetic poles, though XTE J1709-267 is not a known X-ray pulsar.
\\
\\
\\
We thank the referee for their thoughtful comments that have led to the improvement of this work. RML would like to thank Fiona Harrison for approval of the DDT that made this work possible.
This research has made use of the NuSTAR Data Analysis Software (NuSTARDAS) jointly developed by the ASI Science Data Center (ASDC, Italy) and the California Institute of Technology (Caltech, USA). ND is supported by an NWO Vidi grant and a Marie Curie Intra-European fellowship. EMC gratefully acknowledges support from the National Science Foundation through CAREER award number AST-1351222.
|
\section{Introduction}
Throughout this chapter, by $\mathbb{F}$ we refer to any arbitrary field, while by $\mathbb{F}_p$, we only refer to the fields of prime order $p$. We denote the set of non-zero elements by $\mathbb{F}^*$ and $\mathbb{F}^*_p$, respectively. Furthermore, we use the following convention: if the characteristic of $\mathbb{F}$ is positive, then we denote its characteristic by $p$; if the characteristic of $\mathbb{F}$ is zero, then we set $p=\infty$. So a term like $N <p^{5/8}$ is restrictive in positive characteristic, but vacuous for zero one.
For $A\subset \mathbb{F}$, the sum and the product sets are defined as follows:
\[ A+A=\{a+a' : a, a' \in A\}, ~A\cdot A=\{a\cdot a' : a, a' \in A\}.\]
For $A\subset \mathbb{F}_p$, Bourgain, Katz and Tao (\cite{bkt}) proved that if $ p^\delta < |A| < p^{1-\delta}$ for some $\delta > 0$, then we have
\[ \max \left\lbrace |A+A|, |A\cdot A|\right\rbrace \gg |A|^{1+\epsilon},\]
for some $\epsilon = \epsilon(\delta) > 0$. Here, and throughout, by $X \ll Y$ we mean that there exists the constant $C>0$ such that $X\le CY$.
In a breakthrough paper \cite{RRS}, Roche-Newton, Rudnev, and Shkredov improved and generalized this result to arbitrary fields. More precisely, they showed that for $A\subset \mathbb{F}$, the sum set and the product set satisfy
\[\max\left\lbrace |A\pm A|, |A\cdot A|\right\rbrace \gg |A|^{6/5}, ~\max\left\lbrace |A\pm A|, |A : A|\right\rbrace \gg |A|^{6/5}.\]
We note that the same bound also holds for $|A(1+A)|$ \cite{SZ}, and $|A+A^2|$, $\max\left\lbrace |A+A|, |A^2+A^2|\right\rbrace$ \cite{P1}. We refer the reader to \cite{AMRS, BT, RRS, M} and references therein for recent results on the sum-product topic.
Let $G$ be a subgroup of $\mathbb{F}^*$, and $g\colon G\to \mathbb{F}^*$ be an arbitrary function. We define
\[\mu(g)=\max_{t\in \mathbb{F}^*}\left\vert \left\lbrace x\in G\colon g(x)=t\right\rbrace \right\vert.\]
For $A, B\subset \mathbb{F}_p$ and two-variable functions $f(x,y)$ and $g(x,y)$ in $\mathbb{F}_p[x, y]$, Hegyv\'{a}ri and Hennecart \cite{heg}, using graph theoretic techniques, proved that if $|A|=|B|=p^\alpha$, then \[\max\left\lbrace |f(A,B)|, |g(A,B)|\right\rbrace\gg |A|^{1+\Delta(\alpha)},\] for some $\Delta(\alpha)>0$. More precisely, they established the following results.
\begin{theorem}[\textbf{Hegyv\'{a}ri and Hennecart}, \cite{heg}]\label{thm1}
Let $G$ be a subgroup of $\mathbb{F}_p^*$. Consider the function $f(x,y)=g(x)(h(x)+y)$ on $G\times \mathbb{F}_p^*$, where $g,h\colon G\to \mathbb{F}_p^*$ are arbitrary functions. Define $m=\mu(g\cdot h)$. For any subsets $A\subset G$ and $B,C\subset \mathbb{F}_p^*$, we have
\[\left\vert f(A,B)\right\vert \left\vert B\cdot C\right\vert\gg \min\left\lbrace\frac{|A||B|^2|C|}{pm^2}, \frac{p|B|}{m}\right\rbrace.\]
\end{theorem}
\begin{theorem}[\textbf{Hegyv\'{a}ri and Hennecart}, \cite{heg}]\label{2.2}
Let $G$ be a subgroup of $\mathbb{F}_p^*$. Consider the function $f(x,y)=g(x)(h(x)+y)$ on $G\times \mathbb{F}_p^*$, where $g,h\colon G\to \mathbb{F}_p^*$ are arbitrary functions. Define $m=\mu(g)$. For any subsets $A\subset G$, $B,C \subset \mathbb{F}_p^*$, we have
\[|f(A,B)||B+C|\gg \min\left\lbrace \frac{|A||B|^2|C|}{pm^2}, \frac{p|B|}{m}\right\rbrace\]
\end{theorem}
It is worth noting that Theorem $6$ established by Bukh and Tsimerman \cite{BT} does not cover such a function defined in Theorem \ref{2.2}. The reader can also find the generalizations of Theorems \ref{thm1} and \ref{2.2} in the setting of finite valuation rings in \cite{Ham}.
Suppose $f(x,y)=g(x)(h(x)+y)$ with $\mu(g), \mu(h)=O(1)$ and $A=B=C$. Then, it follows from Theorems \ref{thm1} and \ref{2.2} that
\begin{enumerate}
\item If $|A|\gg p^{2/3}$, then we have
\[|f(A, A)||A\cdot A|,|f(A, A)||A+A|\gg p|A|.\]
\item If $|A|\ll p^{2/3}$, then we have
\begin{equation}\label{xxx1}|f(A, A)||A\cdot A|,|f(A, A)||A+A|\gg |A|^4/p.\end{equation}
\end{enumerate}
The main goal of this paper is to improve and generalize Theorems \ref{thm1} and \ref{2.2} to arbitrary fields for small sets. Our first result is an improvement of Theorem \ref{thm1}.
\begin{theorem}\label{thm1*}
Let $f(x,y)=g(x)(h(x)+y)$ be a function defined on $\mathbb{F}^*\times \mathbb{F}^*$, where $g,h\colon \mathbb{F}^*\to \mathbb{F}^*$ are arbitrary functions. Define $m=\mu(g\cdot h)$. For any subsets $A, B,C\subset \mathbb{F}^*$ with $|A|, |B|, |C|\le p^{5/8}$, we have
\[\max\left\lbrace |f(A, B)|, |B\cdot C|\right\rbrace\gg \min\left\lbrace\frac{|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}}{m^{\frac{4}{5}}}, \frac{|B||C|^{\frac{1}{2}}}{m}, \frac{|B||A|^{\frac{1}{2}}}{m}, \frac{|B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}}{m^{\frac{2}{3}}}\right\rbrace.\]
\end{theorem}
The following are consequences of Theorem \ref{thm1*}.
\begin{corollary}\label{co1}
Let $f(x,y)=g(x)(h(x)+y)$ be a function defined on $\mathbb{F}^*\times \mathbb{F}^*$, where $g,h\colon \mathbb{F}^*\to \mathbb{F}^*$ are arbitrary functions with $\mu(g\cdot h)=O(1)$. For any subset $A\subset \mathbb{F}^*$ with $|A|\le p^{5/8}$, we have
\[\max\left\lbrace |f(A, A)|, |A\cdot A|\right\rbrace\gg |A|^{\frac{6}{5}}.\]
\end{corollary}
\begin{corollary}\label{corx}
Consider the subsets $A\subset \mathbb{F}^{*}$, and $ B, C\subset \mathbb{F}$ with $|A|, |B|, |C|\le p^{5/8}$.
\begin{enumerate}
\item By fixing $g(x)=1$ and $h(x)=x^{-1}$, we get
\[\max\left\lbrace |A^{-1}+B|, |B\cdot C|\right\rbrace\gg \min\left\lbrace|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}, |B||C|^{\frac{1}{2}}, |B||A|^{\frac{1}{2}}, |B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}\right\rbrace.\]
\item By fixing $g(x)=x$ and $h(x)=1$, we have
\[\max\left\lbrace |A(B+1)|, |B\cdot C|\right\rbrace\gg \min\left\lbrace|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}, |B||C|^{\frac{1}{2}}, |B||A|^{\frac{1}{2}}, |B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}\right\rbrace.\]
\end{enumerate}
\end{corollary}
It follows from Corollary \ref{corx}(2) that if $B=A$ and $C=A+1$ then we have $|A(A+1)|\gg |A|^{6/5}$, which recovers the result of Stevens and de Zeeuw \cite{SZ}.
Our next result is the additive version of Theorem \ref{thm1*}, which improves Theorem \ref{2.2}.
\begin{theorem}\label{thm2*}
Let $f(x,y)=g(x)(h(x)+y)$ be a function defined on $\mathbb{F}^*\times \mathbb{F}^*$, where $g\colon \mathbb{F}^*\to \mathbb{F}^*$ are arbitrary functions. Define $m=\mu(g)$. For any subsets $A, B,C\subset \mathbb{F}^*$ with $|A|, |B|, |C|\le p^{5/8}$, we have
\[\max\left\lbrace |f(A, B)|, |B+ C|\right\rbrace\gg \min\left\lbrace\frac{|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}}{m^{\frac{4}{5}}}, \frac{|B||C|^{\frac{1}{2}}}{m}, \frac{|B||A|^{\frac{1}{2}}}{m}, \frac{|B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}}{m^{\frac{2}{3}}}\right\rbrace.\]
\end{theorem}
\begin{corollary}\label{co2}
Let $f(x,y)=g(x)(h(x)+y)$ be a function defined on $\mathbb{F}^*\times \mathbb{F}^*$, where $g\colon \mathbb{F}^*\to \mathbb{F}^*$ are arbitrary functions with $\mu(g)=O(1)$. For any subset $A\subset \mathbb{F}^*$ with $|A|\le p^{5/8}$, we have
\[\max\left\lbrace |f(A, A)|, |A+ A|\right\rbrace\gg |A|^{\frac{6}{5}}.\]
\end{corollary}
Let $g(x)=x$ and $h(x)=1$, we have the following corollary.
\begin{corollary}
For $A, B, C\subset \mathbb{F}$ with $|A|, |B|, |C|\le p^{5/8}$, we have
\[\max\left\lbrace |A(B+1)|, |B+ C|\right\rbrace\gg \min\left\lbrace|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}, |B||C|^{\frac{1}{2}}, |B||A|^{\frac{1}{2}}, |B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}\right\rbrace.\]
\end{corollary}
By fixing $g(x)=x$ and $h(x)=0$, we have the following result.
\begin{corollary}
For $A, B, C\subset \mathbb{F}$ with $|A|, |B|, |C|\ll p^{5/8}$, we have
\[\max\left\lbrace |A\cdot B|, |B+C|\right\rbrace \gg \min\left\lbrace|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}, |B||C|^{\frac{1}{2}}, |B||A|^{\frac{1}{2}}, |B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}\right\rbrace\]
\end{corollary}
In the case $A=B=C$, we recover the following result due to Roche-Newton, Rudnev, and Shkredov \cite{RRS}, which says that $\max\left\lbrace|A+A|, |A\cdot A|\right\rbrace\gg |A|^{6/5}$.
It has been shown in \cite{SZ} that if $f(x,y)=x(x+y)$, then $|f(A, A)|\gg |A|^{5/4}$ under the condtion $|A|\le p^{2/3}$. In the following theorem, we show that if either $|A+A|$ or $|A\cdot A|$ is sufficiently small, the exponent $5/4$ can be improved from the polynomials to a larger family of functions on $\mathbb{F}^*\times \mathbb{F}^*$
\begin{theorem}\label{thmx}
Let $f(x,y)=g(x)(h(x)+y)$ be a function defined on $\mathbb{F}^*\times \mathbb{F}^*$, where $g,h\colon \mathbb{F}^*\to \mathbb{F}^*$ are arbitrary functions with $\mu(f), \mu(g)=O(1)$. Consider the subset $A\subset \mathbb{F}^*$ with $|A|\le p^{5/8}$, satisfying
\[\min\left\lbrace|A+A|, |A\cdot A|\right\rbrace\le |A|^{\frac{9}{8}-\epsilon}\]
for some $\epsilon>0$. Then, we have
\[|f(A, A)|\gg |A|^{\frac{5}{4}+\frac{2\epsilon}{3}}.\]
\end{theorem}
\section{Proofs of Theorems \ref{thm1*}, \ref{thm2*}, and \ref{thmx}}
Let $\mathcal{R}$ be a set of points in $\mathbb{F}^3$ and $\mathcal{S}$ be a set of planes in $\mathbb{F}^3$. We write $\mathcal{I}(\mathcal{R},\mathcal{S}) = |\{(r,s)\in \mathcal{R}\times \mathcal{S} : r\in s\}|$ for the number of \emph{incidences} between $\mathcal{R}$ and $\mathcal{S}$. To prove Theorems \ref{thm1*} and \ref{thm2*}, we make use of the following point-plane incidence bound due to Rudnev \cite{R}. A short proof can be found in \cite{Z}.
\begin{theorem}[\textbf{Rudnev}, \cite{R}]\label{thm:rudnev}
Let $\mathcal{R}$ be a set of points in $\mathbb{F}^3$ and let $\mathcal{S}$ be a set of planes in $\mathbb{F}^3$, with $|\mathcal{R}|\ll |\mathcal{S}|$ and $|\mathcal{R}|\ll p^2$.
Assume that there is no line containing $k$ points of $\mathcal{R}$.
Then
\[ \mathcal{I}(\mathcal{R},\mathcal{S})\ll |\mathcal{R}|^{1/2}|\mathcal{S}| +k|\mathcal{S}|.\]
\end{theorem}
\paragraph{Proof of Theorem \ref{thm1*}:}
Define $f(A, B)=\{f(a, b)\colon a\in A, b\in B\}, g(A)=\{g(a)\colon a\in A\}$, $h(A)=\{h(a)\colon a\in A\}$.
For $\lambda\in B\cdot C$, let
\[E_\lambda=\left\vert \left\lbrace \left(f(a, b), c\cdot g(a)^{-1}, c\cdot h(a)\right)\colon (a, b, c)\in A\times B\times C, ~f(a, b)\cdot c\cdot g(a)^{-1}-c\cdot h(a)=\lambda\right\rbrace\right\vert,\]
where by $g(a)^{-1}$ we mean the multiplicative inverse of $g(a)$ in $\mathbb{F}^*$. For a given triple $(x, y, z)\in (\mathbb{F}^*)^3$, we count the number of solutions $(a,b,c)\in A\times B \times C$ to the following system
\[g(a)(h(a)+b)=x, ~c\cdot g(a)^{-1}=y, ~c\cdot h(a)=z.\]
This implies that
\[ g(a)h(a)=zy^{-1}.\]
Since $\mu(g\cdot h)=m$, there are at most $m$ different values of $a$ satisfying the equation $g(a)h(a)=zy^{-1}$, and $b, c$ are uniquely determined in term of $a$ by the first and second equations of the system. This implies that
\[ |A||B||C|/m\le \sum_{\lambda\in B\cdot C}E_\lambda .\]
By the Cauchy-Schwarz inequality, we get
\begin{equation}\label{cn1} \left(|A||B||C|/m\right)^2 \le \left(\sum_{\lambda\in B\cdot C}E_{\lambda}\right)^2\le E\cdot |B\cdot C| ,\end{equation}
where $E=\sum_{\lambda\in B\cdot C}E_{\lambda}^2$.
Define the point set $\mathcal{R}$ as
\[ \mathcal{R} = \left\lbrace\left(c\cdot g(a)^{-1}, c\cdot h(a), g(a')(h(a')+b')\right): a, a'\in A, b'\in B, c\in C\right\rbrace\]
and the set of planes $\mathcal{S}$ as
\[ \mathcal{S} = \left\lbrace g(a)(h(a)+b)X - Y -c'g(a')^{-1}Z = -c'\cdot h(a')\colon a, a'\in A, b\in B, c'\in C\right\rbrace.\]
We have $E \le I(\mathcal{R}, \mathcal{S})$, and $|\mathcal{R}|=|\mathcal{S}|\le |f(A, B)||A||C|$. To apply Theorem \ref{thm:rudnev}, we need to find an upper bound on $k$ which is the maximum number of collinear points in $\mathcal{R}$. The projection of $\mathcal{R}$ into the first two coordinates is the set $\mathcal{T}=\{(c\cdot g(a)^{-1}, c\cdot h(a))\colon a\in A, c\in C\}$. The set $\mathcal{T}$ can be covered by the lines of the form $y=g(a)h(a)x$ with $a\in A$. This implies that $\mathcal{T}$ can be covered by at most $|A|$ lines passing through the origin, with each line containing $|C|$ points of $\mathcal{T}$. Therefore, a line in $\mathbb{F}^3$ contains at most $\max\{|A|,|C|\}$ points of $\mathcal{R}$, unless it is vertical, in which case it contains at most $|f(A, B)|$ points. In other words, we get
\[k\le \max\{|A|, |C|, |f(A, B)|\}.\]
If $|\mathcal{R}|\gg p^2$, then we get $|f(A, B)||A||C|\gg p^2$. Since $|A|, |C|\le p^{5/8}$, we have $|f(A, B)|\gg p^{3/4}\gg |A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}$, and we are done in this case. Thus, we can assume that $|\mathcal{R}|\ll p^2$. Applying Theorem \ref{thm:rudnev}, we obtain
\begin{equation}\label{cn2}I(\mathcal{R}, \mathcal{S})\le |f(A, B)|^{3/2}|A|^{3/2}|C|^{3/2}+k|f(A, B)|A||C|.\end{equation}
Putting (\ref{cn1}) and (\ref{cn2}) together gives us
\[\max\left\lbrace |f(A, B)|, |B\cdot C|\right\rbrace\gg \min\left\lbrace\frac{|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}}{m^{\frac{4}{5}}}, \frac{|B||C|^{\frac{1}{2}}}{m}, \frac{|B||A|^{\frac{1}{2}}}{m}, \frac{|B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}}{m^{\frac{2}{3}}}\right\rbrace.\]
This completes the proof of the theorem. $\square$
\paragraph{Proof of Theorem \ref{thm2*}:}
The proof goes in the same direction as Theorem \ref{thm1*}, but for the sake of completeness, we include the detailed proof. For $\lambda\in B+ C$, let
\[E_\lambda=\left\vert \left\lbrace \left(f(a, b), g(a)^{-1}, c-h(a)\right)\colon (a, b, c)\in A\times B\times C, ~f(a, b)\cdot g(a)^{-1}+(c-h(a))=\lambda\right\rbrace\right\vert.\]
For a given triple $(x, y, z)\in (\mathbb{F}^*)^3$, we count the number of solutions $(a,b,c)\in A\times B \times C$ to the following system
\[g(a)(h(a)+b)=x, ~g(a)^{-1}=y, ~c-h(a)=z.\]
Since $\mu(g)=m$, there are at most $m$ different values of $a$ satisfying the equation $g(a)=y^{-1}$, and $b, c$ are uniquely determined in term of $a$ by the first and third equations of the system.
This implies that
\[|A||B||C|/m \le\sum_{\lambda\in B+C}E_\lambda .\]
By the Cauchy-Schwarz inequality, we have
\begin{equation}\label{cn21} \left(|A||B||C|/m\right)^2 \le \left(\sum_{\lambda\in B+C}E_{\lambda}\right)^2\le E\cdot |B+C| ,\end{equation}
where $E=\sum_{\lambda\in B+C}E_{\lambda}^2$.
Define the point set $\mathcal{R}$ as
\[ \mathcal{R} = \left\lbrace\left(g(a)^{-1}, c-h(a), g(a')(h(a')+b')\right): a, a'\in A, b'\in B, c\in C\right\rbrace,\]
and the collection of planes $\mathcal{S}$ as
\[ \mathcal{S} = \left\lbrace g(a)(h(a)+b)X +Y -g(a')^{-1}Z = c'-h(a')\colon a, a'\in A, b\in B, c'\in C\right\rbrace.\]
It is clear that $|\mathcal{R}|=|\mathcal{S}|\le |f(A, B)||A||C|$, and $E\le I(\mathcal{R}, \mathcal{S})$. To apply Theorem \ref{thm:rudnev}, we need to find an upper bound on $k$ which is the maximum number of collinear points in $\mathcal{R}$. The projection of $\mathcal{R}$ into the first two coordinates is the set $\mathcal{T}=\{\left(g(a)^{-1}, c-h(a)\right)\colon a\in A, c\in C\}$. The set $\mathcal{T}$ can be covered by at most $|A|$ lines of the form $x=g(a)^{-1}$ with $a\in A$, where each line contains $|C|$ points of $\mathcal{T}$. Therefore, a line in $\mathbb{F}^3$ contains at most $\max\{|A|,|C|\}$ points of $\mathcal{R}$, unless it is vertical, in which case it contains at most $|f(A, B)|$ points. So we get
\[k\le \max\{|A|, |C|, |f(A, B)|\}.\]
If $|\mathcal{R}|\gg p^2$, this implies that $|f(A, B)||A||C|\gg p^2$. Since $|A|, |C|\le p^{5/8}$, we have $|f(A, B)|\gg p^{3/4}\gg |A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}$, and we are done. Thus, we can assume that $|\mathcal{R}|\ll p^2$. Applying Theorem \ref{thm:rudnev}, we obtain
\begin{equation}\label{cn22}I(\mathcal{R}, \mathcal{S})\le |f(A, B)|^{3/2}|A|^{3/2}|C|^{3/2}+k|f(A, B)|A||C|.\end{equation}
Putting (\ref{cn21}) and (\ref{cn22}) together gives us
\[\max\left\lbrace |f(A, B)|, |B+C|\right\rbrace\gg \min\left\lbrace\frac{|A|^{\frac{1}{5}}|B|^{\frac{4}{5}}|C|^{\frac{1}{5}}}{m^{\frac{4}{5}}}, \frac{|B||C|^{\frac{1}{2}}}{m}, \frac{|B||A|^{\frac{1}{2}}}{m}, \frac{|B|^{\frac{2}{3}}|C|^{\frac{1}{3}}|A|^{\frac{1}{3}}}{m^{\frac{2}{3}}}\right\rbrace.\]
This completes the proof. $\square$
\paragraph{Proof of Theorem \ref{thmx}:}
One can assume that $|f(A,A)| \le |A|^2$, since otherwise we are done. Now by the proofs of Theorems \ref{thm1*} and \ref{thm2*} for $A\subset \mathbb{F}^*$ with $|A|\le p^{5/8}$, we have
\[|f(A, A)|^{3/2}|A\cdot A|\gg |A|^{3}, ~|f(A, A)|^{3/2}|A+ A|\gg |A|^{3}.\]
Since $\min\left\lbrace |A+A|, |A\cdot A|\right\rbrace\le |A|^{\frac{9}{8}-\epsilon}$, we get $|f(A, A)|^{3/2}\gg |A|^{3-\frac{9}{8}+\epsilon}$, which concludes the proof of the theorem.
$\square$
\section{Acknowledgments}
The authors were partially supported by Swiss National Science Foundation grants 200020-162884 and 200021-175977.
|
\section{Introduction}
\label{sec:intro}
Recent observations of distinct cosmological probes are closing in on the few parameters that enter the standard model of cosmology \citep[see for example][and references therein]{2016A&A...594A..13P}.
Although there is clear evidence that the Universe is well described by the $\Lambda$CDM model,
some tensions are found between probes. For instance, the best fit cosmology
inferred from the observation of the Cosmic Microwave Background (CMB) in \citet{2016A&A...594A..13P} is in tension with some cosmic shear analyses \citep[]{2015MNRAS.451.2877M, KiDS450, 2016arXiv161004606J, 2017MNRAS.465.2033J},
while both direct and strong lensing measurements of today's Hubble parameter $H_0$ are more than 3$\sigma$ away from the values inferred from the CMB \citep{2016JCAP...10..019B, HOLICOW}.
{\color{black} At face value, these discrepancies} either point towards new physics {\color{black}\citep[for a recent example, see][]{2016arXiv161004606J}} or un-modelled systematics in any of those probes.
In this context, cross-correlation of different cosmic probes stands out as a unique tool, as many residual systematics that could contaminate
one data set are unlikely to correlate also with the other (e.g. `additive biases'). This type of measurement can therefore be exempt from un-modelled biases that might otherwise source the tension.
Another point of interest is that the systematic effects that do not fully cancel, for example `multiplicative biases' or the uncertainty on the photometric redshifts,
will often impact differently the cosmological parameters compared to the stand-alone probe, allowing for degeneracy breaking or improved calibration.
In this paper, we present the first tomographic cross-correlation measurement between CMB lensing and galaxy lensing,
based on the lensing map described in \citet{2015arXiv150201591P} and the lensing data from the Kilo Degree Survey\footnote{KiDS: {\tt http://kids.strw.leidenuniv.nl}} presented in \citet[][KiDS hereafter]{KiDSLenS} and in the KiDS-450 cosmic shear analysis \citep{KiDS450}.
The main advantage in this sort of measurement resides in it being free of
uncertainty on galaxy bias, which otherwise dominates the error budget in CMB lensing - galaxy position cross-correlations \citep{2015arXiv150203405O, 2015arXiv150705551G, 2016arXiv160207384B}.
Over the last two years, the first lensing-lensing cross-correlations were used to measure $\sigma_8$ and $\Omega_{\rm m}$ \citep{2015PhRvD..91f2001H, 2015PhRvD..92f3517L},
by combining the CMB lensing data from the Atacama Cosmology Telescope \citep{2014JCAP...04..014D} with the lensing data from the Canada-France-Hawaii Telescope Stripe 82 Survey \citep{2014RMxAC..44..202M}, and from the {\it Planck} lensing data and the Canada-France-Hawaii Telescope Lensing Survey \citep[][CFHTLenS hereafter]{2013MNRAS.433.2545E}.
Since then, additional effects were found to contribute to the measurement, introducing extra complications in the interpretation of the signal.
For instance, \citet{2014MNRAS.443L.119H} and \citet{TroxelIshak14} showed that the measurement is likely to be contaminated
by the intrinsic alignment of galaxies with the tidal field in which they live.
{\color{black} At the same time, \citet{2016PhRvD..93j3508L} argued that this measurement could point instead to residual systematics in the multiplicative shear bias
and proposed that the measurement itself could be used to set constraints on the shear bias \citep[see also][]{2013arXiv1311.2338D}.
Their results showed that large residuals are favoured, despite the calibration accuracy claimed by the analysis of image simulations tailored for the same survey} \citep{2013MNRAS.429.2858M}.
A recent analysis from \citet[][HD16 hereafter]{2016MNRAS.460..434H} suggested instead that the impact of catastrophic redshift outliers could be causing this apparent discrepancy,
since these dominate the uncertainty in the modelling. They also showed that choices concerning the treatment of the masks can lead to biases in the measured signal,
and that the current estimators should therefore be thoroughly calibrated on full light-cone mocks.
Although these pioneering works were based on Fourier space cross-correlation techniques, more recent analyses presented results from configuration-space measurements,
which are cleaner due to their insensitivity to masking.
\citet[][K16 hereafter]{Kirk15} combined the CMB lensing maps from {\it Planck} and from the South Pole Telescope \citep[][SPT]{2012ApJ...756..142V} with
the Science Verification Data from the Dark Energy Survey\footnote{DES: {\tt www.darkenergysurvey.org}}.
Their measurement employed the {\small POLSPICE} numerical tool \citep{spice, polspice},
which starts off with a pseudo-$C_{\ell}$ measurement that is converted into configuration space to deal with masks, then turned back into a Fourier space estimator.
Soon after, HD16 showed consistency between pseudo-$C_{\ell}$ analyses and {\color{black} configuration space analyses of two-point correlation functions}, combining the {\it Planck} lensing maps with both CFHTLenS and the Red-sequence Cluster Lensing Survey \citep[][RCSLenS hereafter]{RCSLenS}.
A similar configuration space estimator was recently used with {\it Planck} lensing and SDSS shear data \citep{2017MNRAS.464.2120S},
although the signal was subject to higher noise levels.
This paper directly builds on the K16 and HD16 analyses, utilising tools and methods described therein, but on a new suite of lensing data.
The additional novelty here is that we perform the first tomographic CMB lensing -- galaxy lensing cross-correlation analysis, where we split the galaxy sample into 5 redshift bins and
examine the redshift evolution. This is made possible by the high quality of the KiDS photometric redshift data, {\color{black} by the extend of the spectroscopic matched sample,
and consequently by the precision achieved on the calibrated source redshift distribution \citep[see][for more details]{KiDS450}}.
It provides a new test of cosmology within the $\Lambda$CDM model, {\color{black} including} the redshift evolution of the growth of structure,
and also offers an opportunity to examine the tension between the KiDS and {\it Planck} cosmologies \citep[reported in][]{KiDS450}.
With the upcoming lensing surveys such as LSST\footnote{\tt www.lsst.org} and Euclid\footnote{\tt sci.esa.int/euclid},
it is forecasted that this type of cross-correlation analysis will be increasingly used to {\color{black} validate the data calibration} \citep{2016arXiv160701761S}
and extract cosmological information in a manner that complements the cosmic shear and clustering data.
The basic theoretical background upon which we base our work is laid out in Section \ref{sec:th}.
We then describe the data sets and our measurement strategies in Sections \ref{sec:data} and \ref{sec:measurement}, respectively.
Our cosmological results are presented in Section \ref{sec:results}. {\color{black} We also describe therein a calibration analysis along the lines of \citet{2016PhRvD..93j3508L},
this time focussing on high redshift galaxies for which the photometric redshifts and shear calibration are not well measured.
Informed on cosmology from lower redshift measurement, this self-calibration technique has the potential to constraint jointly the shear bias and the photo-$z$ distribution,
where other methods fail.} We conclude in Section \ref{sec:conclusion}.
The fiducial cosmology that we adopt in our analysis corresponds to the flat {WMAP9}+SN+BAO cosmology\footnote{Our fiducial cosmology consists of a flat $\Lambda$CDM universe in which the dark energy equation of state is set to $w=-1$.} \citep{2013ApJS..208...19H},
in which the matter density, the dark energy density, the baryonic density, the amplitude of matter fluctuations, the Hubble parameter
and the tilt of the matter power spectrum are described by $(\Omega_{\rm m}, \Omega_{\Lambda}, \Omega_{\rm b}, \sigma_8, h, n_s) = (0.2905, 0.7095, 0.0473, 0.831, 0.6898, 0.969)$.
Aside from {\color{black} determining} the overall amplitude of the theoretical signal from the $[\sigma_8 - \Omega_{\rm m}]$ pair, this choice has little impact on our analysis,
{\color{black} as we later demonstrate}.
{\color{black} Future surveys will have the statistical power to constrain the complete cosmological set, but this is currently out of reach for a survey the size of KiDS-450.}
{\color{black} We note that our fiducial cosmology is a convenient choice} that is consistent within $2\sigma$ with the {\it Planck}, KiDS-450, CFHTLenS, and {WMAP}9+{ACT}+{SPT} analyses in the $[\sigma_8 - \Omega_{\rm m}]$ plane.
As such, it minimizes the impact of residual tension across data sets.
\section{Theoretical Background}
\label{sec:th}
Photons from the surface of last scattering are gravitationally lensed by large scale structures in the Universe before reaching the observer.
{\color{black} Similarly}, photons emitted by observed galaxies are lensed by the low redshift end of the same large scale structures.
The signal expected from a cross-correlation measurement between the two {\color{black} lenses} can be related to the fluctuations in their common foreground matter field,
{\color{black} more precisely by the} matter power spectrum $P(k,z)$. The lensing signal is obtained from an extended first order Limber integration over the past light cone up to the horizon
distance $\chi_{\rm H}$, weighted by geometrical factors $W^{i}(\chi)$, assuming a flat cosmology \citep{Limber1954, LimberExt, NonLimber}:
\begin{eqnarray}
C^{\kappa_{\rm CMB}\kappa_{\rm gal}}_{\ell} = \int_0^{\chi_{\rm H}} d\chi W^{{\rm CMB}}(\chi)W^{{\rm gal}}(\chi) P\left(\frac{\ell+1/2}{\chi};z\right).
\label{eq:limber_cross}
\end{eqnarray}
In the above expression, $\chi$ is the comoving distance from the observer, $\ell$ is the angular multipole, and $z$ is the redshift.
The lensing kernels are given by
\begin{eqnarray}
W^{i}(\chi) = \frac{3 \Omega_{\rm m} H_{0}^{2}}{2 c^2} \chi g^i(\chi) (1 + z),
\label{eq:W-def}
\end{eqnarray}
with
\begin{eqnarray}
g^{\rm gal }(\chi) = \int_{\chi}^{\chi_H} {\rm d} \chi' \tilde{n}(\chi') \frac{\chi' - \chi}{\chi'} \mbox{\hspace{5mm}and} \nonumber
\end{eqnarray}
\begin{eqnarray}
g^{\rm CMB}(\chi) = \left[1 - \frac{\chi}{\chi_*}\right]{\rm H}(\chi_* - \chi).
\label{eq:g-chi-def}
\end{eqnarray}
The constant $c$ is the speed of light in vacuum, $\chi_*$ is the comoving distance to the surface of last scattering.
The term $\tilde{n}(\chi)$ is related to the redshift distribution of the observed galaxy sources, $n(z)$, by
$\tilde{n}(\chi) = n(z) \mbox{d}z/\mbox{d}\chi$, which depends on the depth of the survey.
The Heaviside function ${\rm H}(x)$ guarantees that no contribution comes from behind {\color{black} the surface of last scattering} as the integration in equation \ref{eq:limber_cross} approaches the horizon.
The angular cross-spectrum described by equation \ref{eq:limber_cross} is related to correlation functions in configuration space, in particular
{\color{black} between the CMB lensing map and the tangential shear} \citep{1991ApJ...380....1M}:
\begin{eqnarray}
\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}( \vartheta) = \frac{1}{2\pi} \int_0^\infty \mbox{d} \ell\ \ell\ C^{\kappa_{\rm CMB}\kappa_{\rm gal}}_\ell J_{2}(\ell \vartheta),
\label{eq:xi-prediction}
\end{eqnarray}
where, ${J_2}$ is the Bessel function of the first kind of order 2, and the quantity $\vartheta$ represents the angular separation on the sky.
Details about measurements of $C^{\kappa_{\rm CMB}\kappa_{\rm gal}}_{\ell}$ and the tangential shear $\gamma_{\rm t}$ -- relevant to equations \ref{eq:limber_cross} and \ref{eq:xi-prediction} respectively -- are provided in Section \ref{sec:measurement}.
Our predictions are obtained from the {\small NICAEA}\footnote{\tt www.cosmostat.org/software/nicaea/} cosmological tool \citep{Nicaea}, assuming a non-linear power spectrum described by the \citet{Takahashi2012} revision of the {\small HALOFIT} model \citep{Smith2003}.
\section{The Data Sets}
\label{sec:data}
\subsection{KiDS-450 lensing data}
\label{subsec:KiDSLenS_data}
\begin{figure}
\begin{center}
\hspace{-3mm}
\includegraphics[width=3.4in]{./graphs/fig1.eps}
\caption{Redshift distribution of the selected KiDS-450 sources in the tomographic bins (unnormalized), calibrated using the DIR method described in \citet{KiDS450}.
The $n(z)$ of the broad $Z_B \in [0.1, 0.9]$ bin is shown in black in all panels for reference, while the $n(z)$ for the five tomographic bins are shown in red.
The mean redshift and effective number of galaxy in each tomographic bin are summarized in Table \ref{table:data}. }
\label{fig:nz}
\end{center}
\end{figure}
The KiDS-450 lensing data that we use for our measurements are based on the third data release of {\color{black} dedicated KiDS} observations from the VLT Survey Telescope at Paranal, in Chile,
and are described in \citet{KiDSLenS}, in \citet{KiDS450} and {\color{black} de Jong (2017 in prep.). These references describe the reduction and analysis pipelines leading to the shear catalogues, and present a rigorous and extensive set of systematic verifications}. Referring to these papers for more details, we summarise here the properties of the data that directly affect our measurement.
Although the full area of the KiDS survey will consist of two large patches on the celestial equator and around the South Galactic Pole, the observing strategy was optimized to prioritize the coverage of the GAMA fields \citep{2015MNRAS.452.2087L}.
The footprint of the KiDS-450 data is consequently organized in five fields, G9, G12, G15, G23 and GS, covering a total of 449.7 deg$^2$
{\color{black} While the multiband imaging data are processed by Astro-WISE \citep{KiDS_DR2}},
the lensing $r$-band data are processed by the {\small THELI} reduction method described in \citet{2013MNRAS.433.2545E}.
Shape measurements are determined using the self-calibrated {\it lens}fit algorithm \citep[based on][]{2013MNRAS.429.2858M}
detailed in \citet{2016arXiv160605337F}.
As described in \citet{KiDS450}, each galaxy is assigned a photometric redshift probability distribution provided by the {\color{black} software} {\small BPZ} \citep{2000ApJ...536..571B}. The position of the maximum value of this distribution, labelled $Z_B$, serves only to divide the data into redshift bins.
Inspired by the KiDS-450 cosmic shear measurement, we split the galaxy sample into 5 redshift bins: $Z_B \in [0.1, 0.3]$, $[0.3, 0.5]$, $[0.5, 0.7]$, $[0.7, 0.9]$ and $>0.9$.
We also define a broad redshift bin by selecting all galaxies falling in the range $Z_B \in [0.1, 0.9]$.
The KiDS-450 cosmic shear measurement did not include the $Z_B>0.9$ bin because the photo-$z$ and the shear calibration {\color{black} were poorly constrained therein}.
For this reason, we do not use this bin in our cosmological analysis either. Instead, we estimate these calibration quantities directly from our measurement in Section \ref{subsec:m}.
For each tomographic bin, the estimate of the redshift distribution of our galaxy samples, $n(z)$, is not obtained from the stacked {\small BPZ}-PDF, but from a magnitude--weighted scheme (in 4-dimensional $ugri$ magnitude space) of a spectroscopically matched sub-sample (the `weighted direct calibration' or `DIR' method from the KiDS-450 cosmic shear analysis, demonstrated to be the most precise method covering our redshift range).
Fig. \ref{fig:nz} shows these weighted $n(z)$ distributions, which enter the theoretical predictions through equation \ref{eq:limber_cross}, along with the effective number density per bin.
In order to preserve the full description of the data in the high redshift tail, from where most of the signal originates, we do not fit the distributions with analytical functions,
as was done in previous work \citep[][K16, HD16]{2015PhRvD..91f2001H, 2015PhRvD..92f3517L}. Fitting functions tend to capture well the region where the $n(z)$ is maximal,
however they attribute almost no weight to the (noisy) high redshift tail. This is of lesser importance in the galaxy lensing auto-correlation measurements,
but becomes highly relevant for the CMB lensing cross-correlation.
Instead, we use the actual histograms in the calculation,
recalling that their apparent spikes are smoothed by the lensing kernels in equation \ref{eq:g-chi-def}.
{\color{black} What is apparent from Fig. \ref{fig:nz}, and of importance for this analysis, is that the first tomographic bin has a long tail that significantly overlaps
with the CMB lensing kernel, more than the second tomographic bin. This feature is well captured by the mean redshift distributions, which are listed in Table \ref{table:data}.}
\begin{table}
\caption{Summary of the data properties in the different tomographic bins.
The effective number of galaxy assumes the estimation method of \citet{Heymans2012c}. }
\begin{center}
\begin{tabular}{cccc}
\hline
$Z_B$ cut & $\bar{z}$ & $n_{\rm eff}$ (gal/arcmin$^2$) & $\sigma_{\epsilon}$ \\
\hline
\hline
\multirow{1}{*}{[0.1, 0.9]}
& 0.72 & 7.54 & 0.28\\
\hline
\hline
\multirow{1}{*}{[0.1, 0.3]}
& 0.75 & 2.23 & 0.29 \\
\hline
\multirow{1}{*}{[0.3, 0.5]}
& 0.59 & 2.03 & 0.28 \\
\hline
\multirow{1}{*}{[0.5, 0.7]}
& 0.72 & 1.81 & 0.27 \\
\hline
\multirow{1}{*}{[0.7, 0.9]}
& 0.87 & 1.49 & 0.28 \\
\hline
\multirow{1}{*}{$>$0.9}
& 1.27 & 0.90 & 0.33\\
\hline
\end{tabular}
\end{center}
\label{table:data}
\end{table}%
Based on the quality of the ellipticity measurement, each galaxy is assigned a {\it lens}fit weight $w$,
plus a multiplicative shear calibration factor -- often referred to as the $m$-correction or the shear bias -- that is obtained from image simulations \citep{2016arXiv160605337F}. {\color{black} This calibration is accurate to better than 1\% for objects with $Z_B < 0.9$, but the precision quickly degrades at higher redshifts. As recommended, we do not correct for shear bias in each galaxy, but instead compute the average correction for each tomographic bin (see equation \ref{eq:K}).
In the fifth tomographic bin, we expect to find residual biases in the $m$-correction, but apply it nevertheless, describing in Section \ref{subsec:m} how this correction can be self-calibrated.
To be absolutely clear, we reiterate that we do not include this fifth bin in our main cosmological analysis.
The effective number density and the shape noise in each tomographic bin are also listed in Table \ref{table:data}.
Following \citet{KiDS450}, we apply a $c$-correction by subtracting the weighted mean ellipticity in each field and each tomographic bin,
but this has no impact on our analysis since this $c$ term does not correlate with the CMB lensing data. }
\subsection{{\it Planck} $\kappa_{\rm CMB}$ maps}
\label{subsec:CMB_data}
The CMB lensing data that enter our measurements are {\color{black} the $\kappa_{\rm CMB}$ map} obtained from the 2015 public data release\footnote{{\it Planck} lensing package: {\tt pla.esac.esa.int/pla/\#cosmology}}, thoroughly detailed in \citet{2015arXiv150201591P}. The map making procedure is based on the quadratic estimator described in \citet{2002ApJ...574..566H},
generalized for a suite of multi-frequency temperature and polarization maps. Frequencies are combined such as to remove foreground contamination, while other
sources of secondary signal (mainly emissions from the galactic plane, from point sources and hot clusters) are masked in the CMB maps, prior to the reconstruction.
If some of these are not fully removed from the lensing maps, they will create systematic effects in the $\kappa_{\rm CMB}$ map that show up differently
in the cross-correlation measurement compared to the auto-spectrum analysis.
{\color{black} For example, there could be leakage in the CMB map coming from e.g. residual thermal Sunyaev-Zel'dovich signal that is most likely located near massive clusters.
These same clusters are highly efficient at lensing background galaxies, hence our cross-correlation measurement would be sensitive to this effect.
Indeed, the $\langle {\rm tSZ} \times \gamma_{\rm t}\rangle$, as recently measured in \citet{2016arXiv160807581H}, has a very large signal to noise
and could possibly be detected in a targeted analysis. Although it is difficult to assess the exact level of the tSZ signal in our $\kappa_{\rm CMB}$ map,
the cleaning made possible from the multi-frequency observations from {\it Planck} is thorough, reducing the residual contaminants to a very small fraction.
No quantitative evidence of such leakage has been reported as of yet, and we therefore ignore this in our analysis.
}
Regions from the full sky lensing map that overlap with the five KiDS footprints are extracted, including a 4 degree extension to optimise the signal-to-noise ratio of the measurement (see HD16).
The {\it Planck} release of lensing data also provides the analysis mask, which we apply to the $\kappa_{\rm CMB}$ map prior to carrying out our measurement\footnote{
This procedure does not entirely capture the masking analysis since the mask was applied on the temperature field, not on the lensing map.
The reconstruction process inevitably leaks some of the masked regions into unmasked area, and vice versa. Applying this mask will therefore only remove the most problematic
regions. }.
\section{The measurements}
\label{sec:measurement}
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=4.5in]{./graphs/fig2a.eps}\\
\hspace{-7mm}
\includegraphics[width=4.45in]{./graphs/fig2b.eps}
\end{tabular}
\caption{Cross-correlation measurement between {\it Planck} 2015 $\kappa_{\rm CMB}$ maps and KiDS-450 lensing data.
The upper part presents results from the $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ estimator,
while the lower part shows the estimation of $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$.
Different panels show the results in different tomographic bins, with predictions (solid curve) given by equations \ref{eq:limber_cross} and \ref{eq:xi-prediction} in our fiducial cosmology.
The black squares show the signal, whereas the red circles present the EB null test described in Section \ref{subsec:null}, slightly shifted horizontally to improve the clarity in this figure.
The error bars are computed from 100 CMB lensing simulations. }
\label{fig:cmb_x_shear_KiDS_gamma_t}
\end{center}
\end{figure*}
This section presents the cross-correlation measurements, which are performed with two independent estimators:
$\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ (equation \ref{eq:xi-prediction}) and the {\small POLSPICE} measurement of $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ (equation \ref{eq:limber_cross}).
These were used and rigorously validated in previous work (see HD16, K16 and references therein for more details about these estimators).
{\color{black} The reasons for conducting our analysis with two estimators instead of only one are two fold: they do not probe the same scales, which makes them complementary,
and, being completely independent, residual systematics could be identified through their different effect on these two statistics. }
\subsection{The $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ estimation}
\label{subsec:real_space}
The first estimator presented in this paper, $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$, was recently introduced in HD16, and used later in \citet{2017MNRAS.464.2120S}.
It is a full configuration-space measurement that involves minimal manipulation of the data.
The calculation simply loops over each pixel of the $\kappa_{\rm CMB}$ maps and defines concentric annuli with different radii $\vartheta$, therein measuring the average tangential {\color{black} component of the} shear,
$\gamma_{\rm t}$, from the KiDS galaxy shapes. For this reason, it is arguably the cleanest avenue to perform such a cross-correlation measurement, even though there appears to be a
limit to its accuracy at large angles in some cases due to the finite support of the observation window \citep{2013MNRAS.432.1544M}. That being said, it nevertheless bypasses a number of potential issues that are encountered with
other estimators (see HD16 for a discussion).
The $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ estimator is given by:
\begin{eqnarray}
\label{eq:shear-estimator}
\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}(\vartheta) = \frac{\sum_{ij} \kappa_{\rm CMB}^i e_{\rm t}^{ij} w^j \Delta_{ij}(\vartheta)}{\sum_{ij} w^j \Delta_{ij}(\vartheta)}\frac{1}{1+K(\vartheta)} \,
\end{eqnarray}
where the sum first runs over the $\kappa_{\rm CMB}$ pixels `$i$', then over all galaxies `$j$' found in an annulus of radius $\vartheta$ and width $\Delta$, centered on the pixel $i$.
In this local coordinate system, $e_{\rm t}^{ij}$ is the tangential component of the {\it lens}fit ellipticity from the $j^{\rm th}$ galaxy relative to pixel $i$.
The exact binning scheme is described by $\Delta_{ij}(\vartheta)$, the binning operator:
\begin{eqnarray}
\Delta_{ij}(\vartheta)= \begin{cases}
1 {\rm , if } \left|{\boldsymbol \theta}_i - {\boldsymbol \theta}_j \right | < \vartheta \pm \frac{\Delta}{2} \\
0 {\rm , otherwise}
\end{cases}
\label{eq:bin_operator}
\end{eqnarray}
where ${\boldsymbol \theta}_i$ and ${\boldsymbol \theta}_j$ are the observed positions of the pixel $i$ and galaxy $j$.
{\color{black} Following HD16,} the bin width $\Delta$ is set to 30 arcmin, equally spanning the angular range [1, 181] arcmin with 6 data points.
{\color{black} Larger angular scales capture very little signal with the current level of statistical noise.
We verified that our analysis results are independent of our choice of binning scheme.}
In equation \ref{eq:shear-estimator}, $w^j$ is the {\it lens}fit weight of the galaxy $j$ and $K(\vartheta)$ corrects
for the shape multiplicative bias $m^j$ that must be applied to the lensing data \citep{2016arXiv160605337F}:
\begin{eqnarray}
\frac{1}{1+K(\vartheta)} = \frac{\sum_{ij} w^j \Delta_{ij}(\vartheta)}{\sum_{ij} w^j (1+m^j)\Delta_{ij}(\vartheta)} \ .
\label{eq:K}
\end{eqnarray}
The theoretical predictions for $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ are provided by equation \ref{eq:xi-prediction}.
We apply the same binning as with the data, averaging the continuous theory lines inside each angular bin.
We show in the upper panel of Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t} the measurements in all tomographic bins,
compared to theoretical predictions given by our fiducial WMAP9+BAO+SN cosmology.
The estimation of our error bars is described in Section \ref{subsec:error}.
We also project the galaxy shape components onto $e_{\times}$, which is rotated by 45 degrees compared to $e_{\rm t}$.
This effectively constitutes a nulling operation that can inform us of systematic leakage, in analogy to the EB test performed in the context of cosmic shear.
For this reason, we loosely refer to EE and EB tests in this paper, when we are in fact comparing $\kappa_{\rm CMB} \times e_{\rm t}$ and $\kappa_{\rm CMB} \times e_{\times}$, respectively.
We note that the past literature referred to such a EB measurement as the `B-mode test',
which can be misleading for the non-expert. Indeed, the proper B-mode test refers to the BB measurement in weak lensing analyses,
a non-lensing signal that can be caused by astrophysics and systematics. The EB signal test {\color{black} asserts} something more fundamental:
since B changes sign under parity, and E does not, a non-zero EB means a violation of the parity of the shear/ellipticity field \citep{2003A&A...408..829S}.
That is not expected from astrophysics, so could only come from a systematic effect that does not vanish under averaging.
Our EB measurement is shown with the red symbols in Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t}.
We find by visual inspection that in most tomographic bins, these seem closely centered on zero, but not in all cases.
To quantify the significance of this EB measurement, we estimate the confidence at which these red points deviate from zero.
We detail in Section \ref{subsec:null} how we carry out that test and show that they are consistent with noise.
We have carried out an additional null test presented in HD16, which consists in rotating randomly the shapes of the galaxies before the measurement
($\kappa_{\rm CMB} \times \mbox{random}$). {\color{black} This test is sensitive to the noise levels in the galaxy lensing data and hence affected by the shape noise $\sigma_{\epsilon}$ listed in Table \ref{table:data}. We find that the resulting signal is fully consistent with zero in all tomographic bins. }
\subsection{The $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ estimation}
\label{subsubsec:PolSpice}
The second estimator uses the same data as our $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ analysis, namely the $\kappa_{\rm CMB}$ map and the KiDS shear catalogues,
but requires additional operations on the data, including harmonic space transforms. This is accomplished with the {\small POLSPICE} numerical
code \citep{spice,polspice} running in polarization mode, where the \{T, Q, U\} triplets are replaced by {\color{black} \{$\kappa_{\rm CMB}, 0, 0\}$ and $\{0, -e_1, e_2$\}}.
The code first computes the pseudo-$C_{\ell}$ of the maps and of the masks, then transforms the results into configuration space quantities,
that are finally combined and transformed back into Fourier space. The output of {\small POLSPICE} is therefore an estimate of the cross-spectrum $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$.
While {\small POLSPICE} is frequently used for CMB analyses, it was applied for the first time in the context of CMB lensing $\times$ galaxy lensing by K16 and serves as a good comparison to the configuration estimator described in Section
\ref{subsec:real_space}. One main advantage of this estimator is that in principle different $\ell$-bands are largely uncorrelated, which makes the covariance matrix almost diagonal and hence easier to estimate.
The {\small POLSPICE} measurement\footnote{{\color{black} {\small POLSPICE} has adjustable internal parameters, and we use {\sc thetamax} = 60 deg., {\sc apodizesigma} = 60 deg. and {\sc nlmax} = 3000.}} is presented in the lower panel of Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t}, plotted against the theoretical predictions given by equation \ref{eq:limber_cross}.
The EB data points are directly obtained from the temperature/B-mode output provided by the polarization version of the code,
and are further discussed in Section \ref{subsec:null}.
Note that our choice of the $\gamma_{\rm t}$ and {\small POLSPICE} estimators was motivated by our desire to avoid producing $\kappa_{\rm gal}$ maps in order to reduce the risks of errors and systematic biases
that can arise in the map making stage in the presence of a mask as inhomogeneous as that of the KiDS-450 data.
These two estimators produce correlated measurements, but the scales they are probing differ.
The $\gamma_{\rm t}$ estimator is accurate at the few percent level, as verified on full mock data in HD16, and the {\small POLSPICE} code has been thoroughly
verified and validated on the same mocks as well. We refer the reader to K16 and HD16 for details of these tests.
\subsection{Covariance Estimation}
\label{subsec:error}
The $\kappa_{\rm CMB}$ map reconstructed by the {\it Planck} data is noise dominated for most Fourier modes \citep{2015arXiv150201591P}.
It is only by combining the full sky temperature and polarization maps that the {\it Planck} Collaboration could achieve a lensing detection of $40\sigma$.
Since the noise $N_{\rm CMB}$ is larger than the signal $\kappa_{\rm CMB}$ at every scale included in our analysis (HD16),
we can evaluate the covariance matrix from cross-correlation measurements between the
100 {\it Planck} simulated lensing maps (also provided in their 2015 public data release) and the tomographic KiDS data:
\begin{eqnarray}
{\rm Cov}^{\kappa_{\rm CMB} \kappa_{\rm gal}}_{\ell \ell'} \simeq \langle \Delta \hat{C}^{N_{\rm CMB} \kappa_{\rm gal}}_{\ell} \Delta \hat{C}^{N_{\rm CMB} \kappa_{\rm gal}}_{\ell'} \rangle
\end{eqnarray}
and
\begin{eqnarray}
{\rm Cov}^{\kappa_{\rm CMB} \gamma_{\rm t}}_{\vartheta \vartheta'} \simeq \langle \Delta \hat{\xi}^{N_{\rm CMB} \gamma_{\rm t}}_{\vartheta} \Delta \hat{\xi}^{N_{\rm CMB} \gamma_{\rm t}}_{\vartheta'} \rangle.
\label{eq:Cov_approx}
\end{eqnarray}
where the `hats' refer to measured quantities, $\Delta \hat{x} = \hat{x} - \bar{x}$, and the brackets represent the average over the 100 realizations.
{\color{black} This method assumes that the covariance is completely dominated by the CMB lensing and neglects the contribution from the shear covariance.
This is justified by the fact that the signal from the former is about an order of magnitude larger, and hence completely drives the statistical uncertainty (HD16).}
The error bars shown in Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t} are obtained from these matrices (from the square root of the diagonals).
For each tomographic bin, the ${\rm Cov}^{\kappa_{\rm CMB} \kappa_{\rm gal}}_{\ell \ell'}$ matrix has 25 elements, whereas the ${\rm Cov}^{\kappa_{\rm CMB} \gamma_{\rm t}}_{\vartheta \vartheta'}$
matrix has 36. The 100 realizations are enough to invert these matrices one at a time with a controllable level of noise bias,
and the numerical convergence on this inverse is guaranteed \citep{Lu2010}.
{\color{black}
Note that this strategy fails to capture the correlation between tomographic bins, which are not required by our cosmological analysis presented in Section \ref{subsec:cosmo_broad}.
If needed in a future analysis, these could be estimated from full light-cone mock simulations.}
For both estimators, the covariance matrix is dominated by its diagonal, with most off-diagonal elements of the cross-correlation coefficient matrix being under $\pm10\%$.
Some elements reach larger values, $\pm$40\% correlation at the most, but these are isolated, not common to all tomographic bins,
and are consistent with being noise fluctuations,
given that we are measuring many elements from `only' 100 simulations.
This partly explains why our cosmological results are not based on a joint tomographic analysis.
We keep the full matrices in the analysis, even though we could, in principle, include only the diagonal part in the {\small POLSPICE} measurement.
Nevertheless, we have checked that our final results are only negligibly modified if we use this approximation in the $\chi^2$ calculation, suggesting that one could reliably use a {\color{black} Gaussian approximation to the error estimation in this type of measurement (see equation 23 in HD16)}.
\section{Cosmological Inference}
\label{sec:results}
Given the relatively low signal-to-noise ratio of our measurement (Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t}), we do not fit our signal for the six parameters $\Lambda$CDM cosmological model.
Instead, we follow the strategy adopted by earlier measurements: {\color{black} we compare the measured signal to our fiducial cosmological predictions, treating the normalization as a free parameter `$A$'.
If the assumed fiducial cosmology is correct and in absence of other systematic effects, $A$ is expected to be consistent with unity. }
As discussed in previous studies, $A$ is affected by a number of {\color{black} effects} that can similarly modulate the overall amplitude of the signal.
Aside from its sensitivity to cosmology -- our primary science target -- this rescaling term will absorb contributions from residual systematic errors in the estimation of $n(z)$,
from mis-modelling of the galaxy intrinsic alignments,
from residual systematic bias in the shear multiplicative term $m$ (equation \ref{eq:K}), from astrophysical phenomena such as massive neutrinos and/or baryonic feedback,
and from residual systematics in the cross-correlation estimators themselves (K16 and HD16).
In this section, we first present our constraints on $A$; we then quantify how the different effects listed above can impact our measurements,
and finally present our cosmological interpretation. {\color{black} Our primary results assume the fiducial WMAP9+BAO+SN cosmology, i.e. we first place constraints on $A_{\rm fid}$,
however we also report constraints on $A_{\rm KiDS}$ and $A_{Planck}$, obtained by assuming different baseline cosmologies.}
\subsection{Significance}
\label{subsec:A}
To measure $A$, we first compute the $\chi^2$ statistic:
\begin{eqnarray}
\chi^2 = \Delta{\rm x}^{\rm T}\ \mbox{Cov}^{-1}\ \Delta{\rm x}
\label{eq:chi}
\end{eqnarray}
with
\begin{eqnarray}
\Delta{\rm x} = \hat{\xi}^{\kappa_{\rm CMB}\gamma_{\rm t}} - A {\xi}^{\kappa_{\rm CMB}\gamma_{\rm t}} \mbox{ or } \Delta{\rm x} = \hat{C}^{\kappa_{\rm CMB}\kappa_{\rm gal}} - A {C}^{\kappa_{\rm CMB}\kappa_{\rm gal}}
\label{eq:chi2}
\end{eqnarray}
for the configuration space and {\small POLSPICE} estimators, respectively. {\color{black} As before, quantities with `hats' are measured, and the predictions assume the fiducial cosmology,
unless stated otherwise}.
The signal-to-noise (SNR) ratio is given by the likelihood ratio test,
which measures the confidence at which we can reject the null hypothesis (i.e. that there is no signal, simply noise)
in favour of an alternative hypothesis described by our theoretical model with a single parameter $A$ \citep[see][for a recent derivation in a similar context]{2016arXiv160807581H}.
We can write SNR $= \sqrt{\chi^2_{\rm null} - \chi^2_{\rm min}}$, where $\chi^2_{\rm null}$ is computed by setting $A = 0$,
and $\chi^2_{\rm min}$ corresponds to the best-fit value for $A$.
The error on $A$ is obtained by varying the value of $A$ until $ \chi^2_{A} - \chi^2_{\rm min} = 1$ \citep[see, e.g.][]{2003psa..book.....W}.
We include two additional statistical corrections to this calculation.
The first is a correction factor that multiplies the inverse covariance matrix, $\alpha = (N_{\rm sim}-N_{\rm bin}-2)/(N_{\rm sim}-1) = 0.94$, to account for biases inherent to matrix inversion in the presence of noise \citep{Hartlap2007a}. Here $N_{\rm bin}$ is the number of data bins (5 for $C^{\kappa_{\rm CMB}\kappa_{\rm gal}}$ and 6 for $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$) and $N_{\rm sim}$ is the number of simulations (100) used in the covariance estimation.
There exists an improved version of this calculation based on assuming a $t$-distribution {\color{black} in the likelihood},
however {\color{black} with our values of $N_{\rm bin}$ and $N_{\rm sim}$}, the differences in the inverted matrix would be of order 10-20\% \citep{2016MNRAS.456L.132S},
a correction on the error that we ignore given the relatively high level of noise in our measurement.
The second correction was first used in HD16, and consists of an additional error on $A$
due to the propagated uncertainty coming from the noise in the covariance matrix \citep{2014MNRAS.442.2728T}.
This effectively maps $\sigma_A \rightarrow \sigma_A(1 + \epsilon/2)$, where $\epsilon = \sqrt{2/N_{\rm sim}+ 2(N_{\rm bin}/N_{\rm sim}^2)} = 0.145 $.
These two correction factors are included in the analysis.
The results from our statistical investigation are reported in Table \ref{table:stat}, where we list $\chi^2_{\rm min}$, $\chi^2_{\rm null}$, SNR and $A$
for every tomographic bin. The theoretical predictions provide a good fit to the data given that for our degrees of freedom $\nu = N_{\rm bin}-1$,
$\nu - \sqrt{2\nu} < \chi^2_{\rm min} < \nu + \sqrt{2\nu}$. In other words, all our measured $\chi^2$ fall within the expected 1$\sigma$ error.
{\color{black} We also compute the $p$-value for all these $\chi^2$ measurements at the best-fitting $A$ in order to estimate the confidence at which we can accept or reject the assumed model.
Assuming Gaussian statistics, $p$-values smaller than 0.01 correspond to a 99\% confidence in the rejection of the model (the null hypothesis) by the data, and are considered `problematic'.
Our measured $p$-values, also listed in Table \ref{table:stat}, are always larger than 0.12, meaning that the model provides a good fit to the data in all cases.}
\begin{table}
\caption{Summary of $\chi^2$, SNR {\color{black} and $p$-values} obtained with the two different pipelines.
The $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ measurements have 4 degrees of freedom (5 $\ell$-bins - 1 free parameter),
whereas the configuration space counterpart $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}(\vartheta)$ has one more, with 6 $\vartheta$-bins.
$A_{\rm fid}$ is the best-fit amplitude that scales the theoretical signals in the fiducial cosmology, according to equation \ref{eq:chi2},
also shown in Fig. \ref{fig:A_tomo}.
The numbers listed here include the covariance debiasing factor $\alpha$ and the extra error $\epsilon$ due to the noise in the covariance (see main text of Section \ref{subsec:A} for more details). }
\begin{center}
\begin{tabular}{clccccc}
\hline
$Z_B$ & Estimator & $\chi^2_{\rm min}$ & $\chi^2_{\rm null}$ & SNR & \color{black} $p$-values & $A_{\rm fid}$ \\
\hline
\hline
\multirow{2}{*}{[0.1, 0.9]}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 2.80 & 18.21 & 3.93 & 0.53& 0.77 $\pm$ 0.19 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 2.88 & 22.94 & 4.48 & 0.64 & 0.69 $\pm$ 0.15 \\
\hline
\hline
\multirow{2}{*}{[0.1, 0.3]}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 5.48 & 8.89 & 1.85 & 0.20 & 0.55 $\pm$ 0.30 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 7.93 & 13.38 & 2.34 & 0.12 & 0.53 $\pm$ 0.24 \\
\hline
\multirow{2}{*}{[0.3, 0.5]}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 2.95 & 4.95 & 1.42 & 0.50 & 0.71 $\pm$ 0.51 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 1.44 & 4.19 & 1.66 & 0.84 & 0.60 $\pm$ 0.37 \\
\hline
\multirow{2}{*}{[0.5, 0.7]}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 4.00 & 10.13 & 2.47 & 0.35 & 0.87 $\pm$ 0.35 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 2.00 & 6.45 & 2.11 & 0.77 & 0.55 $\pm$ 0.26 \\
\hline
\multirow{2}{*}{[0.7, 0.9]}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 5.12 & 10.04 & 2.22 & 0.23 & 0.79 $\pm$ 0.36 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 2.78 & 15.41 & 3.55 & 0.65 & 1.02 $\pm$ 0.29 \\
\hline
\multirow{2}{*}{$>0.9$}
& $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ & 4.70 & 12.92 & 2.87 & 0.26 & 0.83 $\pm$ 0.29 \\
& $\xi^{\kappa_{\rm CMB}\gamma_{\rm t}}$ & 4.68 & 22.64 & 4.24 & 0.38 & 0.95 $\pm$ 0.22 \\
\hline
\end{tabular}
\end{center}
\label{table:stat}
\end{table}%
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig3.eps}
\end{tabular}
\caption{Tomographic measurement of $A_{\rm fid}$, defined in equation \ref{eq:chi2}, assuming our fiducial cosmology.
The two panels present results from the two cross-correlation estimators (labelled in the top left corner).
Black symbols assume no IA, while color symbols include correction factors from two IA models ($f_{\rm red}$ in magenta and HT in blue, see Sec. \ref{subsec:IA}).
The horizontal solid lines of a given color enclose the $1\sigma$ region measured in the broad $Z_B \in [0.1, 0.9]$ bin,
while the dotted horizontal lines indicate the fiducial values ($A_{\rm fid}$ = 1).
The mean source redshift in each bin is indicated at the top and summarized in Table \ref{table:data}.
{\color{black} The mean in the first bin is high because of the long tail, visible in Fig. \ref{fig:nz}.
The best fit values in different cosmologies are $A_{\rm fid} = 0.69\pm0.15$, $A_{\rm KiDS} = 0.86\pm0.19$ and $A_{Planck} = 0.68\pm0.15$.}
}
\label{fig:A_tomo}
\end{center}
\end{figure}
These tomographic measurements are re-grouped in Fig. \ref{fig:A_tomo}, where we compare the redshift evolution of $A$ for both estimators.
We mark the $1\sigma$ region of the broad bin $n(z)$ with the solid horizontal lines, and see that all points overlap with this region within $1\sigma$.
This is an indication that the {\color{black} relative growth of structure between the tomographic bins} is consistent with the assumed $\Lambda$CDM model.
For the broad $n(z)$, the signal prefers an amplitude that is $\sim23-31$ percent lower than the fiducial cosmology,
i.e. the $1\sigma$ region shown by the horizontal solid lines in Fig. \ref{fig:A_tomo} is offset from unity by that amount.
The main cosmological result that we quote from the $Z_B \in [0.1, 0.9]$ measurement is that of the $\gamma_{\rm t}$ estimator due to its higher SNR,
as seen from comparing the top two rows of Table \ref{table:stat}.
For our fiducial cosmology, we find:
\begin{eqnarray}
A_{\rm fid} = 0.69\pm0.15
\label{eq:A_fid}
\end{eqnarray}
Varying the cosmology to the best fit KiDS-450
and {\it Planck} cosmologies\footnote{\color{black} Fiducial $(\Omega_{\rm m}, \Omega_{\Lambda}, \Omega_{\rm b}, \sigma_8, h, n_s) = (0.29, 0.71, 0.047, 0.83, 0.69, 0.97)$.}$^,$\footnote{KiDS-450 $(\Omega_{\rm m}, \Omega_{\Lambda}, \Omega_{\rm b}, \sigma_8, h, n_s) = (0.25, 0.75, 0.047, 0.85, 0.75, 1.09)$.}$^,$\footnote{{\it Planck} $(\Omega_{\rm m}, \Omega_{\Lambda}, \Omega_{\rm b}, \sigma_8, h, n_s) = (0.32, 0.68, 0.049, 0.80, 0.67, 0.97)$.}, we obtain
\begin{eqnarray}
A_{\rm KiDS} = 0.86\pm0.19 \mbox{\hspace{5mm} and \hspace{5mm} } A_{Planck} = 0.68\pm0.15
\end{eqnarray}
The relative impact of these different cosmologies on our signal is presented in Fig. \ref{fig:frac_cosmo}, where we see that the KiDS-450 cosmology mostly differ from the other two at large scales.
The signal from our fiducial cosmology agrees with that assuming the best-fit {\it Planck} cosmology to better than 5\% in all tomographic bins.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig4.eps}
\end{tabular}
\caption{Fractional effect on the signal when changing the fiducial cosmology to {\it Planck} or KiDS-450.
Different symbols show the impact in different tomographic bins, relative to the fiducial predictions.
Current measurements are limited to $\ell<2000$.}
\label{fig:frac_cosmo}
\end{center}
\end{figure}
{\color{black} Fig. \ref{fig:A_tomo} demonstrates there are small but noticeable differences between the two estimators at fixed cosmology, especially at high redshift.
As mentioned in Section \ref{subsubsec:PolSpice}, the scales being probed are not identical, and therefore
some differences in the recovered values of $A$ are expected.
Nevertheless, within the current statistical accuracy, the two estimators are fully consistent with one another. }
Visually, the $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}(\vartheta)$ estimator seems to show a mild trend for decreasing values of $A$ in lower redshift bins.
Although such an effect could point towards a number of interesting phenomena suppressing power for source galaxies at $z\lesssim0.7$
(e.g. modification to the growth history compared to the fiducial cosmology, additional feedback processes from baryons or massive neutrinos, or redshift-dependent contamination from IA)
the significance of this redshift dependence is too low to draw any robust conclusions.
What is significantly seen from Fig. \ref{fig:A_tomo} is that the signal is generally low compared to the fiducial and {\it Planck} cosmologies.
Our measurements of the amplitude $A$ prefer instead the KiDS-450 cosmology, which also aligns with the CFHTLenS cosmic shear results \citep{Kilbinger2013}.
We further quantify this comparison in Section \ref{subsec:cosmo_broad}, first presenting results from our set of null tests, and then examining three sources of contamination
and systematic biases that potentially affect our signal.
%
{\color{black} In this work, we neglect the effect of source-lens coupling \citep{1998A&A...338..375B}, which could possibly act as another secondary signal, biasing the signal low.
As it is the case for cosmic shear, this effect should be too small ( $<10\%$) to affect our results significantly, and further investigation will be required to interpret correctly the measurements from future surveys. }
\begin{table}
\caption{$p$-values for the EB test obtained for the 6 tomographic bins. Highlighted in bold is the $p$-value $\le0.01$.
The column labelled $C_{10}$ refers to {\small POLSPICE} measurements
in which the data are organized in 10 bins instead of 5. These calculations assume $t$-distributed likelihoods \citep[following][]{2016MNRAS.456L.132S}.}
\begin{center}
\begin{tabular}{ccccccc}
\hline
$Z_B$ & $C^{t{\rm -dist}}$ & $C_{10}^{t{\rm -dist}}$ & $\xi^{t{\rm -dist}}$ \\
\hline
\hline
\multirow{1}{*}{[0.1, 0.9]}
&0.20 & 0.09 & 0.58 \\
\hline
\hline
\multirow{1}{*}{[0.1, 0.3]}
&\bf 0.01 & 0.09 & 0.52 \\
\hline
\multirow{1}{*}{[0.3, 0.5]}
&0.39 & 0.63 & 0.08 \\
\hline
\multirow{1}{*}{[0.5, 0.7]}
& 0.94 & 0.57 & 0.78\\
\hline
\multirow{1}{*}{[0.7, 0.9]}
& 0.21 & 0.20 & 0.16\\
\hline
\multirow{1}{*}{$>0.9$}
& 0.53 & 0.54 & 0.68\\
\hline
\end{tabular}
\end{center}
\label{table:p_value_B}
\end{table}%
\subsection{Null tests}
\label{subsec:null}
{\color{black} We have shown in Section \ref{sec:measurement} and in Fig. \ref{fig:cmb_x_shear_KiDS_gamma_t} (red circles) that the parity violation EB test seemed consistent with noise in most
tomographic bins, but occasionally that was not obvious. In this section, we investigate the significance of these measurements.}
Statistically, this is accomplished by measuring the confidence at which we can reject the null hypothesis `{\it parity is not violated}'.
We therefore re-run the full $\chi^2$ statistical analysis\footnote{Due to the absence of parameters in the null hypothesis, the EB case has one additional degree of freedom compared to the EE case.} and measure the $p$-value about the model with $A=0$.
Low $p$-values correspond to high confidence of rejection, i.e. that some residual systematic effect might be causing and apparent parity violation.
{\color{black} This type of measurement strongly probes the tail of the $\chi^2$ distribution, hence assuming a Gaussian likelihood would provide inaccurate estimations of the $p$-values,
even when including the \citet{Hartlap2007a} debiasing $\alpha$ factor. Instead, we follow \citet{2016MNRAS.456L.132S} and assume a $t$-distribution for the likelihood,
which better models the tail of the likelihood.
Table \ref{table:p_value_B} lists all these $p$-values, highlighting in bold one that seems slightly problematic ($p$-value $\le$ 0.01).
Since this single low $p$-value is only seen in one of the two estimators, we conclude that it must originate from expected noise fluctuations, and
not from the data itself.
This conclusion is additionally supported by the fact that the level of B-modes in the KiDS data (i.e. the BB measurement)
is consistent with zero on the scales we are probing \citep{KiDS450}.
For the sake of testing the robustness of the EB {\small POLSPICE} measurement, we have additionally investigated the effect of changing the number of bins from 5 to 10.
In the EE case, the recovered values of $A$ and the SNR are similar to those presented in Table \ref{table:stat},
from which we conclude that this comes with no gain.
However, when applied to the EB null test, something interesting happens:
the `problematic' measurement ($p$-value = 0.01 in the $Z_B \in [0.1, 0.3]$ bin, Table \ref{table:p_value_B})
relaxes to 0.09, as seen in the column labelled $C_{10}$.
This is another indication that the cause of the low $p$-value originates from fluctuations in the noise --
which is affected in the binning process -- without pointing to residual systematic effects in the data.}
We have verified that our measurement of $A$ is robust against the removal of some scales.
When we exclude the largest or the smallest angular bin in the $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ measurement,
results change by at most 0.7$\sigma$, generally by less than 0.2$\sigma$. This gives us confidence in the robustness of our measurement.
The same holds when removing the highest $\ell$ bin from the {\small POLSPICE} measurement, but not for the lowest $\ell$ bin,
which captures the peak of the signal, and therefore contributes significantly to the SNR.
{\color{black} At the same time, this test illustrates that we are currently not sensitive to the effect of massive neutrinos nor to baryonic feedback,
which affect mainly these non-linear scales.}
\subsection{Effect of intrinsic alignments}
\label{subsec:IA}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig5.eps}
\end{tabular}
\caption{Strength of the contamination by intrinsic galaxy alignments for different tomographic bins, assuming our fiducial cosmology and
the linear non-linear alignment model.
The difference between lines is caused by changes in $n(z)$ (and in the red fraction in the $f_{\rm red}$-IA model).}
\label{fig:frac_IA}
\end{center}
\end{figure}
Intrinsic alignments (IA) are a known secondary effect to the cross-correlation of galaxy lensing and CMB lensing
that lowers the amplitude of the measured signal \citep{2014MNRAS.443L.119H, TroxelIshak14, 2015MNRAS.453..682C}.
It is therefore important to investigate how much IA could contribute to the observed low values of $A$ reported in Table \ref{table:stat}.
To estimate the contamination level, we compare two different {\color{black} models}, which we then apply equally to both estimators, $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ and $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}(\vartheta)$.
First, we follow \citet[][`HT-IA' model hereafter]{2014MNRAS.443L.119H} in using the `linear non-linear alignment' model of \citet{2007NJPh....9..444B}
with the SuperCOSMOS normalization found in \citet{SuperCosmosIA}. {\color{black} We recall that this prescription comes from constraints at $z = 0.1$ that are independent of galaxy type or colour, and that the effect of IA in this model is to reduce the amplitude of the observed signal, as the galaxies tend to align radially towards each other.}
The scale-dependence of the alignment contribution is similar to the lensing signal, as seen in \citet{2014MNRAS.443L.119H} and in Fig. \ref{fig:frac_IA}, hence we only quote the
percentage of contamination at $\ell = 1000$ for reference.
{\color{black} This also allows us to use with confidence the same IA contamination levels for the configuration space estimator, since
rescaling $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ by a constant rescales $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ by the same constant (as per equation \ref{eq:limber_cross}). }
For each of the five redshift bins considered in this paper, starting from the lower redshift, we estimate
a $\{10,17,10, 8, 5\}$\% contamination to the signal, respectively.
For the broader tomographic bin $Z_B \in [0.1, 0.9]$, we estimate a $11\%$ contamination.
In other words, within the HT-IA model, the measured value of $A_{\rm fid}$ in the broad bin (equation \ref{eq:A_fid}) should be corrected to
\begin{eqnarray}
A_{\rm fid}^{\rm HT} = 0.77\pm0.15
\label{eq:A_fid_HT}
\end{eqnarray}
The error bars are not modified compared to the no-IA case since this contamination signal is additive.
This model is the simplest as it assumes no luminosity or redshift dependence of the alignment normalization, and adopts the same alignment prescription for all galaxies regardless of morphological type/colour.
Second, we estimate the contamination from the alignment model of \citet[][`$f_{\rm red}$-IA' model hereafter]{2015MNRAS.453..682C} that allows for differential contributions based on galaxy colour/morphology.
We assume that blue galaxies do not contribute at all, consistent with observations \citep{Heymans2013a, BlueIA}, even though this null measurement remains poorly constrained.
We estimate the red fraction directly from the data in each redshift bin, using the BPZ template information {\tt T\_B}.
Motivated by \citet{2015MNRAS.449.1505S}, we identify red galaxies as objects with {\tt T\_B} $<1.5$.
For the five tomographic bins, we obtain fractions of red galaxies $f_{\rm red} = \{0.04, 0.12, 0.27, 0.18, 0.04\}$; we estimate $f_{\rm red} = 0.15$ for the broad bin.
We then use the alignment amplitude for the red galaxies from \citet{Heymans2013a} to obtain an estimate of alignment contamination given our red fractions.
These results are presented in the upper panel of Fig. \ref{fig:frac_IA}.
With this method, we estimate a $\{2, 11, 14, 7, 1\}$\% contamination from intrinsic alignments in the tomographic bins,
and a 9\% contamination in the $Z_B \in [0.1, 0.9]$ bin\footnote{
We measured the field-to-field variance in $f_{\rm red}$ and observed that it hardly varies except in the highest redshift bin, where
the scatter could turn the 1\% IA contamination into a 0.0 -- 2.5\% contamination.
This remains small and should have a negligible impact.}.
Then we can estimate
\begin{eqnarray}
A_{\rm fid}^{f_{\rm red}} = 0.75\pm0.15.
\label{eq:A_fid_fred}
\end{eqnarray}
{\color{black} One caveat with this model is that the $K$-correction and evolutionary corrections are uncertain at high redshift,
which could result in biased estimates of the red fraction \citep[see discussion in][]{2015MNRAS.453..682C}.
This has an impact on the exact level of intrinsic alignment contamination by red galaxies, but we neglect this effect in this work.}
Both methods are broadly consistent even though they differ in details, especially in the lowest redshift bin.
For instance, the second method captures the redshift differences observed in the data and takes into account the split in contributions arising from different galaxy types,
which introduces a slightly different redshift dependence of the IA signal.
The overall trends between the HT and the $f_{\rm red}$-IA models are similar though, but that is not the case for all IA models
\citep[see, for example, the tidal torque theory from][in which the sign of the IA effect on the signal is the opposite]{2015MNRAS.452.3369C}.
There remains a large uncertainty in the modelling of the IA contamination, and we do not know which model, if any,
should enter in our cosmological interpretation.
According to these estimations, both the HT-IA and $f_{\rm red}$-IA models help to bring $A$ closer to unity.
From the contamination levels listed above, at most 17\% of the observed cross-correlation signal can be canceled by IA contamination in our tomographic bins.
After correcting for this effect in each tomographic bin, most points agree with $A_{\rm fid}=1$ within $1\sigma$.
This is shown with the color symbols in Fig. \ref{fig:A_tomo}.
Finally, we note that the uncertainty on the level of IA contamination quoted in the section is high,
especially because of the unknown signal from the blue galaxies.
For instance, at the $1\sigma$ level and assuming the linear non-linear alignment model, the IA contamination from blue galaxies could range from
$-\{10, 15, 8, 6, 4 \}$\% to +$\{6, 9, 5, 4, 3\}$\% in each tomographic bin, and from $-10\%$ to +6\% in the $Z_B \in [0.1, 0.9]$ bin.
\subsection{Effect of $n(z)$ errors}
\label{subsec:nz}
We investigate here the impact on our measurement of $A$ from the uncertainty on the source redshift distribution.
This is estimated from 100 bootstrap resamplings of the source catalogue, as detailed in \citet[][the DIR method described therein]{KiDS450}.
These samples consist of internal fluctuations in the $n(z)$, which we turn into fluctuations in the signal with equations \ref{eq:limber_cross} - \ref{eq:g-chi-def}.
We present in the top panel of Fig. \ref{fig:frac_nz} the fractional error on the signal, i.e. $\sigma^{\rm boot}_{C_{\ell}}/C_{\ell}$.
According to this error estimate, the uncertainty on $n(z)$ is up to 8\% of the signal in the first redshift bin, then 4, 2 and 1\% for the others, {\color{black} and about 3\%
for the $Z_B \in$ [0.1, 0.9] tomographic bin.}
Note that this quantity is a measure of how the DIR $n(z)$ varies -- and how it impacts the signal -- across subsamples of the re-weighted spectroscopically matched catalogue.
This catalogue is by itself subsampling the full KiDS sources, and hence subject to sampling variance.
It is therefore likely that the error quoted above slightly under-estimates the true error on the signal due to the $n(z)$,
{\color{black} as discussed in Section C3.1 of \citet{KiDS450}}.
This is sub-dominant compared to our statistical uncertainty and is therefore not expected to affect our results.
For comparison purposes, we also investigated estimates of the redshift distribution determined using the cross-correlation between spectroscopic and photometric samples
\citep[known as the CC method in][]{KiDS450,TheWiZZ}. This scheme has a high level of noise compared to the fiducial DIR method and we find that the error on the recovered $C_\ell$ in our analysis increases from $\sim 5\%$ in the DIR case to $\sim 30\%$ in the CC case. From this we can draw the same conclusion as the KiDS-450 cosmic shear analysis, that determining the redshift distribution using the cross-correlation CC method will remove any discrepancy with a {\it Planck} cosmology through the inflation of the error bars. We believe however that the error on the CC estimate is not representative of our actual knowledge of the $n(z)$ in the KiDS data, and refer instead to the redshift distribution defined using DIR method in the rest of this paper.
Precision on the KiDS source redshift distribution will soon increase thanks to the ongoing processing of near-IR VIKING data (Hildebrandt et al. in prep.),
which primarily impact the high redshift tail so crucial to our measurement. Finally note that in the DIR method we are using a calibrated $n(z)$, estimated from weighted spectroscopic data,
hence we do not have to worry nearly as much about catastrophic photo-$z$ outliers. This was not the case for the analysis presented in HD16, which showed
that for $n(z)$ estimated directly from photometric data (for e.g. CFHTLenS and RCSLenS), these can easily dominate the error budget, with systematic effects on the signal of the order of 15\%.
If our measurement contains more high-redshift objects than our $n(z)$ suggests, our predictions are too low; correcting for this would lower $A$.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=3.2in]{./graphs/fig6.eps}\\
\end{tabular}
\caption{Fractional effect on the $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ signal when varying the $n(z)$ between 100 bootstrap resamplings, for the four tomographic bins with $Z_B<0.9$.
Shown is the $1\sigma$ scatter divided by the signal. }
\label{fig:frac_nz}
\end{center}
\end{figure}
\subsection{Baryon feedback, massive neutrinos and non-linear modelling}
\label{subsec:baryon}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig7.eps}
\end{tabular}
\caption{Fractional effect of the AGN baryon feedback and massive neutrinos on the cross-spectra, for different combinations of source planes.
The red solid line shows the combined effect on the cross-spectrum for sources placed on a single plane at $z_s = 0.5$.
The effect of 0.05eV massive neutrinos and AGN feedback are shown separately by the upper dashed and the dotted line (also in red).
The lower dashed redline shows the impact of 0.4eV neutrinos.
Blue lines show the same quantities, but for sources placed at $z_s = 1.5$.
{\color{black}The dashed black line shows the ratio between the predictions from \citet{Smith2003} and that of \citet{Takahashi2012}.}}
\label{fig:frac_baryons}
\end{center}
\end{figure}
As shown in HD16, baryonic feedback and massive neutrinos can cause an important decrease of the cross-correlation signal,
which would translate into lower values of $A$ when compared to a fiducial dark matter only cosmology.
To investigate how this could affect our cosmological results, we modify the $P(k,z)$ term in equation \ref{eq:limber_cross} to include `massive neutrino bias' and `baryon feedback bias' as detailed in \citet{HWVH2015}.
The baryon bias was extracted from the OWL simulations, assuming the AGN model \citep{vanDaalen2011},
while the neutrino bias was extracted from the recalibrated {\small HALOFIT} code \citep{Takahashi2012} with {\color{black} total} neutrino masses ${\color{black} M_{\nu}}$ = 0.05, 0.2, 0.4 and 0.6eV.
Our results are presented in Fig. \ref{fig:frac_baryons} for two simplified cases,
in which the source galaxy populations are placed on single planes at $z_s = 0.5$ (in red) and at $z_s = 1.5$ (in blue).
The figure focuses on the 0.05eV scenario, showing the suppression of power caused by massive neutrinos (3.5\% effect on $A$ for both $z_s$ planes, averaged over the $\ell$-modes that we measured),
by baryonic feedback (5.0\% for $z_s = 1.5$ and 10.6\% for $z_s = 0.5$), and by the combination of both (8.2\% and 13.7\% for $z_s = 1.5$ and $0.5$ respectively).
The reason why the effect of baryons is larger on the lower redshift source plane is simply a projection effect: the same physical scales subtend different angles on the sky,
which contribute differently to our measurement restricted to the $\ell \in [20 -2000]$ range.
We also show the effect of 0.4eV neutrinos (28\% and 30\% in the two $z_s$ slices), which demonstrates a scaling of 7\% per 0.1eV.
We note that \citet{MeadFit} proposes an alternative method to account simultaneously for baryons and neutrinos based on the halo model,
which might prove useful in future work.
These two effects contribute at some level to the measurement of $A$, {\color{black} but it is too early to put constraints on them based on our measurement. Firstly,}
the cosmology is not guaranteed to be that of {\it Planck}, secondly the exact feedback mechanism that is at play in the Universe remains largely unknown,
and thirdly other effects (e.g. IA contamination or error in the $n(z)$) could explain why our measured $A_{\rm fid}$ is low.
However, if the fiducial cosmology is correct and if the intrinsic alignments are well described by the HT model described in Section \ref{subsec:IA},
then $A_{\rm fid}$ would be brought to unity with $M_{\nu} = 0.33{\color{black}\ \pm\ 0.22}$eV in absence of baryonic feedback, and $M_{\nu} = 0.19{\color{black}\ \pm\ 0.22}$eV within the AGN model.
{\color{black} We have verified that the uncertainty in the non-linear modelling does not affect our measurement of $A$ significantly.
This is mainly because the angles and redshifts probed by our measurement correspond to scales that are mostly in
the linear and mildly non-linear regime. Replacing the non-linear power spectrum from the \citet{Takahashi2012} model
with that of \citet{Smith2003}, a radical change in the non-linear predictions at small scales {\color{black} shown in Fig. \ref{fig:frac_baryons}} (black dashed line), affects our measurement of $A$ by 1-2\% only.
This is well within the statistical uncertainty and can be safely neglected.
Fig. \ref{fig:frac_baryons} shows that there is a clear degeneracy between differences in the two models, and the effect of baryonic feedback.
However, the \citet{Smith2003} predictions are known to suffer from a significant loss of power at small scales,
visible in Fig. \ref{fig:frac_baryons}, and the state-of-the-art precision on the non-linear power spectrum, from e.g. the Cosmic Emulator \citep{Heitmann2013} deviates from the \citet{Takahashi2012} model by less than 5\% \citep{MeadFit}.
This alleviates the degeneracy between modelling and baryonic feedback effects and further supports our (model-dependent) neutrino mass contraints presented above. }
\subsection{Cosmology from broad $n(z)$}
\label{subsec:cosmo_broad}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.5in]{./graphs/fig8.eps}
\end{tabular}
\caption{Constraints on $\sigma_8$ and $\Omega_{\rm m}$ as estimated from the cross-correlation measurement, ignoring potential contamination by intrinsic galaxy alignements
(shown in black). The solid line shows the best fit, while the dashed and dotted lines indicate the 68\% and 95\% confidence level (CL) regions, respectively.
The cross-correlation results can be compared to KiDS-450 (green, where IA are accounted for) {\it Planck} (orange) and WMAP9+SPT+ACT (blue).}
\label{fig:Cosmo}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.5in]{./graphs/fig9.eps}
\end{tabular}
\caption{Same as Fig. \ref{fig:Cosmo}, but here assuming 10\% contamination from IA in the cross-correlation measurement (equation \ref{eq:A_cosmo_HT}), consistent with both the `HT-IA' and the `$f_{\rm red}$-IA' models. }
\label{fig:Cosmo_IA}
\end{center}
\end{figure}
In this section, we investigate how our cross-correlation measurement can constrain cosmology, and specifically compute confidence regions
in the $[\sigma_8- \Omega_{\rm m}]$ plane. For this calculation we assume massless neutrinos, no baryonic feedback, we ignore the error on $n(z)$, but examine
our results for the three IA models described in Section \ref{subsec:IA}.
It was shown in \citet{2015PhRvD..92f3517L} that the amplitude of the cross-correlation signal scales approximately with
$[\sigma_8^{2} \Omega_{\rm m}^{-0.5}]$ at large, linear scales ($\ell<$ few hundred), and as $[\sigma_8^{3} \Omega_{\rm m}^{1.3}]$ at small scales ($\ell>$ 1000).
Most of our constraints come from small scales, but our measurement includes some large modes down to $\ell\sim200$.
For this reason, we strike a compromise: we keep the $\Omega_{\rm m}^{1.3}$ dependence, as suggested by \citet{2015PhRvD..92f3517L},
but use a $\sigma_8^{2.5}$ dependence, to capture the gradual transition between both.
Future measurements will require MCMC algorithms to be run to better capture these dependencies,
but this does not seem to be necessary in this case given the relatively large uncertainty on $A$.
As discussed before, we use the $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ results in the broad $Z_B \in [0.1, 0.9]$ tomographic bin
because it has the highest signal to noise, however our results would not change significantly if we used the $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$
measurement instead. We could also have used the tomographic results, i.e. the $A(z)$ in the four bins.
However these measurements are all correlated, probing common low-redshift lenses.
This would require us to calculate and include cross-correlation coefficients between the different tomographic bins
when solving for the best-fit cosmology. These could be evaluated from mock data, but this is not required when working with a single data point for $A$.
Combining this scaling relation with equations \ref{eq:A_fid} - \ref{eq:A_fid_HT}, we get
\begin{eqnarray}
A = A_{\rm fid} \left(\frac{\sigma_8}{0.831}\right)^{2.5} \left(\frac{\Omega_{\rm m}}{0.2905}\right)^{1.3} = 0.69\pm0.15
\label{eq:A_cosmo}
\end{eqnarray}
and
\begin{eqnarray}
A = A_{\rm fid}^{\rm HT} \left(\frac{\sigma_8}{0.831}\right)^{2.5} \left(\frac{\Omega_{\rm m}}{0.2905}\right)^{1.3} = 0.77\pm0.15,
\label{eq:A_cosmo_HT}
\end{eqnarray}
which we use to propagate the error on $A$ into confidence regions in the $[\sigma_8 - \Omega_{\rm m}]$ plane.
We show in Fig. \ref{fig:Cosmo} and \ref{fig:Cosmo_IA} how these constraints compare to the results from KiDS-450 cosmic shear {\color{black} (with IA)}, {\it Planck} and other CMB experiments\footnote{\color{black} The MCMC chains entering these contour plots can be found on the KiDS-450 website: {kids.strw.leidenuniv.nl/cosmicshear2016.php}. Note also that the WMAP9+SPT+ACT cosmology presented in Fig. \ref{fig:Cosmo} and \ref{fig:Cosmo_IA} differ from the fiducial WMAP9+BAO+SN cosmology.}.
Our cross-correlation measurement has a larger overlap with the KiDS-450 constraints but is still consistent with the {\it Planck} cosmology in the sense that their $\lesssim 95\%$ confidence regions overlap. Including IA reduces the offset from {\it Planck}.
Given that our signal has different dependences on cosmological (e.g. $\Omega_{\rm m}$, $\sigma_8$) and nuisance (e.g. $m$, $n(z)$) parameters,
we can see how this can provide new insights in resolving tensions between the cosmic shear and CMB measurements.
For example, whereas the KiDS-450 and CFHTLenS cosmic shear results scale as $[m^2 n^{2}(z)]$,
our KiDS-450 $\times$ {\it Planck} lensing measurement scales as $[m\,n(z)]$.
This difference could therefore allow us to break the degeneracy in a joint probe analysis.
Also note that in general, we should not exclude possibility that there could be residual systematics left over in a CMB temperature and polarization analysis -- driving the cosmology to higher $[\sigma_8, \Omega_{\rm m}]$ values --
that do not make their way to the CMB lensing map or into the joint probes measurement, in analogy with the cosmic shear $c$-term. This is certainly the case for the additive shear bias (the $c$-correction) described in \citet{KiDSLenS}.
Having this new kind of handle can help to identify the cause of disagreements between different probes, and will be central to the cosmological analyses of future surveys.
We explore further how cross-correlation analyses can be turned into a calibration tool in the next section.
\subsection{Application: {photo-$z$} and $m$-calibration}
\label{subsec:m}
The {\color{black} CMB lensing -- galaxy lensing cross-correlation signal} has been identified as a promising alternative to calibrate the cosmic shear data without relying completely on image simulations
\citep{2013arXiv1311.2338D, 2016PhRvD..93j3508L, 2016arXiv160701761S}. This statement relies on the fact that $A$ absorbs all phenomena that affect the amplitude of the measurement,
i.e. cosmology, intrinsic alignment, $n(z)$, shear calibration, and that we can marginalize over some of these in order to solve for others.
Most of the attention so far has been directed towards the multiplicative term in the cosmic shear calibration -- the $m^{j}$ factor in equation \ref{eq:K} -- which
has an important impact on the cosmological interpretation. In the case of the KiDS-450 data, the shear calibration is known at the percent level from image simulations
for objects with $Z_B<0.9$ (see Fig. \ref{fig:deltam_deltaz_low_z}), but the precision on $m^{j}$ quickly degrades at higher redshift \citep{2016arXiv160605337F}.
Similar conclusions can be drawn from the photometric redshift estimation, which becomes unreliable at high redshift when only using optical bands \citep{KiDS450}.
We see in our cross-correlation measurement a unique opportunity to place a joint-constraint on these two quantities in the highest redshift bin,
informed by our measurement at lower redshift. {\color{black} We ignore the contribution from IA, knowing that this will need to be included for upcoming surveys\footnote{
{\color{black} One might well object that the uncertainty in IA modelling and its evolution is already larger than the uncertainty in shear calibration,
and hence that our strategy is flawed to start with. Instead, we should be placing simultaneous constraints on the photo-$z$, $m$-calibration and IA.
A full MCMC will certainly be required in the future to disentangle these effects, exploiting their different shape dependence to break the degeneracy between these parameters.
As an illustration of this strategy however, we present a simple case and assume no IA contamination in this rest of this section.}}. }
Our approach is to fix the scaling term $A$ to the value preferred by the $Z_B \in [0.1, 0.9]$ data, which we label $A_{{\rm low}z}$ here for clarity, and to
jointly fit for the mean shear bias and mean redshift distribution in the $Z_B>0.9$ bin.
Forcing $A$ to this value in the high redshift bin provides constraints on $\langle m_{{\rm high}z}\rangle$ and $\langle n_{{\rm high}z}(z)\rangle$,
which we extract by varying these quantities in the predictions.
The correction to the shear bias is trivial to implement as it scales linearly with $A$,
so we simply write $A_{{\rm high}z} = A_{{\rm low}z}(1 + \delta_{m}) = 0.95 \pm 0.22$ (from Table \ref{table:stat}) and solve for $\delta_{m}$.
If this was the only correction, we could write $\delta_{m} = A_{{\rm high}z}/A_{{\rm low}z} - 1 = 0.38\pm0.44$,
which is consistent with zero but not well constrained.
Corrections to the photometric distribution can be slightly more complicated as the full redshift distribution that enters our calculation is not simple, as seen in Fig. \ref{fig:nz}.
There are a number ways with which we could alter the $n(z)$ and propagate the effect onto the signal, e.g. by modifying the overall shape, the mean or the tail of the distribution.
We opted for arguably the simplest prescription, which consists in shifting the $n(z)$ along the $z$ direction by applying the mapping $z\rightarrow z + \delta_z$
(thereby shifting $\langle n_{{\rm high}z}(z) \rangle$ by the same amount). We propagate this new $n(z)$ through equation \ref{eq:limber_cross} and solve for values of $\delta_z$ that satisfy constraints on $A$.
In this process, we allow $\delta_z$ to vary by up to 0.5, which is rather extreme.
Following the simple reasoning described for the shear calibration, we can see that if $m$ was trusted at the percent level in this high redshift bin,
constraints on the redshift distribution could be simply derived by computing $A_{{\rm high}z}/A_{{\rm low}z} = C_{\delta_z} / C_{\rm fid} = (1+\delta_z)$.
We therefore obtain the exact same constraints as for $\delta_m$, namely: $\delta_z=0.38\pm0.44$.
We place constraints on the [$\delta_m - \delta_z$] plane by requiring $(1+\delta_m)(1+\delta_z) = {A_{{\rm high}z}}/{A_{{\rm low}z}}$,
and present the $1\sigma$ constraints in Fig. \ref{fig:deltam_deltaz}. The data are still consistent with $\delta_m = \delta_z = 0$,
but these two biases are not currently well constrained.
{\color{black} We also show in Fig. \ref{fig:deltam_deltaz} the results from the $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ estimator (in red dashed),
but these have a lower SNR hence are not included in the analysis. At first sight, the difference observed between the results from the two estimators could seem worrisome.
Given that these constraints on [$\delta_m - \delta_z$] are obtained from the same data, and that the only difference is the analysis method,
it is justified to question whether we could use this measurement for precise self-calibration if two methods on the same data give such different values for $\delta_m$ and $\delta_z$.
We recall that differences are expected since both techniques are probing different scales, however the calibration technique presented here is sensitive to these differences.
The calibration is only weakly sensitive to the fiducial cosmology adopted, as shown in Fig. \ref{fig:deltam_deltaz} with the different line styles.
A significant improvement will come from the future data sets (advanced-ACT, SPT-3G, LSST, Euclid), in which the noise will be much lower, allowing for more accurate measurements of
$\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ and $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ to start with.
In addition, including other measurements in this self-calibration approach will greatly enhance the achievable precision.}
{\color{black} For example, one could measure the galaxy-galaxy lensing signal
from the same KiDS-450 source galaxies, using i.e. the GAMA galaxies as lenses \citep{2016MNRAS.459.3251V},
selecting the sources in the same tomographic bin (i.e. $Z_B \in [0.1, 0.9]$ and $Z_B>0.9$).
Fixing the cosmology from the low redshift bin, one could then similarly constrain [$\delta_m - \delta_z$] in the high redshift bin.
The idea here is that the trend can be made {\it opposite} to that seen in Fig. \ref{fig:deltam_deltaz}: an increase in $\delta_z$ pushes the sources away from the lenses, which,
depending on the geometry, could reduce the signal. To compensate for this, the $m$-calibration would need to increase as well.
In such a setup, the preferred region in parameter space would inevitably intersect with ours, and exploiting this complementarity might lead to competitive constraints.
Further investigation on this combined measurement will be explored in upcoming work.}
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig10.eps}
\end{tabular}
\caption{$1\sigma$ contour regions on the shear calibration correction $\delta_m$ and the redshift distribution correction $\delta_z$ in the bin $Z_B >0.9$, from the cross-correlation measurements.
Black and red correspond to constraints from $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}$ and $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$ respectively.
The multiple lines present the results in three different cosmologies (fiducial is solid, KiDS-450 is dot-dashed, {\it Planck} is dashed), which are shown here to have a small impact on the constraints.
Other independent measurements and improved image simulations could tighten the region of allowed values. The upper red solid and dot-dashed lines perfectly overlap.
}
\label{fig:deltam_deltaz}
\end{center}
\end{figure}
We are aware that our bi-linear modelling of the $m$ and $n(z)$ calibration is an over-simplification of our knowledge (and uncertainty)
about these quantities in the highest redshift bin, and one could envision {\color{black} improving this strategy} in the future.
For instance, the high-redshift objects are often the hardest to measure spectroscopically, hence there are higher chances that the DIR method fails at higher redshifts.
To capture this effect, instead of shifting the $n(z)$, one could modify only the high-redshift tail, moving 1\%, 5\% or 10\% of our source galaxies
from (very) low redshifts to $z>$1, propagating the effect on the signal, and use our measurement of $A$ to constrain the fraction of such `missing' high-redshift galaxies.
However, given the size of our error bars, it is not clear that we would learn more from this approach at the moment.
{\color{black} This situation will improve significantly with future CMB and galaxy surveys.
According to \citet{2016arXiv160701761S}, the lensing data provided by a Stage-4 CMB experiment, combined with
10 tomographic bins for LSST, will enable a $m$-calibration that is accurate to better than 0.5\%.
This is marginalising over a number of nuisance parameters that unfortunately does not include catastrophic photometric redshift outliers,
so the actual accuracy will likely degrade compared to this impressive benchmark.
Nevertheless, this is an avenue that is certainly worth exploiting with the upcoming data.}
{\color{black} The $Z_B<0.9$ redshift data in the KiDS survey has been calibrated on image simulations whose precision on $\delta_m$
largely surpasses that of the cross-correlation technique presented in this section.
Fig. \ref{fig:deltam_deltaz_low_z} shows the $1\sigma$ constraints in the $\left[\delta_m - \delta_z \right]$ plane in the $Z_B \in [0.1, 0.9]$ bin assuming the fiducial cosmology without IA (black),
the KiDS-450 cosmology without IA (blue), and the KiDS-450 with 10\% IA (red), consistent with both the HT-IA and the $f_{\rm red}$-IA models.
For comparison, the 1\% precision on $\delta_m$ obtained from image simulations and the 3\% precision on $\delta_z$ obtained from bootstrap resampling the $n(z)$, described in Section \ref{subsec:nz},
are shown as the pairs of horizontal and vertical lines, respectively. For these redshifts at least, the measurement provides interesting constraints on the cosmology, IA and $\delta_z$, but not on the $m$-calibration. }
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\hspace{-5mm}
\includegraphics[width=3.2in]{./graphs/fig11.eps}
\end{tabular}
\caption{$1\sigma$ contour regions on the shear calibration correction $\delta_m$ and the redshift distribution correction $\delta_z$ derived from the $Z_B \in [0.1, 0.9]$ measurement of $A$ in three different cases.
Results from $A_{\rm fid}$ (no-IA) are shown in solid black, results from $A_{\rm KiDS}$ (no IA) are shown in solid blue,
and results from $A_{\rm KiDS}$ with 10\% IA are shown in solid red.
The pair of solid horizontal lines show the region of $\delta_m$ values allowed from image simulations,
while the pair of dashed vertical lines show the region of $\delta_z$ values allowed from bootstrap resampling the $n(z)$. }
\label{fig:deltam_deltaz_low_z}
\end{center}
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
We perform the first tomographic lensing-lensing cross-correlation by combining the {\it Planck} 2015 lensing map with the KiDS-450 shear data.
Our measurement is based on two independent estimators, the {\small POLSPICE} measurement of $C_{\ell}^{\kappa_{\rm CMB} \kappa_{\rm gal}}$,
and the configuration-space measurement of $\xi^{\kappa_{\rm CMB} \gamma_{\rm t}}(\vartheta)$.
The two techniques agree within $1\sigma$ in all tomographic bins, although the former exhibits a lower signal to noise ratio.
{\color{black} We compare our tomographic results against a two-dimensional lensing analysis of a single broad redshift bin ($Z_B \in [0.1, 0.9]$),
and fit the measured amplitude of the signal with a single multiplicative parameter $A$ that scales the predictions.
We obtain $A_{\rm fid} = 0.69\pm0.15$ in our fiducial cosmology,
and show that the constraints on the $[\sigma_8 - \Omega_{\rm m}]$ plane are consistent with the flat $\Lambda$CDM {\it Planck} cosmology
at the 95\% level, with $A_{\it Planck} = 0.68\pm0.15$, and with all previous results \citep[][K16 \& HD16]{2015PhRvD..91f2001H, 2015PhRvD..92f3517L, 2017MNRAS.464.2120S}.
The KiDS-450 cosmology is preferred however, in which we obtain $A_{\rm KiDS} = 0.86\pm0.19$.}
Photometric redshifts have been examined carefully and are unlikely to be affecting these results significantly ($<8\%$ effect on the signal), unless the spectroscopic sample that is used to estimate the $n(z)$ suffers from significant sampling variance. Multiplicative shear calibration is also highly unlikely to be affecting $A$ since it is known to be accurate at the percent level over the redshift range that enters our cosmological measurement. However, including different models of intrinsic alignment, massive neutrinos and baryon feedback in the predictions all affect the signal by tens of percent, pushing the recovered $A$ {\color{black} to higher values}.
Fixing the cosmology to that favoured by our low-redshift measurements ($Z_B < 0.9$), we calibrate the high-redshift ($Z_B>0.9$) photometric $n(z)$
and the multiplicative shear calibration, which are {\color{black} not robustly constrained}.
We find that the high redshift data is consistent with no residual systematics, but that these are still allowed and only weakly constrained.
Improved results on this high redshift calibration will come in the future from larger data sets, from improved image simulations and from the combination with other independent measurements.
Tomographic measurements such as that presented in this paper are insensitive to galaxy bias,
and hence opening the possibility to obtain cosmological constraints from measurements of the growth factor.
Upcoming and future lensing surveys will have excellent opportunities for combining probes and improving their cosmological
analyses.
\section*{Acknowledgements}
Based on data products from observations made with ESO Telescopes at the La Silla Paranal Observatory under programme IDs 177.A-3016, 177.A-3017 and 177.A-3018, and on data products produced by Target/OmegaCEN, INAF-OACN, INAF-OAPD and the KiDS production team, on behalf of the KiDS consortium. OmegaCEN and the KiDS production team acknowledge support by NOVA and NWO-M grants. Members of INAF-OAPD and INAF-OACN also acknowledge the support from the Department of Physics \& Astronomy of the University of Padova, and of the Department of Physics of Univ. Federico II (Naples).
JHD acknowledges support from the European Commission under a Marie-Sk{\l}odowska-Curie European Fellowship (EU project 656869).
TT, LvW and AH are supported by the NSERC of Canada, TT is additionally funded by CIFAR.
HHi is supported by an Emmy Noether grant (No. Hi 1495/2-1) of the Deutsche Forschungsgemeinschaft.
MV, CH and MA acknowledge support from the European Research Council under FP7 grant number 279396 (MV) and 647112 (CH, MA).
MV is also supported by the Netherlands Organisation for Scientific Research (NWO) through grants 614.001.103.
Parts of this research were conducted by the Australian Research Council Centre of Excellence for All-sky Astrophysics (CAASTRO), through project number CE110001020.
JM has received funding from the People Programme (Marie Curie Actions) of the European Union's Seventh Framework Programme (FP7/2007-2013)
under REA grant agreement number 627288. ENC is supported by a Beecroft fellowship. PS is supported by the Deutsche
Forschungsgemeinschaft by the TR33 `The dark Universe'. LM is supported by STFC grant ST/N000919/1, and FK by a de Sitter Fellowship of the Netherlands Organization for Scientific Research (NWO) under grant number 022.003.013. KK acknowledges support by the Alexander von Humboldt Foundation.
{\footnotesize
{\it Author Contributions:} All authors contributed to the development and writing of this paper.
The authorship list reflects the lead authors of this paper (JHD, TT and ENC),
followed by two alphabetical groups.
{\color{black} The first alphabetical group (CH and LvW) consists of authors who contributed to both the infrastructure of the KiDS data that was used in this work, and to the analysis itself.
Members of the second group are infrastructure contributors whose products are directly used in this work.}
\bibliographystyle{hapj}
|
\section{INTRODUCTION}
In the richest countries, the population pyramid is turning upside down \cite{he2016aging}. In 2015, 8.5 percent of the world's population was aged 65 and over and, by 2050, this older population is projected to represent 16.7 percent of the world total population. To allow people to continue to have active and productive lives as they age, new technologies are being studied. Recently, as far as home robots are concerned, there has been many promising developments. New products like Softbank's Pepper have been introduced into the market and many research platforms, e.g. the healthcare robots Pearl \cite{pollack2002pearl}, ASTRO \cite{cavallo2013design}, Max \cite{gross2015robot}, Hobbit \cite{fischinger2016hobbit} or our prototype O-Robot \cite{carraro2015open}, have been proposed. Not only such robots aim at fostering research to keep the house safe by monitoring and detecting anomalies, but also at being friendly companions able to enhance the elderly people's social lives without invading their privacy. In particular, among all the sources of harm, falls are known to be the major one in elderly people \cite{lord2007falls}. In this work, given that it is unlikely for a robot to capture the act of falling while patrolling, the focus is on detecting people already lying on the floor.
The main contributions in this paper are:
\begin{itemize}
\item a real-time pure-3D approach to detect fallen people suitable for real cluttered scenes;
\item its integration with two basic robot functionalities, 2D mapping and navigation, in order to suppress false positives thanks to the a-priori knowledge of the environment and the availability of multiple view points;
\item our RGB-D dataset of fallen people\footnote{\url{http://robotics.dei.unipd.it/117-fall}} consisting of several static and dynamic sequences with 15 different people acquired in 2 different environments.
\end{itemize}
The remainder of the paper is organized as follows. Section \ref{related_work} reviews the work related to fall detection, people detection and body pose estimation. Section \ref{approach} describes our novel approach, first giving a picture of the entire workflow, then focusing on both the single-view approach and its integration with mapping and robot navigation. In Section \ref{results}, our dataset is described and our methods thoroughly evaluated. Finally, in Section \ref{conclusions}, conclusions are drawn and future directions of research identified.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{images/pipeline3.png}
\caption{The proposed approach is split into two separately running processes. The \textit{single-view detector} detects fallen people on the single frames in a way which proves to be fast and robust to clutter. The \textit{multi-view analyser} fuses the single-view results exploiting the availability of the 2D map and the multiple points of view explorable during the robot navigation. The final map includes also the semantic information about the location of the fallen people, see the red placeholders.}
\label{fig:overview}
\end{figure*}
\section{RELATED WORK}
\label{related_work}
Nowadays the wide adoption of Deep Neural Networks (DNN) is boosting the classification accuracy in many fields. In particular, many recent works~\cite{angelova2015real,cao2016realtime,wei2016cpm} address the person detection and body pose estimation problems showing great results. This kind of algorithms could be used also for detecting people lying on the floor. Nevertheless, their recognition capabilities are limited to RGB images and so they cannot work in dimmer scenes, which are usual in real life houses. In addition, the high complexity of the DNN requires the algorithm to be accelerated by using high-end Graphical Processing Units (GPU) in order to achieve real-time performances useful in real applications. For these reasons, those networks do not fit our application. Indeed, we are proposing techniques which can work also without the presence of the color information (e.g. under different illumination conditions or during the night). Moreover, we want to keep the power requirements at a minimum, given that this is a major issue in the design of mobile robots. Thus, the usage of an high-end GPU is unsuitable. Our approach draws upon two recent methods for the semantic segmentation of scene structures and objects from RGB-D data \cite{wolf2015fast, wolf2016enhancing}. Both approaches are almost real-time and based on fast features calculated on 3D patches or clusters. They also try to learn contextual relations among them, respectively by means of Conditional Random Fields and 3D Entangled Forests.
There exist also more specific approaches addressing the detection of falls. These comprehend wearable devices, whose great popularity is linked to the spread of open-source platforms which are small, powerful and connectible to low-cost sensors \cite{perry2009survey}. In most cases, such sensors include accelerometers \cite{li2009accurate, boyle2008simulated, lindemann2005evaluation}. These technologies suffer from the difficulty of correctly distinguishing falls from common actions like sitting or lying down. Furthermore, the elderlies easily forget to wear them. Other approaches specifically addressing falls need the installation of environmental devices like microphones \cite{popescu2008acoustic}, cameras for person tracking \cite{williams2007aging,cucchiara2007multi,ghidoni2014distributed}, infrared or vibration sensors \cite{yazar2014multi}. Anyway, these approaches are less effective and, being invasive, less accepted.
To the best of our knowledge, there exist just a few approaches trying to detect fallen people already lying on the floor: \cite{wang2012lying, volkhardt2013fallen, nishi2015head}. Both \cite{wang2012lying} and \cite{volkhardt2013fallen} are specifically designed for mobile robots. In \cite{wang2012lying}, the authors propose a pipeline working on just single RGB images extending a deformable part-based model to the multi-view case for viewpoint invariant lying posture detection. Like us, \cite{volkhardt2013fallen} proposes a pipeline working on single depth images. Putative candidates are found by means of a segmentation phase based on an Euclidean clustering. Then, they are layered so as to face with occlusions and classified by means of a SVM using Histograms of Local Surface Normals. The downside of the approach is the Euclidean segmentation, in particular its distance threshold: if people fall on or near furniture, the segmented object may contain the user and parts of the furniture. On the contrary, this work specifically addresses this problem by concatenating two classifiers. Unfortunately, neither the code or dataset of \cite{volkhardt2013fallen} are available making a direct comparison impossible. Finally, in \cite{nishi2015head}, a method for detecting and locating the head of a person lying on the floor by means of a RGB-D sensor is proposed. It would allow to test vital signs on the fallen people, but has not been tested in real cluttered scenarios and requires the head to be visible. Remarkably, none of the previous approaches take advantage of the other functionalities available thanks to the mobile robot like 2D mapping, i.e. the actual knowledge of the environment, and navigation, i.e. the availability of multiple view points.
\section{APPROACH}
\label{approach}
An overview of the proposed approach for detecting people lying on the floor is given in Figure \ref{fig:overview}. It is decoupled into two separately running processes, the \textit{single-view detector} and the \textit{multi-view analyser}. The former process, the \textit{single-view detector}, operates on pure-3D Point Clouds generated by a RGB-D sensor such as the Kinect One V2, which, in our experiments, is mounted on a mobile robot 1.16\,m off the floor and parallel to it. First, the input cloud is preprocessed in order to restrict the subsequent phases to work on a region of interest comprehending all the objects above the floor and below a maximum height. Then, the pre-processed cloud is over-segmented into small patches of voxels with similar appearance. In a two-phase classification step, the patches are classified as part of person or not and gathered together. The use of the Euclidean clustering on the cloud including only the person patches makes it possible to handle also cluttered scenes. Finally, to further improve performances, the latter process, the \textit{multi-view analyser}, rejects all the detections not belonging to the free space of the 2D map and accumulates the detections from several frames by taking into account their 2D map positions and timestamps. Each phase is deeply discussed in the next subsections: Subsection \ref{subsec:patchbaseddetectionoffallenpeople} deals with the description of the \textit{single-view detector} while Subsection \ref{subsec:map_verification} and Subsection \ref{subsec:multiplevantagepoints} describe the \textit{multi-view analyser}.
\subsection{Patch-based Detection of Fallen People}
\label{subsec:patchbaseddetectionoffallenpeople}
Each point cloud is pre-processed to restrict the analysis to a region of interest and reduce the data noise. First of all, the point cloud is truncated to a 3D region containing the floor and the points between it and a maximum height of 0.66\,m. Then, the floor is removed with an approach based on the RANSAC segmentation \cite{fischler1981random}. To improve its robustness to the robot motion, two floor planes are estimated, on a first half of the cloud close to the robot and on a second half far from the robot. In particular, a good split distance proved to be 3\,m. Finally, to reduce the data noise without affecting the running time, a soft statistical outlier removal is applied with the number of neighbours set to 50 and the standard deviation set to 0.3.
The core of the algorithm draws upon two recent works about the semantic segmentation of objects and scene structures \cite{wolf2015fast,wolf2016enhancing}. It comprehends the following 4 phases:
\begin{enumerate}
\item supervoxel over-segmentation in 3D patches;
\item classification of each 3D patch as positive, i.e. part of a fallen person or negative, i.e. not part of a fallen person;
\item clustering of positive patches;
\item classification of each cluster as positive, i.e. a fallen person, or negative, i.e. not a fallen person.
\end{enumerate}
They allow to segment and classify correctly also the people lying close to other objects or scene structures. In the following, each phase is described.
The pre-processed point cloud is over-segmented into homogeneous 3D patches by means of the Voxel Cloud Connectivity Segmentation (VCCS) \cite{papon2013voxel}. An example of over-segmented cloud is reported in Figure \ref{fig:supervoxelsegmentation}. This solution preserves the edges by finding patches not crossing object boundaries and, at the same time, it reduces the noise and the amount of data. The set of parameters used here is: voxel resolution 0.06\,m, seed resolution 0.12\,m, color importance 0.0, spatial importance 1.0 and normal importance 4.0. The voxel resolution is a good trade off between speed and having a sufficient number of points per patch. The seed resolution is a good trade off between having big patches and over-segmenting also the thinner body elements, e.g. arms and legs. The others are suggested in \cite{stein2014convexity}. As the proposed approach does not rely on RGB data, color is not considered at all by setting the color importance to 0.0.
\begin{figure}[!ht]
\centering
{\includegraphics[width=\columnwidth]{images/supervoxels.png}}
\caption{An example of pre-filtered and over-segmented cloud. A random color is assigned to each patch. The person is lying in the center.}
\label{fig:supervoxelsegmentation}
\end{figure}
For each patch generated by the over-segmentation, a feature vector $\mathbf{x_1}$ of length 16 is calculated. The choice of the features is based on the semantic segmentation works \cite{wolf2015fast,wolf2016enhancing}, whose presented features proved to be as fast as effective. Here, the color features are left out and only the geometric features are taken into account. Some of them are calculated from the eigenvalues of the scatter matrix of the patch, $\lambda_0 \leq \lambda_1 \leq \lambda_2$ while others from the Oriented Bounding Box (OBB) including all the patch points. The complete list is given in Table \ref{features1}.
\begin{table}[h]
\caption{List of features calculated for each 3D patch and their dimensionality.}
\label{features1}
\begin{center}
\begin{tabular}{c c}
\hline
Features & Dimensionality\\
\hline
Compactness ($\lambda_0$) & 1\\
Planarity ($\lambda_1 - \lambda_0$) & 1\\
Linearity ($\lambda_2 - \lambda_1$) & 1\\
Angle with floor plane (mean and std. dev.) & 2\\
Height (top, centroid, and bottom point) & 2\\
OBB dimensions (width, height and depth) & 3\\
OBB face areas (frontal, lateral and upper) & 3\\
OBB elongations ($\frac{height}{width}$, $\frac{depth}{width}$, $\frac{height}{depth}$) & 3\\
\hline
Total number of features & 16\\
\hline
\end{tabular}
\end{center}
\end{table}
To calculate the predicted label (part or not part of a fallen person) for each patch, this feature vector is then passed to a binary SVM classifier. After k-fold validation, a Radial Basis Function (RBF) kernel with the misclassification cost $C$ equal to 62.5 and the bandwidth $\gamma$ equal to 0.51 turned out to be the best performing solution. Of course, having each patch classified as part of a person (positive) or not (negative) does not suffice to detect a fallen person. Indeed, as shown in Figure \ref{fig:corerecap}(a), given that this classifier analyses just small patches, there can be false positives and false negatives. Because of this, two further steps, explained in the following and sketched in Figure \ref{fig:corerecap}(b)(c), have been developed in order to find 3D regions with a high density of positive patches and whose size is comparable to that of a person.
\begin{figure}[!ht]
\centering
\subfloat[]
{\includegraphics[width=0.33\columnwidth]{images/greenAndRed.png}}
\subfloat[]
{\includegraphics[width=0.33\columnwidth]{images/onlyGreenBig.png}}
\subfloat[]
{\includegraphics[width=0.33\columnwidth]{images/greenBB.png}}
\caption{The last three steps of the algorithm core: a) The first SVM classifies each patch as a person part (green color) or not (red color); b) Euclidean clustering of the positive patches; c) Calculation of the cluster OBB. The second SVM classifies each cluster as a person or not. Here, the response is positive.}
\label{fig:corerecap}
\end{figure}
In contrast to the methods in \cite{volkhardt2013fallen}, having two sets of patches respectively with the positive and negative ones opens up the possibility to apply the Euclidean cluster extraction without the risk of segmenting a fallen person together with the adjacent scene elements. First of all, some false positive patches can be easily recognized, e.g. all the patches with less than 5 neighbouring positive patches in a radius of 0.5 \,m can be filtered out. Then, the negative patches are pushed aside, and the Euclidean clusters are extracted from the point cloud of the remaining positive patch centroids using a large distance threshold of 1.0\,m.
For each cluster, its OBB is calculated. Thus, depending on the OBB dimensions and the number of positive and negative patches in it, each cluster may be a fallen person or not. For each cluster, a feature vector $\mathbf{x_2}$ of size 9 has been devised. The complete list of features is given in Table \ref{features2}. In particular, the sample distances to the separating hyperplane returned by the former SVM turned out to be really useful. They have been exploited by means of an histogram with 4 bins for the distance intervals $[0, 0.25)$, $[0.25, 0.5)$, $[0.5, 1)$ and $[1, \infty)$. For each cluster, each histogram bin is filled with the positive patches whose distance to the hyperplane falls in the respective interval. Thus, the number of positive patches in each bin/interval gives 4 additional features. The whole feature vector is passed to a binary SVM classifier. After k-fold validation, a RBF kernel with the misclassification cost $C$ equal to 312.5 and the bandwidth $\gamma$ equal to \num{2.25e-3} turned out to be the best performing solution.
\begin{table}[h]
\caption{List of features calculated for each 3D cluster and their dimensionality.}
\label{features2}
\begin{center}
\begin{tabular}{c c}
\hline
Features & Dimensionality\\
\hline
OBB dimensions (width, height and depth) & 3\\
Number of positive patches & 1\\
Percentage of positive patches & 1\\
4-bin histogram of positive patch confidences & 4\\
\hline
Total number of features & 9\\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Map Verification}
\label{subsec:map_verification}
A mobile robot navigates through the environment thanks to the information of two maps: a static one necessary to compute a collision-free plan with static objects, e.g. walls or furniture items, and a dynamic one necessary to avoid moving obstacles, e.g. people. In this work, the static map, which is usually acquired only once and for all, is exploited to implement a false positive rejection phase. Let the static map be defined as a set of cells $S = \{Cell_i, 0 \leq i \leq N\}$, where:
\begin{equation}
Cell_i = \begin{cases}
-1 & \textnormal{unknown content}\\
0 & \textnormal{free space}\\
0 < n \leq 1 & \textnormal{probability to be occupied}\\
\end{cases}
\end{equation}
Thanks to the transformations computable with a 2D SLAM algorithm like \cite{grisetti2005improving,grisetti2007improved}, each single-view detection can be transformed from the camera coordinate system to the map coordinate system and projected to a cell map $Cell_i$. If the $Cell_i$ value is unknown ($-1$) or occupied by a static obstacle ($K \leq Cell_i \leq 1$ with $K = 0.30$), then the detection can be easily rejected.
An example of successful false positive rejection is shown in Figure \ref{fig:mapverification}, in which a single-view detections falls on the static furniture, in this case a tree trunk. Indeed, given its geometric similarity to a lying person, the single-view algorithm may detect it as a person. The map verification allows to reject it, enhancing the final detection performances. This step handles also other challenging situations, like shelf glass surfaces which can be really noisy.
\begin{figure}[!htb]
\centering
\subfloat[\label{subfig:mapverificationa}]{%
\includegraphics[width=.52\columnwidth ]{images/Annotatore4.png}%
}
\subfloat[\label{subfig:mapverificationb}]{%
\includegraphics[width=.48\columnwidth ]{images/trunkrejection.png}%
}
\caption{An example of successful false positive rejection performed by the map verification step: a) shows the furniture item raising some false positives, a tree trunk (in green) b) shows that these detections (the blue squares on the right) are located onto the map occupied space.}
\label{fig:mapverification}
\end{figure}
\subsection{Merging Detections from Multiple Vantage Points}
\label{subsec:multiplevantagepoints}
The map is not the only robot feature that can enhance the detection performances. Indeed, in a typical scenario, the robot is patrolling a known environment. Thus, given that the location of each fallen person is mostly static, all the single-view detections available from the multiple points of view can be easily tracked. A detection may be a false positive from a certain view, while a true negative from many others. Moreover, the false positive detection rate is very low compared to the true positive one. Given these two facts, another contribution of this work is the exploitation of the detections available from the different vantage points.
After the map verification, the single-view detections are already expressed in the map reference system. In this section, an algorithm able to cluster or reject each of them is devised. Its output is a set $\mathfrak{P}$ of validated lying person locations $p_i$ in the map, formally $\mathfrak{P} = \{p_i, 0 \leq i \leq g\}$, where $g$ is the total number of people. Given each new detection $d = (loc, t)$, where $d.loc$ is its location in the map coordinate system and $d.t$ its timestamp, the set of clusters, formally $\mathfrak{C} = \{C_i: 0 \leq i \leq n\}$, is updated with the following rule:
\begin{equation}
C_i = \{d_j: ||d_j.loc - d_m.loc|| < \overline{th}, \forall j,m \in [0,k-1] \} \text{,}
\end{equation}
in which $\overline{th}$ is a user-defined threshold which indicates if a detection is close enough to be considered in the cluster or not, and $k$ the number of detections in the cluster.
The set $\mathfrak{P}$ of fallen people is computed by a fixed-time periodic thread which analyses the set $\mathfrak{C}$. It updates the set $\mathfrak{P}$ by deleting the old detections and analysing the new ones in $\mathfrak{C}$. Indeed, in order to maintain a lightweight representation of $\mathfrak{C}$ and reject the false positives, whose frame rate is typically low, the old detections are discarded and a further check on the timestamp is performed. The pseudo-code of the whole procedure is reported in Algorithm \ref{alg:dofp_callback}, in which $\widehat{f}$ is the minimum detection frequency, $\widehat{t}$ is the maximum detection age and $\widehat{n}$ is the minimum number of detections in a cluster. Lines~\ref{alg:line:removal1}-\ref{alg:line:removal2} handle the time-based rejection on the basis of the maximum allowed age, while Lines~\ref{alg:line:classification1}-\ref{alg:line:classification2} reject the clusters whose detections have a low frame rate or are less than the minimum allowed.
In our implementation, we used $\overline{th}$ equal to 1\,m, $\widehat{t}$ equal to 60\,s, $\widehat{f}$ equal to 1\,Hz and $\widehat{n}$ equal to 5. The use of the frame rate allows to set a low $\widehat{n}$, thus preventing over-fitting. The procedure is invoked by the periodic thread every 10 seconds. In Figure~\ref{fig:detection_multiframe}, the algorithm is shown in action.
\begin{algorithm}
\caption{Cluster validation for detecting fallen people exploiting multiple vantage points}\label{alg:dofp_callback}
\begin{algorithmic}[1]
\Procedure{VALIDATE\textunderscore CLUSTERS}{$\mathfrak{C}$, $\mathfrak{P}$, $\widehat{t}$, $\widehat{f}$, $\widehat{n}$}
\ForEach {$C_i \in \mathfrak{C}$}
\For {$j \in [0,k - 2]$} \label{alg:line:removal1}
\For {$o \in [j + 1, k - 1]$}
\If {$|d_o.t - d_j.t| > \widehat{t}$}
\State {$index \gets $ARG\textunderscore MIN$(d_o.t, d_j.t)$}
\State {$C_i \gets C_i \setminus d_{index}$}
\EndIf
\EndFor
\EndFor \label{alg:line:removal2}
\State {$t_m \gets \min_{d \in C_i}\{d.t\}$} \label{alg:line:classification1}
\State {$t_M \gets \max_{d \in C_i}\{d.t\}$}
\State {$f_i \gets $ $\frac{||\mathfrak{C}||}{t_M - t_m}$}
\If {$f_i \geq \widehat{f} \text{ and } ||C_i|| \geq \widehat{n}$}
\State {$loc_i \gets \sum_{d \in C_i} \frac{d.loc}{||C_i||}$}
\State {$\mathfrak{P} \gets \mathfrak{P} \cup loc_i$}
\State {$\mathfrak{C} \gets \mathfrak{C} \setminus C_i$} \label{alg:line:classification2}
\EndIf
\EndFor
\State {\Return $\mathfrak{C}$, $\mathfrak{P}$}
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{figure}[!htb]
\centering
\subfloat[\label{subfig:detection_multiframe}]{%
\includegraphics[width=\columnwidth]{images/multiviewalgorithm.png}%
}
\\
\caption{The single-view detections projected on the 2D map are analysed by the \textit{multi-view analyser}. If they meet both the distance and time criteria, they are clustered. The white points compose the input point cloud, the blue cubes are the projected detections, here rejected false positives, and the coloured cylinders are the validated detection.}
\label{fig:detection_multiframe}
\end{figure}
\section{RESULTS}
\label{results}
The detection of fallen people is a challenging problem also because of the lack of public datasets. For this reason, another contribution of this work is the release of the IASLAB-RGBD Fallen Person Dataset\footnote{\url{http://robotics.dei.unipd.it/117-fall}}. On it, 4 common metrics, the detection $accuracy$, $precision$, $recall$ and $F_{0.5}$ score, are evaluated for each presented method. If $TP$, $TN$, $FP$ and $FN$ are the true positives, true negatives, false positives and false negatives, then these metrics are defined as in the following:
\begin{equation}
accuracy = \frac{TP + TN}{TP + FP + TN + FN}
\end{equation}
\begin{equation}
precision = \frac{TP}{TP + FP}
\end{equation}
\begin{equation}
recall = \frac{TP}{TP + FN}
\end{equation}
\begin{equation}
F_{0.5} = \frac{(1 + 0.5^2)*precision*recall}{0.5^2*precision + recall} \text{,}
\end{equation}
where the $F_{0.5}$ score, already proposed in \cite{volkhardt2013fallen}, is an harmonic average of precision and recall promoting an high precision, i.e. a low number of false positives. In addition, given the impossibility to compare with other existent and similar approaches, the baseline to which our algorithms are compared is a simple approach based on the Euclidean cluster extraction. This way, it will be clear how important the use of patches is in order to handle cluttered scenes. Finally, a detailed analysis of the running times is provided.
\subsection{IASLAB-RGBD Fallen Person Dataset}
This dataset consists of several RGB-D frame sequences containing 15 different people. It has been acquired in two different laboratory environments, the \textit{Lab A} and the \textit{Lab B}, by means of a Microsoft Kinect One V2, placed on a pedestal or on our mobile robot. The \textit{Lab A} is bigger and useful to test whether the algorithm can find people in the full distance range of the sensor (up to 5\,m). The \textit{Lab B} is smaller and more similar to a real domestic scenario. It is more cluttered and contains a sofa. It comprehends also glass surfaces which can be very noisy. For the sake of explanation, the dataset can be divided into three parts:
\begin{enumerate}
\item Part 1 includes 360 RGB-D frames acquired from 3 static pedestals. It is composed of several views of 10 people, which have been asked to lie in 12 different poses, 6 from the back and 6 from the front. Each person has been manually segmented in 3D;
\item Part 2 includes 4 sequences of RGB-D frames, for a total of 15932 frames, acquired from a mobile robot during its patrolling task in the \textit{Lab A}. People lie in 4 different fixed locations;
\item Part 3 includes 4 sequences of RGB-D frames, for a total of 9391 frames, acquired from a mobile robot during its patrolling task in the \textit{Lab B}. People lie in 4 different fixed locations.
\end{enumerate}
Training and test splits are also available. Some images of the dataset will be reported when discussing the results even if our approach does not exploit the RGB info.
The first classifier of the \textit{single-view detector} has been trained on thousand of patches extracted from the frames in Part 1 and Part 2 and tested on patches extracted from the frames in Part 1 and 3. All the positive samples have been taken from Part 1. The 70-30 train-test split of the segmented fallen people in Part 1 is also available. Negative samples have been taken from the \textit{Lab A} (just 24 frames out of 15932), the \textit{Lab B} (just 32 frames out of 9391) and the NYU Depth Dataset V2~\cite{Silberman2012indoor} (just 35 out of 1449), which contains thousands of indoor scenes for scene understanding. Only some of the negative samples have been used for balancing the number of positive and negative samples.
The second classifier of the \textit{single-view detector} has been trained on clusters extracted from the frames in Part 2 (\textit{Lab A}) and tested on clusters extracted from the frames in Part 3 (\textit{Lab B}). Approximately, for the training, the 15\% of all the available frames has been considered.
Not only the \textit{single-view detector} but also the \textit{multi-view analyser} has been tested on Part 3. Indeed, both Part 2 and 3 comprehend the entire robot transformation tree. Given that the position of the fallen people in the 2D map is known, this allows to calculate the performance indices automatically by checking if the location of the detected cluster centroid is close (at a distance less or equal to 1\,m) to the ground truth centroid of a person position in the 2D map.
\subsection{Validation}
The presented methods have been quantitatively evaluated on the IASLAB-RGBD Fallen Person Dataset. First, the separated evaluation of each classifier is presented. Then, the entire pipeline has been evaluated on both rooms, the \textit{Lab A} and the \textit{Lab B}. As previously explained, both the classifiers have been trained on just a part of the frames in the \textit{Lab A} while they see the \textit{Lab B} for the first time. In particular, we present the results for each of the 3 contributions, the \textit{single-view detector} and the two modules of the \textit{multi-view analyser}: the map validation and the detection merging from multiple vantage points. Furthermore, given the impossibility to compare directly with \cite{volkhardt2013fallen}, the comparison baseline (B) is a simple approach not exploiting patches. It finds putative clusters by means of the Euclidean cluster extraction with a distance threshold of 0.10\,m, which is really low considering a voxel resolution of 0.06\,m and far less than the one required by our approach (1\,m). The baseline classifies then each cluster on the basis of its position and its OBB size.
As previously mentioned, both classifiers of the \textit{single view detector} have been trained and tested on two different dataset splits. In both cases, K-fold validation with $K$ equal to 10 has been performed on the training set in order to find the optimal misclassification cost $C$ and bandwidth $\gamma$ values of the RBF kernel. As a preliminary evaluation, the SVM performances on the respective test sets are reported in Table~\ref{classifierperformances}.
\begin{table}[h]
\caption{Performances of the two classifiers on their test sets.}
\label{classifierperformances}
\begin{center}
\begin{tabular}{c c c c c}
\hline
Method & Accuracy & Precision & Recall & $F_{0.5}$\\
\hline
Classifier 1 (C1) & 0.89 & 0.93 & 0.84 & 0.91\\
Classifier 2 (C2) & 0.93 & 0.86 & 0.95 & 0.88\\
\hline
\end{tabular}
\end{center}
\end{table}
The results of the quantitative comparison between all the methods are shown in Table~\ref{methodcomparison1} and \ref{methodcomparison2}. Thanks to the patches, our methods outperform the baseline, not only in precision but also in recall. Furthermore, the map validation can further improve performances by rejecting some false positives.
\begin{table}[h]
\caption{Performance comparison on the \textit{Lab A}.}
\label{methodcomparison1}
\begin{center}
\begin{tabular}{c c c c c}
\hline
Method & Accuracy & Precision & Recall & $F_{0.5}$\\
\hline
Baseline (B) & 0.88 & 0.65 & 0.33 & 0.54\\
Single-view (SV) & 0.90 & 0.77 & 0.78 & 0.77\\
SV + Map verification (MV) & 0.92 & 0.87 & 0.77 & 0.85 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{Performance comparison on the \textit{Lab B}, never seen before by both classifiers.}
\label{methodcomparison2}
\begin{center}
\begin{tabular}{c c c c c}
\hline
Method & Accuracy & Precision & Recall & $F_{0.5}$\\
\hline
Baseline (B) & 0.84 & 0.64 & 0.26 & 0.50\\
Single-view (SV) & 0.89 & 0.87 & 0.74 & 0.83\\
SV + Map verification (MV) & 0.90 & 0.92 & 0.72 & 0.87 \\
\hline
\end{tabular}
\end{center}
\end{table}
As shown in Table~\ref{multiviewperformance}, also the detection merging from multiple vantage points proved to be useful. It has been tested on each one of the eight frame sequences acquired in the two environments. Each time, even if the environment is the same, the navigation path can differ due to dynamic obstacles and the different positions of the lying people on the floor. After the 4 patrolling tasks of the \textit{Lab A}, each person is always detected and only once, a false positive is still present while, after the 4 patrolling tasks of the \textit{Lab B} (never seen before by both classifiers), each person is always detected and all the false positives are successfully rejected.
\begin{table}[h]
\caption{Performances of the \textit{multi-view analyser} on both environments. Each time, even if the environment is almost the same, the robot path can differ because of dynamic obstacles and different positions of the lying people on the floor.}
\label{multiviewperformance}
\begin{center}
\begin{tabular}{c c c c c}
\hline
Environment & TP/P & FP\\
\hline
Lab A (sequence 1) & 4/4 & 0 \\
Lab A (sequence 2) & 4/4 & 1 \\
Lab A (sequence 3) & 4/4 & 0 \\
Lab A (sequence 4) & 4/4 & 0 \\
Lab B (sequence 1) & 4/4 & 0 \\
Lab B (sequence 2) & 4/4 & 0 \\
Lab B (sequence 3) & 4/4 & 0 \\
Lab B (sequence 4) & 4/4 & 0 \\
\hline
\end{tabular}
\end{center}
\end{table}
Finally, in Figure \ref{fig:qualitative_results}, some qualitative results are reported. They show the ability of the \textit{single-view detector} to find people in cluttered environments, see Figure \ref{fig:qualitative_results}(a)(b)(c)(d). Two difficult cases due to close objects or noisy regions, like glass surfaces, are also reported, see \ref{fig:qualitative_results}(e)(f). Anyway, they are easily handled by the \textit{multi-view detector}, \ref{fig:qualitative_results}(g)(h).
\begin{figure*}[!htb]
\centering
\subfloat[\label{subfig:detection_multiframe_a}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487936502_775865.jpg}%
}
\subfloat[\label{subfig:detection_multiframe_b}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487934667_637942.jpg}%
}
\subfloat[\label{subfig:detection_multiframe_b}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487934554_500764.jpg}%
}
\\
\subfloat[\label{subfig:detection_multiframe_a}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487934683_296158.jpg}%
}
\subfloat[\label{subfig:detection_multiframe_b}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487932496_241913.jpg}%
}
\subfloat[\label{subfig:detection_multiframe_b}]{%
\includegraphics[width=0.33\textwidth]{images/compressed_selection/1487932543_277975.jpg}%
}
\\
\subfloat[\label{subfig:detection_multiframe_a}]{%
\includegraphics[width=0.44\textwidth]{images/cad_1_detections_final_algo_hist.png}%
}
\subfloat[\label{subfig:detection_multiframe_b}]{%
\includegraphics[width=0.33\textwidth]{images/lab_4_detections_final_algo_hist_cropped.png}%
}
\caption{Qualitative results on the IASLAB-RGBD Fallen Person Dataset: (a)(b) even if the lying people can be very close to the wall or other scene elements, the \textit{single-view detector} can find them at a high detection rate; (c)(d) the \textit{single-view detector} can discard fake lying people, see the white circles; (e)(f) the \textit{single-view detector} may find some false positives in the presence of clutter (several close objects) or high noise (glass surfaces); (g)(e) the \textit{multi-view analyser} can reject both FP like in (e) thanks to the low frame rate or in (f) thanks to the map validation.}
\label{fig:qualitative_results}
\end{figure*}
\subsection{Runtime Analysis}
In Table~\ref{runtimes}, the running times of \textit{single-view detector} are reported. The algorithm is very efficient in terms of computing time proving to be an optimal choice for a mobile robot. Even if it is not yet fully parallelized, it can work in real-time at an average speed of 7.72\,fps. The test machine is a Dell Inspiron 15 7000 with an Intel Core i7-6700HQ CPU with 4 cores clocked at 2.60GHz, 16 GB of RAM and Linux Mint 17.3. Given that the \textit{multi-view analyser} is a daemon running in the background, its running times are of no interest and thereby not reported.
\begin{table}[h]
\caption{Average runtimes of the main steps of the proposed algorithm on our test machine (Intel Core i7-6700HQ CPU, 2.60GHz x 4).}
\label{runtimes}
\begin{center}
\begin{tabular}{c c}
\hline
Processing Stage & Runtime\\
\hline
Pre-processing and Oversegmentation & 10.27\,fps\\
Patch Feature Extraction & 105.98\,fps\\
SVM Classification 1 (per patch) & 0.84 \,$\mu$s \\
Cluster Feature Extraction & 2639.56\,fps\\
SVM Classification 2 (per cluster) & 0.04 \,$\mu$s \\
\hline
Total runtime & 7.72\,fps\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{CONCLUSIONS}
\label{conclusions}
This paper presented a real-time and robust approach to detect fallen people lying on the floor in various positions and from different distances. A single-view algorithm, which draws upon recent developments in the semantic segmentation field and does not need restrictive distance thresholds to segment putative clusters, was fully integrated on a mobile robot. The map of the environment and the availability of many different vantage points allowed to reduce the number of false positives, further improving the final performances. The algorithms here presented were thoroughly validated on the IASLAB-RGBD Fallen Person Dataset, which was published online for the benefit of the research community. They clearly outperform a simple method based on a finer distance threshold. In the near future, we would like to validate not only the ability of the algorithm to detect, but also to semantically segment fallen people.
We also plan to extend the test bed with sequences taken from real apartments along with different navigation paths. Finally, it would be interesting to merge close similar patches before their classification in order to analyse bigger segments.
\section*{ACKNOWLEDGEMENT}
The research leading to these results has been partially supported by Omitech Srl. The authors would like to thank M. Munaro and S. Ghidoni for valuable discussions.
|
\section{Introduction}
For many tasks, robots need to understand the 3D structure and semantics of their environment. For example, recognizing the free space and surfaces in a scene helps motion planning in robot navigation and manipulation tasks. Semantic understanding, beyond pure geometry, enables a robot to reason about objects, which is particularly important for manipulation and human robot interaction tasks. Over the last years, various techniques have been proposed for dense 3D scene reconstruction using depth cameras, including RGBD-Mapping, KinectFusion, Kintinuous, and ElasticFusion \cite{henry2012rgb,newcombe2011kinectfusion,Whe12Kin,whelan2015elasticfusion}. These methods jointly reconstruct the 3D scene and track the camera position from RGB-D videos. However, they do not provide semantic information about the scene. In parallel, different approaches for recognizing scene semantics have been proposed. These include methods in object detection \cite{felzenszwalb2008discriminatively,girshick2014rich}, object pose estimation \cite{brachmann2014learning,savarese20073d,xiang2012estimating}, and semantic labeling \cite{ren2012rgb,long2015fully}. Most of these methods focus on detecting specific objects or on recognizing scene elements in individual 2D images.
The goal of our work is to use RGB-D videos to reconstruct and label every observed surface element in a 3D scene, providing dense information about small objects, such as bowls and mugs, and larger objects such as tables and chairs. In such a setting, the key question is how the information from the RGB-D frames can be combined to improve recognition accuracy. Recent approaches handle this by incorporating recognition results from individual RGB-D frames into a 3D model, possibly followed by additional reasoning over the 3D structure~\cite{Lai13Obj,lai2014unsupervised,mccormac2016semanticfusion}. However, in these approaches, the reasoning about individual frames and their information accumulation is only loosely coupled.
In this work, we introduce DA-RNNs, a deep network architecture that tightly connects the analysis of individual RGB-D frames and their integration over time. To do so, we take advantage of \emph{Recurrent} Neural Networks (RNNs), where recurrent units connect information over time. A naive approach for achieving a strong connection between the mapping and the labeling process would be to establish a fixed network structure in 3D and treat each surface element in a KinectFusion or ElasticFusion map as a recurrent unit in the RNN. Unfortunately, such an approach is not feasible since it would quickly exhaust the memory available even on large-scale GPUs. To overcome this problem, our approach performs recurrent reasoning only over those parts of the map that are currently observed by the RGB-D camera. Specifically, we introduce a new recurrent unit inside our RNN called Data Associated Recurrent Unit (DA-RU). Each DA-RU corresponds to a pixel in the input image. The hidden state of the DA-RU accumulates information about that pixel in time. Crucially, the temporal connectivity between the DA-RU states of consecutive frames is not fixed, but depends on the data association provided by the mapping process. As a result, each DA-RU incorporates the hidden state from the associated DA-RU in the previous frame, allowing information to flow in a spatially consistent way.
\begin{figure}
\centering
\includegraphics[height = 0.55\linewidth, width = \linewidth]{framework.pdf}
\caption{Overview of the DA-RNN framework. RGB-D frames are fed into a Recurrent Neural Network. KinectFusion provides the 3D reconstruction and the data associations necessary to connect recurrent units between RGB-D frames. The pixel labels provided by the RNN are integrated into the 3D semantic map. The overall labeling and reconstruction process runs at 5fps.}
\label{fig:framework}
\vspace{-4mm}
\end{figure}
In order to semantically reconstruct the 3D scene, we integrate the outputs of our DA-RNN into the 3D voxels of KinectFusion, which provides a consistent semantic labeling of the 3D scene (similar to~\cite{lai2014unsupervised,mccormac2016semanticfusion}). Fig.~\ref{fig:framework} illustrates an overview of our framework.
We have conducted extensive experiments to test our framework on the RGB-D Scene dataset \cite{lai2014unsupervised} and a synthetic dataset we generated with 3D shapes from the ShapeNet repository \cite{chang2015shapenet}. The experimental results demonstrate that DA-RNNs are able to provide superior semantically labeled 3D scenes from RGB-D videos. Our code and data are available at \url{https://rse-lab.cs.washington.edu/projects/darnn/}.
In summary, our work has the following key contributions:
\begin{itemize}
\item We propose a novel recurrent neural network for semantic labeling on RGB-D videos with a new data associated recurrent unit to capture dependencies across video frames.
\item We introduce a novel updating rule for DA-RU's to perform weighted moving averaging of the hidden state.
\item We integrate DA-RNN's with KinectFusion for semantic 3D scene reconstruction.
\item We contribute pixel-wise semantic labels on the RGB-D Scene dataset \cite{lai2014unsupervised} and a new synthetic dataset which can benefit future research on 3D semantic mapping.
\end{itemize}
This paper is organized as follows. After discussing related work, we introduce DA-RNNs, followed by experimental results and a conclusion.
\section{Related Work}
Our work is mostly related to 3D mapping and semantic labeling methods in the literature.
\subsection{Dense 3D Scene Reconstruction}
3D reconstruction techniques can be roughly classified into point-based methods, voxel-based methods and surfel-based methods. Point-based methods use 3D points to represent 3D scenes~\cite{snavely2008skeletal,crandall2011discrete,henry2012rgb}. Voxel-based methods such as KinectFusion, PatchVolumes, or Kintinuous \cite{newcombe2011kinectfusion,henry2013patch,Whe12Kin} employ a volumetric representation of the 3D space, which reconstruct dense 3D surfaces of the scene. Surfel-based methods \cite{keller2013real,henry2012rgb,whelan2015elasticfusion} make a trade-off between 3D points and voxels, where the 3D scene is represented compactly by 3D disks, i.e., surfels.
In principle, our DA-RNN framework only requires dense data associations between consecutive frames. It is thus independent of the underlying representation and could be combined with any of the reconstruction techniques described above. Here, we use KinectFusion~\cite{newcombe2011kinectfusion} to achieve a volumetric representation for geometry and semantics.
\subsection{Semantic Labeling}
Semantic labeling on images classifies each pixel of an input image into one of the predefined semantic classes. The semantic labeling problem has often been tackled with probabilistic graphical models such as Markov Random Fields (MRFs) or Conditional Random Fields (CRFs) \cite{shotton2006textonboost,krahenbuhl2011efficient}, which model the context around pixels. More recently, convolutional neural networks have been applied to semantic labeling \cite{long2015fully,zheng2015conditional,badrinarayanan2015segnet,chen2016deeplab}, which achieve significant improvement over previous methods. However, all these approaches mainly focus on semantic labeling of a single image. Recurrent neural networks \cite{pavel2015recurrent,shelhamer2016clockwork} have been applied to semantic video segmentation, which exploit the temporal relationship or information provided by multiple viewpoints of a scene. \cite{Lai13Obj,mccormac2016semanticfusion} show how the labels extracted from individual RGB-D frames can be incorporated into a voxel or surfel map, resulting in more stable labeling. Further improvements are achieved by performing MRF or CRF inference in the 3D map. Approaches such as \cite{salas2013slam++,lai2014unsupervised,song2016deep} perform labeling by conducting 3D object detection through the 3D reconstruction, thereby potentially incorporating information that is not available in any single view.
Different from these works, we propose a recurrent neural network architecture that tightly integrates the information contained in multiple viewpoints of an RGB-D video stream. Both individual frame and across frame parameters are learned in a single network structure. In contrast to existing RNNs, DA-RNNs do not assume a fixed relationship between input images and network structure, but rely on data association to generate the connections between recurrent units on the fly. The recurrent layer we introduce in this work could also be used as a standalone layer and plugged into existing CNN-based methods for semantic video labeling.
\section{Method}
\begin{figure*}
\centering
\includegraphics[height = 0.22\linewidth, width = \linewidth]{single_stream.pdf}
\caption{Architecture of our single stream network for semantic labeling.}
\label{fig:single_stream}
\vspace{-3mm}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[height = 0.28\linewidth, width = \linewidth]{double_stream.pdf}
\caption{Architecture of our double stream network for semantic labeling.}
\label{fig:double_stream}
\vspace{-4mm}
\end{figure*}
In this section, we present our framework for 3D semantic mapping using RGB-D videos. We first describe our design of the convolutional neural network for single frame semantic labeling. Then, we extend the single frame network to a recurrent neural network for semantic labeling on videos. Finally, we integrate the recurrent neural network with KinectFusion \cite{newcombe2011kinectfusion} in order to semantically reconstruct the 3D scene.
\subsection{Single Frame Labeling with Fully Convolutional Networks}
The basis of our semantic labeling framework is a Fully Convolutional Network (FCN) for single frame labeling. An influential network architecture for semantic labeling as been introduced by \cite{long2015fully}, which converts a network for image classification into fully convolutional by treating the fully connected layers in the network as $1 \times 1$ convolutional layers. In addition, \cite{long2015fully} uses deconvolutional layers to increase the resolution of the network output. Inspired by \cite{long2015fully}, we design our network architecture for single frame labeling as illustrated in Fig. \ref{fig:single_stream}.
\subsubsection{Single Stream Network}
Our single stream network in Fig. \ref{fig:single_stream} takes a single tensor as input, such as an RGB image or a depth image. It consists of 16 convolutional layers, 4 max pooling layers, 2 deconvolutional layers and 1 addition layer. All the convolutional filters in the network are of size $3 \times 3$ and stride $1$. The max pooling layers are of size $2 \times 2$ and stride $2$. Therefore, each max pooling layer reduces the resolution of its input by a factor of 2. The output of the 4th max pooling layer is 16 times smaller than the input image. The first deconvolutional layer doubles the resolution of its input, while the second deconvolutional layer increases the resolution by 8 times. As a result, the output of the network has the same resolution as the input image, i.e., dense pixel-wise labeling.
We design the network architecture with three phases as in Fig. \ref{fig:single_stream}. The first 13 convolutional layers and the 4 max pooling layers are considered to be the feature extraction phase, which extracts 512-dimensional feature vectors for the input image. The second phase is the embedding phase, which embeds the 512-dimensional features into a 64-dimensional space while increasing the resolution of the feature map using deconvolutional layers. A skip link is used in the embedding phase to combine features from an earlier convolutional layer motivated by \cite{long2015fully} (i.e., the one before the 4th max pooling layer). The last phase of the network classifies each pixel into a semantic class using a convolutional layer. The output of this convolutional layer is treated as the labeling scores for pixels, which has $n$ channels with $n$ the number of the semantic classes. By applying a softmax layer on the labeling scores, we can obtain the class probabilities of the pixels.
\subsubsection{Double Stream Network}
When the input data is multimodal such as color and depth, we have designed the double stream network to fuse RGB-D data (Fig. \ref{fig:double_stream}). In this network, the RGB image and the depth image are processed separately with different convolutional layers for feature extraction. These layers share the same structure as the feature extraction phase in the single stream network. To combine the two types of features, we introduce two concatenation layers, which stack the 512-dimensional features from the RGB image and the depth image and generate 1024-dimensional features. These features are embedded into a 64-dimensional space and classified as in the single stream network. Note that we utilize the ``late fusion'' strategy in this network, where features for color and depth are computed independently and then concatenated.
\subsection{Video Semantic Labeling with DA-RNNs}
In videos, due to the smooth change in camera motion or object motion, information flows across video frames. How to effectively utilize the temporal information for semantic labeling in videos is still an open question. In this work, we propose a Data Associated Recurrent Neural Network (DA-RNN) for video semantic labeling which stores and passes information across frames.
\subsubsection{DA-RNN Architecture}
\begin{figure*}
\centering
\includegraphics[height = 0.5\linewidth, width = \linewidth]{rnn.pdf}
\caption{Architecture of our data associated recurrent neural network for semantic labeling on videos.}
\label{fig:rnn}
\vspace{-2mm}
\end{figure*}
The architecture of our DA-RNN for semantic labeling is illustrated in Fig.~\ref{fig:rnn}. Based on our double stream network for single frame labeling, we introduce a recurrent layer which takes the embedded features of the current frame as input and generates new features for classification. The recurrent layer is designed to combine features from the previous frames and features in the current frame in order to utilize information across frames.
Specifically, the recurrent layer contains one recurrent unit for each pixel location. These recurrent units maintain and update their hidden states, storing information from previous frames. The outputs of the recurrent units depend on their inputs and hidden states. Two widely used recurrent units are the Long Short-Term Memory (LSTM) unit \cite{hochreiter1997long} and the Gated Recurrent
Unit (GRU) \cite{cho2014properties}. Both LSTM and GRU have been shown to perform well in tasks that require capturing long-term dependencies, such as natural language processing, speech recognition and machine translation \cite{sundermeyer2012lstm,graves2013hybrid,sutskever2014sequence}. However, both of them employ the hyperbolic tangent function in updating their hidden states, which makes the gradient back-propagation training inefficient. In DA-RNN, we introduce a new recurrent unit, which is explicitly designed to fuse features across video frames and can be trained more efficiently by using the Rectified Linear Unit (ReLU) as the activation function.
\subsubsection{Data Associated Recurrent Unit (DA-RU)}
\begin{figure}
\centering
\includegraphics[height = 0.55\linewidth, width = \linewidth]{DA-RU.pdf}
\caption{The block diagram of our Data Associated Recurrent Unit (DA-RU).}
\label{fig:daru}
\vspace{-4mm}
\end{figure}
The recurrent layer in our RNN contains $N$ recurrent units, where $N$ is the number of pixels in the input image. At time $t$, the $i$th recurrent unit stores a pair of vectors $\left< \mathbf{h}_t^i, \mathbf{w}_t^i \right>, i = 1,\ldots,N$, where $\mathbf{h}_t^i$ denotes the hidden state of the unit, and $\mathbf{w}_t^i$ indicates the weight vector for the hidden state. $\mathbf{h}_t^i$ and $\mathbf{w}_t^i$ have the same dimension (64-D in our RNN). We can interpret $\mathbf{w}_t^i$ as measuring the importance of the elements of the hidden state. At time $t + 1$, given input $\mathbf{x}_{t+1}^{i}$ from the previous layer (the second deconvolutional layer in our RNN), unit $i$ updates its hidden state and weight vector, and generates its output according to a set of rules described below. Fig. \ref{fig:daru} illustrates the block diagram of the DA-RU.
\textbf{Data association: }
\begin{equation} \label{eq:asso}
\left< \widetilde{\mathbf{h}}_{t+1}^{i}, \widetilde{\mathbf{w}}_{t+1}^{i} \right> =
\left \{
\begin{aligned}
\left< \mathbf{0}, \mathbf{0} \right>, &\text{ if no association}\\
\left< \mathbf{h}_t^{i'}, \mathbf{w}_t^{i'} \right>, &\text{ if $p_{t+1}^i$ associated to $p_{t}^{i'}$},
\end{aligned} \right.
\end{equation}
where $p_{t+1}^i$ and $p_{t}^{i'}$ denote the corresponding pixels of unit $i$ at time $t+1$ and unit $i'$ at time $t$ respectively. Eq. \eqref{eq:asso} indicates that a unit at time $t$ passes its hidden state and weight vector to a unit at time $t+1$ via data association between pixels. If a unit at time $t+1$ is not associated with any previous unit, its hidden state and weight vector are initialized as zeros. All the units in the first frame of a video are initialized with zeros.
\textbf{Computing weights for the input: }
\begin{equation}
\widehat{\mathbf{w}}_{t+1}^{i} = \sigma (\mathbf{W} [ \widetilde{\mathbf{h}}_{t+1}^{i}, \mathbf{x}_{t+1}^{i} ] + \mathbf{b} ), \label{eq:weight}
\end{equation}
where $\widehat{\mathbf{w}}_{t+1}^{i}$ is the weight vector for the input $\mathbf{x}_{t+1}^{i}$, which is a function of the hidden state from the previous frame and the input of the current frame. $\mathbf{W},\mathbf{b}$ are the parameters of the recurrent layer, which are shared by all the units in the layer, $\sigma(\cdot)$ indicates the logistic sigmoid function, $[\cdot, \cdot]$ denotes concatenation of two vectors. $\mathbf{W}$ is a $d \times 2d$ matrix and $\mathbf{b}$ is a $d$ dimensional bias vector, where $d$ is the dimension of the hidden state.
\textbf{Updating weight vector:}
\begin{equation}
\mathbf{w}_{t+1}^{i} = \widetilde{\mathbf{w}}_{t+1}^{i} + \widehat{\mathbf{w}}_{t+1}^{i}.
\end{equation}
The weight vector at time $t+1$ is the sum between the accumulated weight vector from the previous frame and the weight vector for the current input.
\textbf{Updating hidden state:}
\begin{equation} \label{eq:hidden}
\mathbf{h}_{t+1}^{i} = f( (\widetilde{\mathbf{w}}_{t+1}^{i} \oslash \mathbf{w}_{t+1}^{i}) \otimes \widetilde{\mathbf{h}}_{t+1}^{i} + (\widehat{\mathbf{w}}_{t+1}^{i} \oslash \mathbf{w}_{t+1}^{i}) \otimes \mathbf{x}_{t+1}^{i} ),
\end{equation}
where $f(x) = \max(0, x)$ is the Rectified Linear Unit (ReLU) activation function, and $\oslash, \otimes$ denotes element-wise division and element-wise multiplication between vectors respectively. As we can see from Eq. \eqref{eq:hidden}, the new hidden state is computed as a weighted sum between the hidden state from the previous frame and the input for the current frame, where the weights are accumulated in time.
\textbf{Computing Output:}
\begin{equation}
\mathbf{o}_{t+1}^{i} = g(\mathbf{h}_{t+1}^{i}).
\end{equation}
The output of the unit is defined as a function $g(\cdot)$ of the hidden state. In our RNN, we simply use the hidden state as the output of the unit, i.e., $g(\cdot)$ is the identity function.
We name the aforementioned recurrent unit as the Data Associated Recurrent Unit (DA-RU). DA-RU performs weighted moving average of its input in time via data association, where the weights are dynamically generated based on the parameters of the unit and the data it receives, while the parameters are learned during network training. In DA-RNNs, the DA-RUs are used to combine features from the previous frames and features in the current frame for semantic labeling (Fig. \ref{fig:rnn}).
\subsection{Joint 3D Mapping and Semantic Labeling}
In DA-RNNs, data association is needed in order to associate DA-RUs in the recurrent layer across video frames. In general, different data association algorithms can be applied, such as optical flow on RGB images or Iterative Closest Point (ICP) on depth images. In this work, we integrate DA-RNNs with KinectFusion \cite{newcombe2011kinectfusion}, a dense 3D mapping technique using depth camera. KinectFusion estimates the camera poses of the video frames, from which we can compute the data association for the recurrent layer in the RNN. In addition, we fuse the semantic labels of pixels into the volumetric space in KinectFusion. Consequently, our system is able to reconstruct and semantically label the 3D scene.
\subsubsection{Data Association with KinectFusion}
KinectFusion represents the 3D scene with a 3D voxel space which stores the values of the Truncated Signed Distance Function (TSDF). The TSDF value of a voxel indicates the signed distance from the voxel to the closest zero crossing, i.e., surface. Given a stream of depth images, these TSDF values are updated per-frame. In order to fuse the depth images into the voxel space, KinectFusion performs camera tracking by estimating the 6DOF camera pose for each frame. The camera pose estimation is achieved by performing ICP between the 3D points from the current depth image and the 3D points extracted from the surface of the KinectFusion map. Given the camera poses of two consecutive frames from KinectFusion, we compute the data association between the two frames by back-projecting one frame into 3D points in the KinectFusion map and then projecting these 3D points onto the other frame using the estimated camera poses.
\subsubsection{Semantic Fusion}
By combining DA-RNNs with 3D mapping techniques such as KinectFusion, we are able to propagate semantic information into the 3D space. In addition, the semantic labels from different views of the same 3D location are fused in order to obtain a consistent understanding of the 3D space. For each voxel in KinectFusion, we store a probability vector of the semantic label space in addition to the TSDF value. Given a new depth image, the TSDF values of the voxels are first updated as in the traditional KinectFusion. Then, for voxels whose signed distances are smaller than the truncated distance threshold, i.e., voxels around the surface, their probability vectors are updated using the probability map of the semantic labels predicted by the RNN. A running averaging is used for both the TSDF and the probability vector to reduce noise in the fusing process. At any time step, the label of a voxel is predicted as the semantic class with the maximum probability according to the stored probability vector in the voxel. Fig. \ref{fig:framework} illustrates the 3D mapping and semantic labeling pipeline of our framework.
\section{Experiments}
In this section, we conduct experiments to evaluate our proposed system for 3D scene mapping and semantic labeling.
\subsection{Datasets}
\subsubsection{RGB-D Scene Dataset}
Two RGB-D video datasets are used to test our method. The first one is the RGB-D Scene dataset introduced by \cite{lai2014unsupervised}, which consists of 14 RGB-D videos captured by Kinect in indoor scenes. Each scene is reconstructed as a 3D point cloud aligned via the Patch Volumes Mapping method \cite{henry2013patch}. Then these 3D point clouds are labeled by 9 object class labels plus background. However, the RGB-D Scene dataset does not provide pixel-wise labeling for every video frame, which is needed in order to train and test our RNN. We could project the labeled 3D points onto video frames, but the projection only provides sparse labeling of the frames, i.e., not every pixel is labeled. Instead, we use the following procedure to obtain dense labeling on the RGB-D Scene dataset.
Each scene is first reconstructed with KinectFusion. After the 3D reconstruction, we manually annotate the 3D bounding box of the object in the 3D map. For each depth image, we convert it into a 3D point cloud and transform the 3D point cloud into the reconstructed 3D space according to its camera pose estimated from KinectFusion. Finally, 3D points inside a 3D bounding box are labeled with the semantic class of the object inside the box. In this way, we obtain dense labels for all the depth images. Note that pixels with missing depth values are not labeled according to our labeling procedure. We use 7 videos for training (5,808 frames) and the other 7 videos for testing (5,619 frames).
\subsubsection{ShapeNet Scene Dataset}
The second dataset is a synthetic dataset we generated using 3D shapes from the ShapeNet repository \cite{chang2015shapenet}. We selected 3D shapes with high quality and texture in 7 object categories: bottle (110 objects), can (25 objects), cap (23 objects), keyboard (36 objects), monitor (95 objects), mug (65 objects) and table (508 objects). We first compose virtual scenes, each with a table on the ground and five table-top objects among bottle, can, cap, keyboard, monitor and mug. Then, we render each virtual scene into RGB images and depth images from a simulated camera trajectory around the table. To obtain the semantic labels of the rendered images, we color the 3D shapes with distinct colors for each class and render the colored scene again using the same camera trajectory. By checking the color of the pixels in this second-pass rendering, we obtain the class labels of the images.
In total, we generated 100 virtual scenes, i.e., 100 RGB-D videos, by randomly sampling 3D shapes from the 7 object categories. 100 frames are rendered for each scene from a sampled camera trajectory. We use 50 videos for training (5,000 frames) and test on the other 50 videos (5,000 frames). Different from the RGB-D Scene dataset, we make sure that there is no overlapping object instances appearing in both the training set and the test set.
\subsection{Evaluation Metrics}
We evaluate our method on semantic labeling of pixels and 3D points. For pixel labeling, we adopt the pixel Intersection over Union (IoU) as the evaluation metric, which is the standard metric used for image semantic labeling. Pixel IoU computes the intersection over union of the predicated pixel labels and the ground truth pixel labels on the entire dataset for every class. For 3D point labeling, we follow \cite{lai2014unsupervised} and use the precision and recall on 3D points as the evaluation metric in order to achieve a fair comparison.
\subsection{Implementation Details}
DA-RNN is implemented using the TensorFlow library \cite{abadi2016tensorflow} with Python interface for communication with the KinectFusion module. In training, the parameters of the first 13 convolutional layers in the feature extraction phase are initialized with the VGG16 network \cite{simonyan2014very} trained on ImageNet \cite{deng2009imagenet}. Learning is conducted by Stochastic Gradient Descent (SGD) with momentum, where the loss function is the softmax cross entropy loss for pixels. For our single stream FCN and double stream FCN, each SGD mini-batch is a single image, chosen uniformly at random. For DA-RNN, each SGD mini-batch is a video sequence of 3 consecutive frames. In testing, video frames are processed sequentially, and the hidden states of the DA-RNN are passed to the next frame via data association for the entire video sequence. In this way, the DA-RNN captures long term dependencies between pixels.
\subsection{Comparison on Network Architectures}
\begin{table} \setlength{\tabcolsep}{2pt}
{
\centering{
\begin{tabular}{|l||c|c|c|c|}
\hline Methods & FCN \cite{long2015fully} & Our FCN & Our GRU-RNN & Our DA-RNN \\
\hline
\hline Background & 94.3 & 96.1 & 96.8 & \textbf{97.6} \\
\hline Bowl & 78.6 & 87.0 & 86.4 & \textbf{92.7} \\
\hline Cap & 61.2 & 79.0 & 82.0 & \textbf{84.4} \\
\hline Cereal Box & 80.4 & 87.5 & 87.5 & \textbf{88.3} \\
\hline Coffee Mug & 62.7 & 75.7 & 76.1 & \textbf{86.3} \\
\hline Coffee Table & 93.6 & 95.2 & 96.0 & \textbf{97.3} \\
\hline Office Chair & 67.3 & 71.6 & 72.7 & \textbf{77.0} \\
\hline Sofa Can & 73.5 & 82.9 & 81.9 & \textbf{88.7} \\
\hline Sofa & 90.8 & 92.9 & 93.5 & \textbf{95.6} \\
\hline Table & 84.2 & 89.8 & 90.8 & \textbf{92.8} \\
\hline
\hline MEAN & 78.7 & 85.8 & 86.4 & \textbf{90.1} \\
\hline
\end{tabular}
\caption{Comparison in network architectures for image pixel labeling on the RGB-D Scene dataset. The network input is RGB image.}
\label{table:arch}
}
\vspace{-8mm}
}
\end{table}
In this experiment, we fix the network input to RGB images and compare different network architectures for pixel-wise semantic labeling. Table \ref{table:arch} presents the pixel IoU on the RGB-D Scene dataset for four different networks.
i) We compare our single stream FCN (Fig. \ref{fig:single_stream}) with the FCN in \cite{long2015fully} which is fine-tuned on the RGB-D Scene dataset using the same experimental setup as ours. As we can see from the table, our single stream FCN significantly outperforms the FCN in \cite{long2015fully}. \cite{long2015fully} converts the VGG16 network into a fully convolutional network for semantic labeling. There are five max pooling layers and two 4096-dimensional fully connected layers (eventually converted to $1 \times 1$ convolutional layers) in the network, which outputs blob-like segmentations and cannot capture fine-grained details of the objects. In contrast, our FCN uses fewer max pooling layers and embeds the convolutional features into a low dimensional space (64-D) before classification, which is able to generate shaper segmentations of the objects.
ii) We compare two types of recurrent unit in our RNN architecture: Gated Recurrent Unit (GRU) \cite{cho2014properties} and the DA-RU we introduce in this work. From Table \ref{table:arch}, we can see that our DA-RU achieves better labeling performance than GRU. First, our DA-RU can be trained more efficiently since it uses the ReLU function instead of the hyperbolic tangent function in updating its hidden state. Second, the DA-RU is explicitly designed as a weighted moving average unit, which is more effective for video-based applications.
iii) By comparing our DA-RNN with the FCN, DA-RNN achieves better labeling accuracy, thanks to its ability in capturing the temporal information across video frames.
\subsection{Analysis on Network Inputs}
\begin{table*} \setlength{\tabcolsep}{4pt}
{
\centering{
\begin{tabular}{|l||c|c|c|c||c|c|c|c|}
\hline Methods & FCN RGB & FCN Depth & FCN Normal & FCN RGB-D & DA-RNN RGB & DA-RNN Depth & DA-RNN Normal & DA-RNN RGB-D \\
\hline
\hline Background & 96.1 & 97.0 & 95.4 & 97.8 & 97.6 & 98.4 & 98.4 & \textbf{98.7} \\
\hline Bowl & \textbf{97.0} & 94.8 & 86.5 & 89.8 & 92.7 & 89.8 & 91.8 & 93.1 \\
\hline Cap & 79.0 & 86.7 & 86.7 & 82.7 & 84.4 & 88.9 & \textbf{90.5} & 87.0 \\
\hline Cereal Box & 87.5 & 88.1 & 58.3 & 88.5 & 88.3 & 90.6 & 90.3 & \textbf{94.2} \\
\hline Coffee Mug & 75.7 & 81.9 & 83.1 & 82.2 & 86.3 & 83.1 & 86.3 & \textbf{89.4} \\
\hline Coffee Table & 95.2 & 87.2 & 83.7 & 96.3 & 97.3 & 91.8 & 91.7 & \textbf{97.8} \\
\hline Office Chair & 71.6 & 79.0 & 74.8 & 82.4 & 77.0 & 84.2 & 84.5 & \textbf{87.5} \\
\hline Soda Can & 82.9 & 84.4 & 85.7 & 86.1 & 88.7 & 89.9 & 88.1 & \textbf{90.7} \\
\hline Sofa & 92.9 & 94.2 & 92.6 & 96.1 & 95.6 & 95.6 & 96.1 & \textbf{97.9} \\
\hline Table & 89.8 & 69.6 & 68.8 & 92.7 & 92.8 & 81.0 & 81.1 & \textbf{94.5} \\
\hline
\hline MEAN & 85.8 & 85.3 & 81.5 & 89.4 & 90.1 & 89.3 & 89.9 & \textbf{93.1} \\
\hline \multicolumn{5}{|c||}{Improvement of DA-RNN over FCN} & +4.3 & +4.0 & +8.4 & +3.7 \\
\hline
\end{tabular}
\caption{Comparison in network inputs for image pixel labeling on the RGB-D Scene dataset.}
\label{table:rgbd}
}
\vspace{-4mm}
}
\end{table*}
\begin{table*} \setlength{\tabcolsep}{4pt}
{
\centering{
\begin{tabular}{|l||c|c|c|c||c|c|c|c|}
\hline Methods & FCN RGB & FCN Depth & FCN Normal & FCN RGB-D & DA-RNN RGB & DA-RNN Depth & DA-RNN Normal & DA-RNN RGB-D \\
\hline
\hline Background & 99.1 & 99.0 & 98.1 & 99.4 & \textbf{99.5} & 99.3 & 98.8 & \textbf{99.5} \\
\hline Bottle & 79.8 & 80.8 & 76.8 & 81.3 & 84.8 & \textbf{86.1} & 83.1 & 84.5 \\
\hline Can & 64.5 & 83.7 & 53.2 & 67.1 & 65.2 & \textbf{84.6} & 81.1 & 66.9 \\
\hline Cap & 81.3 & 85.3 & 87.4 & 83.1 & 84.6 & 87.9 & \textbf{91.1} & 83.6 \\
\hline Keyboard & 90.2 & 88.9 & 91.3 & 91.3 & 91.2 & 90.6 & 91.6 & \textbf{92.4} \\
\hline Monitor & 87.7 & 90.7 & 90.8 & 92.2 & 91.2 & 92.8 & \textbf{93.9} & 93.2 \\
\hline Mug & 68.9 & 84.9 & 66.4 & 70.7 & 70.5 & \textbf{85.0} & 81.2 & 70.6 \\
\hline Table & 94.9 & 93.7 & 91.5 & 96.0 & 95.8 & 95.1 & 94.2 & \textbf{96.3} \\
\hline
\hline MEAN & 83.3 & 88.4 & 82.0 & 85.1 & 85.3 & \textbf{90.2} & 89.4 & 85.9 \\
\hline \multicolumn{5}{|c||}{Improvement of DA-RNN over FCN} & +2.0 & +1.8 & +7.4 & +0.8 \\
\hline
\end{tabular}
\caption{Comparison in network inputs for image pixel labeling on the ShapeNet Scene dataset.}
\label{table:shapenet}
}
\vspace{-6mm}
}
\end{table*}
We conduct experiments to analyze the effect of different types of network inputs on semantic labeling. These inputs are RGB image, depth image, normal image and RGB-D image. For the depth image, we normalize the depth values between 0 and 255 and copy it three times to feed it into the network whose input has 3 channels. For the normal image, we compute the surface normals from the depth image, and then convert the surface normal coordinates into a 3-channel image. A pair of the RGB image and the depth image is referred as a RGB-D image. Our single stream network (Fig. \ref{fig:single_stream}) is used to process the RGB image, the depth image or the normal image, while the double stream network (Fig. \ref{fig:double_stream}) is used to process the RGB-D image. Both networks can be turned into a DA-RNN by adding a recurrent layer as in Fig. \ref{fig:rnn}.
Table \ref{table:rgbd} presents the results of our FCN and DA-RNN with different inputs on the RGB-D Scene dataset. i) Using RGB image achieves better performance than using depth image or using normal image. Since the RGB-D Scene dataset consists of a few number of object instances such as two specific bowls or cereal boxes, color is more discriminative than depth and normal. ii) By using RGB-D images with our double stream FCN, the labeling accuracy is improved over using RGB image, depth image or normal image only. iii) Our DA-RNN consistently improves over its FCN counterpart, which demonstrates the advantages of DA-RNNs on the semantic video labeling task.
Table \ref{table:shapenet} presents the labeling results on the ShapeNet Scene dataset. i) Depth is more discriminative than color. This is because objects in the ShapeNet Scene dataset are sampled from hundreds of 3D shapes with different colors. The objects in the test set are unseen in the training set, so their general shape from the depth images are more discriminative than their color. ii) Combing RGB image and depth image does not improve over using depth image only. It seems that adding color information confuses the network from differentiating objects in different categories but with similar color. iii) Our DA-RNNs achieve better performance than the FCNs consistently across different input types. Fig. \ref{fig:results} shows some labeling examples from our FCN and DA-RNN on the RGB-D Scene dataset and the ShapeNet Scene dataset.
\begin{figure*}
\centering
\includegraphics[height = 0.62\linewidth, width = 0.9\linewidth]{results.pdf}
\caption{Semantic labeling examples on the RGBS Scene dataset and the ShapeNet Scene dataset.}
\label{fig:results}
\vspace{-2mm}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[height = 0.16\linewidth, width = 0.88\linewidth]{scenes.pdf}
\caption{The semantic 3D mapping built by our method using the RGB-D Scene dataset.}
\label{fig:scenes}
\vspace{-6mm}
\end{figure*}
\begin{table} \setlength{\tabcolsep}{0.5pt}
{
\centering{
\begin{tabular}{|l||c|c|c|c|}
\hline Methods & HMP2D \cite{lai2014unsupervised} & HMP3D \cite{lai2014unsupervised} & HMP2D+3D \cite{lai2014unsupervised} & DA-RNN RGB-D \\
\hline
\hline Background & 42.9 / \textbf{99.6} & \textbf{99.9} / 80.0 & 95.8 / 95.0 & 94.7 / 96.4 \\
\hline Bowl & 74.4 / 85.0 & \textbf{100.0} / \textbf{96.2} & 97.0 / 89.1 & 95.3 / 91.0 \\
\hline Cap & 74.9 / 98.6 & 91.3 / 98.2 & 82.7 / \textbf{99.0} & \textbf{93.5} / 91.1 \\
\hline Cereal Box & 79.9 / 98.6 & 85.1 / \textbf{100.0} & 96.2 / 99.3 & \textbf{98.0} / 93.5 \\
\hline Coffee Mug & 64.4 / 87.8 & 90.0 / \textbf{93.9} & 81.0 / 92.6 & \textbf{90.5} / 86.3 \\
\hline Coffee Table & 11.9 / 17.9 & 96.1 / \textbf{100.0} & \textbf{98.7} / 98.0 & 93.8 / 97.3 \\
\hline Office Chair & 17.7 / 17.2 & 57.6 / \textbf{100.0} & 89.7 / 94.5 & \textbf{96.0} / 96.2 \\
\hline Soda Can & 78.2 / \textbf{98.1} & \textbf{100.0} / 81.9 & 97.7 / 98.0 & 92.0 / 83.4 \\
\hline Sofa & 29.3 / 39.8 & 82.7 / \textbf{100.0} & 92.5 / 92.0 & \textbf{99.6} / 91.3 \\
\hline Table & 16.4 / 23.3 & 95.3 / \textbf{98.7} & 97.6 / 96.0 & \textbf{98.2} / 97.2 \\
\hline
\hline MEAN & 49.0 / 66.6 & 89.8 / 94.9 & 92.8 / \textbf{95.3} & \textbf{95.2} / 92.4 \\
\hline
\end{tabular}
\caption{Precision and recall of 3D point labeling on the RGB-D Scene dataset.}
\label{table:points}
}
}
\vspace{-6mm}
\end{table}
\subsection{3D Scene Labeling Results}
In this experiment, we evaluate our framework on the 3D point labeling task. Our DA-RNN generates pixel-wise labeling of each video frame that is integrated into the KinectFusion map to label the 3D scene. Since the 3D points provided in the RGB-D Scene dataset are not in the same 3D space of our KinectFusion map, we use the following procedure to obtain the labels of these 3D points. For each 3D point, we project it to all the video frames with camera poses provided by the dataset and then check its visibility in the frames. A 3D point is visible in a frame if its projection is inside the frame and the value of the projection in the depth image is within a range of the depth of the 3D point. Then we accumulate all the visible labels of the 3D point, and use the one with the maximum frequency as the final label for it.
Table \ref{table:points} presents the 3D point labeling precision and recall on the RGB-D Scene dataset, where we compare our method with three variations of the method proposed in \cite{lai2014unsupervised}. It is worth to mention that the models in \cite{lai2014unsupervised} are trained with synthetic data generated from rendering 3D shapes only and tested on all the 14 videos in the RGB-D Scene dataset. To test our method on the same videos, we conduct a two-fold cross validation and obtain the results on all the 14 videos. From Table \ref{table:points}, we can see that our method achieves comparable 3D point labeling precision and recall with the HMP2D+3D model. While \cite{lai2014unsupervised} employs several heuristics to remove the ground plane and the table-top, our system processes a RGB-D video automatically without such heuristics. Fig. \ref{fig:scenes} shows some semantic mapping results on the RGB-D Scene dataset. Please see our project website for the result video on the two datasets. The labeling errors are more often caused by confusion between classes with similar 3D shape such as mug and can. Data association accuracy also affects the performance. For example, we sometimes see that the bottom of an object is labeled as table due to wrong assocation to table pixels in the previous frame.
\section{Conclusion}
In this work, we introduce DA-RNNs, a novel framework for joint 3D mapping and semantic labeling on RGB-D videos. DA-RNNs integrate a recurrent neural network for video semantic labeling with KinectFusion. To achieve a compact network representation, recurrent reasoning is only performed over the currently visible part of the environment, using data association to define the connectivity between recurrent units. The labels predicted by the RNN are fused into the KinectFusion map for dense semantic mapping. Experiments are conducted on a real world dataset and a synthetic dataset of RGB-D videos. The experimental results and analyses demonstrate the advantages of our method on video semantic labeling and 3D scene mapping.
A key advantage of DA-RNNs is their flexibility. While this paper focuses on object class labeling, we believe that the same architecture could be applied to train networks for a wide range of semantic labeling problems, including object instance and pose detection, material recognition, and physical support estimation. Data association between frames can also be obtained in different ways such as using optical flow methods. Another promising avenue for improvement is the incorporation of shape information provided by the 3D map.
\vspace{-1mm}
\section*{Acknowledgments}
This work was funded in part by ONR grant N00014-13-1-0720 and by Northrop Grumman. We thank Tanner Schmidt for fruitful discussions and for providing his implementation of KinectFusion.
\bibliographystyle{plainnat}
|
\section{Additional material for the adversarial setting}
\label{app:adversarial}
We first present the pseudocode and proofs for the finite arm setting
and then for the infinite arm setting.
\subsection{Finite arm setting}
Algorithm~\ref{alg:exp3rej} contains the pseudocode for
\textsc{EXP3-ABS}, an algorithm for online learning with abstention
under an adversarial data model that guarantees small
regret. The algorithm itself is a simple adaptation of the ideas in \cite{AlonCesaGentileMannorMansourShamir2014,AlonCesaBianchiDekelKoren2015}, where we incorporate the side information that the loss of an abstaining arm is always observed, while the loss of a predicting arm is observed only if the algorithm actually plays a predicting arm. In the pseudocode and in the proof that follows, $L_t(\xi_j)$ is a shorthand for $L(\xi_j,(x_t,y_t))$.
\begin{algorithm}[t]
\caption{\textsc{EXP3-ABS}}
\label{alg:exp3rej}
\begin{algorithmic}
\INPUT Set of experts $\mathscr E = \{\xi_1,\ldots, \xi_K\}$; learning rate $\eta > 0$~; \\
{\bf Init:} $q_1$ is the uniform distribution over $\mathscr E$~;\\
\FOR{$t\leftarrow 1, 2, \ldots$}
\STATE \textsc{Receive}($x_t$);
\STATE $\xi_{I_t} \gets $ \textsc{Sample}($q_t$);
\IF {$r_{I_t}(x_t) > 0$}
\STATE \textsc{Receive}($y_t$);
\ENDIF
\STATE For all $\xi_j= (h_j,r_j)$, set :
\begin{align*}
&P_t(\xi_j) \leftarrow
\begin{cases}
1 &{\mbox{if $r_j(x_t) \leq 0$}}\\
\sum_{\xi_i\in \mathcal{E}\,:\, r_i(x_t) > 0} q_t(\xi_i) & {\mbox{if $r_j(x_t) > 0$}}~,
\end{cases}\hspace{3.0in}\\
&\widehat{L}_t(\xi_j) \leftarrow \frac{L_t(\xi_j)}{P_t(\xi_j)} \left(1_{r_{I_t}(x_t) \leq 0}1_{r_j(x_t) \leq 0} + 1_{r_{I_t}(x_t) > 0}\right)~,\\
&q_{t+1}(\xi_j) \leftarrow \frac{q_t(\xi_j) \exp(-\eta \widehat{L}_t(\xi_j)) }{\sum_{\xi_i\in \mathcal{E}} q_t(\xi_i ) \exp (-\eta \widehat{L}_t(\xi_i))}~.
\end{align*}
\ENDFOR
\end{algorithmic}
\end{algorithm}
\noindent{\bf Proof of Theorem \ref{th:exp3rej}.}\\
\begin{proof}
By applying the standard regret bound of Hedge (e.g., \cite{Bubeck2012}) to distributions
$q_1,\ldots, q_T$ generated by \textsc{EXP3-ABS}\ and to the
non-negative loss estimates $\widehat{L}_t(\xi_j)$, the following holds:
\begin{align}
&\E \Bigg [ \sum_{t=1}^T \sum_{\xi_j\in \mathcal{E}} q_t(\xi_j) \E\Big[\widehat{L}_t(\xi_j)\Big]- \sum_{t=1}^T \E\Big[\widehat{L}_t(\xi^\star)\Big] \Bigg ] \nonumber \leq \frac{\log K}{\eta} + \frac{\eta}{2}\sum_{t=1}^T \E \left [ \sum_{\xi_j\in \mathcal{E}} q_t(\xi_j) \E\left[\widehat{L}_t(\xi_j)^2\right] \right ] \nonumber
\end{align}
for any fixed $\xi^\star \in \mathcal{E}$. Using the fact that
$\E\Big[\widehat{L}_t(\xi_j)\Big]=L_t( \xi_j)$ and
$ \E\Big[\widehat{L}_t(\xi_j)^2\Big]=\frac{L_t( \xi_j)^2}{P_t(\xi_j)}$, we can write
\begin{align*}
\E \Bigg [ \sum_{t=1}^T \sum_{\xi_j\in \mathscr E} q_t(\xi_j) L_t( \xi_j) - \sum_{t=1}^T L_t( \xi^\star)\Bigg ]
\leq
\frac{\log K}{\eta} + \frac{\eta}{2}\,\sum_{t=1}^T \E \Bigg [ \sum_{\xi_j\in\mathscr E} \frac{q_t(\xi_j)}{P_t(\xi_j)} L_t(\xi_j)^2 \Bigg ].
\end{align*}
For each $t$, we can split the nodes $V$ of $G^{\text{\sc abs}}_t$ into the two subsets $V_{abs,t}$ and $V_{acc,t}$ where if a node $\xi_j$
is abstaining at time $t$ then $\xi_j \in V_{abs,t}$, and otherwise $\xi_j \in V_{acc,t}$. Thus, for any round $t$, we can write
\begin{align*}
\sum_{\xi_j\in \mathscr E} \frac{q_t(\xi_j)}{P_t(\xi_j)} L_t( \xi_j)^2
& = \sum_{\xi_j\in V_{abs,t}} \frac{q_t(\xi_j)}{P_t(\xi_j)} L_t( \xi_j)^2 + \sum_{\xi_j\in V_{acc,t}} \frac{q_t(\xi_j)}{P_t(\xi_j)} L_t(\xi_j)^2 \\
& \leq \sum_{\xi_j\in V_{abs,t}} q_t(\xi_j)\,c^2+ \sum_{\xi_j\in V_{acc,t}} \frac{q_t(\xi_j)}{P_t(\xi_j)} \\
& \leq c^2 +1~.
\end{align*}
The first inequality holds since if $\xi_j$ is an abstaining expert at time $t$,
we know that $L_t(\xi_j)=c$ and $P_t(\xi_j) =1$, while for the accepting experts we know
that $L_t(\xi_j)\leq 1$ anyway. The second inequality holds because if $\xi_j$ is an
accepting expert, we have $P_t(\xi_j) = \sum_{\xi_j\in V_{acc,t}} q_t(\xi_j)$.
\iffalse
************************************************
the abstaining experts form a clique, which implies that
$\sum_{j \in N_t(i)} q_t(\xi_j) =\sum_{\xi_i\in V_{abs,t}} q_t(\xi_i)
$ and so
$\sum_{\xi_i\in V_{abs,t}} \frac{q_t(\xi_i)}{P_t(\xi_i)}
=\sum_{\xi_i\in V_{abs,t}} \frac{q_t(\xi_i)}{ \sum_{j \in N_t(i)}
q_t(\xi_j) } \leq 1$. By a similar reasoning,
$\sum_{\xi_i\in V_{acc,t}} \frac{q_t(\xi_i)}{P_t(\xi_i)} \leq 1$ for
the accepted points.
************************************************
\fi
\ignore{
******************************************************************************
Since all nodes $\xi\in V_{abs,t}$ have self-loops and
$p_t(\xi)\geq \frac{\gamma}{K}$ because we mixed with the uniform
distribution, we can apply Lemma 5 in
\cite{AlonCesaBianchiDekelKoren2015} with
$\epsilon=\frac{\gamma}{K}$. Thus, we have the following
inequalities
\begin{align*}
\sum_{\xi\in V_{abs,t}} \frac{q_t(\xi)}{P_t(\xi)}c^2 &\leq 2 \sum_{\xi\in V_{abs,t}} \frac{p_t(\xi)}{P_t(\xi)}c^2 \leq 8c^2 \alpha_{abs,t} \log \frac{4K K_{abs,t}}{\gamma \alpha_{abs,t}} ,
\end{align*}
where the first inequality is due to the fact that
$p_t(\xi)\geq (1-\gamma)q_t(\xi)\geq \frac{1}{2}q_t(\xi)$ and where
$\alpha_{abs,t}=\alpha(V_{abs,t})$ is the independence number of
$V_{abs,t}$. Similarly,
\begin{align*}
\sum_{\xi\in V_{acc,t}} \frac{q_t(\xi)}{P_t(\xi)} &\leq 2 \sum_{\xi\in V_{acc,t}} \frac{p_t(\xi)}{P_t(\xi)}\leq 8 \alpha_{acc,t} \log \frac{4K K_{acc,t}}{\gamma \alpha_{acc,t}},
\end{align*}
where $\alpha_{acc,t}=\alpha(V_{acc,t})$ is the independence number of $V_{acc,t}$.
For the feedback graphs in this setting, the $\alpha_{abs,t}=\alpha_{acc,t}=1$ for all $t$, which is the best we can hope for. Thus,
\begin{align*}
&\sum_{\xi\in V_{abs,t}} \frac{q_t(\xi)}{P_t(\xi)}c^2+ \sum_{\xi\in V_{acc,t}} \frac{q_t(\xi)}{P_t(\xi)} \leq 8c^2 \log \frac{4K K_{abs,t}}{\gamma } + 8 \log \frac{4K K_{acc,t}}{\gamma}.
\end{align*}
Using then the fact that
\begin{align*}
&\sum_{\xi\in V} p_t(\xi)L_t(\xi) \leq \sum_{\xi\in V} q_t(\xi) L_t(\xi) + \gamma,
\end{align*}
the regret bound in this case can be written as
\begin{align*}
\E \Bigg [ \sum_{t=1}^T \sum_{\xi\in V} p_t(\xi) L_t(\xi) \Bigg ] - \sum_{t=1}^T L_t(\xi^\star) \leq & \gamma T + \frac{\log K}{\eta} + \eta \sum_{t=1}^T \Bigg [ 8c^2 \log \frac{4K K_{abs,t}}{\gamma } + 8 \log \frac{4K K_{acc,t}}{\gamma} \Bigg ],
\end{align*}
which is the bound of the theorem.
***************************************************************************
}
Putting together concludes the proof.
\end{proof}
\subsection{Infinite arm setting}\label{app:adversarialinfinite}
The input space $\mathscr X$ is assumed to be totally bounded, so that
there exists a constant $C_{\mathscr X} > 0$ such that, for all $0 < \varepsilon \leq 1$, $\mathscr X$ can
be covered with at most $C_{\mathscr X}\varepsilon^{-d}$ balls of radius
$\varepsilon$. Moreover, let $\mathscr Y$ be a shorthand for $[-1, 1]^2$, the range
space of the pairs $(h, r)$. An $\varepsilon$-covering $\mathscr Y_{\epsilon}$ of
$\mathscr Y$ with respect to the Euclidean distance on $\mathscr Y$ has size
$K_{\epsilon} \leq C_{\mathscr Y}\varepsilon^{-2}$ for some constant $C_{\mathscr Y}$.
\begin{algorithm}[t]
\caption{\textsc{Cont\EXPABS}.}
\label{ContextualExp3-W}
\begin{algorithmic}
\INPUT Ball radius $\varepsilon > 0$, $\varepsilon$-covering $\mathscr Y_{\varepsilon}$ of $\mathscr Y$ such that $|\mathscr Y_{\varepsilon}| \le C_{\mathscr Y}\,\varepsilon^{-2}$;
\FOR{$t=1,2,\dots$}
\STATE \textsc{Receive}($x_t$);
\STATE If $x_t$ does not belong to any existing ball, create new ball of radius $\varepsilon$ centered on $x_t$, and allocate fresh instance of \textsc{EXP3-ABS};
\STATE Let ``Active \textsc{EXP3-ABS}" be the instance allocated to the existing ball whose center $x_s$ is closest to $x_t$;
\STATE Draw action $\xi_{I_t} \in \mathscr Y_{\varepsilon}$ using Active \textsc{EXP3-ABS};
\STATE Get loss feedback associated with $\xi_{I_t}$ and use it to update state of ``Active \textsc{EXP3-ABS}".
\ENDFOR
\end{algorithmic}
\end{algorithm}
\iffalse
**************************************************
\begin{algorithm}[t]
\caption{\textsc{Exp3} adapted from \citet{CesaBianchiGaillardGentileCerchinovitz2017}}
\label{alg:exp3floor}
\begin{algorithmic}
\INPUT Learning rate $\eta > 0$, uniform distribution $p_1$ over
$\mathscr Y_\varepsilon$ with $K_\varepsilon = |\mathscr Y_\varepsilon| \leq C_{\mathscr Y}\,\varepsilon^{-2}$
\FOR{$t=1,2,\dots$}
\STATE Draw $\hat{\xi}_t \sim p_t$;
\FOR { $\xi \in \mathscr Y_\varepsilon$}
\STATE $\hat {\mathcal{L}}_t(\xi) = \frac{{\mathcal{L}}_t(\xi)}{\Pr({\mathcal{L}}_t(\xi)\,{\mbox{is observed}} )} 1_{{\mathcal{L}}_t(\xi)\,{\mbox{is observed}}}$
\STATE $p_{t+1}(\xi) = \frac{\exp\big(-\eta \sum_{s=1}^t \hat {\mathcal{L}}_s(\xi)\big)}{\sum_{\xi'} \exp\left(-\eta\sum_{s=1}^t \hat {\mathcal{L}}_s(\xi')\right)}~.$
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
*********************************************
\fi
The online learning scenario for the loss ${\tilde{L}}$ under the abstention setting's feedback graphs is as follows. Given an unknown
sequence $z_1, z_2,\dots$ of pairs $z_t = (x_t,y_t) \in \mathscr X\times\{\pm 1\}$, for every round $t = 1,2,\dots$:
\begin{enumerate}
\item The environment reveals input $x_t \in \mathscr X$;
\item The learner selects an action $\xi_{I_t} \in \mathscr Y$ and incurs loss ${\tilde{L}}(\xi_{I_t},z_t)$;
\item The learner obtains feedback from the environment.
\end{enumerate}
Our algorithm is described as Algorithm \ref{ContextualExp3-W}.
Consider the function
\[
{\tilde{L}}(a,r) =
\begin{cases}
c &{\mbox{if $r \leq -\gamma$}}\\
1+\left(\frac{1-c}{\gamma}\right)r &{\mbox{if $r \in (-\gamma,0)$}}\\
1-\left(\frac{1-f_{\gamma}(a)}{\gamma}\right)\,r &{\mbox{if $r \in [0,\gamma)$}}\\
f_{\gamma}(a) &{\mbox{if $r \geq \gamma$}}~,
\end{cases}
\]
where $f_{\gamma}$ is the Lipschitz variant of the 0/1-loss mentioned in Section \ref{sec:adversarial} of the main text (Figure \ref{f:1} (a)).
For any fixed $a$, the function ${\tilde{L}}(a,r)$ is
$1/\gamma$-Lipschitz
when viewed as a function of $r$, and is $1/(2\gamma)$-Lipschitz for any fixed $r$ when viewed as a function of $a$. Hence
\begin{align*}
|{\tilde{L}}(a,r)-{\tilde{L}}(a',r')|
&\leq |{\tilde{L}}(a,r)-{\tilde{L}}(a,r')| + |{\tilde{L}}(a,r')-{\tilde{L}}(a',r')| \\
&\leq \frac{1}{\gamma}\,|r-r'| + \frac{1}{2\gamma}\,|a-a'|\\
&\leq \sqrt{\frac{1}{\gamma^2}+\frac{1}{4\gamma^2}}\,\sqrt{(a-a')^2 + (r-r')^2} \\
&< \frac{2}{\gamma}\,\sqrt{(a-a')^2 + (r-r')^2}~,
\end{align*}
so that ${\tilde{L}}$ is $\frac{2}{\gamma}$-Lipschitz w.r.t. the Euclidean distance on $\mathscr Y$.
Furthermore, a quick comparison to the abstention loss
\[
L(a,r) = f_{\gamma}(a) 1_{r > 0} + c 1_{r \leq 0}
\]
reveals that (recall Figure \ref{f:1} (b) in the main text) :
\begin{itemize}
\item ${\tilde{L}}$ is an upper bound on $L$, i.e.,
\[
{\tilde{L}}(a,r) \geq L(a,r),\quad \forall\ (a,r) \in \mathscr Y~;
\]
\item ${\tilde{L}}$ approximates $L$ in that
\begin{equation}\label{e:approx}
{\tilde{L}}(a,r) = L(a,r),\quad \forall\ (a,r) \in \mathscr Y\,:\, |r| \geq \gamma~.
\end{equation}
\end{itemize}
With the above properties of ${\tilde{L}}$ handy, we are ready to prove Theorem \ref{th:infarms}.
\noindent{\bf Proof of Theorem \ref{th:infarms}.}\\
\iffalse
****************************************
\begin{retheo}~\ref{th:infarms} Fix any sequence of
$(x_1,{\mathcal{L}}_1),(x_2,{\mathcal{L}}_2)\ldots $ of context $x_t\in \mathscr X$ and
loss functions $ {\mathcal{L}}_t$. If the \textsc{ContextualEXP3}\ is run with
parameter
$\varepsilon \simeq
T^{-\frac{1}{2+d}}\,\left(\frac{1}{\gamma}+1\right)^{-\frac{2}{2+d}}$,
then
\begin{align*
& \max_{\xi \in \mathscr F} \E\left[\sum_{t=1}^T \tilde{L}_t(\xi_{I_t},z_t) \right] - \sum_{t=1}^T \tilde{L}_t(\xi,z_t) \leq {\tilde \mathscr O}\left(T^{\frac{d+1}{d+2}}\,\left(\frac{1}{\gamma}+1\right)^{\frac{d}{d+2}}\right) + M^*_{T}(\gamma)~.
\end{align*}
where $M^*_T(\gamma)$ is the number of $x_t$ such that $|r^*(x_t)| \leq \gamma$.
\end{retheo}
****************************************
\fi
\begin{proof}
On each ball $B \subseteq \mathscr X$ that \textsc{Cont\EXPABS}\ allocates during its online functioning, Theorem \ref{th:exp3rej} delivers the following
regret bound for the associated instance of \textsc{EXP3-ABS}:
\[
\frac{\log K_{\varepsilon}}{\eta} + \frac{\eta}{2}\,T_B (c^2+1)\,,
\]
where $T_B$ is the number of points $x_t$ falling into ball $B$. Now, taking into
account that ${\tilde{L}}$ is $\frac{2}{\gamma}$-Lipschitz, and that the functions $h$ and $r$ are assumed to be
$L_{\mathscr E}$-Lipschitz on $\mathscr X$, a direct adaptation of the proof of
Theorem 1 in \cite{CesaBianchiGaillardGentileCerchinovitz2017} gives the bound
\[
\sup_{\xi \in \mathscr E} \E \Bigg[ \sum_{t = 1}^T {\tilde{L}}( \xi_{I_t}, z_t) - \sum_{t = 1}^T {\tilde{L}}(\xi, z_t) \Bigg]
\leq
\frac{N_T\log K_{\varepsilon}}{\eta} + \frac{\eta}{2}\,T (c^2+1) +
L_{\mathscr E}\,\varepsilon\,\frac{2}{\gamma}\,T~,
\]
being $N_T \leq C_{\mathscr X}\varepsilon^{-d}$ the maximum number of balls created by \textsc{Cont\EXPABS}.
Use $c \leq 1$, and set $\eta = \sqrt{\frac{N_T\,\log K_{\varepsilon}}{T}}$ to get
\[
\sup_{\xi \in \mathscr E} \E \Bigg[ \sum_{t = 1}^T {\tilde{L}}( \xi_{I_t}, z_t) - \sum_{t = 1}^T {\tilde{L}}(\xi, z_t) \Bigg]
\leq
2\,\sqrt{T\,N_T\,\log K_{\varepsilon}} + L_{\mathscr E}\,\varepsilon\,\frac{2}{\gamma}\,T~,
\]
then optimize for $\varepsilon$ by setting
$\varepsilon \simeq
T^{-\frac{1}{2+d}}\,\left(\frac{1}{\gamma}\right)^{-\frac{2}{2+d}}$
(and disregarding $L_{\mathscr E}$ and log factors) to achieve
\begin{equation}
\label{e:boundrt}
\sup_{\xi \in \mathscr E} \E \Bigg[ \sum_{t = 1}^T {\tilde{L}}( \xi_{I_t}, z_t) - \sum_{t = 1}^T {\tilde{L}}(\xi, z_t) \Bigg]
=
{\tilde \mathscr O}\left(T^{\frac{d+1}{d+2}}\,\left(\frac{1}{\gamma}\right)^{\frac{d}{d+2}}\right)~.
\end{equation}
Finally, we are left with connecting the above bound on the regret
with a bound on the regret for $L$. To this effect, observe that
\begin{equation}
\label{e:upperloss}
\E\left[\sum_{t=1}^T {\tilde{L}}(\xi_{I_t},z_t) \right] \geq \E\left[\sum_{t=1}^T L(\xi_{I_t},z_t) \right]~,
\end{equation}
due to the fact that ${\tilde{L}}(\xi,z_t)$ is an upper bound on $L(\xi,z_t)$ for any $\xi$ and $z_t$. Moreover, if we set for brevity
\[
\xi^* = (h^*,r^*) = \arg\inf_{\xi\in \mathscr E} \sum_{t=1}^T L(\xi,z_t),\qquad \tilde{\xi}^* = \arg\inf_{\xi\in \mathscr E} \sum_{t=1}^T {\tilde{L}}(\xi,z_t)~,
\]
and denote by $M^*_T(\gamma)$ the number of $x_t$ such that $|r^*(x_t)| \leq \gamma$, we can write
\begin{eqnarray*}
\sum_{t=1}^T {\tilde{L}}(\tilde{\xi}^*,z_t)
&\leq&
\sum_{t=1}^T {\tilde{L}}(\xi^*,z_t)\\
&\leq&
\sum_{t\,:\,|r^*(x_t)| > \gamma}^T {\tilde{L}}(\xi^*,z_t) + M^*_{T}(\gamma)\\
&&{\mbox{(since ${\tilde{L}} \leq 1$))}}\\
&=&
\sum_{t\,:\,|r^*(x_t)| > \gamma}^T L(\xi^*,z_t) + M^*_{T}(\gamma)\\
&&{\mbox{(using (\ref{e:approx}))}}\\
&\leq&
\sum_{t=1}^T L(\xi^*,z_t) + M^*_{T}(\gamma)~.
\end{eqnarray*}
Combining with (\ref{e:boundrt}) and (\ref{e:upperloss}) gives
the following regret bound
\[
\sup_{\xi \in \mathscr E} \E \Bigg[ \sum_{t = 1}^T L( \xi_{I_t}, z_t) - \sum_{t = 1}^T L(\xi, z_t) \Bigg]
\leq
{\tilde \mathscr O}\left(T^{\frac{d+1}{d+2}}\,\left(\frac{1}{\gamma}\right)^{\frac{d}{d+2}}\right) + M^*_{T}(\gamma)~,
\]
thereby concluding the proof.
\end{proof}
\begin{remark}
Though we do not show it here, a better regret bound, of the form $T^{\frac{d}{d+1}}$ instead of $T^{\frac{d+1}{d+2}}$, can be obtained by adopting a hierarchical covering technique of the function space $\mathscr E$, each layer of the hierarchy being a pool of experts for the layer above it, see, e.g., \cite{CesaBianchiGaillardGentileCerchinovitz2017}. However, the resulting algorithm would be of theoretical interest only, since it is computationally infeasible.
\end{remark}
\section{Additional material for the stochastic setting}
\label{app:stochastic}
In this section, we present the proofs of the theoretical guarantees for
\textsc{UCB-NT}\ and \textsc{UCB-GT} as well as the proof of Proposition~\ref{prop:subset}.
An easy extension of the following theorems is to show that they hold when $ S_{j,t}= \sqrt{\frac{2\beta \log t}{Q_{j,t}} }$ for $\beta>2$, which implies slightly better constants in the regret bound, but for simplicity below, we set $\beta=\frac{5}{2}$. Moreover, we prove Theorem~\ref{th:ucbn} for the abstention loss $L$, but it holds for any general loss function.
\ignore{
\begin{align*}
T_j(t)= \sum_{s=1}^t 1_{I_s=j} \hspace{10mm} Q_{j,t} = \sum_{s=1}^t 1_{j\in N_t(I_s)} \hspace{10mm} S_{j,t}= \sqrt{\frac{5 \log t}{Q_{j,t-1}} }
\end{align*}
Please see the different framed boxes for definitions needed for
\textsc{UCB-NT}\ and \textsc{UCB-GT}. Recall that $\ell_s$ is a general
loss function and $L_s$ is the abstention loss function defined in the
Section~\ref{sec:problem}.}
\subsection{Regret of \textsc{UCB-NT}\ }
We now prove the theorem for \textsc{UCB-NT}\ based on the admissible $p$-partitioning of the
time-varying feedback graphs.
\noindent{\bf Proof of Theorem \ref{th:ucbn}.}\\
\begin{proof}
As is standard, the regret can be decomposed according to each arm $i$:
\begin{align*}
&\sum_{t=1}^T \E[L(\xi_{I_t},z_t) - L(\xi_*,z_t)] = \sum_{i=1}^K \sum_{t=1}^T \E[L(\xi_i,z_t) - L(\xi_*,z_t)] \E[1_{I_t = i}] = \sum_{i=1}^K \sum_{t=1}^T \Delta_i \E[1_{I_t = i}]
\end{align*}
so that we can focus on bounding the term $ \sum_{t=1}^T \E[1_{I_t = i}]$ for each $i$.
We split the expectation according to the events $Q_{i,t-1}>s_i$ and $Q_{i,t-1}\leq s_i$, where $s_i$ is a quantity to be determined:
\begin{align*}
\sum_{t=1}^T \E[1_{I_t = i}] &= \sum_{t=1}^T
\E[1_{I_t = i} (1_{Q_{i,t-1} \leq s_i} + 1_{Q_{i,t-1} > s_i})]\\
& \leq s_i + \sum_{t=1}^T \E[1_{I_t = i} 1_{Q_{i,t-1} > s_i}].
\end{align*}
We want to choose $s_i$ sufficiently large so that the second term is bounded while at the same time $s_i$ has a mild dependence on $T$.
Now, whenever $I_t=i$, by the design of the algorithm, it must be the case that the upper confidence bound of $i$ is smaller than that of any other expert. Thus,
\begin{align*}
& \E[1_{I_t = i} 1_{Q_{i,t-1} > s_i}]= \Pr[I_t = i, Q_{i,t-1} > s_i] \leq \Pr[\widehat{\mu}_{i,t-1} - S_{i,t-1} \leq \widehat{\mu}_{*,t-1} - S_{*,t-1}, Q_{i,t-1} > s_i],
\end{align*}
where $*$ is the best-in-class expert. We now use the terms $\mu_{*}$, $\mu_i$ and $S_{i,t-1}$ to reorder the first event in the probability on the right-hand side of the last expression as follows:
\begin{align*}
&0 \leq \widehat{\mu}_{*,t-1} - S_{*,t-1} - \widehat{\mu}_{i,t-1} + S_{i,t-1} \\
&\Leftrightarrow 0 \leq \left(\widehat{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \right) + \left(\mu_i - \widehat{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \right) + \left( \mu_{*} - \mu_i + 2 S_{i,t-1} \right).
\end{align*}
If we can show that the third term is negative, then the first and second term must be positive. Moreover, we will further show that the first and second terms can only be positive with an extremely low probability that is bounded by a constant independent of $T$. Furthermore, the third term will be negative whenever the slack term in the upper confidence bound is small enough, which amounts to choosing $s_i$ large enough.
In particular, by setting $s_i = \frac{20 \log(T)}{\Delta_i^2}$, we ensure that the event $Q_{i,t-1} > s_i$ implies that
\begin{align*}
&Q_{i,t-1} > \frac{20 \log(t)}{\Delta_i^2}
\Leftrightarrow \mu_{*} - \mu_i + 2 S_{i,t-1} < 0.
\end{align*}
As explained above, it then follows that
\begin{align*}
&\Pr[\widehat{\mu}_{i,t-1} - S_{i,t-1} \leq \widehat{\mu}_{*,t-1} - S_{*,t-1}, Q_{i,t-1} > s_i] \\
& \leq \Pr[\widehat{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \geq 0] + \Pr[\mu_i - \widehat{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \geq 0].
\end{align*}
We can bound these last probabilities using the union bound and
a concentration inequality such as Hoeffding's Inequality:
\begin{align*}
& \P[ \mu_i - \widehat{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \geq 0] \\
&= \P\Big[ -\tfrac{1}{Q_{i,t-1}}\sum_{s=1}^{t-1} L(\xi_i,z_s) 1_{i\in N_s(I_s)} + \mu_i - \sqrt{\tfrac{5 \log (t)}{Q_{i,t-1}}} \geq 0\Big].
\end{align*}
Now, the estimate $\widehat{\mu}_{i,t-1}$ is an average of
i.i.d. realizations of the random variable $ L(\xi_i,z)$, with $z \sim \mathscr D$, since the out-neighborhood of the
chosen expert only depends on previous observations. That is,
\begin{align*}
\frac{ \E[ \sum_{s=1}^{t-1}L(\xi_i,z_s) 1_{i\in N_s(I_s)} ]}{\E[ \sum_{s=1}^{t-1} 1_{i\in N_s(I_s)} ]}
&= \frac{ \E[ \sum_{s=1}^{t-1} \E[L(\xi_i,z_s) 1_{i\in N_s(I_s)}| i\in N_s(I_s) ]]}{\E[ \sum_{s=1}^{t-1} 1_{i\in N_s(I_s)} ]}\\
&= \frac{ \E[ \sum_{s=1}^{t-1}1_{i\in N_s(I_s)} \E[L(\xi_i,z_s) | i\in N_s(I_s) ]]}{\E[ \sum_{s=1}^{t-1} 1_{i\in N_s(I_s)} ]}\\
&= \frac{ \E[ \sum_{s=1}^{t-1}1_{i\in N_s(I_s)} \E[L(\xi_i,z_s)]]}{\E[ \sum_{s=1}^{t-1} 1_{i\in N_s(I_s)} ]}\\
&= \E[L(\xi_i,z)].
\end{align*}
Hence,
$\widehat{\mu}_{i,t-1}$ can be turned into an empirical estimate of
$\mu_{i}$ using the union bound as follows:
\begin{align*}
\P\Bigl( -\tfrac{1}{Q_{i,t-1}}\sum_{s=1}^{t-1} L(\xi_i,z_s) 1_{i\in N_s(I_s)} + \mu_i - \sqrt{\tfrac{5 \log (t)}{Q_{i,t-1}}} \geq 0\Bigl)
& \leq \P \Bigl( \exists n \in [1,t] : - \hat{\mu}_{i}^n+\mu_{i} - \sqrt{\tfrac{5 \log (t)}{n}}\Bigl) \\
& \leq \sum_{n=1}^t \frac{1}{t^{\frac{5}{2}}} = \frac{1}{t^{\frac{3}{2}}}~,
\end{align*}
where $ \hat{\mu}_{i}^n = \frac{1}{n}\sum_{s=1}^n L(\xi_i,z_s)$. By the same reasoning, we can also bound the probability of the best arm :
\[
\P\left( \widehat{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \geq 0 \right) \leq \sum_{n=1}^t \frac{1}{t^{\frac{5}{2}}} = \frac{1}{t^{\frac{3}{2}}}~.
\]
By assumption, for each $\mathscr C_{t,k}$,
$\forall i,j \in \mathscr C_{t,k}$, it is the case that
$i \in N_t(j)$. Moreover, for any $i \in [K]$, it must be the case
that for every $s \in [t]$, $i \in \mathscr C_{s,k}$ for some
$k \in [p]$. With these clusters $\mathscr C_{s,k}$,
we can write
\begin{align*}
&Q_{i,t} = \sum_{s=1}^t 1_{i \in N_s(I_s)} = \sum_{s=1}^t \sum_{j=1}^K 1_{i \in N_s(j)} 1_{I_s = j} \geq \sum_{s=1}^t \sum_{j\in \mathscr C_{s,k}} 1_{i \in N_s(j)} 1_{I_s = j}= \sum_{s=1}^t \sum_{j\in \mathscr C_{s,k}} 1_{I_s = j},
\end{align*}
which implies that
\begin{align*}
&\sum_{i=1}^K \sum_{t=1}^T \Delta_i \E[1_{I_t = i} 1_{Q_{i,t-1} \leq s_i}] \leq \sum_{k \in [p]} \left(\max_{\stackrel{t \in [T]}{j \in \mathscr C_{t,k}}} \Delta_j \right) \sum_{t=1}^T \sum_{j \in \mathscr C_{t,k}} \E\Bigg[1_{I_t = j} 1_{Q_{j,t-1} \leq \max_{\stackrel{t \in [T]}{j \in \mathscr C_{t,k}}} s_j}\Bigg] \\
&\leq \sum_{k \in [p]} \left(\max_{t \in [T]} \max_{j \in \mathscr C_{t,k}} \Delta_j \right) \left( \max_{t \in [T]} \max_{j \in \mathscr C_{t,k}} s_j \right). \\
\end{align*}
Combining the above calculations, applying our definition for $s_i$,
and using the fact that the above analysis holds for any such partition shows that
\begin{align*}
&\sum_{t=1}^T \E[L(\xi_{I_t},z_t) -L(\xi_{*},z_t)] \leq \min_{\stackrel{p}{\{\mathscr C_{t,k}\}_{t\in[T], k \in p}}} \sum_{k \in [p]} \left(\max_{\stackrel{t \in [T]}{j \in \mathscr C_{t,k}}} \Delta_j \right) \left( \max_{\stackrel{t \in [T]}{j \in \mathscr C_{t,k}}} \frac{20 \log(T)}{\Delta_j^2} \right) + 5K .\end{align*}
which implies the bound of the theorem.
\end{proof}
\subsection{Regret of \textsc{UCB-GT}\ }
Next, we prove the regret bound for \textsc{UCB-GT}, which demonstrates how one can exploit the bias and feedback structure in the problem.
\noindent{\bf Proof of Theorem \ref{th:ucbabs}.}\\
\begin{proof}
As in the previous proof, the regret can be written as follows:
\begin{align*}
&\sum_{t=1}^T \E[L(\xi_{I_t},z_t) - L(\xi_*,z_t)] = \sum_{i=1}^K \sum_{t=1}^T \E[L(\xi_i,z_t) - L(\xi_*,z_t)] \E[1_{I_t = i}] = \sum_{i=1}^K \sum_{t=1}^T \Delta_i \E[1_{I_t = i}] \\ &= \sum_{i=1}^K \sum_{t=1}^T \Delta_i \E[1_{I_t = i} (1_{Q_{i,t-1} \leq s_i} + 1_{Q_{i,t-1} > s_i})], \\
\end{align*}
where $s_i$ is to be determined. We can decompose this sum and bound the second term as follows:
\begin{align*}
& \E[1_{I_t = i} 1_{Q_{i,t-1} > s_i}]= \Pr[I_t = i, Q_{i,t-1} > s_i] \leq \Pr[\widehat{\mu}_{i,t-1} - S_{i,t-1} \leq \widehat{\mu}_{*,t-1} - S_{*,t-1}, Q_{i,t-1} > s_i].
\end{align*}
Currently, $\widehat{\mu}_{i,t-1}$ is a biased estimate of
$\mu_i$. This is because whenever $x_s$ fell in the region $\{x\,:\,r_i(x_s)>0 \wedge r_{I_s}(x_s) \leq 0\}$, we did not have access to the label $y_s$. In these cases, the \textsc{UCB-GT}\ algorithm updated the expert $i$ optimistically as if the expert was correct at that time step. We can decompose this biased
estimate $\widehat{\mu}_{i,t-1}$ into two terms: $\widehat{\mu}_{i,t-1}= \widetilde{\mu}_{i,t-1} - \epsilon_{i,t}$. The first term, $\widetilde{\mu}_{i,t-1}$, is an unbiased estimate of arm $i$ similar to the estimates in Theorem~\ref{th:ucbn}, and the second term is
the misclassification rate $\epsilon_{i,t-1}$ over the region $r_i(x_s)>0\cap r_{I_s}(x_s) \leq 0$, that is $\epsilon_{i,t-1}= \frac{1}{Q_{i,t-1}} \sum_{s=1}^{t-1} 1_{y_sh_i(x_s) \leq 0}1_{r_i(x_s)>0, r_{I_s}(x_s) \leq 0}$.
Now, by design of the \textsc{UCB-GT}\, if arm $i$ is chosen at time $t$, it must be the case that $\widehat{\mu}_{i,t-1} - S_{i,t-1} \leq \widehat{\mu}_{*,t-1} - S_{*,t-1}$. We can expand and rewrite this expression as follows:
\begin{align*}
&0 \leq \widehat{\mu}_{*,t-1} + \epsilon_{i^*,t-1} - \epsilon_{i^*,t-1} - S_{*,t-1} - \widehat{\mu}_{i,t-1}- \epsilon_{i,t-1} + \epsilon_{i,t-1} + S_{i,t-1} \\
&\Leftrightarrow 0 \leq \left(\widetilde{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \right) + \left(\mu_i - \widetilde{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \right) + \left( \mu_{*} - \mu_i + 3 S_{i,t-1} \right),
\end{align*}
where we used the fact that $ - \epsilon_{i^*,t-1} \leq 0 $, and we bounded $\epsilon_{i,t-1}$ as follows
\begin{align*}
\epsilon_{i,t-1}
& \leq \frac{1}{Q_{i,t-1}} \sum_{s=1}^{t-1} 1_{y_sh_i(x) \leq 0} 1_{r_i(x_s)>0, r_{I_s}(x_s) \leq 0} \\
& \leq \frac{1}{Q_{i,t-1}} \sum_{s=1}^{t-1} 1_{r_i(x_s)>0, r_{I_s}(x_s) \leq 0} = \frac{1}{Q_{i,t-1}} \sum_{s=1}^{t-1} \sum_{\xi_j \in \mathcal{E}-\xi_i} 1_{r_i(x_s)>0, r_{j}(x_s) \leq 0} 1_{I_s=j}\\
& \leq \frac{1}{Q_{i,t-1}} \sum_{\xi_j \in \mathcal{E}-\xi_i} \sum_{s=1}^{t-1} 1_{r_i(x_s)>0, r_{j}(x_s) \leq 0} = \frac{t-1}{Q_{i,t-1}} \sum_{\xi_j \in \mathcal{E}-\xi_i} \frac{1}{t-1} \sum_{s=1}^{t-1} 1_{r_i(x_s)>0, r_{j}(x_s) \leq 0} \\
& \leq \frac{(t-1) (K-1) }{Q_{i,t-1}} \gamma_{i,t-1}\\
& \leq \sqrt{\frac{5 \log (t) }{ Q_{i,t-1} } }.
\end{align*}
\ignore{
Now for $\epsilon_{i^*,t-1}$
\begin{align*}
&|f_{i^*,t-1}| \leq \frac{(t-1) (K-1) }{Q_{i^*}(t-1)} \epsilon_{t-1}
\end{align*}
and want to show that
$ \frac{(t-1) (K-1) }{Q_{i^*}(t-1)} \epsilon_{t-1} \leq \sqrt{\frac{2\beta \log (t) }{ Q_{i,t-1} } }$. For $\epsilon_{t-1}= \min_{i\in [K]} \frac{ Q_{j,t-1}}{ (K-1) (t-1)} \sqrt{\frac{2 \beta \log (t-1) }{ Q_{i,t-1} } }$, this only holds if $Q_{i,t-1} \leq Q_{i^*}(t-1)$}
The rest of the proof now follows by similar arguments as in the proof of Theorem~\ref{th:ucbn}. Specifically, we can choose
$s_i$ such that the term $\mu_{*} - \mu_i + 3 S_{i,t-1}$ is negative, and since now $\widetilde{\mu}_{*,t-1}$ and $\widetilde{\mu}_{i,t-1}$ are unbiased estimates, we can bound the $\Pr [\widetilde{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \geq 0]$ and $\Pr[\mu_i - \widetilde{\mu}_{i,t-1} - S_{i,t-1}\geq 0 ]$ using concentration inequalities.
\ignore{
By definition, $S_{i,t}$ is the slack quantity $S_{i,t} = \sqrt{\frac{2\beta \log(t)}{Q_{i,t}}}$.
By setting $s_i = \frac{18\beta \log(T)}{\Delta_i^2}$, we ensure that the event $Q_{i,t-1} > s_i$ implies that
\begin{align*}
&Q_{i,t-1} > \frac{18\beta \log(t-1)}{\Delta_i^2}
\Leftrightarrow \mu_{*} - \mu_i + 3 S_{i,t-1} < 0.
\end{align*}
It follows that
\begin{align*}
&\Pr[\widetilde{\mu}_{i,t-1} - S_{i,t-1} \leq \widetilde{\mu}_{*,t-1} - S_{*,t-1}, Q_{i,t-1} > s_i] \\
& \leq \Pr[\widetilde{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \geq 0] + \Pr[\mu_i - \widetilde{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \geq 0].
\end{align*}
We can then bound these probabilities by using the union bound and concentration inequality:
\begin{align*}
& \P[ \mu_i - \widetilde{\mu}_{i,t-1} + S_{i,t-1} - 2 S_{i,t-1} \geq 0] \\
&= \P[ -\tfrac{1}{Q_{i,t-1}}\sum_{s=1}^{t-1} \ell_s(\xi_i) 1_{i\in N(I_s)} + \mu_i - \sqrt{\tfrac{2\beta \log (t-1)}{Q_{i,t-1}}} \geq 0]
\end{align*}
The estimate $\widetilde{\mu}_{i,t-1}$ is an average of
i.i.d. observation of $ \ell(\xi_i)$ since the out-neighborhood of the
chosen expert only depends on previous observations. Hence
$\widetilde{\mu}_{i,t-1}$ can be turned into an empirical estimate of
$\mu_{i}$ using the union bound as follows:
\begin{align*}
& \P[ -\tfrac{1}{Q_{i,t-1}}\sum_{s=1}^{t-1} \ell_s(\xi_i) 1_{i\in N(I_s)} + \mu_i - \sqrt{\tfrac{2\beta \log (t-1)}{Q_{i,t-1}}} \geq 0] \\
& \leq \P [ \exists n \in [1,t] : - \widetilde{\mu}_{i}^n+\mu_{i} - \sqrt{\tfrac{2\beta \log (t-1)}{n}} ] \\
& \leq \sum_{n=1}^t \frac{1}{t^\beta} = \frac{1}{t^{\beta-1}}
\end{align*}
where $ \widetilde{\mu}_{i}^n = \frac{1}{n}\sum_{s=1}^n \ell_s(\xi_i)$. By the same reasoning, we can also bound the probability of the best arm : $\P[ \widetilde{\mu}_{*,t-1} - S_{*,t-1} - \mu_{*} \geq 0] \leq \sum_{n=1}^t \frac{1}{t^\beta} = \frac{1}{t^{\beta-1}}$.
By assumption, for each $\mathscr J_{t,k}$,
$\forall i,j \in \mathscr J_{t,k}$, it is the case that $i \in N_t(j)$.
Moreover, for any $i \in [K]$, it must be the case that for every $s \in [t]$, $i \in \mathscr J_{s,k}$ for some $k \in [p]$.
With these $(\mathscr J_{s,k})_{s,k}$, we can write
\begin{align*}
&Q_{j,t} = \sum_{s=1}^t 1_{i \in N_s(I_s)} = \sum_{s=1}^t \sum_{j=1}^K 1_{i \in N_s(j)} 1_{I_s = j} \quad \geq \sum_{s=1}^t \sum_{j\in \mathscr J_{s,k}} 1_{i \in N_s(j)} 1_{I_s = j}= \sum_{s=1}^t \sum_{j\in \mathscr J_{s,k}} 1_{I_s = j},
\end{align*}
which implies that
\begin{align*}
&\sum_{i=1}^K \sum_{t=1}^T \Delta_i \E[1_{I_t = i} 1_{Q_{i,t-1} \leq s_i}] \leq \sum_{k \in [p]} \left(\max_{\stackrel{t \in [T]}{j \in \mathscr J_{t,k}}} \Delta_j \right) \sum_{t=1}^T \sum_{j \in \mathscr J_{t,k}} \E\Bigg[1_{I_t = j} 1_{Q_{j,t-1} \leq \max_{\stackrel{t \in [T]}{j \in \mathscr J_{t,k}}} s_j}\Bigg] \\
&\leq \sum_{k \in [p]} \left(\max_{t \in [T]} \max_{j \in \mathscr J_{t,k}} \Delta_j \right) \left( \max_{t \in [T]} \max_{j \in \mathscr J_{t,k}} s_j \right). \\
\end{align*}
Combining the above calculations, applying our definition for $s_i$,
and using the fact that the above analysis holds for any such
partition shows that
\begin{align*}
&\sum_{t=1}^T \E[L(\xi_{I_t},z_t) - L(\xi_{*},z_t)] \leq \min_{\stackrel{p}{\{\mathscr J_{t,k}\}_{t\in[T], k \in p}}} \sum_{k \in [p]} \left(\max_{\stackrel{t \in [T]}{j \in \mathscr J_{t,k}}} \Delta_j \right) \left( \max_{\stackrel{t \in [T]}{j \in \mathscr J_{t,k}}} \frac{8 \beta \log(T)}{\Delta_j^2} \right) + K \frac{\beta}{\beta - 2}.\end{align*}
which is the bound of the theorem. }
\end{proof}
\subsection{Linear regret without the subset property}
\label{app:subset}
In this section, we prove Proposition~\ref{prop:subset}, which illustrates
that when the subset property does not hold for a feedback graph, then
it is possible to incur linear regret.
\noindent{\bf Proof of Proposition \ref{prop:subset}.}\\
\begin{proof}
Let $p^* \in (0,1)$. We design a setting in which with probability
at least $p^*$, the \textsc{UCB-NT}\ algorithm incurs linear regret.
Since the family of abstention functions induces a feedback graph that
violates the subset property, there exist pairs $(h_i, r_i)$ and $(h_j, r_j)$
and points $x^*$, $\tilde{x}$ for which $x^* \in \mathcal{A}_i \setminus \mathcal{A}_j$,
$\tilde{x} \in \mathcal{A}_i \cap \mathcal{A}_j$, where $\mathcal{A}_i$ and $\mathcal{A}_j$
are the acceptance regions associated with $r_i$ and $r_j$, respectively,
and the feedback graph is designed such that the algorithm updates the pair $(h_i, r_i)$
when the pair $(h_j, r_j)$ is selected.
Now, for some $p \in (0,1)$ to be determined later, consider a distribution with probability $p$ on $(\tilde{x}, \tilde{y})$
and $(1-p)$ on $(x^*, y^*)$.
We choose the set of hypothesis functions $\mathscr H = \{h_i, h_j\}$, the loss function $\ell$ in (\ref{e:absloss}),
and the labels $y^*$ and $\tilde{y}$ in such a way that
$\ell(\tilde{y}, h_i(\tilde{x})) = c - \beta$, $\ell(\tilde{y}, h_j(\tilde{x})) = c - \alpha$,
and $\ell(y^*, h_i(x^*)) = 0$, where $\alpha, \beta$ are values that will
be later specified. For instance, we can consider the hinge loss $\ell(y,\hat{y}) = (1-y \hat{y})_+$,
and $h_i$, $h_j$ such that $h_i(\tilde{x}) = \frac{1-c+\beta}{\tilde{y}}$, $h_j(\tilde{x}) = \frac{1-c+\alpha}{\tilde{y}}$,
and $h_i(x^*) = \frac{1}{y^*}$. Note that since $r_j(x^*) < 0$,
$\ell(y^*, h_j(x^*))$ can be any value.
Now, by construction, $\mu_i = (c - \beta) p$ and $\mu_j = (c-\alpha) p + c (1-p) = c - \alpha p$.
We claim that we can choose $\alpha$, $\beta$ and $p$ such that
(1) $\alpha > \beta$; (2) $\mu_i < \mu_j$; (3) $\mu_j < \ell(\tilde{y}, h_i(\tilde{x}))$.
The first condition is immediate. The second condition is equivalent to
$cp - \beta p < c - \alpha p$, which is itself equivalent to
$\alpha - \beta < \frac{c(1-p)}{p}$. By continuity, we can choose $\alpha$
and $\beta$ close enough such that this is true for any $p \in (0,1)$. The
third condition is equivalent to $c-\alpha p < c - \beta$, which is itself
equivalent to $\beta < \alpha p$. This is true for $p$ close enough
to $1$.
Now let $n \in \mathbb{N}$ be large enough such that $\mu_j < \ell(\tilde{y}, h_i(\tilde{x})) - \sqrt{\frac{5\log(n)}{n}}$. By continuity, we can choose $p$ large enough such that $p > (p^*)^{1/n}$,
and for this choice of $p$, we can choose $\alpha$ and $\beta$ such that
$\alpha > \beta$, $\alpha, \beta < c$, $\alpha - \beta < \frac{c(1-p)}{p}$,
and $\beta < \alpha p$. For instance, if we, without loss of generality,
assume that $p > \frac{1}{2}$, then we can choose, $\alpha = \frac{c(1-p)}{2p}$
and $\beta=\frac{c(1-p)}{4}$.
Then with probability $p^n > p^*$, the point $\tilde{x}$ will be sampled $n$ times
at the start of the game, such that the pair $(h_j, r_j)$ will have a lower
confidence bound than the pair $(h_i, r_i)$ at all time steps. Thus,
\textsc{UCB-NT}\ will choose the pair $(h_j, r_j)$ throughout the entire game, even though
$\mu_i < \mu_j$. Consequently, the regret of the algorithm will be at least $T (\mu_j - \mu_i)$.
\end{proof}
\ignore{
\section{\textsc{UCB-GT}}
This algorithm is based on a time-varying graph that updates as many
experts as possible while following the observability assumptions of
the abstention setting.
The feedback graph at each time $t$ is defined as follows: there is an
edge from expert $j$ to expert $i$ either if acceptance region of
expert $i$ is subset of acceptance region of expert $j$ or if
$\hat{p}_{i,j}^{t-1}\leq \epsilon_{t-1}$ where
$\hat{p}_{i,j}^{t-1}:=\frac{1}{t-1} \sum_{s=1}^{t-1} 1_{r_i(x_s)>0,
r_{j}(x_s) \leq 0}$ and
Option 1: update as if expert $j$ was correct on this region.
$\epsilon_{i,t-1}= \frac{ \sqrt{Q_{i,t-1}}}{ (K-1) (t-1)} \sqrt{2 \beta \log (t-1)} $.
After: use proxy of $\hat{\mu}_{j,t-1}$ for loss of expert $j$ at time $t$ .
$\epsilon_{t-1}= \min_{i\in [K]} \frac{ 1}{ (K-1) (t-1)} \sqrt{\frac{2 \beta \log (t-1) }{ Q_{i,t-1} } }$.
If the chosen expert $I_t$ rejects while expert $i$ accepts and
satisfies $\hat{p}_{i,I_t}^{t-1}\leq \epsilon_{t-1}$, then we don't
have access to the true label. In this case, we update our loss
according to the $\widehat{\mu}_{i,t-1}$ since it is our best
estimate of the loss. This of course will create a bias in our
estimate of $\hat{\mu}_{i,t}$, but we will see that for the given
choice of $\epsilon_{t,i}$ this bias does not affect the regret.
\begin{retheo}~\ref{th:ucbabs}
\textit{
For any $\beta>2$, after $T$ rounds, the regret of \textsc{UCB-GT}\ is
at most:
\begin{align*}
&\sum_{t=1}^T \E[L(\xi_{I_t},z_t) - L(\xi_*,z_t)] \leq K \frac{\beta}{\beta - 2} \\
& + \min_{p_+,\mathcal{F}_{p_+}} \sum_{k \in [p_+]} \bigg(\max_{j\in \mathtt{J}_{+,k} } \Delta_j \bigg) \bigg( \frac{18 \beta \log(T)}{\min_{j\in \mathtt{J}_{+,k} }\Delta_j^2} \bigg)+ \min_{p_-,\mathcal{F}_{p_-}} \sum_{k \in [p_-]} \bigg(\max_{j\in \mathtt{J}_{-,k} } \Delta_j \bigg) \bigg( \frac{18 \beta \log(T)}{\min_{j\in \mathtt{J}_{-,k} }\Delta_j^2} \bigg) ,
\end{align*}
where $\mathtt{J}_{-,k} = \cup_{t\in[T]}\mathscr J_{t,k}^- $ and
$\mathtt{J}_{+,k}= \cup_{t\in[T]}\mathscr J_{t,k}^+ $.}
\end{retheo}
}
\section{Additional experimental results}\label{app:expresults}
\begin{table}[t]
\centering
\begin{tabular}{ c | c}
Dataset & Number of features \\
\hline
{\tt covtype} & 54 \\
{\tt ijcnn} & 22 \\
{\tt skin} & 3 \\
{\tt HIGGS} & 28 \\
{\tt guide} & 4 \\
{\tt phishing} & 68 \\
{\tt cod} & 8 \\
{\tt eye} & 14 \\
{\tt CIFAR} & 25 \\
\end{tabular}
\caption{Table shows the number of features of each dataset.}
\label{tb:features}
\end{table}
In this section, we present several figures with our experimental results. Figure~\ref{fig:fullres1} and Figure~\ref{fig:fullres2} show the regret for different abstention costs $c\in \{0.1,0.2,0.3\}$ for all our datasets. We see that in general \textsc{UCB-GT}\ outperforms \textsc{UCB-NT}\ and \textsc{UCB}\ for all datasets and is even within the standard deviation of the \textsc{FS} 's regret for some of datasets. The figures also indicate that the regret of UCB decreases slowly. This is expected, since there are $2,000$ experts, $10,000$ time steps, and the algorithm only updates a single expert per time step. Figure~\ref{fig:frac1} and Figure~\ref{fig:frac2} shows the fraction of abstained points for all the datasets. Moreover, Figure~\ref{fig:extremec} shows how the fraction of points abstained on varies with abstention cost for two extreme values $c\in\{0.001,0.9 \}$. Again \textsc{UCB-GT}\ has lower regret than \textsc{UCB-NT}\ and \textsc{UCB}\ and, as expected, the fraction of points decreases as the cost of abstention increases. Figure~\ref{fig:confexpert} shows the effect of using confidence-based experts and indicates the choice of experts does not affect the relative performance of the algorithms. Furthermore, we wanted to test the effects of changing the number of experts and so Figure~\ref{fig:fullsmall} shows the regret of three datasets when the number of experts is $K=500$ and $T=5,000$. For this set of experts, we find a similar pattern of performance as above.
c
Next, we describe in more detail the datasets and how they were processed. In Table~\ref{tb:features}, we show the number of features of each dataset. For all datasets, we normalized the features to be in the range of $[-1,1]$. Note that the reason for choosing abstention functions with radius range $(0, \sqrt{d})$ is to cover the entire hypercube $[-1,1]^d$ with our concentric annuli. For the {\tt CIFAR} dataset, we extracted the first twenty-five principal components of the
{\tt horse} and {\tt boat} images, projected the images on these
components, and normalized the range of the projections to $[-1,1]$. The features of the synthetic dataset are drawn from the uniform distribution over $[-1,1]^2$ and the label is determined
by the sign of the projection of a point onto the normal of the
diagonal hyperplane $y=-x$.
The confidence-based abstention function has the form $r=|h(x)| - \theta$. In our experiments (Figure ~\ref{fig:confexpert}), we generated twenty abstention functions with threholds $\theta \in (0,\ldots, 0.25)$, which are paired with each predictor. The predictors are axis-aligned planes along each feature of the dataset. For each dataset, the number of predictors is $\floor{100/d}$ where $d$ is the dimension of the dataset. We chose twenty abstention functions and about 100 prediction functions in order to match the experimental setup of the randomly drawn experts. The total number of experts is then $\floor{100/d}\cdot 20\cdot d$. Note that we only tested some of our datasets since for larger dimensions $d$, the number of experts per feature is too small.
\clearpage
\subsection{Average regret for different abstention costs and datasets}
\label{app:exp_avgreg}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_25.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_25.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_25.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_22.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_22.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_22.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_28.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_28.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_28.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_68.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_68.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_68.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_54.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_54.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_54.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{
A graph of the averaged regret $R_t(\cdot)/t$ with standard deviations as a function of $t$ (log scale) for
{\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost.
Each row is a dataset, starting from the top row we have: {\tt CIFAR}, {\tt ijcnn}, {\tt HIGGS}, {\tt phishing}, and {\tt covtype}. }
\label{fig:fullres1}
\vskip -.1in
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_14.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_14.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_14.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_8.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_8.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_8.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_1.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_1.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_1.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_3.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_3.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_3.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_4.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_4.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/regret_c03_d_4.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{
A graph of the averaged regret $R_t(\cdot)/t$ with standard deviations as a function of $t$ (log scale) for {\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost. Each row is a dataset, starting from the top row we have: {\tt eye}, {\tt cod-ran}, {\tt synthetic}, {\tt skin}, and {\tt guide}.
}
\label{fig:fullres2}
\vskip -.1in
\end{figure*}
\clearpage
\subsection{Average fraction of abstention points for different abstention costs and datasets}
\label{app:exp_avgfracabs}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_25.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_25.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_25.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_22.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_22.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_22.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_28.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_28.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_28.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_68.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_68.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_68.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_54.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_54.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_54.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{
A graph of the averaged fraction of abstained points with standard deviations as a function of $t$ (log scale) for {\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost. Each row is a dataset, starting from the top row we have: {\tt CIFAR}, {\tt ijcnn}, {\tt HIGGS}, {\tt phishing}, and {\tt covtype}. }
\label{fig:frac1}
\vskip -.1in
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_14.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_14.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_14.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_8.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_8.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_8.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_1.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_1.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_1.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_3.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_3.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_3.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_4.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_4.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/counts_c03_d_4.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{ A graph of the averaged fraction of abstained points with standard deviations as a function of $t$ (log scale) for {\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost. Each row is a dataset, starting from the top row we have: {\tt eye}, {\tt cod-ran}, {\tt synthetic}, {\tt skin}, and {\tt guide}.
}
\label{fig:frac2}
\vskip -.1in
\end{figure*}
\clearpage
\subsection{Average regret and fraction of abstention points for extreme abstention costs}
\label{app:exp_extremecost}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c c }
\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c0001_d_25.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c0001_d_25.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c09_d_25.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c09_d_25.pdf} \\
\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c0001_d_22.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c0001_d_22.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c09_d_22.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c09_d_22.pdf} \\
\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c0001_d_68.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c0001_d_68.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c09_d_68.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c09_d_68.pdf} \\
\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c0001_d_54.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c0001_d_54.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/regret_c09_d_54.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.2,trim= 5 10 10 5, clip=true]{exp_results3/counts_c09_d_54.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{
A graph of the averaged regret $R_t(\cdot)/t$ and fraction of points rejected with standard deviations as a function of $t$ (log scale) for
{\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost. The fraction of points decreases as the cost of abstention increases. The \textsc{UCB-GT}\ outperforms \textsc{UCB-NT}\ and \textsc{UCB}\ while approaching the performance of \textsc{FS}\ even at these extreme values of $c$. Each row is a dataset, starting from the top row we have: {\tt CIFAR}, {\tt ijcnn}, {\tt phishing}, and {\tt covtype}. }
\label{fig:extremec}
\vskip -.1in
\end{figure*}
\clearpage
\subsection{Average regret for confidence-based experts }
\label{app:exp_confidence}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c01_d_1.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c02_d_1.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c03_d_1.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c01_d_3.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c02_d_3.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c03_d_3.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c01_d_4.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c02_d_4.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c03_d_4.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c01_d_22.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c02_d_22.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c03_d_22.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c01_d_25.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c02_d_25.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{confidence/regret_c03_d_25.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{ A graph of the averaged regret $R_t(\cdot)/t$ with standard deviations as a function of $t$ (log scale) when using the confidence based experts for {\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} }. Each row is a dataset, starting from the top row we have: {\tt synthetic}, {\tt skin}, {\tt guide}, {\tt ijcnn} and { \tt CIFAR}.
}
\label{fig:confexpert}
\vskip -.1in
\end{figure*}
\clearpage
\subsection{Average regret for a smaller set of experts }
\label{app:exp_fewerexperts}
\begin{figure*}[!ht]
\begin{center}
\begin{tabular}{ c c c }
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c01_d_4.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c02_d_4.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c03_d_4.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c01_d_1.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c02_d_1.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c03_d_1.pdf} \\
\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c01_d_3.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c02_d_3.pdf}&
\hspace*{-5mm}\includegraphics[scale=0.25,trim= 5 10 10 5, clip=true]{exp_results3/exp_experts_small/regret_c03_d_3.pdf} \\
\end{tabular}
\end{center}
\vskip -.15in
\caption{
A graph of the averaged regret $R_t(\cdot)/t$ of abstained points with standard deviations as a function of $t$ (log scale) for
{\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and {\color{blue} \textsc{FS} } for different values of abstentions cost.
Each row is a dataset, starting from the top row we have: {\tt guide}, {\tt synthetic}, and {\tt skin}. We used $K=500$ experts and $T=5,000$ rounds in order to see the effect when changing the number of experts used. }
\label{fig:fullsmall}
\vskip -.1in
\end{figure*}
\section{Experiments}
\vspace{-0.03in}
\label{sec:experiments}
\begin{figure*}[t]
\vskip -.05in
\begin{center}
\begin{tabular}{c@{\hspace{1.5cm}}c@{\hspace{1.5cm}}c}
\includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{exp_results3/regret_c02_d_8.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{exp_results3/counts_c02_d_8.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{smoothedge_c02_d_8} \\
\includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{exp_results3/regret_c01_d_4.pdf} &
\hspace*{-5mm} \includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{exp_results3/counts_c01_d_4.pdf} &
\hspace*{-5mm}\includegraphics[scale=0.182,trim= 5 10 10 5, clip=true]{smoothedge_c01_d_4.pdf} \\
\end{tabular}
\end{center}
\vskip -.18in
\caption{From the left, we show graphs of the average regret
$R_t(\cdot)/t$, fraction of points the chosen expert abstained on, and
the number of edges of the feedback graph as a function of $t$ (log-scale) for
{\color[rgb]{0.16,0.67,0.16}\textsc{UCB-GT} }, \textsc{UCB-NT} , {\color{red} \textsc{UCB} }, and
{\color{blue} \textsc{FS} }. Top row is the results for {\tt cod-rna}
for cost $c=0.2$ and bottom row is the {\tt
guide} for cost $c=0.1$. More results are in Appendix \ref{app:expresults}.}
\label{fig:expmain}
\vskip -.2in
\end{figure*}
\vspace{-0.01in}
In this section, we report the results of several experiments on ten
datasets comparing \textsc{UCB-GT}, \textsc{UCB-NT}\ with feedback graph $G^{\text{\sc sub}}$,
vanilla \textsc{UCB}\ (with no sharing information across experts), as well as Full-Supervision, \textsc{FS}.
\textsc{FS}\ is an algorithm that at each round chooses the expert $\xi_j$ with the smallest abstention
loss so far, ${\widehat \mu}_{j,t-1}$, and even if this expert abstains, the algorithm
receives the true label and can update the empirical abstention loss
estimates for all experts. \textsc{FS}\ reflects an unrealistic and overly optimistic scenario that
clearly falls outside the abstention setting, but it provides an upper
bound for the best performance we may hope for.
\vspace{-0.01in}
We used the following eight datasets from the UCI data repository:
{\tt HIGGS}, {\tt phishing}, {\tt ijcnn}, {\tt covtype}, {\tt eye},
{\tt skin}, {\tt cod-rna}, and {\tt guide}. We also used the {\tt
CIFAR} dataset from \cite{Krizhevsky2009}, where we extracted the
first twenty-five principal components and used their projection
as features, and a synthetic dataset of points drawn according to
the uniform distribution in $[-1,1]^2$.
For each dataset, we generated a total of $K = 2\mathord{,}000$ experts and all
the algorithms were tested for a total of $T = 10\mathord{,}000$ rounds. The
experts, $\xi=(h,r)$, were chosen in the following way. The predictor
functions $h$ are hyperplanes centered at the origin whose normal
vector in $\mathbb{R}^d$ is drawn randomly from the Gaussian
distribution, $\mathcal{N}(0,1)^d$, where $d$ is the dimension of the
feature space of the dataset. The abstention functions $r$ are
concentric annuli around the origin with radii in $(0,
\tfrac{\sqrt{d}}{20},\tfrac{2\sqrt{d}}{20} \ldots, \sqrt{d})$. For each dataset, we generated 100
predictor functions and each predictor function
$h$ is paired with the 20 abstention functions $r$.
For a fixed set of experts, we first calculated the regret by
averaging over five random draws of the data, where the best-in-class
expert was determined in hindsight as the one with the minimum
average cumulative abstention loss. We then repeated this experiments five times
over different sets of experts and averaged the results. We report these results
for
$c \in \{0.1,0.2,0.3\}$.
\vspace{-0.01in}
Figure~\ref{fig:expmain} shows the averaged regret $R_t(\cdot)/t$
with standard deviations across the five repetition for the different
algorithms as a function of $t \in [T]$ for two datasets. In
Appendix~\ref{app:expresults}, we present plots of the regret for all
ten datasets. These results show that \textsc{UCB-GT}\ outperforms both \textsc{UCB-NT}\
and \textsc{UCB}\ on all datasets for all abstention cost values.
Remarkably, \textsc{UCB-GT}'s performance is close to the performance of \textsc{FS}\
for most datasets, thereby implying that \textsc{UCB-GT}\ attains almost the
best regret that we could hope for. We also find that \textsc{UCB-NT}\ performs
better than the vanilla \textsc{UCB}.
\vspace{-0.01in}
Figure~\ref{fig:expmain} also illustrates the fraction of points
in which the chosen expert abstains, as well as the number of edges in
the feedback graph as a function of rounds. We only plot the
number of edges of \textsc{UCB-GT}\, since that is the only graph that varies
with time. For both experiments depicted and in general for the rest of the datasets, the number of edges for
\textsc{UCB-GT}\ is between 1 million to 3 million, which is at least a factor of 5 more than
for \textsc{UCB-NT}, where the number of edges we observed are of the order
200,000. \textsc{FS}\ enjoys the full information property and the number
of edges is fixed at 4 million (complete graph).
The increased information sharing of \textsc{UCB-GT}\ is obviously a strong
contributing factor to the algorithm's improvement in regret relative
to \textsc{UCB-NT}. In general, we find that, provided we are able to keep estimation
bias under control, the higher the number of edges, the smaller the regret.
As far as dependence on $c$ is concerned, in our experiments, we find
that the fraction of points the chosen expert abstains on always
decreases as $c$
increases, but also that this fraction depends on the dataset and the
experts used.
\vspace{-0.01in}
\ignore{Finally, Appendix D also provides more experiments analyzing different aspects of the problem. We tested how the choice of experts and the number of experts might impact our results as well as extreme values of the abstention cost, $c\in \{0.001, 0.9\}$.}
Finally, Appendix~\ref{app:expresults} also provides more experiments analyzing different
aspects of the problem. We tested how the choice of experts (where instead we use confidence-based experts) and the number of experts might
impact our results.
Lastly, we tested extreme values of the abstention costs and as expected, the fraction of abstained points is large for $c=0.001$ and small for
$c=0.9$. In all of these additional experiments, we find again that \textsc{UCB-GT}\ outperforms \textsc{UCB-NT}.
\ignore{(instead
of generating the experts from a Gaussian distribution, we used axis-aligned predictors
and confidence-based abstention functions) and the number of experts might}
\section{Introduction}
\label{sec:introduction}
\vspace{-0.05in}
We consider an online learning scenario, central in many
applications, where the learner is granted the option of abstaining
from making a prediction. For example, in the classification
setting, at each round the learner can choose to make a prediction
and incur a standard zero-one misclassification cost, or elect to
abstain, in which case she incurs an abstention cost, typically
less than one. Abstention can thus represent an attractive option to
avoid a higher cost of misclassification. Note, however, when the
learner abstains, she does not receive the true label (correct class),
which results in loss of information.
\vspace{-0.01in}
This scenario of online learning with abstention is relevant to many
real-life problems. As an example, consider the scenario where a
doctor can choose to make a diagnosis based on the current information
available about a patient, or abstain and request further laboratory
tests, which can represent both a time delay and a financial cost. In this case, the
abstention cost is usually substantially lower than that of a wrong
diagnosis. The online model is appropriate since it captures the
gradual experience a doctor gains by testing, examining and following new
patients.
Another instance of this problem appears in the design of
spoken-dialog applications such as those in modern personal
assistants. Each time the user asks a question, the assistant can
either provide a direct response to the question and risk providing an
inaccurate response, or choose to say ''I am sorry, I do not
understand?'', which results in the abstention cost of a longer dialog
requesting the user to reformulate his question. Similar online
learning problems arise in the context of self-driving cars where, at
each instant, the assistant must determine whether to continue
steering the car or return the control to the driver. Online learning
with abstention also naturally models many problems arising in
electronic commerce platforms such as Ad Exchange, an online
system set up by a publisher where several advertisers bid in order to
compete for an ad slot, the abstention cost being the opportunity
loss of not bidding for a specific ad slot.
\vspace{-0.01in}
In the batch setting, the problem of learning with abstention has been
studied by a number of publications, starting with
\cite{Chow1957,Chow1970}. Its theoretical aspects have been analyzed
by several authors in the last decade:
\citet{YanivWiener2010,YanivWiener2011} studied the trade-off between
the coverage and accuracy of classifiers; \citet{BartlettWegkamp2008}
introduced a loss function including explicitly an abstention cost and
gave a consistency analysis of a surrogate loss that they used to
derive an algorithm; and, more recently,
\citet{CortesDeSalvoMohri2016a,CortesDeSalvoMohri2016b} presented a
comprehensive study of the problem, including an analysis of the
properties of a corresponding abstention (or rejection) loss with a
series of theoretical guarantees and algorithmic results both for
learning with kernel-based hypotheses and for boosting.
\vspace{-0.01in} This paper presents an extensive study of the problem
of online learning with abstentions, in both the adversarial and the
stochastic settings. We consider the common scenario of prediction
with expert advice \citep{LittlestoneWarmuth1994} and adopt the same
general abstention loss function as in
\citep{CortesDeSalvoMohri2016a}, with each expert formed by a pair
made of a predictor and an abstention function.
\vspace{-0.01in}
A key aspect of the problem we investigate, which makes it distinct from
both batch learning with abstentions, where labels are known for
all training points, and standard online learning with full
information, is the following: if the algorithm abstains from making
a prediction for the input point received at a given round, the
true label of that point is not revealed. As a result, the
loss of the experts that would have instead made a prediction on that point
cannot be determined at that round. Thus, we are dealing with an online
learning scenario with partial feedback:
if the algorithm chooses to predict, then the true label is revealed and the
losses of all experts, including abstaining ones, are known. But,
if the algorithm elects to abstain, then only the losses of the
abstaining experts are known, all of them being equal to the same
abstention cost.
\vspace{-0.01in}
\ignore{
To devise algorithms for online learning with abstentions, we draw
from connections with several related online learning problems.
}
As we shall see, our learning problem can be cast as a specific instance of
online learning with a feedback graph, a framework introduced by
\citet{MannorShamir2011} and later extensively analyzed by several
authors
\citep{CaronKvetonLelargeBhagat2012,AlonCesaBianchiGentileMansour2013,AlonCesaGentileMannorMansourShamir2014,AlonCesaBianchiDekelKoren2015,KocakNeuValkoMunos2014,Neu2015,CohenHazanKoren2016}). In our context, the feedback graph varies over time, a scenario to which
most of the existing algorithms and analyses (specifically, in the stochastic
setting) do not readily apply. Our setting is distinct from the KWIK ('knows what it knows')
framework of \citet{LiLittmanWalsh2008} and its later extensions, though
there are some connections, as discussed in Appendix~\ref{sec:relatedworks}.
\vspace{-0.01in}
Our contribution is summarized as follows.
In Section~\ref{sec:adversarial}, we analyze an adversarial setting both in
the case of a finite family of experts and in the infinite case.
We show that the problem can be cast as that of online learning
with a time-varying feedback graph tailored to the problem of learning
with abstentions.
In the finite case, we show how ideas from \citet{AlonCesaGentileMannorMansourShamir2014,AlonCesaBianchiDekelKoren2015}
can be extended and combined with this time-varying feedback graph to devise an algorithm, \textsc{EXP3-ABS},
that benefits from favorable guarantees. In turn, \textsc{EXP3-ABS}\ is used as a subroutine for the infinite case where
we show how a surrogate loss function can be carefully constructed for the abstention loss, while maintaining
its (bandit) observability. We use the structure of this loss function to extend \textsc{ContextualEXP3}\
\citep{CesaBianchiGaillardGentileCerchinovitz2017} to the abstention scenario and prove regret guarantees
for its performance.
\vspace{-0.01in}
However, we will be mainly interested in the stochastic setting
(Section~\ref{sec:stochastic}). Stochastic bandits with a fixed
feedback graph have been previously studied by
\citet{CaronKvetonLelargeBhagat2012} and
\citet{CohenHazanKoren2016}. We first show that an immediate extension of
these algorithms to the time-varying graphs in the abstention scenario
faces a technical \emph{bias problem} in the estimation of the expert
losses. Next, we characterize a set of feedback graphs that can
circumvent this bias problem in the general setting of online learning
with feedback graphs. We further design a new online algorithm, \textsc{UCB-GT},
whose feedback graph is \emph{estimated} based on past
observations. We prove that the algorithm admits more favorable regret
guarantees than the \textsc{UCB-N}\ algorithm
\citep{CaronKvetonLelargeBhagat2012}. Finally, in Section~\ref{sec:experiments}
we report the results of several experiments with both artificial and real-world
datasets demonstrating that \textsc{UCB-GT}\ in practice significantly outperforms an
unbiased, but limited, extension of \textsc{UCB-N}, as well as a standard bandit baseline,
like UCB \cite{acf02}.
\vspace{-0.1in}
\section{Learning Problem}
\label{sec:problem}
\vspace{-0.05in}
Let $\mathscr X$ denote the input space (e.g., $\mathscr X$
is a bounded subset of $\mathbb{R}^d$). We denote by $\mathscr H$ a family of
predictors $h \colon \mathscr X \to \mathbb{R}$, and consider the familiar binary
classification problem where the loss $\ell(y,h(x))$ of $h \in \mathscr H$ on
a labeled pair $(x, y) \in \mathscr X \times \set{\pm 1}$ is defined by
either the 0/1-loss $1_{y h(x) \leq 0}$, or some Lipschitz variant
thereof (see Section \ref{sec:adversarial}). In all cases, we assume
$\ell(\cdot,\cdot) \in [0, 1]$. We also denote by $\mathscr R$ a family of
abstention functions $r\colon \mathscr X \to \mathbb{R}$, with $r(x) \leq 0$
indicating an abstention on $x \in \mathscr X$ (or that $x$ is
\emph{rejected}), and $r(x) > 0$ that $x$ is predicted upon (or that
$x$ is \emph{accepted}).
\vspace{-0.01in}
We consider a specific online learning scenario whose regime lies
between bandit and full information, sometimes
referred to as {\em bandit with side-information} (e.g.,
\citet{MannorShamir2011,CaronKvetonLelargeBhagat2012,AlonCesaBianchiGentileMansour2013,
AlonCesaGentileMannorMansourShamir2014,AlonCesaBianchiDekelKoren2015,KocakNeuValkoMunos2014,Neu2015,CohenHazanKoren2016}).
In our case, the arms are pairs made of a predictor function $h$ and
an abstention function $r$ in a given family
$\mathscr E \subseteq \mathscr H \times\mathscr R$. We will denote by
$\xi_j = (h_j, r_j)$, $j \in [K]$, the elements of $\mathscr E$. In fact,
depending on the setting, $K$ may be finite or (uncountably)
infinite.
Given $h_j$, one natural choice for the associated abstention function
$r_j$ is a confidence-based abstention function of the form
$r_j(x) = |h_j(x)| - \theta$, for some threshold $\theta > 0$. Yet,
more general pairs $(h_j, r_j)$ can be considered here. This
provides an important degree of flexibility in the design of
algorithms where abstentions are allowed, as shown in
\cite{CortesDeSalvoMohri2016a,CortesDeSalvoMohri2016b}.
\vspace{-0.01in}
The online learning protocol is described as follows. The set $\mathscr E$ is
known to the learning algorithm beforehand. At each round $t \in
[T]$, the online algorithm receives an input $x_t \in \mathscr X$ and chooses
(possibly at random) an arm (henceforth also called ``expert" or
``pair") $\xi_{I_t} = (h_{I_t}, r_{I_t}) \in \mathscr E$. If the
inequality $r_{I_t}(x_t) \leq 0$ holds, then the algorithm abstains
and incurs as loss an abstention cost $c(x_t) \in [0,1]$. Otherwise,
it predicts based on the sign of $h_{I_t}(x_t)$, receives the
true label $y_t \in \set{\pm 1}$, and incurs the loss $\ell(y_t,
h_{I_t}(x_t))$. Thus, the overall \emph{abstention loss} $L$ of expert
$\xi = (h, r) \in \mathscr E$ on labeled pair $z = (x, y) \in \mathscr X \times
\set{\pm 1}$ is defined as follows:
\begin{equation}\label{e:absloss}
L( \xi, z) = \ell(y,h(x)) 1_{r(x) > 0} + c(x) 1_{r(x) \leq 0}~.
\end{equation}
For simplicity, we will assume throughout that the abstention cost
$c(x)$ is a (known) constant $c \in [0,1]$, independent of $x$, though
all our results can be easily extended to the case when $c$ is a
(Lipschitz) function of $x$, which is indeed desirable in some
applications.
\vspace{-0.01in}
Our problem can be naturally cast as an online learning problem with
side information in the form of a feedback graph. Online learning
with a feedback graph is a general framework that covers a variety of
problems with partial information, including the full information
scenario, where the graph is fully connected, or the bandit scenario
where all vertices admit only self-loops and are all disconnected
\citep{AlonCesaBianchiGentileMansour2013,AlonCesaGentileMannorMansourShamir2014}.
In our case, we have a directed graph $G^{\text{\sc abs}}_t = (V, E_t)$ that
depends on the instance $x_t$ received by the algorithm at round
$t \in [T]$. Here, $V$ denotes the finite set of vertices of this
graph which, in the case of a finite set of arms, coincides with the
set of experts $\mathscr E$, while $E_t$ denotes the set of directed edges at
round $t$. The directed edge $\xi_i \rightarrow \xi_j$ is in $E_t$
if the loss of expert $\xi_j \in V$ is observed
when expert $\xi_i$ is played by the algorithm at round $t$. In our
problem, if the learner chooses to predict at round $t$ (i.e., if
$r_{I_t}(x_t) > 0$) then she observes the loss $L(\xi_j,z_t)$ of all
experts $\xi_j$, since the label $y_t$ is revealed to her. If instead
she abstains at round $t$ (i.e., if $r_{I_t}(x_t) \leq 0$), then she only
observes $L(\xi_j,z_t)$ for those experts $\xi_j$ that are abstaining
in that round, that is, the set of $j$ such $r_j(x_t) \leq 0$, since
for all such $\xi_j$ we have $L(\xi_j,z_t) = c$. Notice that in both
cases the learner can observe the loss of her own action. Thus,
the feedback graph we are operating with is a nearly fully connected
directed graph with self-loops, except that it admits only one-way
edges from predicting to abstaining vertices (see
Figure~\ref{fig:feedback} for an example). Also observe that
the feedback graph $G^{\text{\sc abs}}_t$ is fully determined by $x_t$.
\begin{figure}[t]
\vspace{-0.1in}
\centering
\includegraphics[scale=0.28]{fig}
\vspace{-0.1in}
\caption{Feedback graph $G^{\text{\sc abs}}_t$ for the scenario of online
learning with abstentions, with $K = 5$.}
\label{fig:feedback}
\vskip -0.2in
\end{figure}
\vspace{-0.01in}
We will consider both an adversarial setting (Section
\ref{sec:adversarial}), where no distributional assumptions are made
about the sequence $z_t = (x_t, y_t)$, $t \in [T]$, and a stochastic
setting (Section \ref{sec:stochastic}), where $z_t$ is assumed to be
drawn i.i.d.\ from some unknown distribution $\mathscr D$ over
$\mathscr X\times\{\pm 1\}$. For both settings, we measure the performance of
an algorithm $\mathscr A$ by its (\emph{pseudo}-)\emph{regret} $R_T(\mathscr A)$, defined as
\vspace{-0.1in}
\begin{equation*}
R_T(\mathscr A) = \sup_{\xi \in \mathscr E} \E \Big[ \sum_{t = 1}^T L( \xi_{I_t}, z_t)
- \sum_{t = 1}^T L(\xi, z_t) \Big]~,
\end{equation*}
\iffalse
*********************************
\begin{equation*}
R_T(\mathscr A) = \sup_{\xi \in \mathscr E}\left(\sum_{t = 1}^T L( \xi_{I_t}, z_t) - \sum_{t = 1}^T L(\xi, z_t)\right).
\end{equation*}
Two other standard benchmarks for measuring the performance
of an algorithm $\mathscr A$ are
its \emph{expected regret} $\E[R_T(\mathscr A)]$ and its
\emph{pseudo-regret} $\overline R_T(\mathscr A)$:
\begin{align*}
\E[R_T(\mathscr A)]
& = \Bigg[\max_{\xi \in \mathscr E} \sum_{t = 1}^T L( \xi_{I_t}, z_t) -
\sum_{t = 1}^T L(\xi, z_t) \Bigg]\\
\overline R_T(\mathscr A)
& = \max_{\xi \in \mathscr E} \E \Bigg[ \sum_{t = 1}^T L( \xi_{I_t}, z_t)
- \sum_{t = 1}^T L(\xi, z_t) \Bigg],
\end{align*}
**********************************
\fi
where the expectation is taken both with respect to the algorithm's
choice of actions $I_t$s and, in the stochastic setting, the random draw of the $z_t$s.
In the stochastic setting, we will mainly be concerned with the case
when $\mathscr E$ is a finite set of experts
$\mathscr E = \set{\xi_1, \ldots, \xi_K}$. We then denote by $\mu_j$ the
expected loss of expert $\xi_j \in \mathscr E$,
$\mu_j = \E_{z \sim \mathscr D}[L(\xi_j, z)]$, by $\mu^*$ the expected loss
of the best expert, $\mu^* = \min_{j \in [K]} \mu_j$, and by
$\Delta_j$ the loss gap to the best, $\Delta_j = \mu_j - \mu^*$. In
the adversarial setting, we will analyze both the finite and infinite
expert scenarios. In the infinite case, since $L$ is non-convex in the
relevant parameters (Eq.~\eqref{e:absloss}), further care is needed.
\vspace{-0.1in}
\section{Adversarial setting}
\label{sec:adversarial}
\vspace{-0.05in}
As a warm-up, we start with the adversarial setting with finitely-many
experts. Following ideas from
\cite{AlonCesaGentileMannorMansourShamir2014,AlonCesaBianchiDekelKoren2015},
we design an online algorithm for the abstention scenario by combining
standard finite-arm bandit algorithms, like \textsc{EXP3}\ \cite{Exp3}, with
the feedback graph described in Section \ref{sec:problem}. We call the
resulting algorithm \textsc{EXP3-ABS}\ (\textsc{EXP3}\ with ABStention). The algorithm is
a variant of \textsc{EXP3}\ where the importance weighting scheme to achieve
unbiasedness of loss estimates is based on the probability of the loss
of an expert being observed, as opposed to the expert being
played. The full pseudocode of \textsc{EXP3-ABS}\ is given in
Appendix~\ref{app:adversarial} (Algorithm~\ref{alg:exp3rej}). We have
the following result.\footnote { Due to space limitations, all proofs
are given in the supplementary material. }
\vspace{-0.05in}
\begin{theorem}
\label{th:exp3rej}
Let \textsc{EXP3-ABS}\ be run with learning rate $\eta$ over a set of $K$
experts $\xi_1,\ldots,\xi_K$. Then, the algorithm admits the
following regret guarantee after $T$ rounds:
\begin{equation*}
R_T(\textsc{EXP3-ABS}) \leq (\log K)/\eta + \eta\,T (c^2 + 1)/2.
\end{equation*}
In particular, if \textsc{EXP3-ABS}\ is run with
$\eta = \sqrt{\frac{2\log K}{(c^2 + 1)T}}$, then
\(
R_T(\textsc{EXP3-ABS}) \leq \sqrt{2(c^2+1)T\log K}.
\)
\end{theorem}
\vspace{-0.05in}
This bound is clearly more favorable than the standard bound for \textsc{EXP3}\
in terms of its dependency on the number of experts ($\sqrt{\log K}$ instead of $\sqrt{K}$). Theorem
\ref{th:exp3rej} is in fact reminiscent of what one can achieve using
the contextual-bandit algorithm
EXP4~\cite{AuerCesaBianchiFreundSchapire2002} run on $K$ experts, each
one having two actions.
\iffalse
**********************
This bound is significantly more favorable than the standard one for
\textsc{EXP3}\ ($O\big (\sqrt{KT\log(K)}\big)$), which shows the benefit the
feedback graph in this context. Theorem~5 in
\citep{AlonCesaGentileMannorMansourShamir2014} provides tight lower
bounds for the general setting of online learning with fixed feedback
graphs. Since one can always design a specific case where the
time-varying feedback graphs are fixed in the abstention scenario, the
result shows that the regret bounds just presented are also tight.
**********************
\fi
\vspace{-0.01in} We now turn our attention to the case of an
uncountably infinite $\mathscr E$. To model this more general framework, one
might be tempted to focus on parametric classes of functions $h$ and
$r$, e.g., the family $\mathscr E$ of linear functions
\[
\set[\big]{(h, r) : h(x) = w^\top x,\,\,r(x) = |w^\top x| - \theta,\,w \in \mathbb{R}^d, \theta > 0},
\]
then introduce some convex surrogate of the abstention loss
(\ref{e:absloss}), and work in the parametric space of $(w, \theta)$
through some Bandit Convex Optimization technique (e.g.,
\cite{hazan16}). Unfortunately, this approach is not easy to put in
place, since the surrogate loss not only needs to ensure convexity and
some form of calibration, but also the ability for the algorithm to
observe the loss of its own action (the self-loops in the graph of
Figure \ref{fig:feedback}).
\iffalse
***********************************
which should not be taken for granted. To see why, assume $h(x), r(x) \in [-1,1]$
for all $x \in \mathscr X$, $\ell(y,h(x))$ is the hinge loss $\ell(y,h(x)) = \max\{\gamma-yh(x),0\}$,
for some $\gamma \in (0,1)$ and, for given $(x_t,y_t)$, let us rewrite (\ref{e:absloss}) as a function of the two arguments
$a = y_t h(x_t)$, and $r = r(x_t)$:
\[
L(a,r) = \max\{\gamma-a,0\} 1_{r > 0} +c 1_{r\leq 0}~.
\]
Consider, for instance, the family of convex surrogate losses studied in \cite{CortesDeSalvoMohri2016a} for the problem of
learning with abstentions.
Despite being convex (specifically, piecewise linear) upper bounds on $L(a,r)$, a feedback graph algorithm operating
with these losses would not grant the observability of the loss values when the algorithm decides to
abstain ($r \leq 0$), for the value of $L(a,r)$ when $r \leq 0$ would also depend on $a = y_t h(x_t)$, which
is indeed not available when the algorithm abstains. As a result, the online problem would not even be a bandit problem, but
a more general partial monitoring problem, with unclear regret guarantees.
***********************************
\fi
\vspace{-0.01in}
We have been unable to get around this problem by just resorting to
convex surrogate losses (and we strongly suspect that it is not
possible), and in what follows we instead introduce a surrogate
abstention loss which is Lipschitz but not convex. Moreover, we take
the more general viewpoint of competing with pairs $(h,r)$ of
Lipschitz functions with bounded Lipschitz constant. Let us then
consider the version of the abstention loss (\ref{e:absloss}) with
$\ell(y, h(x)) = f_{\gamma}(yh(x))$, where $f_{\gamma}$ is the 0/1-loss
with slope $1/(2\gamma)$ at the origin,\,
\( f_{\gamma}(a) = \left(\frac{\gamma+a}{2\gamma}\right) 1_{|a| \leq
\gamma} + 1_{a\leq 0} 1_{|a| > \gamma}\, \)
\iffalse
*********
In this section, instead of considering the standard abstention loss
$L(\xi,z_t)$ as defined in Section~\ref{sec:problem}, we analyze a
surrogate abstention loss $\tilde{L}(\xi,z_t)$, which simply upper
bounds the misclassification loss, $1_{y_t h(x_t) \leq 0}$, of the
abstention loss by any Lipschitz loss $\ell$: $\tilde{L}_t(\xi,z_t) =
c\, 1_{r(x_t) < 0}+ \ell(y_t h(x_t))\,1_{r(x_t) \geq 0}.$
\ignore{Without loss of generality, we assume that $\ell$ has
Lipschitz constant $1$. }
*********
\fi
(see Figure \ref{f:1} (a)), and the class of experts ${\mathscr E} =
\set[\big]{\xi = (h, r)\, |\, h, r\colon \mathscr X \subseteq \mathbb{R}^d \to
[-1, 1]}$. Here, the functions $h$ and $r$ in $\mathscr E$ are Lipschitz
with constant $L_{\mathscr E}$ with respect to an appropriate
distance on $\mathbb{R}^d$.
\ignore{
, e.g., the standard Euclidean norm $||\cdot||_2$, $|h(x)-h(x')| \leq C_{\mathscr F}||x-x'||_2$ for all $x,x' \in
\mathbb{R}^d$, and similarly for the functions $r$.
}
Notice that $L_{\mathscr E}$ determines how big this comparison class is. For brevity, we let $\mathscr Y$ be a shorthand for $[-1, 1]^2$.
\begin{figure}[t]
\centering
\vspace{-0.08in}
\includegraphics[scale=0.35]{liploss}
\vspace{-3.0in}
\caption{(a) The 0/1-loss function with slope $1/(2 \gamma)$ at the origin.
(b) For a given value of $x$ and margin $a = yh(x)$ (which in
turn sets the value of $f = f_{\gamma}(a) \in [0, 1]$), plots of the
abstention loss function $L(a, r)$ (dotted blue curve), and the surrogate
abstention loss $\tilde{L}(a, r)$ (red curve), both as a
function of $r = r(x) \in [-1, 1]$.}
\label{f:1}
\vspace{-0.21in}
\end{figure}
\vspace{-0.03in} Using ideas from
\cite{CesaBianchiGaillardGentileCerchinovitz2017}, we present an
algorithm that approximates the action space by a finite
cover while using the structure of the abstention setting. The crux of
the problem is to define a Lipschitz function ${\tilde{L}}$ that uppers bounds
the abstention loss while maintaining the same feedback assumptions,
namely the feedback graph given in Figure~\ref{fig:feedback}. One
Lipschitz function ${\tilde{L}}\colon \mathscr Y \rightarrow [0, 1]$ that precisely
solves this problem is the following:
\vspace{-0.1in}
\begin{center}
\(
{\tilde{L}}(\xi,z) =
\begin{cases}
c &{\mbox{if $r(x) \leq -\gamma$}}\\
1+\left(\frac{1-c}{\gamma}\right)r(x) &{\mbox{if $r(x) \in (-\gamma,0)$}}\\
1-\left(\frac{1-f_{\gamma}(y h(x))}{\gamma}\right)r(x) &{\mbox{if $r(x) \in [0,\gamma)$}}\\
f_{\gamma}(y h(x)) &{\mbox{if $r(x) \geq \gamma$}}~,
\end{cases}
\)
\end{center}
\vspace{-0.08in}
for $\gamma \in (0,1)$. ${\tilde{L}}(\xi,z)$ is plotted in Figure \ref{f:1}(b).
Notice that this function is consistent with the feedback requirements of Section
\ref{sec:problem}: $r_{I_t}(x_t) \leq 0$ implies that
${\tilde{L}}((h(x_t),r(x_t)),z_t)$ is known to the algorithm (i.e., is
independent of $y_t$) for all $(h,r) \in \mathscr E$ such that
$r(x_t) \leq 0$, while $r_{I_t}(x_t) > 0$ gives complete knowledge of
${\tilde{L}}((h(x_t),r(x_t)),z_t)$ for all $(h,r) \in \mathscr E$, since $y_t$ is
observed.
\vspace{-0.03in}
We can then adapt the machinery from \cite{CesaBianchiGaillardGentileCerchinovitz2017} so as to apply a contextual version of \textsc{EXP3-ABS}\ to
the sequence of losses ${\tilde{L}}(\xi,z_t), t \in [T]$.
The algorithm adaptively covers the input space $\mathscr X$ with a set of balls of a fixed radius $\varepsilon$, each ball hosting
an instance of \textsc{EXP3-ABS}. At each round $t$, if a new incoming input $x_t \in \mathcal{X}$ is not contained in any existing ball,
then a new ball centered at $x_t$ is created, and a new instance of \textsc{EXP3-ABS}\ is allocated to handle $x_t$. Otherwise,
the \textsc{EXP3-ABS}\ instance associated with the closest input so far is used. Each allocated \textsc{EXP3-ABS}\ instance operates on a
discretized action space $\mathscr Y_{\varepsilon} \subseteq \mathscr Y$.
We call this algorithm \textsc{Cont\EXPABS}\ -- see Appendix \ref{app:adversarialinfinite} for a detailed description.
\vspace{-0.05in}
\begin{theorem}
\label{th:infarms}
Consider the abstention loss
\vspace{-0.35in}
\begin{center}
\begin{equation*}
L(\xi,z) = f_{\gamma}(yh(x)) 1_{r(x) > 0} + c 1_{r(x) \leq 0}~,
\end{equation*}
\end{center}
\vspace{-0.18in}
and let
$\xi^* = (h^*, r^*) = \argmin_{\xi \in \mathscr E} \sum_{t = 1}^T L(\xi, z_t)
$, with $\mathscr E = \set{(h, r)}$ made of pairs of Lipschitz functions as
described above. If \textsc{Cont\EXPABS}\ is run with parameter
$\varepsilon \simeq T^{-\frac{1}{2+d}}\,\gamma^{\frac{2}{2+d}}$ and an
appropriate learning rate (see Appendix~\ref{app:adversarial}), then,
it admits the following regret guarantee:
\vspace{-0.05in}
\[
R_T(\textsc{Cont\EXPABS})
\leq
{\tilde \mathscr O}\left(T^{\frac{d+1}{d+2}}\,\gamma^{-\frac{d}{d+2}}\right) + M^*_{T}(\gamma),
\]
where $M^*_T(\gamma)$ is the number of $x_t$ such that $|r^*(x_t)| \leq \gamma$.
\end{theorem}
\vspace{-0.07in}
In the above, $ {\tilde \mathscr O}$ hides constant and
$\ln(T)$ factors, while $ \simeq $ disregards constants like $L_\mathscr E$, and various log factors. \textsc{Cont\EXPABS}\ is also computationally efficient,
thereby providing a compelling solution to the infinite armed case of
online learning with abstention.
\ignore{
Let's close with some quick comments:
\begin{itemize}
\item The alg is efficient (notice that the number of balls and the
the size of the action space within each ball decreases as we
increase $d$).
\item The alg gets suboptimal bounds, a chaining on the function space
would probably get regret $T^{\frac{d}{d+1}}$, instead of
$T^{\frac{d+1}{d+2}}$ but this comes at the cost of an inefficient
algorithm (exponential time or more).
\item There is room for improvements:
\begin{itemize}
\item It might be possible to do transfer learning within each ball
$B$ \textsc{ContextualEXP3} operates on, and leverage the specific structure
of the loss function to get faster rates through an efficient
algorithm. Currently none of the two is done here.
\item The above algorithm needs to know $\gamma$ beforehand, it would
be nice to have in (\ref{e:finalbound}) a min over $\gamma > 0$
instead.
\end{itemize}
\end{itemize}
}
\vspace{-0.1in}
\section{Stochastic setting}\label{sec:stochastic}
\vspace{-0.05in}
We now turn to studying the stochastic setting.
As pointed out in Section~\ref{sec:problem}, the problem can be cast as an instance of
online learning with time-varying feedback graphs $G^{\text{\sc abs}}_t$. Thus, a
natural method for tackling the problem
would be to extend existing algorithms designed for the stochastic setting
with feedback graphs to our abstention scenario~\citep{CohenHazanKoren2016,CaronKvetonLelargeBhagat2012}.
Yet, we cannot benefit from the algorithm of \citet{CohenHazanKoren2016} in
our scenario, even though it is designed for a stochastic setting with
time-varying feedback graphs. This is because at the heart of its
design and theoretical guarantees lies the assumption that the graphs
and losses are \emph{independent}. The dependency of the feedback
graphs on the observations $z_t$, which also define the losses, is
precisely a property that we wish to exploit in our scenario.
An alternative is to extend the \textsc{UCB-N}\ algorithm of
\citet{CaronKvetonLelargeBhagat2012}, for which the authors provide
gap-based regret guarantees. This algorithm is defined for a
stochastic setting with an undirected feedback graph over the arms with
the graph being fixed over time. The algorithm can be straightforwardly
extended to the case of directed time-varying feedback graphs (see
Algorithm~\ref{alg:ucbncaron}). We will denote that extension by
\textsc{UCB-NT}\ to explicitly differentiate it from \textsc{UCB-N}. Let $N_t(j)$ denote
the set of out-neighbors of vertex $\xi_j$ in the directed graph at time
$t$, i.e., the set of vertices $\xi_k$ destinations of an edge from
$\xi_j$. Then, as with \textsc{UCB-N}, the algorithm updates, at each round
$t$, the upper-confidence bound of every expert for which a feedback
is received (those in $N_t(I_t)$), as opposed to updating only the
upper-confidence bound of the expert selected, as in the standard
\textsc{UCB}\ algorithm~(e.g., \cite{acf02}).
\vspace{-0.03in}
In the context of learning with abstentions, the natural feedback
graph $G^{\text{\sc abs}}_t$ at time $t$ depends on the observation $x_t$ and
varies over time. Can we extend the regret guarantees of
\citet{CaronKvetonLelargeBhagat2012} to \textsc{UCB-NT}\ with such graphs?
We will show in Section~\ref{subsec:bias} that
vanishing regret guarantees do not hold for \textsc{UCB-NT}\ run with
graphs $G^{\text{\sc abs}}_t$. This is because of a fundamental
estimation bias problem that arises when the graph at time $t$
depends on the observation $x_t$.
This issue affects more generally any natural method using
the $G^{\text{\sc abs}}_t$ graphs. Nevertheless, we will show in Section~\ref{subsec:ucbn} that \textsc{UCB-NT}\
does benefit from favorable
guarantees,
provided the feedback graph $G^{\text{\sc abs}}_t$ it uses at round $t$ is replaced by one that
only depends on events up to time $t - 1$.
\setlength{\textfloatsep}{5pt}
\begin{algorithm}[t]
\begin{algorithmic}
\FOR{$t \geq 1$}
\STATE \textsc{Receive}($x_t$);
\STATE $\xi_{I_t} \leftarrow \argmin_{\xi_j \in \mathscr E} \Big \{ \widehat{\mu}_{j,t-1} - S_{j,t-1} \Big \}$;
\STATE $Q_{j,t} \leftarrow \sum_{s=1}^{t} 1_{j\in N_s(I_s)}~$;
\STATE $S_{j,t} \leftarrow \sqrt{\frac{5\,\log t}{Q_{j,t}} }$;
\STATE ${\hat \mu}_{j,t} \leftarrow \frac{1}{Q_{j,t}}\sum_{s=1}^{t} L(\xi_j,z_s) 1_{j\in N_s(I_s)}$.
\ENDFOR
\end{algorithmic}
\caption{\textsc{UCB-NT}}
\label{alg:ucbncaron}
\end{algorithm}
\vspace{-0.1in}
\subsection{Bias problem}
\label{subsec:bias}
\vspace{-0.05in}
\iffalse In this section, we show that an estimation bias problem
arises when the feedback graph $G_t$ used at round $t$ depends on
$x_t$. This is a key issue since it limits the applicability of the
\textsc{UCB-NT}\ algorithm to our scenario. This bias issue causes a poor
estimation of the expert losses, thereby resulting in suboptimal
choices of the expert at each round and ultimately in non-vanishing
regret. The problem affects, more generally, any natural method using
the $G^{\text{\sc abs}}_t$ graphs.
By the way \textsc{UCB-NT}\ shares information across actions, the loss
observations at round $t$ depend on the expert $\xi_{I_t}$ selected at
that round as well as on $G^{\text{\sc abs}}_t$ through the out-neighbors
$N_t(I_t)$. Thus, depending on the choice of the expert made by the
algorithm, the feedback graph creates a filter on the loss
observations of the experts. One would expect these loss observations
to be i.i.d., since $G^{\text{\sc abs}}_t$ is a function of the newly drawn $x_t$
only, and the choice of arms depends on empirical estimates computed
from previous times. Yet, this mechanism can bias the empirical
estimates away from the true mean. The following example illustrates
this phenomenon in a simple case.
\fi
\vspace{-0.03in}
Assume there are two experts: $\xi_1$ (red) and $\xi_2$ (blue) with
$\mu_2 < \mu_1$ and $\mathscr X = [0, 1]$ (see
Figure~\ref{fig:biasfeedback}). For $x > \frac{1}{2}$, the red expert
$\xi_1$ is abstaining and incurring a loss $c$, whereas the blue
expert is never abstaining. Assume that the probability mass is
quasi-uniform over the interval $[0, 1]$ but with slightly more mass
over the region $x < \frac{1}{2}$. The algorithm may then start out by
observing points in this region. Here, both experts accept and the
algorithm obtains error estimates corresponding to the solid red and
blue lines for $x < \frac{1}{2}$. When the algorithm observes a point
$x > \frac{1}{2}$, it naturally selects the red abstaining expert since
it admits a better current estimated loss. However, for
$x > \frac{1}{2}$, the red expert is worse than the blue expert
$\xi_2$. Furthermore, it is abstaining and thus providing no updates
for expert $\xi_2$ (which is instead predicting). Hence, the algorithm
continues to maintain an estimate of $\xi_2$'s loss at the level of
the blue solid line indicated for $x < \frac{1}{2}$; it then continues to
select the red expert for all $x$s and incurs a high regret.\footnote
{
For the
sake of clarity, we did not introduce specific real values for the
expected loss of each expert on each of the half intervals, but that
can be done straightforwardly. We have also verified experimentally
with such values that the bias problem just pointed out indeed
leads to poor regret for \textsc{UCB-NT}.
}
\vspace{-0.03in}
This simple example shows that, unlike the adversarial scenario
(Section \ref{sec:adversarial}), $G^{\text{\sc abs}}_t$, here, cannot
depend on the input $x_t$, and that, in general, the indiscriminate
use of feedback graphs may result in biased loss observations. On
the other hand, we know that if we were to avoid using feedback graphs
at all (which is always possible through \textsc{UCB}), we will always be able to construct
unbiased loss estimates. A natural question is then: can we construct
time-varying feedback graphs that lead to unbiased loss observations?
In the next section, we show how to design such a sequence of
auxiliary feedback graphs, which in turn allows us to then extend
\textsc{UCB-NT}\ to the setting of time-varying feedback graphs for general
loss functions. Under this assumption, we can achieve unbiased
empirical estimates of the average losses $\mu_j$ of the experts,
which will allow us to apply standard concentration bounds in the
proof of this algorithm.
\begin{figure}[t]
\centering
\includegraphics[scale=0.22]{reject}
\vspace{-0.1in}
\caption{Illustration of the bias problem.}
\label{fig:biasfeedback}
\end{figure}
\vspace{-0.1in}
\subsection{Time-varying graphs for \textsc{UCB-NT}}
\label{subsec:ucbn}
\vspace{-0.05in}
We now show that \textsc{UCB-NT}\ benefits from favorable guarantees, so long
as the feedback graph $G^{\text{\sc abs}}_t$ it uses at time $t$ depends only on
events up to time $t - 1$. This extension works in general for bounded
losses, and does not only apply to our specific abstention loss $L$.
\ignore{
A natural extension of \textsc{UCB-N}\ for this purpose consists of updating
the upper-confidence bounds of each expert according to the feedback
graph defined at round $t$. That is, t The \textsc{UCB-N}\ algorithm for
time-varying feedback graphs is similar to the standard \textsc{UCB-N}\ of
\citet{CaronKvetonLelargeBhagat2012} except that it updates its
estimate of the losses according to the time-varying
out-neighborhood $N_t(I_t)$ at each round $t$, which depends on the
expert $\xi_{I_t}$ chosen at round $t$.
Algorithm~\ref{alg:ucbncaron} shows the pseudocode where $Q_j(t)$ is
the number of times expert $j$ has been observed and $S_{j,t-1}$ is
a slack term. Please see Appendix~\ref{app:stochastic} for a formal
definition of these quantities.
}
\vspace{-0.02in}
So, let us assume that the feedback graph in round $t$ (and the associated
out-neighborhoods $N_t(\cdot)$) in Algorithm \ref{alg:ucbncaron} only
depends on the observed losses $L(\xi_i,z_s)$ and inputs $x_s$, for
$s = 1, \ldots, t - 1$, and $i \in [K]$, and let us denote this
feedback graph by $G_t$, so as not to get confused with $G^{\text{\sc abs}}_t$.
Under this assumption, we can derive strong regret guarantees for
\textsc{UCB-NT}\ with time-varying graphs. Our guarantees are expressed in
terms of the best sequence of \emph{admissible $p$-partitionings} of
graphs $G_t$. For $p \in [K]$, we say that
$(\mathscr C_{t, k})_{k \in [p]}$ is an admissible $p$-partitioning of
$G_t = (V,E_t)$ if
$V = \set{\xi_j\colon j \in [K]}$ is the union of $p$ (disjoint)
components $\mathscr C_{t, k}$, that is $V = \bigcup_{k \in [p]} \mathscr C_{t, k}$,
and all vertices within a component $\mathscr C_{t, k}$ are neighbors in
$G_t$ with reciprocal edges: if $\xi_i$ and $\xi_j$ are in
$\mathscr C_{t, k}$, then we have $\xi_i \in N_t(\xi_j)$ and
$\xi_j \in N_t(\xi_i)$. Note that admissible $p$-partitionings are
typically
not unique since two neighbor vertices $\xi_i$ and $\xi_j$
may be placed in the same component or not. We denote by $\mathscr S_p$ the
set of all sequences
$((\mathscr C_{1, k})_{k \in [p]}, \ldots, (\mathscr C_{T, k})_{k \in [p]})$ of
admissible $p$-partitionings $(\mathscr C_{t, k})_{k \in [p]}$ of graphs
$G_t$. Moreover, for any sequence
$((\mathscr C_{1, k})_{k \in [p]}, \ldots, (\mathscr C_{T, k})_{k \in [p]})$ in
$\mathscr S_p$, we denote by $\mathtt{C}_k$ the union of the components indexed by
$k$ over all rounds: $\mathtt{C}_k = \bigcup_{t \in [T]}\mathscr C_{t, k}$.
\vspace{-0.05in}
\begin{theorem}
\label{th:ucbn}
Assume that, for all $t \in[T]$, the feedback graph $G_t$ depends only
on information up to time $t - 1$. Then, the regret
of \textsc{UCB-NT}\
is bounded as follows:
\[
R_T(\textsc{UCB-NT}) = \mathscr O \Bigl(\min_{p, \mathscr S_p} \sum_{k \in [p]} \frac{\max_{j\in
\mathtt{C}_k} \Delta_j}{\min_{j \in \mathtt{C}_k}\Delta_j^2} \log(T) + K \Bigl )~.
\]
\end{theorem}
\vspace{-0.1in}
For a sequence $\mathscr S_p$ made up of the same partition
$(\mathscr C_{1, k})_{k \in [p]}$ repeated $T$ times, the theorem gives a
bound on the regret based on this fixed partition, as it is the sum of
$p$ components, one per cluster $\mathscr C_{1,k}$ in the partition. The
minimum over $(p,\mathscr S_p)$ then simply chooses the number of clusters
$p$ and the partitioning of $V$
into $p$ clusters having the smallest regret.
\vspace{-0.01in}
Theorem~\ref{th:ucbn} can be interpreted as an extension of Theorem 2
in \citet{CaronKvetonLelargeBhagat2012} to time-varying feedback
graphs. Its proof involves
showing that the use of feedback graphs $G_t$ that
depend only on information up to $t-1$ can result in unbiased
loss estimates, and it also uses the newly defined notion of admissible
$p$-partitionings to derive a time-varying bound that leverages the shared
updates from the graph.
Moreover, the bound illustrates that if the feedback graphs in a problem admit a
$p$-partitioning for some small $p \ll K$ (e.g. if the feedback graphs
can be decomposed into a small number of components that are
approximately fixed across time) for which
$\max_{j \in \mathtt{C}_k} \Delta_j \approx \min_{j \in \mathtt{C}_k} \Delta_j$,
then this bound can be up to a factor $\frac{p}{K}$ tighter than the
bound guaranteed by the standard UCB algorithm. Moreover, this regret
guarantee is always more favorable than that of the standard UCB since
the (trivial) $K$-partitioning that splits $V$ into $K$ singletons for
all $t$ is an admissible $K$-partitioning for all $G_t$'s.
Furthermore, note that by construction, all vertices within the same component of
an admissible $p$-partitioning are connected to one another. Thus, if
the feedback graph is fixed throughout all rounds, and one interprets
the doubly-directed edges as edges of an undirected graph $G_U$,
we easily
obtain a result that is comparable to Theorem 2 in
\citet{CaronKvetonLelargeBhagat2012}:
\vspace{-0.05in}
\begin{corollary}
If the feedback graph $G_t = G$ is fixed with time,
then the guarantee of Theorem~\ref{th:ucbn} is upper-bounded by:
\vspace{-0.1in}
\[
\mathscr O \Bigl( \min_{\mathcal{C}} \sum_{C \in \mathcal{C}} \frac{\max_{i \in C} \Delta_i }{\min_{
i \in C}\Delta_i^2} \log(T) + K \Bigr )~,
\]
the outer minimum being over all clique coverings $\mathcal{C}$ of $G_U$.
\end{corollary}
\vspace{-0.05in}
\citet{CaronKvetonLelargeBhagat2012} present matching lower bounds for
the case of stochastic bandits with a fixed feedback graph. Since we
can again design abstention scenarios with fixed feedback graphs,
these bounds carry over into our setting.
\vspace{-0.01in}
Now, how can we use the results of this section to design an algorithm
for the abstention scenario? The natural feedback graphs we discussed
in Section \ref{sec:adversarial} are no longer applicable since
$G^{\text{\sc abs}}_t$ depends on $x_t$. How can one define a useful feedback graph
$G_t$ at each round $t$?
We will present two solutions to this problem. In Section
\ref{subsec:subsetfeedback}, we present a solution with a fixed graph
$G$ that closely captures the problem of learning with abstentions. In
some sense, this graph contains the most informative feedback if we
rule out the use of $x_t$-dependent side information, which would lead
to the bias problem. Next, in Section~\ref{subsec:ucbgt}, we will show
how to define and leverage a time-varying graph $G_t$ that is
estimated based on past observations.
\vspace{-0.1in}
\subsection{\textsc{UCB-N}\ with the subset feedback graph}
\label{subsec:subsetfeedback}
\vspace{-0.05in}
In this section, we define a \emph{subset feedback graph}, $G^{\text{\sc sub}}$,
that captures the most informative feedback in the problem of learning
with abstentions and yet is safe in the sense that it does not depend
on $x_t$. The definition of the graph is based on the following simple
observation: if the abstention region associated with $\xi_i$ is a
subset of that of $\xi_j$, then, if $\xi_i$ is selected at some round
$t$ and is abstaining, so is $\xi_j$, for an example see $\xi_i$ and
$\xi_j$ in Figure~\ref{f:aaa}. Crucially, this implication
holds regardless of the particular input point $x_t$ received in the
region of abstention of $\xi_i$. Thus, the set of vertices of $G^{\text{\sc sub}}$
is $\mathscr E$, and $G^{\text{\sc sub}}$ admits an edge from $\xi_i$ to $\xi_j$, iff
$\set{x \in \mathscr X \colon r_i(x) \leq 0} \subseteq \set{x \in \mathscr X \colon r_j(x) \leq 0}$.
Since $G^{\text{\sc sub}}$ does not vary with time, it trivially verifies the
condition of the previous section. Thus, \textsc{UCB-NT}\ run with $G^{\text{\sc sub}}$
admits the regret guarantees of Theorem~\ref{th:ucbn}, where
the admissible $p$-partionnings are those of fixed graph $G^{\text{\sc sub}}$.
\vspace{-0.01in}
The example of Section~\ref{subsec:bias} illustrated a bias problem in
a special case where the feedback graphs $G_t$ were not subgraphs of
$G^{\text{\sc sub}}$. The example indicates that $\textsc{UCB-NT}$ can suffer linear regret.
The following result shows more generally that feedback graphs not
included in $G^{\text{\sc sub}}$ may result in catastrophic regret behavior.
\vspace{-0.1in}
\begin{proposition}
\label{prop:subset}
Assume that \textsc{UCB-NT}\ is run with feedback graphs $G_t$ that are not
subsets of $G^{\text{\sc sub}}$. Then, there exists a family of predictors $\mathscr H$, a
Lipschitz loss function $\ell$ in (\ref{e:absloss}), and a distribution $\mathscr D$ over $z_t$s for
which \textsc{UCB-NT}\ incurs linear regret with arbitrarily high probability.
\end{proposition}
\vspace{-0.1in}
In view of these results, no fixed
feedback graph can be more informative than $G^{\text{\sc sub}}$. We cannot use a
graph $G_t$ depending on $x_t$, but the results of
Section~\ref{subsec:ucbn} suggest that we could use a feedback graph
$G_t$ depending only on information up to time $t - 1$. How can we
leverage past observations (up to time $t - 1$) to derive a feedback
graph that would be more informative than the simple subset graph
$G^{\text{\sc sub}}$? The next section provides a new solution based on feedback
graphs estimated based on past observations and a new algorithm.
\vspace{-0.05in}
\subsection{UCB-GT Algorithm}
\label{subsec:ucbgt}
\vspace{-0.05in}
\begin{algorithm}[t]
\begin{algorithmic}
\FOR{$t \geq 1$}
\STATE \textsc{Receive}($x_t$);
\STATE $\xi_{I_t} \leftarrow \argmin_{\xi_i \in\mathcal{E}} \big \{ \widehat{\mu}_{i,t - 1} - S_{i,t - 1} \big \}$,\\
where $S_{i,t - 1}$ is as in Algorithm \ref{alg:ucbncaron};
\FOR {$\xi_i\in \mathcal{E}$}
\STATE {\bf if} $ \hat{p}^{t - 1}_{I_t, i} \leq \gamma_{i,t - 1} $ {\bf then} $Q_{i,t}\leftarrow Q_{i,t - 1}+1$; \\
\hspace{0.15in}{\bf if} $r_{I_t}(x_t) \leq 0 \wedge r_i(x_s) > 0$ {\bf then}\\
$\hspace{0.26in}\widehat{\mu}_{i,t} \leftarrow \left(1-\frac{1}{Q_{i,t}}\right) \widehat{\mu}_{i,t - 1}$;\qquad\qquad\qquad (*) \\
\hspace{0.15in}{\bf else} $\widehat{\mu}_{i,t} \leftarrow \frac{ L( \xi_i,z_t)}{Q_{i,t}}+ \left(1-\frac{1}{Q_{i,t}}\right) \widehat{\mu}_{i,t - 1}$;\\
{\bf else} $Q_{i,t}\leftarrow Q_{i,t - 1}$,\,\, $\widehat{\mu}_{i,t} \leftarrow \widehat{\mu}_{i,t - 1}$~.
\ENDFOR
\ENDFOR
\end{algorithmic}
\caption{\textsc{UCB-GT}}
\label{alg:ucbtype1}
\end{algorithm}
\vspace{-0.01in}
We seek graphs $G_t$ that admit $G^{\text{\sc sub}}$ as a subgraph. By
Proposition~\ref{prop:subset}, in general, augmenting $G^{\text{\sc sub}}$ with
more edges is not safe since it can lead to the estimation bias
problem pointed out in Section~\ref{subsec:bias}. Instead, we will
show how some edges can be added with high probability, based on past
observations. This leads to a new algorithm, \textsc{UCB-GT}\ (\textsc{UCB}\ with
estimated time-varying graph), whose pseudocode is given in
Algorithm~\ref{alg:ucbtype1}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.4]{graphs.pdf}
\vspace{-0.1in}
\caption{Three experts $\xi_i$, $\xi_j$, and $\xi_k$ on a
one-dimensional input space, along with their prediction and
abstention regions. The feedback graph $G^{\text{\sc abs}}_t$ is shown below
each region when $x_t$ falls in that region. Self-loops are omitted.
To avoid the bias problem affecting the
time-varying graphs $G^{\text{\sc abs}}_t$, one option is to adapt $G^{\text{\sc sub}}$. Yet, as
illustrated, $G^{\text{\sc sub}}$ is minimal and in this example admits only one
edge (excluding self-loops). In contrast, the time-varying graphs
defined and used by \textsc{UCB-GT}\ are richer and more informative (with high
probability). An edge is added from say expert $\xi_j$ to expert $\xi_i$ if
the probability mass of the region where $\xi_j$ is abstaining but $\xi_i$
is accepting is (estimated to be) small.}
\label{f:aaa}
\end{figure}
\vspace{-0.05in}
As illustrated by Figure~\ref{f:aaa}, the key idea of \textsc{UCB-GT}\ is to
augment $G^{\text{\sc sub}}$ with edges from $\xi_j$ to $\xi_i$ where the subset
property $\set{x \colon r_j(x) \leq 0} \subseteq \set{x \colon r_i(x) \leq 0}$ may not hold, but
where the implication $(r_j(x) \leq 0 \Rightarrow r_i(x) \leq 0)$
holds with high probability over the choice of $x \in \mathscr X$, that is, the
region $\set{x \colon r_j(x) \leq 0 \wedge r_i(x) > 0}$ admits low
probability.
Of course, adding such an edge $\xi_j \rightarrow \xi_i$ can cause the
estimation bias of Section~\ref{subsec:bias}. But, if we
restrict ourselves to cases where $p_{j,i} = \Pr[r_j(x) \leq 0 \wedge r_i(x) >
0]$ is upper bounded by some carefully chosen quantity that changes
over rounds, the effect of this bias
will be limited.
\ignore{
Now, if $\Pr[r_i(x) \leq 0 \wedge r_j(x) > 0] > 0$ and we add edge $\xi_i \rightarrow \xi_j$,
then the empirical estimates of the experts may be affected by
the bias problem described in Section~\ref{subsec:bias}.
However, if we only add this edge when
$\Pr[r_i(x) \leq 0 \wedge r_j(x) > 0]$ is upper bounded by
some carefully chosen quantity that changes over rounds,
the effect of this bias on the incurred regret will be
limited.
}
In reverse, as illustrated in Figure~\ref{f:aaa},
the resulting feedback graph can be substantially more
beneficial since it may have many more edges than $G^{\text{\sc sub}}$,
hence leading to more frequent updates of the experts'
losses and more favorable regret guarantees. This benefit
is further corroborated by experimental results (Section
\ref{sec:experiments}).
\vspace{-0.05in}
Since we do not have access to $p_{j,i}$,
we instead
use empirical estimates
$\widehat p_{j, i}^{t - 1} := 1/(t-1) \sum_{s = 1}^{t} 1_{r_j(x_s) \leq 0, r_i(x_s) > 0}$.
At time $t$, if expert $\xi_j$ is selected, we
update expert $\xi_i$ if the condition
$\widehat p_{j, i}^{t - 1} \leq \gamma_{i, t - 1}$ holds
with $\gamma_{i, t - 1} = \sqrt{5 Q_i(t - 1) \log (t )}/ ( (K-1) (t -
1)). $
If the expert $\xi_{I_t}$ chosen abstains while expert $\xi_j$ predicts and
satisfies $\widehat p_{I _ t, j}^{t - 1}\leq \gamma_{j, t - 1}$, then we do
not have access to the true label $y_t$. In that case, we update
optimistically our empirical estimate as if the expert had loss $0$ at
that round (Step (*) in Alg.~\ref{alg:ucbtype1}).
\vspace{-0.05in}
The feedback graph $G_t$ just described can be defined
via the out-neighborhood of vertex $\xi_j$:
$N_t(j) = \set{\xi_i \in \mathscr E \colon\widehat p_{j, i}^{t - 1} \leq \gamma_{i, t - 1}}$.
Let $\mathscr S_p = (\mathscr C_{t, k})_{t \in [T], k \in [p]}$ denote
any sequence of admissible $p$-partitionings of these feedback graphs, then the
following regret guarantee holds for \textsc{UCB-GT}.
\vspace{-0.05in}
\begin{theorem}
\label{th:ucbabs}
For any $t \in [T]$, let the feedback graph $G_t$ be defined via the out-neighborhoods
$N_t(j) = \set{\xi_i \in \mathscr E \colon \widehat p_{j, i}^{t - 1} \leq \gamma_{i,t - 1}}$.
Then, \textsc{UCB-GT}\ admits the following regret guarantee:
\vspace{-0.13in}
\begin{equation*}
R_T(\textsc{UCB-GT})
= \mathscr O \Bigl( \min_{p, \mathscr S_p} \sum_{k\in [p]} \frac{\max_{j\in \mathtt{C}_{k} } \Delta_j}{\min_{j \in \mathtt{C}_{k} }\Delta_j^2} \log(T) + K \Bigr )~.
\end{equation*}
\end{theorem}
\vspace{-0.10in}
Since the graph $G_t$ of \textsc{UCB-GT}\ has more edges than
$G^{\text{\sc sub}}$, it admits more admissible partitionings than $G^{\text{\sc sub}}$,
which leads to a more favorable guarantee than that of $\textsc{UCB-NT}$ run
with $G^{\text{\sc sub}}$. The proof of this result differs from the standard UCB analysis and that of Theorem~\ref{th:ucbn} in that it
involves showing that the \textsc{UCB-GT}\ algorithm can adequately control the amount of
bias introduced by the skewed loss estimates. The experiments in the next section provide an empirical validation of
this theoretical comparison.
\input{experiments}
\vspace{-0.1in}
\section{Conclusion}
\vspace{-0.05in}
We presented a comprehensive analysis of the novel setting of online
learning with abstention, showing that efficient learning is
possible in both the adversarial and stochastic scenarios. We drew
connections between this new online setting and existing work
involving time-varying feedback graphs, generalizing and extending
prior work while resolving the issue of biased loss observations. We
presented a novel algorithm, \textsc{UCB-GT}, that carefully uses
biased estimates and which admits favorable regret guarantees. Finally, we
presented a thorough experimental comparison showing that
\textsc{UCB-GT}\ largely outperforms \textsc{UCB-NT}\ with the fixed graph $G^{\text{\sc abs}}$ (an adaptation
to the abstention scenario of a competitor available in the literature),
and achieves performance that is close to an unrealistic and overly optimistic
full-information benchmark.
\vspace{-0.02in}
The concept of online learning with abstention is general. This work can
be extended to similar problems, including the multi-class and
regression settings, as well as other scenarios, such as online
learning with budget constraints.
|
\section{Introduction and Main Result} \label{introduction}
The basic elements of analysis on Sierpinski-Gasket type fractals were developed, in a more general context, by Kusuoka~\cite{Ku89} and Kigami~\cite{Kig89}. The following definitions are from~\cite{KigMsbleRGeom}, though our presentation is different at some points and was influenced by~\cite{TepHarmCoords,Tep04}; detailed proofs of results not demonstrated here may be found in~\cite{Kigbook}.
\begin{defn}
For $N\geq2$ the classical $N$-Sierpinski Gasket $K_N$ is defined as follows. Let $\{p_j\}_{j=1}^N$ be the vertices of a regular $N$ simplex in $\mathbb{R}^{N-1}$ such that $|p_j-p_k|=1$ if $j\neq k$, and $F_i:\mathbb{R}^{N-1}\to\mathbb{R}^{N-1}$ be $F_j(x)=(x-p_j)/2 +p_j$. Then $K_N$ is the unique non-empty compact set such that $K_N=\cup_1^N F_j(K_N)$. Note that $K_2$ is an interval and $K_3$ is the usual Sierpinski Gasket.
\end{defn}
Fix $N\geq2$. For notational simplicity we write $K$ for $K_N$. Then $K$ is post-critically finite, with post-critical set $V_0=\{p_1,\dotsc,p_N\}$. Let $W_m=\{1,\dotsc,N\}^m$ denote the set of words length $|w|=m$, so $w\in W_m$ is $w=w_1\dotsm w_m$ with each $w_j\in\{1,\dotsc,N\}$. Let $W_*=\cup_m W_m$ and for $w=w_1\dotsm w_m\in W_*$ define $F_w=F_{w_1}\circ\dotsm\circ F_{w_m}$. We set $V_m=\cup_{w\in W_m} F_w(V_0)$, which we call the set of scale~$m$ vertices; evidently this is the level~$m$ critical set. We write $p\sim_m q$ if $p$ and $q$ are both in $F_w(V_0)$ for some $w\in W_m$.
For each $N\geq2$ there is a non-negative definite, symmetric, quadratic form on $K=K_N$ which may be defined as a limit of forms on the sets $V_m$.
\begin{defn}
For $u,v$ continuous functions on $K$ let
\begin{equation}\label{eqn:defnofDFm}
\DF_m(u,v) = \Bigl(\frac{N+2}{N} \Bigr)^m \sum_{p\sim_m q} \bigl( u(p)-u(q)\bigr)\bigl( v(p)-v(q) \bigr).
\end{equation}
Write $\DF_m(u)=\DF_m(u,u)$. It is known that $\DF_m(u)\leq\DF_{m+1}(u)$ for all $m\geq0$, so that $\lim_m\DF_m(u)$ exists. Define $\domDF=\{u:\lim_m\DF_m(u)<\infty\}$ and $\DF(u,v)=\lim_m \DF_m(u,v)$. When $N=2$, $\domDF$ is the space of functions with one derivative in $L^2$ that vanishes at the endpoints, and $\DF$ is the $L^2$ norm of the derivative.
\end{defn}
A proof of the following result is in~\cite{Kigbook}.
\begin{thm}
Let $\mu$ be an atomless Borel probability measure on $K$ with $\mu(O)>0$ if $O$ is non-empty and open. Then $(\DF,\domDF)$ is a local regular Dirichlet form on $L^2(K,\mu)$.
\end{thm}
The analytic structure corresponding to the Dirichlet form $(\DF,\domDF)$ is greatly elucidated by considering the harmonic gasket, or gasket in harmonic coordinates. A function $h\in\domDF$ is harmonic if $\DF_m(h)=\DF_{m+1}(h)$ for all $m$. Since this involves minimization of a quadratic functional the minimizer is given by a linear operator on the values on $V_m$. When $m=0$ we write $H_j$ for the operator taking the values on $V_0$ to those on $F_j(V_0)$, more precisely:
\begin{equation*}
H_j: \bigl(h(p_1),\dotsc,h(p_N)\bigr) \mapsto \bigl(h(F_j p_1),\dotsc,h(F_j p_N)\bigr).
\end{equation*}
By self-similarity we find $H_j: \bigl(h(F_w p_1),\dotsc,h(F_w p_N)\bigr) \mapsto \bigl(h(F_wj p_1),\dotsc,h(F_wj p_N)\bigr)$ for any $w\in W_*$, and therefore the values of $h$ on $V_*$ are determined by composing the $H_j$ (or multiplying the corresponding matrices, which we also denote by $H_j$); this determines $h$ on $K$ by continuity, so the fact that the $H_j$ are invertible ensures the harmonic functions are in one-to-one correspondence with functions on $V_0$ and are an $N$-dimensional space. The latter may be seen by explicitly computing (from~\eqref{eqn:defnofDFm}) that
\begin{equation}
H_1= \frac{1}{N+2} \begin{bmatrix}
N+2 & 0 \\
2 & I+J
\end{bmatrix}
\end{equation}
where $I_{N-1}$ is the identity, $J_{N-1}$ is the size $N-1$ square matrix with all entries equal $1$, and we have written $0$ and $2$ for the length $(N-1)$ vectors with all entries equal $0$ and $2$ respectively. Symmetry implies the other $H_j$ may be obtained by cyclic row and column permutations.
\begin{defn}
Fix $N\geq2$ and for $1\leq j\leq N$ let $\psi_j$ denote the harmonic function on $K_N$ which is $1$ at $p_j$ and $0$ on $V_0\setminus\{p_j\}$. The function $\Psi:K\to\mathbb{R}^N$ given by $\Psi=\bigl((\psi_1,\dotsc,\psi_N)-(1,\dotsc,1)/N\bigr)/\sqrt{2}$ is injective (see~\cite{Kig93}), so it is a homeomorphism onto its image $X_N$, which we call the {\em harmonic N-Sierpinski gasket}.
\end{defn}
Constant functions are harmonic, so $\sum_j \psi_j\equiv1$ and the harmonic N-Sierpinski gasket $X_N$ lies in the subspace $\pl=\bigl\{(x_1,\dotsc,x_N):\sum_j x_j=0\bigr\}$, which we identify with $\mathbb{R}^{N-1}$. Moreover $\Psi(V_0)$ is the set of vertices of a regular $N$-simplex with unit length sides in this subspace, so we identify $\Psi(V_0)$ with $V_0$ via $\Psi(p_j)=p_j$. It is an important fact that $X_N$ is a self-affine set under maps conjugate to the $F_j$ via $\Psi$.
\begin{thm}[\protect{\cite{Kig93}}]
For each $1\leq j\leq N$ let $T_j$ be the linear map on $\pl$ that contracts the direction $\Psi(p_j)$ by the factor $N/(N+2)$ and all orthogonal directions by the factor $1/(N+2)$. Let $S_j(x)=T_j(x-p_j)+p_j$, so $S_j:Y\to Y$. Then $\Psi\circ F_j= S_j\circ\Psi$ for each $j$ and therefore $X_N=\cup_1^N S_j(X_N)$ is self-affine.
\end{thm}
\begin{rmk}
We will sometimes write $S_j$ in the equivalent form
\begin{equation}\label{eq:Sj}
S_j x = \frac{x}{N+2} + \frac{1}{N+2} \Bigl( 2+ \frac{(N-1)x\cdot p_j}{\|p_j\|^2} \Bigr) p_j
\end{equation}
\end{rmk}
It is a special case of results of Kusuoka~\cite{Ku89} that there is a Carr\'{e} du Champs measure $\nu$ for the Dirichlet form $(\DF,\domDF)$ and an associated $\nu$-a.e.\ defined metric $Z$ that can be expressed in terms of the operators $T_j$ as in the following results. Recall that the Hilbert-Schmidt norm on a linear operator on $\mathbb{R}^{N-1}$ may be defined by setting $\|T\|_{\text{HS}}^2$ to be the sum of the squares of the coefficients of the associated matrix with respect to the standard orthonormal basis.
\begin{thm}[\protect{\cite{Ku89}}]
For $w=w_1\dotsm w_m\in W_*$ let $T_w=T_{w_1}\dotsm T_{w_m}$ and $Z_m(w) = T_w T_w^t/\|T_w\|_{\text{HS}}^2$, where the adjoint is taken with respect to the usual inner product on $\pl$. Let $\Sigma=\{1,\dotsc,N\}^\mathbb{N}$ denote the space of infinite words.
\begin{enumerate}[(i)]
\item There is a unique, atomless, Borel regular probability measure $\nu$ on $\Sigma$, called the Kusuoka measure, such that for any $w\in W_*$
\begin{equation*}
\nu( w\Sigma)=\frac{1}{N-1} \Bigl( \frac{N+2}{N}\Bigr)^{|w|} \bigl\| T_w \bigr\|^2_{\text{HS}}.
\end{equation*}
\item The limit $Z(w)=\lim_{m\to\infty} Z_m(w_1\dotsm z_m)$ exists and is the orthogonal projection onto its image for $\nu$-a.e.\ $w=w_1 w_2\dotsm\in\Sigma$.
\end{enumerate}
\end{thm}
Both $\nu$ and $Z$ can be transferred to $K$ using the obvious projection $\pi:\Sigma\to K$, which is defined by $\pi(w)=\lim_m F_{w_1\dotsm w_m}K$, because this projection is injective off the measure-zero set $W_*$. For this reason we abuse notation to use $\nu$ and $Z$ for the measure and a.e.-defined linear operator obtained by pushing forward under $\pi$.
Together, the form $(\DF,\domDF)$, measure $\nu$ and metric $Z$ are analogues of the Riemannian energy, volume and metric in that if $\mathcal{C}$ denotes the $\nu$-measurable functions $K\to\pl$ then the following theorem of Kusuoka holds.
\begin{thm}[\protect{\cite{Ku89}}]
There is $\tilde{\nabla}:\domDF\to\mathcal{C}$ such that for all $u,v\in\domDF$
\begin{equation*}
\DF(u,v) = \int_K \langle \tilde{\nabla}u,Z\tilde{\nabla}v\rangle\,d\nu.
\end{equation*}
\end{thm}
The sense in which this structure is simplified by considering the harmonic gasket $X_N$ is captured by the following theorem which relates $\tilde{\nabla}$ to the classical gradient $\nabla$ on $\pl=\mathbb{R}^{N-1}$. Roughly speaking, it says that the analytic structure we have on $K$ is just the restriction of the usual smooth structure on $\pl$ to the harmonic gasket $X$, transferred to $K$ via $\Psi$, when $K$ is endowed with the Kusuoka measure.
\begin{thm}[\protect{\cite{Ku89,KigMsbleRGeom}}]
Let $\mathcal{D}=\{u=v|_X\circ\Psi: v \text{ is } C^1 \text{ on a neighborhood of $X$ in $\pl$}\}$. Then $\mathcal{D}$ is dense in $\domDF$ with respect to the norm $\|u\|^2=\DF(u)+\|u\|_\infty^2$. Moreover for $u,v\in\mathcal{D}$ we have $\tilde{\nabla}u=Z\nabla u$ $\nu$-a.e.\ and
\begin{equation*}
\DF(u,v) = \int_K \langle \nabla u,Z\nabla v\rangle\, d\nu.
\end{equation*}
\end{thm}
One main result of~\cite{KigMsbleRGeom}, see also Teplyaev~\cite{Tep04}, was that in addition to the energy, measure and metric structure described above, the harmonic Sierpinski gasket $X_3$ admits a geodesic distance analogous to the Riemannian distance. Moreover the distance between two points may be computed by integrating the norm, computed with respect to the metric $Z$, of the directional derivative along a geodesic path joining these points. Kigami states in~\cite{KigMsbleRGeom}, but does not verify, that similar results can be proved for $X_N$, $N>3$ using a similar but more complicated argument. The purpose of the present work is to prove this claim by a slightly different approach involving de~Rham curves. Specifically we prove the following.
\begin{thm}\label{thm:geodesicspace}
If $p,q\in X$ there is a Euclidean geodesic from $p$ to $q$ in $X$. There is a $C^1$ function $g_{pq}:[0,1]\to X$ that parametrizes this curve, and the length of the curve is
\begin{equation}\label{eqn:geodesiceqn}
\int_0^1 \left\langle g_{pq}'(t), Z\bigl(g_{pq}(t)\bigr)\, g_{pq}'(t) \right\rangle^{1/2}\, dt.
\end{equation}
\end{thm}
Knowing Theorem~\ref{thm:geodesicspace} it follows from the work in Section~6 of Kigami's paper~\cite{KigMsbleRGeom} that the heat semigroup associated to the Dirichlet form $\DF$ on $L^2(\nu)$ has a jointly continuous kernel $p(t,x,y)$ which has Gaussian bounds with respect to the geodesic distance $d_\ast$ on $X$. Precisely, we obtain the following.
\begin{cor} There are constants $c_1,c_2,c_3,c_4$ such that if $B(x,r)$ denotes the ball of radius $r$ around $x$ in the metric $d_\ast$ then for $t\in(0,1]$ and $x\in X$
\begin{equation*}
\frac{c_1}{\nu(B(x,\sqrt{t}))} \exp \biggl( \frac{-c_2 d_\ast(x,y)^2}{t} \biggr)
\leq p(t,x,y)
\leq \frac{c_3}{\nu(B(x,\sqrt{t}))} \exp \biggl( \frac{-c_4 d_\ast(x,y)^2}{t} \biggr)
\end{equation*}
\end{cor}
Note that Kajino~\cite{Kajino} has proved more refined estimates for the heat kernel using a related distance estimate on the classical Sierpinski gasket $X_3$. His approach is not applicable to the gaskets with $N>3$, as he points out in~\cite{Kajino2}, and it is not known whether these more precise results are true in this setting.
\section{Geodesics on $X_N$: Proof of Main Result}
Fix $N\geq2$ and write $X=X_N$ for the harmonic N-Sierpinski gasket. Recall that there are affine maps $S_j:\pl\to\pl$ so that $X=\cup_1^N S_j(X)$. To obtain our geodesics we consider the subsets of $X$ that are self-affine under two of the maps $S_j$. By symmetry it is sufficient to consider the unique, non-empty, compact set $\Gamma$ such that $\Gamma=S_1(\Gamma)\cup S_2(\Gamma)$. Our initial goal is to prove the following.
\begin{thm}\label{thm:Gammarectifiable}
$\Gamma$ is a $C^1$ plane curve connecting $p_1,p_2\in V_0$. If $N\geq3$ it is $C^1$ but not $C^2$; in fact it has a parametrization in which the derivative is H\"older continuous with H\"older exponent
\begin{equation*}
\frac{\log N +2\log 2 - 2\log (1+\sqrt{4N+1}) }{\log (1+\sqrt{4N+1}) - \log 2 -\log(N+2) }.
\end{equation*}
Moreover, if $x\in\Gamma$, so $x=S_w(X)$ for some infinite word $w$ with letters from $\{1,2\}$ then $Z(w)$ is projection onto the tangent direction of $\Gamma$ at $w$.
\end{thm}
We give several intermediate results before proving the theorem. Let $P$ be the $2$-dimensional subspace of $\pl$ that contains $p_1$ and $p_2$ and let $\Pi$ denote the orthogonal projection of $\pl$ onto $P$.
\begin{lem}\label{lem:GammainP}
$P$ is invariant under $S_1$ and $S_2$ and therefore contains $\Gamma$. Moreover $\Pi(V_0\setminus\{p_1,p_2\})$ is a single point $p$.
\end{lem}
\begin{proof}
It suffices by symmetry to consider the action of $S_1$. If $x\in P$ we may orthogonally decompose it as $x=\alpha p_1+ (x-\alpha p_1)$. Then \begin{equation*}
S_1(x)=\frac{N}{N+2}\alpha p_1 + \frac{1}{N+2} (x-\alpha p_1)
= \frac{N-1}{N+2} \alpha p_1 + \frac{1}{N+2} x
\end{equation*}
which is a linear combination of $p_1\in P$ and $x\in P$, so lies in $P$. This proves invariance of $P$ under $S_1$ and $S_2$. It follows that the closure of $\{S_w p_1: w\in W_*\}$ is a subset of $P$, but since it is also non-empty, compact, and equal to the union of its images under $S_1$ and $S_2$ it must be equal to $\Gamma$, proving $\Gamma\subset P$. Finally, symmetry ensures that all points in $V_0\setminus\{p_1,p_2\}$ project to the same point $p\in P$, and that $|p_1-p|=|p_2-p|$.
\end{proof}
Many other properties of $\Gamma$ now follow from the fact that $\Gamma$ is an arc on a classical type of curve introduced by de~Rham~\cite{deRham1,deRham2}. This fact was noted by Teplyaev in~\cite{Tep04} for the case of the classical Sierpinski gasket, where he also gave a number of results about the energy and Laplacian in harmonic coordinates.
\begin{defn}\label{def:deRhamcurve}
Fix a polygonal arc $A_0$ and a ratio $r\in(0,1/2)$. Inductively suppose we have defined the polygon $A_{k-1}$, introduce two new vertices on each side of $A_{k-1}$ so as to divide the side into subintervals with length ratios $r:1-2r:r$ and let $A_k$ be the polygon defined by the new vertices ordered as they were on $A_{k-1}$. De Rham proved the $A_k$ converge to a continuous curve $A$; we call $A$ a de~Rham curve.
\end{defn}
To understand the de~Rham curves it helps to recognize that the midpoints of segments of the initial polygon are unchanged by the construction, and that the construction of the arc between two adjacent midpoints is independent of the remainder of the curve, so that it is sufficient to consider the region between two such midpoints. We can therefore perform the construction with an initial polygon having three vertices and two edges to obtain a de~Rham curve between the midpoints of these edges. The significance for our problem is in the following theorem.
\begin{thm}\label{thm:GammaisdeRham}
$\Gamma$ is an arc of a de~Rham curve in $P$ with scale factor $r=\frac{1}{N+2}$.
\end{thm}
\begin{proof}
Let $A_0$ be the line segments from $2p_1$ to $0$ and $0$ to $2p_2$, so that $p_1$ and $p_2$ are the midpoints of the sides. Let $A_k$, $A$ be as in Definition~\ref{def:deRhamcurve} with ratio $r$, except that we retain only the arc between $p_1$ and $p_2$.
Write $[a,b]$ for the line segment from $a$ to $b$ and $D$ for one step of the de~Rham construction. A key observation is that $D$ commutes with $S_1$ and $S_2$ because $D$ depends only on dividing line segments according to fixed proportions and the $S_j$ preserve proportions. We use this to show inductively that $A_{k}=\cup_{|w|=k} S_w\bigl([p_1,0]\cup [p_2,0]\bigr)$. For $k=0$ this is trivial, as the edges of $A_0$ are precisely $[p_1,0]\cup [p_2,0]$, but our induction will actually start at the $k=1$ case. To verify this latter, observe that $S_1$ fixes $p_1$ and scales $[p_1,0]$ by $N/(N+2)$, so it maps $[p_1,0]$ to $[p_1,2p_1/(N+2)]$ which is one edge of $A_1$. Similarly $S_2$ maps $[p_2,0]$ to $[p_2,2p_2/(N+2)]$, which is another edge of $A_1$. By direct computation we then check that $S_1p_2=S_2p_1=(p_1+p_2)/(N+2)$, which is the midpoint of $[2p_1/(N+2),2p_2/(N+2)]$ and conclude that the union $S_2([p_1,0])\cup S_1([p_2,0])$ is the third edge of $A_1$.
The inductive step uses that $A_{k-1}= \cup_{|w|=k-1} S_w\bigl([p_1,0]\cup [p_2,0]\bigr)$ to see that $A_{k-1}=S_1(A_{k-2})\cup S_{2}(A_{k-2})$, and therefore for $k\geq2$
\begin{equation*}
A_k=D(A_{k-1})
=D \bigl( S_1(A_{k-2}) \cup S_2(A_{k-2}) \bigr).
\end{equation*}
All pairs of neighboring segments in $S_1(A_{k-2}) \cup S_2(A_{k-2})$ are internal to either $S_1(A_{k-2})$ or $S_2(A_{k-2})$, with the exception of the segments that meet at $S_1(p_2)=S_2(p_1)$. However, by the inductive hypothesis these segments run from $S_1 S_2^{k-1}(0)$ to $S_1(p_2)=S_2(p_1)$ and from $S_2 S_1^{k-1}(0)$ to $S_1(p_2)=S_2(p_1)$. Moreover there is $c$ so $S_2^{k-1}(0)=cp_2$ and $S_1^{k-1}(0)=cp_1$, because both $S_1$ and $S_2$ contract by the same factor along the respective directions $p_1$, and $p_2$. Using~\eqref{eq:Sj} and computing $(N-1)p_1\cdot p_2/\|p_1\|^2=-1$ we find
\begin{align*}
S_1 S_2^{k-1}(0) = S_1 c p_2 = \frac{cp_2}{N+2} + \frac{2-c}{N+2}p_1\\
S_2 S_1^{k-1}(0) = S_2 c p_1 = \frac{cp_1}{N+2} + \frac{2-c}{N+2}p_2
\end{align*}
so that the midpoint is $(p_1+p_2)/(N+2)=S_1(p_2)=S_2(p_1)$. This shows that the two ending segments of $S_1(A_{k-2})$ and $S_2(A_{k-2})$ form a single line segment in $A_{k-1}$ as soon as $k\geq2$. It follows that all of the de~Rham construction for $D \bigl( S_1(A_{k-2}) \cup S_2(A_{k-2}) \bigr)$ occurs within either $S_1(A_{k-2})$ or $S_2(A_{k-2})$. Then commuting $D$ with the $S_j$, $j=1,2$ gives
\begin{equation*}
A_k=D \bigl( S_1(A_{k-2}) \cup S_2(A_{k-2}) \bigr)
=S_1 \bigl( D(A_{k-2}) \bigr) \cup S_2 \bigl( D(A_{k-2}) \bigr)
= S_1 (A_{k-1}) \cup S_2 (A_{k-1}).
\end{equation*}
This and the inductive hypothesis ensure $A_k=\cup_{|w|=k} S_w\bigl([p_1,0]\cup [p_2,0]\bigr)$, from which we deduce that $A_k$ converges to a set $A$ that is invariant under the iterated function system $\{S_1,S_2\}$. Uniqueness of the attractor of the i.f.s.\ then ensures $A=\Gamma$.
\end{proof}
The de~Rham curves have been fairly extensively studied because they have applications in wavelets, approximation theory, and certain areas in computer science. References for some of these may be found in a paper of Protasov~\cite{Prot04}, which makes a detailed study of the regularity of these curves. It is proved there that the de~Rham curves are affine similar sets, which is the main point in the above proof of Theorem~\ref{thm:GammaisdeRham}; more importantly for the current work he proves (a more general version of) the following result.
\begin{thm}[\protect{\cite{Prot04}} Theorem~2]\label{thm:Protasov}
A de~Rham curve is $C^1$ if $r\in(0,\frac13]$. Moreover if $r\in(0,\frac14)$ the curve is not $C^2$ but has H\"{o}lder continuous derivative with H\"{o}lder exponent
\begin{equation*}
\frac{\log \bigl( r(1-2r)\bigr)} { \log (r+\sqrt{4r-7r^2}) - \log2} - 2.
\end{equation*}
\end{thm}
\begin{proof}[Proof of Theorem~\protect{\ref{thm:Gammarectifiable}}]
From Lemma~\ref{lem:GammainP}, $\Gamma$ is a plane curve. Theorem~\ref{thm:GammaisdeRham} shows that $\Gamma$ is a de~Rham curve, and by Theorem~\ref{thm:Protasov} it is $C^1$.
Now $x\in\Gamma$ is $T_w(X)$ for an infinite word $w$ with letters in $\{0,1\}$. If we write $[w]_m$ for the trunctation to the first $m$ letters then the sequence $T_{[w]_m}([p_1,p_2])$ consist of chords of $\Gamma$ that converge to $x$. Since $\Gamma$ is $C^1$ the normalized sequence $T_{[w]_m}T^t_{[w]_m}/\|T_{[w]_m}\|_{\text{HS}}^2$ converges to the operator of projection onto the tangent direction to $\Gamma$ at $x$.
\end{proof}
Having established the basic regularity properties of $\Gamma$ our next goal is to show that it is the shortest path between $p_1$ and $p_2$ in $X$. In order to proceed we collect some additional features of $\Gamma$ in the next result.
\begin{lem}\label{lem:factsaboutUpsilon}
$\Gamma$ and the line segment from $p_1$ to $p_2$ bound an open convex region. If $\Upsilon$ denotes the complement of this region in the triangle $\Omega$ with vertices $p,p_1,p_2$, then $\Upsilon$ is star-shaped with respect to $0$ and has the property that $(\Upsilon+ \alpha p)\cap \Omega\subset \Upsilon$ for any $\alpha\geq0$.
\end{lem}
\begin{proof}
It is useful to think of constructing the interior of the convex hulls of the approximating curves $A_k$ for our de~Rham curve $A$ as part of the inductive construction. If we denote the $k^{\text{th}}$ hull interior by $C_k$ one easily checks that $C_k$, $k\geq1$ can be obtained as follows: for two adjacent vertices $x,y\in A_k$ the line through $x$ and $y$ divides the plane into an open half-plane $K(x,y)$ that contains both $p_1$ and $p_2$ and its complementary (closed) half-plane $K'(x,y)$; $C_k$ is the intersection of $C_{k-1}$ with all $K(x,y)$ corresponding to adjacent vertices $x,y\in A_k$.
Evidently $\cap_k C_k$ is convex and has $\Gamma$ as a boundary arc; when it is further intersected with the half-plane that is bounded by the line through $p_1,p_2$ and contains $0$ we obtain the convex region asserted in the statement. Then $\Upsilon$ is the intersection of $\Omega$ with the union of the $K'(x,y)$, with the latter taken over all $k$ and all pairs of adjacent vertices from each $A_k$.
Consider the above for three consecutive vertices, $x+v_1,x,x+v_2$ from $A_{k-1}$, $k>1$. The convex set $\{x+\alpha_1v_1+\alpha_2v_2: \alpha_1>0, \alpha_2>0\}$ contains $p_1$ and $p_2$ (for which one of $\alpha_1$ or $\alpha_2$ is at least $1$) and is the intersection $K(x+v_1,x)\cap K(x,x+v_2)$. The vertices of $A_k$ introduced at the $k^{\text{th}}$ step are of the form $y_1=x+\beta_1v_1$, $y_2=x+\beta_2 v_2$ with $\beta_1,\beta_2\in(0,1/2)$, so the line through them has direction $\beta_1 v_1-\beta_2v_2$. At any point $y$ on this line the cone $L(v_1,v_2,y)=\{y-\alpha_1 v_1-\alpha_2v_2: \alpha_1\geq0,\alpha_2\geq0\}$ is disjoint from the line (because it does not contain points of the form $y+\delta(\beta_1 v_1-\beta_2v_2)$), and cannot contain $p_1,p_2$ because these are reached from $x$ using non-negative multiples of $v_1$ and $v_2$, with at least one coefficient being $1$ or more.
We make two observations from the reasoning in the preceding paragraph. The first is that that $L(v_1,v_2,y)\subset K'(y_1,y_2)$ at every $y$ on the segment from $y_1$ to $y_2$. Then, by induction, if $y\in A_{k}$ is between vertices $z,z'$ of $A_{k}$ we see $L(p_1,p_2,y)$ is contained in $K'(z,z')$ and hence $\Omega\cap L(p_1,p_2,y)\subset\Upsilon$. Taking the union over $k$ gives that $L(p_1,p_2,y)\subset\Upsilon$ if $y$ is on any $A_k$. This immediately establishes the assertion about translation of $\Upsilon$ by $\alpha p$ for $\alpha>0$. The second observation is that $L(v_1,v_2,z)\subset K'(y_1,y_2)$ for any $z\in K'(y_1,y_2)$, so in particular $L(v_1,v_2,x)\subset K'(y_1,y_2)$. Induction from this shows that the first such cone, $L(p_1,p_2,0)$ is contained in $K(z,z')$ for each pair of adjacent vertices $z,z'$ in any $A_k$. Since $0$ is in this cone we conclude that $0\in K(z,z')$ for each pair of adjacent vertices $z,z'$ in every $A_k$, and since $K(z,z')$ is convex and hence star-shaped with respect to any of its points we it follows that $\Upsilon$ is a union of regions star-shaped with respect to $0$ and is therefore itself star-shaped with respect to $0$.
\end{proof}
\begin{lem}\label{lem:nonconvexside}
$\Pi(X)\subset\Upsilon$.
\end{lem}
\begin{proof}
We show that the intersection of $\Pi^{-1}(\Upsilon)$ with the simplex having vertices $\{p_j\}_1^N$ is invariant under the iterated function system $\{S_j:j=1,\dotsc,N\}$. It therefore must contain the attractor $X$.
Both $S_1$ and $S_2$ contract every vector that is orthogonal to $P$ by the same factor. It follows that $\Pi\circ S_j\circ \Pi^{-1}$ is well-defined and equal to $S_j\bigr|_P$ for $j=1,2$. Since $\Gamma$ is invariant under $S_1\bigr|_P$ and $S_2\bigr|_P$ it is easy to check that these maps take $\Upsilon$ into itself, whereupon $\Pi\circ S_j\circ \Pi^{-1}=S_j\bigr|_P$ for $j=1,2$ implies $\Pi^{-1}(\Upsilon)$ is invariant under $S_1$ and $S_2$.
By symmetry, to treat the maps $S_j$, $j\geq3$ it suffices to consider $S_3$. If one has $\sum_j \alpha_j=1$ then $x=\sum_j \alpha_j p_j$ is a point in the simplex and its projection to $P$ is $\Pi x=\alpha_1 p_1+\alpha_2 p_2+\sum_3^N \alpha_j p$. We may also compute $S_3x$ using~\eqref{eq:Sj} and the fact that $(N-1)p_j\cdot p_k/\|p_j\|^2=-1$. We obtain
\begin{gather*}
S_3x= \frac{x}{N+2} + \frac{1}{N+2}\Bigl(2+ (N-1)\alpha_j - \sum_{k\neq j}\alpha_k \Bigr)p_j
= \frac{x}{N+2} + \frac{N\alpha_j+1}{N+2}p_j\\
\Pi S_3 x = \frac{\Pi x}{N+2} + \frac{N\alpha_j+1}{N+2}p
\end{gather*}
From Lemma~\ref{lem:factsaboutUpsilon}, $\Upsilon$ is star-shaped with respect to $0$ and therefore $\Pi x\in\Upsilon$ implies $\Pi x/(N+2)\in\Upsilon$. Lemma~\ref{lem:factsaboutUpsilon} further established that the image of $\Upsilon$ under translation by a positive multiple of $p$ followed by intersection with $\Omega$ is contained in $\Upsilon$. As $(N\alpha_j+1)(N+2)>0$ we conclude from this and the previous observation that $\Pi \circ S_3 x\in\Upsilon$ whenever $\Pi x\in\Upsilon$, or equivalently that $S_3$ (and thus, by symmetry, each $S_k$ for $k\geq3$) maps the intersection of $\Pi^{-1}(\Upsilon)$ with the simplex having vertices $\{p_j\}$ to itself.
\end{proof}
We can now prove our main results. Let us write $Z_\ast= Z\circ \pi^{-1}\circ \Psi^{-1}$, which is a well defined map from $X\setminus\Psi(V_\ast)$ to the set of infinite words $\Sigma$.
\begin{thm}\label{thm:mainthm}
Using the restriction of the Euclidean metric from $Y$, $\Gamma$ is a geodesic in $X$ connecting $p_1$ to $p_2$. There is a $C^1$ function $g:[0,1]\to\Gamma$ such that the length of $\Gamma$ is
\begin{equation*}
l(\Gamma) = \int_0^1 \left\langle g'(t), Z_\ast(g(t))\, g'(t) \right\rangle^{1/2} \, dt.
\end{equation*}
Moreover we have
\begin{equation}\label{eqn:lengthbound}
\frac{1}{3}\leq l(\Gamma) \leq 2
\end{equation}
\end{thm}
\begin{proof}
$\Gamma$ is fixed by $S_1$ and $S_2$, so it is a subset of $X$. The fact that that $\Gamma$ is rectifable ensures the infimum of the length of curves in $X$ that contain $p_1$ and $p_2$ is well-defined. Moreover, if $\gamma$ is such a curve then $\Pi(\gamma)$ is rectifiable, contains $p_1$ and $p_2$ and has length $ l(\gamma)\geq l(\Pi(\gamma))$. By Lemma~\ref{lem:nonconvexside} $\Pi(\gamma)\subset\Upsilon$, so $\Pi(\gamma)\cup[p_1,p_2]$ surrounds the convex set $T\setminus\Upsilon$. However it is a result of classical geometry that then
\begin{equation*}
l\bigl(\Pi(\gamma)\cup[p_1,p_2]\bigr)\geq l\bigl(\Gamma\cup[p_1,p_2]\bigr)
\end{equation*}
so $\Gamma$ achieves the infimum and is a geodesic.
The same convexity argument shows~\eqref{eqn:lengthbound} because $\Omega\setminus\Upsilon$ is contained in the triangle with vertices $p_1,p_2,0$ and contains the triangle with vertices $p_1,p_2, S_1(p_2)=S_2(p_1)=(p_1+p_2)/(N+2)$. We need only check that the lengths of the sides are $|p_1|=|p_2|=\sqrt{(N-1)/2N}\leq 2$ and $|(N+1)p_1-p_2|/(N+2)=\sqrt{(N^2+N+2)/2(N+1)^2}\geq1/3$ because $N\geq 2$.
We may parametrize $\Gamma$ in the following way. Take $e_1= p_1-p_2 $ and $e_2=\bigl(N/(N-2)\bigr)^{1/2} (p_1+p_2)$ as a basis for the plane $P$; it is easily checked that these are orthonormal. For $t\in[0,1]$ the fact that $\Gamma$ bounds a convex region ensures there is a unique point $g(t) = p_2+te_1 + se_2$ on $\Gamma$. The function $g(t)$ is $C^1$ because $\Gamma$ is $C^1$, and $g'(t)$ is in the direction of $Z_\ast(g(t))$, so
\begin{equation*}
l(\Gamma) = \int_0^1 \|g't)\|\, dt = \int_0^1 \left\langle g'(t), Z_\ast(g(t))\, g'(t) \right\rangle^{1/2} \, dt. \qedhere
\end{equation*}
\end{proof}
\begin{cor}\label{cor:mainresult}
If $p,q\in F_w(V_0)$ there is a geodesic from $p$ to $q$ with length comparable to the diameter of $F_w(X)$ which has a $C^1$ parametrization $g_{pq}$ for which~\eqref{eqn:geodesiceqn} gives the length of the geodesic.
\end{cor}
\begin{proof}
Given $1\leq j<k\leq N$ one can find an orthogonal transformation of $Y$ that maps $p_1$ to $p_j$ and $p_2$ to $p_k$. Under this map the image of $\Gamma$ is a geodesic curve $\Gamma_{jk}$ from $p_j$ to $p_k$. Evidently $\Gamma_{jk}$ is the invariant set of $\{T_j,T_k\}$. Now there are $j,k$ such that $p=F_w(p_j)$ and $q=F_w(p_k)$ and $F_w(\Gamma_{jk})$ is a geodesic from $p$ to $q$ in $F_w(X)$. The length bound follows from~\eqref{eqn:lengthbound} and the validity of~\eqref{eqn:geodesiceqn} comes from the fact that the restriction of $Z_\ast$ to $F_w(X)$ is $T_w Z_\ast$.
\end{proof}
The proof of Theorem~\ref{thm:geodesicspace} now follows the argument given to prove Theorem~5.1 on page~798 of~\cite{KigMsbleRGeom}, but is included below for the convenience of the reader.
\begin{proof}[Proof of Theorem~\protect{\ref{thm:geodesicspace}}]
Suppose $p,q\in V_\ast$, so there is $m$ so they are both in $V_m$. There is a finite sequence $p=p_1,\dotsc,p_k=q$ with no repetitions such that $p_j$ and $p_{j+1}$ are in the boundary of $F_w(V_0)$ with $|w|=m$. Concatenation of the geodesics from $p_j$ to $p_{j+1}$ that were constructed in Corollary~\ref{cor:mainresult} provides a finite length path from $p$ to $q$.
Conversely, any finite length path from $p$ to $q$ passes through some sequence of points from $V_m$. We may shorten the path by deleting loops, at which point this sequence $p=p_1,\dotsc,p_k=q$ is finite and has no repetitions. The curve constructed above from geodesics joining $p_j$ to $p_{j+1}$ is evidently the shortest path from $p$ to $q$ that passes through these points in this order. Moreover there are finitely many sequences of this type, so by taking the minimum of the lengths of the resulting finite collection of paths we find a geodesic from $p$ to $q$ which we denote $\gamma_{pq}$. This proves the result when both points are in $V_\ast$.
Observe that in the above construction if $p,q\in V_m$ and $\gamma_{pq}$ passes through $p',q'\in V_{m'}$ for some $m'<m$ then replacing the arc of $\gamma_{pq}$ from $p'$ to $q'$ with $\gamma_{p'q'}$ does not increase the length. Thus for sequences $p_j\to q$ and $q_j\to q$ there are geodesics $\gamma_{j}$ from $p_j$ to $q_j$ such that $\gamma_j$ is an arc of $\gamma_k$ for each $j<k$. Then $\cup_j\gamma_j$ is a geodesic from $p$ to $q$ and is made up of arcs as constructed in Corollary~\ref{cor:mainresult}.
To each of the above arcs there is a finite word and an orthogonal transformation as in Corollary~\ref{cor:mainresult} so that the composition of $F_w$, the orthogonal map and the function $g$ from Theorem~\ref{thm:mainthm} is a $C^1$ parametrization of the arc. Concatenating these functions (and arcs) gives a parametrization of the geodesic from $p$ to $q$, and it satisfies~\eqref{eqn:geodesiceqn} because Theorem~\ref{thm:mainthm} shows the parametrizations of individual arcs had this property. Moreover it is $C^1$ because at any point of $V_\ast$ where two arcs from the construction are joined, the one-sided tangents are equal in the same way that they were when we joined the two halves of $\Gamma$ in the proof of Theorem~\ref{thm:GammaisdeRham}.
\end{proof}
\section{Acknowledgements}
The authors thank Alexander Teplyaev for useful discussions and suggestions.
|
\section{Introduction}
Power exhaust is one of the major challenges for a future fusion device.
Future fusion devices are considered to operate in a regime with enhanced confinement, the so called H-Mode.
This regime is usually accompanied by edge localized modes (ELMs) which lead to periodic bursts of energy and particles from the confined plasma towards the plasma facing components.
Applying a non-axisymmetric external magnetic perturbation (MP) is one technique that is studied in order to mitigate or suppress large ELMs in next step fusion devices such as ITER~\cite{Lang2013, Loarte2014}.
The thermal load due to ELMs might limit the lifetime of the divertor of these devices~\cite{Loarte2003}.
Many of today's experiments are equipped with magnetic coils to study the physics and feasibility of ELM mitigation/suppression with an external magnetic perturbation, e.g. ASDEX Upgrade~\cite{Suttrop2009}, DIII-D~\cite{Evans2004}, EAST~\cite{SunEAST2016}, JET~\cite{Liang2007}, KSTAR~\cite{Jeon2012}, MAST~\cite{Kirk2015}, NSTX~\cite{Ahn2010}.
Most of the studies focus on changes of global plasma parameters, e.g. density, or the increase in ELM frequency~\cite{Kirk2015, Thornton2015, SunEAST2016}.\\
Applying external magnetic perturbation brakes the axisymmetry of a tokamak and leads to a 2D heat flux pattern on the divertor target~\cite{Muller2013,Jakubowski2009,Harting2012,Ahn2010,Thornton2014, Faitsch2016}.
It is reported that for ITER it might be necessary to rotate the magnetic perturbation in order to prevent local over-heating due to the toroidally asymmetric heat load~\cite{Loarte2014}.
In this paper the 2D heat flux is studied at ASDEX Upgrade in stationary L-Mode discharges.\\
In section~\ref{Experiment} the plasma parameters and measurements are introduced.
In section~\ref{ExperimentResults} the results obtained from the experiments are shown.
Section~\ref{Modelling} introduces a generic model for the target heat flux pattern using field line tracing in the vacuum field approach.
Section~\ref{ModellingResults} compares this model with experimental data.
Conclusions and a summary are given in section~\ref{Conclusion}.
\section{Experiment}\label{Experiment}
ASDEX Upgrade~\cite{AUG2013} is equipped with two toroidal rows of 8 saddle coils, one below the outer midplane (lower coils) and one above the outer midplane (upper coils)~\cite{Suttrop2009, Suttrop2011}, shown in~\fref{fig:AUGCoils}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{AUGMPCoils.png}
\caption{ASDEX Upgrade magnetic perturbation coils.
Shown is an n\,=\,2 perturbation with a \textit{differential phase} $\Delta\phi\,=\,0$ between the upper row and the lower row.}
\label{fig:AUGCoils}
\end{figure}
The power supplies are able to produce a rigid rotation of a toroidal mode number n\,=\,2 perturbation with various poloidal mode spectra.
The poloidal spectrum is varied by the phase between the currents of the two sets of coils ({\textit{differential phase}, $\Delta \phi$)~\cite{Teschke2015}.
The magnetic perturbation is rotated with 1\,Hz in all discharges discussed in this paper.
This frequency corresponds to the current inside the coils.
With an n\,=\,2 perturbation the magnetic perturbation phase is rotated by $\pi$ within 1\,s.\\
The temperature evolution of the divertor target is measured using an infrared (IR) system~\cite{Sieglin2015}, measuring at a wavelength of $4.7\,\mu\mathrm{m}$ with a FWHM of 125\,nm and a frequency of 800\,Hz, optimized for L-Mode discharges.
The resolution on the outer divertor target is about 0.6\,mm/pixel.
The heat flux to the divertor target is calculated using THEODOR~\cite{Herrmann1995, Sieglin2015}.
The IR system is observing a toroidal location at an angle of $\phi_{IR} = 213^\circ$ in the ASDEX Upgrade coordinate system.
This corresponds to a phase of $\phi_{IR} = 33^\circ$ in relation to the n\,=\,2 perturbation.
The global plasma parameters for the reference shot \#\,32212 are shown in \fref{fig:TimeTraceReferenceShot}.
The same parameters are used for the study of the \textit{differential phase} in section~\ref{diffPhase}.
In section~\ref{densSteps} discharges with higher stationary densities are discussed.
The scenario has a toroidal magnetic field of -2.5\,T and a plasma current of 0.8\,MA.
About 370\,kW of external ECR heating is applied to the center in order to increase the temperature rise at the divertor target and thus increasing the IR signal quality whilst still staying in L-Mode.
The plasma shape and current distribution is fully evolved at around 2.0\,s.
In the reference discharge a constant magnetic perturbation was applied between 4.5\,s and 5.0\,s.
The application of the magnetic perturbation does not change the global plasma parameters.
This allows to study the sole effect of the magnetic perturbation on the heat flux.
Also, the discharges performed with the rigid rotation have a reference phase without the magnetic perturbation in the beginning of the discharges before 2.5\,s.
This is done to have the possibility to check and hence ensure the similarity of the discharges as well as providing a comparison with the axisymmetric target profiles.
This results in about 3.5 rotations throughout every discharge, ensuring the reproducibility of the measurement.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32212_TimeTraces.png}
\caption{Time traces of the global parameters for the reference shot \#\,32212 without magnetic perturbation until 4.5\,s and a steady magnetic perturbation phase until 5.0\,s.
The top panel shows the plasma current that is kept at 0.8\,MA.
The second panel shows the core and edge line integrated density in black and red, respectively.
The third panel shows the stored energy in the plasma.
The bottom panel shows the external heating.}
\label{fig:TimeTraceReferenceShot}
\end{figure}
The heat flux onto the outer divertor target for the discharge \#\,32217 is shown in \fref{fig:RefHeat}.
The heat flux profile obtained in the phase without the magnetic perturbation is shown in black.
The red curve shows the heat flux profile for the phase with the magnetic perturbation.
The heat flux profile without magnetic perturbation is described using the 1D diffusive model presented in~\cite{Eich2011}:
\begin{eqnarray} \label{eq:diffusiveModel}
q(s) &= \frac{q_0}{2} \exp \left(\left(\frac{S}{2 \lambda_q}\right)^2 - \frac{s}{\lambda_q f_x} \right) \cdot \mathrm{erfc} \left(\frac{S}{2 \lambda_q} - \frac{s}{S f_x}\right) & \left[\frac{\mathrm{MW}}{\mathrm{m}^2}\right]
\end{eqnarray}
with s the target location, $S$ the divertor broadening, $\lambda_q$ the power fall-off length and $f_x$ the poloidal flux expansion.
The structure observed in the heat flux profile with magnetic perturbation is referred to as lobe structure, e.g.~\cite{Kirk2015, Ahn2010}, or strike line splitting, e.g.~\cite{Jakubowski2009, Harting2012}, with the characteristic of multiple distinguished peaks as a consequence of the non-axisymmetry.
A hot spot is present, marked in the figures with a grey area.
This hot spot exhibits a larger temperature increase which is not taken into account in the evaluation of the heat flux.
This leads to a too high estimated heat flux.
The area is thus excluded from any further discussions.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32217_MP_ComparisonSingleTimePoint_Without.pdf}
\caption{1D heat flux profile for discharge \#\,32217 with (red) and without (blue) magnetic perturbation.
In the area in grey a hot spot is present.}
\label{fig:RefHeat}
\end{figure}
\subsection{Differential Phase}
The \textit{differential phase} $\Delta\phi$ between the upper and lower coil currents is changed in steps of $\frac{\pi}{2}$ with fixed maximum coil current amplitudes $I_{\mathrm{coil}}$\,=\,1\,kA.
Additional discharges are performed using only the upper (lower) coils.
The \textit{differential phase} of $\Delta\phi\,=\,-\frac{\pi}{2}$ is field line aligned at the edge (q\,=\,5 surface) and is therefore called the \textit{resonant} configuration~\cite{Suttrop2011}.
The \textit{differential phase} with $\Delta\phi\,=\,+\frac{\pi}{2}$ is called \textit{non-resonant} configuration accordingly.
The \textit{resonant} configuration with a reduced amplitude of the perturbation field $I_{\mathrm{coil,red}}\,=\,\frac{1}{\sqrt{2}}~I_{\mathrm{coil}}$ is performed to investigate the influence of the perturbation strength on the 2D structure and changes in the transport properties in the scrape-off layer.
\section{Divertor Heat Loads}\label{ExperimentResults}
In this section the results obtained with the IR thermography system are discussed.
\subsection{2D Heat Flux Profiles}\label{diffPhase}
The heat flux evolution on the outer divertor target for various \textit{differential phases} $\Delta \phi$ between lower and upper coil currents is shown in~\fref{fig:DifferentialPhasesIRTime}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:2DResonant} Resonant $\Delta\phi\,=\,-\frac{\pi}{2}$]{\includegraphics[width=0.3 \textwidth]{32217_HeatFlux_TimeTrace.pdf}}
\subfigure[\label{fig:2DNonResonant} Non-Resonant $\Delta\phi\,=\,+\frac{\pi}{2}$]{\includegraphics[width=0.3 \textwidth]{32218_HeatFlux_TimeTrace.pdf}}
\subfigure[\label{fig:2DUpper} Upper Only]{\includegraphics[width=0.3 \textwidth]{32219_HeatFlux_TimeTrace.pdf}}\\
\subfigure[\label{fig:2DLower} Lower Only]{\includegraphics[width=0.3 \textwidth]{32220_HeatFlux_TimeTrace.pdf}}
\subfigure[\label{fig:2DMid} $\Delta\phi\,=\,0,\,+\pi$]{\includegraphics[width=0.3 \textwidth]{32221_HeatFlux_TimeTrace.pdf}}
\subfigure[\label{fig:2DResonantReduced} Resonant $\Delta\phi\,=\,-\frac{\pi}{2}$ with reduced $I_{\mathrm{coil}}$]{\includegraphics[width=0.3 \textwidth]{32416_HeatFlux_TimeTrace.pdf}}
\caption{Heat flux time traces for the various \textit{differential phases} between the upper and lower coil currents.
White bars indicate start and end time of the external perturbation, respectively.}
\label{fig:DifferentialPhasesIRTime}
\end{figure}
The largest visual influence of the magnetic perturbation on the heat flux profile evolution is observed in the \textit{resonant} configuration (see \fref{fig:2DResonant}).
It is observed that the strike line position, identified as the sharp rise of the heat flux profile along the target, moves in time when the magnetic perturbation is present.
This is due to imperfections of the attitude control in presence of n\,=\,2 external perturbation, e.g.~\cite{Willensdorfer2016}, possibly enlarged by the presence of internal modes due to the plasma response, for example seen with JOREK simulations~\cite{Orain2016}.
For the further analysis this is taken into account by shifting the heat flux profiles to a fixed strike line position.
The \textit{non-resonant} configuration is shown in~\fref{fig:2DNonResonant}.
The deviation between the axisymmetric heat flux without magnetic perturbation (between 2.0 and 2.5\,s) and with magnetic perturbation is reduced compared to the \textit{resonant} configuration.
However, the helical structure is still present.
The discharges with only the upper and lower coils used are shown in~\fref{fig:2DUpper} and \fref{fig:2DLower}, respectively.
The visual heat flux perturbation is more pronounced than in the \textit{non-resonant} configuration but less pronounced than in the \textit{resonant} configuration.
The phases in between the \textit{resonant} and \textit{non-resonant} configuration are both performed in a single discharge, shown in~\fref{fig:2DMid}.
Neither for a \textit{differential phase} of 0 from 2.5-3.7\,s nor a phase of $\Delta\phi\,=\,+\pi$ between 4.0 and 6.0\,s any difference is seen.
The heat flux profile evolution with the \textit{resonant} configuration, but a reduced current in the coils $I_{\mathrm{coil,red}}\,=\,\frac{1}{\sqrt{2}}~I_{\mathrm{coil}}$, is shown in~\fref{fig:2DResonantReduced}.
\subsection{Toroidally Averaged Heat Flux Profiles} \label{section:averaged}
All global plasma parameters are kept constant in the presented discharges, allowing to average the heat flux profile over one or more full rotation periods.
Four different phases in the rotation of the \textit{resonant} configuration are shown in~\fref{fig:HeatFluxDifferentPosResonant}.
The heat flux profiles are, as mentioned in the previous section, shifted along the target location to match in the strike line position.
The before mentioned hot spot is observed (grey area in~\fref{fig:HeatFluxDifferentPosResonant}) leading to an overestimation of the local heat flux at this position but does not affect the analysis of the fitting to the averaged profile.
The maximum heat flux in all phases is close to the axisymmetric maximum and the maximum of the averaged profile.
A more detailed characterisation of the heat flux distribution is presented in section~\ref{ModellingPeakingProfile} with a comparison to a generic heat flux model.\\
In~\fref{fig:DiffPhaseAveraged} heat flux profiles are shown for the shots shown in~\fref{fig:DifferentialPhasesIRTime}.
The heat flux profiles are normalized to the integrated heat flux for the given profile and to the peak heat flux of the reference profile for \#\,32217.
The normalization is performed because the integrated heat flux in the different shots varies in the order of 10\%.
This is within the regularly observed L-Mode heat flux variation.
A normalization to the peak heat flux for the individual profiles would result in the same figure due to the similarity of the profiles.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32217_MP_ComparisonSingleTimePoints_Averaged.pdf}
\caption{Heat flux profiles for different time points in the \textit{resonant} configuration as well as the averaged profile.}
\label{fig:HeatFluxDifferentPosResonant}
\end{figure}
In~\fref{fig:axiWO} heat flux profiles at 2.3\,s are shown.
This is within the reference phase before the magnetic perturbation field is applied at 2.5\,s.
The profiles for all discharges are similar and described by the 1D diffusive model~\eref{eq:diffusiveModel}.
From this it is concluded that the discharges are comparable in terms of edge transport leading to the observed heat flux pattern.
Both transport qualifiers $\lambda_q$ and $S$, obtained by fitting the model to the experimental data, vary only within the fitting uncertainty.
The uncertainty given is the standard deviation for the obtained values.
In~\fref{fig:axiWith} time averaged profiles in the phase with magnetic perturbation are shown.
For the averaging a time window between 3.0\,s and 5.0\,s is chosen.
Except for \#\,32221 where two different \textit{differential phases} are applied and therefore two separate averaged profiles are conducted between 2.7\,s and 3.7\,s for $\Delta \phi \,=\,0$ and between 4.5\,s and 5.5\,s for $\Delta\phi\,=\,+\pi$.
The time windows are chosen to avoid the transient phases during switching of the external magnetic perturbation.
Averaging over 2.0\,s (except for \#\,32221), which corresponds to two periods, ensures that no artificial heat flux variations are present.
The averaged heat flux profiles, in contrast to the single profiles, are described by the 1D diffusive model.
No dependence of the averaged profile on the \textit{differential phase} is observed, in contrast to the different 2D profiles discussed at the beginning of this section.
The parameters describing the heat transport in the scrape-off layer $\lambda_q$ as well as the divertor region $S$ do not show a dependence on the magnetic perturbation.
This is interpreted that there is neither significant additional cross field transport $\chi_{\perp}$ nor significant additional net radial transport along radially deflected field lines due to the change of the radial magnetic field caused by the perturbation coils.
The lobe structure causes a redistribution of heat flux in toroidal direction with an averaged target profile described by the global $\lambda_q$ and $S$.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:axiWO} Heat flux profiles in the reference time window without magnetic perturbation]{\includegraphics[width=0.45 \textwidth]{1DProfile_Reference.pdf}}
\subfigure[\label{fig:axiWith} Time averaged heat flux profiles with different phases between upper and lower coil currents.]{\includegraphics[width=0.45 \textwidth]{1DProfile_Average.pdf}}
\caption{Heat flux profiles with and without magnetic perturbation.}
\label{fig:DiffPhaseAveraged}
\end{figure}
\subsection{Toroidal Heat Flux Variation} \label{section:TimeVariation}
As shown in the previous section, the time averaged profile with rotating magnetic perturbation leads to the same target heat flux profile as the 1D heat flux profile without magnetic perturbation.
Averaging in time with rotating magnetic perturbation is equivalent with averaging in toroidal direction for a static magnetic perturbation or an infinite fast rotation.
Due to the finite rotation frequency in the experiment (1\,Hz) the 2D structure on the target is measured by moving it through the field of view of the IR system.\\
In~\fref{fig:TimeVariation} the heat flux for different target locations in the scrape-off layer, normalized to the averaged heat flux at this position, is shown for the \textit{resonant} and \textit{non-resonant} configuration.
A time window between 3.0\,s and 5.0\,s is chosen containing two complete periods of the rotation.
The target location is expressed in terms of the fitted $\lambda_{q,\mathrm{target}}$ for the averaged heat flux profile.
The black profile (s = 0.21$\cdot \lambda_q$) corresponds to about the peak in the heat flux profile.
Further into the SOL (s \textgreater $\lambda_q$) the IR data becomes more noisy due to the lower signal.
The dots show the single IR measurements, the solid line is smoothed over 125\,ms (100 time points).
The variation between the single time points and the averaged is mainly due to typical heat flux variations in L-Mode and only to a small fraction due to measurement noise.
The heat flux variation for the \textit{resonant} configuration is shown in~\fref{fig:timeVarRes}.
The period in which the heat flux is above the mean value is about the same as the period in which it is below with a nearly sinusoidal structure.
The peak to peak variation varies for the different target locations, with the largest at about 0.67$\cdot \lambda_q$ away from the separatrix with about a factor of 4.
Positions further in the scrape-off layer have less heat flux variation and less averaged heat flux.
Positions closer to the separatrix have less heat flux variation compared to 0.67 $\cdot \lambda_q$ away from the strike line position but with a higher averaged heat flux.\\
This is a direct consequence of the x-point configuration.
The toroidal inclination of the lobes increases towards the strike point which causes a toroidal symmetric profile at the strike line.
This is independent of the heat flux width and set by the unperturbed field configuration.
However, the most critical part for local over-heating along the target of the profile is around $\leq\,\lambda_q$ away from the separatrix, where significant averaged heat flux with a strong variation with the cycle of the magnetic perturbation rotation is observed.\\
The time variation of the heat flux for the \textit{non-resonant} configuration is shown in~\fref{fig:timeVarNon}.
For this configuration the heat flux does not vary significantly in time, as is also seen in the 2D time trace in~\fref{fig:2DNonResonant}.
It has to be noted here, that the structure of the perturbation, although only altering the heat flux marginal, is still observable in the 2D profile.
A comparison between the \textit{toroidal peaking} - maximum value in time in~\fref{fig:TimeVariation} - and the target location is discussed in section~\ref{ModellingPeakingProfile}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:timeVarRes} Resonant]{\includegraphics[width=0.45 \textwidth]{32217_ToroidalPeaking_TimeTrace.pdf}}
\subfigure[\label{fig:timeVarNon} Non-Resonant]{\includegraphics[width=0.45 \textwidth]{32218_ToroidalPeaking_TimeTrace.pdf}}
\caption{Time variation for different target locations normalized to the average heat flux at this position.}
\label{fig:TimeVariation}
\end{figure}
A variation in time with the rotating magnetic perturbation is also observed by divertor Langmuir probes.
The temperature measured by 4 probes in the scrape-off layer is shown in~\fref{fig:LangmuirTe}.
It has to be noted that the spatial distance between the probes is 20\,mm and there was no strike line sweep in the discharge.
Thus, the spatial resolution is rather low and no proper profile can be constructed.
The about sinusoidal oscillation observed in the heat flux measured by the IR thermography is also observed in the electron temperature.
However, compared to the peak to peak variation of the heat flux of up to a factor of 4, the variation in the electron temperature at the target is in the range of 20\,\%.
The position of the two probes in the scrape-off layer is substantially further away from the strike line position than the positions where the heat flux variation measured by the IR thermography is extracted from.
The probe data shown in black is taken at about the peak heat flux position measured by the IR thermography.
The variation is in both, the electron temperature measured by the Langmuir probes as well as the heat flux measured by the IR thermography, in the same order.
Note here, that the IR thermography data is shifted due to a strike line movement, this is not possible for the Langmuir probes due to the limited spatial resolution.
The time window of 2\,s allows the variation that has a periodicity of 1\,s to be attributed to the MP and not to the strike line movement which seems to have only a minor impact.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32217_Langmuir_Temperature.pdf}
\caption{Electron temperature variation on the outer divertor target measured with fixed Langmuir probes.}
\label{fig:LangmuirTe}
\end{figure}
\subsection{Effect of Increasing Density}\label{densSteps}
The \textit{resonant} configuration is used to study the effect of magnetic perturbations on $\lambda_q$ and $S$ on the outer divertor target at different densities.
It was shown in previous studies~\cite{Scarabosio2013, Sieglin2013, Sieglin2016} that in L-Mode discharges at ASDEX Upgrade both, $\lambda_q$ and $S$, depend on the electron density at the plasma edge.
The main mechanism to increase $S$ with density is thought to be the reduction of the electron temperature in the divertor region
and, therefore, the reduction of the parallel heat conduction~\cite{Scarabosio2015, Sieglin2016}.\\
A density scan is performed to be able to choose densities that are low enough to have an attached divertor and no significant divertor radiation.
This condidtions result in $T_e\,>$\,10\,eV, whilst still spanning an as large as possible density range.
To study the toroidally averaged profiles, stable conditions are needed.
Three discharges with different density levels are referred to as \textit{low}, \textit{medium} and \textit{high} density corresponding to a line integrated edge density of $n_{e,edge}\,=\,0.8, 1.5, 1.8 \cdot 10^{19} \mathrm{m}^{-2}$, respectively.
The density is measured with the edge channel of the DCN interferometer~\cite{Mlynek2010}.
The terms of low, medium and high density are a choice of convenience for this paper. They do not refer to operation at high density essential for high fusion performance and high radiative scenarios.
The 2D heat flux profiles are shown in~\fref{fig:density2DHeat}, the \textit{medium} density in~\fref{fig:2DMediumDensity} and the \textit{high} in~\fref{fig:2DHighDensity}.
The \textit{low} density reference is shown in~\fref{fig:2DResonant}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:2DMediumDensity} Medium Density, $n_{e,edge}\,=\,1.5 \cdot 10^{19} \mathrm{m}^{-2}$]{\includegraphics[width=0.45 \textwidth]{32406_HeatFlux_TimeTrace.pdf}}
\subfigure[\label{fig:2DHighDensity} High Density, $n_{e,edge}\,=\,1.8 \cdot 10^{19} \mathrm{m}^{-2}$]{\includegraphics[width=0.45 \textwidth]{32415_HeatFlux_TimeTrace.pdf}}
\caption{Heat flux time traces for the discharges with higher density.
White bars indicate start and end time of the external perturbation, respectively.}
\label{fig:density2DHeat}
\end{figure}
In all three discharges the 2D structure of the heat flux profile is seen.
Comparing the profiles for the elevated densities with the \textit{low} density reference reveals that increasing the density reduces the heat flux variation.
This is shown in~\fref{fig:densityTimeVar} for all three densities.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:timeVarResLow} Low Density, $n_{e,edge}\,=\,0.8 \cdot 10^{19} \mathrm{m}^{-2}$]{\includegraphics[width=0.3 \textwidth]{32217_ToroidalPeaking_TimeTrace.pdf}}
\subfigure[\label{fig:timeVarResMed} Medium Density, $n_{e,edge}\,=\,1.5 \cdot 10^{19} \mathrm{m}^{-2}$]{\includegraphics[width=0.3 \textwidth]{32406_ToroidalPeaking_TimeTrace.pdf}}
\subfigure[\label{fig:timeVarResHigh} High Density, $n_{e,edge}\,=\,1.8 \cdot 10^{19} \mathrm{m}^{-2}$]{\includegraphics[width=0.3 \textwidth]{32415_ToroidalPeaking_TimeTrace.pdf}}
\caption{Time variation for different target locations normalized to the average heat flux at this position.}
\label{fig:densityTimeVar}
\end{figure}
The density difference between the \textit{medium} and \textit{high} density discharges is small and no clear difference in the time variation of the heat flux is seen.
The variation in both cases is around 20\,\%.
In~\fref{fig:density1DHeat} the comparison between the 1D profiles for the reference time window (\fref{fig:DensityStepsWithout}) and the toroidally averaged profile with magnetic perturbation in the \textit{resonant} configuration (\fref{fig:DensityStepsWith}) is shown.
The normalization is the same as in section~\ref{section:averaged}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:DensityStepsWithout} without magnetic perturbation]{\includegraphics[width=0.45 \textwidth]{DensityStepsAxi}}
\subfigure[\label{fig:DensityStepsWith} with magnetic perturbation]{\includegraphics[width=0.45 \textwidth]{DensityStepsAverage}}
\caption{Heat flux profiles with and without magnetic perturbation for the three different densities.}
\label{fig:density1DHeat}
\end{figure}
As mentioned in the beginning of the section, an increase for both, $\lambda_q$ and $S$, is expected for increasing density in ASDEX Upgrade L-Mode.
This is confirmed in the three density steps without magnetic perturbation as the peak heat flux is reduced by a factor of 2.
With increasing density the toroidally averaged profiles are still described by the 1D diffusive model.
The transport qualifiers are similar to the reference values without magnetic perturbation and no significant change with density is observed.
\section{Heat Flux Model}\label{Modelling}
To interpret the experimental results a simple model was developed.
With this model the influence of different quantities, e.g. coil current, divertor broadening $S$ and poloidal spectrum (\textit{resonant} vs. \textit{non-resonant}), is studied.
The intention of this model is to get sufficient agreement with the measured heat flux profiles without treating all the plasma parameters.
This allows to aim for as few as possible (free) input parameters as well as the possibility to change single parameters and study the direct influence they have onto the heat flux distribution.\\
This model is based on the vacuum field approach and a field line tracer, similar approaches are discussed in~\cite{Nguyen1997, Finken1998, Strumberger1996, Eich2000, Cahyna2014}.
The used field line tracer is the 5th order Runge-Kutta GOURDON code~\cite{Gourdon1971, Strumberger1996, Strumberger2000}.
The axisymmetric poloidal flux matrix is calculated using CLISTE~\cite{Schneider2000, McCarthy1999}.
For every step the magnetic field of the perturbation coils is calculated and added to the axisymmetric magnetic field.
The current inside these coils is measured and is corrected taking the conductive passive stabilizing loops (PSL) into account which acts as a low pass filter~\cite{Gruber1993}.
The effective current inside the coils is reduced by 25\% according to FEM calculations at a rotation frequency of 1\,Hz~\cite{Suttrop2009EPS}.
Field lines are traced starting at the outer divertor target and either end at some plasma facing component, e.g. inner divertor, or end up in the confined region and are terminated at a maximum length of \textgreater\,2\,km.
No difference is observed if the field lines are terminated at a length of \textgreater\,200\,m.\\
The following assumptions are made:
\begin{itemize}
\item $R_{sep}(\phi)$ = $A_0 \cdot \sin\left(2\cdot \phi + B_0\right) + A_1 \cdot \sin\left(4\cdot \phi + B_1\right) + R_{sep,axi}$\\
A 2D separatrix at the outer midplane (OMP, z\,=\,0) is defined using the major radius of the axisymmetric separatrix $R_{sep,axi}$ as a mean value (mean($R_{sep}(\phi)$) = $R_{sep,axi}$) and a toroidal $\phi$ sinusoidal component ($\phi\,\epsilon\,[0,2\pi)$).
The periodicity of the sinusoidal is given by the dominant mode number n of the applied magnetic perturbation, which is kept constant at n\,=\,2 for these experiments.
The amplitude of the deformation is fitted using an arbitrary but fixed field line length approaching the separatrix in the unperturbed case.
In the presented results this was fixed to a range between 120-125\,m.
This separatrix is used to define the poloidal flux coordinate $\rho_{pol}(R,\phi)$ also as a 2D quantity.
\item $q_{||}$ = const. along the field line.\\
No perpendicular heat transport. The perpendicular information is covered with the temperature fall-off length. The cross field transport in the divertor region is simplified by convolving the target profile with a Gaussian.
\item $T_e$($\rho_{pol}$), $q_{||}$($\rho_{pol}$) are flux quantities and have an exponential fall-off length.\\
The electron temperature fall-off length $\lambda_{Te}$ is calculated using the two point model~\ref{eq:2Point} and the power-fall-off length $\lambda_q$.
The two-point model is in good agreement for ASDEX Upgrade L-Mode discharges in the density range from 2 to $6\cdot10^{19} \mathrm{m}^{-2}$ using the power fall-off length measured at the outer divertor target and the temperature fall-off length measured with the Thomson scattering diagnostics~\cite{Faitsch2015, Sun2015}.
\item The local field line angle at the target $\alpha$ is calculated for tiles without tilting using.\\
The tilting of the divertor targets is needed to prevent leading edges. A second minor change in the field line angle is the flat tile surface, not following the toroidal direction of the vessel.
The measurement position is not changed within the presented study and thus the effect of the tilting is only a constant attenuation.
Both effects together lead to a variation of the heat flux amplitude in toroidal direction of about 20\,\% and an additional, but less pronounced, difference in direction of the target location.
The angle is calculated with the magnetic perturbation present, although it does not change the angle significantly.\\
\item $L_{\mathrm{OMP}}$ = constant.\\
The length between the OMP and the outer divertor target is approximated to be a constant.
This is done in order to have both an exponential temperature decay and an exponential heat flux decay in the OMP for an axisymmetric configuration.
This reduces the number of free parameters to the separatrix temperature.
\item $T_{e\mathrm{,target}}$\,=\,0\,eV.\\
Using the same approximation as in the derivation for~\eref{eq:2Point}.
\end{itemize}
The following numerical values are used for all calculations:
\begin{itemize}
\item $L_{\mathrm{OMP}}$ = 25\,m.
\item $\lambda_q$\,=\,3.67\,mm, $S = 0.3$\,mm (deduced from measurement at \#\,32217 @ 2.3\,s).
\item $\kappa_0$ = 2000\,$\frac{\mathrm{W}}{\mathrm{m} \cdot \left(\mathrm{eV}\right)^{\frac{7}{2}}}$~\cite{Stangeby2000, Kallenbach2001}
\item $T_{e,sep} = 46, 45, 46$\,eV for the axisymmetric, \textit{resonant}, \textit{non-resonant} case, respectively.
The separatrix temperature is set to match the peak heat flux from the IR measurement.
This leads to reasonable values for a low power L-Mode in ASDEX Upgrade.
\end{itemize}
The parallel heat flux $q_{||}$ is calculated using the two point model with Spitzer-H\"arm electron conduction:
\begin{eqnarray}
q_{||} = -\kappa_0 T^{\frac{5}{2}} \frac{\mathrm{d}T}{\mathrm{d}x}
\end{eqnarray}
with $x$ the coordinate along the field line and $T$ the electron temperature~\cite{Stangeby2000}.
Integrating along $x$ leads to a relation between the upstream temperature $T_u$ and the target temperature $T_t$ with the connection length $L_{\mathrm{OMP}}$
\begin{eqnarray}
T_u = \left(T_t^{\frac{7}{2}} + \frac{7 q_{||} L_{\mathrm{OMP}}}{2 \kappa_0}\right)^{\frac{2}{7}}
\end{eqnarray}
Neglecting the target electron temperature $T_t$ ($T_u\,\gg\,T_t$):
\begin{eqnarray}
T_u \approx \left(\frac{7 q_{||} L_{\mathrm{OMP}}}{2 \kappa_0}\right)^{\frac{2}{7}}
\end{eqnarray}
leading to a ratio between upstream temperature fall-off length~$\lambda_{T_e}$ and power fall-off length~$\lambda_q$ of
\begin{eqnarray} \label{eq:2Point}
\frac{\lambda_{T_e}}{\lambda_q} = \frac{7}{2}
\end{eqnarray}
and a parallel heat flux $q_{||}$ of
\begin{eqnarray}
q_{||} \approx \frac{2}{7} \frac{\kappa_0 T_u^{\frac{7}{2}}}{L_{OMP}}
\end{eqnarray}
The heat flux perpendicular to the target is calculated using the pitch angle $\alpha$ giving:
\begin{eqnarray} \label{eq:perpendicular}
q_{\perp} = \sin\left(\alpha\right) \cdot q_{||}
\end{eqnarray}
which is in the order of $\sin\left(\alpha\right) = \frac{1}{20}$ for ASDEX Upgrade and the present divertor configuration Div\,III.
The field line tracer allows the calculation of the parallel heat flux at a given target position (before the convolution with the Gaussian):
\begin{eqnarray} \label{eq:parallel}
q_{||, target}\left(s,\phi\right) = f\left(T_u\right) = f\left(R_{\mathrm{OMP}}, \phi_{\mathrm{OMP}}\right)
\end{eqnarray}
with the before mentioned assumptions and numerical constants.
The heat flux profile is interpolated in order to get a uniform distribution along the target location s.
This heat flux is then convolved with a Gaussian function g with a width $S \cdot f_x$:
\begin{eqnarray}
q_{\perp}^{*} = (q_{\perp}*g)(s)
\end{eqnarray}
with \eref{eq:parallel}, \eref{eq:perpendicular} and
\begin{eqnarray}
g(s) ={\frac {1}{S \cdot f_x {\sqrt {\pi }}}}e^{-\left({\frac {s}{S \cdot f_x}}\right)^{2}}
\end{eqnarray}
the Gaussian as used in~\eref{eq:diffusiveModel}.
\FloatBarrier
\section{Comparison: Heat Flux Model and Experiment}\label{ModellingResults}
In this section the heat flux profiles obtained from the presented model in section~\ref{Modelling} are compared to the experimental heat fluxes measured with the IR system.
\subsection{2D Heat Flux Structure}
The 2D structure of the heat flux profile is examined in this section.
In order to compare the measured heat flux with the model, the time variation from~\fref{fig:2DResonant} is transferred into a toroidal distribution.
This is justified by the constant background plasma parameters.
The strike line position is corrected as mentioned in section~\ref{ExperimentResults}.
The 2D structure for the \textit{resonant} configuration is shown in~\fref{fig:2DResonantAngle}.
Note that here the scale in toroidal direction and along the divertor target are a factor of 100 different, with the circumference of ASDEX Upgrade is in the order of 10\,m and the extent along the target is in the order of 5\,cm.
The toroidal angle of $\phi\,=\,0$ is arbitrary and set to be the heat flux at $t\,=\,3\,$s.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32217_2DHeatFlux_Angle.pdf}
\caption{2D Heat Flux for the \textit{resonant} configuration in \#\,32217 deduced from IR measurements.}
\label{fig:2DResonantAngle}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32218_2DHeatFlux_Angle.pdf}
\caption{2D Heat Flux for the \textit{non-resonant} configuration in \#\,32218 deduced from IR measurements.}
\label{fig:2DNonResonantAngle}
\end{figure}
\subsubsection{Influence of the Conductive Wall onto the Perturbation Field}
The \textit{resonant} configuration with the rotation in the opposite direction is performed to measure the phase delay due to the conductive passive stabilizing loops (PSL) nearby, acting as a low pass~\cite{Suttrop2009EPS, Rott2009}.
From ref~\cite{Suttrop2009EPS} it is known that the phase delay should be about $\frac{1}{24} \pi$ (15$^{\circ}$) for a rotation frequency of 1\,Hz.
Thus, the difference between the two discharges with opposite rotation directions should be $\frac{1}{12} \pi$ (30$^{\circ}$).\\
A comparison between both rotation directions is shown in~\fref{fig:LobePositionRotationDir}.
The dots represent the local maxima $s_m$ in direction of the target location s, $q(s_m-1) < q(s_m)> q(s_m+1)$.
The agreement obtained without a phase shift for the profile close to the former separatrix position can't be improved by adding a phase shift to one of the profiles.
The distribution close to the former separatrix is nearly toroidal due to the x-point geometry.
The change in the target location along the toroidal direction is less than 5\,mm per $\pi$ up to one $\lambda_q$ away from the former separatrix position.
A phase shift close to the former separatrix is not resolvable with the IR system due to the limited spatial resolution.
Further into the scrape-off layer the change in the target location along the toroidal direction becomes larger.
The determination of the local maxima gets more uncertain with less arriving heat flux.
A phase shift in the expected range of $\frac{1}{12} \pi$ is not resolvable with the heat flux data from the IR system.\\
The low pass filter not only leads to a phase shift but also to an attenuation of the amplitude.
The amplitude is reduced by about 25\,\% compared to a static field according to ref~\cite{Suttrop2009EPS}.
This attenuation is accounted for in the calculation of the magnetic field for the modelling.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{RotationDirection_LobePosition.png}
\caption{Lobe position for both rotation directions without phase shift.}
\label{fig:LobePositionRotationDir}
\end{figure}
The positions of the local maxima in the modelled heat flux distribution as well as in the experiment for the \textit{resonant} configuration are shown in~\fref{fig:LobePosition}.
The modelled data is calculated at 3.5\,s and shifted according to the phase shift due to the PSL and the offset of the IR position.
The position of the local maxima is in agreement within the uncertainty.
However, as already discussed in the previous section, the phase information in the measured data is limited due to the nearly toroidal direction close to the former separatrix due to the x-point shear.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{32217_comparison_lobePosition_shifted.png}
\caption{Lobe position for modelling (black) and experiment (red dots) for the \textit{resonant} configuration \#\,32217.}
\label{fig:LobePosition}
\end{figure}
The local maxima for three different coil configurations is shown in~\fref{fig:LobePositionCurrent}.
It is observed that the relative position of the maxima is independent of the coil current setup.
The local maxima represent two fixed upstream toroidal positions (magnetic perturbation with a toroidal mode number n\,=\,2) at the OMP.
The toroidal angle is fixed by the absolute phase of the magnetic perturbation, the position at the target is - after shifting to the same absolute phase - independent of the magnetic perturbation.
This holds as long as the perturbation is small as well as no significant non ideal plasma response shifts the absolute phase of the perturbation, e.g.~\cite{Liu2011, Orain2016}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{Modelling_LobePosition_MPField.pdf}
\caption{Lobe position for the \textit{resonant} configuration with two different coil currents $I_{\mathrm{coil}}$ and for the \textit{non-resonant} configuration.}
\label{fig:LobePositionCurrent}
\end{figure}
\FloatBarrier
\subsection{Transformation of Time into Toroidal Angle with Rotating Magnetic Perturbation}
In the experimental results we transferred from the time evolution of the heat flux at a fixed toroidal location to a toroidal distribution for a fixed time.
In this section the differences between a toroidal distribution and the variation of the coil currents in order to rotate the magnetic perturbation field are discussed.
In the presented model both approaches can be compared.
One is the \textit{fixed time}, starting field lines at different toroidal positions at the target and computing the 2D heat flux structure for the given fixed magnetic coil currents.
The other is the \textit{fixed position}, starting the field lines at a fixed toroidal position at the target and changing the currents in the magnetic perturbation coils in the same way as in the experiment.
For an infinite number of coils the result would be the same.
For a low number of coils per row, as it is the case for the experiments (8 in the case of ASDEX Upgrade, 9 foreseen for ITER~\cite{Daly2013}), the representation of the sinusoidal perturbation varies for different absolute phases.
For an n\,=\,2 perturbation with 8 toroidal coils the difference is negligible.
The 2D heat flux is shown in~\fref{fig:ModellCompFixTimeLoc2D} for both approaches.
The corrugation of the separatrix is not taken into account here.
The variation in toroidal direction (or time transferred to a toroidal angle) at fixed target locations is shown in\fref{fig:ModellCompFixTimeLocVar}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:2DTimeFix} Fixed time]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_fixedSep_2DHeatFlux.pdf}}
\subfigure[\label{fig:2DTimeVar} Fixed position]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_fixedSep_timeVariation_2DHeatFlux.pdf}}
\caption{2D heat flux profile for fixed time and fixed position, both without taking the separatrix corrugation into account.}
\label{fig:ModellCompFixTimeLoc2D}
\end{figure}
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:torVarTimeFix} Fixed Time]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_fixedSep_ToroidalPeaking_3Positions.pdf}}
\subfigure[\label{fig:torVarTimeVar} Fixed Position]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_fixedSep_timeVariation_ToroidalPeaking_3Positions.pdf}}
\caption{Toroidal (time) variation of the heat flux for fixed time and fixed position, both without taking the separatrix corrugation into account.}
\label{fig:ModellCompFixTimeLocVar}
\end{figure}
\FloatBarrier
\subsection{Radial Displacement of the Plasma Boundary}
With the application of the magnetic perturbation the plasma boundary is radially corrugated, e.g.~\cite{Willensdorfer2016}.
As discussed in the beginning of this section, this deformation can be treated with the presented model.
The model is based on field lines, making it convenient to handle the displacement as the change of the length of field lines at the OMP (\fref{fig:ConnectionLengthOMPFit}).
The black line indicates the used non-axisymmetric separatrix.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{Modelling_32217_350s_FieldLineLengthMidplane_withSep.png}
\caption{Field line length at the OMP with the separatrix obtained as presented in section~\ref{Modelling}.}
\label{fig:ConnectionLengthOMPFit}
\end{figure}
The correction of the radial displacement at the OMP leads to an increase of the heat flux variation in the \textit{resonant} configuration and a decrease in the \textit{non-resonant} configuration in the model.
The 2D heat flux is compared in~\fref{fig:2DCompSep}, different radial positions and the averaged profile in~\fref{fig:1DCompSep} as well as the variation at some target positions in~\fref{fig:TimeDepentendCompSep} for the \textit{resonant} configuration.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:Modell2DFixed} Without separatrix corrugation]{\includegraphics[width=0.45 \textwidth]{Modelling_32217_350s_fixedSep_2DHeatFlux.pdf}}
\subfigure[\label{fig:Modell2DCorr} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.45 \textwidth]{Modelling_32217_350s_2DHeatFlux.pdf}}
\caption{2D heat flux profile for the \textit{resonant} configuration.}
\label{fig:2DCompSep}
\end{figure}
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:ModellAveragedFixed} Without separatrix corrugation]{\includegraphics[width=0.45 \textwidth]{Modelling_32217_350s_fixedSep_HeatFlux_1D_Phases.pdf}}
\subfigure[\label{fig:ModellAveragedCorr} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.45 \textwidth]{Modelling_32217_350s_HeatFlux_1D_Phases.pdf}}
\caption{Averaged heat flux profiles for the \textit{resonant} configuration.}
\label{fig:1DCompSep}
\end{figure}
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:ModellTimeVariationFixed} Without separatrix corrugation]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_fixedSep_ToroidalPeaking_3Positions.pdf}}
\subfigure[\label{fig:ModellTimeVariationCorr} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.3 \textwidth]{Modelling_32217_350s_ToroidalPeaking_3Positions.pdf}}
\subfigure[\label{fig:ExperimentTimeVariation} Measurement]{\includegraphics[width=0.3 \textwidth]{32217_ToroidalPeaking_TimeTrace.pdf}}
\caption{Toroidal (time) variation of the heat flux for the \textit{resonant} configuration.}
\label{fig:TimeDepentendCompSep}
\end{figure}
The 2D structure does not change significantly.
However, the single profiles show a larger variation when taking the corrugation into account.
The maximum heat flux at a given toroidal position is not changed and is at the position of the axisymmetric, or toroidally averaged, maximum where the variation is negligible.
The variation at a given target location changes from a nearly sinusoidal for the complete profile to a more triangular shaped variation far away from the separatrix (blue lines in~\fref{fig:TimeDepentendCompSep}).\\
The same plots for the \textit{non-resonant} configuration are shown in figures~\ref{fig:2DCompSepNon},~\ref{fig:1DCompSepNon} and~\ref{fig:TimeDepentendCompSepNon}.
Without taking the displacement into account the modelled profiles do not differ when varying the \textit{differential phase} (\textit{resonant}, \textit{non-resonant}).
The heat flux variation significantly varies with the correction of the corrugation in the \textit{non-resonant} configuration.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:Modell2DFixedNon} Without separatrix corrugation]{\includegraphics[width=0.45 \textwidth]{Modelling_32218_350s_2DHeatFlux_fixedSep.pdf}}
\subfigure[\label{fig:Modell2DCorrNon} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.45 \textwidth]{Modelling_32218_350s_2DHeatFlux.pdf}}
\caption{2D heat flux profile for the \textit{non-resonant} configuration.}
\label{fig:2DCompSepNon}
\end{figure}
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:ModellAveragedFixedNon} Without separatrix corrugation]{\includegraphics[width=0.45 \textwidth]{Modelling_32218_350s_HeatFlux_1D_Phases_fixedSep.pdf}}
\subfigure[\label{fig:ModellAveragedCorrNon} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.45 \textwidth]{Modelling_32218_350s_HeatFlux_1D_Phases.pdf}}
\caption{Averaged heat flux profiles for the \textit{non-resonant} configuration.}
\label{fig:1DCompSepNon}
\end{figure}
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:ModellTimeVariationFixedNon} Without separatrix corrugation]{\includegraphics[width=0.3 \textwidth]{Modelling_32218_350s_ToroidalPeaking_3Positions_fixedSep.pdf}}
\subfigure[\label{fig:ModellTimeVariationCorrNon} With separatrix corrugation $R_{sep}(\Phi)$]{\includegraphics[width=0.3 \textwidth]{Modelling_32218_350s_ToroidalPeaking_3Positions.pdf}}
\subfigure[\label{fig:ExperimentTimeVariationNon} Measurement]{\includegraphics[width=0.3 \textwidth]{32218_ToroidalPeaking_TimeTrace.pdf}}
\caption{Toroidal (time) variation of the heat flux for the \textit{non-resonant} configuration.}
\label{fig:TimeDepentendCompSepNon}
\end{figure}
In the experiment, a clear difference is observed with the variation of the \textit{differential phase}.
In the model this is only reproduced by taking the plasma boundary displacement into account, although, the \textit{non-resonant} case in the model still exhibits stronger variations than observed in the experiment.
\FloatBarrier
\subsection{Description of Toroidal Heat Flux Peaking}\label{ModellingPeakingProfile}
The \textit{toroidal peaking} is defined as the maximum value along the toroidal direction at a given target location normalized to the mean value at this target location.
The \textit{toroidal peaking} along the divertor target is shown in~\fref{fig:ModellingPeakingToroidalvsTarget}.
\begin{figure}[htb!]
\centering
\subfigure[\label{fig:MeasurementPeakingTarget} Measurement]{\includegraphics[width=0.45 \textwidth]{32217_ToroidalPeakingVsPoloidalPosition.pdf}}
\subfigure[\label{fig:ModelPeakingTarget} Modelling]{\includegraphics[width=0.45 \textwidth]{Modelling_ToroidalPeakingVsPoloidalPosition.pdf}}
\caption{Impact of the \textit{toroidal peaking} along the target location.}
\label{fig:ModellingPeakingToroidalvsTarget}
\end{figure}
The toroidally averaged heat flux profile (mean) is shown in red, the toroidal maximum (max) in blue with the normalization described in section~\ref{section:averaged}.
The max profile reveals the extend along the target which is above a certain heat flux.
The \textit{toroidal peaking}, shown in black.
The measurement shown in~\fref{fig:MeasurementPeakingTarget} shows some deviations from the model~\fref{fig:ModelPeakingTarget}.
The maximum value in the model is never exceeding the peak value of the averaged profile whereas in the experiment the maximum is in the order of 20\,\% larger than the averaged peak value.
However, a hot spot in the measurement close to the peak position leads to an overestimation.
The \textit{toroidal peaking} in the experiment is always about 20\,\% larger, mainly due to the L-Mode heat flux variation of the heat flux, see section~\ref{section:TimeVariation}.
In the modelled data the \textit{toroidal peaking} monotonically increases with increasing distance from the strike line position.
In the measurement the \textit{toroidal peaking} seems to saturate at a distance of 1 $\lambda_q$.
However, since the incident (mean) heat flux decreases with distance to the strike line, the uncertainty becomes larger and the interpretation of this saturation has to be treated with caution.
The proposed explanation is the divertor broadening $S$.
The 1D diffusive model (shown in~\eref{eq:diffusiveModel}) has a single parameter for the diffusion in the divertor region, leading to an effective parameter $S$ for the complete profile.
It was shown that the parameter depends on the electron temperature in the divertor volume, sufficiently characterized by the target electron temperature for attached conditions~\cite{Sieglin2016, Scarabosio2015}. As a result this leads to a variation along the divertor target as the temperature is not constant.
An increased divertor broadening in the far SOL explains the reduction in the \textit{toroidal peaking}.
\FloatBarrier
\subsubsection{Influence of the Magnetic Perturbation Coil Current}
The influence of the magnetic perturbation strength for the \textit{resonant} and \textit{non-resonant} configuration on the variation in toroidal direction is shown in~\fref{fig:toroidalPeakingCurrent}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{Modelling_PeakingVsCoilCurrent_toroidalSep.pdf}
\caption{\textit{Toroidal peaking} for different coil currents at the target position 1 $\lambda_q$ away from the strike line. The current is normalized to the maximum current of the ASDEX Upgrade coils.}
\label{fig:toroidalPeakingCurrent}
\end{figure}
A target location of one $\lambda_q$ away from the strike line is chosen in the figure.
For zero coil current an axisymmetric heat flux is expected, having a \textit{toroidal peaking} of 1.0.
The coil current in the figure is normalized to the maximum coil current of the ASDEX Upgrade coils (1.3\,kA x 5 turns).
The nominal current used in the presented study is 5\,kAt, the PSL attenuates 25\,\% of the magnetic field.
The maximum achievable normalized coil current in the experiment is thus $\approx$\,0.6.
In order to rotate the field with a constant amplitude the effective magnetic perturbation is a square-root reduced for the same coil currents (not all coils have the maximum current at the same time).
This is not reflected in the normalization, all data that is shown is from discharges with rotated field.
The perturbation strength with a static field can be about 1.9 times larger with the same current limit of 5\,kAt.\\
The \textit{toroidal peaking} increases nearly linear with increasing current for a given \textit{differential phase}.
The slope is different for the two configurations, the \textit{resonant} configuration has a larger \textit{toroidal peaking} with about a factor two difference in the slope.
It is seen, that for too high currents in the model (more than 2.0 times the maximum current possible in the experiment) this linear dependence is lost.
This is due to the large influence of the perturbation and might be an artifact of this model.
The red dots in the figure show the \textit{toroidal peaking} for the \textit{resonant} case observed in the experiment.
The agreement with the model is remarkable, keeping in mind the simplifications made for this model.
The experimental data for the \textit{non-resonant} case does not allow to make the same comparison, which is due to the low expected variation and the influence of noise in the measurement.
In~\fref{fig:ExperimentTimeVariationNon} it is seen that the experimental \textit{toroidal peaking} at a normalized current of 0.6 is well below the 1.5 observed in the model.
\FloatBarrier
\subsubsection{Influence of the Divertor Broadening $S$}
In section~\ref{densSteps} it is shown that the \textit{toroidal peaking} decreases with increasing density.
From the axisymmetric reference as well as the toroidally averaged profiles it is observed that both transport qualifiers, $\lambda_q$ and $S$, increase with increasing density.
In the presented model the divertor broadening $S$ is applied after the heat flux calculations and thus can be varied independently.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{Modelling_PeakingVsS_toroidalSep.pdf}
\caption{\textit{Toroidal Peaking} in dependence of the divertor broadening $S$ one $\lambda_q$ away from the strike line.}
\label{fig:ModellingBroadening}
\end{figure}
The \textit{toroidal peaking} decreases with increasing divertor broadening $S$.
The decrease seen in the experiment with increasing density is explained by the increase of the divertor broadening $S$.
The measured \textit{toroidal peaking} at the fitted divertor broadening $S$ for the three different density steps are plotted in red.
The power fall-off length $\lambda_q$ is kept constant in the modelling, although it changes in the experiment with density.
\FloatBarrier
\section{Summary and Conclusions} \label{Conclusion}
The effect of external magnetic perturbation on the divertor heat load is studied in ASDEX Upgrade L-Mode.
ASDEX Upgrade is able to measure the 2D heat flux profile with n\,=\,2 and various poloidal phasing due to the versatile power supply of the saddle coils.
A clear change of the divertor heat flux with the poloidal phasing (\textit{resonant} vs. \textit{non-resonant}) is observed that is in agreement with modelling using the vacuum field approach and field line tracing from the outer divertor target to the outer midplane combined with the two point model.\\
The time averaged heat flux profiles are similar to the axisymmetric reference profiles without MP leading to the same transport qualifiers, power fall-off length $\lambda_q$ and divertor broadening $S$, for both the measurements and the model for all \textit{differential phases}.
No change in the heat transport is observed.\\
The peak heat flux is unchanged and at the same location for all toroidal phases with magnetic perturbation.
This is the same location than in the reference phase without magnetic perturbation.
Although the toroidal averaged heat flux is unchanged, the application of the magnetic perturbation has an affect onto the local heat flux.\\
The \textit{toroidal peaking} is largest for the \textit{resonant} configuration and at lowest density with up to a factor of 2 locally increased heat flux.
The variation decreases with shifting the \textit{differential phase} away from the \textit{resonant} configuration and is for the \textit{non-resonant} configuration within the typical heat flux variation in L-Mode.
Increasing the density increases the divertor broadening $S$ for the outer target of ASDEX Upgrade.
Increasing the density leads to a reduced \textit{toroidal peaking} and a nearly axisymmetric profile in still attached conditions for the discharge parameters used in the presented study.
The reduction of the \textit{toroidal peaking} is explained in the model with the increase of the divertor broadening $S$ solely and is in quantitative agreement with the measurements.
The model suggests a linear increase of the \textit{toroidal peaking} with perturbation strength, obtained by an increase of the current in the saddle coils.\\
The overall agreement between the measurements and the model leads to the conclusion that in these L-Mode conditions plasma response is not a dominant factor for the heat transport.
However, in the foreseen H-Mode regime for ITER this might change.
The influence of this response onto the heat transport in the scrape-off layer is up to now unknown and outside of the scope of the presented study.
The 2D pattern might be less or more extended.
Less due to the more narrow power fall-off length $\lambda_q$ and possible shielding or more due to field amplification from the plasma.
The effect of the divertor broadening $S$ onto the \textit{toroidal peaking} should be similar in H-Mode compared to the presented results in L-Mode.
\section*{Acknowledgements}
This work has been carried out within the framework of the EUROfusion Consortium and has received funding from the Euratom research and training programme 2014-2018 under grant agreement No 633053. The views and opinions expressed herein do not necessarily reflect those of the European Commission.
\section*{References}
\bibliographystyle{nf}
|
\section{Introduction}
Cavity quantum electrodynamics (CQED) \cite{Haroche_24} is the study of the
interaction between matter and photons confined in a cavity. In the
Jaynes-Cummings model \cite{Shore_1195} the matter is described using the
two-level approximation, and only a single cavity mode is taken into account.
The interaction has a relatively large effect on the cavity mode response when
the ratio $E/\hbar\omega_{\mathrm{a}}$ between the energy gap $E$ separating
the two levels and the cavity mode photon energy $\hbar\omega_{\mathrm{a}}$ is
tuned close to unity. Recently, it was experimentally found that the cavity
response exhibits higher order resonances in the nonlinear regime when the
ratio $E/\hbar\omega_{\mathrm{a}}$ is tuned close to an integer value larger
than unity \cite{Buks_033807}
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.3726in,
width=3.4537in
{Fig_device.eps
\caption{The device is made of two $40\operatorname{mm}\times
40\operatorname{mm}\times0.5\operatorname{mm}$ sapphire wafers carrying the
radio frequency omega resonator, and a $5\operatorname{mm}\times
5\operatorname{mm}\times0.5\operatorname{mm}$ silicon wafer carrying the
microwave spiral resonator. The DPPH powder is placed between the omega
inductor and the spiral. The 3 wafers are vertically shifted in the sketch for
clarity. In the assembled device both the top sapphire wafer and the silicon
wafer are placed directly on top of the bottom sapphire wafer. The three
wafers and a loop antenna are assembled together inside a package made of high
conductivity oxygen free copper. Both omega and spiral resonators are made by
DC-magnetron sputtering of a $200\operatorname{nm}$ thick niobium layer. The
radius of the omega inductor is $500\operatorname{\mu m}$ and the linewidth is
$40\operatorname{\mu m}$. The spiral dimensions are: inner radius
$500\operatorname{\mu m}$, outer radius $580\operatorname{\mu m}$, linewidth
$10\operatorname{\mu m}$ and number of turns $4$. The measured frequency of
the omega (spiral) resonator is $\omega_{\mathrm{a}}/2\pi
=0.173\operatorname{GHz}$ ($\omega_{\mathrm{b}}/2\pi=2.00\operatorname{GHz}$),
whereas the value obtained from numerically simulating the structure is
$0.176\operatorname{GHz}$ ($2.07\operatorname{GHz}$).
\label{Fig device
\end{center}
\end{figure}
In the current study we explore the case where $E/\hbar\omega_{\mathrm{a}
\gg1$ \cite{Ates_724}. This is done by investigating the interaction between
an ensemble of spins and a superconducting cavity mode
\cite{Ghirri_063855,Yap_62,Ghirri_184101}. The energy separation between the
spin energy eigenstates, which is given by $E=\hbar\omega_{\mathrm{L}}$, where
$\omega_{\mathrm{L}}$ is the Larmor frequency, is tuned to a value much higher
than the cavity mode photon energy $\hbar\omega_{\mathrm{a}}$. For this case
the CQED interaction is expected to be negligibly small in the regime of weak
driving. On the other hand, with an intense driving at an angular frequency
close to $\omega_{\mathrm{L}}$ we observe a significant change in the cavity
mode response.
In the current experiment the cavity mode effective damping rate is measured
as a function of the spin driving amplitude and detuning frequency. The
observed shift in the effective damping rate is attributed to the retarded
response of the cavity mode to the driven spins. Related effects of Sisyphus
cooling, amplification, lasing and self-excited oscillation have been
theoretically predicted in other systems having a similar retarded response
\cite{Glenn_195454,Grajcar_612,Voogd_1508_07972,Ella_1210_6902,Ramos_193602}.
\section{Experiment}
Significant change in the response of the measured cavity mode of angular
frequency $\omega_{\mathrm{a}}$ is possible only when intense driving is
applied to the spins. In order to allow sufficiently strong driving, the spin
ensemble is coupled to an additional cavity mode having angular frequency
$\omega_{\mathrm{b}}\gg\omega_{\mathrm{a}}$. When the Larmor frequency
$\omega_{\mathrm{L}}$ is tuned to a value close to $\omega_{\mathrm{b}}$, the
additional cavity mode allows enhancing the spin driving amplitude.
A sketch of the device is seen in Fig. \ref{Fig device}. It is made of two
sapphire wafers and a high resistivity silicon wafer that are attached
together to form a dual band resonator. A radio frequency resonator of angular
frequency $\omega_{\mathrm{a}}$ is constructed by integrating an inductor in
the shape of the greek letter $\Omega$ \cite{Twig_104703} made on the bottom
sapphire wafer, and two capacitors in series, which are formed between the two
sapphire wafers. A square hole is made in the upper sapphire wafer in order to
allow inserting the silicon wafer, which carries a spiral shaped microwave
resonator having angular frequency $\omega_{\mathrm{b}}$
\cite{Maleeva_474,Maleeva_064910}.
Both the resonators are designed to be efficiently coupled to the spin
ensemble of diphenylpicrylhydrazyl (DPPH) powder, placed in between them. This
radical, which contains three benzene rings, has a single unpaired electron,
which gives rise to Land\'{e} g-factor of $2.0036$
\cite{Kaplan_1182,Lloyd_1576}. A sketch of the experimental setup is seen in
Fig. \ref{Fig_setup}. A loop antenna is employed for delivering input and
output signals to both resonators
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=1.7591in,
width=1.6337in
{Fig_setup.eps
\caption{The experimental setup. A power combiner (PC) is employed for
combining the injected signals of a signal generator (SG) and a network
analyzer (NA). The combined injected signal is transmitted through an
amplifier (A) and a coupler (C), and feeds the loop antenna (LA), which is
positioned above the device under study (DUS). The back-reflected signal is
splitted by a power splitter (PS) and measured by both a NA and a spectrum
analyzer (SA).
\label{Fig_setup
\end{center}
\end{figure}
The measured reflectivity near the electron spin resonance (ESR) of the omega
and spiral resonators is seen in Fig. \ref{Fig ESR} (a) and (b), respectively
\cite{Schuster_140501}. Fitting the data with theory [e.g. Eq. (4) of Ref.
\cite{Buks_033807}] allows extracting the value of the coupling coefficient
$g_{\mathrm{a}}$ ($g_{\mathrm{b}}$), which characterizes the interaction
between the spin ensemble and the omega (spiral) resonator, and which is found
to be $g_{\mathrm{a}}=1
\operatorname{MHz
$ ($g_{\mathrm{b}}=8
\operatorname{MHz
$)
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.6742in,
width=3.4546in
{Fig_EPS.eps
\caption{ESR of (a) the omega and (b) the spiral resonators. The color coded
plots display the measured reflectivity coefficient $\left\vert S_{11
\right\vert ^{2}$ vs. $\omega_{\mathrm{L}}$ (i.e. vs. static magnetic field)
and the probing frequency $\omega_{\mathrm{NA}}$. Measurements are performed
by a network analyzer at a temperature of $T=3.1\operatorname{K}$, for which
the polarization coefficient $p_{0}$ [see Eq. (\ref{p_0})] is given by
$p_{0}=-1.\,4\times10^{-3}$ ($p_{0}=-1.6\times10^{-2}$) for the omega (spiral)
resonator.
\label{Fig ESR
\end{center}
\end{figure}
The linear response of the decoupled omega resonator is characterized by a
complex angular frequency given by $\omega_{\mathrm{a}}-i\gamma_{\mathrm{a}}$,
where $\gamma_{\mathrm{a}}$ is the mode damping rate. The effect of coupling
on the linear response of the mode can be described in terms of an effective
complex angular frequency $\Omega_{\mathrm{a}}=\omega_{\mathrm{a}
-i\gamma_{\mathrm{a}}+\Upsilon_{\mathrm{a}}$, where $\Upsilon_{\mathrm{a}}$
represents the coupling induced frequency shift. The complex angular frequency
$\Omega_{\mathrm{a}}$ can be extracted from the lineshape of the measured
cavity reflectivity vs. frequency curves. The change in the damping rate
$-\operatorname{Im}\Upsilon_{\mathrm{a}}$ is seen in the color-coded plots of
Fig. \ref{Fig gamma_ba} as a function of the Larmor frequency $\omega
_{\mathrm{L}}$ and the spin driving angular frequency $\omega_{\mathrm{p}}$
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.6742in,
width=3.4546in
{Fig_gamma_ba.eps
\caption{The measured (a) and calculated (b) change in the damping rate
$-\operatorname{Im}\Upsilon_{\mathrm{a}}$ vs. $\omega_{\mathrm{L}}$ and the
pump frequency $\omega_{\mathrm{p}}$. The experimental value is obtained from
the lineshape of the omega resonance. The Larmor frequency $\omega
_{\mathrm{L}}$ is tuned by applying a static magnetic field in a direction
parallel to the wafers. The pump power is set to the value $17{}\mathrm{dBm}$,
which corresponds to a driving amplitude of $\omega_{1}/2\pi
=12\operatorname{MHz}$. The calculated shift (b) is based on Eqs.
(\ref{Upsilon_aL}) and (\ref{Upsilon_ab}). The following parameters are used
in the calculation $\gamma_{\mathrm{b}}=0.4\operatorname{MHz}$ and $\gamma
_{2}=8.3\operatorname{MHz}$ (other parameters are specified above).
\label{Fig gamma_ba
\end{center}
\end{figure}
\section{Theory}
To account for the experimental findings, two possible contributions to
$\Upsilon_{\mathrm{a}}$, which is expressed as $\Upsilon_{\mathrm{a}
=\Upsilon_{\mathrm{aL}}+\Upsilon_{\mathrm{ab}}$, have been theoretically
estimated. While $\Upsilon_{\mathrm{aL}}$ represents the shift induced by the
coupling to the driven spins, the $\Upsilon_{\mathrm{ab}}$ contribution
originates from the coupling to the driven spiral mode.
A magnetic field having two mutually orthogonal components, a static component
and an alternating one at an angular frequency $\omega_{\mathrm{p}}$, is
applied to the spin ensemble. The amplitude of the static (alternating)
component is $\gamma_{\mathrm{g}}^{-1}\omega_{\mathrm{L}}$ ($\gamma
_{\mathrm{g}}^{-1}\omega_{1}$), where $\gamma_{\mathrm{g}}$ is the electron
spin gyromagnetic ratio. The frequency shift $\Upsilon_{\mathrm{aL}}$ is found
to be given by [see appendix A and Eq. (\ref{Lambda_1= V3})
\begin{equation}
\Upsilon_{\mathrm{aL}}=\frac{\frac{\frac{8g_{\mathrm{a}}^{2}\omega_{1}^{2
}{\omega_{\mathrm{a}}^{2}\gamma_{2}}\frac{\Delta_{\mathrm{pL}}}{\gamma_{2
}\left( i-\frac{2\gamma_{2}}{\omega_{\mathrm{a}}}\right) }{1+\frac
{\Delta_{\mathrm{pL}}^{2}}{\gamma_{2}^{2}}+\frac{4\omega_{1}^{2}}{\gamma
_{1}\gamma_{2}}}p_{0}}{\frac{\gamma_{1}}{\omega_{\mathrm{a}}}\left(
\frac{\omega_{\mathrm{R}}^{2}+\eta\omega_{\mathrm{a}}^{2}}{\omega_{\mathrm{a
}^{2}}-1\right) -i\left( \frac{\omega_{\mathrm{R}}^{2}}{\omega_{\mathrm{a
}^{2}}-1\right) }\;, \label{Upsilon_aL
\end{equation}
where $\Delta_{\mathrm{pL}}=\omega_{\mathrm{p}}-\omega_{\mathrm{L}}$ is the
detuning, $\gamma_{1}$ ($\gamma_{2}$) is the longitudinal (transverse) spin
relaxation rate, $p_{0}$ is the spin polarization in thermal equilibrium [see
Eq. (\ref{p_0})], $\omega_{\mathrm{R}}=\sqrt{4\omega_{1}^{2}+\Delta
_{\mathrm{pL}}^{2}}$ is the Rabi frequency of the driven spins and $\eta$ is
given by $\eta=\left( 2\gamma_{2}/\gamma_{1}\right) \left[ 2\omega_{1
^{2}\left( 1-\gamma_{1}/\gamma_{2}\right) /\omega_{\mathrm{a}}^{2}-1\right]
$ [see Eq. (\ref{eta=})]. Note that Eq. (\ref{Upsilon_aL}) is obtained by
assuming that $\left\vert \Delta_{\mathrm{pL}}\right\vert \ll\omega
_{\mathrm{L}}$, $\gamma_{\mathrm{a}}\ll\omega_{\mathrm{a}}$ and $\gamma
_{1},\gamma_{2}\ll\omega_{\mathrm{a}}$.
The real part of $\Upsilon_{\mathrm{aL}}$ is the cavity mode angular frequency
change, induced by the coupling to the driven spins. The imaginary part of
$\Upsilon_{\mathrm{aL}}$ is related to the induced damping rate change
$\gamma_{\mathrm{aL}}$ by $\gamma_{\mathrm{aL}}=-\operatorname{Im
\Upsilon_{\mathrm{aL}}$. The color coded plot in Fig. \ref{Fig gamma_aL}(a)
exhibits the dependence of the normalized change in damping rate
$\gamma_{\mathrm{aL}}/\omega_{\mathrm{a}}$ on the normalized detuning
$\Delta_{\mathrm{pL}}/\omega_{\mathrm{a}}$ and the normalized driving
amplitude $\omega_{1}/\omega_{\mathrm{a}}$. When the driving is red detuned ,
i.e. when $\Delta_{\mathrm{pL}}$ is negative, the change in damping rate
$\gamma_{\mathrm{aL}}$ is positive, and consequently mode cooling is expected
to occur \cite{Aspelmeyer_1391}. The opposite behavior occurs with blue
detuning, i.e. when $\Delta_{\mathrm{pL}}$ is positive. For both cases, large
change in the effective cavity mode damping rate occurs near the overlaid
dotted line in Fig. \ref{Fig gamma_aL}(a), along which the Rabi frequency
$\omega_{\mathrm{R}}$ coincides with the cavity mode frequency $\omega
_{\mathrm{a}}$, i.e. $\Delta_{\mathrm{pL}}=\pm\sqrt{\omega_{\mathrm{a}
^{2}-4\omega_{1}^{2}}$. This behavior can be attributed to the fact that along
the dotted line, i.e. when $\omega_{\mathrm{R}}=\omega_{\mathrm{a}}$, the
imaginary part of the denominator of Eq. (\ref{Upsilon_aL}) vanishes, and
consequently $\left\vert \Upsilon_{\mathrm{aL}}\right\vert $ peaks. The
largest change in damping rate, which is denoted by $\gamma_{\mathrm{aL,\max
}$, can be evaluated by analyzing the expression given by Eq.
(\ref{Upsilon_aL}). In the absence of spin dephasing, i.e. when $\gamma
_{1}/\gamma_{2}=2$, it is found that the largest change, which is given by
$\gamma_{\mathrm{aL,\max}}\simeq0.437\times g_{\mathrm{a}}^{2}p_{0}/\gamma
_{2}$, occurs at the points $\left( \Delta_{\mathrm{pL}}/\omega_{\mathrm{a
},\omega_{1}/\omega_{\mathrm{a}}\right) \simeq\left( \pm0.527,0.425\right)
$, which are labeled by crosses in Fig. \ref{Fig gamma_aL}(a). In the current
experiment, however, these points are not accessible since $\omega_{1
\ll\omega_{\mathrm{a}}$
\begin{figure}
[ptb]
\begin{center}
\includegraphics[
height=2.6862in,
width=3.4546in
{gamma_aL.eps
\caption{The contribution to cavity mode damping rate $\gamma_{\mathrm{aL}}$
due to coupling to the driven spins. (a) The normalized contribution
$\gamma_{\mathrm{aL}}/\omega_{\mathrm{a}}$ vs. normalized detuning
$\Delta_{\mathrm{pL}}/\omega_{\mathrm{a}}$ and normalized driving amplitude
$\omega_{1}/\omega_{\mathrm{a}}$. The calculation is based on Eq.
(\ref{Upsilon_aL}) with the following assumed parameters $\gamma_{1
/\omega_{\mathrm{a}}=2\gamma_{2}/\omega_{\mathrm{a}}=0.05$, $g_{\mathrm{a
}/\omega_{\mathrm{a}}=0.1$ and $p_{0}=-0.1$. (b) Normalized spin polarization
$-p_{z}/p_{0}$ vs. cavity mode amplitude $x_{\mathrm{a}}$ for the case of blue
detuned driving. The black solid line represents the steady state normalized
spin polarization $-p_{z0}/p_{0}$. Retardation in the response of the spins to
periodic oscillation of $x_{\mathrm{a}}$ is illustrated by the blue closed
curved.
\label{Fig gamma_aL
\end{center}
\end{figure}
The underlying mechanism responsible for the change in the effective cavity
mode damping rate that is induced by the coupling to the driven spins can be
described as follows \cite{Aspelmeyer_1391}. As can be seen from the equation
of motion for the cavity mode (\ref{eom a}), the coupling gives rise to a
forcing term acting on the cavity mode, which is proportional to the spin
polarization that is denoted by $p_{z}$. On the other hand, as can be seen
from Eq. (\ref{eom p_+}) below, the same coupling effectively shifts the
Larmor frequency of the spins, and consequently the effective spin driving
detuning, which is given by $\Delta_{\mathrm{pL,eff}}=\Delta_{\mathrm{pL
}+g_{\mathrm{a}}x_{\mathrm{a}}$, becomes dependent on the cavity mode
amplitude $x_{\mathrm{a}}$ [see Eq. (\ref{x_a})]. For any fixed value of the
cavity mode amplitude $x_{\mathrm{a}}$ the spin polarization $p_{z}$ in steady
state, which is denoted by $p_{z0}$, can be calculated using Eq. (\ref{p_z0})
below. The dependence of $p_{z0}$ on $x_{\mathrm{a}}$ is demonstrated by the
solid black line in Fig. \ref{Fig gamma_aL}(b) for the case of blue detuned
spin driving. Consider first the adiabatic case, for which it is assumed that
$\omega_{\mathrm{a}}\ll\gamma_{1,2}$. For this case the dynamics of the cavity
mode is assumed to be relatively slow, and consequently the spin polarization
$p_{z}$ is expected to remain very close to the steady state value given by
$p_{z0}$, i.e. to adiabatically follow the $x_{\mathrm{a}}$ dependent
instantaneous steady state value. Therefore, no change in the cavity mode
damping rate is expected in the adiabatic limit. This behavior is consistent
with the fact that $\operatorname{Im}\Upsilon_{\mathrm{aL}}\rightarrow0$ in
the limit $\omega_{\mathrm{a}}/\gamma_{1,2}\rightarrow0$. Note, however, that
the expression given by Eq. (\ref{Upsilon_aL}) is not valid in the adiabatic limit.
Large deviation between the momentary polarization $p_{z}$ and the steady
state value $p_{z0}$ is possible in the non-adiabatic case, where the response
of the spins to the time evolution of the cavity mode becomes retarded. The
closed curve in Fig. \ref{Fig gamma_aL}(b) represents the periodic time
evolution of $p_{z}$ for the case where the cavity mode oscillates at a fixed
amplitude at its resonance frequency around the point $x_{\mathrm{a}}=0$.
Since $p_{z}$ is proportional to the force acting on the cavity mode, the area
colored in gray in Fig. \ref{Fig gamma_aL}(b) is proportional to the net work
done on the cavity mode per cycle. While the area is positive for the case of
blue detuning, which is demonstrated by Fig. \ref{Fig gamma_aL}(b), red
detuning gives rise to negative values, i.e. to energy flow away from the
cavity mode. These affects of energy flow between the cavity mode and the
driven spins give rise to the above discussed change in the effective cavity
mode damping rate.
The frequency shift due to the driven spiral mode is attributed to an
intermode coupling term in the Hamiltonian of the coupled system, which is
assumed to be given by $K\left( A_{\mathrm{a}}+A_{\mathrm{a}}^{\dag}\right)
\left( A_{\mathrm{b}}+A_{\mathrm{b}}^{\dag}\right) ^{2}$, where
$A_{\mathrm{a}}$ ($A_{\mathrm{b}}$) is an annihilation operator of the omega
(spiral) resonator, and $K$ is the intermode coupling coefficient [see
Eq.(\ref{H_ab})]. The contribution $\Upsilon_{\mathrm{ab}}$ is found to be
given by [see appendix B and Eqs. (\ref{CT1}) and (\ref{CT2})
\begin{align}
\Upsilon_{\mathrm{ab}} & =\frac{4K^{2}\left\vert F_{\mathrm{bf}}\right\vert
^{2}}{\omega_{\mathrm{D}}^{2}+\gamma_{\mathrm{b}}^{2}}\nonumber\\
& \left( \frac{\frac{\omega_{\mathrm{D}}}{\gamma_{\mathrm{b}}^{2}}}{\left(
\frac{i\left( \omega_{\mathrm{a}}-\omega_{\mathrm{D}}\right)
{\gamma_{\mathrm{b}}}-1\right) \left( \frac{i\left( \omega_{\mathrm{a
}+\omega_{\mathrm{D}}\right) }{\gamma_{\mathrm{b}}}-1\right) }+\frac
{1+\frac{i\gamma_{\mathrm{b}}}{\omega_{\mathrm{s}}}}{\omega_{\mathrm{s}
}\right) \;,\nonumber\\
& \label{Upsilon_ab
\end{align}
where $F_{\mathrm{bf}}$ and $\omega_{\mathrm{D}}$ are the amplitude and
angular frequency detuning, respectively, of the spiral mode driving,
$\omega_{\mathrm{b}}$ and $\gamma_{\mathrm{b}}$ are the spiral mode angular
frequency and damping rate, respectively, and $\omega_{\mathrm{s}
=2\omega_{\mathrm{b}}-\omega_{\mathrm{a}}$. Note that when $\gamma
_{\mathrm{b}}\ll\omega_{\mathrm{a}}$ and $\gamma_{\mathrm{b}}\ll
\omega_{\mathrm{s}}$, the first term in the second row of Eq.
(\ref{Upsilon_ab}) becomes negligibly small provided that $\left\vert
\omega_{\mathrm{D}}\right\vert \ll\omega_{\mathrm{a}}^{2}/\omega_{\mathrm{s}}$.
\section{Discussion}
As can be seen from the comparison between Fig. \ref{Fig gamma_ba}(a) and Fig.
\ref{Fig gamma_ba}(b), fair agreement is obtained between data and theory.
Reasonable agreement cannot be obtained unless both contributions
$\Upsilon_{\mathrm{aL}}$ [Eq. (\ref{Upsilon_aL})] and $\Upsilon_{\mathrm{ab}}$
[Eq. (\ref{Upsilon_ab})] are taken into account. The contribution of
$\Upsilon_{\mathrm{ab}}$ is dominated by the second term in the second row of
Eq. (\ref{Upsilon_ab}).
Our results demonstrate the ability to modify the effective damping rate of a
cavity mode by driving spins that are coupled to the mode. Red detuned driving
provides a positive contribution to the damping rate, whereas negative
contribution can be obtained by blue detuned driving. For the former case this
effect can be utilized for cooling down a cavity mode, while the later case of
blue detuning may allow the self excitation of oscillation. Operating close to
the threshold of self-excited oscillation, i.e. close to the point where the
total effective damping vanishes, may be useful for some sensing applications,
since the system is expected to become highly responsive to external
perturbations near the threshold.
As was shown above, relatively large change in the damping rate can be induced
provided that the Rabi frequency $\omega_{\mathrm{R}}$ of the driven spins
becomes comparable to the cavity mode frequency $\omega_{\mathrm{a}}$ (see
Fig. \ref{Fig gamma_aL}). Unfortunately, this region is inaccessible with the
devices that have been investigated in the current experiment. However, in
other CQED systems the condition $\omega_{\mathrm{R}}\simeq\omega_{\mathrm{a
}$ can be more easily satisfied. For example, with superconducting CQED
systems both strong \cite{Wallraff_162,Houck_080502,Blais_062320,Koch_042319
\ and ultra-strong \cite{Niemczyk_772,Forn_237001} coupling is possible . This
together with the ability to drive a Josephson qubit with Rabi frequencies
high in the radio frequency band, may allow satisfying the condition
$\omega_{\mathrm{R}}\simeq\omega_{\mathrm{a}}$ with a strongly coupled cavity
mode. As was shown above, a large change in cavity mode damping rate, on the
order of $g_{\mathrm{a}}^{2}\left\vert p_{0}\right\vert /\gamma_{2}$, is
possible provided that the region where $\omega_{\mathrm{R}}\simeq
\omega_{\mathrm{a}}$ becomes accessible. For a typical superconducting CQED
system, the damping rate of a decoupled cavity mode is far smaller than
$g_{\mathrm{a}}^{2}\left\vert p_{0}\right\vert /\gamma_{2}$, and thus reaching
this region may allow efficiently cooling down cavity modes by off-resonance
qubit driving.
This work was supported by the Israel Science Foundation, the bi-national
science foundation, the Security Research Foundation in the Technion and the
Russell Berrie Nanotechnology Institute. HW acknowledges support by the Shatz
fellowships and by the Viterbi fellowships.
|
\section{introduction}
Spin-orbit coupling (SOC) plays an increasingly important role in modern condensed matter physics. It is essential in many topological phenomena, such as the spin Hall effect\cite{sinova-2015}, the quantum spin Hall effect\cite{maciejko-2011} and topological insulators\cite{hasan-2010,qi-2011}. In recent years, SOC has also attracted much interest in optics\cite{bliokh-2015}. Due to the intrinsic SOC of light, it has been realized that the spin Hall effect naturally occurs in optical systems when there is a gradient in the refractive index, resembling an electric field in electronic systems. As a result, at the interface between two media where the refractive index varies, when transmitted or reflected, a light beam has a transverse shift with the direction depending on the chirality (spin) of the photons\cite{onoda-2004,bliokh-2006}. The transverse shift is named the Imbert-Fedorov (IF) shift after its discoverers\cite{fedorov-1955,imbert-1972}, and has been verified by experiments\cite{hosten-2008,yin-2013}.
In recent works\cite{xie-2015,yang-2015b}, the authors have found that the IF shift also occurs in Weyl semimetals (WSMs). WSMs are recently discovered materials\cite{lv-2015a,xu-2015a,yang-2015a,lv-2015b,xu-2015b}. They host Weyl points, which behave as monopoles of Berry flux. Around the Weyl points, the states are described by the Weyl equation\cite{weyl-1929} and have a linear dispersion\cite{wan-2011,volovik-2009,xu-2011,burkov-2011,yang-2011,halasz-2012,zhang-2014,liu-2014,weng-2015,huang-2015,hirayama-2015}. The IF shift in WSMs is due to the intrinsic coupling between the pseudospin and the orbital degree of freedom. It has been shown that the IF shift depends on the monopole charge (chirality) of the Weyl point which is $\pm1$, similar to that in optical systems.
The IF shift in both optical systems and WSMs has been interpreted semiclassically\cite{onoda-2004,yang-2015b}. The semiclassical equations of motion (EOM) govern the trajectory of wave packets, and dictate that the IF shift of a wave packet is due to its anomalous velocity\cite{xiao-2010}, and hence is an integral of the Berry curvature. If the band structure and hence the Berry curvature vary slowly, then it is an intuitive way to calculate the IF shift. However, if the Berry curvature has an abrupt change at an interface, difficulties arise. Moreover, near the Weyl point where the gap closes, non-Abelian treatment is needed\cite{yang-2015b}. As such, processes like Klein tunneling are not straightforward to account for. Different from the quantum mechanical treatment, the trajectory of a wave packet is fixed, which ignores the possibility that the wave packet may split. Another interpretation of the IF shift is from the conservation of total angular momentum (TAM)\cite{onoda-2004,yang-2015b}, which only works if the two media have the same monopole charge, and the composite system has a rotational symmetry. Therefore, there is a limit in these two methods. In WSMs, the quantum mechanical approach has been applied to the calculation of the IF shift\cite{xie-2015}. It works best if an abrupt interface exists between two WSMs, regardless of the symmetry of the system. Each of the three approaches has its own advantages and drawbacks.
Multi-WSMs that host Weyl points with monopole charge $\pm2$ and $\pm3$ have been discovered\cite{fang-2012}, which are protected by point group symmetries. Namely, $C_4$ and $C_6$ symmetries can protect double-Weyl points while $C_6$ symmetry can protect triple-Weyl points. Therefore, it is natural to extend previous works on the IF shift in single-WSMs to multiple WSMs. In this work, we study the IF shift between WSMs with monopoles of arbitrary integer charges. We apply the quantum mechanical approach to study the problem, and then use the semiclassical approach and the conservation of TAM to confirm and interpret the results in special cases. In addition, we consider the possibility that the Weyl fermions are reflected to another Weyl cone with opposite chirality, in which case the IF shift vanishes under certain symmetry conditions.
\section{IF shift at the interface between WSMs with arbitrary monopole charges}
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{reflection}
\caption{A beam of Weyl fermions is incident from a WSM at $z<0$, with monopole charge $N_1$, to another at $z>0$, with monopole charge $N_2$, and is totally reflected at the interface $z=0$. A step potential is assumed between the two media. During the reflection, the IF shift occurs. }\label{fig:refl}
\end{figure}
The Hamiltonian for a Weyl cone with a positive monopole charge $N$ can be written as
\be
H_0(\bk)=\left(\begin{array}{cc}k_z&(k_x-ik_y)^N\\(k_x+ik_y)^N&-k_z\end{array}\right),\label{eq:ham1}
\ee
where $\bk=(k_x,k_y,k_z)$ and $H_0^*$ carries the opposite charge $-N$. We have chosen the $z$-axis as a high symmetry axis. We set the velocities to 1 hereafter for simplicity. (Actually, the IF shift depends on the velocities\cite{xie-2015,yang-2015b}; we leave the calculation of such dependence to future work.) Note that the generalization of a single-Weyl point with monopole charge 1 to a multi-Weyl point with monopole charge $N$ is strongly reminiscent of the counterpart in two dimensions –- the generalization of a gapless Dirac point with winding number 1 to the case with winding number $N$\cite{zhang-2011}. The eigenenergies are $E=\pm\sqrt{k_\parallel^{2N}+k_z^2}$, associated with the spinor wavefunction
\begin{eqnarray}
\psi(\br) &=& \frac{e^{i(k_xx+k_yy+k_zz)}}{\sqrt{1+\eta^2}}\left(\begin{array}{c}
e^{-iN\al} \\
\eta
\end{array}\right)
\end{eqnarray}
where ${\bf r}=(x,y,z)$, $k_\parallel=\sqrt{k_x^2+k_y^2}$, $\al=\tan^{-1}{(k_y/k_x)}$ and $\eta=\sqrt{\frac{E-k_z}{E+k_z}}$.
We consider the total reflection that occurs at the interface between two WSMs. Assume a beam of Weyl fermions is incident from a WSM with monopole charge $N_1$ to that with monopole charge $N_2$, and the interface is at $z=0$, as shown in Fig.\ref{fig:refl}. A step potential is assumed at the interface, with the form $V=0$ for $z<0$ and $V=V_0>0$ for $z>0$. During the process of reflection, $k_x$ and $k_y$ are conserved, while $k_z$ is not. Using the continuity of the wavefunction at $z=0$ and solving for the reflective coefficient $r$, we get $r=e^{i\phi_r}$
with the phase (see Appendix A for details)
\begin{eqnarray}
\phi_r &=& 2\tan^{-1}\frac{\eta\sin(\phi_\xi-(N_1-N_2)\al)}{1-\eta\cos(\phi_\xi-(N_1-N_2)\al)}\nonumber\\
&+&\phi_\xi-(N_1-N_2)\alpha
\end{eqnarray}
where $\phi_\xi$=$\arg(E-V_0-i\kappa)$ with $\kappa=\sqrt{k_\parallel^{2N_2}-(E-V_0)^2}$. The total reflection only occurs when $\kappa>0$ since the wave function is proportional to $e^{-\kappa z}$ at $z>0$. We assume the incident beam is Gaussian in the $y$-direction, and the central ray lies in the $xz$ plane. Then the $k$-space distribution is also Gaussian, with the center at $\bar k_y=0$. In real space, the center of the incident beam is $N_1\partial_{k_y}\al/(1+\eta^2)|_{k_y=0}$, and the center of the reflected beam is $[(N_1\eta^2\partial_{k_y}\al)/(1+\eta^2)-\partial_{k_y}\phi_r]|_{k_y=0}$, so the IF shift is their difference. Bringing all the variables in, we find the IF shift as follows:
\begin{eqnarray}
\Delta_{IF} &=&-\frac{N_1k_z}{Ek_x}+\frac{(N_1-N_2)k_z}{Ek_x+(V_0-E)k_x^{N_1-N_2+1}}.\label{eq:DIF1}
\end{eqnarray}
Two special cases are noticed. First, if the two WSMs have the same monopole charge, i.e. $N_1=N_2\equiv N$, then $\Delta_{IF}=-N/(E\tan\theta)$, where $\theta=\tan^{-1}k_x/k_z$. Note that $\theta$ is different from the incident angle, since the velocity is not proportional to the momentum in general. (Previous studies on single-Weyl points correspond to $N=\pm1$\cite{xie-2015,yang-2015b} here.)
The IF shift changes sign once the chirality changes, which is termed as the chirality-dependent Hall effect. Here, we extend it to a general monopole charge, and find the IF shift is proportional to the monopole charge. This is like the case in the quantum Hall effect, where the quantized Hall conductance is proportional to the Chern number\cite{thouless-1982}. Second, if the energy of the incident Weyl fermions can be adjusted to $E=V_0$, then $\Delta_{IF}=-N_2/(E\tan\theta)$. Then one can detect the monopole charge of the WSM to which the Weyl fermions are incident.
One disadvantage of the quantum mechanical approach is that we need to assume a specific form of the wave packet to calculate the IF shift. In addition, we have assumed a sharp interface between the two WSMs. However, the Weyl Hamiltonian is actually a low energy, long wavelength Hamiltonian, so the matching wavefunction method only works for the case in which the interface is smooth in the atomic scale. Therefore, we need to confirm that the result obtained above is valid. To this end, we apply the semiclassical approach to calculate the IF shift where the interface is smooth.
\section{Semiclassical approach}
Analogous to the motion of electrons in condensed matter systems, the motion of optical wave packets can be studied by semiclassical EOM\cite{onoda-2004,onoda-2006}, given the condition that the modulation to the band structure is weak and slowly varying. The semiclassical approach gives a simple interpretation to the IF shift: it is the shift due to the anomalous velocity, which appears if the Berry curvature is nonvanishing and an effective electric field exists\cite{xiao-2010}. This approach is used to calculate the IF shift in WSMs in Ref.\cite{yang-2015b}, in the case where the band structure and the Berry curvature are slowly varying. Now we show that the IF shift between two WSMs with the same monopole charge $N$ can be calculated in this way, which further confirms the results obtained above.
The incident momentum of a wave packet is $(k_x,0,k_z)$, and the reflected momentum is $(k_x,0,-k_z)$. The IF shift is the integral of the anomalous velocity\cite{onoda-2004},
\begin{eqnarray}
\Delta_{IF}=\int dt[\dot{\bk}\times{\bf\Omega}(\bk)]_y=\int_{k_z}^{-k_z}dk_z \Omega_x(k_x,0,k_z),\label{eq:DIF}
\end{eqnarray}
where $\dot {\bf k}$ denotes the time derivative of ${\bf k}$, ${\bf \Omega(\bk)}$ is the Berry curvature of the band in which the wave packet resides and $[...]_i$ means the $i$-th component. We see the difference between the Chern number and the IF shift: the former is the integral of the Berry curvature over the planar Brillouin zone, while the latter is the integral of the Berry curvature over a line segment in the Brillouin zone.
The Hamiltonian Eq.(\ref{eq:ham1}) can be written as $H_0(\bk)=\bh(\bk)\cdot {\bm \sigma}$, where components of ${\bm \sigma}$ are Pauli matrices and components of $\bh$ are the coefficients of each Pauli matrix, respectively. Then the $x$-component of the Berry curvature is ${\Omega}_x(\bk)=\bh\cdot(\partial_{k_y}\bh\times\partial_{k_z}\bh)/(2h^3)$, where $h=\sqrt{ \sum_{i=1}^3 h_i^2}$. To obtain $\Omega_x(k_x,0,k_z)$, we expand $\bh(\bk)$ to the first order of $k_y$, $\bh(\bk)=(k_x^{N}, Nk_x^{N-1}k_y,k_z)+O(k_y^2)$, and then find $\Omega_x(k_x,0,k_z)=Nk_x^{2N-1}/[2(k_x^{2N}+k_z^2)^{\frac{3}{2}}]$. Plugging it into Eq.(\ref{eq:DIF}), we have $\Delta_{IF}=-N/(E\tan\theta)$, which agrees with the quantum mechanical result.
An immediate question is whether we can apply the semiclassical approach to the case where the two WSMs carry different monopole charges. A necessary condition to apply the semiclassical approach is one can interpolate the two Hamiltonians smoothly, such that the band structure and the Berry curvature vary slowly. However, since the two Hamiltonians have different topologies, i.e. their monopole charges differ by an integer, they cannot be connected smoothly. Therefore, the semiclassical approach cannot be applied to this case.
\section{Conservation of TAM}
The IF shift in optics has also been understood as the result of the conservation of the total angular momentum of individual photons\cite{onoda-2004}. In WSMs with $N=1$, it has been shown that the IF shift can be interpreted as the result of the conservation of the generalized total angular momentum, which is the sum of the orbital angular momentum and the $1/2$ pseudospin of Weyl fermions\cite{yang-2015b}. Here we generalize further this idea.
If both WSMs are described by $H_0+V$ where $H_0$ is given in Eq.(\ref{eq:ham1}) and $V$ is a step potential, i.e. both have the same monopole charge $N$, there is a continuous rotational symmetry around the $z$-axis, corresponding to the conservation of the $z$-component of the total angular momentum $\bJ$. One can show $[J_z,H]=0$, where $J_z=L_z+S_z$, with $L_z=xk_y-yk_x=-i\partial_\al$ and $S_z=\frac{N}{2}{\bf\si}_z$.
When an incident Weyl fermion is totally reflected at the $z=0$ plane, $J_z$ is conserved, i.e. $J_z^I=J_z^R$, where $I$ and $R$ label incident and reflected, respectively. Since the incidence is in the $xz$ plane, $k_y=0$, $k_x$ is conserved, and $k_z$ changes sign after the reflection. Then the IF shift is
\begin{eqnarray}
\Delta_{IF}=y^R-y^I=\frac{N}{2k_x}(\langle\si_z^R\rangle-\langle\si_z^I\rangle)=-\frac{Nk_z}{E k_x},
\end{eqnarray}
which agrees with the result calculated using quantum mechanics and semiclassical EOM. However, if the monopole charges are different at the two sides, it is not obvious that one can use this argument to calculate the IF shift.
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=4cm]{twocones}\label{fig:twocones}}
\subfigure[]{\includegraphics[width=4cm]{inter}\label{fig:inter}}
\caption{(a) Band structure of the Hamiltonian. (b) An incident particle (with fixed $k_x$) can be reflected to the same Weyl cone or the other. The red dot indicates an incident particle with momentum $k_{z}$, the orange dot indicates the case where it is reflected to the same cone with momentum $k_{z}'$, and the blue dot indicates the case where it is reflected to the other cone with momentum $-k_{z}$.}
\end{figure}
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=4cm]{r1_1}\label{fig:r11}}
\subfigure[]{\includegraphics[width=4cm]{r1_2}\label{fig:r12}}
\caption{Intervalley reflectance $|r_1|^2$ with the energy of the incident Weyl fermions at $E=k_0^2$ (a) and at $E=0.5k_0^2$ (b), where we set $k_0=1$.
}\label{fig:r}
\end{figure}
\section{Intervalley scattering}
Previous discussions of the IF shift in WSMs\cite{xie-2015,yang-2015b} have been restricted to the case where the reflected Weyl fermions are in the same valley as the incident ones. However, since WSMs always hold an even number of Weyl points\cite{nielsen-1983}, there is a certain probability that in total reflection, part of the beam is reflected to another valley. We study this case using a model Hamiltonian with two Weyl cones of monopole charge $N=\pm 1$ located at $(0,0,\pm k_0)$, where $k_0$ is a constant,
\begin{eqnarray}
H(\bk) &=& \left(\begin{array}{cc}
k_z^2-k_0^2 & k_x-ik_y\\
k_x+ik_y & -(k_z^2-k_0^2)
\end{array}\right).
\end{eqnarray}
The dispersion is shown in Fig.\ref{fig:twocones}, and the eigenfunction is
\begin{eqnarray}
\psi(\br) &=& \frac{1}{\sqrt{1+\eta^2}}\left(\begin{array}{c}
e^{-i\al} \\
\eta
\end{array}\right)e^{i(k_xx+k_yy+k_zz)}
\end{eqnarray}
where $k_z^2=\sqrt{E^2-k_x^2-k_y^2}+k_0^2$, $\eta=\sqrt{\frac{E-k_z^2+k_0^2}{E+k_z^2-k_0^2}}$ and $\al$ is defined the same way as above.
The setting is the same as in Fig.\ref{fig:refl} except that two valleys are considered. As shown in Fig.\ref{fig:inter}, the incident wave has $z$-momentum $k_z$ (red point), and by energy conservation, the $z$-momentum of the reflected wave is either $-k_z$ (blue point) or $k_z'=\sqrt{2k_0^2-k_z^2}$ (orange point), corresponding to the intervalley and intravalley scattering.
We have the reflected wavefunction in the region $z<0$ (ignoring the factor $e^{i(k_xx+k_yy)}$)
\begin{eqnarray}
\psi_R({\bf r})= \frac{r_1 e^{-ik_zz}}{\sqrt{1+\eta^2}}\left(\begin{array}{c}
e^{-i\al} \\
\eta
\end{array}\right)
+\frac{r_2 e^{ik_z'z}}{\sqrt{1+\eta^2}}\left(\begin{array}{c}
\eta e^{-i\al} \\
1
\end{array}\right),
\end{eqnarray}
where $r_1$ and $r_2$ are the reflective coefficients associated with the intervalley and intravalley scattering, respectively.
Using the continuity of the wavefunction and of the derivative of the wave function, we can find $r_1$ and $r_2$ (see Appendix B for details). $|r_1|^2$ is the probability that the particle is reflected to the other valley, and $|r_2|^2|v_2|/v_1$ is the probability that it is reflected to the same valley, where $v_1>0$ is the $z$-component of the incident velocity and $-v_1$ and $v_2$ are the $z$-component of the velocities of particles that are reflected to the other and the same valley, respectively. The latter has a factor of $|v_2|/v_1$ because of the conservation of the probability current $v_1=|r_1|^2v_1+|r_2|^2|v_2|$.
We plot $|r_1|^2$ for two energies, $E=k_0^2$ and $E=0.5k_0^2$ in Fig.\ref{fig:r}. The parameters $0<V_0<2E$ and $|E-V_0|<k_x<E$ so that total reflection occurs. $|r_1|^2$ strongly depends on the energy of the particles, since the potential barrier is shallower for higher energy. We see that at smaller incident angles (smaller $k_x$), there is a larger probability of reflecting to the other valley; as $V_0\to E$, $|r_1|^2$ reaches the maximum at $k_x\to 0$. The maximum can be nearly $100\%$ when the energy of the incident particles is high, but decreases to a few percent if the energy decreases to its half. Therefore, the intervalley scattering is weak if the energy is well below the potential barrier, but in general the beam splits into two.
To calculate the shift, we find the initial center of beam is $\partial_{k_y}\al/(1+\eta^2)|_{k_y=0}$, while the final center for the two reflected beams are, respectively, $[\partial_{k_y}\al/(1+\eta^2)-\partial_{k_y}\phi_{r_1}]|_{k_y=0}$ and $[\eta^2\partial_{k_y}\al/(1+\eta^2)-\partial_{k_y}\phi_{r_2}]|_{k_y=0}$. So the shift for the two beams corresponding to intervalley and intravalley reflection are given by
\begin{eqnarray}
\Delta_1 &=& -\partial_{k_y}\phi_{r_1}|_{k_y=0},\\
\Delta_2 &=& -\partial_{k_y}\phi_{r_2}|_{k_y=0} +\frac{（\eta^2-1）}{\eta^2+1}\partial_{k_y}\al|_{k_y=0},
\end{eqnarray}
respectively. Since $k_y$ always appears as $k_y^2$ except in $\al$, we have $\partial_{k_y}\phi_{r_{1,2}}|_{k_y=0}=0$. Therefore $\Delta_1=0$, i.e., the beam reflected to the other valley has no IF shift, and the intravalley reflected beam has the IF shift $\Delta_2=-v_2(k_z-k_0)/(Ek_x)$. Note the replacement of $k_z$ by $k_z-k_0$ due to the finite momentum location of the Weyl point.
The zero IF shift can be easily understood from the semiclassical point of view. If the two Weyl points with monopole charge $\pm N$ are located at $(0,0,\pm k_{0})$, as the $z$-momentum goes from $k_z$ to $-k_z$, the trajectory in $k$-space passes both $k_{0}$ and $-k_{0}$ or none of the two. Then we have
\begin{eqnarray}
\Delta_{IF}&=&\int_{k_z}^{-k_z}dk_z\Omega_x=\int_{k_z}^{-k_z}dk_z(\Omega_x^{(1)}+\Omega_x^{(2)})=0,
\end{eqnarray}
where $\Omega_x^{(1,2)}$ are the Berry curvature contributed by the two Weyl points. Actually, the vanishing of the IF shift is due to the inversion symmetry and the rotational symmetry about the $z$-axis: the inversion symmetry gives ${\bf\Omega}(-\bk)={\bf\Omega}(\bk)$, from which $\Omega_x(-k_x,0,-k_z)=\Omega_x(k_x,0,k_z)$, while from the rotational symmetry, $\Omega_x(-k_x,0,k_z)=-\Omega_x(k_x,0,k_z)$. Combining the two symmetries, we have $\Omega_x(k_x,0,-k_z)=-\Omega_x(k_x,0,k_z)$, which says $\Omega_x$ is an odd function of $k_z$. Thus the integral vanishes, i.e., $\Delta_{IF}=0$.
\section{Discussion}
The IF shift in WSMs could be observed in the confirmed WSMs TaAs and NbP\cite{lv-2015a,xu-2015a,yang-2015a,lv-2015b,xu-2015b}. The dependence of the IF shift on intervalley scattering should be easily detected since there are many Weyl points in these materials. More specific calculations need to be carried out, following the same principles given here. The possible materials that host multi-Weyl points include HgCr$_2$Se$_4$\cite{xu-2011} and SrSi$_2$\cite{huang-2015b}, which provide platforms to observe the monopole charge dependence of the IF shift.
The discovery of WSMs inspires the realization of (multi-)Weyl points in systems besides solid state materials, such as photonic crystals\cite{lu-2013,lu-2015,chen-2016}. Around Weyl points, the states are governed by the same Weyl equation, so we expect that our formalism and results also apply to these systems. However, our two-Weyl cones model breaks time-reversal symmetry, and the two Weyl points are related by mirror symmetry. In a time-reversal invariant photonic Weyl crystal, a minimal number of four Weyl points can be realized\cite{lu-2015,wang-2015}. If an interface exists perpendicular to the axis at which two Weyl points related by time-reversal symmetry reside, intervalley scattering occurs between the two valleys which have the same chirality. In this case, the IF shift contributed by the two valleys are the same rather than opposite, so the shift is approximately twice that of the intravalley reflection. Consequently, our results may be used to detect the topology of various kinds of materials by experimentally measuring their IF shift.
\begin{acknowledgements}
We would like to thank Hong Yao for helpful discussions. This work was supported in part by the NSFC under Grant No. 11474175 at Tsinghua University (LW and SKJ).
\end{acknowledgements}
|
\section{Related Work and Background}\label{sec:relatedwork}
The ability to store, query, and reason about spatiotemporal information is increasingly essential to mobile applications~\cite{biagioni11:easytracker,blanke14:capturing,yan12:fast}. Increasing concerns about location privacy~\cite{sadeh09:understanding} have motivated researchers to explore privacy primitives associated with location sharing~\cite{guha12:koi, toch10:locaccino}, with a focus on determining what location information to share and how to share it in a way that shields the user. Furthermore, on-loading, or moving storage and processing of data onto the device, has become popular for improving user experience in mobile computing environments~\cite{han13:case, vallina12:when}. The contributions in this paper couple the idea of on-loading with the indexing of spatiotemporal data to support valuable contextual data storage of any type. In order to on-load large amounts of context, a proper adaptation of existing spatiotemporal indexing methods is necessary.
Complementary work in mobile context-awareness enables context processing on-device. ACE~\cite{nath12:ace} is a context reasoning framework that executes on top of sensors and the data streams those sensors generate. ACE attempts to reduce the overhead of context sensing by making inferences about what context streams to collect using application-level information across multiple applications. ACE's inference mechanisms only work for boolean context types (e.g., ``is the user running'') and not continuous ones like a trace of the user's spatiotemporal movement. Further, ACE assumes that the applications access only instantaneous context information and does not support historical introspection across stored context information. In contrast, a data store like {\sc \textbf{Paco}}\xspace could be layered {\em underneath} an inference engine like ACE providing a richer store of context information on which to make application decisions.
SeeMon~\cite{kang08:seemon} was an early software enabler of on-device continuous context sensing. Similar approaches to adjusting sensing tasks based on application needs and a device's current situation have resulted in a wealth of similar approaches for continuous context sensing~\cite{kansal13:latency, rachuri12:energy, zhuang10:improving}. As an exemplar, SeeMon intelligently adjusts {\em how} context is acquired, specifically by monitoring context {\em changes} instead of continuously sampling and delivering raw values. Similarly, work related to efficient trajectory sensing dictates how and when to activate various on-board sensors to accurately determine the device's location while incurring low energy overheads~\cite{kjaergaard11:energy}. CoMon~\cite{lee12:comon} extends these ideas beyond a single device and creates a framework by which co-located devices can collaboratively sense the ambient environment, selecting the best sensors to task {\em across} a network of nearby devices. Like ACE, these approaches for intelligent and efficient continuous sensing deliver the instantaneous context; in addition, they focus on selecting the ``best'' set of sensors and sensing parameters to do that. These approaches do not create an on-device contextual {\em history}. As with ACE, {\sc \textbf{Paco}}\xspace directly complements these efforts; as (spatiotemporal) context information is continuously generated by these approaches, it can then be stored efficiently and queried in the future using {\sc \textbf{Paco}}\xspace's API.
In contrast to the wealth of approaches enabling efficient context {\em sensing} on commodity smartphones, existing approaches for {\em storing} that information focus exclusively on off-loading the data to the cloud and enabling efficient and expressive access to it there. That is, existing approaches either assume the sensed context information is shipped off of the device to be stored elsewhere or they assume that the device's applications simply discard any context information that is not immediately consumed. This storage gap is exactly {\sc \textbf{Paco}}\xspace's target. We {\em on-load} the storage of acquired context information, making it available to the device's applications for {\em future} queries while reducing the energy and communication costs of offloading and giving the user direct control over the potentially personal and private spatiotemporal information.
As a starting point for exploring efficient and effective {\em on-device} storage of spatiotemporal data, it is useful to examine the approaches used in server-based systems. Storing spatial data has a rich history in image processing, geographic information systems~(GIS), and robotics. Grid-based approaches, which divide space into regions and insert data into the grid square representative of the data's location, are the most straightforward~\cite{gaede1998multidimensional}. Clever statistical approaches can optimize queries over this data. This type of approach is not well suited for data that is dynamic or data sets with ``hot spots,'' i.e., spatial areas with a high number of data points. In both cases, selecting an optimal grid size is difficult, and grid bounds may not be known {\it a priori}.
The widely used R-tree~\cite{guttman1984r} maintains a balanced structure by representing objects within minimum bounding rectangles and then creating a hierarchical (tree) representation that relates the rectangles to one another. In an R-tree, a rectangle at a node in the tree completely contains any rectangles of any of the node's descendents in the tree; the tree is structured to support queries through simple computations of intersection and containment. An R-tree is especially useful for storing objects encompassing some area, which maps naturally to the abstraction of minimum bounding rectangles. The most complexity in using an R-tree involves developing algorithms to minimize or eliminate bounding rectangle overlap~\cite{beckmann1990r}, thus avoiding worst-case query performance. Existing R-tree optimizations reduce this overlap by reinserting points~\cite{beckmann1990r} or duplicating objects~\cite{gaede1998multidimensional}. Because the R-tree is designed for on-disk storage, we use it as the foundation of {\sc \textbf{Paco}}\xspace's on-device spatiotemporal database, which must be persisted on the device to be shared among applications and to extend its lifetime through application restarts and power cycling of the device\footnote{We use the R*tree, which encompasses the discussed dynamic maintenance optimizations}.
While it is quite beneficial for on-disk storage, the R-tree is not as efficient when used for in-memory computations. The $k$-d tree~\cite{bentley1975multidimensional}, on the other hand, is very well-suited for in-memory computations, especially when all of the data items are known ahead of time, and algorithms can be employed to directly construct a balanced tree. A $k$-d tree can be thought of as a $k$-ary search tree that recursively partitions data points along $k$ coordinate axes alternating between the $k$ dimensions. {\sc \textbf{Paco}}\xspace leverages these existing storage structures by employing an R-tree as the base data structure stored on disk, complemented by a $k$-d tree of three dimensions (i.e., $k=3$) as a utility structure to support sub-queries computed in memory.
We are not the first to consider storing temporal information alongside spatial information; this is a natural extension for both the $k$-d tree and R-tree. More generally, combining spatial and temporal information is particularly prevalent in moving object databases~\cite{erwig1999spatio}. Many optimizations have been explored for improving queries across moving objects such as with the binary string prefix-matching used in~\cite{ganti2016mp}. Temporal aspects may capture a data item's relevance to the state of the database (termed {\it transaction time}) or relevance to the real physical world (termed {\it valid time})~\cite{tansel08:temporal}. We focus on the latter. In contrast to existing work in more general purpose moving object databases, PACO focuses on a data store whose data points all represent just a single moving object: the user who owns the mobile device collects the data.
In support of efficient storage and retrieval of spatiotemporal data for mobile devices, methods can be loosely categorized into those that ``index past positions,'' ``index current positions,'' and ``index current and future positions''~\cite{mokbel03:spatio}. To date, these approaches build almost exclusively on R-trees, and they are thus complementary to our work. However, existing approaches have developed centralized, heavyweight indexes that are not well suited to the lightweight, flexible, and personalizable implementations demanded by mobile devices. Again, this need for a lightweight index that can reside entirely on a user's personal mobile device is {\sc \textbf{Paco}}\xspace's goal.
{\sc \textbf{Paco}}\xspace aims to enable efficient real-time querying of spatiotemporal information by applications on the user's device. Therefore, in designing {\sc \textbf{Paco}}\xspace, it is essential that the data store does not grow too large both because of the limitations of the device and the need to support quick response to applications' queries for spatiotemporal data. To support on-line location-awareness, researchers have investigated approximate query processing~\cite{sun04:querying}, enabling lossiness in data storage~\cite{cao06:spatio, cudremauroux10:trajstore} by storing line segments comprising a trajectory. In contrast, CDR~\cite{lange08:online} uses an on-line trajectory reduction that relies on the moving objects themselves (e.g. mobile embedded sensors) to temporarily store data. These approaches are related to some methods that we use in {\sc \textbf{Paco}}\xspace to intentionally discard redundant data items, however, these approaches continue to rely on a central server or other devices capable of handling significant off-loaded storage and computation. Further, these approaches do not associate the application context with the spatiotemporal data, thus they lack the ability to use this data to respond to complex application-level queries.
\section{Benchmarking}\label{sec:benchmarking}
We performed a series of benchmarks on {\sc \textbf{Paco}}\xspace to better understand its feasibility and efficiency. To our knowledge, we are the first to explore on-loading spatiotemporal data at this scale and density; for this reason, we evaluate {\sc \textbf{Paco}}\xspace across various configurations. We also compare {\sc \textbf{Paco}}\xspace's internal structure to a single table database approach (which we call StdTable). For both internal structure options, we compare using the $k$-d tree reference tree for window computes (as discussed in Section~\ref{sec:structure}) versus not using it. Recall that the $k$-d in-memory reference tree optimizes the otherwise expensive repeated window query that computes over sub-cubes. All evaluations use mobile trace data from CRAWDAD: (1)~a set of 92 traces of 500-2000 on-foot data points, collected at a university in South Korea~\cite{crawdad:kaist} (``Peds''), which is fairly sparse, and (2)~a set of vehicular traces of taxicabs in San Francisco with 500 taxicabs over 30 days~\cite{crawdad:taxis} (``Cabs''), which is fairly dense with distinct highly populated areas.
We used Moto G 1032 Android Devices (Quad-core 1.2 GHz Cortex-A7, Qualcomm Snapdragon 400 chipset, and 1GB RAM). Determining execution times in Java can be unreliable due to the JVM's JIT compiler and variations in system behavior. We mitigate these concerns using industry guidelines\footnote{\url{http://www.ibm.com/developerworks/java/library/j-benchmark1/index.html}}. We perform all benchmarks on quiescent devices and run warmup samples on the JIT compiler before measuring execution times. Energy benchmarking uses Qualcomm's Trepn Profiler\footnote{\url{https://developer.qualcomm.com/mobile-development/increase-app-performance/trepn-profiler}}, and power measurements are reported from a recorded baseline established per test. Despite these measures, execution times and energy levels should not be taken as absolute but rather as reasonable estimates of performance and relative measures across parameter settings.
Our evaluation is framed with the following goals: (1)~maintain fast, responsive query execution for on-device spatiotemporal application queries; (2)~reduce the size of stored contextual data to manageable levels (both for supporting the first goal and for reducing {\sc \textbf{Paco}}\xspace's memory footprint on resource-constrained devices); and (3)~minimize energy consumption of {\sc \textbf{Paco}}\xspace's operations. In this section, we take the elements of {\sc \textbf{Paco}}\xspace's API in turn and evaluate them for these three goals. We also investigate some alternative internal operations for their impact on these goals.
\subsection{Smart Insert}
Recall that {\sc \textbf{Paco}}\xspace's smart insert operation trades accuracy of information for structure size. With decreasing values of {\sc InsThresh}, {\sc \textbf{Paco}}\xspace becomes increasingly selective in inserting new data points. Intuitively, the effect is a decrease in resulting PoK values since there are fewer points to influence region and point queries (i.e., there is less knowledge stored in the structure). Fig.~\ref{fig:eval-smart-insert} shows the various relationships between {\sc InsThresh}, the size of the internal structure, and the resulting PoK values. A decrease in PoK value indicates a decrease in stored data accuracy from the ground truth of storing all of the points. On this graph, the varying thickness of the lines represents the size of the data set (the thickest line belonging to the largest data set). The left axis shows the average PoK achieved as a function of reducing the {\sc InsThresh}. The right axis shows the degree of size reduction of the overall data structure as a function of reducing the {\sc InsThresh}. As an example, for the largest data set, {\sc \textbf{Paco}}\xspace achieves an almost 80\% reduction in the size of the data store, albeit at a decreased PoK of about 0.5.
Larger data sets (such as the trace of 80,000 points) tend to exhibit larger fractional size reductions while still maintaining similar changes in PoK. We empirically choose a default {\sc InsThresh} value of 0.8 as a reasonable compromise that maintains a high fidelity of the ground truth spatiotemporal information (as measured by the change in PoK) relative to the space savings (which ultimately also result in more efficient querying). At this value, most traces maintain PoK within 80\% of the original value with size reductions of an average of over 48\%. Obviously, these curves are directly representative of the particular traces evaluated, and as such, some data sets such as 5k do not follow the general trend as they vary in spacing and density of points in the trace. The aim of this portion of the evaluation is to examine real world data sets as a first effort in empirically establishing guidelines for setting configuration parameters.
\begin{figure}[!t]
\includegraphics[width=\columnwidth]{eval-smart-insert}
\caption{\small Effect of smart insert on PoK and structure size.}
\label{fig:eval-smart-insert}
\end{figure}
\subsection{PoK Window}
We now turn to examining the PoK window query. For this evaluation, we used fixed window query sizes of $10{\it km} \times 10{\it km} \times 7 {\it days}$ for cabs and $500m \times 500m \times 10 {\it hours}$ for peds and ran a large number of window queries for each. We choose larger spatial and temporal ranges for the cabs data since queries about vehicular data is likely to cover greater distances and the data stored will be more sparse. We evaluate using the R-tree and the StdTable (a single table for storing the spatiotemporal data, which represents a conventional baseline) as alternatives for the internal structure and whether or not the $k$-d tree was used for the queries over sub-cubes (utility structure). We also evaluate the final {\sc \textbf{Paco}}\xspace option, which uses the R-tree as the internal structure, the $k$-d tree for the sub-cube queries, and a smart insert {\sc InsThresh} parameter of 0.8. We evaluate the average execution time and consumed energy, shown in Fig.~\ref{fig:eval-struct-compare}, across all traces and windows for the 5 possible structure combinations. The raw time values seem high (relative to desired application responsiveness rates) because the evaluation includes very large window sizes that can encompass a large number of data points. These windows push the limits of the structure; more practical application queries can expect to use smaller windows with fewer than 1000 data points.
The R-tree and StdTable without the $k$-d tree for in memory computation perform the worst (averaging 7 seconds and 1.5 mWh per query). The structures that include the $k$-d reference tree perform significantly better. The R-tree's benefits over the table approach are not as apparent since the sub-cube computation outweighs the initial range query on the internal structure. The R-tree's efficiency is more evident in very large but sparse data sets (i.e., those with many data points but only a small number of those points within a given query window). The fifth combination uses the R-tree with the $k$-d tree but with a smart insert value of 0.8 (as opposed to 1.0 for the other structures). As was previously demonstrated, an {\sc InsThresh} value of 0.8 maintains reasonable PoK values while drastically reducing the size of the structure; this is borne out here as well, where reducing the {\sc InsThresh} has a significant impact on both the query time and energy consumption.
\begin{figure}[!t]
\includegraphics[width=\columnwidth]{eval-struct-compare}
\caption{\small Internal structure and reference tree comparison.}
\label{fig:eval-struct-compare}
\end{figure}
The resources demanded for computing a PoK window query are directly related to the number of data items in the window. Since the window query is a common operation, it is important that it is effective across a large variation of window sizes. Fig.~\ref{fig:eval-supergrid-energy} shows the energy consumed by a PoK window query for various window sizes. We show the best structure combination, R-tree with a $k$-d reference tree, both with and without smart insert. For energy consumption, we set an upper limit of 0.5 mWh as a target for per-query energy consumption, hence the horizontal line in the figure. The Samsung Galaxy S5 has a 2800 mAh battery rated for 3.85 volts. Estimation yields a total capacity of 10,780 mWhs for the total battery, which allows for 21,560 window queries at this upper limit. We would like to keep most queries below this upper limit, and as shown in Fig.~\ref{fig:eval-supergrid-energy}, most realistic queries consume about 0.05 mWh of energy, increasing the number of possible queries in a single charge to over 200,000. Essentially, a user could perform 1,000 PoK window queries per day with less than a 0.5\% effect on battery life.
\begin{figure}[!hbt]
\includegraphics[width=\columnwidth]{eval-supergrid-energy}
\caption{\small Consumed energy vs. window size. Note log axes.}
\label{fig:eval-supergrid-energy}
\end{figure}
We similarly evaluated execution time across various window sizes and the data trends very similarly to Fig.~\ref{fig:eval-supergrid-energy}. We set a realistic target time of 2 seconds since applications should try to keep response times around 1 second~\cite{nielsen1994usability}. Most windows we evaluated maintained window query execution times below this target, with PoK window queries running for about 0.2 seconds on average. Also similar to our energy evaluation, the smart insert {\sc InsThresh} value decreased execution times considerably to assist in meeting the 2 second target.
\subsection{Grid Factor}
As important as {\sc InsThresh} is to smart insert, {\sc GridFactor} is to computing the window query. {\sc GridFactor} determines the number of grid cubes that are used in a window query computation. Larger number of cubes are desirable from an accuracy standpoint but can add significant computational overhead.
For dense query windows (i.e., those with 500+ data points to consider) we found that a {\sc GridFactor} of 2 averages over 37 seconds in execution time, which is unreasonable for mobile applications. For windows with fewer points, a {\sc GridFactor} of 2 is more reasonable (resulting in execution times of ~2 seconds). Ultimately, a {\sc GridFactor} value of 1 is best for default use, with the option to increase the {\sc GridFactor} for small windows or queries where the data is expected to be sparse; one can even consider decreasing the accuracy of the window query result by setting the {\sc GridFactor} to a value of 0.5 when applications require only a lossy representation of a window.
\subsection{Cloud Offload}
We investigate general power consumption of using {\sc \textbf{Paco}}\xspace in comparison to offloading to a server to demonstrate that our entire approach of on-loading is feasible and reasonably comparable to existing cloud-based silo approaches. In particular, we compare using {\sc \textbf{Paco}}\xspace~{\em on-device} to entirely offloading all of the device's spatiotemporal data to a cloud, where we assume queries can be processed quickly. Specifically, implementing the same storage features found in {\sc \textbf{Paco}}\xspace in the cloud in not necessary since we assume that cloud resources can be scaled to serve any desired minimal response time. To mimic minimal server response times, our evaluation sets the service instance to wait for 100ms before returning a response to simulate an optimistically realistic cloud computation. To generate results for the overall energy consumption of insertions and window queries on our cloud instance, we sent a large number of generic HTTP requests and computed the average cost per request. HTTP requests are sent each time a new location is sampled.
\begin{table}[!bt]
\caption{\small System storage and power usage }\label{tab:cloud-offload}
\begin{tabular}{ | p{4.1cm} | p{1.9cm} | p{1.5cm} |}
\hline
{\bf Evaluation } & {\sc \textbf{Paco}}\xspace~{\bf On-load} & {\bf Cloud} \\ \hline\hline
{\sc Walkabout Insert (30 min)} & 9.86 mWh & 11.47 mWh \\ \hline
{\sc Trace Insert (per point)} & 5.87 $\mu$Wh & 7.9 $\mu$Wh \\ \hline
{\sc PoK Window (per window)} & 25.7 $\mu$Wh & 7.9 $\mu$Wh \\ \hline
{\sc 20,000 data storage } & 1.24Mb & N/A \\ \hline
\end{tabular}
\end{table}
To generate the data for this experiment, we performed a one hour ``walkabout'' in which we walked around a WiFi enabled university campus while continuously polling Android location services on the device. As shown in Table~\ref{tab:cloud-offload}, inserting the spatiotemporal data points into {\sc \textbf{Paco}}\xspace was about 14\% more efficient than offloading the same data to the cloud, consuming 9.86 mWhs of energy vs. 11.47 mWhs in the cloud case. To further isolate these results from other effects of running the {\sc \textbf{Paco}}\xspace service while the device was in use, we also compared a controlled insertion of the same data from a captured trace, which resulted in a similar relative result between on-loading and off-loading. The gist of this result, however, is that in a full 16 hour day of continuous location polling, {\sc \textbf{Paco}}\xspace would only consume about 316 mWhs of energy to store a complete trajectory of the user's movement. With modern device battery capacities of over 10,000 mWhs, this is very feasible.
We also compared {\sc \textbf{Paco}}\xspace's average cost per point insertion and per PoK window query with cloud off-loading of the same activities. As also shown in Table~\ref{tab:cloud-offload}, {\sc \textbf{Paco}}\xspace is more energy efficient for data insertion but about three times less efficient for PoK window queries. This is intuitive since the cloud instance has potentially limitless resources to answer expensive queries but has an overhead associated with each request. We limited the queries to windows smaller than 1000 data points to better represent realistic application queries. Because it is expected that insertion will be much more common than window queries, the main take away here is that, while {\sc \textbf{Paco}}\xspace is not designed to save energy, its (overall) energy performance is reasonable in comparison to a conventional approach that relies entirely on a cloud backend for spatiotemporal storage. The last item in Table~\ref{tab:cloud-offload} outlines the storage requirements for 20,000 spatiotemporal data points, for which the {\sc \textbf{Paco}}\xspace representation required 1.24Mb. Devices could insert a new data point every 5 seconds for a week and only use about 7.5 Mb of storage, which is very reasonable given today's device storages of 16-32Gb. Note that this leaves plenty of storage for the additional $z$ context data we also plan to include.
\section{Discussion}\label{sec:discussion}
We have outlined how {\sc \textbf{Paco}}\xspace provides novel types of spatiotemporal queries while intelligently limiting the footprint of the structure on resource constrained mobile devices through the smart insert operation. While checking the usefulness of the data to {\sc \textbf{Paco}}\xspace before inserting it is a good first step, extension of {\sc \textbf{Paco}}\xspace could use a context-aware location sampling technique such as only sampling detailed location when an accelerometer indicates significant movement or by inferring expensive location sampling from cheaper sensors such as in ACE ~\cite{nath12:ace}. Such a step, performed before even considering whether a sample is a candidate for insertion in the {\sc \textbf{Paco}}\xspace data store, has the potential to provide energy savings on resource constrained mobile devices. Furthermore, {\sc \textbf{Paco}}\xspace enables control for the lossiness of the stored structure as well as lossiness in query results. The lossiness of the query results can be dynamically tuned per-application, but the lossiness of the stored structure must match the greatest accuracy requirement of all the applications. Going forward, {\sc \textbf{Paco}}\xspace might support early subscription of accuracy requirements to be able to tune the lossiness of the stored structure. On the other hand, because {\sc \textbf{Paco}}\xspace brings all of the spatiotemporal collection under one roof, each spatiotemporal data point is collected only once. This is in contrast to conventional approaches in which multiple applications may be collecting the same spatiotemporal data independently, which is also potentially wasteful.
The work presented in this paper is a necessary first step in giving mobile users control over vast amounts of personal contextual data that is collected about them without unnecessarily limiting the expressiveness of the queries as exposed to applications. Future work will look to develop a contextual ontology to couple with {\sc \textbf{Paco}}\xspace's current spatiotemporal indexing, further leverage the internal R-Tree structure to provide faster data region summaries, and to build a system of sharing summaries with other devices or determining appropriate cloud offloading to supplement {\sc \textbf{Paco}}\xspace.
\section{Conclusions}\label{sec:conclusions}
The key contribution of this paper is to make it possible to entirely {\em on-load} storage and querying of users' potentially highly personal (and private) spatiotemporal context information. To our knowledge, we are the first to enable such on-loading while providing a detailed historical view of spatiotemporal context to applications. Motivated by real-world applications, we introduced {\sc \textbf{Paco}}\xspace, a programming abstraction that enables efficient on-loading of contextual data storage and querying by indexing crucial spatiotemporal components. We placed {\sc \textbf{Paco}}\xspace in a broader system context, demonstrating its ability to support D2D exchange as well as cloud offloading while maintaining user control over potentially sensitive data. We evaluated {\sc \textbf{Paco}}\xspace over a variety of parameters and compared it to cloud based alternatives, and ultimately demonstrated the efficiency and feasibility of including {\sc \textbf{Paco}}\xspace on modern devices. Not only does {\sc \textbf{Paco}}\xspace give users control of their own spatiotemporally tagged context information, but, by making spatiotemporal context storage and querying a {\em system service}, {\sc \textbf{Paco}}\xspace breaks down the walls between today's existing siloed applications, sharing collected context information among all applications on the device.
As demonstrated in this paper, our implementation of {\sc \textbf{Paco}}\xspace enables on-loading contextual storage, giving users direct control over potentially private information and enabling low-latency responses to spatiotemporally indexed contextual queries for a wide variety of applications.
\section{Introduction}\label{sec:introduction}}
\input{introduction}
\input{systemmodel}
\input{background}
\input{structure}
\input{benchmarking}
\input{usecase}
\input{conclusions}
\section*{Acknowledgments}
This material is based upon work supported by the National Science
Foundation under Grant No. CNS-1218232. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of
the author(s) and do not necessarily reflect the views of the National
Science Foundation.
\bibliographystyle{IEEEtran}
\section{{\sc \textbf{Paco}}\xspace and the {\sc \textbf{Paco}}\xspace API}\label{sec:structure}
As shown in Fig.~\ref{fig:system-architecture}, {\sc \textbf{Paco}}\xspace relies on an internal structure to store spatiotemporal data; because the {\sc \textbf{Paco}}\xspace API shields the application developer from the specifics of this structure, any spatiotemporal data structure can be used that conforms to the appropriate interface. We first overview the requirements of such a data structure and discuss our use of the R-tree to fulfill these requirements. As shown in Fig.~\ref{fig:api}, the internal structure provides basic query operations across time and space. Upon this structure, we build the {\sc \textbf{Paco}}\xspace API as a suite of application layers that effectively extend the internal structure's interface, resulting in the complete {\sc \textbf{Paco}}\xspace abstraction for on-loading a device's spatiotemporal information locally and using that local data storage to answer spatiotemporal queries for applications on the device. Recall that {\sc \textbf{Paco}}\xspace is intended to replace traditional location services and serve as the primary access point for spatiotemporally-indexed contextual data. Note that in some cases single raw location data points may need to be exposed, such as for mapping applications that display your current location. We omit this from our discussion of the {\sc \textbf{Paco}}\xspace API in favor of the more interesting methods that {\sc \textbf{Paco}}\xspace provides.
\begin{figure}[!hbt]
\centering
\includegraphics[width=\columnwidth]{api}
\caption{\small Application layers and internal structure of the {\sc \textbf{Paco}}\xspace API}
\label{fig:api}
\end{figure}
\subsection{Required Internal Data Structure Interface}
The internal structure is responsible for the low-level efficient storage of spatiotemporal data. It must support inserting spatiotemporal data objects in the form $(x, y, t, {\it id})$, where $x$ and $y$ are spatial coordinates, $t$ is a temporal value, and ${\it id}$ is an identifier for the data object that allows {\sc \textbf{Paco}}\xspace to directly associate the entries in the spatiotemporal data store with the appropriate application-level context entries. We only use two spatial dimensions for our initial study, but {\sc \textbf{Paco}}\xspace can be easily extended to support more dimensions. For the remainder of this paper we omit discussion of ${\it id}$ and instead focus on the space and time portions of the indexing; we expect that these will be integral to most future mobile applications and are reasonable targets for storage and query optimizations.
The structure must also support the ability return a list of points matching a range of spatiotemporal values; for example a query may ask for all data points with an area comprising a popular running route from 5-8pm. This is a common feature of data structures supporting spatial and temporal indexing and is essential to the effectiveness of {\sc \textbf{Paco}}\xspace's abstractions; as such, optimizing efficiency of this query is crucial.
In addition, the underlying data structure should support an additional two query operations that will be used in {\sc \textbf{Paco}}\xspace to support expressive application queries. One is the nearest neighbor operation, which takes a piece of candidate spatiotemporal data and finds the closest matching data item stored in the structure. For instance, applications may ask for the stored data item nearest an historical landmark, closest to given time, or closest to a particular place {\em and} time. Finally, the internal structure must allow for sequential retrieval of time-ordered data items between two query points. This query is of the form $Q\{p_{1},p_{2}\}$, where $p_{1}$ and $p_{2}$ are spatiotemporal points previously inserted into the structure, and the query returns the sequential list of data between these two points. An example sequence query may ask for a list of points between a runner's starting and ending location.
\subsection{The {\sc \textbf{Paco}}\xspace API}
Building on the internal data structure, we introduce the {\sc \textbf{Paco}}\xspace API, which establishes a programming interface for answering important spatial and temporal questions and for controlling the contents of the spatiotemporal data store (see the outermost layer in Fig.~\ref{fig:api}). This API is tunable through configuration parameters that can be adjusted to match the current available resources, for different users, and for a particular user's situation or changing environment. While additional operations may be added to {\sc \textbf{Paco}}\xspace in the future, we demonstrate in Section~\ref{sec:casestudy} the effectiveness of those provided in this work to address existing and future application uses. A basic building block for the {\sc \textbf{Paco}}\xspace API is {\sc \textbf{Paco}}\xspace's notion of {\em probability of knowledge}, which we discuss first.
\subsubsection{Probability of Knowledge (PoK)}
Fundamental to {\sc \textbf{Paco}}\xspace is the creation of a model to represent the amount of knowledge stored in the internal structure about a particular place and time. Points in the {\sc \textbf{Paco}}\xspace data store have the form $(x,y,t,z)$ where $x$ and $y$ are spatial dimensions (typically longitude and latitude, respectively), $t$ is the timestamp, and $z$ is an abstraction for any additional context. In representing additional context as a single dimension $z$ is a significant oversimplification, our structure fully supports expressive extensions including incorporating a full context ontology~\cite{wang2004ontology}. We omit discussion of these details as well as computations on $z$ because the process is more of an exercise in developing the semantics of the context ontology and not fundamental to spatiotemporal data storage and querying. In contrast, the aim of this paper is to perform the ``heavy lifting'' related to the most common context aspects, i.e., space and time.
The mobile applications that motivate our work often want to ask questions of the form: ``How well does the data stored in the structure relate to a given reference point?'' or ``What does the structure knows about a point in space or time?'' For example, a contextual chat application might ask whether a user's historical context indicates purchasing items at a specific store at 5pm in order to connect messages with similarly profiled users. For a matching data point, $x$ and $y$ may be the GPS coordinates of the store, $t$ the timestamp for 5pm on the current day, and $z$ additional context that indicates shopping (e.g., a recorded purchase). Given some reference point $r=(r_x, r_y, r_t, r_z)$, we might ask what influence a particular stored data point, $p$, has on $r$. In other words, we want to determine the probability that a point in the structure shares common or related information with the reference point based on their regions of influence. We define this influence on a per-dimension basis, using a spatial influence function $I_s(\Delta x, \Delta y)$ and a temporal influence function, $I_t(\Delta t)$ where $\Delta x, \Delta y$ are the spatial distances from $p$ and $\Delta t$ is the temporal distance from $p$. We say that $p$ {\em influences} any reference point $r$ within this region.
\begin{wrapfigure}{r}{1.65in}
\vspace{-.5cm}
\includegraphics[width=.45\columnwidth]{decay}
\caption{\small Influence function for a point $p$ along a particular dimension (i indicates one of the three dimensions)}\label{fig:decay}
\vspace{-.5cm}
\end{wrapfigure} Fig.~\ref{fig:decay} depicts an influence function for some dimension $i$. Note that the this influence function is similar to modeling a probability density function over space and time, similar to the behavior modeling approach used in \cite{kayacik2014data}.
Using these influence functions, we define the {\em probability of knowledge} (PoK) as the amount of influence a point $p$ has on a reference point $r$ as:
\begin{multline}
{\Delta x = |p_x - r_x|};\:\: {\Delta y = |p_y - r_y|};\:\: {\Delta t = |p_t - r_t|} \\
{{\rm PoK}(p,r) = (I_s(\Delta x, \Delta y) * W_s) * (I_t(\Delta t) * W_t)}
\end{multline}
where $W_s$ and $W_t$ are two parameters termed {\sc SpaceWeight} and {\sc TimeWeight}, respectively. Applications can tune these weights to enable PoK computations to have more or less relative emphasis on spatial or temporal dimensions. In the remainder of this paper, we fix both values at 1.
In our study in Section~\ref{sec:benchmarking}, we use linear decay functions. Specifically, we fix the amplitude of the influence functions at 1 and use the value for which the influence function is 0 (x-intercept) or i-{\sc Range} (where ``i'' is either {\sc Space} or {\sc Time}) to adjust the slope of the influence function. The {\sc Range} values can be conceptually thought of as the physical distance and length of time over which a data point should exert influence on reference points. Note that by using linear influence functions, the decay of PoK behaves quadratically (since the two influences are multiplied in the resulting PoK calculation).
\subsubsection{PoK Window}
Potentially more useful than PoK information about a single point is information about the data structure's aggregate knowledge over a spatiotemporal area (i.e., a {\em PoK window}). For example, rather than querying for purchases made at the store at 5pm, an application may query for purchasing knowledge about a reasonably sized area around the store any time in the afternoon. Given such a spatiotemporal region (visualized by a rectangular cube in three dimensions, as shown in Fig.~\ref{fig:grid}), a PoK window query returns the PoK for the {\em region} as opposed to the PoK of a single point; the region's PoK depends on the spatial and temporal influences of the set of ``nearby'' data items stored in the internal structure.
\begin{wrapfigure}{r}{1.8in}
\includegraphics[width=.5\columnwidth]{cubes}
\caption{\small A PoK window computation highlighting one sub-cube}\label{fig:grid}
\end{wrapfigure}
The computation of a PoK window is an integral over the influence functions across the structure's dimensions. In {\sc \textbf{Paco}}\xspace, we estimate the value by summing the PoK of small {\em sub-cubes} that approximate the influence functions (essentially by performing smaller more tractable window queries), as shown in Fig.~\ref{fig:grid}. The size of the grid sub-cubes trades off computational complexity for the granularity of accuracy of the resulting PoK for the window. To specify the size of the grid sub-cubes, we define a value {\sc GridFactor} that is the number of sub-cubes to include per $i$-{\sc Range} value along any $i$th dimension. For example, given a {\sc SpaceRange} of 100 meters, a {\sc TimeRange} of 30 minutes, and a {\sc GridFactor} of 2, the sub-cube sizes are $50m \times 50m \times 15min$ (2 cubes per $i$-{\sc Range}). The value of the {\sc GridFactor} parameter affects both the accuracy and efficiency of the resulting window query; we briefly evaluate these tradeoffs in Section~\ref{sec:benchmarking}.
To compute the PoK of a window, {\sc \textbf{Paco}}\xspace performs a range query on the internal data structure for each sub-cube. The returned data items are then used to compute the PoK for the sub-cube. Finally, the PoKs for all of the sub-cubes in the window are aggregated to generate the PoK for the window. Since there are often many sub-cubes for which to calculate PoK values, and, for each sub-cube, {\sc \textbf{Paco}}\xspace must execute a smaller range query, we perform an initial range query using the entire window's bounds to generate a {\em reference set} of data points for the smaller sub-queries. The initial range query retrieves all of the candidate points stored in the internal structure that are within the window for the PoK computation. We can further restrict the set of candidate points by joining queries that account for filters relative to the additional context information, i.e., $z$. We then execute each sub-cube's smaller window query only on this reference set of candidate points rather than on the entire internal structure (e.g., in our experiments, reported in Section~\ref{sec:benchmarking}, this pre-query eliminated, on average, 97.5\% of the data points from consideration in the larger of our two data sets). {\sc \textbf{Paco}}\xspace stores this reference set of points in a $k$-d tree that it then uses when calculating each sub-cube's PoK. Our implementation of this approach leverages the benefits of $k$-d trees for in-memory computations and R-trees for maintaining a balanced on-disk database of the entire set of points in the structure. Our evaluation of PoK window queries in Section~\ref{sec:benchmarking} includes exploring the effectiveness of using a $k$-d tree rather than a simple array of points as the reference structure.
To calculate each sub-cube's PoK, we need to resolve the influence of potentially multiple data items. For example, Fig.~\ref{fig:pok2} shows (in only two dimensions) the (overlapping) ranges of influence of three data items C1, C2, and C3. Computing the PoK for a region that contains all or part of any of the intersection areas in Fig.~\ref{fig:pok2} involves resolving the shared influence between the relevant data points.
\begin{wrapfigure}{r}{2in}
\includegraphics[width=.60\columnwidth]{InclusionExclusion.jpg}
\caption{\small Inclusion-exclusion and spatial influence}\label{fig:pok2}
\end{wrapfigure}
Given a sub-cube, identified by a point at its center (the {\em target point}), we retrieve all of the nearby neighbors of the target point (the {\em candidate points}) using the internal structure's range query. We compute the structure's PoK of the target point using the {\em inclusion-exclusion} principle to account for ``double counting.'' This becomes more complicated as an increasing number of candidate points and dimensions are considered. In general, the following expression captures the ${\rm PoK}$ at a reference point $r$, where $1\leq i\leq n$ and $K_i$ is the influence from each of the $n$ nearby candidate points. $\mathbf{P}(K_i)$ reflects the distribution for point $i$ after accounting for both the spatial and temporal decays. Keep in mind that this could be extended to include items from the application-level context (i.e. $z$).
\begin{equation}
{\rm PoK}(r) = \mathbf{P} \left( \bigcup_{i=1}^{n} K_i \right)\label{eq:pok}
\end{equation}
\begin{equation}
\begin{aligned}
{\mathbf{P} \left( \bigcup_{i=1}^{n} i \right)} & =
{\sum_{i=1}^{n} \mathbf{PoK} \left( i \right)} -
{\sum_{i < j}^{} \mathbf{PoK} \left( i \cap j \right)} \\ & \hspace{.5cm} +
{\sum_{i < j < k}^{} \mathbf{PoK} \left( i \cap j \cap k \right)} \\ & \hspace{.5cm} -
{\dotsc} +
{ \left( - 1 \right) ^{n-1} \mathbf{P} \left( \bigcap_{i=1}^{n} i \right)}
\end{aligned}\label{eq:pokdetailed}
\end{equation}
${\rm PoK}$ values are between 0 and 1, inclusive, and the set intersection operator in Equation~\ref{eq:pokdetailed} generates the combined probability for the considered candidate points. Using Equations~\ref{eq:pok} and~\ref{eq:pokdetailed}, {\sc \textbf{Paco}}\xspace computes the ${\rm PoK}$ for each sub-cube. We then compute $T_o$ as the sum of the ${\rm PoK}$ values for all of the sub-cubes and $T_P$ as the maximum possible ${\rm PoK}$ (equivalent to every sub-cube having a ${\rm PoK}$ of 1, or, simply, the number of sub-cubes). We return the PoK for the window as the fraction $T_o/T_p$. For realistic data, our evaluation has show that values of $> 60\%$ indicate high levels of coverage.
Note that Equation~\ref{eq:pok} effectively generates all combinations of the $n$ PoK values. To mitigate state explosion, we sort the candidate points and select the first {\sc TrimThresh} number of points with the most significance and use these points to compute the window's PoK. In our evaluation, we fixed {\sc TrimThresh} at 10 because it empirically allows for a good number of points to be considered without generating too many sets of combinations.
PoK window queries are flexible and, like their underlying range queries, can be performed over any combination of the dimensions. For example, a query might ask for the PoK over a spatial-only region such as on a college campus or a temporal-only region such as from 9am to 5pm on a particular day. This distinction defines only the bounds for the initial query on the internal structure for which the reference structure is constructed. The PoK calculation for each sub-cube is unaffected as it still computes the influence values from all three dimensions. Similarly, the PoK computations could be extended to account for dimensions of the application-level context (i.e., $z$).
\subsubsection{Smart Insert}
Maintaining a data structure that contains detailed historical space-time indexes of contextual data generates large structures that may be expensive to maintain and query. Many data items are redundant in space, time, or both, and maintaining such redundant data items does not add much additional contextual information of use to applications. Most users stay in a relatively confined location throughout a day at work, so updating space-time information on a fixed (e.g., one second) schedule is often overkill. On the other hand, many mobile location services only update observations if the mobile object has moved a significant distance. This does not account for the temporal relevance of data objects, nor does it account for motion paths where an actively mobile user is constantly looping back on already ``covered'' areas.
In {\sc \textbf{Paco}}\xspace, we introduce {\em smart insert}, which uses application tunable guidance to ask ``how well is this information already represented?'' before inserting a data item. Smart insert accomplishes this check by performing a small PoK window query around the candidate point. The result of this window query sample is compared to an application-defined parameter, {\sc InsThresh}. Since both the value of the window query and the value of {\sc InsThresh} are PoK's, they represent the combined spatial {\em and} temporal relevance of the structure relative to the point that is to be inserted. For example, an {\sc InsThresh} value of 80 means that, if ${\rm PoK(p)} \geq 0.8$ for some new point $p$, then $p$ should not be inserted. Lower thresholds result in a greater reduction in the number of data items inserted, trading some degree of accuracy for storage and computational efficiency as depicted in
Fig~\ref{fig:visual1}, which gives heat maps of the PoK values of the resulting structure, given the ground truth in the left most figure (i.e., all data samples are inserted, regardless of their contribution to knowledge) vs. smart insert using an {\sc InsThresh} value of 80\% (in the center) and an {\sc InsThresh} value of 20\% (on the right). We provide a more controlled and detailed study of the impacts of these parameters on smart insert and the resulting available knowledge in the data structure in Section~\ref{sec:benchmarking}.
\begin{figure}[!hbt]
\includegraphics[width=\columnwidth]{visual1}
\caption{\small Effect of smart insert on spatiotemporal trace data.}
\label{fig:visual1}
\end{figure}
Since this operation uses the PoK window operation described above, its performance is partially dependent on the application parameters defined for coverage window. Using the PoK computations defined above, this considers only space and time information when determining whether a data item is redundant; again, PoK computations can be extended to include additional dimensions of context information, also increasing the expressiveness of the determination of redundancy for smart insert.
\subsubsection{Find Path}
{\sc \textbf{Paco}}\xspace also enables queries about a mobile user's {\em trajectories} through space and time. {\sc \textbf{Paco}}\xspace's find path operation takes two points $(x_1,y_1,t_1),(x_2,y_2,t_2)$. Under the hood, the operation first finds the nearest neighbor of each point in {\sc \textbf{Paco}}\xspace's internal data structure, then uses the internal structure's {\sf GetSequence} method to return the path between these two nearest neighbor points. To illustrate, first consider a running/exercise application. A simple example of find path is to query for two known data points, such as the runner's home and a familiar trail he ran two days ago, for which find path can return the sequential list of data points. A more interesting example of find path is to query for two non-exact data points. For example, a user who cannot recall his exact running path could query around a trail head and a nearby lake three weeks ago. While the user may have not passed through those exact points or at that exact time, find path will determine the nearest two points and return the sequential list of items best matching those points.We include the Find Path operation in {\sc \textbf{Paco}}\xspace to complement the other operations in motivating the use cases. Find Path is not dependent on configuration parameters, and as such, we omit an evaluation in favor of more interesting methods within {\sc \textbf{Paco}}\xspace.
A challenge for many of the operations in {\sc \textbf{Paco}}\xspace is setting the parameters to achieve desired application behavior. {\sc \textbf{Paco}}\xspace's programming interface is intentionally designed to make it easy for application developers to use defaults for these parameters or to substitute alternative values on a case-by-case basis. A summary of the parameters as well as concrete values we considered in the evaluation in the next section are summarized in Table~\ref{tab:params}. ``cabs'' and ``peds'' refer to the two data sets we use, which will be introduced in the beginning of the next section. In Table~\ref{tab:params}, {\bf Device Tunable} refers to whether a particular device will adjust the given parameter at run time (and under what conditions). {\bf App Tunable} refers to the likelihood or frequency with which an individual application will adjust the given parameter (e.g., for a particular query or sequence of queries).
\begin{table*}
\small
\caption{{\sc \textbf{Paco}}\xspace parameter values}\label{tab:params}
\begin{tabular}{ | p{1.8cm} | p{5.2cm} | p{3.5cm} | p{4.0cm} | p{1.7cm}|}
\hline
{\bf Parameter} & {\bf Description} & {\bf Study Values} & {\bf Device Tunable} & {\bf App Tunable} \\ \hline\hline
{\sc SpaceRange}& range of a point's spatial influence &50m~(peds); 1000m~(cabs) & No & Infrequently \\ \hline
{\sc TimeRange} & range of a point's temporal influence &5min~(peds); 60h~(cabs) & No & Infrequently \\ \hline
{\sc GridFactor} & grid squares per {\sc Range} value & 1/2, 1, 2 & Subject to energy constraints & Likely \\ \hline
{\sc InsThresh} & value deemed significant PoK about a given area & 0.1 - 1.0 & Subject to energy constraints and device mobility & Infrequently \\ \hline
\end{tabular}
\end{table*}
\subsection{{\sc \textbf{Paco}}\xspace Access Profiles}
Our initial intention is to store spatiotemporal data on device to make it available to applications running on that device. In this sense, the above API is one that is assumed to be ``open'' to any application running on the same device as {\sc \textbf{Paco}}\xspace. An obvious next use of {\sc \textbf{Paco}}\xspace is to enable individual users to control whether and how their personal spatiotemporal contextual data is shared with applications and services on the device and with other individuals or services executing off of the device. To enable this control, {\sc \textbf{Paco}}\xspace defines {\em access profiles} that use the {\sc \textbf{Paco}}\xspace API directly to constrain and filter the spatiotemporal data (and ultimately its derived PoK information) in making it accessible.
In {\sc \textbf{Paco}}\xspace, users employ the {\sc \textbf{Paco}}\xspace API's parameters to control the accessibility of the spatiotemporal data stored within {\sc \textbf{Paco}}\xspace. To provide examples of this process, Table~\ref{tab:profiles} gives a qualitative comparison of different access profiles that could be employed by users or devices to control release of their spatiotemporal information. The {\em open} profile allows complete access to the {\sc \textbf{Paco}}\xspace API and is suitable for trusted applications (i.e., likely those resident on the local device). The {\em guarded} profile outlines a safe middle-ground, for instance only allowing the find path operation for entries more than 24 hours in the past (preventing the user's exact recent movements from being tracked); restricting the grid cube size (effectively blurring the granularity of spatiotemporal information released); and providing a minimum window size (only allowing window query sizes five times larger than the {\sc Range} in any dimension). This guarded profile serves as an example for device-to-device interactions (based on the presumption that users are willing to share with other co-located users more than they are willing to share publicly on the Internet~\cite{jones08:geographic}), the {\em guarded}. At the most conservative end of these examples, the {\em restricted} profile serves well for offloading to a public server or sharing with unknown users by only allowing lossy data representations, ensuring that the data is sufficiently obfuscated to make the user comfortable in releasing it. One could imagine switching between profiles not just for the three different modalities of applications shown in Fig.~\ref{fig:system-architecture} but also based on social relationships with the users of the connected devices, the time of the data, the location of the exchange, the user's mood, etc.
\begin{table}[!h]
\begin{center}
\caption{Qualitative Comparison of Access Profiles}\label{tab:profiles}
\begin{tabular}{ | l | l | l | l |}
\hline
{\bf Profile} & {\sc\bf GridFactor} & {\bf Min. Window Size} &
{\bf Find Path} \\ \hline\hline
{\em open} & 1/2 - 2 & None & Yes \\ \hline
{\em guarded} & 1/2 - 1 & 5X {\sc Range} & $> 24h$ \\ \hline
{\em restricted} & 1/2 & 20X {\sc Range} & No \\ \hline
\end{tabular}
\end{center}
\end{table}
\subsection{Android {\sc \textbf{Paco}}\xspace Implementation}
We implemented the {\sc \textbf{Paco}}\xspace components depicted within the dashed area in Fig.~\ref{fig:system-architecture} on Android\footnote{The source code for this implementation is available at \url{https://github.com/nathanielwendt/LSTAndroid}.}. The current implementation of {\sc \textbf{Paco}}\xspace is an Android service that serves as a sufficient prototype for evaluation. It is our vision that {\sc \textbf{Paco}}\xspace be implemented at the system level to be more natively shared across applications. Our prototype implementation is, however, more suitable to evaluation on commodity devices, especially in a way that can be replicated by others. We use a 3D R-tree as the internal data structure to store each $(x,y,t)$ point; for simplicity of evaluation, our current prototype does not associate the $z$ values with the spatiotemporal samples, but this is a straightforward extension. The default SQLite build contained within Android does not include the needed R-tree module, so we compiled our own version of SQLite and accessed it through JNI in our benchmark test runner. SQLite's R-tree module requires insertion of regions, $(x_{min},x_{max},y_{min},y_{max},t_{min},t_{max})$; for a given $(x,y,t)$ point, we duplicate data so that for each dimension $i$: $i_{min} = i_{max}$. We also adapted the $k$-d tree module from the java machine learning library\footnote{\url{http://java-ml.sourceforge.net/}} for use as our in-memory reference tree. Lastly, to support storing GPS data, we developed a set of GPS library functions for manipulating spatial distance over latitude and longitude while taking into account curvature of the earth, etc.\footnote{Equations: http://www.movable-type.co.uk/scripts/latlong.html}
\section{A Motivating System Model}\label{sec:systemmodel}
This paper describes the {\sc \textbf{Paco}}\xspace API and its supporting infrastructure, which together enable on-loading spatiotemporal data and rich application-level queries over that data. {\sc \textbf{Paco}}\xspace fits in a larger system context that makes contextual data accessible not only to local applications but also to nearby devices, opportunistically connected via device-to-device communication, and to the larger world through cloud off-loading. While this paper does not focus primarily on the elements of the {\sc \textbf{Paco}}\xspace API used for this sharing of the on-loaded spatiotemporal information, the broader system model includes these capabilities.
Fig.~\ref{fig:system-architecture} shows {\sc \textbf{Paco}}\xspace in this larger context. At the base layer of this architecture, we rely on existing implementations that can support a spatiotemporal database (``Internal Structure'' in the figure, which we implement using an {\sf R-Tree}~\cite{guttman1984r}) that can store frequent updates of time- and location-stamped observations of context. Elements are indexed within the internal structure based on their location coordinates and time stamps; each element in the internal structure is also mapped to one or more elements in the {\sf Context Index}, indicating the semantic information associated with the referenced space and time. For example, an observation might be described as: ``at 5:10pm the user is running 2 miles starting from his house (actual gps coordinates)'' or ``at 6:00am the user purchased a coffee at the cafe and gave the cafe a 5 star rating''. For each data item, the spatiotemporal data forms the base to which additional context (e.g., that the current activity is running or that the user purchased coffee or gave a particular rating) can be added. Observations are captured by applications in user-space (e.g., via the {\sf Apps} and {\sf Sensors} components in Fig.~\ref{fig:system-architecture}). Capturing context may require explicit user interaction, e.g., a user inputs a review of a cafe, while other context can be automatically captured in the background through sensor activity, e.g., the ambient nose level at a cafe.
\begin{figure}[!h]
\centering
\includegraphics[width=.75\columnwidth]{system-architecture}
\caption{ System architecture; dotted region is {\em Paco}'s contributions}
\label{fig:system-architecture}
\end{figure}
The {\sc \textbf{Paco}}\xspace API provides access to the stored data in a manner tailored to mobile applications. At the most basic level, the user's personal applications on the device can freely query the stored spatiotemporal data, retrieving any view of the raw data desired. Practically, users will likely want to exercise control over release of their raw contextual data even to on-device apps as well as for device-to-device (D2D) exchange and cloud offloading. For this reason, {\sc \textbf{Paco}}\xspace uses {\em access profiles} to further filter or otherwise alter or constrain the contextual information released from {\sc \textbf{Paco}}\xspace. These access profiles are user- or device-defined and can take into account the user's current situation (e.g., physical environment), his social network, etc. These access profiles can govern the granularity of the information released (e.g., releasing only aggregate ``coverage'' views of large areas) or the spaces and times about which information is released (e.g., releasing fairly detailed information about the user to his colleagues in his work environment or releasing any information that is more than a week old). These access profiles can be applied whether the data is being shared in the local environment (e.g., through D2D exchanges) or to offload contextual data to the cloud, for example offloading sufficiently dated information or lossy summaries that obfuscate the user's exact contextual data.
In this paper, we focus on the {\sc \textbf{Paco}}\xspace API as it relates to on-loading spatiotemporal contextual storage and subsequent on-device application queries; these elements are depicted within the dashed rectangle in Fig.~\ref{fig:system-architecture}. We define the {\sc \textbf{Paco}}\xspace data structure, based on the existing {\sf R-Tree} spatiotemporal index. We create the {\sf Smart Insert} method and the query interface that defines the {\sc \textbf{Paco}}\xspace {\sf API}, and we show how {\sf Access Profiles} can be used to restrict the release of potentially private contextual data. We do not implement communication mechanisms to enable D2D sharing and cloud offloading; we could easily connect {\sc \textbf{Paco}}\xspace to existing mechanisms in these spaces~\cite{kalbarczyk16:xd}. We do not include the {\sf Context Index} in this investigation; for the remainder of this paper, we treat the structure as storing just the space and time of the observations, what we assume to be the ``heavy-lifting'' and most general purpose pieces of contextual indexing. Since our spatiotemporal index is relational, it supports additional context indexing through joins on selection queries. Including the contextual indices as well as the associated contextual ontology is left for future work, and is also addressed in much existing work on context ontologies~\cite{bettini10:survey, wang2004ontology}.
\section{Use Cases}\label{sec:casestudy}
\definecolor{dkgreen}{rgb}{0,0.6,0}
\definecolor{gray}{rgb}{0.5,0.5,0.5}
\definecolor{mauve}{rgb}{0.58,0,0.82}
\lstset{frame=tb,
language=Java,
aboveskip=3mm,
belowskip=3mm,
showstringspaces=false,
columns=flexible,
basicstyle={\scriptsize\ttfamily},
numbers=none,
numberstyle=\tiny\color{gray},
keywordstyle=\color{blue},
commentstyle=\color{dkgreen},
stringstyle=\color{mauve},
breaklines=true,
breakatwhitespace=true
tabsize=1
}
In this section, we provide two case studies for demonstrating applications' use of the {\sc \textbf{Paco}}\xspace service. The first of these examples takes an existing application and refactors it to remove its siloed use of spatiotemporal information to replace it with the use of {\sc \textbf{Paco}}\xspace's abstractions instead. The second shows novel application behavior that is enabled by the existing of {\sc \textbf{Paco}}\xspace. In both uses cases, the device employs {\sc \textbf{Paco}}\xspace's smart insert method to collect spatiotemporal context information, and the device (user) itself manages the insertion thresholds. To provide more application richness, these example applications also bring back the $z$ component of application-level context.
\subsection{Running}
The first example considers a running application that collects location samples for a specified period of time (explicitly started and ended by the user) to collect and report on path information, pacing, and route times. Many such applications exist, including the MapMyRun\footnote{\url{https://www.mapmyrun.com}} family of applications, the Nike+ app\footnote{\url{http://www.nike.com/us/en_us/c/nike-plus/running-app-gps}}, or even apps like Spotify Running\footnote{\url{https://www.spotify.com/running}} that use additional forms of context information (e.g., running tempo) to deliver media content. In all of these applications, context collection is a dedicated part of the app. If a user wants to employ all three of these applications simultaneously (which is reasonable, given that they all provide different capabilities, connections, and statistics), each device maintains its own store of spatiotemporal and context information, and each of these applications pushes that personal context information to a cloud service for processing and storage. Much of the context-related activities of the applications are redundant, as is the cloud offloading, and the latter also releases potentially private information.
In this {\sc \textbf{Paco}}\xspace use case, we therefore show how one would refactor such an application to make use of {\sc \textbf{Paco}}\xspace's spatiotemporal context abstractions. To start with, because {\sc \textbf{Paco}}\xspace continuously collects spatiotemporal context information, the user no longer has to remember to open each of the apps and explicitly select ``begin run.'' Instead, the device passively and unobtrusively records the context information (based on its settings) for later querying by the applications. That is, based on the context collected by {\sc \textbf{Paco}}\xspace on the device, the application could recreate a run {\em post hoc} based on the information stored in the {\sc \textbf{Paco}}\xspace structure. Fig.~\ref{lst:runningapp} shows code that would be launched as startup of a running statistics app (e.g., like mapmyrun) to examine route and pacing information for a previously performed run.
\begin{figure}[!hbt]
\begin{lstlisting}
//on app startup
ContextWindow window =
ContextWindow.Builder().setTimeWindow(lastAppUse, now)
.setContextMobility(``Running'')
.build();
Paco.setGridFactor(0.5);
double pok = Paco.windowPoK(window);
if(pok > 0.1){
List<ContextPoint> points = Paco.findPath(window);
// display map view and/or share running profile
}
\end{lstlisting}
\caption{\small Querying for a activity in a running app.}
\label{lst:runningapp}
\end{figure}
As shown in Fig.~\ref{lst:runningapp}, when the user opens the app, it can access relevant context pieces from {\sc \textbf{Paco}}\xspace's historical record. The app first creates a context window with rough precision. {\sc \textbf{Paco}}\xspace's configurability allows for a lossy representation of PoK for this query since the exact value is not important, and a general indication above some base threshold (0.1) suffices. This allows for a fast qualifying query before the app queries for the exact path information. Once the latter is obtained, the path can be displayed to the user or offloaded to the cloud for use across devices or for archiving. The user can also allow this application access to specific contextual data. For instance, the application may be only allowed to offload context that the user has explicitly provisioned, in this case spatiotemporal data with a mobility profile. Instead of offloading the user's entire trajectory, however, the app can offload exactly only the data from the running activity and then only at a level of abstraction matching a stated access profile.
Given that, for example, accelerometer data is part of the $z$ context collected by the device and associated with the spatiotemporal samples, a more sophisticated running statistics app can give the user historical information about running tempo throughout the entire path. Similarly, an executing app like a tempo-sensitive music playlist could use the information from {\sc \textbf{Paco}}\xspace~{\em as it is collected} to adjust the songs selected as part of the playlist. That is, {\sc \textbf{Paco}}\xspace supports {\em both} applications' needs for instantaneous context information and the ability to query context information retrospectively.
\subsection{ContextChat}
To demonstrate {\sc \textbf{Paco}}\xspace's ability to support next-generation applications by enabling new uses of context, we turn to a {\em ContextChat} example, which connects messages between users not only co-located in time and space but across other contextual attributes, for instance based on their exercise history, their dining patterns, or the spending profiles. To demonstrate the types of interactions that such an application might have with {\sc \textbf{Paco}}\xspace, we consider a use case in which the application aims to send chat messages between drivers who have similar driving experiences. In this example, the selection of similar drivers is based on their spatiotemporally captured paths while engaged in the higher level context activity of ``DRIVING.'' However, using even more expressive notions of context (i.e., $z$), the selection of similar drivers could also be based on other context measures, for instance, of a driver's level of aggressiveness.
Fig.~\ref{lst:contextchatapp} demonstrates how this application would capture the similarity between the drivers (i.e., as measured by PoK) within given space and time bounds. For demonstration purposes, this example uses a static query, but more advanced queries could query over dynamic properties determined over user experience.
\begin{figure}[!hbt]
\begin{lstlisting}
SpaceWindow spaceWin = SpaceWindow.bound(currentLoc, 100);
//100 meter bound
ContextWindow window = ContextWindow.Builder()
.setTimeWindow(now - Time.10MINS, now)
.setSpaceWindow(spaceWin)
.setContextMobility(ContextMobility.Driving);
Paco.setGridFactor(2.0);
double pok = Paco.windowPoK(window);
//retrieve messages
List<Message> messages =
CloudBackend.GetMessages(userId, window, pok);
//send a message
CloudBackend.SendMessage(userId, window, pok, message);
\end{lstlisting}
\caption{\small Sending and receiving messages in the ContexChat app.}
\label{lst:contextchatapp}
\end{figure}
First, the app creates a space window for a 100 meter perimeter around the user's current location. Using this space window, relevant time, and driving context, the app creates a contextual query window. The precision is set to fine since the actual PoK value is important. The PoK value and the query are sent to a backend chat service that retrieves any messages sent by other drivers that match the given context. No specific contextual data is sent to the siloed cloud backend, simply a region and a configurably lossy PoK value representing coverage for a contextual query. This information is abstracted but is sufficient for a backend to index and query to send messages to appropriate users. While this example uses a centralized cloud backend to exchange the chat messages, {\sc \textbf{Paco}}\xspace is also compatible with D2D uses. Devices could share context windows, requesting data queries from nearby users directly to gain information about their surroundings, what they are doing, and where they are heading.
|
\section{Introduction}
\label{sec:intro}
Recently, strong-field ionization in mid-infrared (mid-IR) laser fields has gained a lot of attention for the generation of coherent soft x-rays with high harmonic generation (HHG) \cite{Popmintchev_2012} and for the discovery of a variety of strong field characteristica, like the observation of holographic electron interferences \cite{Huismans_2011} and low-energy structures \citep{Blaga2009strong, Quan_2009,Wolter_2015x}. These processes are typically described through the recollision of the electron wave packet with the residual ion (or parent ion).
In a two-step model \cite{SimpleManModel},
the electron is released to the continuum and subsequently driven back by the laser field towards the ion core where it can recollide.
Upon return, the electron can either recombine, scatter inelastically or scatter elastically \cite{pfeifer_2006}.
Recombination of the electron with the ion leads to HHG \cite{McPherson_1987, Ferray_1988}. Inelastic scattering leads to non-sequential double ionization \cite{Fittinghoff_1992, Walker_1994} and excitation of the ion \cite{Feuerstein_2001}.
However, the vast majority of the rescattering electrons undergoes elastic forward scattering, where the absolute value of momentum of the electron does not change on its trajectory past the parent ion.
Interference of electrons with different ionization paths that end up with the same final momentum leads to many different structures, such as equidistant peaks in energy from above threshold ionization \cite{Agostini_1979}, electron-diffraction patterns \cite{Blaga_2012}, and holographic interference structures \cite{Huismans_2011}.
Elastic rescattering can lead to signatures in photoelectron momentum distributions (PMDs) such as Coulomb-focusing \cite{Brabec_1996,Comtois_2005} and low energy structures \cite{Blaga2009strong, Quan_2009,Wolter_2015x}. These kind of signatures in PMDs have been assigned to forward scattering which is most pronounced at slow recollisions, i.e. recollisions where the momentum of the electron approaches zero at the time of recollision \cite{Sasaki_2009, Liu_2010,Yan_2010,Huang_2010, Liu_2011, Kastner_2012,Lemell_2012,Moller_2014,Kelvich_2016}.
Rescattering effects are the strongest with linear polarization of the laser pulse, however, rescattering is still observed with elliptical polarization, which is possible due to the spread of the returning wave packet \cite{Wang_2009,Mauger_2010,Liu_2012,Shafir_2013,Li_2013}.
The displacement amplitudes within the simple man's model
\cite{SimpleManModel}, i.e. $ E_0/\omega^2$ along the long axis and $\epsilon E_0/\omega^2$ along the short axis of the polarization ellipse, depend linearly on the peak electric field $E_0$ and the inverse square of the laser frequency $\omega$.
Therefore, rescattering effects are expected to become more significant for mid-IR wavelengths even at moderate laser intensities.
Moving from near-IR to mid-IR wavelengths will strongly increase the maximal kinetic energy which allows for diffraction experiments with increased resolution \cite{pullen_2015} and HHG with photon energies up to the order of \mbox{1 keV} \cite{Popmintchev_2012}.
Furthermore, discoveries like low energy structures \cite{Blaga2009strong, Quan_2009,Wolter_2015x} and holographic interference patterns \cite{Huismans_2011} were first observed at mid-IR wavelengths.
\begin{figure}[t]
\includegraphics{figsLatest/SplitSurf.pdf}
\caption{Isosurface of a three-dimensional photoelectron momentum
distribution (3D PMD) with a partial cut in the polarization plane $(p_x,p_y)$. The 3D-PMD is recorded using a
mid-IR laser with a center wavelength of 3.4 $\mu$m, a pulse length of 50 fs, a peak intensity of $6 \cdot 10^{13}$ W/cm$^2$ and an ellipticity of $\epsilon = 0.11$. The sharp thin line-shaped ridge structure around $p_y = 0$ is clearly separated from the lobes of direct electrons.}
\label{fig:Surface}
\end{figure}
Another characteristic phenomenon of strong-field ionization at mid-IR wavelengths is the onset of non-dipole effects. When driven by mid-IR laser pulses, the magnetic field induced Lorentz force
starts to become significant for the electron dynamics during the
ionization process. The occurrence of magnetic field effects at long wavelengths has been theoretically predicted and described as the long-wavelength limit of the dipole approximation \cite{Keitel_1995,Walser_2000a,Milosevic_2000,Kylstra_2001,Reiss_2008,Kohler_2012b}.
The Lorentz force along the laser propagation direction is responsible for the photon's momentum transfer to electrons observed in \cite{Smeenk_2009momentum}, and for the momentum partitioning between the ion and the electron during ionization \cite{Klaiber_2013c,Chelkowski_2014,Cricchio_2015,Chelkowski_2015}.
The laser magnetic field induced drift is known to suppress the recollision and HHG at high laser intensities ($I\gtrsim 10^{17}$ W/cm$^2$ at a laser wavelength $\lambda=800$ nm) \cite{Keitel_1995,Walser_2000a,Milosevic_2000,Kylstra_2001,Dammasch_2001,Klaiber_2005,DiChiara_2008,Ekanayake_2013}, at which the recolliding electron is deflected at the parent ion by more than the electron wave packet size \cite{Walker_2006}.
At lower laser intensities the recollisions are still possible
because Coulomb focusing can compensate the magnetic-field
induced drift for linear polarization \cite{Foerre_2006}.
Strong-field ionization experiments at mid-IR wavelengths revealed another phenomenon beyond the long wavelength limit of the dipole approximation \cite{Ludwig_2014} which is relevant within this paper: an initially surprising shift of the peak of the projection of the PMD onto the beam propagation axis opposite to the beam propagation direction was observed. Also, the magnitude of this momentum shift was practically independent of the parent ion. The magnetic-field-induced lateral displacement of the electron and the successive recollision with the Coulomb potential was identified as the cause for this effect. This initial experiment triggered further experimental and theoretical investigations which is the focus of this paper.
\begin{figure}[t]
\includegraphics{figsLatest/fig15.pdf}
\caption{ Offsets of the peak of the PMDs projected onto the beam propagation axis ($p_z$) using a mid-IR laser (a) Measurement for circular polarization together with the data for linear polarization taken from Ref. \cite{Ludwig_2014}. We compare our data for circular polarization with the radiation pressure picture used in Ref. \cite{Smeenk_2011} (b) Measurements as a function of ellipticity for a peak intensity of $6 \cdot 10^{13}$ W/cm$^2$. We observe a transition from negative to positive values of $p_z$ with the zero crossing at $\epsilon \approx$ 0.12. The vertical dashed line indicates the ellipticity corresponding to the 3D PMD from Fig. \ref{fig:Surface}.}
\label{fig:ellShift}
\end{figure}
Here, we therefore present a more detailed study of ellipticity-dependent strong-field ionization at mid-IR wavelengths taking into account non-dipole effects. We discuss two main results:
1) We observed the creation of a sharp, thin line-shaped ridge structure of low-momentum electrons in the polarization plane at small ellipticities (Fig. \ref{fig:Surface}). Our analytical and numerical model show that
this ridge structure stems mainly from Coulomb-focused electrons
undergoing multiple revisits of the parent ion with at least one
significant rescattering event.
2) We investigated how non-dipole signatures on the PMDs depend on laser ellipticity.
The shift of the peak of PMDs with respect to the beam propagation axis (Fig. \ref{fig:ellShift})is opposite to the beam propagation direction for both linear polarization and small ellipticites, and is directly related to the sharp, thin line-shaped ridge structure created by Coulomb-focusing in the 3D PMDs.
With increasing ellipticity we then observe a shift of the PMD peak into beam propagation direction (Fig.\ref{fig:ellShift} (b)) which gives a direct link between the results observed in Ref. \cite{Ludwig_2014} and the radiation pressure picture in Ref. \cite{Smeenk_2011} that was employed to explain a shift in beam propagation direction.
The relationship between Coulomb focusing and non-dipole effects explored here increases our understanding of electron ionization dynamics in mid-IR laser fields, helps to better understand initially unexpected features of the observed PMDs, and will allow us to further exploit these effects to significantly enhance the resolution of the attoclock, time-resolved holography and strong-field molecular imaging using mid-IR lasers.
The paper is organized as follows: In section \ref{sec:exp}, we present the details of the experiment; in section \ref{sec:PMD3D} we present 3D PMDs recorded with elliptical polarization. We compare the experimental results with classical trajectory Monte-Carlo (CTMC)-simulations; In section \ref{sec:AnaOne} and \ref{sec:AnaTwo}
we present theoretical models to explain the creation of ridge caused by Coulomb-focused electrons. The results of the ellipticity-dependent non-dipole effects on the PMD are presented in section \ref{sec:noDipole}.
\section{Experimental Details}
\label{sec:exp}
PMDs were recorded with a velocity map imaging spectrometer (VMIS) \cite{Eppink_1997, Parker_1997} with the gas nozzle integrated into the repeller to achieve high gas target densities in the interaction region \cite{Ghafur_2009, Weger_2013}. The target was ionized by an optical parametric chirped-pulse amplifier (OPCPA) system based on chirped quasi-phase-matching devices. This system can deliver pulses with duration of \mbox{44 fs} and a pulse energy of \mbox{22 $\mu$J} at a center wavelength of \mbox{3.4 $\mu$m} and a high repetition rate of \mbox{50 kHz} \cite{Mayer_2013, Mayer_2014OSA}. The pulses were focused with a backfocusing dielectric mirror with a focal length of 15 mm into the interaction region.
The polarization of the laser beam was controlled by two custom-made achromatic MgF$_2$ wave plates. A quarter-wave plate induces the ellipticity and the subsequent half-wave plate controls the orientation of the polarization ellipse. The wave plates were fully characterized via polarimetry measurements where the power transmitted through a polarizer was recorded as a function of the angle between the major polarization axis and the polarizer axis. The polarization state at the desired orientation was extracted via a fit and interpolation of the measured values.
The intensity in all experiments was calibrated with reference measurements at close-to-circular polarization. The radial maximum of the torus-shaped momentum distribution was compared with semiclassical Monte-Carlo simulations \cite{Pfeiffer_2012}.
Throughout the article, the following coordinate system will be used: The coordinate $z$ denotes the direction of beam propagation, $x$ the major and $y$ the minor axis of the polarization ellipse and $p_x,p_y,p_z$ the respective electron momenta. $W(p_x,p_y,p_z)$ denotes the PMD, i.e. the amplitude of the photoelectron signal.
The experiments require an accurate determination of the zero momentum spot, in particular on the beam propagation axis. This spot was identified via a sharp point in the center of the PMD recorded with linear polarization that stems from the ionization of atoms that were left in a Rydberg state by the laser pulse and were subsequently ionized by the static electric field of the spectrometer \cite{Nubbemeyer_2008, Smeenk_2011}.
As these electrons do not interact with the pulse, they are guided by the static electric spectrometer field to the position on the detector that corresponds to zero momentum in the ($p_x,p_z$)-plane. The exact position of zero momentum in $p_z$-direction was determined from the projection of a small range of 0.05 a.u. in $p_x$ of the PMD onto the $p_z$-axis. This projection was fitted with a Lorentzian profile. This method was also applied to find the center in $p_y$-direction.
Throughout the article, atomic units (a.u.) are used.
\section{Three Dimensional Photoelectron Momentum Distributions (3D PMDs)}
\label{sec:PMD3D}
\begin{figure*}
\includegraphics{figsLatest/ellPolPlane.pdf}
\caption{Measured PMDs in the polarization plane measured at a peak intensity of $6 \cdot 10^{13}$ W/cm$^2$ for the ellipticites 0.0, 0.03, 0.07, 0.11, 0.15, 0.19, 0.23, 0.26, in (a) to (h), respectively. The central spot that stems from Rydberg states was covered in black for illustration purposes. The shown PMDs are projections from the range $|p_z|<0.06$ a.u. onto the polarization plane. These PMDs for ellipticities of $\epsilon = 0.07$ and $\epsilon = 0.11$ reveal a sharp line structure that disappears for larger ellipticities.}
\label{fig:ellScan3D}
\end{figure*}
We recorded 3D PMDs at various ellipticities at an intensity of $6 \cdot 10^{13}$ W/cm$^2$ to study the ellipticity-dependence of rescattering in mid-IR laser pulses.
When the ellipticity is varied, the electron dynamics changes mainly in the polarization plane, i.e. the $(p_x, p_y)$-plane. To have access to the polarization plane, we record full 3D PMDs from strong-field ionization. The full 3D PMD is obtained by applying a tomographic reconstruction algorithm to the projected PMDs measured with velocity map imaging \cite{Wollenhaupt_2009, Smeenk_2009momentum, Dimitrovski_2014}.
The orientation of the 3D PMD with respect to the detector plane is linked to the orientation of the polarization ellipse, allowing us to rotate the PMD by rotating the polarization ellipse. The beam propagation axis is parallel to the detector plane.
The polarization is rotated in steps of two degrees and a photoelectron image is recorded for each angular step.
Subsequently, for each slice along the beam propagation direction, a filtered back-projection algorithm is applied for the tomographic reconstruction.
An example of a measured 3D PMD is visualized as an isosurface in Fig. \ref{fig:Surface}. The isosurface exhibits two main lobes and a sharp ridge around $p_y = 0$.
In the following we distinguish between two types of photoelectrons:
Type A photoelectrons: The ellipticity dependence of the PMDs in the polarization plane is shown in Fig. \ref{fig:ellScan3D}.
Cuts of the 3D PMDs through the polarization plane
(i.e. $W(p_x, p_y) = \int_{\Omega_z} W(p_x,p_y,p_z) dp_z$) are shown in Fig. \ref{fig:ellScan3D} for ellipticities of 0.03, 0.07, 0.11, 0.15, 0.19, 0.23, 0.26. For the cuts we integrated over a range of $\Omega_z = |p_z|<0.06$ a.u.. The cuts show that, with increasing ellipticity, the cigar-shaped PMD evolves into a torus-like shape that is characteristic for strong field experiments with elliptically polarized pulses, such as typically observed with attoclock experiments \cite{Eckle_2008a, Eckle_2008b}.
The appearance of the two maxima on the short axis of the polarization ellipse can be explained by a simple man's model \cite{SimpleManModel}.
Throughout this article, we will refer to the electrons ending in these maxima as type A electrons, as indicated in Fig. \ref{fig:Traj}.
Within the framework of this model, the maxima are shifted by 90$^\circ$ with respect to the phase at which the maximum of the electric field occurs.
Deviations from 90$^\circ$ that are expected from the simple man's model are due to the Coulomb-interaction of the electron with the ion core \cite{Bashkansky_1988,Eckle_2008b, Pfeiffer_2012}, ionization delay times \cite{Eckle_2008a, Eckle_2008b, Landsman_2014o} and multi-electron effects like the induced dipole moment due to the ion’s polarizability \cite{Pfeiffer_2012}.
Type B photoelectrons: In the evolution of the PMDs, one can observe for small elliptictities, in particular for $\epsilon = 0.07$ and $\epsilon = 0.11$, the appearance of a sharp, thin line-shaped ridge structure around $p_y = 0$.
To the best of our knowledge, no such sharp separated structure has been observed in near-IR-experiments conducted at wavelengths around 800 nm. For the rest of this article, we will refer to these electrons as type B electrons.
To understand the nature of type B electrons, we compare them with CTMC simulations using the two-step model of strong-field ionization. The initial conditions for the photoelectrons (ionization times, positions and momenta) are obtained from tunnel ionization theory in parabolic coordinates \cite{PPT,ADK,Delone_1991, TunnelIonization, Pfeiffer_2012}, while the trajectories of the electrons are obtained by solving Newton’s classical equations of motion in the electromagnetic field of the laser pulse and the Coulomb potential of the parent ion. Subsequently the trajectories are binned in momentum space.
We first compare the outcome of the simulations with our experiments in the polarization plane. The results are shown in Fig. \ref{fig:Compar}. The semiclassical simulations were able to reproduce the appearance of the type B photoelectron signal. Our semiclassical approach does not take into account
the quantum interference of the electron trajectories in the continuum, and thus we can conclude that the appearance of this structure is due to momentum space focusing of photoelectrons and is not created by a pure interference effect.
\begin{figure}
\includegraphics{figsLatest/fig3.pdf}
\caption{Polarization plane PMDs projected from the range $|p_z|<0.06$ a.u.: Comparison of measurement (a) and CTMC calculations (b) from strong field ionization of xenon at an intensity of $6 \cdot 10^{13}$ W/cm$^2$ and an ellipticity of $\epsilon = 0.07$. For both experiment and simulation a sharp line appears around $p_y = 0$. (c) CTMC simulation for which the Coulomb potential was neglected, but the magnetic field component of the laser field was included. The sharp line around $p_y =0$ disappears. (d) CTMC simulation for which the Coulomb potential was included, but the magnetic field component was neglected.}
\label{fig:Compar}
\end{figure}
In order to understand if the creation of the ridge of type B electrons was induced by magnetic field effects, that appear at our beam parameters \cite{Ludwig_2014}, we performed the CTMC-calcuations with and without the inclusion of the magnetic field component (Fig. \ref{fig:Compar} (b) and (d)). The CTMC calculations reproduce the sharp structure of type B-electrons in the $(p_x, p_y)-$plane equally well. Thus, we conclude that for the appearance of type-B electrons in the polarization plane the magnetic field is not essential.
Furthermore, to confirm that the structure was created under the influence of the Coulomb-potential of the ion, we performed CTMC-simulations with and without the Coulomb-potential included. It is obvious from the comparison of Figs. \ref{fig:Compar} (c) (Coulomb-potential not included in simulation) and (d) (Coulomb-potential included) that the inclusion of the Coulomb potential is required to reproduce the experimental data.
Type A and type B electrons have different characteristic position-space trajectories as shown in Fig. \ref{fig:Traj}. Type A electrons travel directly to the detector without revisiting $x=0$ and do not have a point of intersection in the $(x,y)$ plane. However, type B electron trajectories have a point of intersection in the $(x,y)$ plane and furthermore, they cross $x = 0$ multiple times.
\begin{figure}
\includegraphics{figsLatest/fig4.pdf}
\caption{Characteristic trajectories for different parts of the PMD. a) Characteristic trajectories from the central part of the PMD containing the focused photoelectrons (type B). The trajectories revisit $x=0$ multiple times and have a point of intersection in the $(x,y)$-plane. c) The photoelectrons from the outer part (type A) of the PMD go directly to the detector without revisiting $x=0$ and do not have a point of intersection in the $(x,y)$-plane.}
\label{fig:Traj}
\end{figure}
\section{Coulomb focusing at elliptical polarization}
\subsection{Creation of the sharp, thin line-shaped ridge structure in the 3D PMD}
\label{sec:AnaOne}
\begin{figure}
\includegraphics{figsLatest/Points.pdf}
\caption{Comparison of central parts of PMDs obtained by CTMC for different ellipticities: $\epsilon = 0$, $\epsilon = 0.07$ and $\epsilon = 0.11$, respectively.
Characteristic points of the PMD depending on the specific longitudinal momenta, correspond in the linear polarization case, (B2) to the trajectory with two rescatterings, (B3) with three, (B4) with four rescatterings, and (B1) with a single rescattering. (B2) and (B4) correspond to the slow recollision condition ($p_{xr}=0$). (A) corresponds to the center of the main lobe and is originating from recollision-free trajectories.
Horizontal caustics are visible around, e.g., the characteristic points B2 and B4. Furthermore, a vertical caustic due to Coulomb focusing is visible as the line through the points $Bn$.
For $\epsilon = 0.11$, the vertical caustic at point B3 is split.}
\label{fig:01-comparison_of_PESs}
\end{figure}
To understand the nature of this ridge structure in more detail, we analyze the 3D PMD in the polarization plane (i.e. in the $(p_x, p_y)$-plane). We begin the analysis in the simpler case of a linearly polarized laser field and transfer the results to the case of elliptical polarization with $\epsilon>0$. Fig. \ref{fig:01-comparison_of_PESs} shows the results of CTMC simulations in dipole approximation for three different ellipticities: $\epsilon = 0$, $0.07$ and $0.11$. To focus on the mechanism of the creation of the ridge, we consider only electrons starting in the central half-cycle of the laser field to suppress the influence of ionization from multiple half cycles. Furthermore, we neglect the magnetic field effects to disentangle the creation of the ridge in the $(p_x, p_y)$-plane from additional effects in beam propagation $p_z$-direction.
As we saw in section \ref{sec:PMD3D}, the ridge already appears within the dipole approximation because the CTMC simulated PMD in the polarization plane did not depend on the magnetic field.
In the case of linear polarization, the features of the PMD are understood in the terms of laser-driven classical trajectories recolliding with the parent ion. Due to the nature of the Coulomb interaction, a bunching of electrons occurs and is imprinted on the PMD in the form of caustics (Fig.~\ref{fig:01-comparison_of_PESs}). There are two kinds of caustics: horizontal and vertical. Each horizontal caustic line in the PMD corresponds to a certain class of rescattered trajectories. When the longitudinal momentum $p_x$ of the electron at the recollision is vanishing, i.e. the slow recollision condition is fullfilled, longitudinal bunching of electrons occurs \cite{Kastner_2012}. This condition depends on the ionization phase (i.e. the phase of the laser electric field when the electron appears in the continuum).
In Fig. \ref{fig:Traj}, we indicate several characteristic points exhibiting qualitatively different rescattering behavior. Point A is an example of a type A electron, while points $Bn$ indicate type B electrons with n rescattering events. For these cases the sharp peak in the distribution around $p_y=0$, can be seen due to Coulomb focusing. This peak is pronounced for small ellipticities, but starts to disappear for $\epsilon \gtrsim 0.1$.
In order to understand this vertical ridge further, we analyze the set of initial transverse momentum distributions (i.e. at the tunnel exit) corresponding to the final transverse momentum values given by points $Bn$. The resulting distributions are shown in Fig. \ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}. We analyze the initial transverse momentum distribution at the tunnel exit with momentum bins of $0.01\times0.01\times0.01$ a.u. placed at characteristic points of the PMD. We compare the initial PMDs of corresponding points for linear and elliptical polarization (Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}).
\begin{figure}
\centering
\includegraphics{figsLatest/fig6.pdf}
\caption{The initial momentum space distribution for trajectories ending in the momentum bins of $0.01\times0.01\times0.01$ dimensions at the asymptotic values of longitudinal momentum $p_x$: (a) $0.62$, (b) $0.44$ and (c) $0.3$ $a.u.$, corresponding to the points B2, B3, B4 of Fig.~\ref{fig:01-comparison_of_PESs}, respectively. The laser field is linearly polarized in (left column), and elliptically polarized, with $\epsilon = 0.07$, in (right column). Notice the positive offset in the $p_{yi}$ of the structures in the right column due to the ellipticity.}
\label{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}
\end{figure}
The ridge originates from a contraction in momentum transverse to the major polarization axis. In a linearly polarized laser field, the electrons contributing to the ridge are ionized with a nonvanishing transverse momentum at the tunnel exit, and appear after propagation with a vanishing transverse momentum. Their initial distribution is a ring in the ($p_y,p_z$)-initial momentum distribution, see Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}, left column.
in the case of a linearly polarized laser field, the electrons which are initially (i.e. at the tunnel exit) distributed inside this ring obtain a large momentum transfer during recollisions and end up outside of the chosen final momentum bin near the vanishing transverse momentum.
In the case of the elliptically polarized laser field, there are two modifications to this picture. For small ellipticities (quantified in Eq.~(\ref{small-ellipt}) below) the rescattering and Coulomb focusing, similar to the case of linear polarization, takes place for electrons which initially are distributed in a {\it shifted} ring of initial momenta in the ($p_{yi},p_{zi}$)-plane. The size of the shift is discussed further in section \ref{sec:AnaTwo}). The radius of the ring of the initial momentum space distribution is an indicator for Coulomb focusing. It is nearly the same for linear and elliptical polarization, i.e., Coulomb focusing for these trajectories is qualitatively the same.
The points B2 and B4 in Fig.~\ref{fig:01-comparison_of_PESs} corresponding to the slow recollision condition do not change their position in the PMD when changing ellipticity, which is due to the similarity of the underlying trajectories.
Next, we analyze these trajectories analytically to show that the Coulomb focusing dynamics is similar for linear and elliptical polarization up to a certain value of ellipticity (especially in the case of slow recollisions). Furthermore, we show that the recolliding electrons, which create the ridge, end up around $p_{yf} \approx 0$.
The underlying electron trajectories are obtained from the solution of the electron equations of motion in an elliptically polarized laser field assuming that the Coulomb field effect is a perturbation, which affects the electron trajectory near the tunnel exit and at recollisions. The electric field component of the laser field is
\begin{eqnarray}
E_x &=& E_0\cos \eta
\label{eq:laserField1}\\
E_y &=&\epsilon E_0\sin \eta.
\label{eq:laserField2}
\end{eqnarray}
with the phase $\eta $, the ellipticity $0\leq\epsilon\leq 1$, the field amplitude $E_0=\sqrt{I}/\sqrt{1+\epsilon^2}$, and the intensity $I$. The envelope of the pulse is neglected. For the electron dynamics in the laser polarization plane after the ionization, taking into account initial Coulomb momentum transfer at the tunnel exit, we have:
\begin{eqnarray}
p_x &=& -\frac{E_0}{\omega}\left(\sin \eta-\sin\eta_i\right)-\delta p_{xi}^C\label{px}\\
p_y &=& \epsilon\frac{E_0}{\omega}\left(\cos \eta-\cos\eta_i\right)+p_{yi}-\delta p_{yi}^C,
\label{elliptt-mom-y}
\end{eqnarray}
where $\eta_i$ is the ionization phase, $p_{yi}$ is the initial transverse electron momentum, and $\delta p_{xi}^C$ and $\delta p_{yi}^C$ are initial Coulomb momentum transfer.
The electron rescattering and Coulomb focusing in an elliptically polarized laser field will be similar to the case of linear polarization, when the Coulomb momentum transfer during recollision is the same in both cases: $\delta p_{yr}^{C(\epsilon)}=\delta p_{yr}^{C(0)}$.
For the latter it is necessary to have the same impact parameter of the recollision, i.e. the same recollision $y$-coordinate. This condition is fulfilled when the initial momentum at the tunnel exit is shifted with respect to the linear polarization case by a value to compensate the momentum imparted to the electron by the $y$-component of the laser field (as well as a small difference of the initial Coulomb momentum transfers).
The value of this shift at the slow recollision condition is:
\begin{eqnarray}
p_{yi}^{(\epsilon)} -p_{yi}^{(0)}=\epsilon\frac{E_0}{\omega }\cos\eta_i +\delta p_{yi}^{C(\epsilon)}-\delta p_{yi}^{C(0)}+\frac{\epsilon\delta p_{xi}^{C(\epsilon)}}{\eta_r-\eta_i} ,
\label{pyi}
\end{eqnarray}
see the derivation in the appendix \ref{sec:app}. Here, the superscripts $(\epsilon)$ and $(0)$ refer to the cases of elliptical and linear polarization, respectively. The final momentum of the type B electron is found by combining Eqs.~(\ref{elliptt-mom-y}) and (\ref{pyi}):
\begin{eqnarray}
p_{yf}^{(\epsilon)} =-\epsilon\frac{E_0}{\omega }\cos\eta_i+p_{yi}^{(\epsilon)}-\delta p_{yi}^{C(\epsilon)} -\delta p_{yr}^{C(\epsilon)} = \frac{\epsilon\delta p_{xi}^{C(\epsilon)}}{\eta_r-\eta_i },
\label{pyf}
\end{eqnarray}
where we have taken into account that $\delta p_{yr}^{C(\epsilon)}=\delta p_{yr}^{C(0)}$ and $p_{yi}^{(0)}=\delta p_{yi}^{C(0)}+\delta p_{yr}^{C(0)}$, i.e. that in the case of linear polarization the main ridge due to Coulomb focusing is at $p_y\approx 0$. The final momentum of the type B electron has practically a vanishing value. In fact, $\delta p_{xi}^{C(\epsilon)}\approx \pi E(\eta_i)/(2I_p)^{3/2}$ \cite{Shvetsov-Shilovski_2009}, and $\eta_r-\eta_i\sim 3\pi$ at the first slow recollision, and $\epsilon\delta p_{xi}^{C(\epsilon)}/(\eta_r-\eta_i)\sim 10^{-3}$, at $\epsilon\sim 0.1$, where $\eta_r$ is the phase of recollision. All numerical estimations in this section are for $\omega=0.013$ ($\lambda=3400$ nm), $E_0=0.04$ (intensity $5.8\times 10^{13}$ W/cm$^2$).
Thus, neglecting initial Coulomb momentum transfer for simplicity, we can conclude that in the case of elliptical polarization, the electrons with slow recollision condition are initially distributed in momentum space ($p_{yi},p_{xi}$) on the ring centered at
\begin{eqnarray}
p_{yi}^{(\epsilon)}\approx\epsilon\frac{E_0}{\omega }\cos\eta_i,\,\,\,\,
p_{zi}^{(\epsilon)}=0, \label{mshift}
\end{eqnarray}
with a radius $\delta p_{yr}^{C(\epsilon)}$
and finally will end up at the ridge around $p_y \approx 0$.
This is illustrated in Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}, right column, where the top (B2) and bottom (B4) panels correspond to the slow recollision condition.
The radii of the rings in B2 and B4, which indicate the magnitude of Coulomb focusing, do not change significantly due to the change in ellipticity. Thus, we can conclude that Coulomb focusing dynamics is very similar in both cases.
At non-negligible ellipticities the electrons around vanishing initial transverse momentum experience no recollision, no momentum change due to the Coulomb field besides the initial Coulomb momentum transfer when the electron leaves the tunnel exit. They contribute to the lobes of the final $(p_x,p_y)$-distribution, with the final momentum $\textbf{p}_f\approx -\textbf{A}(t_i)+\delta \textbf{p}_{i}^C$, with the initial Coulomb momentum transfer $\delta \textbf{p}_{i}^C$. The latter is mostly along the electric field for the experimental parameters (the transverse component of initial Coulomb momentum transfer is smaller with respect to the longitudinal component by an order of magnitude), i.e., perpendicular to the vector potential, and induces a distortion of the PMD ellipse with respect to the case of the simple man's model [$\textbf{p}_f\approx -\textbf{A}(t_i)$].
The initial momentum space corresponding to the center of the main lobe A is shown in Fig.~\ref{sidelobe} and indicates the absence of Coulomb focusing (the initial and the final phase space are the same).
\begin{figure}
\centering
\includegraphics{figsLatest/fig7.pdf}
\caption{The initial momentum space distribution of direct electrons without Coulomb focusing which are ending in the center of the main lobe, point A in Fig.~\ref{fig:01-comparison_of_PESs}. }
\label{sidelobe}
\end{figure}
Next we want to estimate up to which ellipticities the central sharp ridge of Coulomb focusing electrons still appears in the PMD.
The discussion above is valid if the required initial transverse momentum of type B electrons according to Eq.~(\ref{pyi}), $p_{yi}^{(\epsilon)}\sim \epsilon E_0/\omega+\delta p_{yr}^{C(\epsilon)}$, is within the momentum width tunnelled electron wave packet. The latter reads
\begin{eqnarray}
\epsilon\frac{E_0}{\omega }\lesssim \Delta_\bot,
\label{epsilon-central}
\end{eqnarray}
where $\Delta_\bot= \sqrt{E_0}/(2I_p)^{1/4}$ is the PMD width ($2\sigma$ of the Gaussian distribution) at the tunnel exit according to tunnel-ionization theory\cite{PPT,ADK}
With Eq.~(\ref{epsilon-central}) we conclude that the ridge in the PMD can exist up to ellipticities
\begin{eqnarray}
\epsilon\lesssim \frac{\omega }{\sqrt{E_0}(2I_p)^{1/4}}\approx 0.07.
\label{small-ellipt}
\end{eqnarray}
There are also modifications of the rings in Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}, especially for B3, which we discuss later in the next section based on trajectory analysis of these characteristic points.
Note that the inner rings in the linear case in Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3} are caused by trajectories with multiple significant rescattering events. However, such trajectories are strongly suppressed for increasing ellipticity.
Finally, let us estimate at which ellipticity the side lobes will be separated from the sharp ridge in the PMD. The side lobes appear when the ellipticity is large enough such that the Coulomb momentum transfer at recollision, for the electron with initial $p_{yi}=0$, is negligible with respect to the final momentum. The final electron momentum can be estimated as
\begin{eqnarray}
p_{yf} \approx \epsilon \frac{E_0}{\omega},
\end{eqnarray}
and the Coulomb momentum transfer as
\begin{eqnarray}
\delta p_{yr}^C\approx -\frac{y_r}{R_r^3} \delta t_r\approx \frac{\delta t_r}{y_r^2},
\end{eqnarray}
where $x_r$, $y_r$, $z_r$, $R_r=\sqrt{x_r^2+y_r^2+z_r^2}$ are the electron coordinates, and the distance from the core at the recollision point, respectively. $\delta t_r$ is the recollision duration:
$\delta t_r\approx \sqrt{2y_r/E_x(t_r)}$.
For the latter, we assumed that during the recollision time the electron travels a distance of the order of $R_r$ in x-direction , i.e., $\delta x\approx R_r\approx y_r\approx E_x(t_r)\delta t^2/2$ when we apply the condition for slow recollision. Estimating the recollision coordinate as
$y_r\approx \epsilon E_0/\omega^2$,
we have for the Coulomb momentum transfer
$\delta p_{yr}^C\approx \sqrt{2}\omega^3/(\epsilon^{3/2}E_0^2)$.
Thus, Coulomb focusing at recollision will be negligible, $\delta p_{yr}^C\ll p_{yf}$, if the ellipticity is relatively large
\begin{eqnarray}
\epsilon \gtrsim \left(\frac{\sqrt{2}\omega^4}{E_0^3}\right)^{2/5}\approx 0.05.
\label{eq:cond_on_epsilon_for_side-lobes}
\end{eqnarray}
At ellipticities larger than this value, the lobes in the $(p_x, p_y)$-plane begin to appear.
\begin{figure}
\centering
\includegraphics{figsLatest/traj_P1-1.pdf}
\includegraphics{figsLatest/traj_P2-1.pdf}
\includegraphics{figsLatest/traj_P3-1.pdf}
\caption{Typical photoelectron trajectories in a laser field with $\epsilon = 0.07$, and Coulomb momentum transfers initially and during subsequent recollisions. The trajectories originate on the left (left panels) and right part (right panels) of the initial transverse momentum distribution ring with $p_{zi} = 0$ and end up at the same point B2 (upper group), B3 (middle group), and B4 (bottom group). The middle panels of trajectories show the case of linear polarization.}
\label{fig:03-trajectories-e_0_07-P1}
\end{figure}
\subsection{The structure of the sharp, thin line-shaped ridge in the 3D PMD}
\label{sec:AnaTwo}
To investigate the structure of the sharp ridge, (Fig.~\ref{fig:Surface}), we analyse in this section the corresponding trajectories. The structure of the ridge is very similar to the linear polarization case at the slow recollision condition. The structure slightly deviates from the linear polarization case outside of the slow recollision condition.
In Fig.~\ref{fig:03-trajectories-e_0_07-P1}, we analyze typical trajectories corresponding to each of the points B2, B3, B4, and originating from the left $(-)$ and right $(+)$ parts of the ring of initial momentum distribution (Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}). We choose left and right points which have the same final vanishing momentum,
\begin{eqnarray}
p_{yf}^-=p_{yf}^+.\label{==}
\end{eqnarray}
The difference of the initial momenta of these points determines the diameter of the ring of the initial momentum distribution of Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}, which depends on Coulomb momentum
transfer during recollision, and is an indicator of Coulomb focusing:
\begin{eqnarray}
D\equiv p_{yi}^+-p_{yi}^-=\left(\delta p_{yr}^{C+} + \delta p_{yr}^{C-}\right)\left(1+\frac{2E(\eta_i)}{(2I_p)^2} \right),\label{DDD}
\end{eqnarray}
see the derivations for this section in Appendix \ref{sec:app2}. The center of the ring is given by
\begin{eqnarray}
\frac{p_{yi}^++p_{yi}^- }{2} = \left(\epsilon\frac{E_0}{\omega}\cos\eta_i+\frac{\delta p_{yr}^{C+}}{2}-\frac{\delta p_{yr}^{C-}}{2} \right) \left(1+\frac{2E(\eta_i)}{(2I_p)^2} \right),
\end{eqnarray}
which includes initial Coulomb momentum transfer and Coulomb momentum
transfer during recollision correction into Eq.~(\ref{mshift}), and indicates that in the case of elliptical polarization the ring of the initial momentum distribution for the sharp ridge electrons is shifted due to the drift momentum along the minor axis of polarization, and by initial Coulomb momentum transfer in that direction as $\delta p_{yr}^{C+}\approx \delta p_{yr}^{C-}$:
\begin{eqnarray}
p_{yi}^{(\epsilon)}
\approx \epsilon\frac{E_0}{\omega }\cos\eta_i\left[1 +\frac{2 E_0\cos\eta_i}{(2I_p)^{2}}\right],
\end{eqnarray}
where $p_{yi}^{(\epsilon)}$ describes the shift of the initial momentum along the $p_y$-axis at an ellipticity $\epsilon$ compared to the linear case.
For an ellipticity of $\epsilon = 0.07$, we have $p_{yi}^{(\epsilon)}
\approx 0.23$.
\begin{figure}
\includegraphics{figsLatest/fig9.pdf}
\caption{The initial momentum distributions for $\epsilon = 0.11$ leading to the two distinct ridges [B3(l) and B3(r)] taken at the same value of longitudinal momenta. The pronounced left branch consists of electrons following linear-like trajectories, whereas the faint right branch comes from trajectories strongly influenced by the ellipticity. }
\label{fig:07-initial_phase_space-e_0_11_P2}
\end{figure}
The diameter ($D$) of the ring in the initial momentum distribution, the indicator of Coulomb focusing, depends on Coulomb momentum
transfer during recollision, but not on initial Coulomb momentum transfer $\delta p_{yi}^{C\pm }$ (see appendix \ref{sec:app2}, Eq.~(\ref{DDD})). For all three points B2, B3, B4, the Coulomb momentum
transfer during recollision is approximately the same as in the case of linear polarization, cf. Fig. \ref{fig:03-trajectories-e_0_07-P1}. Therefore, the radius of the ring in the initial transverse momentum distribution and the ridge in the final PMD are also approximately the same.
We next analyze the modifications of the ridge with the variation of ellipticity. For the points B2 and B4 and linear polarization the main Coulomb momentum
transfer during recollision takes place at the slow recollision, $p_{xr}=0$. This corresponds to the second recollision for B2, and the fourth one near B4, although there are multiple recollision points, $x_r=0$, two at B2, and four at B4 (see Fig. \ref{fig:03-trajectories-e_0_07-P1}). In the elliptical polarization case, the Coulomb focusing for the trajectory coming from the left part of the ring resembles the linear case, since the slow recollision has the same impact parameter and the same Coulomb momentum
transfer during recollision. Of the multiple recollisions that occur for these points, the slow one has the dominant effect on the final momentum. Therefore, the Coulomb focusing for the left half-ring in Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3} is the same for both linear and elliptical polarizations, thereby creating the central ridge in the PMD in both cases.
In contrast, the trajectory from right part of the ring at B2 in the elliptical polarization case differs from the linear one, see Fig.~\ref{fig:03-trajectories-e_0_07-P1}, upper panel. The first rescattering for this trajectory, additional to the slow recollision, takes place with shorter impact parameter than in the linear case due to the oscillating part of the $y$-coordinate, yielding to the increase of the total Coulomb momentum transfer. This explains the larger radius of the right half-cycle of the ring structure at B2 in Fig.~\ref{fig:02-initial_phase_space-e_0_and_0_07_P1_P2_P3}. The right-type trajectories are more sensitive to the initial conditions and, consequently, the ring width is significantly smaller. Moreover, the ionization probability is smaller for the right part of the ring because of the lager initial momenta. These explain why the central ridge becomes less pronounced with increasing ellipticity.
However, the right trajectories for B4, and trajectories ending up on the central ridge at $|p_x|$ smaller than for B4, start to resemble the linear case again.
This is because for the ionization phases that are relevant to the latter trajectories, the oscillating part of the $y$-coordinate does not perturb the recollision coordinates significantly. The condition that the recollision coordinate is only weakly perturbed by the ellipticity, i.e. $|y_r^{(\epsilon)}-y_r^{(0)}|\ll y_r^{(0)}$,
leads to an estimate of the threshold ionization phase (see appendix \ref{sec:app3}):
\begin{eqnarray}
\eta_i\ll \pi \left(\frac{\delta p_{y}^{C}}{(\epsilon E_0/\omega)} \right)^2\approx 0.68,
\end{eqnarray}
for $\delta p_{y}^{C} = 0.1$. The numerical value in the equation above is estimated for the parameters of Fig.~\ref{fig:01-comparison_of_PESs}, i.e., at $\eta_i\lesssim 0.068$, which assumes $p_{x}\lesssim E_0\eta_i/\omega \approx 0.2 $, the recollisions and Coulomb focusing in the elliptical polarization case will be similar to the linear one. This estimation fits to the CTMC calculation in Fig.~\ref{fig:01-comparison_of_PESs}.
The ring of the initial momenta of point B3 is deformed in a stronger way then the one for B2. The deformation is due to the fact that both left and right trajectories are perturbed with respect to the linear polarization case because of the quiver motion in the transversal $y$-direction. The perturbed trajectories are more sensitive to the initial conditions, which results in a variable width of the ring in the initial momentum distribution. Moreover, the recollision coordinates for both left and right trajectories are different. Therefore, the Coulomb momentum
transfer during recollision for the left and right trajectories are not symmetric which leads to the bend of the central ridge. Furthermore, at larger ellipticities, e.g. $\epsilon=0.11$, the central ridge at B3 is split, when the left- and right-side trajectories yield to different ridges, see Fig.~\ref{fig:07-initial_phase_space-e_0_11_P2}. However, this splitting is not visible in the experimental data due to focal volume averaging, CEP averaging and the laser pulse shape.
\section{Non-dipole effects}
\label{sec:noDipole}
The PMDs in Ref. \cite{Ludwig_2014} recorded at mid-IR wavelengths showed a strong influence of the Coulomb-potential in combination with the magnetic field component of the laser field due to rescattering.
\begin{figure*}
\includegraphics{figsLatest/ellScanProj.pdf}
\caption{Normalized measured 2D PMDs from xenon recorded for various ellipticities at a peak intensitiy of $6 \cdot 10^{13}$ W/cm$^2$.}
\label{fig:ellScan1}
\end{figure*}
\begin{figure}
\includegraphics{figsLatest/SplitProj.pdf}
\caption{Projection of 3D PMD for $p_x<0.05$ a.u. recorded at a peak intensity of $6 \cdot 10^{13}$ W/cm$^2$ and an ellipticity of $\epsilon =0.11$ (left). The sharp line structure is clearly visible around $p_y = 0$. The ranges of $p_y$ for the projection in are marked with dashed lines. The central spot stemming from ionization of Rydberg atoms was removed prior to the projections (see main text for details).
Projections of the photoelectrons onto the beam propagation axis for two different regions of $p_y$ (right). The magenta crosses show the photoelectron distribution for $|p_y|<0.05$ a.u. The blue crosses show the distribution for $|p_y|>0.05$ . The peak of the inner electrons is shifted opposite to the beam propagation direction whereas the outer electrons are shifted in beam propagation direction. The black lines serve as a guide for the eye.}
\label{fig:Shift3D}
\end{figure}
In order to minimize the influence of rescattering processes, we studied the projection of the PMD onto the beam propagation axis for the case of close-to-circular polarization for helium and xenon. We extract the zero momentum from reference images recorded with linear polarization as described in section \ref{sec:exp}.
The peaks of these projections were extracted via a Gaussian fit to the projection of the PMD onto the $p_z$-axis. The results are shown together with a model from Ref. \cite{Smeenk_2011} for circular polarization in Fig. \ref{fig:ellShift}.
The results for circular polarization are consistent with the results from Ref. \cite{Smeenk_2011} within the error bars. Theoretical studies predicted an additional offset of the order of $\frac{I_p}{3c}$ at the tunnel exit \cite{Klaiber_2013c} as well as in the final momentum distributions \cite{Chelkowski_2014,Cricchio_2015,Chelkowski_2015}.
However, the experiment cannot resolve this additional offset.
In addition we show for comparison the data for linear polarization from Ref. \cite{Ludwig_2014}. The offsets for the case of linear polarization are shifted in contrast to the data for circular polarization opposite to the beam propagation direction.
The results with linear and circular polarization (Fig. \ref{fig:ellShift} (a)) trigger the question what happens when the ellipticity is being changed?
Since our previous results \cite{Ludwig_2014} indicate that the non-dipole effects are independent of the target gas within the accuracy of our measurement. We therefore perform our studies on xenon.
The ellipticity was varied in steps from linear to close-to-circular ($\epsilon=0.97$). We ionized xenon atoms at an intensity of $6 \cdot 10^{13}$\,W/cm$^2$ with 50-fs pulses at a center wavelength of 3.4 $\mu$m. For each ellipticity step, projected momentum images in the ($p_x$, $p_z$) plane were recorded. We would like to point out that during the measurement we kept the intensity constant, not the electric field. That allows us to keep the total momentum transfer per cycle onto a free electron from the field independent of the ellipticity.
Fig. \ref{fig:ellScan1} shows how the projected PMD is evolving from the typical cigar-like shape with the major dimension in $p_x$-direction towards a structure with maxima shifted to high values of $p_x$. Furthermore, the spot in the center stemming from Rydberg-atoms ionized by the spectrometer field disappears with increasing ellipticity because the selection rules do not allow the excitation from the ground state into Rydberg states with circularly polarized light.
Thus, reference measurements with linear polarization were taken right before and after each ellipticity step. The peaks for reference zero were extracted in a similar fashion as the method applied in Ref. \cite{Ludwig_2014} and the one described in section \ref{sec:exp}. For each reference measurement, we recorded $N=10$ photoelectron momentum images. The uncertainties of the determination of zero momentum scale with $\frac{1}{\sqrt{N}}$ due to $N$ repetitive measurements that were weighted with the error for a Gaussian fit of the peak of each projected distribution.
The zero momentum reference was extracted from the momentum images with linear polarization from the peak of the projection of a thin slice of $|p_x| \le 0.05$ a.u. around $p_x = 0$ projected onto the beam propagation. The peak of the projection was identified via a polynomial fit to the central part (i.e. $\Delta p_y \approx 0.05$ a.u.). of the PMD $W(p_z) = \iint_\Omega W(p_x,p_y,p_z) dp_x dp_y$, with $\Omega$ being the integration momentum volume. The offset of that peak from $p_z = 0$ is shown in Fig. \ref{fig:ellShift}(b) as a function of the ellipticity.
We observe an increase of the offset with increasing ellipticity from negative values (i.e. opposite to the beam propagation direction) to positive values (i.e. in beam propagation direction). The transition from negative to positive values occurs at an ellipticity of $\epsilon = 0.12$. Thus, we have observed a zero-crossing along the $p_z$-axis of the peak of the projected PMD as we scan the ellipticity.
Next, we consider the role of Coulomb focusing in the electronic response to non-dipole effects.
Thus we study the different responses of type A and type B electrons. Since we are able to isolate the electrons undergoing Coulomb-focusing in 3D momentum-space, we can study their response separately from the mostly unfocused electrons for $\epsilon\approx 0.1$.
We select type B – electrons by choosing a narrow momentum range of 0.038 a.u. in $p_y-$direction from the 3D PMD recorded at an ellipticity of $\epsilon = 0.11$. The electron signal that lies within this range and the electron signal from outside this range are separately projected onto the $p_z$-axis.
The central spot stemming from ionization of Rydberg atoms was removed prior to the projections by removing the photoelectron signal with $|p|<0.03$ a.u..
The normalized projected electron signals are shown together in Fig. \ref{fig:Shift3D}. The position of the peaks from A and type B electron signals on the $p_z$-axis were identified. We observe that the type A electron signal peaks at a positive value of $p_z$ and the type B electron signal at a negative value of $p_z$.
Increasing ellipticity supresses rescattering and thus, the PMD and its projection becomes dominated by electrons that interact only weakly with the Coulomb-potential of the parent ion. However, for ellipticities $\epsilon \lesssim 0.12$, the sharp ridge consisting of type B electrons starts to dominate the projection of the PMD onto the $p_z$-axis. This creates a peak of electrons that significantly interact with the Coulomb-potential of the parent ion. That peak is shifted opposite to beam propagation direction.
\section{Conclusions}
In conclusion, we are able to separate the electrons that experienced strong Coulomb focusing in momentum space from the unfocused electrons in the case of elliptical polarization.
With a detailed analysis, we identified a subspace in the momentum representation of the tunnel ionized electron wave packet, which in an elliptically polarized laser field shows recollision dynamics similar to the linear polarization case. The electron trajectories originating in this initial PMD subspace have multiple revisits, including at least one significant rescattering event with the parent ion, which causes the large initial momentum subspace to be squeezed into a small final momentum space having near vanishing transverse momentum.
This effect occurs due to Coulomb focusing, and leads to the creation of the sharp ridge structure in momentum space for linear and small elliptical polarization.
The final momenta of Coulomb-focused electrons that underwent slow recollisions, and hence experienced strong Coulomb focusing, are almost unaltered by the introduction of a small ellipticity. The ellipticity modifies the initial momenta of the electrons that end up in the ridge. We also learned that the central part of the tunnel ionized electron wave packet in momentum space provides electrons which are steered away by the elliptical polarization into elliptical side lobes of the final PMD.
Thus we were able to separate the unfocused from the Coulomb-focused electrons.
Dispersion of the electron trajectories in final momentum was already reported in \cite{Shafir_2013}. However, there was no clear separation of the electrons undergoing Coulomb-focusing and the ones that interact only weakly with the Coulomb-field of the parent ion in momentum space.
For small ellipticities we have shown that the electrons undergoing strong Coulomb-focusing accumulate in a sharp ridge in the 3D PMD that is largely separated from the unfocused electrons. This clear separation enables us to analyze non-dipole effects on Coulomb-focused electrons alone.
We use the full 3D PMD to study the different non-dipole response for both electrons with and without strong Coulomb-focusing.
The separation of the electron trajectories allows us to disentangle the response of the photoelectrons from non-dipole effects and to make the connection between prior results \cite{Ludwig_2014, Smeenk_2011}. In Ref. \cite{Smeenk_2011}, a shift of the PMD in beam propagation direction was reported and a simple radiation pressure picture was used as an explanation. In Ref. \cite{Ludwig_2014}, the peak of the projection of the PMD was shifted opposite to the beam propagation direction. Here, we were able to experimentally isolate the electrons that are responsible for this counterintuitive shift and to separately study their non-dipole response.
We measured the ellipticity dependence of the non-dipole effects in PMDs in strong field ionization beyond the long-wavelength limit. We observed an offset of the maximum of the peak of the projection of the PMD on the beam propagation axis. By increasing the ellipticity from linear to circular, this offset shifts from negative to positive values of $p_z$.
Thus, we conclude that the formation of Coulomb-focused structures in PMDs in combination with the magnetic field causes peaks in negative $p_z$-direction. The appearance of a peak on the negative or positive side of $p_z = 0$ can be considered as a competition between two types of photoelectrons which we refer to as type A and type B electrons (Fig. \ref{fig:Traj}).
Type B photoelectrons form the sharp thin line-shaped ridge structure in the 3D PMD and type A photoelectrons are in the clearly separated lobes at higher final momentum. When we analyse the shift of the PMD peak along the laser beam propagation direction we could confirm that the type A electrons result in a positive shift and the type B electrons in a negative shift.
In case that the peak is formed mainly by type A-electrons, the projection of the PMD peaks at positive values of $p_z$ whereas in the case that the type B electrons dominate the peak of the projected PMD, it is shifted towards negative values of $p_z$.
Thus, Coulomb focusing has an essential influence on non-dipole effects and creates an ellipticity-dependent PMD shift along the laser propagation direction on the ellipticity of the laser field.
Our results open up new possibilities to separately study the photoelectrons experiencing strong rescattering from the direct electrons and the electrons that experience only weak Coulomb interaction.
Furthermore, as the occurrence of the Coulomb focusing effects also depend on the width of the returning electron wave packet, our studies open up the possibilities to gain insight into the precise structure of the returning electron wave packet.
|
\section{Formulation of the problem and of the results}
The classical Descartes' rule of signs states that a real polynomial in one
variable has not more real positive roots than the number of sign changes
in the sequence of its coefficients.
Any sequence of $\pm$-signs
$\bar{\sigma}:=(\sigma _0,\sigma _1,\ldots ,\sigma _d)$ is called a
{\em sign pattern}. In the present paper
we consider sign patterns defined by the signs of the coefficients of degree
$d$ polynomials $P$, so in particular $\sigma _d=$sign$(P(0))$.
For a given sign pattern its
{\em Descartes' pair}
$(p_{\bar{\sigma}},n_{\bar{\sigma}})$ is the number of
sign changes and sign preservations in the sequence of coefficients. Denote
by $(pos_P,neg_P)$ the numbers of positive and negative roots of $P$
counted with multiplicity. Hence the following restrictions must hold true:
\begin{equation}\label{pn}
pos_P\leq p_{\bar{\sigma}}~,~neg_P\leq n_{\bar{\sigma}}~,~
pos_P\equiv p_{\bar{\sigma}}({\rm mod}~2)~,~neg_P\equiv ,n_{\bar{\sigma}}({\rm mod}~2)~.
\end{equation}
(The inequality $neg_P\leq n_{\bar{\sigma}}$ follows from Descartes' rule
applied to the polynomial $P(-x)$.) Pairs $(pos,neg)$ satisfying conditions
(\ref{pn}) are called {\em admissible} for the sign pattern $\bar{\sigma}$
(and the latter is {\em admitting} them).
The present paper finishes the study which was begun in
\cite{FoKoSh} of sign patterns and their
admissible pairs for polynomials of degree up to $8$. The present
introduction reproduces with some small
modifications the one of \cite{FoKoSh} and the results
obtained in that paper,
see Theorems~\ref{tmknown}, \ref{tmd7} and~\ref{tmd8}. The new results
are given in Theorem~\ref{tmnew} and then
presented in another way (suitable to be
compared to the previously obtained ones) at the end of this section.
Clearly conditions (\ref{pn}) are only necessary ones, i.e. for a
given sign pattern $\bar{\sigma}$ and an admissible pair
$(p,n)$ it is not a priori clear whether there exists a degree $d$
polynomial with this sign pattern and with exactly $p$ distinct positive and
exactly $n$ distinct negative roots. If such a polynomial exists, then we
say that the given combination of sign pattern and admissible pair
{\em is realizable}.
\begin{nota}
{\rm For a given sign pattern $\bar{\sigma}$ we define its corresponding
{\em reverted} sign pattern $\bar{\sigma}^r$ as $\bar{\sigma}$
read from the back and by
$\bar{\sigma}_m$ the sign pattern obtained from the given one by changing the
signs in second, fourth, etc. position while keeping the other signs the same.
If $\bar{\sigma}$ is
defined by a degree $d$ polynomial
$P(x)$, then $\bar{\sigma}^r$ is the sign pattern of $x^dP(1/x)$ and
$\bar{\sigma}_m$ is the one of $(-1)^dP(-x)$.}
\end{nota}
\begin{ex}
{\rm For $d=4$ the sign pattern $(+,-,-,-,+)$ is equal to
$(+,-,-,-,+)^r$ and one has $(+,-,-,-,+)_m=(+,+,-,+,+)=(+,-,-,-,+)_m^r$.
For $d=8$ the sign pattern $(+,+,-,+,-,-,-,-,+)$ is equal
to $(+,+,-,+,-,-,-,-,+)^r_m$.}
\end{ex}
\begin{rems}
{\rm (1) In what follows we assume that the leading coefficients of
the polynomials are positive, so sign patterns (except in some places of
the proofs) begin with $+$.
(2) It is clear that $(\bar{\sigma}^r)^r=\bar{\sigma}$,
$(\bar{\sigma}_m)_m=\bar{\sigma}$ and $(\bar{\sigma}^r)_m=(\bar{\sigma}_m)^r$
(so we write simply $\bar{\sigma}^r_m$).
(3) The sign patterns and admissible pairs $(\bar{\sigma},(p,n))$,
$(\bar{\sigma}^r,(p,n))$, $(\bar{\sigma}_m,(n,p))$ and
$(\bar{\sigma}^r_m,(n,p))$ are realizable or not simultaneously. Therefore
it makes sense to consider the question of realizability of given
sign patterns with given admissible pairs modulo the standard
$\mathbb{Z}_2\times \mathbb{Z}_2$-action defined by
$\bar{\sigma} \mapsto \bar{\sigma}^r$ and
$\bar{\sigma}\mapsto \bar{\sigma}_m$.}
\end{rems}
It seems that for the first time the question of realizability of sign patterns
with admissible pairs has been asked in \cite{AJS}. In \cite{Gr} Grabiner
has obtained the first example of nonrealizability. Namely, he has shown that
for $d=4$ the sign pattern $(+,-,-,-,+)$ is not realizable with the
admissible pair $(0,2)$ (Descartes' pair of the pattern equals $(2,2)$).
In \cite{AlFu} Albouy and Fu have given the exhaustive answer to this question
of realizability for degrees not greater than $6$. In Theorems~\ref{tmknown},
\ref{tmd7} and \ref{tmd8}
we change at some places (w.r.t. the original formulations in \cite{AlFu}
or \cite{FoKoSh}) a sign pattern $\sigma$ to $\sigma _m$ and the
corresponding pair $(p,n)$ to $(n,p)$ in order to have mostly pairs of the form
$(0,n)$ in the formulations:
\begin{tm}\label{tmknown}
(1) For degree $1$, $2$ and $3$, any sign pattern is realizable with any of its
admissible pairs.
(2) For degree $4$ the only case of nonrealizability (up to the standard
$\mathbb{Z}_2\times \mathbb{Z}_2$-action) is the one of Grabiner's example.
(3) For degree $5$ the only such case is given by the sign pattern
$(+,-,-,-,-,+)$ with the pair $(0,3)$.
(4) For degree $6$ the only such cases are:
$(+,-,-,-,-,-,+)$ with $(0,2)$ or $(0,4)$; $(+,-,+,-,-,-,+)$ with $(0,2)$;
$(+,+,-,-,-,-,+)$ with $(0,4)$.
\end{tm}
The cases $d=7$ and $d=8$ have been considered in \cite{FoKoSh}.
The exhaustive answer to the question of realizability for $d=7$ is
as follows:
\begin{tm}\label{tmd7}
For $d=7$ there are $1472$ cases (modulo the standard
$\mathbb{Z}_2\times \mathbb{Z}_2$-action) of sign pattern and admissible
pair. Of these exactly $6$ are not realizable:
$(+,+,-,-,-,-,-,+)$, $(+,+,-,-,-,-,+,+)$ and $(+,+,+,-,-,-,-,+)$ with
$(0,5)$;
$(+,-,-,-,-,+,-,+)$ with $(0,3)$;
$(+,-,-,-,-,-,-,+)$ with
$(0,3)$ and $(0,5)$.
\end{tm}
For $d=8$ the partial answer from \cite{FoKoSh}
can be summarized by the following theorem. In
\cite{FoKoSh} this result is formulated differently, but equivalently.
In particular, the
authors of \cite{FoKoSh} have not noticed that the number of cases
for which the
answer still remained unknown can be decreased by one due to the standard
$\mathbb{Z}_2\times \mathbb{Z}_2$-action.
\begin{tm}\label{tmd8}
(1) For $d=8$ there are $3648$ possible combinations of sign pattern
and admissible pair (up to the standard
$\mathbb{Z}_2\times \mathbb{Z}_2$-action). Of these exactly $13$ are known to
be nonrealizable:
\noindent $(+,+,-,-,-,-,-,+,+)~~,~~(+,-,-,-,-,-,-,+,+)~~,~~
(+,+,+,+,-,-,-,-,+)$ and $(+,+,+,-,-,-,-,-,+)$ with $(0,6)$;
\noindent $(+,-,+,-,-,-,+,-,+)$ and $(+,-,+,-,+,-,-,-,+)$ with $(0,2)$;
\noindent $(+,-,+,-,-,-,-,-,+)$ and $(+,-,-,-,+,-,-,-,+)$
with $(0,2)$ and $(0,4)$;
\noindent $(+,-,-,-,-,-,-,-,+)$ with $(0,2)$, $(0,4)$ and $(0,6)$.
(2) For exactly another $6$ cases it is not known whether they are realizable
or not (we list the sign patterns and their reverted ones which will be needed
later):
$$\begin{array}{lllll}
{\rm Case~1:}&&\sigma _1:=(+,+,+,-,-,-,-,+,+)&&{\rm with}~~(0,6)\\
&&&&\sigma _1^r=(+,+,-,-,-,-,+,+,+)\\
{\rm Case~2:}&&\sigma _2:=(+,+,-,+,-,-,-,+,+)&&{\rm with}~~(4,0)\\
&&&&\sigma _2^r=(+,+,-,-,-,+,-,+,+)\\
{\rm Case~3:}&&\sigma _3:=(+,+,-,+,-,+,-,-,+)&&{\rm with}~~(4,0)\\
&&&&\sigma _3^r=(+,-,-,+,-,+,-,+,+)\\
{\rm Case~4:}&&\sigma _4:=(+,+,+,-,-,+,-,+,+)&&{\rm with}~~(4,0)\\
&&&&\sigma _4^r=(+,+,-,+,-,-,+,+,+)\\
{\rm Case~5:}&&\sigma _5:=(+,+,+,+,-,+,-,-,+)&&{\rm with}~~(4,0)\\
&&&&\sigma _5^r=(+,-,-,+,-,+,+,+,+)\\
{\rm Case~6:}&&\sigma _6:=(+,+,-,+,-,-,-,-,+)&&{\rm with}~~(4,0)\\
&&&&\sigma _6^r=(+,-,-,-,-,+,-,+,+)~.
\end{array}$$
\end{tm}
The aim of the present paper is to definitely settle the case $d=8$. Namely, we
prove the following theorem:
\begin{tm}\label{tmnew}
The $6$ cases of part (2) of Theorem~\ref{tmd8} are not realizable.
\end{tm}
For Case~1 the proof is given in Section~\ref{seccase1}.
Cases~2-6 are considered in Section~\ref{seccases26}. The proofs of
Lemmas~\ref{basiclemma} and \ref{nextlemma} formulated in
Section~\ref{seccases26} are given in the Appendix.
In the proof of the
theorem we sometimes use sign patterns having as components not only $+$
and/or $-$, but also $0$ (in the sense that the corresponding
coefficient equals $0$), and in some cases $\pm$ meaning that
we consider the cases with $+$ and $-$ together.
As we see, in all cases of nonrealizability one of the components of the
admissible pair equals $0$. The same is true for $d=9$ and $10$,
see \cite{FoKoSh}. To finish this section we list the nonrealizable cases
for $d=8$ by their pairs $(p,n)$; the third column contains the corresponding
Descartes' pair. In order to have only the pairs $(0,2)$, $(0,4)$ and $(0,6)$
as defining the classification we change the sign patterns $\sigma _j$ of
Cases 2-6 of Theorem~\ref{tmd8} to the corresponding patterns $(\sigma _j)_m$.
To find easier Cases 1-6 in the table we give their numbers as indices to
the corresponding sign patterns.
$$\begin{array}{llllll}
(0,2)&(+,-,+,-,-,-,+,-,+)&(6,2)&&(+,-,+,-,+,-,-,-,+)&(6,2)\\
&(+,-,+,-,-,-,-,-,+)&(4,4)&&(+,-,-,-,+,-,-,-,+)&(4,4)\\
&(+,-,-,-,-,-,-,-,+)&(2,6)&&&\end{array}$$
$$\begin{array}{llllll}
(0,4)&(+,-,+,-,-,-,-,-,+)&(4,4)&&(+,-,-,-,+,-,-,-,+)&(4,4)\\
&(+,-,-,-,-,-,-,-,+)&(2,6)&&(+,-,-,-,-,+,-,-,+)_2&(4,4)\\
&(+,-,-,-,-,-,-,+,+)_3&(2,6)&&(+,-,+,+,-,-,-,-,+)_4&(4,4)\\
&(+,-,+,-,-,-,-,+,+)_5&(4,4)&&(+,-,-,-,-,+,-,+,+)_6&(4,4)
\end{array}$$
$$\begin{array}{llllll}
(0,6)&(+,-,-,-,-,-,-,-,+)&(2,6)&&(+,+,-,-,-,-,-,+,+)&(2,6)\\
&(+,-,-,-,-,-,-,+,+)&(2,6)&&(+,+,+,+,-,-,-,-,+)&(2,6)\\
&(+,+,+,-,-,-,-,-,+)&(2,6)&&(+,+,+,-,-,-,-,+,+)_1&(2,6)
\end{array}$$
\begin{rems}\label{remarque1}
{\rm (1) When the sign pattern consists of a sequence of $m_1$
pluses followed by a sequence of $m_2$ minuses and then by a sequence of
$m_3$ pluses,
where $m_1+m_2+m_3=d+1$, then for the pair $(0,d-2)$ this sign pattern is
not realizable if $\kappa :=(d-m_1-1)(d-m_3-1)/m_1m_3\geq 4$ (see Proposition~6
in \cite{FoKoSh}). For the sign patterns with $(0,6)$ in the above table the
quantity $\kappa$ equals respectively $36$, $25/4$, $15$, $9/2$,
$8$ and $20/6<4$. The last inequality shows that Proposition~6 of \cite{FoKoSh}
gives only sufficient, but not necessary conditions for nonrealizability of
the pair $(0,d-2)$ with the sign patterns containing only two sign changes.
(2) In the problem which we consider an important role is played, although
this is not always explicitly pointed out, by the {\em discriminant set} of the
family of monic polynomials. This is the set of values of the coefficients
for which the polynomial has a multiple root.
The number of real roots changes, generically by $2$,
when the tuple of coefficients crosses the
discriminant set. The stratification of the
discriminant set is explained in \cite{KhS}. More about discriminants of
the general
family of univariate polynomials for degree $4$ or $5$ can be found in
\cite{Ko}.}
\end{rems}
{\bf Acknowledgement.} The present paper is a continuation of the research
on sign patterns and admissible pairs which was started by B.~Z.~Shapiro,
J.~Forsg{\aa}rd and the author during the latter's stay at the University
of Stockholm. The author expresses his most sincere gratitude to this
university and to his former coauthors for this fruitful
collaboration.
\section{Case 1 is not realizable\protect\label{seccase1}}
The proof that the sign pattern $\sigma _1$
is not realizable with the pair $(0,6)$ follows
from Lemmas~\ref{RRRRbis} and~\ref{LLL}. The following lemma is
used in the proof of Lemma~\ref{RRRRbis}.
\begin{lm}\label{RRRR}
For any $0<u<v$ there exists a polynomial
$R=x^8+ax^7+bx^6+cx+d$, where $a>0$, $b>0$, $c>0$, $d>0$ and
$R(-u)=R'(-u)=R(-v)=R'(-v)=0$. Hence by Descartes' rule of signs
this polynomial equals $(x+u)^2(x+v)^2S(x)$, where the monic degree $4$
polynomial $S$ has no real roots.
\end{lm}
\begin{proof}
Consider the system of linear equations with unknown variables
$a$, $b$, $c$ and $d$ and parameters $u>0$ and $v>0$:
$$\begin{array}{lll}
u^8-au^7+bu^6-cu+d=0&~~~~~&8u^7-7au^6+6bu^5-c=0\\
v^8-av^7+bv^6-cv+d=0&~~~~~&8v^7-7av^6+6bv^5-c=0~.\end{array}$$
One can solve this system w.r.t. $a$, $b$, $c$ and $d$ (using, say, MAPLE)
and express the solutions as functions of $u$ and $v$. Set
$$g:=35u^4v^4+20u^3v^5+4u^7v+10u^2v^6+u^8+4uv^7+20u^5v^3+v^8+10u^6v^2~.~~~
{\rm Then}$$
$$\begin{array}{lll}a&=&
(2/g)(u^9+4u^8v+10v^2u^7+20v^3u^6+35v^4u^5\\ \\
&&+35v^5u^4+20v^6u^3+10v^7u^2+4v^8u+
v^9)\\ \\
b&=&(1/g)(u^{10}+4vu^9+10u^8v^2+35u^4v^6\\ \\
&&+20u^7v^3+20u^3v^7+35u^6v^4+10u^2v^8+56u^5v^5+4uv^9+v^{10})\\ \\
c&=&(2u^5v^5/g)(5vu^4+6v^2u^3+6v^3u^2+3u^5+5v^4u+3v^5)\\ \\
d&=&(u^6v^6/g)(5u^4+8u^3v+9u^2v^2+8uv^3+5v^4)~.
\end{array}$$
All coefficients being positive, if one gives positive values to $u$ and $v$
($u\neq v$), one obtains positive values of $a$, $b$, $c$ and $d$.
\end{proof}
\begin{lm}\label{RRRRbis}
If the sign pattern $\sigma _1$ is realizable with the pair
$(0,6)$, then there exists a real monic degree $8$ polynomial $H$
having three double negative and one double positive root
and the sign pattern $\sigma _1$.
\end{lm}
\begin{proof}
Suppose that the sign pattern $\sigma _1$
is realizable with the
pair $(0,6)$ by a real degree $8$ polynomial $P$
with six distinct negative roots and a complex conjugate pair. One can
suppose that the values of $P$ at its negative critical points
are all distinct.
One can increase the constant term of $P$ (which does not change the sign
pattern) so that two of the negative roots
coalesce in a double negative root $\alpha$ which is a local minimum of $P$.
Denote by $\tau <0$ and $\kappa <0$ the other two minima of $P$ on the negative
half-axis (one has $P(\tau )<0$ and $P(\kappa )<0$).
Denote by $R_1$ the polynomial of Lemma~\ref{RRRR}
with $u=-\alpha$, $v=-\tau$. Then for $\varepsilon >0$ small
enough the polynomial
$T:=P+\varepsilon R_1$ has five distinct negative roots
(four simple and one double). For some positive value of
$\varepsilon =\varepsilon _0$
the polynomial $T$ has a double root at $\kappa$ as well. As the value of $T$
for each fixed $x>0$ increases with $\varepsilon$,
$T$ has no real positive root.
Consider now the polynomial $T_0:=P+\varepsilon _0R_1$.
Denote by $R_2$ the polynomial of Lemma~\ref{RRRR}
with $u=-\alpha$, $v=-\kappa$. For some positive value of $\eta$ the polynomial
$T^*:=T_0+\eta R_2$ has double roots at $\alpha$, $\kappa$ and $\tau$,
no positive root and has the sign pattern $\sigma _1$.
Set $W:=(x-\alpha )^2(x-\kappa )^2(x-\tau )^2$. Consider the polynomial
$T^*-\mu W$, $\mu >0$. All coefficients of $W$ are positive.
Therefore the sign pattern defined by $T^*-\mu W$ has minuses in the
positions in which $\sigma _1$ has such.
As $T^*-\mu W$ has six negative roots counted with multiplicity, by
Descartes' rule of signs the sign pattern defined by it has at most two
sign changes.
The polynomial $T^*-\mu W$ for $\mu >0$ small enough is of the form
$(x-\alpha )^2(x-\kappa )^2(x-\tau )^2((x-\delta)^2+A)$, $\delta >0$, $A>0$.
Indeed, if $\delta \leq 0$, then all coefficients of $T^*-\mu W$
would be positive and it will not define the sign pattern $\sigma _1$.
Decrease $A$. Denote by $\sigma '$ the sign pattern
defined by $T^*-\mu W$ when $A=0$. When decreasing $A>0$, the signs of the
coefficients of $x^j$ remain negative for $j=2$, $3$, $4$ and $5$. For
$j=0$, $1$ and/or $6$ they might change from $+$ to $-$.
If $\sigma '$ has more minuses than
$\sigma _1$,
then it has a sequence of $m_1$ pluses, $m_1\leq 3$,
followed by a sequence of $m_2$ minuses followed by a sequence of $m_3$ pluses,
$m_3\leq 2$, $m_1+m_2+m_3=9$ (because $T^*-\mu W$ has $6$ negative roots and
the sequence of its coefficients must have at least $6$ sign
preservations, i.e. not more than two sign changes).
One cannot have $m_1<3$ or $m_3<2$ for $A=0$. Indeed, in this case one can
increase slightly $A$ without changing $m_1$, $m_2$ and $m_3$
and obtain a contradiction with Proposition~6 of~\cite{FoKoSh},
see part (1) of Remarks~\ref{remarque1}.
Hence $m_1=3$, $m_3=2$ and for $A=0$ the polynomial
$T^*-\mu W$ defines the sign pattern $\sigma _1$, i.e. $\sigma '=\sigma _1$.
\end{proof}
\begin{lm}\label{LLL}
There exists no real monic degree $8$ polynomial having three double negative
and one double positive root and defining the sign pattern $\sigma _1$.
\end{lm}
\begin{proof}
Assume that such a polynomial exists.
Without loss of generality one can assume that it is the square of the
polynomial
$$L:=(x^3+\alpha x^2+\beta x+\gamma )(x-1)=
x^4+(\alpha -1)x^3+(\beta -\alpha )x^2+(\gamma -\beta )x-\gamma$$
in which the first factor
has three distinct negative roots. Hence $\alpha >0$, $\beta >0$ and
$\gamma >0$. The coefficient of $x^s$ of $L^2$ is denoted by $c_s$. Hence
\begin{equation}\label{ccc}
\begin{array}{ll}c_7=2(\alpha -1)&c_6=2(\beta -\alpha )+(\alpha -1)^2\\
c_5=2((\gamma -\beta )+(\alpha -1)(\beta -\alpha ))&c_2=(\gamma -\beta )^2-
2(\beta -\alpha )\gamma \\
c_1=-2\gamma (\gamma -\beta )&c_0=\gamma ^2~.
\end{array}
\end{equation}
\begin{rems}\label{remremrem}
{\rm (1) As $L^2$ defines the sign pattern $\sigma _1$,
one must have $c_7>0$ and $c_1>0$
from which follows $\alpha >1$ and $\gamma <\beta$. These two inequalities
combined with $c_2<0$ yield $\beta >\alpha$.
(2) The condition $\beta >\alpha$ implies that the absolute value of
at least one of the roots of the polynomial $x^3+\alpha x^2+\beta x+\gamma$
(which are all negative) is $>1$.}
\end{rems}
In what follows we denote by $\mathcal{P}$ the set
$\{ \alpha >1,\beta >0,\gamma >0\}$. For each $\alpha =\alpha _0>1$ fixed
the set $\mathcal{P}|_{\alpha =\alpha _0}$ is the positive quadrant
$\{ \beta >0,\gamma >0\}$.
\begin{lm}\label{severalsets}
Suppose that $\alpha =\alpha _0>1$ is fixed. Then:
(1) The condition
$c_5=0$ defines a straight line $\mathcal{C}_5$.
Its slope $2-\alpha _0$ is positive for $\alpha _0\in (1,2)$,
zero for $\alpha _0=2$ and negative for $\alpha _0>2$. For $\alpha _0>2$
the intersection $(\mathcal{P}|_{\alpha =\alpha _0})\cap \mathcal{C}_5$
is a segment.
(2) The condition $c_2=0$ defines
a hyperbola with centre $(2\alpha _0/3,\alpha _0/3)$
and with asymptotes $\gamma -\alpha _0/3=(2\pm \sqrt{3})(\beta -2\alpha _0/3)$.
One of its branches (denoted by $\mathcal{C}_2$)
belongs to the set $\mathcal{P}|_{\alpha =\alpha _0}$; the other one
is denoted by $\mathcal{C}_2^*$. The point $(0,0)$ belongs to
$\mathcal{C}_2^*$ and the tangent line to $\mathcal{C}_2^*$ at $(0,0)$ is
horizontal. Hence
$\mathcal{C}_2^*\cap (\mathcal{P}|_{\alpha =\alpha _0})=\emptyset$.
(3) For $\alpha _0>\sqrt{3}$
the intersection $\mathcal{C}_5\cap \mathcal{C}_2$ consists of the two
points
$$I_1:=(\alpha _0,\alpha _0)~~~~{\it and}~~~~
I_2:=(\alpha _0(\alpha _0^2-1)/(\alpha _0^2-3),
\alpha _0(\alpha _0-1)^2/(\alpha _0^2-3))~.$$
For $\alpha _0\in (1,\sqrt{3}]$ one has $\mathcal{C}_5\cap \mathcal{C}_2=I_1$.
The tangent line to $\mathcal{C}_2$ at $I_1$ is vertical, at $I_2$ its slope
is negative for $\alpha _0>3$, zero for $\alpha _0=3$ and positive
for $\alpha _0\in (1,3)$. For $\alpha _0>3$ this slope is negative for
the points of $\mathcal{C}_2$ which are between $I_1$ and $I_2$.
(4) The set of hyperbolic polynomials is defined by the condition
\begin{equation}\label{conditionhyperbolicity}
4(\beta -\alpha _0^2/3)^3+27(\gamma +2\alpha _0^3/27-
\alpha _0\beta /3)^2\leq 0~.
\end{equation}
The corresponding equality defines a curve $\mathcal{H}$
having as only singular point a cusp at
$J:=(\alpha _0^2/3,\alpha _0^3/27)$. The set of hyperbolic polynomials
is the closure of the interior of $\mathcal{H}$.
The slope of the tangent lines to $\mathcal{H}$
at its regular points (and the one of
the geometric semi-tangent at its cusp) is positive for
$\beta >0$, $\gamma >0$. The maximal values of the coordinates of the
restriction of $\mathcal{H}$ to $\{ \beta >0,\gamma >0\}$ are attained,
simultaniously for $\beta$ and $\gamma$, at and only at its cusp.
(5) The curve $\mathcal{H}$ intersects the line $\mathcal{C}_5$ exactly when
$\alpha _0\geq u_0:=3.787042615\ldots$. For $\alpha _0<u_0$ the cusp point
$J$ lies below the line $\mathcal{C}_5$. The point $I_2$
does not define a hyperbolic polynomial for any $\alpha _0>1$.
\end{lm}
\begin{figure}[htbp]
\centerline{\hbox{\includegraphics[scale=0.7]{c2c5dd.eps}}}
\caption{The sets $\mathcal{C}_2$, $\mathcal{C}_5$,
$\{ \beta =\gamma \}$ and $\mathcal{H}$.}
\label{c2c5dd}
\end{figure}
Before proving Lemma~\ref{severalsets} we finish the proof of Lemma~\ref{LLL}.
On Fig~\ref{c2c5dd} we show the sets $\mathcal{C}_2$ (branch of a hyperbola),
$\mathcal{C}_5$ (straight line with negative slope), the straight line
$\{ \beta =\gamma \}$ and $\mathcal{H}$ (curve with a cusp point)
for $\alpha _0=5$. The set $\{ c_2<0\}$
is the interior of the branch $\mathcal{C}_2$ and the set $\{ c_2<0,c_5<0\}$
is the lens-shaped domain between $\mathcal{C}_2$ and
$\mathcal{C}_5$. The point $I_1$ is the triple intersection of
$\mathcal{C}_2$,
$\mathcal{C}_5$ and
$\{ \beta =\gamma \}$.
\begin{rem}
{\rm For $\alpha _0\in (1,\sqrt{3}]$ the set $\{ c_2<0,c_5<0\}$ is not compact
and for $\alpha _0\in (1,\sqrt{3})$ the point $I_2$
belongs not to $\mathcal{C}_2$, but to
$\mathcal{C}_2^*$; $I_2$ is at $\infty$ for $\alpha _0=\sqrt{3}$.
Indeed, the slopes of the asymptotes of
the hyperbola $\{ c_2=0\}$ equal $2\pm \sqrt{3}$ while the slope of
$\mathcal{C}_5$ equals $2-\alpha _0$,
see parts (1) and (2) of Lemma~\ref{severalsets}.}
\end{rem}
There exists a unique point $Z\in \mathcal{C}_2$ the tangent to $\mathcal{C}_2$
at which is horizontal. Indeed, the branches $\mathcal{C}_2$ and
$\mathcal{C}_2^*$ of the hyperbola $\{ c_2=0\}$ are symmetric w.r.t. its centre
$(2\alpha _0/3,\alpha _0/3)$, see part (2) of Lemma~\ref{severalsets}. The only
point of $\mathcal{C}_2^*$ at which the tangent line is horizontal is the
origin, see part (2) of Lemma~\ref{severalsets} (the fact that $(0,0)$ is the
only such point follows from the convexity of the hyperbola).
Hence $Z=(4\alpha _0/3,2\alpha _0/3)$.
Compare the $\gamma$-coordinates of the points $Z$ and $J$ (see part (4)
of Lemma~\ref{severalsets}). For $\alpha _0<3\sqrt{2}=4.2\ldots$ one has
$2\alpha _0/3>\alpha _0^3/27$. The point $Z$ has the least possible
$\gamma$-coordinate of the points of $\mathcal{C}_2$ whereas $J$ has the
largest possible $\gamma$-coordinate of the points of
$\mathcal{H}\cap \mathcal{P}_{\alpha =\alpha _0}$, see part (4) of
Lemma~\ref{severalsets}. Hence for $\alpha _0\in (1,3\sqrt{2})$ one has
$$\mathcal{C}_2\cap (\mathcal{H}\cap \mathcal{P}_{\alpha =\alpha _0})=\emptyset
~~~\, \, {\rm and}~~~\, \,
\{ c_2<0,c_5<0\} \cap (\mathcal{H}\cap \mathcal{P}_{\alpha =\alpha _0})=
\emptyset ~.$$
Recall that $u_0<3\sqrt{2}$, see part (5) of Lemma~\ref{severalsets}.
Hence for $\alpha _0=u_0$ the cusp $J$ of $\mathcal{H}$ has a smaller
$\gamma$-coordinate than $I_2$. As $I_2$ does not belong to $\mathcal{H}$
(for any $\alpha _0>1$, see part (5) of Lemma~\ref{severalsets}), for
$\alpha _0>u_0$ the points $I_1$ and $I_2$ are above the two intersection points
$K_1$ and $K_2$ of $\mathcal{H}$ with $\mathcal{C}_5$ (``above''
means ``have larger $\gamma$-coordinates''); $K_1$ is presumed to be
above $K_2$. Denote by $L^*$ and $L^{**}$
the vertical straight lines passing through $I_2$ and $K_1$. Hence for $a>3$
the domain $\{ c_2<0,c_5<0\}$ lies to the left of $L^*$ and above $I_2$, see
parts (1) and (3) of Lemma~\ref{severalsets}. At the same time the part of
$\mathcal{H}\cap \mathcal{P}_{\alpha =\alpha _0}$ which is to the left of
$L^*$ (hence to the left of $L^{**}$ as well) lies below $K_1$ hence below
$I_2$, so the domain $\{ c_2<0,c_5<0\}$ contains no hyperbolic polynomial.
This proves Lemma~\ref{LLL} and Theorem~\ref{tmnew}.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{severalsets}]
The first two statements of part (1) are to be checked directly. To prove the
third statement it suffices to compute the intersection points of the line
$\mathcal{C}_5$ with the $\beta$- and $\gamma$-axes. These points are
$(0,\alpha _0(\alpha _0-1))$ and
$(\alpha _0(\alpha _0-1)/(\alpha _0-2),0)$.
Prove part (2). The determinants of the matrices $M_1=\left( \begin{array}{rrr}
1&-2&0\\ -2&1&\alpha _0\\ 0&\alpha _0&0\end{array}\right)$ and
$M_2=\left( \begin{array}{rr}
1&-2\\ -2&1\end{array}\right)$
(defined after the quadric $c_2|_{\alpha =\alpha _0}$)
are nonzero and $M_2$ has one
positive and one negative eigenvalue. Hence the equation $c_2=0$ defines
a hyperbola.
To find its centre one sets $\beta \mapsto \beta +\mu$,
$\gamma \mapsto \gamma +\nu$ and one looks for $(\mu ,\nu )$ such that the
linear terms in the equation $c_2=0$ disappear. This yields the system
$$ -4\mu +2\nu +2\alpha _0=0~~~~,~~~~2\mu-4\nu =0$$
whose solution is $(\mu ,\nu )=(2\alpha _0/3,\alpha _0/3)$. The slopes of the
asymptotes are solutions to the equation $\lambda ^2-4\lambda +1=0$ deduced
from the matrix $M_2$. The branch
$\mathcal{S}_2$ occupies the upper right sector defined by the asymptotes.
The equation $c_2=0$ is satisfied for $(\beta ,\gamma )=(0,0)$. To compute the
equation of the tangent line to the hyperbola $\{ c_2=0\}$ one writes
\begin{equation}\label{tangent}
(-4\beta +2\gamma +2\alpha )d\gamma +(2\beta -4\gamma )d\beta =0
\end{equation}
in which the coefficient of $d\beta$ is $0$ for $(\beta ,\gamma )=(0,0)$.
The tangent at $(0,0)$ being horizontal the branch $\mathcal{S}_2^*$ belongs
entirely to the lower half-plane and does not intersect the set
$\mathcal{P}|_{\alpha =\alpha _0}$.
Prove part (3).
Set $B:=\gamma -\beta$, $A:=\beta -\alpha _0$. The conditions $c_5=0$ and
$c_2=0$ read (see (\ref{ccc})):
$$B=-(\alpha _0-1)A~~,~~-2(B+A+\alpha _0)A+B^2=0$$
from which one finds that either $A=0$ (hence $B=0$ and
$\beta =\gamma =\alpha _0$, this defines the point $I_1$)
or $-2(-(\alpha _0-1)A+A+\alpha _0)+(\alpha _0-1)^2A=0$.
The last equality implies
$A=2\alpha _0/(\alpha _0^2-3)$. Hence
\begin{equation}\label{firstequa}
\beta =\alpha _0(\alpha _0^2-1)/(\alpha _0^2-3)~,
\end{equation}
so $B=2\alpha _0(1-\alpha _0)/(\alpha _0^2-3)$ and
\begin{equation}\label{secondequa}
\gamma =\alpha _0(\alpha _0-1)^2/(\alpha _0^2-3)~.
\end{equation}
which gives the point $I_2$. To show that the tangent line to $\mathcal{C}_2$
at $I_1$ is vertical it suffices to observe that for $\beta =\gamma =\alpha _0$
equation (\ref{tangent}) reduces to $d\beta =0$. At $I_2$ the tangent line
to $\mathcal{C}_2$ is defined by the equation
$$(2\alpha _0^2/(\alpha _0^2-3))d\gamma +
(\alpha _0(\alpha _0-1)(\alpha _0-3)/(\alpha _0^2-3))d\beta =0~.$$
Its slope equals $-(\alpha _0-1)(\alpha _0-3)/2\alpha _0$. The last statement
of part (3) follows from the convexity of the hyperbola $\{ c_2=0\}$.
To prove part (4) one has to recall that the real polynomial
$x^3+px+q$ is hyperbolic if and only
if $4p^3+27q^2\leq 0$ (this means, in particular, that $p\leq 0$). As
$$x^3+\alpha x^2+\beta x+\gamma =(x+\alpha /3)^3+
(\beta -\alpha ^2/3)(x+\alpha /3)+\gamma +2\alpha ^3/27-\alpha \beta /3~,$$
the polynomial $L|_{\alpha =\alpha _0}$ is hyperbolic if and only if
condition (\ref{conditionhyperbolicity}) holds true.
Set $\beta \mapsto \alpha _0^2\beta$ and $\gamma \mapsto \alpha _0^3\gamma$
in the equation of $\mathcal{H}$ (see (\ref{conditionhyperbolicity})).
In the new variables
$(\beta ,\gamma )$ the equation of $\mathcal{H}$ (after division by
$\alpha _0^3$) coincides with its equation
for $\alpha _0=1$:
\begin{equation}\label{ch1}
4(\beta -1/3)^3+27(\gamma +2/27-\beta /3)^2=0~.
\end{equation}
One can parametrize this curve by setting $\beta =1/3-3t^2$,
$\gamma =1/27+2t^3-t^2=2(t-1/3)^2(t+1/6)$.
It has a cusp for $t=0$, i.e. at $(1/3,1/27)$.
Its tangent vector equals $(-6t,6t^2-2t)$. For $t<0$ its components are
both positive and its slope is also positive. For $t\in (0,1/3)$ they are
both negative and again the slope is positive. One has $\beta >0$ and
$\gamma >0$ exactly when $t\in (-1/6,1/3)$ (i.e. only for values of
$t$ for which the slope is positive). The coordinate $\beta$ attains its
global maximal value $1/3$ only for $t=0$. For $t\in (-1/6,1/3)$ the coordinate
$\gamma$ attains its maximal value $1/27$ only for $t=0$.
Prove part (5). The equation of $\mathcal{H}$ with
$\gamma =\beta -(\alpha _0-1)(\beta -\alpha _0)$ reads:
\begin{equation}\label{Uu}
\begin{array}{ccl}\mathcal{U}(\alpha _0,\beta )&:=&
4\beta ^3+44\beta ^2\alpha _0^2-4\beta \alpha _0^4-108\alpha _0\beta
+27\alpha _0^2-54\alpha _0^3\\ &&
+108\beta ^2+180\beta \alpha _0^2-
144\alpha _0\beta ^2-64\beta \alpha _0^3+23\alpha _0^4+4\alpha _0^5=0~.
\end{array}
\end{equation}
One has
$${\rm Res}(\mathcal{U},\partial \mathcal{U}/\partial \beta ,\beta )=
-64\alpha _0^3(\alpha _0-1)(2\alpha _0^2-7\alpha _0+8)(10\alpha _0^2-
45\alpha -0+27)^3~.$$
The first quadratic factor has no real roots. The roots of the second
one equal $0.7129573851\ldots <1$ and $u_0:=3.787042615\ldots$.
For $\alpha _0=u_0$ the cusp point of $\mathcal{H}$ is on $\mathcal{C}_5$.
For $\alpha _0<u_0$ the curve $\mathcal{H}\cap \mathcal{P}|_{\alpha =\alpha _0}$
lies entirely below the line $\mathcal{C}_5$ (this can be deduced from the
last statement of part (4) of the lemma and from the fact that for
$\alpha _0>0$ small enough the cusp point $J$ is close to the origin);
for $\alpha _0>u_0$ it intersects this line at two points.
\begin{rem}
{\rm Equation (\ref{Uu}) is of degree $3$ w.r.t. $\beta$. On Fig.~\ref{c2c5dd}
one sees two of the solutions (the points $K_1$ and $K_2$, see the proof
of Lemma~\ref{LLL}). The third solution is an intersection point of
$\mathcal{H}$ with $\mathcal{C}_5$, with $\beta <0$ and $\gamma >0$. Such
an intersection point exists because the $\gamma$-coordinate of a point of
$\mathcal{C}_5$ grows linearly in $|\beta |$ as $|\beta |$ increases
($\beta$ being negative) while the $\gamma$-coordinate of a point of
$\mathcal{H}$ grows as $|\beta |^{3/2}$.}
\end{rem}
To prove the last statement of part (5) we substitute the right-hand sides of
(\ref{firstequa}) and (\ref{secondequa})
for $\beta$ and $\gamma$
in (\ref{conditionhyperbolicity}) and we multiply by
$(\alpha _0^2-3)^3/\alpha _0^2(\alpha _0-1)^2>0$. This yields
the equivalent condition
$$3\alpha _0^6-16\alpha _0^5+13\alpha _0^4+24\alpha _0^3-23\alpha _0^2+
104\alpha _0-81\leq 0~.$$
However the left-hand side has no
roots greater than $1$ and the leading coefficient is positive. Hence
the last inequality fails for $\alpha _0>1$.
\end{proof}
\section{Cases 2 - 6 are not realizable\protect\label{seccases26}}
\subsection{Preliminaries}
The following two lemmas are proved in the Appendix.
They allow to simplify
the proof of Theorem~\ref{tmd8} by decreasing the number of parameters.
\begin{lm}\label{basiclemma}
Suppose that there exists a monic degree $8$ polynomial $P$ realizing
Case $j$, $2\leq j\leq 6$. Then there exists a monic degree $8$ polynomial
$U$ having a quadruple root at $1$ and no other real roots, and whose
coefficients define the same sign pattern as the one of Case~$j$.
\end{lm}
\begin{rem}
{\rm One can observe that roots at $1$ remain invariant under reverting of
sign patterns.}
\end{rem}
\begin{lm}\label{nextlemma}
(1) Suppose that a monic polynomial $U=(x-1)^4V$ realizes one of
the sign patterns
$$\begin{array}{lllllll}
\sigma _2^r&=&(+,+,-,-,-,+,-,+,+)&~~~~~~~~~&
\sigma _4&=&(+,+,+,-,-,+,-,+,+)\\
&&~~~~~{\rm or}&&\sigma _6^r&=&(+,-,-,-,-,+,-,+,+)~,
\end{array}$$
where $V$ is a real monic polynomial with no real root.
Then there exists a polynomial of the form $U_t:=U-t(x-1)^4$, $t\geq 0$,
defining the same sign pattern and having one or two negative roots
of even multiplicity, hence a polynomial of the form
\begin{equation}\label{polyW}
W:=(x-1)^4(x^2+Sx+S^2/4)(x^2+Mx+N)~,~~{\rm where}~~S>0~~{\rm and}~~
N\geq M^2/4~.
\end{equation}
(2) If the polynomial $U$ realizes the sign pattern
$\sigma _3^r=(+,-,-,+,-,+,-,+,+)$, then in the family of polynomials
$U^*_t:=U+tx(x-1)^4$, $t>0$, there exists a polynomial
defining the sign pattern $\sigma _3^r$ and of the form (\ref{polyW}).
(3) If the polynomial $U$ realizes the sign pattern
$\sigma _5^r=(+,-,-,+,-,+,+,+,+)$, then in the family of polynomials
$U^*_t:=U+tx(x-1)^4$, $t>0$, there exists a polynomial defining
one of the sign patterns
$\sigma _3^r$, $\sigma _5^r$ or $\sigma ^*:=(+,-,-,+,-,+,0,+,+)$
and of the form (\ref{polyW}).
\end{lm}
In what follows we set $W:=\sum _{j=0}^8w_jx^j$, $w_8=1$, and
$$Q:=3S^2/2-4S+1~~,~~R:=S^2-6S+4~~{\rm and}~~P:=S^2/4-4S+6~.$$
The roots of these three polynomials are real. We denote them by
$$\begin{array}{ccccccccccc}
0.27\ldots &=&(4-\sqrt{10})/3&=&q_1&<&q_2&=&(4+\sqrt{10})/3&=&
2.38\ldots \\
0.76\ldots &=&3-\sqrt{5}&=&r_1&<&r_2&=&3+\sqrt{5}&=&
5.23\ldots \\
1.67\ldots &=&8-\sqrt{40}&=&p_1&<&p_2&=&8+\sqrt{40}&=&
14.32\ldots \end{array}$$
The coefficients
$w_j$, $j=0$, $\ldots$, $7$ are expressed by the following formulae:
\begin{equation}\label{wj}\begin{array}{ccccccc}
w_0&=&S^2N/4&&w_1&=&(S/4)(MS+4N(1-S))\\
w_2&=&QN+(S-S^2)M+S^2/4&&w_3&=&QM-RN+S-S^2\\
w_4&=&Q-RM+PN&&w_5&=&-R+PM+(S-4)N\\
w_6&=&P+(S-4)M+N&&w_7&=&M+S-4
\end{array}
\end{equation}
\subsection{Cases 2 and 4}
In Cases 2 and 4 we are using the sign patterns
$\sigma _2^r=(+,+,-,-,-,+,-,+,+)$
and $\sigma _4=(+,+,+,-,-,+,-,+,+)$. They can be united
in a single sign pattern $\pi _{\pm}:=(+,+,\pm ,-,-,+,-,+,+)$.
If the polynomial $W$ (see (\ref{polyW}) defines
the sign pattern $\pi _{\pm}$, then one must have $w_j>0$ for $j=0$, $1$, $3$
and $7$ and $w_j<0$ for $j=2$, $4$ and $5$.
One has $M>0$. Indeed, $w_7=M+S-4>0$, hence $S>4-M$. Suppose that $M\leq 0$.
Then one has $S>4$, $MS\leq 0$
and $4N(1-S)\leq 0$, i.e. $w_1\leq 0$ -- a contradiction.
Suppose that $S>1$. Then the condition $w_1>0$ is equivalent to
$N<MS/4(S-1)$. On the other hand, as $N\geq M^2/4$, the last two inequalities
together imply $M<S/(S-1)$ hence $N<S^2/u$, where $u=4(S-1)^2$.
For $S\in [p_1,p_2]$ (recall that $p_1>1$) one has
$P\leq 0$, $PN\geq PS^2/u$ and $4-S<M<S/(S-1)$.
Therefore
$$w_4\geq \min
(~Q(S)-R(S)(4-S)+P(S)S^2/u~,~Q(S)-R(S)S/(S-1)+P(S)S^2/u~)~.$$
This minimum is $>5$ hence $>0$ (the numerical check of this is easy)
and the inequality $w_4<0$
fails for $S\in [p_1,p_2]$.
For $S>p_2$
one has $P\geq 0$, $PN\geq PM^2/4\geq 0$ and $0<M<S/(S-1)$, so
$$w_4\geq \min
(~Q(S)-R(S)S/(S-1)~,~Q(S)~)~.$$
This minimum is also positive and again $w_4<0$ fails.
Let now $S\in (0,p_1)$.
The inequality $w_4<0$ can be rewritten as
$N<(RM-Q)/P$ which together with $M^2/4\leq N$ implies
$PM^2-4RM+4Q<0$. This is a quadratic inequality w.r.t. $M$, with $P>0$.
The discriminant
of the quadratic polynomial $Y(M,S):=P(S)M^2-4R(S)M+4Q(S)$ equals
$4(R^2(S)-P(S)Q(S))$.
It is positive for all
$S\in (0,p_1)$ (this is easy to check). Hence for $S\in (0,p_1)$ the polynomial
$Y$ has two real roots $M'<M''$ which depend continuously on $S$ and
one must have $M\in (M',M'')$.
For each $S\in (0,p_1)$ fixed both these roots are smaller than $4-S$. Indeed,
set $M:=4-S$. The polynomial $Y(4-S,S)$ is positive on $(0,p_1)$
(easy to check). For $S=1\in (0,p_1)$ one has $Q=-3/2<0$, i.e.
one of the roots is
negative and the other is positive. Hence for $S\in (0,p_1)$ the number
$4-S$ lies
outside the interval $[M',M'']$, and as $4-S>0$, one has $M'<4-S$,
$M''<4-S$ and $M\in (M',M'')$. But one must have $M>4-S$,
so the inequalities $w_7>0$
and $w_4<0$ cannot hold simultaneously for $S\in (0,p_1)$.
\subsection{Cases 3, 5 and 6}
In Cases 3, 5 and 6 we use the sign patterns
$$\begin{array}{lllllll}\sigma _3^r&=&(+,-,-,+,-,+,-,+,+)&~~~~,&
\sigma _5^r&=&(+,-,-,+,-,+,+,+,+)\\ &&{\rm and}&&
\sigma _6^r&=&(+,-,-,-,-,+,-,+,+)~.\end{array}$$
and formulae (\ref{wj}). The proof of Theorem~\ref{tmnew} in these cases
results from Lemmas~\ref{not356r1}, \ref{notp14} and~\ref{notr1p1}.
\begin{lm}\label{M>0}
In Cases 3, 5 and 6 one has $M>0$.
\end{lm}
\begin{proof}
One must have $w_1>0$ and $w_6<0$.
For $S\geq 1$ the product $N(1-S)$ is negative (see formulae
(\ref{wj})), so for $S\geq 1$ the condition $w_1>0$ implies that one must have
$SM>0$, i.e. $M>0$. Consider for $S\in (0,1)$ the condition $w_6<0$, (i.e.
$P+(S-4)M+N<0$).
One has $P(S)>0$, $N\geq 0$ and $S-4<0$, so the inequality $w_6<0$ is possible
only for $M>0$.
\end{proof}
\begin{lm}\label{not356r1}
Cases 3, 5 and 6 are not realizable for $S\in (0,r_1]$.
\end{lm}
\begin{proof}
In Cases 3, 5 and 6 one has $w_3>0$, i.e. $QM+S-S^2>RN$, see (\ref{wj}).
For $S\in (0,r_1]$ one has $R(S)\geq 0$ and
$QM+S-S^2>RN\geq RM^2/4$, hence
\begin{equation}\label{aster}
L(S,M):=R(S)M^2/4-Q(S)M-S+S^2<0~.
\end{equation}
The inequalities
(\ref{aster}), $0<S\leq r_1$ and $0\leq M<4-S$ have no common solution. Indeed,
$L(S,4-S)=(S-2)^2((S-2)^2+8)/4$. This means that for $S=2$
the line $M+S=4$ has an
ordinary tangency with the curve $L(S,M)=0$, and this is
their only common point in the domain $\{ S>0,M>0\}$. For $S=M=1/2$ one has
$L(S,M)=9/64>0$ and $S+M-4<0$. Hence
below the line $M+S=4$ in the domain $\{ S>0,M>0\}$ one has $L(S,M)>0$.
\end{proof}
\begin{rem}
{\rm (1) The inequalities $S>0$, $M>0$ (see Lemma~\ref{M>0}) and
$S+M<4$ (this follows from $w_7<0$ in Cases 3, 5 and 6) imply $S<4$.}
\end{rem}
\begin{convention}\label{conv1}
{\rm (1) In what follows we interpret an equality of the form $w_j=0$
(see (\ref{wj})) as the equation of a straight line (denoted by $\ell _j$)
in the space $(M,N)$
with coefficients depending on $S$ as on a parameter. Most often
we need equations
of the form $A(S)N+B(S)M+C(S)=0$, and we care to have a positive coefficient
of $N$. E.g. we prefer the equation of the line $\ell _1$
(see the quantity $w_1$ in
formulae (\ref{wj})) to be of the form $4(1-S)N+SM=0$ for $S<1$ and
$4(S-1)N-SM=0$ for $S>1$.
(2) We denote by $\ell _j^+$ (resp. $\ell _j^-$) the
upper (resp. lower) half-plane defined by the line $\ell _j$. In the case
of $\ell _1$ one has $\ell _1^+:4(1-S)N+SM>0$ for $S<1$ and
$\ell _1^+:4(S-1)N-SM>0$ for $S>1$. For $S=1$ this line is vertical
and we do not define
the half-planes $\ell _1^{\pm}$. By $s(\ell _j)$ we denote the {\em slope} of
the line $\ell _j$, i.e. the quantity $-B(S)/A(S)$ for $A(S)\neq 0$.
For $\ell _1$ it equals $S/4(S-1)$.
(3) When in the proofs of the lemmas rational functions appear, it is
presumed that the factors of degree $2$ have no real roots (so
their sign coincides with the one of their leading coefficient). Factorizations
are performed by means of MAPLE.}
\end{convention}
\begin{lm}\label{notp14}
Cases 3, 5 and 6 are not realizable for $S\in [p_1,4)$.
\end{lm}
\begin{proof}
Consider the four conditions $M>0$, $w_1>0$, $w_3>0$ and $w_4<0$.
The second of them defines the half-plane $\ell _1^-$
(recall that $\ell _1:4(S-1)N-SM=0$). The last two of them
read
$$(-R(S))N+Q(S)M+S-S^2>0\hspace{6mm}{\rm and}\hspace{6mm}
(-P(S))N+R(S)M-Q(S)>0~.$$
The straight line $\ell _3:(-R(S))N+Q(S)M+S-S^2=0$ intersects the $N$-axis
at the point $A:=(0,N_A)$ with $N_A:=S(S-1)/(-R(S))>0$. The lines
$\ell _3$ and $\ell _4:(-P(S))N+R(S)M-Q(S)=0$ intersect at the point $B$
with coordinates
$$\begin{array}{lllll}
M_B&:=&(2/5)(5S^4-35S^3+84S^2-64S+16)/K(S)&,&\\
N_B&:=&(2/5)(5S^4-20S^3+36S^2-16S+4)/K(S)&,&{\rm where}\\
K(S)&:=&S^4-8S^3+30S^2-32S+16&.\end{array}$$
and both numerators and the denominator $K$ have no real roots.
This point lies above the straight line $\ell _1$. Indeed, the coefficient
of $N$ in the equation of $\ell _1$ is positive. Substituting
$(M_B,N_B)$ for $(M,N)$ in the left-hand side of this equation
yields the expression
$$\mu :=\frac{6(S^2-2.5\ldots S+3.8\ldots )(S^2-0.5\ldots S+0.2\ldots )
(S-1.2\ldots )}{(S^2-6.6\ldots S+20.4\ldots )(S^2-1.3\ldots S+0.7\ldots )}~$$
which is positive, see Convention~\ref{conv1}.
For the slopes $s(\ell _4)$ and $s(\ell _1)$ one has
$s(\ell _4)>s(\ell _1)>0$.
The first inequality follows from $R(S)/P(S)-S/4(S-1)>0$ which is equivalent to
$$\frac{15(S^2-1.7\ldots S+0.9\ldots )(S-4.6\ldots )}{4P(S)(S-1)}>0$$
and this results from $S-4.6\ldots <0$, $S-1>0$ and $P(S)<0$.
Hence the set defined by the conditions $M>0$, $w_3>0$ and $w_4<0$ is the
domain of $\mathbb{R}^2\simeq (M,N)$ to the right of the $N$-axis, to the
above of the segment $AB$ and to the above of the half-line starting at $B$,
which is part of the line $\ell _4$ and which goes to the right and upward.
This domain does not intersect the half-plane $\ell _1^-$ and the four
conditions $M>0$, $w_1>0$, $w_3>0$ and $w_4<0$ cannot hold true simultaneously.
\end{proof}
\begin{lm}\label{notr1p1}
Cases 3, 5 and 6 are not realizable for $S\in (r_1,p_1)$.
\end{lm}
\begin{proof}
Consider the conditions $w_3>0$ and $w_6<0$. They read
$$(-R(S))N+Q(S)M+S-S^2>0\hspace{6mm}{\rm and}\hspace{6mm}
N+(S-4)M+P(S)<0~.$$
Consider the point $\Pi :=\ell _3\cap \ell _6$. Its coordinates equal
$$(-(S^4-22S^3+120S^2-204S+96)/2Y(S),-3(S^4-16S^3+54S^2-64S+16)/4Y(S))~,$$
where $Y(S):=2S^3-17S^2+48S-30$ has a single real root
$y_0:=0.8609094817\ldots$. For $S\in (r_1,y_0)$ (resp. for $S\in (y_0,p_1)$)
one has $s(\ell _3)>s(\ell _6)$ (resp. $s(\ell _3)<s(\ell _6)$).
This follows from
$$Q(S)/R(S)-(4-S)=(S^2-7.6\ldots S+17.4\ldots )(S-y_0)/R(S)$$
with $R(S)<0$. The second coordinate of $\Pi$ equals
$$-3(S-0.3\ldots )(S-11.9\ldots )(S^2-3.7\ldots S+4.0\ldots )/4Y(S)~.$$
Hence it changes sign from $-$ to $+$ when $S$ passes from $y_0^-$ to $y_0^+$.
For $S\in (r_1,y_0)$ one has
$\{ w_3>0\} \cap \{ w_6<0\} =\ell _3^+\cap \ell _6^-$. For $S=y_0$ the
lines $\ell _3$ and $\ell _6$ are parallel, $\ell _3$ is
above $\ell _6$ and $\{ w_3>0\} \cap \{ w_6<0\} =\emptyset$.
Thus for $S\in (r_1,y_0)$ the sector $\ell _3\cap \ell _6$ belongs to
the domain $N<0$ and if some of Cases 3, 5 or 6 is realizable, it can be
realizable only for $S\in (y_0,p_1)$.
For $S=y_0^+$ the intersection $\{ w_3>0\} \cap \{ w_6<0\}$ is a sector whose
vertex has both coordinates positive because
the first coordinate of $\Pi$ equals
$$-(S-0.7\ldots )(S-1.8\ldots )(S-4.6\ldots )(S-14.7\ldots )/2Y(S)>0~.$$
The point $\Pi$ lies above the line $\ell _4:P(S)N-R(S)M+Q(S)=0$ for
$S\in (y_0,y_1)$, where $y_1:=1.471576286\ldots$.
Indeed, substituting the coordinates of $\Pi$ for $(M,N)$ in the left-hand
side of the equation of $\ell _4$ yields
$$ \frac{5(S^2-5.2\ldots S+20.3\ldots )(S^2-1.2\ldots S+1.2\ldots )
(S-7.9\ldots )(S-y_1)}{32(S^2-7.6\ldots S+17.4\ldots )(S-y_0)}>0~.$$
Moreover, $s(\ell _4)<0<s(\ell _3)<s(\ell _6)$.
Hence for
$S\in (y_0,y_1)$ the three conditions $w_3>0$, $w_4<0$ and $w_6<0$ cannot hold
true simultaneously.
In order to prove the lemma for $S\in [y_1,p_1)$ we consider the conditions
$$w_1>0~~{\rm ,~~i.e.}~~4(S-1)N-MS<0~~~{\rm and}~~~w_3>0~~{\rm ,~~i.e.}
~~-R(S)N+Q(S)M+S-S^2>0~.$$
The point $\Gamma :=\ell _1\cap \ell _3$ has coordinates
$(M_{\Gamma},N_{\Gamma})$ which equal
$$(4(S-1)^2S/(5S^3-16S^2+16S-4)~,
~S^2(S-1)/(5S^3-16S^2+16S-4))~.$$
Both coordinates are positive for $S\in [y_1,p_1)$ (the only real zero of
the denominator equals $0.3\ldots$). The point $\Gamma$ lies above the
straight line $\ell _4$. Indeed, substituting $(M_{\Gamma},N_{\Gamma})$ for $(M,N)$
in the left-hand side of the equation of $\ell_4:P(S)N-R(S)M+Q(S)=0$ with
$P(S)>0$ yields
$$\frac{3(S^2-2.5\ldots S+3.8\ldots )(S^2-0.5\ldots S+0.2\ldots )
(S-1.2\ldots )}{4(S^2-2.8\ldots S+2.1\ldots )(S-0.3\ldots )}>0~.$$
One has $s(\ell _4)<0<s(\ell _3)<s(\ell _1)$;
the last inequality follows from
$$\frac{S}{4(S-1)}-\frac{Q(S)}{R(S)}=
-\frac{5(S^2-2.8\ldots S+2.1\ldots )(S-0.3\ldots )}
{4(S-5.2\ldots )(S-1)(S-0.7\ldots )}>0~.$$
Hence for
$S\in [y_1,p_1)$ the sector $\{ w_1>0\} \cap \{ w_3>0\}$ does not intersect
the half-plane $\{ w_4<0\} =\ell _4^-$, i.e. the three
conditions $w_1>0$, $w_3>0$ and $w_4<0$ do not hold simultaneously.
\end{proof}
\section{Appendix. Proofs of
Lemmas~\protect\ref{basiclemma}
and \protect\ref{nextlemma}}
\begin{proof}[Proof of Lemma~\protect\ref{basiclemma}]
Denote by $0<x_1<x_2<x_3<x_4$ the real roots of $P$.
We are looking first for a polynomial
$U^0(x)$ of the form $(P(x)+ax^8-bx^k+c)/(1+a)$
having a quadruple root $x_0>0$,
where $k=1$ in Cases 3, 5 and 6, $k=3$ in Case~2, $k=5$ in Case~4, and
$a>0$, $b>0$, $c>0$. The signs of $a$, $b$ and $c$ imply that
$U^0$ defines the same sign
pattern as $P$. The polynomial $U$ is obtained from $U^0$ by suitable rescaling
and multiplication by a positive constant
which does not change the sign pattern.
For $x=x_0$ the polynomial $U^0$ satisfies the conditions
$(U^0)'=(U^0)''=(U^0)'''=0$
which read:
\begin{equation}\label{EQQ}
\begin{array}{lll}
~~k=1&&\\
P'(x)+8ax^7-b=0&P''+56ax^6=0&P'''+336ax^5=0\\
~~k=3&&\\
P'(x)+8ax^7-3bx^2=0&P''+56ax^6-6bx=0&P'''+336ax^5-6b=0\\
~~k=5&&\\
P'(x)+8ax^7-5bx^4=0&P''+56ax^6-20bx^3=0&P'''+336ax^5-60bx^2=0
\end{array}
\end{equation}
Consider first Cases 5 and 6, hence $k=1$. One eliminates $a$
from the last two equations which gives
$xP'''(x)=6P''(x)$. The polynomial $P'$ has exactly three positive roots
$\mu _1<\mu _2<\mu _3$, $\mu _j\in (x_j,x_{j+1})$.
Indeed, by Rolle's theorem it has at least three and by
Descartes' rule of signs it has at most three of them. So for $x>x_4$
(resp. $x>\mu _3$) the polynomial $P$ (resp. $P'$) is positive.
The polynomial $P''$ has at least two real roots
$\xi _1<\xi _2$, $\xi _j\in (\mu _j,\mu _{j+1})$ (again by Rolle's theorem).
By Descartes' rule of signs the polynomial $P''$
has at most three positive roots. The sign of the coefficient of $x^2$ in $P$
is negative, therefore
$P''$ has exactly three positive roots. The third of them $\xi _3$
is in $(0,\xi _1)$. Indeed, to the right of $\xi _2$ the number of
positive roots of $P''$ must be even because for $x>0$ sufficiently large
$P$ is convex. So $0<\xi _3<\xi _1<\xi _2$.
The polynomial $P'''$ has real roots
$\zeta _1 \in (\xi _3,\xi _1)$ and $\zeta _2\in (\xi _1,\xi _2)$. By
Descartes' rule of signs it has at most three positive roots in Case~6
and at most two in Case~5. In Case~6, as $P'''$ must have an even number of
roots to the right of $\xi _2$ ($P'$ is convex for $x>0$ sufficiently large),
the three positive roots $\zeta _3<\zeta _1<\zeta _2$ of $P'''$ belong
respectively to the intervals
$(0,\xi _3)$, $(\xi _3,\xi _1)$ and $(\xi _1,\xi _2)$.
Hence the signs of $P'''(\xi _1)$ and $P'''(\xi _2)$ are opposite and
$xP'''-6P''$ changes sign at some
point $x_0\in (\xi _1,\xi _2)$.
In Case~3 one has again $k=1$. The sign patterns $\sigma _3$ and
$\sigma _5$ differ only in their third position. The proof resembles the one
in Cases~5 and 6 yet Descartes' rule of signs allows more positive roots
for $P'$, $P''$ and $P'''$.
Denote by $p(P')$ the number of positive roots of $P'$.
Combining Rolle's theorem and Descartes' rule of signs
one understands that it is possible to encounter
only one of the following triples $(p(P'),p(P''),p(P'''))$:
$$i)~(3,5,4)\hspace{1cm}ii)~(3,3,4)\hspace{1cm}iii)~(3,3,2)\hspace{1cm}
iv)~(5,5,4)~.$$
In case {\em iii)} the proof is carried out in exactly the same way as for
Case~5. In the other cases one performs analogous reasoning with only
difference the two more positive roots of $P''$ and $P'''$ in case {\em i)},
of $P'''$ in case {\em ii)} or of $P'$, $P''$ and $P'''$ in case {\em iv)}.
For parity reasons the two more roots of the corresponding derivative $P^{(j)}$
(compared to their number in the proof of Case~5) must belong to one and the
same interval of $[0,\infty )$ defined by $0$, $\infty$ and the positive
roots of $P^{(j-1)}$. One proves as for Case~5 that the signs of $P'''$
at two consecutive roots of $P''$ are opposite, hence $xP'''-6P''$
changes sign at some
point $x_0$ from the interval between these two roots.
Consider Case~2, hence $k=3$. Eliminating $b$ from equations (\ref{EQQ}) yields:
$$2P'-xP''=40ax^7~~{\rm and}~~P''-xP'''=280ax^6~.$$
Eliminating $a$ from the last two equations gives the equation
$$14P'-8xP''+x^2P'''=(14P'-2xP'')-(x/2)(14P'-2xP'')'=0~.$$
The polynomial $P'$ has at most four
positive roots (by Descartes' rule of signs),
and at least three of them (denoted by $\mu _j$) belong to
the intervals $(x_j,x_{j+1})$,
$j=1$, $2$ and $3$, hence the fourth one $\mu _0$ is in $(0,x_1)$
(because $P'(0)>0$).
The polynomial
$P''$ has positive roots $\xi _{\nu}\in (\mu _{\nu},\mu _{\nu +1})$,
$\nu =1$, $2$, and $\xi _0\in (\mu _0,\mu _1)$. Hence the
polynomial $S:=14P'-2xP''$ has different signs at $\mu _{\nu}$ and
$\mu _{\nu +1}$ for $\nu =1$ and $2$, hence it has roots
$\delta _{\nu}\in (\mu _{\nu},\mu _{\nu +1})$, its derivative has opposite
signs at $\delta _1$ and $\delta _2$, so $S-(x/2)S':=14P'-8xP''+x^2P'''$
has a real root $x_0\in (\mu _1,\mu _3)$.
Consider Case~4, hence $k=5$.
One first eliminates $b$ (see equations (\ref{EQQ})):
$$4P'-xP''=24ax^7~~\, \, {\rm and}~~\, \, 3P''-xP'''=168ax^6~.$$
Eliminating after this $a$ results in
$$28P'-10xP''+x^2P'''=(28P'-4xP'')-(x/4)(28P'-4xP'')'=0~.$$
Similarly to the proof in Case~2 one shows that the polynomial
$28P'-10xP''+x^2P'''$
has a positive root $x_0$.
After the number $x_0$ is found, one finds first $a$ and then $b$
from system (\ref{EQQ}). Now we have to justify the positive signs of $a$ and
$b$ (and after this the one of $c$ as well). To this end
we set $a=ta_*$, $b=tb_*$, where $t>0$, and
we consider the family
of polynomials $R_t(x):=P(x)+t\psi _k(x)$ with
$\psi _k:=a_*x^8-b_*x^k$, $k=1$, $3$ or $5$. We suppose that for some $t>0$
the polynomial $R_t$ has a triple critical point at $x_0$.
Hence for a suitably chosen
$c$ the polynomial $R_t+c$ has a quadruple root at $x_0$.
Consider the function $\psi _k$ for $x>0$. For
$a_*\geq 0$, $b_*\leq 0$ and $a_*-b_*>0$ it is increasing and convex,
for $a_*\leq 0$, $b_*\geq 0$ and $a_*-b_*<0$ it is
decreasing and concave (for $a_*=0$ and $k=1$ it is linear, i.e. convex and
concave at the same time). For $a_*>0$ and $b_*>0$ (resp. for $a_*<0$ and
$b_*<0$)
it has a minimum (resp. a maximum) at
$\lambda _k:=(kb_*/8a_*)^{1/(8-k)}$ with $\psi _k(x)<0$ for $x\in (0,\lambda _k]$
(resp.
with $\psi _k(x)>0$ for $x\in (0,\lambda _k]$).
Consider the family of polynomials $R_t$, where $t$ is
supposed to belong to an interval $[0,\alpha )$ such that the sign pattern
defined by the coefficients of $R_t$ is the one of $P$. We keep the
same notation for the positive roots of $R_t$ and its derivatives
as the one for $P$. Then:
A) If $\psi _k$ is decreasing on $[\mu _2,\mu _3]$, then as $t$ increases,
$\mu _2$ moves to the left and $\mu _3$ to the right;
B) If $\psi _k$ is increasing on $[\mu _1,\mu _2]$, then as $t$ increases,
$\mu _1$ moves to the left and $\mu _2$ to the right.
In both cases A) and B) it is impossible to have the three positive
roots of $R_t'$ coalescing into a single critical point of $R_t$.
If $a_*\geq 0$, $b_*\leq 0$ and $a_*-b_*>0$, then case B) takes place.
If $a_*\leq 0$, $b_*\geq 0$ and $a_*-b_*<0$, then case A) takes place.
If $a_*<0$ and $b_*<0$, then at least one of cases A) or B) takes place.
Hence only for $a_*>0$ and $b_*>0$ can one have a critical point of $R_t$
of multiplicity $3$. This implies that $a>0$ and $b>0$.
Besides, $\lambda _k\in (\mu _1,\mu _3)$. Hence
$R_t(\mu _1)<0$ (because $P(\mu _1)<0$ and $\psi _k(\mu _1)<0$)
and to have $U^0(x_0)=0$ one has to choose $c>0$.
\end{proof}
\begin{proof}[Proof of Lemma~\protect\ref{nextlemma}]
Prove part (1). Consider the one-parameter family
of polynomials $U_t:=U-t(x-1)^4$, $t\geq 0$. The first four
coefficients do not depend on $t$ (they are the same as the ones of $U$).
The signs of the five coefficients of $-(x-1)^4$ are $(-,+,-,+,-)$.
Hence the first $8$ components of the sign pattern of $U_t$
do not depend on $t$
and in the family $U_t$ for some $t>0$, due to the decreasing of the value
of $U_t$ as $t$ increases, one of the two things
takes place first:
a) one has $U_t(0)=0$ or
b) $U_t$ has one or two negative roots,
each of them of even multiplicity.
One can notice that
the family $U_t$ contains no polynomial
with six positive roots (counted with multiplicity) because there are
four or five sign changes in the sign pattern of $U_t$
(the sign pattern of $U_t$ is obtained
from $\sigma _2^r$, $\sigma _4$ or $\sigma _6^r$ by replacing the last
component by $+$, $0$ or $-$).
If a) takes place for $t_0>0$, then as $U_t'(0)>0$, the root of $U_t$ at $0$ is
simple and $U_t$ has one or several negative roots
whose total multiplicity is odd.
Hence
for some $t_1\in (0,t_0)$, b) has taken place. Therefore in the family
$U_t$ there exists (for some $t>0$) a polynomial of the form (\ref{polyW})
which realizes the pattern $\sigma _2^r$, $\sigma _4$ or $\sigma _6^r$.
Prove part (2) of the lemma.
Suppose that the polynomial $U$ realizes the sign pattern
$\sigma _3^r=(+,-,-,+,-,+,-,+,+)$. Consider the family
$U^*_t=U+tx(x-1)^4$, $t>0$. The signs of the coefficients of
$x(x-1)^4$ are $(0,0,0,+,-,+,-,+,0)$, so the sign pattern of $U^*_t$ is
$\sigma _3^r$ for any $t>0$. The value of $U^*_t$ increases
(linearly with $t$) for each
$x>0$, $x\neq 1$ fixed, and decreases for each $x<0$ fixed.
Hence for some $t>0$
the polynomial $U^*_t$ has one or two negative roots each
of even multiplicity. For this value of $t$ the polynomial $U^*_t$ has the
form (\ref{polyW}).
The proof of part (3) resembles the one of part (2).
Suppose that the polynomial $U$ realizes the sign pattern
$\sigma _6^r=(+,-,-,+,-,+,+,+,+)$. The difference between $\sigma _6^r$ and
$\sigma _3^r$ is in the sign of the coefficient of $x^2$. Hence in the family
$U^*_t$ there is a polynomial with a quadruple root at $1$, with one or two
negative roots of even multiplicity and with coefficients defining
either one of the sign patterns $\sigma _3^r$, $\sigma _6^r$ or the
sign pattern $\sigma ^*$ (the sign of the coefficient of $x^2$ in $U^*_t$
might change for some value of $t$). In all three cases this is a polynomial
of the form (\ref{polyW}).
\end{proof}
|
\section{Introduction}
Deceptive opinion spam refers to illegitimate activities, such as writing fake reviews, giving fake ratings, etc., to mislead consumers. While the problem has been researched from both linguistic \cite{ott2011finding,feng2012syntactic} and behavioral \cite{mukherjee2013spotting,lim2010detecting} aspects, the case of sockpuppets still remains unsolved. A sockpuppet refers to a physical author using multiple aliases (user-ids) to inflict opinion spam to avoid getting filtered. Sockpuppets are particularly difficult to detect by existing opinion spam detection methods as a sockpuppet invariably uses a user-id only a few times (often once) thereby limiting context per user-id. Deceptive sockpuppets may thus be considered as a new frontier of attacks in opinion spam.\par
However, specific behavioral techniques such as Internet Protocol (IP) and session logs based detection in \cite{li2015analyzing} and group spammer detection in \cite{mukherjee2012spotting} can provide important signals to probe into few ids that form a potential sockpuppet. Particularly, some strong signals such as using same IP and session logs, abnormal keystroke similarities, etc. (all of which are almost always available to a website administrator) can render decent confidence that some reviews are written by one author masked behind a sockpuppet. This can render a form of “training data” for identifying that sockpuppeter; and the challenge is to find other fake reviews which are also written by the same author but using different aliases in future. Hence, the problem is reduced to an author verification problem. Given a few instances (reviews) written by a (known) sockpuppet author $a$, the task is to build an Author Verifier, $AV_a$ (classifier) that can determine whether another (future) review is also written by $a$ or not.
This problem is related to authorship attribution (AA)\cite{stamatatos2009survey} where the goal is to identify the author of a given document from a closed set of authors. However, having short reviews with diverse topics render traditional AA methods, that mostly rely on content features, not very effective (see section 7). While there have been works in AA for short texts such as tweets in \cite{layton2010authorship} and with limited training data \cite{luyckx2008authorship}, the case for sockpuppets is different because it involves deception. Further, in reality sockpuppet detection is an open set problem (i.e., it has an infinite number of classes or authors) which makes it very difficult if not impossible to have a very good representative sample of the negative set for an author. In that regard, our problem bears resemblance with authorship verification \cite{koppel2004authorship}.\\
In this work we first find that under traditional attribution setting, the precision of a verifier $AV_a$ degrades with the increase in the diversity and size of $\neg a$, where $\neg a$ refers to the negative set authors for a given verifier $AV_a$. This is detailed in section 4.1. This shows that the verifier struggles with higher false positive and cannot learn $\neg a$ well. It lays the ground for exploiting the unlabeled test set to improve the negative set in training. Next, we improve the performance by learning verification models in lower dimensions (section 5). Particularly, we employ a feature selection scheme, $\Delta$KL Parse Tree Features (henceforth abbreviated as $\Delta$KL-PTFs) that exploits the KL-Divergence of the stylistic language models (computed using PTFs) of $a$ and $\neg a$.
Lastly, we address the problem by taking advantage of transduction (section 6). The idea is to simply put a carefully selected subset of positive samples, reviews authored by $a$ (referred to as a spy set) from the training set to the unlabeled test set (i.e., the test set without seeing the true labels) and extract the nearest and farthest neighbors of the members in the spy set. These extracted neighbors (i.e., samples in the unlabeled test set which are close and far from the samples in the spy set) are potentially positive and negative samples that can improve building the verifier $AV_a$. This process is referred to as \textit{spy induction}. The basic rationale is that since all samples retain their identity, a good distance metric should find hidden positive and negative samples in the unlabeled test set. The technique is particularly effective for situations where training data is limited in size and diversity. Although both spy induction and traditional transduction \cite{vapnik2013nature} exploit the assumption of implicit clusters in the data \cite{chapelle2005semi}, there is a major difference between these two schemes; Spy induction focuses on sub-sampling the unlabeled test set for potential positive and negative examples to grow the training set whereas traditional transduction uses the entire unlabeled test set to find the hyperplane that splits training and test sets in the same manner \cite{joachims1999transductive}. Our results show that for the current task, spy induction significantly outperforms traditional transduction and other baselines across a variety of classifiers and even for cross domains.
\section{Related Work}
\textbf{Authorship Attribution (AA):} AA solves the attribution problem on a closed set of authors using text categorization. Supervised multi-class classification algorithms with lexical, semantic, syntactic, stylistic, and character n-gram features have been explored in \cite{graham2005segmenting,gamon2004linguistic,sapkota2015not}. In \cite{qian2016tri}, a tri-training method was proposed to solve AA under limited training data that extended co-training using three views: lexical, character and syntactic. The method however assumes that a large set of unlabeled documents authored by the same given closed set of authors are available which is different from our sockpuppet verification. In \cite{seroussi2012authorship}, latent topic features were used to improve attribution. This method also requires larger text collection per author to discover the latent topics for each author which is unavailable for a sockpuppet.\par
\textbf{Authorship Verification (AV):} In AV, given writings of an author, the task is to determine if a new document is written by that author or not. Koppel and Schler, (2004)\cite{koppel2004authorship} explored the problem on American novelists using one-class classification and ``unmasking" technique. Unmasking exploits the rate of deterioration of the accuracy of learned models as the best features are iteratively dropped. In \cite{koppel2014determining}, the task was to determine whether a pair of blogs were written by the same author. Repeated feature sub-sampling was used to determine if one document of the pair allowed selecting the other among a background set of ``imposters" reliably. Although effective unmasking requires a few hundred word texts to gain statistical robustness and was shown to be ineffective for short texts (e.g., reviews) in \cite{sanderson2006short}.\par
\textbf{Sockpuppet Detection:} Sockpuppets were studied in \cite{solorio2013case} for detecting fake identities in Wikipedia content providers using an SVM model with word and Part Of Speech (POS) features. In \cite{qian2013identifying}, a similarity space based learning method was proposed for identifying multiple userids of the same author. These methods assume reasonable context (e.g., 30 reviews per userid). These may not be realistic in opinion spamming (e.g., \cite{mukherjee2012spotting,jindal2008opinion,fusilier2015detection}) as the reviews per userid are far less and often only one, as shown in singleton opinion spamming \cite{xie2012review}.
\section{Dataset}
\cite{gokhman2012search} reports that crowdsourcing is a reasonable method for soliciting ground truths for deceptive content. Crowdsourcing has been successfully used for opinion spam generation in various previous works \cite{ott2011finding,li2014towards,li2013identifying,Banerjee2014keystroke}. In this work, our focus is to garner ground truth samples of multiple fake reviews written by one physical author (sockpuppet). To our knowledge, there is no existing dataset available for opinion spam sockpuppets. Hence, we used Amazon Mechanical Turk.\par
Participating turkers were led to a website for this experiment where responses were captured. To model a realistic scenario such as singleton opinion spamming \cite{xie2012review}, Turkers were asked to act as a sockpuppet having access to several user-ids and each user-id was to be used exactly once to write a review as if written by that alias. The core task required writing 6 positive and 6 negative deceptive reviews, each had more than 200 words, on an entity (i.e., 12 reviews per entity). Each entity belonged to one of the three domains: hotel, restaurant and product. We selected 6 entities across each domain for this task. Each turker had to complete the core task for two entities each per domain (i.e., 24 reviews per domain). The entities and domains were spread out evenly across 17 authors (Turkers). It took us over a month to collect all samples and the mean writing time per review was about 9 minutes.\par
To ensure original content, copy and paste was disabled in the logging website. We also followed important rubrics in \cite{ott2011finding} (e.g., restricted to US Turkers, maintaining an approval rating of at least $ 90\% $) and Turkers were briefed with the domain of deception with example fake reviews (from Yelp). All responses were evaluated manually and those not meeting the requirements (e.g., overly short, incorrect target entity, unintelligible, etc.) were discarded resulting in an average of 23 reviews per Turker per domain. The data and code of this work is available at this link \footnote{https://www.dropbox.com/sh/xybjmxffmype3u2/AAA95vdkDp6z5fnTHxqjxq5Ga?dl=0} and will be released to serve as a resource for furthering research on opinion spam and sockpuppet detection. \par
Throughout the paper, for single domain experiments, we focus on the hotel domain which had the same trends to that of product and restaurant domains. However, we report results on all domains for cross domain analysis (section 7.4).
\begin{figure}[t!]
\includegraphics[scale=0.4]{fig1}
\caption{Precision, recall and F-Score (y-axis) for different author diversity, $\lambda=25\%, 50\%, 75\%, 100\%$ (x-axis) under in-training setting.}
\end{figure}
\section{Hardness Analysis}
This section aims to understand the hardness of sockpuppet verification via two schemes.
\subsection{Employing Attribution}
An ideal verifier (classifier) for an author $a$ requires a representative sample of $\neg a$. We can approximate this by assuming a pseudo author representing $\neg a$ and populating it by randomly selecting reviews of all authors except $a$. Under the AA paradigm, this is reduced to binary classification. We build author verifiers for each author $a_i \in A=\{a_1,...,a_{17}\}$. As in AA paradigm, we use in-training setting, i.e., negative samples ($\neg a$) in both training and test sets are authored by the same closed set of 16 authors although the test and training sets are disjoint. Given our task, since there are not many documents per author to learn from, the effect of author diversity on problem hardness becomes relevant. Hence, we analyze the effect of the diversity and size of the negative set. Let $\lambda \in \{25\%,50\%,75\%,100\%\} $ be the fraction of total authors in $\neg a$ that are used in building the verifier $AV_a$. Here $\lambda$ refers to author diversity under in-training setting. We will later explore the effect of diversity under out-of-training setting (section 5). For e.g., when $\lambda=50\%$, we randomly choose 8 authors, 50\% of total 16 authors, from $\neg a$ to define the negative set for $AV_a$. Note that since we have a total of 16 authors in $\neg a$ for each $a$ and all $\lambda$ values, the class distribution is imbalanced with the negative class $\neg a$ in majority. We keep the training set balanced throughout the paper as recommended in \cite{mukherjee2013yelp} to avoid learning bias due to data skewness. We use 5-fold Cross Validation (5-fold CV) so, the training fold consists of 80\% of the positive ($a$) and equal sized negative ($\neg a$) samples. But the test fold includes the rest 20\% of positive and remaining negative samples except those in training. Under this scheme, since $\neg a$ is the majority class in the test set, accuracy is not an effective metric. For each $AV_a$, we first compute the precision, recall and F-Score (on the positive class $a$) using 5-fold CV. Next, we average the results across all authors using their individual verifiers (Figure 1). This scheme yields us a robust measure of performance of sockpuppet verification across all authors and is used throughout the paper.\par We report results of Support Vector Mechine (SVM), Logistic Regression (LR) and k-Nearest Neighbor (kNN) classifiers (using the libraries LIBSVM \cite{CC01a} for SVM with RBF kernel, LIBLINEAR\cite{REF08a} for LR with L2 regularization and WEKA \footnote{http://www.cs.waikato.ac.nz/ml/weka/} for kNN with k=3 whose parameters were learned via CV). The feature space consists of lexical units (word unigram) and Parse Tree Features (PTF) extracted using Stanford parser \cite{klein2003accurate} with normalized term frequency for feature value assignment. Unless otherwise stated we use this feature set as well as the classifires setting for all experiments in this paper. We followed some rules from \cite{feng2012characterizing} in computing PTFs. The rules are generated by traversing a parse tree in three ways i) a parent node to the combination of all its non-leaf nodes, ii) an internal node to its grandparent, iii) a parent to its internal child. We also add all interior nodes to the feature space (Table 1). From Figure 1, we note:
\begin{table}[t!]
\caption{Parse Tree Feature (PTF) Types}
\label{tab:table1}
\setlength{\tabcolsep}{0.15em}
\centering
\begin{tabular}{lll}
\toprule
\multicolumn{3}{c}{Parse tree for: ``The staff were friendly."} \\
\midrule
\multirow{4}[8]{*}{{\includegraphics[width=3cm,height=4cm,keepaspectratio]{table2}}}\\ & PTF(I) & S $\rightarrow$ NP VP \\
& PTF(II) & JJ \^ \,ADJP $\rightarrow$ VP\\
& PTF (III) & S$\rightarrow$ NP \\
& Interior nodes & DT, NP \\
\end{tabular}%
\end{table}%
\begin{itemize}[leftmargin=*]
\item With increase in diversity of negative samples, $\lambda$ of $\neg a$, the test set size and variety also increase and we find significant drops in precision across all classifiers. This shows a significant rise in false positives. In other words, as the approximated negative set approaches the universal negative set ($\widetilde{\neg a}\rightarrow\neg a$ with increase in diversity of $\neg a$), learning $\neg a$ becomes harder.
\item Recall, however, does not experience major changes with
increase in the diversity of negative set as it is concerned with retrieving the positive class ($a$).
\item F-Score being the harmonic mean of precision and recall, aligns with the precision performance order. We also note that F-Score in SVM and LR behave similarly followed by kNN.
\end{itemize}
Thus, sockpuppet verification is non-trivial and the hardness increases with the increase in $\neg a$ diversity.
\subsection{Employing Accuracy and F1 on Balanced Class Distribution}
Under binary text classification and balanced class distribution, if accuracy or F1 are high, it shows that the two classes are well separated. This scheme was used in \cite{koppel2004authorship} for authorship verification. In our case, we adapt the method as follows. We consider two kinds of balanced data scenarios for a verifier for author $a$, $AV_a$: $S_1$ and $S_2$. Under $S_1$, we have the positive class $P$ that consists of half of all reviews authored by $a$ $R_a$, i.e., $P=\{r_i \in R_a;|P|=1/2 |R_a |\}$. The negative class $N_{S_1}$ comprises of the other half, $N_{S_1}=\{r_i \in R_a-P;|N_{S_1} |=|P|\}$ and $S_1=P\cup N_{S_1}$. Under $S_2$, we keep $P$ intact but use a random sampling of $\neg a$ for its negative class, $N_{S_2}=\{r_i \in R_{\neg a};|N_{S_2} |=|P|\}$ yielding us $S_2=P\cup N_{S_2}$. Essentially, with this scheme, we wish to understand the effect of negative training set when varied from false negative $(N_{S_1})$ to approximated true negative $(N_{S_2})$. Using lexical and parse tree features and 5-fold CV we report performance under each scenario $S_1$ and $S_2$ in Table 2. We note the following:
\begin{itemize}[leftmargin=*]
\item The precision, recall, F1 and accuracy of all models under $S_2$ is higher than $S_1$. While this is intuitive, it shows for deceptive sockpuppets, writings of an author ($P$) bear separation from other sockpuppeters ($N_{S_2}$).
\item Sockpuppet verification is a difficult problem because under balanced binary classification ($S_2$), there is just 5-10\% gain in accuracy than random (50\% accuracy). Yet it does show the models are learning some linguistic knowledge that separate $a$ and $\neg a$ and using writings of authors other than $a$ is a reasonable approximation for universal $\neg a$.
\end{itemize}
\begin{table}[t!]
\caption{Classification results P: Precision, R: Recall, Acc: Accuracy, F1: F-Score under two balanced data scenarios $S_1$ and $S_2$ for different classifiers.}
\centering
\scalebox{1}{
\begin{tabular}{lrrrrrrrr}
\toprule
\multirow{2}[4]{*}{Model} & \multicolumn{4}{c|}{$S_1$} & \multicolumn{4}{c}{$S_2$} \\
& \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{Acc} & \multicolumn{1}{c|}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{Acc} & \multicolumn{1}{c}{F1} \\
\midrule
SVM & \multicolumn{1}{c}{47.1} & \multicolumn{1}{c}{48.4} & \multicolumn{1}{c}{49.0} & \multicolumn{1}{c}{45.6} & \multicolumn{1}{c}{62.5} & \multicolumn{1}{c}{66.5} & \multicolumn{1}{c}{61.8} & \multicolumn{1}{c}{61.1} \\
LR & \multicolumn{1}{c}{47.4} & \multicolumn{1}{c}{46.4} & \multicolumn{1}{c}{49.5} & \multicolumn{1}{c}{44.6} & \multicolumn{1}{c}{63.5} & \multicolumn{1}{c}{67.4} & \multicolumn{1}{c}{61.7} & \multicolumn{1}{c}{62.1} \\
kNN & \multicolumn{1}{c}{41.9} & \multicolumn{1}{c}{57.4} & \multicolumn{1}{c}{49.8} & \multicolumn{1}{c}{44.9} & \multicolumn{1}{c}{51.0} & \multicolumn{1}{c}{68.9} & \multicolumn{1}{c}{56.1} & \multicolumn{1}{c}{53.8} \\
\bottomrule
\end{tabular}%
}
\label{tab:addlabel}%
\end{table}%
\section{Learning in Lower Dimensions}
From the previous experiment, it hints that in the case of deceptive sockpuppets, only a small set of features differentiate $a$ and $\neg a$. As explored in \cite{feng2012characterizing}, there often exists discriminative author specific stylistic elements that can characterize an author. However, the gamut of all PTFs per author (greater than 2000 features in our data) may be overlapping across authors (e.g., due to native language styles). To mine those discriminative PTFs, we need a feature selection scheme. We build on the idea of linguistic KL-Divergence in \cite{mukherjee2013yelp} and model stylistic elements to capture \textit{how} things are said as opposed to \textit{what} is said. The key idea is to construct the stylistic language model for author, $a$ and its pseudo author $\neg a$. Let $A$ and $\neg A$ denote the stylistic language models for author $a$ and $\neg a$ comprising the positive and negative class of $AV_a$ respectively, where $A(t)$ and $\neg A(t)$ denote the probability of the PTF, $t$ in the reviews of $a$ and $\neg a$. $KL(A||\neg A)=\sum_t (A(t) \log_2{(A(t)/\neg A(t))})$ provides a quantitative measure of stylistic difference between $a$ and $\neg a$. Based on its definition, PTF $t$ that appears in $A$ with higher probability than in $\neg A$, contributes most to $KL(A||\neg A)$. Being asymmetric, it also follows that PTF $t^\prime $ that appears in $\neg A$ more than in $A$ contributes most to $KL(\neg A||A)$. Clearly, both of these types of PTF are useful for building $AV_a$. They can be combined by computing the per feature, $f$, $\Delta KL^f$ as follows:
\begin{align}
\Delta KL_t^f=KL_t(A_t||\neg A_t)-KL_t(\neg A_t||A_t)\\
KL_t(A_t||\neg A_t)=A(t) \log_2{(A(t)/\neg A(t))}\\
KL_t(\neg A_t|| A_t)=\neg A(t) \log_2{(\neg A(t)/A(t))}
\end{align}
Discriminative features are found by simply selecting the top PTF $t$ based on the descending order of $|\Delta KL_t^f |$ until $|\Delta KL_t^f |<0.01 $. This is a form of sub-sampling the original PTF space and lowers the feature dimensionality. Intuitively, as $KL_t$ is proportional to the relative difference between the probability of PTF $t$ in positive ($a$) and negative ($\neg a$) classes, the above selection scheme provides us those PTF $t$ that contribute most to the linguistic divergence between stylistic language models of $a$ and $\neg a$. \par
To evaluate the effect of learning in lower dimensions, we consider a more realistic ``out-of-training" setting instead of the in-training setting as in previous experiments. Under out-of-training setting, the classifier cannot see the writings of those authors that it may encounter in the test set. In other words test and training sets of a verifier $AV_a$ are completely disjoint with respect to $\neg a$ which is realistic and also more difficult than in-training setting. Further, we explore the effect of author diversity under out-of-training setting, $\delta$ for the negative set (not to be confused with $\lambda$ as in section $4$). For each experiment, the reviews from $\delta \%$ of all authors except the intended author, $\neg a$ participate in the training of a verifier $AV_a$ while the rest ($100-\delta \%$) authors make the negative test set. We also consider standard lexical units (word unigram) (L), L + PTF, and top $k=20\%$ (tuned via CV) PTF selected using $\chi ^2$ metric (L + PTF $\chi ^2$) as baselines. We examine different values of $\delta \in \{25\%,50\%,75\%\}$ but not $\delta=100\%$ as that leaves no test samples due to out-of-training setting. From Table 3, we note:
\begin{itemize}[leftmargin=*]
\item For each feature space, as the $\neg a$ diversity ($\delta$) increases, across each classifier, we find gains in precision with reasonably lesser drops in recall resulting in overall higher F1. This shows that with increase in diversity in training, the verifiers reduced false positives improving their confidence. Note that verification gets harder for smaller $\delta$ as the size and skewness of the test set increases. This trend is different from what we saw in Figure 1 with $\lambda$ which referred to diversity under in-training setting.
\item Average F1 based on three classifiers (column AVG, Table 3) improves for $\delta=25\%,50\%$ using L+PTF than L showing parse tree feature can capture style. However feature selection using $\chi ^2$ (L+PTF $\chi ^2$) is not doing well as for all $ \delta$ values there is reduction in F1 for SVM and LR. L+$\Delta KL$ PTF feature selection performs best in AVG F1 across different classifiers. It recovers the loss of PTF $\chi ^2$ and also improves over the L+PTF space by about 2-3$\%$.
\end{itemize}
\begin{table}[t!]
\centering
\scalebox{1}{
\begin{tabular}{lcccccccccr}
\toprule
\multirow{2}[3]{*}{} & \multicolumn{10}{c}{\bf{ $\delta$=25\%}} \\
& \multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \multicolumn{1}{c}{\bf{AVG}} \\
Feature Set & P & R & F1 & P & R & F1 & P & R & F1 & \multicolumn{1}{c}{F1} \\\midrule
L & 23.6 & 82.0 & 34.3 & 23.1 & 74.7 & 30.8 & 19.4 & 84.6 & 25.8 & \multicolumn{1}{c}{30.3} \\
L+PTF & 25.6 & 73.4 & 35.2 & 22.9 & 82.5 & 33.4 & 24.8 & 66.7 & 24.5 & \multicolumn{1}{c}{31.0} \\
L+PTF$\chi^2$ & 21.7 & 73.5 & 30.8 & 14.8 & 53.5 & 21.3 & 22.6 & 75.3 & 25.9 & \multicolumn{1}{c}{26.0} \\
L+$\Delta KL$ PTF & 25.6 & 79.2 & 36.3 & 21.7 & 80.2 & 32.1 & 22.3 & 81.5 & 27.8 & \multicolumn{1}{c}{\textbf{32.1}} \\
\multicolumn{11}{c}{(a)} \\
\midrule
\multirow{2}[4]{*}{} & \multicolumn{10}{c}{\bf{$\delta$ =50\%}} \\
& \multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \multicolumn{1}{c}{\bf{AVG}} \\
Feature Set & P & R & F1 & P & R & F1 & P & R & F1 & \multicolumn{1}{c}{F1} \\
\midrule
L & 30.7 & 83.6 & 41.8 & 28.7 & 83.1 & 38.7 & 21.1 & 85.1 & 27.1 & \multicolumn{1}{c}{35.9} \\
L+PTF & 33.2 & 73.4 & 42.7 & 30.6 & 78.1 & 40.9 & 28.0 & 73.8 & 28.8 & 37.5 \\
L+PTF$\chi^2$ & 24.8 & 69.2 & 33.7 & 21.0 & 47.8 & 26.9 & 23.4 & 81.6 & 30.2 & 30.3 \\
L+$\Delta KL$ PTF & 33.7 & 75.9 & 42.8 & 31.1 & 79.4 & 41.9 & 26.9 & 79.5 & 30.3 & \textbf{38.3} \\
\multicolumn{11}{c}{(b)} \\
\midrule
\multirow{2}[4]{*}{} & \multicolumn{10}{c}{\bf{$\delta$ =75\%}} \\
& \multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \multicolumn{1}{c}{\bf{AVG}} \\
Feature Set & P & R & F1 & P & R & F1 & P & R & F1 & \multicolumn{1}{c}{F1} \\
\midrule
L & 47.1 & 77.7 & 55.1 & 44.4 & 80.4 & 52.7 & 28.7 & 83.5 & 37.8 & \multicolumn{1}{c}{48.5} \\
L+PTF & 51.4 & 72.6 & 56.0 & 43.7 & 78.8 & 53.0 & 28.1 & 64.8 & 31.7 & 46.9 \\
L+PTF$\chi^2$ & 42.4 & 71.2 & 49.5 & 33.9 & 49.6 & 36.4 & 35.6 & 79.8 & 40.0 & 42.0 \\
L+$\Delta KL$ PTF & 50.5 & 71.9 & 56.2 & 46.3 & 79.4 & 54.9 & 42.2 & 80.8 & 46.1 & \textbf{52.4} \\
\multicolumn{11}{c}{(c)} \\
\bottomrule
\end{tabular}%
}
\caption{P: Precision, R: Recall, F1: F-Score for out-of-training with different values of $\delta$ for three classifiers. AVG reports the average F1 across three classifiers. Feature Set: L: Lexical unit (word unigram), PTF: Parse tree feature, PTF $\chi^2$ : PTF selected by $\chi^2$ , $\Delta KL$ PTF: PTF selected via $\Delta KL$}
\label{tab:addlabel}%
\end{table}%
\section{Spy Induction}
We recall from section 1 that our problem suffers with limited training data per author as sockpuppets only use an alias few times. To improve verification, we need a way to learn from more instances. Also from section 4, we know that precision drops with increase in diversity of $\neg a$. This can be addressed by leveraging the unlabeled test set to improve the $\neg a$ set in training under transduction. \par
Figure 2 provides an overview of the scheme. For a given training set and a test set for $AV_a$, spy induction has three main steps. First is spy selection where some carefully selected positive samples are sent to the unlabeled test set. The second step is to find certain Nearest and Farthest Neighbors (abbreviated NN, FN henceforth) of the positive spy samples in the unlabeled test set. As the instances retain their original identity, a good distance metric should be able to retrieve potentially hidden positive (using common NN across different positive spies) and negative (using common FN across different positive spies) samples in the unlabeled test set. These newly retrieved samples from unlabeled test set are used to grow the training set. The previous step can have some label errors in NN and FN as they may not be true positive ($a$) and negative ($\neg a$) samples, which can be harmful in training. These are shown in Figure 2(B) by $\alpha_-$ and $\beta_+$ samples. To reduce such potential errors, a third step of label verification is employed where the labels of the newly retrieved samples from unlabeled test set are verified using agreement of classifiers on orthogonal feature spaces. with this step, we benefit from the extended training data without suffering from the possible issue of error propagation. Lastly, the verifier undergoes improved training with additional samples and optimizes the F-Score on the training set.\par
\begin{figure}[t!]
\includegraphics[scale=.95]{fig2}
\caption{Spy Induction : (A)Spies (Red plus signs) selected based on positive class centrality being put to the unlabeled test set. (B) Common nearest and farthest neighbors (Green plus and minus signs) across different spies’ neighborhood shown by oval boundaries found in unlabeled test set being put back in the training set.}
\end{figure}
\subsection{Spy Selection}
This first step involves sending highly representative spies that can retrieve new samples to improve training. For a given verification problem, $AV_a$, let $D=D.Train\cup D.Test$ denotes the whole data. Although any positive instance in $D.Train$ can be a spy sample, only few of them might satisfy the representativeness constraint. Hence, we select the spies as those positive samples that have maximum similarity with other positive instances. In other words, the selection respects class based centrality and employs minimum overall pairwise distance (OPD) as its selection criterion:
\begin{equation}
OPD(s)=argmin_{s\in P} (\sum_{x\in P}d(s,x))
\end{equation}\par
where $P$ is the positive class of training set, $s$ denotes a potential spy sample and $d(\cdot)$ is distance function. Our spy set, $S=\{s\}$ consists of different spies that have the least pairwise distance to all other positive samples. We also consider different sizes of the spy set $|S|=n_S$ and experiment with different values of $n_S\in N_S=\{1,3,5,7\}$. The method $SelectSpy(\cdot)$ (line 4, Algorithm 1) implements this step.
\subsection{New Instance Retrieval via Nearest and Farthest Neighbors}
After the selected spies are put into the unlabeled test set, the goal is to find potential positive and negative samples. Intuitively, one would expect that the closest data points to positive spy samples belong to the positive class while those that are farthest are likely negative samples. For each spy, $s\in S$, we consider $n_Q$ nearest neighbors forming the likely positive set $Q_s$ and $n_R$ farthest neighbors forming the likely negative set $R_s$ specific to $s$. Then, we find the common neighbors across multiple spies to get confidence on the likely positive or negative samples which yields us the final set of potentially $Q$ positive and $R$ negative samples,
\begin{equation}
Q=\cap_{s\in S}\, Q_s ;\quad R=\cap_{s\in S}\, R_s
\end{equation}\par
This is implemented by the methods $ObtainNN(\cdot),ObtainFN(\cdot)$ (lines 5, 6, Algorithm 1). In most cases, we did not find the common neighbors $Q,R$ to be empty, but if it is null, it implies no reliable samples were found. Further, like $n_S$ (in section 6.1), we try different values for $|Q_s |=n_Q; n_Q \in N_Q=\{1,3\}$ and $|R_s |=n_R; n_R \in N_R=\{5,10,25,40,50,60\}$. These values were set based on pilot experiments.
The above scheme of new sample retrieval works with any distance metric. \par We consider two distance metrics on the feature space L+$\Delta KL$ PTF to compute all pairwise distances in the methods $SelectSpy(\cdot)$, $ObtainNN(\cdot)$ and $ObtainFN(\cdot)$ (line 4-6, Algorithm 1): (1) Euclidean, (2) Distance metric learned from data. Specifically, we use the large margin method in \cite{weinberger2005distance} which learns a Mahalanobis distance metric $d_M(\cdot)$ that optimizes kNN classification in the training data using $d_M$. The goal is to learn $d_M(\cdot)$ such that the k-nearest neighbors (based on $d_M(\cdot)$) of each sample have the same class label as itself while different class samples are separated by a large margin.
\begin{figure}[t!]
\hrule
{\strut\footnotesize Algorithm 1: Spy induction}
\hrule
\footnotesize
\begin{flushleft}
$ SpyInduction(D,N_S,N_Q,N_R)$\\
$ 1:P\leftarrow \{x \in D.Train, x.label>0\}//positive\,class$\\
$ 2 :I\leftarrow \{(n_S,n_Q,n_R)|n_S\in $ $ N_S,n_Q\in N_Q,n_R\in N_R\} $\\
$ 3 :for\ each (i=(n_S,n_Q,n_R )\in$ $ I)$\\
$ 4 :\quad S\leftarrow SelectSpy(P,n_S)$\\
$ 5 :\quad Q\leftarrow ObtainNN(D.Test,S,n_Q)$\\
$ 6 :\quad R\leftarrow ObtainFN(D.Test,S,n_R)$\\
$ 7 :\quad (Q^v,R^v)\leftarrow CoLabelingVerification(Q,R,D.Train)$\\
$ 8 :\quad F1(i)\leftarrow CVImprovedTraining(D.Train,Q^v,R^v)$\\
$ 9 :endfor$\\
$ 10:(n_S,n_Q,n_R)^*\leftarrow {argmax}\ _{i\in I}{(F1(i))}$\\
$ 11:AV\leftarrow Classifier(D.Train,D.Test,(n_S,n_Q,n_R)^*)$\\
\end{flushleft}
\normalsize
\hrule
\caption{Spy induction}
\label{fig:algorithm1}
\end{figure}
\subsection{Label Verification via Co-Labeling}
As it is not guaranteed that the distances between samples can capture the notion of authorship, the previous step can have errors, i.e., there may be some positive samples in $R$ and negative samples in $Q$. To solve this, we apply co-labeling \cite{xu2015co} for label verification. In co-labeling, multiple views are considered for the data and classifiers are built on each view. Majority voting based on classifier agreement is used to predict labels of unlabeled instances. In our case, we consider $D.Train$ to train an SVM on five feature spaces (views): i) unigam, ii) unigram+bigram, iii) PTF, iv) POS , v) $\Delta KL$ PTF+unigram+bigram as five different label verification classifiers. Then, the labels of samples in $Q$ and $R$ are verified based on agreements of majority on classifier prediction. Samples having label discrepancies are discarded to yield the verified retrieved samples, $(Q^v,R^v)$ (line 7, Algorithm 1). The rationale here is that it is less probable for majority of classifiers (each trained on a different view) to make the same mistake in predicting the label of a data point than a single classifier.
\subsection{Improved Training}
The retrieved and verified samples from the previous steps are put back into the training set. However, the key lies in estimating the right balance between the amount of spies sent, and the size of the neighborhood considered for retrieving potentially positive or negative samples, which are governed by the parameters $n_S,n_Q,n_R$. To find the optimal parameters, we try different values of the parameter triple, $i=(n_S,n_Q,n_R )\in I$ (lines 2, 3 Algorithm 1) and record the F-Score of 5-fold CV on $D.Train\cup Q^v\cup R^v$ as $F1(i)$ (line 8, Algorithm 1). This step is carried out by the method $CVImprovedTraining(.)$. Finally, the parameters that yield the highest $F1$ in training are chosen (line 10, Algorithm 1) to yield the output spy induced verifier (line 11, Algorithm 1).
\section{Experimental Evaluation}
This section evaluates the proposed spy method. We keep all experiment settings same as in section $5$ (i.e., use out-of-training with varying author diversity $\delta$). We fix our feature space to L+$\Delta KL$ PTF as it performed best (see Table 3). As mentioned earlier, we report average verification performance across all authors. Below we detail baselines, followed by results and sensitivity analysis.
\subsection{Baselines and Systems}
We consider the following systems:\\
\textbf{MBSP} runs the Memory-based shallow parsing approach \cite{luyckx2008authorship} to authorship verification that is tailored for short text and limited training data.\\
\textbf{Base} runs classification without spy induction and dovetails with Table 3 (last row) for each $\delta$.\\
\textbf{TSVM} uses the transductive learner of SVMLight \cite{joachims1999transductive} and aims to leverage the unlabeled (test) set by classifying a fraction of unlabeled samples to the positive class and optimizes the precision/recall breakeven point.\\
\textbf{Spy (Eu.) \& Spy (LM)} are spy induction systems without co-labeling but use Euclidean (Eu.) and learned distance metric (LM) to compute neighbors.\\
\textbf{Spy (EuC) \& Spy (LMC)} are extensions of previous models that consider label verification via co-labeling approach.
\begin{table}[t!]
\centering
\scalebox{1}{
\begin{tabular}{lrrrrrrrrrc}
\toprule
\multicolumn{10}{c}{ \bf{$\delta$=25\%}} & \\
& \multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \bf{AVG} \\
Model & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & F1 \\
\midrule
MBSP & 22.9 & 84.1 & 32.0 & 22.1 & 82.1 & 31.1 & 20.7 & 77.5 & 23.5 & 28.9 \\
Base & 25.6 & 79.2 & 36.3 & 21.7 & 80.2 & 32.1 & 22.3 & 81.5 & 27.8 & 32.1 \\
TSVM & 30.6 & 43.9 & 34.3 & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & 34.3 \\
Spy(Eu.) & 39.1 & 51.2 & 40.6 & 51.6 & \underline{42.3} & 43.7 & 43.3 & \underline{52.5} & 39.4 & 41.2 \\
Spy(LM) & 42.2 & 49.3 & 42.0 & 44.4 & \underline{49.6} & 43.9 & 34.9 & \underline{62.8} & 33.7 & 39.9 \\
Spy(EuC) & 42.0 & \underline{57.8} & 42.7 & 51.2 & \underline{61.3} & \colorbox{gray!50}{52.5} & 41.5 & \underline{57.9} & 38.4 & \textbf{44.5} \\
Spy(LMC) & 38.1 & \underline{60.5} & 40.6 & 42.9 & \underline{64.9} & 47.1 & 35.3 & \underline{68.1} & 36.0 & 41.2 \\
\multicolumn{10}{c}{(A)} & \\
\midrule
\multicolumn{10}{c}{ \bf{$\delta$=50\%}} & \\
&\multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \bf{AVG} \\
Model & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & F1 \\
\midrule
MBSP & 31.9 & 85.3 & 42.1 & 25.0 & 81.6 & 34.6 & 21.1 & 84.4 & 28.7 & 35.1 \\
Base & 33.7 & 75.9 & 42.8 & 31.1 & 79.4 & 41.9 & 26.9 & 79.5 & 30.3 & 38.3 \\
TSVM & 20.2 & 83.6 & 31.1 & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & 31.1 \\
Spy(Eu.) & 39.1 & 71.2 & 45.9 & 38.1 & 67.9 & 46.2 & 45.1 & \underline{58.7} & 41.7 & 44.6 \\
Spy(LM) & 40.1 & 68.5 & 45.9 & 44.7 & 55.5 & 45.6 & 42.7 & \underline{66.2} & 40.2 & 43.9 \\
Spy(EuC) & 62.3 & \underline{52.0} & 52.3 & 62.5 & \underline{64.6} & \colorbox{gray!50}{61.0} & 46.6 & \underline{62.3} & 43.2 & \textbf{52.2} \\
Spy(LMC) & 46.8 & \underline{60.9} & 48.0 & 51.5 & \underline{67.4} & 53.7 & 40.7 & \underline{67.6} & 39.2 & 47.0 \\
\multicolumn{10}{c}{(B)} & \\
\midrule
\multicolumn{10}{c}{\bf{$\delta$=75\%}} & \\
& \multicolumn{3}{c}{\bf{SVM}} & \multicolumn{3}{c}{\bf{LR}} & \multicolumn{3}{c}{\bf{kNN}} & \bf{AVG} \\
Model & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & \multicolumn{1}{c}{P} & \multicolumn{1}{c}{R} & \multicolumn{1}{c}{F1} & F1 \\
\midrule
MBSP & 49.9 & 80.4 & 57.2 & 53.9 & 81.9 & 59.1 & 33.8 & 82.2 & 38.6 & 51.6 \\
Base & 50.5 & 71.9 & 56.2 & 46.3 & 79.4 & 54.9 & 42.2 & 80.8 & 46.1 & 52.4 \\
TSVM & 34.4 & 80.4 & 45.8 & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & 45.8 \\
Spy(Eu.) & 55.6 & 70.8 & 58.2 & 50.9 & 77.5 & 57.9 & 57.7 & 57.6 & 50.7 & 55.6 \\
Spy(LM) & 53.1 & 62.8 & 54.3 & 51.1 & 69.6 & 56.1 & 51.8 & 57.7 & 45.8 & 52.1 \\
Spy(EuC) & 71.9 & \underline{59.1} & 62.4 & 68.9 & 75.6 & \colorbox{gray!50}{70.2} & 63.6 & \underline{59.0} & 54.7 & \textbf{62.4} \\
Spy(LMC) & 55.6 & \underline{72.3} & 58.4 & 60.8 & 68.5 & 61.4 & 53.4 & \underline{60.8} & 48.7 & 56.2 \\
\multicolumn{10}{c}{(C)} & \\
\bottomrule
\end{tabular}%
}
\caption{P: Precision, R: Recall, F1: F-Score results for spy induction under out-of-training with different values of $\delta$ for three classifiers. AVG reports the average F1 across three classification models. Feature Set: L+$\Delta KL$ PTF. Gains in AVG F1 using spy (EuC) and (LMC) over baselines are significant at p$<$0.001 using a t-test}
\end{table}%
\subsection{Results}
Table 4 reports the results. We note the following:
\begin{itemize}[leftmargin=*]
\item Except for two cases (F1 of SVM and kNN for Spy(LM) with $\delta=75\%$), almost all spy models are able to achieve significantly higher F1 than base (without spy induction) and TSVM for all classifiers SVM, LR, kNN and across all diversity values $\delta$. MBSP performs similarly as Base showing memory based learning does not yield a significant advantage in sockpuppet verification. TSVM is not doing well on F1 but improves recall. One reason could be that due to class imbalance, TSVM has some bias in classifying unlabeled examples to positive class that improves recall but suffers in precision.
\item The AVG F1 column shows that on average, across three classifiers spy induction yields at least 4\% gain or more. The gains in AVG F1 are pronounced for $\delta =25\%$ with gains upto 12\% with spy (EuC). For $\delta =75\%$, we find gains of about 10\% in F1 with spy (EuC). Note that we employ out-of-training setting with varying author diversity ($\delta$) so the test set is imbalanced (i.e., the random baseline is no longer 50\%). Across all classifiers, the relative gains in F1 for spy methods over base reduce with increase in author diversity $\delta$ which is due (a) better $\neg a$ samples in training that raise the base result and (b) test set size and variety reduction limiting spy induction. Nonetheless, we note that for $\delta=25\%$ (harder case of verification), spy induction does well across all classifiers.
\item Anchoring on one distance metric (Eu./LM), we find that spy induction with co-labeling does markedly better than spy induction without co-labeling across all $\delta$ in AVG F1 across three classifiers. This shows label verification using co-labeling is helpful in filtering label noise and an essential component in spy induction.
\item Between Euclidean and distance metric learned via large-margin (LM), Euclidean does better than LM in AVG F1 for both spy induction with and without co-labeling. However, using the LM metric yields higher recall than Euclidean in certain cases (underlined) which shows LM metric can yield gains in F1 beyond base with relatively lesser drops in recall which is again useful.
\end{itemize}\par
In summary, we can see that spy induction works in improving the F1 across different classifiers and author diversity and distance metrics. Overall, the scheme LR+Spy (EuC) does best across each $\delta$ (highlighted in gray) and is used for subsequent experiments to compare against Base.
\subsection{Spy Parameter Sensitivity Analysis}
To analyze the sensitivity of the parameters, we plot the range of precision, recall and F1 values as spy induction learns the optimal values in training. We focus on the variation for $\delta =25\%,75\%$ capturing both extremes of diversity. Figure 3 shows the
performance curves for different spy parameter triples $(n_S,n_Q,n_R)$ sorted in the increasing order of F1. We find that for both $\delta =25\%,75\%,$ the spy induction steadily improves precision with the increase in likely $\neg a$ samples $(n_R$). Although the recall drops more and has more fluctuations for the harder case of $\delta =25\%$, it stabilizes early for $\delta =75\%$ with much lesser drop in recall. This shows that the spy induction scheme is robust in optimizing F1 with only a few (5-7) spy samples $(n_S)$ sent to unlabeled test set.
\begin{table}[t!]
\centering
\scalebox{1}{
\begin{tabular}{lcccccc}
\toprule
\multirow{2}[4]{*}{} & \multicolumn{2}{c}{$\delta$ =25\%} & \multicolumn{2}{c}{ $\delta$=50\%} & \multicolumn{2}{c}{$\delta$ =75\%} \\
\midrule
& Base & Spy & Base & Spy & Base & Spy \\
Test Domain & F1 & F1 & F1 & F1 & F1 & F1 \\
Hotel & 30.3 & 36.4 & 40.0 & 47.0 & 50.3 & 52.6 \\
Product & 29.5 & 36.6 & 34.0 & 40.8 & 51.5 & 53.5 \\
Restaurant & 30.1 & 41.1 & 41.7 & 51.5 & 55.2 & 59.3 \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
}
\caption{ Cross domains results of LR + Spy (EuC). Gains in F1 using spy induction over base are significant at p$<$0.01 for all test domains and each $\delta$ using a t-test.}
\end{table}%
\begin{figure}
\includegraphics[scale=.275]{fig3}
\caption{Spy Parameter Sensitivity. Variation of precision, recall and F1 across different parameter triples $(n_S,n_Q,n_R)$.}
\end{figure}
\subsection{Domain Adaptation}
We now test the effective of spy induction under domain transfer. As mentioned in section 3, we obtained reviews of Turkers for hotel, restaurant and product domains. Keeping all other settings same as in Table 4, Table 5 reports results for cross domain performance by training the verifiers ($AV_a$) using two domains and testing on the third domain. We compare sockpupet verification using LR+Spy (EuC) vs. base (LR without spy induction). We report the F1 scores as the trends of precision and recall for cross domain were similar to the trends in Table 4. The F1 of base in cross domain (Table 5, Hotel row) is lower than corresponding LR results with base (Table 4) for all $\delta$ showing cross domain verification is harder. Nonetheless, spy induction is able to render statistically significant gains in F1 for all $\delta$ (see Table 5).
\begin{table}[t!]
\centering
\begin{tabular}{lcccccc}
\toprule
\multirow{2}[4]{*}{} & \multicolumn{2}{c}{$\delta$ =25\%} & \multicolumn{2}{c}{ $\delta$=50\%} & \multicolumn{2}{c}{$\delta$ =75\%} \\
\midrule
& Base & Spy & Base & Spy & Base & Spy \\
Classifier & F1 & F1 & F1 & F1 & F1 & F1 \\
SVM & 50.7 & 57.6 & 59.6 & 62.9 & 68.0 & 70.3 \\
LR & 40.4 & 42.4 & 49.8 & 51.1 & 60.0 & 61.5 \\
kNN & 23.9 & 29.5 & 32.0 & 37.1 & 43.0 & 51.1 \\
\bottomrule
\end{tabular}%
\label{tab:addlabel}%
\caption{ Performance gains of Spy (EuC) in F1 over Base on Wikipedia Sockpuppet Dataset. Gains are significant (p$<$0.01) except for LR δ =50\%, 75\% }
\end{table}%
\subsection{Performance on Wikipedia Sockpuppet (WikiSock) Dataset}
In \cite{SOLORIO14.1007}, a corpus of Wikipedia sockpuppet authors was produced. It contains 305 authors with an average of 180 documents per author and 90 words per document which we use as another benchmark for evaluating our method.\\ It is important to note that the base results reported in \cite{SOLORIO14.1007} are not directly comparable to this experiment (Table 6). This is because \cite{SOLORIO14.1007} used all 623 cases that were found as candidates but we focus on only 305 of them which were actually confirmed sockpuppets by Wikipedia administrators. Next, we perform experiments under realistic out-of-training setting and varying the author diversity (as in Table 4) which is different from \cite{SOLORIO14.1007}. This explains the rather lower F1 as reported in \cite{SOLORIO14.1007} for Base. We focus on F1 performance of spy (EuC) versus base (without spy) as the precision and recall trends were same as in Table 4. Compared to Table 4 base results, base does better for SVM and LR on WikiSock dataset that hints the data to be slightly easier. The relative gains of spy over base although are a bit lower than those in Table 4, spy induction consistently outperforms base.
\section{Conclusion}
This work performed an in-depth analysis of deceptive sockpuppet detection. We first showed that the problem is different from traditional authorship attribution or verification and gets more difficult with the increase in author diversity. Next, a feature selection scheme based on KL-Divergence of stylistic language models was explored that yielded improvements in verification beyond baseline features. Finally, a transduction scheme, spy induction, was proposed to leverage the unlabeled test set. A comprehensive set of experiments showed that the proposed approach is robust across both (1) different classifiers, (2) cross domain knowledge transfer and significantly outperforms baselines. Further, this work produced a ground truth corpus of deceptive sockpuppets across three domains.
\section*{Acknowledgments}
This work is supported in part by NSF 1527364. We also thank anonymous reviewers for their helpful feedbacks.
\bibliographystyle{splncs}
|
\chapter{Introduction}
\label{Chapter1}
\lhead{Chapter 1. \emph{Introduction}}
Two rich people want to determine which one of them is richer so that he would pay the bill for the dinner. However, none of them is willing to permit the other learn more information about his personal wealth than what the mere knowledge of who is richer does. They start discussing how they could achieve this just by talking to each other. They are quite sure that both will always tell the truth since they are honourable businessmen who cannot take the risk of being caught while lying. On the other hand, both suspect that the other may try to deduce information about his wealth from the conversation. After some time of discussion, they come to the conclusion that it is impossible to decide who is richer under these conditions since they do not know much about \textit{secure computation} techniques.
This famous problem is known as ``millionaires' problem'' proposed by Andrew Yao \cite{Yao82}. He has also proposed a cryptographic solution for this problem, and generalized it to the secure computation of any function \cite{Yao82, Yao86}. His later work has showed that any function that can be computed by a polynomial-size circuit can be computed securely \cite{Yao86}. The problem has further widened and solved for the case of more than two parties \cite{Yao82, GMW87}. Yao's research is followed by many others' in constituting an active subfield of cryptography known as \textit{secure multi-party computation} (MPC) or \textit{secure function evaluation} (SFE), which aims solving the problem of two or more parties computing a function jointly without revealing their secret inputs to each other.
There are many real-life examples where MPC techniques can be applied, including financial systems \cite{BTW12}, cooperation of intelligence agencies, companies and governments \cite{LP08, HLOW16}, electronic elections \cite{CGS97}, electronic auctions \cite{BCD+09, NPS99}, secure biometric identification \cite{KN13, BCP13, KGK15}, secure e-mail filtering \cite{LADM14}, \textit{etc.} In fact, there is no bound for the areas where MPC may be used, and it can be adopted in any case some parties are required to compute a function on their private data.
Various methods have been proposed for MPC, including generic methods and function specific methods. Although function specific methods usually run more efficiently, they are limited in use due to the fact that each of them works for only one function. It is quite inefficient to design a method and to prove its security for each different function unless the function will be used many times. An example of frequently used functions is the Hamming distance calculation which is used in many scenarios, including biometric checks \cite{KN13} $etc.$ Hence, designing a specific protocol for it while proving its security makes sense \cite{BCP13, KGK15}. However, general research approach is towards the generic methods which can be applied to arbitrary functions.
Generic methods have been developed for usage in an unlimited set of functions. Usually one method is better than the other for different computational settings. For instance, \textit{homomorphic encryption} will be a very good fit for arithmetic circuits if an efficient fully homomorphic encryption scheme become available in the future \cite{Sch11}. However, currently the proposed fully homomorphic encryption schemes are inefficient for practical secure computation.
The most efficient methods for secure computation of functions represented as boolean circuits include \textit{GMW protocol} \cite{GMW87} and \textit{Yao's garbled circuit protocol} (Yao's protocol). The former usually gives better results in the presence of at least three parties, while the latter is usually better for two-party case.
Yao's protocol remains one of the most important paradigms for MPC, especially in the case of \textit{secure two-party computation} (2PC) \cite{ZRE15}. In particular, it is valuable for its constant round complexity. Since the time it was proposed by Andrew Yao in \cite{Yao86}, it has become one of the major fields in modern cryptographic research. It is constantly being optimized in terms of communication complexity and computation complexity.
While the research for optimizing Yao's protocol scheme continues, various practical applications using Yao's protocol have also been developed. These applications demonstrate that it is a promising cryptographic primitive for a wide range of applications, including privacy preserving data mining, efficient secure two-party computation, private function evaluation \textit{etc}.
In this thesis, we first describe the Yao's protocol in detail, and include the complete list of optimizations over the Yao's protocol. We also {compare their advantages in terms of communication and computation complexities, and analyse their compatibility with each other}. We also look into generic Yao implementations (including garbled RAM) to demonstrate the use of this powerful tool in practice. We compare those generic implementations in terms of their use of garbled circuit optimizations. We also cover the specific real-world applications for further illustration. Moreover, in some scenarios, the functionality itself may also need to be kept private which leads to an ideal solution of secure computation problem. In this direction, we finally cover the problem of {Private Function Evaluation}, in particular for the 2-party case where garbled circuits have an important role. We finally analyse the generic mechanism of Mohassel \textit{et al.} and contribute to it by proposing {a new technique for the computation of the number of possible circuit mappings}.
\section{Overview of the Thesis} \label{sec:overviewofthesis}{}
\begin{large}
\textbf{Research goal:}
\end{large}
\begin{framed}
Our goal in this thesis is to compare the advantages of currently known Yao's protocol optimizations in terms of communication and computation complexities, to analyse their compatibility with each other, to demonstrate their role with a view towards its practical and real-world applications and in private function evaluation. We intend to describe the current state of the art for Yao's protocol, since it is hard to find many comprehensive works about it. We believe that this work will be quite useful to cryptography community as a study material as well.
\end{framed}
\begin{large}
\textbf{Organization of the thesis:}
\end{large}
\textbf{Chapter \ref{Chapter1}: Introduction} \\
Chapter \ref{Chapter1} is dedicated to introduction and overview of the thesis.
\textbf{Chapter \ref{Chapter2}: Preliminaries} \\
Chapter \ref{Chapter2} is dedicated to generic MPC methods, and to cryptographic basis. We also included a section for circuit concepts which is assumed to be helpful for the people with potentially different backgrounds.
\textbf{Chapter \ref{Chapter3}: Yao's Garbled Circuit Protocol} \\
Chapter \ref{Chapter3} includes general description and formal definition of Yao's protocol, as well as the generic Yao's protocol template together with its security properties.
\textbf{Chapter \ref{Chapter4}: Garbled Circuit Optimizations} \\
Chapter \ref{Chapter4} presents known garbled circuit optimizations in a chronological order (\textit{i.e.}, P\&P (\ref{sec:PP}), GRR3 (\ref{sec:GRR3}{}), free \texttt{XOR} (\ref{sec:freeXOR}), GRR2 (\ref{sec:GRR2}{}), fle\texttt{XOR} (\ref{sec:fleXOR}{}), half gates (\ref{sec:halfgates}{})). We analyze these optimizations in terms of their relations and contradictions as well as their compatibility with each other. One of our aims is to give a clear overview, therefore, we did not get involved with proofs and other related complex formulas.
\textbf{Chapter \ref{chap:implementations}: Practical Implementations of Yao's Protocol} \\
Chapter \ref{chap:implementations} composes of generic Yao's protocol applications and some real-world examples, including pipelining method, garbled RAM, MPC for satellite collusion probability, and privacy preserving data mining.
\textbf{Chapter \ref{Chapter6}: Private Function Evaluation} \\
Chapter \ref{Chapter6} is dedicated to private function evaluation. We intend to describe Mohassel \textit{et al.}'s generic PFE scheme, which is the most efficient to date, and its application to Yao's protocol. We contribute to it by proposing {a new technique for the computation of the number of possible circuit mappings}.
\textbf{Chapter \ref{chap:conclusion}: Conclusion and Discussions} \\
Chapter \ref{chap:conclusion} concludes with general discussions of garbled circuit optimization techniques, Yao's protocol applications and private function evaluation.
\chapter{Preliminaries}
\label{Chapter2}
\lhead{Chapter 2. \emph{Preliminaries}}
In this chapter, we will present the basic concepts of secure computation techniques. First, we will show the required properties for a secure computation scheme. We will continue with general adversary models in cryptographic protocols. This will be followed by circuit concepts useful for MPC techniques which, we suppose, will be quite helpful for people new to the area. Then, we will present general cryptographic primitives. We will also give the summary of oblivious transfer protocol, homomorphic encryption, and GMW protocol.
\section{Requirements of Secure Multi-Party Computation} \label{sec:mpc}{}
To formally claim and prove the security of an MPC protocol, some general security properties are required \cite{LP08}. The most central of these properties are described in \cite{LP08} by Lindell \textit{et al.} as follows:
\begin{enumerate}
\item{\textit{Correctness:} The output that is delivered to each party (\textit{i.e.} each participant of the MPC protocol) is guaranteed to be correct.}
\item{\textit{Privacy}: None of the participants is allowed to learn anything more about other participants' inputs than what he can learn from the output itself.}
\item{\textit{Independence of inputs:} The protocol may not allow any of the parties to choose his input based on other parties' inputs. This property is different from privacy since choosing an input dependent on another party's unknown input is possible .}
\item{\textit{Guaranteed output delivery:} In the end of the protocol, honest parties should receive their outputs no matter how hard corrupt parties try to prevent it.}
\item{\textit{Fairness:} A party whether he is corrupt or not can receive his output if all of the parties receive their outputs. For detailed information about how to achieve efficient fair MPC, we refer the reader to \cite{KiS08, Kir08}.}
\end{enumerate}
Lindell \textit{et al.} stress that this list does not define security, but rather compose of the requirements that any secure protocol must conform \cite{LP08}.
\section{Adversary Models} \label{sec:adversarymodels}{}
Security of cryptographic protocols are formalized and proved against adversaries with different capabilities \cite{Sch11}.
\subsection{Semi-Honest Adversaries} \label{sub:semi-honest}{}
The \textit{semi-honest} (also known as passive, or honest-but-curious) threat model is the standard adversary model for MPC. Here parties typically follow the protocol as they are supposed to but may try to deduce information about another party's input from the protocol transcript \citep{HEKM11}. If a protocol is secure against semi-honest adversaries, it does not allow them to learn any extra information from the protocol.
\subsection{Covert Adversaries} \label{sub:covert}{}
\textit{Covert adversaries} constitute the type of adversaries that are allowed to deviate from the protocol with a restriction that they must evade being caught while they are doing so \cite{Sch11}. It can be safely assumed that in many political, social and business scenarios, the gain from cheating is overweighted by the results of being caught. If those deviations are detected with a certain frequency (\textit{e.g.}, 1 out of 10 times), such a protocol can be considered secure enough. If a protocol is secure against covert adversaries, it allows catching those adversaries with a certain probability if they deviate from the protocol.
\subsection{Malicious Adversaries} \label{sub:malicious}{}
The strongest type of adversaries is the \textit{malicious adversaries} (also known as active adversaries), which may deviate from the protocol arbitrarily so that they can extract the other parties private inputs or alter the computation outcome \cite{Sch11}. If a protocol is secure against malicious adversaries, a corrupt party will be caught whenever he deviates from the protocol.
Throughout this thesis, we focus on the security against semi-honest adversaries due to the following reasons \citep{HEKM11}:
\begin{enumerate}
\item{There are many real-world situations where modelling the parties as semi-honest adversaries is appropriate:
\begin{enumerate}
\item{where parties are legitimately trusted but there is a legal need for preventing them from divulging information, or for protection against break-ins in the future.}
\item{where the software used for MPC can hardly be changed by participants without being detected, either due to software attestation use or the fact that internal controls are in place (\textit{e.g.}, when parties are government agencies, or large corporations).}
\end{enumerate}}
\item{Securing protocols against semi-honest adversaries is an important step toward construction of secure protocols against stronger adversaries. There are generic ways of altering them to achieve security against covert or malicious adversaries \cite{Kir08, KiS06}.}
\end{enumerate}
\section{Corruption Models} \label{sec:corruptionmodels}{}
Apart from the above adversary models, there also exist static and adaptive corruption models.
\textit{Static corruption model:} This model implies that if a party is honest in the beginning, he always remains honest; whereas if a party is corrupted in the beginning, he always remains corrupted \cite{LP08}.
\textit{Adaptive corruption model:} Instead of including a fixed number of corrupted parties, adaptive corruption model suggests that the number of corrupted parties may increase during the computation. However, if a party gets corrupted, it remains that way from then on \cite{LP08}. Therefore, there may never be a decrease in the number of corrupted parties.
\section{Circuit Concepts}\label{sec:circuitconsepts}{}
For a generic MPC protocol to take place, first a function must be written as a combination of common building blocks, \textit{i.e.}, they must be represented as \textit{circuits}. The number of types of building blocks is limited. Therefore, by showing how to compute each building block, a generic MPC scheme permits calculation of unlimited functions. Standard circuit representations generally used in MPC protocols are boolean circuits and arithmetic circuits \cite{Sch11}.
\subsection{Boolean circuits}
In engineering and computer science, functions are classically represented as \textit{Boolean circuits} \cite{Sch11}. A boolean circuit basically composes of \textit{logic gates} and \textit{wires} connecting them \cite{Ana14}. Figure \ref{fig:booleancirc} shows an example boolean circuit whose wires are $a$, $b$, $c$, $d$, $e$, $f$, $h$, $k$, and $o$, and gates are $g1$, $g2$, $g3$, $g4$, and $g5$.
$a$, $b$, and $c$ are the \textit{input}s of the circuit in Figure \ref{fig:booleancirc}, $d$, $e$, $f$, $h$, and $k$ are the \textit{intermediate wires}, and $o$ is the \textit{output} wire. A boolean circuit may have more than one output as well. A wire is exactly 1 bit that may have one of the two truth values, \textit{i.e.}, either \texttt{TRUE} (also denoted as 1 or \texttt{High}) or \texttt{FALSE} (also denoted as 0 or \texttt{Low}). When 2 wires cross each other, they are connected if there is a big dot in the connection point, otherwise they are not connected. For example $a$ and $b$ cross each other but not connected (the same applies to $d$ and $e$ in Figure \ref{fig:booleancirc}).
\begin{figure}
\includegraphics[height=6.2cm, angle=0]{booleancirc}
\caption{ An example boolean circuit.}
\label{fig:booleancirc}
\end{figure}
A logic or boolean gate generally takes 1 or 2 wires as input (although there is no certain limitation) and outputs exactly 1 wire. Formally a $d$-input gate $G_d$ is a boolean function mapping $d>0$ bits input to 1-bit output, \textit{i.e.} \cite{Sch11}:
\begin{equation}
G_d: (in_1,\ldots,in_d) \in \{0,1\}^d \rightarrow (out)\in \{0,1\}
\end{equation}
For the gates of the circuit in Figure \ref{fig:booleancirc}, the left sides are used for the input, the right side are used for the output. For example $g1$ in Figure \ref{fig:booleancirc} takes $a$ and $b$ as inputs and outputs $d$. However, gates may be rotated in a different circuit. In this case, one needs to look at the two asymmetric sides of a gate. Generally, the larger assymetric side of the gate is the side of inputs and the narrower assymetric side is for the output. \textit{A wire can only be an output of exactly 1 gate}, although it can be input to multiple gates \cite{Ana14}.
In Figure \ref{fig:booleancirc}, $g1$ is an \texttt{OR} gate ($d\gets a\vee b$), $g4$ is an \texttt{XOR} gate ($h\gets d\oplus e$), and $g5$ is an \texttt{AND} gate ($k\gets d\wedge f$). If there is a bubble on the wire, its truth value is inverted after the bubble. For example, $g2$ would have been an \texttt{AND} gate without the bubble on its output. But the bubble means the output is inverted. Actually, there is a special name for the type of $g2$, it is a \texttt{NAND} gate ($e\gets (a\wedge b)'$). $g3$ would have been a \texttt{buffer} gate without the bubble on its output. A \texttt{buffer} gate outputs the input as it is. However, with the bubble $g2$ is a \texttt{NOT} gate ($f=c'$). $g6$ would have been an \texttt{OR} gate without the bubbles on its inputs. Now, it takes the inputs inverted, and \texttt{OR}s them afterwards ($o\gets h'\vee k'$). Actually $g6$ is another representation of a \texttt{NAND} gate due to the logic identity $h'\vee k'= (h\wedge k)'$. There also exist \texttt{NOR} gates represented as an \texttt{OR} gate with a bubble on its output.
The truth table of a gate shows the relation between its possible inputs and its possible outputs. The truth table of a gate has $2^k$ rows where $k$ is the number of its input wires. The truth table of the \texttt{AND} gate $g5$ in Figure \ref{fig:booleancirc} can be seen in Table \ref{tab:truthtable}.
\begin{table}[]
\centering
\caption{Truth table of an \texttt{AND} gate ($g5$ in Figure \ref{fig:booleancirc}).}
\label{tab:truthtable}
\begin{tabular}{|cc|c|}
\hline
$d$ & $f$ & $k=d\wedge f$ \\ \hline
0 & 0 & 0 \\
0 & 1 & 0 \\
1 & 0 & 0 \\
1 & 1 & 1 \\ \hline
\end{tabular}
\end{table}
In fact, there are basically $2^4$ different 2-input gates in total. However, some of them are trivial (\textit{i.e.}, the ones whose output depends only one of the inputs and the ones whose output depends none of the inputs). Those gates can be replaced by more efficient representations, \textit{e.g.}, wires, \texttt{NOT} gates, \textit{etc.} The remaining non-trivial gates fall into the category of either even gates or odd gates \cite{PSSW09}.
\begin{definition}{Even gates}
are the 2-input gates whose truth table has 2 \texttt{FALSE} outputs and 2 \texttt{TRUE} outputs.
\end{definition}
\begin{definition}{Odd gates}
are the 2-input gates whose truth table has either 3 \texttt{FALSE} outputs and 1 \texttt{TRUE} output or 1 \texttt{FALSE} output and 3 \texttt{TRUE} outputs.
\end{definition}
There are only 2 non-trivial even gates which are \texttt{XOR} and \texttt{XNOR}, and 8 non-trivial odd gates, including \texttt{OR}, \texttt{AND}, \texttt{NOR}, \texttt{NAND}, \textit{etc} \cite{PSSW09}.
The \textit{size} of a boolean circuit means the number of its gates \cite{Vol99}. The \textit{depth} of a boolean circuit means the number of gates in the longest path that must be taken from any input to any output \cite{Vol99}. The \textit{topology} of a boolean circuit means the connections between its gates \cite{Vol99}. A boolean circuit can uniquely be defined by its topology and its gates.
The \textit{topological order} of a boolean circuit is that when its gates are indexed as $G_1,\ldots , G_n$, $i^{th}$, a gate $G_i$ does not get the output of a succeeding gate $G_{j>i}$ as its input \cite{Sch11}. Intuitively, in order to compute a gate, all of its input wires must be known, which can be ensured by computing the gates in topological order. By computing the gates one-by-one in topological order the whole boolean circuit can be computed. The topological order is not necessarily unique for a given boolean circuit \cite{Sch11}.
A group of gate types ($G_1,\ldots,G_n$) is \textit{Turing-complete}, if and only if any {probabilistic polynomial time} algorithm can be represented by a combination of those gates \cite{Sip96}. Examples are (\texttt{AND},\texttt{XOR}) and (\texttt{NAND}). Building a \texttt{NAND} gate from a group of gates is an easy way to see whether that group of gates is Turing-complete or not.
A decrease in the number of gates in a circuit also means a decrease in overall cost of an MPC protocol in terms of computation complexity, and communication complexity. There are various techniques for circuit optimizations. Some circuit optimization techniques intend to reduce the number of odd gates at the cost of increasing the even gates. They could also be useful in some MPC techniques \cite{KK10,BPP00}.
\subsection{Arithmetic circuits} \label{sub:arth}
A more compact representation for functions is \textit{arithmetic circuits} \cite{Sch11}. Unlike boolean circuits where wires are chosen from $ \mathbb{Z}_2$, here wires have values chosen from $\mathbb{Z}_{m\geq2}$. The gates operations are either modular addition $+$ or modular multiplication $\times$. Figure \ref{fig:arithmeticcirc} shows an example arithmetic circuit.
\begin{figure}
\includegraphics[height=3.5cm, angle=0]{arithmeticcirc}
\caption{ An example arithmetic circuit \cite{Sch11}.}
\label{fig:arithmeticcirc}
\end{figure}
One can express any boolean circuit as an arithmetic circuit over $ \mathbb{Z}_2$. However, if $\mathbb{Z}_m$ has a modulus $m$ which is sufficiently large, then the resulting arithmetic circuit representation of a function will probably have much lower size than its boolean circuit representation, since a single operation will be enough for each integer addition or multiplication \cite{Sch11}.
Computations on both positive and negative integers $x$ can be simulated by arithmetic circuits, since one can map them into elements of $\mathbb{Z}_m: \mathbb{Z}\rightarrow\mathbb{Z}_m, x\rightarrow x \texttt{ mod } m$ \cite{Sch11}.
\section{Cryptographic Basis} \label{sec:cryptographicalbasis}{}
As the cryptographic basis for this thesis, we present symmetric and public key encryptions, cryptographic hash functions, pseudo-random functions and message authentication codes. We will include only a brief summaries of them due to the fact that the details of them are not necessary for understanding protocols and that vast majority of our readers will probably have an acquaintance with them. However, at the end of this section we present dual-key ciphers in more detail because of their intensive use in Yao's protocol and supposed unfamiliarity of some readers with them.
\subsection{Symmetric Encryption} \label{sub:SymetricEncryptions}{}
A \textit{symmetric encryption} scheme uses the same cryptographic key $k$ for both encryption of plaintext and decryption of ciphertext \cite{AES-FIPS}. A well-known example is \texttt{AES} encryption \cite{AES-FIPS}. The notation $c\gets E_k(m)$ means that a plaintext message $m$ is encrypted with a key $k$ resulting in a ciphertext $c$.
Decryption is generally denoted as either $m\gets D_k(c)$ or the inverse of $E$, namely $m\gets E_k^{-1}(c)$ .
\subsection{Public Key Encryption} \label{sub:publickeycryptography}{}
A \textit{public key encryption} scheme uses different keys for encryption and decryption. \textit{Public keys} which are known publicly as their name implies are used for encryption, while private keys which are known only to their owners are used for decryption \cite{RSA78}. Any user can encrypt a message with the public key of the receiver, but the resulting ciphertext can be decrypted only with the receiver's private key. The notation $c\gets E_{pk_i}(m)$ means that a plaintext message $m$ encrypted with a public key $pk_i$ of $i^{th}$ person results in a ciphertext $c$.
Decryption with the secret key $sk_i$ of the $i^{th}$ person is denoted as either $m\gets D_{sk_i}(c)$ or $m\gets E_{pk_i}^{-1}(c)$. The well-known public key cryptosystems are \texttt{ElGamal} \cite{ElGamal85} and \texttt{RSA} \cite{RSA78}.
\subsection{Cryptographic Hash Function} \label{sub:cryptographichashfunction}{}
A \textit{cryptographic hash function} $H(m)$ maps an arbitrary size message $m$ to a fixed size $\ell$-bit string $c\gets H(m)$ \cite{RS04}. Throughout this thesis when we say \textit{hash function}, we refer to a cryptographic hash function.
Hash functions are ideally modelled in the \textit{random oracle model} \cite{KM15}. A random oracle is a theoretical black-box responding to every unique query with a true random number picked from its output domain. It records its responses to unique queries so that it can respond a query the same way every time it is repeated. A well-known hash function scheme is $\texttt{SHA}256$ \cite{Han11}.
\subsection{Pseudo-Random Function} \label{sub:pseudorandomfunction}{}
A \textit{pseudo-random function} (\texttt{PRF}) is a function that can be used for pseudo-random generation, \textit{i.e.}, it can be modelled as random oracle. It is denoted as $\texttt{PRF}(x)$ on an input $x$. Its representation can be extended as $ \texttt{PRF}_k(x)$ to include the use of a private key $k$ \cite{Sch11}.
An instantiation of \texttt{PRF} can be achieved with a block cipher, \textit{e.g.}, \texttt{AES}, or a hash function, \textit{e.g.}, $\texttt{SHA}256$. In case a \texttt{PRF} with the same key $k$ is repeatedly used, the \texttt{AES} instantiation would be more efficient since its key schedule needs to be run just once \cite{Sch11}.
\subsection{Message Authentication Code (\texttt{MAC})} \label{sub:messageauthenticationcode}{}
A \textit{message authentication code} (\texttt{MAC}) is a fixed-sized data that is used for authentication of a message. It is denoted as $\texttt{MAC}_k(m)$ on an input message $m$ that needs to be authenticated and a private key $k$ \cite{Sch11}.
The \texttt{MAC} value provides protection for both data integrity and authenticity of a message since it allows the detection of any changes in the message content by the verifiers possessing the private key $k$.
\subsection{Dual-Key Cipher} \label{part:dkcschemes}{}
A \textit{dual-key cipher} (\texttt{DKC}) is a cryptographic notion proposed by Bellare \textit{et al.} in \cite{BHR12}. A \texttt{DKC} formally represents a two-key lockbox where both keys are required for openning the box. A \texttt{DKC} is a function $E$ associating a security parameter $k \in N$ where $N$ is the set of positive integers and keys $A, B \in \{0, 1\}^k$ with a $k$-bit pseudo-random number $E_{A,B} : \{0, 1\}^k \rightarrow \{0, 1\}^k$. Let $D_{A,B} : \{0, 1\}^k \rightarrow \{0, 1\}^k$ denote the inverse of this function \cite{BHR12}.
Decryption of \texttt{DKC} may also be denoted by the inverse function notation $E^{-1}_{A,B}: \{0, 1\}^k \rightarrow \{0, 1\}^k$ instead of $D_{A,B} : \{0, 1\}^k \rightarrow \{0, 1\}^k$.
Throughout this thesis an encryption with two keys mean a \texttt{DKC} unless it is stated otherwise.
So far, a variety of \texttt{DKC} schemes have been proposed. Among them, an earlier one is Equation~\eqref{eq:encscheme2hashes} proposed by Naor \textit{et al.} in \citep{NPS99}. For every encryption, \texttt{PRF} is called twice. \texttt{PRF} may be implemented as a keyed hash.
\begin{equation}
\label{eq:encscheme2hashes}
E_{A,B}(C)\rightarrow \texttt{PRF}(A, \texttt{gateID})\oplus \texttt{PRF}(B, \texttt{gateID}) \oplus C
\end{equation}
Lindell \textit{et al.} proposed a more efficient \texttt{DKC} scheme Equation~\eqref{eq:encscheme1hash} in \cite{LPS08}. It requires one hash per encryption, which reduces the computational cost significantly.
\begin{equation}
\label{eq:encscheme1hash}
E_{A,B}(C)\rightarrow H(A||B||\texttt{gateID})\oplus C
\end{equation}
Kreuter \textit{et al.} proposed the \texttt{DKC} scheme Equation~\eqref{eq:encschemeaes} in \citep{KsS12}. An $\texttt{AES}256$ encryption is used instead of a hash function. Kreuter \textit{et al.} shows that this improvement reduces the computational cost around 25\%.
\begin{equation}
\label{eq:encschemeaes}
E_{A,B}(C)\rightarrow\texttt{AES}256(A||B||\texttt{gateID})\oplus C
\end{equation}
Bellare \textit{et al.} proposed the \textit{state-of-the-art} \texttt{DKC} scheme Equation~\eqref{eq:encschemeaesconstkey}\footnote{$K=2A\oplus4B\oplus gateID$} in \citep{BHKR13} which eliminates the need for key precessing in each AES encryption by using a constant key $k_c$ for all of them.
\begin{equation}
\label{eq:encschemeaesconstkey}
E_{A,B}(C)\rightarrow\texttt{AES}128_{k_c}(K)\oplus K \oplus C
\end{equation}
\section{Secret Sharing} \label{sec:secretshare}{}
\textit{Secret sharing} refers to the methods where a secret value is distributed amongst a group of parties, each having a share from the secret \cite{Sha79}. To reconstruct the secret, parties need to combine a sufficient number of shares together; since individual share of a party is useless on its own. There have been various secret sharing schemes proposed so far. Here we will introduce only some of them which will be helpful throughout this thesis.
\subsection{\texttt{XOR} Sharing} \label{sub:xorshare}{}
\texttt{XOR} \textit{sharing} (also known as \textit{boolean sharing}) is a secret sharing type where for an $\ell$-bit value $x$ shared by $m$ parties, the share of a party $i$ is an $\ell$-bit value $x_i$, and when the shares of all $m$ parties \texttt{XOR}ed bitwise together the result is $x$, \textit{i.e.}, $x=x_1\oplus\ldots\oplus x_m$ \cite{DSZ15}. There is no number limit for parties in \texttt{XOR} sharing. However, if any of the parties keeps his share, the rest of the parties cannot even get close to learning the shared value.
\subsection{Arithmetic Sharing} \label{sub:arithshare}{}
\textit{Arithmetic sharing} is similar to \texttt{XOR} \textit{sharing} in that there is no number limit for parties and that if any of the parties keeps his share, the rest of the parties cannot even get close to learning the shared value \cite{DSZ15}. It is a secret sharing type where for an $\ell$-bit value $x$ shared by $m$ parties, the share of a party $i$ is an $\ell$-bit value $x_i$, and when the shares of all $m$ parties added together in a modulus $n$ which conforms $2\leq n\leq 2^\ell$ the result is $x$, \textit{i.e.}, $x=x_1+\ldots+ x_m \texttt{ mod } n$.
\subsection{Yao Sharing} \label{sub:yaoshare}{}
\textit{Yao sharing} is a secret sharing type where 1 bit is shared by 2 parties \cite{DSZ15}. In order to share a bit $b$, the first party $P_1$ picks 2 random $\ell$-bit strings $B_0$ and $B_1$. The second party $P_2$, without knowing $b$, keeps only $B_b$. $P_1$ does not know which of the 2 strings kept by $P_2$, and $P_2$ does not know the other string picked by $P_1$. Only together, they can evaluate $b$. Although keeping costly strings for a bit does not look very efficient at first, Yao sharing has certain advantages for 2PC which will be obvious when we describe Yao's protocol in \ref{Chapter3}.
\subsection{Shamir's Secret Sharing} \label{sub:shamirshare}{}
\textit{Shamir's secret sharing} is an effective secret sharing scheme proposed by Adi Shamir \cite{Sha79, NS10} where a group of $n$ users share a secret data $D$. The scheme permits any predefined $(k + 1)\leq n$ or more users to reconstruct the secret. However, no information about $D$ can be recovered by $k$ or less users. This scheme can also be referred to as $(k + 1,n)$-threshold secret sharing scheme, where $(k + 1)$ is the threshold and $n$ is the number of users sharing the secret.
All users have a different point in two-dimensional plane, $(x_1,y_1),\ldots,(x_n,y_n)$. All of the points must be chosen such that they are on a $k$-degree polynomial. Therefore, any $k + 1$ of these shares suffices for Lagrange's interpolation. The secret value is the evaluation of the polynomial on axis $x=0$.
\section{Oblivious Transfer} \label{sec:oblivioustrans}{}
An \textit{1-out-of-$m$ oblivious transfer} (1-out-of-$m$ OT) protocol is a two-party asymmetric\footnote{An asymmetric protocol means that parties play different roles during the protocol.} protocol where one of the parties is the sender, and the other one is the receiver \cite{Pil15}. The sender has the set of values $\{x_1,\ldots,x_m\}$ and the receiver has an index $i$. At the end of the protocol, the receiver should only learn one of the sender's inputs, which is $x_i$; whereas the sender should not learn anything about the index $i$. An efficient 1-out-of-$m$ OT technique can be found in \cite{CO15}.
The high computational complexity of OT is a major source of inefficiency. In order to reduce this cost, some optimizations (\textit{e.g.}, extended OT \cite{YKNP03}) have been proposed.
There also exist OT protocols for settings with more parties, known as \textit{multi-party oblivious transfer}. A multi-party OT is a protocol where one of the parties holds the values $x_1, \ldots, x_m$, but multiple parties secret share the choice index $i$. At the end of the protocol, the parties learn shares of $x_i$ instead of learning it as a whole. The party holding the initial values is called the sender, whereas the other ones are called the receivers.
\section{Homomorphic Encryption} \label{sec:homoenc}{}
\textit{Homomorphic Encryption} (HE) schemes are used for secure evaluation of arithmetic circuits since they permit computation of multiplication and addition on ciphertexts \cite{Sch11}. An \textit{additively} HE scheme allows only unlimited addition on encrypted data; whereas a \textit{multiplicative} HE scheme allows only unlimited multiplication on it. An encryption scheme having both multiplicatively and additively HE property is called \textit{fully homomorphic encryption} (FHE).
There was a wide-spread belief that FHE does not exist until recently. Gentry has been the inventor of the first FHE scheme \cite{Gen09}. Unfortunately, huge sizes and computational costs of current FHE schemes make them too inefficient to be used in practical applications no matter how much effort has been given for improving their performances. The problem is that a FHE scheme must allow algebraic operations while providing strong security assumptions, which makes the costs grow substantially.
\section{Goldreich-Micali-Wigderson (GMW) Protocol for MPC} \label{sec:gmwprotocol}{}
One of the commonly used MPC schemes is \textit{Goldreich-Micali-Wigderson} (GMW) protocol that uses \texttt{XOR} sharing (\ref{sec:secretshare}), and is proposed in \cite{GMW87}. It proposes MPC of boolean circuits with gates \texttt{AND} and \texttt{XOR} against semi-honest adversaries (\ref{sub:semi-honest}).
\texttt{XOR} gates can be computed locally and are communication free \cite{Pil15}. To illustrate, to compute $c=a \oplus b$, each party $i$ only needs to use its shares $c_i = a_i \oplus b_i$ in order to receive his output share $c_i$. However, to compute an \texttt{AND} gate, parties are required to communicate for 1-out-of-4 OT (\ref{sec:oblivioustrans}). In the case of $2$ parties, to compute their output shares of $a \wedge b$, $P_1$ constructs the evaluation table for both input shares of $P_2$ and they engage in a 1-out-of-4 OT (\ref{sec:oblivioustrans}) where $P_2$'s inputs are used as the choice index. To extend the protocol for $m$ parties, ${m\choose 2}$ runs of the OT protocol is required. One can also see it as one run of a multi-party 1-out-of-4 OT protocol where the choice indices are $a$ and $b$ \cite{Pil15}.
\chapter{Yao's Garbled Circuit Protocol}
\label{Chapter3}
\lhead{Chapter 3. \emph{Yao's Garbled Circuit Protocol}}
Even though Yao's protocol has more than two-party applications, its use will be held limited to 2PC. It is an asymmetric protocol, which means that parties play different roles while the protocol is running. One of the parties has the role of the \textit{garbler}, whereas the other one becomes the \textit{evaluator}. The protocol is intended to be secure in the semi-honest model (\ref{sub:semi-honest}). It runs on boolean functions, so first a function must be converted to a boolean circuit. Figures \ref{fig:abooleancircuit}, \ref{fig:anencryptedcircuit} and \ref{fig:yaosgcpflow} have been taken from Mike Rosulek's presentation in Simons Institute, University of California, Berkeley, namely \textit{A Brief History of Practical Garbled Circuit Optimizations}.
\textbf{A gentle introduction}. Yao's garbled circuit protocol is briefly as follows (later we propose it in a more formal model):
Assume Alice and Bob are trying to compute a function $f$ whose boolean circuit is given in Figure \ref{fig:abooleancircuit}. Throughout this thesis, Alice will be the {garbler}, Bob will be the {evaluator}. Alice's input is $x$ including bits $a$ and $c$, and Bob's input is $y$ including bits $b$ and $d$.
\begin{figure}[b]
\includegraphics[height=5cm, angle=0]{abooleancircuit}
\caption{ A boolean circuit of a function $f$ with the truth table of the gates included.}
\label{fig:abooleancircuit}
\end{figure}
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice picks random and computationally indistinguishable masking values for possible truth values \texttt{FALSE} and \texttt{TRUE} of each wire.}
\item{She encrypts the output masking values of each gate using their corresponding input masking values as the \texttt{DKC} key (\ref{part:dkcschemes}). This way she gets four ciphertexts for each gate in the circuit as in Figure \ref{fig:anencryptedcircuit}.}
\end{enumerate}
\texttt{Input Transfer:}
\begin{enumerate}
\setcounter{enumi}{2}
\item{She sends all ciphertexts for each gate, as well as her masked input values for $a$ and $c$ to Bob. He takes his own masked input values from Alice using 1-out-of-2 OT (\ref{sec:oblivioustrans}).}
\end{enumerate}
\texttt{Evaluating:}
\begin{enumerate}
\setcounter{enumi}{3}
\item{Bob decrypts the related ciphertext (we will come to this later) gate-by-gate in \textit{topological order}, reaching the output masking values of the circuit. Topological order means from the inputs to the output. The rule is that if the output of a gate $g_1$ is input to another gate $g_2$, $g_1$ must be evaluated before $g_2$. In this case the gate order might be chosen as 1, 2, 3, 4, 5.}
\end{enumerate}
\texttt{Output Reveal:}
\begin{enumerate}
\setcounter{enumi}{4}
\item{Bob tells Alice the output masking values, and Alice sends the output of the function $f(x,y)$ to Bob.}
\end{enumerate}
\begin{figure}
\includegraphics[height=12cm, angle=-90]{anencryptedcircuit}
\caption{Garbling the circuit in Figure \ref{fig:abooleancircuit}}
\label{fig:anencryptedcircuit}
\end{figure}
The flow of communication between the garbler and the evaluator is summed up in Figure \ref{fig:yaosgcpflow}.
\begin{figure}[b]
\includegraphics[height=10cm, angle=-90]{yaosgcpflow}
\caption{ Communication flow in the semi-honest Yao's protocol.}
\label{fig:yaosgcpflow}
\end{figure}
\section{Formal Definiton of Yao's Protocol} \label{sec:FormalDefinitonandSecurityofYao'sProtocol}{}
The Yao's protocol scheme proposed by Bellare \textit{et al.} in \citep{BHR12} brought a significant jump by defining the procedures involved in a secure Yao's protocol. A conventional circuit can be defined as $f = (n, m, q, A, B, G)$ where the numbers of its inputs, its outputs, and its gates are $n \geq 2$, $m \geq 1$, and $q \geq 1$, respectively. The number of its wires is denoted as $r = n + q$. The sets of the circuit \texttt{Inputs}, \texttt{Wires}, \texttt{OutputWires} and \texttt{Gates} are defined as $\texttt{Inputs}=\{1,\ldots,n\}$, $\texttt{Wires}=\{1,\ldots,n+q\}$, $\texttt{OutputWires}= \{n+q-m+1,\ldots,n+q\}$, and $\texttt{Gates}= \{n + 1,\ldots, n + q\}$. Then the function identifying each gate's first incoming wire is $A : \texttt{Gates} \rightarrow \texttt{Wires}\backslash \texttt{OutputWires}$. The function identifying each gate's second incoming wire is $B : \texttt{Gates} \rightarrow \texttt{Wires}\backslash \texttt{OutputWires}$. The function determining the functionality of each gate is $G : \texttt{Gates} \times \{0, 1\}^2 \rightarrow \{0, 1\}$. The requirement is that $A(g) < B(g) < g$ for all $g \in \texttt{Gates}$.
Bellare \textit{et al.} defines the generic garbling scheme consisting of \texttt{\textsc{Gb}}, \texttt{\textsc{En}}, \texttt{\textsc{Ev}}, and \texttt{\textsc{De}} algorithms which are described as folows (see also Figure \ref{fig:gcpprocflow} and Algorithm \ref{alg:yaosgcpproc}) \citep{BHR12}:
\begin{figure}[t]
\includegraphics[height=2.8cm, angle=0]{gcpprocflow}
\caption{The flow of procedures in Yao's protocol in \citep{BHR12}.}
\label{fig:gcpprocflow}
\end{figure}
\begin{enumerate}
\item{\textbf{Garble} (\texttt{\textsc{Gb}}): \texttt{\textsc{Gb}} procedure takes $1^k$ and a boolean circuit $f$ as input, and outputs $(F,e,d)$, where $F$ is a garbled circuit, $e$ is the encoding information, and $d$ is the decoding information. The \textbf{for}-loop on Line \ref{lst:assigningmaskingvalue} of Algorithm \ref{alg:yaosgcpproc} assigns masking values for every wire in the circuit for both \texttt{TRUE} and \texttt{FALSE}. It also assures that the last bits of the assigned masking values for a wire, which we call label bits, differ from each other. The \textbf{for}-loop on Line \ref{lst:encryptingoutputs} of Algorithm \ref{alg:yaosgcpproc} encrypts the possible output masked with their corresponding input masking values for each gate. It also orders the ciphertexts with respect to the label bits (\texttt{lsb}) of input masking values so that the order does not leak information (we will call this technique point and permute).}
\item{\textbf{Encode} (\texttt{\textsc{En}}): \texttt{\textsc{En}} procedure takes $(e,x)$ as input, where $e$ is as we mentioned above and $x$ is a suitable input for $f$, and outputs a garbled input $X$. In this scheme, encoding is directly assigning the pre-known masking values for the inputs.}
\item{\textbf{Evaluate} (\texttt{\textsc{Ev}}): \texttt{\textsc{Ev}} procedure takes $(F, X)$ as input, and outputs a garbled output $Y$. The \textbf{for}-loop on Line \ref{lst:evaluategates} of Algorithm \ref{alg:yaosgcpproc} decrypts only one ciphertext related to a gate with its input masking values and with respect to their label bits.}
\item{\textbf{Decode} (\texttt{\textsc{De}}): \texttt{\textsc{De}} procedure takes $(d, Y )$ as input, and outputs a plain output $y$. In this scheme, decoding is directly assigning the pre-known outputs for the masking values obtained by the \texttt{\textsc{Ev}} procedure.}
\end{enumerate}
\begin{algorithm}
\scriptsize
\caption{Garbled circuit scheme \citep{BHR12}.}
\label{alg:yaosgcpproc}
\begin{algorithmic}[1]
\Procedure{\texttt{Gb}}{$1^k,f$} \Comment \texttt{Garbling phase}
\State $(n,m,q,A',B',G)\gets f$
\For{$i\in\{1,\ldots,n+q\}$} \label{lst:assigningmaskingvalue}
\State $t\gets \{0,1\}$, $X_i^0\gets \{0,1\}^{k-1}t$, $X_i^1\gets \{0,1\}^{k-1}\overline{t}$
\EndFor
\For{$(g,i,j)\in\{n+1,\ldots,n+q\}\times \{0,1\}\times \{0,1\}$} \label{lst:encryptingoutputs}
\State $a \gets A'(g)$, $b \gets B'(g)$, $A \gets X_a^i$, $\text{a} \gets \texttt{lsb}(A)$, $B \gets X_b^i$, $\texttt{b} \gets \texttt{lsb}(B)$
\State $T \gets g\parallel\texttt{a}\parallel\text{b}$, $P[g,\texttt{a},\texttt{b}] \gets E_{A,B}(X_g^{G_g(i,j)})$
\EndFor
\State $F\gets(n,m,q,A',B',P)$
\State $e\gets(X_1^0,X_1^1,\ldots,X_n^0,X_n^1)$
\State $d\gets(X_{n+q-m+1}^0,X_{n+q-m+1}^1,\ldots,X_{n+q}^0,X_{n+q}^1)$
\State \textbf{return} $(F,e,d)$
\EndProcedure
\Statex
\Procedure{\texttt{En}}{$e,x$} \Comment \texttt{Encoding phase}
\State $(X_1^0,X_1^1,\ldots,X_n^0,X_n^1)\gets e$
\State $x_1\ldots x_n\gets x$, $X\gets (X_1^{x_1},\ldots,X_n^{x_n})$
\State \textbf{return} $X$
\EndProcedure
\Statex
\Procedure{\texttt{Ev}}{$F,X$} \Comment \texttt{Evaluating phase}
\State $(n,m,q,A',B',P)\gets F$, $(X_1,\ldots,X_n)\gets X$
\For{$g\gets n+1 \textbf{ to } n+q$} \label{lst:evaluategates}
\State $a \gets A'(g)$, $b \gets B'(g)$, $A \gets X_a^i$, $\text{a} \gets \texttt{lsb}(A)$, $B \gets X_b^i$, $\text{b} \gets \texttt{lsb}(B)$
\State $T \gets g\parallel\texttt{a}\parallel\texttt{b}$, $X_g \gets D_{A,B}(P[g,\texttt{a},\texttt{b}])$
\EndFor
\State \textbf{return} $(X_{n+q-m+1},\ldots,X_{n+q})$
\EndProcedure
\Statex
\Procedure{\texttt{De}}{$d,Y$} \Comment \texttt{Decoding phase}
\State $(Y_1,\ldots,Y_m)\gets Y$, $(Y_1^0,Y_1^1,\ldots,Y_m^0,Y_m^1)\gets d$
\For{$i\in\{1,\ldots,m\}$}
\If{$Y_i=Y_i^0$} $y_i\gets 0$
\ElsIf{$Y_i=Y_i^1$} $y_i\gets 1$
\Else \textbf{ return} $\bot$
\EndIf
\EndFor
\State \textbf{return} $y\gets y_1\ldots y_m$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\textit{Correctness} property is that Equation~\eqref{eq:correctnessofGCP} holds for all possible input $x$ where $(F,e,d)\gets\texttt{\textsc{Gb}} (1^k,f)$.
\begin{equation}
\label{eq:correctnessofGCP}
\texttt{\textsc{De}}(d, \texttt{\textsc{Ev}}(F, \texttt{\textsc{En}}(e, x))) = f (x)
\end{equation}
\section{Security Properties of Yao's Protocol}\label{sec:yaosec}{}
We need some parameters in order to appreciate the security of a garbling scheme. The security parameters defined by Bellare \textit{et al.} are \textit{privacy}, \textit{obliviousness}, and \textit{authenticity} \citep{BHR12}.
\subsection{Privacy}\label{sub:Privacy}{}
\textit{Privacy} is achieved by a garbling scheme if no more information about the input $x$ must be revealed by the collection $(F,X,d)$ than that is revealed by $f (x)$ \citep{ZRE15, BHR12}. Let $(f, x)$ be chosen by the adversary. Then either the circuit is garbled to $(F,e,d)\gets\texttt{\textsc{Gb}}(1^k,f)$, the input is encoded as $X\gets\texttt{\textsc{En}}(e,x)$, the adversary getting $(F,X,d)$; or the simulator $S$ devises a \textit{fake} $(\bar{F}, \bar{X}, \bar{d})$ depending solely on the security parameter $k$, the side information\footnote{Side-information means any information about the circuit which the protocol does not intend to hide, like its size or its topology. $\Phi(f)$ is the side-information function which maps $f$ to $\phi$. \label{ftnt:si}} $\phi = \Phi(f )$, and the output $y = \texttt{\textsc{Ev}}(f, x)$. The $(\bar{F}, \bar{X}, \bar{d})$ produced by the simulator must be indistinguishable from the ones coming from the actual garbling scheme.
\subsection{Obliviousness} \label{sub:Obliviousness}{}
\textit{Obliviousness} is achieved by a garbling scheme if $(F, X)$ reveals nothing more than the side information\textsuperscript{\ref{ftnt:si}} $\Phi(f )$ about $f$ or $x$ \citep{ZRE15, BHR12}. To compare obliviousness with privacy (\ref{sub:Privacy}), where the output is learned by the evaluator, here, he does not learn that since $d$ is kept hidden. The output can be revealed by a private scheme even without $d$, while $x$ can be revealed by an oblivious scheme once $d$ is exposed. Let $(f, x)$ be chosen by the adversary. Either the circuit is garbled to $(F,e,d)\gets\texttt{\textsc{Gb}}(1^k,f)$, the input is encoded as $X\gets\texttt{\textsc{En}}(e,x)$, and the adversary getting $(F,X)$; or the simulator $S$ to devises a \textit{fake} $(\bar{F}, \bar{X})$ depending solely on $k$, and $\phi = \Phi(f )$. The $(\bar{F}, \bar{X})$ produced by the simulator must be indistinguishable from from the ones coming from the actual garbling scheme.
\subsection{Authenticity} \label{sub:Authenticity}{}
\textit{Authenticity} is achieved by a garbling scheme if from $(F, X)$, an adversary cannot construct a garbled output $\bar{Y}$ which is not authentic, \textit{i.e.} $\texttt{\textsc{De}}(d, \bar{Y} )\neq \bot$ only if $\bar{Y}= \texttt{\textsc{Ev}}(F, X)$, except for negligible probability \citep{ZRE15, BHR12}.
\chapter{Garbled Circuit Optimizations}
\label{Chapter4}
\lhead{Chapter 4. \emph{Garbled Circuit Optimizations}}
Since we have introduced the generic garbled circuit framework, it is time to present the optimizations on it in detail. We start with describing the parameters of a garbled circuit scheme that can be optimized and their relevant importance. We then continue with optimization techniques, along with comparing them with each other and presenting the relations between them. At the end, we have included a useful table to show the compatibility of various garbling techniques. Figures \ref{fig:pointandpermute}, \ref{fig:GRR3}, \ref{fig:freexor}, \ref{fig:freexorand}, \ref{fig:grr2} and \ref{fig:flexor} have been taken from Mike Rosulek's presentation in Simons Institute, University of California, Berkeley, namely \textit{A Brief History of Practical Garbled Circuit Optimizations}.
Mainly, there are three parameters related to Yao's protocol that can be optimized: the \textit{size} of the garbled circuit which limits the communication complexity cost, the \textit{computation time} required both for encryption and decryption, and the \textit{security} of the protocol \cite{KMR14}. The size of the garbled circuit is important because it usually needs to be transmitted to the evaluator over a limited channel. Clearly, the computation time required is also an important parameter for both parties.
\section{General Focus}
\subsection{The Size Parameter} \label{sec:size}{}
The size of the garbled circuit is usually the primary parameter due to the limits of the communication channel. The most effort in the garbled circuit research has been dedicated to make it smaller. Reducing it even in the expense of worse computation times or weaker hardness assumptions is often preferable \cite{KMR14}.
A reduction in the size of a garbled circuit generally comes from a decrease in the number of ciphertexts needed per gate. Circuits can grow to contain billions of gates, meaning each garbled circuit can be gigabytes in size. Our primary goal in this chapter is to cover garbled gate size optimization techniques.
\subsection{The Computation Time Parameter} \label{sec:comptime}{}
Computation time is related to time consumptions of \texttt{\textsc{Gb}} and \texttt{\textsc{Ev}} procedures. Naturally the research aims to make them shorter. The computation time may be even more important when the \texttt{CPU} resource of a party is restricted, such as a mobile device. The improvements in \texttt{DKC} schemes (\ref{part:dkcschemes}) schemes proposed are also for this parameter. The gate garbling techniques may also improved for this parameter as well \cite{KMR14}.
\subsection{Security Parameter} \label{sec:hardassump}{}
A garbling scheme must conform the security properties (\ref{sec:yaosec}) although in some cases authenticity parameter may be omitted. If the hardness assumptions of the building blocks of a scheme (\textit{e.g.}, \texttt{DKC} scheme (\ref{part:dkcschemes}), gate garbling technique) is stronger, the protocol will also be more secure \cite{KMR14}.
The rest of this chapter is especially dedicated to the techniques related to the optimizations in the size parameter. However, the techniques will also be compared for the other parameters whenever it is necessary. After the description of each technique, there will be a size and computation time scoreboard for comparing that technique with the previous ones (see Tables \ref{tab:scrbrdpp}, \ref{tab:scrbrdgrr3}, \ref{tab:scrbrdfrxor}, \ref{tab:scrbrdfrgrr2}, \ref{tab:scrbrdfrflxor}, \ref{tab:scrbrdhg}). The time for encryptions and decryptions for both \texttt{DKC} schemes and symmetric schemes assumed to be the same and denoted as \texttt{edt} (for encryption/decryption time). ct stands for ciphertexts.
\begin{figure}
\includegraphics[height=5cm, angle=0]{pointandpermute}
\caption{ (a) The gate to be evaluated. (b) Label assignment. (c) Rearrangement of ciphertexts canonically with respect to input labels.}
\label{fig:pointandpermute}
\end{figure}
\section{Point and Permute} \label{sec:PP}{}
The evaluator needs to know which one of the ciphertexts for a gate must be decrypted during the evaluation process. However, he cannot be allowed to deduce the truth value of any of inputs or outputs. The oldest and yet secure method achieving is \textit{point and permute} (P\&P), and suggested by Beaver \textit{et al.} in \cite{BMR90}.
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of the gate in Figure \ref{fig:pointandpermute} (a) where $a$ and $b$ is the input $c$ is the output.}
\item{Alice chooses masking values of wires such that each masking value has one of the two possible labels (the one for $a$ is either $A_0$ or $A_1$, the one for $b$ is either $B_0$ or $B_1$, and the one for $c$ is either $C_0$ or $C_1$), and for a given wire both masking values have different labels (see Figure \ref{fig:pointandpermute} (b)). The label needs to be something that can be directly detectable from the masking value (\textit{e.g.}, its last bit). For example, if the masking value on the wire $a$ corresponding to the truth value \texttt{FALSE} ($A_0$) has 0 on the last bit, then the masking value for the truth value \texttt{TRUE} ($A_1$) must have 1 on the last bit. The truth value cannot be detected from the label of the masking value. Alice encrypts the possible output masking values of the gate with the corresponding input masking values ($E_{A_0,B_0}(C_0)$, $E_{A_0,B_1}(C_1)$, $E_{A_1,B_0}(C_0)$, and $E_{A_1,B_1}(C_0)$).}
\item{Alice rearrange the ciphertexts with respect to the input labels, as in Figure \ref{fig:pointandpermute} (c). During the evaluation, Bob will know which ciphertext he must decrypt from the labels of the inputs. This way, ciphertexts are ordered unrelated to the truth values of wires and any information leakage is prevented.}
\end{enumerate}
The number of ciphertexts per gate that needs to be transmitted is 4 in this method. 4 encryption and 1 decryption are the computational cost for each gate (see Table \ref{tab:scrbrdpp}).
\begin{table}[b]
\centering
\caption{Optimization Scoreboard (P\&P)}
\label{tab:scrbrdpp}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
\begin{figure}
\includegraphics[height=6.5cm, angle=0]{GRR3}
\caption{Garbled row reduction 3 ciphertexts.}
\label{fig:GRR3}
\end{figure}
\section{Garbled Row Reduction 3 Ciphertexts} \label{sec:GRR3}{}
Instead of choosing the masking values of the output of a gate randomly as in P\&P (\ref{sec:PP}), in \cite{NPS99} Naor \textit{et al.} suggested a smarter way, called \textit{garbled row reduction 3 ciphertexts} (GRR3).
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of the gate in Figure \ref{fig:GRR3}.}
\item{Alice choose the masking value of the first output in label order such that all bits of the resulting ciphertext is 0 (\textit{i.e.}, by decrypting all 0, $C_1\gets E_{A_0,B_1}^{-1}(0^n)$). The masking value reached will still be pseudo-random.}
\item{Since there is no need to send the first ciphertext, sending 3 ciphertexts per gate suffices.}
\end{enumerate}
Although GRR3 results in smaller-sized garbled circuits than the ones resulted from P\&P (\ref{sec:PP}), it has little affect on the computation cost since the gain coming from one less encryptions goes to the decryption of the first ciphertext (see Table \ref{tab:scrbrdgrr3}).
\begin{table}[b]
\centering
\caption{Optimization Scoreboard (GRR3)}
\label{tab:scrbrdgrr3}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P (\ref{sec:PP})} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{GRR3} & \small{3 ct / 3 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
\section{Free \texttt{XOR}} \label{sec:freeXOR}{}
One of the greatest jumps in the garbled circuit technology has been the free \texttt{XOR} technique, which is proposed by Kolesnikov and Schneider in \cite{KS08}. It basically eliminates the need for any ciphertext transmission and any calculation for \texttt{XOR} gates. The function can be compiled such that the number other gates are minimized. Usually they are just \texttt{AND} gates, since $(\texttt{XOR},\texttt{AND})$ is Turing complete.\footnote{The number of \texttt{AND} gates in the Boolean functions is called \textit{multiplicative complexity}. Reducing it at the expense of increasing \texttt{XOR}s is already an active research topic \cite{KK10}.}
\begin{figure}
\includegraphics[height=5.5cm, angle=0]{freexor}
\caption{(a) \texttt{XOR} gate with masked values on its wires. (b) \texttt{XOR} gate whose masked values interpreted with offsets. (c) \texttt{XOR} with the same offset in the masked values on wires. (d) \texttt{XOR} gate arranged for free \texttt{XOR} technique}
\label{fig:freexor}
\end{figure}
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of the \texttt{XOR} gate in Figure \ref{fig:freexor} (a).}
\item{The masking value for \texttt{TRUE} in a wire $a$ can be written as the one for \texttt{FALSE} in that wire $A$ \texttt{XOR}ed with some offset $\triangle_A$, which is a random value having the same number of bits as $A$ and $B$, as in Figure \ref{fig:freexor} (b). The masking value for \texttt{FALSE} becomes $A$, and the masking value for \texttt{TRUE} becomes $A\oplus\triangle_A$. Alice also writes the masking values of $b$ and $c$ the same way.}
\item{Alice sets the offsets of all wires be the same secret value $\triangle$ as in Figure \ref{fig:freexor} (c). Even if there are more than one gate in a circuit, all wires must be given the same offset so that the free \texttt{XOR} method can be applied. Offset must be kept as a secret by the garbler.}
\item{Alice choose the masking value for \texttt{FALSE} in the output, \texttt{XOR} of those for \texttt{FALSE} in the inputs as in Figure \ref{fig:freexor} (d). This makes transmitting any ciphertext for an \texttt{XOR} gate unnecessary.}
\end{enumerate}
\texttt{Evaluating:}
\begin{enumerate}
\setcounter{enumi}{4}
\item{Bob just \texttt{XOR}s the masking value of the inputs to calculate the masking value of the output. }
\end{enumerate}
\texttt{AND} gates can be encrypted as in GRR3 (\ref{sec:GRR3}), and 3 cipher texts needs to be transmitted (see Figure \ref{fig:freexorand}). Labels still exist, and ciphertexts must be ordered accordingly. The offset must be chosen such that for a given wire both masking values have different labels (\textit{e.g.}, its \texttt{lsb} must be 1 if the label is the last bit). Since the same offset is used in both inputs and the payload, there is a need for a circularity assumption for the encryption scheme used \cite{CKK12}.
\begin{figure}
\includegraphics[height=4.5cm, angle=0]{freexorand}
\caption{Encryptions of a gate other than \texttt{XOR} in the free \texttt{XOR} technique.}
\label{fig:freexorand}
\end{figure}
Free \texttt{XOR} technique, makes \texttt{XOR}s completely free for transmission and computation in both the garbler's side and the evaluator's side. This has a huge impact, not just for freeing \texttt{XOR}s but also permitting the minimization of the other gates at the expense of increasing \texttt{XOR}s (see Table \ref{tab:scrbrdfrxor}).
\begin{table}[b]
\centering
\caption{Optimization Scoreboard (Free \texttt{XOR})}
\label{tab:scrbrdfrxor}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P (\ref{sec:PP})} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{GRR3 (\ref{sec:GRR3}{})} & \small{3 ct / 3 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Free \texttt{XOR}} & \small{3 ct / free} & \small{4 \texttt{edt} / free} & \small{1 \texttt{edt} / free} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
\section{Garbled Row Reduction 2 Ciphertexts} \label{sec:GRR2}{}
Pinkas \textit{et al.} proposed a method called \textit{garbled row reduction 2 ciphertexts} (GRR2) in order to reduce the number of transferred ciphertexts in \citep{PSSW09}. GRR2 is based on Shamir's secret sharing (\ref{sec:secretshare}). It is especially good for reducing the size in case of abundant \texttt{AND} gates \citep{PSSW09}.
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of the odd gate in Figure \ref{fig:grr2} (a).}
\item{Alice calculates $K_{1}$, $K_{2}$, $K_{3}$, and $K_{4}$ by decrypting all 0 for all possible input combinations (\textit{e.g.}, $K_{1}\gets E_{A_0,B_0}^{-1}(0^n)$, $K_{2}\gets E_{A_0,B_1}^{-1}(0^n)$, $K_{3}\gets E_{A_1,B_0}^{-1}(0^n)$, $K_{4}\gets E_{A_1,B_1}^{-1}(0^n)$).}
\item{Using the rows which give the same output (in this case the rows 1, 3, 4) Alice plots a $2^{nd}$ degree polynomial $P(x)$ (\textit{e.g.}, the red parabolas in Figure \ref{fig:grr2} (b)).}
\item{Alice also plots another $2^{nd}$ degree polynomial $Q(x)$ from the excluded row (here the row 2), $P(5)$, and $P(6)$ (\textit{e.g.}, the blue parabolas in Figure \ref{fig:grr2} (b)).}
\end{enumerate}
\texttt{Evaluating:}
\begin{enumerate}
\setcounter{enumi}{4}
\item{Alice sends only the intersection points $P(5)$ and $P(6)$. Bob will get another point by decrypting all 0 with the masking values that he gets in the input. He will be able to reaching only one of the polynomials, not knowing which one. The output masking value will be the evaluation of this polynomial at $x=0$ (\textit{e.g.}, $C_{0}=P(0)$ and $C_{1}=Q(0)$).}
\end{enumerate}
\begin{figure}
\centering
\begin{subfigure}[b]{0.35\textwidth}
\centering
\includegraphics[width=\textwidth]{grr2a}
\caption{}
\label{fig:sub1}
\end{subfigure}%
\quad
\begin{subfigure}[b]{0.57\textwidth}
\centering
\includegraphics[width=\textwidth]{grr2b}
\caption{}
\label{fig:sub2}
\end{subfigure}
\caption{(a) The odd gate to be garbled. (b) Plots of two polynomials obtained from $K_{1}$, $K_{2}$, $K_{3}$, and $K_{4}$.}
\label{fig:grr2}
\end{figure}
The position in this scheme leaks information. Moreover, since the wire masking values are not chosen but calculated pseudo-random values, it is impossible to directly use the P\&P (\ref{sec:PP}) technique. Instead, Pinkas \textit{et al.} proposed adding a one bit \textit{external value} $c_{i}$ for each wire. External values, like labes, are different for the \texttt{TRUE} and \texttt{FALSE} truth values unrelated to the truth value. Just like labels, external values are used for ordering. To calculate the external value of the output of a gate, 4 additional $M_{r}$ bits are sent. The evaluator, then, just needs to \texttt{XOR} the first bits of both input masking values and the related $M_{r}$ bit to find out the output external value. Since he does not know the masking values for other truth values of the input wires, he cannot find out the external values for the other output.
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{For an even gate, Alice similarly calculates $K_{1}$, $K_{2}$, $K_{3}$, and $K_{4}$ as in the odd gate case, in order of the external values.}
\item{Somewhat differently from the previous procedure, she plots the two $1^{st}$ degree polynomials each passing through the two points which correspond to the same output value. For instance, if both $K_{1}$ and $K_{3}$ are for the rows corresponding to \texttt{TRUE}, she plots $P(x)$ passing through $(1,K_{1})$ and $(3,K_{3})$ and $Q(x)$ passing through $(2,K_{2})$ and $(4,K_{4})$. She sends $P(5)$ and $Q(5)$, along with the 4 additional $M_{r}$ bits. She makes sure that ordering $P(5)$ and $Q(5)$ is according to the external value of the output of the gate just like using them the same as label bits, so that the evaluator know which one to use.}
\end{enumerate}
\texttt{Evaluating:}
\begin{enumerate}
\setcounter{enumi}{2}
\item{The evaluator decrypts all 0 with the masking values of the inputs. With two points in hand he plots the $1^{st}$ degree polinomial evaluate it at $x=0$ and reaches the output masking value.}
\end{enumerate}
Referring to Shamir's secret sharing (\ref{sec:secretshare}), two $t$-length values and 4 $M_r$ bits ($2t+4$) are needed to be sent per gate. For the sake of simplicity, we can take it as 2 ciphertexts per gate (see Table \ref{tab:scrbrdfrgrr2}).
\begin{table}[]
\centering
\caption{Optimization Scoreboard (GRR2)}
\label{tab:scrbrdfrgrr2}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P (\ref{sec:PP})} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{GRR3 (\ref{sec:GRR3}{})} & \small{3 ct / 3 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Free \texttt{XOR} (\ref{sec:freeXOR})} & \small{3 ct / free} & \small{4 \texttt{edt} / free} & \small{1 \texttt{edt} / free} \\ \hline
\small{GRR2} & \small{2 ct / 2 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
Although GRR2 is good for reducing the sizes of odd gates, it has a major drawback: incompatibility with free \texttt{XOR} (\ref{sec:freeXOR}). This is because the output masking values of the gates garbled with the GRR2 technique are pseudo-random numbers which cannot be set to the same offset.
\section{Fle\texttt{XOR}}\label{sec:fleXOR}{}
The incompatibility of free \texttt{XOR} (\ref{sec:freeXOR}) and GRR2 (\ref{sec:GRR2}) causes an inconvenient situation where both may be better for different circuits depending on the proportion of \texttt{XOR} and \texttt{AND} gates. To solve this issue, Kolesnikov \textit{et al.} proposed the fle\texttt{XOR} technique in \cite{KMR14}. Fle\texttt{XOR} may reduce the number of ciphertexts for an \texttt{XOR} gate even if it has different offsets on its wires. With this technique, \texttt{XOR} gates requires 1 or less ciphertext most of the time. It may cost 2 ciphertexts, only if the output masked value of the XOR gate has different offset from its inputs. Actually, most of the time, the output masked value may be chosen such that it has the same offset at least one of the inputs
\begin{figure}
\includegraphics[height=6.5cm, angle=0]{flexor}
\caption{ (a) An \texttt{XOR} gate with different offsets in its inputs and output. (b) A buffer gate to carry the offset of a wire. (c) An \texttt{XOR} gate offsets of whose inputs are carried to the offset of its output by two imaginary buffer gates. (d) An \texttt{XOR} gate the offset of whose an input is carried to the offset of its output by an imaginary buffer gate.}
\label{fig:flexor}
\end{figure}
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of the \texttt{XOR} gate in Figure \ref{fig:flexor} (a).}
\item{The idea is that if it was possible to carry the input wires to the same offset level with the output wire, which is $\triangle_{C}$, the \texttt{XOR} gate would be free. Figure \ref{fig:flexor} (b) depicts an imaginary buffer gate which can be used to carry the offset of a wire. Alice encrypts the output masked values with their corresponding inputs as $E_{A}(A^*)$ and $E_{A\oplus\triangle_{1}}(A^*\oplus\triangle_{2})$. She order them by P\&P (\ref{sec:PP}), and since $A^*$ can be any random value, she can let the first one in order all 0. Therefore, sending just one cipher text for a buffer gate suffices.}
\item{Alice needs at most two imaginary buffer gates for an \texttt{XOR} gate to carry the inputs to the same offset level as the outputs (see Figure \ref{fig:flexor} (c)).}
\item{Most of the time, one imaginary buffer per \texttt{XOR} gate will be enough since Alice can let the offset of the output the same as one of the inputs (see Figure \ref{fig:flexor} (d)). And if the inputs and the output have the same offset, the \texttt{XOR} gate will be free.}
\end{enumerate}
fle\texttt{XOR} technique can be combined with GRR2 (\ref{sec:GRR2}) in order to reduce the number of ciphertexts for \texttt{AND} gates (see Table \ref{tab:scrbrdfrflxor}). The combined scheme proposed by Kolesnikov \textit{et al.} can be seen in Algorithm \ref{alg:combflexor}.
\begin{table}[t]
\centering
\caption{Optimization Scoreboard (Fle\texttt{XOR})}
\label{tab:scrbrdfrflxor}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P (\ref{sec:PP})} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{GRR3 (\ref{sec:GRR3}{})} & \small{3 ct / 3 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Free \texttt{XOR} (\ref{sec:freeXOR})} & \small{3 ct / free} & \small{4 \texttt{edt} / free} & \small{1 \texttt{edt} / free} \\ \hline
\small{GRR2 (\ref{sec:GRR2}{})} & \small{2 ct / 2 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Fle\texttt{XOR}} & \small{2 ct / \{0,1,2\} ct} & \small{4 \texttt{edt} / \{0,2,4\} \texttt{edt}} & \small{1 \texttt{edt} / \{0,1,2\} \texttt{edt}} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
\begin{algorithm}
\scriptsize
\caption{The combined fle\texttt{XOR} and GRR2 scheme proposed by Kolesnikov \textit{et al.} in \cite{KMR14}.}
\label{alg:combflexor}
\begin{algorithmic}[1]
\Procedure{\texttt{Gb}}{$1^k,f$} \Comment \texttt{Garbling phase}
\State $(n,m,q,A',B',G)\gets f$
\For{$i\in\{1,\dots,n\}$} $t\gets \{0,1\}$, $X_i^0\gets \{0,1\}^{k-1}t$, $X_i^1\gets \{0,1\}^{k-1}\overline{t}$
\EndFor
\For{$g\in\{n+1,\ldots,n+q\}$ in a \textbf{safety-respecting} order} $a \gets A'(g)$, $b \gets B'(g)$
\If {$g\in \texttt{XORGates}(f)$}
\If{$X_a^0\oplus X_a^1=X_b^0\oplus X_b^1$} $X_g^0\gets X_a^0\oplus X_b^0$, $X_g^1\gets X_a^0\oplus X_b^1$, $P[g]\gets \bot$
\Else
\If{$\texttt{C}_{X_a^0}=0$} $X_a^{\bar{0}}\gets H(X_a^0,g\parallel 00)$, $X_a^{\bar{1}}\gets X_a^{\bar{0}}\oplus X_b^0\oplus X_b^1$
\State $X_g^0\gets X_a^{\bar{0}}\oplus X_b^0$, $X_g^1\gets X_a^{\bar{0}}\oplus X_b^1$, $P[g]\gets H(X_a^1,g\parallel 00) \oplus X_a^ {\bar{1}}$
\Else \texttt{ } $X_a^{\bar{1}}\gets H(X_a^1,g\parallel 00)$, $X_a^{\bar{0}}\gets X_a^{\bar{1}}\oplus X_b^0\oplus X_b^1$
\State $X_g^0\gets X_a^{\bar{0}}\oplus X_b^0$, $X_g^1\gets X_a^{\bar{0}}\oplus X_b^1$, $P[g]\gets H(X_a^0,g\parallel 00) \oplus X_a^ {\bar{0}}$
\EndIf
\EndIf
\State $\texttt{C}_{X_g^0}\gets \texttt{C}_{X_a^0}\oplus \texttt{C}_{X_b^0}$, $\text{C}_{X_g^1}\gets \bar{\texttt{C}_{X_g^0}}$
\Else
\State $\text{C}_{X_g^0}\gets\{0,1\}$ $\texttt{C}_{X_g^1}\gets \bar{\texttt{C}_{X_g^0}}$
\For{$(i,j)\in \{0,1\}^2$} $V_{ij}\parallel m_{ij} \gets H(X_{ai},X_{bj},g\parallel i\parallel j)$, $c_{ij}\gets \texttt{C}_{X_g^{w_{ai}\wedge w_{bj}}}\oplus m_{ij}$
\EndFor
\State $Q \gets interp \{(2i+j,V_{ab})\texttt{ } | \texttt{ }w_{ai}\wedge w_{bj}=0\}$
\State $R \gets interp \{(2i+j,V_{ab}\text{ } |\texttt{ }w_{ai}\wedge w_{bj}=1),(4,Q(4)),(5,Q(5))\}$
\State $X_g^0\gets Q(-1)$, $X_g^1\gets R(-1)$, $P[g]\gets (Q(4),Q(5),c_{00},c_{01},c_{10},c_{11})$
\EndIf
\EndFor
\State $F\gets(n,m,q,A',B',P)$, $e\gets(X_1^0,X_1^1,\ldots,X_n^0,X_n^1)$, $d\gets(X_{n+q-m+1}^0,X_{n+q-m+1}^1,\ldots,X_{n+q}^0,X_{n+q}^1)$
\State \textbf{return} $(F,e,d)$
\EndProcedure
\Statex
\Procedure{\texttt{En}}{$e,x$} \Comment \texttt{Encoding phase}
\State $(X_1^0,X_1^1,\ldots,X_n^0,X_n^1)\gets e$, $x_1\ldots x_n\gets x$, $X\gets (X_1^{x_1},\ldots,X_n^{x_n})$, \textbf{return} $X$
\EndProcedure
\Statex
\Procedure{\texttt{Ev}}{$F,X$} \Comment \texttt{Evaluating phase}
\State $(n,m,q,A',B',P)\gets F$, $(X_1,\ldots,X_n)\gets X$
\For{$g\gets n+1 \textbf{ to } n+q$} $a \gets A'(g)$, $b \gets B'(g)$
\If {$g\in \texttt{XORGates}(f)$}
\If{$P[g]\gets \bot$} $X_g\gets X_a\oplus X_b$
\Else
\If{$\texttt{C}_{X_a}=0$} $X_a^{\bar{}}\gets H(X_a,g\parallel 00)$
\Else \texttt{ } $X_a^{\bar{}}\gets P[g]\oplus H(X_a,g\parallel 00)$
\EndIf
\EndIf
\State $\texttt{C}_{X_g}\gets \texttt{C}_{X_a}\oplus \texttt{C}_{X_b}$
\Else
\State$V*\parallel m* \gets H(X_{a},X_{b},g\parallel {C}_{X_a}\parallel {C}_{X_b})$
\State $R* \gets interp \{(2{C}_{X_a}+{C}_{X_b},V*),(4,Q(4)),(5,Q(5))\}$
\State $X_g\gets R*(-1)$, $\texttt{C}_{X_g}\gets c_{\texttt{C}_{X_a}\texttt{C}_{X_b}}\oplus m*$
\EndIf
\EndFor
\State \textbf{return} $(X_{n+q-m+1},\ldots,X_{n+q})$
\EndProcedure
\Statex
\Procedure{\texttt{De}}{$d,Y$} \Comment \texttt{Decoding phase}
\State $(Y_1,\ldots,Y_m)\gets Y$, $(Y_1^0,Y_1^1,\ldots,Y_m^0,Y_m^1)\gets d$
\For{$i\in\{1,\ldots,m\}$}
\If{$Y_i=Y_i^0$} $y_i\gets 0$
\ElsIf{$Y_i=Y_i^1$} $y_i\gets 1$
\Else \textbf{ return} $\bot$
\EndIf
\EndFor
\State \textbf{return} $y\gets y_1\ldots y_m$
\EndProcedure
\end{algorithmic}
\end{algorithm}
The notation used in Algorithm \ref{alg:combflexor} is similar to the one in Algorithm \ref{alg:yaosgcpproc}. $\texttt{XORGates}(f)$ denotes the set of $\texttt{XOR}$ gates in $f$. $\texttt{C}_{X_i}$ denotes the external value of the wire whose masking value is $X_i$. $V_{ij}$ denotes the value used in the interpolation related to the order $ij$. $m_{ij}$ denotes the one bit value used to mask the external value. $X_{ai}$ denotes the masking value on the wire $a$, $i$ being the external value. $w_{ai}$ denotes the truth value on the wire $a$, $i$ being the external value. $c_{ij}$ denotes the bits sent for the calculation of the external value of the output of a gate, $ij$ being the order coming from the input external values \cite{KMR14}.
\section{Half Gates} \label{sec:halfgates}{}
The \textit{half gates} method, which is proposed by Zahur \textit{et al.} in \cite{ZRE15}, proves that sending 2 ciphertexts can be enough for an \texttt{AND} gate while \texttt{XOR} gates are still free. The same offset is kept throughout the whole circuit wires, like the free \texttt{XOR} (\ref{sec:freeXOR}). It is based upon the idea that if one of the sides knows the truth value on an input wire of an \texttt{AND} gate, it is enough to send just one ciphertext. The method divides the \texttt{AND} gate into two \texttt{AND} gates where one of the parties knows the truth value on an input wire. The name of the method comes from this division.
$A$, $B$, $C$, $C_1$, and $C_2$ are the masking values for the wires $a$, $b$, $c$ (output of the \texttt{AND} gate), $c_1$ (output of the garbler half gate), and $c_2$ (output of the evaluator half gate), respectively. $\triangle$ denotes the common offset as in free \texttt{XOR} (\ref{sec:freeXOR}).
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Alice and Bob want to compute the output of an \texttt{AND} gate whose inputs are $a$ and $b$.}
\item{An \texttt{AND} gate can be written as an \texttt{XOR} of two \texttt{AND} gates as in Equation~\eqref{eq:andgatedivided} where $r$ is a randomly chosen bit only known to Alice. Alice chooses it to be the label bit of the $B$, which is the masked value for \texttt{FALSE} on the wire $b$. $r$ is still unknown to Bob.}
\end{enumerate}
\begin{equation}
\label{eq:andgatedivided}
a\wedge b=(a\wedge r)\oplus [a\wedge(b\oplus r)]
\end{equation}
\texttt{Garbler Half Gate:}
\begin{enumerate}
\setcounter{enumi}{2}
\item{$a\wedge r$ is the garbler half gate, whereas $a\wedge(b\oplus r)$ is the evaluator half gate. For the output of the garbler half gate $c_{1}\gets a\wedge r$, Alice needs to send $E_{B}(C_{1})$ and $E_{B\oplus\triangle} (C_{1}\oplus r\triangle)$. Since she knows the value of $r$, there is just 2 input combinations. She orders the ciphertexts with respect to the label bit of $b$. Row reduction (\ref{sec:GRR3}) is also possible by letting the $1^{st}$ ciphertext in all 0. She calculates the $2^{nd}$ ciphertext from the value she reaches by decrypting the first one. Thus, sending just 1 ciphertext is enough for the garbler half gate.}
\item{During the evaluation of the garbler half gate, Bob decrypts the related cipher text depending on the label bit of the masking value on the wire $b$. Since the order is by labels he can not learn the truth value of $b$.}
\end{enumerate}
\texttt{Evaluator Half Gate:}
\begin{enumerate}
\setcounter{enumi}{4}
\item{For the evaluator half gate, Alice needs to let Bob learn $q=b\oplus r$ without learning $b$ or $r$. Actually, it is whatever Bob gets as the label bit of the masked value on wire $b$. This was the main reason why $r$ was chosen as the label bit of $B$ in the beginning.}
\item{To garble the evaluator half gate $c_{2}\gets a\wedge q$, there are two ways Alice may go depending on the value of $r$. If $r$ is \texttt{FALSE}, Alice sends two ciphertexts $E_{B}(C_{2})$ and $E_{B\oplus\triangle} (C_{2}\oplus A)$ in this order strictly. Otherwise, Alice sends two ciphertexts $E_{B\oplus\triangle} (C_{2})$ and $E_{B}(C_{2}\oplus A)$ in this order strictly. Moreover, the $1^{st}$ ciphertext can be let all 0 and the $2^{nd}$ one can be calculated from it. Therefore, sending only one ciphertext for the evaluator half gate also suffices.}
\item{If Bob gets \texttt{FALSE} as $q$, he decrypts the first ciphertext using the masking value on the wire $b$, arriving at the masking value of the output of the evaluator half gate. Otherwise, he decrypts the second ciphertext using the value on the wire $b$, and \texttt{XOR}s the result with the masking value on the wire $a$, arriving at the masking value of the output of the half gate.}
\end{enumerate}
The evaluator does not learn the truth values of $a$, $b$, $r$, or $c_{2}$ (if $q= 1$, of course, otherwise he learns $c_{2}$). In the end, the results of the half gates must be \texttt{XOR}ed, in order to obtain the final output of the \texttt{AND} gate.
With the half gates technique, an \texttt{AND} gate costs 2 cipher texts and \texttt{XOR}s are free, which makes the half gates technique the optimum from size point of view among the methods developed so far (see Table \ref{tab:scrbrdhg}). Zahur \textit{et al.} have also proven that decreasing the size of an \texttt{AND} gate further is impossible.
\begin{table}[b]
\centering
\caption{Optimization Scoreboard (Half Gates)}
\label{tab:scrbrdhg}
\begin{tabular}{l|l|l|l}
\small{\textbf{Method}} & \small{\textbf{Odd / Even Gate Size}} & \small{\textbf{Enc. Time per}} & \small{\textbf{Dec. Time per}} \\
\small{ } & \small{ } & \small{\textbf{Odd / Even Gate}} & \small{\textbf{Odd / Even Gate}} \\ \hline
\small{P\&P (\ref{sec:PP})} & \small{4 ct / 4 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{GRR3 (\ref{sec:GRR3}{})} & \small{3 ct / 3 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Free \texttt{XOR} (\ref{sec:freeXOR})} & \small{3 ct / free} & \small{4 \texttt{edt} / free} & \small{1 \texttt{edt} / free} \\ \hline
\small{GRR2 (\ref{sec:GRR2}{})} & \small{2 ct / 2 ct} & \small{4 \texttt{edt} / 4 \texttt{edt}} & \small{1 \texttt{edt} / 1 \texttt{edt}} \\ \hline
\small{Fle\texttt{XOR} (\ref{sec:fleXOR}{})} & \small{2 ct / \{0,1,2\} ct} & \small{4 \texttt{edt} / \{0,2,4\} \texttt{edt}} & \small{1 \texttt{edt} / \{0,1,2\} \texttt{edt}} \\ \hline
\small{Half Gates} & \small{2 ct / free} & \small{4 \texttt{edt} / free} & \small{2 \texttt{edt} / free} \\
\multicolumn{4}{c}{ }\\
\multicolumn{4}{c}{{\scriptsize ct: ciphertexts; \texttt{edt}: total encryption and/or decryption time}}\\
\end{tabular}
\end{table}
\textbf{The Complete Scheme.} For a boolean circuit $f$, a numeric index is assigned to each wire in the circuit. The sets of input wires, output wires, output wires of \texttt{XOR} gates in $f$$ are denoted as \texttt{Inputs}(f)$, $\texttt{Outputs}(f)$, and $\texttt{XORGates}(f)$, respectively. These functions can also be applied to garbled version $F$ of $f$ as $\texttt{Inputs}(F)$, $\texttt{Outputs}(F)$, and $\texttt{XORGates}(F)$. $v_{i}$ denotes the one bit truth value on the $i^{th}$ wire in a circuit. If the output wire of a gate has index $i$, that gate is named as $i^{th}$ gate. The wire masking values for \texttt{FALSE} and \texttt{TRUE} on the $i^{th}$ wire is denoted as $W_{i}^0, W_{i}^1 \in \{0, 1\}^k$, respectively. The security parameter of the scheme is denoted as $k$. For each wire masking value $W$, the label bit is its least significant bit $\texttt{lsb}W$. For the $i^{th}$ wire, define $p_{i}= \texttt{lsb}W_{i}^0$. Being named as the permute bit of the wire, that value is a secret kept by the generator. Intuitively, if label bit a masking value on a wire is $s_{i}$, that masking value is $W _{i}^{s_{i}\oplus p_{i}} $, and corresponds to the truth value $s_{i}\oplus p_{i}$. $W_{i}$ implies that the evaluator does not know $v_{i}$. The free \texttt{XOR} offset is denoted as $R \in \{0, 1\}^k$. We have $\texttt{lsb}R = 1$ so that $\texttt{lsb}W_{i}^0\neq\texttt{lsb}W_{i}^1$, and the complementary masking values on wires have different label bits. Sometimes $\wedge$ is omitted and two symbols is juxtaposed to imply \texttt{AND} ($ab = a\wedge b$). $H:\{0,1\}^k\times \mathbb{Z} \rightarrow \{0,1\}^k$ denotes a hash-function that is usable in garbled circuits.
\begin{equation}
\label{eq:oddgategeneralization}
(v\textsubscript{a},v\textsubscript{b}) \rightarrow(a\textsubscript{a} \oplus v\textsubscript{a})\wedge(a\textsubscript{b} \oplus v\textsubscript{b})\oplus a\textsubscript{c}
\end{equation}
The technique can be further generalized such that it can be applied any odd gate (\texttt{OR}, \texttt{NOR}, \texttt{NAND}, etc.), since all of them can be written as in Equation~\eqref{eq:oddgategeneralization} where $a\textsubscript{a}$, $a\textsubscript{b}$, $a\textsubscript{c}$ are constants. For example, an \texttt{AND} gate results from setting all to \texttt{FALSE}, an \texttt{OR} gate results from setting all to \texttt{TRUE}. The construction of half gate is shown step-by-step in Table \ref{tab:halfgates}. Note that the $a$ values does not affect what the evaluator does.
\vspace{2mm}
\begin {table}[H]
\caption {The construction of half gates for computing Equation~\eqref{eq:oddgategeneralization} \cite{ZRE15}.} \label{tab:halfgates}
\scriptsize
\begin{center}
\begin{tabular}{ >{\arraybackslash}m{2.7in} | >{\arraybackslash}m{2.7in}}
\toprule[1.5pt]
{\bf Generator half gate: $p_b$ known to generator} & {\bf Evaluator half gate: $v_b\oplus p_b$ known to evaluator} \\
\underline{Computes:}\newline $f_G(v_a,p_b)\gets(v_a\oplus a_a)(p_b\oplus a_b) \oplus a_c$ & \underline{Computes:}\newline $f_E(v_a,v_b\oplus p_b)\gets(v_a\oplus a_a)(v_b\oplus p_b)$ \\
\underline{Before GRR and Permutation:}\newline $H(W_a^0)\oplus f_G(0,p_b)R\oplus W_{Gc}^0$\newline $H(W_a^1)\oplus f_G(1,p_b)R\oplus W_{Gc}^0$ & \underline{Before GRR:}\newline $H(W_b^{p_b})\oplus W_{Ec}^0$\newline $H(W_b^{p_b\oplus 1})\oplus W_{Ec}^0\oplus W_a^{a_a}$ \\
\underline{After GRR and permutation:}\newline $T_{Gc}\gets H(W_a^0)\oplus H(W_a^1) \oplus (p_b\oplus a_b)R$\newline $W_{Gc}^0\gets H(W_a^{p_a})\oplus f_G(p_a,p_b)R$& \underline{After GRR (permutation not needed):}\newline $T_{Ec}\gets H(W_b^0)\oplus H(W_b^1) \oplus W_a^{a_a}$\newline $W_{Ec}^0\gets H(W_b^{p_b})$ \\
\underline{Generator sends $T_{Gc}$}&\underline{Generator sends $T_{Ec}$ } \\
\bottomrule[1.25pt]
\end {tabular}
\end{center}
\end {table}
The complete garbling procedure for an entire circuit proposed by Zahur \textit{et al.} is shown in Algorithm \ref{alg:comphalfgates} \cite{ZRE15}. All gates are assumed to be either an \texttt{AND} or an \texttt{XOR} gate. Since \textsc{\texttt{De}} never returns $\bot$, this scheme does not satisfy the authenticity criterion. In order to make it authentic, Zahur \textit{et al.} propose the following changes:
\begin{itemize}
\item{The \textbf{for}-loop on Line \ref{lst:labelarrange} of Algorithm \ref{alg:comphalfgates} must be changed as:}
\end{itemize}
\begin{algorithm}
\scriptsize
\label{alg:dehalfgates}
\begin{algorithmic
\For{$i\in \texttt{Outputs}(f)$}
\State $j\gets \texttt{NextIndex}()$
\State $d_i\gets (H(W_i^0,j),H(W_i^1,j))$
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{itemize}
\item{The \textbf{for}-loop on Line \ref{lst:decode} of Algorithm \ref{alg:comphalfgates} must be changed as:}
\end{itemize}
\begin{algorithm}[H]
\scriptsize
\label{alg:dehalf-gates}
\begin{algorithmic
\For{$d_i\in d$} $j\gets \texttt{NextIndex}()$, parse $(h_0,h_1)\gets d_i$
\If{$H(Y_i,j)=h_0$} $y_i\gets 0$
\ElsIf{$H(Y_i,j)=h_1$} $y_i\gets 1$
\Else \textbf{ return} $\bot$
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\scriptsize
\caption{The complete half gates garbling scheme proposed by Zahur \textit{et al.} in \cite{ZRE15}.}
\label{alg:comphalfgates}
\begin{algorithmic}[1]
\Procedure{\texttt{Gb}}{$1^k,f$} \Comment \texttt{Garbling phase}
\State $R\twoheadleftarrow \{0,1\}^{k-1}1$
\For{$i\in \texttt{Inputs}(f)$}
\State $W_i^0\twoheadleftarrow \{0,1\}^k$, $W_i^1\gets W_i^0\oplus R$, $e_i\gets W_i^0$
\EndFor
\For{$i\notin \texttt{Inputs}(f)$} \{\textit{in topo. order}\}
\State $\{a,b\} \gets \texttt{GateInputs}(f,i)$
\If{$i\in \texttt{XORGates}(f)$} $W_i^0\gets W_a^0\oplus W_b^0$
\Else \texttt{ }$(W_i^0,T_{Gi},T_{Ei})\gets \textsc{\texttt{GbAnd}}(W_a^0,W_b^0)$, $F_i\gets T_{Gi},T_{Ei}$
\EndIf
\State $W_i^1\gets W_i^0\oplus R$
\EndFor
\For{$i\in \texttt{Outputs}(f)$} \label{lst:labelarrange}
\State $d_i\gets \texttt{lsb} (W_i^0)$
\EndFor
\State \textbf{return} $(F,e,d)$
\EndProcedure
\Statex
\algrenewcommand\algorithmicprocedure{\textbf{private procedure}}
\Procedure{\texttt{GbAnd}}{$W_a^0,W_b^0$} \Comment \texttt{Garbling AND gates}
\State $p_a\gets \texttt{lsb} (W_a^0)$, $p_b\gets \texttt{lsb} (W_b^0)$
\State $j\gets \texttt{NextIndex}()$, $j'\gets \texttt{NextIndex}()$
\State \{\textit{First half gate}\}
\State $T_G\gets H(W_a^0,j)\oplus H(W_a^1,j)\oplus p_bR$
\State $W_G^0\gets H(W_a^0,j)\oplus p_aT_G$
\State \{\textit{Second half gate}\}
\State $T_E\gets H(W_b^0,j')\oplus H(W_b^1,j')\oplus W_a^0$
\State $W_E^0\gets H(W_b^0,j')\oplus p_b(T_E\oplus W_a^0)$
\State \{\textit{Combine two halves}\}
\State $W_0\gets W_G^0\oplus W_E^0$
\State \textbf{return} $(W^0,T_G,T_E)$
\EndProcedure
\Statex
\algrenewcommand\algorithmicprocedure{\textbf{procedure}}
\Procedure{\texttt{En}}{$e,x$} \Comment \texttt{Encoding phase}
\For{$e_i\in e$} $X_i \gets e_i\oplus x_iR$
\EndFor
\State \textbf{return} $X$
\EndProcedure
\Statex
\Procedure{\texttt{Ev}}{$F,X$} \Comment \texttt{Evaluating phase}
\For{$i\in \texttt{Inputs}(F)$}
\State $W_i\gets X_i$
\EndFor
\For{$i\notin \texttt{Inputs}(F)$} \{\textit{in topo. order}\}
\State $\{a,b\} \gets \text{GateInputs}(F,i)$
\If{$i\in \texttt{XORGates}(F)$} $W_i\gets W_a\oplus W_b$
\Else \text{ } $s_a\gets \texttt{lsb} (W_a)$, $s_b\gets \texttt{lsb} (W_b)$, $j\gets \texttt{NextIndex}()$, $j'\gets \texttt{NextIndex}()$
\State $T_{Gi},T_{Ei}\gets F_i$, $W_{Gi}\gets H(W_a,j)\oplus s_aT_{Gi}$, $W_{Ei}\gets H(W_b,j')\oplus s_b(T_{Ei}\oplus W_a)$
\State $W_i\gets W_{Gi}\oplus W_{Ei}$
\EndIf
\State $W_i^1\gets W_i^0\oplus R$
\EndFor
\For{$i\in \texttt{Outputs}(F)$} $Y_i\gets W_i$
\EndFor
\State \textbf{return} $Y$
\EndProcedure
\Statex
\Procedure{\texttt{De}}{$d,Y$} \Comment \texttt{Decoding phase}
\For{$d_i\in d$} $y_i \gets d_i\oplus \text{lsb}Y_i$ \label{lst:decode}
\EndFor
\State \textbf{return} $y$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\section{Our Compatibility Analysis of Garbled Circuit Optimizations}
We conclude this chapter with a useful table which reflects the compatibility of garbled circuit optimizations with each other (see Table \ref{tab:gcoptcompatibility}). \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \normalsize and $\mathsf{X}$ stand for compatible and non-compatible, respectively. For the use of external value (Ext. Val.), see Section \ref{sec:GRR2}{}.
\vspace{7mm}
\begin{table}[H]
\centering
\caption{Compatibility of Garbled Circuit Optimization Techniques.}
\label{tab:gcoptcompatibility}
\begin{tabular}{c|c|c|c|c|c|c}
\cellcolor{black!50} & \footnotesize{\textbf{P\&P}} & \footnotesize{\textbf{GRR3}} & \footnotesize{\textbf{Free \texttt{XOR}}} & \footnotesize{\textbf{GRR2}} & \footnotesize{\textbf{Fle\texttt{XOR}}} & \footnotesize{\textbf{Half Gates}} \\ \hline
\footnotesize{\textbf{P\&P}} &\cellcolor{black!50} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle; (Ext. Val.)} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle; (Ext. Val.)} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \\ \hline
\footnotesize{\textbf{GRR3}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \cellcolor{black!50} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \\ \hline
\footnotesize{\textbf{Free \texttt{XOR}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \cellcolor{black!50} & $\mathsf{X}$ & $\mathsf{X}$ & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \\ \hline
\footnotesize{\textbf{GRR2}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle; (Ext. Val.)} & {$\mathsf{X}$} & $\mathsf{X}$ & \cellcolor{black!50} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ \\ \hline
\footnotesize{\textbf{Fle\texttt{XOR}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle; (Ext. Val.)} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \cellcolor{black!50} & $\mathsf{X}$ \\ \hline
\footnotesize{\textbf{Half Gates}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ &$\mathsf{X}$ & \cellcolor{black!50} \\
\multicolumn{7}{c}{ }\\
\multicolumn{7}{c}{{\scriptsize Ext. Val.: External Value (\ref{sec:GRR2})}}\\
\end{tabular}
\end{table}
\chapter{Practical Implementations of Yao's Protocol} \label{chap:implementations}
\lhead{Chapter \ref{chap:implementations}. \emph{Practical Implementations of Yao's Protocol}}
Various implementations have been developed so far based on Yao's protocol. Many of them utilize Yao's protocol for MPC applications, although some targets \textit{Private Function Evaluation} (PFE). A comprehensive catalogue of them would have been far from the reach of just a master's thesis work. So, we will explain only some of them which are supposed to be helpful for people to see Yao's protocol in practise. They also reflect the importance of Yao's protocol and the areas it can be applied in the future. First, we will start with introducing some of the generic MPC solutions that use Yao's protocol. We compare those generic implementations in terms of their use of garbled circuit optimizations. At the end, we will present some real-world applications.
\section{Generic Usage of Yao's Protocol in Practice} \label{sec:GenericUsageofYao'sProtocolinPractice}{}
\subsection{Pipelined Implementation (\texttt{FastGC})}\label{sub:PipelinedImplementationFastGC}{}
The memory required to store the entire garbled circuit is generally a limitation. Huang \textit{et al.} proposed pipelining optimization in their framework in \cite{HEKM11} to reduce the required memory. The garbled circuit generation and evaluation procedures can be done simultaneously, eliminating the need for keeping the entire garbled circuit in memory and the need for preparation of the entire garbled circuit before its transmission to the evaluator, which results in a decrease in total Yao's protocol time. \texttt{FastGC} framework automates pipelined implementation, so that the only need remaining is the construction of the desired circuit \cite{HEKM11}.
At the beginning of the computation the circuit structure is instantiated by both the garbler and the evaluator. While the protocol is being executed, the generator garbles each gate in topological\footnote{Safety-respecting if the garbling method is fle\texttt{XOR}.} order, and transmits it over the network as soon as it is produced. When a garbled gate is received by the evaluator, it is associated with the corresponding gate of the circuit and evaluated. A gate is eliminated as soon as it has been evaluated, so that the memory use would be minimal. This technique is called \textit{pipelined implementation}. Note that it also reduces total Yao's protocol time of at the expense that both parties needs to be online at the same time.
\subsection{Garbled RAM} \label{sub:GarbledRAM}{}
The notion of \textit{garbled RAM} was introduced by Lu and Ostrovsky in \cite{LO13}. Gentry \textit{et al.} have later improved it using \textit{identity-based encryption} (IBE)\footnote{Identity based encryption (IBE) is a form of public key encryption (\ref{sub:publickeycryptography}) where a user's public key is his identity. In generic public key cryptosystems, private keys are chosen randomly and public keys are produced from them. However, in IBE the private keys are generated from users' public keys \cite{BF01}. \label{ftnt:ibe}} in \cite{GHRW14} for provable security. It differs from Yao's garbled circuits in that it permits direct garbling of a RAM program, \textit{without converting it into a boolean circuit}. A RAM program whose run-time is $T$ can be converted into a Turing Machine whose run-time is $O(T^3)$ resulting in a boolean circuit of size $O(T^3\texttt{log}T)$, whereas the size and computation time of a garbled RAM program is only proportional to its running time on a RAM \cite{GHRW14}. The inefficiency is even more prominent in the setting of \textit{big data} \cite{GHRW14}. In this case, efficient programs, such as binary search, run in sub-linear time with the size of the data, however their boolean circuit representations run in linear time with the size of the data.
Just like garbled circuits, garbled RAM includes a garbler who garbles the program, and sends it to the evaluator. Evaluator evaluates the garbled program using the garbled inputs and, unlike the case of Yao's protocol, outputs the actual output of the RAM program. Like the garbled circuits, garbled RAM targets security againist semi-honest adversaries ((\ref{sub:semi-honest}). Gentry \textit{et al.}'s scheme of garbled RAM is explained in detail below \cite{GHRW14}.
The notation $P^D (x)$ denotes a RAM program $P$ which accesses a memory containing data $D$ and takes an input $x$. Imagining $D$ as a \textit{huge database} controlled by the evaluator and $P$ as a \textit{database query} that has read or write access to the database and whose parameter is a value $x$ (like $P$ searches $x$ in $D$) would help for understanding the notions.
A garbled RAM scheme can be used to garble $P$, $D$, $x$ into $\bar{P}$, $\bar{D}$, $\bar{x}$, such that $\bar{P}$, $\bar{D}$, $\bar{x}$ reveals only $P^D(x)$. Furthermore, the sizes of $\bar{P}$, $\bar{D}$, $\bar{x}$ are only proportional to their corresponding plain texts. Similar to Yao's the garbled circuits, garbling $x$ consists of providing a subset of masking values.
A RAM program $P$ can be represented as a colleciton of \texttt{CPU}-Step Circuits which execute a single \texttt{CPU} step. Equation~\eqref{eq:cpustepj} shows the execution of \texttt{CPU} step $j$. The input to the circuit $C_\texttt{CPU}^P$ is the current \texttt{CPU} $\texttt{state}_j$ and a bit $b_j^\texttt{read}$ which resides in the memory location assigned in the previous cycle. Its outputs are an updated $\texttt{state}_{j+1}$, the next reading location $i_(j+1)^\texttt{read}$, a location $i_{j+1}^\texttt{write}$ for writing to (maybe $\bot$), a bit $b_{j+1}^\texttt{write}$ to write into that location. The start of the computation $P^D(x)$ is in the initial state $\texttt{state}_1=x$ and $b_0^\texttt{read}=0$, and it proceeds step-by-step. In each step $j$, first $b_j^\texttt{read}$ is set to $D[i_j^\texttt{read}]$, and if $i_j^\texttt{write}\neq\bot$, $D[i_j^\texttt{write}]$ is set to $b_j^\texttt{write}$. The output of the last \texttt{CPU} step is the output of the computation $y=P^D(x)$ as \texttt{state}. $P$ has \textit{read-only} memory access if it never overwrites any values in memory $D$ (\textit{i.e.}, $i_j^\texttt{write}$ is always $\bot$).
\begin{equation}
\label{eq:cpustepj}
C_\texttt{CPU}^P (\texttt{state}_j,b_j^\texttt{read},i_j^\texttt{write},b_j^\texttt{write}) = (\texttt{state}_{j+1},i_{j+1}^\texttt{read},i_{j+1}^\texttt{write},b_{j+1}^\texttt{write})
\end{equation}
Gentry \textit{et al.} propose their scheme with security against unprotected memory access (UMA) in which the initial contents of the memory $D$ and the complete memory access pattern of \texttt{MemAccess} (including the contents) may be learned by the intruder, \cite{GHRW14}. They also propose that encrypting the memory contents and applying oblivious RAM is enough for transforming any garbled RAM scheme with UMA security into one providing full security.
\textbf{Read-only Solution.} \label{subsub:ReadonlySolution}{}The garbled memory is made of $\bar{D}[i]$'s, each containing an IBE\textsuperscript{\ref{ftnt:ibe}} secret key $sk_{(i,b)}$ for the public key $(i,b)$ where $i$ is the location and $b$ is the data bit $D[i]$. Another future of $\bar{D}[i]$ is that it can remain and be used by the future programs. The garbled input $\bar{x}_j$ to the \texttt{CPU} step $j$ is the masking value for the $\texttt{state}_j$, and $\bar{x}_0$ is the masking input $\bar{x}$. The \texttt{CPU} step in Equation~\eqref{eq:cpustepj} simply becomes the one in Equation~\eqref{eq:cpustepjro}.
\begin{equation}
\label{eq:cpustepjro}
C_\texttt{CPU}^P (\texttt{state}_j,b_j^\texttt{read}) = (\texttt{state}_{j+1},i_{j+1}^\texttt{read})
\end{equation}
\eqref{eq:garbledcpucircuitj} shows the garbled circuit $\bar{C}_\texttt{CPU,j}^P$ of the step $j$. The problem with garbling the \texttt{CPU} step $j$ is that the location of $b_j^\texttt{read}$ is not pre-known since it is the output of the previous cycle. Let $\bar{b}_{0j}^\texttt{read}$ denote the masking value of $b_j^\texttt{read}$ for \texttt{FALSE} and $\bar{b}_{1j}^\texttt{read}$ denote the masking value of $b_j^\texttt{read}$ for \texttt{TRUE}. Each garbled step $j$ outputs a translation mapping $\texttt{translate}_{j+1}= (ct_{0(j+1)}, ct_{1(j+1)})$ where $ct_{b(j+1)}=E((i_{j+1}^\texttt{read},b),r_{bj},\bar{b}_{b(j+1)}^\texttt{read})$ calculated\footnote{$r_{bj}$ is the randomization value to provide semantic security.} by using IBE so that the evaluator can only learn the masking value of $D[i_{j+1}^\texttt{read}]$ using the key $\bar{D}[i_{j+1}^\texttt{read}]$. $\bar{b}_{0(j+1)}^\texttt{read}$, $\bar{b}_{1(j+1)}^\texttt{read}$, $r_{0j}$ and $r_{1j}$ are hardcoded in the step circuit $j$ and cannot be learned directly by the evaluator due to the garbling process. $i_{j+1}^\texttt{read}$ is not private since the target is UMA security, and so it does not require a masking value.
\begin{equation}
\label{eq:garbledcpucircuitj}
(\bar{x}_{j+1},i_{j+1}^\texttt{read},\texttt{translate}_{j+1}) \gets \bar{C}_\texttt{CPU,j}^P(\bar{x}_j,\bar{b}_{bj}^\texttt{read})
\end{equation}
Each garbled cycle $j$ starts with the decryption of $ct_{bj}$ (the evaluator may know which one to decrypt due to UMA security) to get $\bar{b}_{bj}^\texttt{read}$, except for the first cycle where $\bar{b}_{bj}^\texttt{read}=\bot$. The last cycle directly outputs $y = P^D(x)$.
\textbf{Writing to the Memory.} \label{subsub:WritingtotheMemory}{}Similar to the read-only case, the garbled memory is made of $\bar{D}[i]$'s, each containing an \textit{timed IBE} secret key $sk_{(u,i,b)}$ for the public key $(u,i,b)$ where $u$ is the cycle that $i$ is written last time. The full step given in Equation~\eqref{eq:cpustepj} needs to be evaluated. Equation~\eqref{eq:garbledcpuwritecircuitj} shows the garbled circuit $\bar{C}_\texttt{CPU,j}^P$ of the step $j$. Unlike the read-only case, each step $j$ writes $sk_{(j,i,b)}$ to the garbled memory address $i_j^\texttt{write}$ (if they are not $\bot$), and outputs $sk_{(j+1,i,b)}$ and $i_{j+1}^\texttt{write}$ for writing in the next cycle. Each garbled step $j$ outputs a translation mapping $\texttt{translate}_{j+1}= (ct_{0(j+1)}, ct_{1(j+1)})$ where $ct_{b(j+1)}=E((u_{j+1},i_{j+1}^\texttt{read},b),r_{bj},\bar{b}_{b(j+1)}^\texttt{read})$ calculated by using timed IBE. Here, the assumption is that there exists a polynomial size circuit \texttt{WriteTime} such that $u_{j+1} = \texttt{WriteTime}(j,\bar{x}_j,i_{j+1}^\texttt{read})$, and step $j$ can call it. Just like the read-only case, the evaluator can only learn the masking value of $D[i_{j+1}^\texttt{read}]$ using the key $\bar{D}[i_{j+1}^\texttt{read}]$. $\bar{b}_{0(j+1)}^\texttt{read}$, $\bar{b}_{1(j+1)}^\texttt{read}$, $r_{0j}$ and $r_{1j}$ are hardcoded in the step circuit $j$ and cannot be learned directly by the evaluator due to the garbling process. $i_{j+1}^\texttt{read}$ and $i_{j+1}^\texttt{write}$ are not private since the target is UMA security.
\begin{equation}
\label{eq:garbledcpuwritecircuitj}
(\bar{x}_{j+1},i_{j+1}^\texttt{read},\texttt{translate}_{j+1},i_{j+1}^\texttt{write},sk_{(j+1,i,b)}) \gets \bar{C}_\texttt{CPU,j}^P(\bar{x}_j,\bar{b}_{bj}^\texttt{read},i_j^\texttt{write},sk_{(j,i,b)})
\end{equation}
\textbf{Full Security.} \label{subsub:FullSecurity}{}Gentry \textit{et al.} propose that any garbled RAM scheme that only provides UMA security and only supports program executions with \texttt{WriteTime} calls can be transformed into a fully secure garbled RAM scheme for arbitrary programs \cite{GHRW14}. This transformation uses \textit{oblivious RAM} (ORAM)\footnote{Oblivious RAM (ORAM), first proposed by Goldreich and Ostrovsky \textit{et al.} \cite{GO96}, permit a user to hide its access pattern to a remote storage. Although the physical storage locations accessed can be observed by an adversary, it is ensured by ORAM that anything about the real access pattern may not be learned \cite{SDS+12}.} to first compile the original program $P$ into a new program $P^*$ that stores/accesses its memory using ORAM. This ensures that the memory contents and access pattern of the compiled program do not reveal anything about those of the original program. Some ORAM schemes already ensure that the compiled program provides \texttt{WriteTime} calls.
\subsection{\texttt{JustGarble}} \label{sub:JustGarble}{}
In \cite{BHKR13}, Bellare \textit{et al.} proposed \texttt{JustGarble} framework, which targets optimized garbling of any circuit. It is entirely open-source and can be freely downloaded from \url{http://cseweb.ucsd.edu/groups/justgarble}. It implements Ga (P\&P (\ref{sec:PP})), GaX (Free \texttt{XOR} (\ref{sec:freeXOR}) without GRR3 (\ref{sec:GRR3})), and GaXR (Free \texttt{XOR} with GRR3), using constant key 128-bit \texttt{AES} as the \texttt{DKC} (\ref{part:dkcschemes}) as in Equation~\eqref{eq:encschemeaesconstkey}. It works both ways: garble a boolean circuit, and evaluate a garbled circuit.
\texttt{JustGarble} uses a circuit representation called Simple Circuit Description (SCD). It is based on the circuit formulation from \cite{BHR12}. An SCD file consists of values $n$, $m$, $q$, and arrays $A$, $B$, and $G$. If $G$ is not present the file is a topological circuit representation. In \texttt{JustGarble}, there are modules for building circuits, garbling boolean circuits, and evaluating garbled circuits. The Build module is useful for constructing circuits, permitting working at the individual gate level or higher. SCD files are written with constructed circuits. The Garble module is utilized for realizing the \textsc{\texttt{Gb}} algorithm of the three garbling schemes given. Garble takes a circuit $f = (n, m, q, A, B, G)$ described in an SCD as input and outputs the garbled tables $P$ that compose of the related garbled circuit $F = (n, m, q, A, B, P )$. The inputs to the Evaluate module are a topological circuit$ \bar{f} = (n,m,q,A,B)$, the garbled tables $P$ needed for evaluating, and a garbled input $X$. The garbled output $Y$ is produced. JustGarble also composes of procedures to realize \texttt{De}, mapping the garbled output $Y$ to the plain output $y$ \cite{BHKR13}.
The \texttt{JustGarble} implementation of GaXR (Free \texttt{XOR} (\ref{sec:freeXOR}) with GRR3 (\ref{sec:GRR3})) for 36.5K gate optimized \texttt{AES} boolean circuit whose 82\% are \texttt{XOR} gates has resulted in 5.40 bytes per gate (bpg) as the size, 35.0 cycles per gate (cpg) as the evaluation time, and 63.3 cpg as the garbling time. The JustGarble implementation of GaX for the same circuit, however, has yielded 23.2 cpg as the evaluation time, 55.6 cpg as the garbling time, and 11.5 as the size \cite{BHKR13}. (With a 3.201 GHz processor, evaluating the garbled circuit is 7.25 nsec/gate and garbling it is 17.4 nsec/gate.)
\subsection{\texttt{ABY}} \label{sub:ABY}{}
\texttt{ABY} is a framework for 2PC, proposed by Demmler \textit{et al.} in \cite{DSZ15}. Most of the time, a mixture of MPC primitives (GMW protocol (\ref{sec:gmwprotocol}), Yao, HE (\ref{sec:homoenc})$\ldots$) may yield more efficient implementations than what would have been if just one of them is used. Based on this idea, \texttt{ABY} uses \textbf{A}rithmetic sharing, \textbf{B}oolean sharing, and \textbf{Y}ao sharing (\ref{sec:secretshare}). The framework aims security in the semi-honest model (\ref{sub:semi-honest}). \texttt{ABY} works like a virtual machine, and high-level languages can be compiled to it. Variables may be either in Cleartext (i.e. one of the parties knows its value, e.g. inputs and outputs) or secret shared among the two parties. \texttt{ABY} also allows efficient conversion between the different types of sharings. The user of the framework may decide which sharings to be used depending on the application.
\subsection{\texttt{Obliv-C}}
\texttt{Obliv-C} is built by Zahur and Evans as an extension of C programming language with secure computation infrastructure \cite{ZE15}. It supports various C features like \texttt{pointers}, \texttt{typedef}, \texttt{struct}, \textit{etc.}, and provides new data types and constructions so that programs would run on private inputs. It is especially designed for scalable MPC protocols, and to enhance research on new MPC techniques by easing implementation such that just writing a new library is enough instead of building a new compiler for each technique. The source code for \texttt{Obliv-C} can be found at \url{https://oblivc.org}.
\subsection{\texttt{ObliVM}}
Liu \textit{et al.} proposed \texttt{ObliVM} as a programming framework for MPC \cite{DBLP:conf/sp/2015}. It offers a domain-specific language (\texttt{ObliVM-lang}) useful for compilation of programs into suitable representations required for MPC protocols. It also provides high-level programming constructions for MPC infrastructure which can be adapted by non-specialist programmers on security as well. The source code for \texttt{ObliVM} can be found at \url{http://oblivm.com}.
\subsection{\texttt{Frigate}} \label{sub:Frigate}{}
\texttt{Frigate} is designed by Mood \textit{et al.} as a compiler and a circuit interpreter for MPC \cite{MGC+16}. It can implement any function that can be written as a boolean circuit and run any MPC primitive that operates on boolean circuits. \texttt{Frigate} permits the use of C-like language with constructs and operators specifically designed for representing Boolean circuit efficiently. To improve the efficiency, the compiler is designed to favor \texttt{XOR} gates, utilizing structures like Boyar \textit{et al.}'s full adder with four \texttt{XOR} and one \texttt{AND} \cite{BPP00}. \texttt{Frigate} is also \textit{significantly fast} in terms of compilation, interpretation and execution times. The source code for \texttt{Frigate} can be found at \url{https://bitbucket.org/bmood/frigaterelease}.
\subsection{Comparison Based on Garbling Optimizations Used}
Now, we compare the generic frameworks for Yao's protocol based on their use of garbled circuit optimizations (see Table \ref{tab:comparefwgcopt}). \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \normalsize and $\mathsf{X}$ stand for compatible and non-compatible, respectively.
\vspace{7mm}
\begin{table}[H]
\centering
\caption{Comparison of Generic Frameworks Techniques Based on Their Use of Garbled Circuit Optimizations.}
\label{tab:comparefwgcopt}
\begin{tabular}{c|c|c|c|c|c|c}
& \footnotesize{\textbf{P\&P}} & \footnotesize{\textbf{GRR3}} & \footnotesize{\textbf{Free \texttt{XOR}}} & \footnotesize{\textbf{GRR2}} & \footnotesize{\textbf{Fle\texttt{XOR}}} & \footnotesize{\textbf{Half Gates}} \\ \hline
\footnotesize{\textbf{\texttt{JustGarble} (2013) \cite{BHKR13}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ & $\mathsf{X}$ & $\mathsf{X}$ \\ \hline
\footnotesize{\textbf{\texttt{ABY} (2015) \cite{DSZ15}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ & $\mathsf{X}$ & $\mathsf{X}$ \\ \hline
\footnotesize{\textbf{\texttt{Obliv-C} (2015) \cite{ZE15}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \\ \hline
\footnotesize{\textbf{\texttt{ObliVM} (2015) \cite{DBLP:conf/sp/2015}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & $\mathsf{X}$ & $\mathsf{X}$ & $\mathsf{X}$ \\ \hline
\footnotesize{\textbf{\texttt{Frigate} (2016) \cite{MGC+16}}} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} & \footnotesize{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;} \\
\end{tabular}
\end{table}
\texttt{Obliv-C} and \texttt{Frigate} make the use of any garbled circuit optimization possible since they permit the alterations of garbling schemes although those garbling circuit techniques are not built-in. On the other hand, \texttt{JustGarble}, \texttt{ABY}, and \texttt{ObliVM} do not allow changing the built-in garbling constructions, therefore, is limited for the use of state-of-the-art garbled circuit optimization techniques. All of the frameworks allow compilations optimized for reducing the number of odd gates. We can deduce that \texttt{Frigate} is the optimum for working with garbled circuits since it offer maximum optimization options while being the most efficient one.
\section{Real-World Applications} \label{sec:SomePracticalApplicationsinRealWorld}
We give two real-world examples indicating the importance of Yao's protocol in practise.
\subsection{Secure Computation of Satellite Collusion Probabilities} \label{sub:SecureComputationofSatelliteCollusionProbabilities}{}
Satellite operators are very eager to protecting their satellites since they are extremely costly. One of the issues that operators are interested in is preventing collisions with other satellites. However, the operators also want to keep the trajectories of their satellites private, which makes coordination between different operators difficult. Hemenway \textit{et al.} proposed an 2PC framework that combines GMW protocol (\ref{sec:gmwprotocol}) and Yao's protocol for high-precision computation of satellite collusion probabilities in \cite{HLOW16}. The framework does not target just the semi-honest model (\ref{sub:semi-honest}) since in the case of satellite operators, it does not provide sufficient security. Instead, first, they prove the security of the protocol in semi-honest (\ref{sub:semi-honest}) setting. Then, they strengthen their construction by using standard arithmetic \texttt{MAC}s against malicious adversaries (\ref{sub:malicious}).
For the sake of simplicity, the model of each satellite is a spherical object on a linear path in any short time window. Each satellite may deviate from its position, \textbf{p}, and the distribution of these deviations are assumed to be covariance matrix$^2$ \textbf{C}. The private input of a satellite $a$ includes four parts: its position \textbf{p}$_a$ in $\mathbb{R}^3$, its velocity \textbf{v}$_a$ in $\mathbb{R}^3$, the covariance matrix \textbf{C}$_a$ in $\mathbb{R}^{3\times3}$, and its radius $R_a$ in $\mathbb{R}$. The algorithm which needs to be calculated securely for satellites $a$ and $b$ is the conjunction analysis calculation, which returns the collision probability $p$ (see Algorithm \ref{alg:conjunctionanalysis}).
\begin{algorithm}
\scriptsize
\caption{The conjunction analysis calculation proposed in \cite{HLOW16}.}
\label{alg:conjunctionanalysis}
\begin{algorithmic}[1]
\State \textbf{Inputs: } $\{\textbf{v}_i,\textbf{C}_i,\textbf{p}_i,R_i\}_{i\in a,b}$
\State $\textbf{v}_r\gets \textbf{v}_b - \textbf{v}_a$, $\textbf{i}\gets \dfrac{\textbf{v}_r}{|\textbf{v}_r|}$, $\textbf{j}\gets \dfrac{\textbf{v}_b\times \textbf{v}_a}{|\textbf{v}_b\times \textbf{v}_a|}$, $\textbf{k}\gets\textbf{i}\times\textbf{j}$, $\textbf{Q}\gets \begin{bmatrix} \textbf{j} & \textbf{k} \end{bmatrix}$, $\textbf{C} \gets \textbf{Q}^T (\textbf{C}_a+\textbf{C}_b)\textbf{Q}$
\State $(\textbf{u},\textbf{v})\gets \texttt{Eigenvectors}(\textbf{C})$, $(\sigma_x^2,\sigma_y^2)\gets \texttt{Eigenvalues}(\textbf{C})$, $\sigma_x\gets\sqrt{\sigma_x^2}$, $\sigma_y\gets\sqrt{\sigma_y^2}$
\State $\textbf{u}\gets\dfrac{\textbf{u}}{|\textbf{u}|}$, $\textbf{v}\gets\dfrac{\textbf{v}}{|\textbf{v}|}$, $\textbf{U}\gets\begin{bmatrix} \textbf{u} & \textbf{v} \end{bmatrix}$, $\begin{bmatrix} x_m \\ y_m \end{bmatrix}\gets \textbf{U}^T\textbf{Q}^T(\textbf{p}_b-\textbf{p}_a)$
\State $p \gets \dfrac{1}{2\pi\sigma_x\sigma_y} \displaystyle\int_{-R}^R\int_{-\sqrt{R^2-x^2}}^{\sqrt{R^2-x^2}}f(x,y)dydx$ where $f(x,y)= \texttt{exp}\left[\dfrac{-1}{2}\left[\left(\dfrac{x-x_m}{\sigma_x}\right)^2+\left(\dfrac{y-y_m}{\sigma_y}\right)^2\right]\right]$ \label{lst:integral}
\State \textbf{return } $p$
\end{algorithmic}
\end{algorithm}
Hemenway \textit{et al.} propose GMW protocol (\ref{sec:gmwprotocol}) for computing integer addition and multiplications \cite{HLOW16}. To compute comparison and shift operations, they are represented as Boolean circuits and then evaluated using Yao's protocol. For compatibility with GMW protocol the garbled circuit must take secret inputs of both parties and the output of the gate must be computed as an arithmetic secret sharing (\ref{sec:secretshare}) among both sides.
\begin{itemize}
\item{A shift operation is computed as follows: $x_0$ and $x_1$ are Alice and Bob's arithmetic shares, respectively. $(x_0 + x_1)$ needs to be shifted by an amount $N$ which is known publicly. This can be accomplished by using Yao's protocol to compute Algorithm \ref{alg:shiftop}, where Alice is the garbler, and Bob is the evaluator. Bob uses OT (\ref{sec:oblivioustrans}) to get the masking values for his inputs.}
\end{itemize}
\begin{algorithm}
\caption{The shift operation computation proposed in \cite{HLOW16}.}
\label{alg:shiftop}
\begin{algorithmic}[1]
\State \textbf{Hardwired:} $M = 2m$ and $c$, which are a modulus and a shift constant, respectively.
\State \textbf{Inputs:} $x_0$ and $x_1$, held by Alice and Bob, respectively. In addition a random $R$ is provided by Alice.
\State $x \gets x_0 + x_1\text{ } (\texttt{mod }M )$ using standard $m$-bit addition circuit.
\State $y \gets x >> c$ by dropping $c$ rightmost wires.
\State \textbf{Return:} $z_1 \gets y + R\text{ } (\texttt{mod }M )$ to Bob. She sets $z_0 = -R \text{ }(\texttt{mod }M )$ for herself.
\end{algorithmic}
\end{algorithm}
\begin{itemize}
\item{A shift comparison is computed as follows: $x_0$ and $x_1$ are Alice and Bob's arithmetic shares, respectively. They would like to detect whether $(x_0 + x_1)$ is positive or not. This can be done by using Yao's protocol to compute Algorithm \ref{alg:comparisonop}, where Alice is the garbler, and Bob is the evaluator. Bob uses OT (\ref{sec:oblivioustrans}) to get the masking values for his inputs.}
\end{itemize}
\begin{algorithm}
\caption{The comparison operation computation proposed in \cite{HLOW16}.}
\label{alg:comparisonop}
\begin{algorithmic}[1]
\State \textbf{Hardwired:} $M = 2m$, which is a modulus.
\State \textbf{Inputs:} $x_0$ and $x_1$, held by Alice and Bob, respectively. In addition a random $R$ is provided by Alice.
\State $x \gets x_0 + x_1\text{ } (\texttt{mod }M )$ using standard $m$-bit addition circuit.
\State $b\gets sgn(x)$
\State \textbf{Return:} $z_1 \gets b + R\text{ } (\texttt{mod }M )$ to Bob. She sets $z_0 = -R \text{ }(\texttt{mod }M )$ for herself.
\end{algorithmic}
\end{algorithm}
Now, we return the computation of Algorithm \ref{alg:conjunctionanalysis}. In the rest of this section, we provide the methods proposed by Hemenway \textit{et al.} for the implementation of functions in Algorithm \ref{alg:conjunctionanalysis} \cite{HLOW16}.
\textbf{Circuit Representation for Division:} Integer division is implemented by repeated subtractions.
\textbf{Circuit Representation for} \texttt{exp}(): The function \texttt{exp}() must be implemented by representing it as a degree-24 Taylor series. Then the Taylor coefficients can be hard-coded constants in the circuit \cite{HLOW16}.
\textbf{Circuit Representation for} $\sqrt{\cdot}$: Iterative Babylonian Algorithm can be used to approximate a square root. The Babylonian Algorithm computes Equation~\eqref{eq:babylonianalg} on an input $S$, and an initial estimate $x_0$ \cite{HLOW16}.
\begin{equation}
\label{eq:babylonianalg}
x_{n+1}=\dfrac{1}{2}\left( x_n+ \dfrac{S}{x_n}\right)
\end{equation}
The double integral on Line \ref{lst:integral} of Algorithm \ref{alg:conjunctionanalysis} can be written as in Equation~\eqref{eq:integral} where $g(x)$ is a sum of \texttt{erf}s. Simpson's Rule approximation to this integral (\textit{i.e.}, using arcs of parabola) is suggested by Alfano in \cite{Alf05}.
\begin{equation}
\label{eq:integral}
p=\dfrac{3}{\sqrt{8\pi}\sigma_x}\displaystyle\int_{-R}^R g(x)dx
\end{equation}
\textbf{Circuit Representation for} \texttt{erf}(): approximate $1 - \texttt{erf}(x)$ using the degree 96 rational function in Equation~\eqref{eq:erfapprox} where $a_1 = .3275911$, $a_2 = .254829592$, $a_3 = .0092705272$, $a_4 = .0001520143$, $a_5 = .0002765672$, and $a_6 = .0000430638$.
\begin{equation}
\label{eq:erfapprox}
1 - \texttt{erf}(x) \approx \dfrac{1}{\left(1+a_1x+a_2x^2 +a_3x^3 +a_4x^4 +a_5x^5 +a_6x^6\right)^{16}}
\end{equation}
Hemenway \textit{et al.} demonstrate that their framework is highly efficient. The collision probability calculation scheme proposed requires numerical estimation of a complicated integral. The work of Hemenway \textit{et al.} proves that evaluating very complex functions is now possible by using MPC technology \cite{HLOW16}.
\subsection{Privacy-Preserving Data Mining} \label{sub:ppdatamining}{}
\textit{Privacy-preserving data mining} deals with the problem of how to run data mining algorithms on private data \cite{LP08}. Mainly, privacy-preserving data mining is applied to two classic settings \cite{LP08}:
\begin{enumerate}
\item{Instead of a single party having the whole data set, two or more parties hold different parts of it. Running a data mining algorithm on the union of the parties' databases is aimed while each party's input is being kept private \cite{LP08}.}
\item{Some part of statistical data that needs to be released may be confidential. Hence, it can be first altered so that }
\begin{enumerate}
\item{no one's privacy is compromised by it,}
\item{Data mining algorithms can be run on the modified data set to obtain meaningful results \cite{LP08}.}
\end{enumerate}
\end{enumerate}
Although both privacy problems are important, we will only deal with the first one where MPC techniques suit better. An example of the first type problem occurs in the field of medical research \cite{LP08}. A group of hospitals would like to mine their patient data jointly for the medical research purposes but they also need to keep their patients' personal data private. Another example would be a cooperation scenario of intelligence agencies. These agencies cannot grant each other free access to their confidential databases because of the high security standards they must obey \cite{LP08}.
The relationship of privacy preserving techniques and MPC is so wide that we cannot cover it here comprehensively. Instead here we will examine and explain common notions of classification problem and ID3 algorithm and their relationships with MPC.
\textbf{Classification problem.} The input of a \textit{classification problem} is a database structured such that each of its rows is a \textit{transaction} and each of its columns is an \textit{attribute} which may have different values (\textit{e.g.}, each row may be a patient, and each column may be a different type of symptoms that is found in the patient) \cite{LP08}. One of those attributes in the database is the main one, named as the class attribute (\textit{e.g.}, it may represent whether the patient has lung cancer or not) \cite{LP08}. We aim to use the database for prediction of the class of a new transaction by examining only its non-class attributes \cite{LP08}.
Another example would be credit risk analysis of a bank that wishes to identify which customers are likely to be profitable before giving them a loan \cite{LP00}. Then the class attribute is defined as \texttt{Profitable-customer} (its values may be \texttt{YES} or \texttt{NO}) by the bank. The database attributes used for prediction include: \texttt{Home-Owner}, \texttt{Income}, \texttt{Years-of-Credit}, and \texttt{Other-Delinquent-Accounts}. In order to ensure proper decision making, various rules are defined by the bank. For example \cite{LP00}:
\textbf{If} ($\texttt{Other-Delinquent-Accounts} = 0$) and ($\texttt{Income} > 30k$ or $\texttt{Years-of-Credit} > 3$)
\textbf{then} \texttt{Profitable-customer} = \texttt{YES} [\texttt{accept credit-card application}]
The collection of those rules that cover all possible transactions can be used for classification of a customer as profitable or not. The classification may include a probability of error \cite{LP00}.
\textbf{Decision tree.} Being a rooted tree, a \textit{decision tree} has internal nodes, each corresponding to an attribute, and the edges leaving each node, corresponding to the possible values of the attribute \cite{LP08}. The tree also has leaves, each containing the expected class value for a transaction that has the attribute values in the path from the root to that leaf. By using a decision tree, the class of a new transaction can be predicted by following the nodes from the root until the leaf. \cite{LP08}.
\begin{figure}
\centering
\label{fig:credit}
\begin{tikzpicture}
[
grow = right,
sibling distance = 6em,
level distance = 10.6em,
edge from parent/.style = {draw, -latex},
every node/.style = {font=\footnotesize},
sloped
]
\node [env] {\texttt{Other-}\\\texttt{Delinquent-}\\\texttt{Accounts}}
child { node [env] {\texttt{Profitable-}\\\texttt{customer}\\ $=$ \texttt{NO}}
edge from parent node [below] {$>0$?} }
child { node [env] {\texttt{Income}}
child { node [env] {\texttt{Years-of-}\\\texttt{Credit}}
child { node [env] {\texttt{Profitable-}\\\texttt{customer}\\ $=$ \texttt{NO}}
edge from parent node [below] {$<1$ year?} }
child { node [env] {\texttt{Profitable-}\\\texttt{customer}\\ $=$ \texttt{NO}}
edge from parent node [above] {1-3}
node [below] {years?} }
child { node [env] {\texttt{Profitable-}\\\texttt{customer}\\ $=$ \texttt{YES}}
edge from parent node [above] {$>3$ years?} }
edge from parent node [below] {$<30k?$} }
child { node [env] {\texttt{Years-of-}\\\texttt{Credit}$\ldots$}
edge from parent node [above] {$>30k?$}
}
edge from parent node [above] {0?} };
\end{tikzpicture}
\caption{A decision tree for credit eligibility.}
\end{figure}
Figure 5.1 shows an example decision tree for identifying profitable customers as in the previous scenario. However, it reflects only a small portion of the tree. The whole tree would have many more nodes, edges, and leaves.
\textbf{ID3 Algorithm.} One of the well-known ways for designing decision trees is the use of \textit{ID3 algorithm} \cite{LP08}. The construction of the tree starts from the root node, goes top-down recursively. At each node the attribute is chosen based on its ability of classifying the transactions on its own. If an attribute is chosen for a node, the remaining transactions are partitioned by it, resulting in a smaller database which composes of the related transactions \cite{LP08}.
The main principle of ID3 is choosing the attribute which is best at predicting the class of the transaction. This is done by searching the attribute that decreases the information of the class to the maximum degree \cite{LP08}. Namely, by choosing the attribute maximizing the \textit{information gain}, which is the difference between the entropies of the class attribute for all transactions and for ones having the same value for a give attribute \cite{LP08}. The resulting decision tree is a smaller one consistent with the database due to the greedy algorithm used in searching \cite{LP08}.
\textbf{Privacy preserving distributed computation of ID3.} We include a setting involving two parties, each having a database with different transactions to which the same set of attributes applies \cite{LP08}. The parties aim at computing a decision tree of the union of their databases by using the ID3 algorithm \cite{LP08}. Lindell and Pinkas describe an efficient privacy preserving protocol to solve this problem in \cite{LP00}.
According to Lindell and Pinkas, direct application of Yao's protocol faces some major problems, mainly the large sizes of input databases require too many OTs (\ref{sec:oblivioustrans}), resulting in huge communication and computation costs \cite{LP08}. Moreover, the boolean circuit conversion of ID3 results in a very large circuit, because of myriad repetitions of information gain calculation which is the basic step of the algorithm \cite{LP08}.
Lindell and Pinkas observe that MPC of each node can be done separately \cite{LP00,LP08}. Starting with the root node, for each node a secure computation is invoked. Its output is revealed to both parties and the computation goes with the next node in the path. This does not compromise the protocol security since the assigned attribute to each node is also a part of the final output. Just like the non-privacy preserving implementation of ID3, both parties separately partition the rest of their transactions after an attribute is assigned to a node. This way, Lindell and Pinkas reduce the whole protocol to proper attribute assignments for node, namely the ones resulting in the highest information gains \cite{LP08}. They also show how to apply Yao's protocol to proper attribute assignment \cite{LP08}.
\chapter{Private Function Evaluation}
\label{Chapter6}
\lhead{Chapter 6. \emph{Private Function Evaluation}}
Consider the case that one invents an algorithm which can be used for efficiently diagnosing various diseases based on some information about a person's general health \cite{Pil15}. It is obvious that this algorithm would be precious, and healthcare institutions would volunteer to pay millions in order to use it. However, the inventor of the algorithm would prefer keeping it as a secret since he is regularly payed for it a lot of money. The problem is that medical institutions generally prefer keeping their patients' data private, preventing them from just giving it to the algorithm owner. Here the following question might be asked: How can those parties compute an algorithm which is known by only one of the parties while its input is known by only the other one? This problem is known as \textit{private function evaluation} (PFE) \cite{Pil15}. The problem may also be widened to involve the case that the algorithm owner may also have his private inputs.
PFE is a special case of MPC in which $n$ participants needs to compute a private function $f$ using their private inputs ($x_1,\ldots, x_n)$), resulting in $f(x_1,\ldots, x_n)$. One of the parties $P_1$ holds a boolean circuit $\mathcal{C}_f$ of the function $f$, while each party $P_i$ holds a private input $x_i$, and the parties aim to learn only the output of the circuit $\mathcal{C}_f(x_1,\ldots, x_n)$ while $f$ or all other partys' inputs remain unknown to each of them except for $P_1$ who already knows $f$ \cite{MS13}. The difference of this scheme from the standard MPC setting is that here the function $f$ and its boolean circuit representation $\mathcal{C}$ are not known publicly. There are many situations where such a PFE scheme would be useful, \textit{e.g.} the ones where the function itself contains private information, or reveals security weaknesses, or the ones where service providers may prefer hiding their function or its specific implementation as their Intellectual Property. Design of efficient special or generic PFE protocols is considered in a variety of papers in literature \cite{MS13}.
Most generic PFE solutions target the MPC of a \textit{universal circuit} $U_g$ taking the circuit $\mathcal{C}$ with a number of gates less than $g$ and the inputs $x_1,\ldots, x_n$ of parties as input, and outputing $f(x_1,\ldots, x_n)$. The works based on this approach mainly aim to reduce the size of universal circuits, and to optimize their implementations with the help of various MPC techniques, such as Yao's protocol. However, they have a main source of inefficiency the massive sizes of known universal circuits. The complexity in their designs and implementations also increases the need for searching better alternatives.
In this section, we will explain the concepts and constructions for PFE proposed by Mohassel and Sadeghian in \cite{MS13}, especially for two-party case where Yao's protocol is involved.\footnote{For a clear explanation of multi-party case where GMW protocol is privately evaluated, we must refer the reader to \cite{Pil15}.} The target security is in the semi-honest (\ref{sub:semi-honest}) setting. Their work remains the most efficient PFE scheme to this date.
\section{Mohassel and Sadeghian's Generic PFE Scheme \cite{MS13}}\label{sec:genericpfe}
Mohassel and Sadeghian present a generic PFE framework in \cite{MS13}. In addition to the private inputs of parties which is hidden by any proper MPC scheme, hiding the topology of a boolean circuit $\mathcal{C}$ and the functionality of its gates suffices for hiding a circuit completely \cite{MS13}.
There are three types of information that Mohassel and Sadeghian's PFE scheme does not intend to hide about a circuit \cite{MS13}:
\begin{enumerate}
\item{The number of its inputs,}
\item{The number of its outputs,}
\item{The number of its gates.}
\end{enumerate}
Mohassel and Sadeghian suggest two different functionalities that make up the complete task of PFE \cite{MS13}:
\begin{enumerate}
\item{\textbf{Circuit Topology Hiding} (\texttt{CTH}) \textbf{Functionality.} The full description of the topology of a circuit $\mathcal{C}$ can be accomplished with the use of a mapping $\pi_{\mathcal{C}} : \texttt{OW}\rightarrow \texttt{IW}$. Let $g$, $n$ and $m$ denote the size, the number of inputs and the number of outputs of $\mathcal{C}$, respectively. \texttt{OW} (outgoing wires) is the union of the input wires of the circuit and the output wires of its non-output gates: $ \{ \texttt{ow}_1 = x_1,\ldots,\texttt{ow}_n = x_n,\texttt{ow}_{n+1}=\texttt{Output}(G_1),\ldots,\texttt{ow}_{n+g-m}=\texttt{Output}(G_{g-m})\}$. \texttt{IW} (incoming wires) is the set of input wires to all the gates in the circuit: $\{\texttt{iw}_1,\ldots,\texttt{iw}_{2g}\}$. $\pi_{\mathcal{C}}$ maps $i$ to $j$ (\textit{i.e.}, $\pi_{\mathcal{C}}(i) \rightarrow j)$, if and only if $\texttt{ow}_i \in \texttt{OW}$ and $\texttt{iw}_j \in \texttt{IW}$ correspond to the same wire in the circuit $\mathcal{C}$. Because an outgoing wire can correspond to more than one incoming wire, $\pi_{\mathcal{C}}$ is rarely a function. However, its inverse $\pi_{\mathcal{C}}^{-1}$ is a function since a wire can be either an output of only one gate or an input. Figure \ref{fig:mapping} shows an example circuit (a) and its mapping $\pi_{\mathcal{C}}$ (b). The main target of the \texttt{CTH} \textit{functionality} is the oblivious application of this mapping $\pi_{\mathcal{C}} : \texttt{OW}\rightarrow \texttt{IW}$.}
\end{enumerate}
\begin{figure}
\includegraphics[height=5cm, angle=0]{mapping}
\caption{(a) An example circuit \cite{Pil15}. (b) The mapping of the circuit \cite{Pil15}.}
\label{fig:mapping}
\end{figure}
It is useful to include a computation of the number of possible mapping since it is directly related to the security of the PFE scheme. Although one may expect the number of possible mappings to be $M^N$ due to the ability of any $\texttt{ow}_i$ to go to any $\texttt{iw}_j$, the exact value is smaller since an $\texttt{ow}_i$ must go at least one $\texttt{iw}_j$. Since $\pi^{-1}$ is an onto function, computing the number of possible onto functions suffices. Applying the inclusion-exclusion principle, we get Equation~\eqref{eq:numberofmappings} which shows the number ($\rho$) of possible mappings for a circuit where \texttt{OW} has $M$ elements and \texttt{IW} has $N$ elements \cite{Maz10}.
\begin{equation}
\label{eq:numberofmappings}
\rho=\sum_{i=0}^{M} (-1)^i {M\choose i} (M-i)^N
\end{equation}
\begin{enumerate}
\setcounter{enumi}{1}
\item{\textbf{Private Gate Evaluation} (\texttt{PGE}) \textbf{Functionality.} The \texttt{PGE} \textit{functionality} deals with hiding the functionality of each gate in a circuit. It can be seen as a black-box gate mechanism where only one of the parties ($P_1$) knows its functionality. The input of the mechanism is the shares of all parties for both inputs of the hidden gate, and it returns to the parties their shares for the output of the gate.}
\end{enumerate}
\section{\texttt{CTH} Functionality Realization}\label{sec:cthrealization}
Before describing Mohassel and Sadeghian's construction in more detail, the concept of an \textit{extended permutation} needs to be explained \cite{MS13}. A mapping $\pi : \{1\ldots M\} \rightarrow \{1\ldots N\}$ can be regarded as a permutation if it is one-to-one and onto (\textit{i.e.} a bijection). This notion can be generalized to an extended permutation as follows: Given the positive integers $M$ and $N$, a mapping $\pi : \{1\ldots M\} \rightarrow \{1\ldots N\}$ is called as an extended permutation (\texttt{EP}) if and only if there exists exactly one $x \in \{1\ldots M\}$ for every $y \in \{1\ldots N\}$ such that $\pi(x) = y$. $x$ is often denoted by $\pi^{-1}(y)$. Unlike the mapping of a standard permutation, the mapping of an \texttt{EP} may also replicate or discard elements in the domain, allowing the domain to be larger or smaller than the range.
$n+q-m$ \textit{oblivious mapping} (\texttt{OMAP}) queries and $2q$ \texttt{Reveal} queries are needed to be implemented in order to realize the \texttt{CTH} functionality (an \texttt{OMAP} query for each $\texttt{ow}_i$, and a \texttt{Reveal} query for each $\texttt{iw}_i$). These \texttt{OMAP}/\texttt{Reveal} queries can be combined to construct a problem known as \textit{oblivious evaluation of the extended permutation} (\texttt{OEP}) to which Mohassel and Sadeghian's \texttt{CTH} scheme mainly address.
\textbf{\texttt{OEP} Definition.} Two-party \texttt{OEP} Problem 2-$\texttt{OEP}(\vec{\pi},\vec{x},\vec{t})$ is defined as follows: The first party $P_1$ holds an \texttt{EP} $\pi : \{1\ldots M\} \rightarrow \{1\ldots N\}$, and a blinding vector for outputs $\vec{t} = (t_1,\ldots,t_N)$; whereas the other party $P_2$ holds a vector of inputs $ \vec{x} = (x_1,\ldots,x_M)$. Both the $x_i$s and $t_i$s are $\ell$-bit strings. The protocol ends in $P_2$ learning $(x_{\pi^{-1}(1)}\oplus t_1,\ldots,x_{\pi^{-1}(N)}\oplus t_N)$, while $P_1$ learning nothing.
Mohassel and Sadeghian construct a solution for \texttt{OEP} from switching networks which they observe as more efficient than the previous constructions.
\textbf{Switching Networks.} A \textit{switching network} \texttt{SN} composes of \textit{2-switches} which are interconnected. Its inputs are $N$ $\ell$-bit strings and a set of selection bits of each switches, while its outputs are $N$ $\ell$-bit strings. Each switch takes two $\ell$-bit strings and two selection bits as input, outputting two $\ell$-bit strings. Each of the outputs may get the value of any of the input strings depending on the selection bits. This means for input values $(x_0,x_1)$ and output values $(y_0,y_1)$, there are four different switch output possibilities. The two selection bits $s_0$ and $s_1$ are used for determining the switch output. In particular, the switch will output $y_0 = x_{s_0}$, and $y_1 = x_{s_1}$.
The mapping $\pi : \{1\ldots N\} \rightarrow \{1\ldots N\}$ of an \texttt{SN} is defined as $\pi(i)=j$ if and only if after the \texttt{SN} is evaluated, the string on the output wire $j$ becomes that on the input wire $i$. There is no need for the mapping $\pi$ to be a function because the value of any input wire can be mapped to any number of output wires. However, its inverse $\pi^{-1}$ must always be a function.
A \textit{permutation network} \texttt{PN} is a switching network whose mapping is a permutation of its inputs. In contrast to switching networks, permutation networks compose of \textit{1-switches}. Unlike 2-switches, they have only one selection bit $s$. For an input $(x_0,x_1)$, a 1-switch outputs one of the two possible outputs: $(x_0,x_1)$ if $s=0$, and $(x_1,x_0)$ if $s=1$. 1-switch may also be called a \textit{permutation cell}.
Waksman proposed an efficient construction for a permutation network in \cite{Wak68}. Mainly, his work suggests that a permutation network with $N=2^k$ can be constructed with $N \texttt{log}_2 N - N + 1 $ switches, that the switch depth of the constructed \texttt{PN} will be $2 \texttt{log}_2N - 1$, and that its computational complexity will be $O(N \texttt{log}_2 N )$.
\textbf{Extended Permutation from Switching Networks.} Mohassel and Sadeghian propose the general method for construction of an extended permutation from switching and permutation networks \cite{MS13}. However, extended permutations differ from switching networks in that the number of their inputs $M$ and that of their outputs $N$ need not be equal ($M \leq N$) \cite{MS13}. $N - M$ additional dummy inputs are added to the real inputs of an \texttt{EP} $\pi : \{1\ldots M\} \rightarrow \{1\ldots N\}$ in order to simulate it as an \texttt{SN}.
\begin{figure}[]
\includegraphics[height=6cm, angle=0]{EP}
\caption{The switching network for \texttt{EP} of the circuit in Figure \ref{fig:mapping} \cite{Pil15}.}
\label{fig:ep}
\end{figure}
Mohassel and Sadeghian divide a switching network into three components \cite{MS13}:
\begin{enumerate}
\item{\textbf{Dummy-value placement component.} This component takes $N$ input strings composing of real and dummy ones. For each real input that $\pi$ maps to $k$ different outputs, the dummy-value placement component's output is the real string followed by $k-1$ dummy strings. An efficient implementation of this process can be via a Waksman permutation network \cite{Wak68}.}
\item{\textbf{Replication component.} This component takes the output of the dummy-value placement component as input. If a value is real, it goes unchanged. If it is a dummy value, it is replaced by the real value which precedes it. This can be computed by a series of $N - 1$ 2-switches whose selection bits $(s_0,s_1)$ are either (0,0) or (0,1). If the selection bits are (0,0), that means $x_1$ is dummy, and $x_0$ goes both of the outputs. If they are (0,1), that means both inputs are real, and both are kept on the outputs in the same order. At the end of this step, all the dummy inputs are replaced by the necessary copies of the real inputs.}
\item{\textbf{Permutation component.} This component takes the output wires of the replication component as input and outputs a permutation of them so that each string is placed on its final location according to the prescription of mapping $\pi$. An efficient implementation of this process can also be via a Waksman permutation network \cite{Wak68}.}
\end{enumerate}
Adding up the three components, the number of switches needed for implementation of \texttt{EP} is $2(N \texttt{log}_2 N - N + 1) + N - 1 = 2N \texttt{log}_2 N - N + 1$. The topology of the whole switching network is the same for all $N$ input \texttt{EP}s and the output depends on the selection bits.
\textbf{Oblivious Evaluation of Switching Networks (OSN).} Now, we can return to our \texttt{OEP} problem. If the \texttt{EP} construct from switched and permutation networks can be evaluated oblivously, we have a solution. Mohassel and Sadeghian propose a method for oblivious evaluation of their building blocks, \textit{i.e.}, 1-switches and 2-switches \cite{MS13}.
Recall that $P_1$ holds the selection bits of the switching network, and an output blinding vector $\vec{t}$ while $P_2$ holds the input vector $\vec{x}$. $P_2$ must learn the switching network's blinded output which is the \texttt{EP} of her input vector blinded with the vector $\vec{t}$; while $P_1$ learns $\bot$.
\textbf{Secure evaluation of a single 2-switch.} The express the general idea of the secure computation of whole network, Mohassel and Sadeghian describe the secure evaluation of its building block, a single 2-switch $u$ \cite{MS13}. Let the input wires of the 2-switch be $w_i$ and $w_j$, and its output wires be $w_k$ and $w_l$. $P_2$ assigns four uniformly random values $r_i$, $r_j$, $r_k$, $r_l$ to the four wires of the switch. $P_1$ has the blinded values $x_i \oplus r_i$ and $x_j \oplus r_j$ as his shares for the two input wires. The aim is letting $P_1$ obtain his output shares which is the blinded values on the output wires (see Table \ref{tab2swtichout}). In fact, there are four possible output pairs ($x_i \oplus r_k$, $x_i \oplus r_l$), ($x_i \oplus r_k$, $x_j\oplus r_l$), ($x_j \oplus r_k$, $x_i \oplus r_l$), or ($x_j\oplus r_k$, $x_j \oplus r_l$) which $P_1$ may obtain based on the values of his selection bits $s_{0u}$ and $s_{1u}$.
\begin{table}[]
\centering
\caption{$P_1$ must learn one of these ($y_0$,$y_1$) according to his selection bits.}
\label{tab2swtichout}
\begin{tabular}{c|c|c}
{($s_{0u}$,$s_{1u}$)} & {$y_0$} & {$y_1$} \\ \hline
(0,0) & $x_i \oplus r_k$ & $x_i\oplus r_l$ \\
(0,1) & $x_i \oplus r_k$ & $x_j\oplus r_l$ \\
(1,0) & $x_j \oplus r_k$ & $x_i\oplus r_l$ \\
(1,1) & $x_j \oplus r_k$ & $x_j\oplus r_l$
\end{tabular}
\end{table}
\begin{table}[b]
\centering
\caption{$P_1$ gets one of these ($T_0$,$T_1$) by engaging in 1-out-of-4 OT (\ref{sec:oblivioustrans}) with $P_2$.}
\label{tab2swtichmasks}
\begin{tabular}{c|c|c}
{($s_{0u}$,$s_{1u}$)} & {$T_0$} & {$T_1$} \\ \hline
(0,0) & $r_i \oplus r_k$ & $r_i\oplus r_l$ \\
(0,1) & $r_i \oplus r_k$ & $r_j\oplus r_l$ \\
(1,0) & $r_j \oplus r_k$ & $r_i\oplus r_l$ \\
(1,1) & $r_j \oplus r_k$ & $r_j\oplus r_l$
\end{tabular}
\end{table}
$P_2$ prepares a table with four rows: ($r_i\oplus r_k$, $r_j\oplus r_l$), ($r_i\oplus r_k$, $r_i\oplus r_l$), ($r_j\oplus r_k$, $r_i\oplus r_l$), and ($r_j \oplus r_k$, $r_j \oplus r_l$) as shown in Table \ref{tab2swtichmasks}. Then, $P_1$ and $P_2$ engage in a 1-out-of-4 OT (\ref{sec:oblivioustrans}) in which $P_2$ inputs the four rows that he just prepared, and $P_1$ inputs his selection bits for the switch $u$. Suppose that $P_1$'s selection bits are (0,0). This means $P_1$ retrieves the first row, \textit{i.e.}, ($r_i \oplus r_k$, $r_j \oplus r_l$). He then \texttt{XOR}s $x_i \oplus r_i$ and $r_i \oplus r_k$, as well as $x_j \oplus r_j$ and $r_i \oplus r_l$, reaching his output shares $x_i \oplus r_k$ and $x_i \oplus r_l$.
\textbf{Constant round protocol.} Using the OT-based protocol proposed for 2-switches, the entire switching network can be securely computed in constant round since the protocol permits parallel OT (\ref{sec:oblivioustrans}) runs \cite{MS13}. In an \textit{offline} stage, a set of random strings for each wire is generated, and a table for each switch is prepared by $P_2$. Then $P_1$ and $P_2$ run the parallel OTs (\ref{sec:oblivioustrans}) as described above, leading to that a single row of each table is learned by $P_1$ according to his selection bits.
In the \textit{online} stage, $P_2$ blinds his input vector with the blinding strings on the inputs of the input switches before sending them to $P_1$. $P_1$ is now able to compute the entire switching network. He just need to perform sequential \texttt{XOR}s (in topological order) to reach the blinded values on the output wires. He then applies his own blinding vector $\vec{t}$ and sends the result to $P_2$. $P_2$ removes her blinding, and obtains the output of the \texttt{OEP} \cite{MS13}.
\textbf{Efficiency of the Mohassel and Sadeghian's \texttt{OEP}.} As we mentioned before, to implement an extended permutation $\pi: {1\ldots M} \rightarrow {1\ldots N}$, $2N \texttt{log} N - N + 1$ switches are needed. In fact, 1-out-of-2 OTs (\ref{sec:oblivioustrans}) are enough to implement \texttt{PN}s which consist of 1-switches. Moreover, 2-switches in replication component can also be implemented with 1-out-of-2 OTs (\ref{sec:oblivioustrans}) since their outputs have 2 possibilities unlike the generic 2-switches \cite{MS13}. To sum up, this protocol costs $2N \texttt{log} N - N + 1$ 1-out-of-2 OTs (\ref{sec:oblivioustrans}). Mohassel and Sadeghian suggests the use of OT extension \cite{IKNP03}, which reduce total number of public key operations for their \texttt{OEP} to a constant value depending on the security parameter of protocol, \textit{i.e.} $O(k)$ \cite{MS13}. In this case, the number of symmetric key operations will be twice the number of OTs, which is $4N \texttt{log} N - 2N + 2$ \cite{MS13}.
\begin{figure}
\includegraphics[height=3cm, angle=0]{oep}
\caption{Basic procedures of topology hiding: (1)The function $f$ known by $P_1$. (2) Circuit representation of $f$. (3) Circuit mapping of $f$. (4) \texttt{OEP} for $P_2$ learning blinded values. (5) The blinded values learnt by $P_2$. (6) Yao's protocol with the blinded values.}
\label{fig:oep}
\end{figure}
Figure \ref{fig:oep} summarizes the basic procedures of topology hiding via \texttt{OEP}. $P_1$ owns a function $f$ (1). He converts $f$ to a circuit representation $C_f$ (2). Then he extracts the circuit mapping $\pi_f$ (3). $P_1$ and $P_2$ engage in an $\texttt{OEP}$ (4) of $\pi_f$ where $P_2$ learns the blinded values of her input masking values (5) which she will later use in Yao's protocol (6).
Mohassel and Sadeghian show applications of their framework to arithmetic circuits, GMW protocol (\ref{sec:gmwprotocol}), and Yao's protocol. Since the main topic of this thesis Yao's protocol and two-party cases, we will continue with its application to Yao's protocol \cite{MS13}.
\section{Two-Party PFE of Yao's Protocol}\label{twopartypfe}
Alice and Bob would like to compute a function $f(x_0,x_1)$, where $x_0$ is Alice's input, $x_1$ is Bob's input. Bob acts like $P_1$, and Alice acts like $P_2$ in Mohassel and Sadeghian's generic PFE scenario \cite{MS13}. So, only Bob knows $f$, and the topology of the circuit $\mathcal{C}$ of $f$. Since (\texttt{NAND}) is Turing complete, all gates in the circuit are let to be a \texttt{NAND} gate, so that the need for \texttt{PGE} functionality can be eliminated. Alice may learn the number of gates, but she should know the circuit topology. Now, one may ask the following question: How can someone garble a circuit which she does not know? Well, cryptography can achieve many incredible things.
The protocol goes as follows \cite{MS13}:
\texttt{Offline Preparation:}
\begin{enumerate}
\setcounter{enumi}{0}
\item{Bob sorts the gates topologically and computes the extended permutation $\pi_\mathcal{C}$ corresponding to circuit $\mathcal{C}$.}
\item{Alice randomly generates a masking value pair ($W_{i}^0$,$W_i^1$) for each $\texttt{ow}_i \in \texttt{OW}$. This yields a total of $M = n + g - o$ pairs. Each masking value is $k$ bits long, where $k$ is the security parameter. The \texttt{lsb} of 2 masking values belonging to the same pair must be different so that they have different labels.}
\item{Alice generates a bit vector $\vec{v} = (v_1,\ldots,v_M)$ where $v_i = \texttt{lsb} W_i^0$. She arranges each masking value pair with respect to their labels. So, they become ($W_i^{v_i}$,$W_i^{\bar{v_i}}$). This arrangement will be important during the garbled circuit evaluation. Moreover, she assigns those pairs to 2 vectors $\vec{p}=(p_1,\ldots,p_M)$ and $\vec{q}=(q_1,\ldots,q_M)$ where $p_i=W_i^{v_i}$ and $q_i=W_i^{\bar{v_i}}$.}
\item{Bob generates a random bit vector $\vec{v}' = (v_1',\ldots,v_N')$ where $v_j'$ is a random bit. This yields a total of $N = 2 g$ bits. He also generates random blinding pairs ($t_j^0$,$t_j^1$) for each $\texttt{iw}_j \in \texttt{IW}$ such that $\texttt{lsb}(t_j^0)=\texttt{lsb}(t_j^1)$. He assigns those pairs to 2 blinding vectors $\vec{t^0}=(t_1^0,\ldots,t_N^0)$ and $\vec{t^1}=(t_1^1,\ldots,t_N^1$).}
\end{enumerate}
\texttt{Oblivious Evaluation of Switching Networks:}
\begin{enumerate}
\setcounter{enumi}{4}
\item{Alice and Bob engage in an \texttt{OEP} protocol where his input is the extended permutation $\pi_\mathcal{C}$ and $\vec{v}'$, while her input is $\vec{v}$. As a result, Alice learns $v''= (v_1'',\ldots,v_N'')$ where $v_j'' = v_{\pi^{-1}(j)} \oplus v_j'$.}
\item{Alice and Bob engage in a slightly modified \texttt{OEP} protocol where his input is the extended permutation $\pi_\mathcal{C}$ and $\vec{t^0}$, while her input is $\vec{p}$. The output is a vector $p'=(p_1',\ldots,p_N')$ where $p_j' = p_{\pi^{-1}(j)} \oplus t_j^0$. The modification is that the output is not learned by Alice but fed to a new permutation network $\hat{\texttt{SN}}$.}
\item{Alice and Bob engage in a slightly modified \texttt{OEP} protocol where his input is the extended permutation $\pi_\mathcal{C}$ and $\vec{t^1}$, while her input is $\vec{q}$. The output is a vector $q'=(q_1',\ldots,q_N')$ where $q_j' = q_{\pi^{-1}(j)} \oplus t_j^1$. The modification is that the output is not learned by Alice but fed to $\hat{\texttt{SN}}$ as well.}
\item{$\hat{\texttt{SN}}$ is a switching network including $N$ 1-swiches whose switch depth is 1. Each 1-switch $u_j$ takes ($p_j'$,$q_j'$) as input, $v_j'$ as the selection bit and outputs either ($p_j'$,$q_j'$) if $v_j'=0$ or ($q_j'$,$p_j'$) if $v_j'=1$.}
\item{After oblivious evaluation of $\hat{\texttt{SN}}$, Alice learns the output which is a set of $N$ pairs whose $j^{th}$ element is either ($p_j'$,$q_j'$) if $v_j'=0$ or ($q_j'$,$p_j'$) if $v_j'=1$.}
\end{enumerate}
\texttt{Garbling:}
\begin{enumerate}
\setcounter{enumi}{9}
\item{Alice needs to arrange the blinded pairs into their original position since the truth values must be known for garbling. This can be done by using $v''$. If $v_j''=0$, the pair remains unchanged, otherwise it is swapped. $j^{th}$ element of the output will be ($\hat{W}_{\pi^{-1}(j)} ^0$,$\hat{W}_{\pi^{-1}(j)}^1$) where $\hat{W}_i^b$ means a blinded value for $W_i^b$.}
\item{For all gates, Bob tells Alice which two of the incoming wires and which one of the outgoing wires belong to the same gate. He also tell her the outgoing wires corresponding to his and her input bits.}
\item{Alice garbles each gate by encrypting the masking values on the outgoing wires using the blinded values on the incoming wires as the keys. She sends Bob the garbled gates and the masking values for her inputs in \texttt{OW}. Bob gets his input masking values from her using 1-out-of-2 OT (\ref{sec:oblivioustrans}).}
\end{enumerate}
\texttt{Evaluating:}
\begin{enumerate}
\setcounter{enumi}{12}
\item{Using the circuit mapping $\pi_\mathcal{C}$, his blinding vectors $\vec{t^0}=(t_1^0,\ldots,t_N^0)$ and $\vec{t^1}=(t_1^1,\ldots,t_N^1$), and the garbled gates told by Alice, Bob evaluates the whole garbled circuit in topological order. When an outgoing wire $i$ is mapped to an incoming wire $j$, the masking value $W_i$ on that outgoing wire is \texttt{XOR}ed with $t_j^{\texttt{lsb}W_i}$ on the incoming wire $j$. These \texttt{XOR}ed (blinded) values are used as the decryption keys in the corresponding garbled gates to reveal the next masking value on the outgoing wire of the gate.}
\item{In the end, Bob reaches the output masking values. He tells Alice those output masking values. She decodes them and reaches $f(x_0,x_1)$. Alice tells Bob the output.}
\end{enumerate}
\textbf{Complexity}. The steps 5, 6 and 7 can be combined for only one \texttt{OEP}. Hence, this protocol requires $2N \texttt{log}_2 N - N + 1$ OTs for \texttt{OEP} and $N$ OTs for $\hat{\texttt{SN}}$, \textit{i.e.} $2N \texttt{log}_2 N + 1$ OTs in total. OTs for Bob's input masking values increases the total OT requirement of complete two-party PFE protocol but they do not change its round complexity since they can be implemented in parallel with the OTs for \texttt{OEP}.
\chapter{Conclusion and Discussions}\label{chap:conclusion}
\lhead{Chapter \ref{chap:conclusion}. \emph{Conclusion and Discussions}}
In this thesis, we were interested in surveying all known Yao's protocol optimizations and showing practical applications of Yao's protocol.
We have presented P\&P (\ref{sec:PP}), GRR3 (\ref{sec:GRR3}), free \texttt{XOR} (\ref{sec:freeXOR}), GRR2 (\ref{sec:GRR2}), fle\texttt{XOR} (\ref{sec:fleXOR}), and half gates (\ref{sec:halfgates}) techniques in the descending order for size of garbled gates. We have compared those optimizations in terms of communication and computation complexities, and showed their compatibilities with each other.
What else can be done for optimization? Well, in science, especially in cryptography there is no end. Although Zahur \textit{et al.} have proved that the half gates method gives the most size-optimum technique for an odd gate and yet compatible with free \texttt{XOR} \cite{ZRE15}, there are still two more optimization parameters that can be improved. There may be faster and/or more secure garbling techniques in the future. To improve on the size parameter there is a need for a revolutionary change in the traditional approach. This improvement may be a method which garbles a group of gates together instead of garbling each gate separately, resulting in a lower size.
We have also presented some generic applications as well as some real-world application examples. The generic applications include pipelining method (\ref{sub:PipelinedImplementationFastGC}) which is useful for reducing total protocol time when the both parties of a garbling scheme is online at the same time. We also included garbling RAM (\ref{sub:GarbledRAM}) which is a quite useful technique especially for applications within the realm of big data. Some generic MPC tools \texttt{JustGarble} (\ref{sub:JustGarble}), \texttt{ABY} (\ref{sub:ABY}), \texttt{Obliv-C}, \texttt{ObliVM}, and \texttt{Frigate} (\ref{sub:Frigate}) are also introduced briefly. We compared them in terms of the use of garbling optimization techniques. At the end of the chapter, we have given some real-world applications, including MPC of satellite collusion probabilities (\ref{sub:SecureComputationofSatelliteCollusionProbabilities}) and privacy preserving data mining.
We have explained private function evaluation, and Mohassel \textit{et al.}'s PFE scheme. It is the most efficient PFE scheme known. Although their PFE scheme is limited for use right now, we know that cryptography is one of the fastest fields in computing science. It is hard to say whether it will be in use soon but someday generic PFE schemes will be in every day use for many applications, where one of the parties is also willing to hide her function since the path to developing such a technique is already open.
\addtocontents{toc}{\vspace{0em}}
|
\section{Introduction}
Diabetes is a chronic disease affecting over sixty million people \citep{wild04}. Diabetics are unable to correctly regulate blood glucose concentrations which, if not succesfully managed, leads to multiple adverse complications. Typically, management involves subcutaneous administration of insulin to minimise plasma glucose concentration whilst keeping it above a lower bound to avoid hypoglycaemia. Current treatment is invasive and often leads to poor outcomes. Hence, much recent effort has been devoted to developing an artificial pancreas which automates treatment \citep{harv10} and provides better control of glucose concentrations. The development of such systems and further treatment improvements requires an understanding of the dynamics of glucose regulation and pharmacokinetics of insulin. A number of models of glucose regulation have been proposed \citep{makr06}. One of these, the \emph{Bergman Minimal Model} (\cite{berg05, good15, kand09}), is a non-linear continuous-time model for glucose regulation. The model comprises a set of first order linear ordinary differential equations which govern the concentration and effectiveness of insulin:
\begin{align*}
\frac{d}{dt} I_{sc}(t) &= -\frac{1}{\tau_1} I_{sc}(t) + \frac{1}{\tau_1} \frac{ ID (t)}{C_l} \\
\frac{d}{dt} I_p (t) &= -\frac{1}{\tau_2} I_p (t) + \frac{1}{\tau_2} I_{sc} \\
\frac{d}{dt} I_{eff}(t) &= - p_2 I_{eff}(t) + p_2 S_I I_p (t)
\end{align*}%
\newline \quad \newline
and a non-linear ordinary differential equation which governs the plasma glucose concentration $g(t)$:
\begin{align*}
\frac{dg}{dt} = -g(t) \cdot (I_{eff} (t) + G) + r(t) + E
\end{align*}%
where:
\begin{itemize}
\item{$ID(t), I_{sc}(t), I_{p}(t)$ and $I_{eff}(t)$ -- are the delivery, subcutaneuos concentration, plasma concentration and insulin effectiveness, respectively. }
\item{$\tau_1$ and $\tau_2$ -- are time constants.}
\item{$C_l, S_I$ and $p_2$ -- are the clearance rate, insulin sensitivity and the insulin motility \citep{roy07}.}
\item{$g(t)$ -- is the plasma glucose concentration.}
\item{$E$ and $G$ -- are the endogenous glucose production and the effect of glucose on the uptake of plasma glucose and the suppression of endogenous glucose production, respectively.}
\item{$r(t)$ -- is the glucose absorption from meals.}
\end{itemize}
A variety of physiological values for the above are derived from \citep{kand09} and given in Table 1 of \citep{good15}. For notational convenience, we rewrite the Bergman minimal model as the following system of differential equations:
\begin{align}
\label{eq:eqs}
\begin{split}
\dot z &= -d z + dk u \\
\dot y &= -c y + cz \\
\dot x &= -ax + aby\\
\dot g &= - hg + w
\end{split}
\end{align}%
where all variables and constants are positive, $u(t) = ID(t)$ is the input function,
\begin{align}
\label{eq:wandh}
\begin{split}
h = x + G\\
w = r + E
\end{split}
\end{align}%
and the function $r$ is a given bounded function.
We develop necessary and sufficient conditions, given in Theorem \ref{thm:cream}, for the glucose response $g(t)$ to a pulse input function, $u(t)$, to be minimised. Additionally, these conditions give a non-linear version of the fundamental control limitation explored in Theorem 2 of \cite{medi15}, for the specific case of a single input pulse. In Figure \ref{fig:opteg} we show an example of the glucose responses to two pulse inputs delivered at times $t'_2 > t'_1$. The response which satisfies the conditions of Theorem \ref{thm:cream} has a lower maximum glucose concentration for the system whilst still maintaining the glucose concentration above a specified minimum concentration.
\begin{figure}[H]
\begin{center}
\resizebox{0.35\textwidth}{!}{%
\begin{tikzpicture}
\draw[.] (-2.5,3.73) -- (-2.5,3.73) node[above] {$t'_1$};
\draw[.] (-1,3.73) -- (-1,3.73) node[above] {$t'_{2}$};
\draw[-, dashed] (-2.5,2) -- (7,2) node[below] {$g(\infty)$};
\draw[-, dashed,orange] (-2.5,3.73) -- (7,3.73) node[below,black] {$\gamma_1$} ;
\draw[-, dashed,green] (-2.5,2.75) -- (7,2.75) node[below,black] {$\gamma_2$};
\draw[-, dashed,purple] (-2.5,0.28) -- (7,0.28) node[below,black] {$\lambda$};
\draw[scale=1,domain= -2.5:7,smooth,variable=\x, blue] plot ({\x},{2*(1.05*exp(-(\x)^2) - 1.6*exp(-(\x-0.61)^2))+ exp(-(\x-2)^2)+2});
\draw[scale=1,domain= -2.5:7,smooth,variable=\x, brown] plot ({\x},{1.75*(exp(-(\x-2)^2) - exp(-(\x)^2))+ 2});
\end{tikzpicture}}
\caption{Glucose responses given a fixed function $w$ and two different injection times, where $t'_1$ corresponds to the injection time of the brown response and $t'_2$ the injection time of the blue response.}
\label{fig:opteg}
\end{center}
\end{figure}
\section{Outline and Notation}
In Section \ref{sec:prelim} we state our assumptions and the constraints on the system. Additionally we prove some facts about the system and develop tools necessary for the subsequent Sections. In Section \ref{sec:opt} we prove necessary and sufficient conditions on the plasma glucose response to inputs $w$ and $u$ for the input $u$, comprising a single pulse, to be optimal. Finally, in Section \ref{sec:mult} we develop a sufficient condition for an input function, comprising multiple pulses, to be optimal.
\begin{table}[htbp]\caption{Notation}
\centering
\begin{tabular}{r c p{5cm}}
\toprule
$\overline u, \hat u$ and $\hat U (\lambda)$ & -- & the basal input, the magnitude of the bolus input and an expression for $\hat u$ given in \eqref{eq:bolusbound}\\
$w,h$ and $g$ & -- & the bounded functions, see \eqref{eq:wandh}, and plasma glucose concentration, respectively.\\
$\lambda$ and $\gamma$ & -- & the global minimum glucose concentration and the global maximum glucose concentration.\\
$t', t_{\max}, t_{\min}$ and $\tau$ & -- & the delivery time, a time when the glucose concentration is at its global maximum, a time when the glucose concentration is at its minimum and the duration of the interval over which the bolus is delivered, respectively.\\
$\mathcal{Y} (t)$ and $Y(t)$ & -- & the response of $x$ to the input $\hat u = 0$ and $\overline u =1$ and $\hat u =1$ and $\overline u =0$ respectively.\\
$g(\infty)$ & -- & the steady-state glucose concentration $g(\infty) := \lim_{t \to \infty} g(t)$.\\
$g(h(u),w)$ & -- & the reponse of $g$ to the functions $h$ and $w$, where $h(u)$ is the response of $h$ to the input $u$.\\
\bottomrule
\end{tabular}
\label{tab:notation}
\end{table}
\section{Preliminaries, Assumptions and Constraints}
\label{sec:prelim}
\subsection*{Assumptions}
Throughout we impose the following initial conditions: $z(0) = y(0) = k u(0)$, $x(0) = bk u(0)$ and $g(0) > 0$. We assume the function $w$ is positive and bounded. We also assume the input $u(t)$ is positive and bounded and of the form:
\begin{equation}
\label{eq:u}
u(t) = \bar{u}+ \hat{u} \chi_A (t)
\end{equation}
where the constant $\bar u$ is the \emph{basal} input, $\hat u$ is the magnitude of the \emph{bolus} input applied at some time $t'$, known as the \emph{delivery time}. The bolus input is held constant over the interval $A=[t',t' + \tau]$ and $\chi_A$ is the characteristic function of the interval $A$. The boundedness and positivity of $u(t)$ implies that $h$, given by \eqref{eq:eqs} and \eqref{eq:wandh}, is a continuous, positive and bounded function. We desire that there exist $\lambda > 0$ such that $g(t) \geq \lambda$ for all $t$. This is achieved if $\lambda$ is a global minimum of $g(t)$. We denote by $t_{min} \in \mathbb{R}_+$ a point such that $g(t_{min}) = \lambda$. Note, from \eqref{eq:eqs} that by setting $\dot g = 0$ at $t_{\min}$, we have $w(t_{\min} ) = \lambda h(t_{\min})$ at such $t_{\min}$.
\begin{defn}[Proper Input]
For some $\lambda \leq g(0)$, an input function, $u(t)$, is \emph{proper}, if there exists $t_{\min}$ such that $g(h(u(t_{\min})),w) = \lambda$, $g(t) \geq \lambda$ for all $t$.
\end{defn}
The existence of a proper input, of the form \eqref{eq:u}, is established in Theorem \ref{thm:bolus}. Finally, unless otherwise stated we assume that $t_{\max} := \argmax_t {g(t)} < \infty$. The maximal time $t_{\max}$ exists as shown in Corollary \ref{cor:gammain}.
\subsection*{Bounds and System Properties}
\begin{lem}[Bounds]
\label{lem:bounds}
Suppose $h$ and $w$ are bounded positive real-valued functionals, and $g$ is as in \eqref{eq:eqs}. Then there exist $\Gamma, \Lambda \in \mathbb{R}_+$ and constants $c_1$ and $c_2$ depending on the initial condition such that $\Gamma \geq \Lambda$ and:
\[
c_1 \exp\l (-\int h \r ) + \Lambda \leq g(t) \leq c_2 \exp\l (-\int h \r ) + \Gamma
\]
\end{lem}
\begin{pf}
A solution for $g$ is given by:
\[
g(t) = \exp\l (-\int h \r ) \l ( c_3 + \int w \exp\l (\int h \r )\r)
\]%
where $c_3$ is the value of $g$ at the lower extreme of integration. Choose $\Gamma \in \mathbb{R}_+$ such that $w \leq \Gamma h$. Note such $\Gamma$ always exists since $w$ and $h$ are bounded positive functions. We obtain:
\begin{align*}
g(t)& = \exp\l (-\int h \r ) \l ( c_3 + \int w \exp\l (\int h \r )\r) \\
& \leq \exp\l (-\int h \r ) \l ( c_3 + \Gamma \int h \exp\l (\int h \r )\r) \\
& = c_2 \exp\l (-\int h \r ) + \Gamma
\end{align*}
where $c_2 = c_3 -\Gamma$. Similarly, for $\Lambda$ such that $\Lambda h \leq w$, the lower bound on $g(t)$ is obtained and $c_1 = c_3 - \Lambda$.
\qed \end{pf}
\begin{rem}
\label{rem:upplow}
The bounds, $\Gamma \geq \sup_t\l \{ \frac{w(t)}{h(t)} \r \}$ and $\Lambda \leq \inf_t \l \{ \frac{w(t)}{h(t)} \r \}$ in Lemma \ref{lem:bounds}, may be improved at any $t$ by taking a finite ordered partition $\mathcal{P} := \{t_0, \cdots, t_n, t\}$ of the interval $[0,t]$ and defining $\Gamma_i$ and $\Lambda_i$ such that $\Gamma_i h(t) \geq w(t)$ and $\Lambda_i h(t) \leq w(t)$ for all $t\in[t_i , t_{i+1}]$.
\end{rem}
\begin{lem}
\label{lem:tend}
Let $u(t)$ be as in \eqref{eq:u} and choose $\lambda \leq g(0)$. Then:
\begin{enumerate}
\item{ The response $x(t)$ is separable into its basal and bolus responses i.e.
\begin{equation}
\label{eq:x}
x(t) = x(\bar u) + x(\hat u) := \bar u \mathcal{Y}(t) + \hat u Y(t)
\end{equation}
where $\mathcal{Y}(t) $ is the response of $x$ to the input with $\bar u = 1$, $\hat u =0$ and $Y(t)$ is the response of $x$ to the input~\eqref{eq:u} with $\bar u = 0$, $\hat u=1$.}
\item{Under the assumed initial conditions, $x(\overline u) = bk\overline u$ for all $t$ i.e. $\mathcal{Y} (t) = bk$. Furthermore, if $A$ is bounded. Then:
\[
\lim_{t \to \infty} x(t) = bk \overline u
\]}
\item{if $A$ is empty and the basal input $\overline u \leq \frac{E - \lambda G}{\lambda bk}$ then $g(t) \geq \lambda$ for all $t$.}
\end{enumerate}
\end{lem}
\begin{pf}
\begin{proofpart}
This follows by linearity of the $(z,y,x)$-system in~\eqref{eq:eqs}.
\end{proofpart}
\begin{proofpart}
Follows from the solutions to the first order linear ordinary differential equations in \eqref{eq:eqs} and $\mathcal{Y}(t)= bk$ from the specified initial conditions.
\end{proofpart}
\begin{proofpart}
By Lemma \ref{lem:bounds} there exists $\Lambda \in (0, g(0)]$ satisfying $\Lambda h \leq w$, which guarantees that $g(t) \geq \Lambda$ for all $t$. As, from \eqref{eq:wandh}, $w(t) \geq E$ and $h(t) = h(\overline u) = x(\overline u) + G \leq \overline u bk +G$ by Lemma \ref{lem:tend}, such $\Lambda$ may be chosen to satisfy the inequality:
\[
\Lambda \leq \frac{E}{bk \overline u + G}
\]%
Therefore, to ensure that $\Lambda \geq \lambda$ we require:
\begin{align*}
\lambda \leq \frac{E}{bk \overline u + G}
\iff \overline{u} \leq \frac{E- \lambda G}{\lambda bk}
\end{align*}
\end{proofpart}
\hspace{8cm} \qed
\end{pf}
\begin{rem}
\label{rem:hu}
We define the \emph{steady-state} of $g$ to be $g(\infty):=\lim_{t \to \infty} g(t)$, when $\lim_{t \to \infty}Y(t) = 0$ and $\lim_{t \to \infty} w(t) = E$ i.e. it is the limit of the response of $g(t)$ when the only input is the constant input $\overline u$. The steady-state may be set to be any positive real number. Since, if $\overline u$ is constant then $\dot g \to 0$ as $t \to \infty$. Indeed, assuming that $A$ is bounded, setting:
\begin{align}
\label{eq:basal}
\overline u = \frac{1}{kb}\l (\frac{E}{g(\infty)} - G \r)
\end{align}%
gives the result, where $g(\infty) \geq \lambda$ is some specified value. In this case:
\[
h(\overline u) := x(\overline u) + G \leq \frac{E}{g(\infty)}
\]%
We note, the constant, $c_3$, from Lemma \ref{lem:bounds} corresponds to the initial plasma glucose concentration, $g(0)$. Henceforth, we fix $g(\infty) := g(0)$ and $\overline u$ to be as in~\eqref{eq:basal}.
\end{rem}
The nature of the system dynamics \eqref{eq:eqs}--\eqref{eq:wandh} and the positivity of the inputs induces a monotonic relationship between the insulin input and glucose concentration. This property is proven in the intermediate Lemma \ref{lem:forprooftowork}. Note, we define $u_1 (t) > u_2 (t)$ if there exists $s$ such that $u_1(s) > u_2(s)$ and $u_1 (t) \geq u_2(t)$ for all $t$.
\begin{lem}
\label{lem:forprooftowork}
Suppose $w$ is fixed. Then $g(t)$ is a strictly monotone function of the input $u(t)$.
\end{lem}
\begin{pf}
Fix $w$ and let $u_1$ and $u_2$ be two inputs with delivery time $t'$ such that $\hat u_1 < \hat u_2$. Denote by $h_1, g_1$ and $h_2$ and $g_2$ their respective responses. Since $h$ is a monotone function of the input $u$ we have that $h_1 < h_2$ for all $t > t'$.
A solution for $g(t)$ for $t \geq t'$ is given by:
\begin{align*}
g(t) & = g(t')\exp\l (-\int_{t'} ^t h(s) \, ds \r ) \\ & \quad + \int_{t'} ^t w(s) \exp\l (- \int_{s} ^t h(\xi) \, d \xi \r ) \, ds
\end{align*}%
As $g_1 (t') = g_2 (t')$, because the inputs are identical before $t'$, and:
\begin{align*}
\exp\l (-\int_{t'} ^t h_1 (s) \, ds \r ) > \exp\l (-\int_{t'} ^t h_2 (s) \, ds \r ) \\
\exp\l (- \int_{l} ^t h_1(\xi) \, d \xi \r ) \geq \exp\l (- \int_{l} ^t h_2(\xi) \, d \xi \r )
\end{align*}%
for all $t > t'$ and $l \leq t$, we have that $g_1 (t) > g_2(t)$ for all $t> t'$.
\qed \end{pf}
Theorem \ref{thm:bolus} proves the existence of a bolus input delivered at any $t'$ which achieves a specified minimum $\lambda > 0$ and thus proves the existence of proper inputs of the form \eqref{eq:u}.
\begin{thm}[Insulin Bolus]
\label{thm:bolus}
Suppose $u(t)$ is of the form \eqref{eq:u}. Fix $\tau$ and $t'$ -- the input time i.e. $A:= [t', t'+ \tau]$, choose $\lambda \in (0, g(t')]$ and suppose $\bar u$ is as in Remark \ref{rem:hu}. Then there exists $\hat u$ such that $u(t)$ is proper.
\end{thm}
\begin{pf}
Denote by $g(\hat u)$ the response of $g(t)$ to the input $u(t) := \overline u + \hat u \chi_{A}$. By Lemmas \ref{lem:bounds} and \ref{lem:forprooftowork}, there exist $\hat o$ and $\hat v$ such that $\min_{t\geq t'} g(\hat o) \geq \lambda$ and $\min_{t\geq t'} g(\hat v) \leq \lambda$.
Suppose $\min_{t\geq t'} g(\hat v) < \min_{t\geq t'} g(\hat o)$. We recursively define the sequences $\overline o := (\hat o _i)_{i=0} ^\infty$ and $\overline v := (\hat v _i)_{i=0} ^\infty$ by $\hat o _0 = \hat o$ and $\hat v_0 = \hat v$ and $\hat o_i$ the greatest element of the following finite ordered partition of the interval $[\hat o_{i-1}, \hat v_{i-1}]$:
\begin{align*}
{L}_i := &\l \{\hat o_{i-1}, \frac{ (n-1) \hat o_{i-1} + \hat v_{i-1}}{n}, \cdots, \r . \\ &\quad \cdots, \l. \frac{k_i \hat o_{i-1} + (n-k_i) \hat v_{i-1}}{n}, \cdots, \hat v_{i-1} \r \}
\end{align*}%
where $n \in \mathbb{N}$ is arbitary and $k_i \leq n$, such that the response:
\[
g \l (\frac{k_i \hat o_{i-1} + (n-k_i) \hat v_{i-1}}{n} \r ) \geq \lambda
\
for all $t \geq t'$. Similarly, $\hat v_i$ is defined to be the least element of $L_i$ such that, for all $t \geq t'$:
\[
g \l (\frac{k_j \hat o_{i-1} + (n-k_j) \hat v_{i-1}}{n} \r ) \leq \lambda
\
The sequence $\overline o$ is a monotone increasing sequence bounded above by $\hat v_i$ for all $\hat v_i \in \overline v$ and therefore has a limit $o$. Similarly, $\overline v$ is a monotone decreasing sequence bounded below by $\hat o_i$ for all $\hat o_i \in \overline o$ and thus has a limit $v$. It remains to show that these two limits are equal. If either sequence is eventually constant then both are constant and equal. As either sequence is constant only if $\min_{t \geq t'} g(t) = \lambda$, in which case, by construction of the sequences, both sequences would have the same value. Suppose, instead, for all $i$ that $\hat o_i < \hat v_i$. We see that if:
\[
\hat o _{i+1} = \frac{k_i \hat o_{i} + (n-k_{i})\hat v_i}{n}
\]%
Then, by Lemma \ref{lem:forprooftowork}, $\hat v_{i+1}$ must be the next element of $L_i$, that is:
\[
\hat v_{i+1} = \frac{(k_i-1) \hat o_{i} + (n-k_{i}+1)\hat v_i}{n}
\
Thus:
\begin{align*}
\hat v_{i+1} - \hat o_{i+1} &= \frac{1}{n} \l (\hat v_i - \hat o_i \r) \\
&\qquad \vdots\\
&= \frac{1}{n^{i+1}} \l (\hat v_0 - \hat o_0 \r)
\end{align*}%
i.e. $\lim_{i \to \infty} \l (\hat v_{i+1} - \hat o_{i+1} \r ) = 0$ i.e. $v = o$. Thus $\lambda \leq \min_{t' \geq t} g(v) = \min_{t' \geq t} g(o) \geq \lambda$. Setting $\hat u = v$ gives $\min_{t' \geq t} g(\hat u) = \lambda$ and $g(\hat u) \geq \lambda$ for all $t\geq t'$.
\qed \end{pf}
Corollary \ref{cor:bolus} provides an explicit expression for the magnitude of the bolus input which achieves the gloal minimum of the glucose concentration.
\begin{cor}[Insulin Bolus Bound]
\label{cor:bolus}
Fix $t'$ and choose $\lambda > 0$ and let $u= \overline{u} + \hat{u} \chi_A$, see \eqref{eq:u}, where $\bar u$ is as in Remark \ref{rem:hu}. Suppose the input $\hat u$ is as in Theorem \ref{thm:bolus}. Then the input satisfies, $\hat u \leq \hat U (\lambda)$, where:
\begin{align}
\label{eq:bolusbound}
\hat U (\lambda) := \l ( \frac{w(t_{min})}{\lambda} - G - x(\overline u, t_{\min})\r) \l( \frac{1}{Y(t_{min})} \r)
\end{align}
and $g(t) \geq \lambda$ for all $t$. In particular, if $\hat u = \hat U(\lambda)$ then $g(t_{\min}) = \lambda$.
\end{cor}
\begin{pf}
We have the following:
\begin{align}
\begin{split}
\label{eq:2}
\dot g(t_{min}) = & -g(t_{min}) h(t_{min})+w(t_{min}) = 0 \\ & \iff g(t_{min}) = \frac{w(t_{min})}{h(t_{min})}
\end{split}
\end{align
Suppose $g(t_{min}) < \lambda$. Then, from \eqref{eq:wandh}, \eqref{eq:x}, \eqref{eq:2}, and as $h(t) > 0$:
\begin{align*}
w(t_{min}) & < \lambda h(t_{min})= \lambda (x(\bar u, t_{\min}) + \hat u Y(t_{min}) +G) \\ & \iff \hat{u} > \hat{U}(\lambda).
\end{align*}
Thus, \eqref{eq:bolusbound} implies $g(t) \ge \lambda$ for all $t$ and the result follows.
\qed \end{pf}
Corollary \ref{cor:gammain} proves the existence of a finite global maximum for glucose concentration responses to proper inputs.
\begin{cor}[An Upper Bound]
\label{cor:gammain}
Choose $\lambda \leq g(0)$. Suppose $u(t)$ is proper. Then there exists $t_{\max} \in \mathbb{R}_+ ^* := \mathbb{R}_+ \cup \{\infty \}$ such that $g(t) \leq g(t_{\max}) =: \gamma$ for all $t$ and $\gamma = \lambda$ if and only if $g(t) = \lambda$ for all $t$. Furthermore, if $t_{\max} < \infty$. Then:
\begin{align}
\label{eq:maxgam}
\gamma = \frac{w(t_{\max})}{h({t_{\max}})} = \frac{\alpha_1 \lambda}{\alpha_2 + \alpha_3 \lambda}
\end{align}
where $\alpha_1 := w(t_{\max})$, $\alpha_2 := w(t_{\min}) \l (\frac{Y(t_{\max})}{Y(t_{\min})} \r)$, $\alpha_3 := \l (G+ x(\overline u, t_{\max}) \r) \l( 1 - \frac{Y(t_{\max})}{Y(t_{\min})} + x(\overline u, t_{\max}) - x(\overline u, t_{\min})\r)$.
\end{cor}
\begin{pf}
If there is $s \in \{t : \dot g(t) = 0\}$ such that $g(s) \geq g(t)$ for all $t$. Then $t_{\max} = s$. Otherwise $g(t)$ must increase as $t \to \infty$. We may take a monotone increasing sequence $(g(t_i))_{i=0} ^\infty$, where $t_0 \geq 0$ and $g(t_k) \in (g(t_i))_{i=0} ^\infty$ only if $g(t_k) > g(t)$ for all $t \in [0, t_k)$ i.e. $g(t_k)$ is the \emph{peak} of the function $g$. By Lemma \ref{lem:bounds} the sequence, $(g(t_i))_{i=0} ^\infty$, is bounded above and thus converges to $\overline g \leq \Gamma$. By construction $g(t) \leq \overline g$ for all $t$. We see that $\gamma = \lambda$ if and only if $g(t) = \lambda$ for all $t$ follows by definition of $\gamma$.
Suppose $t_{\max} < \infty$. Then \eqref{eq:maxgam} follows from rearranging the differential equation $\dot g = -gh + w$ evaluated at $t_{\max}$ and substituing in the formula for $\hat u$ given by \eqref{eq:bolusbound}.
\qed \end{pf}
Corollary \ref{cor:seesaw} shows that, the higher the minimum glucose concentration, the higher the peak glucose concentration.
\begin{cor}
\label{cor:seesaw}
Choose $\lambda < \lambda' \leq g(0)$. Let $u(t)$ and $u'(t)$ be inputs, of the form \eqref{eq:u}, with common delivery time $t'$, which are proper for $\lambda$ and $\lambda'$ respectively. Then $\gamma < \gamma'$.
\end{cor}
\begin{pf}
Note $\gamma' \geq \lambda'$ and $\gamma > \lambda$. Denote by $t_{min}$ and $t'_{min}$ the times at which $g(t) = \lambda$ and $g'(t) =\lambda'$ respectively. If for example $h' > h$ for some $t>t'$. Then $h' > h$ for all $t > t'$. This is because $u(t)$ and $u'(t)$ are of the form \eqref{eq:u} and have common delivery time. Suppose that $h' \geq h$ for all $t \geq t'$. This implies that $g' \leq g$ for all $t \geq t'$. In particular at $t_{\min} \ge t'$ we have that $\lambda = g(t_{\min}) \geq g'(t_{\min}) \geq \lambda'$ contradicting $\lambda < \lambda'$. Thus $h > h'$ for all $t > t'$. Finally, as $h$ is a monontone function of the input $u(t)$, we see that $\hat u > \hat u'$. This implies that $g(t) < g'(t)$ for all $ t > t'$.
\qed \end{pf}
\section{Optimal Inputs}
\label{sec:opt}
In this section we give necessary and sufficient conditions on the delivery time of a proper input such that the glucose response is optimal i.e. the maximum glucose concentration $\gamma$ is minimised. Throughout, we fix the length, $\tau$, of the interval over which the bolus is delivered. This ensures that the function $Y(t - t')$ in \eqref{eq:x} is invariant under translation by $t'$ -- the delivery time. Under these conditions we establish a property of the function $h$, defined in \eqref{eq:wandh}, in Lemma \ref{lem:morelemmas}.
\begin{lem}
\label{lem:morelemmas}
For any two distinct proper inputs $u_1$ and $u_2$, delivered at times $t'_1$ and $t'_2$ respectively, with responses $h_1$ and $h_2$, there exists $t_i$ such that either:
\begin{align}
\label{eq:firstcase}
\l \{ \begin{array}{lc}
h_1 > h_2,& t \in (\min \{t'_1, t'_2\}, t_i) \\
h_1 = h_2,& t = t_i \\
h_1 < h_2,& t > t_i
\end{array}\r .
\end{align}%
Or:
\begin{align}
\label{eq:secondcase}
\l \{ \begin{array}{lc}
h_1 < h_2,& t \in (\min \{t'_1, t'_2\}, t_i) \\
h_1 = h_2,& t = t_i \\
h_1 > h_2,& t > t_i
\end{array}\r .
\end{align}%
\end{lem}
\begin{pf}
Indeed, if $h_2 (t) > h_1(t)$ or $h_2 (t) < h_1(t)$ for all $t > \min \{t'_1, t'_2\}$ then, by Lemma \ref{lem:forprooftowork} there would exist $t$ such that $g_2 (t) < \lambda$ or $g_1 (t) < \lambda$. Implying that either $\hat u _1$ or $\hat u_2$ are not proper.
\qed \end{pf}
In the subsequent proofs of Lemmas \ref{lem:anotherone}, \ref{lem:optimaltime} and \ref{lem:optglu}, we only present the first case, \eqref{eq:firstcase}, as the other case, \eqref{eq:secondcase}, follows by a similar argument.
\begin{rem}
\label{rem:labelit}
For fixed $w$ the functions $h_1$ and $h_2$ satisfy \eqref{eq:firstcase} only if $t'_1 < t'_2$ i.e. only if $\hat u_1$ is delivered before $\hat u_2$ as the function $Y(t)$ is independent of the magnitude $\hat u$.
\end{rem}
We define the optimal, minimised, glucose response as the input control strategy that ensures that the maximum glucose concentration is minimised given the control and system limitations. Formally:
\begin{defn}[Minimised Response]
We say the response $g(t) = g(w,h)$ is \emph{minimised} if $\max_t g(w,h') > \max_t g(w,h)$, for all $h' \neq h$,
\end{defn}
\begin{defn}[Optimal Delivery Time]
We say a delivery time $t'$ of a bolus input $\hat u \chi_{[t',t' + \tau]}$, where $\hat u$ is given by (\ref{eq:bolusbound}), is \emph{optimal} if the response $g(t)$ is minimised.
\end{defn}
\begin{lem}
\label{lem:anothernotherone}
Suppose $w(t)$ is a continuous and bounded positive functional and $h(t)$ is the response to a proper input of the form \eqref{eq:u}. Then, for any $\varepsilon \in [0, g(0) - \lambda)$, there exist at most finitely many $t$ such that the response $g(t) < \lambda + \varepsilon$ and $\dot g(t) = 0$.
\end{lem}
\begin{pf}
This follows as $w(t)$ is bounded below and the response $Y(t)$ is continuous and converges to $0$.
\qed \end{pf}
\begin{lem}
\label{lem:anotherone}
Suppose $g (h_1, w) = g_1(t)$ is a response to a proper input with bolus $\hat u_1$ delivered at time $t'_1$ such that there is a unique minimum $t_{1,\min}$ i.e. $g_1 (t_{1,\min}) = \lambda$ and $g_1 (t) > \lambda$ for all $t \neq t_{1,\min}$. Then there exists a proper bolus input $\hat u_2$ delivered at $t'_2$ and a time $t_i \geq \max\{t'_1,t'_2\}$ at which $h_1(t_i) = h_2(t_i)$ such that the response $g_2(h_2, w) := g_2(t)$ attains its minimum at $\lambda$ and satisfies:
\[
\l \{ \begin{array}{lc}
g_1 < g_2,& t < t_g \\
g_1 = g_2,& t = t_g \\
g_1 > g_2,& t > t_g
\end{array}\r .
\]%
for some time $t_g \in [t_i, t_{1, \max})$, where $t_{1,\max} := \min \{s > t_{1,\min} : g_1(s) \geq g_1(t) \, \forall \, $t$\}$.
\end{lem}
\begin{pf}
As $g$ is a continuous function of $h$, for all $\varepsilon > 0$ we may find $\delta > 0$ such that $|h_1 - h_2| < \delta$ implies $|g_2 - g_1| < \varepsilon$, for all $t$. Such $h_2$ exists and is of the form $x(t) + G$, where $x(t)$ is as in \eqref{eq:x}, as $x(t)$ is a continuous function of the input $u(t)$. Thus for all $\delta > 0$ there exists $\delta' > 0$ such that $\max |u_1 - u_2 | < \delta'$ implies $|h_1 - h_2| < \delta$. Furthermore, by Theorem \ref{thm:bolus}, we may assume $u_2$ is proper and of the form \eqref{eq:u}. Thus, by Lemma \ref{lem:morelemmas} and Remark \ref{rem:labelit}, $u_2$ may be chosen such that there exists $t_i$ for which:
\[ \l \{ \begin{array}{lc}
h_1 > h_2,& t \in (t_1 ', t_i) \\
h_1 = h_2,& t = t_i \\
h_1 < h_2,& t > t_i
\end{array}\r .
\]%
Define $\overline \lambda := \min\{g_1(t) : t \neq t_{1,min} \wedge \dot g_1 (t) = 0\}$ or $g(0)$ if this minimum does not exist. Such $\overline \lambda > \lambda$ exists by Lemma \ref{lem:anothernotherone} and $\overline \lambda \leq g(0)$, by construction.
Choosing $\varepsilon < \min\{\gamma_1 - g(0), \overline \lambda - \lambda\}$ or $\varepsilon < \overline \lambda - \lambda$ if $\gamma_1 - g(0)= 0$, where $\gamma_1 := \max(g_1(t))$, implies that $t_{2,min} < t_{1,\max}$. Note that $t_{2, \min} > t_i$ as if it were not there would exist $t$ such that $g_1 (t) < \lambda$ since $g_1 < g_2$ for all $t \in (t_1', t_i)$. Also, $g_2(t_{2, \min}) = \lambda < g_1(t_{2,\min})$, by assumption. By the Intermediate Value Theorem there is a $t_g \in [t_i, t_{1, \max} )$ such that $g_1 (t_g) = g_2 (t_g)$. As $h_2 > h_1$ for all $t > t_i$ we see that $g_2 (t) < g_1 (t)$ for all $t > t_g$.
\qed \end{pf}
\begin{rem}
Similarly to Lemma \ref{lem:anotherone}, we may show that if $t_{1, \max} < t_{1, \min}$ then there exists $h_2$ with response $g_2$ and $t_g \in [t_i,t_{1, \min})$ such that:
\[
\l \{ \begin{array}{lc}
g_1 > g_2,& t < t_g \\
g_1 = g_2,& t = t_g \\
g_1 < g_2,& t > t_g
\end{array}\r .
\]%
\end{rem}
\begin{lem}[Single Minimum]
\label{lem:optimaltime}
Suppose $g(h,w) = g(t)$ is a response, to a proper input $u$, for which there is a unique $t_{\min}$ such that $g(t_{\min}) = \lambda$. Then $g(t)$ is minimised if and only if $\max_{t < t_{\min}} g(t) = \max_{t > t_{\min}} g(t)$.
\end{lem}
\begin{pf}
Without loss of generality, suppose that for some $h_1$ we have that:
\[
\overline \gamma_1 := \max_{t < t_{1,\min}} g_1(t) < \max_{t > t_{1,\min}} g_1 (t) =: \gamma_1
\]%
where $g_1 (t) := g(h_1,w)$. The existence of a unique $t_{1,\min}$ implies there is a proper input with non-zero bolus $\hat u_1$ delivered at some time $t'_1$. Define $\bar t _1 := \argmax_{t < t_{1,\min}} g_1(t)$ and $\hat t_1 := \argmax_{t > t_{1,\min}} g_1 (t)$. By Lemma \ref{lem:anotherone} there exists $h_2$ such that:
\[
\l \{ \begin{array}{lc}
g_1 < g_2,& t < t_g \\
g_1 = g_2,& t = t_g \\
g_1 > g_2,& t > t_g
\end{array}\r .
\]%
Proceeding as in the proof of Lemma \ref{lem:anotherone}, choosing $\varepsilon < \min\{\gamma_1 - g(\infty), \overline \lambda - \lambda, \gamma_1 - \overline \gamma_1\}$ implies that $g_2 < \gamma_1$ for all $t$, since the choice of $\varepsilon$ ensures that $t_g < \hat t_{1}$.
Now, suppose that $\max_{t < t_{1,\min}} g_1(t) = \max_{t > t_{1,\min}} g_1 (t) := \gamma_1$ and $g_1$ is not minimised. We observe that for $g_1$ not to be minimised there must exist $h_2$ with response $g_2$ such that $g_2 (\bar t _1) < \gamma_1$ and $g_2 (\hat t_1 ) < \gamma_1$, which implies $t_g < \overline t_1$. So $g_2 (t_{1,\min}) < g_1 (t_{1,\min}) = \lambda$. This condradicts the constraint on $g_2$. Hence no such $h_2$ exists.
\qed \end{pf}
\begin{figure}[H]
\begin{center}
\resizebox{0.35\textwidth}{!}{%
\begin{tikzpicture}
\draw[.] (-0.8,3.1) -- (-0.8,3.1) node[above] {$\overline{t}_1$};
\draw[.] (3,3.1) -- (3,3.1) node[above] {$\hat t_{1}$};
\draw[.] (0.8,0.28) -- (0.8,0.28) node[below] {$t_{1,\min}$};
\draw[-, dashed] (-2.5,2) -- (7,2) node[below] {$g(\infty)$};
\draw[-, dashed,orange] (-2.5,3.0) -- (7,3.0) node[right,black] {$\gamma_1$} ;
\draw[-, dashed,green] (-2.5,2.35) -- (7,2.35) node[right,black] {$\overline \gamma_1$};
\draw[-, dashed,purple] (-2.5,0.28) -- (7,0.28) node[below,black] {$\lambda$};
\draw[scale=1,domain= -2.5:7,smooth,variable=\x, blue] plot ({\x},{2*(1.05*exp(-(\x)^2) - 1.65*exp(-(\x-0.4)^2))+ exp(-(\x-3)^2)+2});
\end{tikzpicture}}
\caption{Glucose response for functions $w$ and $h_1$ showing two unequal maxima about a single minimum, where $\overline \gamma_1 := \max_{t < t_{1,\min}} g_1(t)$ and $\gamma_1 := \max_{t > t_{1,\min}} g_1(t)$.}
\label{fig:unequalglu}
\end{center}
\end{figure}
\begin{lem}[Single Maximum]
\label{lem:optglu}
Suppose $g(h,w) = g(t)$ is a response, to a proper input, for which there exist distinct $\underline{t}$ and $\underaccent{\check} t$ such that $g(\underline{t}) = g(\underaccent{\check} t) = \lambda$ but a single $t_{\max} := \argmax\{g(t)\}$. Then $g(t)$ is minimised if and only if $\argmax_t \{g(t)\} \in (\underline{t},\underaccent{\check} t)$.
\end{lem}
\vspace{10pt}
\begin{pf}
Suppose, for some $h_1$, that:
\[
\min_{t < t_{1, \max}} \{ g_1 (t) \} = \min_{t > t_{1, \max}} \{ g_1(t) \} = \lambda
\] Define $\underline{t}_1 := \min \{t < t_{1, \max} : g_1 (t) = \lambda \}$ and ${\underaccent{\check}{t}}_1 := \max \{t > t_{1, \max} : g_1(t) = \lambda \}$. Suppose $h_2 \neq h_1$ is a response to a proper input $u_2$ as in Lemma \ref{lem:anotherone}. As $g_2 \geq \lambda$ for all $t$, the crossing time $t_g$ satisfies either $t_g < \underline t_1$ or $t_g > \underaccent{\check}{t}_1$. In both cases $g_2(t) > g_1(t)$ for all $t \in [\underline t_1, \underaccent{\check}{t}_1]$ which implies $\max\{g_2(t)\} > \max\{g_1 (t) \}$.
Now, suppose that there are at least two distinct $t$ such that $g_1(t) = \lambda$ and, without loss of generality, that:
\[
\min_{t < t_{1, \max}} \{ g_1(t)\} < \min_{t > t_{1, \max}} \{g_1(t)\}
\]%
As in Lemma \ref{lem:anotherone} there exists $h_2$ and $t_i$ with response $g_2$ such that:
\[
\l \{ \begin{array}{lc}
g_1 < g_2,& t < t_g \\
g_1 = g_2,& t = t_g \\
g_1 > g_2,& t > t_g
\end{array}\r .
\]%
where $t_g \in [t_i, t_{1, \max})$. In this case, proceeding as in Lemma \ref{lem:optimaltime}, there exists $h_2$ such that $g_2 (t) < \gamma_1$, where $\gamma_1 := \max\{g_1(t)\}$.
\qed \end{pf}
Lemmas \ref{lem:optimaltime} and \ref{lem:optglu} show that for a given $w$ and input of the form $u(t) := \overline u + \hat u \chi_{[t', t'+ \tau]}$, see \eqref{eq:u}, the maximum glucose concentration $g(t_{\max})$ is minimised if and only if the maximum occurs between two minima, where $g(t) = \lambda$, or the minimum between to equal maxima. We state this formally in Theorem \ref{thm:cream}.
\begin{thm}[Multiple Extrema]
\label{thm:cream}
Suppose $g(h,w) = g(t)$ is a response to an input of the form \eqref{eq:u}. Then $g(t)$ is minimised if and only if there exists $t_{\min}$ such that $g(t_{\min}) = \lambda$ and either: $\argmax_t \{g(t)\} \in (\underline{t},\underaccent{\check} t)$ where $\underline t, \underaccent{\check} t \in \argmin\{g(t)\}$ or there is $\underaccent{\check} t \in \argmin\{g(t)\}$ such that $\max_{t< \underaccent{\check} t} \{g(t) \}= \max_{t< \underaccent{\check} t} \{g(t) \}$.
\end{thm}
\section{Multiple Pulses}
\label{sec:mult}
The optimality conditions given in Theorem \ref{thm:cream} apply to an input $u(t)$ with only a single bolus input delivered at some $t'\in \mathbb{R}_+$. We provide a sufficient condition for optimality of an input with finitely many bolus inputs i.e. we consider inputs of the form:
\begin{align}
\label{eq:genu}
u(t,T) := \overline u + \sum_{i=0} ^N \hat u_i \chi_{[t'_i, t'_i + \tau]}
\end{align}%
where $T := (t_i)_{i=0} ^N$ is a finite sequence of delivery times and the magnitude $\hat u_i$ of each bolus input is given sequentially by Theorem \ref{thm:bolus}\footnote{Theorem \ref{thm:bolus} does not explicitly need the input to have the form~\eqref{eq:u}}, if there exists $t > t'_i$ such that $g(t) > g(0)$ or if $i = 0$. Otherwise $\hat u_i := 0$. This ensures that each bolus input, and therefore the input, $u(t,T)$, is proper and avoids unnecessary inputs. Thus we may assume that each $\hat u_i > 0 $. By abuse of notation, we denote $u(t,N) = u(t,T)$, where $N$ is the length of the sequence $T$.
\begin{lem}
\label{lem:extend}
Suppose $u_1(t,N)$ and $u_2(t,N)$ are distinct inputs of the form \eqref{eq:genu}. Then there exist at most $2N$, $t_{g,i}$ such that $g_1 (t_{g,i}) = g_2(t_{g,i})$ and for which one of the following, with either direction or order of the inequalities, is satisfied:
\[
\l \{ \begin{array}{lc}
g_1 < g_2,& t \in (t_{g,i-1}, t_{g,i}) \\
g_1 = g_2,& t = t_{g,i} \\
g_1 \geq g_2,& t \in (t_{g,i}, t_{g,i+1})
\end{array}\r .
\]%
Or:
\[
\l \{ \begin{array}{lc}
g_1 < g_2,& t \in (t_{g,i-1}, t_{g,i}) \\
g_1 = g_2,& t = t_{g,i} \\
g_1 < g_2,& t \in (t_{g,i}, t_{g,i+1})
\end{array}\r .
\]%
where $t_{g,N+1} := \infty$. Furthermore, $t_{g,0} := \inf\{t : g_1 (t) \neq g_2(t) \}$ must exist.
\end{lem}
\begin{pf}
This follows as $g$ is a monotonic function of $h$ which is a monotonic function of the input $u$.
\qed \end{pf}
\begin{defn}
Let $u_1(t,N)$ and $u_2(t,N)$ be two distinct inputs of the form \eqref{eq:genu}. The points $\{t_{g,0}, t_{g,1}, \cdots, t_{g,2N -1} \}$ defined in Lemma \ref{lem:extend} are the \emph{intersection points} of the responses $g(h(u_1(t,N)),w) = g_1(t)$ and $g(h(u_2(t,N)),w) = g_2(t)$.
\end{defn}
\begin{thm}
\label{thm:mult}
Suppose $u(t,N)$ is proper, the sum of the number of global maxima and minima, of the response $g(h(u(t,N)),w)$, is $2N+1$ and these minima and maxima are interlaced. Then $u(t,N)$ is optimal, over inputs of the same $N$.
\end{thm}
\begin{pf}
Suppose $g_1$ is a response as in the statement of the Theorem and that there is an input $u_2(t,N) \neq u_1(t,N)$ such that $\max \{ g_2 (t) \} < \max \{ g_1(t)\}$. Additionally, assume that $t_{g,i} \neq t_{1,i,\min}$, where $t_{1,i,\min}$ is the $i^{\text{th}}$ minimum of $g_1$ i.e. the intersection points of $g_1$ and $g_2$ do not occur at the minima of $g_1$. Under this assumption we see that $\max \{ g_2 (t) \} < \max \{ g_1(t)\}$ if and only if each $t_{1, i , \max} \in (t_{g,i}, t_{g_, i+1})$, where $t_{g,i}$ and $t_{g_,i+1}$ are two subsequent intersection points of the reponses $g_1$ and $g_2$ and $t_{1,i,\max}$ is the $i^{\text{th}}$ maximum of $g_1$. If this condition were not satisfied there would exist $t$ such that $g_2(t) < \lambda$ or $g_2 (t) > \gamma_1$. Thus we see to ensure that $g_2 < g_1$ at $n$ maxima followed by $n$ minima $g_1$ and $g_2$ must have $2n$ intersection points. If the maximum is not followed by a minimum then it requires one intersection point. By assumption, only the final maximum may not be followed by a minimum.
Suppose the first minimum occurs before the first maximum. Either $t_{g,0} < t_{1,1, \min}$ or $t_{g,0} \in (t_{1,1,\min}, t_{1,1,\max})$ i.e. $u_1 (t) = u_2(t)$ for all $t < t_{g,0} - \varepsilon$, $\varepsilon \in (0,t_{g,0})$. This reduces to the case where the first maximum of $g_1$ occurs before the first minimum of $g_1$with distinct inputs $v_1(t,N-1)$ and $v_2(t,N-1)$, of the form \eqref{eq:genu}.
Suppose $g_1$ has $N+1$ minima, and therefore $N$ maxima. By the above $t_{g,0} < t_{1,1,\min}$ thus from Lemma \ref{lem:extend} $2N -1$ intersection points remain but $2N$ intersection points are required as each maximum is followed by a minimum.
Conversely, suppose $g_1$ has $N+1$ maxima, and therefore $N$ minima, this implies $g_2$ must intersect $g_1$ at $2N+1$ points. But by Lemma \ref{lem:extend} there exist at most $2N$ intersection points.
Finally, suppose that $m$ of the $t_{g,i} = t_{1,i,\min}$. This would imply that at least one of the maxima require one fewer intersection points. Indeed, $n$ maxima followed by $n$ minima would require $2n-m$ intersection points. At such $t_{g,i}$ we have that $h_1 = h_2$ and $g_1 = g_2$. Each intersection point after which $g_2 < g_1$ corresponds to a pulse of $u_2(t,N)$. Thus each $t_{1,i,\max}$ corresponds to a pulse of $u_2(t,N)$. Either there are $N+1$ maxima of $g_1$, in which case $u_2 (t,N)$ does not have sufficiently many pulse inputs, or there are $N$ maxima and $N+1$ minima of $g_1$. In this case $g_1$ must start with a minimum. As above we may assume that $t_{g,0} < t_{1,1,\min}$, which reduces to the case of $N$ maxima followed by $N$ minima and only $N-1$ pulses of $u_2(t,N)$ which we may apply.
Hence, there exists no such response $g_2$ and therefore no input $u_2(t,N)$ exists. Thus $u_1 (t)$ is optimal.
\qed \end{pf}
\begin{rem}
The converse of Theorem \ref{thm:mult} does not hold for all $w$. As rapid changes in $w$ will outpace the response time of $h$. However the results may be applied over specified bounded intervals, each of which have different maxima, to find the optimal input for each interval.
\end{rem}
\begin{lem}
\label{lem:thelastone}
Suppose $a < b < \infty$. Then there exists $M \geq 1$ such that the minimised response $\max\{g(h_n, w)\} =\max\{g(h_{n+1}, w)\}$ for all $t \in [a,b]$ and for all $n \geq M$, where $h_n$ is the response to the input $u(t , n)$.
\end{lem}
\section{Example}
In the example presented in Figure \ref{fig:minmaxmin}, we consider the following values for the parameters in \eqref{eq:eqs} and \eqref{eq:wandh}:
$d=0.0204$, $k=497.5124$, $c=0.0213$, $a=0.0106$, $b=8.11 \times 10^{-4}$, $G=0.0032$, $E=1.3$, and $r(t) = 0.0018 f_1(t)$, where $f_1(t)$ is the response of the system of linear differential equations:
\[
\begin{pmatrix}
\dot f_1(t) \\
\dot f_2(t)
\end{pmatrix}
=
\l ( \frac{1}{47} \r)
\begin{pmatrix}
-1 & 1 \\
0 & -1
\end{pmatrix}
\begin{pmatrix}
f_1(t) \\
f_2(t)
\end{pmatrix}
+
\begin{pmatrix}
0 \\
\delta(t)
\end{pmatrix}
\]
where $\delta(t)$ is an impulse of magnitude $120$ applied at time $500$. We take the initial conditions to be as in Section \ref{sec:prelim} and set $g(\infty) = g(0) = 100 \mathrm{mgdl}^{-1} \, (5.2 \mathrm{mmolL}^{-1})$. The input is of the form \eqref{eq:u} with $\bar{u}$ computed as in \eqref{eq:basal} and duration $\tau=10$. The bolus magnitude $\hat u$ is computed as in Theorem~\ref{thm:bolus}, for the optimal injection time $t'$ that minimises $\gamma$, the global maximum of $g$. The minimum glucose concentration $\lambda$ is chosen to be $80 \mathrm{mgdl}^{-1} \, (4.4 \mathrm{mmolL}^{-1})$.
The first plot in Figure \ref{fig:minmaxmin} shows the plasma glucose response to the function, $w(t)$, shown in the second plot of the same Figure, and an optimal bolus input $\hat u$ delivered at time $445$. Two minima occur at times $500$ and $800$ bounding the unique maximum which occurs at time $574$. The final plot of Figure \ref{fig:minmaxmin} shows the maximum glucose concentration, $\gamma$, and the magnitude of a proper input bolus as a function of the input time $t'$. We see that $\gamma$ is minimised at the optimal input time $445$.
\begin{figure}[H]
\begin{flushleft}
\includegraphics[width=10cm, height=9.7cm, keepaspectratio]{MinMaxMin.pdf}
\end{flushleft}
\caption{The optimal glucose response to the functions $w(t)$ and $h(t)$, which are shown in the second plot, and the magnitude of a proper bolus and the maximum glucose concentration as a function of the delivery time $t'$, shown in the third plot.}
\label{fig:minmaxmin}
\end{figure}
\section{Conclusions}
Current research aims to generalise the presented results to any bounded input function $u(t)$. We are also interested in studing the effect of $\tau$ -- the length of the bolus delivery interval $A$, on the response $g(t)$.
We do not know whether similar results to those presented may be shown for other models of glucose metabolism. In particular, those which include other factors such as exercise \citep{roy07} or free fatty acid metabolism \citep{roy06}. Given the general nature of the proofs of the current results we believe it is likely that similar results may hold for other models.
Additionally, we aim to derive a formula for the maximum plasma glucose concentration which is independent of the times $t_{\min}$ and $t_{\max}$ and depends solely on $\lambda$ and the set $A$ i.e. to find $f: \mathbb{R}_+ \times A \to \mathbb{R}_+$ such that $\gamma = f(\lambda,A)$. This may allow us to extend the results of Corollary \ref{cor:seesaw} by specifying the rate at which the maximum concentration increases with respect to increases in the fixed minimum concentration $\lambda$.
Finally, we desire to prove that there is an optimal partition of $\mathbb{R}_+$ into intervals so that the converse of Theorem~\ref{thm:mult} holds over each interval and no other partition will produce a lower maximum glucose concentration. Such a result may follow by extending Lemma \ref{lem:anotherone} to cover inputs of the form \eqref{eq:genu}. This would also allow us, in conjuction with Lemma \ref{lem:thelastone} to specify the minimum number of pulses required to achieve the lowest possible maximum glucose concentration over some bounded interval.
|
\section{Introducton}
Unconventional superconductivity is driven by electron-electron interactions,
instead of electron-phonon couplings~\cite{Anderson}.
It occurs in a variety of strongly correlated electron systems, with the
iron-based superconductors
(FeSCs)
representing a prototype case
~\cite{Hosono, Johnston2011, Si2016,Wang_Lee2011,HosonoKuroki.2015,Hirschfeld2016}.
The field of
FeSC started with most of the efforts being directed toward
the iron pnictide class.
The normal state was found to be a bad metal,
with room-temperature resistivity reaching the Mott-Ioffe-Regel limit \cite{Johnston2011, Qazilbash:Nat_Phys_2009},
suggesting the importance of electron correlations \cite{Si_Abrahams:PRL_2008,Yin_Haule_Kotliar:Nat_Phys_2011}.
More recently, the focus has been shifted to iron selenide systems. The reasons are
manifold.
They
have
the highest
$T_c$~\cite{Xue.2012, Shen.2014},
they show
even stronger electron correlations,
and, as we discuss here, their superconductivity is highly unusual.
The puzzle
of the superconducting pairing state is highlighted
by the ``122" alkaline iron selenides.
These systems have
a
$T_c$
of about
$31$ K at ambient pressure.
They have only
electron Fermi pockets,
lacking the hole pockets that exist in the iron pnictides
at the center of the Brillouin Zone (BZ) \cite{Mou_et_al:PRL_2011,Zhang_et_al:PRL_2014,Wang_et_al:EPL_2011}.
Angle-resolved photoemission spectroscopy (ARPES)
experiments
show that the quasiparticle dispersion is fully gapped
on
all the parts of the
FS~\cite{Mou_et_al:PRL_2011,Zhang_et_al:PRL_2014,Wang_et_al:EPL_2011},
including a small electron Fermi
pocket at the center of the BZ \cite{Xu_et_al:PRB_2012,Wang_et_al:EPL_2012}.
This is compatible with the usual $s$-wave $A_{1g}$ pairing state,
but not with
the usual
$d$-wave $B_{1g}$ state
(which would produce nodes on the small electron Fermi pocket near the center of
the BZ).
On the other hand,
inelastic
neutron scattering experiments \cite{Park_etal:2011,Friemel_etal:2012}
observe a sharp resonance peak around the
wavevector $(\pi, \pi/2)$.
It is
consistent
with a pairing function that changes sign~\cite{Eschrig:Adv_Phys_2006}
between the two Fermi pockets at the edge of the BZ,
such as would occur in a $d$-wave $B_{1g}$ state, but
not in the usual $s$-wave $A_{1g}$ case.
In this work, we demonstrate how an orbital-selective pairing state, dubbed $s \tau_{3}$,
exhibits properties that are commonly associated with a
$d$-wave $B_{1g}$ state
\emph{or}
a $s$-wave $A_{1g}$ state. The key to the emergence of this superconducting state is the multiband nature
of the FeSCs.
This is associated with the multiplicity of $3d$ electron orbitals,
whose conceptual importance
follows the tradition wherein new physics develops out of extra degrees of freedom,
similar, for instance, to the way the so-called valley quantum number
in the electronic structure introduces new
topological properties~\cite{Schaibley_2016}.
It is important for the FeSCs that there are multiple orbitals at play in the neighborhood of the Fermi level.
Thus there is reason to expect that correlation effects will be different for different orbitals.
In fact, there is evidence for orbitally-selective Mott behavior
in the iron selenides ~\cite{MYi.2015, Yi_et_al:PRL_2013,Wang2014,Ding2014,Li2014}
and, thus, orbital selectivity is to be expected for pairing as well.
For strongly correlated superconductivity,
Cooper pairing
is naturally considered in an orbital basis
due to the
tendency of the electrons to avoid the dominating Coulomb repulsions.
Considering a basis
formed from all
five
$3d$-orbitals, the $s \tau_{3}$
state has an $s$-wave form factor, but
transforms
as a $d$-wave $B_{1g}$ state. As such, it represents an energetically-favored reconstruction of the
conventional $s$-wave and $d$-wave pairing states
when they are quasi-degenerate, due to
frustrated antiferromagnetic interactions \cite{Yu_Nat_Comm:2013}.
The pairing function incorporates
a matrix $\tau_3$ in the $3d_{xz}, 3d_{yz}$ subspace, which does not commute with the kinetic term of the Hamiltonian.
Consequently, in the band basis, it
must also have
a matrix structure,
which contains both
intraband and interband terms.
This allows
the intraband pairing component
to have
a $d$-wave sign change,
while the addition in quadrature
of the intraband
and interband pairing terms is nonzero everywhere on the FS.
Thereby,
the spin excitations show
a $(\pi, \pi/2)$ resonance
while
the quasiparticle excitations as measured by ARPES
are fully gapped on the Fermi surface.
\section{Result}
{\it Orbital selectivity in the normal state of iron selenides:}~~
In the normal state,
ARPES has provided evidence not only for the existence of the orbital degree of freedom but also for strong orbital-selective correlation effects
in the iron selenides.
These materials include the
alkaline iron selenides, the Te-doped ``11'' iron selenides FeSe,
and the monolayer FeSe on the SrTiO$_{\rm 3}$ substrate~\cite{MYi.2015, Yi_et_al:PRL_2013,Wang2014,Ding2014,Li2014}.
The effective quasiparticle mass normalized by its non-interacting counterpart, $m^*/m_{band}$
is on the order of $3-4$ for the $3d_{xz,yz}$ orbitals, but is as large as $20$ for the $3d_{xy}$ orbital~\cite{MYi.2015,Yi_et_al:PRL_2013,Liu2015}.
Such orbital selectivity has also
been the subject of extensive recent theoretical studies
~\cite{Yu_Si:2011, Yu_Zhu_Si:PRL_2013, deMedici:PRL_2014}.
All of these aspects make it
natural to study orbital dependent
~\cite{Yu_Zhu_Si:2014,Yin:Nat_Phys_2014, Ong_Coleman_Schmalian:2014}
and related ~\cite{HaoHu:2014}
superconducting pairing.
We are thus motivated to
address
the hitherto unexplored question, {\it viz.} whether there exists an orbital-selective pairing state which can
reconcile the seemingly contradictory properties observed in the iron-selenide superconductors.
We also
examine
the stability of such a pairing state
at the level of an effective Hamiltonian for studying superconductivity,
in which we
incorporate
the
orbital-selectivity
in the short-range
exchange interactions
(see Supplementary Information (SI)).
{\it Orbital-selective $s\tau_3$ pairing state -- a simplified case:}~~
We first discuss the structure and properties of the $s\tau_3$ pairing state in
a simplified two-orbital $d_{xz}, d_{yz}$ system. This
illustrates
how features typically associated
with {\it both} standard structure-less $s$- and $d$-wave states can
{\it simultaneously} arise.
The salient features of the two-orbital model are illustrated in Fig.~\ref{Fig:Empty_BZ}.
We consider
spin-singlet pairing in the orbital basis,
in the case of two orbitals
$3d_{xz}$, $3d_{yz}$
~\cite{Raghu_et_al:2008}.
The
Hamiltonian, incorporating
the
$s\tau_{3}$
pairing
term,
is given by
\noindent \begin{align}
\label{Eq;Hamilt_orbital}
\hat{H}=& \sum_{\pmb{k}} \psi^{\dagger}_{\pmb{k}} \big( \hat{H}_{\text{Kinetic}}(\pmb{k})
+ \hat{H}_{\text{Pair}}(\pmb{k}) \big) \psi_{\pmb{k}} \notag \\
\hat{H}_{\text{Kinetic}}= & \big( \xi_{+}(\pmb{k}) \otimes \tau_{0} + \xi_{-}(\pmb{k}) \otimes \tau_{3}
+ \xi_{xy}(\pmb{k}) \otimes \tau_{1} \big) \otimes \sigma_{0} \otimes \gamma_{3} \notag \\
\hat{H}_{\text{Pair}}= & \Delta_{0} g_{x^2y^2}(\pmb{k}) \otimes \tau_{3} \otimes \sigma_{0} \otimes \gamma_{1},
\end{align}
\noindent where $\psi^{\dagger}
_{\pmb{k}}
=(c^{\dagger}_{\pmb{k} i \sigma}, c_{-\pmb{k} j \sigma'} (i\sigma_{2})_{\sigma' \sigma} )$ is equivalent to a Nambu spinor
where $i,j$ are orbital indices
(SI Section).
The
$\tau_{i}, \sigma_{i}$, and $\gamma_{i}, (i=0, \ldots, 4)$
2 x 2 Pauli matrices
represent orbital iso-spin, spin, and Nambu indices, respectively.
The $\xi_{+}, \xi_{-}$, and $\xi_{xy}$ factors appearing
in the kinetic part
belong to the $A_{1g}, B_{1g}$, and $B_{2g}$ irreducible
representations
of the $D_{4h}$ point-group.
Their exact forms, as well as
the resulting electron bands
are given in the SI.
\noindent \begin{figure}[!htb]
\centering
\includegraphics[width=0.75\columnwidth]{Fig1}
\caption{
\emph{Schematic} illustration of the two-orbital
$s\tau_3$ pairing in a 1-Fe Brillouin Zone (BZ), which is
obtained by unfolding the 2D crystallographic BZ cell in the conventional fashion~\cite{Nica_Yu_Si:Unpublished_2015}.
The solid lines indicate typical Fermi pockets for the Fe-based superconductors.
The dotted, red lines indicate the zeroes specific to the intraband pairing (~$\xi_{-}$) while the dashed, blue lines mark the zeroes specific to the interband pairing (~$\xi_{xy}$).
The intra- and inter- band components do not vanish at the same subset of $\pmb{k}$, ensuring there is always a non-zero pairing given by either of the two components
on the entire Fermi surface.
For $\text{max}(\xi_{-}) \approx \text{max}(\xi_{xy})$ the angle $\phi(\pmb{k})$
(Eqs.~\ref{Eq:Balian_Wert_dispersion}-\ref{Eq:Ang_dep})
can be roughly identified with twice the winding angle shown
for fixed $|\pmb{k}|$.
In addition, there is a sign change between the intraband pairing along the two pockets at the edge of the BZ, a condition necessary to the formation of a resonance in the spin excitation spectrum at the wavevector $\pmb{q}=(\pi, \pi/2)$ observed in experiment \cite{Dai:arxiv_2015}.
}
\label{Fig:Empty_BZ}
\end{figure}
The even-parity, spin-singlet candidate
$s\tau_{3}$
pairing function
with non-trivial orbital structure
is included in the $\hat{H}_{\text{Pair}}$
term in Eq.~\ref{Eq;Hamilt_orbital}.
While $\Delta_{0}$ is a (generally) complex number,
we choose a real amplitude for convenience.
The form factor
$g_{x^{2}y^{2}}(\pmb{k})$ is parity-even and belongs to
the $A_{1g}$ representation of the $D_{4h}$ point group.
In the absence of spin-orbit coupling, the rotational properties of the $s\tau_{3}$ pairing
are of $B_{1g}$ symmetry. The latter
is entirely determined by the
tensor product of the $g_{x^{2}y^{2}}(\pmb{k})$
(s-wave)
form factor and the $\tau_{3}$ orbital matrix.
To illustrate,
under a $C_{4z}$ rotation, the form-factor is invariant, while the $\tau_{3}$ matrix
transforms as a rank-two $B_{1g}$ tensor representation of the point-group, i.e. it changes sign.
We note that the anti-symmetry under exchange is guaranteed by the spin-singlet nature, together with the even-parity of the form factor.
Since the spin-structure is not essential
for the following arguments, we shall henceforth omit
the explicit $\sigma_{0}$ matrix.
The non-trivial characteristics of this pairing
are consequences of the commutator $\left[\hat{H}_{\text{Kinetic}}, \hat{H}_{\text{Pair}} \right] \neq 0$
for \emph{general} momentum $\pmb{k}$.
We use the notation of Ref.~\citenum{Ong_Coleman_Schmalian:2014},
and rewrite the Hamiltonian Eq.~\ref{Eq;Hamilt_orbital}
as follows
\noindent \begin{align}
\label{Eq:Balian_Wert_form}
\hat{H}=& \sum_{\pmb{k}} \psi^{\dagger}({\pmb{k}}) \big[ \left( \xi_{+}(\pmb{k}) \tau_{0} +\vec{B}({\pmb{k}}) \cdot \vec{\tau} \right) \otimes \gamma_{3} \notag \\
& + \left( \vec{d}({\pmb{k}}) \cdot \vec{\tau} \right) \otimes \gamma_{1} \big] \psi({\pmb{k}}),
\end{align}
\noindent where
\noindent \begin{align}
\vec{B}(\pmb{k})= & \left( \xi_{xy}(\pmb{k}), 0, \xi_{-}(\pmb{k}) \right)
\nonumber
\\
\vec{d}(\pmb{k})= & \left( 0, 0, \Delta_{0} g_{x^2y^2}(\pmb{k}) \right)
\label{Eq:Pairing_form}.
\end{align}
\noindent This is formally similar to
a Balian-Werthamer form
~\cite{Balian_PR_1963,Leggett1975,Sigrist_Ueda:1991}
(see SI for more details),
with the $\vec{B}(\pmb{k})$ factor
being
analogous to a $\pmb{k}$-dependent spin-orbit coupling.
To account for the non-commuting $\hat{H}_{\text{Kinetic}}$ and $\hat{H}_{\text{Pair}}$,
we write
the square of the Hamiltonian matrix:
\noindent \begin{align}
\label{Eq:Squared_Hamilt}
\hat{H}^2= & \sum_{\pmb{k}} \left[ \xi_{+}(\pmb{k}) \tau_{0} + \left( \vec{B}_{\pmb{k}} \cdot \vec{\tau} \right) \right]^2 \otimes \gamma_{0} + \left| \vec{d}(\pmb{k}) \right|^2 \tau_{0} \otimes \gamma_{0} \notag \\
& + 2i \left( \vec{B}(\pmb{k}) \times \vec{d}(\pmb{k}) \right) \cdot \vec{\tau} \otimes i \gamma_{2}.
\end{align}
\noindent where the well-known relation
$\big( \vec{a} \cdot \vec{\tau} \big) \big( \vec{b} \cdot \vec{\tau} \big)= \vec{a} \cdot \vec{b}+i \big( \vec{a} \times \vec{b} \big) \cdot \vec{\tau}$
was used. The first two terms,
proportional to the $\gamma_{0}$ Nambu matrix,
are the squares of the
kinetic
Hamiltonian
and of a pairing contribution with no essential structure in orbital space, given by
$\left| \vec{d}(\pmb{k}) \right|^2$.
The latter is an effective amplitude of
the pairing interactions and, as such, is proportional to the square of the s-wave like $g_{x^{2}y^{2}}$ form factor, as can be seen from Eq.~\ref{Eq:Pairing_form}.
Together with the
kinetic
part,
it amounts to the usual (and sole) contribution to the
Bogoliubov-de Gennes (BdG) quasiparticle spectrum,
whenever $\left[\hat{H}_{\text{Kinetic}}, \hat{H}_{\text{Pair}} \right] = 0$
for \emph{all} $\pmb{k}$.
The last term
in Eq.~\ref{Eq:Squared_Hamilt}
reflects the non-commuting $\hat{H}_{\text{Kinetic}}$ and $\hat{H}_{\text{Pair}}$.
Since the Nambu matrices $\gamma_{0}$ and $i \gamma_{2}$ commute,
$\hat{H}^{2}$ in Eq. \ref{Eq:Squared_Hamilt}
can be easily expressed in block diagonal form
(SI).
The resulting Bogoliubov-de Gennes (BdG) bands are given by
\noindent
\begin{widetext}
\begin{equation}
\label{Eq:Balian_Wert_dispersion}
E_{\pm}(\pmb{k}) = \sqrt{ \left ( \sqrt{\xi^{2}_{+}(\pmb{k}) + \left| \vec{d}(\pmb{k}) \right|^2 \text{sin}^2\phi(\pmb{k}) } \pm \left| \vec{B}(\pmb{k}) \right| \right)^2 + \left| \vec{d}(\pmb{k}) \right|^2 \left( 1- \text{sin}^2\phi(\pmb{k}) \right) }
\end{equation}
\noindent where
\noindent
\begin{equation}
\label{Eq:Winding_angle}
\text{sin}\phi(\pmb{k})=\frac{\xi_{xy}(\pmb{k})}{\left| \vec{B}(\pmb{k}) \right|}= \frac{\xi_{xy}(\pmb{k})}{\sqrt{ \xi^2_{-}(\pmb{k})+ \xi^2_{xy}(\pmb{k})}}.
\end{equation}
\end{widetext}
\noindent
The terms proportional to $\sin \phi(\pmb{k})$
reflects the non-Abelian aspect of the pairing state.
Note that
Eq.~\ref{Eq:Balian_Wert_dispersion} corresponds
to the sum of two positive semi-definite terms.
For \emph{general} $\vec{d}(\pmb{k})$ we see that
nodes can appear only when both terms in the square root vanish.
The second of these goes to zero when either $\text{sin}\phi(\pmb{k})=1$
or, trivially, when $\left| \vec{d}(\pmb{k}) \right| = 0$.
This latter case occurs when the FS intersects
the lines of zeros of the $g_{x^{2}y^{2}}$ form factor.
With the FeSCs in mind, we ignore this simple case in the following. Alternately,
when $\text{sin}\phi(\pmb{k})=1$, the dispersion reduces to
\noindent \begin{equation}
\label{Eq:Ang_dep}
E_{\pm}(\pmb{k}) = \left| \sqrt{\xi^{2}_{+}(\pmb{k}) + \left| \vec{d}(\pmb{k}) \right|^2 } \pm \left| \vec{B}(\pmb{k}) \right| \right|.
\end{equation}
\noindent \emph{On the FS}, we have
$\xi^{2}_{+}(\pmb{k})=\left| \vec{B}(\pmb{k}) \right|^2$
(see SI).
Thus,
there are no nodes \emph{on the FS}.
We note that \emph{away from the FS},
Eq.~\ref{Eq:Ang_dep} does not in general guarantee the absence of nodes.
However, because the lifetime of quasiparticles away from the FS will be finite,
the corresponding contributions to thermodynamical properties will be much weaker compared to the case of nodes on the FS.
In the band basis, the kinetic part of the Hamiltonian is diagonalized.
Given that the kinetic and pairing parts do not commute with each other,
the two cannot be simultaneously diagonalized.
Thus, the pairing part must contain an interband component.
To see this,
we apply a
canonical transformation which diagonalizes the kinetic part (see the SI),
but which also transforms the pairing into
\noindent
\begin{align}
\label{Eq:Pairing_band}
\hat{H}_{\text{Pair}}(\pmb{k})=\Delta_1(\pmb{k}) \alpha_3 + \Delta_2 (\pmb{k}) \alpha_1
\end{align}
\noindent where $\alpha_{1,3}$ are
Pauli matrices
corresponding to inter- and intra-band pairing terms. The two components are given by
\noindent \begin{align}
\Delta_1(\pmb{k}) = &
-\Delta_0 g_{x^2y^2}(\pmb{k}) \frac{\xi_{-}(\pmb{k})}{\sqrt{\xi^2_{-}(\pmb{k})+\xi^2_{xy}(\pmb{k})}}
\nonumber
\\
\Delta_2(\pmb{k}) = &
-\Delta_0 g_{x^2y^2}(\pmb{k}) \frac{\xi_{xy}(\pmb{k})}{\sqrt{\xi^2_{-}(\pmb{k})+\xi^2_{xy}(\pmb{k})}}
\label{Eq:Delta-1-2}.
\end{align}
\noindent
The band-diagonal $\alpha_3$ and band off-diagonal $\alpha_1$
pairing components have $d(x^2-y^2)$ and $d(xy)$ form factors, respectively.
As illustrated in
Fig.~\ref{Fig:Empty_BZ},
these
have nodes along
the diagonals and axes of the BZ,
respectively.
Because the two matrices $\alpha_{1,3}$
anti-commute, the single-particle excitation energy depends on the addition in quadrature of
the two pairing amplitudes $\Delta_1(\pmb{k})$ and $\Delta_2(\pmb{k})$.
This ensures that the excitation gap is nodeless on the entire Fermi surface.
As can be seen from Eqs.~\ref{Eq:Pairing_band},~\ref{Eq:Delta-1-2},
the band-index diagonal term changes sign about the diagonals
($k_x=\pm k_y$) of the BZ,
as dictated by the $d(x^2-y^2)$ nature of the intraband component.
Thus, the intraband pairing component does indeed change sign
between the two electron Fermi pockets at the BZ boundaries.
It ensures that this type of pairing is conducive
to the formation of a resonance with a wavevector that connects the two electron Fermi pockets.
We stress that the two main features of the $s\tau_{3}$ pairing, i.e.
the formation of a gap on the FS and the sign-change in the intraband component, cannot be reconciled by the more typical pairing candidates,
which lack an orbital structure. In the context of our two-orbital model, the $s \otimes \tau_{0}$ and $d \otimes \tau_{0}$ candidate states, corresponding to the typical orbitally-trivial s and d-wave pairings, commute with $\hat{H}_{\text{Kinetic}}$. Consequently, they are associated with intraband pairing only. As such, neither of the two types can induce a nodeless gap \emph{and} account for the sign change required for the spin-resonance.
{\it Orbital-selective $s\tau_3$ pairing state -- the case of iron selenides:}~~
Superconductivity in the alkaline iron selenides, like in the related case of the iron pnictides, involves all five Fe-$3d$ orbitals.
Thus, it is important to consider
the five-orbital case to address i) whether
the $s\tau_{3}$ pairing state is energetically favored compared to the more conventional pairing states and ii) whether it captures
the essential properties of this pairing state as they pertain to the iron selenide
superconductors.
To
study the stability of
the $s\tau_{3}$ pairing state, we start
from
two
previously
discussed aspects
of the FeSCs.
We do so in terms of
a strong-coupling approach to superconductivity, in light of
the strong correlation effects
~\cite{Si_Abrahams:PRL_2008,Yin_Haule_Kotliar:Nat_Phys_2011,Fang08,Xu08,KSeo,Moreo.2009,WQChen.2009,DHLee.2013,Berg.2009,Lv,Bascones,deMedici:PRL_2014}
that are especially clear-cut for the iron selenides
~\cite{MYi.2015,Yi_et_al:PRL_2013,Liu2015}.
This approach is described in the SI, with superconductivity
driven by short-range interactions.
The latter include the antiferromagnetic interactions between
the nearest-neighbor (NN, $J_{1}^{\alpha}$)
and next-nearest-neighbor (NNN,$J_{2}^{\alpha}$) Fe sites on their square lattice,
for the three
most relevant orbitals,
$\alpha=$ $3d_{xz}$, $3d_{yz}$, and $3d_{xy}$.
We reiterate that we will analyze the model in the 1-Fe unit cell and the corresponding BZ.
One
of the known
aspects of the FeSCs is the large parameter regime
where
the conventional
$d$-wave $B_{1g}$ and $s$-wave $A_{1g}$
pairing states
are quasi-degenerate
~\cite{Yu_Nat_Comm:2013,Graser.2009}.
In terms of a model
with short-range antiferromagnetic interactions, this occurs in the regime of magnetic
frustration with $J_{2}$ being comparable to $J_{1}$~\cite{Yu_Nat_Comm:2013}, a condition that is evidenced by both
theoretical considerations and experimental measurements
~\cite{Si2016,Dai:arxiv_2015}.
To quantify this effect, we introduce the ratio
$A_{L} \equiv
J_2/J_1$ to describe the relative strength of these two interactions.
For a proof-of-concept demonstration, we analyze the phase diagram by taking the $J_2/J_1$ axis to be a cut in the parameter
space along which $A_L$ is the same for the different $3d$ orbitals.
The quasi-degeneracy arises
when $A_{L} \sim 1$.
The second well-known property of the FeSCs
is orbital selectivity, as
described
above.
Our effective model incorporates
an exchange orbital-anisotropy factor
$A_{O}=J_{1}^{xy}/J_{1}^{xz/yz}=J_{2}^{ xy}/J_{2}^{xz/yz}$,
and reflects the orbital selectivity
by $A_{O}$'s deviation from $1$.
For the iron selenides, $A_O$ is expected to be considerably smaller than $1$ (see SI).
We are now in position to discuss how the
$s\tau_{3}$
pairing state emerges in a range of parameters where the $s-$ and $d-$wave pairing channels are quasi-degenerate.
Within the 5-orbital $t-J_{1}-J_{2}$ model, we focus on the case with a kinetic part appropriate for the alkaline iron selenides
K$_{y}$Fe$_{2-x}$Se$_{2}$ although similar behavior emerges in the cases appropriate for the iron pnictides and single-layer FeSe
(see SI).
We present our results for the case of orbital-diagonal exchange interactions. The inter-orbital exchange interactions have only negligible
effects on the pairing amplitudes, as demonstrated in the SI.
\begin{figure}[htb!]
\centering
\subfloat[]{\includegraphics[width=0.8\columnwidth]{Fig2a}}\\
\subfloat[]{\includegraphics[width=0.8\columnwidth]{Fig2b}}
\caption{
Phase diagrams based on the leading pairing amplitudes given by self-consistent calculations with fixed $J_2=1$ and tight-binding parameters appropriate to (a) alkaline iron selenides, and (b) iron pnictides. The tight-binding parameters used can be found in Ref. \citenum{Yu_Nat_Comm:2013}. The blue shaded areas correspond to dominant pairing channels with an $s_{x^2y^2}$ form factor while the red shading covers those with a $d_{x^2-y^2}$ form factor. The continuous line separates regions where the pairing belongs to the $A_{1g}$ and the $B_{1g}$ representations respectively. The $1 \times 1$ matrix in the $d_{xy}$ subspace is represented by $\pmb{1}_{xy}$. The orbital-selective $s\tau_{3}$ pairing occurs for $A_O < 1$, $A_L$ near 1 in all cases.
}
\label{Fig:Phase_diagrams_K_Fe_Se}
\end{figure}
The phase diagram for the alkaline iron selenides is shown in Fig.~\ref{Fig:Phase_diagrams_K_Fe_Se} (a).
In the absence of orbital selectivity, $A_O=1$, it is known that
small and large $A_L$
promote the $s_{x^2y^2} \otimes \tau_{0} , {A_{1g}}$
and $d_{x^2-y^2} \otimes \tau_{0}, B_{1g}$,
both defined in the $d_{xz}$, $d_{yz}$ subspace~\cite{Yu_Nat_Comm:2013}.
Increasing the orbital selectivity, with $A_O$ decreasing from $1$,
these two limiting regimes
remain essentially unchanged.
However, in the magnetically frustrated regime
$A_L \sim 1$,
the $s_{x^2y^2} \otimes \tau_{0} , {A_{1g}}$
and $d_{x^2-y^2} \otimes \tau_{0}, B_{1g}$
become quasi-degenerate.
When $A_O$ is sufficiently smaller than $1$, the $s\tau_{3}$ pairing state becomes the dominant channel
in the intermediate regime.
Similar phase diagrams are obtained for the iron pnictides and single-layer FeSe
shown in Figs.~\ref{Fig:Phase_diagrams_K_Fe_Se} (b) and~S1
(SI), respectively.
A typical dominant
$s\tau_{3}$ pairing case is shown in
Fig.~S2 in the
SI
for a number of subleading symmetry-allowed channels~\cite{Goswami_Nikolic_Si:EPL_2010} for alkaline iron selenide dispersion with fixed $J_2/J_1=1.5$, $A_O=0.3$ and varying $A_L$ (horizontal axis).
Having established the stability of the $s\tau_3$ pairing state, we now address its salient properties. We first consider the spin-excitation spectrum.
In Fig.~\ref{Fig:K_spin_spectrum} we show the
dynamical spin susceptibility
at wave-vector $\pmb{q}=(\pi, \pi/2)$ for $J_2=1.5$. We note the complicated frequency behavior which can be traced to the anisotropy in the effective gap affecting both the coherence factors and the position of minimum in quasi-particle energy. We show the minimum and maximum particle-hole (p-h)
thresholds corresponding to twice the minimum and twice the maximum gaps. As suggested by Figs.~\ref{Fig:K_FS_pair_sign} (a) and (b), states connected by $\pmb{q}=(\pi, \pi/2)$ would correspond to a p-h threshold given roughly by the sum of the minimum and maximum gap.
A sharp feature appears below this threshold,
confirming the existence of the resonance for $\pmb{q}=(\pi, \pi/2)$ as found in experiments on the alkaline iron selenides \cite{Park_etal:2011,Friemel_etal:2012,Dai:arxiv_2015}.
The resonance at this wavevector originates from the sign change of the intraband pairing component across
the two Fermi pockets at the edge of the BZ, around $(\pm \pi,0)$ ($\delta$) and $(0, \pm \pi)$, as illustrated in Fig.~\ref{Fig:K_FS_pair_sign} (a),
and further discussed in the SI.
Without such a sign change, there cannot be a sharp resonance
{\it below} the p-h threshold energy.
\noindent \begin{figure}[!htb]
\centering
\includegraphics[width=0.7\columnwidth]{Fig3}
\caption{
The imaginary part of the dynamical spin susceptibility for the alkaline iron selenides at wave-vector $\pmb{q}=(\pi, \pi/2)$, for a dominant $s\tau_{3}$ pairing for parameters $J_2=1.5$, $A_O=0.3, A_L=0.9$.
The arrows show twice the minimum and maximum gaps (see Fig. \ref{Fig:K_FS_pair_sign} (b)). There is a sharp feature ar $\omega \approx 0.36$ within the bounds of twice the effective gap and below the p-h threshold of roughly 0.41 associated with this wavevector.
}
\label{Fig:K_spin_spectrum}
\end{figure}
We next turn to the quasiparticle excitation spectrum.
Fig.~\ref{Fig:K_FS_pair_sign} (b) shows the gap at the FS as a function of winding angle $\theta$.
It clearly illustrates the node-less dispersion as the gap is nonzero for all $\theta$.
The electron dispersion considered here does not produce any Fermi pockets close to $\Gamma$ in the BZ. This is in contrast to ARPES experiments on K$_{y}$Fe$_{2-x}$Se$_{2}$~\cite{Zhang_et_al:2011,Liu-et-al:2012}
which show a small electron pocket near $\Gamma$. Because this
electron pocket has very small spectral weight, it is to be expected that even if such a pocket were included, the dominant
$s\tau_{3}$ pairing will still arise; moreover, the gap on this Fermi pocket will be node-less as discussed in the two-orbital case. To substantiate this,
we consider the results for the iron pnictides class, which do have significant ({\it albeit} hole) Fermi pockets
at the zone center yet exhibit a full gap. In Figs.~\ref{Fig:A_FS_gap} (a), (b) we show the FS and the gaps as functions of winding angle $\theta$ for $A_O=0.5$ and $A_L=1.3$ corresponding to a dominant
$s\tau_{3}$ pairing. The gap along $\beta$ is finite and exhibits an anisotropy consistent with the two orbital results in
Eq.~\ref{Eq:Balian_Wert_dispersion}. In the latter case, at winding angle $\theta=0$, $\text{sin}\phi=0$ and the spectrum has a minimum/maximum gap for $E_{+/-}$. As $\theta$ is increased the $\left| \vec{B}(\pmb{k}) \times \vec{d}(\pmb{k}) \right|^2$ term increases reaching a maximum at $\theta=\pi/4$. Here the gap is maximum/minimum for $E_{+/-}$. This is consistent with the anisotropy in the gap shown in Fig.~\ref{Fig:A_FS_gap}.
\noindent \begin{figure*}[htb!]
\centering
\subfloat[]{\includegraphics[width=0.75\columnwidth]{Fig4a}}
\qquad
\subfloat[] {\includegraphics[width=0.75\columnwidth]{Fig4b}}
\caption{
(a) The FS (solid line) and the real intra-band pairing for the band generating the $\delta$ pockets at the edge of the BZ for a dispersion typical of the alkaline iron selenides. Note the clear change in sign between pockets separated by the BZ diagonal. The dashed arrow indicates the $\pmb{q}=(\pi,\pi/2)$ wave-vector associated with the resonance in the spin spectrum found in experiment \cite{Dai:arxiv_2015}. (b) The size of the gap along the $\delta$ pocket. Both figures are for $J_2=1.5$, $A_O=0.3, A_L=0.9$ with dominant $s\tau_{3}$ pairing.
}
\label{Fig:K_FS_pair_sign}
\end{figure*}
\noindent \begin{figure*}[htb!]
\centering
\subfloat[]{\includegraphics[width=0.75\columnwidth]{Fig5a}}
\qquad
\subfloat[]{\includegraphics[width=0.75\columnwidth]{Fig5b}}
\caption{
(a) FS for the iron pnictides which includes hole pockets around the $\Gamma$ point
with dominant $s\tau_{3}, B_{1g}$
for $J_2=1, A_L=1.3, A_O=0.5 $.The tight-binding parameterscan be found in Ref.~\citenum{Yu_Nat_Comm:2013}. (b) The gaps along the $\beta$ and $\delta$ pockets close to the center and edge of the BZ. A similar gap forms around the $\alpha$ pocket.
}
\label{Fig:A_FS_gap}
\end{figure*}
\section{Discussion}
\label{Sec:Dscssn}
Several remarks are in order.
First, the full gap and the sign change of the intraband pairing component discussed above provide evidence that, with strong orbital selectivity, the $s \tau_{3}$ pairing in a realistic five-orbital
model
has a behavior very similar to that of the two-orbital case.
Second, with the short-range $J_1-J_2$ interactions driving superconductivity,
pairing involves the electronic states over an extended range of energy
about the Fermi energy.
The energy window can be determined from the zone-boundary spin excitation energies,
which
are
on the order of $200$ meV for most iron selenides (and pnictides)~\cite{Dai:arxiv_2015}.
This is important for the consideration of the quasiparticle excitation gap at the small
electron pocket
of K$_{y}$Fe$_{2-x}$Se$_{2}$
near the origin of the Brillouin zone.
According to the ARPES experiments~\cite{Zhang_et_al:2011, Liu-et-al:2012}
this Fermi pocket
contains Fe $3d_{xy}$ and Se $4p_z$ orbitals ($\alpha$ band),
while
the hole ($\beta$) bands containing both $3d_{xz}$ and $3d_{yz}$ orbitals and are
only about $60$-$80$ meV below the Fermi energy.
We therefore expect that both the intraband and interband pairing components will be significant
for this part of the Brillouin zone and the mechanism advanced here will make the quasiparticle excitations to be fully gapped for this small electron pocket.
Third, within our approach both the iron selenides and
pnictides are bad metals in the regime of quasi-degenerate $s-$ and $d-$wave pairings. However, the iron selenides have stronger correlations,
which will lead to a larger ratio of the exchange interaction to renormalized kinetic energy
(note that the {\it renormalized} bandwidth goes to zero when a bad metal approaches the electron localization transition) and,
correspondingly~\cite{Yu_Nat_Comm:2013},
larger pairing amplitudes. We expect
this will contribute to the larger maximum $T_c$ observed in the iron selenides than in the iron pnictides.
Relatedly,
the alkaline iron selenides have a stronger orbital selectivity than the iron pnictides,
and we thus expect that the $s\tau_3$ pairing
is more likely
realized in the former than in the latter.
Fourth, it is instructive to
compare the mechanism advanced here with
a conventional means of relieving quasi-degenerate $s-$ and $d-$wave pairing states with trivial orbital structure, which
consists in linearly superposing the two into an $s+id$ state.
The latter, breaking
the time-reversal symmetry, would be stabilized at temperatures sufficiently below the superconducting transition temperature.
By contrast, the $s\tau_{3}$ pairing state
preserves the time-reversal symmetry. It is an
irreducible representation of the point group, and
is therefore
stabilized as the temperature is lowered immediately below the superconducting transition.
Thus, the emergence of the intermediate $s\tau_{3}$ pairing state
represents a new means to relieve the quasi-degeneracy through the development of orbital selectivity.
Finally, the nodeless $d$-wave nature of $s\tau_{3}$ may shed new light on other strongly correlated multi-band superconductors. For instance, one of the striking puzzles emerging in heavy fermion superconductors is the simultaneous exhibition of a variety of $d-$wave characteristics and of a gap in the lowest-energy excitation spectrum~\cite{Kittaka-2014}. Whether a multiband pairing state such as $s\tau_{3}$ provides a systematic understanding of such properties is an intriguing open question for future studies.
To summarize,
we have demonstrated that an orbital-selective $s\tau_{3}$ pairing state exhibits properties
that would appear mutually exclusive from the conventional perspective where the orbital degrees of freedom are ignored.
It provides a natural understanding of the enigmatic properties observed in the alkaline iron selenides.
These include the single-particle excitations which are fully gapped on the entire Fermi surface,
as observed in ARPES experiments, and a pairing function which changes sign across the electron Fermi
pockets at the Brillouin-zone
boundary, as indicated by the resonance peak seen near $(\pi,\pi/2)$ in the inelastic neutron scattering experiments.
In addition, we have shown that the pairing state is energetically competitive in an orbital-selective model of short-range
antiferromagnetic exchange interactions, in the regime where the conventional $s-$ and $d-$wave pairing channels are quasi-degenerate.
As such, our understanding of the properties of the iron-selenide superconductors provides evidence
that the high-T$_{\rm c}$ superconductivity in the iron-based materials originates from the antiferromagnetic correlations
of strongly correlated electrons. More generally, our work highlights how new classes of unconventional superconducting pairing state
emerge in the presence of additional internal degrees of freedom, with properties that cannot otherwise be expected.
This new insight may well be important for the understanding of a variety of other strongly correlated superconductors,
including the heavy fermion and organic systems.
\section{Acknowledgments}
\label{Acknow}
We acknowledge useful discussions with E. Abrahams, A. V. Chubukov, G. Kotliar and P. J. Hirschfeld.
The authors declare that they have no competing interests.
All authors contributed to the research of the work and the writing of the paper.
The work has been supported in part by the NSF Grant No.\ DMR-1611392 and the Robert A.\ Welch Foundation Grant No.\ C-1411
(E.M.N. \& Q.S.). R.Y. was partially supported by the National Science Foundation of China Grant number 11374361,
and the Fundamental Research Funds for the Central Universities and the Research Funds of Renmin University of China.
All of us acknowledge the support provided in part by the NSF Grant No. NSF PHY11-25915 at KITP, UCSB.
Correspondence and requests for materials should be addressed to E.M.N. (<EMAIL>)
or Q.S. (<EMAIL>).
\section{Two-orbital model}
\label{Sec:Dscssn_tw_orbtl_mdl}
\subsection{Tight-binding details}
The components of the tight-binding part
of the two-orbital Hamiltonian discussed in the main text are given by
\noindent \begin{align}
\xi_{\pmb{k}+} = & -(t_1+t_2)(\cos k_x+\cos k_y) \notag \\
& -4t_3\cos k_x \cos k_y, \\
\xi_{\pmb{k}-}= & -(t_1-t_2)(\cos k_x-\cos k_y), \\
\xi_{\pmb{k}xy}= & -4t_4\sin k_x \sin k_y,
\end{align}
\noindent where $t_1$,$t_2$ and $t_3$ are tight-binding parameters.
Details can be found in Ref.~\onlinecite{S_Raghu_et_al:2008}.
The corresponding band dispersion is in general given by
\noindent \begin{align}
\label{Eq:Free-band}
\epsilon_{\pm}(\pmb{k})= & \xi_{+}(\pmb{k}) \pm \sqrt{\xi^2_{-}(\pmb{k})+ \xi^2_{xy}(\pmb{k})} \notag \\
= & \xi_{+}(\pmb{k}) \pm \left| \vec{B}(\pmb{k})\right|
\end{align}
\noindent The Fermi surface is determined by the condition
\noindent \begin{equation}
\epsilon_{\pm} (\pmb{k}_{FS}) = 0,
\end{equation}
\noindent which is equivalent to
\noindent \begin{equation}
\xi_{+}(\pmb{k}_{FS}) = \mp \left| \vec{B}(\pmb{k}_{FS}) \right|.
\end{equation}
\subsection{Nambu form}
The pairing part
written as $\hat{H}_{\text{Pair}} \sim \vec{d} \cdot \vec{\tau}$
is equivalent to a more-conventional Balian-Werthamer form
$ \left( \tilde{\pmb{d}} \cdot \vec{\tau} \right) \left(i \tau_{2} \right)$
which is conventionally used for pairing functions with non-trivial spin structure.
This is so provided that $d_{2}=\tilde{d}_{2}=0$, which is the case for $s \otimes \tau_{3}$ pairing,
together with ${d}_{1} \rightarrow \tilde{d}_{3}, {d}_{3}\rightarrow -\tilde{d}_{1}$.
Formally, this transforms $2i \left( \vec{B} \times \vec{d}\right) \cdot \vec{\tau}$
in the expression for $\hat{H}^{2}$ (Eq.~4 in the main text)
to
$ 2 \left( \vec{B} \cdot \pmb{\tilde{d}} \right) i\tau_{2}$. The resulting BdG bands are
identical, as can be seen by expanding the direct products.
Note that, in contrast to the typical \emph{spin}-triplet pairing,
both $\vec{d}$ and $\tilde{d}$ orbital iso-spin vectors
are parity-even $\left(\vec{d}(-\pmb{k})=\vec{d}(\pmb{k})\right)$.
Together with the spin-singlet nature,
this ensures that the Cooper pairs are anti-symmetric under exchange.
In order to better illustrate the effects of the non-trivial orbital structure,
we incorporate the spin-singlet nature of the pairing Hamiltonian into a transformed Nambu spinor:
\noindent \begin{align}
\psi^{\dagger}(\pmb{k})= \psi^{\dagger}_{\text{Nambu}}~U^{\dagger}
\end{align}
\noindent where
\noindent \begin{equation}
\psi^{\dagger}_{\text{Nambu}}(\pmb{k})=(c^{\dagger}_{\pmb{k} i \sigma}, c_{-\pmb{k} j \sigma}),
\end{equation}
\noindent is the canonical Nambu spinor and
\noindent \begin{equation}
U=\frac{1}{2} \big( \sigma_{0} \otimes (\gamma_{0}+\gamma_{3})
+ i\sigma_{2} \otimes (\gamma_{0}-\gamma_{3}) \big) \otimes \tau_{0}.
\end{equation}
\subsection{BdG spectrum}
$\hat{H}^{2}$ in Eq.~4 in the main text
\noindent \begin{align}
\hat{H}^2= & \sum_{\pmb{k}} \left[ \xi_{+}(\pmb{k}) \tau_{0} + \left( \vec{B}_{\pmb{k}}
\cdot \vec{\tau} \right) \right]^2 \otimes \gamma_{0} + \left| \vec{d}(\pmb{k}) \right|^2 \tau_{0} \otimes \gamma_{0} \notag \\
& + 2i \left( \vec{B}(\pmb{k}) \times \vec{d}(\pmb{k}) \right) \cdot \vec{\tau} \otimes i \gamma_{2}.
\end{align}
\noindent can be brought to a block-diagonal form in the Nambu indices by applying the transformation
\noindent \begin{equation}
\tilde{U}=e^{-i \gamma_{1} (\pi/4)} \otimes \tau_{0} \times \sigma_{0}
\end{equation}
\noindent such that
\noindent \begin{align}
\label{Eq: H_sqrd_explct}
\tilde{U} \left( \hat{H}^{2} \right) \tilde{U}^{\dagger}=
\begin{pmatrix}
\tilde{H} & 0 \\
0 & \tilde{H}^{T}
\end{pmatrix}
\end{align}
\noindent where
\noindent \begin{align}
\tilde{H}=
\begin{pmatrix}
\xi^{2}_{+}+ | \vec{B} |^{2} + \left|d \right|^{2} + 2 B_{3} \xi_{+} & 2 B_{1}( \xi_{+} - i d_{3}) \\
2 B_{1} ( \xi_{+} + i d_{3}) & \xi^{2}_{+}+ | \vec{B} |^{2} + \left|d \right|^{2} - 2 B_{3} \xi_{+} .
\end{pmatrix}
\end{align}
\noindent From this expression,
one can easily check that the eigenvalues of $\hat{H}$
are given by
\noindent
\begin{equation}
\label{Eq:Gnrl_2_orbtl_dsprsn}
E_{\pm}(\pmb{k})= \sqrt{ \xi^{2}_{+}(\pmb{k}) + \left| \vec{B}(\pmb{k}) \right|^2
+ \left| \vec{d}(\pmb{k}) \right|^2 \pm \sqrt{ 4\xi^{2}_{+}(\pmb{k})\left| \vec{B}(\pmb{k}) \right|^2 + 4\left| \vec{B}(\pmb{k})
\times \vec{d}(\pmb{k}) \right|^2 }. }
\end{equation}
\noindent
The explicitly positive semi-definite form of Eq.~5 in the main text
was obtained by writing
\noindent \begin{align}
\label{Eq:Expndd_sqrt_rt}
\sqrt{ 4\xi^{2}_{+}(\pmb{k})\left| \vec{B}(\pmb{k}) \right|^2 + 4\left| \vec{B}(\pmb{k}) \times \vec{d}(\pmb{k}) \right|^2 }
= & \sqrt{ 4\xi^{2}_{+}(\pmb{k})\left| \vec{B}(\pmb{k}) \right|^2 + 4 B^{2}_{1}(\pmb{k}) d^{2}_{3}(\pmb{k}) } \notag \\
= & 2 \left| \vec{B}(\pmb{k}) \right| \sqrt{ \xi^{2}_{+}(\pmb{k})
+ \frac{B^{2}_{1}(\pmb{k})}{\left| \vec{B}(\pmb{k}) \right|^{2}} d^{2}_{3}(\pmb{k}) } \notag \\
= & 2 \left| \vec{B}(\pmb{k}) \right| \sqrt{ \xi^{2}_{+}(\pmb{k}) + \sin^{2}\phi(\pmb{k}) \left| \vec{d}(\pmb{k}) \right|^{2} }.
\end{align}
The square can be completed by adding and subtracting
$\sin^{2}\phi(\pmb{k}) \left| \vec{d}(\pmb{k}) \right|^{2}$.
Alternately, a more conventional form for the BdG dispersion
can be obtained from Eq.~\ref{Eq:Gnrl_2_orbtl_dsprsn}
by adding and subtracting $2 \xi_{+}(\pmb{k}) \left| \vec{B}(\pmb{k}) \right|$ to Eq.~\ref{Eq:Expndd_sqrt_rt},
and completing the square for the non-interacting bands $\epsilon^{2}_{\pm}$.
The result is:
\noindent \begin{equation}
\label{Eq:Effct_prng_gps}
E_{\pm}(\pmb{k})= \sqrt{\epsilon^{2}_{\pm}(\pmb{k}) + |\vec{d}|^{2}(\pmb{k}) \pm |Q|(\pmb{k})},
\end{equation}
\noindent where
\noindent \begin{align}
\label{Eq:Fr_dsprsn}
\epsilon_{\pm}(\pmb{k})= & \xi_{+}(\pmb{k}) \pm |\vec{B}|(\pmb{k}) \notag \\
= & \xi_{+}(\pmb{k}) \pm \sqrt{\xi_{-}^{2}(\pmb{k}) + \xi^{2}_{xy}(\pmb{k})},
\end{align}
\noindent are the electron bands, and the additional $|Q|$ factor is given by
\noindent \begin{equation}
\label{Eq:Efftv_gp}
|Q|(\pmb{k})= 2|\vec{B}|(\pmb{k}) \left( \sqrt{\xi^{2}_{+}(\pmb{k}) + \left| \vec{d}(\pmb{k}) \right|^2 \text{sin}^2\phi(\pmb{k}) }
- \xi_{+}(\pmb{k}) \right).
\end{equation}
\noindent
The presence of this additional
contribution, due to the non-commuting aspect discussed
in the main text,
induces a splitting between the two
conventionally-gapped BdG bands.
\noindent
Indeed, if $\left[\hat{H}_{\text{Kinetic}}, \hat{H}_{\text{Pair}} \right] \sim \vec{B} \times \vec{d}$
(Eq.~4 in the main text)
were to vanish for all $\pmb{k} \in BZ$,
the splitting given by $|Q|$ term would be absent as well.
This can occur for a $\vec{B}$ vector which is either
identically zero or aligned parallel/anti-parallel to
$\vec{d}$ for all momenta. In such cases,
the remaining first two terms in Eq.~\ref{Eq:Effct_prng_gps}
would correspond to a quasiparticle spectrum
with gaps determined by the amplitude of the pairing,
or by the square of the $g_{x^{2}y^{2}}$ form factor in our case.
The resulting BdG bands would be identical to
those for a simpler $s_{x^{2}y^{2}} \otimes \tau_{0}$ state,
which is an example of the $s\pm$ pairing.
As in this latter case, nodes would appear only
when the form factor vanishes along the
$\{\pm \pi/2, k_{y} \}, \{k_{x}, \pm \pi/2 \}$ lines.
A FS which does not intersect these lines would consequently be
completely gapped.
The presence of the last term in Eq.~4 in the main text
modifies this simple picture, by introducing
the additional splitting of the two conventionally-gapped BdG
bands.
Furthermore, it is possible that this splitting can be sufficiently strong
to induce nodes for the $\epsilon_{-}$ band. As shown by
Eq.~5 in the main text,
these can emerge
along the diagonals $|k_{x}|=|k_{y}|$ of the BZ.
However, we stress that, along the FS, this cannot occur, as explained above.
We also briefly mention that terms similar to $|\vec{Q}|$ are also known in the context
of non-unitary, spin-triplet, time-reversal-symmetry breaking pairings~\cite{S_Sigrist_Ueda:1991}.
\subsection{Band basis}
The pairing Hamiltonian ($\hat{H}_{\text{pair}}$) in the band-basis (Eq.~8 in the main text)
was obtained from
\noindent \begin{equation}
\left( V(\pmb{k}) \otimes \sigma_{\sigma_{0}}\right) \hat{H}_{\text{pair}}(\pmb{k}) \left( V^{T}(\pmb{k}) \otimes \sigma_{0}\right),
\end{equation}
\noindent where
\noindent \begin{equation}
V(\pmb{k})=
\begin{pmatrix}
\frac{\xi_{-}-\sqrt{\xi_{-}^2+\xi_{xy}^2}}{\sqrt{ \xi^2_{xy}+ \left( \xi_{-} - \sqrt{ \xi_{- }^2+ \xi_{xy}^2 } \right)^2 }} & \frac{\xi_{-}+ \sqrt{\xi_{-}^2+\xi_{xy}^2}}{\sqrt{ \xi^2_{xy}+ \left( \xi_{-} + \sqrt{ \xi_{- }^2+ \xi_{xy}^2 } \right)^2 }}\\
\frac{\xi_{xy}}{\sqrt{ \xi^2_{xy}+ \left( \xi_{-} - \sqrt{ \xi_{- }^2+ \xi_{xy}^2 } \right)^2 }} & \frac{\xi_{xy}}{\sqrt{ \xi^2_{xy}+ \left( \xi_{-} + \sqrt{ \xi_{- }^2+ \xi_{xy}^2 } \right)^2 }}.
\end{pmatrix}.
\end{equation}
\noindent is chosen such that $ V \hat{H}_{\text{Kinetic}} V^{\dagger}$ is diagonal. It can be recast as
\noindent \begin{equation}
\label{}
V(\pmb{k})= \frac{1}{\sqrt{2}}
\begin{pmatrix}
- \sqrt{1-\cos{\phi(\pmb{k})}} & \sqrt{1+\cos{\phi(\pmb{k})}}\\
\sqrt{1+\cos{\phi(\pmb{k})}} & \sqrt{1- \cos{\phi(\pmb{k})}}
\end{pmatrix},
\end{equation}
\noindent where
\noindent \begin{equation}
\cos \phi(\pmb{k})= \frac{\xi_{-}(\pmb{k})}{\sqrt{\xi_{-}^{2}+ \xi_{xy}^{2}}}.
\end{equation}
\noindent The transformation on $\hat{H}_{\text{Pair}}$ is formally equivalent to the improper rotation
\noindent \begin{equation}
\label{}
\vec{d}'(\pmb{k})=
\begin{pmatrix}
-\cos(\theta) & 0 & \sin(\theta) \\
0 & -1 & 0 \\
\sin(\theta) & 0 & \cos{\theta}
\end{pmatrix}
\begin{pmatrix}
0 \\
0 \\
d
\end{pmatrix}
\end{equation}
\noindent of $\vec{d}(\pmb{k})$ provided that $\theta(\pmb{k})= \phi(\pmb{k})+ \pi$.
\section{The five-orbital model and its solution}
\label{SI:5_orbtl}
\subsection{Model}
We proceed to describe the effective $t-J_{1}-J_{2}$ model we used in our calculations.
These were done for an effective 1-Fe unit cell or equivalently in an unfolded BZ~\cite{S_Nica_Yu_Si:Unpublished_2015}.
To simplify our analysis,
we consider the kinetic part for all $d$ orbitals but restrict the exchange couplings
and hence the pairing interactions to $d_{xz}, d_{yz}$, and $d_{xy}$ orbitals only.
Specifically, the Hamiltonian in the orbital basis is given by
\begin{align}
\label{Eq:t_J_Hamiltonian}
H= - & \sum_{i<j} (t_{ij}^{\alpha\beta}c_{\alpha}^{\dagger}c_{\beta} + H.C.) + \sum_{i,\alpha} \left( \epsilon_{i\alpha} - \mu \right) n_{i}+ \sum_{<ij>,\alpha, \beta}J_{1}^{\alpha\beta} \left( \pmb{S}_{i\alpha} \cdot \pmb{S_{j\beta}} - \frac{1}{4}n_{i\alpha}n_{j\beta} \right) + \notag \\
& + \sum_{<<ij>>,\alpha, \beta}J_{2}^{\alpha\beta} \left( \pmb{S}_{i\alpha} \cdot \pmb{S_{j\beta}} - \frac{1}{4}n_{i\alpha}n_{j\beta} \right)
\end{align}
\noindent \begin{equation}
\label{Eq:Orbital_aniso}
J^{xz/yz}_{1,2} \neq J^{xy}_{1,2}
\end{equation}
\noindent where $\alpha,\beta \in \{1,2,3,4,5 \}$ are orbital indices representing all five $d_{xz}$, $d_{yz}$, $d_{x^2-y^2}$, $d_{xy}$, and $d_{3z^2-r^2}$ orbitals, $\epsilon_{i}$ are the on-site energies, and $\mu$ is the chemical potential. The local moments can be written as $\pmb{S}_{i \alpha}=\sum_{ss'}\frac{1}{2}c_{i \alpha s}^{\dagger}\pmb{\sigma}_{ss'}c_{i \alpha s'}$ in terms of the conduction electrons. We first consider only intra-orbital exchange ($\alpha=\beta$) and set $J^{x^2-y^2}_{1(2)}=J^{3z^2-r^2}_{1(2)}=0$. We consider general exchange couplings which reflect the possible orbital selectivity by allowing $J_{xz,xz}=J_{yz,yz} \neq J_{xy,xy}$ (Eq. \ref{Eq:Orbital_aniso}).
The density of states projected onto the $3d_{xy}$ orbital is considerably narrower than that projected onto the $3d_{xz}/3d_{yz}$ orbitals
(with a ratio of about $0.6$ for the alkaline iron selenides)
~\cite{S_Yu_Zhu_Si:PRL_2013}.
Using the square of this ratio as a rough guide, we can expect $A_O
=J_{1}^{xy}/J_{1}^{xz/yz}=J_{2}^{ xy}/J_{2}^{xz/yz}$ to be significantly smaller than $1$ in the iron selenides.
\subsection{Solution method and superconducting pairing phase diagram}
The interactions in Eq. \ref{Eq:t_J_Hamiltonian} can be decomposed into nearest-neighbor (NN) and next-nearest neighbor (NNN) singlet pairing terms. The double occupancy constraint can be incorporated in practice through a band renormalization by the doping factor $\delta= \left|\sum_{i,s} n_{i\alpha s}- 2 \right|$. The pairing Hamiltonian can be solved numerically in a 1-Fe unit cell calculation by varying the exchange couplings.
For more details on the method, we refer the reader to Refs. \onlinecite{Yu_Nat_Comm:2013,Yu_Zhu_Si:2014}.
As specified above,
an exchange orbital anisotropy factor
is defined as $A_{O}=J_{1}^{xy}/J_{1}^{xz/yz}=J_{2}^{ xy}/J_{2}^{xz/yz}$
and an orbital-independent NN-NNN exchange anisotropy factor $A_{L}=J_{1}^{\alpha}/J_{2}^{\alpha}$
for all three non-zero intra-orbital exchange couplings for $d_{xz}$, $d_{yz}$, and $d_{xy}$.
To explore the zero-temperature superconducting phases corresponding to different classes of Fe-based materials
we consider the associated electron dispersions for
$ \textnormal{K}_{\textnormal{y}}\textnormal{Fe}_{\textnormal{2-x}}\textnormal{Se}_{\textnormal{2}} $, iron pnictides
and single-layer FeSe. We subsequently tune the exchange couplings for various NN-NNN and orbital anisotropy ratios
($A_L$ and $A_O$) and determine the real-space pairing functions.
This leads to the pairing phase diagram in the $A_L-A_O$ parameter space. The results for the electronic dispersions of the alkaline iron selenides and iron pnictides are shown in the main text as Figs.~2 (a) and (b), respectively. Those for the case of the single-layer FeSe is shown here, in Fig.~\ref{Phase_diagram_single_layer_FeSe_PNAS_11_10_2016}. For the case of the alkaline iron selenides, a cut along the $A_L$ axis for a fixed $A_O=0.3$ is shown in Fig.~\ref{Fig:Pairing_amplitude_A_O=0.5_A_L=1.3}.
\subsection{Effects of inter-orbital exchange interactions}
Throughout the main text, the discussion has been centered on cases with only
intra-orbital
$J$'s and their consequence on the pairing amplitudes.
To analyze the robustness of our results, we turn to
calculations which allow for
\emph{inter-orbital} NN and NNN ($J_{1}$ and $J_{2}$, respectively) exchange interactions
between the dominant $d_{xz}, d_{yz}$, and $d_{xy}$ orbitals, in addition to the intra-orbital interactions
considered in Eq.~\ref{Eq:t_J_Hamiltonian}. More specifically, we introduce
\noindent \begin{align}
J^{xz/yz}_{1} = & J^{yz/xz}_{1} = 0 \notag \\
J^{xz/xy}_{1} = & \sqrt{ J^{xz/xz}_{1} \times J^{xy/xy}_{1} } \notag \\
J^{yz/xy}_{1} = & J^{xz/xy}_{1}
\end{align}
\noindent and
\noindent
\begin{align}
J^{xz/yz}_{2} = & J^{xz/xz}_{2} = J^{yz/yz}_{2} \notag \\
J^{xz/xy}_{2} = & \sqrt{ J^{xz/xz}_{2} \times J^{xy/xy}_{2} } \notag \\
J^{yz/xy}_{2} = & J^{xz/xy}_{2}
\end{align}
Crucially, these conditions allow the inter-orbital coupling constants to be consistent with the underlying
super-exchange
mechanism. Thus, the absence of NN hopping between $d_{xz}$ and $d_{yz}$ orbitals~\cite{S_Yu_Nat_Comm:2013,Yu_Zhu_Si:2014} implies vanishing $J^{xz/yz}_{1}, J^{yz/xz}_{1}$.
Similarly, the $xz-xy$ and $yz-xy$ super-exchange coupling constants, involving the square-root terms, reflect the influence of orbitally-selective correlations.
In Figs.~\ref{Fig:K_Prng_ampltds_intra} (a) and (b), we show the amplitudes for the leading \emph{intra-orbital} pairing channels in the case of the alkaline iron selenides, for $A_{O}=0.2$ and $J_{2}=1$, with and without inter-orbital exchange interactions. As these figures clearly show, no significant changes occur. Similar pictures emerge for virtually all values of $A_{O}$ and $A_{L}$ shown in the phase diagram in Fig.~2 (a) in the main text.
In Figs.~\ref{Fig:K_Prng_ampltds_inter} (a) and (b), we plot
one of the leading \emph{inter-orbital}
pairing amplitudes for the alkaline iron selenides,
in the $d_{xy} \otimes \tau_{1}, A_{1g}$ channel,
with and without inter-orbital exchange couplings. In either case, the leading inter-orbital pairing amplitude is roughly two orders of magnitude smaller than the leading intra-orbital amplitude.
(The numerical accuracy of our calculation for the pairing amplitudes is about $10^{-4}$.)
The same conclusion is drawn throughout the phase diagram.
Based on these results and similar ones for the Fe-pnictide cases, we conclude that
the
inter-orbital exchange interactions
have
a negligible effect on the pairing amplitudes within our model.
\section{Dynamical spin susceptibility and neutron resonance}
\label{Sec:Grnl_rqrmnt_spn_rsnnc}
\subsection{General formulation}
In the single-band BCS case, the
bare
contribution to the dynamical spin susceptibility (see Eq.~\ref{Eq:RPA_suscept} for the multi-orbital case) depends~\cite{S_Eschrig:Adv_Phys_2006, S_Fong_et_al:PRL_1995} on terms like
\begin{align}
\label{Eq:BCS_RPA_suscept}
\chi_{0}(\pmb{q},\omega)=\frac{1}{N}\sum_{\pmb{k}} \bigg[ \frac{1}{2} \bigg( 1- \frac{\epsilon_{\pmb{k}+\pmb{q}} \epsilon_{\pmb{k}} + \Delta_{\pmb{k}+\pmb{q}} \Delta_{\pmb{k}} }{E_{ \pmb{k} + \pmb{q} } E_{\pmb{k}} } \bigg) \frac{ f( E_{ \pmb{k} + \pmb{q} }) + f(E_{\pmb{k}}) -1 }{ \omega - ( E_{ \pmb{k} + \pmb{q} }+ E_{ \pmb{k} } ) + i0^{+} } + ... \bigg] ,
\end{align}
\noindent where $\epsilon$'s and $E$'s are the free particle and the BdG quasi-particle dispersions respectively.
The existence of a sharp feature in the RPA
dynamical spin susceptibility
below the particle-hole threshold
(given roughly by twice the characteristic gap magnitude $2\Delta$) is related to the sign
of the $\Delta_{\pmb{k}+\pmb{q}} \Delta_{\pmb{k}}$ term
in the spin (time-reversal-odd) coherence factor in Eq.~\ref{Eq:BCS_RPA_suscept}.
Close to the Fermi surface, when the sign is positive,
the coherence factor suppresses
$\chi_{0}(\pmb{q},\omega)$ and, consequently,
inhibits the appearance of a resonance.
By contrast, when $\Delta_{\pmb{k}+\pmb{q}}$ and $\Delta_{\pmb{k}}$ have opposite signs, the resonance can form
at an energy below $2\Delta$.
In the present multi-orbital model,
the bare dynamical spin susceptibility is defined as
\noindent \begin{equation}
\label{Eq:RPA_suscept}
\chi_0(\pmb{q}, i\omega_n)=\sum_{\alpha \beta} \chi_{0 \alpha \beta}(\pmb{q}, i\omega_n),
\end{equation}
\noindent where
\noindent \begin{equation}
\chi_{0, \alpha \beta}(\pmb{q}, i\omega_n)=\int_{1}^{1/T} d\tau e^{i \omega_{n} \tau } \left< \mathcal{T}_{\tau}
\left[ S^{-}_{ \pmb{q}\alpha}(\tau)S^{+}_{ \pmb{-q}\beta}(0) \right] \right>.
\nonumber
\end{equation}
\noindent The interaction corrected susceptibility is then
\noindent \begin{align}
\chi_{\alpha \beta}(\pmb{q}, i\omega_n)=\sum_{\gamma} & \left[ \textbf{I} + J(\pmb{q}) \sum_{\delta \mu}
\chi_{0, \delta \mu} (\pmb{q}, i\omega_n) \right]^{-1}_{\alpha \gamma} \times \notag \\
& \times \chi_{0, \gamma \beta} (\pmb{q}, i\omega_n),
\nonumber
\end{align}
\noindent where
\noindent \begin{equation}
J(\pmb{q})=\frac{J_1}{2} \left( \text{cos} q_x + \text{cos} q_y \right) + J_2 \text{cos} q_x \text{cos} q_y.
\nonumber
\end{equation}
In our case, the intraband pairing component has a sign change across the electron Fermi pockets at the BZ edges.
This implies that the corresponding component of the bare susceptibility $\chi_0$ will dominate the final contribution to
the imaginary part of the renormalized spin susceptibility, ${\rm Im}\chi$, at the wavevector $(\pi,\pi/2)$,
which spans across the two electron Fermi pockets. We
discuss this issue further in the next subsection.
\subsection{Spin resonance in the alkaline iron selenides}
Here
the dynamical spin susceptibility of interest is near the wave vector $\pmb{q}$
which connects the two electron pockets near the BZ boundaries $(\pi,0)$ and $(0,\pi)$ [Fig.~4a, main text].
While both the intraband and interband components of the pairing function are crucial for the overall properties of the $s \tau_3$ pairing state,
as far as the spin resonance is concerned, the involved electron Fermi pockets
near $(\pi,0)$ and $(0,\pi)$
belong to only one band.
We can then treat the ratio of the interband pairing amplitude to the separation of the energies between the neighboring (normal state)
energy bands as a perturbation. In this way, we obtain a simplified expression for the leading term of the dynamical susceptibility,
which links the spin resonance with the sign change of the intraband
component of the pairing function.
In the band basis, the bare dynamical spin susceptibility is written as
\noindent \begin{equation}
\label{Eq:chi_0_multiband}
\chi_{0}(\pmb{q}, i\omega_n)=\frac{1}{N} \sum_{\pmb{k}} \sum_{a,b} \mathfrak{F}_{\pmb{k},\pmb{q},a,b} \frac{f(E_{\pmb{k},a})+f(E_{\pmb{k+q},b})-1}{i\omega_n-E_{\pmb{k},a}-E_{\pmb{k}+\pmb{q},b}},
\end{equation}
\noindent where $a$ and $b$ run over all the
BdG quasiparticle bands, and $\mathfrak{F}_{\pmb{k},\pmb{q},a,b}$ is a
prefactor
with the following generic expression,
\noindent \begin{equation}
\label{Eq:Coherence_Factor}
\mathfrak{F}_{\pmb{k},\pmb{q},a,b} = \left[ \sum_{B,D}\tilde{V}_{BD,\pmb{k},\pmb{q}} \bar{U}^\star_{\pmb{k}(B\downarrow,a)} \bar{U}^\star_{\pmb{k+q}(D\uparrow,b)} \right]
\left[ \sum_{A,C} \tilde{V}_{AC,\pmb{k},\pmb{q}} \bar{U}_{\pmb{k}(A\downarrow,a)} \bar{U}_{\pmb{k+q}(C\uparrow,b)} - \tilde{V}^\star_{AC,\pmb{k},\pmb{q}} \bar{U}_{\pmb{k}(A\uparrow,a)} \bar{U}_{\pmb{k+q}(C\downarrow,b)} \right] .
\end{equation}
\noindent Here $A$-$D$ are the indices of the bands in the normal state.
$\tilde{V}_{AC,\pmb{k},\pmb{q}}=\sum_{\alpha} V_{\alpha A}(\pmb{k}) V^\star_{\alpha C}(\pmb{k+q})$ is a
factor
associated with the canonical transformation $V(\pmb{k})$ from the orbital basis to the band basis.
This factor describes the band-dependent contribution to the spin operator and
have the same form in the normal and superconducting states.
$\bar{U}_{\pmb{k}(A\sigma,a)}$ is a matrix element of the Bogoliubov transformation that diagonalizes the pairing Hamiltonian in the band basis (see below).
We denote by ``$+$"
the (normal state) band that crosses the Fermi energy near $(\pi,0)$ and $(0,\pi)$ [{\it cf.} Fig.~4a, main text].
In general, the interband pairing amplitude will be small compared to the separations of this band from the other bands near that part of the
1-Fe BZ.
We can then simplify the analysis
by considering the $+$ band along with only a second band, denoted by ``$-$".
The effective Hamiltonian reads
\begin{equation}
\hat{H}=
\begin{pmatrix}
\epsilon_+ & 0 & \Delta_{++} & \Delta_{+-} \\
0 & \epsilon_- & \Delta_{+-} & \Delta_{--} \\
\Delta_{++} & \Delta_{+-} & -\epsilon_+ & 0 \\
\Delta_{+-} & \Delta_{--} & 0 & -\epsilon_-
\end{pmatrix},
\end{equation}
\noindent where $\epsilon_{+/-}$ are the energies of the bands in the normal state;
$\Delta_{++}$ and $\Delta_{--}$ are the intraband pairing components;
$\Delta_{+-}$ is the interband pairing component,
satisfying the condition $ | \Delta_{+-} | \ll |\epsilon_{+}-\epsilon_{-}|$.
The Hamiltonian can be diagonalized by a Bogoliubov transformation $\bar{U}_{\pmb{k}}$,
and we obtain the excitation energies of the
BdG energy dispersion
\noindent \begin{equation}
E_{\pm}(\pmb{k})=\sqrt{\frac{1}{2} \left\{ \epsilon_+^2+\epsilon_-^2+\Delta_{++}^2+\Delta_{--}^2+2\Delta_{+-}^2 \pm \sqrt{\left[ \epsilon_+^2-\epsilon_-^2+\Delta_{++}^2-\Delta_{--}^2 \right]^2 + 4\Delta_{+-}^2 \left[ (\epsilon_+-\epsilon_-)^2 +(\Delta_{++}-\Delta_{--})^2 \right]} \right\}}.
\end{equation}
\noindent We stress again that
we are focusing on the pairing near the electron pockets centered at $(\pi,0)$ and $(0,\pi)$ only.
Because, in the 1-Fe BZ, only one band crosses the Fermi level at these electron Fermi pockets~\cite{S_Yu_Nat_Comm:2013} and the energy separation between this band and nearby hole band is about $100$ meV,\cite{S_MYi.2015} which is much larger than the pairing functions,
there is a strong constraint to the summations in Eqs.~\ref{Eq:chi_0_multiband} and \ref{Eq:Coherence_Factor}: In Eq.~\ref{Eq:chi_0_multiband},
the relevant term
of the dynamical spin susceptibility
is now the one with $a=b=+$, and in Eq.~\ref{Eq:Coherence_Factor}
the term with $A=B=C=D=+$ contributes the most to the
prefactor because the energy separation of the bands $|\epsilon_+-\epsilon_-|$ is much larger than the pairing components.
As a result, the leading term of the bare dynamical spin susceptibility
reads
\noindent \begin{equation}
\chi_{0}(\pmb{q}, i\omega_n) \sim \frac{1}{N} \sum_{\pmb{k}}
\tilde{V}^2_{++,\pmb{k},\pmb{q}} \bar{U}^\star_{\pmb{k}(+\downarrow,+)} \bar{U}^\star_{\pmb{k+q}(+\uparrow,+)} \left[ \bar{U}_{\pmb{k}(+\downarrow,+)} \bar{U}_{\pmb{k+q}(+\uparrow,+)} - \bar{U}_{\pmb{k}(+\uparrow,+)} \bar{U}_{\pmb{k+q}(+\downarrow,+)} \right]
\frac{f(E_+(\pmb{k}))+f(E_+(\pmb{k+q}))-1}{i\omega_n-E_+(\pmb{k})-E_+(\pmb{k}+\pmb{q})}.
\end{equation}
We define the small parameter $\eta \equiv
\Delta_{+-}/\sqrt{\epsilon^2_- - \epsilon^2_+ + \Delta_{--}^2 -\Delta_{++}^2} $,
and expand
the BdG energy dispersion
$E_{\pm}(\pmb{k})$ and the matrix elements of the Bogoliubov transformation $\bar{U}_{\pmb{k}(A\sigma,a)}$ in terms of $\eta$. We obtain,
\noindent\begin{align}
& E_+(\pmb{k}) = E_{+0} + \frac{\eta^2\left[ (\epsilon^2_+ + \Delta^2_{++})-(\epsilon^2_- + \Delta^2_{--}) + (\epsilon_+-\epsilon_-)^2 +(\Delta_{++}-\Delta_{--})^2 \right]}{2E_{+0}} + O(\eta^3)\\
& \bar{U}_{\pmb{k}(+\uparrow,+)} = \sqrt{\frac{E_{+0}+\epsilon_+}{2E_{+0}}} + O(\eta^2)\\
& \bar{U}_{\pmb{k}(+\downarrow,+)} = \sqrt{\frac{E_{+0}-\epsilon_+}{2E_{+0}}} + O(\eta^2)\\
& \bar{U}_{\pmb{k}(-\uparrow,+)} = \frac{\eta}{\sqrt{\epsilon^2_- - \epsilon^2_+ + \Delta_{--}^2 -\Delta_{++}^2}}\left[ (\epsilon_- - E_{+0}) \sqrt{\frac{E_{+0}-\epsilon_+}{2E_{+0}}} -\Delta_{--} \sqrt{\frac{E_{+0}+\epsilon_+}{2E_{+0}}} \right]+ O(\eta^2)\\
& \bar{U}_{\pmb{k}(-\downarrow,+)} = \frac{\eta}{\sqrt{\epsilon^2_- - \epsilon^2_+ + \Delta_{--}^2 -\Delta_{++}^2}}\left[ -\Delta_{--} \sqrt{\frac{E_{+0}-\epsilon_+}{2E_{+0}}} -(\epsilon_- + E_{+0}) \sqrt{\frac{E_{+0}+\epsilon_+}{2E_{+0}}} \right]+ O(\eta^2),
\end{align}
\noindent where $E_{+0}(\pmb{k})=\sqrt{\epsilon_+^2(\pmb{k})+\Delta_{++}^2(\pmb{k})}$.
This leads to the following form for the leading term of $\chi_0$:
\noindent \begin{equation}
\label{Eq:chi0_final_form}
\chi_{0} (\pmb{q}, i\omega_n) \sim \frac{1}{N} \sum_{\pmb{k}}
\tilde{V}^2_{++,\pmb{k},\pmb{q}} \frac{1}{2} \bigg( 1- \frac{\epsilon_{+,\pmb{k}+\pmb{q}} \epsilon_{+,\pmb{k}} + \Delta_{++,\pmb{k}+\pmb{q}} \Delta_{++,\pmb{k}} }{E_{+0,\pmb{k} + \pmb{q} } E_{+0,\pmb{k}} } \bigg)
\frac{f(E_{+0}(\pmb{k}))+f(E_{+0}(\pmb{k+q}))-1}{i\omega_n-E_{+0}(\pmb{k})-E_{+0}(\pmb{k}+\pmb{q})} + O(\eta).
\end{equation}
Here, the prefactor $\tilde{V}^2_{++,\pmb{k},\pmb{q}}$ is the same as in the normal state;
it simply weighs the contribution of this particular band to the p-h excitation in the spin channel at these wave vectors.
In Eq.~\ref{Eq:chi0_final_form}, the effect of superconductivity appears through the factor in the big brackets,
which is essentially the same as the spin coherence factor
of the 1-band case, given in Eq.~\ref{Eq:BCS_RPA_suscept} (an analytical continuation $i\omega_n\to\omega+i0^+$ is needed to compare the two
equations).
Similar to the usual case
~\cite{S_Eschrig:Adv_Phys_2006},
a sharp resonance appears in the imaginary part of the dynamical spin susceptibility $\chi''(\pmb{q}, \omega)$
when there is a sign change in the intraband pairing components $\Delta_{++}(\pmb{k})$ between the two electron pockets.
This conclusion is consistent with our numerical result for $\chi''(\pmb{q}, \omega)$, shown in Fig.~3 of the main text.
\noindent \begin{figure}[!h]
\centering
\includegraphics[width=0.4\columnwidth]{Phase_diagram_single_layer_FeSe_PNAS_11_10_2016}
\caption{ Phase diagram based on the leading pairing amplitudes given by self-consistent calculations using tight-binding parameters appropriate to single-layer FeSe. The tight-binding parameters used can be found in
Ref.~\onlinecite{Yu_Nat_Comm:2013}. The blue shaded areas correspond to dominant pairing channels with an $s_{x^2y^2}$ form factor while the red shading covers those with a $d_{x^2-y^2}$ form factor. The continuous line separates regions where the pairing belongs to the $A_{1g}$ and the $B_{1g}$ representations respectively. The $1 \times 1$ matrix in the $d_{xy}$ subspace is represented by $\pmb{1}_{xy}$.}
\label{Phase_diagram_single_layer_FeSe_PNAS_11_10_2016}
\end{figure}
\noindent \begin{figure}[!h]
\centering
\includegraphics[width=0.4\columnwidth]{K_Pairing_Amplitudes_J_2=1_5_A_O=0_3_A_L=0_9_complete}
\caption{Leading pairing amplitudes (vertical axis) for a dispersion typical of alkaline iron selenides for fixed $J_2=1.5, A_O=0.3$ and varying NN-NNN ratio $A_L$ (horizontal axis). The $\tau$ label indicates a dominant $d_{xz}, d_{yz}$ sector while $ \mathbf{1}_{xy}$ marks a $d_{xy}$ dominant pairing. For $ 0.8 \leq A_L \leq 0.94 $ the leading pairing is in the $s\tau_{3}$ channel shown in dark filled squares. Note that the reduced parameter space for the $s\tau_{3}$ is due to the proximity to the phase boundary and for lower values of $A_O$ the range over which this pairing leads is increased.}
\label{Fig:Pairing_amplitude_A_O=0.5_A_L=1.3}
\end{figure}
\noindent \begin{figure}[h!]
\centering
\subfloat[]{\includegraphics[width=0.4\columnwidth]{K_Pairing_Amplitudes_J_2=1_A_O=0_2_01_11_2017_inter_ON}} \\
\subfloat[]{\includegraphics[width=0.4\columnwidth]{K_Pairing_Amplitudes_J_2=1_A_O=0_2_01_11_2017_inter_OFF}}
\caption{Leading intra-orbital pairing amplitudes (vertical axis) for a dispersion typical of alkaline iron selenides for fixed $J_2=1, A_O=0.2$ and varying NN-NNN ratio $A_L$ (horizontal axis) \emph{with (a) and without (b) inter-orbital exchange interactions}. As mentioned in the discussion above, no significant changes are observed.
}
\label{Fig:K_Prng_ampltds_intra}
\end{figure}
\noindent \begin{figure}[h!]
\centering
\subfloat[]{\includegraphics[width=0.4\columnwidth]{K_Pairing_Amplitudes_J_2=1_A_O=0_2_01_09_2017_largest_inter_ON}} \\
\subfloat[]{\includegraphics[width=0.4\columnwidth]{K_Pairing_Amplitudes_J_2=1_A_O=0_2_01_09_2017_largest_inter_OFF}}
\caption{Leading inter-orbital pairing amplitude (vertical axis) for a dispersion typical of alkaline iron selenides for fixed $J_2=1, A_O=0.2$ and varying NN-NNN ratio $A_L$ (horizontal axis) \emph{with (a) and without (b) inter-orbital exchange interactions}. As mentioned in the discussion above, no significant changes are observed.
}
\label{Fig:K_Prng_ampltds_inter}
\end{figure}
|
\section{Introduction}
\label{sec: Introduction}
\subsection{Motivation}
\label{subsec: Motivation}
In recent decades, topological order has emerged as a novel paradigm
for describing states of matter. Motivated by the study of the
fractional quantum Hall effect and chiral spin liquids, theoretical
investigations uncovered a rich landscape of topologically ordered
phases in two spatial dimensions. The unifying features common to all
phases in this landscape are 1) the degeneracy of the ground state
when the system is defined on a manifold with nonzero
genus~\cite{Wen89b}, and 2) the (intimately related) existence of
fractionalized excitations in the gapped bulk~\cite{Oshikawa06}. The
theoretical understanding of these topologically ordered phases has
been placed on a firm mathematical footing rooted in the apparatus of
modular tensor
categories%
~\cite{Friedan87,Froehlich88,Moore1989,Froehlich90a,Etingof2015}.
While numerous problems remain open to investigation, such as the
inclusion of symmetries%
~\cite{Barkeshli14,Teo15,Barkeshli16,Thorngren16} and the
description of topological phases
starting from interacting electrons%
~\cite{Gu14,Cheng15,Kapustin15,Gaiotto16,Ware16,Tarantino16,Williamson16},
this mathematical framework provides an indispensible point of
reference in the ongoing effort to understand strongly interacting
topological states of matter in two spatial dimensions.
The theoretical proposal~\cite{Fu07,Moore07} and experimental
discovery~\cite{Hsieh08,Chen09,Hsieh09,Hsieh09b} of three-dimensional
topological insulators protected by time-reversal symmetry (TRS)
underscores the natural question of whether a similar understanding of
topological order in three spatial dimensions could be achieved.
Numerous examples of topologically ordered phases in three spatial
dimensions are known, of which discrete gauge theories and their
twisted counterparts are perhaps the most
elementary~\cite{Dijkgraaf90,Wang15,Wan15,Else17}. There also exists
a procedure, the Crane-Yetter/Walker-Wang
construction~\cite{Crane93,Walker12,Wang16,Williamson17}, that can be
used to build certain topological phases in three spatial dimensions.
Despite this progress, the question of what kinds of strongly
interacting topological phases can exist in
(3+1)-dimensional spacetime (3D)
is far from settled. This is especially true of non-Abelian topological
orders. Furthermore, it is (for the most part) unclear how such
topological phases emerge as low-energy descriptions of condensed
matter systems, which are conventionally made of electrons and spins
that interact in decidedly non-exotic ways.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Figures1}
\caption{
(Color online)
Analogy between the quantum Hall effect (QHE)
in (2+1)-dimensinal spacetime
and the $\mathbb{Z}^{\,}_{2}$ topological insulator (TI)
in (3+1)-dimensional spacetime.
In the quantum-Hall setting (left), the
boundary between the topological phase (blue) and the vacuum (beige)
hosts an edge state that realizes one chiral sector of a conformal
field theory (red). In the TI setting (right), a domain wall
separating two regions with opposite time-reversal breaking fields
(labeled ``$\uparrow$'' and ``$\downarrow$'') hosts the same chiral mode
that appeared on the edge of the quantum Hall system. }
\label{fig: QHE-TI analogy}
\end{figure}
In this paper, we propose a family of non-Abelian topological phases
in (3+1)-dimensional spacetime. This family of phases can be viewed as
completing the following series of analogies between topological
phases in two- and three-dimensional space. We begin with the integer
quantum Hall effect (IQHE)~\cite{QHE-Book}.
This is a topological phase in (2+1)-dimensional spacetime (2D)
whose electromagnetic response is encoded by
a $U(1)$ Chern-Simons effective action at level
1~\cite{Wen89,Froehlich91a,Froehlich91b}. It features a unique ground
state on the torus, and, on a manifold with boundary, has gapless
chiral Dirac fermion edge states~\cite{Halperin82} described by the
affine Lie algebra $u(1)^{\,}_{1}$
(see Fig.\ \ref{fig: QHE-TI analogy}).
The noninteracting
$\mathbb{Z}^{\,}_{2}$ topological insulator (TI) in 3D
can be viewed as inheriting many of its defining
properties from the IQHE. For example, although the noninteracting TI
respects time-reversal symmetry (TRS) while the IQHE does not, the
gapped surface states that emerge when TRS is broken on the surface of
the TI feature a Hall response that is exactly half of what is
expected in the IQHE case~\cite{Fu07b}. Thus, a ``magnetic domain
wall'' that separates regions with opposite TRS-breaking fields on the
TI surface binds the same gapless chiral Dirac fermion mode that
constitutes the edge state of the IQH system (see Fig.\
\ref{fig: QHE-TI analogy}). This is a direct consequence of the axion
electromagnetic response (signaled by a $\theta$ term in the effective
action) in 3D that
characterizes the bulk of the TI~\cite{Wilczek87,Qi08,Essin09}.
When TRS is preserved, the noninteracting $\mathbb{Z}^{\,}_{2}$
TI features a single massless Dirac fermion on its surface. In pure 2D,
the existence of a TRS theory of a single massless Dirac fermion is
forbidden by the fermion doubling theorem~\cite{Nielsen81}. However,
on the surface of a $\mathbb{Z}^{\,}_{2}$ TI, its presence is necessary
to ensure that the TRS-breaking surface contribution of the $\theta$
term does not spoil TRS on the surface. If TRS is broken on the surface,
then a mass term for the Dirac fermion is symmetry-allowed,
and the aforementioned surface quantum Hall effect develops.
In this sense, the Dirac fermion surface states
of the $\mathbb{Z}^{\,}_{2}$ TI are anomalous,
and their gaplessness is protected by TRS.
Recent work on so-called fractional TIs (FTIs)
in 3D has borne out this analogy to the interacting setting.
Indeed, these FTIs can be \textit{defined}
as systems in (3+1)-dimensional spacetime with
TRS whose bulk axion electromagnetic response is characterized by axion angles
$\theta$ that are rational multiples of $\pi$. Consistency with TRS
then demands the presence of topological order in the bulk%
~\cite{Maciejko10,Swingle11}. In the case $\theta=\pi/k$ with
$k\in\mathbb{Z}$, one finds that breaking TRS on the surface of an FTI
yields a gapped surface state with Hall conductivity
$\sigma^{\,}_{\mathrm{xy}}=(1/2k)\,e^{2}/h$.
Consequently, a magnetic domain wall on the
surface binds a chiral Luttinger-liquid mode described by the affine
Lie algebra $u(1)^{\,}_{k}$, which is precisely the edge state of the
$\nu=1/k$ Laughlin state in the fractional quantum Hall effect (FQHE)%
~\cite{Wen91c}. Moreover, preserving TRS on the surface necessitates
the presence of fractionalized gapless excitations
on the surface~\cite{Swingle11}.
Hence, FTIs feature fractionalized analogues of the anomalous gapless
surface states of the noninteracting $\mathbbm Z^{\,}_{2}$ TI.
Given the two preceding analogies between the (F)QHE
in 2D
and the (F)TI in 3D,
the following natural question arises.
Are there TI-like analogues in 3D
of the 2D non-Abelian quantum Hall states?
One can focus, for example, on asking this question for the
(bosonic) Read-Rezayi quantum Hall sequence in (2+1)-dimensional
spacetime~\cite{Moore91,Read99}. These non-Abelian topological
phases are described by an $SU(2)$ Chern-Simons term at level $k$, and
feature chiral edge states described by the affine Lie algebra
$su(2)^{\,}_{k}$~\cite{Wen91b,Wen91c}. Hence, the analogous
topological phase in (3+1)-dimensional spacetime
would need to satisfy the following three properties.
First, it should be time-reversal invariant in the bulk.
Second, it should be topologically ordered, in the sense
that the ground state manifold on the three-torus must have dimension
greater than one. Third, a domain wall between regions on the surface
in which TRS is broken in opposite ways should bind a chiral $su(2)^{\,}_{k}$
mode. Fourth, there should be gapless surface states protected by TRS.
Is it possible to construct such a topological phase?
If so, what is the nature of the bulk topological order?
These are the questions we address in this paper.
We address the question of the existence of 3D analogues
of the $su(2)^{\,}_{k}$ fractional quantum Hall states
in 2D
by attempting to build them from scratch. In particular, we employ a coupled-wire
construction based on non-Abelian current algebras to construct a
topological phase with the desired properties. In this approach, the
topological phase in (3+1)-dimensional spacetime
is constructed by coupling together many sub-systems,
each of which lives in (1+1)-dimensional spacetime (1D),
with appropriate many-body interactions.
Coupled-wire constructions have been used to construct a
variety of strongly-correlated phases
in 2D,
including non-Fermi liquids~\cite{Emery00,Mukhopadhyay01,Vishwanath01}
as well as Abelian and non-Abelian quantum Hall states%
~\cite{Poilblanc87,Yakovenko91,Lee94,Sondhi01,Kane02,Teo14,Mong14,Neupert14}.
Moreover, this approach has recently been generalized to
3D, yielding a variety of phases including Weyl
semimetals~\cite{Vazifeh13,Meng15b},
fractional topological insulators~\cite{Sagi15b},
and strongly-correlated phases described by
emergent Abelian gauge theories~\cite{Iadecola16}. The utility of
this approach lies in the fact that numerous analytic techniques exist
for quantum field theories in (1+1)-dimensional spacetime,
enabling the description of a wide variety of strongly interacting states
of matter in a controlled manner. We also argue by example in this work
that the coupled-wire approach can be used as a means to search for and
characterize candidates for new topological phases of matter,
like the family of 3D $su(2)^{\,}_{k}$ phases constructed here.
\subsection{Outline and summary of results}
\label{subsec: Outline and summary of results}
We now provide an overview of the organization of the paper and
summarize the results.
In Sec.~\ref{sec: Non-Abelian bosonization of a single wire}, we
review how to bosonize a multi-flavor fermionic wire in terms of the
currents associated with the non-Abelian internal symmetry group of
the wire~\cite{Witten84}. This bosonization scheme has been used to
address a wide variety of physical problems in 1D,
including the multichannel Kondo
effect~\cite{Affleck90,Affleck91a,Affleck91b} and marginally-perturbed
conformal field theories (CFTs)~\cite{Tsvelik14}.
In Ref.~\cite{Huang16a} it was also used as a starting point for the
construction of a series of non-Abelian topological phases in
2D.
In Sec.~\ref{subsec: Intra-wire interactions}, we show
how to add intra-wire interactions to drive the fermionic wire to a
strong-coupling fixed point described by an $su(2)^{\,}_{k}$ CFT.
This treatment is crucial for what follows, as these CFTs are used as
building blocks for the coupled-wire constructions of the subsequent
sections; the non-Abelian topologically ordered phases
in 2D and 3D
that we construct later in the paper inherit their
non-Abelian character from the $su(2)^{\,}_{k}$ CFTs.
Next, in Sec.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions},
we describe how to construct non-Abelian topological
phases of matter in 2D
starting from a one-dimensional array of decoupled $su(2)^{\,}_{k}$ CFTs.
This section serves as a prelude to Sec.%
~\ref{sec: Non-Abelian topological order in three dimensions},
where the 3D topological phase
is constructed. While the $su(2)^{\,}_{k}$ topological phases constructed
in Sec.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
are not new, this section serves two important
purposes, on which we now elaborate.
First, Sec.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
establishes the approach we later take to construct the 3D
topological phases
of the following section.
This approach can be described as follows. We use $su(2)^{\,}_{k}$
current-current interactions to couple channels in neighboring wires
that have opposite chirality. These couplings can be viewed
as arising from continuum limits of microscopic
interactions between the spin sectors of neighboring
wires (see, e.g., Refs.~\cite{Huang16b} and \cite{chen17}).
Focusing on the specific example of $su(2)^{\,}_{2}$,
we argue that, in the strong-coupling limit, these interactions
gap the bulk of
the array of coupled wires, leaving chiral $su(2)^{\,}_{2}$ modes on
the boundaries when the model is defined on a cylinder. Once we have
shown how to gap the bulk of the array, we move on to characterize the
bulk topological order within the coupled-wire construction.
(In the quantum Hall
parlance, this topological phase is related to the Moore-Read state
for bosons at filling factor $\nu=1$.) The procedure for doing so
hinges on using the primary operators of the unperturbed CFTs in each
wire to construct nonlocal ``string operators''
that commute with the
interaction term and satisfy a nontrivial algebra among themselves.
These string operators can then be used to determine the topological
ground-state degeneracy of the coupled-wire theory on the torus. More
specifically, these string operators can be used to
construct a representation of the ground-state manifold of the
coupled-wire theory at strong coupling.
\begin{figure}[t]
\includegraphics[width=0.4\textwidth]{Figures2}
\caption{
(Color online)
Two-dimensional cross-section of the array of quantum wires used to
construct the family of non-Abelian topological phases
in (3+1)-dimensional spacetime
studied in this work. Each grey disc centered on a site of the
square lattice $\Lambda$ represents a quantum wire aligned along the
$z$-direction of the ambient three-dimensional space with Cartesian
coordinates $x$, $y$, and $z$. The colored discs within each grey
disc represent normal modes, dispersing along the $z$-direction, that
correspond to different chiral sectors of the $su(2)^{\,}_{k}$
conformal field theory. The binary color code distinguishes between
left- and right-handed sectors. In this example, each wire contains
two right-moving and two left-moving normal modes.
}
\label{fig: general wire setup}
\end{figure}
Second, the calculation of the ground-state degeneracy of the
$su(2)^{\,}_{2}$ topological phase
in 2D
constructed in Sec.\
\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
serves to clarify precisely what is meant by a
``non-Abelian'' topological phase in the context of this paper.
(Moreover, such a calculation was not presented in previous
coupled-wire constructions of similar topological phases, see
e.g., Refs.\ \cite{Teo14} and \cite{Huang16a}.) In particular, we will
show that the algebra of the nonlocal string operators mentioned
previously suggests the algebra of Wilson loops in a
$\mathbb{Z}^{\,}_{2}$ gauge theory. Namely, there are four nonlocal string
operators that break into two sets of anticommuting operators.
Naive intuition derived from
Abelian gauge theory then suggests that the ground-state
degeneracy on the torus should be fourfold. However, one finds that
one of these four putative ground states cannot reside in the
ground-state manifold. The reason for this has deep connections to
the non-Abelian algebra of primary operators in the
CFT~\cite{Moore1989}, and has come up before in less microscopic
studies of related topological phases~\cite{Oshikawa07}. In this way,
we conclude that the topological degeneracy of the $su(2)^{\,}_{2}$
topological phase in 2D
is three, rather than four.
This exclusion of states from the ground-state manifold based on
non-Abelian operator algebras is at the heart of what distinguishes
non-Abelian topological phases from Abelian ones, and appears again
with a vengeance in the (3+1)-dimensional case.
In Sec.~\ref{sec: Non-Abelian topological order in three dimensions},
we attempt a naive generalization of the results of
Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
to (3+1)-dimensional spacetime. The starting point for the 3D
wire construction is a two-dimensional array of
parallel interacting quantum wires that realize $su(2)^{\,}_{k}$ CFTs
at low energies (see Fig.~\ref{fig: general wire setup}). We then couple
the $su(2)^{\,}_{k}$ CFTs in neighboring wires with current-current interactions,
exactly as in
Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}.
When these current-current interactions open a gap in the array of wires,
the result is a quantum phase of matter that supports
\textit{weak} topological order, being simply a stack of many copies of the 2D
phases considered in
Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}.
The $su(2)^{\,}_{2}$ case features an extensive ground-state degeneracy
that is inherited directly from its 2D precursor.
Section~\ref{sec: Non-Abelian topological order in three dimensions}
nevertheless introduces a new ingredient that is important in the subsequent
analysis carried out in
Sec.~\ref{sec: Possible strong non-Abelian topological order in three dimensions},
namely a symmetry that plays the role of TRS in a TI.
In this case, the relevant symmetry is a cousin of TRS.
While the individual interwire couplings
do not respect TRS, they do so up to a translation by one half of
either of the two lattice vectors
that define the two-dimensional array.
[Note that the 2D $su(2)^{\,}_{k}$ topological phases
constructed in Sec.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
break the one-dimensional analogue of this symmetry,
just like the associated fractional quantum Hall states break TRS.]
This ``antiferromagnetic'' realization of time-reversal
symmetry has appeared in other coupled-wire constructions of
TRS topological phases%
~\cite{Mross15,Sahoo15}.
This symmetry plays a similar role to that of TRS
in the TI (see also Refs.~\cite{Mong10,Fang13});
in particular, it protects anomalous gapless surface states,
as we will discuss below.
In Sec.%
~\ref{sec: Possible strong non-Abelian topological order in three dimensions},
we propose a different interwire interaction that precludes the possibility
of viewing the system as a stack of decoupled planes. This interaction is
still based on current-current interactions, but is defined on
\textit{plaquettes} of the square lattice, rather than bonds as in Secs.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions} and
\ref{sec: Non-Abelian topological order in three dimensions}.
We move on to consider the $su(2)^{\,}_{2}$ case in detail,
building a set of nonlocal operators that can be used
to define a manifold of topologically degenerate ground states.
Most of the important aspects of the analysis of Sec.%
~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
carry through to Sec.%
~\ref{sec: Non-Abelian topological order in three dimensions},
albeit with a few crucial modifications
that we discuss in detail as they arise.
\begin{figure}[t]
\includegraphics[width=0.35\textwidth]{Figures3}
\caption{
(Color online)
Visual summary of results for the $su(2)^{\,}_{2}$
topological phase in (3+1)-dimensional spacetime
constructed in this paper. The blue membrane and
orange string in the bulk represent membrane and string operators
constructed in Sec.%
~\ref{sec: Possible strong non-Abelian topological order in three dimensions}.
These membrane and string operators are related to the
``spin-$\frac{1}{2}$'' and ``spin-$1$'' primary fields in the
$su(2)^{\,}_{2}$ conformal field theory. The boundary
of the ``spin-$\frac{1}{2}$'' membrane
$\widehat{\Sigma}^{(\frac{1}{2})}$ is a stringlike excitation,
while the boundaries of the ``spin-$1$'' string
$\widehat{\Gamma}^{(1)}$
represent pointlike excitations.
The surface features a ``magnetic'' domain wall separating two regions in
which the analogue of time-reversal symmetry (TRS) is broken in two
different ways (shown here as a purple and a green region of the surface,
denoting the case where the surface couplings $\lambda>0$ and $\lambda'=0$,
and the case where $\lambda=0$ and $\lambda'>0$, respectively).
The domain wall binds a chiral $su(2)^{\,}_{2}$ mode (red).
}
\label{fig: su(2)_k visual summary}
\end{figure}
The most important modification involved in generalizing the
2D results of Sec.\
\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
to the 3D setting of Sec.\
\ref{sec: Possible strong non-Abelian topological order in three dimensions}
is that one must construct nonlocal ``membrane''
operators'' (see Fig.~\ref{fig: su(2)_k visual summary}), in addition
to the string operators of the two-dimensional case, in order to
characterize the resulting topological order. These membrane
operators can be viewed as the two-dimensional worldsheets of deconfined
stringlike excitations, while the string operators can be viewed as
the one-dimensional worldlines of deconfined pointlike excitations. The
existence of stringlike excitations is crucial for topological order
in 3D, since deconfined point particles in three-dimensional space
must have trivial braiding~\cite{Leinaas77}. (Indeed, this is the
case for the present topological phase: all of the string operators
will be shown to commute with one another.)
The non-Abelian algebra of string
operators in the 2D example generalizes to a
non-Abelian algebra of string and membrane operators in
3D. Moreover, there is also a non-Abelian algebra among the
membranes. Carrying through the naive Abelian gauge theory counting
for the (3+1)-dimensional case and removing states that are excluded
based on the non-Abelian operator algebra, we find 20 degenerate ground
states of the $su(2)^{\,}_{2}$ topological phase in (3+1)-dimensional spacetime
when space is a three-torus.
Finally, we investigate in Sec.\ \ref{sec: Surface theory} the surface
states of the 3D topological phase
constructed in Sec.\
\ref{sec: Possible strong non-Abelian topological order in three dimensions}.
It is readily seen that, when open boundary conditions are imposed in one
of the two directions of the array of wires, there are gapless
$su(2)^{\,}_{k}$ modes left on the exposed two-dimensional surfaces.
The goal of Sec.~\ref{sec: Surface theory} is to better understand the
fate of these ``dangling'' modes when they are coupled by marginally
relevant local interactions. When these interactions break the TRS
analogue, it is straightforward to see that a gapped surface phase
results. It is similarly straightforward to see that a ``magnetic''
domain wall (i.e., a domain wall between two regions of the surface in
which the TRS analogue is broken in different ways) binds a chiral
$su(2)^{\,}_{k}$ current (see Fig.~\ref{fig: su(2)_k visual summary}).
To probe the phase diagram of the surface when the TRS analogue is
preserved, we perform a one-loop renormalization group (RG) analysis
of the surface. We find that all couplings on the surface can flow to
strong coupling simultaneously, even though neighboring couplings do
not commute in general. Furthermore, we allow for surface couplings
that break the $SU(2)$ spin-rotation symmetry, and find
that the surface couplings nevertheless flow towards an
$SU(2)$-symmetric strong-coupling fixed point in a large region of
parameter space. Next, we investigate the nature of the
$SU(2)$-symmetric strong-coupling fixed point by an explicit
self-consistent mean-field calculation for the $su(2)^{\,}_{2}$ case.
We find that the surface states are indeed gapless when the TRS
analogue is imposed, and that they do not break the symmetry
spontaneously (at least at the level of mean-field theory).
While the question of whether or not there exist
gapped, symmetry-preserving surface states of these
topological phases, which would then likely
exhibit surface topological
order~\cite{Keyserlingk13,Vishwanath13,Wang13a,Wang13b,Burnell14,Chen14,Metlitski15,Mross15},
is interesting, we do not
pursue it in this work. However, some investigation along
these lines has been carried out in Ref.~\cite{Sahoo15}.
In summary, we are able to demonstrate in this paper that it is
possible to construct (3+1)-dimensional analogues of the
$su(2)^{\,}_{k}$ non-Abelian (bosonic) quantum Hall states. For the
special case of $su(2)^{\,}_{2}$, we partially characterize the non-Abelian
topological order by explicit calculation, in addition
to characterizing the associated surface states. Thus, we arrive at
the surprising result that non-Abelian topological states of matter based on
conformal field theory can be constructed
in (3+1)-dimensional spacetime.
\section{Non-Abelian bosonization of a single wire}
\label{sec: Non-Abelian bosonization of a single wire}
\subsection{Free-fermion wire}
\label{subsec: Free-fermion wire}
Consider a one-dimensional wire containing $N^{\,}_{\mathrm{c}}$
``colors'' of spinful fermions.
Its action $S^{\,}_{0,\mathrm{wire}}$
is the integral over time $t$ and the coordinate $z$ along the
wire of the Lagrangian density
\begin{equation}\label{eq: single non-abelian wire}
\begin{split}
\mathcal{L}^{\,}_{0,\mathrm{wire}}\:=
2
\sum_{\sigma=\uparrow,\downarrow}
\sum^{N^{\,}_{\mathrm{c}}}_{\alpha=1}
\Big(&
\chi^{*}_{\mathrm{L},\sigma,\alpha}\,
\mathrm{i}\partial^{\,}_{\mathrm{L}}
\chi^{\,}_{\mathrm{L},\sigma,\alpha}
\\
&
+
\chi^{*}_{\mathrm{R},\sigma,\alpha}\,
\mathrm{i}\partial^{\,}_{\mathrm{R}}
\chi^{\,}_{\mathrm{R},\sigma,\alpha}
\Big).
\end{split}
\end{equation}
The derivatives $\partial^{\,}_{\mathrm{M}}\equiv\partial^{\,}_{z^{\,}_{\mathrm{M}}}$
($\mathrm{M}=\mathrm{L},\mathrm{R}$) are taken with respect to the chiral
(light-cone) coordinates
\begin{equation}
z^{\,}_{\mathrm{L}}\equiv t+z,
\qquad
z^{\,}_{\mathrm{R}}\equiv t-z.
\label{eq: def z of M}
\end{equation}
We assume periodic boundary conditions along the wire, i.e.,
in the $z$-direction. The four Grassmann-valued fields
$\chi^{*}_{\mathrm{L},\sigma,\alpha}$,
$\chi^{\,}_{\mathrm{L},\sigma,\alpha}$,
$\chi^{*}_{\mathrm{R},\sigma,\alpha}$,
$\chi^{\,}_{\mathrm{R},\sigma,\alpha}$
are independent of each other.
Such a wire has the internal symmetry
$U(2N^{\,}_{\mathrm{c}})^{\,}_{\mathrm{L}}\times U(2N^{\,}_{\mathrm{c}})^{\,}_{\mathrm{R}}$.
The central idea of the series of coupled-wire constructions presented in this
paper is to decompose the Lie algebra associated with this symmetry using the following identity (or ``conformal embedding'')~\cite{DiFrancesco97},
\begin{equation}
u(2N^{\,}_{\mathrm{c}})^{\,}_{1}=
u(1)\oplus
su(2)^{\,}_{N^{\,}_{\mathrm{c}}}
\oplus
su(N^{\,}_{\mathrm{c}})^{\,}_{2},
\label{eq: conformal embedding}
\end{equation}
where we have employed the notation $g_{k}$ for the affine Lie algebra
at level $k$ associated with the
connected, compact, and simple Lie group $G$.
(For a review of affine Lie algebras, see, e.g., Ref.~\cite{DiFrancesco97}.)
Equation \eqref{eq: conformal embedding}
tells us that the theory~\eqref{eq: single non-abelian wire} has three conserved currents
$j^{\,}_{\mathrm{R}}$,
$J^{a}_{\mathrm{R}}$,
and $\mathsf{J}^{\mathsf{a}}_{\mathrm{R}}$
corresponding to the affine Lie algebras
$u(1)$,
$su(2)^{\,}_{N^{\,}_{\mathrm{c}}}$,
and $su(N^{\,}_{\mathrm{c}})^{\,}_{2}$, respectively.
(Note that, of course, there are analogous conserved currents
$j^{\,}_{\mathrm{L}},J^{a}_{\mathrm{L}},$ and $\mathsf{J}^{\mathsf{a}}_{\mathrm{L}}$
for the left-handed sector.) We use indices
$a=1,2,3$ to label the generators of $SU(2)$ and $\mathsf{a} =
1,\cdots,N^{2}_{\mathrm{c}}-1$ to label the generators of
$SU(N^{\,}_{\mathrm{c}})$. In terms of the complex fermions, these
currents are given by
\begin{subequations}
\label{eq: non-Abelian currents}
\begin{align}
&
j^{\,}_{\mathrm{M}}\:=
\sum_{\sigma=\uparrow,\downarrow}
\sum^{N^{\,}_{\mathrm{c}}}_{\alpha=1}
\chi^{*}_{\mathrm{M},\sigma,\alpha}\,\chi^{\,}_{\mathrm{M},\sigma,\alpha},
\label{eq: non-Abelian currents a}
\\
&
J^{a}_{\mathrm{M}}\:=
\frac{1}{2}
\sum_{\sigma,\sigma^{\prime}=\uparrow,\downarrow}
\sum^{N^{\,}_{\mathrm{c}}}_{\alpha=1}
\chi^{*}_{\mathrm{M},\sigma,\alpha}\,
\sigma^{a}_{\sigma\sigma^{\prime}}\,
\chi^{\,}_{\mathrm{M},\sigma^{\prime},\alpha},
\label{eq: non-Abelian currents b}
\\
&
\mathsf{J}^{\mathsf{a}}_{\mathrm{M}}\:=
\sum_{\sigma=\uparrow,\downarrow}
\sum^{N^{\,}_{\mathrm{c}}}_{\alpha,\alpha^{\prime}=1}
\chi^{*}_{\mathrm{M},\sigma,\alpha}\,
T^{\mathsf{a}}_{\alpha\alpha^{\prime}}\,
\chi^{\,}_{\mathrm{M},\sigma,\alpha^{\prime}},
\label{eq: non-Abelian currents c}
\end{align}
\end{subequations}
with $\mathrm{M}=\mathrm{L},\mathrm{R}$.
The $U(1)$ currents $j^{\,}_{\mathrm{M}}$
with $\mathrm{M}=\mathrm{L},\mathrm{R}$
are associated with charge conservation. The $SU(2)$ currents
$J^{a}_{\mathrm{M}}$
with $\mathrm{M}=\mathrm{L},\mathrm{R}$ and $a=1,2,3$
are associated with the spin-rotation symmetry.
The $SU(N^{\,}_{\mathrm{c}})$ currents $\mathsf{J}^{\mathsf{a}}_{\mathrm{M}}$
with $\mathrm{M}=\mathrm{L},\mathrm{R}$
and $\mathsf{a}=1,\cdots,N^{2}_{\mathrm{c}}-1$
are associated with the color isospin-rotation symmetry. The generators
$\sigma^{a}/2$ of $SU(2)$ and $T^{\mathsf{a}}$ of $SU(N^{\,}_{\mathrm{c}})$ obey
the normalizations and the independent algebras
\begin{subequations}
\begin{align}
\mathrm{tr}\,\left(\sigma^{a}\,\sigma^{b}\right)=
2\delta^{ab},
\qquad
[\sigma^{a},\sigma^{b}]=
2\mathrm{i}\,\epsilon^{abc}\,\sigma^{c},
\\
\mathrm{tr}\,\left(T^{\mathsf{a}}\,T^{\mathsf{b}}\right)=
\frac{1}{2}\delta^{\mathsf{ab}},
\qquad
[T^{\mathsf{a}},T^{\mathsf{b}}]=
\mathrm{i}\,f^{\mathsf{abc}}\,T^{\mathsf{c}},
\end{align}
\end{subequations}
where $\epsilon^{\,}_{abc}$ is the Levi-Civita symbol and
$f^{\,}_{\mathsf{abc}}$ are the structure constants of
$SU(N^{\,}_{\mathrm{c}})$. With these definitions, one can build the
energy-momentum tensor for the free theory defined by
the Lagrangian density
\eqref{eq: single non-abelian wire}
using the Sugawara construction~\cite{Sugawara68,Affleck90,Affleck91a,Affleck91b}
for the energy-momentum tensor
in the M-moving sector,
\begin{subequations}
\label{eq: fermionic conformal embedding energy momentum tensor}
\begin{align}
T^{\,}_{\mathrm{M}}[u(2N^{\,}_{\mathrm{c}})^{\,}_{1}]=
T^{\,}_{\mathrm{M}}[u(1)]
+
T^{\,}_{\mathrm{M}}[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]
+
T^{\,}_{\mathrm{M}}[su(N^{\,}_{\mathrm{c}})^{\,}_{2}].
\end{align}
Here,
\begin{align}
&
T^{\,}_{\mathrm{M}}[u(2N^{\,}_{\mathrm{c}})^{\,}_{1}]\:=
\frac{1}{\pi}
\sum_{\sigma=\uparrow,\downarrow}
\sum^{N^{\,}_{\mathrm{c}}}_{\alpha=1}
\chi^{*}_{\mathrm{M},\sigma,\alpha}\,
\mathrm{i}\partial^{\,}_{\mathrm{M}}\chi^{\,}_{\mathrm{M},\sigma,\alpha},
\\
&
T^{\,}_{\mathrm{M}}[u(1)]\:=
\frac{1}{4N^{\,}_{\mathrm{c}}}\,
j^{\,}_{\mathrm{M}}\,j^{\,}_{\mathrm{M}},
\\
&
T^{\,}_{\mathrm{M}}[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]\:=
\frac{1}{N^{\,}_{\mathrm{c}}+2}
\sum^{3}_{a=1}
J^{a}_{\mathrm{M}}\,J^{a}_{\mathrm{M}},
\\
&
T^{\,}_{\mathrm{M}}[su(N^{\,}_{\mathrm{c}})^{\,}_{2}]\:=
\frac{1}{2+N^{\,}_{\mathrm{c}}}
\sum^{N^{2}_{\mathrm{c}}-1}_{\mathsf{a}=1}
\mathsf{J}^{\mathsf{a}}_{\mathrm{M}}\,\mathsf{J}^{\mathsf{a}}_{\mathrm{M}}\!\!\!\!.
\end{align}
\end{subequations}
With these definitions, it follows that the Hamiltonian density associated with
the free Lagrangian density~\eqref{eq: single non-abelian wire} is given by
\begin{equation}
\mathcal{H}^{\,}_{0,\mathrm{wire}}
\!\!\:=\!\!
2\pi
\!\!\!\!\!
\sum_{\mathrm{M}=\mathrm{L},\mathrm{R}}
\!\!\!\!
\left(
T^{\,}_{\mathrm{M}}[u(1)]
\!+\!
T^{\,}_{\mathrm{M}}[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]
\!+\!
T^{\,}_{\mathrm{M}}[su(N^{\,}_{\mathrm{c}})^{\,}_{2}]
\right)\!.
\end{equation}
Rewriting the free theory~\eqref{eq: single non-abelian wire}
in terms of the currents~\eqref{eq: non-Abelian currents}
amounts to performing a non-Abelian bosonization of the
free theory. This rewriting highlights the fact that a theory of
multiple flavors of free fermions can be broken up into independent
charge [$u(1)$], spin [$su(2)^{\,}_{N^{\,}_{\mathrm{c}}}$], and color
[$su(N^{\,}_{\mathrm{c}})^{\,}_{2}$] sectors.
\subsection{Intra-wire interactions}
\label{subsec: Intra-wire interactions}
Having rewritten the free theory~\eqref{eq: single non-abelian wire}
in terms of the non-Abelian currents~\eqref{eq: non-Abelian currents},
we now wish to isolate the $su(2)^{\,}_{N^{\,}_{\mathrm{c}}}$ spin degrees of freedom
by removing the $u(1)$ charge and $su(N^{\,}_{\mathrm{c}})^{\,}_{2}$ color degrees of
freedom from the low-energy sector of the theory. We accomplish this
by adding interactions that gap out the latter pair of degrees of freedom.
To gap out the charge sector, we add to the free
Lagrangian density~\eqref{eq: single non-abelian wire}
the interaction term
\begin{subequations}\label{eq: umklapp}
\begin{align}\label{eq: umklapp a}
\mathcal{L}^{\,}_{\mathrm{int}}[u(1)]\:=
-
\lambda^{\,}_{u(1)}
\cos\left(\sqrt{2}\,(\phi^{\,}_{\mathrm{R}}+\phi^{\,}_{\mathrm{L}})\right).
\end{align}
The chiral bosonic fields $\phi^{\,}_{\mathrm{M}}$ are defined by the Abelian
bosonization identity
\begin{align}\label{eq: umklapp b}
j^{\,}_{\mathrm{M}}=
-
\frac{1}{\sqrt{2}\,\pi}\,
\partial^{\,}_{\mathrm{M}}\phi^{\,}_{\mathrm{M}}.
\end{align}
\end{subequations}
In the fermionic language, the interaction~\eqref{eq: umklapp a} is interpreted
as an Umklapp process. It is marginally relevant in the renormalization
group (RG) sense, i.e., it flows to strong coupling under RG
and gaps the charge sector when $\lambda^{\,}_{u(1)}>0$.
To gap out the color sector, we add to the free
Lagrangian density~\eqref{eq: single non-abelian wire}
the interaction term
\begin{align}\label{eq: su(N_c) current-current}
\mathcal{L}^{\,}_{\mathrm{int}}[su(N^{\,}_{\mathrm{c}})^{\,}_{2}]\:=
-
\lambda^{\,}_{su(N^{\,}_{\mathrm{c}})^{\,}_{2}}
\sum^{N^{2}_{\mathrm{c}}-1}_{\mathsf{a}=1}
\mathsf{J}^{\mathsf{a}}_{\mathrm{L}}\,
\mathsf{J}^{\mathsf{a}}_{\mathrm{R}},
\end{align}
where the currents $\mathsf{J}^{\mathsf{a}}_{\mathrm{M}}$
are defined in Eqs.~\eqref{eq: non-Abelian currents}. This
current-current interaction is also marginally relevant,
flowing to strong coupling for $\lambda^{\,}_{su(N^{\,}_{\mathrm{c}})^{\,}_{2}}>0$.
At the strong-coupling fixed point dominated by the interactions
\eqref{eq: umklapp} and \eqref{eq: su(N_c) current-current}, the
effective Hamiltonian density for the low-energy sector becomes
\begin{align}
\mathcal{H}^{\,}_{0,\mathrm{eff}}
\:=
2\pi
\left(
T^{\,}_{\mathrm{L}}[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]
+
T^{\,}_{\mathrm{R}}[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]
\right).
\end{align}
This is nothing but the Hamiltonian description of the
$su(2)^{\,}_{N^{\,}_{\mathrm{c}}}$
Wess-Zumino-Witten (WZW) CFT
\cite{Wess71,Witten84}
with the central charge
\begin{align}
c[su(2)^{\,}_{N^{\,}_{\mathrm{c}}}]=
\frac{3\,N^{\,}_{\mathrm{c}}}{2+N^{\,}_{\mathrm{c}}}.
\end{align}
Thus, by adding the interactions \eqref{eq: umklapp} and
\eqref{eq: su(N_c) current-current} to the free theory
\eqref{eq: single non-abelian wire}, we can convert a quantum
wire containing $N^{\,}_{\mathrm{c}}$ colors of spinful fermions
into a highly nontrivial conformal field theory.
The coupled-wire constructions presented in this paper
use arrays of these $su(2)^{\,}_{N^{\,}_{\mathrm{c}}}$ WZW theories
as building blocks for non-Abelian topological phases.
\section{Non-Abelian topological order in two dimensions}
\label{sec: Warm-up: Non-Abelian topological order in two dimensions}
As preparation for Sec.\
\ref{sec: Non-Abelian topological order in three dimensions},
we construct a class of $su(2)^{\,}_{k}$ quantum
liquids in two spatial dimensions and show,
for the case of $k=2$, how to compute their
topological degeneracy on the torus.
This analysis yields insights that are interesting and useful
in their own right.
\subsection{Definition of the class of models}
\label{subsec: Definition of the class of models 2D}
\begin{figure}[tb]
\includegraphics[width=0.49\textwidth]{Figures4}
\caption{
(Color online)
Schematic of the coupled-wire construction for $su(2)^{\,}_{k}$ non-Abelian
topological orders in two spatial dimensions. Grey ovals represent quantum
wires, while red and blue circles represent chiral $su(2)^{\,}_{k}$ currents.
}
\label{fig: 2D system}
\end{figure}
We begin with a one-dimensional array $\Lambda$ of parallel nonchiral spinful fermionic
quantum wires aligned along the $z$-direction, each of which is
described by the Lagrangian density
\eqref{eq: single non-abelian wire} (see Fig.\ \ref{fig: 2D system}).
The cardinality of the one-dimensional lattice $\Lambda$ is
\begin{equation}
|\Lambda|\equiv
L^{\,}_{y}+1.
\end{equation}
We set $N^{\,}_{\mathrm{c}}=k$, where $N^{\,}_{\mathrm{c}}$ is the number of
``colors'' of fermions in each wire. Each wire has an internal
symmetry $U(2k)^{\,}_{\mathrm{L}}\times U(2k)^{\,}_{\mathrm{R}}$, with
respect to which we carry out the bosonization procedure of
Sec.~\ref{sec: Non-Abelian bosonization of a single wire}.
We then gap the $u(1)$ and $su(k)^{\,}_{2}$ sectors with the
intra-wire interactions discussed in Sec.~\ref{subsec: Free-fermion wire},
leaving behind an $su(2)^{\,}_{k}$ current algebra for
each of the left- and right-moving chiral sectors in every wire.
In the Heisenberg picture and in two-dimensional Minkowski space,
we denote the chiral $su(2)^{\,}_{k}$ currents by
$\widehat{J}^{a}_{\mathrm{M},y}(z^{\,}_{\mathrm{M}})$
where $\mathrm{M}=\mathrm{L},\mathrm{R}$ labels the chirality, $a=1,2,3$ labels
the $SU(2)$ generators, $y$ labels the wire,
and $z^{\,}_{\mathrm{M}}$ is defined in Eq.~\eqref{eq: def z of M}.
We couple nearest-neighbor wires with the
$su(2)^{\,}_{k}$
interaction (see Fig.~\ref{fig: 2D system})
\begin{subequations}
\begin{align}
\widehat{\mathcal{L}}^{\,}_{\mathrm{bs}}\equiv
-
\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}
\:=
-
\frac{\lambda}{2}
\sum_{y=0}^{L^{\,}_{y}-\sigma^{\,}_{\mathrm{BC}}}
\widehat{J}^{+}_{\mathrm{L},y+1}\,
\widehat{J}^{-}_{\mathrm{R},y}
+
\mathrm{H.c.},
\label{eq2Dcase: 2D interaction}
\end{align}
where $\sigma^{\,}_{\mathrm{BC}}=0,1$
for periodic and open boundary conditions,
respectively.
In Eq.~\eqref{eq2Dcase: 2D interaction}, we have introduced
the linear combinations
\begin{align}
\widehat{J}^{\pm}_{\mathrm{M},y}\:=
\widehat{J}^{1}_{\mathrm{M},y}\pm\mathrm{i}\,\widehat{J}^{2}_{\mathrm{M},y}.
\label{eq: def chiral parafermion and boson rep su(2)k b}
\end{align}
\end{subequations}
When
periodic boundary conditions are imposed in the $y$-direction,
i.e., when $\sigma^{\,}_{\mathrm{BC}}=0$,
each chiral current is paired with exactly one current of the opposite
chirality in a neighboring wire, and, hence,
the full array of quantum wires
may become gapped in the strong-coupling limit $|\lambda|\gg0$.
When open boundary conditions are imposed in the
$y$-direction,
i.e., when $\sigma^{\,}_{\mathrm{BC}}=1$,
there is a left-moving $su(2)^{\,}_{k}$ current at
$y=0$ and a right-moving $su(2)^{\,}_{k}$ current at $y=L^{\,}_{y}$ that are
fully decoupled from the bulk. This edge structure is reminiscent of
that of the $su(2)^{\,}_{k}$ non-Abelian Chern-Simons
theories~\cite{Witten89,Cabra00} and that of the
$\mathbb{Z}^{\,}_{k}$ Read-Rezayi quantum Hall states~\cite{Read99}.
\subsection{Parafermion representation of the interwire interactions}
\label{subsec: Parafermion representation of the interwire interactions 2D}
The interaction \eqref{eq2Dcase: 2D interaction}
can be better understood by rewriting
the $su(2)^{\,}_{k}$ currents in terms of auxiliary degrees of freedom.
This rewriting must preserve the $su(2)^{\,}_{k}$ current algebra,
which is encoded in the operator product expansion (OPE)~\cite{DiFrancesco97}
\begin{align}\label{eq: current-current OPEs}
\widehat{J}^{a}_{\mathrm{L},y}(v)\,\widehat{J}^{\tilde{a}}_{\mathrm{L},\tilde{y}}(w)\sim
\delta^{\,}_{y,\tilde{y}}\,
\(
\frac{
(k/2)\,\delta^{a\tilde{a}}
}
{
v^{2}-w^{2}}
+
\frac{\mathrm{i}\,
\epsilon^{a\tilde{a}b}\,
\widehat{J}^{b}_{\mathrm{L},y}(w)
}
{
v-w
}
\),
\end{align}
for the holomorphic sector $\mathrm{M}=\mathrm{L}$,
and similarly for the antiholomorphic sector $\mathrm{M}=\mathrm{R}$.
(Here, we employ complex coordinates
$v\equiv t+\mathrm{i}\,z$, obtained from the chiral coordinate
$z^{\,}_{\mathrm{L}}$
defined in Eq.~(\ref{eq: def z of M})
by the analytic continuation $z\to\mathrm{i}\,z$,
and $\bar{v}\equiv t-\mathrm{i}\,z$, obtained from the chiral coordinate
$z^{\,}_{\mathrm{R}}$ also defined in Eq.~(\ref{eq: def z of M})
by the same analytic continuation.)
The group indices $a,\tilde{a}=1,2,3$, and summation over the repeated index $b=1,2,3$
is implied.
The symbol $``\sim''$ denotes equality up to nonsingular terms
in the limit $v\to w$.
As shown by Zamolodchikov and Fateev~\cite{Zamolodchikov85}
(see Appendix \ref{appsec: The parafermion current algebra}),
the current algebra \eqref{eq: current-current OPEs}
can be represented in terms of $\mathbb{Z}^{\,}_{k}$
parafermion and chiral boson operators as follows~\cite{DiFrancesco97}:
\begin{subequations}
\label{eq: def chiral parafermion and boson rep su(2)k}
\begin{align}
&
\widehat{J}^{+}_{\mathrm{M},y}\=:
\sqrt{k}\
\widehat{\Psi}^{\,}_{\mathrm{M},y}\,
\bm{:}
e^{+\mathrm{i}\sqrt{1/k}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:},
\label{eq: def chiral parafermion and boson rep su(2)k 0a}
\\
&
\widehat{J}^{-}_{\mathrm{M},y}\=:
\sqrt{k}\
\bm{:}
e^{-\mathrm{i}\sqrt{1/k}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:}\,
\widehat{\Psi}^{\dag}_{\mathrm{M},y},
\label{eq: def chiral parafermion and boson rep su(2)k 0b}
\\
&
\widehat{J}^{3}_{\mathrm{M},y}\=:
\mathrm{i}\,
\frac{\sqrt{k}}{2}\,
\partial^{\,}_{\mathrm{M}}\widehat{\phi}^{\,}_{\mathrm{M},y},
\label{eq: def chiral parafermion and boson rep su(2)k 0c}
\end{align}
where $\bm{:}\cdot\bm{:}$ denotes normal ordering with respect to the
many-body ground state of $\widehat{H}^{\,}_{0,\rm eff}$ within
each wire.
Here, the $\mathbb{Z}^{\,}_{k}$ parafermions
$\widehat{\Psi}^{\,}_{\mathrm{M},y}$ satisfy the equal-time algebra
\begin{widetext}
\begin{equation}
\begin{split}
&
\widehat{\Psi}^{\,}_{\mathrm{M},y}(t,z)\,
\widehat{\Psi}^{\,}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})=
\widehat{\Psi}^{\,}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})\,
\widehat{\Psi}^{\,}_{\mathrm{M},y}(t,z)\,
e^{
-\mathrm{i}\,\frac{2\pi}{k}\,\delta^{\,}_{y,y^{\prime}}
\[
(-1)^{\mathrm{M}}\,\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,\mathrm{sgn}(z-z^{\prime})
+
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}
\]
+
\mathrm{i}\,\frac{2\pi}{k}\,\mathrm{sgn}(y-y^{\prime})
},
\end{split}
\label{eq: def chiral parafermion and boson rep su(2)k c}
\end{equation}
\begin{equation}
\begin{split}
&
\widehat{\Psi}^{\dag}_{\mathrm{M},y}(t,z)\,
\widehat{\Psi}^{\dag}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})=
\widehat{\Psi}^{\dag}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})\,
\widehat{\Psi}^{\dag}_{\mathrm{M},y}(t,z)\,
e^{
-\mathrm{i}\,\frac{2\pi}{k}\,\delta^{\,}_{y,y^{\prime}}
\[
(-1)^{\mathrm{M}}\,\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,\mathrm{sgn}(z-z^{\prime})
+
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}
\]
+
\mathrm{i}\,\frac{2\pi}{k}\,\mathrm{sgn}(y-y^{\prime})},
\end{split}
\label{eq: def chiral parafermion and boson rep su(2)k cc}
\end{equation}
\begin{equation}
\begin{split}
&
\widehat{\Psi}^{\,}_{\mathrm{M},y}(t,z)\,
\widehat{\Psi}^{\dag}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})=
\widehat{\Psi}^{\dag}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})\,
\widehat{\Psi}^{\,}_{\mathrm{M},y}(t,z)\,
e^{
+\mathrm{i}\,\frac{2\pi}{k}\,\delta^{\,}_{y,y^{\prime}}
\[
(-1)^{\mathrm{M}}\,\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,\mathrm{sgn}(z-z^{\prime})
+
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}
\]
-
\mathrm{i}\,\frac{2\pi}{k}\,\mathrm{sgn}(y-y^{\prime})}.
\end{split}
\label{eq: def chiral parafermion and boson rep su(2)k ccc}
\end{equation}
The sign function above is defined such that $\mathrm{sgn}(0)=0$.
The left- and right-moving labels $\mathrm{M}=\mathrm{L},\mathrm{R}$ are defined with
the convention that
$\epsilon^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}$ is the antisymmetric
Levi-Civita symbol obeying
$\epsilon^{\,}_{\mathrm{L},\mathrm{R}}=-\epsilon^{\,}_{\mathrm{R},\mathrm{L}}=-1$.
Moreover, $(-1)^{\mathrm{R}}=-(-1)^{\mathrm{L}}\equiv1$.
The algebra of the $su(2)^{\,}_{k}$ currents holds so long as the
equal-time algebra
\begin{align}
\label{eq: def chiral parafermion and boson rep su(2)k d}
\begin{split}
&
\[
\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z),
\widehat{\phi}^{\,}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})
\]=
-\mathrm{i}\,
2\pi
\[
(-1)^{\mathrm{M}}\,
\delta^{\,}_{y,y^{\prime}}\,
\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,
\mathrm{sgn}(z-z^{\prime})
+
\delta^{\,}_{y,y^{\prime}}\,
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}
-
\mathrm{sgn}(y-y^{\prime})
\],
\end{split}
\end{align}
\end{widetext}
\end{subequations}
is imposed in the chiral bosonic sector. In particular, one verifies
that currents defined in different wires commute with one another at
equal times when the definitions
\eqref{eq: def chiral parafermion and boson rep su(2)k} are imposed.
Furthermore, one can show that all
equal-time commutators between $su(2)^{\,}_{k}$ currents differing by
their $L$ and $R$ labels also vanish. Finally, the chiral
parafermions commute with the chiral bosons at equal times.
The representation \eqref{eq: def chiral parafermion and boson rep su(2)k}
of the $su(2)^{\,}_{k}$ current algebra provides a convenient interpretation of
the interactions \eqref{eq2Dcase: 2D interaction}
in terms of fractionalized degrees of freedom, as we discuss below.
However, there are several caveats to keep in mind. Chief among these is
the fact that the factorization
\eqref{eq: def chiral parafermion and boson rep su(2)k 0a}--%
\eqref{eq: def chiral parafermion and boson rep su(2)k 0c}
of the $su(2)^{\,}_{k}$ currents re-expresses a set of local operators
(the currents) in terms of products of auxiliary degrees of freedom
(the parafermions and the chiral bosons).
While the $su(2)^{\,}_{k}$ currents admit a local expression
[Eq.~\eqref{eq: non-Abelian currents b}] in terms of
the original degrees of freedom used to define the theory (the electrons)
these auxiliary degrees of freedom do not.
This fact will be important
when we construct the nonlocal string operators that allow us to
calculate the topological degeneracy in Sec.~\ref{subsec: Case study: su(2)_2}.
Furthermore, we note that this parafermion representation
is not unique in two ways.
First, as it factorizes a local (observable) operator
into the product of two operators,
there is an ambiguity with the choice of the phase
assigned to each operator-valued factor.
(This is an explicit manifestation of the nonlocality of
the auxiliary degrees of freedom.)
The choice for this phase cannot have observable consequences.
Second, the dependence on the labels $y\neq y'$
of the equal-time algebra is not unique since
many distinct choices accomodate the fact that any two currents
belonging to two distinct wires $y$ and $y'$ must always commute.
Hence, the dependence
on the labels $y\neq y'$ of the parafermion equal-time algebra
cannot have observable consequences.
We demonstrate that this is true for the case of $su(2)^{\,}_{2}$ in Appendix\
\ref{sec: Independence of string operator algebra on arbitrary phase factors}.
We work with the normalization convention for which
the operator $\exp(\mathrm{i}\,a\,\widehat{\phi}^{\,}_{\mathrm{M}})$,
for $a$ any real-valued number,
has the anomalous scaling dimension $a^{2}$. With this convention,
the chiral vertex operator
$\exp(\mathrm{i}\,\sqrt{1/k}\,\widehat{\phi}^{\,}_{\mathrm{M}})$,
which annihilates a chiral Abelian quasiparticle,
has anomalous scaling dimension $1/k$. In turn,
the chiral parafermion operator
$\widehat{\Psi}^{\,}_{\mathrm{M}}$ must have the
anomalous scaling dimension $1-(1/k)$,
as the current operators have scaling dimension 1.
The expressions
\eqref{eq: def chiral parafermion and boson rep su(2)k}
for the currents are equivalent to the identity~\cite{DiFrancesco97}
\begin{subequations}
\begin{align}
\label{eq: su2 level k decomposition}
su(2)^{\,}_{k}\simeq u(1)^{\,}_{k}\oplus\mathbb{Z}^{\,}_{k},
\end{align}
where
\begin{align}
\mathbb{Z}^{\,}_{k}\equiv \frac{su(2)^{\,}_{k}}{u(1)^{\,}_{k}},
\end{align}
\end{subequations}
which states that an $SU(2)$ WZW theory at level $k$ can be
interpreted as a direct product of a chiral boson and a
$\mathbb{Z}^{\,}_{k}$ parafermion conformal field theory.
With these definitions, the interactions~\eqref{eq2Dcase: 2D interaction}
take the form
\begin{widetext}
\begin{align}
\widehat{\mathcal L}^{\,}_{\mathrm{bs}}\equiv
-
\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}&=\,
-\lambda\,
\frac{k}{2}\,
\sum^{L^{\,}_{y}}_{y=0}\,
\left(
\widehat{\Psi}^{\,}_{\mathrm{L},y}\,
\bm{:}
e^{+\mathrm{i}\sqrt{\frac{1}{k}}\,\widehat{\phi}^{\,}_{\mathrm{L},y}}
\bm{:}
\,
\bm{:}
e^{-\mathrm{i}\sqrt{\frac{1}{k}}\,\widehat{\phi}^{\,}_{\mathrm{R},y+1}}
\bm{:}\,
\widehat{\Psi}^{\dag}_{\mathrm{R},y+1}\,
+
\mathrm{H.c.}
\right),
\label{eq2Dcase: parafermion representation su(2)_k}
\end{align}
\end{widetext}
(We employ periodic boundary conditions for the remainder of this section.)
Written this way, the current-current interactions~\eqref{eq2Dcase: 2D interaction}
can be reinterpreted as correlated hoppings of (nonlocal) fractionalized degrees of freedom
between wires. Indeed, viewing $\widehat{\Psi}^{\dag}_{\mathrm{M},y}$ as the
creation operator for a parafermion with chirality $\mathrm{M}$ in wire $y$,
and viewing the vertex operator $\bm{:}
e^{-\mathrm{i}\sqrt{\frac{1}{k}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:}$ as the creation operator for an Abelian quasiparticle, we can interpret
Eq.~\eqref{eq2Dcase: parafermion representation su(2)_k}
as allowing parafermions to hop between wires so long
as an Abelian quasiparticle hops at the same time.
Since the composite of these two fractionalized excitations is a boson,
per
Eqs.~\eqref{eq: def chiral parafermion and boson rep su(2)k},
this correlated hopping process forbids isolated fractionalized degrees
of freedom from hopping between wires.
When periodic boundary conditions are imposed, the interaction
\eqref{eq2Dcase: parafermion representation su(2)_k} gaps out the array of wires if the current-current
coupling on each bond in the lattice $\Lambda$ acquires a finite vacuum expectation value.
Such a scenario is possible in the limit $\lambda\to\infty$.
We will see an explicit example of this gapping mechanism in the next
section, where we study the case $k=2$ in detail.
\subsection{Case study: $su(2)^{\,}_{2}$}
\label{subsec: Case study: su(2)_2}
In this section, we work through the example of $k=2$ in detail.
First, we will show how the interaction \eqref{eq2Dcase: parafermion representation su(2)_k}
leads to a gapped state of matter.
Next, we will characterize the topological order in this gapped state of matter
by imposing periodic boundary conditions in the $y$- and
$z$-directions and constructing nonlocal \textit{string operators} that
commute with the interaction
$\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
defined by Eq.~(\ref{eq2Dcase: 2D interaction}).
These string operators will label the topologically degenerate ground states in the limit
$\lambda\to\infty$.
The Lagrangian density in this case is (omitting the normal ordering of the
vertex operators)
\begin{widetext}
\begin{subequations}
\label{eq2Dcase: su(2)_{2} backscattering}
\begin{align}
\widehat{\mathcal{L}}^{\,}_{\mathrm{bs}}\equiv
-
\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}
&\:=
-\lambda
\sum_{y=0}^{L^{\,}_{y}}
\(
e^{+\mathrm{i}\sqrt{1/2}\,\(\widehat{\phi}^{\,}_{\mathrm{L},y}-\widehat{\phi}^{\,}_{\mathrm{R},y+1}\)}\,
\widehat{\psi}^{\,}_{\mathrm{L},y}\,
\widehat{\psi}^{\,}_{\mathrm{R},y+1}
+
\mathrm{H.c.}
\)
\label{eq2Dcase: su(2)_{2} backscattering a}
\\
&=
-2\lambda
\sum_{y=0}^{L^{\,}_{y}}
\(\mathrm{i}\, \widehat{\psi}^{\,}_{\mathrm{L},y}\,
\widehat{\psi}^{\,}_{\mathrm{R},y+1}\)
\sin\[\sqrt{\frac{1}{2}}\(\widehat{\phi}^{\,}_{\mathrm{L},y}-\widehat{\phi}^{\,}_{\mathrm{R},y+1}\)\],
\label{eq2Dcase: su(2)_{2} backscattering b}
\end{align}
\end{subequations}
\end{widetext}
which should be compared with
Eq.~\eqref{eq2Dcase: parafermion representation su(2)_k}.
The chiral operators
\begin{subequations}
\begin{equation}
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z)\equiv
\widehat{\Psi}^{\,}_{\mathrm{M},y}(t,z)\equiv
\widehat{\Psi}^{\dag}_{\mathrm{M},y}(t,z)
\end{equation}
with $\mathrm{M}=\mathrm{L},\mathrm{R}$
are Majorana operators
(i.e., $\mathbb{Z}^{\,}_{2}$ parafermions).
Their equal-time exchange algebra is given by
Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k c}
with $k=2$. We also impose the normalization
\begin{equation}\label{eq2Dcase: normalization}
\lim_{z'\to z}
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z)\,
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z')\equiv
\lim_{z'\to z}\delta(z-z')\:=\mathcal{N}^{\,}_{\delta},
\end{equation}
\end{subequations}
where $\mathcal{N}^{\,}_{\delta}$ is a constant with
dimension [1/length].
The chiral bosons $\widehat{\phi}^{\,}_{\mathrm{M},y}$
obey the equal-time algebra
\eqref{eq: def chiral parafermion and boson rep su(2)k d},
as before.
Furthermore, the
chiral Majorana operators and the chiral bosons commute at equal
times:
\begin{align}
\[
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z),
\widehat{\phi}^{\,}_{\mathrm{M}^{\prime},y^{\prime}}(t,z^{\prime})
\]=
0.
\end{align}
The rewriting of the interaction \eqref{eq2Dcase: su(2)_{2} backscattering a}
presented in Eq.~\eqref{eq2Dcase: su(2)_{2} backscattering b} provides
an intuitive illustration of the discussion in
Sec.~\ref{subsec: Parafermion representation of the interwire interactions 2D}
of how the interaction \eqref{eq2Dcase: parafermion representation su(2)_k}
leads to a gap when periodic boundary conditions are imposed. In this case,
when the bosonic field
$\widehat{\phi}^{\,}_{\mathrm{L},y}-\widehat{\phi}^{\,}_{\mathrm{R},y+1}$
becomes locked to an extremum of the sine potential, a Majorana mass term is
induced for the fermionic degrees of freedom. The simultaneous gapping of the
Majorana modes
and locking of the bosonic fields is consistent due to the
independence of the $u(1)^{\,}_{2}$ and $\mathbb{Z}^{\,}_{2}$ sectors of the
$su(2)^{\,}_{2}$ theory.
\subsubsection{Quasilocal chirality-resolved $\mathbb{Z}^{\,}_{2}$ gauge
symmetry}
Observe that the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}
is invariant under the M- and $y$-resolved
$\mathbb{Z}^{\,}_{2}$ gauge transformation
\begin{subequations}
\label{eq2Dcase: quasilocal Z_{2} symmetry}
\begin{align}
&
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z)\mapsto
e^{\mathrm{i}\alpha^{\,}_{\mathrm{M},y}}\,
\widehat{\psi}^{\,}_{\mathrm{M},y}(t,z),
\label{eq2Dcase: quasilocal Z_{2} symmetry a}
\\
&
\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)\mapsto
\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)
+
\sqrt{2}\,\alpha^{\,}_{\mathrm{M},y},
\label{eq2Dcase: quasilocal Z_{2} symmetry b}
\end{align}
where the assignments
\begin{equation}
\alpha^{\,}_{\mathrm{M},y}\in\{0,\pi\}
\label{eq2Dcase: quasilocal Z_{2} symmetry c}
\end{equation}
for all chiralities $\mathrm{M}=\mathrm{L},\mathrm{R}$
and all wires $y$ define the map
\begin{equation}
\alpha:
\{\mathrm{M}=\mathrm{L},\mathrm{R}\}
\times
\{y=0,\cdots,L^{\,}_{y}\}\to
\{0,\pi\}.
\label{eq2Dcase: quasilocal Z_{2} symmetry d}
\end{equation}
\end{subequations}
This transformation is implemented by the operator
\begin{equation}
\widehat{\Gamma}^{\,}_{\alpha}(t)\equiv
\prod_{\mathrm{M}=\mathrm{L},\mathrm{R}}
\prod_{y=0}^{L^{\,}_{y}}
\widehat{\Gamma}^{\,}_{\alpha^{\,}_{\mathrm{M},y}}(t)\:=
\widehat{\mathcal{U}}^{\,}_{\alpha}(t)\,
\widehat{\mathcal{Z}}^{\,}_{\alpha}(t),
\label{eq: def generator M and y resolved Z2 gauge trsf}
\end{equation}
where the operator
\begin{widetext}
\begin{align}
\widehat{\mathcal{U}}^{\,}_{\alpha}(t)\equiv
\prod_{\mathrm{M}=\mathrm{L},\mathrm{R}}\,
\prod_{y=0}^{L^{\,}_{y}}\,
\widehat{\mathcal{U}}^{\,}_{\alpha^{\,}_{\mathrm{M},y}}(t)
\:=
\prod_{\mathrm{M}=\mathrm{L},\mathrm{R}}\,
\prod_{y=0}^{L^{\,}_{y}}\,
\exp
\left(
(-1)^{\mathrm{M}}\,
\frac{
\mathrm{i}\alpha^{\,}_{\mathrm{M},y}
}
{
2\pi\sqrt{2}
}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)
\right)
\label{eq2Dcase: U_alpha definition}
\end{align}
\end{widetext}
acts only on the chiral boson sector of the theory and
implements the transformation \eqref{eq2Dcase: quasilocal Z_{2} symmetry b},
and where the operator
\begin{align}
\label{eq2Dcase: Z_alpha def}
\widehat{\mathcal{Z}}^{\,}_{\alpha}(t)
&=
\prod_{\mathrm{M}=\mathrm{L},\mathrm{R}}
\prod_{y=0}^{L^{\,}_{y}}
\widehat{\mathcal{Z}}^{\,}_{\alpha^{\,}_{\mathrm{M},y}}(t)
\end{align}
acts only
on the Ising (i.e., $\mathbbm Z^{\,}_{2}$) sector and implements
the transformation \eqref{eq2Dcase: quasilocal Z_{2} symmetry a}. The action
of the operator $\widehat{\mathcal{U}}^{\,}_{\alpha}(t)$ on
the chiral bosons follows from the fact that
\begin{equation}
\begin{split}
\widehat{\mathcal{U}}^{\,}_{\alpha^{\,}_{\mathrm{M},y}}(t)\,
\widehat{\phi}^{\,}_{\mathrm{M}',y'}(t,z)\,
\widehat{\mathcal{U}}^{\dag}_{\alpha^{\,}_{\mathrm{M},y}}(t)=&\,
\widehat{\phi}^{\,}_{\mathrm{M}',y'}(t,z)
\\
&\,
+
\sqrt{2}\,
\alpha^{\,}_{\mathrm{M},y}\,
\delta^{\,}_{y,\pri y}\,
\delta^{\,}_{\mathrm{M},\pri M}
\end{split}
\label{eq2Dcase: action widehat mathcal U on phi}
\end{equation}
holds for any pair of chiralities
$\mathrm{M},\mathrm{M}'=\mathrm{L},\mathrm{R}$,
for any pair of wires $y,y'$, and for any
$t$ and $z$ [see Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}].
The action of the operator $\widehat{\mathcal{Z}}^{\,}_{\alpha}(t)$ follows from
the definition of $\widehat{\mathcal{Z}}^{\,}_{\alpha^{\,}_{\mathrm{M},y}}(t)$
in terms of the fermion parity operator in the wire $y$, which is somewhat
involved and will not be presented here.
\subsubsection{$su(2)^{\,}_{2}$ primary fields}
\label{subsec: twist operators}
To construct the excitations of the coupled-wire theory, we will use
the primary operators of the underlying $su(2)^{\,}_{2}$ theory
defined on each quantum wire in Fig.~\ref{fig: 2D system}. In
accordance with the identity \eqref{eq: su2 level k decomposition},
the primary operators of the $su(2)^{\,}_{k}$ theory can be expressed
in terms of the primary operators of the $u(1)^{\,}_{k}$ and
$\mathbb{Z}^{\,}_{k}$ theories for any $k$~\cite{Zamolodchikov85}. For
$k=2$, there are three primary operators labeled by the ``angular
momenta'' $0$, $\frac{1}{2}$, and $1$ with scaling dimensions $0$,
$\frac{3}{16}$, and $\frac{1}{2}$, respectively. The ``spin-$0$''
primary, with scaling dimension $0$, is simply the identity operator
in the $\mathrm{M}$-moving channel of wire $y$. The
``spin-$\frac{1}{2}$'' primary, with scaling dimension $\frac{3}{16}$,
is defined to be the product~\cite{Zamolodchikov85}
\begin{subequations}
\begin{align}
\label{eq: spin-1/2 primary}
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{M},y}(t,z)
\:=
\widehat{\sigma}^{\,}_{\mathrm{M},y}(t,z)\,
\bm{:}
e^{+\mathrm{i}\frac{1}{2\sqrt{2}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)}
\bm{:}\, ,
\end{align}
where the operator $\widehat{\sigma}^{\,}_{\mathrm{M},y}$ is the chiral
``twist field'' from the $\mathbb{Z}^{\,}_{2}$ sector
(c.f.\ Appendix~\ref{appsec: Z_{k} CFT}), which has scaling dimension
$1/16$. Adding this scaling dimension to that of the vertex operator
$\bm{:}
e^{+\mathrm{i}\frac{1}{2\sqrt{2}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:}$
gives the appropriate scaling dimension $\frac{3}{16}$
(c.f.\ Appendix~\ref{appsubsec: Gaussian algebra}). Finally,
the ``spin-$1$'' primary, with scaling dimension $\frac{1}{2}$, can be
written as \cite{Zamolodchikov85}
\begin{align}
\label{eq: spin-1 primary}
\widehat{\Phi}^{(1)}_{\mathrm{M},y}(t,z)
\:=
\bm{:}
e^{+\mathrm{i}\frac{1}{\sqrt{2}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)}
\bm{:}\, .
\end{align}
Note that the scaling dimension of the vertex operator
$\bm{:}
e^{+\mathrm{i}\frac{1}{\sqrt{2}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:}$
is already $\frac{1}{2}$, so the only operator from the $\mathbb{Z}^{\,}_{2}$ sector
that can appear in this expression is the identity.
\end{subequations}
It is important to note that the above expressions for the primary operators
carry chirality labels $\mathrm{M}=\mathrm{L},\mathrm{R}$ and wire labels
$y=0,\cdots,L^{\,}_{y}$. In other words, these operators are defined
within either the holomorphic ($\mathrm{L}$)
or antiholomorphic ($\mathrm{R}$) sector of a single wire $y$.
Consequently, the primary operators \eqref{eq: spin-1/2 primary}
and \eqref{eq: spin-1 primary} have nonvanishing conformal spin, equal to their
scaling dimensions, and are hence nonlocal
(i.e., they cannot be regularized on a 1D lattice by an operator
with finite support). However, the operators
$
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{L},y}\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{R},y}
$
and
$\widehat{\Phi}^{(1)}_{\mathrm{L},y}\,\widehat{\Phi}^{(1)}_{\mathrm{R},y}$,
being products of holomorphic and antiholomorphic operators with the same
scaling dimensions, have vanishing conformal spin and are local. We will
use these local building blocks to construct the nonlocal string operators
that encode the ground-state degeneracy of the coupled-wire theory.
In order to compute commutators of these string operators with the
Hamiltonian \eqref{eq2Dcase: su(2)_{2} backscattering}
and with each other, we need to establish the algebra of the primary operators
\eqref{eq: spin-1/2 primary} and \eqref{eq: spin-1 primary}. We can obtain
this by considering the $u(1)^{\,}_{2}$ and $\mathbb{Z}^{\,}_{2}$ sectors
separately. The algebra of the $u(1)^{\,}_{2}$ vertex operators is obtained
directly from Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}.
The algebra of operators in the $\mathbbm Z^{\,}_{2}$ sector is determined
as follows.
For any pair of wires $y$ and $y'$, we posit the
OPEs (using the complex coordinates
$v\equiv t+\mathrm{i}\, z$
and
$v^{\prime}\equiv t^{\prime}+\mathrm{i}\, z^{\prime}$)
\begin{subequations}
\label{eq2Dcase: Z_{2} psi sigma OPE}
\begin{align}
&
\widehat{\psi}^{\,}_{\mathrm{L},y}(v)\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(v^{\prime})=
\delta^{\,}_{y,y^{\prime}}\,
\frac{C^{\sigma}_{\psi\sigma}}{(v-v^{\prime})^{1/2}}\,\widehat{\sigma}^{\,}_{\mathrm{L},y}(v)
+
\cdots,
\label{eq2Dcase: Z_{2} psi sigma OPE a}
\\
&
\widehat{\psi}^{\,}_{\mathrm{R},y}(\bar{v})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\bar{v}^{\prime})=
\delta^{\,}_{y,y^{\prime}}\,
\frac{C^{\sigma}_{\psi\sigma}}{(\bar{v}-\bar{v}^{\prime})^{1/2}}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y}(\bar{v})
+
\cdots,
\label{eq2Dcase: Z_{2} psi sigma OPE b}
\\
&
\widehat{\psi}^{\,}_{\mathrm{L},y}(v)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\bar{v}^{\prime})=
\widehat{\psi}^{\,}_{\mathrm{R},y}(\bar{v})\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(v^{\prime})=
0
+
\cdots,
\label{eq2Dcase: Z_{2} psi sigma OPE c}
\end{align}
where the structure constants obey the symmetry condition
\begin{align}
C^{\sigma}_{\psi\sigma}=
C^{\sigma}_{\sigma\psi},
\label{eq2Dcase: Z_{2} psi sigma OPE d}
\end{align}
and $\cdots$ stands for nonsingular terms.
\end{subequations}
Determining the equal-time algebra of the twist fields and the
Majorana fields requires one to restrict the above OPE to the real
line in the complex plane. Because of the symmetry condition
(\ref{eq2Dcase: Z_{2} psi sigma OPE d})
on the structure constants, exchanging the order of the fields
$\widehat{\psi}^{\,}_{\mathrm{L},y}(v)$
and
$\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(v^{\prime})$
on the left-hand side of Eqs.\
(\ref{eq2Dcase: Z_{2} psi sigma OPE a}), say,
is equivalent to exchanging $v$ and $v^{\prime}$.
However, when we restrict the OPE to equal times,
information about the handedness of this exchange is lost,
see Fig.\ \ref{fig: monodromy}.
We therefore adopt the following convention for
their equal-time operator algebra in two-dimensional Minkowski space.
We make the choice
\begin{subequations}
\label{eq2Dcase: psi sigma algebra}
\begin{align}
&
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,z^{\prime})=
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,z^{\prime})\,
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)
\nonumber\\
&
\hphantom{
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,z^{\prime})=
}
\times
e^{-\mathrm{i}\,\frac{\pi}{2}\,\delta^{\,}_{y,y^{\prime}}\,\mathrm{sgn}(z-z^{\prime})},
\label{eq2Dcase: psi sigma algebra a}
\\
&
\widehat{\psi}^{\,}_{\mathrm{R},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z^{\prime})=
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z^{\prime})\,
\widehat{\psi}^{\,}_{\mathrm{R},y}(t,z)
\nonumber\\
&
\hphantom{
\widehat{\psi}^{\,}_{\mathrm{R},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z^{\prime})=
}
\times
e^{+\mathrm{i}\, \frac{\pi}{2}\,\delta^{\,}_{y,y^{\prime}}\, \mathrm{sgn}(z-z^{\prime})},
\label{eq2Dcase: psi sigma algebra b}
\\
&
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z^{\prime})=
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z^{\prime})\,
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z),
\end{align}
\end{subequations}
for any pair of wires $y$ and $y'$ and for any
$z\neq z'$. The appearance of the phase $\pi/2$ is fixed by the OPE
\eqref{eq2Dcase: Z_{2} psi sigma OPE a}
and
\eqref{eq2Dcase: Z_{2} psi sigma OPE b}
and the sign $\mathrm{sgn}(z-z^{\prime})$
is used to keep track of the handedness of
the exchange. This choice of sign convention for the phase $\pi/2$ is
equivalent to a choice of analytic continuation into the complex plane
in order to regularize the equal-time exchange of the two operators.
\begin{figure}[t]
\includegraphics[width=0.4\textwidth]{Figures5}
\caption{
Counterclockwise monodromy of two operators $\widehat{\mathcal O}^{\,}_{1}(t^{\,}_{0},z^{\,}_{1})$ and
$\widehat{\mathcal O}^{\,}_{2}(t^{\,}_{0},z^{\,}_{2})$ in the complex plane. When the operators
$\widehat{\mathcal O}^{\,}_{1}$ and $\widehat{\mathcal O}^{\,}_{2}$ are evaluated at equal times,
their exchange can be viewed as monodromy in the complex plane, provided
that the handedness of the monodromy is specified. We adopt the convention that the (holomorphic)
operator with the larger value of $z$ is passed counterclockwise around the
operator with the smaller value of $z$, resulting in the factors of $\ensuremath \text{sgn}(z-z^{\prime})$
that appear in the exchange algebras for the primary operators in this section.
}
\label{fig: monodromy}
\end{figure}
The equal-time algebra of two twist operators is more subtle.
For any pair of wires $y$ and $y'$,
the OPE of two twist fields in the complex plane is given by
[c.f.\ Eq.~\eqref{appeq: general Z_{k} OPE}]
\begin{subequations}
\label{eq2Dcase: sigma sigma OPE}
\begin{align}
\widehat{\sigma}^{\,}_{\mathrm{L},y}(v)\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(v^{\prime})=&\,
\delta^{\,}_{y,y^{\prime}}\,
\frac{C^{\mathbbm{1}}_{\sigma\sigma}}{(v-v^{\prime})^{1/8}}
\nonumber\\
&
+
\delta^{\,}_{y,y^{\prime}}\,
C^{\psi}_{\sigma\sigma}\,(v-v^{\prime})^{3/8}\,\psi^{\,}_{\mathrm{L},y}(v),
\label{eq2Dcase: sigma sigma OPE a}
\\
\widehat{\sigma}^{\,}_{\mathrm{R},y}(\bar{v})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\bar{v}^{\prime})=&\,
\delta^{\,}_{y,y^{\prime}}\,
\frac{C^{\mathbbm{1}}_{\sigma\sigma}}{(\bar{v}-\bar{v}^{\prime})^{1/8}}
\nonumber\\
&\,
+
\delta^{\,}_{y,y^{\prime}}\,
C^{\psi}_{\sigma\sigma}\,
(\bar{v}-\bar{v}^{\prime})^{3/8}\,
\psi^{\,}_{\mathrm{R},y}(\bar{v}),
\label{eq2Dcase: sigma sigma OPE b}
\\
\widehat{\sigma}^{\,}_{\mathrm{L},y}(v)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\bar{v}^{\prime})=&\,
\widehat{\sigma}^{\,}_{\mathrm{R},y}(\bar{v})\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(v^{\prime})=
0
+
\cdots.
\label{eq2Dcase: sigma sigma OPE c}
\end{align}
\end{subequations}
Since there are two singular terms appearing on the right-hand side of
Eqs.\ (\ref{eq2Dcase: sigma sigma OPE a}) and
(\ref{eq2Dcase: sigma sigma OPE b}),
the product of two chiral twist fields must be defined with care.
In particular, correlation functions involving multiple
chiral twist fields are not well-defined unless the fusion channel
$\mathbbm{1}$ or $\psi$ is specified~\cite{Fendley07}.
We choose an equal-time operator algebra that reflects this ambiguity
in the definition of chiral correlation functions involving the twist field.
Thus, we define the equal-time algebra
\begin{subequations}
\label{eq2Dcase: sigma sigma algebra}
\begin{align}
&
\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,z')=
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,z')\,\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z)
\nonumber\\
&\qquad
\times
\begin{cases}
e^{-\mathrm{i}\,\frac{\pi}{8}\,\delta^{\,}_{y,y^{\prime}}\,\mathrm{sgn}(z-z^{\prime})},
&
\mathrm{if}\ \sigma\times\sigma=\mathbbm{1},
\\
e^{+\mathrm{i}\,\frac{3\pi}{8}\,\delta^{\,}_{y,y^{\prime}}\,\mathrm{sgn}(z-z^{\prime})},
&
\mathrm{if}\ \sigma\times\sigma=\psi,
\end{cases}
\\
&
\widehat{\sigma}^{\,}_{\mathrm{R},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z')=
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z')\,\widehat{\sigma}^{\,}_{\mathrm{R},y}(t,z)
\nonumber\\
&\qquad
\times
\begin{cases}
e^{+\mathrm{i}\,\frac{\pi}{8}\,\delta^{\,}_{y,y^{\prime}}\,\mathrm{sgn}(z-z^{\prime})},
&
\mathrm{if}\ \sigma\times\sigma=\mathbbm{1},
\\
e^{-\mathrm{i}\,\frac{3\pi}{8}\,\delta^{\,}_{y,y^{\prime}}\,\mathrm{sgn}(z-z^{\prime})},
&
\mathrm{if}\ \sigma\times\sigma=\psi
\end{cases}
\\
&
\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z)\,\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z')=
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(t,z')\,\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z),
\end{align}
\end{subequations}
in two-dimensional Minkowski space
for any pair of wires $y$ and $y'$ and for any $z\neq z'$.
We have used the shorthand notation $\sigma\times\sigma=\mathbbm{1}$
and $\sigma\times\sigma=\psi$ to distinguish the two possible fusion outcomes.
It is important to stress here that this equal-time algebra is not well-defined
unless one specifies a fusion channel.
This ambiguity is essential. Its origin is physical,
and it reflects the non-Abelian nature of the twist field.
We will see in the next section that this ambiguity has
important consequences for the topological degeneracy.
\subsubsection{String operators and topological degeneracy on the two-torus}
\label{subsec: String operators and topological degeneracy}
We shall consider two distinct wires $y$ and $y'$
and a coordinate $z$ along any one of these wires.
Periodic boundary conditions are imposed both along the $y$-direction and
along the $z$-direction. Hence, the one-dimensional array of wires
has the topology of a torus.
We are going to construct the equal-time algebra
\begin{align}
\left\{
\widehat{\Gamma}^{(\frac{1}{2})}_{1},
\widehat{\Gamma}^{(1)}_{2}
\right\}=0
\label{eq2Dcase: Gamma (1/2) 1 anticommutes with Gamma (1) 2}
\end{align}
for a first pair of nonlocal operators
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$
and
$\widehat{\Gamma}^{(1)}_{2}$.
This pair will be shown to commute with the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}.
The nonlocal, nonunitary operator
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$
can be thought of as creating a pair of pointlike
``spin-$\frac{1}{2}$'' excitations,
transporting them in opposite directions around
a noncontractible cycle of the torus along the $y$-direction,
and then annihilating them.
Likewise, the nonlocal operator $\widehat{\Gamma}^{(1)}_{2}$
can be thought of as implementing a similar process for a
pair of pointlike ``spin-$1$'' excitations
around a noncontractible cycle of the torus along the $z$-direction.
Similarly, we are going to construct the equal-time algebra
\begin{align}
\left\{
\widehat{\Gamma}^{(1)}_{1},
\widehat{\Gamma}^{(\frac{1}{2})}_{2}
\right\}=0
\label{eq2Dcase: Gamma psi 1 anticommutes with Gamma sigma 2}
\end{align}
for a second pair of nonlocal operators
$\widehat{\Gamma}^{(1)}_{1}$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$.
This pair will also be shown to commute with the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering},
modulo appropriate regularization
of the operator $\widehat{\Gamma}^{(\frac{1}{2})}_{2}$, as we will discuss.
The nonlocal, unitary operator
$\widehat{\Gamma}^{(1)}_{1}$
can be thought of as creating a pair of ``spin-$1$'' excitations,
transporting them in opposite directions around
a noncontractible cycle of the torus along the $y$-direction,
and then annihilating them.
The nonlocal, nonunitary operator $\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
can be thought of as implementing the same process for a pair
of ``spin-$\frac{1}{2}$'' excitations
around a noncontractible cycle of the torus along the $z$-direction.
If we denote a ground state of the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}
by $|\Omega\rangle$, we shall demonstrate that the three states
\begin{align}
|\Omega\rangle,
\quad
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}\:=
\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
|\Omega\rangle,
\quad
&
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\:=
\widehat{\Gamma}^{(\frac{1}{2})}_{2}\,
|\Omega\rangle,
\label{eq2Dcase: three ground states}
\end{align}
are linearly independent ground states of the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}. The proof of this
claim relies on the vanishing equal-time commutators
\begin{align}
&
\left[
\widehat{\Gamma}^{(1)}_{2},
\widehat{\Gamma}^{(1)}_{1}
\right]=0,
\label{eq2Dcase: (1) 2-cyle commutes with (1) 1-cycle}
\end{align}
\begin{align}
&
\left[
\widehat{\Gamma}^{(\frac{1}{2})}_{1},
\widehat{\Gamma}^{(1)}_{1}
\right]=0,
\label{eq2Dcase: spin-1/2 1-cyle commutes with spin-1 1-cycle}
\end{align}
and
\begin{align}
&
\left[
\widehat{\Gamma}^{(1)}_{2},
\widehat{\Gamma}^{(\frac{1}{2})}_{2}
\right]=0.
\label{eq2Dcase: psi 2-cyle commutes with sigma 2-cycle}
\end{align}
Crucially, however, the exchange algebra of the nonlocal operators
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
suffers from the same ambiguity as that found on the right-hand side of
Eq.~\eqref{eq2Dcase: sigma sigma algebra}.
This is why one cannot infer from
Eqs.\ (\ref{eq2Dcase: Gamma (1/2) 1 anticommutes with Gamma (1) 2})--%
(\ref{eq2Dcase: psi 2-cyle commutes with sigma 2-cycle})
that the state
\begin{equation}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}\,
|\Omega\rangle
\end{equation}
is linearly independent from the states (\ref{eq2Dcase: three ground states}).
(See also Appendix%
~\ref{appendix: Commutation between string operators and the Hamiltonian}.)
\begin{proof}
The proof consists of three steps.
\textit{Step 1: ``Spin-$1$'' string operators.}
The first string operators that
we will construct are the ``spin-$1$'' string operators.
We begin with strings running along the $y$-direction,
perpendicular to the wires. These strings are built from the local
bilinears
\begin{subequations}
\begin{align}
\begin{split}
\widehat{\mathcal{O}}^{(1)}_{y}(t,z):=&\,
\widehat{\Phi}^{(1)\dagger}_{\mathrm{L},y}(t,z)\,
\widehat{\Phi}^{(1)}_{\mathrm{R},y}(t,z)
\\
=&\,
e^{-\mathrm i\frac{1}{\sqrt{2}}\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)}\,
e^{+\mathrm i\frac{1}{\sqrt{2}}\widehat{\phi}^{\,}_{\mathrm{R},y}(t,z)},
\end{split}
\end{align}
\end{subequations}
for any $0<z<L^{\,}_{z}$ (hereafter, we suppress the normal
ordering of the vertex operators).
Being a product of local unitary operators
[the $u(1)^{\,}_{2}$ vertex operators],
$\widehat{\mathcal{O}}^{(1)}_{y}(t,z)$
is also a local unitary operator.
Using Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}
for $k=2$,
we see that a product of ``spin-$1$'' bilinears in neighboring
wires commutes with the part of the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}
that connects the two wires, since
\begin{widetext}
\begin{align}
\label{eq2Dcase: psi y-string commutators}
\widehat{\mathcal{O}}^{(1)}_{y}(t,z)\,
\widehat{\mathcal{O}}^{(1)}_{y+1}(t,z)\,
e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z^{\prime})
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z^{\prime})\]}=
e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z^{\prime})
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z^{\prime})\]}\,
\widehat{\mathcal{O}}^{(1)}_{y}(t,z)\,
\widehat{\mathcal{O}}^{(1)}_{y+1}(t,z),
\end{align}
and because $\widehat{\mathcal{O}}^{(1)}_{y}(t,z)$
commutes with any operator from the $\mathbb{Z}^{\,}_{2}$ sector of the theory.
Thus, the nonlocal string operator
\begin{align}
\widehat{\Gamma}^{(1)}_{1}(t,z)\:=
\prod_{y=0}^{L^{\,}_{y}}
\widehat{\mathcal{O}}^{(1)}_{y}(t,z)
\label{eq: def Gamma psi 1 at z}
\end{align}
commutes with the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}
for any value of $0\leq z<L^{\,}_{z}$
when periodic boundary conditions are imposed in the $y$-direction.
The nonlocal operator (\ref{eq: def Gamma psi 1 at z})
is a member of the family
\begin{align}
\widehat{\Gamma}^{(1)}_{1}(t,z^{\,}_{1},\cdots z^{\,}_{L^{\,}_{y}})\:=
\widehat{\Phi}^{(1)\dagger}_{\mathrm{L},1}(t,z^{\,}_{1})\,
\widehat{\Phi}^{(1)}_{\mathrm{R},1}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(1)\dagger}_{\mathrm{L},2}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(1)}_{\mathrm{R},2}(t,z^{\,}_{3})\,
\cdots\,
\widehat{\Phi}^{(1)\dagger}_{\mathrm{L},L^{\,}_{y}}(t,z^{\,}_{L^{\,}_{y}})\,
\widehat{\Phi}^{(1)}_{\mathrm{R},L^{\,}_{y}}(t,z^{\,}_{1})
\label{eq: general ``spin-1'' string operator}
\end{align}
\end{widetext}
of operators, which all commute with the Hamiltonian
defined by Eq.~(\ref{eq2Dcase: 2D interaction})
for any values of $0\leq z^{\,}_{1},\cdots,z^{\,}_{L^{\,}_{y}}<L^{\,}_{z}$
when periodic boundary conditions are imposed in the $y$-direction.
Any ``spin-$1$'' string operator from the family
(\ref{eq: general ``spin-1'' string operator})
can be viewed as creating a pair of ``spin-$1$'' excitations and transporting
one of them around a noncontractible loop that encircles the torus in the
$y$-direction
(a noncontractible cycle along the $y$-direction),
before annihilating it with its partner.
To construct a ``spin-$1$'' string running along the $z$-direction,
parallel to the wires, we consider the unitary operator
\begin{subequations}
\begin{align}
\begin{split}
&\widehat{\mathcal{O}}^{(1)}_{\mathrm{M},y}(t,z^{\,}_{1},z^{\,}_{2})
\:=
\widehat{\Phi}^{(1)\dagger}_{\mathrm{M},y}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(1)}_{\mathrm{M},y}(t,z^{\,}_{1})
\\
&\qquad=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{\sqrt{2}}
\int\limits^{z^{\,}_{2}}_{z^{\,}_{1}}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{M},y}(t,z)
\Bigg)
\end{split}
\label{eq2Dcase: spin-1 z string operator}
\end{align}
for any $0\leq z^{\,}_{1},z^{\,}_{2}<L^{\,}_{z}$ and $\mathrm{M}=\mathrm{L},\mathrm{R}$..
\end{subequations}
Hence, $\widehat{\mathcal{O}}^{(1)}_{\mathrm{M},y}(t,z^{\,}_{1},z^{\,}_{2})$
is a bilocal unitary operator that also obeys
\begin{widetext}
\begin{align}
\widehat{\mathcal{O}}^{(1)}_{\mathrm{L},y}(t,z^{\,}_{1},z^{\,}_{2})\,
e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z)\]}
=&\,
e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z)\]}\,
\widehat{\mathcal{O}}^{(1)}_{\mathrm {L},y}(t,z^{\,}_{1},z^{\,}_{2})
\times e^{+\mathrm{i}2\pi\int\limits^{z^{\,}_{2}}_{z^{\,}_{1}}\mathrm{d}z^{\prime}\,
\delta(z-z^{\prime})},
\label{eq2Dcase: su(2)_{2} z-string commutator}
\end{align}
\end{widetext}
as a result of Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d} for $k=2$.
(A similar expression holds for $\mathrm{M}=\mathrm{R}$.)
Now define the nonlocal operator
\begin{align}
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)\:=
\widehat{\mathcal{O}}^{(1)}_{\mathrm{M},y}(t,0,L^{\,}_{z}),
\label{eq: def Gamma (1) 2 at y}
\end{align}
which commutes with the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering}
by Eq.~\eqref{eq2Dcase: su(2)_{2} z-string commutator}.
This ``spin-$1$'' string operator can be
viewed as transporting a ``spin-$1$'' excitation around a noncontractible
loop that encircles the torus in the $z$-direction
(a noncontractible cycle along the $z$-direction).
The equal-time commutation relation between the string operators
(\ref{eq: def Gamma psi 1 at z})
with $0<z<L^{\,}_{z}$ and
(\ref{eq: def Gamma (1) 2 at y})
is computed using
Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}
for $k=2$.
It is simply the commutative rule
\begin{align}
\widehat{\Gamma}^{(1)}_{1}(t,z)\,
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)
=&\,
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)\,
\widehat{\Gamma}^{(1)}_{1}(t,z),
\label{eq2Dcase: (1) (1) braiding}
\end{align}
for any $\mathrm{M}=\mathrm{L},\mathrm{R}$.
This result reflects the fact
that the spin-1 primary operator in the $su(2)^{\,}_{2}$ has
trivial self-monodromy.
We have established
Eq.~(\ref{eq2Dcase: (1) 2-cyle commutes with (1) 1-cycle})
provided we make the identifications
\begin{equation}
\widehat{\Gamma}^{(1)}_{1}(t,z)\to
\widehat{\Gamma}^{(1)}_{1},
\quad \mathrm{and} \quad
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)\to
\widehat{\Gamma}^{(1)}_{2},
\end{equation}
for some choice of chirality $\mathrm{M}$ and wire $y$.
\textit{Step 2: ``Spin-$\frac{1}{2}$'' string operators.}
We next construct string operators associated with the spin-$\frac{1}{2}$
primary of the $su(2)^{\,}_{k}$ theory.
We proceed according to a strategy similar to the one used for the
``spin-$1$'' strings. To construct a ``spin-$\frac{1}{2}$'' string along
the $y$-direction,
let $0<z,z^{\prime}<L^{\,}_{z}$ and consider the local ``spin-$\frac{1}{2}$''
bilinears
\begin{subequations}
\begin{align}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{y}(t,z)
&\:=
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{L},y}(t,z)\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{R},y}(t,z)
\nonumber
\\
&=
e^{-\mathrm i\frac{1}{2\sqrt{2}}\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)}
e^{+\mathrm i\frac{1}{2\sqrt{2}}\widehat{\phi}^{\,}_{\mathrm{R},y}(t,z)}\\
&\qquad\times
\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y}(t,z),
\nonumber
\end{align}
where we have defined the operator
\begin{align}
\label{eq2Dcase: phi (1/2) adjoint def}
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{L},y}(t,z)
\:=
e^{-\mathrm i\frac{1}{2\sqrt{2}}\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)}\,
\widehat{\sigma}^{\,}_{\mathrm{L},y}(t,z),
\end{align}
\end{subequations}
in which the adjoint operation pertains only to the $u(1)^{\,}_{k}$ vertex operator.
Using Eqs.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}
and \eqref{eq2Dcase: psi sigma algebra}, we find that
the equal-time product of such bilinears over all wires, namely
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)\:=
\prod_{y=0}^{L^{\,}_{y}}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{y}(t,z),
\label{eq: def Gamma (1/2) 1}
\end{align}
commutes with the interaction \eqref{eq2Dcase: su(2)_{2} backscattering}
for any value $0<z<L^{\,}_{z}$
when periodic boundary conditions are imposed in the
$y$-direction. This nonlocal, nonunitary operator is a member of the family
\begin{widetext}
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z^{\,}_{1},\cdots,z^{\,}_{L^{\,}_{y}})\:=
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{L},1}(t,z^{\,}_{1})\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{R},1}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{L},2}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{R},2}(t,z^{\,}_{3})\,
\cdots\,
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{L},L^{\,}_{y}}(t,z^{\,}_{L^{\,}_{y}})\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{R},L^{\,}_{y}}(t,z^{\,}_{1})
\label{eq: general ``spin-1/2'' string operator}
\end{align}
\end{widetext}
of operators that commute with the Hamiltonian
defined by Eq.~(\ref{eq2Dcase: 2D interaction})
for any values of $0<z^{\,}_{1},\cdots,z^{\,}_{L^{\,}_{y}}<L^{\,}_{z}$
when periodic boundary conditions are imposed in the $y$-direction.
Any ``spin-$\frac{1}{2}$'' string operator from the family
(\ref{eq: general ``spin-1/2'' string operator})
can be interpreted as creating a pair of ``spin-$\frac{1}{2}$'' excitations and transporting
one of them around a noncontractible cycle along the $y$-direction,
before annihilating it with its partner.
We first observe that the operators
$\widehat{\Gamma}^{(1)}_{1}(t,z)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z^{\prime})$
commute with one another for any $z$
and $z^{\prime}$, as one can show using the equal-time algebra
\eqref{eq: def chiral parafermion and boson rep su(2)k d},
\begin{equation}
\widehat{\Gamma}^{(1)}_{1}(t,z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z^{\prime})=
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z^{\prime})\,
\widehat{\Gamma}^{(1)}_{1}(t,z).
\end{equation}
We have established Eq.\
(\ref{eq2Dcase: spin-1/2 1-cyle commutes with spin-1 1-cycle})
provided we make the identifications
\begin{equation}
\widehat{\Gamma}^{(1)}_{1}(t,z)\to
\widehat{\Gamma}^{(1)}_{1},
\qquad
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z^{\prime})\to
\widehat{\Gamma}^{(\frac{1}{2})}_{1}.
\end{equation}
We claim that the ``spin-$\frac{1}{2}$'' string $\widehat{\Gamma}^{(\frac{1}{2})}_{1}$ can be
interpreted as an operator that ``twists,'' from
antiperiodic to periodic, the boundary conditions of a
``spin-$1$'' excitation that encircles the torus in the $z$-direction.
To see that this is the case, we use the chiral boson algebra of
Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}
to show that the equal-time operator algebra
\begin{align}
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)
=&\,
-
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)\,
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)
\label{eq2Dcase: Gamma^(1)_{2},y and Gamma^(1/2)_{1} commutator}
\end{align}
holds for any choice of chirality $\mathrm{M}=\mathrm{L},\mathrm{R}$ and
wire $y$. We further recall that the operator
$\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)$ transports
a ``spin-$1$'' excitation around the torus along the $z$-direction.
Thus, Eq.~\eqref{eq2Dcase: Gamma^(1)_{2},y and Gamma^(1/2)_{1} commutator}
shows that the amplitude for transporting a ``spin-$1$'' excitation around the torus
and then applying the operator $\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)$ differs by
a minus sign from the amplitude for applying the operator $\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)$
and then transporting a ``spin-$1$'' excitation around the torus. This is precisely the
action of an operator that twists the boundary conditions of a``spin-$1$'' excitation.
In deriving Eq.~\eqref{eq2Dcase: Gamma^(1)_{2},y and Gamma^(1/2)_{1} commutator},
we have established
Eq.~(\ref{eq2Dcase: Gamma (1/2) 1 anticommutes with Gamma (1) 2})
provided that we make the identifications
\begin{equation}
\widehat{\Gamma}^{(1)}_{2,\mathrm{M},y}(t)\to
\widehat{\Gamma}^{(1)}_{2},
\qquad
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(t,z)\to
\widehat{\Gamma}^{(\frac{1}{2})}_{1}.
\end{equation}
for some choice of chirality $\mathrm{M}$ and wire $y$.
Next, we seek an operator that twists the boundary conditions
of a ``spin-$1$'' excitation encircling the torus along the $y$-direction.
We proceed in direct analogy with Eq.~\eqref{eq2Dcase: spin-1 z string operator}
by defining the (nonlocal, nonunitary) operator
\begin{align}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\mathrm{M},y^{\prime}}(t,z^{\,}_{1},z^{\,}_{2})
&\:=
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\mathrm{M},y^{\prime}}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(\frac{1}{2})}_{\mathrm{M},y^{\prime}}(t,z^{\,}_{1})
\nonumber
\\
&=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{z^{\,}_{2}}_{z^{\,}_{1}}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{M},y^{\prime}}(t,z)
\Bigg)
\nonumber
\\
&\qquad
\times
\widehat{\sigma}^{\,}_{\mathrm{M},y^{\prime}}(t,z^{\,}_{2})\,
\widehat{\sigma}^{\,}_{\mathrm{M},y^{\prime}}(t,z^{\,}_{1}).
\label{eq2Dcase: sigma_M,y' sigma_M,y' definition}
\end{align}
We seek
to define a string operator by taking $z^{\,}_{1}\to 0$
and $z^{\,}_{2}\to L^{\,}_{z}$. However, one must be careful in taking
these limits since Eq.~\eqref{eq2Dcase: sigma_M,y' sigma_M,y' definition}
contains two \textit{chiral} $\mathbb{Z}^{\,}_{2}$ twist fields in the
\textit{same} wire.
Due to the ambiguity of the OPE \eqref{eq2Dcase: sigma sigma algebra},
such a product is ill-defined unless a fusion channel is specified.
[Meanwhile, the product of $u(1)^{\,}_{2}$ vertex operators is unambiguous.]
By analogy with the construction of
$\widehat{\Gamma}^{(1)}_{2}$
in Eq.~\eqref{eq2Dcase: spin-1 z string operator}, we would like to define
the string operator $\widehat{\Gamma}^{(\frac{1}{2})}_{2}$ in such a way as
to leave the system in the vacuum sector. Hence, the natural choice is to
specify that the two $\widehat{\sigma}^{\,}_{M,y^{\prime}}$ operators in
Eq.~\eqref{eq2Dcase: sigma_M,y' sigma_M,y' definition}
fuse to the identity operator $\mathbbm{1}$.
In addition to providing
a sensible parallel with the construction of
$\widehat{\Gamma}^{(1)}_{1}$, this choice agrees with the choice
made in the construction of the operator that tunnels an $e/4$
quasiparticle across a quantum point contact in the Moore-Read
state~\cite{Fendley07}.
This motivates the definition of the ``spin-$\frac{1}{2}$'' string operator
\begin{align}
\begin{split}
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)
&\:=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{M},y^{\prime}}(t,z)
\Bigg)
\\
&\qquad
\times
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\mathrm{M},y^{\prime}}(t,0)\,
\widehat{\sigma}^{\,}_{\mathrm{M},y^{\prime}}(t,\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}},
\end{split}
\label{eq2Dcase: Gamma^(1/2)_2,y' definition}
\end{align}
where
$\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$
is the projection operator onto the fusion channel
$\sigma\times\sigma=\mathbbm{1}$.
One can show that this projector does not affect the
algebra of twist operators $\widehat{\sigma}^{\,}_{\mathrm{M},y}$
and Majorana operators
$\widehat{\psi}^{\,}_{\mathrm{M},y}$.
We claim that the string operator
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)$
defined in this way commutes with the interaction
\eqref{eq2Dcase: su(2)_{2} backscattering} in the limit $\epsilon\to 0$.
To see this, note that
\begin{widetext}
\begin{align}
\lim_{\epsilon\to0}
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)\,
\widehat{\psi}^{\,}_{\mathrm{R},y+1}(t,z)\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,0)\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,\epsilon)
=
\lim_{\epsilon\to0}
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,0)\,
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(t,\epsilon)\,
\widehat{\psi}^{\,}_{\mathrm{L},y}(t,z)\,
\widehat{\psi}^{\,}_{\mathrm{R},y+1}(t,z)
\times
\begin{cases}
+1, & y\neq y^{\prime},
\\
-1, & y=y^{\prime},
\end{cases}
\end{align}
follows from the algebra \eqref{eq2Dcase: psi sigma algebra},
while
\begin{align}
\begin{split}
&e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z)\]}\,
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{L},y^{\prime}}(t,z)
\Bigg)
\\
&\qquad\qquad\qquad
=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{L},y^{\prime}}(t,z)
\Bigg)\,
e^{+\mathrm{i}\sqrt{1/2}\,\[\widehat{\phi}^{\,}_{\mathrm{L},y}(t,z)
-\widehat{\phi}^{\,}_{\mathrm{R},y+1}(t,z)\]}
\times
\begin{cases}
+1, & y\neq y^{\prime},\\
-1, & y=y^{\prime},
\end{cases}
\end{split}
\end{align}
\end{widetext}
follows from the algebra
\eqref{eq: def chiral parafermion and boson rep su(2)k d}.
(Similar expressions hold for $\mathrm{M}=\mathrm{R}$.)
Consequently, $\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)$
commutes with the interaction \eqref{eq2Dcase: su(2)_{2} backscattering}
in the limit $\epsilon\to 0$~\cite{foot1}.
Moreover, we can also show that
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)$
twists the boundary conditions of a ``spin-$1$'' excitation encircling the torus
along the $y$-direction. To do this, we use the algebra
\eqref{eq: def chiral parafermion and boson rep su(2)k d} to compute
the exchange relation (in the limit $\epsilon\to0$)
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)\,
\widehat{\Gamma}^{(1)}_{1}(t,z)\,
=&\,
-
\widehat{\Gamma}^{(1)}_{1}(t,z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon),
\label{eq2Dcase: Gamma^(1/2)_2,y' and Gamma^(1)_{1} commutator}
\end{align}
which holds for any chirality $\mathrm{M}$ and wire $y^{\prime}$.
This exchange relation has an interpretation similar to
Eq.~\eqref{eq2Dcase: Gamma^(1)_{2},y and Gamma^(1/2)_{1} commutator}.
Thus, we have established Eq.\
(\ref{eq2Dcase: Gamma psi 1 anticommutes with Gamma sigma 2})
provided we make the identifications
\begin{equation}
\widehat{\Gamma}^{(1)}_{1}(t,z)\to
\widehat{\Gamma}^{(1)}_{1},
\qquad
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(t,\epsilon)
\to
\widehat{\Gamma}^{(\frac{1}{2})}_{2},
\end{equation}
for infinitesimal $\epsilon>0$.
By assumption $y\neq y^{\prime}$. Hence, the operators
$\widehat{\Gamma}^{\psi}_{2,y}\to\widehat{\Gamma}^{(1)}_{2}$
and
$\widehat{\Gamma}^{\sigma}_{2,\mathrm{M},y^{\prime}}\to
\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
commute with one another in a trivial way. This establishes
Eq.~(\ref{eq2Dcase: psi 2-cyle commutes with sigma 2-cycle}).
\textit{Step 3: The topological degeneracy.}
There exists a many-body ground state
\begin{subequations}
\label{eq: Ansatz for three ground states}
\begin{equation}
\ket{\Omega}\equiv\ket{\mathbbm{1}}
\label{eq: Ansatz for three ground states a}
\end{equation}
of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
defined in Eq.~(\ref{eq2Dcase: su(2)_{2} backscattering})
from which
we can obtain two additional many-body states by acting
with the ``spin-$\frac{1}{2}$'' string operators along the
$y$- and $z$-directions,
respectively,
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}\:=
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\ket\Omega
\label{eq: Ansatz for three ground states b}
\end{align}
and
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\:=
\lim_{\epsilon\to0}
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)
\ket\Omega,
\label{eq: Ansatz for three ground states c}
\end{align}
\end{subequations}
for any $z$, $y^{\prime}$, and $z^{\,}_{1}$.
It is important to point out that not all
choices of $\ket{\Omega}$ are equal. As argued in Appendix%
~\ref{appendix: Commutation between string operators and the Hamiltonian},
depending on the topological sector in which the state
$\ket{\Omega}$ resides, one or both of the states
\eqref{eq: Ansatz for three ground states b}
and
\eqref{eq: Ansatz for three ground states c}
could have norm zero or infinity.
We will first prove that the many-body states
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
share the same eigenvalue of $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
as $\ket{\Omega}$.
Second, we will prove that the many-body states
(\ref{eq: Ansatz for three ground states})
are linearly independent. In doing so, we will have established
that the ground state degeneracy on the torus
of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
is threefold.
First, we recall that
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$
commutes with the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
defined in Eq.~(\ref{eq2Dcase: su(2)_{2} backscattering}).
Hence, the many-body state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$ defined in
Eq.~\eqref{eq: Ansatz for three ground states b}
is a ground state of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$.
Making sure to treat
the limit $\epsilon\to0$ with care,
we show in Appendix%
~\ref{appendix: Commutation between string operators and the Hamiltonian}
that the many-body state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$ defined
in Eq.~\eqref{eq: Ansatz for three ground states b}
is also a ground state of the interaction
$\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$.
Now, we are going to show that the three many-body states
(\ref{eq: Ansatz for three ground states})
are linearly independent.
The operators $\widehat{\Gamma}^{(1)}_{1}$ and
$\widehat{\Gamma}^{(1)}_{2}$
commute with the interaction
$\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
and with each other
[recall Eq.~(\ref{eq2Dcase: (1) 2-cyle commutes with (1) 1-cycle})].
They are thus simultaneously diagonalizable.
Consequently, we can choose $\ket{\Omega}$
to be a simultaneous eigenstate of the pair of operators
$\widehat{\Gamma}^{(1)}_{1}$
and
$\widehat{\Gamma}^{(1)}_{2}$.
Both
$\widehat{\Gamma}^{(1)}_{1}$
and
$\widehat{\Gamma}^{(1)}_{2}$
are unitary, i.e., there should exist the pair of unimodular complex numbers
$\omega^{(1)}_{1}\neq0$ and $\omega^{(1)}_{2}\neq0$ such that
\begin{subequations}
\begin{align}
\widehat{\Gamma}^{(1)}_{1}\,\ket{\Omega}=
\omega^{(1)}_{1}\,\ket{\Omega},
\end{align}
and
\begin{align}
\widehat{\Gamma}^{(1)}_{2}\,\ket{\Omega}=
\omega^{(1)}_{2}\,\ket{\Omega},
\end{align}
\end{subequations}
respectively.
Because of the anticommutator
(\ref{eq2Dcase: Gamma (1/2) 1 anticommutes with Gamma (1) 2}),
we find the eigenvalue
\begin{align}
&
\widehat{\Gamma}^{(1)}_{2}\,\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
=
-
\omega^{(1)}_{2}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}.
\label{eq2Dcase: proof degeneracy of at least three a}
\end{align}
Hence, $\ket{\Omega}$ and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
are simultaneous eigenstates of the unitary operator
$\Gamma^{(1)}_{2}$ with distinct eigenvalues.
As such, $\ket{\Omega}$ and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$ are othogonal.
Similarly, because of the anticommutator
(\ref{eq2Dcase: Gamma psi 1 anticommutes with Gamma sigma 2}),
we find the eigenvalue
\begin{align}
&
\widehat{\Gamma}^{(1)}_{1}\,\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
=
-
\omega^{(1)}_{1}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}.
\label{eq2Dcase: proof degeneracy of at least three b}
\end{align}
Hence, $\ket{\Omega}$ and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are simultaneous eigenstates of the unitary operator
$\Gamma^{(1)}_{1}$ with distinct eigenvalues.
As such, $\ket{\Omega}$ and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$ are othogonal.
To complete the proof that
$\ket{\Omega}$,
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$,
and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are linearly independent, it suffices to show that
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are orthogonal.
Because of the commutator
(\ref{eq2Dcase: spin-1/2 1-cyle commutes with spin-1 1-cycle}),
we find the eigenvalue
\begin{align}
&
\widehat{\Gamma}^{(1)}_{1}\,\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
=
+
\omega^{(1)}_{1}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}.
\label{eq2Dcase: proof degeneracy of at least three c}
\end{align}
Hence, $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$ and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are simultaneous eigenstates of the unitary operator
$\widehat{\Gamma}^{(1)}_{1}$
with the pair of distinct eigenvalues
$+\omega^{(1)}_{1}$
and
$-\omega^{(1)}_{1}$.
As such,
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are orthogonal.
We note that the commutator
(\ref{eq2Dcase: psi 2-cyle commutes with sigma 2-cycle})
could equally well have been used to show that
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$ and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
are simultaneous eigenstates of the unitary operator
$\widehat{\Gamma}^{(1)}_{2}$
with the pair of distinct eigenvalues
$+\omega^{(1)}_{2}$
and
$-\omega^{(1)}_{2}$.
As promised, we have shown that the ground-state
manifold of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
on the torus
is threefold degenerate.
\end{proof}
It is useful to pause at this stage to interpret this lower
bound on the ground state degeneracy and how it comes about.
Naively, given two pairs of anticommuting nonlocal operators, all of which
commute with the Hamiltonian, [i.e., given
Eqs.\ \eqref{eq2Dcase: Gamma (1/2) 1 anticommutes with Gamma (1) 2}
and \eqref{eq2Dcase: Gamma psi 1 anticommutes with Gamma sigma 2}]
there are at most four degenerate ground states. In the case of
Kitaev's toric code~\cite{Kitaev03}, the dimensionality of the ground
state manifold saturates this upper bound. However, in the case of the
two-dimensional state of matter that we have constructed here, we argue
that this is not the case. The reason for this is intimately related to the
nonunitarity of the string operators $\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$
and $\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$.
In particular, we assert that neither of the naively-expected
fourth states, namely
\begin{subequations}
\label{eq2Dcase: definition of ket sigma_1,2 and ket sigma_2,1}
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\:=
\lim_{\epsilon\to0}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\ket{\Omega},
\label{eq2Dcase: definition of ket sigma_1,2}
\end{align}
and
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}}\:=
\lim_{\epsilon\to0}
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\ket{\Omega},
\label{eq2Dcase: definition of ket sigma_2,1}
\end{align}
\end{subequations}
belongs to the ground-state manifold
of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$.
Note that the limit $\epsilon\to0$ above is to be taken after forming
the products $\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$, as discussed in
Footnote~\cite{foot1} and Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}.
If the operator products
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$ were to commute
with the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
in the limit $\epsilon\to0$, as they would in an Abelian topological phase,
then there would be no obstruction
to the states $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$ and
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
belonging to the ground-state manifold.
The proof that such an obstruction exists in the present (non-Abelian)
case is undertaken in two complementary ways in the present work.
The first, which we call the ``algebraic'' approach,
relies on diagrammatic techniques developed in Appendix
\ref{appendix: Diagrammatics for operator algebra in the Ising CFT},
and is presented below.
The second, which we call the ``analytic'' approach,
is carried out in Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}.
Both the ``algebraic'' and ``analytic'' proofs rely on the fact,
discussed in Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian},
that the
operator products
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$
are not bound to
commute with the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$
in the limit $\epsilon\to0$.
We now proceed with the ``algebraic'' version of the proof,
and refer the reader to Appendices
\ref{appendix: Diagrammatics for operator algebra in the Ising CFT}
and
\ref{appendix: Commutation between string operators and the Hamiltonian}
for more details.
\begin{proof}[Proof (``algebraic'')]
We introduce the projection operator
\begin{equation}\label{eq2Dcase: projection operator}
\begin{split}
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}\:=&\,
\mathcal{N}^{-1}_{\mathbbm 1}\,
\ket{\mathbbm 1}\bra{\mathbbm 1}
\\
&\,
+
\mathcal{N}^{-1}_{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}\bra{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
\\
&\,
+
\mathcal{N}^{-1}_{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\bra{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
\\
&\,
+
\cdots
\end{split}
\end{equation}
onto the ground state manifold. Here,
$\mathcal{N}^{\,}_{\mathbbm 1}$ is the squared norm of the state
$\ket{\mathbbm 1}\equiv\ket{\Omega}$,
$\mathcal{N}^{\,}_{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$
is the squared norm of the state
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$,
$\mathcal{N}^{\,}_{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
is the squared norm of the state
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$,
and
$\cdots$ is a sum over any remaining elements from the orthonormal basis
of the ground state manifold.
By definition, any one of the three states
$\ket{\mathbbm 1}$,
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$,
and $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
defined in Eq.~(\ref{eq: Ansatz for three ground states})
is invariant under the action of
\begin{equation}
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=
\widehat{\mathcal{P}}^{2}_{\mathrm{GSM}}.
\end{equation}
Hence, we may write
\begin{subequations}
\label{eq2Dcase: act with P_GSM on new ground states}
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
=&\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
=
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\ket{\Omega},
\label{eq2Dcase: act with P_GSM on new ground states a}
\\
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
=&\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
=
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\lim_{\epsilon\to0}
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\ket{\Omega}.
\end{align}
\end{subequations}
On the other hand,
\begin{align}
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}\,
\widehat{\mathcal{O}}\,
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}=0
\end{align}
must hold for any operator $\widehat{\mathcal{O}}$
such that $\widehat{\mathcal{O}}$
returns an excited state when applied to any state from the
ground-state manifold.
We are first going to show that the operators
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$
do not commute in the limit $\epsilon\to0$.
After that, we will
elaborate on why the state
$
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
$
does not belong to the ground-state manifold
of the interaction $\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}$.
We begin by considering the exchange algebra of the string operators
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)$
defined in Eqs.~(\ref{eq: def Gamma (1/2) 1})
and (\ref{eq2Dcase: Gamma^(1/2)_2,y' definition}),
respectively.
Specifically, we consider the product
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)
\propto&\,
\(\prod^{L^{\,}_{y}}_{y=0}
\widehat{\sigma}^{\,}_{\mathrm{L},y}(z)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y}(z) \)
\label{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 product a}
\\
&\qquad
\times
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(0)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}},\nonumber
\end{align}
where $\epsilon>0$ is infinitesimal and
we have also omitted the operators in the $u(1)^{\,}_{2}$
sector appearing in the
definition \eqref{eq2Dcase: Gamma^(1/2)_2,y' definition},
as these operators commute with all operators in the $\mathbb{Z}^{\,}_{2}$ sector.
Using the fact that
twist operators in different wires
(and in different chiral sectors of the same wire) commute,
we deduce that
\begin{align}
&
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)
\propto\,
\(\prod^{\,}_{y\neq y^{\prime}}
\widehat{\sigma}^{\,}_{\mathrm{L},y}(z)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y}(z) \)
\widehat{\sigma}^{\,}_{\mathrm{L},y^{\prime}}(z)
\nonumber
\\
&\qquad
\times
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(0)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}.
\label{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 product b}
\end{align}
Since all operators in the first line of the right-hand side above
commute with all operators in the second line, computing the exchange algebra
of the operators $\widehat{\Gamma}^{(\frac{1}{2})}_{1}$
and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
boils down to considering the following product of operators,
\begin{align}
\lim_{\substack{z^{\,}_{2}\to z^{\,}_{1}+\epsilon\\ z^{\,}_{1}\to0}}
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{1})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{2})\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}.
\end{align}
Using the prescriptions of Appendix
\ref{appendix: Diagrammatics for operator algebra in the Ising CFT},
we find that the process of
commuting the leftmost operator, $\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z)$,
past the remaining two operators is represented by the diagram
\begin{align}
\includegraphics[width=.12\textwidth]{Figures6}.
\end{align}
Untwisting the legs of this fusion diagram, we find
\begin{widetext}
\begin{align}\label{eq2Dcase: 3-sigma diagram untwisting}
\includegraphics[width=.65\textwidth]{Figures7}
\end{align}
\end{widetext}
where the $F$- and $R$-symbols are given in Appendix
\ref{appendix: Diagrammatics for operator algebra in the Ising CFT}.
The diagrammatic relation expressed in Eq.\
\eqref{eq2Dcase: 3-sigma diagram untwisting}
can be rewritten as the algebraic statement
\begin{align}\label{eq2Dcase: 3-sigma algebraic untwisting}
&
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{1})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{2})\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}
\\
&\qquad\qquad
=
e^{+\mathrm{i} \frac{\pi}{4}}\,
\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{1})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{2})\,
\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z),
\nonumber
\end{align}
where $\widehat{\mathcal{P}}^{\,}_{\psi}$ is a projection operator
that projects the product
$\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{1})\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(z^{\,}_{2})$
into the fusion channel $\sigma\times\sigma=\psi$. Taking the limits
$z^{\,}_{2}\to z^{\,}_{1}+\epsilon$
and
$z^{\,}_{1}\to 0$
and restoring the operators
$\widehat{\sigma}^{\,}_{\mathrm{M},y}(z)$ present in
Eq.~\eqref{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 product b}
(as well as the operators from the $u(1)^{\,}_{2}$ sector
that were omitted there), we arrive at the relation
\begin{subequations}
\begin{equation}
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{R},y^{\prime}}(\epsilon)=
e^{+\mathrm{i} \frac{3\pi}{4}}\,
\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2,\mathrm{R},y^{\prime}}
(\epsilon)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}(z),
\label{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 exchange algebra}
\end{equation}
in the limit $\epsilon\to0$,
where we have defined the operator
\begin{equation}
\begin{split}
\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2,\mathrm{R},y^{\prime}}
(\epsilon)
&\:=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\mathrm{R},y^{\prime}}(t,z)
\Bigg)
\\
&\qquad
\times
\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(0)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\psi},
\end{split}
\end{equation}
\end{subequations}
which is identical to the operator $\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
defined in Eq.~\eqref{eq2Dcase: Gamma^(1/2)_2,y' definition},
except that the product
$\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(0)\,
\widehat{\sigma}^{\,}_{\mathrm{R},y^{\prime}}(\epsilon)$
is evaluated in the fusion channel $\psi$ rather than the fusion channel
$\mathbbm{1}$. This difference is fundamental.
Since the two twist operators
entering the operator
$\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2}$
fuse to $\psi$, this operator can be interpreted as adding an extra Majorana
fermion to the state on which it acts. Acting with
$\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2}$
on any of the states
$\ket{\mathbbm 1},\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}},
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}},\cdots$
in the ground-state manifold
of the interaction $\widehat{\mathcal H}^{\,}_{\rm bs}$
can then be viewed as
creating an \textit{excited state}
of the interaction $\widehat{\mathcal H}^{\,}_{\rm bs}$
with one extra fermion.
In other words, we have
\begin{align}
\label{eq2Dcase: projection of tilde Gamma^sigma_2 into GSM vanishes}
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\lim_{\epsilon\to0}
\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{2,\mathrm{R},y^{\prime}}
(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=
0.
\end{align}
This relation is crucial in what follows. Note also
the difference between the phase on
the RHS of
Eq.~\eqref{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 exchange algebra}
and that on the RHS of
Eq.~\eqref{eq2Dcase: 3-sigma algebraic untwisting},
which comes from commutators in the $u(1)^{\,}_{2}$
sector.
We are now prepared to exclude the state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
from the ground-state manifold of the interaction
$\widehat{\mathcal H}^{\,}_{\rm bs}$. Applying Eq.\
\eqref{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 exchange algebra}
to the definition
\eqref{eq2Dcase: definition of ket sigma_1,2} of the state
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$, we obtain
\begin{align}
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}
=
e^{+\mathrm{i}\, \frac{3\pi}{4}}\,
\lim_{\epsilon\to0}
\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2,\mathrm{R},y^{\prime}}
(\epsilon)\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}.
\end{align}
If the state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
\textit{is} in the ground-state manifold
of the interaction $\widehat{\mathcal H}^{\,}_{\rm bs}$,
then it cannot be a null vector of $\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}$.
However, using Eqs.\ \eqref{eq2Dcase: act with P_GSM on new ground states}
and \eqref{eq2Dcase: projection of tilde Gamma^sigma_2 into GSM vanishes},
we find that
\begin{align}
\begin{split}
&\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\, \ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}\\
&\quad=\,
e^{+\mathrm{i}\, \frac{3\pi}{4}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\lim_{\epsilon\to0}
\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2,\mathrm{R},y^{\prime}}
(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}}
\\
&\quad=\,
0.
\end{split}
\label{eq2Dcase: exclude ket sigma_1,2}
\end{align}
Thus, the state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{2}}$ does not lie in the ground-state
manifold of the interaction $\widehat{\mathcal H}^{\,}_{\rm bs}$.
Similarly, the
state $\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{2}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{1}}$ defined in Eq.\
\eqref{eq2Dcase: definition of ket sigma_2,1}
is excluded from the ground-state manifold.
We note in passing that a related line of
reasoning was used in Ref.~\cite{Oshikawa07} to exclude certain
states from the ground-state manifold of the gauged $p+\mathrm{i}\, p$
superconductor (see also Ref.~\cite{Read00}).
\end{proof}
\begin{table}[t]
\begin{tabular}{|c||c|c||c|c|}
\hline
\ &
$\widehat{\Gamma}^{(1)}_{1}$ &
$\widehat{\Gamma}^{(1)}_{2}$ &
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$ &
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
\\
\hline\hline
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$ &
$+$ &
$-$ &
$+$ &
$\text{\ding{55}}$
\\
\hline
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$ &
$-$ &
$+$ &
$\text{\ding{55}}$ &
$+$
\\
\hline
\end{tabular}
\caption{
Summary of the algebra of the string operators
$\widehat{\Gamma}^{(1)}_{1,2}$ and $\widehat{\Gamma}^{(\frac{1}{2})}_{1,2}$.
Entries corresponding
to a pair of operators that commute are labeled with a $+$.
Entries corresponding
to a pair of operators that anticommute are labeled with a $-$.
Entries corresponding
to a pair of operators that neither commute nor anticommute
are labeled with a $\text{\ding{55}}$.
\label{table2Dcase: summary of algebra}
}
\end{table}
In summary, we have shown that the
$su(2)^{\,}_{2}$ coupled-wire construction
in (2+1)-dimensional spacetime
has a threefold topological degeneracy on the two-torus. The proof that
this topological degeneracy is threefold and not fourfold
relied on the observation that the ``spin-$\frac{1}{2}$'' string operators
obey the non-Abelian exchange algebra
\eqref{eq2Dcase: Gamma^sigma_1 Gamma^sigma_2 exchange algebra}.
This algebra, whereby exchanging the two operators does not simply
produce a phase factor, but instead enacts a nontrivial transformation
on the operators themselves, is the essence of what it means to be
a non-Abelian topological phase. We will see that a similar, albeit
richer, algebra arises in the (3+1)-dimensional case discussed in the
next section. For future comparison with the
(3+1)-dimensional case, we summarize
the exchange algebra of the ``spin-$1$'' and ``spin-$\frac{1}{2}$''
string operators in
Table~\ref{table2Dcase: summary of algebra}. In both the (2+1)-
and (3+1)-dimensional cases, the non-Abelian algebra that encodes
the topological degeneracy is induced by the algebra of the
primary operators of the corresponding CFT.
\section{Weak non-Abelian topological order in three dimensions}
\label{sec: Non-Abelian topological order in three dimensions}
In this section and the next, we consider 3D generalizations of
the class of 2D models defined in Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}.
In Sec.~\ref{sec: Non-Abelian topological order in three dimensions}
we construct a family of weak topological phases
that can be viewed as stacks of the topological phases discussed in
Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions}
In Sec.%
~\ref{sec: Possible strong non-Abelian topological order in three dimensions},
we consider a more complicated interaction that may yield a truly 3D
non-Abelian topological phase.
These two families of phases share a common physical setup, which
we describe below.
\subsection{Physical setup for 3D wire arrays}
\label{subsec: Definition of the class of models 3D}
\begin{figure}[t]
\centering
\includegraphics[width=0.385\textwidth]{Figures8}
\caption{(Color online)
Schematic of the couplings between chiral $SU(2)$ sectors
in a four-wire unit cell.
}
\label{fig: Schematic of the su(2)_{k} couplings}
\end{figure}
Consider a square lattice $\Lambda$ of wires,
each described by the Lagrangian density
\eqref{eq: single non-abelian wire}.
We want to break the degrees of freedom in any one of the
identical wires up into four groups, two of which
contain only right-moving degrees of freedom and two of which contain
only left-moving degrees of freedom
(see Fig.\ \ref{fig: general wire setup}).
Consequently, let each wire \eqref{eq: single non-abelian wire}
contain $N^{\,}_{\mathrm{c}}=2k$ colors of fermions, so that the
full symmetry group of each wire is
$U(4k)^{\,}_{\mathrm{L}}\times U(4k)^{\,}_{\mathrm{R}}$.
However, we wish to employ the conformal
embedding \eqref{eq: conformal embedding} to write down the couplings
in our theory in terms of currents. Thus, let us consider only
couplings that are symmetric under the subgroup
$U(2k)^{\,}_{\mathrm{M}}\times U(2k)^{\,}_{\mathrm{M}}\subset U(4k)^{\,}_{\mathrm{M}}$
with $\mathrm{M}=\mathrm{L},\mathrm{R}$.
[Note that the central charges associated with the groups
$U(2k)^{\,}_{\mathrm{M}}\times U(2k)^{\,}_{\mathrm{M}}$
and $U(4k)^{\,}_{\mathrm{M}}$ are identical~\cite{DiFrancesco97}. Thus,
we can use couplings with either symmetry to fully gap the
theory.]
Then, we can use the identity
\begin{align}\label{eq: conformal embedding k}
u(2k)^{\,}_{1}=
u(1)\oplus
su(2)^{\,}_{k}\oplus
su(k)^{\,}_{2}
\end{align}
to define the $\mathrm{M}$-moving chiral currents
$\widehat{j}^{\,}_{\mathrm{M}}$, $\widehat{J}^{a}_{\mathrm{M}}$, and
$\widehat{\mathsf{J}}^{\mathsf{a}}_{\mathrm{M}}$, which are given by
Eqs.\ \eqref{eq: non-Abelian currents}
with the substitution $N^{\,}_{\mathrm{c}}\to k$.
Because we are considering couplings that are symmetric under
rotations in
$\big[U(2k)\times U(2k)\big]^{\,}_{\mathrm{L}}\times
\big[U(2k)\times U(2k)\big]^{\,}_{\mathrm{R}}$,
there are actually two copies of each of the chiral currents
$\widehat{j}^{\,}_{\mathrm{M}},\widehat{J}^{a}_{\mathrm{M}},$ and
$\widehat{\mathsf{J}}^{\mathsf{a}}_{\mathrm{M}}$
with $\mathrm{M}=\mathrm{L},\mathrm{R}$ in each wire.
Therefore, we adopt an additional label $\gamma=1,2$ to
distinguish the chiral currents
$\widehat{j}^{\,}_{\gamma,\mathrm{M}}$, $\widehat{J}^{a}_{\gamma,\mathrm{M}}$, and
$\widehat{\mathsf{J}}^{\mathsf{a}}_{\gamma,\mathrm{M}}$ from one another.
The label $\gamma$ is somewhat redundant
in that it will always transform trivially
under all the symmetries that we shall impose.
We only use it to keep track of the two
independent copies of each set of currents.
\subsection{Interwire couplings for weak non-Abelian topological order}
\label{subsec: Interwire couplings for weak non-Abelian topological order}
Next, as in Sec.~\ref{subsec: Definition of the class of models 2D},
we gap out the $u(1)$ and $su(k)^{\,}_{2}$ degrees of freedom
by turning on intra-wire interactions of the
form~\eqref{eq: umklapp} and \eqref{eq: su(N_c) current-current},
respectively, for each $\gamma=1,2$. The remaining $su(2)^{\,}_{k}$
degrees of freedom are coupled in the following way.
First, we define a square lattice $\tilde{\Lambda}$, whose unit cell
is enlarged with respect to that of the square lattice $\Lambda$.
Let the unit cell of $\tilde{\Lambda}$ contain four quantum wires,
which we label by an index $J=A,B,C,D$.
This enlarged unit cell is depicted in
Fig.\ \ref{fig: Schematic of the su(2)_{k} couplings}.
Each $su(2)^{\,}_{k}$ current operator
$\widehat{J}^{a}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}$
thus carries the labels $\gamma=1,2$ and
$\mathrm{M}=\mathrm{L},\mathrm{R}$,
as well as a label $\tilde{\bm{r}}\in\tilde{\Lambda}$
to specify the unit cell and a label
$J=A,B,C,D$ to specify a wire within a unit cell.
We then write down the many-body ``backscattering''
current-current interactions
encoded by the Lagrangian density
\begin{subequations}
\label{eq: su(2) level k weak topo order interactions}
\begin{widetext}
\begin{align}
\widehat{\mathcal{L}}^{\,}_{\mathrm{bs}}[su(2)^{\,}_{k}]\equiv
-\widehat{\mathcal{H}}^{\,}_{\mathrm{bs}}[su(2)^{\,}_{k}]\:=&
-
\lambda^{\,}_{su(2)^{\,}_{k}}
\sum_{\tilde{\bm{r}}\in\tilde{\Lambda}}
\Big(
\widehat{\mathcal{L}}^{\,}_{2,\mathrm{R},A,\tilde{\bm{r}}|1,\mathrm{L},D}
+
\widehat{\mathcal{L}}^{\prime}_{2,\mathrm{L},D,\tilde{\bm{r}}|1,\mathrm{R},A}
+
\widehat{\mathcal{L}}^{\,}_{2,\mathrm{L},B,\tilde{\bm{r}}|1,\mathrm{R},C}
+
\widehat{\mathcal{L}}^{\prime}_{2,\mathrm{R},C,\tilde{\bm{r}}|1,\mathrm{L},B}
\nonumber
\\
&\,
+
\widehat{\mathcal{L}}^{\,}_{2,\mathrm{L},A,\tilde{\bm{r}}|1,\mathrm{R},B}
+
\widehat{\mathcal{L}}^{\prime\prime}_{2,\mathrm{R},B,\tilde{\bm{r}}|1,\mathrm{L},A}
+
\widehat{\mathcal{L}}^{\,}_{2,\mathrm{R},D,\tilde{\bm{r}}|1,\mathrm{L},C}
+
\widehat{\mathcal{L}}^{\prime\prime}_{2,\mathrm{L},C,\tilde{\bm{r}}|1,\mathrm{R},D}
\Big),
\label{eq: su(2) level k inter-wire gapping}
\end{align}
\end{widetext}
where we have assigned to each of the eight nearest-neighbor
bonds shown in Fig.\ \ref{fig: Schematic of the su(2)_{k} couplings}
the bond operators
\begin{align}
\widehat{\mathcal{L}}^{\,}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}|\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime}}
&\:=
\sum^{2}_{a=1}
\widehat{J}^{a}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}\,
\widehat{J}^{a}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},\tilde{\bm{r}}},
\\
\widehat{\mathcal{L}}^{\prime}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}|\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime}}
&\:=
\sum^{2}_{a=1}
\widehat{J}^{a}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}\,
\widehat{J}^{a}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},\tilde{\bm{r}}+2\hat{\bm{x}}},
\\
\widehat{\mathcal{L}}^{\prime\prime}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}|\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime}}
&\:=
\sum^{2}_{a=1}
\widehat{J}^{a}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}\,
\widehat{J}^{a}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},\tilde{\bm{r}}+2\hat{\bm{y}}},
\label{eq: su(2) level k local bond terms}
\end{align}
and where the lattice vectors $2\hat{\bm{x}}$ and $2\hat{\bm{y}}$
connect neighboring unit cells along the $x$- and $y$-directions,
respectively.
\end{subequations}
The interactions described by these bond operators have the same form
as those appearing in Eq.~\eqref{eq2Dcase: 2D interaction}, which were used
in the 2D case. The answer to the question of whether or not the
interactions~\eqref{eq: su(2) level k inter-wire gapping}
yield a fully gapped spectrum is thus the same as in 2D. In particular, the
argument for the existence of a gap in the case $k=2$ is identical.
The current-current interactions \eqref{eq: su(2) level k inter-wire gapping}
possess an antiunitary involutive symmetry that
effectively plays the same role as that of time-reversal symmetry
for TIs. We call this symmetry an antiferromagnetic
time-reversal symmetry.
(This symmetry will also appear in
Sec.%
~\ref{sec: Possible strong non-Abelian topological order in three dimensions}.)
To see this, note that
Fig.\ \ref{fig: Schematic of the su(2)_{k} couplings}
is invariant under interchanging
the colors red and blue, and then translating
by either of the \textit{half}-lattice vectors
$\hat{\bm{x}}$ or $\hat{\bm{y}}$.
Formally, we define the symmetry operations
\begin{subequations}
\label{eq: T_eff definitions}
\begin{align}
\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}}}&\:=
\mathcal{T}\times \mathsf{T}^{\,}_{\hat{\bm{x}}},
\label{eq: T_eff definitions a}
\\
\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}&\:=
\mathcal{T}\times \mathsf{T}^{\,}_{\hat{\bm{y}}},
\label{eq: T_eff definitions b}
\end{align}
where the time-reversal operation $\mathcal{T}$ acts in the
usual way on the spinful fermions (i.e., $\mathcal{T}^{2}=-1$),
and acts on the $su(2)^{\,}_{k}$ currents
[see Eqs.~\eqref{eq: non-Abelian currents}] as
\begin{align}
\mathcal{T}\,\widehat{J}^{a}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}\mathcal{T}^{-1}=
-\widehat{J}^{a}_{\gamma,\overline{\mathrm{M}},J,\tilde{\bm{r}}},
\end{align}
for any $a=1,2,3$, $\gamma=1,2$, $\mathrm{M}=\mathrm{L},\mathrm{R}$,
$J=A,B,C,D$, and $\tilde{\bm{r}}\in\tilde{\Lambda}$,
with $\overline{\mathrm{L}}\:=\mathrm{R}$ and
$\overline{\mathrm{R}}\:=\mathrm{L}$,
and where the half-lattice translation operators $\mathsf{T}^{\,}_{\hat{\bm{x}}}$
and $\mathsf{T}^{\,}_{\hat{\bm{y}}}$ act as
\begin{align}
\mathsf{T}^{\,}_{\hat{\bm{x}}}
\widehat{J}^{\,}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}
\mathsf{T}^{-1}_{\hat{\bm{x}}}
&=
\begin{cases}
\widehat{J}^{\,}_{\gamma,\mathrm{M},B,\tilde{\bm{r}}}, & J=A,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},A,\tilde{\bm{r}}+2\hat{\bm{x}}}, & J=B,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},D,\tilde{\bm{r}}+2\hat{\bm{x}}}, & J=C,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},C,\tilde{\bm{r}}}, & J=D,
\end{cases}
\end{align}
and
\begin{align}
\mathsf{T}^{\,}_{\hat{\bm{y}}}
\widehat{J}^{\,}_{\gamma,\mathrm{M},J,\tilde{\bm{r}}}
\mathsf{T}^{-1}_{\hat{\bm{y}}}
&=
\begin{cases}
\widehat{J}^{\,}_{\gamma,\mathrm{M},D,\tilde{\bm{r}}}, & J=A,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},C,\tilde{\bm{r}}}, & J=B,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},B,\tilde{\bm{r}}+2\hat{\bm{y}}}, & J=C,\\
\widehat{J}^{\,}_{\gamma,\mathrm{M},A,\tilde{\bm{r}}+2\hat{\bm{y}}}, & J=D,
\end{cases}
\end{align}
respectively.
\end{subequations}
The full Lagrangian in the presence of the current-current interactions
\eqref{eq: su(2) level k inter-wire gapping} is invariant under the
symmetry operations $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}}}$
and $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ when periodic
boundary conditions are imposed in the $x$- and $y$-directions.
If we use mixed boundary conditions, such as ones that are
open along the $x$-direction and periodic along the $y$-direction
(or vice versa),
then only the symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$
(or $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}}}$)
remains. In Sec.~\ref{sec: Surface theory},
we will see that the remaining symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ protects gapless surface
states on the boundaries at $x=0$ and $x=L^{\,}_{x}$. Analogous
non-onsite implementations of time reversal symmetry
have arisen in studies of antiferromagnetic TIs~\cite{Mong10,Fang13}
and in coupled-wire models of topological-insulator and
topological-superconductor surfaces~\cite{Mross15,Sahoo15}.
\begin{figure}[t]
\centering
\includegraphics[width=0.385\textwidth]{Figures8-5}
\caption{
(Color online)
Schematic grouping of the interactions
\eqref{eq: su(2) level k inter-wire gapping}
into a sum over decoupled planes. As in
Fig.\ \ref{fig: Schematic of the su(2)_{k} couplings},
purple bonds denote current-current interactions of the form
\eqref{eq: su(2) level k local bond terms}.
Regions between any pair of adjacent dashed black lines constitute a plane.
The shaded and unshaded planes are related by the symmetries
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}},\hat{\bm{y}}}$ defined in
Eqs.\ \eqref{eq: T_eff definitions}
when periodic boundary conditions
are imposed in the $\hat{\bm{x}}$- and $\hat{\bm{y}}$-directions.
}
\label{fig: weak phase}
\end{figure}
One can interpret the coupled-wire array with the
interaction~\eqref{eq: su(2) level k inter-wire gapping} as a stack of
2D topological phases like the ones defined and studied in
Sec.~\eqref{sec: Warm-up: Non-Abelian topological order in two
dimensions}. To see this, one need only group the terms in the sum
in Eq.~\eqref{eq: su(2) level k inter-wire gapping} in an appropriate
way, as indicated in Fig.\ \ref{fig: weak phase}. Then, one sees that
Eq.~\eqref{eq: su(2) level k inter-wire gapping} breaks up into a sum
over decoupled planes, each consisting of a set of wires coupled by
interactions of the form \eqref{eq2Dcase: 2D interaction}. Each plane
thus forms its own 2D non-Abelian topological phase.
More specifically, any ground-state wavefunction of the coupled-wire
array can be written in the form
\begin{align}
\label{eq: decoupled layer wavefunction}
\ket{\Psi^{\,}_{\mathrm{gs}}}\:=\bigotimes_{i}\ket{\Psi^{i}_{\mathrm{gs}}},
\end{align}
where the index $i$ runs over the layers in the stack. Here,
$\ket{\Psi^{i}_{\mathrm{gs}}}$ is a ground-state wavefunction of the 2D
topological phase supported by layer $i$ of the stack
depicted in Fig.\ \ref{fig: weak phase}. Since each
layer carries with it a threefold topological degeneracy when periodic
boundary conditions are imposed in all directions, the full 3D
system then carries a weak topological degeneracy of
\begin{align}
3^{\ \#\, \text{of layers}}.
\end{align}
This subextensive degeneracy reflects the direct-product structure of
the ground-state wavefunction \eqref{eq: decoupled layer wavefunction}.
If one uses this wavefunction to calculate the entanglement entropy between
adjacent layers in the stack by partitioning the stack along a plane
parallel to both layers, one inevitably obtains zero.
For any pair $i$ and $j$ of shaded layers from the stack depicted in Fig.\
\ref{fig: weak phase}, the ground state wavefunctions
$\ket{\Psi^{i}_{\mathrm{gs}}}$
and
$\ket{\Psi^{j}_{\mathrm{gs}}}$
are not correlated.
The result is a topological ground-state degeneracy that scales with the
number of layers.
When periodic boundary conditions are imposed, each plane of coupled wires
defined in this way is related to its neighbors by the symmetries
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}},\hat{\bm{y}}}$ defined in
Eq.~\eqref{eq: T_eff definitions}.
Thus, one can view this 3D stack of 2D non-Abelian topological phases as a
weak topological phase respecting the symmetries
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}},\hat{\bm{y}}}$.
\section{Strong non-Abelian topological order in three dimensions}
\label{sec: Possible strong non-Abelian topological order in three dimensions}
The interaction \eqref{eq: su(2) level k inter-wire gapping}
depicted in Fig.~\ref{fig: Schematic of the su(2)_{k} couplings}
can be decomposed into a sum
of interactions between wires living in decoupled planes. We now propose an alternative
interaction that avoids this problem, thereby realizing a truly 3D non-Abelian
topological phase.
The proposed interaction preserves the structure of the individual wires
in the decoupled limit, as described in
Sec.~\ref{subsec: Definition of the class of models 3D} and depicted in
Fig.~\ref{fig: general wire setup}. It is given by
\begin{widetext}
\begin{subequations}
\label{eq: plaquette interaction}
\begin{align}
\widehat{\mathcal{H}}^{\,}_{\square}\:=
-\lambda
\(
\sum_{p^{\,}_{A}\in\Lambda}
\widehat{\mathcal{H}}_{p^{\,}_{A}}
+
\sum_{p^{\,}_{B}\in\Lambda}
\widehat{\mathcal{H}}_{p^{\,}_{B}}
+
\sum_{p^{\,}_{C}\in\Lambda}
\widehat{\mathcal{H}}_{p^{\,}_{C}}
+
\sum_{p^{\,}_{D}\in\Lambda}
\widehat{\mathcal{H}}_{p^{\,}_{D}}
\),
\label{eq: plaquette interaction a}
\end{align}
where
\begin{align}
&
\widehat{\mathcal{H}}_{p^{\,}_{A}}
\:=
\sum^{2}_{a,b,c,d=1}
\(\widehat{J}^{a}_{2,\mathrm{L},A,p^{\,}_{A}}\, \widehat{J}^{a}_{1,\mathrm{R},B,p^{\,}_{A}}\)
\(\widehat{J}^{b}_{2,\mathrm{L},B,p^{\,}_{A}}\, \widehat{J}^{b}_{1,\mathrm{R},C,p^{\,}_{A}}\)
\(\widehat{J}^{c}_{2,\mathrm{R},D,p^{\,}_{A}}\, \widehat{J}^{c}_{1,\mathrm{L},C,p^{\,}_{A}}\)
\(\widehat{J}^{d}_{2,\mathrm{R},A,p^{\,}_{A}}\, \widehat{J}^{d}_{1,\mathrm{L},D,p^{\,}_{A}}\),
\\
&
\widehat{\mathcal{H}}_{p^{\,}_{B}}
\:=
\sum^{2}_{a,b,c,d=1}
\!
\(\widehat{J}^{a}_{2,\mathrm{R},B,p^{\,}_{B}}\, \widehat{J}^{a}_{1,\mathrm{L},A,p^{\,}_{B}}\)
\(\widehat{J}^{b}_{2,\mathrm{R},A,p^{\,}_{B}}\, \widehat{J}^{b}_{1,\mathrm{L},D,p^{\,}_{B}}\)
\(\widehat{J}^{c}_{2,\mathrm{L},C,p^{\,}_{B}}\, \widehat{J}^{c}_{1,\mathrm{R},D,p^{\,}_{B}}\)
\(\widehat{J}^{d}_{2,\mathrm{L},B,p^{\,}_{B}}\, \widehat{J}^{d}_{1,\mathrm{R},C,p^{\,}_{B}}\),
\\
&
\widehat{\mathcal{H}}_{p^{\,}_{C}}
\:=
\sum^{2}_{a,b,c,d=1}
\!
\(\widehat{J}^{a}_{2,\mathrm{L},C,p^{\,}_{C}}\, \widehat{J}^{a}_{1,\mathrm{R},D,p^{\,}_{C}}\)
\(\widehat{J}^{b}_{2,\mathrm{L},D,p^{\,}_{C}}\, \widehat{J}^{b}_{1,\mathrm{R},A,p^{\,}_{C}}\)
\(\widehat{J}^{c}_{2,\mathrm{R},B,p^{\,}_{C}}\, \widehat{J}^{c}_{1,\mathrm{L},A,p^{\,}_{C}}\)
\(\widehat{J}^{d}_{2,\mathrm{R},C,p^{\,}_{C}}\, \widehat{J}^{d}_{1,\mathrm{L},B,p^{\,}_{C}}\),
\\
&
\widehat{\mathcal{H}}_{p^{\,}_{D}}
\:=
\sum^{2}_{a,b,c,d=1}
\!
\(\widehat{J}^{a}_{2,\mathrm{R},D,p^{\,}_{D}}\, \widehat{J}^{a}_{1,\mathrm{L},C,p^{\,}_{D}}\)
\(\widehat{J}^{b}_{2,\mathrm{R},C,p^{\,}_{D}}\, \widehat{J}^{b}_{1,\mathrm{L},B,p^{\,}_{D}}\)
\(\widehat{J}^{c}_{2,\mathrm{L},A,p^{\,}_{D}}\, \widehat{J}^{c}_{1,\mathrm{R},B,p^{\,}_{D}}\)
\(\widehat{J}^{d}_{2,\mathrm{L},D,p^{\,}_{D}}\, \widehat{J}^{d}_{1,\mathrm{R},A,p^{\,}_{D}}\).
\end{align}
\end{subequations}
\end{widetext}
Here, the labels $1,2$, $\mathrm{L},\mathrm{R}$, and $A,\cdots,D$ are
as in Fig.~\ref{fig: Schematic of the su(2)_{k} couplings}.
Moreover, we have defined a new label $p^{\,}_{J}$, $J=A,\cdots,D$ as follows.
The label $p^{\,}_{A}$ denotes a four-wire plaquette in the lattice
$\Lambda$ in which the four labels $J=A,B,C,D$ cycle through $A,B,C,$
and $D$ clockwise from the top left (this is the boxed plaquette in
Fig.~\ref{fig: general wire setup}). The remaining plaquette labels
are defined analogously, with $p^{\,}_{B}$ labeling a plaquette whose
top-left wire carries the label $B$ and so on.
The plaquette interaction \eqref{eq: plaquette interaction} avoids the
problem of the current-current interaction
\eqref{eq: su(2) level k inter-wire gapping}
in that it is not possible to rewrite the former
as a sum over decoupled planes. The crucial difference between the two
interactions that allows this to happen is the fact that the
interaction \eqref{eq: plaquette interaction} is a sum of
\textit{products} of current-current bilinears, whereas the
interaction \eqref{eq: su(2) level k inter-wire gapping} is merely a
sum of such bilinears.
In the remainder of this section, we will analyze the consequences of
the plaquette interaction \eqref{eq: plaquette interaction} in the
strong-coupling limit $\lambda\to\infty$ when periodic boundary
conditions are imposed in all directions. We will focus our attention
on the case $k=2$, although generalizations to $k>2$ along the lines
of Sec.\
\eqref{subsec: Parafermion representation of the interwire interactions 2D}
are possible. First, we will show that the
interaction \eqref{eq: plaquette interaction} opens a gap at strong
coupling. Second,
we will define a set of string and membrane operators
associated with excitations of the gapped theory, and argue on the
basis of their algebra that the ground state of the gapped system is
topologically degenerate. The algebra obeyed by
these operators will reveal the non-Abelian character of the phase.
\subsection{Analyzing the opening of a gap when $k=2$}
\label{subsec: Parafermion representation of the interwire interactions 3D}
To see how the plaquette interaction \eqref{eq: plaquette interaction}
opens a gap at strong coupling when $k=2$, it is
useful to rewrite it in terms of the identities
\eqref{eq: def chiral parafermion and boson rep su(2)k}.
In particular, we decompose the $su(2)^{\,}_{2}$ currents as~\cite{DiFrancesco97}
\begin{subequations}
\label{eq: def bulk chiral parafermion and boson rep su(2)k}
\begin{align}
&
\widehat{J}^{+}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\=:
\sqrt{2}\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\,
\bm{:}
e^{+\mathrm{i}\sqrt{1/2}\,\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}}
\bm{:},
\label{eq: def bulk chiral parafermion and boson rep su(2)k a}
\\
&
\widehat{J}^{-}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\=:
\sqrt{2}\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\,
\bm{:}
e^{-\mathrm{i}\sqrt{1/2}\,\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}}
\bm{:},
\label{eq: def bulk chiral parafermion and boson rep su(2)k aa}
\\
&
\widehat{J}^{3}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\=:
\mathrm{i}
\frac{1}{\sqrt{2}}\,
\partial^{\,}_{\mathrm{M}}\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}},
\label{eq: def bulk chiral parafermion and boson rep su(2)k aaa}
\end{align}
\end{subequations}
where
$\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}$
is a chiral Majorana fermion and
$\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}$
is a chiral bosonic field,
for any $\gamma=1,2$, $\mathrm{M}=\mathrm{L},\mathrm{R}$,
and $J,K=A,B,C,D$.
On the Majorana operators, we impose the equal-time algebra
\vspace{2cm}
\begin{widetext}
\begin{subequations}
\label{eq: def equal time algebra three dim bul parafermions}
\begin{align}
\label{eq: def bulk chiral parafermion and boson rep su(2)k b}
\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z)\,
\widehat{\psi}^{\,}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},p^{\prime}_{K^{\prime}}}
(t,z^{\prime})=&\,
\widehat{\psi}^{\,}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},p^{\prime}_{K^{\prime}}}
(t,z^{\prime})\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z)
\nonumber\\
&\,
\times
e^{
-\mathrm{i}\,\pi\,\delta^{\,}_{\bm{r},\bm{r}^{\prime}}
\[
(-1)^{\mathrm{M}}\,
\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,
\delta^{\,}_{\gamma,\gamma^{\prime}}\,
\mathrm{sgn}(z-z^{\prime})
\]
}
\\
&\,
\times
e^{
+\mathrm{i}\,\pi
\left\{
\(1-\delta^{\,}_{\mathrm{M},\mathrm{M}^\prime}\,\delta^{\,}_{\gamma,\gamma^\prime}\,\delta^{\,}_{\bm{r},\bm{r}^\prime}\)
\[
(-1)^{\mathrm{M}}\, \delta^{\,}_{\mathrm{M},\mathrm{M}^\prime}
+
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^\prime}
\]
\right\}
}
\nonumber.
\end{align}
This equal-time algebra is a generalization of Eqs.\
\eqref{eq: def chiral parafermion and boson rep su(2)k c}--%
\eqref{eq: def chiral parafermion and boson rep su(2)k ccc}.
As before, we use the conventions $\ensuremath \text{sgn}(0)=0$
and $\epsilon^{\,}_{\mathrm{L},\mathrm{R}}=-\epsilon^{\,}_{\mathrm{R},\mathrm{L}}=-1$.
Note that, on the one hand, the Majorana operator
$\widehat{\psi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}$
is labeled by the tuple
$(J,p^{\,}_{K})\in\tilde{\Lambda}$,
which is the square lattice with a four-wire unit cell depicted in
Fig.\ \ref{fig: Schematic of the su(2)_{k} couplings}.
On the other hand, the coordinates $\{\bm{r}\:=(x,y)\in\Lambda\}$,
which label
the square lattice with a single-wire unit cell,
enter the phase factors in
Eqs.\
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k b}.
This notation is consistent because the tuple
$(J,p^{\,}_{K})$, $J,K=A,B,C,D$, uniquely specifies an $\bm{r}\in\Lambda$.
Finally, we impose the equal-time algebra
\begin{align}
\[
\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z),
\widehat{\phi}^{\,}_{\gamma^{\prime},\mathrm{M}^{\prime},J^{\prime},p^{\prime}_{K^{\prime}}}
(t,z^{\prime})
\]=&\,
-\mathrm{i}\,
2\pi
\Big\{
(-1)^{\mathrm{M}}\,
\delta^{\,}_{\mathrm{M},\mathrm{M}^{\prime}}\,
\delta^{\,}_{\gamma,\gamma^{\prime}}\,
\delta^{\,}_{\bm{r},\bm{r}^{\prime}}\,
\mathrm{sgn}(z-z^{\prime})
\nonumber
\\
&\qquad\qquad
-
\(1-\delta^{\,}_{\mathrm{M},\mathrm{M}^\prime}\,\delta^{\,}_{\gamma,\gamma^\prime}\,\delta^{\,}_{\bm{r},\bm{r}^\prime}\)
\[
(-1)^{\mathrm{M}}\, \delta^{\,}_{\mathrm{M},\mathrm{M}^\prime}
+
\epsilon^{\,}_{\mathrm{M},\mathrm{M}^\prime}
\]
\Big\},
\label{eq: def bulk chiral parafermion and boson rep su(2)k c}
\end{align}
\end{subequations}
on the chiral bosons. This algebra is a generalization of
Eq.~\eqref{eq: def chiral parafermion and boson rep su(2)k d}.
The interaction \eqref{eq: plaquette interaction} is amenable to analysis upon
substituting in the decompositions \eqref{eq: def bulk chiral parafermion and boson rep su(2)k}.
For instance, the first term in Eq.~\eqref{eq: plaquette interaction a} becomes
\begin{align}
\label{eq: bulk interactions su(2)_k}
\widehat{\mathcal{H}}_{p^{\,}_{A}}
\!
&=
16\
\widehat{\psi}^{\,}_{2,\mathrm{L},A,p^{\,}_{A}}
\widehat{\psi}^{\,}_{1,\mathrm{R},B,p^{\,}_{A}}
\widehat{\psi}^{\,}_{2,\mathrm{L},B,p^{\,}_{A}}
\widehat{\psi}^{\,}_{1,\mathrm{R},C,p^{\,}_{A}}
\widehat{\psi}^{\,}_{1,\mathrm{L},C,p^{\,}_{A}}
\widehat{\psi}^{\,}_{2,\mathrm{R},D,p^{\,}_{A}}
\widehat{\psi}^{\,}_{1,\mathrm{L},D,p^{\,}_{A}}
\widehat{\psi}^{\,}_{2,\mathrm{R},A,p^{\,}_{A}}
\\
&
\quad
\times
\sin\!\(\frac{\widehat{\phi}^{\,}_{2,\mathrm{L},A,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{1,\mathrm{R},B,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{2,\mathrm{L},B,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{1,\mathrm{R},C,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{1,\mathrm{L},C,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{2,\mathrm{R},D,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{1,\mathrm{L},D,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{2,\mathrm{R},A,p^{\,}_{A}}}{\sqrt{2}}\!\)\!.
\nonumber
\end{align}
\end{widetext}
Similar expressions hold for the remaining three plaquette terms
in the summand in Eq.~\eqref{eq: plaquette interaction a}.
Each such plaquette term factorizes into a $u(1)^{\,}_{2}$ piece (the
product of four sine potentials) and a $\mathbb{Z}^{\,}_{2}$ piece (the
eight-Majorana interaction). By analogy with Eq.%
~\eqref{eq2Dcase: su(2)_{2} backscattering}
in the 2D case, we would like to argue
that the $u(1)^{\,}_{2}$ and $\mathbb{Z}^{\,}_{2}$ sectors can
simultaneously acquire expectation values and become gapped. To see
that this is indeed possible, it is useful to consider the
$u(1)^{\,}_{2}$ and $\mathbb{Z}^{\,}_{2}$ sectors separately.
First, one can show by direct calculation that the four eight-Majorana
plaquette terms appearing in Eq.~\eqref{eq: plaquette interaction a}
commute with one another. Hence, it is possible for each
eight-Majorana plaquette term to acquire an expectation value such
that the total energy in the Majorana sector is minimized. When this
occurs, we can replace the product of eight Majorana operators that
enters each plaquette term with a constant. What remains is an
effective interaction for the $u(1)^{\,}_{2}$ sector consisting of
four types of plaquette terms, each a product of four sines.
Next, we address the question of whether this effective interaction is
capable of opening a gap in the $u(1)^{\,}_{2}$ sector. To do this,
we successively apply the trigonometric identities
\begin{align}
\begin{split}
&
2\,\sin a\, \sin b = \cos(a-b)-\cos(a+b),
\\
&
2\,\cos a\, \cos b = \cos(a-b)+\cos(a+b),
\end{split}
\end{align}
until each product of four sines becomes a sum of \textit{eight}
cosines. For instance, applying this procedure to the $u(1)^{\,}_{2}$
part of the plaquette term \eqref{eq: bulk interactions su(2)_k}
yields
\begin{widetext}
\begin{subequations}
\label{eq: cosine representation}
\begin{align}
\begin{split}
&
\sin\!\(\frac{\widehat{\phi}^{\,}_{2,\mathrm{L},A,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{1,\mathrm{R},B,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{2,\mathrm{L},B,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{1,\mathrm{R},C,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{1,\mathrm{L},C,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{2,\mathrm{R},D,p^{\,}_{A}}}{\sqrt{2}}\!\)
\sin\!\(\frac{\widehat{\phi}^{\,}_{1,\mathrm{L},D,p^{\,}_{A}}\!\!-\widehat{\phi}^{\,}_{2,\mathrm{R},A,p^{\,}_{A}}}{\sqrt{2}}\!\)\\
&\qquad\qquad\qquad\qquad
=\frac{1}{8}\sum^{8}_{i=1}s^{\,}_{i}\, \cos\(\frac{1}{\sqrt{2}}\,\mathcal{T}^{\mathsf{T}}_{i}\mathcal{K}\,\widehat{\varphi}_{p^{\,}_{A}}\),
\end{split}
\end{align}
where $s^{\,}_{i}=\pm1$ is an $i$-dependent sign whose value is unimportant for the purposes of this argument.
Here, we have defined the eight-component vector
\begin{align}
\widehat{\varphi}^{\,}_{p^{\,}_{A}}
\:=
\begin{pmatrix}
\widehat{\phi}^{\,}_{2,\mathrm{L},A,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{2,\mathrm{R},A,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{2,\mathrm{L},B,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{1,\mathrm{R},B,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{1,\mathrm{L},C,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{1,\mathrm{R},C,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{1,\mathrm{L},D,p^{\,}_{A}}
&
\widehat{\phi}^{\,}_{2,\mathrm{R},D,p^{\,}_{A}}
\end{pmatrix}^\mathsf{T}
\end{align}
and the $8\times8$ diagonal matrix
\begin{align}
\mathcal{K}\:=
\mathrm{diag}(+1,-1,+1,-1,+1,-1,+1,-1).
\end{align}
The eight-component vectors $\mathcal{T}^{\,}_{i}$, $i=1,\cdots,8$ with entries $\pm1$ specify the
linear combinations of bosonic fields that enter each cosine term.
\end{subequations}
\end{widetext}
One can verify using the algebra \eqref{eq: def bulk chiral parafermion and boson rep su(2)k c}
that the cosine terms appearing in
Eq.~\eqref{eq: cosine representation} do not commute with one another. One might worry that this
implies that these eight cosines cannot be minimized simultaneously, so that the $u(1)^{\,}_{2}$
sector is not gapped, even when the $\mathbb{Z}^{\,}_{2}$ sector is. However, Haldane showed~\cite{Haldane95}
that a weaker condition than strict commutation is sufficient for a set of cosine potentials to be
simultaneously minimizable. For the cosine terms in Eq.~\eqref{eq: cosine representation}, this
condition reads
\begin{align}
\label{eq: Haldane criterion}
\mathcal{T}^{\mathsf{T}}_{i}\mathcal{K}\,\mathcal{T}^{\,}_{j}=0,
\qquad
i,j=1,\cdots,8.
\end{align}
Once the eight vectors $\mathcal{T}^{\,}_{i}$ have been determined, it
is straightforward to check that the above condition indeed holds for
the interaction \eqref{eq: cosine representation}. Thus, the eight
linear combinations
$\mathcal{T}^{\mathsf{T}}_{i}\,\mathcal{K}\,\widehat{\varphi}^{\,}_{p^{\,}_{A}}$
can simultaneously assume classical
configurations that minimize the expectation value of the effective
interaction \eqref{eq: cosine representation}. One can repeat this
exercise for the three plaquettes
$p^{\,}_{B}$,
$p^{\,}_{C}$,
and
$p^{\,}_{D}$
and verify that they,
too, can each assume classical configurations that minimize the energy
in the $u(1)^{\,}_{2}$ sector. Finally, one checks that a version of
the criterion \eqref{eq: Haldane criterion} holds for bosonic fields
that are shared between adjacent plaquettes, ensuring that the
adjacent plaquette terms do not interfere with one another in this
minimization process. (The procedure defined in this paragraph is
essentially the one used in coupled-wire constructions of gapped
Abelian topological phases in 2D~\cite{Neupert14} and
3D~\cite{Iadecola16}.)
Thus, in this section we have argued that,
despite being more
complicated than the ``naive'' interaction
\eqref{eq: su(2) level k inter-wire gapping},
the plaquette interaction \eqref{eq: plaquette interaction}
opens a gap in the array of coupled wires in the limit
$\lambda\to\infty$ when $k=2$.
Moreover, because this interaction cannot be
rewritten as a sum of terms living in decoupled planes, it does not
manifestly result in a weak topological phase, unlike the interaction
\eqref{eq: su(2) level k inter-wire gapping}. In fact, we conjecture
that the phase realized by the plaquette interaction
\eqref{eq: plaquette interaction}
is a truly 3D non-Abelian topological phase.
We provide evidence in support of this conjecture in the next section.
One heuristic argument in favor of this conjecture is the following.
If one ignores the $z$-direction of space, the
plaquette interaction \eqref{eq: plaquette interaction} resembles the
eight-Majorana interaction of the Wen-Plaquette model~\cite{Wen03},
which realizes $\mathbb Z^{\,}_{2}$ topological order in 2D.
While the Wen-Plaquette model does not contain anything resembling the
$u(1)^{\,}_{2}$ part of Eq.\ \eqref{eq: plaquette interaction},
this sector is necessary in the context of the wire construction
in order to obtain a local interaction.
Nevertheless, the similarity between the $\mathbb{Z}^{\,}_{2}$ part of
the interaction~\eqref{eq: plaquette interaction}
and the Wen-Plaquette model suggests that one should be
able to construct string operators in the $x$-$y$ plane that are in one-to-one
correspondence with deconfined pointlike excitations.
(This reasoning is similar to the reasoning that guided our work in
Ref.~\cite{Iadecola16}, which constructed Abelian 3D topological phases
from coupled wires with interactions meant to resemble those of the toric code.)
Sec.\ \ref{subsec: String and membrane operators}
is devoted to the construction of these operators,
and others associated with the $z$-direction,
as well as membrane operators corresponding to stringlike excitations.
Of course, the fact that the Majorana operators in the interaction
\eqref{eq: plaquette interaction} arise from a conformal
field theory, and not from a representation of the Pauli algebra as in
the Wen-Plaquette model, implies that the topological order (if any)
obtained from the interaction \eqref{eq: plaquette interaction} must
be richer than ``simple" $\mathbb Z^{\,}_{2}$ topological order. In
particular, the conformal field theories that furnish the Majorana
operators also furnish more complicated non-Abelian operators like the
Ising twist field, from which the coupled-wire model inherits its
non-Abelian character.
Finally, we note that applying the above reasoning to the
interaction~\eqref{eq: su(2) level k weak topo order interactions} in
Sec.~\ref{subsec: Interwire couplings for weak non-Abelian topological order}
does not allow one to draw comparisons with a truly 2D lattice model.
Instead, the $\mathbb Z^{\,}_{2}$ sector of the
interaction~\eqref{eq: su(2) level k weak topo order interactions}
resembles an array of decoupled 1D Kitaev chains~\cite{Kitaev01}.
This is yet another way to see that the interaction
\eqref{eq: su(2) level k weak topo order interactions}
cannot yield strong topological order.
\subsection{String and membrane operators when $k=2$}
\label{subsec: String and membrane operators}
In this section we build string and membrane operators to characterize
the different topological sectors of the coupled-wire theory. This
discussion parallels the analysis of the 2D case presented in
Sec.~\ref{subsec: Case study: su(2)_2}, albeit with complications
owing to the increase in dimensionality.
\subsubsection{Primary operators in a single wire}
\label{subsubsec: Primary operators in a single wire}
Since each wire consists of four independent chiral $su(2)^{\,}_{2}$
CFTs in the decoupled limit (see, e.g., Fig.\
\ref{fig: Schematic of the su(2)_{k} couplings}),
there are a number of chiral primary
operators in each wire. In particular, there is a ``spin-$0$'', a
``spin-$\frac{1}{2}$'', and a ``spin-$1$'' primary for each set of
labels $(\gamma,\mathrm{M},J,p^{\,}_{K})$, with $\gamma=1,2$,
$\mathrm{M}=\mathrm{L},\mathrm{R}$, and $J,K=A,B,C,D$. Fixing some
$J$, $K$, and plaquette $p^{\,}_{K}$ specifies a single wire, and the
remaining labels $\gamma$ and $\mathrm{M}$ enumerate the four chiral
CFTs defined within that wire. Each chiral sector of the wire is equipped
with a ``spin-$0$'' primary operator, which is simply the identity operator in that
sector. The nontrivial primary operators in each
CFT are the ``spin-$\frac{1}{2}$'' primary,
\begin{align}
\label{eq: spin-1/2 primary 3D}
\widehat{\Phi}^{(\frac{1}{2})}_{\gamma,\mathrm{M},J,p^{\,}_{K}}
\!
(t,z)
\:=
\widehat{\sigma}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}
\!
(t,z)
\bm{:}
\!
e^{+\mathrm{i}\frac{1}{2\sqrt{2}}\,\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\!(t,z)}
\!
\bm{:}\, ,
\end{align}
and the ``spin-$1$'' primary
\begin{align}
\label{eq: spin-1 primary 3D}
\widehat{\Phi}^{(1)}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\!(t,z)
\:=
\bm{:}
\!
e^{+\mathrm{i}\frac{1}{\sqrt{2}}\,\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}\!(t,z)}
\!
\bm{:}\,
\end{align}
[compare with Eqs.\ \eqref{eq: spin-1/2 primary} and
\eqref{eq: spin-1 primary} in
Sec.\ \ref{subsec: twist operators}].
The algebraic properties of these primary fields, including their OPEs
and exchange algebras, are summarized in Sec.~\ref{subsec: twist
operators}. The key information to retain from that discussion is
that all non-Abelian properties of these primary operators stem from
the presence of the Ising twist operator
$\widehat{\sigma}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}$ in the
definition of the ``spin-$\frac{1}{2}$'' primary. The vertex operators
from the $u(1)^{\,}_{2}$ sector have Abelian fusion rules.
As in Sec.~\ref{subsec: Case study: su(2)_2}, our strategy for
building string operators will be to find appropriate nonchiral
products of primary fields from which to build nonlocal string and
membrane operators that characterize the topological order. We proceed
with this program below.
\begin{figure}[tb]
\centering
\flushleft(a)
\includegraphics[width=0.45\textwidth]{Figures9}\\
\flushleft(b)
\hspace{.5cm}
\includegraphics[width=0.35\textwidth]{Figures10}
\caption{
(Color online) Depictions of the paths $\mathcal{P}^{\,}_{\hat{x}}$
and $\mathcal{P}^{\,}_{\hat{y}}$. The grey circles represent wires, as in
Fig.\ \ref{fig: general wire setup},
and each white oval contains a pair of chiral modes L and R labeled
by the indices $\gamma$, $J$, and $p^{\,}_{K}$.
In panel (a), the paths $\mathcal{P}^{\,}_{\hat{x}}$
and $\mathcal{P}^{\,}_{\hat{y}}$ are
depicted at different values of the coordinate $z$,
while in panel (b), the projection of the two paths into the $x$-$y$
plane is shown.
}
\label{fig: P_x and P_y defs}
\end{figure}
\begin{figure}[tb]
\centering
(a)
\includegraphics[width=0.35\textwidth]{Figures10-25a}\\
\vspace{.5cm}
(b)
\includegraphics[width=0.35\textwidth]{Figures10-25b}
\caption{(Color online)
Pairs of defective plaquettes as pointlike excitations.
Panel (a) is a pictorial representation of the action of an operator like
\eqref{eq: O (1) def} on the interaction
\eqref{eq: plaquette interaction}.
The application of such an operator creates solitons
in linear combinations of scalar fields,
defined on the two purple links.
These solitons enter three plaquette terms in the interaction
\eqref{eq: plaquette interaction}. The two diagonally-opposite
plaquettes (each marked with a dashed box) each contain one soliton,
while the plaquette between the two purple links (marked with a pair
of concentric dashed boxes) contains two solitons. Panel (b)
demonstrates that acting with additional operators like \eqref{eq: O
(1) def} in neighboring wires allows one to separate \textit{pairs}
of these defective plaquettes. Thus, we should view these pairs of
defective plaquettes as the pointlike excitations of the coupled-wire
array. }
\label{fig: strings and plaquettes}
\end{figure}
\subsubsection{``Spin-1'' string and membrane operators}
\label{subsubsec: ``Spin-1'' string and membrane operators}
In direct analogy with Sec.\
\ref{subsec: Case study: su(2)_2}, we build ``spin-1'' string operators
acting parallel to the $x$-$y$ plane of the square lattice $\Lambda$
by taking products of the unitary operators
\begin{align}
\label{eq: O (1) def}
\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}(t,z)
&\:=
\widehat{\Phi}^{(1)\dagger}_{\gamma,\mathrm{L},J,p^{\,}_{K}}(t,z)\,
\widehat{\Phi}^{(1)}_{\gamma^{\prime},\mathrm{R},J,p^{\,}_{K}}(t,z)
\\
&=
e^{-\mathrm i\frac{1}{\sqrt{2}}\widehat{\phi}^{\,}_{\gamma,\mathrm{L},J,p^{\,}_{K}}(t,z)}\,
e^{+\mathrm i\frac{1}{\sqrt{2}}\widehat{\phi}^{\,}_{\gamma^{\prime},\mathrm{R},J,p^{\,}_{K}}(t,z)},
\nonumber
\end{align}
for $\gamma,\gamma^{\prime}=1,2$, $J,K=A,B,C,D$, and for any
$0<z<L^{\,}_{z}$ (we continue to suppress the normal ordering of
vertex operators). For any choice of $\gamma,\gamma^{\prime},J,$ and
$K$, the operator
$\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$ fails
to commute with exactly three plaquette terms in the interaction
\eqref{eq: plaquette interaction}
[see Fig.~\ref{fig: strings and plaquettes}(a)].
We refer to these plaquettes as ``defective.''
A calculation analogous to Eq.~\eqref{eq2Dcase: psi y-string commutators}
shows that acting with an additional operator
$\widehat{\mathcal{O}}^{(1)}_
{\gamma^{\prime\prime}\gamma^{\prime\prime\prime\prime},J^\prime,p^{\prime}_{K^{\prime}}}$
in a neighboring wire heals some of these defective plaquettes,
while creating others.
By repeating this calculation, one can verify that
it is possible to
separate \textit{pairs} of defective plaquettes arbitrarily
far from one another without creating
any additional defective plaquettes
[see Fig.~\ref{fig: strings and plaquettes}(b)].
Thus, we can interpret each \textit{pair} of defective plaquettes
as a pointlike ``spin-$1$'' excitation.
The fact that one can separate these excitations arbitrarily far
from one another without creating additional
defective plaquettes indicates that they are deconfined.
(For a more detailed discussion of deconfinement and the energetics of
these excitations, see Sec.~\ref{subsubsec: Aside: Comments on energetics}.)
For the purposes of building string operators, we need to identify a
set of three orthogonal cycles of the three-torus. The cycle winding
around the torus in the $z$-direction consists of traversing the
torus along a single wire, as was done in the 2D case. For cycles
parallel to the $x$-$y$ plane, we need to choose a canonical set of
paths through the square lattice $\Lambda$. We call these paths
$\mathcal{P}^{\,}_{\hat{x}}$ and $\mathcal{P}^{\,}_{\hat{y}}$. They are
depicted in Fig.\ \ref{fig: P_x and P_y defs}. Although two particular
choices of path are depicted in
Fig.\ \ref{fig: P_x and P_y defs}, one should view these
paths as members of two families
$[\mathcal{P}^{\,}_{\hat{x}}]$ and $[\mathcal{P}^{\,}_{\hat{y}}]$,
each consisting of of parallel paths through the
lattice. Members of the family $[\mathcal{P}^{\,}_{\hat{x}}]$
can be deformed into members of the family $[\mathcal{P}^{\,}_{\hat{y}}]$
without creating additional defective plaquettes
(see Fig.\ \ref{fig: deforming paths}).
Moreover, one can act with products of
operators \eqref{eq: O (1) def} along closed loops
such that no defective plaquettes are created
(see Fig.\ \ref{fig: closed loops}).
These observations are consistent
with the aforementioned deconfinement of the ``spin-$1$'' excitations.
\begin{figure}[tb]
\centering
\includegraphics[width=0.425\textwidth]{Figures10-5}
\caption{
(Color online) Example of a deformation of a path in the family
$[\mathcal{P}^{\,}_{\hat{x}}]$ (see Fig.~\ref{fig: P_x and P_y defs})
that runs parallel to paths in the family $[\mathcal{P}^{\,}_{\hat{y}}]$
in some segments. One can check that a ``spin-$1$''
or ``spin-$\frac{1}{2}$'' string defined along this path still commutes
with the interaction \eqref{eq: plaquette interaction}. }
\label{fig: deforming paths}
\end{figure}
\begin{figure}[tb]
\centering
\includegraphics[width=0.375\textwidth]{Figures10-75}
\caption{
(Color online)
Two examples of closed paths. One can show that a
``spin-$1$'' or ``spin-$\frac{1}{2}$'' string defined along any such path
commutes with the interaction \eqref{eq: plaquette interaction}, and thereby
introduces no excitations.
}
\label{fig: closed loops}
\end{figure}
The ``spin-$1$'' string operators acting along the paths
$\mathcal{P}^{\,}_{\hat{x}}$ and $\mathcal{P}^{\,}_{\hat{y}}$
are given by
\begin{align}
\widehat{\Gamma}^{(1)}_{\hat{a}}(t,z)\:=
\prod_{(\gamma,\gamma^{\prime},J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}}}
\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,\tilde{\bm{r}}}(t,z),
\label{eq: Gamma (1) x,y defs}
\end{align}
where $\hat{a}=\hat{x},\hat{y}$.
A calculation analogous to Eqs.\
\eqref{eq2Dcase: psi y-string commutators}
in Sec.\
\ref{subsec: Case study: su(2)_2}
shows that both families of string operators commute with the interaction \eqref{eq: plaquette interaction}
when periodic boundary conditions are imposed in all directions.
Like their 2D counterpart defined in Eq.~\eqref{eq: def Gamma psi 1 at z}, these string operators
can be interpreted
as creating a pair of pointlike ``spin-$1$'' excitations, transporting them around either the
$\mathcal{P}^{\,}_{\hat{x}}$- or $\mathcal{P}^{\,}_{\hat{y}}$-cycle of the three-torus, and then
annihilating them. Also like their 2D counterpart, these string operators can be rewritten as products
of primary operators at different $z$ points, similar to Eq.~\eqref{eq: general ``spin-1'' string operator}
in the 2D case. We can thus view these ``spin-$1$'' excitations as being free to move in all three
spatial dimensions.
To construct a ``spin-$1$'' string operator running along the $z$-direction, we define the
unitary operator
\begin{align}
&\widehat{\mathcal{O}}^{(1)}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z^{\,}_{1},z^{\,}_{2})
\:=
\widehat{\Phi}^{(1)\dagger}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z^{\,}_{2})\,
\widehat{\Phi}^{(1)}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z^{\,}_{1})
\nonumber
\\
&\qquad=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{\sqrt{2}}
\int\limits^{z^{\,}_{2}}_{z^{\,}_{1}}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z)
\Bigg).
\label{eq: spin-1 z string operator}
\end{align}
This definition is in direct parallel with Eq.~\eqref{eq2Dcase: spin-1 z string operator} in the 2D case.
Similar to the 2D case, one can verify by direct calculation that the nonlocal operator
\begin{align}
\label{eq: def Gamma (1) z}
\widehat{\Gamma}^{(1)}_{\hat{z},\gamma,\mathrm{M},J,p^{\,}_{K}}(t)
\:=
\widehat{\mathcal{O}}^{(1)}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,0,L^{\,}_{z})
\end{align}
commutes with the interaction \eqref{eq: plaquette interaction}. This completes the definitions of the
``spin-$1$'' string operators we will consider.
\begin{figure}[tb]
\includegraphics[width=0.45\textwidth]{Figures11}
\caption{
(Color online)
A cartoon representation of the ``spin-$1$'' membrane operator
$\widehat{\Sigma}^{(1)}_{\hat{x}}$ defined in Eq.~(\ref{eq: Sigma (1) x,y defs}).
}
\label{fig: membrane def}
\end{figure}
In 3D, we can also build nontrivial \textit{membrane} operators whose
algebra with the string operators can indicate the presence of
topological order. Our general strategy for defining membrane
operators parallel to the $x$-$z$ and $y$-$z$ planes is to apply
$z$-string operators of the form \eqref{eq: def Gamma (1) z} along
paths in the families $[\mathcal{P}^{\,}_{\hat{x}}]$ and
[$\mathcal{P}^{\,}_{\hat{y}}$], respectively. This yields the
``spin-1'' membrane operators
\begin{align}
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)
\:=
\prod_{(\gamma,\gamma^{\prime},J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}}
\widehat{\mathcal{O}}^{(1)}_{\gamma^{\prime},\mathrm{R},J,p^{\,}_{K}}(t,0,L^{\,}_{z}),
\label{eq: Sigma (1) x,y defs}
\end{align}
where $\hat{a}=\hat{x},\hat{y}$ and $\hat{a}^{\,}_{\perp}$ is defined such that
$\hat{x}^{\,}_{\perp}=\hat{y}$ and $\hat{y}^{\,}_{\perp}=\hat{x}$.
The choice of chirality
$\mathrm{M}=\mathrm{R}$ above is arbitrary.
Here, we have adopted a convention whereby any membrane carries the label
of a path normal to the membrane.
A depiction of one such membrane is shown in Fig.\ \ref{fig: membrane def}.
Constructing a membrane parallel to the $x$-$y$ plane
is achieved by simply acting with the bilinears \eqref{eq: O (1) def}
in all wires according to the prescription
\begin{align}
\label{eq: Sigma (1) z def}
\widehat{\Sigma}^{(1)}_{\hat{z}}(t,z)\:=
\prod_{p^{\,}_{A}\subset\Lambda}
\prod^{D}_{J=A}
\widehat{\mathcal{O}}^{(1)}_{11,J,p^{\,}_{A}}(t,z)\,
\widehat{\mathcal{O}}^{(1)}_{22,J,p^{\,}_{A}}(t,z).
\end{align}
A partial implementation of any of these membrane operators, obtained
by restricting its support to an open surface, leaves a line of
defective plaquettes along the boundary of the surface. Thus, we can
interpret the membranes on which these operators act as worldsheets of
stringlike excitations of the coupled-wire theory.
Using the definitions
\eqref{eq: Gamma (1) x,y defs},
\eqref{eq: def Gamma (1) z},
\eqref{eq: Sigma (1) x,y defs},
and
\eqref{eq: Sigma (1) z def},
and the equal-time algebra
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c},
one can show that the
``spin-$1$'' string operators $\widehat{\Gamma}^{(1)}_{\hat{a}}$
and membrane operators
$\widehat{\Sigma}^{(1)}_{\hat{b}}$
commute with one another for all
$\hat{a},\hat{b}=\hat{x},\hat{y},\hat{z}$.
Moreover, one verifies that this
equal-time algebra is independent of the details of how
one defines the paths and surfaces on which the string and membranes
act, i.e., deforming the path along which a string operator acts,
or the surface on which a membrane operator acts, has no effect on the
equal-time algebra as long as these deformations leave
the intersection of the path and surface intact.
\subsubsection{``Spin-$\frac{1}{2}$'' string and membrane operators}
\label{subsubsec: ``spin=1/2'' string and membrane operators}
String operators corresponding to the ``spin-$\frac{1}{2}$'' primary operator can
be constructed as follows. Similarly to the case of ``spin-$1$'' strings,
``spin-$\frac{1}{2}$'' strings acting along paths in the $x$-$y$ plane are built out
of the bilinear operators
\begin{align}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}(t,z)\:=
\widehat{\Phi}^{(\frac{1}{2})\dagger}_{\gamma,\mathrm{L},J,p^{\,}_{K}}(t,z)\,
\widehat{\Phi}^{(\frac{1}{2})}_{\gamma^{\prime},\mathrm{R},J,p^{\,}_{K}}(t,z),
\end{align}
where the operator
$\widehat{\Phi}^{(\frac{1}{2})}_{\gamma^{\prime},\mathrm{R},J,p^{\,}_{K}}(t,z)$
is defined in Eq.\ \eqref{eq: spin-1/2 primary 3D}.
The ``spin-$\frac{1}{2}$'' string operators along the
paths $\mathcal{P}^{\,}_{\hat{x},\hat{y}}$ are then defined by
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}(t,z)\:=
\prod_{(\gamma,\gamma^{\prime},J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}}}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}(t,z),
\label{eq: Gamma (1/2) x,y defs}
\end{align}
where $\hat{a}=x,y$. ``Spin-$\frac{1}{2}$'' string operators
along the $z$-direction are again defined by analogy with the 2D case,
\begin{align}
\begin{split}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(t,\epsilon)
&\:=
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,z)
\Bigg)
\\
&\quad
\times
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,0)\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(t,\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}},
\end{split}
\label{eq: Gamma (1/2) z def}
\end{align}
where the choice of $\gamma,\mathrm{M},J,$ and $p^{\,}_{K}$
is arbitrary. The operator $\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$ above is the
counterpart to the
projector onto the fusion channel $\sigma\times\sigma=\mathbbm{1}$
that appears in Eq.~\eqref{eq2Dcase: Gamma^(1/2)_2,y' definition}.
This definition of the operator
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z},\gamma,\mathrm{M},J,p^{\,}_{K}}(t,\epsilon)$
is subject
to the same caveats as its 2D analogue,
which was defined in Eq.~\eqref{eq2Dcase: Gamma^(1/2)_2,y' definition}.
In particular, the limit $\epsilon\to0$ must be taken carefully, as discussed
in footnote \cite{foot1} and Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}.
As in the 2D case, we will only take the limit $\epsilon\to0$
at the end of calculations.
With these definitions, one verifies
using Eq.~\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c},
that these ``spin-$\frac{1}{2}$'' string operators have the following equal-time algebra
with the ``spin-$1$'' membrane operators. Any ``spin-$\frac{1}{2}$'' string operator
defined along a
noncontractible cycle parallel to the $x$-$y$ plane anticommutes with a
``spin-$1$'' membrane operator orthogonal to the noncontractible cycle,
\begin{subequations}
\label{eq: equaltime algebra 3d between string and membranes part I}
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}(t,z)\,
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)
=&\,
-
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}(t,z),
\end{align}
for any $\hat{a}=\hat{x},\hat{y}$.
In contrast, any ``spin-$\frac{1}{2}$'' string operator defined along a
a noncontractible cycle in the $x$-$y$ plane commutes with any
``spin-$1$'' membrane operator such that the noncontractible cycle and membrane
are not pairwise orthogonal, i.e.,
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}(t,z)\,
\widehat{\Sigma}^{(1)}_{\hat{b}}(t)
=&\,
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{b}}(t,z),
\end{align}
\end{subequations}
for $(\hat{a},\hat{b})=(\hat{x},\hat{y})$ or $(\hat{a},\hat{b})=(\hat{y},\hat{x})$.
This equal-time algebra holds independently of local deformations
of the paths and surfaces on which the string and membrane operators are
defined, so long as these deformations leave the intersections of these paths
and surfaces unchanged. The ``spin-$\frac{1}{2}$'' string operator acting
along the $\hat{z}$-direction anticommutes with any
``spin-$1$'' membrane operator
acting on a surface that is orthogonal to the $\hat{z}$-direction,
\begin{subequations}
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(t,\epsilon)\,
\widehat{\Sigma}^{(1)}_{\hat{z}}(t,z)=
-
\widehat{\Sigma}^{(1)}_{\hat{z}}(t,z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(t,\epsilon),
\end{align}
for any infinitesimal $\epsilon>0$.
The ``spin-$\frac{1}{2}$'' string operator acting
along the $\hat{z}$-direction commutes with
any $\psi$-membrane operator
acting on a surface orthogonal to the $x$- or $y$-directions
(for simplicity, we assume that the $z$-string does not
intersect with the $x$- and $y$-membranes),
\begin{align}
&
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(t,\epsilon)\,
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)=
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(t,\epsilon),
\end{align}
\end{subequations}
for any $\hat{a}=\hat{x},\hat{y}$
and for any infinitesimal $\epsilon>0$.
Once we have constructed the ``spin-$\frac{1}{2}$'' string
operators, we can also investigate the braiding statistics of
pointlike particles in the coupled-wire theory. For example, the
mutual statistics of ``spin-$\frac{1}{2}$'' and ``spin-$1$''
excitations can be deduced from exchange relations like
\begin{widetext}
\begin{align}
\widehat{\Gamma}^{(1)}_{\hat{y}}(t,z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})=
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})\,
\widehat{\Gamma}^{(1)}_{\hat{y}}(t,z)\,
e^{-\mathrm{i}\,\frac{\pi}{2}\,\mathrm{sgn}(z-z^{\prime})}\,
e^{+\mathrm{i}\,\frac{\pi}{2}\,\mathrm{sgn}(z-z^{\prime})}=
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})\,
\widehat{\Gamma}^{(1)}_{\hat{y}}(t,z),
\end{align}
where we used the equal-time algebra
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c}, which
demonstrates that ``spin-$1$'' and ``spin-$\frac{1}{2}$'' particles braid trivially in the
three-dimensional model. Likewise, the self-statistics of ``spin-$\frac{1}{2}$''
excitations can be deduced from exchange relations like
\begin{align}\label{eq: Gamma^sigma_y commutes with Gamma^sigma_x}
\begin{split}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}(t,z)\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})
&=
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}(t,z)
\times
\begin{cases}
e^{-\mathrm{i}\,\frac{3\pi}{8}\,\mathrm{sgn}(z-z^{\prime})}\,
e^{+\mathrm{i}\,\frac{3\pi}{8}\,\mathrm{sgn}(z-z^{\prime})},
&
\mathrm{if } \sigma\times\sigma=\mathbbm{1},
\\
e^{+\mathrm{i}\,\frac{\pi}{8}\,\mathrm{sgn}(z-z^{\prime})}\,
e^{-\mathrm{i}\, \frac{\pi}{8}\,\mathrm{sgn}(z-z^{\prime})},
& \mathrm{if } \sigma\times\sigma=\psi
\end{cases}\\
&=
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}(t,z^{\prime})\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}(t,z),
\end{split}
\end{align}
\end{widetext}
where we used \eqref{eq: def bulk chiral parafermion and boson rep su(2)k c}
and the counterparts to the equal-time algebra
\eqref{eq2Dcase: sigma sigma algebra}.
The meaning of the two cases distinguished above, namely the cases
$\sigma\times\sigma=\mathbbm{1}$ and $\sigma\times\sigma=\psi$, is
as follows. When two ``spin-$\frac{1}{2}$'' strings act along the
noncontractible cycles
$\mathcal{P}^{\,}_{\hat{x}}$ and $\mathcal{P}^{\,}_{y}$,
they necessarily coincide in exactly two chiral
channels located astride a bond of the square lattice
[see Fig.\ \ref{fig: P_x and P_y defs}(b)].
Each of these chiral channels
is acted upon by two $\sigma$ operators from the $\mathbb{Z}^{\,}_{2}$ sector,
one from the $\mathcal{P}^{\,}_{\hat{x}}$ string and one from the
$\mathcal{P}^{\,}_{\hat{y}}$ string. The outcome of fusing the two
$\sigma$ fields in each of the two channels is correlated. If one pair
of $\sigma$s fuses to $\mathbbm{1}$ or $\psi$, then the other pair
must fuse in this channel as well. Otherwise, extra excitations are
created. The upshot of this discussion is that all pointlike
particles in the three-dimensional theory have trivial braiding with
one another. This fact is consistent with the
fact that any deconfined point particle in three spatial dimensions must be
either a fermion or a boson. As we will see below, however, there is
no such restriction for the braiding of a pointlike excitation with a
linelike excitation.
The logic for the construction of ``spin-$\frac{1}{2}$'' membranes parallels the
logic for ``spin-$1$'' membranes. A ``spin-$\frac{1}{2}$'' membrane parallel to the $x$-$y$ plane
is defined by
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}(t,z)\:=
\prod_{J,p^{\,}_{A}}
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{11,J,p^{\,}_{A}}(t,z)\,
\widehat{\mathcal{O}}^{(\frac{1}{2})}_{22,J,p^{\,}_{A}}(t,z),
\label{eq: def sigma membranes a}
\end{align}
while the membrane operators orthogonal to the $x$-$y$ plane can be
chosen to be
\begin{widetext}
\begin{align}
&
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon)\:=
\prod_{(\gamma,J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}}
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}(t,z)
\Bigg)
\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(t,0)\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(t,\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}},
\label{eq: az sigma membrane def}
\end{align}
\end{widetext}
for any $\hat{a}=\hat{x},\hat{y}$.
The choice of chirality $\mathrm{M}=\mathrm{R}$ is arbitrary.
The definition \eqref{eq: az sigma membrane def}
of the operator $\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon)$
is subject to the same caveats as the definition
\eqref{eq: Gamma (1/2) z def}
of the operator $\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}(\epsilon,t)$.
As before, we refer the reader
to Footnote~\cite{foot1} and to Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}
for details.
Using the algebra
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c}
and the definitions (\ref{eq: def sigma membranes a}) and (\ref{eq: az sigma membrane def}),
one can show that the equal-time algebra
between any pair of ``spin-$\frac{1}{2}$'' membrane ``spin-$1$''-string operators
is mere commutations except for the three anticommuting exceptions
\begin{subequations}
\begin{align}
&
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon)\,
\widehat{\Gamma}^{(1)}_{\hat{a}}(t,z)=
-
\widehat{\Gamma}^{(1)}_{\hat{a}}(t,z)\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon),
\\
&
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}(t,z)\,
\widehat{\Gamma}^{(1)}_{\hat{z}}(t)=
-
\widehat{\Gamma}^{(1)}_{\hat{z}}(t)\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}(t,z),
\end{align}
\end{subequations}
for $\hat{a}=\hat{x},\hat{y}$,
and for infinitesimal $\epsilon>0$.
Thus, any of the ``spin-$\frac{1}{2}$'' membrane operators can be interpreted
as twisting the boundary conditions of a pointlike ``spin-$1$'' excitation encircling the
three-torus along any noncontractible cycle orthogonal
to the membrane.
Finally, we also have the algebra between ``spin-$\frac{1}{2}$'' membranes
and ``spin-$1$'' membranes given by
\begin{subequations}
\begin{align}
&
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}^{\,}_{\perp}}(t,\epsilon)=
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}^{\,}_{\perp}}(t,\epsilon)\,
\widehat{\Sigma}^{(1)}_{\hat{a}}(t),
\\
&
\widehat{\Sigma}^{(1)}_{\hat{a}}(t)\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}(t,z)=
(-1)^{N^{\,}_{\hat{a}}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}(t,z)\,
\widehat{\Sigma}^{(1)}_{\hat{a}}(t),
\\
&
\widehat{\Sigma}^{(1)}_{\hat{z}}(t,z)\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon)=
(-1)^{N^{\,}_{\hat{a}}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}(t,\epsilon)\,
\widehat{\Sigma}^{(1)}_{\hat{z}}(t,z),
\end{align}
\end{subequations}
for any $\hat{a}=\hat{x},\hat{y}$
(recall that $\hat{x}^{\,}_{\perp}=\hat{y}$ and $\hat{y}^{\,}_{\perp}=\hat{x}$),
and for any infinitesimal $\epsilon>0$.
The system-size-dependent integers $N^{\,}_{\hat{x}}$ and
$N^{\,}_{\hat{y}}$ are the number of wires contained in the path
$\mathcal{P}^{\,}_{\hat{x}}$ and $\mathcal{P}^{\,}_{\hat{y}}$,
respectively. One can show that $N^{\,}_{\hat{x}}$ and
$N^{\,}_{\hat{y}}$ are even for paths $\mathcal{P}^{\,}_{\hat{x}}$
and $\mathcal{P}^{\,}_{\hat{y}}$ that encompass the entire system,
so long as the system contains an integer number of unit cells.
\subsubsection{An aside on energetics}
\label{subsubsec: Aside: Comments on energetics}
Before moving on to discuss the derivation of the topological
degeneracy from the algebra of the string and membrane operators
constructed in Secs.\
\ref{subsubsec: ``Spin-1'' string and membrane operators} and
\ref{subsubsec: ``spin=1/2'' string and membrane operators},
it is necessary to address some aspects of the energetics of the
excitations associated with these operators. We will focus first on
the case of the pointlike ``spin-$1$'' excitations, and then on the
case of the pointlike ``spin-$\frac{1}{2}$'' excitations.
The plaquette excitations generated by the ``spin-$1$'' operator
$\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
are solitons in a particular linear combination of scalar fields in the
$u(1)^{\,}_{2}$ sector. For example, the operator
$\widehat{\mathcal{O}}^{(1)}_{11,C,p^{\,}_{A}}$ acts nontrivially on
only two scalar fields,
\begin{widetext}
\begin{subequations}
\begin{align}
\frac{1}{\sqrt{2}}\, \widehat{\mathcal{O}}^{(1)}_{11,C,p^{\,}_{A}}(z^\prime)\,
\widehat{\phi}^{\,}_{1,R,C,p^{\,}_{A}}(z)\,
\widehat{\mathcal{O}}^{(1)\dagger}_{11,C,p^{\,}_{A}}(z^\prime)
&=\frac{1}{\sqrt{2}}\,
\widehat{\phi}^{\,}_{1,R,C,p^{\,}_{A}}(z)
-
2\pi\,
\Theta(z-z^\prime)
+
\text{const.},
\\
\frac{1}{\sqrt{2}}\, \widehat{\mathcal{O}}^{(1)}_{11,C,p^{\,}_{A}}(z^\prime)\,
\widehat{\phi}^{\,}_{1,L,C,p^{\,}_{A}}(z)\,
\widehat{\mathcal{O}}^{(1)\dagger}_{11,C,p^{\,}_{A}}(z^\prime)
&=\frac{1}{\sqrt{2}}\,
\widehat{\phi}^{\,}_{1,L,C,p^{\,}_{A}}(z)
-
2\pi\,
\Theta(z-z^\prime)
+
\text{const.},
\end{align}
\end{subequations}
\end{widetext}
where we have used the algebra
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c}.
Here, $2\pi\,\Theta(z-z^\prime)=\pi\[\text{sgn}(z-z^\prime)+1\]$
is the Heaviside step function that represents a sharp soliton of height
$2\pi$. The additive constant terms arise from the algebra in Eq.\
\eqref{eq: def bulk chiral parafermion and boson rep su(2)k c} and
are multiples of $2\pi$.
The introduction of these soliton profiles costs infinite energy after
the limit $\lambda\to\infty$ of the interaction
\eqref{eq: plaquette interaction},
holding the kinetic energy fixed, has been taken.
This is consistent with the expectation that the
theory has an infinite energy gap in this limit.
Acting in neighboring wires with additional operators
$\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$ in
the manner of Fig.~\ref{fig: strings and plaquettes} leads to
cancellations of soliton profiles in any linear combination of scalar
fields that is affected by two successive such operators.
For this reason, a string of operators
$\widehat{\mathcal{O}}^{(1)}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
costs energy only at its end points (end wires).
It is the property that this energy cost
(possibly infinite in the limit of infinitly strong interactions
relative to the kinetic energy)
is localized around the two end points (end wires) of the string
that justifies interpreting the associated ``spin-1'' excitations
as being deconfined.
The case of the ``spin-$\frac{1}{2}$'' excitations is similar, but
complicated by the apperance of operators from the
$\mathbb{Z}^{\,}_{2}$ sector in the operators
$\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$.
In this case, the $u(1)^{\,}_{2}$ part of the operator
$\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
also creates a soliton, but one with height $\pi$ rather than
$2\pi$. Such an object cannot, by itself, be created by a local
operator. However, the combination of $u(1)^{\,}_{2}$ and
$\mathbb{Z}^{\,}_{2}$ operators entering
$\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
is local. While an explicit calculation of the energy cost that
results from acting with the operator
$\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
in a single wire for any value of the coupling $\lambda$ is beyond the
scope of this work, we expect that it will be infinite in the limit
$\lambda\to\infty$, judging from the action of this operator on the
$u(1)^{\,}_{2}$ sector.
If this energy cost is localized around
the two end points (end wires) of the string, we may then
interpret the associated ``spin-$\frac{1}{2}$'' excitations
as being deconfined. Although proving deconfinement of ``spin-$\frac{1}{2}$'' excitations
is beyond the scope of this work because of the $\mathbb{Z}^{\,}_{2}$ sector of
the theory, we expect it judging from the action of
$\widehat{\mathcal{O}}^{(\frac{1}{2})}_{\gamma\gamma^{\prime},J,p^{\,}_{K}}$
on the $u(1)^{\,}_{2}$ sector.
The interaction \eqref{eq: plaquette interaction} possesses
an extensive number of symmetries. String operators
acting along closed loops do not leave any solitons in
the linear combinations of scalar fields that enter the interaction
\eqref{eq: plaquette interaction} and hence create no excitations.
Indeed, such operators commute with the interaction
\eqref{eq: plaquette interaction}.
As such, they play a role similar to
the one played by the local gauge symmetry of the toric code.
In fact, in the coupled-wire construction of 3D Abelian
topological phases caried out in Ref.~\cite{Iadecola16}, it was shown
explicitly that such symmetries were present
for a class of interactions studied there.
However, a complicating difference between the interaction
\eqref{eq: plaquette interaction}
and the ones studied in Ref.\ \cite{Iadecola16} is that the latter
consisted of sums over local terms that were all
pairwise commuting, whereas this is not the case for
the plaquette interactions entering
Eq.\ \eqref{eq: plaquette interaction}.
Consequenlty, it is not possible to find a closed form for the ground states
of Eq.\ \eqref{eq: plaquette interaction} as was done in
Ref.\ \cite{Iadecola16}.
When the interaction \eqref{eq: plaquette interaction}
is weakly perturbed by the kinetic energy of the chiral modes
in the quantum wires, i.e., when $0<\lambda<\infty$, the string operators
acting along closed loops like the ones depicted in Fig.\
\ref{fig: closed loops} fail to commute with the Hamiltonian.
However, because the kinetic energy is a local perturbation,
it cannot lift the topological degeneracy,
which we derive in Sec.\
\ref{subsubsec: Topological degeneracy on the three-torus}
when the interaction is infinitely stronger than the kinetic energy,
to any finite order in perturbation theory. A quantitative effect of
weakly perturbing interaction \eqref{eq: plaquette interaction}
by the kinetic energy of the chiral modes is to render the excitation energy
of open strings finite. This finite excitation energy is shared between
the end points (end wires) and a string tension. However, the string tension
is not strong enough to confine the excitations localized around the
end points (end wires) for a sufficiently weak kinetic energy,
as there is a one-to-one correspondence between the topological
degeneracy and the deconfinement
of excitations carrying fractional quantum numbers
(defective plaquettes on the boundaries of strings and membranes).
\subsubsection{Topological degeneracy on the three-torus}
\label{subsubsec: Topological degeneracy on the three-torus}
\begin{table*}[t]
\begin{tabular}{|c||c|c|c|c|c|c||c|c|c|c|c|c|}
\hline
\ &
$\widehat{\Gamma}^{(1)}_{\hat{x}}$ &
$\widehat{\Gamma}^{(1)}_{\hat{y}}$&
$\widehat{\Gamma}^{(1)}_{\hat{z}}$&
$\widehat{\Sigma}^{(1)}_{\hat{x}}$ &
$\widehat{\Sigma}^{(1)}_{\hat{y}}$&
$\widehat{\Sigma}^{(1)}_{\hat{z}}$&
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}$&
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}$ &
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}$&
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$ &
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$&
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\\
\hline
\hline
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}$ &
$+$ &
$+$ &
$+$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$+$ &
$+$
\\
\hline
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}$ &
$+$ &
$+$ &
$+$ &
$+$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$+$
\\
\hline
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$
\\
\hline
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$\text{\ding{55}}$
\\
\hline
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$ &
$+$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$+$ &
$\text{\ding{55}}$ &
$+$ &
$\text{\ding{55}}$
\\
\hline
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}$ &
$+$ &
$+$ &
$-$ &
$+$ &
$+$ &
$+$ &
$+$ &
$+$ &
$\text{\ding{55}}$ &
$\text{\ding{55}}$ &
$\text{\ding{55}}$ &
$+$
\\
\hline
\end{tabular}
\caption{
Summary of the algebra of the string and membrane operators
\eqref{eq: string and membrane operators}. Entries corresponding
to a pair of operators that commute are labeled with a $+$.
Entries corresponding
to a pair of operators that anticommute are labeled with a $-$.
Entries corresponding
to a pair of operators that neither commute nor anticommute
are labeled with a $\text{\ding{55}}$.
(Compare with Table~\ref{table2Dcase: summary of algebra}.)
The operator algebra contained in the left $6\times 6$
subblock of the table is derived in
Secs.\ \ref{subsubsec: ``Spin-1'' string and membrane operators}
and \ref{subsubsec: ``spin=1/2'' string and membrane operators}.
The operator algebra contained in the right $6\times 6$
subblock of the table is derived in
Sec.~\ref{subsubsec: Topological degeneracy on the three-torus}.
\label{table: summary of algebra}
}
\end{table*}
Using the results of the previous sections, we now derive
the topological ground-state degeneracy of the array of quantum wires
coupled by the interwire interactions
\eqref{eq: bulk interactions su(2)_k},
for the case of $su(2)^{\,}_{2}$ current-current interactions.
We assume periodic boundary conditions in $x$, $y$, and
$z$, so that the array of coupled wires has the topology of a
three-torus ($\mathbb T^{3}$). The logic of our derivation of this
lower bound follows
closely the logic of the corresponding derivation in the
two-dimensional case discussed in Sec.\
\ref{subsec: Case study: su(2)_2}.
It hinges on the exchange algebra
of the following set of nonlocal operators,
which is summarized in Table~\ref{table: summary of algebra}.
There are three nonlocal and \textit{unitary} ``spin-$1$'' string operators
(\ref{eq: Gamma (1) x,y defs}) and
(\ref{eq: def Gamma (1) z}),
for which we use the short-hand notation
\begin{subequations}\label{eq: string and membrane operators}
\begin{align}
\widehat{\Gamma}^{(1)}_{\hat{x}},\quad
\widehat{\Gamma}^{(1)}_{\hat{y}},\quad
\widehat{\Gamma}^{(1)}_{\hat{z}},
\label{eq: family of psi string operators}
\end{align}
respectively.
There are three nonlocal and \textit{nonunitary} ``spin-$\frac{1}{2}$'' string operators
(\ref{eq: Gamma (1/2) x,y defs})
and
(\ref{eq: Gamma (1/2) z def})
for which we use the short-hand notation
\begin{align}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}},\quad
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}},\quad
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}},
\label{eq: family of sigma string operators}
\end{align}
respectively.
There are three nonlocal and \textit{unitary} ``spin-$1$'' membrane operators
(\ref{eq: Sigma (1) x,y defs})
and
(\ref{eq: Sigma (1) z def})
for which we use the short-hand notation
\begin{align}
\widehat{\Sigma}^{(1)}_{\hat{z}},\quad
\widehat{\Sigma}^{(1)}_{\hat{y}},\quad
\widehat{\Sigma}^{(1)}_{\hat{x}},
\label{eq: family of psi membrane operators}
\end{align}
respectively.
There are three nonlocal and \textit{nonunitary} ``spin-$\frac{1}{2}$'' membrane operators,
defined in Eqs.~(\ref{eq: def sigma membranes a}) and
(\ref{eq: az sigma membrane def}),
for which we use the short-hand notation
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}},\quad
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}},\quad
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}},
\label{eq: family of sigma membrane operators}
\end{align}
\end{subequations}
respectively.
Each of these twelve nonlocal operators commutes with the
plaquette interaction \eqref{eq: plaquette interaction},
except for the three operators $\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}$,
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$, and $\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$,
defined in Eqs.~\eqref{eq: Gamma (1/2) z def} and
\eqref{eq: az sigma membrane def}, respectively.
These three operators are regularized by the parameter $\epsilon$,
and therefore
must be treated in a manner similar to the operator
$\widehat{\Gamma}^{(\frac{1}{2})}_{2,\mathrm{M},y^{\prime}}(\epsilon)$
in the 2D case. Nevertheless, an analysis along the lines of the
one presented in Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}
for the 2D case reveals that these three $\epsilon$-regularized
operators can be used to define states in the ground-state
manifold of the interaction \eqref{eq: plaquette interaction}.
We will elaborate on this statement below.
The derivation of the topological degeneracy
begins by observing that the ``spin-$1$'' string and ``spin-$1$'' membrane operators
appearing in Eqs.\ \eqref{eq: family of psi string operators} and
\eqref{eq: family of psi membrane operators} all commute with one another.
Thus, we can choose a many-body ground state
\begin{align}
\ket{\Omega}\equiv
\ket{\mathbbm{1}}
\end{align}
that is a simultaneous eigenstate of all
``spin-$1$'' string and ``spin-$1$'' membrane operators, namely
\begin{subequations}
\label{eq: string eigenvalues for vacuum}
\begin{align}
\widehat{\Gamma}^{(1)}_{\hat{z}}\,\ket{\mathbbm{1}}=
\omega^{\Gamma}_{\hat{z}}\,\ket{\mathbbm{1}},\
\widehat{\Gamma}^{(1)}_{\hat{y}}\,\ket{\mathbbm{1}}=
\omega^{\Gamma}_{\hat{y}}\,\ket{\mathbbm{1}},\
\widehat{\Gamma}^{(1)}_{\hat{x}}\,\ket{\mathbbm{1}}=
\omega^{\Gamma}_{\hat{x}}\,\ket{\mathbbm{1}},
\end{align}
on the one hand, and
\begin{align}
\widehat{\Sigma}^{(1)}_{\hat{z}}\,\ket{\mathbbm{1}}=
\omega^{\Sigma}_{\hat{z}}\,\ket{\mathbbm{1}},\
\widehat{\Sigma}^{(1)}_{\hat{x}}\,\ket{\mathbbm{1}}=
\omega^{\Sigma}_{\hat{x}}\,\ket{\mathbbm{1}},
\widehat{\Sigma}^{(1)}_{\hat{y}}\,\ket{\mathbbm{1}}=
\omega^{\Sigma}_{\hat{y}}\,\ket{\mathbbm{1}},
\end{align}
on the other hand, must hold for the nonvanishing eigenvalues
\begin{equation}
\label{eq: string and membrane eigenvalues}
\omega^{\Gamma}_{\hat{z}},\omega^{\Gamma}_{\hat{y}},\omega^{\Gamma}_{\hat{x}},
\omega^{\Sigma}_{\hat{z}},\omega^{\Sigma}_{\hat{y}},\omega^{\Sigma}_{\hat{x}}
\in U(1).
\end{equation}
\end{subequations}
Not all choices of $\ket{\Omega}$ are equivalent.
Similarly to the argument presented in Appendix%
~\ref{appendix: Commutation between string operators and the Hamiltonian}
for the 2D case,
depending on the topological sector in which the state
$\ket{\Omega}$ resides, it is possible for the state created by acting upon
$\ket{\Omega}$ with certain combinations of the nonlocal, nonunitary
operators $\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x},\hat{y},\hat{z}}$ and
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x},\hat{y},\hat{z}}$ to have norm zero or infinity.
In other words, not all combinations of the
eigenvalues~\eqref{eq: string and membrane eigenvalues}
label states in the ground-state manifold.
\begin{table*}[t]
\centering
\begin{tabular}{|c|c||c|c||c|c|}
\hline
State & Eigenvalues & & & &
\\
\hline
\hline
$\ket{\mathbbm{1}}$
&
$(+,+,+,+,+,+)$
&
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,+,-,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,-,+,+,+)$
\\
\hline
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(+,+,+,-,+,+)$
&
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,+,+,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(+,-,+,-,+,+)$
\\
\hline
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,+,+,+,-,+)$
&
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,+,-,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(+,+,-,-,+,+)$
\\
\hline
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,+,+,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(-,+,+,+,+,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(-,+,+,+,-,+)$
\\
\hline
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,+,+,-,-,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,-,+,+,+,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,+,-,+,-,+)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,+,+,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,+,+,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,+,-,-,-,+)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,+,-,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,+,+,-,-)$
&
&
\\
\hline
\end{tabular}
\caption{
The 20 orthogonal states
of the form (\ref{eq: Ansatz GSM})
that span the ground-state manifold of the
$su(2)^{\,}_{2}$ coupled-wire theory in $(3+1)$-dimensional spacetime,
as well as the eigenvalues of these states under
the ``spin-$1$'' string and membrane operators.
The states are labeled according to the notation
$\ket{\widehat{\mathcal{O}}}=\widehat{\mathcal{O}}\,\ket{\Omega}$.
The 6-tuple of signs $\pm$ indicating the eigenvalues
of a state $\ket{\widehat{\mathcal{O}}}$ is obtained by evaluating the
list of matrix
elements
$\bra{\widehat{\mathcal{O}}}
(
\widehat{\Gamma}^{(1)}_{\hat{x}},
\widehat{\Gamma}^{(1)}_{\hat{y}},
\widehat{\Gamma}^{(1)}_{\hat{z}},
\widehat{\Sigma}^{(1)}_{\hat{x}},
\widehat{\Sigma}^{(1)}_{\hat{y}},
\widehat{\Sigma}^{(1)}_{\hat{z}}
)
\ket{\widehat{\mathcal{O}}}$
and dividing each element in the list by its magnitude.
\label{table: 3D ground states}
}
\end{table*}
Next, we define a set of many-body states obtained by acting on the
state $\ket{\mathbbm{1}}$ with the ``spin-$\frac{1}{2}$'' string and ``spin-$\frac{1}{2}$'' membrane
operators from Eqs.\
(\ref{eq: family of sigma string operators})
and
(\ref{eq: family of sigma membrane operators}),
respectively. There are
\begin{equation}
4^{3}=64
\label{eq: naive counting GSM degeneracy}
\end{equation}
states, since for any choice of a
noncontractible cycle of the three-torus ($\hat{x},\hat{y},$ or
$\hat{z}$), there are four nonlocal operators we can apply to the
state $\ket{\Omega}$. For example, fixing the
noncontractible cycle $\hat{x}$, we can insert the identity
operator $\mathbbm{1}$, the ``spin-$\frac{1}{2}$'' string operator
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}$,
the ``spin-$\frac{1}{2}$'' membrane operator
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$,
or the product
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$.
Alternatively, we can label all $2^{6}=64$ states according to the rule
\begin{widetext}
\begin{equation}
\Ket{
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{x}}\)^{\sigma^{\Sigma}_{x}}
\!\!
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{y}}\)^{\sigma^{\Sigma}_{y}}
\!\!
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{z}}\)^{\sigma^{\Sigma}_{z}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{x}}\)^{\sigma^{\Gamma}_{x}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{y}}\)^{\sigma^{\Gamma}_{y}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{z}}\)^{\sigma^{\Gamma}_{z}}
}
\!\:=\!
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{x}}\)^{\sigma^{\Sigma}_{x}}
\!\!
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{y}}\)^{\sigma^{\Sigma}_{y}}
\!\!
\(\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{z}}\)^{\sigma^{\Sigma}_{z}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{x}}\)^{\sigma^{\Gamma}_{x}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{y}}\)^{\sigma^{\Gamma}_{y}}
\!\!
\(\widehat{\Gamma}^{(\!\frac{1}{2}\!)}_{\hat{z}}\)^{\sigma^{\Gamma}_{z}}
\!
\ket{\Omega},
\label{eq: Ansatz GSM}
\end{equation}
\end{widetext}
where
$\sigma^{\Gamma}_{x},
\sigma^{\Gamma}_{y},
\sigma^{\Gamma}_{z},
\sigma^{\Sigma}_{x},
\sigma^{\Sigma}_{y},
\sigma^{\Sigma}_{z}=0,1$.
Any of the above states
involving one or more of the $\epsilon$-regularized
operators $\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}$,
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$, and
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$
carries an implicit limit $\epsilon\to0$. As in the 2D case
(see Footnote~\cite{foot1} and Appendix
\ref{appendix: Commutation between string operators and the Hamiltonian}),
this limit should be taken after forming the product of the relevant
string and/or membrane operators.
Not all states of the form (\ref{eq: Ansatz GSM})
belong to the ground state manifold, as we are going to show explicitly.
The counting based on Eq.~(\ref{eq: naive counting GSM degeneracy})
is ``naive'' because it is based purely on the number
of noncontractible cycles of
the manifold on which the theory is defined, and on the number of
string or membrane operators that can act along each
noncontractible cycle. In the following, we are going to show
that a majority of the states (\ref{eq: Ansatz GSM})
must be excluded from the
ground-state manifold, on grounds similar to the reason
for which we had to exclude the ``extra'' state
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\, \widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
that appeared in the two-dimensional example discussed
in Sec.\ \ref{subsec: String operators and topological degeneracy}. In
the end, there will be a total of
\begin{align}
D^{\,}_{\mathbb T^3} \:= 20
\end{align}
states that survive projection
into the ground-state manifold. These states are listed, along with
their eigenvalues under the ``spin-$1$'' string and membrane operators, in
Table~\ref{table: 3D ground states}. We emphasize that all of these
20 ground states are mutually orthogonal, as they are simultaneous
eigenstates of the unitary ``spin-$1$'' string and membrane operators with
different eigenvalues.
The $20$ states in Table~\ref{table: 3D ground states} have in
common the fact that they are created by acting on the state
$\ket{\Omega}$ with a product of commuting operators (these
correspond to entries marked with a $+$ in the right
$6\times6$ block of Table~\ref{table: summary of algebra}).%
~\cite{foot2}
Conversely,
the $64-20=44$ excluded states not appearing in Table
\ref{table: 3D ground states}
are created by acting on the state
$\ket{\Omega}$ with a product of noncommuting operators
(these correspond to entries marked with a $\text{\ding{55}}$ in the right
$6\times6$ block of Table~\ref{table: summary of algebra}). In the
two-dimensional case studied in
Sec.~\ref{subsec: String operators and topological degeneracy},
it was precisely the noncommutativity of the string operators
$\widehat{\Gamma}^{(\frac{1}{2})}_{1}$ and
$\widehat{\Gamma}^{(\frac{1}{2})}_{2}$
that led to the exclusion of the state
$\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{1}\, \widehat{\Gamma}^{(\frac{1}{2})}_{2}}$
from the ground-state manifold. In the proof below, we
follow this ``algebraic'' approach, identifying
all noncommuting pairs of ``spin-$\frac{1}{2}$'' string and ``spin-$\frac{1}{2}$'' membrane
operators from Eqs.~\eqref{eq: family of sigma string operators} and
\eqref{eq: family of sigma membrane operators}, respectively.
The noncommuting ``spin-$\frac{1}{2}$'' string-membrane pairs
consist of strings and membranes that are perpendicular to
one another, intersecting in a point.
The noncommuting ``spin-$\frac{1}{2}$'' membrane-membrane
pairs consist of perpendicular membranes, whose intersection is a line.
Whenever a noncommuting pair of operators acts on one
of the states in Table~\ref{table: 3D ground states},
we will show that the resulting state must be excluded from
the ground-state manifold.
(A complementary ``analytic'' proof that these states must be excluded,
along the lines of Appendix%
~\ref{appendix: Commutation between string operators and the Hamiltonian},
could also be undertaken, but we do not do this here.)
\begin{table*}[t]
\centering
\begin{tabular}{|c|c||c|c||c|c|}
\hline
State & Eigenvalues & & & &
\\
\hline
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(-,+,+,-,+,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(-,+,-,-,-,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(-,+,+,-,-,+)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,-,+,+,-,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,-,+,-,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,+,-,+,-)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,-,+,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,-,-,-,-,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,-,+,-,-,+)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,+,-,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,-,+,+,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,+,+,-,-)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,+,-,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,-,-,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,-,-,+,-)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,-,-,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,-,+,-,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,+,-,+,-,-)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(-,-,+,-,+,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}$
&
$(-,+,-,-,+,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(-,-,+,+,-,+)$
\\
\hline
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}$
&
$(+,-,-,+,-,+)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(-,+,-,+,+,-)$
&
$\ket{\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}$
&
$(+,-,-,+,+,-)$
\\
\hline
\end{tabular}
\caption{
The 24 orthogonal states
of the form (\ref{eq: Ansatz GSM})
that are excluded based on
Eqs.~\eqref{eq: 3-sigma exclusion algebra}
and
\eqref{eq: projection kills single tilde ops},
as well as the eigenvalues of these states under
the ``spin-$1$'' string and membrane operators.
The notation for states and eigenvalues
is as in Table~\ref{table: 3D ground states}.
\label{table: 3D 24 excluded states}
}
\end{table*}
We now proceed with the proof. Of key importance
is the projection operator
\begin{widetext}
\begin{equation}\label{eq: projection operator}
\begin{split}
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}\:=&\,
\mathcal{N}^{-1}_{\mathbbm{1}}\,
\ket{\mathbbm{1}}\bra{\mathbbm{1}}
+
\mathcal{N}^{-1}_{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}\bra{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}}
+
\mathcal{N}^{-1}_{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}\bra{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}}
+
\mathcal{N}^{-1}_{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}\,
\ket{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}\bra{\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}}
+
\cdots
\end{split}
\end{equation}
\end{widetext}
onto the ground state manifold, c.f.\
Eq.~\eqref{eq2Dcase: projection operator}.
Here, $\mathcal{N}^{\,}_{\widehat{\mathcal{O}}}$ is the
squared norm of the state $\ket{\widehat{\mathcal{O}}}$, and $\cdots$
is a sum over the remaining elements of the orthonormal basis of the
ground state manifold, including the states listed in
Table \ref{table: 3D ground states}. By construction,
$\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}$ leaves any state in
Table~\ref{table: 3D ground states} invariant, and, being a projector,
satisfies
\begin{align}
\widehat{\mathcal{P}}^{2}_{\rm{GSM}}=\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}.
\end{align}
Furthermore, the projector $\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}$ satisfies
\begin{align}\label{eq: projection kills O}
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}\,
\widehat{\mathcal{O}}\,
\widehat{\mathcal{P}}^{\,}_{\rm{GSM}}=0
\end{align}
for any operator $\widehat{\mathcal{O}}$ whose action on any
of the states in Table~\ref{table: 3D ground states} produces an excited state.
To prove that the $44$ states
in question must be excluded from the ground state manifold, we will
show for two particular classes of operators $\widehat{\mathcal{O}}$
that Eq.~\eqref{eq: projection kills O} holds. This will turn out to
be sufficient to exclude the offending states.
The first class of operators arises when we consider products of
perpendicular ``spin-$\frac{1}{2}$'' strings and ``spin-$\frac{1}{2}$'' membranes.
This includes the three operators
\begin{subequations}\label{eq: 3-sigma ops}
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
,\qquad
\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
,\qquad
\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
,
\label{eq: 3-sigma ops a}
\end{align}
as well as products of the form
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}\,
\widehat{\mathcal{O}},
\label{eq: 3-sigma ops b}
\end{align}
\end{subequations}
for $\hat{a}=\hat{x},\hat{y},\hat{z}$ and any string or membrane operator $\mathcal{O}$ that
commutes with $\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}$ and
$\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{a}}$.
The operators in Eq.~\eqref{eq: 3-sigma ops a} share the common trait
that the domain of intersection of the string and membrane operators
contains three $\widehat{\sigma}$ operators per chiral channel.
The exchange algebra relevant to this case was computed in
Sec.\ \ref{subsec: String operators and topological degeneracy}.
By small variations on the calculation presented in Eqs.\
\eqref{eq2Dcase: 3-sigma diagram untwisting} and
\eqref{eq2Dcase: 3-sigma algebraic untwisting},
one verifies the relations
\begin{subequations}
\label{eq: 3-sigma exclusion algebra}
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}=
e^{+\mathrm{i}\, \frac{3\pi}{4}}\,
\widehat{\widetilde{\Gamma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}},
\\
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}\,\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}=
e^{+\mathrm{i}\,\frac{3\pi}{4}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}},
\\
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}\,\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}=
e^{+\mathrm{i}\,\frac{3\pi}{4}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{y}}\,
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}},
\end{align}
\end{subequations}
where the operators
$\widehat{\widetilde{\Gamma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{i}}$
and
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{i}}$
are defined in the same way as the operator
$\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2}$
appearing in Eq.~\eqref{eq2Dcase: 3-sigma algebraic untwisting}, i.e.,
one replaces any appearance of
$\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$
in the \textit{intersection} of the string/membrane pair with
$\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\psi}$,
and leaves all other appearances of
$\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$
unchanged. Each of Eqs.~\eqref{eq: 3-sigma exclusion algebra}
carries an implicit limit $\epsilon\to0$.
Whether the tilde appears on the
string or the membrane operator above depends on which operator
contains a product of two $\widehat{\sigma}$ operators in the same
wire. By direct analogy with the two-dimensional case, we have
\begin{subequations}
\label{eq: projection kills single tilde ops}
\begin{align}
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\widetilde{\Gamma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{z}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=&\,0,
\\
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=&\, 0,\\
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{y}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=&\, 0,
\end{align}
\end{subequations}
in the limit $\epsilon\to0$.
Combining Eqs.~\eqref{eq: 3-sigma exclusion algebra}
and
\eqref{eq: projection kills single tilde ops},
one can show, in direct analogy
with Eq.~\eqref{eq2Dcase: exclude ket sigma_1,2} in the
two-dimensional case,
that
any state created by acting with any operator of the $3\times8$ operators
of the form
\eqref{eq: 3-sigma ops}
on $\ket{\mathbbm{1}}$
must be excluded from the ground-state manifold.
This is sufficient to
eliminate 24 of the 44 ``extra'' states
of the form (\ref{eq: Ansatz GSM}) that are not in the Table
\ref{table: 3D ground states}. These eliminated states are listed
in Table~\ref{table: 3D 24 excluded states}.
\begin{table*}[t]
\centering
\begin{tabular}{|c|c||c|c|}
\hline
State &
Eigenvalues &
&
\\
\hline
\hline
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
}$ &
$(-,+,-,+,+,+)$ &
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
}$ &
$(+,-,-,+,+,+)$
\\
\hline
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
}$ &
$(-,-,+,+,+,+)$ &
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
\widehat{\Gamma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
}$ &
$(-,+,-,+,-,+)$
\\
\hline
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
\widehat{\Gamma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
}$ &
$(+,-,-,-,+,+)$ &
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
\widehat{\Gamma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
}$ &
$(-,-,+,+,+,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{x}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{y}}
\widehat{\Sigma}^{\raisebox{-6pt}{\scriptsize$\sigma$}}_{\hat{z}}
}$ &
$(-,-,-,+,+,+)$ &
&
\\
\hline
\end{tabular}
\caption{
The seven orthogonal states
of the form (\ref{eq: Ansatz GSM})
that are excluded based on
Eqs.~\eqref{eq: Sigma_z Sigma_x,y algebra},
\eqref{eq: Sigma_x Sigma_y algebra},
\eqref{eq: projection kills 2-membrane states 1},
and
\eqref{eq: projection kills 2-membrane states 2},
as well as the eigenvalues of these states under
the ``spin-$1$'' string and membrane operators.
The notation for states and eigenvalues
is as in Table~\ref{table: 3D ground states}.
\label{table: 3D 7 excluded states}
}
\end{table*}
The second class of operators arises when we consider products of two
perpendicular membranes, e.g.,
\begin{subequations}
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}},\quad
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}},\quad
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\,\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}},
\end{align}
as well as products of the form
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{b}}\,
\widehat{\mathcal O},
\end{align}
\end{subequations}
where $\hat{a},\hat{b}=\hat{x},\hat{y},\hat{z}$, for $\hat{a}\neq \hat{b}$
and any operator $\widehat{\mathcal O}$ that
commutes with
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}$
and
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{b}}$.
It turns out that the operators
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$
and
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$
can also be handled using minor
variations on the calculation presented in Eqs.\
\eqref{eq2Dcase: 3-sigma diagram untwisting} and
\eqref{eq2Dcase: 3-sigma algebraic untwisting} of Sec.\
\ref{subsec: String operators and topological degeneracy}.
Specifically, one can show the relations
\begin{subequations}
\label{eq: Sigma_z Sigma_x,y algebra}
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{a}}=&\,
e^{+\mathrm{i}\,N^{\,}_{\hat{a}^{\,}_{\perp}}\frac{3\pi}{4}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{a}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}},
\end{align}
\end{subequations}
in the limit $\epsilon\to0$,
for $\hat{a}=\hat{x},\hat{y}$, and for
$\hat{x}^{\,}_{\perp}=\hat{y}$ and $\hat{y}^{\,}_{\perp}=\hat{x}$.
The system-size-dependent integers $N^{\,}_{\hat{x}}$ and
$N^{\,}_{\hat{y}}$ are the number of wires contained in the path
$\mathcal{P}^{\,}_{\hat{x}}$ and $\mathcal{P}^{\,}_{\hat{y}}$,
respectively. The operators
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{x}}$
and
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{y}}$
are again defined by direct analogy with the operator
$\widehat{\widetilde\Gamma}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{2}$
appearing in Eq.~\eqref{eq2Dcase: 3-sigma algebraic untwisting}, i.e.,
by replacing \textit{any} appearance of
$\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$
with
$\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\psi}$.
The reason for using the primes here is to distinguish these operators from
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}}$
and
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{y}}$,
where only \textit{some} of
the appearances of
$\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}$
are replaced by
$\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(0)\,
\widehat{\sigma}_{\gamma,\mathrm{M},J,p^{\,}_{K}}(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\psi}$.
Regardless of these slight differences in definition, the operators
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{x}}$
and
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{y}}$
create excited states when acting on the vacuum
$\ket{\Omega}$. Consequently, we have
\begin{subequations}
\label{eq: projection kills 2-membrane states 1}
\begin{align}
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})\prime$}}_{\hat{a}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=&\,0,
\end{align}
\end{subequations}
in the limit $\epsilon\to0$
for $\hat{a}=\hat{x},\hat{y}$.
The operator
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$, which involves four
$\widehat{\sigma}$ operators in the same chiral channel contained in the
intersection of the two membranes, can be treated similarly. The
exchange algebra of the membrane operators
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}$
and
$\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}$
can be determined by considering the diagram
\begin{align}
\includegraphics[width=.1325\textwidth]{Figures12},
\end{align}
which obeys
\begin{align}\label{eq: 4-sigma diagram untwisting}
\includegraphics[width=.375\textwidth]{Figures13}.
\end{align}
The algebraic interpretation of this diagrammatic statement is
\begin{align}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}\,
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}=
e^{+\mathrm{i}\,\frac{3\pi}{4}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{y}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}},
\label{eq: Sigma_x Sigma_y algebra}
\end{align}
in the limit $\epsilon\to0$,
where the operators
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{y}}$
and
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}}$
also appear in Eqs.\ \eqref{eq: 3-sigma exclusion algebra}.
Explicitly, we have [compare Eq.~\eqref{eq: az sigma membrane def}]
\begin{widetext}
\begin{subequations}\label{eq: tilde Sigma^sigma_y def}
\begin{align}
\begin{split}
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{a}}\:=&\,
\(
\prod_{
(
\gamma,J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}
\!\!\!\setminus(\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}\!\!\!\cap\mathcal{P}^{\,}_{\hat{a}}
)
}
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}(t,z)
\Bigg)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\)
\\
&\qquad\qquad
\times
\prod_{(\gamma,J,p^{\,}_{K}) \in \mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}\!\!\!\cap \mathcal{P}^{\,}_{\hat{a}}}
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}(t,z)
\Bigg)\,
\widehat{\mathcal{P}}^{\,}_{\psi}\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\psi}
\end{split}
\\
\begin{split}
\sim&\,
\(
\prod_{
(
\gamma,J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}
\!\!\!\setminus(\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}\!\!\!\cap
\mathcal{P}^{\,}_{\hat{a}})
}
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}(t,z)
\Bigg)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)\,
\widehat{\sigma}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(\epsilon)\,
\widehat{\mathcal{P}}^{\,}_{\mathbbm{1}}
\)
\\
&\qquad\qquad
\times
\prod_{(\gamma,J,p^{\,}_{K})\in\mathcal{P}^{\,}_{\hat{a}^{\,}_{\perp}}\!\!\!\cap\mathcal{P}^{\,}_{\hat{a}}}
\exp
\Bigg(
-
\mathrm{i}
\frac{1}{2\sqrt{2}}
\int\limits^{L^{\,}_{z}}_{0}\mathrm{d}z\,
\partial^{\,}_{z}\widehat{\phi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}(t,z)
\Bigg)\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)
+
\cdots,
\end{split}
\end{align}
\end{subequations}
\end{widetext}
for $\hat{a}=\hat{x},\hat{y}$ and for infinitesimal $\epsilon>0$
(recall that $\hat{x}^{\,}_{\perp}=\hat{y}$ and $\hat{y}^{\,}_{\perp}=\hat{x}$),
where in the second
line we have performed the OPE in the wires where the two membranes
intersect.
Thus, the chiral
channel in which the two membranes intersect [see Fig.\
\ref{fig: P_x and P_y defs}(b)] contains a pair of fermion excitations due to
the product
$\widehat{\psi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)$.
Note that both fermion operators are evaluated at $z=0$ purely
due to an (arbitrary) choice made in the definitions
\eqref{eq: az sigma membrane def}
of the ``spin-$\frac{1}{2}$'' membrane operators.
All of our algebraic results would still
hold if we had replaced
$z=0\to z=z^{\,}_{\hat{x}}$ and $z=\epsilon\to z=z^{\,}_{\hat{x}}+\epsilon$
the definition of
$\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\frac{1}{2})$}}_{\hat{x}}$,
say.
Hence, the two fermion excitations in the product
$\widehat{\psi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)\,
\widehat{\psi}^{\,}_{\gamma,\mathrm{R},J,p^{\,}_{K}}
(0)$
can actually be separated by arbitrarily large distances along the $z$-axis.
Consequently, we have
\begin{align}
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{x}}\,
\widehat{\Sigma}^{(\!\frac{1}{2}\!)}_{\hat{y}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=
e^{+\mathrm{i}\,\frac{3\pi}{4}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\!\frac{1}{2}\!)$}}_{\hat{y}}\,
\widehat{\widetilde{\Sigma}}^{\raisebox{-6pt}{\scriptsize$(\!\frac{1}{2}\!)$}}_{\hat{x}}\,
\widehat{\mathcal{P}}^{\,}_{\mathrm{GSM}}=0,
\label{eq: projection kills 2-membrane states 2}
\end{align}
in the limit $\epsilon\to0$.
Using Eqs.\ \eqref{eq: Sigma_z Sigma_x,y algebra},
\eqref{eq: Sigma_x Sigma_y algebra},
\eqref{eq: projection kills 2-membrane states 1},
and
\eqref{eq: projection kills 2-membrane states 2},
one can show that the seven states listed in
Table~\ref{table: 3D 7 excluded states} are eliminated
from the ground-state manifold.
\begin{table*}[t]
\centering
\begin{tabular}{|c|c||c|c|}
\hline
State
&
Eigenvalues
&
&
\\
\hline
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
}$ &
$(-,-,-,-,+,+)$ &
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
}$ &
$(-,-,-,+,-,+)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
}$ &
$(-,-,+,-,-,+)$ &
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,+,-,-,+,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,-,+,-,-,-)$ &
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,+,-,-,-,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
}$ &
$(-,-,-,-,-,+)$ &
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,-,-,-,+,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,-,-,-,-,-)$
&
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,-,-,+,+,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(+,-,-,+,-,-)$
&
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(+,-,-,-,-,-)$
\\
\hline
$\ket{
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{x}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Sigma}^{(\frac{1}{2})}_{\hat{z}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{y}}
\widehat{\Gamma}^{(\frac{1}{2})}_{\hat{z}}
}$ &
$(-,-,-,+,-,-)$
&
&
\\
\hline
\end{tabular}
\caption{
The 13 orthogonal states
of the form (\ref{eq: Ansatz GSM})
that are excluded based on
appropriate combinations of Eqs.\
\eqref{eq: 3-sigma exclusion algebra},
\eqref{eq: projection kills single tilde ops},
\eqref{eq: Sigma_z Sigma_x,y algebra},
\eqref{eq: Sigma_x Sigma_y algebra},
\eqref{eq: projection kills 2-membrane states 1},
and \eqref{eq: projection kills 2-membrane states 2},
as well as the eigenvalues of these states under
the ``spin-$1$'' string and membrane operators.
The notation for states and eigenvalues
is as in Table~\ref{table: 3D ground states}.
\label{table: 3D 13 excluded states}
}
\end{table*}
Finally, the $44-24-7=13$ remaining ``extra'' states
of the form (\ref{eq: Ansatz GSM})
can also be eliminated
using appropriate combinations of Eqs.\
\eqref{eq: 3-sigma exclusion algebra},
\eqref{eq: projection kills single tilde ops},
\eqref{eq: Sigma_z Sigma_x,y algebra},
\eqref{eq: Sigma_x Sigma_y algebra},
\eqref{eq: projection kills 2-membrane states 1},
and \eqref{eq: projection kills 2-membrane states 2}.
These states are listed in Table~\ref{table: 3D 13 excluded states}.
In all cases, the reason for elimination is the same.
Each state is created by acting on one of the states in
Table~\ref{table: 3D ground states}
with an operator that creates an excess of
fermion excitations.
To summarize, we have shown that of the $2^{6}=64$ states
labeled by the eigenvalues of the ``spin-$1$'' string or
``spin-$1$'' membrane operators, only the $20$ listed in
Table~\ref{table: 3D ground states} truly reside in the ground-state
manifold once the exchange algebra of the ``spin-$\frac{1}{2}$'' string and
``spin-$\frac{1}{2}$'' membrane operators is taken into account.
This exchange algebra is highly nontrivial,
because reordering a product of ``spin-$\frac{1}{2}$'' string and/or ``spin-$\frac{1}{2}$'' membrane
operators not only produces simple multiplicative phase factors,
but enacts nontrivial unitary operations within the space spanned by the
operator products. As in the two-dimensional case discussed in Sec.\
\ref{subsec: String operators and topological degeneracy},
this reduction of the number
of states in the ground-state manifold from the naive value lies at
the heart of the distinction between Abelian and non-Abelian
topological states of matter.
\section{Surface theory of the 3D non-Abelian phase}
\label{sec: Surface theory}
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{Figures14}
\caption{
(Color online)
Surface of the coupled-wire theory with the plaquette interaction
(\ref{eq: plaquette interaction})
for the case of mixed periodic and open boundary conditions.
The graphical notation of Fig.~\ref{fig: Schematic of the su(2)_{k} couplings}
is used here.
When open boundary conditions are imposed along the $x$-direction,
while periodic boundary conditions are imposed along
the $y$- and $z$-directions, the plaquette interaction
(\ref{eq: plaquette interaction})
leaves gapless modes on the surfaces at
$x=0$ and $x=L^{\,}_{x}$.
The purple bonds connecting chiral modes on the surface represent
the interactions \eqref{eq: non-abelian surface current current interactions}
in the $SU(2)$-symmetric limit
\eqref{eq: SU(2) isotropic limit for Sbs x=0}.
When $\lambda=\lambda^{\prime}$, the symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ is present.
}
\label{fig: non-abelian surface states}
\end{figure}
Let us now remove the periodic boundary conditions imposed in the
previous section and replace them with boundary conditions that are
open along the $x$-direction and periodic along the $y$-direction.
The bulk of the coupled-wire theory in (3+1)-dimensional spacetime
then possesses the ``time-reversal'' symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ defined in
Eq.~\eqref{eq: T_eff definitions b}.
[The operation $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}}}$ defined in
Eq.~\eqref{eq: T_eff definitions a} is not a symmetry anymore
under open boundary conditions along the $x$-direction.]
In this case, the plaquette interaction
(\ref{eq: plaquette interaction})
is then not sufficient to gap all gapless $su(2)^{\,}_{k}$ modes;
even if this interaction succeeds in gapping the bulk,
there must remain gapless modes that are confined to the surfaces
at $x=0$ and $x=L^{\,}_{x}$.
In this section, we investigate the fate of these gapless surface
modes when they are coupled by marginally relevant
current-current interactions.
From now on, we shall only consider the surface at $x=0$.
The surface at $x=0$ supports gapless modes that
can be represented by a quadratic form for the currents that generate
the copy $\gamma=1$ of the $su(2)^{\,}_{k}$ affine Lie algebra
with the M-moving currents $J^{a}_{\gamma=1,\mathrm{M},y}$
(the label $\gamma=2$ applies to the surface $x=L^{\,}_{x}$,
see Fig.\ \ref{fig: non-abelian surface states}).
From now on, we will drop the explicit reference to the
label $\gamma=1$. Hence, the Hamiltonian density for the gapless
modes on the surface at $x=0$ is the linear combination
\begin{subequations}
\label{eq: def decoupled surface non Abelian theory}
\begin{equation}
\mathcal{H}^{\,}_{x=0}\:=
2\pi
\left(
T^{\,}_{\mathrm{L},A,x=0}[su(2)^{\,}_{k}]
+
T^{\,}_{\mathrm{R},B,x=0}[su(2)^{\,}_{k}]
\right),
\end{equation}
where
\begin{equation}
T^{\,}_{\mathrm{M},J,x=0}[su(2)^{\,}_{k}]=
\frac{1}{2+k}
\prisum{L^{\,}_{y}}{y=0}\,
\sum^{3}_{a=1}
J^{a}_{\mathrm{M},J,y}\,
J^{a}_{\mathrm{M},J,y}
\end{equation}
\end{subequations}
is the energy-momentum tensor for the M-moving
mode on sublattice $J=A,B$.
Here, the priming of the sum over $y$
indicates that only \textit{even} wires are to be summed over.
Summing only over even $y$ and over $J=A,B$ accounts for
the ``dangling'' gapless modes on the $x=0$ surface that do
not couple to any neighbors via the couplings depicted in Fig.\
\ref{fig: Schematic of the su(2)_{k} couplings} when open
boundary conditions are imposed in the $x$-direction.
We assume that $L^{\,}_{y}$ is odd, so that the total number
of wires (i.e., $L^{\,}_{y}+1$) is even.
The surface theory at $x=0$, whose energy-momentum
tensor has the chiral components
(\ref{eq: def decoupled surface non Abelian theory}),
can be viewed as a conformal field theory in $(1+1)$-dimensional spacetime
with an extensive central charge.
We would like to decrease this central charge to a finite number
in the thermodynamic limit ($L^{\,}_{y}\to\infty$).
To this end, we perturb the gapless theory
(\ref{eq: def decoupled surface non Abelian theory})
with the interactions
\begin{subequations}
\begin{equation}\label{eq: non-abelian surface current current interactions}
\mathcal{L}^{\,}_{\mathrm{bs},x=0}\:=
\!-\!
\prisum{L^{\,}_{y}}{y=0}\,\!
\sum^{3}_{a=1}\!
\(
\lambda^{a}
J^{a}_{\mathrm{R},y}\,J^{a}_{\mathrm{L},y+1}
\!\!+\!
\lambda^{\prime\,a}
J^{a}_{\mathrm{L},y+1}\,J^{a}_{\mathrm{R},y+2}
\)\!.
\end{equation}
To investigate the nature of the surface
more closely, we allow the
possibility that this surface interaction breaks explicitly
the $SU(2)$ symmetry. The choices
\begin{equation}\label{eq: SU(2) isotropic limit for Sbs x=0}
\lambda\equiv\lambda^{a},
\qquad
\lambda^{\prime}\equiv\lambda^{\prime\,a},
\qquad
a=1,2,3,
\end{equation}
\end{subequations}
restore the explicit $SU(2)$ symmetry. These
couplings are depicted in Fig.~\ref{fig: non-abelian surface states}.
For the isotropic point (\ref{eq: SU(2) isotropic limit for Sbs x=0}),
it is readily shown that there are two gapped phases, one for
$\lambda>\lambda^{\prime}\geq0$ and one for $0\leq\lambda<\lambda^{\prime}$,
that are related to one another by the ``time-reversal'' symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ defined in Eq.\
\eqref{eq: T_eff definitions b}.
Indeed, when $\lambda^{\prime}=0$ and $\lambda>0$ (or vice versa) the
interactions
\eqref{eq: non-abelian surface current current interactions}
are marginally relevant, flowing to strong coupling and opening a gap,
as they do in the bulk.
Furthermore, if we define a ``magnetic'' domain wall at $y=0$
by allowing $\lambda$ and $\lambda'$ to acquire the $y$-dependent
profiles
\begin{subequations}
\begin{equation}
\lambda^{\,}_{y}=
\begin{cases}
\lambda^{\,}_{-\infty}>0,
&
\hbox{ if $y<0$,}
\\
0,
&
\hbox{ otherwise,}
\end{cases}
\end{equation}
and
\begin{equation}
\lambda^{\prime}_{y}=
\begin{cases}
0,
&
\hbox{ if $y<0$,}
\\
\lambda^{\prime}_{-\infty}>0,
&
\hbox{ otherwise,}
\end{cases}
\end{equation}
\end{subequations}
respectively, one finds that a single chiral $su(2)^{\,}_{k}$
current is localized at the domain wall. This is reminiscent
of the surface physics of the usual 3D TI, where a domain
wall between different TRS-broken regions on the surface binds
a chiral $u(1)^{\,}_{1}$ current (i.e., a chiral Dirac fermion mode).
In the present setting, the role of TRS is played by the non-onsite
symmetry operation $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$.
This non-onsite implementation of an effective TRS is
common to other coupled-wire construction.
For example, counterparts to such an effective nonlocal TRS can be found
in the coupled-wire constructions presented in
Refs.\ \cite{Mross15} and \cite{Sahoo15}.
The remainder of this section is devoted to elucidating the nature
of the surface theory for the $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-symmetric
(but not necessarily $SU(2)$-symmetric) case
\begin{align}
\lambda^a=\lambda^{\prime a},\indent a=1,2,3.
\end{align}
First, we present a one-loop renormalization group (RG) analysis,
valid for small magnitudes of $\lambda^{a}$ and $\lambda^{\prime\,a}$
with $a=1,2,3$. This analysis sheds light on the phase diagram of the
surface, particularly on the response of the surface theory to
$SU(2)$-breaking perturbations. Second, we present a mean-field
analysis of the surface theory for the
case $k=2$. This analysis demonstrates that the point
$\lambda=\lambda^{\prime}>0$,
a strongly interacting quantum field theory
when expressed in terms of the original fermionic modes,
is a continuous quantum critical point
that can be described by two noninteracting modes.
The first mode is a gapless complex-valued fermion realizing a single Dirac
cone in the low-energy limit.
The second mode is a gapless real-valued fermion realizing a single
Majorana cone in the low-energy limit. These low-energy modes
are surface states of the $su(2)^{\,}_{2}$ coupled-wire theory
that are protected by the symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$.
\subsection{One-loop RG analysis}
\label{subsec: One-loop RG analysis}
We now perform a one-loop RG analysis of the
surface interaction
\eqref{eq: non-abelian surface current current interactions}
in the presence of both $\lambda^{a}$ and $\lambda^{\prime\,a}$
with $a=1,2,3$ under the assumption that these couplings are small.
Hence, the bare surface interaction
\eqref{eq: non-abelian surface current current interactions}
is a small perturbation to the critical surface theory with
the energy-momentum tensor
(\ref{eq: def decoupled surface non Abelian theory}).
The RG calculation itself is standard, and makes use of the
current-current OPEs~\eqref{eq: current-current OPEs}
(see, e.g.,~\cite{Cardy96}). The resulting RG
equations describing the flow of the couplings $\lambda^{a}$
and $\lambda^{\prime a}$ as functions of the cutoff length scale
$\ell$ are
\begin{subequations}\label{eq: renormalization group equations}
\begin{align}
\frac{\mathrm{d}\lambda^{a}}{\mathrm{d}\ell}&=
+2\pi\,
\lambda^{b}\,\lambda^{c},\\
\frac{\mathrm{d}\lambda^{\prime\,a}}{\mathrm{d}\ell}&=
+2\pi\,
\lambda^{\prime\,b}\,\lambda^{\prime\,c},
\end{align}
\end{subequations}
for $1\leq a<b<c\leq3$ and cyclic permutations thereof.
Note that at the $SU(2)$-symmetric
point~\eqref{eq: SU(2) isotropic limit for Sbs x=0}, the
RG flows \eqref{eq: renormalization group equations}
indicate that the couplings $\lambda$ and $\lambda^{\prime}$
are marginally relevant, as is the case in the bulk.
The one-loop renormalization-group flows for the triplet
$\lambda^{1}$\, $\lambda^{2}$, and $\lambda^{3}$ have decoupled from those of
the triplet $\lambda^{\prime\,1}$, $\lambda^{\prime\,2}$, and
$\lambda^{\prime\,3}$.
In fact, they are identical at the one-loop level.
We expect this to be true to all orders in perturbation theory,
since the interaction
\eqref{eq: non-abelian surface current current interactions}
is form-invariant under composing the transformation
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$
with the interchange of $\lambda^{a}$ and $\lambda^{\prime a}$.
Thus, if the flow starts from an initial condition such
that $\lambda^{a}=\lambda^{\prime a}$ for all $a=1,2,3$
(as must be the case for a surface that does not explicitly
break the symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$), the
asymmetry $\lambda-\lambda^{\prime}=0$ for all $\ell>0$.
If initial conditions are chosen such that $\lambda^{a},\lambda^{\prime a}>0$
for all (or even, as we shall see below, for only some) $a=1,2,3$,
then all couplings $\lambda^{a}$ and $\lambda^{\prime a}$
flow to infinity.
For initial conditions that do not respect the symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$,
we then expect this strong-coupling
fixed point to break the symmetry as well.
We now focus on the $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-symmetric case
($\lambda^{a}=\lambda^{\prime a}$ for all $a=1,2,3$)
and investigate the fate of $SU(2)$ symmetry under the RG flows
\eqref{eq: renormalization group equations}. By analyzing vector-field
plots for the differential equations
\eqref{eq: renormalization group equations},
one can convince oneself that the strong-coupling fixed point reached
from initial conditions $\lambda^{a}>0$ for $a=1,2,3$ is in fact
$SU(2)$-symmetric. Thus, even if the initial conditions do not satisfy
the conditions
\eqref{eq: SU(2) isotropic limit for Sbs x=0}, the strong-coupling
fixed point does.
We will illustrate this below for the $U(1)$-symmetric case,
which is easier to visualize as the phase diagram is then two-
rather than three-dimensional.
\begin{figure}[t]
\includegraphics[width=0.4\textwidth]{Figures15}\\
\caption{
One-loop renormalization group flows
(\ref{eq: KT renormalization group equations}).
}
\label{fig: KT renormalization group equations}
\end{figure}
Let us analyze in greater detail the $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-
and $U(1)$-symmetric case
\begin{subequations}
\begin{equation}
\lambda^{a}=\lambda^{\prime\,a},
\qquad
\lambda^{1}=\lambda^{2}\equiv\lambda^{\,}_{\perp},
\qquad
\lambda^{3}\equiv\lambda^{\,}_{\parallel},
\end{equation}
for which the one-loop renormalization-group flows
(\ref{eq: renormalization group equations}) simplify to
\begin{equation}\label{eq: KT renormalization group equations}
\frac{\mathrm{d}X}{\mathrm{d}\ell}=
+2\pi\,
Y^{2},
\qquad
\frac{\mathrm{d}Y}{\mathrm{d}\ell}=
+2\pi\,
X\,Y,
\end{equation}
\end{subequations}
where either $(X,Y)=(\lambda^{\,}_{\parallel},\lambda^{\,}_{\perp})$
or $(X,Y)=(\lambda^{\prime}_{\parallel},\lambda^{\prime}_{\perp})$.
(Recall that the RG flows of $\lambda^{a}$ and $\lambda^{\prime\, a}$
are decoupled.)
These one-loop renormalization-group flows are shown in Fig.\
\ref{fig: KT renormalization group equations}.
The separatrix $X^{2}-Y^{2}=0$ is typical of the Kosterlitz-Thouless
renormalization group flows.
The RG flow diagram in Fig.~\ref{fig: KT renormalization group equations}
indicates that the fixed points for the interacting surface modes
when $\lambda^{a}=\lambda^{\prime\,a}$
are (i) the $SU(2)$-symmetric strong-coupling fixed point
\begin{equation}
\lambda^{a}=\lambda^{\prime\,a}\equiv\lambda\to\infty,
\end{equation}
(ii) the strong-coupling fixed point
\begin{equation}
-\lambda^{1}=-\lambda^{\prime\,1}=-\lambda^{2}=-\lambda^{\prime\,2}=
\lambda^{3}=\lambda^{\prime\,3}
\equiv\lambda\to\infty,
\end{equation}
that follows from performing a global $SU(2)$ rotation by $\pi/2$
about the quantization axis, and (iii) the line of fixed points
\begin{equation}
\lambda^{1}=\lambda^{\prime\,1}=\lambda^{2}=\lambda^{\prime\,2}=0,
\qquad
\lambda^{3}=\lambda^{\prime\,3}
\equiv\lambda^{\,}_{\parallel}<0.
\end{equation}
Case (i) [and, upon making a global $SU(2)$ rotation, case (ii)] is the
$SU(2)$-isotropic strong-coupling fixed point discussed earlier. Cases
(i) and (ii) dominate the phase diagram in the sense that any initial conditions
in three of the four quadrants of the $X$-$Y$ plane will lead to one of those
strong-coupling fixed points.
The line of stable
fixed points in case (iii) constitutes what we call a ``sliding parafermion
liquid'' (SPL).
This set of fixed points is gapless--even if the wires are initially
coupled with some finite values of $\lambda^{\,}_{\perp}$ and
$\lambda^{\,}_{\parallel}$, the wires decouple as the theory flows to
the infrared.
(The use of the term ``parafermion'' refers to the fact that the
decoupled chiral $su(2)^{\,}_{k}$ CFTs contain parafermion degrees of
freedom.) Thus, the SPL resembles the so-called
``sliding Luttinger liquids,'' which are another class of non-Fermi
liquid in (2+1)
dimensions~\cite{Ohern99,Emery00,Sondhi01,Mukhopadhyay01,Vishwanath01}.
These sliding phases have in common the fact that
certain classes of perturbations are either irrelevant or marginally
irrelevant and hence flow to zero in the infrared.
It would be interesting
to investigate the SPL phase in more detail, but at present such a study
is beyond the scope of this work.
In summary, we have shown that, for a variety of initial conditions on the
couplings $\lambda^{a}=\lambda^{\prime a}$, the interactions
\eqref{eq: non-abelian surface current current interactions} lead to an
$SU(2)$-symmetric strong-coupling RG fixed point, even if the
interaction itself breaks $SU(2)$ symmetry explicitly. The
$SU(2)$-broken fixed points (like the SPL) may constitute
interesting strongly-correlated gapless phases (i.e.,
non-Fermi liquids).
\subsection{Mean-field theory for $k=2$}
\label{subsec: Mean-field theory for k=2}
Having established the stability of the $SU(2)$-symmetric strong-coupling
fixed point, we now wish to investigate the nature of this fixed point.
Specifically, we would like to know whether this fixed point is gapped or
gapless when
the symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$
is not explicitly broken.
This is equivalent to asking whether the phase transition
between the two
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-conjugate gapped phases
$\lambda>\lambda^{\prime}\geq0$ and $0\leq\lambda<\lambda^{\prime}$ is
discontinuous or continuous.
Moreover, we would like to determine whether
the symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ might be spontaneously
broken by the interacting surface theory.
For general $k=3,4,\dots$,
the answers to these questions are difficult to determine.
Rewriting the interaction
\eqref{eq: non-abelian surface current current interactions}
at the $SU(2)$-symmetric point \eqref{eq: SU(2) isotropic limit for Sbs x=0}
in terms of the parafermion representation
\eqref{eq: def chiral parafermion and boson rep su(2)k} recasts it as a
correlated hopping process like
\eqref{eq2Dcase: parafermion representation su(2)_k}.
However, unlike in the $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-breaking
case studied in
Sec.~\ref{sec: Warm-up: Non-Abelian topological order in two dimensions},
the current-current interactions on neighboring bonds do not
commute for general $k$, owing to the presence of the nonzero couplings
$\lambda^{\prime}=\lambda$. Furthermore, since not all $su(2)^{\,}_{k}$
CFTs admit a free-field description, performing detailed calculations
is intractable in general. (Although it may be possible to make progress
using certain methods from the theory of integrable systems,
like the thermodynamic Bethe ansatz, which has been used to study
perturbed $su(2)^{\,}_{k}$ and $\mathbb{Z}^{\,}_{k}$
CFTs~\cite{Fateev91,Tsvelik14}.)
However, the case $k=2$ is special, for it
is the simplest nontrivial example in which the $su(2)^{\,}_{k}$
current algebra has a free-fermion representation.
In the $k=2$ case, we can rewrite the $su(2)^{\,}_{2}$ current operators as
[see also Eqs.%
~\eqref{eq: def chiral parafermion and boson rep su(2)k}
for $k=2$]
\begin{subequations}
\label{eq: k=2 currents}
\begin{align}
&
\widehat{J}^{+}_{\mathrm{M},y}=
\sqrt{2}\,
\widehat{\psi}^{\,}_{\mathrm{M},y}\,
\widehat{\xi}^{\,}_{\mathrm{M},y},
\\
&
\widehat{J}^{-}_{\mathrm{M},y}=
\sqrt{2}\,
\widehat{\xi}^{\ \dag}_{\mathrm{M},y}\,
\widehat{\psi}^{\,}_{\mathrm{M},y},
\\
&
\widehat{J}^{3}_{\mathrm{M},y}=
\mathrm{i}\,
\frac{1}{\sqrt{2}}\,
\partial^{\,}_{\mathrm{M}}\,\widehat{\phi}^{\,}_{\mathrm{M},y}.
\end{align}
On the one hand, the operator
$\widehat{\psi}^{\,}_{\mathrm{M},y}$
either creates or annihilates an M-moving Majorana mode
with $\mathrm{M}=\mathrm{L},\mathrm{R}$
standing for left and right, respectively.
On the other hand, the creation operators
$\widehat{\xi}^{\ \dag}_{\mathrm{M},y}$
and
annihilation operators
$\widehat{\xi}^{\,}_{\mathrm{M},y}$
create and annihilate M-moving complex Dirac modes,
respectively.
Moreover, they are related to the
chiral boson operators $\widehat{\phi}^{\,}_{\mathrm{M},y}$
through the vertex operator
\begin{align}
\widehat{\xi}^{\,}_{\mathrm{M},y}\=:
\bm{:}
e^{+\mathrm{i}\,\sqrt{1/2}\,\widehat{\phi}^{\,}_{\mathrm{M},y}}
\bm{:},
\end{align}
where $\mathrm{M}=\mathrm{L},\mathrm{R}$.
Finally,
the chiral currents $\widehat{J}^{3}_{\mathrm{M},y}$ can be reexpressed in terms
of the pair of Dirac fermion operators
$\widehat{\xi}^{\dag}_{\mathrm{M},y}$
and
$\widehat{\xi}^{\,}_{\mathrm{M},y}$
using the bosonization identity
\begin{align}
\begin{split}
\widehat{\rho}^{\,}_{\mathrm{M},y}\:=&\,
\widehat{\xi}^{\dag}_{\mathrm{M},y}\,
\widehat{\xi}^{\,}_{\mathrm{M},y}
\\
\equiv&\,
-
\frac{1}{2\pi\sqrt{2}}\,
\partial^{\,}_{\mathrm{M}}
\widehat{\phi}^{\,}_{\mathrm{M},y},
\end{split}
\end{align}
\end{subequations}
where $\mathrm{M}=\mathrm{L},\mathrm{R}$ and
$(-1)^{\mathrm{R}}=-(-1)^{\mathrm{L}}\equiv1$,
that defines the two chiral fermionic densities.
Note that the wire-dependence of the chiral bosonic commutation relations
\eqref{eq: def chiral parafermion and boson rep su(2)k d}
was chosen so as to ensure that fermionic vertex operators
in different wires anticommute at equal times.
The same is true for the Majorana operators,
owing to Eq.\
\eqref{eq: def chiral parafermion and boson rep su(2)k c}
with $k=2$.
The representation (\ref{eq: k=2 currents})
is a free-fermion representation because it can
be used to rewrite the kinetic contribution
\eqref{eq: def decoupled surface non Abelian theory}
in terms of two decoupled sectors of noninteracting
modes,
one for the Majorana modes
and one for the Dirac modes.
This is to say that one can use
Grassmann coherent states
to represent the partition function associated with the Hamiltonian
\eqref{eq: def decoupled surface non Abelian theory}
as a path integral over Grassmann variables with the free Lagrangian
density~\cite{Fradkin13}
\begin{align}\label{eq: free fermion re kinetic surface contribution}
\begin{split}
\mathcal L^{\,}_{0,x=0}\:=&\,
\prisum{L^{\,}_{y}}{y=0}\,
2
\(
\psi^{\,}_{\mathrm{L},y+1}\,
\mathrm{i}\partial^{\,}_{\mathrm{L}}
\psi^{\,}_{\mathrm{L},y+1}
+
\psi^{\,}_{\mathrm{R},y}\,
\mathrm{i}\partial^{\,}_{\mathrm{R}}
\psi^{\,}_{\mathrm{R},y}
\)
\\
&\,
+
\prisum{L^{\,}_{y}}{y=0}\,
2
\(
\xi^{*}_{\mathrm{L},y+1}\,\mathrm{i}\partial^{\,}_{\mathrm{L}}\xi^{\,}_{\mathrm{L},y+1}
+
\xi^{*}_{\mathrm{R},y}\,\mathrm{i}\partial^{\,}_{\mathrm{R}}\xi^{\,}_{\mathrm{R},y}
\).
\end{split}
\end{align}
Here, we have set the velocities in the $z$-direction to unity.
Passing to a Lagrangian formulation of the problem provides an enormous
simplification relative to the case of general $k>2$, where there is no such
formulation.
With the free-fermion representation \eqref{eq: k=2 currents} in hand,
we can now embark on a ``traditional'' mean-field analysis of the
interacting problem in which neighboring wires are coupled via the
current-current interactions. This will allow us to address the
question of whether the surface theory is truly critical, as well as
that of whether the interacting surface breaks spontaneously the
symmetry $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$.
In the fermionic representation
\eqref{eq: free fermion re kinetic surface contribution},
the current-current
interaction \eqref{eq: non-abelian surface current current interactions}
between nearest-neighbor wires on the surface takes the form
\begin{widetext}
\begin{align}
\mathcal{L}^{\,}_{\mathrm{bs},x=0}
=&\,
-
\lambda\,
\prisum{L^{\,}_{y}}{y=0}\,
\Big[
\(
\xi^{*}_{\mathrm{L},y+1}\,
\xi^{\,}_{\mathrm{R},y}\,
\psi^{\,}_{\mathrm{L},y+1}\,
\psi^{\,}_{\mathrm{R},y}
+
\xi^{*}_{\mathrm{R},y}\,
\xi^{\,}_{\mathrm{L},y+1}\,
\psi^{\,}_{\mathrm{R},y}\,
\psi^{\,}_{\mathrm{L},y+1}
\)
-
(2\pi)^{2}\,
\xi^{*}_{\mathrm{L},y+1}\,
\xi^{\,}_{\mathrm{R},y}\,
\xi^{*}_{\mathrm{R},y}\,
\xi^{\,}_{\mathrm{L},y+1}
\Big]
\nonumber\\
&\,
+
\(
\lambda\to\lambda',\ \mathrm{R}\leftrightarrow\mathrm{L},\ y\to y+1
\),
\label{eq: k=2 interaction before decoupling}
\end{align}
\end{widetext}
where we have set $\lambda^{\,}_{\perp}=\lambda^{\,}_{\parallel}\equiv\lambda$
and $\lambda^{\prime}_{\perp}=\lambda^{\prime}_{\parallel}\equiv\lambda^{\prime}$
as we are considering the $SU(2)$-symmetric limit.
Next, we decouple this interaction with a Hubbard-Stratonovich transformation.
That is to say, for each directed bond $\langle y,y+1\rangle$,
we introduce the complex-valued auxiliary field
$\Delta^{\,}_{\xi,y}(t,z)$,
together with the real-valued auxiliary field
$\Delta^{\,}_{\psi,y}(t,z)$.
Similarly,
for each directed bond $\langle y+1,y+2\rangle$,
we introduce the complex-valued auxiliary field
$\Delta^{\prime}_{\xi,y}(t,z)$,
together with the real-valued auxiliary field
$\Delta^{\prime}_{\psi,y}(t,z)$.
We then introduce the auxiliary Lagrangian density
\begin{widetext}
\begin{align}
\mathcal{L}^{\mathrm{aux}}_{\mathrm{bs},x=0}\:=&\,
\frac{1}{\lambda}\,
\prisum{L^{\,}_{y}}{y=0}\,
\left[
\(
\overline{\Delta}^{\,}_{\xi,y}
+
\Delta^{\,}_{\xi,y}
\)
\Delta^{\,}_{\psi,y}
-
(2\pi)^{2}\,
\left(
\overline{\Delta}^{\,}_{\xi,y}\,
\Delta^{\,}_{\xi,y}
+
\overline{\Delta}^{\,}_{\xi,y}\,
\mathrm{i}\,\lambda\,
\xi^{*}_{\mathrm{L},y+1}\,
\xi^{\,}_{\mathrm{R},y}
-
\Delta^{\,}_{\xi,y}\,
\mathrm{i}\,\lambda\,
\xi^{*}_{\mathrm{R},y}\,
\xi^{\,}_{\mathrm{L},y+1}
\right)
\right]
\nonumber\\
&\,
-
\prisum{L^{\,}_{y}}{y=0}\,
\left[
\left(
\mathrm{i}\,
\xi^{*}_{\mathrm{R},y}\,
\xi^{\,}_{\mathrm{L},y+1}
-
\mathrm{i}\,
\xi^{*}_{\mathrm{L},y+1}\,
\xi^{\,}_{\mathrm{R},y}
\right)
\Delta^{\,}_{\psi,y}
+
\left(
\overline{\Delta}^{\,}_{\xi,y}\,
+
\Delta^{\,}_{\xi,y}
\right)
\mathrm{i}\,
\psi^{\,}_{\mathrm{L},y+1}
\psi^{\,}_{\mathrm{R},y}
\right]
\nonumber\\
&\,
+
\(
\lambda\to\lambda',\
\Delta^{\,}_{\xi,\psi}\to\Delta^{\prime}_{\xi,\psi},\
\mathrm{i}\to-\mathrm{i},\
\mathrm{R}\leftrightarrow\mathrm{L},\
y\to y+1
\).
\label{eq: k=2 Lagrangian after Hubbard-Stratonovich}
\end{align}
\end{widetext}
The auxiliary Lagrangian
density~\eqref{eq: k=2 Lagrangian after Hubbard-Stratonovich}
reduces to the original Lagrangian density
\eqref{eq: k=2 interaction before decoupling}
when the equations of motion for the auxiliary fields,
namely,
\begin{subequations}
\label{eq: constraints on auxiliary fields when k=2}
\begin{align}
&
\Delta^{\,}_{\xi,y}=
-
\mathrm{i}\,
\lambda\,
\xi^{*}_{\mathrm{L},y+1}\,
\xi^{\,}_{\mathrm{R},y},
\label{eq: constraints on auxiliary fields when k=2 a}
\\
&
\Delta^{\,}_{\psi,y}=
+\mathrm{i}\,
\lambda\,
\psi^{\,}_{\mathrm{L},y+1}\,
\psi^{\,}_{\mathrm{R},y},
\label{eq: constraints on auxiliary fields when k=2 b}
\end{align}
\end{subequations}
and similarly for the primed fields,
are imposed. Note that the phases of the complex
auxiliary fields
$\Delta^{\,}_{\xi,y}$
and
$\Delta^{\prime}_{\xi,y}$ can be removed by a
gauge transformation, e.g.
$\xi^{\,}_{\mathrm{R},y}\to e^{\mathrm{i}\,\theta}\, \xi^{\,}_{\mathrm{R},y}$.
Imposing the symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ forces the constraints
$\Delta^{\,}_{\xi}=\Delta^{\prime}_{\xi}$ and
$\Delta^{\,}_{\psi}=\Delta^{\prime}_{\psi}$ among the
Hubbard-Stratonovich fields. However, it is important to
remember that, within a mean-field treatment of the
theory with the interaction
(\ref{eq: k=2 Lagrangian after Hubbard-Stratonovich}),
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$-symmetry can
be broken spontaneously if Eqs.\
(\ref{eq: constraints on auxiliary fields when k=2})
develop vacuum expectation values that are not symmetric under
$\Delta^{\,}_{\xi,\psi}\leftrightarrow\Delta^{\prime}_{\xi,\psi}$.
Checking the self-consistency of such spontaneous-symmetry-breaking
solutions is one of the primary goals of the present mean-field calculation.
At this point, the standard way to proceed is to integrate out both
the Dirac fields $\xi^{\,}_{\mathrm{M}}$ and the Majorana fields
$\psi^{\,}_{\mathrm{M}}$ and then to solve for the saddle point of the
effective action involving only the Hubbard-Stratonovich fields.
We first focus on the Majorana contribution to
Eq.~\eqref{eq: k=2 Lagrangian after Hubbard-Stratonovich}.
Taking the continuum limit in the $y$-direction and linearizing around $k^{\,}_{y}=\pi/2$
yields the full Euclidean action
\begin{subequations}
\begin{widetext}
\begin{align}
S^{\mathrm{aux}\,\mathrm{mf}\,\psi}_{x=0}\:=
\int\frac{\mathrm{d}\omega}{2\pi}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\begin{pmatrix}
c^{*}_{\bm{k}}
&
c^{\,}_{-\bm{k}}
\end{pmatrix}
\begin{pmatrix}
\mathrm{i}\,\omega+v^{\,}_{\xi}\,k^{\,}_{y}
&
k^{\,}_{z}-\mathrm{i}\,m^{\,}_{\xi}
\\
k^{\,}_{z}+\mathrm{i}\,m^{\,}_{\xi}
&
\mathrm{i}\,\omega-v^{\,}_{\xi}\,k^{\,}_{y}
\end{pmatrix}
\begin{pmatrix}
c^{\,}_{\bm{k}}
\\
c^{*}_{-\bm{k}}
\end{pmatrix},
\end{align}
\end{widetext}
where we have transformed to Fourier space along each wire
and defined the velocity and mass
\begin{align}\label{eq: psi m and v}
v^{\,}_{\xi}\:=
2\(
\Delta^{\,}_{\xi}
+
\Delta^{\prime}_{\xi}
\),
\qquad
m^{\,}_{\xi}\:=
2\(
\Delta^{\,}_{\xi}
-
\Delta^{\prime}_{\xi}
\).
\end{align}
The action $S^{\mathrm{aux}\,\mathrm{mf}\,\psi}_{x=0}$ is expressed
in terms of complex fermions
\begin{align}
c^{\,}_{\bm k}\:= c^{\,}_{k^{\,}_{y}}(k^{\,}_{z})
\end{align}
that are defined in terms of the Majorana modes by
\label{eq: complex bond fermion representation}
\begin{equation}
\psi^{\,}_{\mathrm{R},y}=
\mathrm{i}\,
\(
c^{*}_{y}
-
c^{\,}_{y}
\),
\qquad
\psi^{\,}_{\mathrm{L},y+1}=
c^{*}_{y+1}
+
c^{\,}_{y+1},
\end{equation}
for any directed bond $\langle y,y+1\rangle$ with $y$ even, and
\begin{equation}
\psi^{\,}_{\mathrm{L},y+1}=
c^{*}_{y}
+
c^{\,}_{y},
\quad
\psi^{\,}_{\mathrm{R},y+2}=
\mathrm{i}\,
\(
c^{*}_{y+1}
-
c^{\,}_{y+1}
\),
\end{equation}
for any directed bond $\langle y+1,y+2\rangle$ with $y$ even,
followed by taking the Fourier transform
\begin{align}
c^{\,}_{y}(k^{\,}_{z})\:=
\frac{1}{\sqrt{\mathcal{N}}}
\sum_{k^{\,}_{y}}
e^{+\mathrm{i}\,k^{\,}_{y}\,y}\,
c^{\,}_{k^{\,}_{y}}(k^{\,}_{z}).
\label{eq: Fourier expansion for c's}
\end{align}
\end{subequations}
A similar treatment of the Dirac contribution to
Eq.~\eqref{eq: k=2 Lagrangian after Hubbard-Stratonovich}
yields the full Euclidean action
\begin{subequations}
\begin{widetext}
\begin{align}
S^{\mathrm{aux}\,\mathrm{mf}\,\xi}_{x=0}\:=
\int\frac{\mathrm{d}\omega}{2\pi}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\begin{pmatrix}
\xi^{*}_{\mathrm{R},\bm{k}}
&
\xi^{*}_{\mathrm{L},\bm{k}}
\end{pmatrix}
\begin{pmatrix}
\mathrm{i}\,\omega\,+\,k^{\,}_{z}
&
+
\mathrm{i}\,v^{\,}_{\psi}\,k^{\,}_{y}
+
m^{\,}_{\psi}
\\
-
\mathrm{i}\,v^{\,}_{\psi}\,k^{\,}_{y}
+
m^{\,}_{\psi}
&
\mathrm{i}\,\omega\,-\,k^{\,}_{z}
\end{pmatrix}
\begin{pmatrix}
\xi^{\,}_{\mathrm{R},\bm{k}}
\\
\xi^{\,}_{\mathrm{L},\bm{k}}
\end{pmatrix},
\end{align}
\end{widetext}
where we have defined the velocity and mass
\begin{align}
&
v^{\,}_{\psi}\:=
\(
\Delta^{\,}_{\psi}
+
\Delta^{\prime}_{\psi}
\)
-
\frac{(2\pi)^{2}}{2}\,
v^{\,}_{\xi},
\label{eq: chi m and v a}
\\
&
m^{\,}_{\psi}\:=
\(
\Delta^{\,}_{\psi}
-
\Delta^{\prime}_{\psi}
\)
-
\frac{(2\pi)^{2}}{2}\,
m^{\,}_{\xi}
\label{eq: chi m and v b},
\end{align}
\end{subequations}
where $v^{\,}_{\xi}$ and $m^{\,}_{\xi}$ are defined in
Eq.~\eqref{eq: psi m and v}.
At this point, it is already possible to see that the surface theory
\begin{align}
S^{\mathrm{aux}\,\mathrm{mf}}_{x=0}\:=
S^{\mathrm{aux}\,\mathrm{mf}\,\xi}_{x=0}+
S^{\mathrm{aux}\,\mathrm{mf}\,\psi}_{x=0}
\end{align}
for the
case $k=2$ will be gapless so long as $\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$
is not broken spontaneously. This is because the masses
$m^{\,}_{\xi}$ and $m^{\,}_{\psi}$ vanish when
$\Delta^{\,}_{\xi}=\Delta^{\prime}_{\xi}$ and $\Delta^{\,}_{\psi}=\Delta^{\prime}_{\psi}$
[see Eqs.~\eqref{eq: psi m and v} and \eqref{eq: chi m and v b}].
Thus, what remains to be checked is that, upon integration over the fermions,
the saddle point of the resulting effective action has self-consistent solutions
such that
$\Delta^{\,}_{\xi}=\Delta^{\prime}_{\xi}$ and $\Delta^{\,}_{\psi}=\Delta^{\prime}_{\psi}$.
Integrating out the real and complex fermions, we find the following set of
four self-consistency equations for the masses and velocities:
\begin{widetext}
\begin{subequations}\label{eq: full set of coupled self-consistency equations}
\begin{align}
&
m^{\,}_{\psi}=
\frac{4\,\lambda\lambda^{\prime}}{\lambda+\lambda^{\prime}}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\frac{m^{\,}_{\xi}}{\sqrt{v^{2}_{\xi}\,k^{2}_{y}+k^{2}_{z}+m^{2}_{\xi}}}
-
2\,\pi^{2}\,m^{\,}_{\xi}
+
\frac{
\lambda-\lambda^{\prime}
}
{
\lambda+\lambda^{\prime}
}
\(v^{\,}_{\psi}+2\,\pi^{2}\,v^{\,}_{\xi}\),
\\
&
m^{\,}_{\xi}=
\frac{4\,\lambda\lambda^{\prime}}{\lambda+\lambda^{\prime}}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\frac{m^{\,}_{\psi}}{\sqrt{v^{2}_{\psi}\,k^{2}_{y}+k^{2}_{z}+m^{2}_{\psi}}}
+
\frac{\lambda-\lambda^{\prime}}{\lambda+\lambda^{\prime}}\,v^{\,}_{\xi},
\\
&
v^{\,}_{\psi}=
\frac{4\,\lambda\lambda^{\prime}}{\lambda+\lambda^{\prime}}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\frac{v^{\,}_{\xi}\,k^{2}_{y}}{\sqrt{v^{2}_{\xi}\,k^{2}_{y}+k^{2}_{z}+m^{2}_{\xi}}}
-
2\,\pi^{2}\,v^{\,}_{\xi}
+
\frac{\lambda-\lambda^{\prime}}{\lambda+\lambda^{\prime}}\,
\(m^{\,}_{\psi}+2\,\pi^{2}\,m^{\,}_{\xi}\),
\\
&
v^{\,}_{\xi}=
\frac{4\,\lambda\lambda^{\prime}}{\lambda+\lambda^{\prime}}
\int\frac{\mathrm{d}^{2}k}{(2\pi)^{2}}\,
\frac{
v^{\,}_{\psi}\,k^{2}_{y}
}
{
\sqrt{v^{2}_{\psi}\,k^{2}_{y}+k^{2}_{z}+m^{2}_{\psi}}
}
+
\frac{
\lambda-\lambda^{\prime}
}
{
\lambda+\lambda^{\prime}
}\,m^{\,}_{\xi}.
\end{align}
\end{subequations}
\end{widetext}
Equations \eqref{eq: full set of coupled self-consistency equations}
constitute a set of four coupled self-consistency equations that must
be solved simultaneously for the four unknowns
$m^{\,}_{\xi}$,
$m^{\,}_{\psi}$,
$v^{\,}_{\xi}$,
and
$v^{\,}_{\psi}$.
For general values of $\lambda$ and $\lambda^{\prime}$, this must be
done numerically. We find that nontrivial solutions of
Eqs.\ \eqref{eq: full set of coupled self-consistency equations} exist,
and that they exhibit the following general features. When
$\lambda=\lambda^{\prime}$, we find that
$m^{\,}_{\xi}=m^{\,}_{\psi}=0$ for all $\lambda>0$. Thus, at the
mean-field level, the surface of the $su(2)^{\,}_{2}$ non-Abelian coupled-wire
construction is a gapless liquid with both Dirac and
Majorana degrees of freedom, so long as the symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ is not broken explicitly. When
$\lambda\neq\lambda^{\prime}$, we find solutions where the masses
$m^{\,}_{\xi}$ and $m^{\,}_{\psi}\neq 0$. This agrees with our
earlier hypothesis that the surface develops a gap when the symmetry
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{y}}}$ is broken explicitly.
\section{Conclusions}
\label{sec: Conclusions}
In this paper, we have proposed a means of constructing a family
of non-Abelian topological phases in (3+1)-dimensional spacetime.
These phases inherit their non-Abelian
character from the underlying $su(2)^{\,}_{k}$ CFTs that describe the
constituent interacting fermionic quantum wires in the decoupled limit. For
the special case of $su(2)^{\,}_{2}$, we showed explicitly how to
construct a set of nonlocal operators that can be used to label a set of
degenerate ground states and to cycle between states in this set,
thus shedding light on some aspects of the topological order.
This calculation relies on the operator algebra of the underlying CFTs
that furnish the low-energy degrees of freedom for the coupled-wire
construction, thus making explicit the connection between these CFTs
and the emergent topological phase. We also examined the phase
diagram of the surface for this family of topological phases, and
showed explicitly for the case of $su(2)^{\,}_{2}$ that they are
gapless and protected by a nonlocal analogue of TRS.
There are many open questions to be pursued in light of this work.
First and foremost, a deeper study of the precise nature of the
topological order in the $su(2)^{\,}_{2}$ example is necessary in
order to fully specify the phase. (Indeed, the question of what are
the minimal data necessary in order to uniquely determine an arbitrary
topological order in 3D is itself not settled.) For example, one can
determine the topological spins of the pointlike excitations, and
study the braiding of multiple looplike excitations in this model.
One can also ask whether the presence of the time-reversal analogues
$\mathcal{T}^{\,}_{\mathrm{eff},\hat{\bm{x}},\hat{\bm{y}}}$
\textit{enriches} the topological order in this phase, i.e., whether
the action of the symmetry on excitations provides additional
topological information~\cite{Barkeshli16}. Moreover, one could
consider cases where $k>2$ in more detail, and ask whether these also
yield candidates for non-Abelian topological order in 3D. More
broadly, it would be interesting to determine whether and how such
phases could be represented within the Crane-Yetter/Walker-Wang
construction, or the formalism of discrete non-Abelian gauge theory.
Another avenue to pursue would be to try to construct these phases
via parton constructions like the ones that have been carried out for
some non-Abelian quantum Hall states~\cite{Wen99,Barkeshli10}, and for
the Abelian FTIs in (3+1)-dimensional spacetime%
~\cite{Maciejko10,Swingle11}.
Coupled-layer constructions of non-Abelian topological phases,
extending the work of Ref.~\cite{Jian14}, could also be considered.
It would also be interesting to investigate how to construct
microscopic lattice models that yield the
$su(2)^{\,}_{k}$ topological phases in (3+1)-dimensional spacetime
proposed in this work. One clue for how to proceed relies on the fact that
$su(2)^{\,}_{k}$ CFTs can be obtained as continuum limits of certain
spin-$\frac{k}{2}$ chains~\cite{Affleck86}. Microscopic spin-spin interactions
can then be derived whose continuum limits give rise to the current-current
interactions used to gap out an array of such spin chains; for
example. Such interactions were constructed for the (2+1)-dimensional
case in Ref.~\cite{Huang16b}.
Another line of inquiry is to investigate more deeply the nature of
the $su(2)^{\,}_{k}$ surface states for $k>2$. While spontaneous
breaking of the nonlocal TRS analogue
(and the concomitant opening of a gap on
the surface) is always a possibility, it could be that these surface
states constitute novel stable fractionalized non-Fermi liquid phases.
The investigation of this class of models would likely need to rely on
nonperturbative techniques, and could provide insights into conformal
field theories in (2+1)-dimensional spacetime.
\begin{acknowledgments}
We thank D.~Aasen, M.~Barkeshli, P.~Bonderson, F.~Burnell, M.~Cheng,
M.~Metlitski, M.~Oshikawa, Z.~Wang, X.-G.~Wen, and D.~Williamson for
helpful discussions. T.I. gratefully acknowledges the hospitality of
the KITP, where a significant portion of this work was completed, and
thanks the organizers of the ``Symmetry, Topology, and Quantum Phases
of Matter: From Tensor Networks to Physical Realizations'' and
``Synthetic Quantum Matter'' programs, both of which were supported in
part by the National Science Foundation under Grant No.~NSF
PHY11-25915. T.I. was supported by a KITP Graduate Fellowship,
the National Science Foundation Graduate Research Fellowship Program
under Grant No.~DGE-1247312, the Laboratory for Physical Sciences,
Microsoft, and a JQI Postdoctoral Fellowship.
C.C. was supported by DOE Grant DE-FG02-06ER46316.
\end{acknowledgments}
|
\section{Introduction}
Networks arise in various applications in economics, engineering and social sciences. In a typical social science application, the vertices of the network represent individuals, while their relationships are represented by the edges. The study of structural properties of these networks, and algorithms to find these structures are extremely important in this context. Various random graph models have been introduced to study such real-life networks (see, e.g., \cite{holland83sbm})--- and questions about networks translate directly into questions about random graphs under this approach.
Graph partition problems are natural class of algorithmic questions which arise in this context. In these problems, the goal is to
find a partition of the vertex set that maximizes some objective function,
typically given by a function of the edges. Graph partition problems are of interest in applications as diverse as community detection \cite{decelle2011sbm} and VLSI design \cite{kahng2011partition}.
In this paper, we focus on the Max $\kappa$-cut, an important example in this class.
For any graph $G= (V,E)$, the Max $\kappa$-cut problem (henceforth denoted as ${\sf MaxCut }_\kappa$)
seeks to divide the vertices, $V$, into $\kappa$ (not necessarily equal) parts such that
the number of edges between distinct parts is maximized.
For $\kappa=2$ this reduces to the well known ${\sf{MAXCUT}}$ problem
(see \cite{poljaktuza1993maxcut} for a survey of the $\sf{MAXCUT}$ problem).
From the point of view of complexity theory, these questions are usually NP hard in the worst
case. This motivates a study of average case complexity, often formalized by studying this problem on random graph instances. As a first attempt, one seeks to determine the typical behavior of these quantities on a random graph--- this provides a valuable benchmark for comparing the performance of specific algorithms on random instances.
Such questions have been studied in classical settings, such as the \ER random graph and random regular graph ensembles.
The key insight in this setting is a connection between statistical physics and the ${\sf MaxCut }_\kappa$ problem, enunciated as follows.
Any $\kappa$-cut can be represented by an assignment of spins $\sigma \in [\kappa]^N$ to the vertices of the graph. Further, setting $A= (A_{ij})$ to be the adjacency matrix of the random graph $G$, we have,
\begin{align}
\frac{{\sf MaxCut }_\kappa(G)}{N} = \frac{1}{2N} \max_{\sigma \in [\kappa]^N} \sum_{i, j =1 }^N A_{i,j} \1(\sigma_i \neq \sigma_j). \label{eq:representation}
\end{align}
In statistical physics parlance, \eqref{eq:representation} establishes a direct relation between the ${\sf MaxCut }_\kappa$ and the ground state of the antiferromagnetic Potts model on the graph. Connections between
graph partition problems and statistical physics are, by now, classical \cite{fu-anderson}. For a
textbook introduction to the physical perspective on these questions, we refer the reader to \cite{mezard2009information,MPV}.
Physicists predict that the antiferromagnetic nature of the Max $\kappa$-cut should force the quantity to behave as the ground state of a disordered spin glass, and its behavior in graphs with large degrees should be well approximated by properties of ground states in mean field spin glasses.
For the ${\sf {MAXCUT} }$ problem on sparse \ER and random regular graphs, this idea was partially formalized in \cite{dembo2016extremal} and \cite{sen2016optimization}. The authors of \cite{dembo2016extremal} deduced that for $G\sim G(N, \frac{c}{N})$, as $N \to \infty$, we have,
\begin{align}
\frac{{\sf MAXCUT}(G)}{N} = \frac{c}{4} + {\sf{P}_*} \sqrt{\frac{c}{4}} + o_c(\sqrt{c}). \label{eq:prev_result}
\end{align}
Here, ${{\sf P}}_*$ is the limiting ground state energy of the Sherrington-Kirkpatrick model \cite{P3}.
Here and henceforth in the paper, we say that a sequence of random variables, $(X_N)$, satisfies $X_N = o_c(\sqrt{c})$ if and only if there is a deterministic function $g(c) = o(\sqrt{c})$ such that $\P[|X_N| \leq g(c)] \to 1$ as $N \to \infty$. The first term in the right hand side comes from the standard observation that a typical partition of the vertices will contribute $Nc/4$ edges to the cut in expectation. The second term is the leading order correction, and specifies the difference in size between a typical cut and the MAXCUT. An analogous formula for the ${\sf MaxCut }_\kappa$ on sparse \ER and random regular graphs was derived in \cite{sen2016optimization}.
For practical applications, it is thus of natural interest to determine the typical value of these quantities on
random graph ensembles that capture natural properties of realistic networks. In practice, networks are typically observed to be sparse and ``inhomogeneous" \cite{albert02review,doro2002survey}.
The simplest random graph models, such as \ER and random regular graphs, lead to instances where the degree distributions are relatively concentrated--- a feature seldom observed in real networks. To address this issue, a plethora of models have been introduced, which faithfully capture some of the observed characteristics of real networks.
In this paper, we seek to establish formulae similar to \eqref{eq:prev_result} for a general family of graph models using the framework of \cite{dembo2016extremal} and \cite{sen2016optimization}. Our approach leads naturally to the study of an inhomogeneous Potts spin glass model, which has yet to be studied rigorously in the mathematical literature.
Let us first explain the class of random graph models that we study.
Our framework will be similar to the one introduced by Soderberg \cite{soderberg2002inhomogeneous} and adopted by Bollobas, Janson and Riordan \cite{BJR2007inhomogeneous}. Furthermore, this model has natural connections to the theory of graphons for dense sequences of random graphs \cite{BC1,BC2}.
Consider a symmetric
kernel $K:[0,1]^2\to[0,\infty)$. We will assume throughout that $K \in L^1([0,1]^2,{\rm d} x)$.
Given such a kernel, consider the following model for a sequence of inhomogeneous random graphs $\mathcal{G}_N=(V_N,E_N)$.
For all $N\geq 1$, we let the vertex set be $V_N =[N]$. The edges will then be added independently with probability
\begin{align}
\prob[\{i,j\}\in E_N] = \min\Big\{c\frac{\tilde K(i,j)}{N},1 \Big\}, \label{eq:connection}
\end{align}
where $\tilde K$ is the average of $K$ within blocks,
\begin{align}
\tilde{K}_N(i,j) = N^2 \int_{ [\frac{i-1}{N}, \frac{i}{N}] \times [\frac{j-1}{N}, \frac{j}{N}] } K(x,y)\, {\rm d} x {\rm d} y.
\end{align}
This specifies the random graph model. The parameter $c$ controls the degree of the vertices.
We note that this model is more restricted compared to that of Bollobas, Janson and Riordan \cite{BJR2007inhomogeneous}. In the notation of \cite{BJR2007inhomogeneous}, we restrict ourselves to the case where the ground space $\mathcal{S} = [0,1]$ and the measure $\mu$ is the Lebesgue measure. Further, the model introduced in \cite{BJR2007inhomogeneous} is governed by the value of the kernel $K$ on a set of measure zero. Here we average over small partitions of the kernel, so that
we may avoid technical subtleties on sets of measure zero. For reasonable kernels, such as continuous ones, this distinction will be negligible.
To state our main result in a concrete setting, let us first work in the case when $K$ is block constant.
That is, we assume that there are numbers
\[
0= t_0 < t_1 < t_2 < \ldots < t_{M-1} < 1= t_M
\]
such that $K$ is constant on each square of the form $[t_{j-1}, t_j] \times [ t_{k-1}, t_k]$ for $0\leq j, k \leq M$. Further, we set $\rho^{s} = t_{s}- t_{s-1}$, $s= 1, \ldots, M$. For any such block kernel $K$, let $\mathbf{K}$ denote the $M\times M$ matrix of the values of the kernel on the blocks. For technical reasons, we will work with block kernels such that the matrix $\mathbf{K}$ is positive definite. Finally, we note that any block constant kernel with finitely many blocks is almost surely bounded. By a standard application of the Efron-Stein inequality \cite{BLM2013conc}, it suffices to study the asymptotic behavior of $\E[{\sf MaxCut }_\kappa(G_N)]/N$.
To analyze this quantity, we introduce the following notation. For any finite set $\mathscr{S}$, let $\mathscr{D}$ be the space of proportions, given by
\begin{equation}
\mathscr{D} = \Big\{ (d_1^s, \dots, d_{\kappa}^s)_{s \in \mathscr{S}} \mathrel{}\Big| d_k^{s} \geq 0, \mathrel{} \sum_{k = 1}^{\kappa} d_k^s = 1 ~\forall s \in \mathscr{S} \Big\}.\label{eq:props}
\end{equation}
In our setting, $\mathscr{S}=[M]$. Any $d \in \mathscr{D}$ can be expressed as $d = (d^s)_{s \in \mathscr{S}}$, where $(d^s)_{s \in \mathscr{S}}$ is a collection of probability measures on $[\kappa]$. The distribution $d^s$ governs the proportion of vertices in block $s$ which belong to the partition $i$, $1\leq i \leq \kappa$. We will refer to the elements $d \in \mathscr{D}$ as proportions. The following theorem characterizes the value of the ${\sf MaxCut }_\kappa$ problem for inhomogeneous graphs $G_N$ with block constant kernels and large degrees, up to corrections which are $o(\sqrt{c})$.
\begin{theo}
\label{thm:discretization}
We have, as $N\to \infty$,
\begin{align}
\lim_{N \to \infty} \E\Big[ \frac{{\sf MaxCut }_\kappa(G_N)}{N} \Big]
&= \sup_{d \in \mathscr{D}} \Big[ \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s, t) \rho^{s} \rho^{t} \bigl(1- \langle d^s, d^t \rangle \bigr) + \frac{\sqrt{c}}{2} \mathcal{P}(d) \Big] + o(\sqrt{c}). \nonumber
\end{align}
\end{theo}
\begin{remark}
We take this opportunity to comment on the positive definite assumption on the matrix $\mathbf{K}$. One prominent example where $\mathbf{K}$ is not positive definite is the random bipartite graph, where the kernel consists of two off-diagonal blocks. However, note that for $\kappa=2$, the behavior of the ${\sf{MAXCUT}}$ on this graph is trivial, and very different from that established in Theorem \ref{thm:discretization}.
\end{remark}
\noindent
Note that the leading term in Theorem \ref{thm:discretization} is a variational problem involving the empirical distribution of spins within each block. This variational problem has two terms: the first term governs the expected cut-size, while the second term, of order $\sqrt{c}$, governs the extra contribution which is attained by optimization. It remains to introduce $\mathcal{P}(d)$.
It turns out that $\mathcal{P}(d)$ is the limiting ground state energy of the
inhomogeneous Potts spin glass model, subject to constraints on the composition of spins within each block. We introduce this model in the rest of the section, and define the constant $\mathcal{P}(d)$ rigorously using a Parisi type formula for the limiting free energy.
\subsection{The Inhomogeneous Potts Model}
We consider a natural
generalization of the Potts spin glass model that allows for inhomogeneous coupling interactions
between species. The configuration space for this model is $\Sigma_N=[\kappa]^N$ for some $\kappa\geq2$.
Let $\mathscr{S}$ be the finite set in \eqref{eq:props}, each element of which is called a species.
For each $N$, we are given a partition of $[N]$ indexed by the species as
\[
[N]=\cup_{s\in\mathscr{S}} I_s.
\]
We say that $i$ belongs to species $s$ if $i\in I_s$. Conversely, we denote by $s(i)$ the species to which $i$ belongs.
Let $N_s = |I_s|$. Naturally, this quantity varies in $N$. To obtain a reasonable
limiting structure, we assume that the proportions converge:
\begin{equation}\label{eq:proportion-def}
\rho_N^s = \frac{N_s}{N}\to\rho^s\in(0,1).
\end{equation}
The Hamiltonian for this model, $H_N$, is the centered Gaussian process
\begin{equation}\label{eq:ham-def}
H_N(\sigma)=\frac{1}{\sqrt{N}} \sum_{i,j=1}^{N} g_{i,j} \1(\sigma_i = \sigma_j),
\end{equation}
where $g_{i,j}$ are independent, centered Gaussian random variables with covariance
\begin{equation}
\E g_{i,j}^2 = \Delta_{s,t}^2, \qquad s,t\in\mathscr{S}, i\in I_s, j\in I_t.
\end{equation}
We assume, following \cite{Barra2015,P1}, that the matrix
$\Delta := \Delta_{s,t}^2$ is symmetric and positive definite in $s$ and $t$.
Observe that if we define, for $\sigma^1,\sigma^2\in\Sigma_N$,
\begin{equation}
R_{1,2}^{s}(k, {k'})=\frac{1}{N_s} \sum_{i\in I_s} \1 (\sigma_i^1=k)\1 (\sigma_i^2=k')
\end{equation}
and define the $\kappa\times\kappa$ species overlap matrix
\begin{equation}\label{eq:overlap-def}
R_{1,2}^s=\bigl(R_{1,2}^s (k, {k'}) \bigr)_{k,k'\leq\kappa}
\end{equation}
then $H_N$ has covariance
\begin{equation}
\Cov(\sigma^1,\sigma^2)= N\sum_{s,t\in\mathscr{S}}\Delta_{s,t}^2\rho^s_N\rho_N^t(R_{1,2}^s,R_{1,2}^t)
\end{equation}
where $(\cdot,\cdot)$ denotes the Frobenius (or Hilbert-Schmidt) inner product.
This model is an inhomogeneous extension of the Potts spin glass model, which has been studied extensively in the physics literature \cite{EldSher83b,EldSherr83a, nishimori1983gauge,caltagirone2012dynamical}, and analyzed rigorously in \cite{P2, PVS}. In particular, we break the symmetry between sites.
When $\kappa=2$, this type of inhomogeneity was introduced in an equivalent form by Barra, Contucci, Mignone and Tantari
in \cite{Barra2015} where a Guerra-type \cite{Guerra} upper bound for the free energy was obtained.
The matching lower bound was obtained by Panchenko in \cite{P1}.
Our goal is to compute constrained free energies of the type
\begin{equation}\label{eq:const-free-def}
F_N(A) = \frac{1}{N}\E \log \sum_{\sigma\in A} e^{\beta H_N(\sigma)}
\end{equation}
for a specific choices of $A$. Recall the space of proportions $\mathscr{D}$ defined in \eqref{eq:props}.
Given a $d \in \mathscr{D}$, we have the associated \emph{constrained state space}
\begin{equation}
\Sigma^\epsilon_N(d) = \Big\{ \sigma \in \Sigma_N \mathrel{}\Big|\mathrel{} \sum_{i \in I_s} \frac{\1(\sigma_i = k)}{N_s} \in [d^s_k - \epsilon, d_k^s + \epsilon] \Big\}. \label{eq:constrainedspace}
\end{equation}
We will use the notation $\Sigma_N(d) := \Sigma_N^0(d)$ to denote the constrained state space where the proportions of spins
within species are exactly equal to the proportion $d$. Let $\mathscr{D}_N \subset \mathscr{D}$ be the space of feasible constraints for configuration spaces of $N$ coordinates
\begin{equation}\label{eq:feasible-constraint-space}
\mathscr{D}_N = \{ d \in \mathscr{D} \mathrel{}\mid\mathrel{} \Sigma_N(d) \neq \emptyset \}.
\end{equation}
We are particularly interested in computing constrained free energies as in \prettyref{eq:const-free-def} with
$A=\Sigma(d)$.
Note that since the space of feasible configurations is at most polynomial growth,
classical concentration arguments show that the free energy of the whole system,
$F_N\bigl(\Sigma_N\bigr)$, is asymptotically given by
the maximum of $F_N\bigl(\Sigma_N(d)\bigr)$ over $\mathscr{D}$.
As in the Potts spin glass model, among others, the overlap \prettyref{eq:overlap-def} will play a key role.
In particular, we will find that the array of overlaps from i.i.d. draws of configurations from (a perturbation of) the Gibbs measure,
will be determined by a path $\pi:[0,1]\to \Gamma_\kappa^{\abs{\mathscr{S}}}$. Here $\Gamma_\kappa$ is the
space of $\kappa \times \kappa$ positive definite matrices and $\pi$ is effectively the family of quantile
transform of the limiting law of the overlap of two independent copies $\sigma^1,\sigma^2$ from the Gibbs measure.
We now turn to the main result. To this end, we denote
the space of \emph{left-continuous monotone functions} on $\Gamma_\kappa$ as
\begin{equation}
\Pi = \left\{ \pi: [0,1] \to \Gamma_\kappa : \text{$\pi$ is left-continuous, $\pi(x) \leq \pi(x')$ for $x \leq x'$} \right\},
\end{equation}
where $\pi(x) \leq \pi(x')$ means that $ \pi(x') - \pi(x) \in \Gamma_\kappa$. Similarly, we let
\begin{equation}
\bm\Pi =\left\{ (\pi^s)_{s \in \mathscr{S}} :[0,1]\to\Gamma_\kappa^{\abs{\mathscr{S}}} \mathrel{}\big|\mathrel{} \pi^s\in\Pi\right\}.
\end{equation}
We also have the following metric on $\bm \Pi$
\begin{equation}
\Delta(\bm \pi, \bm {\tilde \pi}) = \int_0^1 \max_{s \in \mathscr{S}} \| \pi^s(x) - \tilde\pi^s(x) \|_1 \, {\rm d} x.
\end{equation}
For $d \in \mathscr{D}$ and $r > 0$ we define the following sequences of parameters. Let $(x_i)_{i=1}^r$ be a strictly increasing sequence of numbers
\begin{equation}\label{eq:RPC6}
0=x_{-1}<x_{0}<\ldots<x_{r}=1.
\end{equation}
For each species, let $(Q^s_i)_{i=1}^r$ be an increasing sequence of $\kappa \times \kappa$ positive semi-definite matrices
\begin{equation}\label{eq:RPC5}
0=Q_{0}^{s}\leq Q_{1}^{s}\leq\ldots\leq Q_{r}^{s}= \operatorname{diag}(d_{1}^{s},\ldots,d_{\kappa}^{s})=D^{s}.
\end{equation}
Given these sequences, for each species we can define the Gaussian vector $(z_p^s) \in \mathbb{R}^{\kappa}$ such that
\begin{equation}
\E z_p^s (z_q^s)^T = 2 \delta_{p,q} \bigg( \sum_{t \in \mathscr{S}} \Delta_{s,t}^2 \rho^t Q_p^t - \sum_{t \in \mathscr{S}} \Delta_{s,t}^2 \rho^t Q_{p-1}^t \bigg).
\end{equation}
The non-random value $X_0^s$ is defined recursively as
\[X_r^s = \log \sum_{k \leq \kappa} \exp \bigg( \sum_{1 \leq p \leq r} z_p^s(k) + \lambda^s_{k} \bigg), \]
\begin{equation}\label{eq:recursion}
X^s_k = \frac{1}{x_k} \log \E_k \exp(x_k X^s_{k+1}), \text{ for }0 \leq k < r,
\end{equation}
where $\E_k$ denotes the expectation with respect to only $z_{k + 1}$. Finally, define the functional
\begin{multline}\label{eq:parisiformula}
\mathscr{P}(r, x, d, ((\lambda^s), (Q^s))_{s\in \mathscr{S}}) = \sum_{s \in \mathscr{S}} \rho^s X_0^s - \sum_{s \in \mathscr{S}}\sum_{k \leq \kappa}\rho^s \lambda_k^s d^s_k \\- \frac{1}{2} \sum_{\ell = 0}^{r-1} x_{\ell} \sum_{s,t \in \mathscr{S}} \Delta^2_{st} \rho^s \rho^t \left( (Q^s_{\ell+1},Q^t_{\ell+1}) - (Q^s_{\ell},Q^t_{\ell}) \right).
\end{multline}
The next result characterizes the limiting free energy in these models.
\begin{theo}\label{thm:ParisiUnConstrained}
For any $\kappa \geq 2$, set of species $\mathscr{S}$, and sequences $\rho_N^s\rightarrow \rho^s$
we have the following:
\begin{enumerate}
\item For any $d \in \mathscr{D}$ and $\epsilon_N \to 0$ sufficiently slowly, the limit of the constrained free energy is given by
\begin{equation}\label{eq:ParisiConstrained}
\lim_{N \to \infty} F_N\bigl(\Sigma^{\epsilon_N}_N(d) \bigr) = \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d,(\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
\item The limit of the unconstrained free energy is given by
\begin{equation}\label{eq:ParisiUnConstrained}
\lim_{N \to \infty} F_N\bigl(\Sigma_N\bigr) = \sup_{d \in \mathscr{D}} \, \inf_{x, r, (\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
\end{enumerate}
\end{theo}
Before moving forward, we explain the non-trivial obstacles encountered in the inhomogeneous Potts model. First, the symmetry between sites is broken. As a result, interactions both within the species and between the species must be considered. Secondly, the natural overlap structure are matrices \eqref{eq:overlap-def} and are a priori not necessarily positive definite in the limit.
Each of these issues has been studied in the past in \cite{P1} and \cite{P2}. The synchronization property in \cite{P1} connected the species overlaps with the average of the overlaps over the entire system. Similarly, the synchronization property in \cite{P2} proved the overlaps concentrated on the space of Gram matrices in the limit. Another consequence of this result implied the overlap matrices could also be recovered from the trace of the matrix. At the heart of both of these synchronization arguments were generalized Ghirlanda-Guerra type identities that implied an ultrametric underlying structure of the overlaps \cite{PUltra}. The synchronization combines local and global ultrametric properties forcing a rigid distribution structure.
We prove an analogue of the Ghirlanda-Guerra identities which combines those in the inhomogeneous SK and Potts models. This results in a simultaneous synchronization mechanism of the overlap matrices both within and between species. In our setting, we will be able recover the structure of the overlap matrices $R^s_{\ell, \ell'}$ deterministically from the average of the traces of overlaps
\[
\sum_{s \in \mathscr{S}} \rho^s \operatorname{tr} \bigl( R^s_{\ell, \ell'} \bigr).
\]
These techniques will reduce the problem to a familiar setting, allowing us to derive a formula for the free energy using the Guerra upper bound \cite{Guerra} and the Aizenman-Sims-Starr scheme \cite{AS2} using the characterization method introduced in \cite{AA} and formalized further in \cite{Pspins,PPF}. The resulting functional order parameter is a vector of monotone paths of $\kappa \times \kappa$ matrices.
In applications, we will be interested, not only in the free energy, but also the maximum of
\eqref{eq:ham-def} subject to the same constraints.
To understand the connection between the two, we recall the classical fact from statistical mechanics
that the ground state energy of a system can be obtained as the ``zero temperature limit'' of the free energy of this system---
a limit usually referred to as annealing \cite{kirkpatrick1983optimization}. In this case, we take the beaten path, and define free energies of the form
\begin{equation}
F^\beta_N(A) = \frac{1}{N}\E \log \sum_{\sigma\in A} e^{\beta H_N(\sigma)}, \label{eq:free_energy_beta}
\end{equation}
where $H_N$ is as in \eqref{eq:ham-def} for some fixed $\Delta$. Here $\beta$ is the inverse temperature, and the zero temperature limit corresponds to sending $\beta \to \infty$.
We note that the free energy \eqref{eq:free_energy_beta} corresponds to the Hamiltonian $H_N(\sigma)$ defined as in \eqref{eq:ham-def}, with respect to $\Delta^\beta=\beta^2 \Delta$ instead of $\Delta$. Therefore, a straightforward modification of Theorem \ref{thm:ParisiUnConstrained} implies that for any $d \in \mathscr{D}$,
\begin{equation}
\lim_{N \to \infty} F^\beta_N\bigl(\Sigma^{\epsilon_N}_N(d) \bigr) = \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}_\beta\bigl(r, x, d,(\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
The functional $\mathscr{P}_\beta(r, x, d, (\lambda^s, Q^s))_{s \in \mathscr{S}})$ is identical to \eqref{eq:parisiformula} with $\Delta^\beta$ in place of $\Delta$. As a corollary to the theorem above, we obtain an expression for the limiting constrained ground state energies.
\begin{cor}
\label{cor:ground_state}
For any configuration $d \in \mathscr{D}$, we have, as $N\to \infty$, for some sequence $\varepsilon_N \to 0$ sufficiently slowly,
\begin{align}
\lim_{N \to \infty} \frac{1}{N} \E\Big[ \max_{\sigma \in \Sigma_N^{\varepsilon_N}(d)} H(\sigma) \Big] &= \lim_{\beta \to \infty} \frac{1}{\beta} \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}_{\beta} \bigl(r, x, d,(\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr)
:= \mathcal{P} (d). \nonumber
\end{align}
\end{cor}
\noindent Note that $\mathcal{P}(d)$ is the relevant constant for the ${\sf MaxCut }_\kappa$ in Theorem \ref{thm:discretization}.
Exact variational formulas for ground state energies like $\mathcal{P}(d)$ have been obtained recently in several models \cite{auffchen2017,chensen2017,JagTob17}. However, in this setting this remains an interesting question.
\subsection{Applications}
We return to study of graph partitioning problems in this section and study some examples .
We first note that given a general kernel $K \in L^1$, to determine the ${\sf MaxCut }_\kappa$ up to $o(\sqrt{c})$ corrections, it is enough to restrict ourselves to block-constant kernels. To this end, given a kernel $K$ and $M \geq 1$, we construct the kernel $K_1$ by ``coarsening" the kernel $K$,
\begin{align}
K_1(x,y) = M^2 \sum_{i,j=1}^{M} \1 \Big(x \in \Big[\frac{i-1}{M}, \frac{i}{M} \Big], y \in \Big[\frac{j-1}{M}, \frac{j}{M} \Big] \Big) \int_{\Big[\frac{i-1}{M}, \frac{i}{M}\Big]\times \Big[\frac{j-1}{M}, \frac{j}{M} \Big]} K(s,t ) \, {\rm d} s {\rm d} t.
\end{align}
\noindent Let $\tilde{G}_N$ denote the sequence of graphs formed from the kernel $K_1$ using \eqref{eq:connection}.
Then we have,
\begin{lem}
\label{lemma:discretization}
For any kernel $K$ and $c$, $0 <\delta <1/2 $, we can choose $M := M(c)$ such that for all $N$ sufficiently large,
\begin{align}
\Big| \E\Big[ \frac{{\sf MaxCut }_\kappa(G_N)}{N} \Big] - \E \Big[ \frac{{\sf MaxCut }_\kappa(\tilde{G}_N)}{N} \Big] \Big| \leq c^{1/2- \delta}. \nonumber
\end{align}
\end{lem}
\noindent
Any kernel $K$ is naturally associated with the integral operator $T_K (f ) (x) := \int K(x,y) f(y) {\rm d} y$. We assume that for $f \in L^\infty([0,1])$, $\iint f(x) f(y ) K(x,y) {\rm d} x {\rm d} y \geq 0$. We note that in this case, the corresponding discretized kernel $K_1$, described in Lemma \ref{lemma:discretization}, inherits the positive definite character. Therefore, ${\sf MaxCut }_\kappa(\tilde{G}_N)$ can be determined by an application of Theorem \ref{thm:ParisiUnConstrained}. We now turn to some examples to which our results apply.
\begin{exam}[Finite species block model]
The first example concerns the simple case when the kernel $K$ has an explicit block structure. This model has been proposed
and studied intensely by S\"{o}derberg \cite{soderberg2002inhomogeneous} and Bollobas, Janson and Riordan \cite{BJR2007inhomogeneous}. These models have also been studied as ``Stochastic Block Models" in Statistics, Machine Learning, and Theoretical Computer Science in connection to the community detection problem \cite{decelle2011sbm,massoullie2014sdp,mossel2013sbm}. Our results apply directly to this model in case the kernel is positive semidefinite.
\end{exam}
\begin{exam}[Rank 1 model]
The next example concerns the Rank 1 model for random graphs. In this model, we have a function $\psi : [0,1] \to \mathbb{R}^+$ such that
$K (x,y) = \psi(x) \psi(y)$. $\psi(x)$ governs the ``activity" of the vertex and the probability of a connection is determined by the product of the activities of the two vertices. This model has been extensively studied, see, e.g.,\cite{BDMl2006degree,chunglu2002degree,norrosreittu2006degree}. Prominent features of interest include the existence and size of a giant component, the degree distribution, the typical distances between the vertices etc. We refer to \cite[Section 16]{BJR2007inhomogeneous} for an extensive survey of the related family of models and connections to earlier results. The kernel is positive semidefinite in this case. Further, $\int K < \infty$ whenever
$\psi \in L^1$. In this case, the kernel is in our framework and our result applies.
We note that for this example, if $\psi>0$ is constant on blocks, then we reduce to the example of block kernels discussed earlier. The approximation scheme for this example basically approximates the function $\psi$ by a piecewise constant function on $[0,1]$. From an algorithmic viewpoint, our result says that for evaluating the first order correction term, we can coarsen the model to a setup where there are finitely many species with the same activity.
\end{exam}
\begin{exam}[Dubins's model]
Consider the Dubins kernel $K(x,y) = \frac{1}{\max\{x,y\}}$. Observe that $K\in L^1$, and is symmetric and positive definite as a function so that our results apply. (The associated integral operator is a bounded operator from $L^2([0,1])$ to $L^2([0,1])$ as shown in \cite{BJR2007inhomogeneous}.) In this case, $\P[ \{i,j \} \in E_N ] = c/j \wedge 1$ for $j \geq i, c$. This corresponds to the situation where the graph is formed by a sequential addition of vertices, and the $j^{th}$ vertex joins to the existing vertices independently with probability $c/j$. This model is naturally inhomogeneous, in that the older vertices usually have higher degrees and play a crucial role in determining the structure of the graph. An infinite version of this model was introduced by Dubins in 1984, who wished to determine the critical $c$ such that the graph would have an infinite path (see \cite{kalikow88weiss, shepp89connectedness}). The critical constant $c=1/4$ was partially determined by Kalikow and Weiss \cite{kalikow88weiss} and finally determined by Shepp \cite{shepp89connectedness}. Durett \cite{durrett03rigorous} determined that $c=1/4$ is also the critical threshold for the emergence of a giant component in the finite graph. We refer the reader to \cite{BJR2007inhomogeneous} for a detailed survey of the model and related results.
\end{exam}
\subsection*{Acknowledgements}
This research was conducted while A.J. was supported by NSF OISE-1604232 and J.K. was partially supported by
NSERC grant RGPIN-2015-04637.
\subsection{Outline}
Before turning to the body of the paper, let us briefly outline the structure of the paper and the proof of
the main results. Theorem \ref{thm:discretization} follows using the interpolation idea introduced in \cite{dembo2016extremal} and generalized in \cite{sen2016optimization}. Its proof, included in Section \ref{sec:cut_proofs}, compares the ${\sf MaxCut }_\kappa$ value on inhomogeneous random graphs to the ground state of the inhomogeneous Potts spin glass.
The proof of Theorem \ref{thm:ParisiUnConstrained} follows the method outlined above.
The starting point of the proof is
the characterization of a family of arrays that follow a natural generalization of the Ghirlanda-Guerra
identities \cite{GG} for this setting. This combines the synchronization mechanisms of \cite{P1,P2} and is included in
\prettyref{sec:characterization}. In \prettyref{sec:perturbation}, we construct a perturbation of the system that
does not affect the limiting free energy but allows us to use the derived invariance properties.
With these results in hand, we prove the upper bound in \prettyref{sec:Guerra-UB} using a Guerra-type interpolation
and the matching lower bound using an Aizenman-Sims-Starr scheme in \prettyref{sec:Aizenman-LB}.
Before proving the lower bound, we briefly study the continuity of certain functionals
used in the lower bound in \prettyref{sec:continuity}.
\section{Invariant arrays and their Characterization}\label{sec:characterization}
In this section, we study an invariance property that combines the
multispecies and vector spin Ghirlanda-Guerra type identities \cite{P1, P2} for the limit points
of doubly infinite array of overlaps, \prettyref{eq:overlap-def}, of independent
draws from the Gibbs measure. This will allow us to characterize
these limit points.
Let $\mathcal{R}_{\kappa}$
be the space of arrays of the form $R_{\ell,\ell'}^s$ such
that there is a collection of vectors $\bigl(v_{i}(\ell,s)\bigr)_{i,\ell\geq1,s\in\mathscr{S}}$
in $\mathbb{R}^{\kappa}$ such that
\begin{equation}\label{eq:arraydefinition}
R_{\ell,\ell'}^s=\sum_{i}v_{i}(\ell,s)\otimes v_{i}(\ell',s),
\end{equation}
and such that
\begin{equation}
R_{\ell,\ell'}^s(e_{k},e_{k})\in[0,1]\quad\forall k\in[\kappa].
\end{equation}
We equip $\mathcal{R}_{\kappa}$ with the induced topology from the product
topology on a countable product of $[0,1]$ with itself so that, in particular, it is compact Polish.
Observe that the overlap array \prettyref{eq:overlap-def} is in $\mathcal{R}_{\kappa}$ for each $N$.
We now introduce the aforementioned invariance property. For any $m,p\geq1$, $(\nu_{s}^{k})_{s\in\mathscr{S},k\in m}\in\mathbb{R}^{\kappa}$, and $\phi:\mathbb{R}^{m}\to\mathbb{R}$,
let
\begin{equation}
\mathcal{Q}_{\ell,\ell'}=\varphi\left( \bigl[ \bigl((R_{\ell,\ell'}^s) ^{\circ p}\nu_{1}^{s},\nu_{1}^{s} \bigr),\dots,\bigl( (R_{\ell,\ell'}^s)^{\circ p}\nu_{m}^{s},\nu_{m}^{s} \bigr) \bigr]_{s \in \mathscr{S}} \right).
\end{equation}
The $(R_{\ell,\ell'}^s)^{\circ p}$ term appearing above is the Hadamard $p^{\mathrm{th}}$ power of $R_{\ell,\ell'}^s$. We say that a random variable with values in $\mathcal{R}_{\kappa}$ is weakly
exchangeable if
\begin{equation}
R_{\pi_{s}(\ell)\pi_{s}(\ell')}^s\stackrel{(d)}{=} R_{\ell,\ell'}^s
\end{equation}
for all collections $(\pi_{s})$ of permutations of $\mathbb{N}$ of finitely
many coordinates. We say that a random variable in $\mathcal{R}$ is $IP$-invariant
if for all $n\geq2$, bounded $f$, and choice of $\mathcal{Q}$ as above, we
have
\begin{equation}
\E f(R^{n})\mathcal{Q}_{1,n+1}=\frac{1}{n}\E f(R^{n})\E \mathcal{Q}_{1,2}+\frac{1}{n}\sum_{\ell=2}^{n}\E f(R^{n})\mathcal{Q}_{1,\ell},\label{eq:MSPGGI}
\end{equation}
where $R^{n}$ is the array $(R_{\ell,\ell'}^s)_{\ell,\ell'\in[n],s\in\mathscr{S}}$.
We now turn to the main result of this section. Recall that $\rho^{s}$
is a probability measure on $\mathscr{S}$, which we denote equivalently by
${\rm d}\rho$. Let
\begin{equation}
\bar{R}_{\ell,\ell'}=\int R_{\ell,\ell'}^s\,{\rm d}\rho.
\end{equation}
Our goal is to prove that random arrays in $\mathcal{R}$ that satisfy \prettyref{eq:MSPGGI}
synchronize, in the sense that $R_{\ell,\ell'}^s$ is actually a Lipschitz
function of the trace.
\begin{theo}
\label{thm:IP-characterization}Suppose that $R$ is a $\mathcal{R}_\kappa$-valued
random variable that is $IP$-invariant and weakly exchangeable. Then
there are deterministic, Lipschitz functions $\Psi_{s}$, depending
on the law of $R$, such that
\begin{equation}
R_{\ell,\ell'}^s =\Psi_{s} \bigl(\operatorname{tr} (\bar{R}_{\ell,\ell'}) \bigr)
\end{equation}
almost surely.
\end{theo}
The proof of this result is essentially by composition of the synchronization
theorems from \cite{P1,P2}.
\begin{lem}
Suppose that $R$ in $\mathcal{R}$, is weakly exchangeable and $IP$-invariant.
Then there exists deterministic, Lipschitz functions, depending on
the law of $\operatorname{tr} (\bar{R}_{\ell,\ell'})$, such that
\begin{equation}
\operatorname{tr}(R_{\ell,\ell'}^s)=L_{s}\bigl(\operatorname{tr} (\bar{R}_{\ell,\ell'}) \bigr)
\end{equation}
\end{lem}
\begin{proof}
We begin by observing that for any element of $\mathcal{R}$, the array of
traces $\bigr(\operatorname{tr} (R_{\ell,\ell'}^s ) \bigr)_{\ell, \ell' \geq 1}$ is a Gram matrix for each $s$. To
see this simply observe that if we let
\[
V_{\ell}(s)=\sum_{i}v_{i}(\ell,s)\otimes e_{i}
\]
then
\[
\operatorname{tr} (R_{\ell,\ell'}^s)=\left(V_{\ell}(s),V_{\ell'}(s)\right)_{HS}.
\]
Thus $T=\bigl(\operatorname{tr} (R_{\ell,\ell'}^s) \bigr)$ is a Gram-De Finetti array for each
$s\in\mathscr{S}.$ Applying \eqref{eq:MSPGGI} with $f$ and $\varphi$ being
functions of $T$, we see that $T$ satisfies the Multispecies Ghirlanda-Guerra
Identities \cite[Eq. (36)]{P1}
\[
\E f(T^{n}) \mathcal{Q}_{1,n+1}=\frac{1}{n}\E f(T^n)\cdot\E \mathcal{Q}_{1,2}+ \frac{1}{n} \sum_{\ell=2}^{n}\E f(T^{n})\mathcal{Q}_{1,\ell}.
\]
Recall from \cite[Section 4]{P1} that for such arrays there exist
Lipschitz functions $L_{s}$, that depend on the law of $T^{n}$,
such that
\[
T_{\ell,\ell'}^s=L_{s}\bigg(\int T_{\ell,\ell}^s\,{\rm d}\rho\bigg)
\]
almost surely. Applying this result to our setting yields a family
of Lipschitz functions depending on the law of $\operatorname{tr} (\bar{R}_{\ell,\ell'}) $
such that
\[
\operatorname{tr} (R_{\ell,\ell'}^s) =L_{s}\bigl(\operatorname{tr} (\bar{R}_{\ell,\ell'}) \bigr)
\]
almost surely, as desired.
\end{proof}
\begin{lem}
Suppose that $R$ in $\mathcal{R}_\kappa$, is weakly exchangeable and $IP$-invariant.
Then there exist deterministic, Lipschitz, monotone functions $\Theta^{s}:\mathbb{R}_{+}\to\Gamma_{\kappa}$
which depend on the law of $R$ such that
\[
R_{\ell,\ell'}^s=\Theta^{s}\bigl(\operatorname{tr} (R_{\ell,\ell'}^s) \bigr)
\]
almost surely.
\end{lem}
\begin{proof}
For fixed $s\in\mathscr{S}$, we may apply \eqref{eq:MSPGGI}, with $\mathcal{Q}$ of
the form
\[
\mathcal{Q}_{\ell,\ell'}=\varphi\Big( \bigl((R_{\ell,\ell'}^s) ^{\circ p}\nu_{1}^{s},\nu_{1}^{s} \bigr),\dots,\bigl( (R_{\ell,\ell'}^s)^{\circ p}\nu_{m}^{s},\nu_{m}^{s} \bigr) \Big).
\]
As a result, taking $f$ to be a function of this species as well,
we see that the array satisfies
\[
\E f\bigl(R^{n}(s)\bigr)\mathcal{Q}_{1,n+1}=\frac{1}{n}\E f\bigl(R^{n}(s) \bigr)\E \mathcal{Q}_{1,2}+\frac{1}{n}\sum_{\ell=2}^{n}\E f\bigl(R^{n}(s)\bigr)\mathcal{Q}_{1,\ell},
\]
where $R^n(s) = (R^s_{\ell, \ell'})_{\ell, \ell' \in [n]}$. It was shown in \cite[Theorem 3]{P2} that for such arrays,
there is a Lipschitz, monotone function $\Theta^{s}$ depending on
the law of $R^{n}(s)$ such that
\[
R_{\ell,\ell'}^s=\Theta^{s} \bigl( \operatorname{tr} (R_{\ell,\ell'}^s) \bigr)
\]
almost surely, as desired.
\end{proof}
\begin{proof}[\textbf{\emph{Proof of \prettyref{thm:IP-characterization}}}]
Applying the previous two lemmas we obtain families $(\Theta^{s})_{s \in \mathscr{S}}$ and $\left(L_{s}\right)_{s \in \mathscr{S}}$. The result then follows by taking
\[
\Psi_{s}=\Theta^{s}\circ L_{s}.
\]
\end{proof}
\section{Perturbation for Invariance}\label{sec:perturbation}
In this section, we show that after a small perturbation, the limiting overlap array will satisfy a generalized form of the Ghirlanda Guerra identities appearing in \cite{P1} and \cite{P2}. This argument is standard and can be safely skipped by the expert reader.
For completeness we include it here. The key observation is that, as with the Potts model, it is crucial that we restrict ourselves to configurations with fixed proportions of states.
Let $h_\theta(\sigma)$ be a Gaussian process with covariance
\begin{equation}\label{eq:CovPerturbed}
C^\theta_{\ell, \ell'} = \Cov \bigl( h_\theta(\sigma^{\ell})h_\theta(\sigma^{\ell'}) \bigr)= \prod_{s \in \mathscr{S}} \prod_{j \leq m} \Big( \bigl(( R_{\ell, \ell'}^s)^{\circ p } \nu_j^s, \nu_j^s \bigr) \Big)^{n_j^s}.
\end{equation}
where $\theta = (m , p_s, n_1^s, \dots, n_m^s, \nu_1^s, \dots, \nu_m^s)_{s \in \mathscr{S}}$ are the parameters in the covariance. The Gaussian process $h_\theta(\sigma)$ can be constructed explicitly using a similar construction as in \cite[Section 5]{P2}. We will provide a brief non-constructive existence proof here.
\begin{lem}
The covariance structure $C^\theta_{\ell, \ell'}$ is positive semidefinite for all $(R^s_{\ell,\ell'})_{s \in \mathscr{S}} \in \mathcal{R}_{\kappa}$.
\end{lem}
\begin{proof}
Clearly if $R^s_{\ell,\ell'} \in \mathcal{R}_{\kappa}$, then $( R_{\ell, \ell'}^s)^{\circ p } \in \mathcal{R}_{\kappa}$ for all $p \geq 1$. By the definition on \eqref{eq:arraydefinition}, we can find some collection of vectors $\bigl(v_{i}(\ell,s)\bigr)_{i,\ell\geq1,s\in\mathscr{S}}$ such that
\[
( R_{\ell, \ell'}^s)^{\circ p } = \sum_{i}v_{i}(\ell,s)\otimes v_{i}(\ell',s) = \sum_{i}v_{i}(\ell,s) v^{\mathrm{T}}_{i}(\ell',s).
\]
Given $\nu^s_j \in \mathbb{R}^\kappa$, we have
\begin{equation}\label{eq:gram}
\bigl(( R_{\ell, \ell'}^s)^{\circ p } \nu_j^s, \nu_j^s \bigr)_{\ell, \ell' \geq 1} = \Bigl( \sum_{i}v_{i}(\ell,s) v^{\mathrm{T}}_{i}(\ell',s) \nu_j^s, \nu_j^s \Bigr)_{\ell, \ell' \geq 1} = \sum_{i} \bigl( v_{i}(\ell',s)^{\mathrm{T}} \nu_j^s, v_{i}(\ell,s)^{\mathrm{T}} \nu_j^s \bigr)_{\ell, \ell' \geq 1}
\end{equation}
is a Gram array and hence positive semidefinite. Since Hadamard products preserves positivity,
\[
\prod_{s \in \mathscr{S}} \prod_{j \leq m} \Big( \bigl(( R_{\ell, \ell'}^s)^{\circ p } \nu_j^s, \nu_j^s \bigr) \Big)^{n_j^s}
\]
is positive semidefinite because it is the Hadamard product of finitely many arrays of the form \eqref{eq:gram}. Hence there exists a Gaussian process indexed with $\sigma^\ell$ with covariance given by \eqref{eq:CovPerturbed}.
\end{proof}
\iffals
\begin{proof}
For $p \geq 1$, we will use the notation
\[
e^s = (i_1, \dots, i_p) \in I_s^p, ~ \sigma_{e^s} = (\sigma_{i_1}, \dots, \sigma_{i_p})
\]
for a given $\sigma \in [\kappa]^N$. This notation is the same as \cite[Section 5]{P2}, but with indices restricted to species $s$. We will denote
\[
v_{k}(\sigma^\ell_{e^s}) = \frac{\1(\sigma^\ell_{i_1} = k) \cdots \1(\sigma^\ell_{i_p} = k)}{N_s^{p/2}}.
\]
We define the following $N^{p} \times \kappa$ matrices. Let $(e_i^s)_{i = 1}^{N_s^p}$ be an enumeration of $e^s$. If we define the following $N_s^p \times \kappa$ matrix
\begin{equation}
V^s_\ell = \begin{bmatrix}
v_{1}\bigl(\sigma^\ell_{e_1^s}\bigr) & \cdots & v_{\kappa}\bigl(\sigma^\ell_{e_1^s}\bigr)\\
\vdots & \ddots & \vdots\\
v_{1}\bigl(\sigma^\ell_{e_{N_s^p}^s}\bigr) & \cdots & v_{\kappa} \bigl(\sigma^\ell_{e_{N_s^p}^s} \bigr)
\end{bmatrix}
\end{equation}
then a direct computation will show
\[
\bigl( R^s_{\ell, \ell'} \bigr)^{\circ p_s} = \bigl( V^s_{\ell'} \bigr)^T V^s_{\ell}
\]
where $\bigl( R^s_{\ell, \ell'} \bigr)^{\circ p_s}$ is the Hadamard product of the overlap matrix \eqref{eq:overlap-def}. Given $\nu^s_j \in \mathbb{R}^\kappa$, we have
\begin{equation}
\bigl(( R_{\ell, \ell'}^s)^{\circ p } \nu_j^s, \nu_j^s \bigr)_{\ell, \ell' \geq 1} = \bigl(\bigl( V^s_{\ell'} \bigr)^T V^s_{\ell} \nu_j^s, \nu_j^s \bigr)_{\ell, \ell' \geq 1} = \bigl( V^s_{\ell} \nu_j^s, V^s_{\ell'} \nu_j^s \bigr)_{\ell, \ell' \geq 1}
\end{equation}
is positive semidefinite because it is a Gram array. Since Hadamard products preserves positivity,
\[
\prod_{s \in \mathscr{S}} \prod_{j \leq m} \Big( \bigl(( R_{\ell, \ell'}^s)^{\circ p } \nu_j^s, \nu_j^s \bigr) \Big)^{n_j^s}
\]
is positive semidefinite because it is the Hadamard product of finitely many matrices of the form \eqref{eq:gram}. Hence there exists a Gaussian process indexed with $\sigma^\ell$ with covariance given by \eqref{eq:CovPerturbed}.
\end{proof}
\fi
Let $\nu^s_j$ take rational values in $[-1,1]^\kappa$ and define the space of parameters
\begin{equation}
\Theta := \left\{ \theta : m,p, n_1^s, \dots, n_m^s \in \mathbb{N}, \nu_1^s, \dots, \nu_m^s \in \mathbb{Q} \cap [-1,1] \text{ for all $s \in \mathscr{S}$} \right\}.
\end{equation}
Since $\Theta$ is countable, we can find a enumeration map $j(\theta) : \Theta \to \mathbb{N}$. Let $(u_\theta)_{\theta \in \Theta}$ be i.i.d uniform random variables on $[1,2]$ and let $(h_\theta)_{\theta \in \Theta}(\sigma)$ of be pairwise independent copies of $h_\theta$. Finally, define
\begin{equation}
h_N(\sigma) = \sum_{\theta \in \Theta} \frac{1}{2^{j(\theta)}} u_{\theta} h_{\theta} (\sigma).
\end{equation}
Let $d_N \in \mathscr{D}_N$ be such that $d_N \to d \in \mathscr{D}$, and consider the perturbed Gibbs measure on $\Sigma_N(d_N)$ given by
\begin{equation}\label{eq:Perturbed-Gibbs-Measure}
G^{pert}_{d_N} = \frac{\exp H_N^{pert} (\sigma)}{Z_N(d_N)}, ~H_N^{pert} = H_N(\sigma) + s_N h_N(\sigma),
\end{equation}
where $\sigma \in \Sigma_N(d_N)$ and $s_N = N^\alpha$ for $1/4 < \alpha < 1/2$. We then have the following.
\begin{theo}\label{thm:MSPOTTSGG}
There is a choice of $(u_\theta)$ such that the following holds
\begin{itemize}
\item The perturbation is small in the sense that
\[
\lim_{N \to \infty}\abs{\frac{1}{N} \E \log \sum_{\sigma \in \Sigma_N(d)} \exp(H_N(\sigma))-\frac{1}{N} \E \log \sum_{\sigma \in \Sigma_N(d)} \exp(H^{pert}_N(\sigma))}=0
.\]
\item If $ R^N = (R^s_{\ell,\ell'})_{s \in \mathscr{S}, \ell,\ell' \geq 1}$ is the overlap array drawn from $\E (G^{pert}_{d_N})^{\infty}$, then any weak limit point, $ R^\infty$, satisfies \eqref{eq:MSPGGI}.
\end{itemize}
\end{theo}
\begin{proof}
The proof of this fact is almost identical to Chapter 3.2 in \cite{P3}, so we omit most details. The only essential difference
is the same as that in \cite{P2}, namely to point out why restricting the configuration space is important.
This is because the main integration by parts step in the proof of \cite[Theorem 3.2]{P3} uses in an essential
way that the self overlap, $R_{\ell, \ell}$, and thus the variance of the field $h_N$, is constant. In our setting,
the relevant term, namely $C_{\ell,\ell}^\theta$ is plainly constant on $\Sigma_N(d_N)$ by inspection of \prettyref{eq:CovPerturbed}.
\end{proof}
\section{Upper Bound - Guerra Interpolation}\label{sec:Guerra-UB}
We now turn to proving the upper bound for the restricted free energy in \eqref{eq:ParisiConstrained} by a Guerra
interpolation argument \cite{Guerra}. Recall the definition of $\Sigma_N^\epsilon(d)$ \eqref{eq:constrainedspace} and let the corresponding partition function
be denoted by $Z_N^\epsilon(d) = \sum_{\sigma \in \Sigma_N^\epsilon(d)} \exp\bigl(H_N(\sigma)\bigr)$. We will prove
\begin{equation}
\limsup_{N \to \infty} \frac{1}{N} \E \log Z_N^\epsilon(d) \leq \mathscr{P}\bigl(r, x, d, (\lambda^s,Q^s)_{s \in \mathscr{S}}\bigr)+ O(\epsilon).
\end{equation}
Given a sequence of strictly increasing $(x_i)_{i=1}^r$ as in \eqref{eq:RPC6}, let $(v_\alpha)_{\alpha \in \mathbb{N}^r}$ be the weights of the Ruelle probability cascades associated with that sequence. For $\alpha, \beta \in \mathbb{N}^r$, define
\begin{equation}
|\alpha \wedge \beta| = \min\{ 0 \leq p \leq r-1 \mathrel{}\mid\mathrel{} \alpha_1 = \beta_1, \dots, \alpha_p = \beta_p, \alpha_{p+1} \neq \beta_{p+1} \}
\end{equation}
and $|\alpha \wedge \beta| =r$ if $\alpha = \beta$. For each species, let $(Z_{s}^{\alpha}(k))_{k\leq\kappa}$ be the centered Gaussian vector with covariance
\begin{equation}\label{eq:z-def}
\E Z_{s}^{\alpha}Z_{s}^{\beta} = 2\sum_{t \in \mathscr{S}}\Delta_{st}^{2}\rho_N^{t} Q_{\abs{\alpha\wedge\beta}}^{t}.
\end{equation}
Similarly let
\begin{equation}\label{eq:y-def}
\E Y^{\alpha}Y^{\beta}=\sum_{s,t \in \mathscr{S}}\Delta_{st}^{2}\rho_N^{s}\rho_N^{t}\left(Q_{\abs{\alpha\wedge\beta}}^{s},Q_{\abs{\alpha\wedge\beta}}^{t}\right).
\end{equation}
For each $s\in\mathscr{S}$ and each $i\in I_s$, let $Z_i^\alpha$ be an independent copy of $Z_s^\alpha$. The processes $Z_i^\alpha$, $Y^\alpha$, and $H_N(\sigma)$ are all independent. Finally
define the interpolating Hamiltonian,
\begin{equation}
H_{N}(\sigma,\alpha;t)=\sqrt{t}H_{N}(\sigma)+\sqrt{t}\sqrt{N}Y^{\alpha}+\sqrt{1-t}\sum_{i \leq N}Z_{i}^{\alpha}(\sigma_{i}),
\end{equation}
and the corresponding interpolating free energy function
\begin{equation}
\phi^\epsilon_{N}(t)=\frac{1}{N}\E\log\sum_{\alpha \in \mathbb{N}^r }v_{\alpha}\sum_{\sigma\in \Sigma^\epsilon_N(d)}e^{H_{N}(\sigma,\alpha;t)}.
\end{equation}
We then have the following result.
\begin{lem}\label{lem:GuerraInterpolation}
For any $\epsilon > 0$ and $N \geq \frac{1}{\epsilon}$,
\[
\partial_t \phi^\epsilon_N(t) \leq C \epsilon^2.
\]
for some constant $C(\kappa, \Delta, |\mathscr{S}|)$, uniformly in $N$.
Furthermore, if $d\in\mathscr{D}_N$ and $\epsilon=0$, then
\[
\partial_t \phi_N^0(t)\leq 0.
\]
\end{lem}
\begin{proof}
For any $d \in \mathscr{D}$, the set $\Sigma_N^\epsilon(d)$ is non-empty for $N \geq \frac{1}{\epsilon}$. Recall that by Gaussian integration by parts,
\begin{align}
\partial_t \phi^\epsilon_N(t) & = \frac{1}{N}\E\Big\langle \partial_{t}H(\sigma,\alpha)\Big\rangle \notag\\
& = \frac{1}{N} \E\Big\langle \E \partial_{t}H(\sigma^{1},\alpha^{1})H(\sigma^{1},\alpha^{1})- \E\partial_{t}H(\sigma^{1},\alpha^{1})H(\sigma^{2},\alpha^{2}) \Big \rangle \label{eq:upbdtemp}
\end{align}
where $\langle \cdot \rangle$ is with respect to the Gibbs measure $G(\sigma, \alpha)\propto v_\alpha \exp(H_N(\sigma, \alpha;t))$ on $\Sigma_N^\epsilon(d) \times \mathbb{N}^r$. If we write $Z^\alpha_i(\sigma_i) = \sum_{k \leq \kappa}Z_i^\alpha(k) \1 (\sigma_i = k)$ then
\begin{align*}
\E\partial_{t}H(\sigma^{1},\alpha^{1})H(\sigma^{2},\alpha^{2}) & =\frac{1}{2}\Big(\E H_{N}(\sigma^{1})H_{N}(\sigma^{2})+N\E Y^{\alpha^{1}}Y^{\alpha^{2}}-\sum_{s \in \mathscr{S}}\sum_{i\in I_{s}}\E Z_{s}^{\alpha^1}(\sigma_{i}^{1})Z_{s}^{\alpha^2}(\sigma_{i}^{2})\Big)\\
& =\frac{N}{2}\sum_{s,t \in \mathscr{S}} \rho_N^s \rho_N^t\Delta_{st}^{2}\Big[\big( R_{1,2}^s, R_{1,2}^t\big)+\big(Q_{\abs{\alpha^1\wedge\alpha^2}}^{t},Q_{\abs{\alpha^1\wedge\alpha^2}}^{s}\big)- 2\big(Q_{\abs{\alpha^1\wedge\alpha^2}}^{t}, R^s_{1,2}\big)\Big]\\
& =\frac{N}{2}\sum_{s,t\in \mathscr{S}}\Delta_{st}^{2}\left( R^s_{1,2}-Q_{\abs{\alpha^1\wedge\alpha^2}}^{s}, R^t_{1,2}-Q_{\abs{\alpha^1\wedge\alpha^2}}^{t}\right)\rho_N^{s}\rho_N^{t}\geq0.
\end{align*}
The last quantity is non-negative comes from the fact that if we
define the matrix
\[
A_{1,2,\alpha^1,\alpha^2}=\left(\left( R^s_{1,2} -Q_{\abs{\alpha^1\wedge\alpha^2}}^{s}, R^t_{1,2}-Q_{\abs{\alpha^1\wedge\alpha^2}}^{t}\right)\right)_{s,t}
\]
then it is positive definite as it is a Gram-Matrix. The Hadamard product of this matrix with the positive definite matrix $\Delta = (\Delta^2_{s,t})_{s,t}$, is still positive definite by the Schur product theorem. Thus the above expression can be written as
\[
\frac{N}{2}\left(\Delta^{2}\circ (A_{1,2,\alpha^1,\alpha^2})\rho,\rho\right)\geq0,
\]
where $\circ$ denotes the Hadamard product.
For $\sigma \in \Sigma_N^\epsilon(d)$, the self overlap matrix $R^s_{1,1}$ is a diagonal matrix with entries $R^s_{1,1}(k,k) \in [d^s_k - \epsilon, d_k^s + \epsilon]$. Therefore the diagonal terms of $R^s_{1,1} - Q^s_{|\alpha^1, \alpha^1|}$ satisfy
\[
|R^s_{1,1}(k,k) - Q^s_{r}(k, k)| \leq \epsilon, \quad \forall s \in \mathscr{S}, k \leq \kappa.
\]
Since there are $\kappa$ non-zero terms in the inner product $\left( R^s_{1,1} -Q_{\abs{\alpha^1\wedge\alpha^1}}^{s}, R^t_{1,1}-Q_{\abs{\alpha^1\wedge\alpha^1}}^{t}\right)$, our upper bound of the diagonals imply
\[
\E\partial_{t}H(\sigma^{1},\alpha^{1})H(\sigma^{1},\alpha^{1}) = \frac{N}{2}\left(\Delta^{2}\circ (A_{1,1,\alpha^{1}\alpha^{1}})\rho,\rho\right) \leq \frac{N\|\Delta\| |\mathscr{S}|^2\kappa}{2} \epsilon^2.
\]
Putting this back in \eqref{eq:upbdtemp}, implies
\[
\partial \phi^\epsilon_N (t) = \E\left\langle\frac{1}{2}\left(\Delta^{2}\circ (A_{1,1,\alpha^{1},\alpha^{1}})\rho,\rho\right)\right\rangle -\E\left\langle\frac{1}{2}\left(\Delta^{2}\circ (A_{1,2,\alpha^{1},\alpha^{2}})\rho,\rho\right)\right\rangle\leq \frac{\|\Delta\| |\mathscr{S}|^2\kappa}{2} \epsilon^2.
\]
In the case $d \in \mathscr{D}_N(d)$, the set $\Sigma_N(d) = \Sigma_N^0(d)$ is non-empty. For $\sigma \in \Sigma_N(d)$, the overlap array $R^s_{\ell, \ell'} = Q^s_{r}$ for all $s \in \mathscr{S}$. In particular, $A_{1,1,\alpha^{1},\alpha^{1}} = 0$, which implies
\[
\partial \phi^0_N (t) = -\E\left\langle\frac{1}{2}\left(\Delta^{2}\circ (A_{1,2,\alpha^{1},\alpha^{2}})\rho,\rho\right)\right\rangle\leq 0.
\]
\end{proof}
As a consequence $\frac{1}{N}\E\log Z^\epsilon_{N}(d)$ is bounded above by
\begin{equation}\label{eq:upbdtemp2}
\frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma^\epsilon_N(d)} \exp\sum_{i \leq N} Z^\alpha_{i} (\sigma_i) - \frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \exp \sqrt{N} Y^\alpha + O(\epsilon^2).
\end{equation}
We now introduce the Lagrange multipliers $(\lambda_k^s)_{k \leq \kappa} \in \mathbb{R}^k$, which are
dual to the proportions $\sum_{i\in I_s} \1(\sigma_i=k)$.
If we add and subtract $\sum_{i \leq N}\sum_{k \leq \kappa}\lambda^{s(i)}_{k} \1(\sigma_i = k)$ in the exponent of the first term in \eqref{eq:upbdtemp2}, then the first term is bounded by
\[
-\sum_{s \in \mathscr{S}}\sum_{k \leq \kappa} \rho_N^s d^s_k \lambda^s_k + \frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma^\epsilon_N(d)} \exp\sum_{i \leq N} \Big(Z^\alpha_{i} (\sigma_i) + \sum_{k \leq \kappa} \1(\sigma_i = k) \lambda_k^{s(i)} \Big) + O(\epsilon).
\]
Since $\Sigma_N^\epsilon(d) \subset \Sigma_N$, summing over $\sigma \in \Sigma_N$ in the larger set only increases our upper bound. We can now factor into species using the basic properties of the Ruelle Probability Cascades (see the discussion after Theorem 2.9 in \cite{P3}) to see
\begin{align}
&\quad \frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma_N} \exp\sum_{i \leq N} \Big( Z^\alpha_{i} (\sigma_i) + \sum_{k \leq \kappa} \1(\sigma_i = k) \lambda_k^{s(i)} \Big) \notag\\
&= \frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \prod_{i \leq N} \sum_{\sigma_i \leq \kappa} \exp \Big(Z^\alpha_{i} (\sigma_i) + \sum_{k \leq \kappa} \1(\sigma_i = k) \lambda_k^{s(i)} \Big) \notag\\
&= \sum_{s \in \mathscr{S}} \rho_N^s \E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \leq \kappa} \exp \Big(Z^\alpha_s (\sigma) + \lambda_\sigma^s \Big)\notag\\
&= \sum_{s \in \mathscr{S}} \rho_N^s X_0^s.\label{eq:upbdzterm}
\end{align}
where $X_0^s$ was defined in \eqref{eq:recursion}. Similarly, we see
\begin{equation}\label{eq:upbdyterm}
\frac{1}{N}\E\log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \exp \sqrt{N} Y^\alpha = \frac{1}{2} \sum_{\ell = 0}^{r-1} x_{\ell} \sum_{s,t \in \mathscr{S}} \Delta^2_{st} \rho_N^s \rho_N^t \left( (Q^s_{\ell+1},Q^t_{\ell+1}) - (Q^s_{\ell},Q^t_{\ell}) \right).
\end{equation}
Referring back to \eqref{eq:upbdtemp2}, equations \eqref{eq:upbdzterm} and \eqref{eq:upbdyterm} imply
\begin{equation}\label{eq:guerra-ub-thick}
\limsup_{N \to \infty} \frac{1}{N} \E \log Z_N^\epsilon(d) \leq \mathscr{P}\bigl(r, x, d, (\lambda^s,Q^s)_{s \in \mathscr{S}}\bigr)+ O(\epsilon).
\end{equation}
In the case $d_N \in \mathscr{D}_N$, $\sum_{i \in I_s} \1(\sigma_i = k) = N_s d_k^s$ for all $s \in \mathscr{S}$ and $k \leq \kappa$. The above computation implies
\begin{equation}\label{eq:guerra-ub}
\frac{1}{N} \E \log Z_N(d_N) \leq \mathscr{P}\bigl(r, x, d, (\lambda^s,Q^s)_{s \in \mathscr{S}}\bigr).
\end{equation}
\section{Continuity and Decoupling Theorems}\label{sec:continuity}
Before we turn to the proof of the matching lower bounds, we briefly pause to study
the analytical properties of some of the functionals used in the upper bound,
as well as relevant functionals for the lower bound.
These functionals will be in terms of the Gaussian processes $(Z_s^\alpha)$ and $(Y^\alpha)$
from \eqref{eq:z-def} and \eqref{eq:y-def} respectively.
Many of the proofs in this section are essentially identical to arguments either from
\cite[Section 3]{P2}, or are standard arguments and can be seen, for example, in \cite{P3}. Thus to make the presentation
concise, we explain only the parts where these arguments deviate from standard arguments
and outline the rest.
\subsection{Decoupling Size of the Constraints}\label{thm:decouple}
When we compute the lower bound for the free energy, we will find that the cavity
method will naturally impose an additional constraint on free energy, namely,
that the cavity coordinates satisfy the additional constraint that they lie in some $\Sigma(d)$.
As this constraint does not appear in the upper bound from \prettyref{sec:Guerra-UB}, we will need to remove this to obtain
the matching lower bound. To this end, define the functional on $\mathscr{D}_M$
\begin{equation}\label{eq:func-constrained}
f^s_{M}(d) := \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma_M(d)} \exp \sum_{i \leq M} Z_i^\alpha(\sigma_i),
\end{equation}
where $Z_i^\alpha$ are i.i.d. copies of $Z_s^\alpha$. Furthermore, we fix the covariance structure of $Z_i^\alpha$ and make the dependence of $X_0^s$ on the parameter $\lambda$ explicit
\begin{equation}\label{eq:decouplefunctional}
X_0^s\bigl(\lambda\bigr) = \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_{\alpha} \sum_{\sigma \in \Sigma_M} \exp \sum_{i \leq M} Z_{s}^\alpha(\sigma_i) + \sum_{k \leq \kappa} \lambda^s_k\1(\sigma_i = k).
\end{equation}
We have the following result.
\begin{theo}
If $d_M^s \in \mathscr{D}_M$ and $\lim_{M \to \infty} d_M^s \rightarrow d^s$ then
\begin{equation}
\lim_{M \to \infty} f^s_M(d_M) = \inf_{\lambda^s} \bigg(- \sum_{k \leq \kappa} \lambda_k^s d_k^s + X_0^s\bigl(\lambda\bigr) \bigg).
\end{equation}
\end{theo}
\begin{proof}
We begin by observing that there is a constant $L$ such that
\begin{equation}\label{eq:epsilon-dilation-distance}
\sup_{d \in \mathscr{D}_M} |f^s_{M, \epsilon} (d) - f^s_M(d) | \leq L \sqrt{\epsilon},
\end{equation}
where $f^s_{M, \epsilon} (d)$ is the same functional as in \eqref{eq:func-constrained} but summed over $\sigma \in \Sigma^{\epsilon}_M(d)$. This is the analogue of \cite[Lemma 3]{P2}, adapted to the covariance structure of the Gaussian processes $Z_i$ defined in \eqref{eq:z-def}. We also fix the covariance structure of $Z_i$ to remove its dependence on $d$. Let $\sigma\in \Sigma_M^\epsilon(d)$ and $\tilde\sigma$ be a vector in $\Sigma_M(d)$
with the minimal number of different coordinates from $\sigma$.
Let
\[
\tilde f^s_{M,\epsilon} = \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma^\epsilon_M (d)} \exp \sum_{i \in S_M} Z_i^\alpha(\tilde \sigma_i).
\]
Finally, let $\tilde Z_i^\alpha$ be independent copies of $Z_i^\alpha$, and consider the ``smart path''
\[
Z_t(\alpha,\sigma)= \sum_{i\leq M} (\sqrt{t} Z_i^\alpha(\sigma)+\sqrt{1-t}\tilde Z_i^{\alpha}(\tilde\sigma_i).
\]
We will then show that
\[
\phi(t) = \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in \Sigma^\epsilon_M(d)} \exp Z_t(\alpha, \sigma),
\]
has a derivative that is bounded by $C\epsilon$. To see this, let
\begin{align*}
C\left((\sigma^1, \alpha^1), (\sigma^2, \alpha^2)\right) &= \frac{1}{M} \frac{\partial Z_{t} (\sigma^1, \alpha^1)}{\partial_t} Z_{t} (\sigma^2, \alpha^2)
\\&= \frac{1}{M} \sum_{i \leq M} \left(\sum_{t \in \mathscr{S}}\Delta_{st}^{2}\rho^{t} \Big( Q_{\abs{\alpha\wedge\beta}}^{t} (\sigma_i^1, \sigma_i^2) - Q_{\abs{\alpha\wedge\beta}}^{t} (\tilde \sigma_i^1, \tilde \sigma_i^2) \Big)\right)
\\&\leq \|\Delta\| \cdot |\mathscr{S}| \cdot \kappa\cdot \epsilon.
\end{align*}
In the last line, we use that $Q^t\leq D^t$ and that $\norm{D^t}\leq 1$ along with the observation that
\[
\sum \1\left(\sigma_i\neq\tilde{\sigma}_i\right) \leq \lceil \kappa M \epsilon\rceil.
\]
Differentiating and integrating by parts, we obtain
\[
\abs{\phi'(t)} =\abs{\E\left\langle C((\sigma^1,\alpha^1),(\sigma^1,\alpha^1))-C((\sigma^1,\alpha^1),(\sigma^2,\alpha^2))\right\rangle}
\leq 2L\epsilon.
\]
for some constant $L=L(\kappa,\mathscr{S},\Delta)$. Integrating this inequality, we can conclude
\begin{equation}\label{eq:thm11temp}
|f^s_{M, \epsilon} - \tilde f^s_{M, \epsilon}| \leq 2 L \epsilon.
\end{equation}
For $\sigma\in \Sigma(d)$, let us denote by $\mathcal{N}(\sigma)$ the number of configurations $\rho\in\Sigma_\epsilon(d)$ such that $\tilde{\rho} = \sigma.$ Then we can rewrite and bound $\tilde{f}^s_{M,\epsilon}(d)$ as follows,
\begin{align*}
\tilde{f}_{M,\epsilon}(d)
&=
\frac{1}{M} \E \log \sum_{\alpha\in\mathbb{N}^r} v_\alpha \sum_{\sigma\in \Sigma(d)} \mathcal{N}(\sigma)\exp \beta \sum_{i\leq N} Z_{i}^\alpha({\sigma}_i).
\\
&\leq f^s_{M}(d) +\frac{1}{N}\max_{\sigma\in \Sigma(d)} \log \mathcal{N}(\sigma).
\end{align*}
Using a combinatorial argument, the term containing $\mathcal{N}(\sigma)$ can be made arbitrarily small by choosing $\epsilon$ small enough. For any $\sigma\in \Sigma(d)$, the number $\mathcal{N}(\sigma)$ is bounded by the number of configurations $\rho$ such that $\sum_{i\leq M}I(\rho_i\not= {\sigma}_i) \leq L M\epsilon$. By the classical large deviation estimate for Bernoulli random variables, a number of different ways to choose $LM\epsilon$ coordinates is bounded by $2^M \exp(-M I(1-L\epsilon)),$ where
$$
I(x)=\frac{1}{2}\bigl((1+x)\log(1+x)+(1-x)\log(1-x)\bigr),
$$
and there are $\kappa^{LM\epsilon}$ ways to choose $\rho_i$ different from $\sigma_i$ on these coordinates. Therefore,
\begin{align*}
\frac{1}{M}\max_{\sigma\in \Sigma(d)} \log \mathcal{N}(\sigma)
&\leq
L\epsilon\log\kappa+
\log 2 - I(1-L\epsilon)
\\
&=
L\epsilon\log\kappa+
\log\Bigl(1+\frac{L\epsilon}{2-L\epsilon}\Bigr)+\frac{L\epsilon}{2}\log
\frac{2-L\epsilon}{\epsilon}
\leq L\sqrt{\epsilon},
\end{align*}
for small enough $\epsilon$. We showed that ${f}_{M,\epsilon}(d)\leq f_M(d)+L\sqrt{\epsilon}$.
Combining this with \eqref{eq:thm11temp} yields the estimate
\[
|f^s_M - f^s_{M,\epsilon}|\leq 2 L \sqrt{\epsilon}.
\]
Note that \eqref{eq:epsilon-dilation-distance}, implies that the map $d\mapsto f_M^s(d)$ is
H\"older 1/2.
The remaining steps to complete the proof are identical to \cite[Section 3]{P2}. An additivity argument and \eqref{eq:epsilon-dilation-distance} will imply the following.
\begin{lem}\label{lem:decouple1}
\cite[Lemma 4,5]{P2} If $d_M \in \mathscr{D}_M$ and $\lim_{M \to \infty} d_M = d \in \mathscr{D}$ then the limit
\[
f(d) := \lim_{M \to \infty} f_M(d_M) = \lim_{\epsilon \to 0} f_{\epsilon}(d)
\]
exists and is concave. In addition, for all $d^1, d^2 \in \mathscr{D}$
\[
| f(d^1) - f(d^2)| \leq L \|d^1 - d^2\|_\infty^{1/2}
\]
for some constant $L$ that depends $\kappa, \mathscr{S}, \Delta$.
\end{lem}
Next, a direct computation using the recursive property of the Ruelle Probability cascades will imply the following result.
\begin{lem}\label{lem:decouple2} \cite[Lemma 6]{P2} For any $\lambda = (\lambda_k)_{k \leq \kappa} \in \mathbb{R}^\kappa$
\begin{equation}\label{eq:biconjugation}
X_0^s(\lambda) = \max_{d \in \mathscr{D}} \Bigl( f^s(d) + \sum_{k \leq \kappa} \lambda_k^s d_k^s \Bigr).
\end{equation}
\end{lem}
Notice $f^s_M(d)$ is continuous and bounded on $\mathscr{D}$. Since it is also concave by \prettyref{lem:decouple1}, we can take the Legendre transform of \eqref{eq:biconjugation} to obtain
\[
\lim_{N \to \infty} f^s_M (d) = \inf_\lambda \bigg( - \sum_{k \leq \kappa} \lambda^s_kd_k^s + X_0^s(\lambda) \bigg).
\]
\end{proof}
\subsection{Continuity Theorems}
Along with the decoupling theorem, we will also need to prove continuity
of the functionals appearing in the Aizenman-Sims-Starr scheme for this system.
Define the two functionals on a subset $S \subset [\kappa]^M$,
\begin{equation}\label{eq:func-pi-z}
f^Z_{M}(S,s; \bm \pi) := \frac{1}{M_s} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in S} \exp \sum_{i \leq M_s} Z_i^\alpha(\sigma_i),
\end{equation}
\begin{equation}\label{eq:func-pi-y}
f^Y_{M}(\bm \pi) := \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \exp \sqrt{M} Y^\alpha,
\end{equation}
where $(Z_i^\alpha)_{i \leq M}$ are \text{i.i.d. } copies of $Z_s^\alpha$ defined in \eqref{eq:z-def}. Observe that this functional depends on
$\bm \pi = (\pi^s)_{s \in \mathscr{S}}$ is through the covariance structures \eqref{eq:z-def} and \eqref{eq:y-def}.
\begin{lem}\label{lem:ctyofz}
For every $S \subset [\kappa]^M$, the functional $f^Z_{M}(S,s; \bm \pi)$ is Lipschitz in $\bm \pi$
\[
|f^Z_{M}(S,s; \bm \pi) - f^Z_{M}(S,s; \bm {\tilde \pi}) | \leq L \int_{0}^{1} \max_{s \in \mathscr{S}} \| \pi^s(x) - \tilde\pi^s(x) \|_1 \, {\rm d} x.
\]
\end{lem}
\begin{proof}
Observe that any two monotone paths $\bm \pi$ and $\tilde{\bm \pi}$,
can be associated with a single sequence
\[
x_{-1} = 0 \leq x_0 \leq \dots \leq x_{r-1} \leq x_r = 1\\
\]
and, for every $s\in\mathscr{S}$, two sequences
\begin{align*}
0 < Q^s_0 \leq \dots < Q^s_{r-1} < Q^s_r = \operatorname{diag}(d_1, \dots, d_\kappa)\\
0 < \tilde Q^s_0 \leq \dots < \tilde Q^s_{r-1} < \tilde Q^s_r = \operatorname{diag}(d_1, \dots, d_\kappa).
\end{align*}
Consider the Gaussian processes $Z_s^\alpha$ and $\tilde Z_s^\alpha$ with the covariance
\prettyref{eq:z-def} with $Q$ and $\tilde Q$ respectively, and consider the smart path between these
two processes
\[
Z_{t,i}^\alpha =\sqrt{t}Z_i^\alpha + \sqrt{1-t}\tilde Z_i^\alpha.
\]
If we take $M$ copies of this process, and let $(v_\alpha)$ be the Ruelle Probability Cascade \cite{Ruelle}
associated to $(x_k)$, then if we define
\[
\varphi(t) := \frac{1}{M_s} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \sum_{\sigma \in S} \exp \sum_{i \leq M_s} Z_{t,i}^\alpha (\sigma_i),
\]
we have $\phi(1)=f_Z(S,s;\bm\pi)$ and $\phi(0)=f_{Z}(S,s;\tilde{\bm{\pi}})$. If $H_{M,t}(\alpha, \sigma) = \sum_{i \leq M} Z_{t,i}^\alpha(\sigma_i)$, let $\langle \cdot \rangle_t$ denote the average with respect to the Gibbs measure
\begin{equation}\label{eq:gibbs-smartpath}
G_t(\sigma, \alpha) \sim v_\alpha \exp H_{M,t}(\alpha, \sigma).
\end{equation}
We can now compute $\varphi'(t)$ using integration by parts. We first note that the covariance
\[
\frac{1}{M_s} \E \frac{\partial H_{M,t} (\sigma^1, \alpha^1)}{\partial_t} H_{M,t} (\sigma^2, \alpha^2) = \frac{1}{M_s} \sum_{i \leq M}
\sum_{s \in \mathscr{S}} \Delta^2_{s, s(i)} \rho^s \Bigl(Q^s_{\alpha^1 \wedge \alpha^2}(\sigma_i^1, \sigma_i^2) - (\tilde Q^s_{\alpha^1 \wedge \alpha^2}(\sigma_i^1, \sigma_i^2) \Bigr).
\]
The term on the right is bounded in absolute value by $\norm{\Delta} \cdot \abs{\mathscr{S}}\max_{s\in\mathscr{S}}
\norm{Q_{\alpha^1\wedge\alpha^2}^s-\tilde Q_{\alpha^1\wedge\alpha^2}^s}_1$. Recalling that the marginals of $G_t(\alpha)$ \eqref{eq:gibbs-smartpath} on $\mathbb{N}^r$ has the same distribution as the weights of $v_\alpha$ \cite[Theorem 4.4]{P3}, a standard Gaussian integration by parts argument will show
\begin{align*}
\abs{\varphi'(t)} &\leq \E \left\langle \norm{\Delta}\cdot\abs{\mathscr{S}}\max_{s\in\mathscr{S}}
\norm{Q_{\alpha^1\wedge\alpha^2}^s-\tilde Q_{\alpha^1\wedge\alpha^2}^s}_1\right\rangle_t\\
&\leq \|\Delta\||\mathscr{S}| \sum_{0 \leq p \leq r} \max_{s \in \mathscr{S}} \| Q^s_{\alpha^1 \wedge \alpha^2} - \tilde Q^s_{\alpha^1 \wedge \alpha^2} \|_1 \E \sum_{\alpha^1 \wedge \alpha^2 = p} v_{\alpha^1} v_{\alpha^2}\\
&= \|\Delta\||\mathscr{S}| \int_{0}^{1} \max_{s \in \mathscr{S}} \| \pi^s(x) - \tilde\pi^s(x) \|_1 \, {\rm d} x.
\end{align*}
Integrating this inequality yields the result.
\end{proof}
The following argument is well known and follows by a direct computation using properties of
Ruelle Probability Cascades, see \cite{P2,P3}.
\begin{lem}\label{lem:ctyofy} The functional $f^Y_{M}(\bm \pi)$ is Lipschitz in $\bm \pi$,
\[
\abs{f^Y_{M}(\bm \pi) - f^Y_{M}( \bm{\tilde \pi})} \leq L \int_{0}^{1} \max_{s \in \mathscr{S}} \| \pi^s(x) - \tilde\pi^s(x) \|_1 \, {\rm d} x.
\]
\end{lem}
We finally observe here that discrete paths are dense in $\pi$.
\begin{lem}\label{thm:discretize}
For any path $\bm \pi \in \bm \Pi$ and $\epsilon >0$, there exists a finite sequence of points $(x_p)_{p=1}^r$ such that the discrete path
\[
\bm \pi^* (x) := \bm \pi(x_p) \text{ for } x_{p-1} \leq x \leq x_p
\]
satisfies
\[
\Delta(\bm \pi, \bm {\pi} ^*) < \epsilon.
\]
\end{lem}
We now state a general continuity theorem regarding functionals of this form. Such results are completely standard, see, e.g., \cite{P2,P3}. Let $S \subset [\kappa]^M$, $(w_\alpha)_{\alpha \in \mathscr{A}}$ be the weights (possibly random) of a probability density distribution on a countable set $\mathscr{A}$, and
$$R_{\mathscr A} = (R^s_{\alpha^1, \alpha^2})_{s \in \mathscr{S}, \alpha^1, \alpha^2 \in \mathscr{A}},$$
where each $R^s_{\alpha^1, \alpha^2}$ is a $\kappa \times \kappa$ matrix. We think of this array as fixed and non-random. It will be the values that some abstract overlap structure can take. Finally, let
$(I_s)$ be partition of $[M]$ into species.
We define the functionals
\begin{equation}
f_{1,M} = \frac{1}{M} \E \log \sum_{\alpha \in \mathscr{A}} w_\alpha \sum_{\epsilon \in S} \exp \sum_{i \leq M} Z_i^\alpha(\epsilon_i)
\end{equation}
and
\begin{equation}
f_{2,M} = \frac{1}{M} \E \log \sum_{\alpha \in \mathscr{A}} w_\alpha \exp \sqrt{M} Y^\alpha.
\end{equation}
Here, $Z^\alpha_i(\sigma)$ is a centered Gaussian process is such that, if $i\in I_s\subset [M] $, the covariance structure of the Gaussian vector $Z^\alpha_i = (Z^\alpha_i(k))_{k \leq \kappa}$ is given by
\begin{equation}
\Cov (Z_i^{\alpha^1}, Z_i^{\alpha^2}) = C^{s}_Z((R^s_{\alpha^1, \alpha^2} )_{s \in \mathscr{S}}).
\end{equation}
for some $C_z^s$, a continuous function of the overlaps $R^s_{\alpha^1, \alpha^2}$.
Similarly, $Y$ is a centered Gaussian process with covariance given by
\begin{equation}
\Cov (Y^{\alpha^1}, Y^{\alpha^2}) = C_Y((R^s_{\alpha^1, \alpha^2} )_{s \in \mathscr{S}}).
\end{equation}
The following result is standard and follows from basic properties of Gaussian processes, and the fact that
log-sums of exponentials have at most linear growth at infinity. Let $(\alpha(\ell))_{\ell\geq1}$ be i.i.d. drawn from
$\mathscr{A}$ with law $w_\alpha$, and denote
\[
R^n =\bigl( R^s_{\alpha(\ell), \alpha(\ell')} \bigr)_{ \ell, \ell' \in [n], s \in \mathscr{S}}.
\]
We have the following continuity property
\begin{lem}\label{thm:approximation}
For any $\epsilon > 0$, there are continuous bounded functions $g^Z_{\epsilon}$ and $g^Y_\epsilon$
such that
\begin{equation}
| f_{1,M} - \E g^Z_\epsilon (R^n)| \leq \epsilon, \qquad | f_{2,M} - \E g^Y_\epsilon (R^n)| \leq \epsilon.
\end{equation}
These functions depend at most on $M,S,C_z, C_y$, and $\epsilon$.
\end{lem}
\section{Lower Bound via an Aizenman-Sims-Starr Scheme}\label{sec:Aizenman-LB}
For fixed $d \in \mathscr{D}$ and a sequence of realizable proportions $d_N \in \mathscr{D}_N$ \eqref{eq:feasible-constraint-space} converging to $d$, we prove the matching constrained lower bound,
\begin{equation}\label{eq:lowerbound}
\liminf_{N \to \infty} \frac{1}{N} \E \log Z_N(d_N) \geq \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}}\bigr).
\end{equation}
By part (1) of \prettyref{thm:MSPOTTSGG}, computing the lower bound of the free energy with respect to
\begin{equation}\label{eq:perthamiltonian}
H^{pert}_N(\sigma) = H_N(\sigma) + s_N h_N(\sigma)
\end{equation}
introduced in Section \ref{sec:perturbation} and the corresponding constrained partition function
\begin{equation}\label{eq:perturbed-partition-function}
Z_N(d_N) = \sum_{\sigma \in \Sigma_N(d_N)} \exp\bigl(H^{pert}_N(\sigma) \bigr).
\end{equation}
is equivalent to computing the lower bound in \eqref{eq:lowerbound}. We will continue to work with the perturbed Hamiltonian \eqref{eq:perthamiltonian} throughout the remainder of this section.
For $M \geq 1$, our starting point is the following inequality,
\begin{equation}\label{eq:ASSSCHEME}
\liminf_{N \to \infty}\frac{1}{N} \E\log Z_N(d_N) \geq \liminf_{N \to \infty} \frac{1}{M} \bigl( \E \log Z_{N + M} (d_{N + M}) - \E \log Z_N(d_N) \bigr).
\end{equation}
We will write the free energy in terms of the Gibbs measure $G_{d_N}^{pert}$ defined in \eqref{eq:Perturbed-Gibbs-Measure} using the cavity method via an Aizenman-Sims-Starr scheme \cite{AS2}. To this end, let us denote a configuration $\tilde \rho\in[\kappa]^{N+M}$ by $\tilde \rho =(\epsilon, \sigma)$, where $\epsilon = (\epsilon_1, \dots, \epsilon_M) \in [\kappa]^M$ are called the cavity coordinates and $\sigma = (\sigma_{M + 1}, \dots, \sigma_{N + M}) \in [\kappa]^N$ are called the bulk coordinates. We define $\mathcal{M}_s$, $\mathcal{N}_s$ to be the subset of the respective cavity and bulk coordinates that belong to species $s$. Let $M_s$, $N_s$ be the cardinality of $\mathcal{M}_s$ and $\mathcal{N}_s$.
We control the rate of convergence of $d_N$ so that for some constant $L_\kappa$,
\begin{equation}\label{eq:convergence-proportions}
| d^s_{N,k} - d_k^s | \leq \frac{L_\kappa}{N_s}~\text{for all $k \leq \kappa$, $s \in \mathscr{S}$} \text{ and $d^s_{N,k} = 0$ if $d_k^s = 0$}.
\end{equation}
A generalized version of \cite[Lemma 11]{P2} will allow us to split constrained configuration space into a product set of the bulk coordinates and the species wise cavity coordinates along a subsequence.
We first introduce some more notation. Let $\mathcal{A} \subset N + M$ and $\mathcal{A}_s$ is the subset of $\mathcal{A}$ in species $s$. If $A := |\mathcal{A}|$ is the cardinality of the set $\mathcal{A}$, we define
\begin{equation}\label{def:subset-proportions}
\Sigma_{A}(d) := \Bigl \{ (\sigma_i)_{i \in A} \mathrel{}\Bigm{|} \mathrel{} \sum_{i \in \mathcal{A}_s} \1(\sigma_i = k) = A_s d^s_{k}, \forall k \leq \kappa, s \in \mathscr{S} \Bigr\}
\end{equation}
to be the configurations of spins in $\mathcal{A}$ that satisfy the constraint $d$. If $\mathcal{A} = N + M$, then \eqref{def:subset-proportions} coincides with \eqref{eq:constrainedspace}. The definition in \eqref{def:subset-proportions} naturally implies
\begin{equation}
\Sigma_A(d) = \prod_{s \in \mathscr{S}} \Sigma_{A_s}(d^s) \label{lem16temp}.
\end{equation}
If we use the sets $\mathcal{N}$ and $\mathcal{M}$ in place of $\mathcal{A}$ in \eqref{lem16temp}, the observations means we can break the cavity and bulk coordinates into a product set over species. We will now show that the entire constrained system $\Sigma_{N+ M}(d)$ contains a product set over the bulk and cavity coordinates.
\begin{lem}
For every $M > 0$, there exists a constraint $\delta_{M} \in \mathscr{D}_{M}$ such that
\begin{equation}
|\delta^s_{M,k} - d_k^s| \leq \frac{2 L_\kappa}{M_s} \text{ for all $k \leq \kappa$, $s \in \mathscr{S}$},
\end{equation}
and we can find a subsequence of $N$ such that
\begin{equation}\label{eq:Separate_into_product_sets}
\Sigma_{N + M}(d_{N + M}) \supseteq \Sigma_{N}(d_{N}) \times \Sigma_{M} (\delta_{M}).
\end{equation}
\end{lem}
\begin{proof}
We first fix $s \in \mathscr{S}$ and apply \cite[Lemma 11]{P2} to the subset of spins in $\Sigma_{N + M}$ belonging to species $s$. There exists a sequence $\delta^s_{M} \in \mathscr{D}_{M}$ such that $|\delta^s_{M,k} - d^s_k| \leq \frac{2 L_\kappa}{M_s}$ for all $k \leq \kappa$, and
\begin{equation}\label{eq:additive}
N_s d^s_N + M_s \delta^s_{M} = (N_s + M_s) d^s_{N + M}
\end{equation}
for infinitely many $N_s$. Therefore, we can find a subsequence of $N$ such that that associated $N_s$ satisfies \eqref{eq:additive} and
\[
\Sigma_{N_s + M_s} (d_{N + M}) \supseteq \Sigma_{N_s}(d_N) \times \Sigma_{M_s}(\delta_{M}).
\]
Repeating the argument over each species and extracting a further subsequence each iteration, we can conclude
\[
\prod_{s \in \mathscr{S}} \Sigma_{N_s + M_s} (d_{N + M}) \supseteq \prod_{s \in \mathscr{S}} \Sigma_{N_s}(d^s_N) \times \prod_{s \in \mathscr{S}} \Sigma_{M_s}(\delta^s_M).
\]
Writing the product sets using the observation in \eqref{lem16temp} completes the proof.
\end{proof}
With this observation, we can restrict the first sum in \eqref{eq:ASSSCHEME} to the product set \prettyref{eq:Separate_into_product_sets}, so that \eqref{eq:ASSSCHEME} is bounded below by
\begin{equation}\label{eq:LBstep1}
\frac{1}{M} \bigg( \E \log \sum_{\sigma \in \Sigma_N(d_N)} \sum_{\epsilon \in \Sigma_{M}(\delta_{M})} \exp \bigl(H^{pert}_{N + M} (\sigma, \epsilon) \bigr) - \E \log \sum_{\sigma \in \Sigma_N(d_N)} \exp \bigl(H^{pert}_N(\sigma) \bigr) \bigg).
\end{equation}
We will now separate the unperturbed portion of the Hamiltonians into its cavity fields,
\begin{align}
H_{N + M} (\epsilon,\sigma) &= H_N'(\sigma) + \sum_{i \leq M} Z_{N,i}^\sigma(\epsilon_i) + r (\epsilon), \\
H_{N} (\sigma) &= H_N'(\sigma) + \sqrt{M} Y_N^\sigma.
\end{align}
These cavity fields are the same as those appearing in \cite[Equation (107) and (109)]{P2}, except with different covariance structure because of the inhomogeneity. In our case, these fields are independent Gaussian processes with covariance
\begin{alignat}{2}
\E H'_N(\sigma^\ell)H'_N(\sigma^{\ell'}) &= \frac{N^2}{N + M} \sum_{s,t \in \mathscr{S}} \Delta^2_{s,t} \rho_N^s \rho_N^t ( R^s_{\ell,\ell'}, R^t_{\ell,\ell'}),\\
\E Z_{N,i}^{\sigma^\ell}(\epsilon) Z_{N,j}^{\sigma^{\ell'}}(\epsilon') &= 2 \delta_{\{i = j\}} \sum_{s \in \mathscr{S}} \Delta^2_{s(i),s} \rho_N^s R^s_{\ell,\ell'}(\epsilon, \epsilon') + O(N^{-1}) , \\
\E Y_N^{\sigma^\ell}Y_N^{\sigma^{\ell'}} &= \sum_{s,t \in \mathscr{S}} \Delta^2_{s,t} \rho_N^s \rho_N^t ( R^s_{\ell,\ell'}, R^t_{\ell,\ell'}) + O(N^{-1}).
\end{alignat}
Here $\rho_N^s := \frac{\sum_{M + 1 = 1}^{N + M} \1(i \in I_s)}{N}$ is the proportion of the bulk coordinates that belong to species $s$. This distinction is not critical because this proportion converges to the same $\rho^s$ defined in \eqref{eq:proportion-def}. The $r(\epsilon)$ term is $O(N^{-1})$ and can be omitted without affecting \eqref{eq:lowerbound}.
By a standard interpolation argument, see for example \cite[Theorem 3.6]{P3}, if $(z^\sigma_{N,i}(\epsilon))_{i \leq M}$ and $y_N^\sigma$ are centered Gaussian processes with covariances
\begin{align}
\E z_{N,i}^{\sigma^\ell}(\epsilon) z_{N,j}^{\sigma^{\ell'}}(\epsilon') &= 2 \delta_{\{i = j\}} \sum_{s \in \mathscr{S}} \Delta^2_{s(i),s} \rho^s R^s_{\ell,\ell'}(\epsilon, \epsilon'),\label{eq:cov-gibbs-z}\\
\E y_N^{\sigma^\ell}y_N^{\sigma^{\ell'}} &= \sum_{s,t \in \mathscr{S}} \Delta^2_{s,t} \rho^s \rho^t ( R^s_{\ell,\ell'}, R^t_{\ell,\ell'})\label{eq:cov-gibbs-y},
\end{align}
then \eqref{eq:LBstep1} is bounded below by
\begin{equation}\label{eq:LBstep2}
\frac{1}{M}\E \log \bigg \langle \sum_{\epsilon \in \Sigma_{M}(\delta)} \exp{\sum_{i \leq M} z^\sigma_{N,i} (\epsilon)} \bigg \rangle_{G'_N} -\frac{1}{M} \E \log \bigg\langle \exp{\sqrt{M} y^\sigma_N} \bigg\rangle_{G'_N} + o(1).
\end{equation}
Here $\langle \cdot \rangle_{G_N'}$ is the average in $\sigma$ with respect to the perturbed Gibbs measure \eqref{eq:Perturbed-Gibbs-Measure}.
By \prettyref{thm:approximation}, the functionals appearing in \eqref{eq:LBstep2} are continuous functionals of the distribution of the overlap matrix array
\begin{equation}
R^N = \bigl(R^s_{\ell, \ell'}\bigr)_{\ell, \ell' \geq 1, s \in \mathscr{S}}
\end{equation}
of i.i.d. draws from the perturbed Gibbs measure $G_{d_N}^{pert}$. We will now relate \eqref{eq:LBstep2} to the Ruelle Probability Cascades allowing us to compute its value explicitly.
In the following, we take a subsequence along which the limit inferior of \prettyref{eq:LBstep2} is achieved. Since $\mathcal{R}$ is compact we may take a subsequential weak limit of $R^N$ along our minimizing subsequence, which we denote by $R^\infty$. For ease of notation, we will continue to denote this subsequence with $N$. By the choice of the perturbation Hamiltonian, \prettyref{thm:MSPOTTSGG}, we have the limiting array $R^\infty$ satisfies equation \eqref{eq:MSPGGI}. By the characterization theorem, \prettyref{thm:IP-characterization}, we see the order parameter for this system, that is the quantity which determines the law of the system, will be the law of $\operatorname{tr} (\bar R^\infty_{12})$.
With this in mind, we make the following approximation. The array, $(\operatorname{tr}(\bar R^\infty_{\ell,\ell'}))_{\ell,\ell' \geq 1}$, by definition of $\mathcal{R}$, is a Gram-De Finetti array. Furthermore, taking $\nu^s_n = \sqrt{\rho^s} e_n$, $p_s = 1$, and $\phi := \sum_{s \in \mathscr{S}} \sum_{i \leq \kappa} (\nu^s_i, R^{s,\infty}_{\ell,\ell'} \nu^s_i) $
in \prettyref{eq:MSPGGI} the array $\bigl(\operatorname{tr}(\bar R^\infty_{\ell,\ell'}) \bigr)_{\ell,\ell' \geq 1}$ also satisfies the classical Ghirlanda-Guerra identities. By \cite[Thereom 2.13]{P3}, the law of $(\operatorname{tr}(\bar R^\infty_{\ell,\ell'}))_{\ell, \ell' \geq 1}$ is uniquely determined by $\zeta$, the law of $\operatorname{tr} \bigl(\bar R^\infty_{1,2} \bigr)$.
Let $\zeta^n\to\zeta$ weakly in $\Pr[0,1]$, such that $\zeta^n$ consists of a finite number of atoms. This yields sequences
\begin{align}
x^n_{-1}= 0 &< x^n_0 < \ldots < x^n_{r-1} < x^n_r = 1,
\label{eq:RPC-sequences}
\\
0 &=q^n_0 \, < \ldots < q^n_{r-1} < q^n_r = 1,
\nonumber
\end{align}
such that $\zeta^n([0,q_p^n])=x_p^n$. By \cite[Theorem 2.17]{P3}, if $\operatorname{tr} \bigl(\bar R^n_{1,2} \bigr)$ has distribution $\zeta^n$, then the approximating array of traces $ \bigl( \operatorname{tr} (\bar R^n_{1,2}) \bigr)_{\ell, \ell' \geq 1}$ converge weakly to $(\operatorname{tr}(\bar R^\infty_{\ell,\ell'}))_{\ell, \ell' \geq 1}$.
Let $v_\alpha$ be the weights of the Ruelle probability cascades associated with the sequence $(x^n)$ in \eqref{eq:RPC-sequences}. If $(a_\ell)_{\ell \geq 1}$ are i.i.d samples from $\mathbb{N}^r$ according to $v_\alpha$ then it is well known \cite[Section 3.6]{P3} that
$T^n_{\ell,\ell'} = q^n_{\alpha^\ell \wedge \alpha^{\ell'}}$ will be close in distribution to $ \bigl(\operatorname{tr}( \bar R^\infty_{\ell,\ell'}) \bigr)_{\ell,\ell' \geq 1}$. More precisely,
\begin{equation}\label{eq:approx-trace}
\bigl(T^n_{\ell,\ell'} \bigr)_{\ell,\ell' \geq 1} \stackrel{d}{\to} \bigl( \operatorname{tr}( \bar R^\infty_{\ell,\ell'}) \bigr)_{\ell,\ell' \geq 1} .
\end{equation}
Since the limiting array $R^\infty$ is IP-Invariant \eqref{eq:MSPGGI}, by \prettyref{thm:IP-characterization} we can find a family of Lipschitz functions $(\Phi_s)_{s \in \mathscr{S}}$ on $[0,1]$ such that
\[
R^\infty = (R_{\ell,\ell'}^{s,\infty})_{\ell,\ell' \geq 1, s \in \mathscr{S}} = \bigl( \Phi_s \bigl( \operatorname{tr}( \bar R^\infty_{\ell,\ell'}) \bigr) \bigr)_{\ell,\ell' \geq 1, s \in \mathscr{S}}.
\]
Since $\Phi_s$ is Lipshitz, \eqref{eq:approx-trace} implies the $\kappa \times \kappa$ matrices $Q^{s,n}_{\ell,\ell'} := \Phi_s (T^n_{\ell,\ell'} )$ will be close in distribution to $R^{s,\infty}_{\ell,\ell'}$. If we let $ Q^n := \bigl(Q^{s,n}_{\ell,\ell'} \bigr)_{\ell,\ell' \geq 1, s \in \mathscr{S}}$ denote the approximating array generated from \text{i.i.d. } samples under Ruelle probability cascades corresponding to order parameter $\zeta^n$ then \eqref{eq:approx-trace} also implies
\begin{equation}\label{eq:convergence-Q}
Q^n \stackrel{d}{\to} R^\infty.
\end{equation}
For $n$ sufficiently large, we will bound \eqref{eq:LBstep2} arbitrarily closely with functionals of the infinite array $Q^n$, which we will now show. Let $\big( (Z_{i,n}^{\alpha}(k))_{k\leq\kappa} \big)_{i \leq M}$ and $Y_n^\alpha$ be
centered Gaussian processes with covariances
\begin{align}
\E Z_{i,n}^{\alpha}Z_{j,n}^{\beta} &= 2 \delta_{\{i = j\}} \sum_{t \in \mathscr{S}}\Delta_{s(i),t}^{2}\rho^{t}Q_{\abs{\alpha\wedge\beta}}^{t,n}, \label{eq:cov-RPC-z}\\
\E Y_n^{\alpha}Y_n^{\beta} &=\sum_{s,t \in\mathscr{S}}\Delta_{s,t}^{2}\rho^{s}\rho^{t}
\left(Q_{\abs{\alpha\wedge\beta}}^{s,n},Q_{\abs{\alpha\wedge\beta}}^{t,n}\right) \label{eq:cov-RPC-y}.
\end{align}
Notice that the covariance structure in \eqref{eq:cov-RPC-z} and \eqref{eq:cov-RPC-y} depend on the overlap array in exactly the same way as \eqref{eq:cov-gibbs-z} and \eqref{eq:cov-gibbs-y}. A direct application of the continuity in \prettyref{thm:approximation} and the convergence in distribution of $ Q^n$ in \eqref{eq:convergence-Q} will imply the following result.
\begin{lem}\label{lem:RPCLWBD}
For every $\tilde \epsilon > 0$, there is an $n > 0$ such that
\begin{align}
& \liminf_{N \to \infty} \frac{1}{M}\E \log \bigg \langle \sum_{\epsilon \in \Sigma_{M}(\delta)}
\exp{\sum_{i \leq M} z^\sigma_{N,i} (\epsilon)} \bigg \rangle_{G'_N}
-\frac{1}{M} \E \log \bigg\langle \exp{\sqrt{M} y^\sigma_N} \bigg\rangle_{G'_N} \notag\\
&\qquad\geq \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha
\sum_{\epsilon \in \Sigma_M(\delta)} \exp \sum_{i \leq M} Z_{i,n}^\alpha (\epsilon_i)
- \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha \exp \sqrt{M} Y_n^\alpha- \tilde \epsilon. \label{eq:temp4}
\end{align}
\end{lem}
\begin{proof}
The covariance structure of \eqref{eq:cov-RPC-z} and \eqref{eq:cov-RPC-y} are identical to \eqref{eq:cov-gibbs-z} and \eqref{eq:cov-gibbs-y}. By \prettyref{thm:approximation} the functionals appearing in \eqref{eq:temp4} can be approximated by the same bounded continuous functions $\E g_\epsilon^Z( R)$ and $\E g_\epsilon^{Y}( R)$ of the overlap arrays. Since the distribution of the arrays $ R^N$ and $ Q^n$ both converge weakly to $ R^\infty$, by first taking a subsequence $N$ along which $ R^N$ converges to $ R^\infty$ and then approximating with $ Q^n$ we have for any $\tilde \epsilon > 0$,
\[
|\E g_\epsilon^Z( R^\infty) - \E g_\epsilon^{Y}( R^\infty) - \E g_\epsilon^Z( Q^n) + \E g_\epsilon^{Y}( Q^n)| \leq \tilde \epsilon.
\]
by choosing $n$ sufficiently large. Applying the triangle inequality will complete the proof.
\end{proof}
Since $\Sigma_M(\delta_M) = \prod_{s \in \mathscr{S}} \Sigma_{M_s} (\delta_{M_s})$,
we express the first term in \eqref{eq:temp4} as a weighted average. By the properties of the Ruelle probability cascades (see \cite[Theorem 2.9]{P3}),
\begin{align}
\frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha
\sum_{\epsilon \in \Sigma_M(\delta)} \exp \sum_{i \leq M} Z_{i,n}^\alpha (\epsilon_i) &= \frac{1}{M} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha
\prod_{s \in \mathscr{S}} \sum_{\epsilon \in \Sigma_{M_s}(\delta_{M_s})} \exp \sum_{i \in M_s} Z_{i,n}^\alpha (\epsilon_i) \nonumber\\
&= \sum_{s \in \mathscr{S}} \rho^s_M \frac{1}{M_s} \E \log \sum_{\alpha \in \mathbb{N}^r} v_\alpha
\sum_{\epsilon \in \Sigma_{M_s}(\delta_{M_s})} \exp \sum_{i \in M_s} Z_{i,n}^\alpha (\epsilon_i). \nonumber
\end{align}
Every sequence $(x^n_p)_{p=0}^r$ and $(Q^{n,s}_p)_{p=0}^r$ defines a discrete path $\bigl(\pi_n^s(x) \bigr)_{s \in \mathscr{S}} \in \Pi$. Using the notation of the functionals $f^Z_{M_s}(\Sigma_{M_s}(\delta_{M_s}), s; \bm \pi_n)$ and $f^Y_M(\bm \pi_n)$ defined on \eqref{eq:func-pi-z} and \eqref{eq:func-pi-y}, we have shown that
\[
\liminf_{N \to \infty}\frac{1}{N} \E \log Z_N(d_N) \geq \Big( \sum_{s \in \mathscr{S}} \rho_M^s f^Z_{M_s}(\Sigma_{M_s}(\delta_{M_s}), s; \bm \pi_n) - f^Y_M (\bm \pi_n) \Big) - \epsilon.
\]
\prettyref{lem:ctyofz} and \prettyref{lem:ctyofy} imply the functionals $f^Z_{M_s}(\Sigma_{M_s}(\delta_{M_s}), s;\bm \pi_n)$ and $f^Y_M (\bm \pi_n)$ are Lipschitz. Sending $\epsilon \to 0$ and noticing that the paths $\bm \pi_n \to \bm \pi_\infty$ in $\Delta$, we have shown
\begin{equation}\label{eq:temp2}
\liminf_{N \to \infty}\frac{1}{N} \E \log Z_N(d_N) \geq \Big( \sum_{s \in \mathscr{S}} \rho_M^s f^Z_{M_s}(\Sigma_{M_s}(\delta_{M_s}),s;\bm \pi_\infty) - f^Y_M (\bm \pi_\infty) \Big) , \text{ for any $M > 0$}.
\end{equation}
All that remains is to remove the dependence on $\Sigma_M(\delta_M)$. This will be a direct application of the decoupling proved in \prettyref{thm:discretize}.
\begin{lem} There exists a path $\bm \pi^* \in \Pi$ such that
\[
\lim_{M \to \infty} \Big( \sum_{s \in \mathscr{S}} \rho_M^s f^Z_{M_s}(s,\bm \pi^*) - f^Y_M (\bm \pi^*) \Big)
\geq \mathscr{P}(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}}) - \epsilon.
\]
\end{lem}
\begin{proof}
We now make the paths $\bm \pi^M_\infty$ dependence on $M$ explicit. Since $(\bm \pi^M_\infty)_{M \geq 1}$ is a countable collection of bounded monotone paths of $\kappa \times \kappa$ matrices, there exists a subsequence in $M$ such that $\bm \pi^M_\infty \to \bm \pi^*$. Given $\epsilon > 0$, by \prettyref{thm:discretize} we can find a discrete path $\bm \pi_\epsilon$ such that $\Delta(\bm \pi_\epsilon, \bm \pi^*) < C\epsilon$, where $C$ is the maximum Lipshitz constant over all $f^Z_{M_s}(s)$. Since $\bm \pi^\epsilon$ is discrete, applying \prettyref{thm:decouple} to the summation appearing in \eqref{eq:temp2} shows
\[
\lim_{M \to \infty} \sum_{s \in \mathscr{S}} \rho_M^s f^Z_{M_s}(s, \bm \pi^*) \geq \lim_{M \to \infty} \sum_{s \in \mathscr{S}} \rho_M^s f^Z_M(\bm \pi_{\epsilon}) - C\epsilon \geq \inf_{\lambda} \sum_{s \in \mathscr{S}} \rho^s \bigg( -\sum_{k \leq \kappa}\lambda^s_k d_k^s + X_0^s(\bm \pi^\epsilon) \bigg) - C\epsilon
\]
since $\rho^s_M \to \rho^s$, $\delta^s_{M,k} \to d_k^s$ and $f_M^z$ is Lipschitz.
The term $f^Y_M(\bm \pi_\infty)$ is actually independent of $M$, and a similar computation using the properties of Ruelle Probability Cascades like in \eqref{eq:upbdyterm} shows
\[
\lim_{M \to \infty} f^Y_M(\bm \pi^*) = \frac{1}{2} \sum_{\ell = 0}^{r-1} x_{\ell} \sum_{s,t \in \mathscr{S}} \Delta^2_{st} \rho^s \rho^t \left( (Q^s_{\ell+1},Q^t_{\ell+1}) - (Q^s_{\ell},Q^t_{\ell}) \right).
\]
Combining the computations of the two terms above, we now take $\epsilon \to 0$ and get
\begin{equation}\label{eq:lowerbd-end}
\liminf_{N \to \infty} \frac{1}{N}\E \log Z_N(d_N) \geq \mathscr{P}\bigl(d, \lambda, \bm \pi^*\bigr)
\end{equation}
since $\pi_\epsilon \to \pi^*$ as $\epsilon \to 0$.
\end{proof}
Since the path $\bm \pi^*$ can be described as the limit of the discrete approximating sequences \eqref{eq:RPC6} and \eqref{eq:RPC5}, we have shown
\[
\liminf_{N \to \infty} \frac{1}{N}\E \log Z_N(d_N) \geq \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}}\bigr).
\]
\section{Proofs of Main Theorem}
\begin{proof}[Proof of \prettyref{thm:ParisiUnConstrained}]
We begin by proving part (1) of the Theorem.
We start with the lower bound of $F_N \bigl(\Sigma_N^{\epsilon_N}(d) \bigr)$. For any $d \in \mathscr{D}$ let us choose $\epsilon_N = \frac{L_\kappa}{N}$. We choose a sequence of $d_N \in \mathscr{D}_N$ such that $\|d_N - d\|_\infty = \epsilon_N$ to satisfy the condition in \eqref{eq:convergence-proportions}. Since $\Sigma^\epsilon_N(d) \supseteq \Sigma_N(d_N)$, the lower bound \eqref{eq:lowerbound} computed in \prettyref{sec:Aizenman-LB} implies
\begin{equation}\label{eq:mainthm-cons-lb}
\liminf_{N \to \infty} \frac{1}{N} \E \log Z_N^{\epsilon_N}(d) \geq \liminf_{N \to \infty} \frac{1}{N} \E \log Z_N(d_N) \geq \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
We now obtain the matching upper bound of $F_N \bigl(\Sigma_N^{\epsilon_N}(d) \bigr)$. Since $\epsilon_N \to 0$, the $O(\epsilon)$ term in the upper bound \eqref{eq:guerra-ub-thick} computed in \prettyref{sec:Guerra-UB} vanishes yielding
\begin{equation}\label{eq:mainthm-cons-ub}
\limsup_{N \to \infty} \frac{1}{N} \E \log Z_N^{\epsilon_N}(d) \leq \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
Combining the inequalities in \eqref{eq:mainthm-cons-lb} and \eqref{eq:mainthm-cons-ub}, we arrive at the formula for the constrained free energy
\begin{equation}\label{eq:mainthm-cons-functional}
\lim_{N \to \infty} \frac{1}{N} \E \log Z_N(d) = \inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
Part (2) of the Theorem is a direct consequence of part (1). By classical Gaussian concentration inequalities, observe that the limit of the free energy $F_N(\Sigma_N)$ is asymptotically given by the supremum of $F_N\bigl(\Sigma_N(d)\bigr)$ over $d \in \mathscr{D}$. By taking the supremum over $d \in \mathscr{D}$ in \eqref{eq:mainthm-cons-functional}, we arrive at the formula for the free energy
\begin{equation}
\lim_{N \to \infty} \frac{1}{N} \E \log Z_N = \sup_{d \in \mathscr{D}} ~\inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}\bigl(r, x, d, (\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr).
\end{equation}
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:ground_state}]
We include a proof of the Corollary for the sake of completeness. By H\"{o}lder's inequality,
$\frac{F_N^{\beta}(\Sigma_N^{\varepsilon_N}(d))}{\beta}$ is increasing in $\beta$ and therefore
\begin{align}
\lim_{\beta \to \infty} \frac{1}{\beta}\inf_{x, r,(\lambda^s, Q^s)_{s \in \mathscr{S}}} \mathscr{P}_\beta\bigl(r, x, d,(\lambda^s, Q^s)_{s \in \mathscr{S}} \bigr)
\end{align}
exists. Moreover, we have,
\begin{align}
\frac{1}{N} \E \Big[ \max_{\Sigma_{N}^{\varepsilon_N}(d)} H_N(\sigma) \Big] \leq \frac{F_N^{\beta}(\Sigma_N^{\varepsilon_N}(d))}{\beta} \leq \frac{\log q}{\beta} + \frac{1}{N} \E \Big[ \max_{\Sigma_{N}^{\varepsilon_N}(d)} H_N(\sigma) \Big]. \nonumber
\end{align}
Taking the limit as $N\to \infty$ and then as $\beta \to \infty$ completes the proof.
\end{proof}
\section{Proofs regarding Cuts Problems}
\label{sec:cut_proofs}
In this section, we prove Theorem \ref{thm:discretization}. We start with a proof of Lemma \ref{lemma:discretization}.
\begin{proof}[Proof of Lemma \ref{lemma:discretization}]
The proof of this lemma proceeds by a direct coupling argument. We will denote the adjacency matrix corresponding to $\tilde{G}_N$ as $\tilde{A} = ( \tilde{A}_{i,j})$. We will realize the two graphs on the same probability space and couple the adjacency matrices such that
\begin{align}
\P[ A_{i,j} \neq \tilde{A}_{i,j}] &= | \P[ A_{i,j} =1] - \P[\tilde{A}_{i,j} =1 ] | \nonumber \\
&\leq c N \Big| \int_{[\frac{i-1}{N}, \frac{i}{N}] \times [\frac{j-1}{N},\frac{j}{N}]} K(x,y) \, {\rm d} x {\rm d} y - \int_{[\frac{i-1}{N}, \frac{i}{N}] \times [\frac{j-1}{N},\frac{j}{N}]} K_1(x,y) \, {\rm d} x {\rm d} y \Big|. \nonumber
\end{align}
We note that for any two graphs $G_N$ and $\tilde{G}_N$, we have,
\begin{align}
\Big| \frac{{\sf MaxCut }_\kappa(G_N)}{N}- \frac{{\sf MaxCut }_\kappa(\tilde{G}_N)}{N} \Big| \leq \frac{1}{2N} \sum_{i,j = 1}^N |A_{i,j} - \tilde{A}_{i,j}|. \nonumber
\end{align}
This implies that
\begin{align}
&\Big| \E \Big[ \frac{{\sf MaxCut }_\kappa(G_N)}{N} \Big]- \E\Big[\frac{{\sf MaxCut }_\kappa(\tilde{G}_N)}{N} \Big] \Big| \nonumber \\
&\leq \frac{c}{2} \sum_{i,j = 1}^N \Big| \int_{[\frac{i-1}{N}, \frac{i}{N}] \times [\frac{j-1}{N},\frac{j}{N}]} K(x,y) \, {\rm d} x {\rm d} y - \int_{[\frac{i-1}{N}, \frac{i}{N}] \times [\frac{j-1}{N},\frac{j}{N}]} K_1(x,y) \, {\rm d} x {\rm d} y \Big| \nonumber \\
&= \frac{c}{2} \bigl( \| K - K_1 \|_1 + o(1) \bigr). \nonumber
\end{align}
Thus the proof is complete once we choose $M$ sufficiently large such that $\|K - K_1 \|_1 \leq 1/ c^{1/2 + \delta}$.
\end{proof}
To complete the proof of Theorem \ref{thm:discretization}, we first introduce the following Gaussian optimization problem. Set $J = (J_{i,j})$ a symmetric matrix such that $\{J_{i,j}: i \leq j \}$ are independent $N(0, \frac{\tilde{K}_N(i,j)}{N})$ random variables. We define
\begin{align}
\tilde{Z}_N = \frac{1}{2N} \max_{\sigma \in [\kappa]^N} \Big[ \frac{c}{N} \sum_{i,j = 1}^N \tilde{K}_N(i,j) \1(\sigma_i \neq \sigma_j) + \sqrt{c}\sum_{i,j = 1}^N \frac{J_{i,j}}{\sqrt{N}} \1(\sigma_i \neq \sigma_j) \Big].
\end{align}
The following lemma establishes that in the ``large degree" limit, we can study the asymptotic behavior of the ${\sf MaxCut }_\kappa$ problem via that of the Gaussian optimization problem $\tilde{Z}_N$.
\begin{lem}
\label{lemma:comp}
As $N \to \infty$, we have,
\begin{align}
\E\Big[ \frac{{\sf MaxCut }_\kappa(G_N)}{N} \Big] = \E[\tilde{Z}_N] + o(\sqrt{c}). \nonumber
\end{align}
\end{lem}
\begin{proof}
The lemma follows from a direct application of \cite[Theorem 1.1]{sen2016optimization}. Specifically, we note that
\begin{align}
\frac{{\sf MaxCut }_\kappa(G_N)}{N} = \frac{1}{2N} \max_{\sigma \in [\kappa]^N} \sum_{i,j = 1}^N A_{i,j} \1(\sigma_i \neq \sigma_j). \nonumber
\end{align}
The proof follows directly upon an application of the result in \cite{sen2016optimization}.
\end{proof}
Thus finally, it comes down to the study of $\E[Y_n]$ as $N\to \infty$. Fix a probability distribution on $[\kappa]$, $d^{s} = (d^{s}_1, \ldots, d^{s}_\kappa)$, $s= 1, \ldots, M$. Consider the Hamiltonian
\begin{align}
H(\sigma) = \sum_{i,j =1}^N \frac{J_{i,j}}{\sqrt{N}} \1(\sigma_i = \sigma_j),
\end{align}
where $J = (J_{i,j})$ is as described above.
For $d = \bigl(d^{1}, \ldots, d^{M} \bigr)$, probability measures on $[\kappa]$, and a sequence $\varepsilon_N$ decaying to zero sufficiently slowly, recall the restricted configuration space \eqref{eq:constrainedspace}
\begin{align}
\Sigma_N^{\varepsilon_N}(d) = \Big\{ \sigma \in [\kappa]^N \mathrel{}\Big|\mathrel{} \sum_{i \in I_s} \frac{\1(\sigma_i = k )}{N_s} \in [ d^{s}_k -\varepsilon_N, d^{s}_k + \varepsilon_N], s \in [M] \Big\}.
\end{align}
Recall the restricted ground state energy $\mathcal{P}(d)$, introduced in Corollary \ref{cor:ground_state}.
This will allow us to deduce the following lemma.
\begin{lem}
\label{lemma:exp}
\begin{align}
\lim_{N \to \infty} \E[\tilde{Z}_N] = \sup_{d} \Big[ \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s, t) \rho^{s} \rho^{t} (1- \langle d^{s}, d^{t} \rangle) + \frac{\sqrt{c}}{2} \mathcal{P}(d) \Big] + o(\sqrt{c}). \nonumber
\end{align}
\end{lem}
\begin{proof}
We start with the lower bound.
We define
\begin{align}
\tilde{Z}_N(d) &= \frac{1}{2N} \max_{\sigma \in \Sigma_N^{\varepsilon_N}(d)} \Big[ \frac{c}{N} \sum_{i,j = 1}^N \tilde{K}_N(i,j) \1(\sigma_i \neq \sigma_j) + \sqrt{c}\sum_{i,j = 1}^N \frac{J_{i,j}}{\sqrt{N}} \1(\sigma_i \neq \sigma_j)\Big] \nonumber\\
&= \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s,t) \rho^{s} \rho^{t} \bigl(1 - \langle d^{s}, d^{t} \rangle \bigr) + \frac{\sqrt{c}}{2N}\max_{\sigma \in \Sigma_N^{\varepsilon_N}(d)} \sum_{i,j = 1}^N \frac{J_{i,j}}{\sqrt{N}} \1(\sigma_i \neq \sigma_j) + o(1). \nonumber
\end{align}
For fixed probability vectors $d^{1}, \ldots, d^{M}$, we have,
\begin{align}
\liminf_{N \to \infty} \E[\tilde{Z}_N] &\geq \liminf_{N \to \infty} \E[\tilde{Z}_N(d)] \nonumber \\
&= \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s,t) \rho^s \rho^t \bigl(1 - \langle d^s, d^t \rangle \bigr) + \frac{\sqrt{c}}{2} \mathcal{P}(d). \nonumber
\end{align}
We take the supremum over all possible probability vectors $d^{1}, \ldots , d^{M}$ to get the requisite lower bound. To establish the upper bound, we define
\begin{align}
\tilde{M} = \sup_{d} \Big[ \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s, t) \rho^s \rho^t \bigl(1- \langle d^{s}, d^{t} \rangle \bigr)+ \frac{\sqrt{c}}{2} \mathcal{P}(d) \Big].
\end{align}
Therefore, we have,
\begin{align}
\E[\tilde{Z}_N(d)] &\leq \frac{c}{2} \sum_{s,t = 1}^M \mathbf{K}(s,t) \rho^s \rho^t \bigl(1 - \langle d^{s}, d^{t} \rangle \bigr) + \frac{\sqrt{c}}{2} \mathcal{P}(d) + o(1). \nonumber \\
&\leq \tilde{M} + o(1), \nonumber
\end{align}
uniformly over all choices of $d^{1}, \ldots, d^{M}$. We note that empirical distributions within each block may assume only finitely many values and hence, summing over these values, we have,
\begin{align}
\P[ \tilde{Z}_N > \tilde{M} + t] = \sum_{d^1, \ldots, d^M} \P \bigl[ \tilde{Z}_N(d) > \tilde{M} +t \bigr] . \nonumber
\end{align}
Now,
\begin{align*}
\P[\tilde{Z}_N(d)> \tilde{M} +t ] &\leq \P[ \tilde{Z}_N (d) - \E[\tilde{Z}_N(d)] > t]
\leq \exp[- C N t^2],
\end{align*}
where the last inequality follows by Gaussian concentration. Finally, plugging this tail bound, we have,
\begin{align}
\P[\tilde{Z}_N > \tilde{M} + t ] \leq A N^{\kappa \tilde{M}} \exp[-CN t^2]. \nonumber
\end{align}
We note that
\begin{align}
\E[\tilde{Z}_N] &\leq \int_{0}^{\infty} \P [ \tilde{Z}_N > x] \, {\rm d} x \leq \tilde{M} + \delta_N + \int_{ \delta_N}^{\infty} \P[ \tilde{Z}_N > \tilde{M} + t] \, {\rm d} t \nonumber \\
&\leq \tilde{M} + \delta_N + A N^{\kappa \tilde{M}} \frac{\exp[- CN \delta_N^2]}{\sqrt{N}\delta_N} . \nonumber
\end{align}
Finally, we choose $\delta_N = C_0 \sqrt{\frac{\log N}{N}}$ for some constant $C_0$ sufficiently large. This establishes that
\begin{align}
\E[\tilde{Z}_N] \leq \tilde{M} + o(1), \nonumber
\end{align}
completing the proof of the upper bound.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:discretization}]
The Theorem follows directly upon combining Lemma \ref{lemma:comp} and Lemma \ref{lemma:exp}.
\end{proof}
|
\section{Introduction}\label{Sect:intro}
Cosmology is currently in a golden era of surveys probing an ever larger fraction of the Hubble volume. In the next decade we will obtain data covering most of the extragalactic sky out to a redshift of two or more, thanks to surveys like the Dark Energy Survey \citep[DES,][]{Abbott:2005bi}, the Dark Energy Spectroscopic Instrument \citep[DESI,][]{Aghamousa:2016zmz}, Euclid \citep{Laureijs:2011gra}, the Large Synoptic Sky Telescope \citep[LSST,][]{Abell:2009aa}, and the Square Kilometer Array \citep[SKA,][]{Maartens:2015mra}, to name just a few. These surveys will try to answer some of the most challenging questions in cosmology and fundamental physics, about the nature of dark matter and dark energy and the origin of the large-scale structure of the Universe \citep[e.g.\ ][and the survey references given above]{Amendola:2012ys}. Thanks to the sheer number of objects that will be observed in these surveys, the statistical uncertainties will be strongly reduced. However, in order to obtain accurate statistical inferences, the systematic errors also need to be under control. A key quantity for the likelihood function is the survey covariance matrix of the observables.
Incorrect covariance matrices, for example including only Gaussian contributions, can significantly overestimate the information content of galaxy surveys observables \citep{Repp2015}.
As galaxy surveys become deeper and larger, traditional sources of statistical errors such as shot noise are beaten down. The effect of super-survey modes, called super-sample covariance (SSC), however, decreases more slowly with the survey volume. These modes change the average matter density within the survey volume to which the probes react through linear (for 1 point probes such as cluster counts) or non-linear (for 2+ point probes such as the galaxy power spectrum) perturbation theory. The SSC effect was first pointed out for cluster counts by \cite{Hu2003}, and then for the matter power spectrum by \cite{Hamilton2006}. Super-sample covariance is particularly important when combining probes as it couples the observables together \citep{TakadaBridle2007, Takada2014, Krause2016, Lacasa2016}. Careful estimation of covariances including this term is thus necessary for single probe and combined probes analyses with current and future surveys.
Survey covariances can be estimated from the data itself through jackknife or bootstrap methods. However \cite{Hoffmann2015} showed that the standard jackknife estimator failed to recover correctly the 3D cluster counts of the MICE simulations and proposed an improved estimator instead. For galaxy weak lensing, \cite{Friedrich2016} showed that bootstrap and jackknife recover only partially the correct volume of cosmological parameter constraints, and \cite{Shirasaki2016} further pointed out the limitations of the jackknife covariance in the regime dominated by super-sample covariance (SSC).
In their standard form these methods fail because they intrinsically assume that the subsamples are independent draws of a given random field. However, in reality the subsamples are not independent because they are correlated by large-scale fluctuations, which is exactly the SSC effect.
This article gives the formalism to exactly predict how biased the jackknife and bootstrap covariance estimators are, further driving home the point that they should not be used for cosmological parameter inference. It also examines the case of the subsampling method, which was designed to estimate the covariance of the subsample area instead of the full survey area. This case arises for example when cutting down a large simulation into survey-shaped pieces, and we show that this also produces a biased covariance. The capacity of predicting analytically this bias is of importance when validating a prediction code against a single numerical simulation or a limited set of them. Indeed, we do not now possess the resources to produce a large number of full observable universe sized simulations with all the necessary cluster, galaxy, and weak-lensing physics for current and future surveys, though smaller simulations with adequate rescaling have recently been proposed as an alternative \citep{Klypin2017}.
This article is organised as follows: we review the internal covariance estimators in Sect. \ref{Sect:intern-cov-estim}; we derive analytically the impact of the SSC on these estimators in Sect. \ref{Sect:effect-on-SSC}; we apply this formalism numerically to quantify the importance of the effect in Sect. \ref{Sect:num-results}; and finally we discuss our results and their perspectives in Sect. \ref{Sect:discussion}.
In all numerical computations, we take a flat $\Lambda$CDM cosmology with parameters
$h=0.67$, $\Omega_b h^2=0.022$, $\Omega_c h^2=0.12$, $n_S=0.96$ , $\sigma_8=0.83$.
The cluster counts are computed in two bins of redshift in the range $z\in[0.9,1.1]$ with a width $\Delta z=0.1$, and four bins of mass in the range $\log [M/(h^{-1}M_{\odot})] \in[14,16]$
with a width $\Delta\log [M/(h^{-1}M_{\odot})]=0.5$. The corresponding covariance is thus an $8\times8$ matrix composed of $4\times4$ auto- and cross-redshift blocks. The halo mass function is taken from \cite{Tinker2008}, with the corresponding halo bias from \cite{Tinker2010}, and the linear matter power spectrum is from the transfer function by \cite{Eisenstein1998}. We quote these details here for completeness, but the precise values do not affect the conclusions of this paper in a qualitative way.
\section{Internal covariance estimators}\label{Sect:intern-cov-estim}
Let us denote by an index $\alpha$ the bins of cluster mass and redshift: $\alpha\sim(i_M,i_z)$. The survey area is divided in $N_\mathrm{sub}$ equal-geometry subsamples with a latin index $i$. As will be shown in the subsections below, any natural covariance estimators built from these subsamples can be written in the form
\begin{equation}\label{Eq:Cov-wij}
\widehat{\mr{Cov}}(N_\alpha,N_\beta) = \sum_{i,j=1}^{N_\mathrm{sub}} w_{i,j} \ N_\alpha(i) \ N_\beta(j)
\end{equation}
with the requirement that
\begin{equation}
\sum_{i,j=1}^{N_\mathrm{sub}} w_{i,j} = 0
,\end{equation}
which implements the constraint that $\widehat{\mr{Cov}}=0$ if $N(i)=\mathrm{cst}$, i.e. the covariance estimator should yield zero if all samples are identical.
The main point is that all covariance estimators below are purely `survey internal' and are unable to see modes on scales larger than the survey volume. Because of this limitation they are unable to correctly take into account the super-sample covariance.
\subsection{Subsampling}\label{Sect:subsamp}
The subsampling method tries to estimate the covariance of the subsample area through the simple empirical covariance of the subsample counts:
\begin{equation}
\widehat{\mr{Cov}}_{\rm sub}(N_\alpha,N_\beta) = \frac{1}{N_\mathrm{sub}\,\text{-}\, 1} \sum_{i=1}^{N_\mathrm{sub}} \left(N_\alpha(i) \,\text{-}\, \overline{N}_\alpha\right) \left(N_\beta(i) \,\text{-}\, \overline{N}_\beta\right)
\end{equation}
with
\begin{equation}
\overline{N}_\alpha \equiv \frac{1}{N_\mathrm{sub}} \sum_{i=1}^{N_\mathrm{sub}} N_\alpha(i) \, .
\end{equation}
This leads to an estimator of the form of Eq.\ (\ref{Eq:Cov-wij}) with
\begin{equation}
w_{i,j}^\mathrm{sub} = \frac{\delta_{i,j} - \frac{1}{N_\mathrm{sub}}}{N_\mathrm{sub}-1} \, .
\end{equation}
\subsection{Jackknife}\label{Sect:jackknife}
The jackknife method, however, aims to recover the covariance of the full survey area. To do so, it defines a jackknife sample as the total survey minus one subsample. Denoting $N_\alpha^{\rm jack}(i) = N_\alpha^{tot} - N_\alpha(i)$ such jackknife sample, we have
\begin{eqnarray}
\lefteqn{\widehat{\mr{Cov}}_{\rm jack}(N_\alpha,N_\beta) =} && \nonumber \\
&&\frac{N_\mathrm{sub}}{N_\mathrm{sub}\,\text{-}\, 1} \sum_{i=1}^{N_\mathrm{sub}} \left(N^{jk}_\alpha(i) \,\text{-}\, \overline{N}^{\rm jack}_\alpha\right) \left(N^{jk}_\beta(i) \,\text{-}\, \overline{N}^{\rm jack}_\beta\right)
,\end{eqnarray}
where
\ba
\overline{N}^{\rm jack}_\alpha &\equiv \frac{1}{N_\mathrm{sub}} \sum_{i=1}^{N_\mathrm{sub}} N^{jk}_\alpha(i)\\
&= N_\alpha^{\rm tot} - \overline{N}_\alpha \, .
\ea
It easily follows that
\ba
\widehat{\mr{Cov}}_{\rm jack}(N_\alpha,N_\beta) &= \frac{N_\mathrm{sub}}{N_\mathrm{sub} \,\text{-}\, 1} \sum_{i=1}^{N_\mathrm{sub}} \left(N_\alpha(i) \,\text{-}\, \overline{N}_\alpha\right) \left(N_\beta(i) \,\text{-}\, \overline{N}_\beta\right)\\
&= N_\mathrm{sub} \times \widehat{\mr{Cov}}_{\rm sub}(N_\alpha,N_\beta)
\ea
and thus the jackknife estimator is just a rescaling of the subsampling estimator. It can thus be put in the form Eq.~(\ref{Eq:Cov-wij}) with
\begin{equation}
w_{i,j}^\mathrm{jack} = N_\mathrm{sub} \ \frac{\delta_{i,j} - \frac{1}{N_\mathrm{sub}}}{N_\mathrm{sub}-1} \, .
\end{equation}
\subsection{Bootstrap}\label{Sect:bootstrap}
Like the jackknife method described in Sect. \ref{Sect:jackknife}, the bootstrap method attempts to estimate the covariance of the full survey area. To do so, it first defines a bootstrap sample $b$ as the union of $N_\mathrm{sub}$ possibly repeated subsamples. For example with $N_\mathrm{sub}=4$, a possible sample is \{3,1,4,1\}. A high number of samples $N_\mathrm{bs}$ is drawn, typically 1,000 or 10,000, and the natural empirical covariance estimator is applied on these samples:
\ba
\lefteqn{\widehat{\mr{Cov}}_{\rm boot}(N_\alpha,N_\beta) =} && \nonumber \\
&&\frac{1}{N_\mathrm{boot} \,\text{-}\, 1} \sum_{b=1}^{N_\mathrm{boot}} \left(N^{bs}_\alpha(b) \,\text{-}\, \overline{N}^{\rm boot}_\alpha\right) \left(N^{\rm boot}_\beta(b) \,\text{-}\, \overline{N}^{bs}_\beta\right) \, .
\ea
In Appendix \ref{App:bootstrap}, it is shown that when $N_\mathrm{boot}\rightarrow\infty$, the bootstrap method basically asymptotes to the jackknife one, giving an estimator of the form Eq.~(\ref{Eq:Cov-wij}) with
\begin{equation}
w_{i,j}^\mathrm{boot} = \delta_{i,j} - \frac{1}{N_\mathrm{sub}}
,\end{equation}
which is the same as the jackknife weight, up to a factor $\frac{N_\mathrm{sub}}{N_\mathrm{sub}-1}\simeq 1$.\\
We note that this factor can be cancelled by applying an alternative normalisation to the jackknife estimator, and it has an impact smaller than the effects discussed in the following sections of this article. Thus, we will not be mentioning the bootstrap method again, just the subsampling and jackknife methods.\\
It should be emphasized that the estimators discussed here all work correctly in their respective domains of applicability. The main issue is that in the presence of SSC the different subsamples are not independent and identically distributed, which breaks an important assumption that underlies these estimators.
\section{Effect on super-sample covariance}\label{Sect:effect-on-SSC}
Following \cite{Lacasa2016b}, the super-sample covariance of the number counts in a given subsample $i$ can be predicted as
\ba\label{Eq:Cov-counts-wcovell}
\nonumber \left\langle N_\alpha(i) \, N_\beta(i) \right\rangle_c &\equiv \left\langle N_\alpha(i) \, N_\beta(i) \right\rangle - \left\langle N_\alpha(i)\right\rangle \left\langle N_\beta(i) \right\rangle\\
\nonumber &= \frac{1}{f^2_\mathrm{sky}(i)} \sum_\ell \frac{2\ell+1}{4\pi} \ C_\ell\left(M(i)\right) \\
& \qquad \times \mr{Cov}_\ell^\mathrm{SSC}\left(N_\alpha , N_\beta\right),
\ea
where $M(i)$ is the mask corresponding to the $i$th subsample and $f_\mathrm{sky}(i)$ is the corresponding sky fraction. The factor $\mr{Cov}_\ell^\mathrm{SSC}$, which is independent of geometry but does depend on the cosmology and general modelling, is plotted in Fig. \ref{Fig:Cov_ell^SSC}. The plot shows the $\ell$ dependence of this factor for an arbitrary and unimportant choice of mass bins, for the auto-redshift covariance and the cross-redshift correlation (i.e. covariance normalised by its diagonal).
\begin{figure}[!th]
\begin{center}
\includegraphics[width=.49\linewidth]{Cov_ell_SSC_autoz.pdf}
\includegraphics[width=.49\linewidth]{Corr_ell_SSC_crossz.pdf}
\caption{Multipole dependence of $\mr{Cov}_\ell^\mathrm{SSC}$. \textit{Left:} Auto-redshift covariance. \textit{Right:} Cross-redshift correlation.}
\label{Fig:Cov_ell^SSC}
\end{center}
\end{figure}
For the auto-redshift covariance, $\mr{Cov}_\ell^\mathrm{SSC}$ exhibits a behaviour similar to the matter power spectrum which sources it, first increasing with multipole up to a maximum corresponding to the matter-radiation equality angular scale at that redshift ($\ell_\mathrm{eq} = k_\mathrm{eq}/r(z)$) and then decreasing monotonically. The cross-redshift covariance shows a large anti-correlation on large scales (reaching -20\% for a full-sky survey for which $C_\ell(M)\propto \delta_{\ell,0}$), then changes sign and reaches a maximum of 7\% at $\ell=50$, and then decreases to zero reaching 2.5\% at $\ell=300$.
Formula~(\ref{Eq:Cov-counts-wcovell}) can straightforwardly be generalised to predict the cross-covariance between two different subsamples $i$ and $j$, involving the cross-spectrum of their respective masks $C_\ell\left(M(i),M(j)\right)$:
\ba
\nonumber \left\langle N_\alpha(i) \, N_\beta(j) \right\rangle_c &= \frac{1}{f_\mathrm{sky}(i) \, f_\mathrm{sky}(j)} \sum_\ell \frac{2\ell+1}{4\pi} \ C_\ell\left(M(i),M(j)\right) \\
& \qquad \times \mr{Cov}_\ell^\mathrm{SSC}\left(N_\alpha , N_\beta\right) \, .
\ea
In the following, we assume that all subsamples have the same area, denoted $f_\mathrm{sky}^\mathrm{sub}$.
It follows from the previous equations that the average value of a covariance estimator of the form of Eq.~(\ref{Eq:Cov-wij}) is
\ba\label{Eq:avgcov-cleff}
\nonumber \left\langle \widehat{\mr{Cov}}(N_\alpha,N_\beta) \right\rangle &= \sum_{i,j=1}^{N_\mathrm{sub}} w_{i,j} \left( \left\langle N_\alpha(i) \ N_\beta(j) \right\rangle_c + \overline{N}_\alpha \, \overline{N}_\beta\right)\\
&= \frac{1}{\left(f^\mathrm{sub}_\mathrm{sky}\right)^2} \sum_\ell \frac{2\ell+1}{4\pi} \ C_\ell^\mathrm{eff} \ \mr{Cov}_\ell^\mathrm{SSC}\left(N_\alpha , N_\beta\right),
\ea
where the effective power spectrum is
\begin{equation}\label{Eq:def-cleff}
C_\ell^\mathrm{eff} = \sum_{i,j=1}^{N_\mathrm{sub}} w_{i,j} \ C_\ell\left(M(i),M(j)\right) \, .
\end{equation}
The difference between the effective power spectrum and the real power spectrum of the survey area is the exact reason that makes internal covariance estimators biased for SSC. This can be regarded as the main point of this article.
\section{Numerical results}\label{Sect:num-results}
In this section we evaluate Eq. (\ref{Eq:avgcov-cleff}) numerically with the effective power spectrum given by Eq. (\ref{Eq:def-cleff}) and compare it to the true covariance using the true power spectrum in order to quantify the possible bias of the covariance estimator.
\subsection{Setup}
We defined two subsampling setups for illustration in this article. In both cases, we chose the subsamples as Healpix low resolution pixels as it allows for fast Healpix operations and gives the subsamples the exact same geometry and area.
In the first case, hereafter called the large-scale setup, we selected the maximum number of Healpix pixels at $N_\mathrm{side}=4$ that fit strictly in the octant $0<\theta<\pi/2$ $0<\phi<\pi/2$, and further removed the northernmost pixel\footnote{This pixel is unfit for $C_\ell$ computations because the Healpix pixelisation has poor sampling of $\phi$ near the poles. Other pixels are removed because they do not reside entirely within the octant, giving the total area the X shape visible in Fig.~\ref{Fig:subsamples-illustration}.}. This setup is visible in the top panel of Fig. \ref{Fig:subsamples-illustration} and gives subsamples of area $\sim 215$ deg$^2$. This situation arises for covariance estimations from a numerical simulation. Indeed, with an N-body simulation, light cone observables are usually computed on a sky octant, which corresponds to putting the observer at one corner of the simulation box \cite{Sehgal2010,Fosalba2015}.\\
In the second case, herafter called the small-scale setup, we divided a Healpix pixel at $N_\mathrm{side}=32$ in 16 subsamples corresponding to Healpix subpixels at $N_\mathrm{side}=128$. This setup is visible in the bottom panel of Fig. \ref{Fig:subsamples-illustration} and gives subsamples of area $\sim 0.2$ deg$^2$. This situation is a toy model of a small-scale rectangular survey.
\begin{figure}[!th]
\begin{center}
\includegraphics[width=.9\linewidth]{subsamples-illustration-ns4-256.png}
\includegraphics[width=.6\linewidth]{subsamples-illustration-ns32-128-512.png}
\caption{Illustration of the subsamples used in the analysis. The colour code is arbitrary, and only serves to differentiate between the subsamples. \textit{Top:} Large-scale case. \textit{Bottom:} Small-scale case.}
\label{Fig:subsamples-illustration}
\end{center}
\end{figure}
\subsection{Subsampling}
Figure \ref{Fig:cl-subvseff} shows the effective power spectrum of the subsampling method, compared to the power spectrum of the subsample mask that is needed for the correct SSC prediction.
\begin{figure}[!th]
\begin{center}
\includegraphics[width=0.49\linewidth]{Cl-subvseff-ns4-256.pdf}
\includegraphics[width=0.49\linewidth]{Cl-subvseff-ns32-128-512.pdf}
\caption{True power spectrum of the subsample compared to the effective power spectrum that enters the subsampling covariance estimate. \textit{Left:} Large-scale case. \textit{Right:} Small-scale case.}
\label{Fig:cl-subvseff}
\end{center}
\end{figure}
We see a lack of power in the effective power spectrum on scales comparable to or larger than the survey size. This corresponds to $\ell \lesssim 4$ in the large-scale case, and extends to $\ell \lesssim 150$ in the small-scale case.
We computed the corresponding cluster count covariances, and Fig.\ \ref{Fig:ratio-cov-eff/true-subsampling} shows the ratio of the subsampling covariance to the true one for both cases.
\begin{figure}[!th]
\begin{center}
\includegraphics[width=0.49\linewidth]{ratio-cov-truesub-vs-effsub-autoz.pdf}
\includegraphics[width=0.49\linewidth]{ratio-cov-truesub-vs-effsub-crossz.pdf}
\caption{Ratio of the effective covariance coming from subsampling to the true covariance of the subsample. \textit{Left:} Auto-redshift case. \textit{Right:} Cross-redshift case.}
\label{Fig:ratio-cov-eff/true-subsampling}
\end{center}
\end{figure}
In the large-scale setup, the subsampling method correctly estimates the auto-redshift covariance. However the cross-redshift covariance is dominated by small multipoles, visible in the right panel of Fig. \ref{Fig:Cov_ell^SSC}. The lack of power at small multipoles leads to an underestimate of the true cross-redshift covariance by 10\% to 40\% using the subsampling method. In this setup, the cross-$z$ covariance is negative and ranges between -1.5\% and -7.5\% when the covariance matrix is normalised to its diagonal.
In the small-scale setup, the lack of power on a large portion of the multipole range makes the subsampling method underestimate the covariance by $\sim15\%$ for auto-redshift and $\sim23\%$ for cross-redshift. In this setup, the cross-$z$ covariance is positive and ranges between 1.5\% and 5.5\% when the covariance matrix is normalised to its diagonal.
We note that in most cases, the difference between the estimated and true covariances is comparable to or larger than the 10\% systematic uncertainty due to the 5\% precision on halo bias predictions \citep{Hoffmann2017}.
\subsection{Jackknife}
Figure \ref{Fig:cl-totvseff} shows the effective power spectrum of the jackknife method, this time compared to the full survey area power spectrum.
\begin{figure}[!th]
\begin{center}
\includegraphics[width=.49\linewidth]{Cl-totvseff-ns4-256.pdf}
\includegraphics[width=.49\linewidth]{Cl-totvseff-ns32-128-512.pdf}
\caption{True power spectrum of the survey area compared to the effective power spectrum that enters the jackknife covariance estimate. \textit{Left:} Large-scale case. \textit{Right:} Small-scale case.}
\label{Fig:cl-totvseff}
\end{center}
\end{figure}
In the large-scale setup, the jackknife power spectrum is larger than the true power spectrum of the survey on most of the multipole range, except for a lack of power in the few first multipoles. In the small-scale setup, however, the lack of power dominates most of the multipole range; the jackknife power spectrum becomes larger than the true spectrum only on small scales.
The corresponding cluster count covariances were again computed, and Fig.\ \ref{Fig:ratio-cov-eff/true-jackknife} shows the ratio of the jackknife covariance to the true one for both setups.
\begin{figure}[!th]
\begin{center}
\includegraphics[width=0.49\linewidth]{ratio-cov-tot-vs-jk-autoz.pdf}
\includegraphics[width=0.49\linewidth]{ratio-cov-tot-vs-jk-crossz.pdf}
\caption{Ratio of the effective covariance coming from jackknife to the true covariance of the total survey area. \textit{Left:} Auto-redshift case. \textit{Right:} Cross-redshift case.}
\label{Fig:ratio-cov-eff/true-jackknife}
\end{center}
\end{figure}
In the large-scale setup, the jackknife method overestimates the auto-redshift covariance by $\sim6\%$, due to the power spectrum excess. The cross-redshift covariance is dominated by small multipoles (see right panel of Fig. \ref{Fig:Cov_ell^SSC}), and we find that the jackknife method recovers less than half of the true covariance. In this setup, the cross-$z$ covariance is negative and ranges between -9\% and -15\% when the covariance matrix is normalised to its diagonal. We note that cross-$z$ correlation can have a significant impact for constraints on cosmological parameters sensitive to the redshift evolution of the probe(s), such as the dark energy equation of state.
In the small-scale setup, the lack of power on most of the multipole range makes the jackknife underestimate the covariance by a factor 3-4 for auto-redshift and 5-6 for cross-redshift. In this setup, the cross-$z$ covariance is positive and ranges between 2\% and 8\% when the covariance matrix is normalised to its diagonal.
We note that in most cases, the difference between the estimated and true covariances is again comparable to or larger than the 10\% systematic uncertainty due to the 5\% precision on halo bias predictions \citep{Hoffmann2017}.
\section{Discussion}\label{Sect:discussion}
\subsection{Results}
As has been shown previously in the literature \citep{Hoffmann2015,Shirasaki2016,Friedrich2016}, the standard jackknife method gives a biased estimate of the covariance of the survey area. This arises because the jackknife is in fact a constant rescaling of the natural estimator of the subsample area. This rescaling is calibrated to give an unbiased estimate of the Poissonian shot-noise; however, it cannot reproduce super-sample covariance because the latter has a non-trivial dependence on the mask geometry which depends on cosmological parameters \cite{Lacasa2016b}.
We have given here an analytical understanding of this phenomenon and developed the formalism to predict the covariance bias quickly and without any simulation. We emphasize that the numerical calculations were performed on a simple dual-core laptop in at most tens of seconds. This article thus offers a fast way for survey analysts to check how biased their favourite internal covariance estimator is, and to estimate the true covariance.
Furthermore, we showed that the subsampling method also gives a biased estimate of the covariance, this time for the subsample area. The reason for this is that the $C_\ell^{\rm eff}$ estimated from a survey or simulation of finite size always lack power on the largest scales, and this lack of power is then transferred to all covariance elements through the convolution of Eq.~(\ref{Eq:avgcov-cleff}).
At least on small scales, this sounds the death knell to estimating the covariance from a single simulation larger than the survey by carefully cutting different survey areas into it. On large scales, the subsampling method actually performs quite well, especially for the auto-redshift covariance, although it performs a bit less well for the cross-redshift covariance. The subsampling method works relatively well for the auto-redshift covariance
because the lack of power affects only the smallest multipoles and because $\mr{Cov}_\ell^\mathrm{SSC}$ is flatter or even turning around at low $\ell$ (see left panel of Fig. \ref{Fig:Cov_ell^SSC}), giving less weight to large scales. On the contrary for cross-redshift, $\mr{Cov}_\ell^\mathrm{SSC}$ is still steep at low multipoles, explaining why the subsampling lack of power affects it more. We tested setups other than those presented in Fig. \ref{Fig:subsamples-illustration}, and found that the subsampling method actually gives a very good covariance estimate in cases with a larger number of subsamples or when subsamples have a larger area.
In some cases, covariance estimation from a single simulation is unavoidable. One such case arises when analytical covariance predictions need to be compared to a carefully crafted and CPU-expensive simulation in order to validate the prediction code. Then, the formalism developed here allows the analytical predictions to be tweaked in order to match how biased the covariance from the simulation is. In that case, we recommend comparing the covariance of the subsample area from the prediction and the estimate from the subsampling method. Indeed, the latter performs much better than the jackknife, and the bias may even be negligible in large-scale cases or when there is a large enough number of subsamples.
Although this article focused on the case of cluster number counts, the formalism developed and the qualitative conclusions are also valid for other probes, including the two-point correlation function or power spectrum for galaxy clustering and weak-lensing shear.
\subsection{Debiasing (im)possibility}
\cite{Hoffmann2015} developed a method for debiasing the standard jackknife estimator by first removing the shot-noise contribution, then debiasing the SSC contribution using the ratio of the variances of the matter density field in the survey volume and subsample volume, $r_\sigma \equiv \sigma^2_m(V_\mathrm{tot})/\sigma^2_m(V_\mathrm{JKcell})$, and finally re-adding the shot-noise. This convoluted method is possible because they worked with 3D volumes at a single fixed redshift, but it becomes theoretically impossible in the case considered in this article of a lightcone observable, due to the redshift evolution within the bin and because we consider different bins. An effective rescaling, e.g. using $r_\sigma$ at the centre of the redshift bin, is possible and may perform in a satisfactory manner for the auto-redshift covariance; however, we see that the ratios plotted in Figs. \ref{Fig:ratio-cov-eff/true-subsampling} and \ref{Fig:ratio-cov-eff/true-jackknife} are not constant with mass and redshift. Hence, it is not possible to debias simultaneously all covariance elements together, in particular not the auto-$z$ and cross-$z$ elements at the same time.\\
We emphasize that the covariances were predicted from analytical equations, not estimated from simulations, so that the ratios plotted are exact (up to machine precision), and their variations are thus genuine.
Another debiasing method would be additive instead of multiplicative. It can be computed as
\ba\label{Eq:bias-cov}
\nonumber \Delta \, \mr{Cov}(N_\alpha,N_\beta) &\equiv \frac{1}{\left(f^\mathrm{sub}_\mathrm{sky}\right)^2} \sum_\ell \frac{2\ell+1}{4\pi} \ \left(C_\ell(M) - C_\ell^\mathrm{eff}\right) \\
& \qquad \times\mr{Cov}_\ell^\mathrm{SSC}\left(N_\alpha , N_\beta\right)
\ea
and used to debias the estimated covariance:
\begin{equation}
\widetilde{\mr{Cov}}(N_\alpha,N_\beta) \equiv \widehat{\mr{Cov}}(N_\alpha,N_\beta) + \Delta \, \mr{Cov}(N_\alpha,N_\beta) \, .
\end{equation}
However, the advantages of this debiasing are few: computing Eq. (\ref{Eq:bias-cov}) has exactly the same analytical requirements as computing the full prediction Eq. (\ref{Eq:Cov-counts-wcovell}), including a theoretical prediction of the halo mass function and bias (while the ratio $r_\sigma$ from \citealt{Hoffmann2015} only depends on the linear matter power spectrum and the volumes considered). As such, it is far preferable to predict the covariance entirely analytically, avoiding simulation noise and thus the need for a Hartlap's correction \citep{Hartlap2007} or a Bayesian marginalisation \citep{Sellentin2016}.
\subsection{Catalogue-based covariance estimators}
Another class of often-used covariance estimators are jackknife or bootstrap estimators based on object catalogues instead of subsamples. These estimators are used for example for the covariance of galaxy clustering or weak-lensing correlation functions or power spectra. At first sight, a catalogue-based estimator may seem like the asymptote of the corresponding subsample-based estimator in the limit of infinitely small subsamples, but the presence of empty subsamples means that this is not the case. It also does not correspond to the case of subsamples carefully crafted to contain a unique object; this construction is doomed by non-uniqueness and enters the realm of a posteriori statistics.
It can be seen that catalogue-based estimators have intrinsic limitations. We argue that these limitations limit their use for estimating covariance of large-scale structure observables. First, they are a tautological nonsense for number counts, e.g. a jackknife would always have $N_\mathrm{obj}-1$ objects and thus the estimated covariance is always zero. Though this may sound trivial, it means that it does not permit their use for galaxy and/or cluster number counts, and thus for combining probes. Combined probes is a main objective for current and future galaxy surveys, and thus a careful prediction of the joint probe covariance, including the cross-covariance between probes, is of prime importance. Second, catalogue-based estimators give weight to dense regions, but regions devoid of galaxies will not change between jackknife samples (nor between bootstrap samples); this is problematic for cosmological applications as void properties do change with cosmological parameters. Finally, in a field where all galaxies would be grouped in a single pixel, a catalogue-based jackknife/bootstrap estimator would find the same correlation function or power spectrum in all jackknife/bootstrap samples and thus a zero covariance, which would not be the correct answer at all.
Though it is not the purpose of this article to examine thoroughly the case of catalogue-based estimators, the arguments presented above allow us to cast doubt on the validity of these estimators already for a Gaussian matter density field, and even more so for super-sample covariance.
\section{Conclusion}\label{Sect:conclusion}
We have developed the formalism necessary to predict the effect of SSC on standard covariance estimators based on spatial subsamples of the data. We demonstrated that internal covariance estimation with the jackknife estimator is significantly biased, calling into question the applicability of the method for real galaxy surveys. Another application is the case of covariance estimation using survey-sized subsamples extracted from a large simulation. We showed that in this case the estimated covariance is also biased, due to the non-independence of the subsamples, though this bias is less important than in the jackknife case.
The main reason for these difficulties is that the large-scale modes that underlie super-sample covariance are not accessible from within the survey, and the largest ones are also inaccessible to simulations.
Although this challenges the use of single simulations for actual covariance estimation, the formalism allows the validation of theoretical prediction codes against single simulations by tweaking the predictions to reproduce the subsampling effect.
Finally, while the formalism and the qualitative conclusions from this article were developed in the case of cluster number counts, they are also valid for other probes, such as galaxy clustering or cosmic shear.
\section*{Acknowledgements}
\vspace{0.2cm}
We thank Camille Bonvin and Kai Hoffmann for discussions that improved this article.
We acknowledge the use of the {\tt Healpix} package by \cite{Gorski2005}.
F.L. and M.K. acknowledge support by the Swiss National Science Foundation.
\bibliographystyle{aa}
|
\section{Introduction}
The Casimir effect is associated with a force which emerges between neutral objects due to vacuum fluctuations of quantum fields~\cite{ref:Casimir,ref:Bogdag,ref:Milton}. The presence of the physical objects affects the spectrum of quantum fluctuations around them, and, therefore, leads to a modification in the (vacuum) energy of these quantum fluctuations. Since the latter depends on mutual positions and orientations of the objects, the fluctuations of the vacuum fields naturally give rise to a force (called sometimes Casimir-Polder force) which acts on these objects~\cite{Casmir:1947hx}.
The simplest and best known example of the Casimir effect is given by interaction of two neutral parallel plates in the vacuum. The zero-point fluctuations of electromagnetic field give rise to attraction of idealized perfectly conducting plates. If the plates are made of real materials and/or are immersed in a medium, then the force between the plates becomes dependent on intrinsic properties of the corresponding materials such as permittivity, permeability and conductivity~\cite{Lifshitz:1956zz}. In certain cases the Casimir forces can even be made repulsive~\cite{ref:repulsive:Casimir} which is of immense practical interest for assembling frictionless (nano)mechanical machines.
Even in free noninteracting theories the calculation of the Casimir forces between objects of general shapes is a difficult task. There are not so many geometrical configurations for which the Casimir energy is known analytically. In order to compute of the Casimir interactions between objects of arbitrary geometries, and in the case if they are made of imperfect materials, a set of dedicated numerical methods has been developed~\cite{Johnson:2010ug,Gies:2006cq}.
The zero-point forces are generally affected by interactions between quantum fields. In the case of Quantum Electrodynamics the one-loop perturbative correction to the Casimir effect is extremely small so that is can safely be neglected~\cite{ref:Milton}. On the other hand, in the fermionic version of the Casimir effect the four-point interaction modifies significantly the Casimir force~\cite{Flachi:2013bc}. In this case the virtual fermions of the vacuum are confined in between ``reflective'' plates with the MIT boundary conditions. The constrained fermionic excitations lead to appearance of the vacuum forces which are acting on the plates. On the other hand, the presence of the plates affects the structure of the fermionic vacuum which, in this model, may exist in the phases with spontaneously broken and restored chiral symmetry. At small separations between the plates the spontaneously broken phase ceases to exist~\cite{Flachi:2013bc,Tiburzi:2013vza}. A similar effect appears also in a cylindrical geometry which may by further enforced by a uniform rotation of the cylinder around its axis~\cite{Chernodub:2016kxh}.
In Ref.~\cite{Chernodub:2016owp} we proposed a general numerical method to study Casimir forces using first-principle simulations of lattice (gauge) theories. These methods, which are widely used in particle physics, are especially convenient for investigation of interacting theories as well as for studying nonperturbative features such as, for example, the phase structure of a given theory in a finite Casimir geometry or nonperturbative corrections to the Casimir force. In order to demonstrate the reliability of the method, in Ref.~\cite{Chernodub:2016owp} we have computed the Casimir energy between thin dielectric wires of finite permittivity in an Abelian gauge theory in two spatial dimensions. We have also shown that in the special case of straight wires of infinite permittivity our approach gives rise to a known analytic result. A similar numerical method for the case of ideal conductors has also been developed in Ref.~\cite{ref:Oleg}.
The method of Ref.~\cite{Chernodub:2016owp} allows us to address straightforwardly nonperturbative properties of the gauge fields. There are two interesting effects associated with nonperturbative dynamics of the gauge fields, these are charge confinement and mass gap generation. Both these phenomena exist in non-Abelian gauge theories, of which is of particular interest is Quantum Chromodynamics. A useful toy Abelian model which possesses both these effects is the compact U(1) gauge theory in two space dimensions. Although there is no matter fields in the theory, it is often called ``compact electrodynamics'' (compact QED) for shortness. We will use this terminology below. Before proceeding further we would like to notice that the compact QED serves as a toy model not only for particle physics, but it is also a viable effective model in a number of condensed matter applications~\cite{ref:Herbut}.
The compactness of the Abelian group leads to the appearance of certain topological objects (often called ``topological defects'') which carry magnetic charges and thus are associated with monopoles. Contrary to the usual particles, in two space time the world-trajectories of these magnetic defects are points, so that the monopoles are, in fact, instantons. At zero temperature these monopoles for a gas, which lead to both nonperturbative properties: to the linear confinement and to the mass gap generation. The confinement reveals itself in the total energy of a pair of particles, which carry positive and negative unit of electric charge. The energy of the pair rises linearly with the distance at large enough separations of the constituents of the pair. The particles in the pair are thus confined since in order to separate them to infinite distance one would need infinite energy. The mass gap generation is revealed via a finite mass of the photon. Both the confining force and the mass of the photon are proportional to the square root of the monopole density, and they cannot be derived in a standard perturbation theory in powers of electric coupling of the model~\cite{Polyakov:1976fu}.
The aim of this paper is to study nonperturbative effects of the monopoles on the Casimir forces between two parallel wires in the vacuum of compact electrodynamics. In addition, we also investigate how the vacuum affects the phase structure of the theory. We concentrate our efforts on zero-temperature case leaving investigation of the finite-temperature effects for a future.
The structure of the paper is as follows. We review compact QED in continuum spacetime in Sect.~\ref{sec:cQED:review}. Then we calculate the influence of the monopole on Casimir energy in the dilute gas approximation in Sect.~\ref{sec:compact:QED:monopoles}. The nonperturbative numerical calculations of the Casimir effect are in the lattice formulation of the model in Sect.~\ref{sec:compact:QED:monopoles}. We summarize our conclusions in the last section.
\section{(2+1) compact electrodynamics}
\label{sec:cQED:review}
In this section we review in details certain well-known features of compact electrodynamics -- i.e. an Abelian theory with compact gauge group -- in two spatial dimensions in continuum spacetime. These properties will later be relevant to our studies of the nonperturbative Casimir effect both in the continuum spacetime (Sect.~\ref{sec:compact:QED:monopoles}) and in the lattice formulation of the model (Sect.~\ref{sec:compact:lattice:QED}).
\subsection{Monopoles and photons}
\subsubsection{Lagrangian of (2+1) compact electrodynamics}
The compact electrodynamics is basically a pure U(1) gauge theory with monopoles. In (3+1) dimensions the monopoles are particle-like objects and their world trajectories are closed lines that share similarly with usual pointlike particles. In (2+1) dimensions the ``trajectories'' of the monopoles are represented as a set of localized points, so that the monopoles are instanton-like objects. We will consider the (2+1) dimensional compact electrodynamics where most calculations can be done analytically. Since we are working in the thermal equilibrium and study stationary phenomena, it is convenient to perform a Wick rotation and consider the compact electrodynamics in three-dimensional Euclidean spacetime. In this section we follow Ref.~\cite{Polyakov:1976fu}.
The Lagrangian of the compact electrodynamics has the same form as the Lagrangian of the usual free U(1) gauge theory:
\begin{eqnarray}
{\cal L} = \frac{1}{4} F^2_{\mu\nu}\,,
\label{eq:L}
\end{eqnarray}
where $F_{\mu\nu}$ is the field strength tensor. However, contrary to a free U(1) gauge theory, the field strength tensor consists of two parts:
\begin{eqnarray}
F_{\mu\nu} = F_{\mu\nu}^{{\mathrm {ph}}} + F_{\mu\nu}^{{\mathrm {mon}}}\,.
\label{eq:F:munu}
\end{eqnarray}
The first, perturbative term is expressed via the vector photon field $A_\mu$,
\begin{eqnarray}
F_{\mu\nu}^{{\mathrm {ph}}}[A] = \partial_\mu A_\nu - \partial_\nu A_\mu\,,
\label{eq:F:ph}
\end{eqnarray}
while the second, nonperturbative monopole part,
\begin{eqnarray}
F_{\mu\nu}^{{\mathrm {mon}}}(\boldsymbol {x}) = - g_{\mathrm {mon}} \epsilon_{\mu\nu\alpha} \partial_\alpha \int d^3 y D(\boldsymbol {x}-\boldsymbol {y}) \rho(\boldsymbol {y})\,,
\label{eq:F:mon}
\end{eqnarray}
is determined by the density of the instanton-like monopoles:
\begin{eqnarray}
\rho(\boldsymbol {x}) = \sum_a q_a \delta^{(3)} \left(\boldsymbol {x} - \boldsymbol {x}_a\right)\,.
\label{eq:rho}
\end{eqnarray}
Here ${\boldsymbol x}_a$ is the position of the $a$-th monopole, $q_a$ is its charge in units of the usual, Dirac-quantized elementary monopole charge,\footnote{
Here we use the standard Dirac quantization~\eq{eq:g:m} as compared to the quantization $g g_{\mathrm {mon}}/(4 \pi ) \in {\mathbb Z}$ which is the more appropriate for the monopole charges of the 't~Hooft-Polyakov monopoles in non-Abelian, grand unified theories~\cite{tHooft:1974kcl,Polyakov:1974ek}.}
\begin{eqnarray}
g_{\mathrm {mon}} = \frac{2 \pi}{g}\,,
\label{eq:g:m}
\end{eqnarray}
which, in turn, is expressed via the elementary electric charge $g$. In three-dimensional Euclidean spacetime the Greek indices run through $\mu,\nu,\alpha, \ldots = 1,2,3$. Notice that in (2+1) Minkowski (or, equivalently) 3 Euclidean dimensions the electric charge is a dimensional quantity, $[g] = {\mathrm{mass}}^{1/2}$, so that the monopole charge is also a dimensionful quantity, $[g_{\mathrm {mon}}] = {\mathrm{mass}}^{-1/2}$.
The monopole field strength tensor~\eq{eq:F:mon} depends nonlocally on the monopole density $\rho(\boldsymbol {x})$. The quantity $D$ in Eq.~\eq{eq:F:mon} is the scalar propagator,
\begin{eqnarray}
D({\boldsymbol x}) = \int \frac{d^3 k}{(2 \pi)^3} \frac{e^{i {\boldsymbol k} \boldsymbol {x}}}{k^2} = \frac{1}{4 \pi | \boldsymbol {x} |}\,,
\label{eq:D}
\end{eqnarray}
which obeys the second-order differential equation:
\begin{eqnarray}
- \Delta D(\boldsymbol {x}) = \delta(\boldsymbol {x})\,,
\end{eqnarray}
where $\Delta \equiv \partial_\mu^2$ is the three-dimensional Laplacian.
The action of the model,
\begin{eqnarray}
S [A,\rho] = \frac{1}{4} \int d^3 x \, F^2_{\mu\nu}
\label{eq:S:continuum}
\end{eqnarray}
decouples into a sum
\begin{eqnarray}
S [A,\rho] & = & \frac{1}{4} \int d^3 x \, \bigl(F^{\mathrm {ph}}_{\mu\nu}[A] + F^{\mathrm {mon}}_{\mu\nu}[\rho] \bigr)^2 \nonumber \\
& \equiv &
S^{\mathrm {ph}}[A] + S^{\mathrm {mon}}[\rho]\,,
\label{eq:S:A:rho}
\end{eqnarray}
of the perturbative photon part
\begin{eqnarray}
S^{\mathrm {ph}}[A] & = & \frac{1}{4} \int d^3 x \, \bigl(F^{\mathrm {ph}}_{\mu\nu}[A]\bigr)^2\,,
\label{eq:S:A}
\end{eqnarray}
and monopole part
\begin{eqnarray}
S^{\mathrm {mon}}[\rho] & = & \frac{g^2_{\mathrm {mon}}}{2} \int d^3 x \, \int d^3 y \, \rho({\boldsymbol x}) D(x-y) \rho({\boldsymbol y}), \quad
\label{eq:S:rho}
\end{eqnarray}
where we have implemented integration by parts and used the explicit forms of the photon and monopole field strengths given in Eqs.~\eq{eq:F:ph} and \eq{eq:F:mon}, respectively. The photon-monopole cross-term disappears due to the identity coming from the explicit form of the monopole field strength~\eq{eq:F:mon}:
\begin{eqnarray}
\partial_\mu F^{\mu\nu}_{\mathrm {mon}} \equiv 0\,.
\label{eq:dF:1}
\end{eqnarray}
Using the explicit expression for the monopole density~\eq{eq:rho} one can rewrite the monopole action~\eq{eq:S:rho} in terms of the Coulomb gas of the monopoles:
\begin{eqnarray}
S^{\mathrm {mon}}[\rho] = \frac{g^2_{\mathrm {mon}}}{2} \sum_{\stackrel{{a,b=1}}{a \neq b}}^N q_a q_b \, D({\boldsymbol x}_a-{\boldsymbol y}_b) + N S_0\,,
\label{eq:S:Coulomb:gas}
\end{eqnarray}
where
\begin{eqnarray}
S_0 = \frac{g^2_{\mathrm {mon}}}{2} D({\boldsymbol 0})\,,
\label{eq:S:0}
\end{eqnarray}
is a divergent term which will be renormalized below. According to Eq.~\eq{eq:D} the Coulomb term in the action~\eq{eq:S:Coulomb:gas} describes long-ranged interaction between the monopole instantons.
A variation of the action~\eq{eq:S:A:rho} with respect to the photon field $A_\mu$ provides us with the usual Maxwell equations for photons:
\begin{eqnarray}
\partial_\mu F^{\mu\nu}_{\mathrm {ph}} = 0\,.
\label{eq:dF:2}
\end{eqnarray}
The physical content of the model is given by the photons and the monopoles. The model does not possesses any matter fields that bear electric charges. Indeed, the electric and magnetic charges are defined via the Maxwell equation and its dual, respectively:
\begin{eqnarray}
\partial_\mu F^{\mu\nu} & = & j^\nu_{\mathrm{ch}}\,,
\label{eq:Maxwell:real}\\
\partial_\mu {\widetilde F}^{\mu} & = & j_{\mathrm {mon}}\,,
\label{eq:Maxwell:dual}
\end{eqnarray}
where $j^\mu_{\mathrm{ch}}$ is the electric (charged) current, $j_{\mathrm {mon}}$ is the magnetic charge density and
\begin{eqnarray}
{\widetilde F}^{\mu} = \frac{1}{2} \epsilon^{\mu\alpha\beta} F_{\alpha\beta}\,,
\end{eqnarray}
is the dual field strength tensor (in 3 spacetime dimensions this tensor is, in fact, a pseudovector).
Substituting~\eq{eq:F:munu} with \eq{eq:F:ph} and \eq{eq:F:mon} into the Maxwell equations~\eq{eq:Maxwell:real} and \eq{eq:Maxwell:dual}, one gets:
\begin{subequations}
\begin{eqnarray}
j^\mu_{\mathrm{ch}}(\boldsymbol {x}) & = & 0\,, \\
j_{\mathrm {mon}}(\boldsymbol {x}) & = & g_{\mathrm {mon}} \rho(\boldsymbol {x})\,.
\end{eqnarray}
\label{eq:J:ch:mon}
\end{subequations}
We have used the equation of motion~\eq{eq:dF:1} along with identity~\eq{eq:dF:2} in order to get the first relation in Eq.~\eq{eq:J:ch:mon}. The second relation in \eq{eq:J:ch:mon} comes from the identity $\partial_\mu {\widetilde F}^{\mu}_{\mathrm {ph}} \equiv 0$ and the explicit form of the monopole field strength tensor~\eq{eq:F:mon}. Thus, we conclude that the spectrum of the model consists of vector photons and instanton-like monopoles only.
The presence of the monopoles is related to the compactness of the gauge group. This property will be evident in the lattice formulation of the theory which will be considered in Sect.~\ref{sec:compact:lattice:QED}.
\subsubsection{Decoupling of photons and monopoles}
The photons and monopoles are the only dynamical degrees of freedom in the model. The partition function
\begin{eqnarray}
Z = \int {\cal D} A \hskip 3mm {\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} e^{ - S[A,\rho]}
\label{eq:Z:0}
\end{eqnarray}
involves the integration over the photon configurations $A$ and the sum over all monopole configurations encoded via the monopole densities $\rho$. The integration measure over the monopole configurations can be written as follows:
\begin{eqnarray}
{\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} = \sum_{N=0}^\infty \frac{1}{N!} \prod_{a=1}^N \left(\sum_{q_a = \pm 1} \zeta \int d^3 x_a \right).
\label{eq:sumint}
\end{eqnarray}
Here the sum goes over the total number of monopoles $N$ (the term with $N=0$ corresponds to a unity in the above sum). The parameter $\zeta$ is the so-called ``fugacity'' which controls the monopole density. We consider the dilute monopole gas so that the monopoles possess a unit charge in elementary magnetic charge~\eq{eq:g:m}, $q_a = \pm 1$. The overlaps between the monopoles are rare and therefore are neglected here. Basically, in Eq.~\eq{eq:sumint} we integrate over positions $\boldsymbol {x}_a$ of all $N$ monopoles, sum over all their magnetic charges $q_a$ and then sum the total monopole number $N$ taking into account the combinatoric degeneracy factor $1/N!$.
Due to the decoupling of the photon and monopole parts of the action~\eq{eq:S:A:rho}, the photon and monopole parts of the partition function~\eq{eq:Z:0} are also decoupled:
\begin{eqnarray}
Z = Z_{\mathrm {ph}} \cdot Z_{\mathrm {mon}}\,,
\label{eq:Z}
\end{eqnarray}
where the photon and monopole actions
\begin{eqnarray}
Z_{\mathrm {ph}} = \int {\cal D} A \, e^{ - S_{\mathrm {ph}}[A]}\,,
\label{eq:Z:ph}\\
Z_{\mathrm {mon}} = \hskip 3mm {\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} e^{ - S_{\mathrm {mon}}[\rho]}\,,
\label{eq:Z:mon}
\end{eqnarray}
are given in Eqs.~\eq{eq:S:A} and \eq{eq:S:rho}, respectively.
\subsection{Dual formulation of monopole dynamics}
\subsubsection{Coulomb gas as a sign-Gordon model}
The partition function of the photon part has the perturbative Gaussian form~\eq{eq:Z:ph}. The monopole partition function~\eq{eq:Z:mon}, which describes nonperturbative effects, can be reformulated in terms of a nonlinear sine-Gordon model. Following Ref.~\cite{Polyakov:1976fu}
we rewrite the monopole partition function~\eq{eq:Z:mon} as follows:
\begin{widetext}
\begin{subequations}
\begin{eqnarray}
Z_{\mathrm {mon}} & & = \hskip 3mm {\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} e^{ - S_{\mathrm {mon}}[\rho]}
\ \mathop{\equiv}_{{\color{black}{\mathrm{(a)}}}} \ \sum_{N=0}^\infty \frac{1}{N!} \prod_{a=1}^N \left(\sum_{q_a = \pm 1} \zeta \int d^3 x_a \right) \exp\left[- \frac{g^2_{\mathrm {mon}}}{2} \int d^3 x \, \int d^3 y \, \rho(x) D(x-y) \rho(y)\right]
\label{eq:Z:mon:derivation:a}\\
& & \mathop{\equiv}_{{\color{black}{\mathrm{(b)}}}} \ C \int {\cal D} \chi \sum_{N=0}^\infty \frac{1}{N!} \prod_{a=1}^N \left(\sum_{q_a = \pm 1} \zeta \int d^3 x_a \right) \exp\left\{ - \int d^3 x \, \biggl[\frac{1}{2 g^2_{\mathrm {mon}}} \bigl(\partial_\mu \chi(x) \bigr)^2 + i \chi(x) \rho (x) \biggr]\right\} \\
& & \mathop{\equiv}_{{\color{black}{\mathrm{(c)}}}} \ C \int {\cal D} \chi \exp\left\{ - \frac{1}{2 g^2_{\mathrm {mon}}} \int d^3 x \, \bigl(\partial_\mu \chi(x) \bigr)^2\right\}
\sum_{N=0}^\infty \frac{1}{N!} \prod_{a=1}^N \left(\sum_{q_a = \pm 1} \zeta \int d^3 x_a \, e^{ - i q_a \chi(x_a) } \right) \\
& & \mathop{\equiv}_{{\color{black}{\mathrm{(d)}}}} \ C \int {\cal D} \chi \exp\left\{ - \frac{1}{2 g^2_{\mathrm {mon}}} \int d^3 x \, \bigl(\partial_\mu \chi\bigr)^2\right\}
\sum_{N=0}^\infty \frac{1}{N!} \left(2 \zeta \int d^3 x \, \cos \chi(x) \right)^N \\
& & \mathop{\equiv}_{{\color{black}{\mathrm{(e)}}}} \ C \int {\cal D} \chi \exp\left\{ - \int d^3 x \, \left[\frac{1}{2 g^2_{\mathrm {mon}}} \bigl(\partial_\mu \chi \bigr)^2
- 2 \zeta \cos \chi \right] \right\} \equiv C \int {\cal D} \chi \exp\left\{ - \int d^3 x \, {\cal L}_{s}(\chi) \right\} \,,
\label{eq:Z:mon:derivation:e}
\end{eqnarray}
\label{eq:Z:mon:derivation}
\end{subequations}
\end{widetext}
where
\begin{eqnarray}
{\cal L}_s = \frac{1}{2 g^2_{\mathrm {mon}}} \bigl(\partial_\mu \chi \bigr)^2 - 2 \zeta \cos \chi\,,
\label{eq:cL:sine}
\end{eqnarray}
is the Lagrangian of the sine-Gordon model. Here and below $C$ stands for an inessential constant parameter.
The steps (a)--(e), which are marked in the chain of relations~\eq{eq:Z:mon:derivation} are as follows: (a) the expression under the exponential of the monopole partition function~\eq{eq:Z:mon}, \eq{eq:S:rho} and \eq{eq:sumint} can be linearized with the help of a real-valued scalar field $\chi$ introduced by the Gaussian integration~(b). Notice that the divergent monopole (self) action~\eq{eq:S:0} can be absorbed (renormalized) into the definition of the fugacity, $\zeta \to \zeta e^{S_0}$. Then the explicit expression for the monopole density~\eq{eq:rho} can be used (c) to perform the sum (d) over the monopole charges $q_a = \pm 1$. Finally, the the sum over the total monopole number $N$ is converted into the exponent in the last step (e).
Thus, we have represented the theory of monopoles in terms of the field theory with the Lagrangian~\eq{eq:cL:sine}. The latter corresponds to a dual formulation of the Coulomb gas of the monopoles~\eq{eq:S:rho}, in which the dynamics of the original monopole density~\eq{eq:rho} is described by a scalar real-valued field $\chi$.
\subsubsection{Monopole density, photon mass and confinement}
The dual theory~\eq{eq:cL:sine} is very convenient in investigation of nonperturbative properties of associated with the monopole dynamics. First of all, we notice that it is very easy to calculate the mean monopole density $\varrho_{\mathrm {mon}} \equiv \avr{|\rho|}$ in the Coulomb monopole gas~\eq{eq:Z:0}. We notice that
\begin{eqnarray}
\varrho_{\mathrm {mon}} \equiv \avr{N} = \frac{\partial \ln Z_{\mathrm {mon}}}{\partial \ln \zeta}\,,
\label{eq:rho:mon}
\end{eqnarray}
where we used the explicit form of the monopole partition function in Eq.~\eq{eq:Z:mon:derivation:a}. Repeating all steps down to Eq.~\eq{eq:Z:mon:derivation:e} and using the explicit form of the dual Lagrangian~\eq{eq:cL:sine} we get for the monopole density~\eq{eq:rho:mon} the following expression $\varrho_{\mathrm {mon}} = 2 \zeta \avr{\cos \chi}$. In the leading order the mean monopole density is,
\begin{eqnarray}
\varrho_{\mathrm {mon}} = 2 \zeta\,,
\label{eq:density}
\end{eqnarray}
where we have neglected the quantum corrections due to fluctuations of the dual field $\chi$.
The field $\chi$ in Eq.~\eq{eq:cL:sine} has the mass
\begin{eqnarray}
m_{\mathrm {ph}} = g_{\mathrm {mon}} \sqrt{2 \zeta} \equiv \frac{2\pi \sqrt{2 \zeta}}{g}\,,
\label{eq:m:ph}
\end{eqnarray}
where we have used Eq.~\eq{eq:g:m} and expanded the sine-Gordon Lagrangian~\eq{eq:cL:sine} over small fluctuations of the dual field~$\chi$:
\begin{eqnarray}
{\cal L}_s = \frac{1}{2 g^2_{\mathrm {mon}}} \left[\bigl(\partial_\mu \chi \bigr)^2 + m^2_{\mathrm {ph}} \, \chi^2 \right] + O(\chi^4) \,.
\label{eq:cL:sine:exp}
\end{eqnarray}
Despite the field $\chi$ is associated with the monopoles, the mass~\eq{eq:cL:sine:exp} eventually becomes the mass of the photon (gauge) field. The gauge field $A_\mu$ can be considered to be composed of the regular photon field and singular monopole field which leads to the decomposition of the field strength tensor~\eq{eq:F:munu}. One can show that in the photon-mediated interactions the massless pole of the regular part cancels out and the massive pole of the field $\chi$ determines the interaction range.
Notice that according to Eq.~\eq{eq:density} the photon mass~\eq{eq:m:ph} can be directly expressed, in the leading order, via the monopole density $\varrho_{\mathrm {mon}}$ and the electric charge $g$:
\begin{eqnarray}
m_{\mathrm {ph}} = \frac{2\pi \sqrt{\varrho_{\mathrm {mon}}}}{g}\,.
\label{eq:m:ph:rho}
\end{eqnarray}
Thus, the monopole gas leads to the nonperturbative mass gap generation in the system. The emergent mass of the photon field is proportional to the square root of the monopole density~\eq{eq:m:ph:rho}.
Equations~\eq{eq:density} and \eq{eq:m:ph:rho} and subsequent relations are valid provided fluctuations of the sine-Gordon field $\chi$ are small, $\avr{\chi^2} \ll 1$, and the relevance of higher-than-quadratic terms in the sine-Gordon model~\eq{eq:cL:sine} is negligible. In terms of monopoles themselves, this condition is realized provided the fluctuations of individual monopoles can be neglected. Since a monopole may affect another monopole at a typical distance of the Debye length $\lambda_D = m_{\mathrm {ph}}^{-1}$, the fluctuations of individual monopoles are negligible provided the number of monopoles in a unit Debye volume is sufficiently high, $\varrho_{\mathrm {mon}} \lambda_D^3 \gg 1$. Using Eq.~\eq{eq:m:ph:rho} the applicability requirement may be reformulated as follows:
\begin{eqnarray}
\varrho^{1/2}_{\mathrm {mon}} \ll \frac{g^3}{(2\pi)^3}, \qquad\ \mbox{or} \qquad\ \varrho^{1/2}_{\mathrm {mon}} \; g_{\mathrm {mon}}^3 \ll 1\,.
\label{eq:applicability}
\end{eqnarray}
In other words, our estimations are valid provided the monopole density, expressed in units of magnetic charge~\eq{eq:g:m}, is very small~\eq{eq:applicability}. Therefore Eq.~\eq{eq:applicability} is often called the dilute gas approximation.
Another important property of the compact electrodynamics is the linear confinement of electric charges. It turns out that a pair of static, electrically charged particle and antiparticle separated by sufficiently large distance $R \gg \lambda_D$ experiences the confining potential $V(R) = \sigma R$ which grows linearly with the distance $R$.
The string tension is given by the following formula:
\begin{eqnarray}
\sigma = \frac{8 \sqrt{2 \zeta}}{g_{\mathrm {mon}}} \equiv \frac{4 g \sqrt{\varrho_{{\mathrm {mon}}}}}{\pi}
\label{eq:string:tension}
\end{eqnarray}
The compact electrodynamics is one of a few field-theoretical models where the linear confinement property may be proved analytically. Since the confinement property is not a central topic of our study, we leave out the derivation of Eq.~\eq{eq:string:tension} and refer an interested reader to Ref.~\cite{Polyakov:1976fu} for further details.
\section{Casimir effect and monopoles: analytical arguments}
\label{sec:compact:QED:monopoles}
In this section we calculate analytically the effect of dynamical monopoles on the Casimir interaction between two perfectly conducting wires in the limit when the monopole gas is sufficiently dilute~\eq{eq:applicability}. In our derivation we follow the general line of Ref.~\cite{Chernodub:2016owp}. Our analytical calculations of this section will be supplemented by the results of numerical simulations described in the next section.
\subsection{Casimir boundary conditions in integral form}
In two spatial dimensions the basic Casimir problem is formulated for one-dimensional objects. We call these objects as ``wires''. These wires are similar to canonical plates used in the studies of the Casimir (zero-point) interactions in three spatial dimensions.
The effect of a perfectly conducting metallic wire on electromagnetic field is simple. A static and infinitely thin wire makes the tangential component of the electric field vanishing at every point $\boldsymbol {x}$ of wire:
\begin{eqnarray}
E_\| (\boldsymbol {x})
= 0\,.
\label{eq:E:parallel}
\label{eq:F01:3d}
\end{eqnarray}
In order to generalize the Casimir boundary condition~\eq{eq:E:parallel} to the most general case of non-static (moving) wires of an arbitrary shape, let us describe the two-dimensional world surface $S$ of the wire by a vector ${\bar \boldsymbol {x}} = {\bar \boldsymbol {x}}(\tau,\xi)$ parameterized by the timelike ($\tau$) and spacelike ($\xi$) parameters. The surface element of $S$ can be described by the singular asymmetric tensor function
\begin{eqnarray}
s_{\mu\nu}(\boldsymbol {x}) = \int d \tau \int d \xi \frac{\partial {\bar x}_{[\mu,}}{\partial \tau} \frac{\partial {\bar x}_{\nu]}}{\partial \xi}
\delta^{(3)}\left(\boldsymbol {x} - {\bar \boldsymbol {x}}(\tau,\xi)\right), \quad
\label{eq:s:munu:gen}
\end{eqnarray}
where $a_{[\mu,} b_{\nu]} = a_\mu b_\nu - a_\nu b_\mu$. Consequently, the boundary condition~\eq{eq:E:parallel} can be rewritten, for any point $\boldsymbol {x}$, in the explicitly covariant form:
\begin{eqnarray}
F^{\mu\nu}(\boldsymbol {x}) s_{\mu\nu}(\boldsymbol {x}) = 0\,,
\label{eq:F:0:cov}
\end{eqnarray}
where $F_{\mu\nu}$ is the field strength tensor~\eq{eq:F:munu}.
In this paper we are interested in zero-point interactions between two parallel wires. The geometry of our problem is illustrated in Fig.~\ref{fig:geometry:plane}.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.375,clip=true]{wires2d_boundaries.pdf}
\end{center}
\vskip -2mm
\caption{The Casimir problem in two spatial dimensions: the wires $l_1$ and $l_2$ are separated by the distance $R$. The two-dimensional space is compactified into a torus due to periodic boundary conditions.}
\label{fig:geometry:plane}
\end{figure}
In our calculation below the wires are assumed to be perfectly conducting so that they require the tangential component $E_\parallel(x)$ of the electric field to vanish at every point of the wire~$x$.
A pair of two static wires placed at $x_1 = \pm R/2$ and parallel to the $x_2$ axis can be parametrized with a help of the pair of vector:
\begin{eqnarray}
{\bar \boldsymbol {x}}_\pm(\tau,\xi) \equiv (x_1, x_2, x_3) = \left(\pm \frac{R}{2}, \xi,\tau\right)\,,
\label{eq:bx:pm}
\end{eqnarray}
where the subscript ``$\pm$'' corresponds to the right/left wire, respectively. The wires are static with respect to the time direction $x_3$. The tensors~\eq{eq:s:munu:gen} for their world surfaces are as follows:
\begin{eqnarray}
s^\pm_{\mu\nu}(\boldsymbol {x}) = \left(\delta_{\mu,2} \delta_{\nu,3} - \delta_{\nu,3} \delta_{\mu,2} \right) \delta(x_1 \mp R/2)\,.
\label{eq:s:pm}
\end{eqnarray}
The boundary condition for our static straight wires can be read off from Eqs.~\eq{eq:s:pm} and \eq{eq:F:0:cov}:
\begin{eqnarray}
F_{23}(\pm R/2, x_2, x_3) = 0\,.
\label{eq:F23:0}
\end{eqnarray}
This relation coincides with Eq.~\eq{eq:E:parallel} because $F_{23} \equiv E_2$ is the component of the electric field that is parallel to the wires.
In the path-integral formalism the Casimir condition~\eq{eq:F:0:cov} may conveniently be implemented with the help of the following $\delta$ functional:
\begin{eqnarray}
\delta_S[F] = \prod_{\boldsymbol {x}} \delta\Bigl(F^{\mu\nu}(\boldsymbol {x}) s_{\mu\nu}(\boldsymbol {x})\Bigr)\,.
\end{eqnarray}
The infinite product of the $\delta$ functions may be rewritten with the help of the functional integration over the Lagrange multiplier $
\lambda({\boldsymbol x})$:
\begin{eqnarray}
\delta_S[F] & = & \int {\cal D} \lambda \exp\left[ \frac{i}{2} \int d^3 x \, \lambda(\boldsymbol {x}) F^{\mu\nu}(\boldsymbol {x}) s_{\mu\nu}(\boldsymbol {x}) \right]
\nonumber \\
& \equiv & \int {\cal D} \lambda \exp\left[ \frac{i}{2} \int d^3 x \, F^{\mu\nu}(\boldsymbol {x}) J_{\mu\nu}(\boldsymbol {x};\lambda) \right],
\label{eq:S:F}
\end{eqnarray}
where the ``surface tensor''
\begin{eqnarray}
J_{\mu\nu}(\boldsymbol {x};\lambda) = \lambda(\boldsymbol {x}) s_{\mu\nu}(\boldsymbol {x}),
\label{eq:J:munu}
\end{eqnarray}
is the product of the Lagrange multiplier and the surface tensor~\eq{eq:s:munu:gen}. In our case of the two parallel Casimir plates~\eq{eq:bx:pm} we have:
\begin{eqnarray}
\delta_S[F] = \int {\cal D} \lambda_+ \int {\cal D} \lambda_- \exp\biggl[ i \int d x_2 \int d x_3 \, \nonumber\\
\sum_{a=\pm 1} \lambda_a(x_2,x_3) F_{23}\left(a \frac{R}{2},x_2,x_3\right) \biggr].
\label{eq:delta:S:plates}
\end{eqnarray}
The integration under the exponent is taking place along the two-dimensional world surface, and the integrations over the Lagrange multipliers $\lambda_+$ and $\lambda_-$ enforce the Casimir conditions~\eq{eq:F23:0} at the flat world surfaces for the right ($x_1 = + R/2$) and left ($x_1 = - R/2$) wire, respectively.
The partition function~\eq{eq:Z}, \eq{eq:Z:ph} and \eq{eq:Z:mon} in the presence of the Casimir surface $S$ is then given by following compact formula:
\begin{eqnarray}
Z_S = \int {\cal D} A \hskip 3mm {\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} \, e^{ - S_{\mathrm {ph}}[A] - S_{\mathrm {mon}}[\rho]} \, \delta_S[F]\,.
\label{eq:Z:S}
\end{eqnarray}
\subsection{Casimir boundaries in dual sine-Gordon theory}
Since the field strength tensor contains both the photon and monopole parts, $F = F_{\mathrm {ph}} + F_{\mathrm {mon}}$, the decoupling of the photon and monopole partition functions~\eq{eq:Z} is no longer possible in the presence of the Casimir surfaces~\eq{eq:Z:S}. It is clearly seen from the partition function~\eq{eq:Z:S} which can be represented in the following functional form
\begin{eqnarray}
Z_S & = & \int {\cal D} \lambda\, Z_{\mathrm {ph}}[\lambda] Z_{\mathrm {mon}}[\lambda], \quad
\label{eq:Z:S:lambda}\\
Z_{\mathrm {ph}}[\lambda] & = & \int {\cal D} A e^{ - S_{\mathrm {ph}}[A] + \frac{i}{2} \int d^3 x \, F^{\mu\nu}_{\mathrm {ph}}(\boldsymbol {x}) J_{\mu\nu}(\boldsymbol {x};\lambda)}, \quad
\label{eq:Z:ph:lambda}\\
Z_{\mathrm {mon}}[\lambda] & = & \hskip 3mm {\mathclap{\displaystyle\int}\mathclap{\textstyle\sum}}_{{\mathrm {mon}}} \, e^{- S_{\mathrm {mon}}[\rho] + \frac{i}{2} \int d^3 x \, F^{\mu\nu}_{\mathrm {mon}}(\boldsymbol {x}) J_{\mu\nu}(\boldsymbol {x};\lambda)}, \quad
\label{eq:Z:mon:lambda}
\end{eqnarray}
where the surface current $J_{\mu\nu}$ is given in Eq.~\eq{eq:J:munu}.
Performing the Gaussian integration over the photon field $A_\mu$, we get for the photon part~\eq{eq:Z:ph:lambda} the following expression~\cite{Chernodub:2016owp}:
\begin{eqnarray}
& & Z_{\mathrm {ph}}[\lambda] {=} \int {\cal D} A \, \exp\left[ \int d^3 x \, \left(- \frac{1}{4} F^2_{\mu\nu} + i A_\mu J^\mu\right)\right] \nonumber \\
& & {=} \, C \exp\left[ - \frac{1}{2} \int d^3 x \, d^3 y \, J_\mu(\boldsymbol {x}; \lambda) D(\boldsymbol {x} - \boldsymbol {y}) J_\mu(\boldsymbol {y}; \lambda) \right]\!, \qquad
\label{eq:Z:ph:lambda:1}
\end{eqnarray}
where $C$ stands for an inessential constant which will be omitted below.
The surface tensor~\eq{eq:J:munu} enters Eq.~\eq{eq:Z:ph:lambda:1} via the conserved vector
\begin{eqnarray}
J_\mu(\boldsymbol {x}; \lambda) = \partial^\nu J_{\mu\nu}(\boldsymbol {x};\lambda)\,, \qquad \partial^\mu J_\mu(\boldsymbol {x};\lambda) = 0\,.
\label{eq:J:mu}
\end{eqnarray}
Using Eqs.~\eq{eq:s:pm} and \eq{eq:J:munu} we find that for the two parallel straight wires~\eq{eq:bx:pm} the current~\eq{eq:J:mu} is given by the following expression:
\begin{eqnarray}
J_\mu = \sum_{a = \pm } \delta\biggl(x_1 - \frac{a R}{2} \biggr) \biggl(\delta_{\mu2} \frac{\partial \lambda_a}{\partial x_3} - \delta_{\mu3} \frac{\partial \lambda_a}{\partial x_2} \biggr)\,,
\label{eq:J:mu:flat}
\end{eqnarray}
where $\lambda_\pm = \lambda_\pm (x_2,x_3)$ are the Lagrange multipliers associated with the left and right plates, respectively.
The monopole part~\eq{eq:Z:mon:lambda} can be evaluated in analogy with the chain of transformations~\eq{eq:Z:mon:derivation}. Using the explicit form of the monopole field strength tensor~\eq{eq:F:mon} we get the following representation of the monopole partition function in the sine-Gordon form:
\begin{eqnarray}
Z_{\mathrm {mon}}[\lambda] = \int {\cal D} \chi \exp\left\{ - \int d^3 x \, {\cal L}_{s}(\chi;\lambda) \right\} \,,
\label{eq:Z:mon:chi}
\end{eqnarray}
where the kinetic term in the sine-Gordon action
\begin{eqnarray}
{\cal L}_s(\chi;\lambda) = \frac{1}{2 g^2_{\mathrm {mon}}} \Bigl[\partial_\mu \bigl(\chi(\boldsymbol {x}) - q(\boldsymbol {x},\lambda)\bigr)\Bigr]^2 {-} 2 \zeta \cos \chi (\boldsymbol {x}), \qquad
\label{eq:cL:lambda:sine}
\end{eqnarray}
is modified by a ``curl'' of the surface tensor~\eq{eq:J:munu}:
\begin{eqnarray}
q(\boldsymbol {x},\lambda) = \frac{g_{\mathrm {mon}}}{2} \int d^3 y \, D(\boldsymbol {x} - \boldsymbol {y}) \epsilon_{\alpha\mu\nu} \partial^\alpha J^{\mu\nu} (\boldsymbol {y};\lambda)\,.
\label{eq:q:general}
\end{eqnarray}
For two parallel straight lines~\eq{eq:bx:pm} the explicit form of the scalar function~\eq{eq:q:general} is as follows:
\begin{eqnarray}
q(\boldsymbol {x}) & = & g_{\mathrm {mon}} \int d y_2 d y_3 \sum_{a=\pm} \lambda_a(y_2,y_3) \\
& & \cdot \frac{\partial}{\partial x_1} D\left(x_1 - \frac{a R}{2},x_2 - y_2,x_3 - y_3\right). \nonumber
\label{eq:q:flat}
\end{eqnarray}
Finally, we substitute the photon~\eq{eq:Z:ph:lambda:1} and monopole~\eq{eq:Z:mon:chi} partition functions into Eq.~~\eq{eq:Z:S:lambda} and get for the total partition function in the presence of the Casimir boundaries:
\begin{widetext}
\begin{eqnarray}
Z_S {=} \int {\cal D} \lambda {\cal D} \chi
\exp\left[ - \frac{1}{2} \int d^3 x \, d^3 y \, J_\mu(\boldsymbol {x}; \lambda) D(\boldsymbol {x} - \boldsymbol {y}) J_\mu(\boldsymbol {y}; \lambda)
{-} \int d^3 x \, \left(\frac{1}{2 g^2_{\mathrm {mon}}} \Bigl[\partial_\mu \bigl(\chi(\boldsymbol {x}) - q(\boldsymbol {x},\lambda)\bigr)\Bigr]^2 {-} 2 \zeta \cos \chi (\boldsymbol {x})\right)\right]\!,
\qquad
\label{eq:Z:lambda:tot}
\end{eqnarray}
\end{widetext}
where the current~$J_\mu$ and the scalar $q$ are given in Eqs.~\eq{eq:J:mu} and \eq{eq:q:flat}, respectively.
It is interesting to notice the effect of the Casimir boundary condition on the dual sine-Gordon field $\chi$ is quite nontrivial. One could naively expect that the requirement of vanishing of the electric field~\eq{eq:E:parallel} at the surface of the wire would lead either to Dirichlet or to Neumann boundary condition for the dual field $\chi$ at the worldsheet of the wire (the field $\chi$ or its normal derivative, respectively, would vanish at the position of the wire). According to our result~\eq{eq:Z:lambda:tot} this naive expectation is not true.
In the partition function~\eq{eq:Z:lambda:tot} the nonlinear term in sine-Gordon field may be expanded in powers of the scalar field $\chi(\boldsymbol {x})$ while the local interaction terms $\chi^{n}(\boldsymbol {x})$ with $n \geqslant 4$ may be neglected in the leading order of the dilute gas approximation~\eq{eq:applicability}. Therefore the functional in the exponential becomes quadratic both in the field $\chi$ and in the field $\lambda$ so that the corresponding integrals become Gaussian and thus can be easily evaluated. Below we calculate them explicitly for the case of two parallel static wires.
\subsection{Casimir potential for parallel wires}
For two static straight wires separated by the distance $R$ the density of the Casimir energy $V(R)$ per unit length of the wire is given by the following formula
\begin{eqnarray}
V(R) = - \frac{1}{{\cal A}} \ln Z_{W_R}\,,
\label{eq:Z:W:R}
\end{eqnarray}
where ${\cal A} = T L$ is the area of the worldsheet of each of the wires. Both the length of the wire $L$ and the time of their existence $T$ are assumed to be very (infinitely) long. The partition function $Z_{W_R}$ is explicitly given by Eq.~\eq{eq:Z:lambda:tot} where the surface $S = W_R \equiv W_{R_+} \cup W_{R_-}$ is represented by two flat sheets corresponding to the parallel wires. For the sake of convenience, below we repeat a known derivation of the Casimir potential~\eq{eq:Z:W:R} in the absence of the monopoles and then we evaluate the effect of the monopole gas.
\subsubsection{Casimir energy in the absence of monopoles}
The density of monopoles~\eq{eq:density} is proportional to the fugacity parameter $\zeta$. By setting $\zeta=0$ we remove the monopoles from the ensembles. As a consequence, the nonlinear term in the sine-Gordon Lagrangian~\eq{eq:cL:sine} disappears, and
the partition function~\eq{eq:Z:lambda:tot} becomes independent of the functional $q$ because it can now be absorbed in the sine-Gordon field $\chi$ by the shift $\chi \to \chi + q$. Then in Eq.~\eq{eq:Z:lambda:tot} the field $\chi$ can be integrated out exactly:
\begin{eqnarray}
Z_S = \int {\cal D} \lambda \, e^{ - \frac{1}{2} \int d^3 x \, d^3 y \, J_\mu(\boldsymbol {x}; \lambda) D(\boldsymbol {x} - \boldsymbol {y}) J_\mu(\boldsymbol {y}; \lambda)}\,.
\label{eq:Z:lambda:nomon}
\end{eqnarray}
Next, using the explicit form~\eq{eq:J:mu:flat} of the current $J_\mu(\boldsymbol {x}; \lambda)$ one gets for the partition function~\eq{eq:Z:lambda:nomon} of the plates:
\begin{eqnarray}
Z_{W_R} = \int {\cal D} \Lambda \, e^{ - \frac{1}{2} \int d^2 x \, d^2 y \, \Lambda^T(\vec {x}) {\widehat K}(\vec {x} - \vec {y}) \Lambda(\vec {y})}\,,
\label{eq:Z:lambda:nomon:1}
\end{eqnarray}
where we introduced the two-dimensional vector on the worldsheets of the wires, $\vec {x} = (x_2,x_3)$. We also introduced the vector field:
\begin{eqnarray}
\Lambda(\vec {x}) =
\left(
\begin{array}{c}
\lambda_+ (\vec {x}) \\
\lambda_- (\vec {x})
\end{array}
\right)\,,
\end{eqnarray}
and the matrix
\begin{eqnarray}
{\widehat K}(\vec {x}) & = & \left(\frac{\partial^2}{\partial x_2^2} + \frac{\partial^2}{\partial x_3^2}\right)
\label{eq:K} \\
& & \left(
\begin{array}{ll}
D(0, x_2, x_3) & D(- R, x_2, x_3) \\
D(+ R, x_2, x_3) & D(0, x_2, x_3)
\end{array}
\right)\,, \nonumber
\end{eqnarray}
where the function $D(\boldsymbol {x})$ given in Eq.~\eq{eq:D}.
The Gaussian integral~\eq{eq:Z:lambda:nomon:1} is given (up to a multiplicative constant) by the determinant of the operator~\eq{eq:K}:
\begin{eqnarray}
Z_{W_R} = {\mathrm{det}}^{-1/2}\, {\widehat K}\,.
\label{eq:Z:lambda:nomon:2}
\end{eqnarray}
The density of the Casimir energy is given by Eqs.~\eq{eq:Z:W:R}:
\begin{eqnarray}
V(R) = \frac{1}{2 {\cal A}} {\mathrm{Tr}} \log {\widehat K}.
\label{eq:A:V:R}
\end{eqnarray}
Formally, Eq.~\eq{eq:A:V:R} can be written as a sum over all eigenvalues~$\kappa_i$
\begin{eqnarray}
{\mathrm{Tr}} \log {\widehat K} \equiv \sum_i \log \kappa_i\,,
\label{eq:Tr:log:formal}
\end{eqnarray}
of the integral operator ${\widehat K}$:
\begin{eqnarray}
{\widehat K} L_i = \kappa_i L_i\,,
\label{eq:eigen:K}
\end{eqnarray}
which can be rewritten in the explicit form as follows:
\begin{eqnarray}
\int d^2 y \, {\widehat K}(\vec {x} - \vec {y}) L(\vec {y}) = \kappa L(\vec {x})\,.
\label{eq:K:L:xy}
\end{eqnarray}
It is convenient to represent the operator $\widehat K$, given by Eqs.~\eq{eq:K} and \eq{eq:D}, and its eigenvalues $L_i(\vec {x})$ as the Fourier integrals,
\begin{eqnarray}
{\widehat K}(\vec {x}) & = &- \int \frac{d^3 k}{(2\pi)^3} \frac{p_2^2 + p_3^2}{p_1^2 + p_2^2 + p_3^2}
\left(
\begin{array}{ll}
1 & e^{-i p_1 R} \\
e^{i p_1 R} & 1
\end{array}
\right),\qquad
\label{eq:K:p} \\
L_i(\vec {x}) & = & \int \frac{d^2 q}{(2 \pi)^2} L_i(\vec {q}) e^{i \vec {q} \vec {x}}\,.
\label{eq:L:x}
\end{eqnarray}
Then we substitute Eqs.~\eq{eq:K:p} and \eq{eq:L:x} into Eq.~\eq{eq:K:L:xy}, integrate over $\vec {y}$ and get the following eigenvalue equation:
\begin{eqnarray}
\int \frac{d^2 q}{(2 \pi)^2} \left[{\widehat Q}(R,{\vec q}) - \kappa_i \right] L_i(\vec {q}) e^{i \vec {q} \vec {x}} = 0\,,
\label{eq:eigen:1}
\end{eqnarray}
where
\begin{eqnarray}
{\widehat Q}(R,{\vec q}) = - \frac{|\vec q|}{2}
\left(\begin{array}{cc}
1 & e^{- |{\vec q}| R} \\
e^{- |{\vec q}| R} & 1
\end{array}\right)\,.
\label{eq:hat:Q}
\end{eqnarray}
Since Eq.~\eq{eq:eigen:1} should be valid for all vectors $\vec {q}$, we arrive to the following equation for the eigenmodes:
\begin{eqnarray}
\left[{\widehat Q}(R,{\vec q}) - \kappa_i \right] L_i(\vec {q}) = 0\,,
\label{eq:eigen:2}
\end{eqnarray}
which has the following eigenvalues $\kappa_i \equiv \kappa_\pm (\vec {q})$:
\begin{eqnarray}
\kappa_\pm (\vec {q}) = - \frac{|\vec {q}|}{2} \left(1 \pm e^{- |\vec {q}| R}\right)\,.
\label{eq:kappa:pm}
\end{eqnarray}
The solutions are characterized by the discrete index $\pm$ and continuous parameter $\vec {q}$. The phase space associated with these variables is
\begin{eqnarray}
{\mathrm{Tr}}_{\vec {q}} \equiv {\cal A} \int \frac{d^2 q}{(2\pi)^2} \sum_\pm\,,
\label{eq:Tr:q}
\end{eqnarray}
where ${\cal A}$ is the area in the transverse $\vec {x} \equiv (x_2,x_3)$ plane.
Substituting Eqs.~\eq{eq:kappa:pm} and \eq{eq:Tr:q} into \eq{eq:A:V:R} and evaluating the integrals explicitly we get the known result for the density of the Casimir energy in the absence of monopoles:
\begin{eqnarray}
V_{\mathrm{Cas}}(R) & = & \frac{1}{2} \int \frac{d^2 q}{(2\pi)^2} \log \left(1 - e^{- 2 |\vec {q}| R}\right) \nonumber \\
& & = - \frac{\zeta(3)}{16 \pi} \frac{1}{R^2}\,,
\label{eq:V:Cas:R}
\end{eqnarray}
where $\zeta(x)$ is the zeta-function with $\zeta(3) \approx 1.20206$. As usual, in our calculation a divergent $R$-independent contribution to the potential $V(R)$ has been been omitted.
\subsubsection{Casimir energy in the presence of monopoles}
Now let us consider the case of the finite monopole density which is given by a nonzero fugacity parameter $\zeta$ in Eq.~\eq{eq:Z:lambda:tot}.
We expand in Eq.~\eq{eq:Z:lambda:tot} the sine-Gordon action over the small fluctuations of the dual field~\eq{eq:cL:sine:exp}, and arrive to the following representation of the partition function:
\begin{eqnarray}
Z_S & = & \int {\cal D} \lambda {\cal D} \chi \exp\biggl\{ \frac{1}{2} \left(J_\mu(\lambda), \Delta^{-1} J_\mu(\lambda)\right) \nonumber\\
&& - \frac{1}{2 g^2_{\mathrm {mon}}} \int d^3 x \, \Bigl[\Bigl(\partial_\mu \bigl(\chi - q(\lambda) \bigr)\Bigr)^2 + m_{\mathrm {ph}}^2 \chi^2 \Bigr]\biggr\}\,.
\qquad \label{eq:Z:lambda:2}
\end{eqnarray}
Here and below we use the following shorthand notations:
\begin{eqnarray}
& & D = - \Delta^{-1}\,, \\
& & \int d^3 y \, D(\boldsymbol {x} - \boldsymbol {y}) A(\boldsymbol {y}) = - \Delta^{-1} A\,, \\
& & \int d^3 x \, d^3 y \, A(\boldsymbol {x}) D(\boldsymbol {x} - \boldsymbol {y}) B(\boldsymbol {y}) = - \left(B, \Delta^{-1} A \right).
\end{eqnarray}
In Eq.~\eq{eq:Z:lambda:2} we omit the $O(\chi^4)$ interaction terms which, in the dilute gas approximation~\eq{eq:applicability}, represent next to the leading order corrections.
Integrating out the Gaussian field $\chi$ in Eq.~\eq{eq:Z:lambda:2} we get
\begin{eqnarray}
Z_S & = & \int {\cal D} \lambda {\cal D} \chi \exp\biggl\{ \frac{1}{2} \left(J_\mu(\lambda), \Delta^{-1} J_\mu(\lambda)\right) \nonumber\\
&& - \zeta \biggl(q(\lambda), \frac{\Delta}{-\Delta + m_{\mathrm {ph}}^2} q(\lambda) \biggr) \Bigr]\biggr\}\,.
\qquad \label{eq:Z:lambda:3}
\end{eqnarray}
Then, for the case of the flat world surfaces, we use Eqs.~\eq{eq:J:mu:flat} and \eq{eq:q:flat} and rewrite Eq.~\eq{eq:Z:lambda:3} in the form similar to Eq.~\eq{eq:Z:lambda:nomon:1}:
\begin{eqnarray}
Z_{W_R} = \int {\cal D} \Lambda \, e^{ - \frac{1}{2} \int d^2 x \, d^2 y \, \Lambda^T(\vec {x}) {\widehat K}_{m_{\mathrm{ph}}}(\vec {x} - \vec {y}) \Lambda(\vec {y})}\,,
\label{eq:Z:lambda:mon:1}
\end{eqnarray}
where the operator ${\widehat K}_{m_{\mathrm{ph}}}$ is a massive analogue of the ${\widehat K}$ operator in Eq.~\eq{eq:K}:
\begin{eqnarray}
{\widehat K}_{m_{\mathrm {ph}}}(\vec {x}) & = & \left(\frac{\partial^2}{\partial x_2^2} + \frac{\partial^2}{\partial x_3^2}\right)
\label{eq:K:m:ph} \\
& & \left(
\begin{array}{ll}
D_{m_{\mathrm {ph}}}(0, x_2, x_3) & D_{m_{\mathrm {ph}}}(- R, x_2, x_3) \\
D_{m_{\mathrm {ph}}}(+ R, x_2, x_3) & D_{m_{\mathrm {ph}}}(0, x_2, x_3)
\end{array}
\right). \nonumber
\end{eqnarray}
The function ,
\begin{eqnarray}
D_{m_{{\mathrm {ph}}}}({\boldsymbol x}) = \int \frac{d^3 k}{(2 \pi)^3} \frac{e^{i {\boldsymbol k} \boldsymbol {x}}}{k^2+m_{\mathrm {ph}}^2}\,.
\label{eq:D:mph}
\end{eqnarray}
is a Green function of a massive scalar field:
\begin{eqnarray}
(- \Delta + m_{\mathrm {ph}}^2) D_{m_{\mathrm {ph}}}(\boldsymbol {x}) = \delta(\boldsymbol {x})\,.
\end{eqnarray}
The photon mass $m_{\mathrm {ph}}$ is given in Eqs.~\eq{eq:m:ph} and \eq{eq:m:ph:rho}.
Following all steps made of previous section we arrive to the following expression for the Casimir energy density in the presence of monopoles:
\begin{eqnarray}
V^{\mathrm {mon}}_{\mathrm{Cas}}(R,m_{\mathrm {ph}}) & = & \frac{1}{2} \int \frac{d^2 q}{(2\pi)^2} \log \left(1 - e^{- 2 \sqrt{\vec {q}^{\;2} + m_{\mathrm {ph}}^2} R}\right) \nonumber\\
& = & - \frac{\zeta(3)}{16 \pi} \frac{1}{R^2} \, f_{\mathrm {mon}} \left(m_{\mathrm {ph}} R\right) \,.
\label{eq:V:Cas:R:massive}
\end{eqnarray}
where the function
\begin{eqnarray}
f_{\mathrm {mon}}(x) = - \frac{2 x^2}{\zeta(3)} \int_0^\infty d y \, \log \left( 1 - e^{- 2 x \sqrt{y+1}} \right)\,.
\label{eq:f:x}
\end{eqnarray}
is shown in Fig.~\ref{fig:f:x}.
\begin{figure}[!thb]
\begin{center}
\vskip 5mm
\includegraphics[scale=0.5,clip=true]{f.png}
\end{center}
\vskip -5mm
\caption{The function $f_{\mathrm {mon}}$ in Eq.~\eq{eq:f:x}.}
\label{fig:f:x}
\end{figure}
The monopole density $\varrho_{\mathrm {mon}}$ enters the Casimir energy via the photon mass~\eq{eq:m:ph:rho} in terms of the function
\begin{eqnarray}
f_{\mathrm {mon}}\left(m_{\mathrm {ph}} R\right) \equiv \frac{V_{\mathrm{Cas}}(R,m_{\mathrm {ph}})}{V_{\mathrm{Cas}}(R,0)}\,,
\end{eqnarray}
where $V_{\mathrm{Cas}}(R) \equiv V_{\mathrm{Cas}}(R,0)$ is the Casimir energy density in the absence of the monopoles~\eq{eq:V:Cas:R}.
At small distances between the wires (or, equivalently, at small monopole density, $R \, m_{\mathrm {ph}} \ll 1$) the function $f_{\mathrm {mon}}$ is close to unity, $f_{\mathrm {mon}}(x) = 1 + O(x^2)$. Therefore in this case the Casimir energy~\eq{eq:V:Cas:R:massive} is close to the Casimir energy in the standard non-compact electrodynamics where the monopoles are absent~\eq{eq:V:Cas:R}. However the function $f_{\mathrm {mon}}$ decays exponentially at large value of its argument,
\begin{eqnarray}
f_{\mathrm {mon}}(x) = \frac{2 x}{\zeta(3)} e^{- 2 x} + \dots \,, \qquad x \gg 1\,,
\label{eq:f:mon:expansion}
\end{eqnarray}
indicating that at large monopole densities and/or at large wire separations, $R m_{\mathrm {ph}} \gg 1$, the Casimir energy density should be exponentially suppressed.
Thus we come to the conclusion that in the dilute gas approximation the presence of the dynamical monopoles leads to suppression of the Casimir effect at large separation between the wires. This nonperturbative effect effect does not come totally unexpected in view of the mass gap generation in the Coulomb gas of monopoles.
In the next Section we study the Casimir interaction numerically in the lattice formulation of the theory. The numerical approach allows us to explore certain unexpected monopole effects that are beyond the dilute gas approximation~\eq{eq:applicability}.
\section{Compact lattice electrodynamics}
\label{sec:compact:lattice:QED}
\subsection{Action, photons and monopoles}
The lattice version of the action~\eq{eq:L} of the compact electrodynamics in three space-time dimensions is given by the sum over elementary plaquettes $P $ of the lattice:
\begin{eqnarray}
S[\theta] = \beta \sum_P \left(1 - \cos \theta_P \right)\,,
\label{eq:S}
\end{eqnarray}
A plaquette $P \equiv P_{x,\mu\nu}$ is determined by a position~$x$ of one of its corners and its orientation given by two orthogonal vectors $\mu < \nu$ in the plaquette plane with $\mu,\nu = 1, 2, 3$. The plaquette angle
\begin{eqnarray}
\theta_{P_{x,\mu\nu}} = \theta_{x,\mu} + \theta_{x+\hat\mu,\nu} - \theta_{x+\hat\nu,\mu} - \theta_{x,\nu}\,,
\label{eq:theta:P}
\end{eqnarray}
is constructed from elementary angles
\begin{eqnarray}
\theta_{x,\mu} \in [-\pi,+\pi)\,,
\label{eq:angles:lattice}
\end{eqnarray}
which belong to the links of the lattice starting at the point $x$ and pointing towards the direction $\mu$. The lattice coupling constant
\begin{eqnarray}
\beta = \frac{1}{g^2 a}\,,
\label{eq:beta:3D}
\end{eqnarray}
is determined by the length of the elementary lattice link (lattice spacing) $a$ and the electric charge $g$. The dimensionality of the continuum coupling $g$ in $3$ dimensional spacetime $[g] = {\text{mass}}^{1/2}$, so that the lattice coupling $\beta$ in Eq.~\eq{eq:beta:3D} is a dimensionless quantity. The relation~\eq{eq:beta:3D} is valid in the regime of the weak coupling $g$ (large $\beta$) where the expansion of the action in terms of small field fluctuations, and consequently the comparison with the continuum action~\eq{eq:L}, is possible.
The angular variable $\theta_{x\mu}$ is a lattice version of the continuum Abelian gauge field $A_\mu$, $\theta_{x\mu} = a A_{\mu}(x)$. In the continuum limit the lattice spacing approaches zero, $a \to 0$, and the plaquette variable~\eq{eq:theta:P} reduces, for finite values of the gauge field $A_\mu$, to the continuum field strength tensor~\eq{eq:F:munu}:
\begin{eqnarray}
\theta_{P_{x,\mu\nu}} = a^2 F_{\mu\nu}(x) + O(a^4)\,.
\label{eq:theta:P:continuum}
\end{eqnarray}
Substituting this equation into the lattice action~\eq{eq:S}, expanding over the powers of the lattice spacing $a$ and keeping the leading term only, we get the action of the continuum U(1) gauge theory~\eq{eq:S:continuum}.
The partition function of the model,
\begin{eqnarray}
{\cal Z} = \prod_{l} \int_{-\pi}^\pi d \theta_l \, e^{-S[\theta]}\,,
\label{eq:Z:lat}
\end{eqnarray}
includes integration over all angular link variables~\eq{eq:angles:lattice}. The action~\eq{eq:S} is invariant under the $2 \pi$ shifts of the plaquette variable,
\begin{eqnarray}
\theta_P \to \theta_P + 2 \pi n\,, \qquad \quad n \in {\mathbb Z}\,,
\label{eq:theta:P:shifts}
\end{eqnarray}
indicating that the lattice field strengths $\theta_P$ and $\theta_P' = \theta_P + 2 \pi$ that are different by a $2 \pi$ shift~\eq{eq:theta:P:shifts} are physically equivalent indicating that the Abelian gauge group is a compact manifold (hence the name, ``compact electrodynamics'' or ``compact gauge theory''). In the continuum limit~\eq{eq:theta:P:continuum} these $2 \pi$ shifts become singular functions proportional to $2 \pi/a^2$ where $a \to 0$ is a vanishing lattice spacing. These shifts corresponds to the Dirac sheets which are world-lines of the Dirac strings attached to the Abelian monopoles. The positions of the Dirac strings are gauge-dependent so that the Dirac strings themselves are not gauge invariant objects. However the ends of the Dirac strings, monopoles, are physical, gauge-invariant topological defects. Thus, the compactness of the model leads to singular configurations which appear as Abelian monopoles and lead to the decomposition of the continuum field-strength tensor to the regular (photon) and singular (monopole) parts~\eq{eq:F:munu}. A comprehensive review on compact (gauge) fields and topological defects can be found in Ref.~\cite{ref:book:Kleinert}.
In the three-dimensional compact electrodynamics the monopoles are pointlike (instanton-like) objects. In the continuum limit their density is given by Eq.~\eq{eq:rho}. On the lattice, the monopoles are living on three-dimensional cubes $C_x$. Their local magnetic charge density
\begin{eqnarray}
\rho_x = \frac{1}{2\pi} \sum_{P \partial C_x} {\bar \theta}_P\,,
\label{eq:rho:lattice}
\end{eqnarray}
is nothing but a divergence of the physical part of the lattice field-strength tensor~\eq{eq:theta:P}
\begin{eqnarray}
{\bar \theta}_P = \theta_P + 2 \pi k_P \in [-\pi,\pi), \qquad k_P \in {\mathbb Z},
\label{eq:bar:theta}
\end{eqnarray}
where the integer number $k_P$ is chosen in such a way that the plaquette angle ${\bar \theta}_P$ belongs to the canonical interval. One can show that the lattice monopole density~\eq{eq:rho:lattice} is an integer number, $\rho_x \in {\mathbb Z}$. In fact, Eq.~\eq{eq:rho:lattice} is a divergence of a curl field which is zero for the regular (noncompact) gauge fields and non-zero for singular (compact) fields. The monopoles were studied intensively both in Abelian and non-Abelian lattice gauge theories~\cite{Chernodub:1997ay}.
\subsection{Casimir boundary conditions on the lattice}
The Casimir boundary conditions on the lattice were formulated in various dimensions for Abelian and non-Abelian gauge theories in Ref.~\cite{Chernodub:2016owp}. Here we briefly mention the results relevant to our context.
In (2+1) dimensions the Casimir boundary conditions force a tangential component of the electric field to vanish at each wire~\eq{eq:F:0:cov}. In the lattice gauge theory this boundary condition corresponds to the vanishing of the field strength tensor~\eq{eq:theta:P} -- up to the discrete compact transformations~\eq{eq:theta:P:shifts} -- at the set of the plaquettes $P \in {\cal P}_{\cal S}$ that belongs to the world-surfaces of the wires.
We consider two static straight wires directed along the $x_2$ axis and separated along the $x_1$ direction, $x_1 = l_1$ and $x_1 = l_2$, Fig.~\ref{fig:geometry:plane}. The $x_3$ axis is associated with the Euclidean ``time'' direction. In the case of an ideal metal, the corresponding boundary condition is given by the lattice version of Eq.~\eq{eq:F01:3d}:
\begin{eqnarray}
\cos\theta_{x,23} = 1\,,
\label{eq:F01:latt:3D}
\end{eqnarray}
where $x = (x_1,x_2,x_3)$ with fixed $x_1=l_1,l_2$ and all possible $x_2$ and $x_3$.
The simplest way to implement the boundary condition~\eq{eq:F01:latt:3D} is to add a set of Lagrange multipliers which will force the plaquettes belonging to the plane ${\cal P}_{\cal S}$ to vanish. To this end the standard $U(1)$ action~\eq{eq:S} can be changed as follows:
\begin{eqnarray}
S_{\varepsilon}[\theta;{\cal P}_{\cal S}] = \sum_P \beta_P(\varepsilon) \cos \theta_P\,,
\label{eq:S:beta}
\end{eqnarray}
where the plaquette-dependent gauge coupling is
\begin{eqnarray}
\beta_{P_{x,\mu\nu}} (\varepsilon) = \beta \bigl[1 + (\varepsilon - 1)\, & & (\delta_{\mu,2} \delta_{\nu,3} + \delta_{\mu,3} \delta_{\nu,2}) \nonumber \\
& & \cdot \left(\delta_{x,l_1} + \delta_{x,l_2}\right)\bigr]\,.
\label{eq:beta:P:3d}
\end{eqnarray}
is a function of the dielectric permittivity $\varepsilon$. At $\varepsilon = 1$ the wires are absent while in the limit $\varepsilon \to \infty$ the components of the physical lattice field-strength tensor~\eq{eq:bar:theta} vanish at the word-surfaces of the wires, ${\bar \theta}_{x,23} = 0$ as required by Eq.~\eq{eq:F01:latt:3D}.
The partition function~\eq{eq:Z:lat} of the model in the presence of the Casimir plates becomes as follows:
\begin{eqnarray}
{\cal Z}[{\cal P}_{\cal S}] & = & \lim_{\varepsilon \to + \infty} {\cal Z}_\varepsilon[{\cal P}_{\cal S}]\,,
\label{eq:Z:Casimir:1}
\\
{\cal Z}_\varepsilon[{\cal P}_{\cal S}] & = & \int {\cal D} \theta \, e^{-S_\varepsilon[\theta;{\cal P}_{\cal S}]}\,.
\label{eq:Z:Casimir:2}
\end{eqnarray}
In our simulations we realize the case of perfectly conducting wires by considering the limit of large permittivity $\varepsilon \to \infty$ in which a component of the electric field parallel to the wire vanishes~\eq{eq:F01:3d} thus mimicking an ideal metal. In two spatial dimensions the magnetic permeability does not exist and a wire with infinite static dielectric permittivity affects the electromagnetic field in the same way as an ideal metal (cf. Section 5.1 of Ref.~\cite{ref:Bogdag}).
Our formulation also allows us to consider the system at finite permittivity $\varepsilon$ described by the partition function~\eq{eq:Z:Casimir:2}.
\section{Casimir effect and monopoles: numerical simulation}
\label{sec:cQED:3D}
\subsection{Numerical setup}
Following our previous study~\cite{Chernodub:2016owp} we consider a symmetric $24^3$ cubic lattice which corresponds to a zero-temperature theory. We impose the periodic boundary conditions at the opposite sides of the lattice along all three directions. The two parallel static straight wires are located at the positions $x_1 = l_1$ and $x_1 = l_2$ separated by the distance $R = |l_2 - l_1|$, Fig.~\ref{fig:geometry:plane}. The wires are implemented via the space/orientation-dependent gauge coupling~\eq{eq:beta:P:3d} in the action of the theory~\eq{eq:S:beta}. The wires divide the $x_1$ axis into two, generally inequivalent, intervals, $R$ and $L_s - R$. Due to the periodic boundary conditions all calculated $R$-dependent quantities (potentials, densities, etc) should be invariant under the spatial flip in the $x_1$ direction, $R \to L_s - R$.
\vskip 2mm
\begin{table}[!ht]
\centering
\begin{tabular}{|l|c|}
\hline Trajectories per one value of $\varepsilon$ & $2 \times 10^5$ \\
\hline Trajectories for thermalization & $2 \times 10^4$\\
\hline Overrelaxation steps between trajectories & $5$\\
\hline Lattice size & $24^3$ \\
\hline Range of gauge coupling & $\beta = 1.0\, \sim\, 1.9$ \\
\hline Values of permittivity $\varepsilon$ per single value of $\beta$ & $\approx 33$ \\
\hline
\end{tabular}
\caption{Basic simulation parameters.}
\label{tabl:simparam}
\end{table}
\vskip 2mm
Our numerical tools are the same as in Ref.~\cite{Chernodub:2016owp}. We generate gauge-field configurations using a Hybrid Monte Carlo algorithm which combines the molecular dynamics approach with standard Monte-Carlo methods~\cite{ref:Gattringer}. The molecular-dynamics component utilizes a second-order minimum norm integrator~\cite{ref:Omelyan} with several time scales~\cite{ref:Sexton}. The use of different timescales allows us to equilibrate the integration errors accumulated at and outside worldsheets of the wires at which the Casimir boundary conditions are imposed. Long autocorrelation lengths in Markov chains are eliminated by overrelaxation steps which separate gauge field configurations far from each other~\cite{ref:Gattringer}. We use a self-tuning adaptive algorithm in order to control the acceptance rate of the Hybrid Monte-Carlo in a reasonable range between 0.70 and 0.85. The parameters of our simulations are presented in Table~\ref{tabl:simparam}. Other details of simulations may be found in Ref.~\cite{Chernodub:2016owp}.
\subsection{Monopoles in the absence of wires}
The monopole density falls off very quickly with increase of the lattice coupling $\beta$ (or, equivalently, with decrease of the coupling $g$ in the continuum limit). In our previous calculations~\cite{Chernodub:2016owp} we were working in the region of sufficiently large lattice gauge coupling $\beta \geqslant 3$ where the density of the monopoles was extremely small. In this region the theory becomes effectively noncompact since strong values of the lattice strength tensor $\theta_P \sim \pi + 2 \pi n$ (with $n \in {\mathbb Z}$) are practically unaccessible at so large coupling $\beta$. Therefore we have found no measurable monopole effects and we have proven reliability of our lattice methods by demonstrating reproducibility of well-known theoretical results. In the present we would like to study the effect of Abelian monopoles on the Casimir forces and therefore we choose the region of small lattice gauge coupling~$\beta$ which corresponds, according to Eq.~\eq{eq:beta:3D}, to the strong coupling region in term of the continuum electric charge~$g$.
In Fig.~\ref{fig:monopole:density} we show the monopole density $\rho_{\mathrm {mon}}^{\mathrm{lat}\,} = a^3 \rho_{\mathrm {mon}}$ (in lattice units) in the strong coupling region given by small values of $\beta$. Below we will be working in the region $\beta = 1 \sim 2$ where the monopole density is rather high and the monopole effects are expected to be rather strong. For comparison, at $\beta=3$ (the smallest coupling of our previous study~\cite{Chernodub:2016owp}) the density of monopoles in lattice units is $\rho = 1.8(1)\times 10^{-5}$ which means that on average we have less than one monopole at the whole $24^3$ lattice. The monopole effects were therefore very small at $\beta \geqslant 3$.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{mon_dens_no_plates_applicability.pdf}
\end{center}
\vskip -5mm
\caption{The density of lattice monopoles (in lattice units) vs the lattice coupling $\beta$. The inset demonstrates the inapplicability of the
dilute gas approximation~\eq{eq:applicability}.}
\label{fig:monopole:density}
\end{figure}
The inset in Fig.~\ref{fig:monopole:density} shows the quantity $\varrho^{1/2}_{\mathrm {mon}} g^3_m$ which is expected to be small in the dilute gas approximation~\eq{eq:applicability}. However, in the strong-coupling region this quantity is rather large indicating that the dilute gas approach should theoretically break down. Nevertheless we will see below that at relatively large $\beta$ (but still within the strong coupling region) certain features of the dilute monopole gas are manifestly visible.
\subsection{Effect of wires on monopoles}
\subsubsection{Monopole densities}
The Casimir effect and monopoles influence each other in both directions: the finite geometry imposed by the wires in the Casimir problem affects the dynamics of the monopoles while the presence of the dynamical monopoles modify the Casimir forces between the wires. In this section we discuss the former question and then in the next section we will address the latter.
Firstly, let us make a comment on our terminology. The wires are one-dimensional objects in two spatial dimensions. Their world-surfaces are flat planes that are parallel to each other. The monopoles, in turn, are instanton-like objects, the positions of which are marked by the points in three-dimensional spacetime. Therefore it is more suitable to discuss the monopoles positioned in between the plates (the mentioned flat planes) rather than monopoles in between the wires. Below, we will alternatively speak about wires and plates (the latter ones are the world-surfaces of the wires themselves).
Secondly, let us discuss what could be an expected effect of the plates on the monopoles in between them? The plates affect the electromagnetic flux emanating from the volume confined in between the plates. In the perfect-metal limit, $\varepsilon \to \infty$, the plates should make the electromagnetic flux going into or out of the volume between the plate to vanish. According to the Gauss law the total magnetic charge density of monopoles should therefore be zero. Moreover, the closely-spaced plates should squeeze the spherical $3D$ configuration of magnetic flux around monopoles which should form a $2D$ configuration which is definitely more energetically costly. Therefore we expect that the overall effect of the plates is likely to suppress the density of monopoles and antimonopoles in between the plates. One can also presume that the stronger permittivity $\varepsilon$ the stronger is the suppression of the monopole density. Finally, the larger distance between the plates $R$ the smaller the effect of the monopole suppression is expected be.
\begin{figure*}[!thb]
\begin{center}
\vskip 3mm
\begin{tabular}{cc}
\includegraphics[scale=0.5,clip=true]{mon_dens_beta_1_0.pdf} & \hskip 5mm
\includegraphics[scale=0.5,clip=true]{mon_dens_beta_1_5.pdf}
\end{tabular}
\end{center}
\vskip -5mm
\caption{Density of monopoles $\rho_{\mathrm {mon}}$ in between the plates vs. permittivity $\varepsilon$ for various values of the lattice gauge coupling $\beta=1.0,1.1,1.2,1.3$ and distances between the plates $R = 1, 2, \dots, 8$ (in lattice units). The lines are the fits by the function~\eq{eq:fit:function}.}
\label{fig:monopole:casimir}
\end{figure*}
All mentioned features are well seen in Fig.~\ref{fig:monopole:casimir} which shows the monopole density $\rho_{\mathrm {mon}}$ between the plates vs. permittivity $\varepsilon$ for various values of the lattice coupling $\beta$ and plate separation $R$. The closer the plates (wires) the smaller is the monopole density in between the wires. Increase of the permittivity results in diminishing of the monopole density. Qualitatively, all these properties are universal as they are largely independent of the lattice gauge coupling~$\beta$. The effect of the monopole suppression is especially strong at the smallest separation between the plates (wires), $R = a$.
At the closest separation $R=1a$, the world-surfaces of the wires touch the two sides of a monopole in between the wires, Fig.~\ref{fig:monopole:squeezed}. At these sides, marked by ``$B$'' in the Figure, the magnetic flux is vanishing due to the lattice version~\eq{eq:F01:latt:3D} of the Casimir boundary condition~\eq{eq:F01:3d}. As the result, the magnetic flux of the monopole may only penetrate the into the subspace in between the plates via other four sides of the lattice monopole that are marked by ``$A$'' in Fig.~\ref{fig:monopole:squeezed}. The dynamics of the monopoles become, essentially, two-dimensional.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.45,clip=true]{plates3d_monopole_labels.png}
\end{center}
\vskip -5mm
\caption{Illustration of the monopole squeezed in between the plates (worldsheets of the wires) at the minimal distance $R=a$ between the wires.}
\label{fig:monopole:squeezed}
\end{figure}
According to Fig.~\ref{fig:monopole:casimir} the monopole density is a rather smooth function of the permittivity $\varepsilon$ at fixed coupling $\beta$ and inter-wire distance $R$. As we will see below, the $\varepsilon$ dependence of all quantities, including the monopole density, can be described with a very good accuracy by the following heuristic function:
\begin{eqnarray}
{\cal O}^{\mathrm{fit}}(\varepsilon) = {\cal O}_\infty + \frac{a_1}{\varepsilon + b_1} + \frac{a_2}{\varepsilon^2 + b_2}\,,
\label{eq:fit:function}
\end{eqnarray}
where the fitting parameters $a_{1,2}$ and $b_{1,2}$ control the slope of the $\varepsilon$ dependence while $ {\cal O}_\infty$ corresponds to the value of the quantity ${\cal O}$ in the limit $\varepsilon \to \infty$. Using Eq.~\eq{eq:fit:function} we extrapolate the monopole density (in this case ${\cal O} \equiv \rho_{\mathrm{mon}}$) in between the plates to the perfect metal limit $\varepsilon \to \infty$. The result is shown in Fig.~\ref{fig:mon:dens:inside}.
Notice that in Ref.~\cite{Chernodub:2016owp} in a related investigation of the Casimir effect at weak coupling it was found that various expectation values can be excellently described the function~\eq{eq:fit:function} with three fitting parameters ${\cal O}_\infty$, $a_1$ and $b_1$ while the second term in Eq.~\eq{eq:fit:function} may be put to zero with $a_2 = b_2 = 0$. At the strong coupling studied in the present paper we need both terms in Eq.~\eq{eq:fit:function} to successfully describe the data.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{mon_dens_inside.pdf}
\end{center}
\vskip -5mm
\caption{Monopole density $\rho_{\mathrm{mon}}$ in between the wires as the function of the inter-wire distance $R$ (both are shown in the lattice units) extrapolated to the perfect metal limit for a set of fixed lattice gauge couplings $\beta$. The arrows show the asymptotic values $R \to \infty$ for each $\beta$. The latter are given by the monopole density in the absence of the wires (Fig.~\ref{fig:monopole:density}).}
\label{fig:mon:dens:inside}
\end{figure}
We see from Fig.~\ref{fig:mon:dens:inside} that in the perfect-metal limit $\varepsilon \to \infty$ the monopole density at the closest separation between the wires is zero. As the inter-wire distance $R$ increases the monopole density increases as well. The steepest slope of increase is achieved at stronger coupling $g$ [at smaller lattice coupling $\beta$, Eq.~\eq{eq:beta:3D}]. The latter fact is natural since in the absence of the wires the monopole medium is denser at stronger coupling $g$.
\subsubsection{Dimensional reduction and monopole transition}
What happens to the monopoles in between the plates? Surely, the approaching plates make the monopole density smaller due to squeezing of the magnetic flux and making monopoles heavier. This leads, as we discussed, to the overall suppression of the monopole density in between the plates. However, the very same effect of the restriction of the total magnetic flux of the monopoles from the three-dimensional space to the two-dimensional subspace, Fig.~\ref{fig:monopole:squeezed}, also affects the interactions between the monopoles and antimonopoles. In particular, the three dimensional attracting $1/r$ Coulomb potential~\eq{eq:D} transforms into a two dimensional logarithmic potential:
\begin{eqnarray}
D_{3D}({\boldsymbol x}) = \frac{1}{4 \pi | \boldsymbol {x} |} \to D_{2D}({\boldsymbol x}) = \frac{2}{R} \log \frac{| \boldsymbol {x} |}{R} \,.
\label{eq:dim:red}
\end{eqnarray}
The dimensional reduction should have a strong effect on the monopole dynamics since the rapidly decreasing potential becomes the slowly rising logarithmic function~\eq{eq:dim:red}. Moreover, the logarithmic function represents a confining potential and therefore we expect that at sufficiently large values of the permittivity $\varepsilon$ the individual monopoles should be confined into the magnetic dipole pairs that consist of closely spaced monopoles and anti-monopoles. This transition is indeed seen in our numerical simulations: in Figure~\ref{fig:monopoles} we visualize typical monopole configurations in between and outside the wires for a large value of permittivity $\varepsilon$. At small values of $\varepsilon$ the monopoles in between the wires resemble a gas of monopoles rather than the gas of magnetically neutral dipoles. Thus the increasing permittivity leads to a transition of the monopole gas into the dipole gas in the space between the wires.
The transition from monopole gas to the magnetic dipole gas should lead to the absence of the mass-gap generation and to a confinement-deconfinement transition in the region between the plates. As we discuss later, the absence of the mass gap should give rise to a certain enhancement of the zero-point potential between the wires. Moreover, one could expect on general grounds that the physical picture should be similar to the finite-temperature deconfining transition in three-dimensional compact electrodynamics~\cite{Chernodub:2001ws,Chernodub:2000uv} . However, in our numerical simulations we found no conclusive evidence of the suggested phase transition from the confinement to deconfinement phases. In order to study the deconfinement we have calculated the vacuum expectation value of the Polyakov loop
\begin{eqnarray}
L(x_1,x_2) = \exp\left\{i \sum_{x_3=0}^{L_t -1} \theta_{3}(x_1,x_2,x_3)\right\}\,,
\label{eq:Polyakov:Loop}
\end{eqnarray}
inside and outside the plates. This quantity is the order parameter of confinement: it is zero in the confinement phase and nonzero otherwise.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\begin{tabular}{cc}
\includegraphics[scale=0.3,clip=true]{monconf1.pdf} &
\includegraphics[scale=0.3,clip=true]{monconf2.pdf} \\[3mm]
\includegraphics[scale=0.3,clip=true]{monconf3.pdf} &
\includegraphics[scale=0.3,clip=true]{monconf4.pdf}
\end{tabular}
\end{center}
\vskip -2mm
\caption{Examples of two configurations of monopoles (the red dots) and anti-monopoles (the blue dots) in the space between the wires (the upper plots) and outside the wires (the lower plots) at strong coupling regime ($\beta = 1$) and at large permittivity ($\varepsilon = 59$).}
\label{fig:monopoles}
\end{figure}
We found that at zero temperature the Polyakov loop is consistent with a zero value inside and outside plates. The absolute values of the bulk sum of the Polyakov loop~\eq{eq:Polyakov:Loop} over all the points inside and, separately, outside of the plates do not show the existence of the phase transition at all studied values of parameters (coupling constants $\beta$, permittivity $\varepsilon$ and interwire separations $R$). Most plausibly this negative result comes due to the fact that the Polyakov loop is not a convenient variable to study confinement of charges at zero temperature. We are planning to readdress this question at finite-temperature theory in our next study.
\subsection{Monopoles, energy and pressure}
\subsubsection{Stress energy tensor on the lattice}
The zero-point energy of the quantum field fluctuations is affected by the vicinity of the wires. Since the zero-point (Casimir) energy depends on the interwire distance $R$, the modification of the vacuum fluctuations leads to appearance of a force applied to the wires. This phenomenon is known as the Casimir effect.
The Casimir effect can be calculated by a direct evaluation of the energy density of vacuum fluctuations which is given by is the ``temporal-temporal'' component $T^{00}$ of the Lorentz-covariant energy-momentum (stress\--energy) tensor of the gauge field~\eq{eq:S:continuum},
\begin{eqnarray}
T^{\mu\nu} = - \frac{1}{g^2}F^{\mu\alpha} F^\nu_{\ \alpha} + \frac{1}{4 g^2} \eta^{\mu\nu} F_{\alpha\beta} F^{\alpha\beta}.
\label{eq:T:munu}
\end{eqnarray}
In Minkowski spacetime the diagonal components of the energy-momentum tensor~\eq{eq:T:munu} of the gauge field are as follows:
\begin{subequations}
\begin{eqnarray}
T^{00} & = & \frac{1}{2 g^2} \left(E_x^2 + E_y^2 + B_z^2\right),
\label{eq:T00:M} \\
T^{11} & = & \frac{1}{2 g^2} \left(- E_x^2 + E_y^2 + B_z^2\right),
\label{eq:T11:M} \\
T^{22} & = & \frac{1}{2 g^2} \left(E_x^2 - E_y^2 + B_z^2\right).
\label{eq:T22:M}
\end{eqnarray}
\label{eq:T:munu:M}
\end{subequations}
\!\!In the spacetime with the metric $(+,-,-)$ the components of the field-strength tensor~\eq{eq:F:munu} are $F_{01} = E_x$, $F_{02} = E_y$ and $F_{12} = - B_z$. The former two ones are the components of the electric fields acting along $x$ and $y$ axis, while the remaining component $B_z$ has a formal sense of the magnetic field although there is no $z$ axis in the (2+1) dimensional Minkowski spacetime.
In Euclidean spacetime the components~\eq{eq:T:munu:M} of the energy-momentum tensor are as follows:
\begin{subequations}
\begin{eqnarray}
T^{00}_E & = & \frac{1}{2 g^2} \left( - E_x^2 - E_y^2 + B_z^2\right),
\label{eq:T00:E} \\
T^{11}_E & = & \frac{1}{2 g^2} \left( - E_x^2 + E_y^2 - B_z^2\right),
\label{eq:T11:E} \\
T^{22}_E & = & \frac{1}{2 g^2} \left( E_x^2 - E_y^2 - B_z^2\right),
\label{eq:T22:E}
\end{eqnarray}
\label{eq:T:munu:E}
\end{subequations}
\!\!\!\!where we took into account that as we pass from Minkowski to Euclidean spacetime the terms with electric field in Eq.~\eq{eq:T:munu:M} change their signs, $E_x^2 \to - E_x^2$ and $E_y^2 \to - E_y^2$, while the magnetic field remains intact, $B_z^2 \to B_z^2$. Moreover, $T^{00}\to T^{00}_E$ while $T^{11}\to -T^{11}_E$ and $T^{22}\to - T^{22}_E$ .
At zero temperature the system is invariant under discrete rotations by the angle $\pm \pi/2$ around the $x \equiv x_1$ axis. Since after these rotations the fields $E_x$ and $B_z$ are interchanged, one has $E_x \leftrightarrow \pm B_z$, and, consequently,
\begin{eqnarray}
\avr{E_x^2} = \avr{B_z^2}\,.
\label{eq:invariance}
\end{eqnarray}
Therefore we conclude from Eqs.~\eq{eq:T:munu:E} and \eq{eq:invariance} that
\begin{eqnarray}
\avr{T^{00}_E} & = & \avr{T^{22}_E} = - \frac{1}{2 g^2} \avr{E_y^2}\,,
\label{eq:T00:T22:E}\\
\avr{T^{11}_E} & = & \frac{1}{2 g^2} \bigl(\avr{E_y^2} - 2 \avr{E_x^2} \bigr)\,.
\label{eq:T11:E:2}
\end{eqnarray}
In Minkowski spacetime the temporal component $T^{00}$ of the strength-energy tensor has the sense of the energy density while the spatial components $T^{11}$ and $T^{22}$ are the stresses that are usually associated with the (local) pressure. Notice that in the presence of the closely spaced wires the directions $x \equiv x_1$ and $y \equiv x_2$ are not equivalent, and therefore the local stresses in these directions may be different. Making the Wick rotation to the Euclidean spacetime, we get that the local (stress) pressure along the wires is equal to the energy density~\eq{eq:T00:T22:E} while it is not equal to the stress in the transversal direction with respect to the wires~\eq{eq:T11:E:2}.
In the ultraviolet limit the expectation value of the energy density~\eq{eq:T00:T22:E} is a divergent quantity both in the presence and in the absence of the wires. The difference between these expectation values is the physical, ultraviolet-finite quantity which has a sense of an excess in the energy density of quantum fluctuations due to the presence wires. Therefore the physical (measurable) effect of the wires can be quantified in terms of the normalized energy density:
\begin{eqnarray}
{\cal E}_R(x) &
= & \avr{T^{00}_E(x)}_{R} - \avr{T^{00}_E(x)}_{0},
\label{eq:E:norm}
\end{eqnarray}
where the subscripts ``$R$'' and ``0'' indicate that these expectation values are taken, respectively, in the presence of the wires separated by the distance $R$ and in the absence of the wires.
Since the wires are parallel to each other the energy density~\eq{eq:E:norm} depends only on the coordinate $x_1$, which is transversal to the wires themselves. Therefore it is natural to introduce the total (Casimir) energy density with respect to the unit length of the wires:
\begin{eqnarray}
V_{{\mathrm{Cas}\,}}(R) = \int\limits_{-\infty}^{+\infty} d x_1\, {\cal E}_R(x_1) \equiv - \frac{1}{2 g^2} \aavr{E_y^2}\,,
\label{eq:V:Cas}
\end{eqnarray}
where we used Eq.~\eq{eq:T00:T22:E} and, following Ref.~\cite{Chernodub:2016owp}, introduced the average
\begin{eqnarray}
\avr{\!\avr{{\cal O}(x)}\!} = \int d x_1 \left[ \avr{{\cal O}(x)}_R - \avr{{\cal O}}_0 \right]\,,
\label{eq:avr:O}
\end{eqnarray}
which corresponds to the excess of the expectation value of the quantity ${\cal O}$ evaluated per unit length of the wires. In the lattice regularization the average~\eq{eq:avr:O} has the following form:
\begin{eqnarray}
\avr{\!\avr{{\cal O}(x)}\!}_{{\mathrm{lat}\,}} =\sum_{x_1 = 0}^{L_s - 1} \left[ \avr{{\cal O}(x_1)}_R - \avr{{\cal O}}_0 \right]\,.
\label{eq:avr:O:lat}
\end{eqnarray}
Thus, the lattice Casimir energy density per unit length of the wires~\eq{eq:V:Cas} takes the following compact form:
\begin{eqnarray}
V_{{\mathrm{Cas}\,}}(R) = \beta \aavr{\cos \theta_{23}},
\label{eq:V:Cas:lat}
\end{eqnarray}
where the lattice gauge coupling $\beta$ is given in Eq.~\eq{eq:beta:3D}. Here we naturally redefined the Casimir energy in the lattice units, $V_{{\mathrm{Cas}\,}} \to a^2 V_{{\mathrm{Cas}\,}}$.
For our purposes it is also convenient to consider an unnormalized version of the Casimir potential
\begin{eqnarray}
V^w_{{\mathrm{Cas}\,}}(R) = \beta \avr{\cos \theta_{23}}_R.
\label{eq:V:Cas:lat:unnorm}
\end{eqnarray}
Similarly to the energy density~\eq{eq:V:Cas:lat}, one can also define the (cummulative) transversal pressure of the quantum fluctuations,
\begin{eqnarray}
P_{x}(R) = \beta \bigl(2 \aavr{\cos \theta_{13}} - \aavr{\cos \theta_{23}} \bigr),
\label{eq:P:Cas:lat}
\end{eqnarray}
which is associated with the $T^{11}$ component of the energy-momentum tensor~\eq{eq:T11:E:2}. The quantity~\eq{eq:P:Cas:lat} is the energy-momentum stress in the $x_1$ direction integrated over the transverse spatial direction in a manner of Eq.~\eq{eq:V:Cas}.
It is important to notice that the pressure experienced by the wires due to the zero-point fluctuations is not given by the integrated value of the $T^{11}$ stress~\eq{eq:P:Cas:lat}. Naturally, the pressure experienced by each wire is equal to the difference of the values of the $T^{11}$ component of the energy-momentum tensor at inner and outer sides of the wire:
\begin{eqnarray}
P_\pm = T^{11}(\pm R/2 + \epsilon) - T^{11}(\pm R/2 - \epsilon)\,,
\end{eqnarray}
where $\epsilon \to 0$ is a vanishing positive quantity. It turns out that~\cite{ref:Milton}
\begin{itemize}
\item[(i)] the pressure which is experienced by the left wire $P_-$ is opposite in sign with respect to the pressure $P_+$ which is felt by the right wire, $P_+ = - P_-$;
\item[(ii)] the force of attraction, that is pressure times length of each wire, $F = P L$, is equal to the minus derivative of the Casimir energy,
\begin{eqnarray}
F(R) = - E'_{\mathrm{Cas}\,}(R)\,.
\label{eq:F:R}
\end{eqnarray}
\end{itemize}
Summarizing, the energy density of the zero-point fluctuations~\eq{eq:T00:M} and their pressure (cumulative normal stress) in the direction normal to the wires~\eq{eq:T11:M}, integrated over the separation between the wires, are given by the lattice expressions in, respectively, Eq.~\eq{eq:V:Cas:lat} and Eq.~\eq{eq:P:Cas:lat}. Belo we study their properties numerically.
\subsubsection{Casimir energy and monopoles}
Our numerical data indicates that the Casimir potential is a smooth function of the wire permittivity $\varepsilon$. In Figure~\ref{fig:Vcas:extrapolation} we show the raw data unnormalized potential~\eq{eq:V:Cas:lat:unnorm} for the minimal separation between the wires, $R = 1a$ and for various values of the lattice coupling $\beta$. It is clearly seen that the increase in the permittivity $\varepsilon$ of the wires leads to the increase of the vacuum energy for all values of $\beta$. This property is quite natural since as at a higher value of $\varepsilon$ the wires are more ``visible'' to the vacuum fluctuations compared to the lower-$\varepsilon$ wires.
Following the general strategy of Ref.~\cite{Chernodub:2016owp} we extrapolate the potential to the ideal metal limit $\varepsilon \to \infty$ by fitting the finite-$\varepsilon$ data by the function~\eq{eq:fit:function}. Due to the nonlinear nature of the function~\eq{eq:fit:function} we fit separately both terms~\eq{eq:avr:O:lat} in the normalized potential~\eq{eq:V:Cas:lat}. The fits of the first term in Eq.~\eq{eq:avr:O:lat} corresponding to the unrenormalized potential~\eq{eq:V:Cas:lat:unnorm} are also shown in Fig.~\ref{fig:Vcas:extrapolation}.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{Vcas_extrapolation_R1.pdf}
\end{center}
\vskip -5mm
\caption{The unnormalized lattice Casimir potential~\eq{eq:V:Cas:lat:unnorm} as function of permittivity $\varepsilon$ at a set of lattice couplings $\beta$. The extrapolation to the perfect-metal limit, shown by the continuous lines, is done by the fitting function~\eq{eq:fit:function}. The thickness of the lines corresponds to the errors of extrapolation.}
\label{fig:Vcas:extrapolation}
\end{figure}
In Figure~\ref{fig:Vcas:R1:R2} we show the (unnormalized) Casimir potential~\eq{eq:V:Cas:lat:unnorm} as a function of the lattice coupling constant $\beta$ at two interwire separations. We see that the larger permittivity $\varepsilon$ the larger is the (unrenormalized) Casimir potential and the stronger is the effect of the wires on vacuum fluctuations. The unrenormalized Casimir potential is a non-monotonic function of the coupling constant $\beta$ with $\varepsilon$-dependent positions of maxima. As the permittivity $\varepsilon$ increases, the maximum of the unrenormalized vacuum energy shifts towards smaller values of $\beta$. In the $\varepsilon \to \infty$ limit, which is obtained with the help of function~\eq{eq:fit:function}, the maximum of the unnormalized Casimir energy is achieved at $\beta \simeq 1.2$. At approximately the same value of the coupling constant the physical density of the monopoles achieves its maximum as it is shown in the inset of Fig.~\ref{fig:monopole:density}.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{Vcas_vs_beta_R1.pdf}\\
(a) \\[5mm]
\includegraphics[scale=0.55,clip=true]{Vcas_vs_beta_R2.pdf}\\
(b)
\end{center}
\vskip -5mm
\caption{The unnormalized Casimir potential~\eq{eq:V:Cas:lat:unnorm} vs. the lattice coupling $\beta$ at interwire separations $R=1a$ (the upper plot) and $R=2a$ (the lower plot) and various values of permittivity $\varepsilon$.}
\label{fig:Vcas:R1:R2}
\end{figure}
The physical (normalized) Casimir potential is a monotonic function of the distance $R$. In Figure~\ref{fig:Vcas:lattice} we show the potential in the limit of infinite permittivity (at finite permittivity $\varepsilon$ the potential resembles the limiting case $\varepsilon \to 0$ albeit smaller values of the amplitude). For all studied values of the couplings and distances the potential is a negative quantity. From Fig.~\ref{fig:Vcas:lattice} we can deduce a few interesting features of the Casimir energy:
\begin{enumerate}
\item At a weaker gauge coupling, at the lattice coupling $\beta \simeq 2$, the potential has a moderate strength and it is relatively long-ranged in similarity with our previous results~\cite{Chernodub:2016owp}.
\item In the stronger gauge coupling regime, as $\beta$ decreases, the potential becomes more short-ranged while its strength at short distances increases.
\item As the coupling becomes even stronger and the lattice coupling approaches $\beta = 1$, the potential becomes even more short-ranged and the strength of the potential decreases.
\end{enumerate}
Most these numerical findings may be understood from our analytical results obtained in Sect.~\ref{sec:compact:QED:monopoles} in the dilute gas approximation to the dynamics of the monopoles. Indeed, the monopoles lead to the mass gap generation which results in the finite mass of photon~\eq{eq:m:ph:rho}. As the photon becomes massive, the Casimir effect becomes naturally smaller and its effective radius of interaction decreases, as the Casimir interaction between the wires gets an additional exponential factor according to Eqs.~\eq{eq:V:Cas:R:massive}, \eq{eq:f:x} and \eq{eq:f:mon:expansion}. Since the interaction is very short-ranged we were not able to fit the numerically obtained results by the theoretical formula given in Eqs.~\eq{eq:V:Cas:R:massive} and \eq{eq:f:x}.
In addition, our theoretical analysis indicates that the numerically observed enhancement of the Casimir potential at intermediate coupling cannot be explained by simple analytical calculations in the dilute gas approximation. However, one may strongly believe that the formation of the monopole-antimonopole pairs in between the plates -- discussed earlier and illustrated in Fig.~\ref{fig:monopoles} -- makes the Casimir potential long-ranged due to the absence of the mass gap. On the other the increased density of the monopoles outside the plates contributes to the pressure and increase the Casimir potential.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{Vcas_lattice_extended.pdf}
\end{center}
\vskip -5mm
\caption{The physical (normalized) Casimir potential~\eq{eq:V:Cas:lat} vs. the interwire distance $R$ at various values of the lattice coupling constant~$\beta$ (in lattice units).}
\label{fig:Vcas:lattice}
\end{figure}
We would like to notice that in the strong coupling regime there is no physical scaling of the Casimir potential contrary to the weak coupling regime considered in Ref.~\cite{Chernodub:2016owp}. This is the expected property because the lattice formulation of the compact electrodynamics represents inherently lattice gauge theory defined at the specific cutoff $a$. For example the relation~\eq{eq:beta:3D} between the physical gauge coupling $g$ and the lattice spacing $a$ is valid only in the weak coupling regime. We demonstrate the absence of the formal scaling in Fig.~\ref{fig:Vcas:scaling}. The results are shown in the limit $\varepsilon \to \infty$.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{Vcas_scaling_pure.pdf}
\end{center}
\vskip -5mm
\caption{Absence of scaling of the Casimir potential~\eq{eq:V:Cas:lat} in the strong coupling regime. All quantities are shown in physical units.}
\label{fig:Vcas:scaling}
\end{figure}
For the sake of completeness, we also show in Fig.~\ref{fig:T11:extrapolation} the normal stress~\eq{eq:P:Cas:lat} as the function of permittivity $\varepsilon$ (at $R = 1a$ the integrated normal stress is equal to the normal stress). The stress in our definition is the positive quantity which is a monotonically increasing function of~$\varepsilon$. It gets maximal values at $\beta \simeq 1.1$ which naturally, in view of Eq.~\eq{eq:F:R}, corresponds to a steepest slope of the Casimir potential shown Fig.~\ref{fig:Vcas:lattice}.
\begin{figure}[!thb]
\begin{center}
\vskip 3mm
\includegraphics[scale=0.55,clip=true]{T11_extrapolation_R1_lat.pdf}
\end{center}
\vskip -5mm
\caption{Normal stress~\eq{eq:P:Cas:lat} vs. permittivity~$\varepsilon$ at minimal distance between the wires, $R = 1 a$. The lines are drawn to guide the eye.}
\label{fig:T11:extrapolation}
\end{figure}
\section{Conclusions}
In our article we discussed nonperturbative features of the Casimir effect which arise due to the presence of dynamical topological defects. As an interesting example that can be treated both analytically and numerically, we considered the zero-point potential between two dielectric wires in a compact version of Abelian gauge theory in two spatial dimensions at zero temperature. The spectrum of this theory possesses topological defects, instanton-like monopoles, which are known to be responsible for a number of nonperturbative features in the model including the effect of mass gap generation.
We calculated the zero-point Casimir potential in an analytical approach based on a duality transformation: the model is first transformed to a dual representation where a perturbative treatment is possible, and then the Casimir potential is evaluated directly in the dual model. We have shown that the dynamical monopoles make the Casimir potential short-ranged, Eqs.~\eq{eq:V:Cas:R:massive} and \eq{eq:f:x}, with the radius of interaction given by the inverse mass of the photon~\eq{eq:m:ph:rho}. The Casimir potential becomes dependent both on the monopole density and on the gauge coupling of the model. The calculation is valid in the dilute gas approximation where the number of monopoles in a unit Debye volume is small.
Our numerical simulations complement the analytical calculations. Using the numerical method developed in Ref.~\cite{Chernodub:2016owp} we have shown that in the region of strong gauge coupling, where the monopole density is high, the Casimir potential is weak and it falls down very quickly with increase of the distance, Fig.~\ref{fig:Vcas:lattice}. Both these findings are consistent with the effect expected from a massive photon field (the photon becomes massive due to the nonperturbative mass generation caused by the presence of the dynamical monopoles). In the region of an intermediate gauge coupling the Casimir potential becomes enhanced by the monopoles. At a weak gauge coupling the density of monopoles becomes negligibly small and the potential reduces to the standard coupling-independent analytical result.
The behavior of zero-point potential at the intermediate coupling may be related to the effects of the wires on the monopole dynamics.
Firstly, as we argued analytically and then found numerically, the monopole density is suppressed in between the wires. The monopole suppression is especially strong in the limit of large permittivity $\varepsilon$ of the wires.
Secondly, the wires effectively squeeze the flux of the monopoles in between them thus making inter-monopole interactions two-dimensional. In two-dimensions the inter-monopole interaction is governed by a confining logarithmic Coulomb potential. Therefore the monopoles tend to form dilute monopole-antimonopole pairs (magnetic dipoles) which cannot generate the mass gap, and, consequently, cannot effectively inhibit the zero-point interactions between the wires. Thus, the formation of the magnetic dipoles in between the plates makes the Casimir potential long-ranged. Moreover, the density of the monopoles outside the plates is much larger compared to the density in between the plates. This difference in monopole densities increases the external pressure on the wires and enhances the Casimir potential. At stronger gauge coupling the mentioned mechanism does not work due to large density of monopoles because the intermonopole distance become smaller than the typical size of the monopole-antimonopole pair so that the magnetic dipole picture is no more applicable.
Summarizing, we have found that the dynamical topological defects modify nonperturbatively the zero-point Casimir interaction between dielectric bodies. In general, this finding may be relevant to a large class of effective infrared theories in condensed matter physics.
\acknowledgments
The work was supported by the Federal Target Programme for Research and Development in Priority Areas of Development of the Russian Scientific and Technological Complex for 2014-2020 (the unique identifier of the Contract RFMEFI58415X0017, contract number 14.584.21.0017). The numerical simulations were performed at the computing cluster Vostok-1 of Far Eastern Federal University.
|
\section{Algorithm \ensuremath{\mathsf{RSP}}\xspace}
\label{sec:algo}
This section is devoted to the presentation of our algorithm,
Algorithm \ensuremath{\mathsf{RSP}}\xspace (which stands for {\em Rooted Shortest-Path}). The
code of Algorithm \ensuremath{\mathsf{RSP}}\xspace is given in Algorithm~\ref{alg:A}.
\begin{algorithm}[htpb]
\footnotesize
\begin{center}
Macro of \ensuremath{\mathsf{RSP}}\xspace for any process $u$
\end{center}
\begin{tabular}{lll}
$\variable{children}(u)$ & $=$ & $\{v
\in \Gamma(u) \mid \variable{st}_u \neq I \wedge \variable{st}_v \neq I
\wedge \variable{par}_v = u \wedge \variable{d}_v \geq \variable{d}_u + \omega(v,u) \wedge
(\variable{st}_v = \variable{st}_{u} \vee \variable{st}_{u} = EB)\}$\\
\end{tabular}\\
\smallskip
\dotfill
\begin{center}
Code of \ensuremath{\mathsf{RSP}}\xspace for the root process $r$
\end{center}
\noindent \textbf{Constants:}\\
\begin{tabular}{lll}
$\variable{st}_r$ & $=$ & $C$\\
$\variable{par}_r$ & $=$ & $\perp$\\
$\variable{d}_r$ & $=$ & $0$\\
\end{tabular}\\
\smallskip
\dotfill
\begin{center}
Code of \ensuremath{\mathsf{RSP}}\xspace for any process~$u \neq r$
\end{center}
\noindent \textbf{Variables:}\\
\begin{tabular}{lll}
$\variable{st}_u$ & $\in$ & $\{I,C,EB,EF\}$\\
$\variable{par}_u$ & $\in$ & $Lbl$\\
$\variable{d}_u$ & $\in$ & $\mathds R^+$
\end{tabular}\\
\noindent \textbf{Predicates:} \\
\begin{tabular}{lll}
$abRoot(u)$ & $\equiv$ & $\variable{st}_u \neq I \wedge \big[ \variable{par}_u
\notin \Gamma(u) \vee \variable{st}_{\variable{par}_u} = I \vee \variable{d}_u <
\variable{d}_{\variable{par}_u} + \omega(u,\variable{par}_u) \vee$\\
&& $(\variable{st}_u \neq
\variable{st}_{\variable{par}_u} \wedge \variable{st}_{\variable{par}_u} \neq EB) \big]$\\
$\ensuremath{P\_reset}\xspace(u)$ & $\equiv$ & $\variable{st}_u = EF \wedge abRoot(u)$ \\
$\ensuremath{P\_correction}\xspace(u)$ & $\equiv$ &
$(\exists v \in \Gamma(u) \mid \variable{st}_v = C
\wedge \variable{d}_v+ \omega(u,v) < \variable{d}_u)$
\end{tabular}\\
\noindent \textbf{Macro:} \\
\begin{tabular}{lll}
$computePath(u)$ & : & $\variable{par}_u := \argmin_{(v \in \Gamma(u)
\And \variable{st}_v = C)}(\variable{d}_v + \omega(u,v))$; \\
&& $\variable{d}_u := \variable{d}_{\variable{par}_u} + \omega(u, \variable{par}_u)$;\\
&& $\variable{st}_u := C$
\end{tabular}\\
{\textbf{Rules}} \\
\begin{tabular}{lllll}
$\mathbf{R_C}(u)$ & : & $\variable{st}_u = C \wedge \ensuremath{P\_correction}\xspace(u)$ & $\to$ & $computePath(u)$\\
$\mathbf{R_{EB}}(u)$ & : & $\variable{st}_u = C \wedge \neg\ensuremath{P\_correction}\xspace(u) \wedge$ & $\to$ & $\variable{st}_u := EB$ \\
& & \qquad ($abRoot(u)$ $\vee$ $\variable{st}_{\variable{par}_u}=EB$) & &\\
$\mathbf{R_{EF}}(u)$ & : & $\variable{st}_u = EB \wedge (\forall v \in \variable{children}(u) \mid \variable{st}_v = EF)$ & $\to$ & $\variable{st}_u := EF$\\
$\mathbf{R_I}(u)$ & : & $\ensuremath{P\_reset}\xspace(u) \And (\forall v \in \Gamma(u) \mid \variable{st}_v \neq C)$ & $\to$ & $\variable{st}_u := I$\\
$\mathbf{R_R}(u)$ & : & $(\ensuremath{P\_reset}\xspace(u) \vee \variable{st}_u = I) \And (\exists v \in \Gamma(u) \mid \variable{st}_v = C)$ & $\to$ & $computePath(u)$
\end{tabular}
\caption{Code of \ensuremath{\mathsf{RSP}}\xspace
\label{alg:A}}
\end{algorithm}
\subsection{Variables}
In \ensuremath{\mathsf{RSP}}\xspace, each process $u$ maintains three variables: $\variable{st}_u$,
$\variable{par}_u$, and $\variable{d}_u$.
Those three variables are constant for the root process\footnote{We
should emphasize that the use of constants at the root is not a
limitation, rather it allows to simplify the design and proof of the
algorithm. Indeed, these constants can be removed by adding a rule
to correct all root's variables, if necessary, within a single
step.}, $r$: $\variable{st}_r = C$, $\variable{par}_r = \perp$\footnote{$\perp$
is a particular value which is different from any value in $Lbl$.},
and $\variable{d}_r = 0$.
For each non-root process $u$, we have:
\begin{itemize}
\item $\variable{st}_u \in \{I,C,EB,EF\}$, this variable gives the {\em
status} of the process. $I$, $C$, $EB$, and $EF$ respectively
stand for {\em Isolated}, {\em Correct}, {\em Error Broadcast}, and
{\em Error Feedback}. The two first states, $I$ and $C$, are
involved in the normal behavior of the algorithm, while the two last
ones, $EB$ and $EF$, are used during the correction
mechanism. Precisely, $\variable{st}_u = C$ (resp. $\variable{st}_u = I)$ means
that $u$ believes it is in $V_r$ (resp. not in $V_r$). The meaning
of status $EB$ and $EF$ will be further detailed in
Subsection~\ref{err:corr}.
\item $\variable{par}_u \in Lbl$, a {\em parent pointer}. If $u \in V_r$,
$\variable{par}_u$ should designate a neighbor of $u$, referred to as its
{\em parent}, and in a terminal configuration, the parent pointers
exhibit a shortest path from $u$ to $r$.
Otherwise ($u \notin V_r$), the variable is meaningless.
\item $\variable{d}_u \in \mathds{R}^+$, the {\em distance} value. If $u \in V_r$, then
in a terminal configuration, $\variable{d}_u$ gives the weight of the
shortest path from $u$ to $r$.
Otherwise ($u \notin V_r$), the variable is meaningless.
\end{itemize}
\subsection{Normal Execution}
Consider any configuration, where every process $u \neq r$ satisfies
$\variable{st}_u = I$, and refer to such a configuration as a {\em normal
initial configuration}. Each configuration reachable from a {\em
normal initial configuration} is called a {\em normal
configuration}, otherwise it is an {\em abnormal
configuration}. Recall that $\variable{st}_r = C$ in all
configurations. Then, starting from a normal initial configuration,
all processes in a connected component different from $V_r$ are
disabled forever. Focus now on the connected component $V_r$. Each
neighbor $u$ of $r$ is enabled to execute $\mathbf{R_R}(u)$. A process
eventually chooses $r$ as parent by executing this rule, which in
particular sets its status to $C$. Then, executions of rule
$\mathbf{R_R}$ are asynchronously propagated in $V_r$ until all its
processes have status $C$: when a process $u$ with status $I$ finds
one of its neighbor with status $C$ it executes $\mathbf{R_R}(u)$, {\em i.e.} $u$
takes status $C$ and chooses as parent its neighbor $v$ with status
$C$ such that $\variable{d}_v + \omega(u,v)$ is minimum, $\variable{d}_u$ being updated
accordingly. In parallel, rules $\mathbf{R_C}$ are executed to reduce
the weight of the tree rooted at $r$: when a process $u$ with status
$C$ can reduce $\variable{d}_u$ by selecting another neighbor with status $C$
as parent, it chooses the one allowing to minimize $\variable{d}_u$ by
executing $\mathbf{R_C}(u)$. Hence, eventually, the system reaches a
terminal configuration, where the tree rooted at $r$ is a
shortest-path tree spanning all processes of $V_r$.
\subsection{Error Correction}\label{err:corr}
Assume now that the system is in an abnormal
configuration. Thanks to the predicate $abRoot$, some
non-root processes locally detect that their state is inconsistent
with that of their neighbors. We call {\em abnormal roots} such
processes. Informally (see Algorithm \ensuremath{\mathsf{RSP}}\xspace for the formal
definition), a process~$u\neq r$ is an {\em abnormal root} if~$u$
is not isolated ({\em i.e.}, $\variable{st}_u \neq I$) and satisfies one of
the following four conditions:
\begin{enumerate}
\item its parent pointer does not designate a neighbor,
\item its parent has status~$I$,
\item its distance value~$\variable{d}_u$ is inconsistent with
the distance value of its parent, or
\item its status is inconsistent with
the status of its parent.
\end{enumerate}
Every non-root process $u$ that is not an abnormal root satisfies one
of the two following cases. Either $u$ is {\em isolated}, {\em i.e.},
$\variable{st}_u = I$, or $u$ points to some neighbor ({\em i.e.},
$\variable{par}_u \in \Gamma(u)$) and the state of $u$ is coherent {\em
w.r.t.} the state of its parent. In this latter case, $u \in
\variable{children}(\variable{par}_u)$, {\em i.e.}, $u$ is a ``real'' child of its
parent (see Algorithm \ensuremath{\mathsf{RSP}}\xspace for the formal definition). Consider a path
$\mathcal P = u_1, \ldots, u_k$ (with $k \geq 1$) such that $u_1$ is either $r$ or an
abnormal root, and $\forall i, 1 \leq i < k, u_{i+1} \in
\variable{children}(u_i)$. $\mathcal P$ is acyclic and called a {\em branch}
rooted at $u_1$. Hence, we define the normal tree $T(r)$ (resp. an
abnormal tree $T(v)$, for any abnormal root $v$) as the set of all
processes that belong to a branch rooted at $r$ (resp. $v$).
Then, the goal is to remove all abnormal trees so that the system
recovers a normal configuration. For each abnormal tree $T$, we have
two cases. In the former case, the abnormal root $u$ of $T$ can join
another tree $T'$ using rule $\mathbf{R_C}(u)$, making $T$ a
subtree of $T'$. In the latter case, $T$ is entirely removed in a
top-down manner starting from its (abnormal) root $u$. Now, in that
case, we have to prevent the following situation: $u$ leaves $T$; this
removal creates some trees, each of those is rooted at a previous
child of $u$; and later $u$ joins one of those (created) trees.
Hence, the idea is to freeze $T$, before removing it. By freezing we
mean assigning each member of the tree to a particular state, here
$EF$, so that (1) no member $v$ of the tree is allowed to execute
$\mathbf{R_C}(v)$, and (2) no process $w$ can join the tree by
executing $\mathbf{R_R}(w)$. Once frozen, the tree can be safely
deleted from its root to its leaves.
The freezing mechanism (inspired from~\cite{BlinCV03}) is achieved
using the status $EB$ and $EF$, and the rules $\mathbf{R_{EB}}$
and $\mathbf{R_{EF}}$. If a process is not involved into any
freezing operation, then its status is $I$ or $C$. Otherwise, it has
status $EB$ or $EF$ and no neighbor can select it as its parent. These
two latter states are actually used to perform a ``Propagation of
Information with Feedback'' \cite{C82j,S83j} in the abnormal
trees. This is why status $EB$ means ``Error Broadcast'' and $EF$
means ``Error Feedback''. From an abnormal root, the status $EB$ is
broadcast down in the tree using rule $\mathbf{R_{EB}}$. Then,
once the $EB$ wave reaches a leaf, the leaf initiates a convergecast
$EF$-wave using rule $\mathbf{R_{EF}}$. Once the $EF$-wave
reaches the abnormal root, the tree is said to be {\em dead}, meaning
that all processes in the tree have status $EF$ and, consequently, no
other process can join it. So, the tree can be safely deleted from its
abnormal root toward its leaves. There is two possibilities for the
deletion. If the process $u$ to be deleted has a neighbor with status
$C$, then it executes rule $\mathbf{R_R}(u)$ to directly join another
``alive'' tree. Otherwise, $u$ becomes isolated by executing rule
$\mathbf{R_I}(u)$, and $u$ may join another tree later.
Let $u$ be a process belonging to an abnormal tree of which it is not
the root. Let $v$ be its parent. From the previous explanation, it
follows that during the correction, $(\variable{st}_v,\variable{st}_u) \in
\{(C,C),$ $(EB,C),$ $(EB,EB),$ $(EB,EF),$ $(EF,EF)\}$ until $v$ resets
by $\mathbf{R_R}(v)$ or $\mathbf{R_I}(v)$. Now, due to the arbitrary
initialization, the status of $u$ and $v$ may not be coherent, in this
case $u$ should also be an abnormal root. Precisely, as formally
defined in Algorithm~\ref{alg:A}, the status of $u$ is incoherent {\em
w.r.t} the status of its parent $v$ if $\variable{st}_u \neq \variable{st}_v$
and $\variable{st}_v \neq EB$.
Actually, the freezing mechanism ensures that if a process is the root
of an abnormal alive tree, it is in that situation since the initial
configuration (see Lemma~\ref{lem:pseudoRoots}, page
\pageref{lem:pseudoRoots}). The polynomial step complexity mainly
relies on this strong property.
\subsection{Example}
An example of synchronous execution of \ensuremath{\mathsf{RSP}}\xspace is given in
Figure~\ref{fig:ex}. We consider the network topology given on the
top left of the figure. The names $v_1, \ldots, v_{10}$ are only given
to ease the explanation (recall that we consider semi-anonymous
networks where only the root $r$ is distinguished). The network
contains eleven processes divided into two connected components. Let $v_i$
be a process. In the synchronous execution described from configuration
a) to configuration m), the color of $v_i$ indicates its status
$\variable{st}_{v_i}$, according to the legend on the top of the figure. The
number next to $v_i$ gives its distance value, $\variable{d}_{v_i}$. If there
is an arrow outgoing from $v_i$, this arrow designates the neighbor
$u$ of $v_i$ pointed as parent, {\em i.e.}, $\variable{par}_{v_i} =
u$. Otherwise, this means that $\variable{par}_{v_i} \notin \Gamma(v_i)$.
In the initial configuration a), there are two abnormal roots: $v_2$
and $v_9$, indeed $\variable{par}_{v_2} \notin \Gamma(v_2)$ and
$\variable{par}_{v_9} \notin \Gamma(v_9)$. The status of $v_2$ is already
equal to $EB$ and this value should be broadcast down in its
subtree. In contrast, $v_9$ has status $C$ and, consequently, should
initiate the broadcast of $EB$. Note also that $v_{10}$ can reduce its
distance value by modifying its parent pointer. Hence, in the step
a) $\mapsto$ b), $v_9$ takes status $EB$ (rule $\mathbf{R_{EB}}(v_9)$),
$v_{10}$ selects $v_9$ as parent (rule $\mathbf{R_C}(v_{10})$), and
finally $v_3$ the unique child of $v_2$ takes status $EB$ (rule
$\mathbf{R_{EB}}(v_3)$).
In the step b) $\mapsto$ c), $EB$ is propagated down the two abnormal
trees: $v_5$, $v_7$, $v_8$, and $v_{10}$ execute $\mathbf{R_{EB}}$.
In configuration c), the value $EB$ has reached three leaves: $v_5$,
$v_7$, and $v_{10}$. These processes are then enabled to initiate a
convergecast $EF$-wave. Hence, in the step c) $\mapsto$ d),
$v_5$, $v_7$, and $v_{10}$ execute $\mathbf{R_{EF}}$, while the last
leaf $v_4$ takes status $EB$ ($\mathbf{R_{EB}}(v_4)$).
In configuration d), all children of $v_9$ have status $EF$, so $v_9$
is enabled to take status $EF$ too ($\mathbf{R_{EF}}(v_9)$). In
contrast, $v_3$ should wait until its child $v_8$ takes status $EF$.
Hence, in the step d) $\mapsto$ e), $v_9$ takes status $EF$
($\mathbf{R_{EF}}(v_9)$), its abnormal tree becomes frozen, while the
last leaf $v_4$ of the second abnormal tree initiates a convergecast
$EF$-wave (rule $\mathbf{R_{EF}}(v_4)$).
In the step e) $\mapsto$ f), $v_9$ leaves its tree and becomes
isolated by rule $\mathbf{R_I}(v_9)$, while $v_8$ takes status $EF$ by
$\mathbf{R_{EF}}(v_8)$.
Since all its children have now status $EF$, $v_3$ can take status $EF$
by $\mathbf{R_{EF}}(v_3)$ in step f) $\mapsto$ g), while $v5$ and $v_{10}$
become isolated by rule $\mathbf{R_I}$ in the same step.
Remark then that in g), all processes in the connected component $\{v_5,
v_9,v_{10}\}$ are isolated and, since $r$ is not part of this
component, they are disabled forever. In the step g) $\mapsto$ h), the
abnormal root $v_2$ of the remaining abnormal tree takes status $EF$
($\mathbf{R_{EF}}(v_2)$). So, the abnormal tree rooted at $v_2$ is
frozen in configuration h). In the step h) $\mapsto$ i), $v_2$ leaves
its tree and becomes isolated by rule $\mathbf{R_I}(v_2)$. Then, $v_3$
becomes isolated in step i) $\mapsto$ j) (rule
$\mathbf{R_I}(v_3)$). In step j) $\mapsto$ k), $v_8$ becomes isolated
(rule $\mathbf{R_I}(v_8)$), while $v_7$ joins the normal tree (the
tree rooted at $r$) by rule $\mathbf{R_R}(v_7)$. In the last two
steps, $v_2$, $v_3$, $v_8$, and then $v_4$ successively join the
normal tree by rule $\mathbf{R_R}$, and configuration m) is terminal.
\begin{figure}
\centering
\includegraphics[scale=0.72]{IPE-execution-example.pdf}
\caption{A synchronous execution of \ensuremath{\mathsf{RSP}}\xspace\label{fig:ex}}
\end{figure}
\section{Conclusion}\label{ccl}
In this paper, we have proposed a silent self-stabilizing
algorithm for the DCDSPM problem. This algorithm is written in the
composite atomicity model, assuming a distributed unfair daemon (the
weakest scheduling assumption of the model).
Its stabilization time in rounds is at most~$3{n_{\texttt{maxCC}}}+D$, where ${n_{\texttt{maxCC}}}$
is the maximum number of non-root processes in a connected component
and $D$ is the hop-diameter of $V_r$.
Furthermore, if we additionally assume that edge weights are positive
integers, then it stabilizes in a polynomial number of steps: namely,
we exhibit a bound in $O(\texttt{W}_{\max} {n_{\texttt{maxCC}}}^3 n)$, where $\texttt{W}_{\max}$ is the
maximum weight of an edge and $n$ is the number of processes.
To obtain this stabilization time polynomial in steps, the key idea
was to freeze the growth of abnormal trees before removing them in a
top-down manner. This freezing mechanism is implemented as a
propagation of information with feedback in the tree. This technique
is general. In particular, it can be used in other spanning tree or
forest constructions.
The stabilization time is, by definition, evaluated from an arbitrary
initial configuration, and so is drastically impacted by worst case
scenarios. Now, in many cases, transient faults are sparse and their
effect may be superficial. For example, a topological change in a
network commonly consists of a single link failure. Some
specializations of self-stabilization, such as
superstabilization~\cite{DH97j}, self-stabilization with service
guarantee~\cite{JM14}, or gradual stabilization~\cite{ADDP16c} have
been proposed to target recovery from such favorable cases as a
performance issue.
Proposing silent algorithms for the DCDSPM problem implementing one
of these aforementioned stronger properties, while achieving polynomial step
complexity, is an interesting perspective of our work.
\section{Introduction}
Given a connected undirected edge-weighted graph $G$, a {\em
shortest-path (spanning) tree rooted at node $r$} is a spanning tree
$T$ of $G$, such that for every node $u$, the unique path from $u$ to
$r$ in $T$ is a shortest path from $u$ to $r$ in $G$. This data
structure finds applications in the networking area ({\em n.b.}, in
this context, nodes actually represent processes), since many
distance-vector routing protocols, like {\em RIP} ({\em Routing
Information Protocol}) and {\em BGP} ({\em Border Gateway
Protocol}), are based on the construction of shortest-path
trees. Indeed, such algorithms implicitly builds a shortest-path tree
rooted at each destination.
From time to time, the network may be split into several connected
components due to the network dynamics. In this case, routing to
process $r$ correctly operates only for the processes of its connected
component, $V_r$. Consequently, in other connected components,
information to reach $r$ should be removed to gain space in routing
tables, and to discard messages destined to $r$ (which are unable to
reach $r$ anyway) and thus save bandwidth.
The goal is then to make the network converging to a configuration
where every process of~$V_r$ knows a shortest path to~$r$ and every
other process detects that~$r$ is not in its own connected
component. We call this problem the {\em Disconnected Components
Detection and rooted Shortest-Path tree Maintenance}
(DCDSPM) problem. Notice that a solution to this problem allows to prevent the
well-known {\em count-to-infinity} problem \cite{LgW04}, where the
distances to some unreachable process keep growing in routing tables
because no process is able to detect the issue.
When topological changes are infrequent, they can be considered as
transient faults~\cite{T01} and self-stabilization~\cite{D74j} --- a
versatile technique to withstand \textit{any} finite number of
transient faults in a distributed system --- becomes an attractive
approach. A self-stabilizing algorithm is able to recover
without external ({\em e.g.}, human) intervention a correct behavior
in finite time, regardless of the \emph{arbitrary} initial
configuration of the system, and therefore, also after the occurrence
of transient faults, provided that these faults do not alter the code
of the processes.
A particular class of self-stabilizing algorithms is that of silent
algorithms. A self-stabilizing algorithm is {\em
silent}~\cite{DolevGS96} if it converges to a global state where the
values of communication registers used by the algorithm remain fixed.
Silent (self-stabilizing) algorithms are usually proposed to build
distributed data structures, and so are well-suited for the problem
considered here.
As quoted in~\cite{DolevGS96}, the silent property usually implies
more simplicity in the algorithm design, moreover a silent algorithm
may utilize less communication operations and communication bandwidth.
For the sake of simplicity, we consider here a single destination process
$r$, called the {\em root}. However, the solution we will propose can
be generalized to work with any number of destinations, provided that
destinations can be distinguished. In this context, we do not require
the network to be fully identified. Rather, $r$ should be
distinguished among other processes, and all non-root processes are
supposed to be identical: we consider semi-anonymous networks.
In this paper, we propose a silent self-stabilizing algorithm, called
Algorithm \ensuremath{\mathsf{RSP}}\xspace, for the DCDSPM problem with a single destination process
in semi-anonymous networks. Algorithm \ensuremath{\mathsf{RSP}}\xspace does not require any {\em a
priori} knowledge of processes about global parameters of the
network, such as its size or its diameter. Algorithm \ensuremath{\mathsf{RSP}}\xspace is written in
the locally shared memory model with composite atomicity introduced by
Dijkstra~\cite{D74j}, which is the most commonly used model in
self-stabilization. In this model, executions proceed in (atomic)
steps, and a self-stabilizing algorithm is silent if and only if all
its executions are finite. Moreover, the asynchrony of the system is
captured by the notion of {\em daemon}. The weakest ({\em i.e.}, the
most general) daemon is the {\em distributed unfair daemon}. Hence,
solutions stabilizing under such an assumption are highly desirable,
because they work under any other daemon
assumption. Interestingly, self-stabilizing algorithms designed
under this assumption are easier to compose (composition techniques
are widely used to design and prove complex self-stabilizing
algorithms). Moreover, time complexity (the {\em stabilization
time}, mainly) can be bounded in terms of steps only if the
algorithm works under an unfair daemon. Otherwise ({\em e.g.}, under a
weakly fair daemon), time complexity can only be evaluated in terms of
rounds, which capture the execution time according to the slowest
process. There are many self-stabilizing algorithms proven under the
distributed unfair daemon, {\em
e.g.},~\cite{ACDDP14,CDDLR15,DLP11,DLV11,GHIJ14}. However, analyses
of the stabilization time in steps is rather unusual and this may be
an important issue. Indeed, this complexity captures the amount
of computations an algorithm needs to recover a correct behavior.
Now, recently, several self-stabilizing algorithms, which work
under a distributed unfair daemon, have been shown to have an
exponential stabilization time in steps in the worst
case. In~\cite{ACDDP14}, silent leader election algorithms
from~\cite{DLP11,DLV11} are shown to be exponential in steps in the
worst case. In~\cite{DJ16}, the Breadth-First Search (BFS) algorithm
of Huang and Chen~\cite{HC92} is also shown to be exponential in
steps. Finally, in~\cite{GHIJ16} authors show that the first silent
self-stabilizing algorithm for the DCDSPM problem (still assuming a
single destination) they proposed in~\cite{GHIJ14} is also exponential
in steps.
\subsection{Contribution}
Algorithm \ensuremath{\mathsf{RSP}}\xspace proposed here is proven assuming the distributed unfair daemon.
We also study its stabilization time in rounds. We establish a
bound of at most~$3{n_{\texttt{maxCC}}}+D$ rounds, where ${n_{\texttt{maxCC}}}$ is the maximum number of
non-root processes in a connected component and $D$ is the hop-diameter of
$V_r$ (defined as the maximum over all pairs $\{u, v\}$ of nodes in
$V_r$ of the minimum number of edges in a shortest path from $u$ to
$v$).
Furthermore, \ensuremath{\mathsf{RSP}}\xspace is the first silent self-stabilizing algorithm for the
DCDSPM problem which, assuming that the edge weights are positive
integers, achieves a polynomial stabilization time in steps. Namely,
in this case, the stabilization time of \ensuremath{\mathsf{RSP}}\xspace is at most
$(\texttt{W}_{\max}{n_{\texttt{maxCC}}}^3+(3-\texttt{W}_{\max}){n_{\texttt{maxCC}}}+3)(n-1)$, where $\texttt{W}_{\max}$ is the maximum
weight of an edge and $n$ is the number of processes. ({\em N.b.},
this stabilization time is less than or equal to $\texttt{W}_{\max} n^4$, for all
$n \geq 3$.)
Finally, notice that when all weights are equal to one, the DCDSPM
problem reduces to a BFS tree maintenance and the step complexity
becomes at most $({n_{\texttt{maxCC}}}^3 +2{n_{\texttt{maxCC}}}+3)(n-1)$, which is less than or
equal to $n^4$ for all $n \geq 2$.
\subsection{Related Work}
To the best of our knowledge, only one self-stabilizing algorithm for
the DCDSPM problem has been previously proposed in the
literature~\cite{GHIJ14}. This algorithm is silent and works under the
distributed unfair daemon, but, as previously mentioned, it is
exponential in steps. However,
it has a slightly better
stabilization time in rounds, precisely at most $2({n_{\texttt{maxCC}}}+1)+D$
rounds\footnote{In fact, \cite{GHIJ14} announced $2n+D$ rounds, but it
is easy to see that this complexity can be reduced to
$2({n_{\texttt{maxCC}}}+1)+D$.}.
There are several shortest-path spanning tree algorithms in the
literature that do not consider the problem of disconnected components
detection. The oldest distributed algorithms are inspired by the
Bellman-Ford algorithm~\cite{bellman,ford}. Self-stabilizing
shortest-path spanning tree algorithms have then been proposed
in~\cite{CS94,HL02}, but these two algorithms are proven assuming a
central daemon, which only allows sequential executions. However, in
\cite{huang2005self}, Tetz Huang proves that these algorithms actually
work assuming the distributed unfair daemon. Nevertheless, no upper
bounds on the stabilization time (in rounds or steps) are given. More
recently, Cobb and Huang \cite{CH09} proposed an algorithm constructing
shortest-path trees based on any maximizable routing metrics. This
algorithm does not require a priori knowledge about the network but it
is proven only for the central weakly-fair daemon. It runs in a
linear number of rounds and no analysis is given on the number of steps.
Self-stabilizing shortest-path spanning tree algorithms are also given
in \cite{AGH90,CG02,JT03}. These algorithms additionally ensure the
loop-free property in the sense that they guarantee that a spanning
tree structure is always preserved while edge costs change
dynamically. However, none of these papers consider the unfair daemon, and
consequently their step complexity cannot be analyzed.
Whenever all edges have weight one, shortest-path trees correspond to
BFS trees. In~\cite{DDL12}, the authors introduce the {\em
disjunction} problem as follows. Each process has a constant input
bit, 0 or 1. Then, the problem consists for each process in computing
an output value equal to the disjunction of all input bits in the
network. Moreover, each process with input bit 1 (if any) should be
the root of a tree, and each other process should join the tree of the
closest input bit 1 process, if any. If there is no process with input
bit~1, the execution should terminate and all processes should
output~0. The proposed algorithm is silent and
self-stabilizing. Hence, if we set the input of a process to 1 if and
only if it is the root, then their algorithm solves the DCDSPM problem
when all edge-weights are equal to one, since any process which is not
in $V_r$ will compute an output 0, instead of 1 for the processes in
$V_r$. The authors show that their algorithm stabilizes in $O(n)$
rounds, but no step complexity analysis is given. Now, as their approach
is similar to \cite{DLV11}, it is not difficult to see that their
algorithm is also exponential in steps.
Several other self-stabilizing BFS tree algorithms have been proposed, but
without considering the problem of disconnected components detection.
Chen {\em et al.} present the first self-stabilizing BFS tree
construction in \cite{CYH91} under the central daemon. Huang and
Chen present the first self-stabilizing BFS tree construction in
\cite{HC92} under the distributed unfair daemon, but recall that this
algorithm has been proven to be exponential in steps
in~\cite{DJ16}. Finally, notice that these two latter
algorithms~\cite{CYH91,HC92} require that the processes know the exact
number of processes in the network.
According to our knowledge, only the following works
\cite{CDV09,CRV11} take interest in the computation of the number of
steps required by their BFS algorithms. The algorithm in \cite{CDV09}
is not silent and has a stabilization time in $O(\Delta n^3)$ steps,
where $\Delta$ is the maximum degree in the network. The silent
algorithm given in \cite{CRV11} has a stabilization time~$O(D^2)$
rounds and~$O(n^6)$ steps.
Silent self-stabilizing algorithms that construct spanning trees of
arbitrary topologies are
given in~\cite{Cournier09,KK05}. The solution proposed
in~\cite{Cournier09} stabilizes in at most $4n$ rounds and
$5n^2$ steps, while the algorithm given in~\cite{KK05}
stabilizes in $nD$ steps (its round complexity is not analyzed).
Several other papers propose self-stabilizing algorithms stabilizing
in both a polynomial number of rounds and a polynomial number of
steps, {\em e.g.},~\cite{ACDDP14} (for the leader
election),~\cite{CDPV06,CDV05} (for the DFS token circulation). The
silent leader election algorithm proposed in~\cite{ACDDP14} stabilizes
in at most $3n + D$ rounds and $O(n^3)$ steps. DFS token
circulations given in~\cite{CDPV06,CDV05} execute each wave in $O(n)$
rounds and $O(n^2)$ steps using $O(n\log n)$ space per process
for the former, and $O(n^3)$ rounds and $O(n^3)$ steps using $O(\log
n)$ space per process for the latter.
\subsection{Roadmap} In the next section, we present the computational
model and basic definitions. In Section~\ref{sec:algo}, we describe
Algorithm \ensuremath{\mathsf{RSP}}\xspace. Its proof of correctness and a complexity analysis in
steps are given in Section~\ref{sec:correctness}, whereas an analysis
of the stabilization time in rounds is proposed in
Section~\ref{sect:rounds}. Finally, we make concluding remarks in
Section~\ref{ccl}.
\section{Preliminaries}
We consider {\em distributed systems} made of $n \geq 1$ interconnected
processes. Each process can directly communicate with a subset of
other processes, called its {\em neighbors}. Communication is assumed
to be bidirectional. Hence, the topology of the system can be
represented as a simple undirected graph $G = (V,E)$, where $V$ is the
set of processes and $E$ the set of edges, representing communication
links. Every process $v$ can distinguish its neighbors using a {\em
local labeling} of a given datatype $Lbl$. All labels of $v$'s
neighbors are stored into the set $\Gamma(v)$. Moreover, we assume
that each process $v$ can identify its local label in the set
$\Gamma(u)$ of each neighbor $u$. Such labeling is called {\em
indirect naming} in the literature~\cite{Sloman:1987}. By an abuse
of notation, we use $v$ to designate both the process $v$ itself, and
its local labels.
Each edge~$\{u,v\}$ has a strictly positive
{\em weight},
denoted~by $\omega(u,v)$. This notion naturally extends to paths: the
weight of a path in $G$ is the sum of its edge weights. The weighted
distance between the processes~$u$ and~$v$, denoted by~$d(u,v)$, is the
minimum weight of a path from~$u$ to~$v$. Of course, $d(u,v) = \infty$
if and only if $u$ and $v$ belong to two distinct connected components
of $G$.
We use the {\em composite atomicity model of
computation}~\cite{D74j,D00b} in which the processes communicate
using a finite number of locally shared registers, called {\em
variables}. Each process can read its own variables and those of
its neighbors, but can write only to its own variables. The
\textit{state} of a process is defined by the values of its local
variables. A {\em configuration} of the system is a vector consisting
of the states of every process.
A \emph{distributed algorithm\/} consists of one local program per
process. We consider semi-uniform algorithms, meaning that all
processes except one, the {\em root} $r$, execute the same program. In
the following, for every process $u$, we denote by $V_u$ the set of
processes (including $u$) in the same connected component of $G$ as
$u$. In the following $V_u$ is simply referred to as the connected
component of $u$. We denote by ${n_{\texttt{maxCC}}}$ the maximum number of non-root
processes in a connected component of $G$. By definition, ${n_{\texttt{maxCC}}} \leq
n-1$.
The \textit{program} of each process consists of a finite set of
\textit{rules} of the form\ $label\ :\ guard \to\ action$. {\em
Labels} are only used to identify rules in the reasoning. A
\textit{guard} is a Boolean predicate involving the state of the
process and that of its neighbors. The {\em action\/} part of a rule
updates the state of the process. A rule can be executed only if its
guard evaluates to {\em true}; in this case, the rule is said to be
{\em enabled\/}. A process is said to be enabled if at least one of
its rules is enabled. We denote by $\mbox{\it Enabled}(\gamma)$ the
subset of processes that are enabled in configuration~$\gamma$.
When the configuration is $\gamma$ and
$\mbox{\it Enabled}(\gamma) \neq \emptyset$,
a non-empty set $\mathcal X \subseteq \mbox{\it Enabled}(\gamma)$ is selected; then
every process of $\mathcal X$ {\em atomically} executes one of its
enabled rules, leading to a new configuration $\gamma^\prime$, and so
on. The transition from $\gamma$ to $\gamma^\prime$ is called a {\em
step}. The possible steps induce a binary
relation over $\mathcal C$, denoted by $\mapsto$. An
{\em execution\/} is a maximal sequence of
configurations $e=\gamma_0\gamma_1\ldots \gamma_i\ldots$ such that
$\gamma_{i-1}\mapsto\gamma_i$ for all $i>0$. The term ``maximal''
means that the execution is either infinite, or ends at a {\em
terminal\/} configuration in which no rule is
enabled at any process.
Each step from a configuration to another is driven by a daemon. We
define a daemon as a predicate over executions. We said that an
execution $e$ is {\em an execution under the daemon $S$}, if $S(e)$
holds. In this paper we assume that the daemon is {\em distributed}
and {\em unfair}. ``Distributed'' means that while the configuration
is not terminal, the daemon should select at least one enabled
process, maybe more. ``Unfair'' means that there is no fairness
constraint, {\em i.e.}, the daemon might never select an enabled
process unless it is the only enabled process. In other words, the
distributed unfair daemon corresponds to the predicate $true$, {\em
i.e.}, this is the most general daemon.
In the composite atomicity model, an algorithm is {\em silent\/} if
all its possible executions are finite. Hence, we can define silent
self-stabilization as follows.
\begin{definition}
[Silent Self-Stabilization]
\label{def:selfstabilization}
Let~$\mathcal{L}$ be a non-empty subset of configurations, called set
of legitimate configurations. A distributed system is silent and
self-stabili\-zing under the daemon~$S$ for~$\mathcal{L}$ if and only
if the following two conditions hold:
\begin{itemize}
\item all executions under~$S$ are finite, and
\item all terminal configurations belong to~$\mathcal{L}$.
\end{itemize}
\end{definition}
We use the notion of \textit{round}~\cite{DIM93} to measure the time
complexity. The definition of round uses the concept of {\em
neutralization}: a process~$v$ is \textit{neutralized} during a
step~$\gamma_i \mapsto \gamma_{i+1}$, if~$v$ is enabled in~$\gamma_i$
but not in configuration~$\gamma_{i+1}$. Then, the rounds are
inductively defined as follows. The first round of an execution~$e =
\gamma_0, \gamma_1, \cdots$ is the minimal prefix~$e' = \gamma_0,
\cdots, \gamma_j$, such that every process that is enabled
in~$\gamma_0$ either executes a rule or is neutralized during a step
of~$e'$. Let~$e''$ be the suffix $\gamma_j, \gamma_{j+1}, \cdots$
of~$e$. The second round of~$e$ is the first round of~$e''$, and so
on.
The {\em stabilization time} of a silent self-stabilizing algorithm is
the maximum time (in steps or rounds) over every execution possible
under the considered daemon $S$ (starting from any initial
configuration) to reach a terminal (legitimate) configuration.
\section{Correctness and Step Complexity of Algorithm \ensuremath{\mathsf{RSP}}\xspace}
\label{sec:correctness}
\subsection{Definitions}\label{pred}
Before proceeding with the proof of correctness and the step
complexity analysis, we define some useful concepts.
\begin{definition}[Abnormal Root]
Every process $u \neq r$ that satisfies $abRoot(u)$ is said to
be an {\em abnormal root}.
\end{definition}
\begin{definition}[Alive Abnormal Root]
A process~$u \neq r$ is said to be an \emph{alive abnormal root}
(resp. a \emph{dead abnormal root}) if~$u$ is an abnormal root and
has a status different from~$EF$ (resp. has status $EF$).
\end{definition}
\begin{definition}[Branch]
A \emph{branch} is a
sequence of processes~$v_1, \cdots ,v_k$ for some integer~$k \geq
1$, such that~$v_1$ is~$r$ or an abnormal root and, for every~$1
\leq i < k$, we have~$v_{i+1} \in \variable{children}(v_{i})$.
The process~$v_i$ is said to be at \emph{depth}~$i$ and $v_i, \cdots
,v_k$ is called a {\em sub-branch}.
If~$v_1 \neq r$, the branch is said to be \emph{illegal}, otherwise,
the branch is said to be \emph{legal}.
\end{definition}
\begin{observation}\label{obs:size}
A branch depth is at most~${n_{\texttt{maxCC}}}$.
A process~$v$ having status~$I$ does not belong to any branch.
If a process~$v$ has status~$C$ (resp.~$EF$), then all processes of a
sub-branch starting at~$v$ have status~$C$ (resp.~$EF$).
\end{observation}
\begin{definition}[Legitimate State]\label{def:leg:state}
A process~$u$ is said to be in a \emph{legitimate state} if $u$
satisfies one of the following three conditions:
\begin{enumerate}
\item $u = r$,
\item\label{def:2} $u \neq r$, $u \in V_r$, $\variable{st}_u = C$,
$\variable{d}_u = d(u,r)$, and $\variable{d}_u =
\variable{d}_{\variable{par}_u}+\omega(u,\variable{par}_u)$, or
\item $u \notin V_r$ and $\variable{st}_u = I$.
\end{enumerate}
\end{definition}
\begin{observation}
\label{obs:0}
Every process $u \neq r$ such that $\variable{st}_u = C$ and $\variable{d}_u \neq
\variable{d}_{\variable{par}_u}+\omega(u,\variable{par}_u)$ is enabled.
\end{observation}
\begin{definition}[Legitimate Configuration]\label{def:leg}
A {\em legitimate configuration} is any configuration where every
process is in a legitimate state. We denote by $\mathcal{LC}_{\ensuremath{\mathsf{RSP}}\xspace}$
the set of all legitimate configurations of Algorithm \ensuremath{\mathsf{RSP}}\xspace.
\end{definition}
Let $\gamma$ be a configuration. Let $T_\gamma = (V_r,E_{T_\gamma})$
be the subgraph, where $E_{T_\gamma} = \{\{p,q\} \in E\ |\ p \in V_r
\setminus \{r\} \wedge \variable{par}_p = q\}$. By Definition~\ref{def:leg:state} (point~\ref{def:2}), we deduce the following
observation.
\begin{observation}
In every legitimate configuration $\gamma$, $T_\gamma$ is a
shortest-path tree spanning all processes of $V_r$.
\end{observation}
\subsection{Partial Correctness}
We now prove that the set of terminal configurations is exactly the
set of legitimate configurations. We start by proving the following
intermediate statement.
\begin{lemma}\label{lem:IorC}
In any terminal configuration, every process has either status~$I$ or~$C$.
\end{lemma}
\begin{proof}
This is trivially true for the root process,~$r$. Assume that there
exists a non-root process with status~$EB$ in a terminal
configuration~$\gamma$. Consider the non-root process~$u$ with
status~$EB$ having the largest distance value~$\variable{d}_u$ in~$\gamma$.
In~$\gamma$, no process~$v$ with status~$C$ can be a child
of $u$, otherwise either $\mathrm{R_{EB}}$ or~$\mathrm{R_{C}}$ is
enabled at $v$ in~$\gamma$, a contradiction. Moreover, by maximality
of $\variable{d}_u$,~$u$ cannot have a child with status~$EB$ in
$\gamma$. Therefore, in $\gamma$ process~$u$ has no child or it has
only children with status~$EF$, and thus rule~$\mathrm{R_{EF}}$ is
enabled at $u$, a contradiction. Thus, every process has
status~$C$,~$I$, or~$EF$ in $\gamma$.
Assume now that there exists a non-root process with status~$EF$ in a
terminal configuration~$\gamma$. Consider the process $u$ with
status~$EF$ having the smallest distance value~$\variable{d}_u$ in~$\gamma$.
By construction, $u$ is an abnormal root in $\gamma$.
So, either~$\mathrm{R_{I}}$ or~$\mathrm{R_R}$ is enabled at $u$ in
$\gamma$, a contradiction.
\end{proof}
The next lemma, Lemma~\ref{lem:leg1}, deals with the connected
components that do not contain $r$, if any. Then, Lemma~\ref{lem:leg2}
deals with the connected component~$V_r$.
\begin{lemma}\label{lem:leg1}
In any terminal configuration, every process that does not belong
to~$V_r$ is in a legitimate state.
\end{lemma}
\begin{proof}
Consider, by contradiction, that there exists a process $u$ that
belongs to the connected component $CC$ other than $V_r$ which is
not in a legitimate state in some terminal configuration~$\gamma$.
By definition, $u$ is not the root, moreover it has status
$C$ in $\gamma$, by Lemma~\ref{lem:IorC}.
So, consider the process $v$ of $CC$ with status~$C$ having
the smallest distance value~$\variable{d}_v$ in $\gamma$. By
construction,~$v$ is an abnormal root in $\gamma$. Thus,
rule~$\mathrm{R_{EB}}$ is enabled at $v$ in $\gamma$, a
contradiction.
\end{proof}
\begin{lemma}\label{lem:leg2}
In any terminal configuration, every process of~$V_r$ is in a
legitimate state.
\end{lemma}
\begin{proof}
Assume, by contradiction, that there exists a terminal
configuration $\gamma$ where at least one process in the connected
component~$V_r$ is not in a legitimate state.
Assume also that there exists some process of~$V_r$ that has
status~$I$ in $\gamma$. Consider now a process~$u$ of~$V_r$ such
that in $\gamma$, $u$ has status~$I$ and at least one of its
neighbors has status~$C$. Such a process exists because no process
has status~$EB$ or~$EF$ in $\gamma$ (Lemma~\ref{lem:IorC}), but at
least one process of $V_r$ has status~$C$, namely~$r$. Then,
$\mathrm{R_R}$ is enabled at~$u$ in $\gamma$, a contradiction. So,
every process in~$V_r$ must have status~$C$ in $\gamma$. Moreover,
for all processes in~$V_r$, we have~$\variable{d}_u = \variable{d}_{\variable{par}_u} +
\omega(\variable{par}_u,u)$ in $\gamma$, otherwise $\mathrm{R_{C}}$ is
enabled at some process of $V_r$ in $\gamma$.
Assume now that there exists a process~$u$ such that~$\variable{d}_u <
d(u,r)$ in $\gamma$. Consider a process~$u$ of~$V_r$ having the
smallest distance value $\variable{d}_u$ among the processes in~$V_r$ such
that~$\variable{d}_u < d(u,r)$ in $\gamma$. By definition, $u \neq r$ and
we have~$\variable{d}_u > \variable{d}_{\variable{par}_u}$ in $\gamma$, so
$\variable{d}_{\variable{par}_u} \geq d(\variable{par}_u,r)$ in $\gamma$. Hence, we can
conclude that $\variable{d}_u \geq d(u,r)$ in $\gamma$, a contradiction.
So, every process~$u$ in~$V_r$ satisfies~$\variable{d}_u \geq d(u,r)$ in
$\gamma$.
Finally, assume that there exists a process~$u$ such that~$\variable{d}_u >
d(u,r)$ in $\gamma$. Consider a process~$u$ in~$V_r$ having the
smallest distance to~$r$ among the processes in~$V_r$ such
that~$\variable{d}_u > d(u,r)$ in $\gamma$. By definition, $u \neq r$ and
there exists some process~$v$ in~$\Gamma(u)$ such that $d(u,r) =
d(v,r) + \omega(u,v)$ in $\gamma$. Thus, we have~$\variable{d}_v = d(v,r)$
in $\gamma$. So,~$\mathrm{R_C}$ is enabled at $u$ in $\gamma$, a
contradiction.
\end{proof}
After noticing that any legitimate configuration is a terminal one (by
construction of the algorithm), we deduce the following corollary from
the two previous lemmas.
\begin{corollary}\label{coro:legterm}
For every configuration $\gamma$, $\gamma$ is terminal if and only
if $\gamma$ is legitimate.
\end{corollary}
\subsection{Termination}
In this section, we establish that every execution of Algorithm~$\ensuremath{\mathsf{RSP}}\xspace$
under a distributed unfair daemon is finite. Furthermore, we
compute the following bound on the number of steps of every
execution:~$[\texttt{W}_{\max}{n_{\texttt{maxCC}}}^3+(3-\texttt{W}_{\max}){n_{\texttt{maxCC}}}+3](n-1)$, where $n$ is the
number of processes, $\texttt{W}_{\max}$ is the maximum weight of an edge, and
${n_{\texttt{maxCC}}}$ is the maximum number of non-root processes in a connected
component, when all weights are strictly positive integers.
\begin{lemma}\label{lem:pseudoRoots}
No alive abnormal root is created along any execution.
\end{lemma}
\begin{proof}
Let $\gamma \mapsto \gamma'$ be a step. Let~$u$ be a non-root process
that is not an \emph{alive abnormal root} in~$\gamma$, and let~$v$
be the process such that $\variable{par}_{u}=v$ in~$\gamma'$.
If the status of~$u$ is~$EF$ or~$I$ in~$\gamma'$, then~$u$ is not an
alive abnormal root in~$\gamma'$. So, let us assume now that the
status of~$u$ is either~$EB$ or~$C$ in~$\gamma'$.
Consider then the case where~$u$ has status~$EB$ in~$\gamma'$. The
only rule $u$ can execute in~$\gamma \mapsto \gamma'$
is~$\mathrm{R_{EB}}$. So, $\variable{st}_{u} \in \{C,EB\}$
in~$\gamma$. Moreover, whether $u$ executes $\mathrm{R_{EB}}$ or
not, $\variable{par}_u = v$ in~$\gamma$. Since $\variable{st}_{u} \in \{C,EB\}$
and $u$ is not an alive abnormal root in~$\gamma$, we can deduce
that $u$ is not an abnormal root in $\gamma$ (whether dead or alive). So, if $\variable{st}_{u} =
EB$ in~$\gamma$, then $\variable{st}_{v} = EB$ in~$\gamma$ too. Otherwise,
$u$ has status $C$ in $\gamma$ while not being an abnormal root in
$\gamma$: it executes $\mathrm{R_{EB}}(u)$ in~$\gamma \mapsto
\gamma'$ because $\variable{st}_{v} = EB$ in~$\gamma$. Hence, in either
case $v$ has status~$EB$ in~$\gamma$, and this in particular means
that $v \neq r$ (this status does not exist for $r$). Moreover, $u$
belongs to $\variable{children}(v)$ in~$\gamma$ (again because $\variable{par}_u = v$
and $u$ is not an abnormal root in~$\gamma$). So, $v$ is not
enabled in~$\gamma$ and $u \in \variable{children}(v)$ remains true in
$\gamma'$. Hence, we can conclude that~$u$ is still not an alive
abnormal root in~$\gamma'$.
Consider now the other case, {\em i.e.}, $u$ has status~$C$
in~$\gamma'$. During~$\gamma \mapsto \gamma'$, the only rules
that~$u$ may execute are~$\mathrm{R_R}$ or~$\mathrm{R_C}$.
%
If $u$ executes $\mathrm{R_R}$ or~$\mathrm{R_C}$, we have
$\variable{st}_{v}=C$ in~$\gamma$ (because it is a requirement to execute
any of these rules) and consequently, the only rules that~$v$ may
execute in~$\gamma \mapsto \gamma'$ are~$\mathrm{R_C}$
or~$\mathrm{R_{EB}}$.
%
Otherwise ({\em i.e.}, $u$ does not execute any rule in $\gamma
\mapsto \gamma'$), $\variable{par}_{u} = v$ and $\variable{st}_{u}=C$ already
hold in~$\gamma$. In this case, $u$ being not an alive abnormal root
and $\variable{st}_{u}=C$ in~$\gamma$ implies that $u \in \variable{children}(v)$ and
thus $\variable{st}_{v} \in \{C,EB\}$ in $\gamma$, which further implies
that the only rules that~$v$ may execute in~$\gamma \mapsto \gamma'$
in this case are~$\mathrm{R_C}$ or~$\mathrm{R_{EB}}$.
Thus, in either case, during~$\gamma \mapsto \gamma'$,~$v$ either
takes the status~$EB$, decreases its distance value, or does not
change the value of its variables. Consequently, $u$ belongs
to $\variable{children}(v)$ in~$\gamma'$, which prevents $u$ from being an
alive abnormal root in~$\gamma'$.
\end{proof}
Let $AAR(\gamma)$ be the set of alive abnormal roots in any
configuration $\gamma$. From the previous lemma, we know that, for
every step $\gamma \mapsto \gamma'$, we have $AAR(\gamma') \subseteq
AAR(\gamma)$ (precisely, for every process $u$ and every step $\gamma \mapsto
\gamma'$, $u \notin AAR(\gamma) \Rightarrow u \notin
AAR(\gamma')$). So, we can use the notion of \emph{$u$-segment}
(inspired from~\cite{ACDDP14}) to bound the total number of steps in
an execution.
\begin{definition}[$u$-Segment]
Let $u$ be any non-root process. Let $e = \gamma_0, \gamma_1,
\cdots$ be an execution.
If there is no step $\gamma_i \mapsto \gamma_{i+1}$ in $e$, where
there is a non-root process in $V_u$ which is an alive abnormal root in
$\gamma_i$, but not in $\gamma_{i+1}$, then the {\em first
$u$-segment} of $e$ is $e$ itself and there is no other $u$-segment.
Otherwise, let $\gamma_i \mapsto \gamma_{i+1}$ be the first step of
$e$, where there is a non-root process in $V_u$ which is an alive
abnormal root in $\gamma_i$, but not in $\gamma_{i+1}$. The {\em
first $u$-segment} of $e$ is the prefix $\gamma_0, \cdots,
\gamma_{i+1}$. The {\em second $u$-segment} of~$e$ is the first
$u$-segment of the suffix~$\gamma_{i+1}, \gamma_{i+2}, \cdots$, and so
forth.
\end{definition}
By Lemma~\ref{lem:pseudoRoots}, we have
\begin{observation}\label{obs:seg}
For every non-root process $u$, for every execution $e$, $e$ contains at
most~${n_{\texttt{maxCC}}}+1$ $u$-segments, because there are initially at
most~${n_{\texttt{maxCC}}}$ alive abnormal roots in $V_u$.
\end{observation}
\begin{lemma}
Let $u$ be any non-root process. During a $u$-segment, if $u$
executes the rule~$\mathrm{R_{EF}}$, then $u$ does not execute any
other rule in the remaining of the $u$-segment.
\end{lemma}
\begin{proof}
Let $\mathsf{seg}_u$ be a $u$-segment. Let~$s_1$ be a step of $\mathsf{seg}_u$ in
which $u$ executes~$\mathrm{R_{EF}}$. Let~$s_2$ be the next step in
which $u$ executes its next rule. (If $s_1$ or $s_2$ do not exist,
then the lemma trivially holds for $\mathsf{seg}_u$.)
Just before~$s_1$, all branches containing~$u$ have an alive
abnormal root, namely the non-root process~$v$ at depth~$1$ in any
of these branches. (Note that we may have $v=u$.) On the other hand,
just before~$s_2$, $u$ is the dead abnormal root of all branches it
belongs to. This implies that $v$ must have executed the
rule~$\mathrm{R_{EF}}$ in the meantime and thus is not an alive
abnormal root anymore when the step $s_2$ is executed. Therefore,
$s_1$ and~$s_2$ belong to two distinct $u$-segments of the
execution.
\end{proof}
\begin{corollary}\label{cor:language}
Let $u$ be a non-root process. The sequence of rules executed by
$u$ during a $u$-segment belongs to the following language:
$(\mathrm{R_I} + \varepsilon)(\mathrm{R_R} +
\varepsilon)\mathrm{R_C}^*(\mathrm{R_{EB}} +
\varepsilon)(\mathrm{R_{EF}} + \varepsilon)$.
\end{corollary}
We use the notion of \emph{maximal causal chain} to further analyze
the number of steps in a $u$-segment.
\begin{definition}[Maximal Causal Chain]
Let $u$ be a non-root process and~$\mathsf{seg}_u$ be any $u$-segment. A
{\em maximal causal chain of $\mathsf{seg}_u$} rooted at~$u_0 \in V_u$ is a
maximal sequence of actions $a_1,a_{2},\cdots,a_k$ executed
in~$\mathsf{seg}_u$ such that the action~$a_1$ sets $\variable{par}_{u_1}$ to~$u_0
\in V_u$ not later than any other action by~$u_0$ in~$\mathsf{seg}_u$, and
for all~$2 \leq i \leq k$, the action~$a_i$ sets $\variable{par}_{u_i}$
to~$u_{i-1}$ after the action~$a_{i-1}$ but not later than
$u_{i-1}$'s next action.
\end{definition}
\begin{observation}\label{obs:5} ~
\begin{itemize}
\item An action $a_i$ belongs to a maximal causal chain if and only if
$a_i$ consists in a call to the macro $computePath$ by a non-root
process.
\item Only actions of Rules~$\mathrm{R_R}$ and~$\mathrm{R_C}$ contain
the execution of~$computePath$.
\end{itemize}
Let $u$ be a non-root process and~$\mathsf{seg}_u$ be any $u$-segment. Let~ $a_1,a_{2},\cdots,a_k$ be a maximal
causal chain of $\mathsf{seg}_u$ rooted at~$u_0$.
\begin{itemize}
\item For all~$ 1 \leq i \leq k$, $a_i$ consists in the execution of
$computePath$ by~$u_i$ ({\em i.e.},~$u_i$ executes the
rule~$\mathrm{R_R}$ or~$\mathrm{R_C}$) where $u_i \in V_u$.
\item Denote by $ds_{\mathsf{seg}_u,v}$ the distance value of process~$v$ at
the beginning of~$\mathsf{seg}_u$. For all~$1 \leq i \leq k$, $a_i$
sets~$\variable{d}_{u_i}$ to
$ds_{\mathsf{seg}_u,u_0}+\sum_{j=1}^{j=i}w(u_j,u_{j-1})$, where $u_i$
is the process that executes $a_i$.
\end{itemize}
\end{observation}
For the next lemmas and theorems, we recall that ${n_{\texttt{maxCC}}} \leq n-1$ is
the maximum number of non-root processes in a connected component of
$G$.
\begin{lemma}\label{lem:causalchain}
Let $u$ be a non-root process. All actions in a maximal causal chain
of a $u$-segment are caused by different non-root processes of
$V_u$. Moreover, an execution of~$computePath$ by some non-root
process~$v$ never belongs to any maximal causal chain rooted at~$v$.
\end{lemma}
\begin{proof}
First note that any rule~$\mathrm{R_C}$ executed by a process~$v$
makes the value of~$\variable{d}_v$ decrease.
Assume now, by the contradiction, that there exists a process $v$
such that, in some maximal causal chain $a_1,a_{2},\cdots,a_k$
of a $u$-segment, $v$ is used as parent in some action $a_i$
and executes the action $a_j$, with $j>i$. The value of~$\variable{d}_{v}$
is strictly larger just after the action~$a_j$ than just before the
action~$a_i$. This implies that process $v$ must have executed the
rule~$\mathrm{R_R}$ in the meantime. So, $a_i$ and $a_j$ are
executed in two different $u$-segments by Corollary~\ref{cor:language}
and the fact that $v$ has status $C$ just before the
action~$a_i$. Consequently, they do not belong to the same maximal
causal chain, a contradiction.
Therefore, all actions in a maximal causal chain are caused by
different processes, and a process never executes an action in a
maximal causal chain it is the root of. As all actions in a maximal
causal chain are executed by processes in the same connected
component, we are done.
\end{proof}
\begin{definition}[$S_{\mathsf{seg}_u,v}$]
Given a non-root process $u$ and a $u$-segment~$\mathsf{seg}_u$, we define
$S_{\mathsf{seg}_u,v}$ as the set of all the distance values obtained after
executing an action belonging to any maximal causal chain of~$\mathsf{seg}_u$ rooted
at~process $v$ ($v \in V_u$)).
\end{definition}
Note that, from Observation~\ref{obs:5} and Lemma~\ref{lem:causalchain}, we have the following observation:
\begin{observation}\label{obs:finite}
The size of the set $S_{\mathsf{seg}_u,v}$ is bounded by a function of the
number of processes in $V_u$.
\end{observation}
\begin{lemma}\label{lem:Xnmax}
Given a non-root process $u$ and a $u$-segment~$\mathsf{seg}_u$, if the size
of~$S_{\mathsf{seg}_u,v}$ is bounded by~$X$ for all~process $v \in V_u$,
then the number of~$computePath$ executions done by $u$ in~$\mathsf{seg}_u$ is
bounded by~$X({n_{\texttt{maxCC}}}-1)$.
\end{lemma}
\begin{proof}
Except possibly the first, all $computePath$ executions done by a $u$
in a $u$-segment~$\mathsf{seg}_u$ are done through the
rule~$\mathrm{R_C}$. For all these, the variable $\variable{d}_u$ is always
decreasing. Therefore, all the values of $\variable{d}_u$ obtained by the
$computePath$ executions done by~$u$ are different. By definition of
$S_{\mathsf{seg}_u,v}$ and by Lemma~\ref{lem:causalchain}, all these values
belong to the set $\bigcup_{v \in V_u \setminus \{u\}}S_{\mathsf{seg}_u,v}$,
which has size at most $X({n_{\texttt{maxCC}}}-1)$.
\end{proof}
By definition, each step contains at least one action, made by a
non-root process. Let $u$ be any non-root process. Assume that, in
any $u$-segment $\mathsf{seg}_u$, the size of~$S_{\mathsf{seg}_u,v}$ is bounded
by~$X$ for all~process $v \in V_u$. So, the number of step of $u$ in
$\mathsf{seg}_u$ is bounded by~$X({n_{\texttt{maxCC}}}-1)+3$, by Lemma~\ref{lem:Xnmax} and
Corollary~\ref{cor:language}. Moreover, recall that each execution
contains at most ${n_{\texttt{maxCC}}}+1$ $u$-segments
(Observation~\ref{obs:seg}). So, $u$ executes in at most $X{n_{\texttt{maxCC}}}^2
+3{n_{\texttt{maxCC}}}-X+3$ steps. Finally, as $u$ is an arbitrary non-root process
and there are $n-1$ non-root processes, follows.
\begin{theorem}\label{theo:compl}
If the size of~$S_{\mathsf{seg}_u,v}$ is bounded by~$X$ for all~non-root
process $u$,~for all $u$-segment $\mathsf{seg}_u$, and for all process $v$
in $V_u$, then the total number of steps during any execution, is
bounded by $(X{n_{\texttt{maxCC}}}^2 +3{n_{\texttt{maxCC}}}-X+3)(n-1)$.
\end{theorem}
Let $\texttt{W}_{\max} = \max_{\{u,v\} \in E} \omega(u,v)$. If all weights are
strictly positive integers, then the size of any~$S_{\mathsf{seg}_u,v}$, where
$u$ is a non-root process and $v \in V_u$, is bounded by~$\texttt{W}_{\max}{n_{\texttt{maxCC}}}$,
because~$S_{\mathsf{seg}_u,v} \subseteq [ds_{\mathsf{seg}_u,v} + 1, ds_{\mathsf{seg}_u,v} +
\texttt{W}_{\max}(n_{cc}-1)]$, where $n_{cc} \leq {n_{\texttt{maxCC}}}+1$ is the number of
processes in $V_u$. Hence, we deduce the following theorem from
Theorem~\ref{theo:compl}, Observation~\ref{obs:finite}, and
Corollary~\ref{coro:legterm}.
\begin{theorem}
Algorithm~$\ensuremath{\mathsf{RSP}}\xspace$ is silent self-stabilizing under the distributed
unfair daemon for the set $\mathcal{LC}_{\ensuremath{\mathsf{RSP}}\xspace}$ and, when all weights
are strictly positive integers, its stabilization time in steps is
at most~$[\texttt{W}_{\max}{n_{\texttt{maxCC}}}^3+(3-\texttt{W}_{\max}){n_{\texttt{maxCC}}}+3](n-1)$, {\em i.e.},
$O(\texttt{W}_{\max}{n_{\texttt{maxCC}}}^3n)$.
\end{theorem}
If all edges in~$G$ have the same weight~$w$, then the size
of~$S_{\mathsf{seg}_u,v}$, where $u$ is a non-root process and $v \in V_u$, is bounded by~${n_{\texttt{maxCC}}}$. Indeed, in such a case, we
have $S_{\mathsf{seg}_u,v} \subset \{ ds_{\mathsf{seg}_u,v} + i.w ~|~ 1 \leq i \leq
n_{cc}-1\}$, where $n_{cc} \leq {n_{\texttt{maxCC}}}+1$ is the number of
processes in $V_u$. Hence, we obtain the following
corollary.
\begin{corollary}
If all edges have the same weight, then the stabilization time in
steps of Algorithm~$\ensuremath{\mathsf{RSP}}\xspace$ is at most $({n_{\texttt{maxCC}}}^3 +2{n_{\texttt{maxCC}}}+3)(n-1)$, which
is less than or equal to $n^4$ for all $n \geq 2$.
\end{corollary}
\section{Round Complexity of Algorithm \ensuremath{\mathsf{RSP}}\xspace}\label{sect:rounds}
We now prove that every execution of Algorithm \ensuremath{\mathsf{RSP}}\xspace lasts at
most~$3{n_{\texttt{maxCC}}}+D$ rounds, where~${n_{\texttt{maxCC}}}$ is the maximum number of
non-root processes in a connected component and $D$ is the
hop-diameter of the connected component containing~$r$, $V_r$.
The first lemma essentially claims that all processes that are in
illegal branches progressively switch to status~$EB$ within~${n_{\texttt{maxCC}}}$
rounds, in order of increasing depth.
\begin{lemma}\label{lem:propagatingEB}
Let $i \in \mathds{N}^*$. Starting from the beginning of round~$i$,
there does not exist any process both in state~$C$ and at depth less
than~$i$ in an illegal branch.
\end{lemma}
\begin{proof}
We prove this lemma by induction on~$i$. The base case~($i=1$) is
vacuum, so we assume that the lemma holds for some integer~$i\geq
1$. From the beginning of round~$i$, no process can ever choose a
parent which is at depth smaller than~$i$ in an illegal branch
because those processes will never have status~$C$, by induction
hypothesis. Moreover, no process with status~$C$ can have its depth
decreasing to $i$ or smaller by an action of one of its ancestors at
depth smaller than $i$, because these processes have status~$EB$ and
have at least one child not having status~$EF$. Thus, they cannot
execute any rule. Therefore, no process can take state~$C$ at depth
smaller or equal to~$i$ in an illegal branch.
Consider any process~$u$ with status~$C$ at depth~$i$ in an illegal
branch at the beginning of the round~$i$. $u \neq r$. Moreover, by
induction hypothesis,~$u$ is an abnormal root, or the parent of~$u$
is not in state~$C$ ({\em i.e.}, it is in the state~$EB$). During
round~$i$,~$u$ will execute rule~$\mathrm{R_{EB}}$
or~$\mathrm{R_{C}}$ and thus either switch to state~$EB$ or join
another branch at a depth greater than $i$. This concludes the
proof of the lemma.
\end{proof}
\begin{corollary}\label{cor:propagatingEB}
After at most ${n_{\texttt{maxCC}}}$ rounds, the system is in a configuration from
which no process in any illegal branch has status~$C$ forever.
Moreover, once such a configuration is reached, each time a process
executes a rule other than~$\mathrm{R_{EF}}$, this process is
outside any illegal branch forever.
\end{corollary}
The next lemma essentially claims that, once no process in an illegal
branch has status~$C$ forever, processes in illegal branches
progressively switch to status~$EF$ within~at most ${n_{\texttt{maxCC}}}$ rounds, in
order of decreasing depth.
\begin{lemma}\label{lem:propagatingEF}
Let $i \in \mathds{N}^*$. Starting from the beginning of
round~${n_{\texttt{maxCC}}}+i$, there does not exist any process at depth larger
than~${n_{\texttt{maxCC}}}-i+1$ in an illegal branch having the status~$EB$.
\end{lemma}
\begin{proof}
We prove this lemma by induction on~$i$. The base case~($i=1$) is
vacuum (by Observation~\ref{obs:size}), so we assume that the lemma
holds for some integer~$i\geq 1$. At the beginning of
round~${n_{\texttt{maxCC}}}+i$, no process at depth larger than~${n_{\texttt{maxCC}}}-i+1$ has the
status~$EB$ (by induction hypothesis) or status~$C$ (by
Corollary~\ref{cor:propagatingEB}). Therefore, processes with
status~$EB$ at depth~${n_{\texttt{maxCC}}}-i+1$ in an illegal branch can execute
the rule~$\mathrm{R_{EF}}$ at the beginning of round~${n_{\texttt{maxCC}}}+i$.
These processes will thus all execute within round~${n_{\texttt{maxCC}}}+i$ (they
cannot be neutralized as no children can connect to them). We
conclude the proof by noticing that, from
Corollary~\ref{cor:propagatingEB}, once round~${n_{\texttt{maxCC}}}$ has
terminated, any process in an illegal branch that executes either
gets status~$EF$, or will be outside any illegal branch forever.
\end{proof}
The next lemma essentially claims that, after the propagation of
status $EF$ in illegal branches, the maximum length of illegal
branches progressively decreases until all illegal branches vanish.
\begin{lemma}\label{lem:propagatingI}
Let $i \in \mathds{N}^*$. Starting from the beginning of
round~$2{n_{\texttt{maxCC}}}+i$, there does not exist any process at depth larger
than~${n_{\texttt{maxCC}}}-i+1$ in an illegal branch.
\end{lemma}
\begin{proof}
We prove this lemma by induction on~$i$. The base case~($i=1$) is
vacuum (by Observation~\ref{obs:size}), so we assume that the lemma
holds for some integer~$i\geq 1$. By induction hypothesis, at the
beginning of round~$2{n_{\texttt{maxCC}}}+i$, no process is at depth larger than or
equal to~${n_{\texttt{maxCC}}}-i+1$ in an illegal branch. All processes in an
illegal branch have the status~$EF$. So, at the beginning of
round~$2{n_{\texttt{maxCC}}}+i$, any abnormal root satisfies the
predicate~$\ensuremath{P\_reset}\xspace$, they are enabled to execute
either~$\mathrm{R_I}$, or~$\mathrm{R_R}$.
So, all abnormal roots at the beginning of the round~$2{n_{\texttt{maxCC}}}+i$ are
no more in an illegal branch at the end of this round: the maximal
depth of the illegal branches has decreased, since by
Corollary~\ref{cor:propagatingEB}, no process can join an illegal
tree during the round~$2{n_{\texttt{maxCC}}}+i$.
\end{proof}
\begin{corollary}\label{cor:propagatingI}
After at most round~$3{n_{\texttt{maxCC}}}$, there are no illegal branches forever.
\end{corollary}
Note that in any connected component that does not contain the
root~$r$, there is no legal branch. Then, since the only way for a
process to be in no branch is to have status~$I$, we obtain the
following corollary.
\begin{corollary}\label{cor:convergeI}
For any connected component~$H$ other than $V_r$, after at
most~$3{n_{\texttt{maxCC}}}$ rounds, every process of~$H$ is in a legitimate state
forever.
\end{corollary}
In the connected component~$V_r$, Algorithm~\ensuremath{\mathsf{RSP}}\xspace may need additional
rounds to propagate the correct distances to~$r$.
In the next lemma, we use the notion of hop-distance to~$r$
defined below.
\begin{definition}[Hop-Distance and Hop-Diameter]
A process~$u$ is said to be at \emph{hop-distance}~$k$ from~$v$ if
the minimum number of edges in a shortest path from $u$ to $v$ is~$k$.
The \emph{hop-diameter} of a graph $G$ (resp. of a connected
component $H$ of the graph $G$) is the maximum hop-distance between
any two nodes of $G$ (resp. of $H$).
\end{definition}
\begin{lemma}\label{lem:lastDrounds}
Let~$i \in \mathds{N}$. In every execution of Algorithm~\ensuremath{\mathsf{RSP}}\xspace, starting
from the beginning of round~$3{n_{\texttt{maxCC}}}+i$, every process at hop-distance
at most~$i$ from~$r$ is in a legitimate state.
\end{lemma}
\begin{proof}
We prove this lemma by induction on $i$. First, by definition, the
root $r$ is always in a legitimate state, so the base case~($i=0$)
trivially holds.
Then, after at most $3{n_{\texttt{maxCC}}}$ rounds, every process either belongs to
a legal branch or has status~$I$ (by
Corollary~\ref{cor:propagatingI}), thus any non-isolated process~$v
\in V_r$ always stores a distance $d$ such that $d \geq d(v,r)$, its
actual weighted distance to~$r$.
%
By induction hypothesis, every process at hop-distance at most~$i$
from~$r$ has converged to a legitimate state within at most
$3{n_{\texttt{maxCC}}}+i$ rounds. Therefore, at the beginning of
round~$3{n_{\texttt{maxCC}}}+i+1$, every process~$v$ at hop-distance~$i+1$ from~$r$
which is not in a legitimate state is enabled for executing
rule~$\mathrm{R_C}$. Thus, at the end of round~$3{n_{\texttt{maxCC}}}+i+1$, every
process at hop-distance at most~$i+1$ from~$r$ is in a legitimate
state (such processes cannot be neutralized during this round).
Also, these processes will never change their state since there are
no processes that can make them closer to~$r$.
\end{proof}
Summarizing all the results of this section, we obtain the following
theorem.
\begin{theorem}\label{theo:main-fair}
Every execution of Algorithm \ensuremath{\mathsf{RSP}}\xspace lasts at most~$3{n_{\texttt{maxCC}}}+D$ rounds,
where~${n_{\texttt{maxCC}}}$ is the maximum number of non-root processes in a
connected component and $D$ is the hop-diameter of the connected
component containing~$r$.
\end{theorem}
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{T}{here} is an increasing demand for multimedia streaming applications thanks to the ubiquity of internet access, the availability of the online content and the growing number of wireless hand-held devices. The predictions of Cisco Visual Networking Index \cite{cisco} indicate that IP video traffic will constitute 82 percent of all consumer internet traffic by 2021. For instance, in 2016, YouTube and Netflix account for up to 53 percent of fixed access Internet traffic in North America \cite{sandvine}. Moreover, 21 percent of the mobile internet traffic in North America is solely based on YouTube \cite{sandvine}.
\begin{figure*}
\centering
\includegraphics[scale=0.4]{systemarc.pdf}
\caption{Multiuser video streaming system}
\label{fig1}
\end{figure*}
In this work, we derive a Dynamic Adaptive Streaming over HTTP (DASH)-compatible multiuser resource control policy, which we call blind deadline-based resource allocation (BDRA) scheme, operating at an edge server. The aim of the BDRA scheme is to perform slot-based resource allocation to users in order to minimize the probability of a \emph{stalling} event at a client. When the amount of data at the buffer of a client is insufficient to continue to display the video stream, a stalling event occurs, and the client begins a re-buffering period during which it fills its buffer without displaying the video stream. The BDRA scheme utilizes Media Presentation Description (MPD) files of clients and HTTP-GET requests, which are sent in compliant to DASH protocol, in order to define and update the deadline of each media file displayed by a client. Then, it allocates slots to the users in the order of their deadlines. We formally prove that this algorithm minimizes the stalling probability for a network with links associated with fixed loss rates.
In conventional applications of DASH framework, the client is the only agent that manages the video streaming process in order to maximize the subjective video quality \cite{csca1,csca2,csca3,csca4,csca5,csca6}. In particular, the main promise of DASH is that the clients dynamically select among different representations of the same media stream differing with respect to video encoding rates based on the estimated network throughput. However, while each client has access to only its own MDP file, the edge server has access to MDP files of all clients it is serving. Hence, the edge server has a better view of the overall operation of the network, and it is in a position to \emph{proactively} take resource allocation decisions to prevent stalling events, whereas individual clients can only react after a stalling event occurs.
The DASH protocol has several benefits over push-based media streaming protocols such as Real-time Transport Protocol (RTP) \cite{dash1,dash2,dash3,dash4}. First, the infrastructure of the Internet has evolved to efficiently support HTTP, and HTTP offers ubiquitous connectivity.
Second, DASH is a pull-based protocol, so it traverses the firewalls. Third, the underlying TCP/IP protocol is widely deployed and provide reliable data transmission. Fourth, a client does not have to maintain a session state on the server to stream over the HTTP reducing overhead at the server.
The importance of the problem is well established as indicated by past and ongoing studies, e.g., \cite{main1} and \cite{main2}. In \cite{main1}, the authors examined the joint optimization of network resource allocation and video quality adaptation. The authors propose a resource allocation algorithm that aims to prevent the stalling event by employing a parameter that reflects the risk of stalling according to the duration of the video in the client’s buffer. A larger rate is assigned to a user that has a high stalling risk.
In \cite{main2}, the authors introduce the notion of {\em playout lead}, which is defined as the duration of the additional time a client can play the video by using its currently buffered data. The authors propose an algorithm that aims to prevent the stalling occurrences by maximizing the playout lead for all clients. To this end, the resource (time slots) is allocated so that the minimum of the playout lead among all users is maximized. Besides \cite{main1} and \cite{main2}, a buffer-aware approach is considered in \cite{bufferaware} where the video streaming traffic is shaped by SDN controllers according to clients' buffer status and the buffer occupancy trends.
Our work improves the current state-of-the-art in two ways. First, we prove that the derived policy of serving the clients in the order of deadlines is optimal in the sense that it minimizes the stalling event probability of the network when the link loss rates are fixed.\footnote{ When the instantaneous channel state information is available, this information can be further utilized to modify the scheduling algorithm in order to improve the network performance as in \cite{channelbased}.} Second, our policy relies only on the acknowledgment (ACK) feedback from the clients taken in the form of HTTP-GET requests for the subsequent byte ranges of the media file, and thus, significantly reducing the implementation complexity.
Note that in HTTP adaptive streaming, the quality of experience (QoE) depends on the selection of different system parameters such as initial setup delay, re-buffer duration, average video quality, video quality fluctuations and the number of stalling events \cite{vq1,vq2,vq3,vq4,vq5,vq6}. In practical DASH implementations, on the other hand, the client is only responsible for the video quality selection process. In this paper, considering the initial setup delay and re-buffer duration as predefined system parameters, we focus on minimizing the stalling probability with our server side algorithm in order to improve the QoE of users. To the best of our knowledge, this paper is the first study that proposes a systematic approach based on Markov decision processes (MADP) in order to investigate the performance of DASH based multiuser video streaming systems with network assistance. The proposed MADP framework enables us to take into account the effect of resource allocation decisions in the current time-slot on the stalling likelihood in future time-slots over a finite time horizon. Our main contributions in the paper are summarized as follows.
\begin{itemize}
\item Using dynamic programming, we show that the {\em optimum} algorithm minimizing the system-wide stalling probability in DASH based multiuser video streaming systems when only statistical knowledge of channels is available at the server-side is a {\em blind} deadline based algorithm, which we call the BDRA algorithm.
\item Having a simple structure with polynomial-time computational complexity, the BDRA algorithm is easy to implement as a server-side add-on solution for the existing DASH architecture in order to reduce the frequency of stalling events. We further provide a particular implementation of the BDRA algorithm that prioritizes the users with small GoP sizes in order to achieve fairness among the streaming users with varying bit-rates for video files.
\item Thanks to its operation oblivious to the quality adaptation mechanism at the client side, the BDRA algorithm can operate together with any choice of quality adaptation scheme such as buffer-based adaptation (BBA) and rate-based adaptation (RBA), which further increases the utility of the derived BDRA algorithm.
\item We perform NS3 simulations in order to illustrate the {\em optimality gap} between the BDRA scheme and four other blind resource allocation schemes.
\end{itemize}
The remainder of the paper is organized as follows. In Section \ref{sec:background}, we provide a detailed background on the operation of the DASH protocol.
Section \ref{Section: System Model} provides the analytical model for our system as well as the optimum slot-based resource allocation problem to be solved. The BDRA algorithm is formally introduced and its optimality is formally established in Section \ref{sec:optimum}.
Implementation and design issues regarding the BDRA algorithm are explained in Section \ref{sec:Implementation}. Performance of the BDRA algorithm, in comparison to commonly used {\em rate-fair resource allocation} schemes, is numerically investigated in Section \ref{sec:numerical}. Section \ref{sec:related} provides a detailed discussion on the previous work that is most relevant to our findings in this paper, by first describing the current state-of-the-art and then explaining the differences between our solution and these previous solutions in detail. Finally, we conclude the paper with a summary of findings and future research directions in Section \ref{sec:conclusion}.
\section{DASH Video Streaming}
\label{sec:background}
As illustrated in Fig. \ref{fig1}, the studied video streaming system consists of two main sections: A Long Distance Network (LDN) and an Access Network (AN). The LDN may involve both a main server and a content delivery network (CDN), and it has the responsibility of delivering the requested video files to the edge servers in the AN. In general, the bottleneck of the end-to-end connection is at the edge servers, so we focus on the resource allocation strategies operating at the edge servers to alleviate this bottleneck.
Due to their significant advantages over push-based media streaming protocols such as RTP, HTTP-based streaming protocols have been widely adopted by most of the on-demand video service providers. In particular, DASH protocol is developed to provide a common set of functionality among different HTTP-based streaming protocols \cite{dash1,dash2,dash3,dash4}.
In DASH, a video file is encoded with multiple different bit-rates into different \emph{representations}, where each representation corresponds to a different level of quality of the same video stream. Each representation is broken into \emph{segments} of duration 2-10 seconds \cite{seg}. Segments may be further subdivided into \emph{sub-segments}, each of which contains a whole number of complete access units. The video content providers employing DASH often use video files encoded according to Advanced Video Coding (AVC) (e.g., H264.AVC) standard. In this video encoding format, the smallest meaningful bit-chunk is called {\em Group of Pictures} (GoP) since the frames of the same GoP are encoded and decoded together \cite{gop,dash4}. Thus, an AVC encoded video file is considered as a combination of mutually exclusive fragments that contain different frames of the same video file. Each GoP contains a fixed number of frames and has a fixed video display duration. We note that although each GoP has a fixed video display duration, their sizes might be different due to video content. To display a frame, all information related to the corresponding GoP needs to be available at the client buffer. \\%Although each GoP has a fixed duration, each may consist of different number of bits.
\begin{figure*}
\centering
\includegraphics[scale=0.5]{proposed.pdf}
\caption{ End-to-end video streaming system.}
\label{fig2}
\end{figure*}
\indent
The DASH client behavior can be summarized as follows. The client first accesses the Media Presentation Description (MPD) file. The MPD file contains metadata required by a DASH client to construct appropriate HTTP-URLs to access segments and to provide the streaming service to the user. In particular, an MPD file provides information for the earliest presentation time and presentation duration for each segment in the representation.\footnote{This information will be used by the derived optimum algorithm to perform resource allocation among multiple DASH clients.} The client selects an appropriate video representation, typically based on an estimate of the available bandwidth to the server but also on the rendering capabilities of the client. Then, the client creates a list of accessible segments for each representation. The client accesses the content by requesting entire segments or byte ranges of segments via HTTP-GET command. Once the presentation has started, the client continues consuming the media content by continuously requesting segments or parts of segments.
The client may switch representations taking into account updated information from its environment, e.g., change of observed throughput. In this paper, we focus on the resource allocation at the edge server. Hence, DASH clients can use any adaptive video quality selection algorithm to select an appropriate representation based on the observed throughput and client capabilities.
\section{Analytical Model, Definitions and the Optimum Scheduling Problem } \label{Section: System Model}
In this section, we will introduce the details of our analytical model (following the standard terminology of the {\em MADP} literature \cite{dynamicprog}), the definitions that go with this model and the {\em optimum scheduling problem} that we solve to minimize the number of stalling events in DASH based multiuser video streaming systems.
\subsection{Receiver and Playout Curves}
The data arrival process of client $i$ is denoted by $R_{i}(t)$, which we call the \emph{receiver curve} of client $i$. The receiver curve $R_{i}(t)$ indicates the total amount of error free data in unit of packets that is delivered to client $i$ up to time $t$. For each client $i$, $R_{i}(t)$ is a non-decreasing function of $t$. The video of client $i$ is displayed according to $p_{i}(t)$, which is called the \emph{playout curve}. The playout curve describes the minimum amount of data in units of packets that needs to be decoded up to time $t$ to perform uninterrupted video display.
\begin{figure}
\centering
\includegraphics[scale=0.4]{systemparam.pdf}
\caption{Receiver and Playout curves}
\label{systemparameter}
\end{figure}
The GoP based structure of the video files implies that all playout curves are right continuous functions as illustrated by Fig. \ref{systemparameter}. A time instant $t>0$ at which there is a jump in the playout curve, i.e., $p(t^-) \neq p(t)$ for any $t^- < t$, is called an {\em increment point}.
We consider a time-slotted video streaming system with fixed slot length equal to $\Delta$ so that the edge server can serve only one user in each slot duration. Hence, the receiver curve $R_{i}(t)$ increases by one unit at the end of a time-slot if and only if user $i$ is scheduled at the beginning of the corresponding time-slot and the transmitted packet is received successfully. As a result of this operation, all receiver curves are right continuous functions as well, an example of which is illustrated in Fig. \ref{systemparameter}. We further assume that GoP duration is also an integer multiple of the slot duration $\Delta$. Since both playout curves and receiver curves remain constant during a slot duration, we can discretize these functions and use time index $k=\floor{t/\Delta}$, where $\floor{\cdot}$ is the {\em floor} function that produces the largest integer smaller than or equal to its argument. Throughout the paper, we will normalize $\Delta$ to {\em one} time unit to simplify notation. To ensure continuous displaying of a video at client $i$, there should be sufficient number of packets in the client buffer so that the following inequality holds
\begin{equation}\label{neccond}
R_{i}(t) \geq p_{i}(t)
\end{equation}
for any time instant $t$.
\subsection{Analytical Model and Definitions}
Our primary aim is to discover the structure of the optimum scheduling policy (at the edge server side) that will minimize the stalling event probability for multiuser video streaming over stochastically varying wireless channels. To this end, we focus on minimizing the stalling probability per segment, where each segment spans $T \in \N$ consecutive slots of time.\footnote{The main reason for us to consider {\em only} the segment stalling probability in this paper is the technological constraint introduced by the DASH protocol. In particular, the DASH protocol determines the representation level of the next segment only after the current segment requests are provisioned, and we cannot state our optimum scheduling problem without knowing the representation levels of the forthcoming segments.} Hence, without loss of generality, we model our optimum slot-based resource allocation problem as a {\em finite} horizon stochastic dynamic programming problem over time interval $[0, T]$ below.
The classical packet {\em erasure} channel is used to model wireless channels between the end users and the edge server, as such a packet sent for user $i$ is either successfully received with probability $\beta_i$ or lost with probability $1-\beta_i$ in each time slot.
We assume that channel statistics $\vecbold{\beta}=(\beta_{1},\ldots,\beta_{N})$ are known at time $t=0$ and remain the same over the time interval $\sqparen{0, T}$. Similarly, we also assume that playout curves (or, alternatively called representation levels) $\vecbold{p}(t)=\paren{p_{1}(t),\ldots, p_{N}(t)}$ are known by the edge server at time $t=0$, which is a standard assumption of the DASH protocol. Here, $\vecbold{p}(t)$ is a vector valued function that describes the amount of data (measured in terms of number of packets) required by each user up to time $t$ to display its video without any interruptions.
The edge server can serve {\em only one} user in each time slot. Hence, a scheduling decision must be made at the beginning of each time slot to select an {\em appropriate} user (i.e., usually the one that optimizes the system performance) for data transmission based on the current system state that summarizes the data reception history. In this paper, we represent the system states by the $N$ dimensional vector $\vecbold{s}=\paren{s_{1},\ldots,s_{N}}$, where $s_{i}$ is equal to the number of packets received by user $i \in \brparen{1, \ldots, N}$. We will often use states with time index $\floor{t}$ (or, by using the discrete time index $k \in \brparen{0, \ldots, T-1}$), i.e., $\vecbold{s}_{\floor{t}}=\paren{s_{1,\floor{t}},\ldots,s_{N,\floor{t}}}$, to denote the number of packets received by the users at the beginning of time slot $\floor{t}$. The set $\mathcal{S} = \brparen{0,1,2,\ldots,T}^{N}$ defines the set of all state vectors.\footnote{We note that $\mathcal{S}$ is larger than the set of all admissible states. If needed to be more precise, we can write $\mathcal{S}^{'}=\brparen{\vecbold{s}\in\mathcal{S}: \sum_{i=1}^N s_{i}\leq T}$.}
In this setting, we define the {\em action set} $\mathcal{A}$ to be $\mathcal{A}= \left\{1,2,\ldots,N\right\}$, and each action $a$ belonging to $\mathcal{A}$ denotes the index of the user scheduled for video streaming in the current time-slot. We note that $\mathcal{A}$ is a state-independent action set that remains the same for all $\vecbold{s}\in\mathcal{S}$. Consider now a specific time-slot $k$. An important quantity of interest that describes how the video streaming system in question evolves in time is the {\em transition probability function} $P_{k}\paren{\vecbold{z} \vert \vecbold{s}, a}$ that represents the transition probability of the video streaming system to another system state $\vecbold{z}$ at the beginning of the next time-slot given that the system state in the current time-slot $k$ is $\vecbold{s}$, i.e., $\vecbold{s}_k = \vecbold{s}$, and the action taken in this time-slot is $a$. Using the wireless channel model between the edger server and the users, $P_{k}\paren{\vecbold{z} \vert \vecbold{s}, a}$ can be more formally written as
\begin{eqnarray*}
P_{k}\paren{\vecbold{z} \vert \vecbold{s}, a} = \left\{\begin{array}{cc} 1-\beta_{a} & \mbox{ if } \vecbold{z} = \vecbold{s} \\ \beta_{a} & \mbox{ if } \vecbold{z} = \paren{s_{1},\ldots,s_{a}+1,\ldots, s_{N}}\\ 0 & \mbox{ otherwise} \end{array}\right..
\end{eqnarray*}
In addition to the analytical framework introduced above, two other major components of our model that operate on this framework are {\em decision rules} and the {\em scheduling policy}, which are what we define next. Considering the fact that packet success or failure events are independent from time-slot to time-slot in our wireless channel model\footnote{This assumption implies that knowing the transmission history and associated success or failure events do not give us any information about the channel conditions in the current time-slot.}, knowledge of the current system state is sufficient to predict current channel conditions and to construct remaining playout curves, i.e., remaining demand for data for uninterrupted video streaming. Hence, without loss of generality, we focus on Markovian and deterministic decision rules defined as functions that map the set of states $\mathcal{S}$ to the set of actions $\mathcal{A}$. More specifically, the decision rule $d_k$ for time-slot $k$ takes the system state $\vecbold{s}_k$ in the beginning of this time-slot as an input, and produces an action $a$ belonging to $\mathcal{A}$, i.e., $d_k\paren{\vecbold{s}_k} = a \in \mathcal{A}$, that represents the user index scheduled for video streaming in this time-slot.
Utilizing the definition of decision rules, we next state the definition of scheduling policy and {\em tail} scheduling policy below, which will conclude the description of our analytical model.
\begin{definition}
A scheduling policy $\vecbold{\pi} = \paren{d_{0}, \ldots, d_{T-1}}$ is a sequence of decision rules as such the $k$th element of $\vecbold{\pi}$ determines the index of the user scheduled for the $k$th time-slot based on the observed system state at the beginning of this time-slot for $k \in \brparen{0, \ldots, T-1}$. Similarly, a tail scheduling policy $\vecbold{\pi}_k = \paren{d_k, \ldots, d_{T-1}}$ is a sequence of decision rules that determines the indices of the users scheduled for the time-slots from $k$ to $T-1$.
\end{definition}
\subsection{The Optimum Scheduling Problem}
Having introduced our analytical model above, we are now ready to state the optimum scheduling problem. To this end, we first need to define {\em total expected reward} that is obtained when the user scheduling policy $\vecbold{\pi} = \paren{d_{0}, \ldots, d_{T-1}}$ is employed to determine scheduling decisions for each time slot.
\begin{definition}
The total expected reward $u^{\boldsymbol{\pi}}_{k}: \mathcal{S}\mapsto \R$ collected from time-slot $k$ to $T-1$ under the scheduling policy $\vecbold{\pi} = \paren{d_0, \ldots, d_{T-1}}$ is a function that maps the initial system state $\vecbold{s}_k$ at the beginning of the time-slot $k$ to a real number.
\end{definition}
We note that $u^{\boldsymbol{\pi}}_{k}$ can be easily expressed recursively as
\begin{equation}\label{receq1}
u^{\boldsymbol{\pi}}_{k}\paren{\vecbold{s}_{k}} = r_{k}\paren{\vecbold{s}_{k}, a} + \sum_{ \vecbold{s} \in \mathcal{S}} P_{k}\paren{\vecbold{s} \vert \vecbold{s}_{k},a} u^{\vecbold{\pi}}_{k+1}\paren{\vecbold{s}}
\end{equation}
for any $\vecbold{s}_k \in \mathcal{S}$, where $r_{k}\paren{\vecbold{s}_{k}, a}$ denotes the reward obtained by the scheduling decision $a = d_{k}\paren{\vecbold{s}_{k}}$ at time-slot $k$ if the current system state is $\vecbold{s}_{k}$, and the summation term in \eqref{receq1} represents the total expected reward obtained from time-slot $k+1$ onwards. It should be noted that $u^{\boldsymbol{\pi}}_{k}\paren{\vecbold{s}_{k}}$ in \eqref{receq1} depends on $\vecbold{\pi}$ only through its tail policy $\vecbold{\pi}_k = \paren{d_k, \ldots, d_{T-1}}$. For the sake of completeness, we set $u^{\vecbold{\pi}}_{T}\paren{\vecbold{s}_{T}} = r_{T}\paren{\vecbold{s}_{T}}$, where it is understood that $\vecbold{s}_{T}$ is the system state reached at the end of the video segment of interest, $r_{T}\paren{\vecbold{s}_{T}}$ is the reward collected due to the occurrence of $\vecbold{s}_{T}$, and no action is allowed at this termination time, which is a standard assumption of the finite horizon stochastic control problems \cite{dynamicprog}. The notion of {\em optimality} for a scheduling policy is introduced in the following definition.
\begin{definition} \label{Def: Scheduling Optimality}
Let $\Pi$ be the set of all scheduling policies. Then, we say that a scheduling policy $\vecbold{\pi}^\star$ is {\em optimum} if it solves the optimization problem below
\begin{eqnarray}
\max_{\vecbold{\pi} \in \Pi} u^{\boldsymbol{\pi}}_{k}\paren{\vecbold{s}} \label{Eqn: Scheduling Optimality}
\end{eqnarray}
for all time-slots $k \in \brparen{0, \ldots, T-1}$ and initial state vectors $\vecbold{s} \in \mathcal{S}$.\footnote{The maximum value in \eqref{Eqn: Scheduling Optimality} is always achieved since $\Pi$ is a finite set, and hence there is no ambiguity in this definition.}
\end{definition}
We note that the condition of optimality introduced in Definition \ref{Def: Scheduling Optimality} is a strong one since we do not only want a given scheduling policy is optimum itself considering time-slots from $0$ to $T-1$ but also want all of its tail policies to be optimum and achieve the {\em best} possible total expected reward starting from any time-slot and initial system state. To put it in another way, we want an optimal scheduling policy $\vecbold{\pi}^\star$ to satisfy the following equality
\begin{equation}
u^{\vecbold{\pi}^{\star}}_{k}\paren{\vecbold{s}}=u^{\star}_{k}\paren{\vecbold{s}}
\end{equation}
for all $k \in \brparen{0, \ldots, T-1}$ and $\vecbold{s} \in \mathcal{S}$, where $u^{\star}_{k}\paren{\vecbold{s}} = \max_{\vecbold{\pi} \in \Pi} u^{\vecbold{\pi}}_{k}\paren{\vecbold{s}}$.
We will derive the structure of $\vecbold{\pi}^\star$ by considering a specific but practically relevant total expected reward function, which is the system-wise segment {\em non}-stalling probability, i.e., none of the users experiences stalling throughout a particular segment duration. Indeed, our problem formulation lends itself to readily calculate the segment non-stalling probability if we set the per-slot reward functions $r_k\paren{\vecbold{s}, a}$ to zero for all $k \in \brparen{0, \ldots, T-1}$, set $r_{T}\paren{\vecbold{s}_{T}}$ to zero (one) if a stalling event does (not) occur at the end of time-slot $T-1$ (i.e., the termination time).\footnote{The condition to check if a stalling event occurs or not at the end of time-slot $T-1$ is equivalent to checking the inequality $s_{i, T} \geq p_i\paren{T}$ for all $i \in \brparen{1, \ldots, N}$. If this inequality is not satisfied for a user, we say that a stalling event occurs at the termination time $T$.} Accordingly, the total expected reward in \eqref{receq1} for the segment stalling probability can be written as
\begin{eqnarray} \label{receq2}
\lefteqn{u^{\vecbold{\pi}}_{\floor{t}}\paren{\vecbold{s}_{\floor{t}}}} \hspace{8cm} \nonumber \\
\lefteqn{ = \left\{\begin{array}{cc} \sum_{\vecbold{s} \in\mathcal{S}} P_{\floor{t}}\paren{\vecbold{s} \vert\vecbold{s}_{\floor{t}}, a}u^{\vecbold{\pi}}_{\floor{t}+1}\paren{\vecbold{s}} & \mbox{ if } \vecbold{s}_{\floor{t}} \succeq \vecbold{p}(t) \\ 0 & \mbox{otherwise} \end{array}\right.,} \hspace{8cm}
\end{eqnarray}
where ``$\succeq$" represents element-wise vector inequality and $a$ is the action taken in time-slot $\floor{t}$ by the scheduling policy $\vecbold{\pi} = \paren{d_0, \ldots, d_{T-1}}$, i.e., $a =d_{\floor{t}}\paren{\vecbold{s}_{\floor{t}}}$. It should be noted that a given scheduling policy $\vecbold{\pi}$ induces a probability distribution over the set of system states $\mathcal{S}$, which in turn determines a probability distribution for {\em random} receiver curves $R_i\paren{t}$ for $i \in \brparen{1, \ldots, N}$ and $t \in [0, T]$. Hence, $u^{\vecbold{\pi}}_{\floor{t}}\paren{\vecbold{s}_{\floor{t}}}$ can also be written as the probability that all random receiver curves to be above all playout curves over the time interval $[t, T]$ starting from the initial system state $\vecbold{s}_{\floor{t}}$. That is, $u^{\vecbold{\pi}}_{\floor{t}}\paren{\vecbold{s}_{\floor{t}}}$ is equal to
\begin{eqnarray}
\lefteqn{F_{\floor{t}}\paren{\vecbold{p},\vecbold{\pi}_{\floor{t}},\vecbold{s}_{\floor{t}}}} \nonumber \hspace{8cm}\\
= \lefteqn{\PRP{ \bigcap_{i=1}^N \brparen{R_i(\tau) \geq p_i(\tau), \forall \tau \in\sqparen{t, T}} \Big\vert \ \vecbold{s}_{\floor{t}}}.} \label{Eqn: Stalling Probability} \hspace{7cm}
\end{eqnarray}
Above representation of $u^{\vecbold{\pi}}_{k}\paren{\vecbold{s}}$ in \eqref{Eqn: Stalling Probability} that shows the dependence of segment stalling (or, non-stalling to this effect) probability on receiver and playout curves explicitly will be helpful in our derivation to determine the structure of the optimum scheduling policy in the next section.
\section{The Optimum Scheduling Policy}
\label{sec:optimum}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{scheduling3.pdf}
\caption{Schematic representation of playout curves at the beginning of time-slot $k$.}
\label{fig3}
\end{figure}
In this section, we derive the structure of the optimum scheduling policy that solves the optimum scheduling problem introduced in Section \ref{Section: System Model} for maximizing the non-stalling event probability in multiuser video streaming systems. In particular, it will be shown that a simple but practical greedy scheme that schedules users according to packet deadlines maximizes the segment non-stalling probability $u_k^{\vecbold{\pi}}\paren{\vecbold{s}}$ for all initial system states $\vecbold{s} \in \mathcal{S}$ as well as for time-slots $k \in \brparen{0,\ldots, T-1}$. We call this scheme the {\em blind deadline-based resource allocation} (BDRA) scheme. Before we formally state the optimality of the BDRA scheme in Theorem \ref{Thm: Optimality of BDRA scheme}, which is the main analytical result of this paper, it would be helpful to explain the operational details of the BDRA scheme through a particular situation for facilitating the upcoming discussion and the exposition of the proof of its optimality.
To this end, consider the case where the current time-slot index is $k$ and assume that there are $M$ jumps in the playout curves of users at time-slots $k+l_{1},\ldots,k+l_{M}$, which is illustrated in Fig. \ref{fig3}. These are the ordered time instants increasing from the smallest one to the biggest one with the last time instant $k+l_{M}$ coming no later than $T$. We recall that such a jump occurring in the playout curve of a user corresponds to the additional data demanded by this user (in terms of number of packets) for smooth displaying of its video, and this data demand must be provisioned by the edge server in order to avoid video stalling at this user.
We let $q_{i, m}$ denote the height of the jump at time-slot $k + l_{m}$ occurring at the playout curve of user $i$. Here, $q_{i, m}$ corresponds to the number of additional data packets requested by user $i$ between the deadlines $m-1$ and $m \leq M$. Therefore, we can consider the delivery of $q_{i, m}$ packets to user $i$ as a {\em task} with a {\em deadline} $k+l_{m}$. If this task is accomplished by the edge server for all deadlines, then no stalling event occurs at user $i$. The BDRA scheme simply prioritizes all such tasks based on their deadlines by instructing the edge server to conclude the tasks with the earliest deadlines first before proceeding to those with deadlines coming at later times. If there are two or more users with the same deadline, the BDRA scheme can choose any one of such users without any loss of optimality
\begin{theorem} \label{Thm: Optimality of BDRA scheme}
For given playout curves $\boldsymbol{p}$ and channel statistics $\boldsymbol{\beta}$, the BDRA scheme produces an optimal scheduling policy $\boldsymbol{\pi}^{bdra}$ i.e.,
\begin{equation}
u^{\boldsymbol{\pi}^{bdra}}_{k}(\boldsymbol{s})=u^{\star}_{k}(\boldsymbol{s})
\end{equation}
holds for all $k=0, \ldots,T-1$ and $\vecbold{s} \in \mathcal{S}$.
\end{theorem}
\begin{IEEEproof}
We will prove this theorem by induction.
\emph{Base Case:} We first consider the base case in which the optimum scheduling problem is solved for the last time-slot $T-1$. If there are two or more deadlines in the beginning of time-slot $T-1$, no scheduling policy can achieve stalling-free video streaming for all users, and therefore all scheduling policies are the same in terms of their segment stalling probability performances in such cases. On the other hand, if there is only one deadline in the beginning of time-slot $T-1$, the user associated with this deadline must be served to avoid a possible stalling event. This discussion shows that the BDRA scheme minimizes the segment stalling probability for the last time-slot.
\emph{Induction Step:} Secondly, we consider a time-slot with index $k+1 \leq T-2$ and assume that $u^{\vecbold{\pi}^{bdra}}_{k+1}\paren{\vecbold{s}} = u^{\star}_{k+1}\paren{\vecbold{s}}$ for all $\vecbold{s} \in\mathcal{S}$. Then, it is well-known from \cite{dynamicprog} that the optimal decision for time-slot $k$ must satisfy the following condition
\begin{equation}\label{optconst}
d^{\star}_{k}\paren{\vecbold{s}_k} \in \argmax_{a \in \mathcal{A}}\brparen{\sum_{\boldsymbol{s} \in \mathcal{S}} P_{k}\paren{ \boldsymbol{s} \vert \boldsymbol{s}_k, a} u^{\star}_{k+1}\paren{\vecbold{s}}}
\end{equation}
for all system states $\vecbold{s}_k \in \mathcal{S}$ in the beginning of time-slot $k$.
Since the induction hypothesis asserts that $u^{\boldsymbol{\pi}^{bdra}}_{k+1}\paren{\boldsymbol{s}} = u^{\star}_{k+1}\paren{\boldsymbol{s}}$, (\ref{optconst}) can also be expressed as
\begin{equation}\label{optconst2}
d^{\star}_{k}\paren{\boldsymbol{s}_{k}}\in \argmax_{a\in A} F_{k}\paren{\boldsymbol{p},\left(a,\boldsymbol{\pi}_{k+1}^{bdra}\right),\boldsymbol{s_{k}}}.
\end{equation}
Note that the term $\paren{a,\boldsymbol{\pi}_{k+1}^{bdra}}$ in (\ref{optconst2}) is a tail scheduling policy that is obtained by concatenating an action $a$ and the tail policy $\boldsymbol{\pi}_{k+1}^{bdra}$. Next, we will show that $\vecbold{\pi}_k^{bdra} = \paren{d^\star_k, \vecbold{\pi}_{k+1}^{bdra}}$. To this end, we will provide an alternative expression for $F_{k}\paren{\boldsymbol{p},\boldsymbol{\pi}_{k},\boldsymbol{s}_{k}}$ for any tail scheduling policy $\vecbold{\pi}_k$.
Let there be $M$ deadlines at $k+l_{1},\ldots,k+l_{M}$ for a given playout curve $\boldsymbol{p}$ and system state $\vecbold{s}_k \in \mathcal{S}$ after the time-slot $k$, an example of which is illustrated in Fig. \ref{fig3}. Let also the random variable $\lambda_{m}$ denote the first time-slot when all packets belonging to the first $m$ deadlines are delivered successfully. We note that $\lambda_{m}$ depends on the tail scheduling policy $\boldsymbol{\pi}_k$ and $\vecbold{p}$, and $F_{k}(\boldsymbol{p},\boldsymbol{\pi}_{k},\boldsymbol{s}_{k})$ can be expressed in terms of $\left\{\lambda_{m}\right\}^{M}_{m=1}$ as
\begin{equation}\label{sfvsp}
F_{k}(\boldsymbol{p},\boldsymbol{\pi}_{k},\boldsymbol{s}_{k})=\PRP{\bigcap^{M}_{m=1} \brparen{\lambda_{m}\leq k+l_{m}} \Big \vert \ \vecbold{s}_k}.
\end{equation}
Consider now the random variable $\tau_{m}$, which denotes the total number of time-slots required to send all $\sum^{N}_{i=1}q_{i,m}$ packets associated with the deadline at $k+l_{m}$ successfully. Under the BDRA scheme, the relationship between $\lambda_{m}$ and $\brparen{\tau_{i}}_{i=1}^m$ is $\lambda_{m}=k+\sum^{m}_{i=1}\tau_{i}$. Hence, using \eqref{sfvsp}, we obtain
\begin{equation}\label{sfvsp2}
F_{k}\paren{\boldsymbol{p},\boldsymbol{\pi}^{bdra}_{k},\boldsymbol{s}_{k}}=\PRP{ \bigcap^{M}_{m=1} \brparen{\sum^{m}_{i=1}\tau_{i} \leq l_{m}} \Big \vert \ \vecbold{s}_k}.
\end{equation}
Assume now that we choose an action $a \neq d_k^{bdra}\paren{\vecbold{s}_k}$ and form a tail scheduling policy $\paren{a, \vecbold{\pi}_k^{bdra}}$. For this tail scheduling policy, we will show that $F_k\paren{\vecbold{p}, \paren{a, \vecbold{\pi}_{k+1}^{bdra}}, \vecbold{s}_k} \leq F_{k}\paren{\boldsymbol{p},\boldsymbol{\pi}^{bdra}_{k},\boldsymbol{s}_{k}}$. Let the scheduled user $a$ has the first deadline at $k+l_{j}$ for some $j \geq 2$. Since time-slot $k$ is allocated for user $a$, and the slot allocation is done according to the tail policy $\boldsymbol{\pi}^{bdra}_{k+1}$ in the remaining time slots, we can write $F_k\paren{\vecbold{p}, \paren{a, \vecbold{\pi}_{k+1}^{bdra}}, \vecbold{s}_k}$ as
\begin{eqnarray}\label{sfvsp3}
\lefteqn{F_k\paren{\vecbold{p}, \paren{a, \vecbold{\pi}_{k+1}^{bdra}}, \vecbold{s}_k}} \hspace{8cm}\nonumber\\
\lefteqn{= \PRP{\bigcap_{m=1}^M \brparen{\sum_{i=1}^m \tau_i \leq l_m - \I{m < j} } \Big \vert \ \vecbold{s}_k},} \hspace{6.5cm}
\end{eqnarray}
where $\I{m<j}$ is an indicator function that returns $1$ if the inequality $m<j$ holds.\footnote{The random variables appearing in \eqref{sfvsp2} and \eqref{sfvsp3} must be considered to be equal in distribution.} Comparing \eqref{sfvsp2} and \eqref{sfvsp3}, we conclude that $F_{k}\paren{\boldsymbol{p},\boldsymbol{\pi}^{bdra}_{k},\boldsymbol{s}_{k}} \geq F_k\paren{\vecbold{p}, \paren{a, \vecbold{\pi}_{k+1}^{bdra}}, \vecbold{s}_k}$ for any $a \in \mathcal{A}$. This result implies that $\vecbold{\pi}_k^{bdra}$ is the optimum tail scheduling policy starting from any time-slot $k$ onwards, and hence $\vecbold{\pi}^{bdra}$ is the solution of the optimum scheduling problem given by \eqref{Eqn: Scheduling Optimality}.
\end{IEEEproof}
An important corollary of Theorem \ref{Thm: Optimality of BDRA scheme} is that the optimum scheduling policy minimizing the segment stalling probability does not depend on the statistical knowledge $\vecbold{\beta} = \paren{\beta_1, \ldots, \beta_N}$ of the wireless channel between the edge server and the users. This observation may seem counter-intuitive at a first glance. In particular, it can be conjectured that we should always perform better if we take channel statistics into account while giving scheduling decisions in each time-slot. However, the particular solution constructed for the optimum scheduling problem in Theorem \ref{Thm: Optimality of BDRA scheme}, i.e., the BDRA scheme, shows that we cannot improve the segment stalling probability even if we utilize the statistical channel knowledge.\footnote{The solution for the optimum scheduling problem in \eqref{Eqn: Scheduling Optimality} is not necessarily unique, and there may exist other resource allocation policies utilizing wireless channel statistics and achieving the same performance with the BDRA scheme. The determination of the complete solution set for \eqref{Eqn: Scheduling Optimality} is outside the scope of the current paper.} The point here is that the dynamic playout curve updating procedure embedded in the BDRA scheme already includes the effect of the packet drop probabilities of the users, and this is sufficient to make the BDRA scheme an optimum scheduling policy for multiuser video streaming systems.
This observation has some important practical ramifications. Firstly, implementation of the BDRA scheme avoids any channel estimation issues to learn channel conditions before it starts its operation. In particular, implementation of a channel estimation algorithm suited for the particular requirements of video streaming coupled with an efficient and high-throughput feedback protocol design (for frequency-division-duplexing systems) from users to the edge server may become an onerous task for delay sensitive video traffic.
Secondly, perhaps the most importantly, the BDRA scheme can be implemented as an {\em add-on} solution to the existing video streaming systems, especially to the DASH based systems, for improving their efficiency. Therefore, it must be backward-compatible with them for all practical purposes, rather than necessitating a substantial re-design of a video streaming system. Besides improving the efficiency of video streaming systems by minimizing the stalling event probability, its simple and channel statistics invariant nature makes the BDRA scheme an ideal backward-compatible solution for serving this purpose. Finally, the BDRA scheme has only polynomial-time computational complexity due to ordering users according to corresponding deadlines, and hence easy to execute in real-time. In the next section, we present a particular NS-3 implementation of the BDRA scheme integrated into a DASH based video streaming system to illustrate its aforementioned benefits.
\section{Implementation and Design Issues}
\label{sec:Implementation}
\subsection{Implementation}
Another important corollary of Theorem \ref{Thm: Optimality of BDRA scheme} is that the optimum scheduling policy, minimizing the segment stalling probability, allocates the time slots to the current client until all the packets in the corresponding GoP are sent. This is because, the BDRA scheme allocates the time slots to the clients in the order of upcoming deadlines, and all the packets belonging to the same GoP has the same deadline. An important practical consequence of this fact is that BDRA scheme can be implemented at the application layer completely oblivious of the operation of the lower layer protocols. The only information required by BDRA scheme when implemented at the application layer is the acknowledgment of completion of GoP, which can be effectively inferred when the client sends a new HTTP-GET message for the subsequent GoP. We note that GoP based video transmission methods are already known in the literature \cite{GoPbased}. However, in this work we show that by utilizing certain features of DASH structure an optimal GoP based policy can be constructed without using an additional feedback mechanism between the server and clients.
The operation of DASH based video streaming can be further conceptualized as follows. The client begins the streaming period by first requesting the associated MDP file. The edge server acts as a {\em web proxy} for the client, requesting the MDP file from the video content delivery server on its behalf. A copy of the received MDP file is stored at the edge server, whereas another copy is forwarded to the client. Based on the received MDP file and the estimated network throughput, the client requests the first segment among all available representations with HTTP-GET command and the video streaming from the main server starts. At the same time, the edge server observes the HTTP-GET command for the first segment and defines the deadline of the user according to initial buffer duration. The received files from the main server stored in the edge server in a sequence of GoPs via utilizing the MDP file. The edge server controls the deadlines of the existing users and executes the BDRA scheme. Whenever a client receives a GoP file successfully, it sends a HTTP-GET \footnote{The smallest size of the sub-segment requested can be equal to the size of one GoP and it is requested via HTTP partial GET command.} command for the next GoP and this command is not conveyed to the main server but utilized by the edge server to update the deadline of the client i.e., the deadline of the client is extended by the GoP duration. If the deadline of the client expires, the edge server senses that a stalling event has occurred and extends the deadline of the client by the rebuffer duration. Since the HTTP-GET commands are utilized to update deadlines, the BDRA scheme executed in the edge server does not need to trace the client buffer constantly, which is critical to reduce the feedback load between the client and the edge server. Note that the download times of each GoP, and thus, the arrival time of the next HTTP-GET command may vary with respect to the size of the GoP and the conditions of the channel between the edge server and the client.
\subsection{Design Issues}
There are three important design issues for integrating the BDRA scheme with the existing protocol stack and DASH protocol. We discuss them below, starting with the issue of secure HTTP connection requests from the clients.
\subsubsection{BDRA Scheme with Secure Video Streaming}
HTTP secure (HTTPS) is a newly emerging variant of the HTTP protocol in order to offer increased levels of privacy and security to end users on-demand \cite{Secure1}, \cite{Secure2}, \cite{Secure3}. In particular, video streaming services such as YouTube and Netflix already provide the secure end-to-end connection option through HTTPS connection. The secure connection in HTTPS is established through an authentication process in which a third party certification authority ensures the authenticity of the presented certificate by the streaming server \cite{Secure2}. One challenge that arises with HTTPS connection to implement the BDRA scheme at the access point is that only servers with a certain certificate can observe user video demands and corresponding GoP statistics \cite{Secure3}. This implies that the access point, if not equipped with the correct certificate, cannot observe the GoP deadlines for prioritizing the scheduling decisions. A potential resolution of this problem with HTTPS connections for video streaming is to notify the BDRA execution point with deadline information of the streaming video files without compromising the security of the content of these files. This can be achieved by assigning a unique identifier to each GoP and its corresponding deadline, and then relaying this information from the main server to the access point. The resource allocation mechanism at the access point utilizes only this information to allocate resources over multiple end users requesting different video files over shared communication resources.
\subsubsection{BDRA Scheme with GoP Size Adaptation and Fairness}
The main promise of the derived BDRA scheme is its ability to minimize system-wide stalling event probability. This property of the BDRA scheme holds correct independent of GoP sizes of multiple users streaming the video files simultaneously. In addition, the simple structure of the BDRA scheme can be further utilized to improve end user experience with widely varying GoP sizes. In particular, the resource allocation mechanism at the access point can minimize the total number of streaming interruptions aggregated over all end users by considering GoP sizes of the files having the same deadline. For example, one can envision a scenario in which there are three users with the same deadline having GoP sizes of $8$, $5$ and $3$ packets, and the access point is able to send $8$ packets up to the corresponding deadline. Hence, the overall system experiences a stalling event at this deadline since there will be at least one end user whose packets cannot be delivered on time. However, if we start allocating time slots to the users in the order of increasing GoP sizes, then only the user with the largest GoP size will experience a service interruption, whereas there will be two of them experiencing such an interruption in the reverse order.
In our implementation of the BDRA scheme for obtaining its performance figures, we assumed that the BDRA scheme assigns a priority to users according increasing GoP sizes if there are multiple users with the same deadline.\footnote{ We note that if the channel rates are known, then instead of ordering clients according to GoP sizes, clients with the same deadline can be ordered according to GoP transmission time.} Another variation of the BDRA implementation one can consider here is to choose a user randomly when they have the same deadline to achieve a degree of fairness among them. The probability distribution for prioritizing the end users can even depend on their respective GoP sizes to strike a balance between fairness and stalling performance. As this discussion makes it clear, the simple structure we obtained for the optimum slot-based resource allocation lends itself to various modifications, and performance enhancements depending on the objectives to optimize are possible. Since the computing capabilities of edge servers are increasing rapidly, the more complex versions of the BDRA scheme than the one considered in this paper can be implemented at the edge server with new product roll-outs.
One final remark we would like to make here is about the GoP durations of the video files being streamed. When we use the term ``GoP duration" in the paper, we refer to the time duration measured in terms of the ratio of GoP sizes to the video frame rates. To the best of our knowledge, the on-demand video streaming services keep the GoP duration fixed in practical implementations. For instance, YouTube recommends GoP-Size-to-Frame-Ratio to be $0.5$. Hence, our experiments in the paper assume an identical GoP duration for the video files being streamed, which implies that an increase in the GoP size of a video file results in a corresponding increase in the video frame rate without a change in the GoP duration.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace1_1.pdf}
\caption{Tokyo Olympics}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace2_1.pdf}
\caption{Silence of the Lambs}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace3_1.pdf}
\caption{Star Wars, $QP=10$}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace4_1.pdf}
\caption{Star Wars, $QP=16$}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace5_1.pdf}
\caption{NBC News}
\end{subfigure}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[scale=0.3]{trace6_1.pdf}
\caption{Sony Demo}
\end{subfigure}
\caption{Variation of GoP size over time.}\label{fluc}
\end{figure*}
\subsubsection{ BDRA Scheme with Redundant Chunk and Multiple Segment Requests}
An important adaptive feature of the existing DASH implementations is the redundant requests \cite{redundant}. In particular, if an end user senses that the network is lightly loaded (via bandwidth estimation), it may request a higher quality version of GoP/segment that is already buffered but not played. These redundant chunk requests are responded in a best effort way so that there is no guarantee that the user receives the higher quality version before starting to play the buffered version.
In our implementation of the BDRA scheme for simulations, we do not consider such redundant GoP requests. More specifically, these redundant chunk requests can be easily detected and discarded at the edge server since they will point to a deadline which has already been served. The reason for selecting this BDRA design is to identify the effects of network load on the frequency of stalling events experienced by the end users. However, similar to the case of GoP size adaptation and fairness above, a modified version of the BDRA scheme that does not discard redundant chunk requests can also be implemented by prioritizing such requests after serving the end users with the current deadline according to the BDRA rule. In these implementations, the end users will experience more frequent streaming interruptions at the expense of having higher streaming video quality due to elevated levels of network load.
Another notable design issue regarding the existing implementations of the DASH protocol is the ability of users to request multiple segments/chunks by means of a single range-request. There can be around $20$ GoPs in a single range-request. The structure of the derived BDRA scheme also exhibits agility against such multiple GoP requests from an implementation point of view. In particular, the sole purpose of sending a HTTP-GET request for each GoP in the derived BDRA implementation is to inform the access point about the successful delivery of the requested GoP so that the deadlines of the corresponding user can be updated accordingly. With multiple GoP requests, the access point will need to wait until the next such request before updating the deadlines of the corresponding user. To put it another way, multiple GoP requests transform the notion of ``GoP" in the derived BDRA implementation into a notion of ``super-GoP", and the reception of a super-GoP request triggers the access point to update the deadlines for the subsequent GoPs of the streamed video file.
An important remark here is the possibility of such super-GoP requests giving rise to a deterioration in the performance of the BDRA scheme to minimize the frequency of stalling events experienced by the users. The introduced BDRA implementation in this paper depends on a HTTP-GET request for each GoP, which requires a minimal modification at the client side, with a substantially improved video streaming experience in terms of the number of service interruptions. This performance boost is not available without the obtained BDRA-DASH integration, and hence super-GoP requests are beneficial in such a setting from the perspective of minimizing communication overhead between the users and the content distribution servers. However, with an BDRA scheme implementation integrated into the DASH protocol, it is an extra design problem to determine whether or not super-GoPs are still beneficial and if they are so, to decide about the number of GoPs to be included in each super-GoP request. Last but not least, we can always consider other more demanding but useful alternative implementations of the BDRA scheme in order to accommodate super-GoP requests such as having an BDRA-assistant link layer control mechanism for conveying the GoP acknowledgment messages to update the deadlines at the access point.
\section{Numerical Results}
\label{sec:numerical}
In this section, we demonstrate the performance of the BDRA scheme as compared to other blind resource allocation schemes under realistic channel and network conditions. All simulations are performed in NS-3 simulation environment. By virtue of our proof in Section \ref{sec:optimum}, we know that the BDRA scheme is the optimum algorithm in order to minimize the frequency of video streaming interruptions for cases in which only the statistical knowledge of channels is available at the server-side. From this perspective, our main intention with NS-3 simulations in this part of the paper is to illustrate the {\em optimality gap} between the BDRA algorithm and other selected rate-fair resource allocation schemes. As a result, with this intention in the paper, we only compare the performance of the BDRA algorithm with other potential resource allocation mechanisms whose operation does not require knowledge about either network throughput rates, or channel quality indicators, or detailed client operation as different from most existing work in the literature \cite{proxy1, proxy2, channelbased, Colonnese15, networkassist2, sdn1, sdn2, bufferaware, networkassist}.
Recall that our protocol and its subsequent analysis is oblivious to the operation of lower layer networking stacks, but considers only whether the video packets are delivered to the end-user by their deadlines or not. An important question arises on how the performance of this application layer protocol is affected by the operation of the lower layer protocols, i.e., specifically TCP congestion control protocol, and under general channel loss models. Hence, in our simulations, we first considered a general Markov modulated channel model with packet loss varying among the states. We also considered both an ideal cross-layer mechanism, which provides perfect and instantaneous feedback to our application layer protocol, and a realistic TCP protocol that performs retransmissions and adjusts the congestion window size based on packet losses.
We consider two different types of experimental setups. The goal of the first set of experiments is to verify the predictions of our theoretical results in Section IV by focusing on small time intervals (i.e., $20$ GoPs corresponding approximately to $10$ seconds). This first set of experiments are repeated $1000$ times with different NS-3 seeds, which corresponds to a long time interval of approximately $2.8$ hours in the ergodic limit sense. In the second type of experiments, on the other hand, we consider various video files with the number of GoPs ranging from $480$ (i.e., corresponding approximately to $4$ minutes) to $1200$ (i.e., corresponding approximately to $10$ minutes). Our experiments indicate that the video duration does not have an impact on the performance of the BDRA scheme. Hence, considering the video durations in on-demand streaming services such as YouTube as well as the observation of video duration having minimal effect on the performance of the BDRA scheme , video file durations ranging from $4$ to $10$ minutes provide substantive evidence for the performance improvements to be gained through the BDRA scheme in DASH based video streaming services. We relegate the implementation of a prototype platform with real clients dynamically joining to and leaving the system over longer time horizons on the order of weeks to a future study. As a final note, although it can be easily implemented along with the BDRA scheme, we do not consider the client-side quality selection mechanism for subsequent video segments in the simulations until subsection \ref{sec: adaptive} . That is, all subsequent segments (and sub-segments) are of the same quality in our simulations until subsection \ref{sec: adaptive}. This allows us to more clearly demonstrate the improvement in the segment stalling probability provided by the derived BDRA scheme.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{seg1.pdf}
\caption{$\rho^{-1}= 1.3$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{seg2.pdf}
\caption{$\rho^{-1}=1.35$}
\end{subfigure}
\caption{The number of segment stallings. }\label{segstall}
\end{figure*}
\subsection{Experimental Setup}
In the experiments, we use H.264/AVC video traces that are accessible on the internet \cite{trace1,trace2}. All video traces have CIF resolution (352 $\times$ 288) at $30$ frames per second, frame configuration of $1$ B frames in between I/P key pictures and GoP size of $16$ frames. The pool of videos considered in the simulations are named Tokyo Olympics, Silence of the Lambs, Star Wars IV, NBC News and Sony Demo. For each video file except Star Wars IV, we add video trace with quantization parameter (QP) of 10 and for Star Wars IV we add video traces with QP of 10 and 16.\footnote{A quantization parameter is used to determine the quantization level of transform coefficients in H.264/AVC. An increase of 1 unit in the quantization parameter means an increase of quantization step size by approximately 12 percent, which in turn means 12 percent reduction in the video-rate \cite{trace3}.} The segment size is assumed to be 10.66 seconds (i.e., 20 GoPs).
Due to the AVC encoding, although the GoP duration is fixed, GoPs in the same video file may have different sizes measured in terms of the number of bits contained in each GoP segment. Variation in the GoP sizes over time is demonstrated for the video files used in our experiments in Fig \ref{fluc}. In parallel to GoP size variation, data requirement of the client also fluctuates over time. Note that one key advantage of the GoP based BDRA scheme is that it can respond to the fluctuations in the data rate requirements from the end users. This adaptive feature of the BDRA scheme leads to more significant performance improvements over the rate-fair resource allocation schemes when high bit-rate video files are requested for streaming since the GoP size variation in high bit-rate video files is significantly higher than that in low bit-rate ones.\footnote{For instance, the variance of GoP size in Star Wars IV with QP 10 is almost three times larger than the variance of GoP size in Star Wars IV with QP 16, although the GoP size variation patterns are identical.}
The main motivation to use the CIF resolution format as opposed to using QCIF in this paper is the ease of accessing to CIF statistics for several video files through publicly available databases such as http://trace.eas.asu.edu. Since the main feature of the BDRA scheme is its ability to exploit the GoP size variation, its operation does not depend on the particular choice of the resolution format (i.e., CIF or QCIF) of the video signals. As long as our BDRA scheme is supplied with the GoP variation trends, it can exploit these variations in order to minimize the system-wide stalling event probability in a DASH-based video streaming platform.
The data transmission channel between the edge server and a user is characterized by a data rate and the error model. For the error model, we use {\em RateErrorModel} class of NS-3. In the NS-3 environment, rate error model is implemented under the transport layer. Hence, TCP packets are dropped according to an underlying probability distribution. In the literature, packet error rate (PER) is considered to be in the range of $\left[10^{-4},10^{-2}\right]$ for the TCP simulations \cite{per}. In \cite{per2}, the authors analyze the relationship between the PER and quality-of-service by using the video traces to calculate the peak signal-to-noise ratio (PSNR) of the received video files. These PSNR indicators are used to evaluate the mean opinion score (MOS). Their analysis reveals that $\left[10^{-4},10^{-3}\right]$, $\left[10^{-3},3\times10^{-3}\right]$ and $\left[3\times10^{-3},10^{-2},\right]$ correspond to the quality-of-service levels {\em excellent}, {\em good} and {\em satisfactory}, respectively. In order to conform with these existing results, we consider a Markov modulated link model, where there are three states with packet drop probabilities $\left[0.001,0.002,0.005\right]$, respectively, with each state corresponding to a different level of quality-of-service. The state transition probability matrix $\mathbf{\Gamma}$ is taken as
\[
\mathbf{\Gamma}=
\begin{bmatrix}
0.3 & 0.6 & 0.1\\
0.2 & 0.6 & 0.2 \\
0.1 & 0.6 & 0.3
\end{bmatrix}.
\]
A state transition occurs at every 0.5 seconds, and the packet loss probabilities remain constant in between state transitions.
We note that although the derived BDRA scheme will achieve similar performance gains for different physical layer telecommunication technologies (i.e., its operation is independent of the particular physical layer implementation as long as the packet losses can be modeled in a probabilistic manner at the upper layers), this particular NS-3 implementation is closer to a $4$G/$5$G scenario in which all communication resources are allocated to the user with the earliest deadline. In this scenario, the physical layer outage events due to fading at various time scales (e.g., fast and slow fading) will be observed as packet losses at higher network and transport layers. These packet losses will, in turn, determine which video files to be sent based on the updated deadlines to minimize the probability of stalling event occurrences. Further, the derived Markov modulated link model captures the time-varying nature of wireless channels in this scenario.
Let $\lambda_{i}$ (packets/sec) be the average rate of video packets generated for user $i$, which is calculated as the ratio of total size of the requested video file and the duration of the video. Also, let $r$ (bytes/sec) and $L$ (bytes/packet) be the fixed channel data rate for successful transmissions and fixed packet size, respectively. Then, the {\em inverse} utilization rate $\rho^{-1}$ is the ratio of the channel data rate and the cumulative video source rate, which is defined as,
\begin{equation}
\rho^{-1}=\frac{r}{L\sum^{n}_{i=1}\lambda_{i}}.
\end{equation}
Values of $\rho^{-1}$ close to $1$ correspond to a highly loaded network, whereas large values of $\rho^{-1}$ correspond to a lightly loaded network. In the following, we only consider an {\em underloaded} network scenario in the sense of having $\rho^{-1} > 1$ since those are the cases in which stalling events can be avoided and the efficiency of a scheduling algorithm is more clearly observed.\footnote{We use inverse utilization rates in order to obtain a parametrization for describing how heavily loaded the network in question is through numbers larger than $1$ in our simulations below.}
\subsection{Segment Stalling Probability Distribution}
In this subsection, we analyze the distribution of the stalling events per segment when the BDRA scheme and a rate-fair resource allocation (RFRA) scheme that divides available communication resources equally among the streaming clients are employed. We consider a network with six users with each one requesting a different video file, e.g., Tokyo Olympics with QP=10, Silence of the Lambs with QP=10, Star Wars IV with QP=10, NBC news with QP=10, Sony Demo with QP=10, and Star Wars IV with QP=16. The duration of the simulation is taken as 10.666 seconds, which is also the duration of a segment. The experiment is repeated for 1000 times with different random seeds for the RateErrorModel class, which corresponds to a long time interval of approximately $2.8$ hours in the ergodic limit sense.
For the purpose of discovering segment stalling probability distribution, we set the average rate of packet losses to $0.2$ for each user, which corresponds to the packet loss rate experienced at the link layer. Whenever a packet is lost, we assume that there is a perfect and instantaneous feedback sent to the transmitter. Although high for a realistic experiment, the rationale behind fixing the packet loss rate at $0.2$ in this set of experiments is to increase the channel randomness to observe a wider spectrum of stalling events, and thereby to verify our analytical results. We note that the experiments demonstrating the system performance for longer video durations approximately ranging from $4$ minutes to $10$ minutes are conducted according to the above Markovian channel model in the next subsection. The simulations are performed for $\rho^{-1}$ values of $1.3$ and $1.35$.
The results are summarized as histogram plots of the number of stalling events per segment for each $\rho^{-1}$ value in Fig. \ref{segstall}. We first note that users experience a single stalling event per segment with very high probability when the BDRA scheme is employed with $\rho^{-1} = 1.3$. However, when the RFRA scheme is used with the same inverse channel utilization rate, the users experience six stalling events per segment approximately $70\%$ of the time and they never experience less than four stalling events per segment. Secondly, when $\rho^{-1}=1.35$, i.e., the network is more lightly loaded, the performance of BDRA scheme improves further as expected. In particular, with approximately $30\%$ of the time, the users experience no stalling events, and they experience only one stalling event per segment for the rest of the time. For the same case, although the performance of the RFRA scheme also improves, it is still far away from the BDRA scheme, with users experiencing four stalling events $80\%$ of the time and five stalling events $20\%$ of the time.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{sc11.pdf}
\caption{Video duration is $4\times16/15$ minutes.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{sc12.pdf}
\caption{Video duration is $6\times16/15$ minutes.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{sc13.pdf}
\caption{Video duration is $8\times16/15$ minutes.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.45]{sc14.pdf}
\caption{Video duration is $10\times16/15$ minutes.}
\end{subfigure}
\caption{Average number of stalling events per minute versus inverse utilization rate $\rho^{-1}$.}\label{sec1}
\end{figure*}
\subsection{Average Number of Stalling Events per Minute}
In this section, we investigate the average number of stalling events per minute with respect to the network utilization rate and video duration. The average number of stalling events per minute is defined as the ratio of the total number of segment stalling events of all users and the total number of users multiplied by the video duration. In our simulations, we assume that the clients have infinite size buffers used for storing incoming video packets. Whenever a stalling event occurs, the client freezes the display of the video through a certain prescribed time duration.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc21.pdf}
\caption{Rebuffer duration is 2 seconds, and $\rho^{-1}=1$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc22.pdf}
\caption{Rebuffer duration is 2 seconds and $\rho^{-1}=1.1$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc23.pdf}
\caption{Rebuffer duration is 3 seconds and $\rho^{-1}=1$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc24.pdf}
\caption{Rebuffer duration is 3 seconds and $\rho^{-1}=1.1$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc25.pdf}
\caption{Rebuffer duration is 4 seconds and $\rho^{-1}=1$.}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[scale=0.4]{sc26.pdf}
\caption{Rebuffer duration is 4 seconds and $\rho^{-1}=1.1$.}
\end{subfigure}
\caption{Average number of stalling events per minute with respect to video duration (scaled by 15/16) and rebuffer duration.}\label{sc2}
\end{figure*}
The performance of the BDRA scheme is compared with that of the RFRA scheme which is in-cognizant of the temporal properties of the video file. In particular, the RFRA scheme allocates communication resources to users equally in a time division manner, i.e., channel resources are allocated to users sequentially until certain number of packets is transmitted.
We first set the \emph{initial buffer} duration to 0 second, which means that the client starts displaying the video immediately upon the arrival of the first video packet. The \emph{rebuffer} duration is set to be 2 seconds, which means that the client freezes the display of the video during 2 seconds upon a stalling event. There are three clients, where each requesting the video files for Tokyo Olympics $QP=10$, Star Wars IV $QP=10$, NBC News $QP=10$ respectively. We vary the length of the videos between $4\times16/15$ and $10\times16/15$ minutes (i.e., 24 to 60 segment), and for each duration we vary the value of inverse utilization rate $\rho^{-1}$ from 1 to 1.5 in steps of 0.1. Each experiment scenario is repeated for 10 times with different random seeds for the random channel loss model. The results presented are the averages of these experiments.
As illustrated in Fig. \ref{sec1}, the average number of stalling events per minute decreases and ultimately approaches to zero as the channel data rate becomes much higher than the total requested video rate (i.e., as the inverse utilization rate increases). Note that the average number of stalling events per minute with the BDRA scheme is at least $1.75$ times lower than that of the RFRA scheme, when TCP transport layer protocol is used. Also, as the inverse utilization rate increases, the average number of stalling events per minute decreases much rapidly for the BDRA scheme.
In fact, the BDRA scheme can provide service with no stallings if the inverse utilization rate is more than $1.3$ when implemented with TCP and $1.1$ when implemented with an ideal transport layer. The RFRA scheme cannot provide service with no stallings when the inverse utilization rate is less than $1.5$ for both implementations (TCP and ideal transport layer). The RFRA algorithm is much more adversely affected by the TCP implementation than the BDRA scheme, with its average stalling events per minute remain in the range of four stallings per minute even when inverse utilization rate is more than $1.5$. We also observe that the video length has almost no effect on the outcome of the experiments for both schemes.
The transport layer has significant impact on the performance. In the simulations, both algorithms are implemented at the application layer, and they wait until certain number of packets are send to switch to another user. Once there is a packet loss, the TCP time-out mechanism is provoked if an ACK is not received after \emph{Retransmission Timeout (RTO)} duration. As per the specifications given in RFC6298 \cite{rfc6298}, the minimum RTO duration is $1$ second, even though this duration can be optimized to improve efficiency \cite{optrto}. Note that during an RTO duration, no new packets are sent and the link becomes under-utilized. This affects not only the ongoing transmission, but also the subsequent GoP transmissions to other clients by limiting the amount of time that can be used to deliver video packets before their deadlines.
In the next set of simulations, we investigate the effect of rebuffer duration and video length. We implement the obtained BDRA scheme together with TCP layer only. We set the initial buffer duration to $4$ seconds, and the rebuffer durations are taken $2$, $3$, and $4$ seconds. The duration of the video is $\left\{ 4, 6, 8, 10\right\}\times16/15$ minutes. We performed the simulations for inverse utilization rates $\rho^{-1}$ of 1 and 1.1. And each experiment scenario is repeated for 10 times with different random seeds for the random channel loss model, and we then take average of them. Fig. \ref{sc2} indicates that the average number of stalling events per minute stays approximately the same with increasing video length for a given rebuffer duration and $\rho^{-1}$. We also observe that the rebuffer duration is another important factor for decreasing the number of stalling events with its impact more prominent for larger $\rho^{-1}$ values. Also note that the improvement in video stalling events is more significant when the rebuffering duration is increased from $2$ seconds to $3$ seconds, but this improvement gets smaller for higher rebuffer durations.
In Fig. \ref{sc2}, we depict the average number of stallings events per minute averaged over all clients, as well as for the client which has the highest the number of stallings among the three in the network. Although it is not identified as one of our main initial objective, we observe that the BDRA scheme performs much more fairly than the RFRA scheme in this aspect, too. The performance of the worst performing client, who requests the video stream with the highest source rate, is much closer to the average performance in the network as compared to that with the RFRA scheme.
\begin{table}
\caption{Comparison of WRFRA scheme with BDRA scheme}\label{WRR}
\centering
\begin{tabular}{|M{1cm}|M{3cm}|M{3.5cm}|}
\hline
$\rho^{-1}$ & \textbf{Resource Allocation Scheme} & \textbf{Average Number of Stallings per Minute per User} \\ \hline
\multirow{3}{*}{1.1} & BDRA(ideal) & 0.531 \\
& BDRA(Super-GoP) & 1.781 \\
& WRFRA & 6.062 \\ \hline
\multirow{3}{*}{1.2} & BDRA(ideal) & 0.0218\\
& BDRA(Super-GoP) & 0.025\\
& WRFRA & 4.312\\ \hline
\multirow{3}{*}{1.3} & BDRA(ideal) & 0\\
& BDRA(Super-GoP) & 0\\
& WRFRA & 1.937\\ \hline
\end{tabular}
\end{table}
\begin{table*}
\caption{Simulation results for quality adaptation}\label{adaptiveresults}
\centering
\begin{tabular}{|M{1cm}|M{4cm}|M{2cm}|M{3cm}|M{3cm}|}
\hline
$\rho^{-1}$ & \textbf{Resource Allocation Scheme} & \textbf{Number of Stallings per User} & \textbf{Average Quality} & \textbf{Average Quality of the Worst User} \\ \hline
\multirow{4}{*}{1} & BDRA(ideal) & 11 & 5.858 & 5.675\\
& BDRA(Super-GoP) & 14 & 5.785 & 5.666\\
& RFRA allocation & 33.666 & 5.758 & 5.358\\
& WRFRA & 34.666 & 5.733 & 5.675\\
& DWRFRA & 20.33 & 5.83 & 5.808\\ \hline
\multirow{4}{*}{1.1} & BDRA(ideal) & 5.666 & 5.941 & 5.85\\
& BDRA(Super-GoP) & 7.333 & 5.922 & 5.883\\
& RFRA & 28 & 5.791 & 5.408\\
& WRFRA& 26.666 & 5.819 & 5.775\\
& DWRFRA & 10 & 5.9 & 5.875\\ \hline
\end{tabular}
\end{table*}
\subsection{Comparison with Weighted Rate-Fair Resource Allocation}
In our simulations above, we only considered the RFRA scheme. Recall that the clients are served with equal average rates in the RFRA scheme. However, when the clients request video files with different bit-rates, their rate requirements will not be equal. Hence, the RFRA scheme does not guarantee a fairness among users in terms of the QoE. In that case, the resource allocation to users can be performed in proportion to the bit-rate requirements of the requested video files.
We call this type of resource allocation scheme {\em weighted rate-fair resource allocation} (WRFRA) . While we still expect the BDRA scheme to perform better than the WRFRA scheme since the BDRA scheme also takes bit-rate fluctuations into account, the WRFRA scheme is expected to perform better than the RFRA scheme. To test this hypothesis, we conducted the second set of simulations with WRFRA scheme for video files with duration $10 \times 16/15$ minutes (i.e., 120 segments) and initial buffer and rebuffer durations are set to $3$ and $2$ seconds respectively. We remark that average bit-rates of the requested video files are approximately proportional to the weights of $2$, $1$ and $4$. Hence, the WRFRA scheme allocates the channel resources to users according to these weights.\\
\indent In the simulations, we also implement the BDRA scheme with super-GOP approach, where the clients send HTTP-GET request for only segments (super-GOPs) instead of each GoP, in addition to ideal BDRA scheme. In the super-GOP implementation, the access point update the deadlines of the packets when there is a request for the next segment. Since the access point is not able to identify the GoPs in the requested segment in this case, we use a deadline for the segment instead of each GoP. The deadline of each segment is set to the deadline of the first GoP of the corresponding segment\\
\indent Table \ref{WRR} shows the statistics for stalling frequency both with the BDRA scheme and WRFRA scheme. It can be observed that such a weight assignment improves the performance of the RFRA scheme, but the BDRA scheme still performs much better than the WRFRA scheme, especially under heavily loaded network conditions. Indeed, our BDRA scheme can be considered as a WRFRA scheme with weights updated dynamically throughout the video streaming session in an optimal way. One can also observe that when $\rho^{-1}>1.1$, the ideal BDRA scheme and the super-GOP based BDRA scheme perform very close, however when $\rho^{-1}=1.1$ the difference between the performances of two different implementation of the BDRA scheme is more visible. Hence, this observation simply implies that the deadline information is more critical when the network is highly loaded.
\subsection{BDRA Scheme with Video Quality Adaptation}
\label{sec: adaptive}
In order to assess the performance of the BDRA scheme with adaptive bit-rate streaming, we also implemented a simple bit-rate selection mechanism along with the BDRA scheme. Our results are summarized in Table \ref{adaptiveresults}. The basic simulation set-up to obtain the results in this table is similar to those above: $3$ users request the video files Tokyo Olympics $QP = 10$, Start Wars IV $QP = 10$ and NBC News $QP = 10$ with durations of $10 \times 16 / 15$ minutes (i.e., 120 segments) and initial buffer and rebuffer durations are set to $3$ and $2$ seconds, respectively. For the video quality adaptation, we consider $L=6$ different quality levels, i.e., $\left\{1,2,3,4,5,6\right\}$, which corresponds to different resolutions, i.e., $240$p, $360$p, $480$p, $720$p, $1080$p and $1440$p. The bit-rate of the corresponding resolutions/quality levels are proportional to weights $Q_{1}=0.05$, $Q_{2}=0.08$, $Q_{3}=0.13$, $Q_{4}=0.26$, $Q_{5}=0.47$ and $Q_{6}=1$. For the video quality adaptation, we assume that the quality level $6$ indicates the same bit-rate with the original video file and quality level $l<6$ indicates the bit-rate level which is equal to the bit-rate of the original video file weighted by $Q_{l}$.\\
\indent The switching mechanism between quality levels is as follows. If a user experiences an interruption in its video streaming service while on quality level $l>1$, then the next segment is requested at the quality level $l-1$. On the other hand, if a user does not experience any stalling event in the current
segment while on quality level $l<6$ and the lead time between the last GoP delivery time and the deadline is above a threshold value (i.e., we take this threshold value as $3$ GoP durations in
our simulations), then the user requests the next segment with a higher quality level of $l+1$. We note that, although the quality level is decreased when there is a stalling event, a threshold based policy, like the one used for the quality level increment, can be employed to pro-actively reduce the quality level of the streaming video in order to prevent a possible jitter event in the future. We also want to remark that we assume the same threshold value for each quality level increment step. However, through more advanced ways of selecting threshold values (e.g., assigning higher threshold values to higher quality levels or dynamically assigning threshold values according to $\rho^{-1}$ and the number of users), the algorithm performance can even be improved further. This direction is not within the scope of this paper and will be considered as a future work. For the numerical results, we also consider a metric representing the average quality level of video streaming sessions, which is defined according to
\begin{equation}
\sum_{i=1}^{N} \sum_{l=0}^{L-1}l\frac{S_{i,l}}{N S_i},
\end{equation}
where, for $i \in \{1, \ldots, N\}$, $S_{i,l}$ and $S_{i}$ are the number of segments received by user $i$ at quality level $l\in\left\{1,\ldots,6\right\}$ and the total number of segments received by user $i$, respectively.
We consider the performance results for each value of $\rho^{-1}\in\left\{1, 1.1\right\}$. For each $\rho^{-1}$, we obtain three different performance metrics: (i) the average number of stalling events per user, (ii) the average quality level and (iii) the average quality level of the worst case user. We compare $5$ different blind resource allocation schemes. These are the BDRA scheme with GoP requests, the BDRA scheme with segment requests (i.e., super-GoP requests), RFRA, WRFRA and dynamic weighted rate-fair resource allocation (DWRFRA). The DWRFRA scheme is an extension of the WRFRA scheme where the weights dynamically change according to GoP sizes. The fundamental difference between the DWRFRA scheme and the BDRA scheme is that the BDRA scheme further utilizes the deadlines of the GoPs and prioritizes users with the small GoP sizes to minimize the number of stalling events. Performance comparison of these $5$ resource allocation algorithms is illustrated in Table \ref{adaptiveresults}.
As expected, we observe that DWRFRA outperforms the WRFRA algorithm in all performance metrics since DWRFRA changes the weights dynamically. We also observe that although the RFRA scheme achieves higher average quality level compared to the WRFRA when $\rho^{-1}=1$, the average quality of the worst case user is higher in the case of WRFRA. The root cause for this observation is that the RFRA scheme favors the users requesting low bit-rate files, which leads to a deterioration in the performance of the users requesting high bit-rate files. Hence, the observed high service quality of the RFRA scheme stems from the selection of higher quality levels for low rate videos.
In terms of the fairness, DWRFRA, WRFRA and BDRA schemes obtain a balanced average quality selection over all users (i.e., average quality level is close to the average quality level of the worst case user) without biasing any user predominantly unlike RFRA. From this perspective, the BDRA scheme can be thought to be more fair than the plain RFRA scheme. An interesting observation is that the highest average quality level of the worst case user is achieved by the DWRFRA scheme. We remark that the BDRA scheme prioritizes the GoPs with smaller sizes. Hence, the clients streaming higher bit-rate video files are more prone to stalling events. Although this design leads to unfairness between the clients in terms of the received video quality level, it still implies certain level of fairness in terms of the bandwidth usage. In addition, one can observe from Table \ref{adaptiveresults} that less number of stalling events is observed with the BDRA scheme when compared to the DWRFRA scheme by virtue of its prioritization. Hence, when the clients are streaming video files with different bit-rates, there is a trade-off between the number of system-wide stalling events and the average quality level of the worst case user.
Finally, we also observe that with increasing $\rho^{-1}$, the system performance improves faster with the BDRA scheme when compared to the RFRA and WRFRA schemes according to all three quality measures (i.e., number of stallings per user, average quality, and the average quality of the worst case-user). These benefits of the BDRA scheme are due to its two main properties. First, the BDRA scheme strives to improve the system-wide stalling frequency performance through its deadline based structure. Hence, it does not favor any one of the users requesting video streaming service since such a biasing will degrade the collective system performance. Second, our implementation of the BDRA scheme, as explained above, prioritizes the users with smaller GoP sizes. This implementation helps the BDRA scheme to prevent high bit-rate files from overwhelming the network, as observed through more balanced quality selection and service interruptions over all users.
\section{Related Work}
\label{sec:related}
\subsection{Adaptive Video and DASH}
There is a plethora of work on the adaptation of quality of video with respect to the network conditions. In particular, many prior studies investigate the use of scalable video coding (SVC) for this purpose \cite{svc1,svc2,svc3,svc4,svc5,svc6}. In almost all of these works, a rate-distortion metric is first constructed by considering the specific structure of the video. Then, a utility function, which is defined with respect to this metric, is optimized by developing different scheduling, rate allocation, and admission control policies. Although there is still an ongoing interest on the use of SVC along side DASH \cite{svcdash}, in most commercial applications this alternative is forgone, mainly because SVC requires a control mechanism at the server side that introduces an added complexity contradicting with the initial premise of the DASH structure.
In \cite{dash1,dash2,dash4}, the process of video streaming over HTTP is explained in detail and the specifications of the DASH protocol are introduced. The rate-adaptation mechanisms of commercial players are investigated in \cite{exp1}. After the standardization of DASH protocol, a great deal of attention is devoted to the client side control algorithms for bit-rate (quality) selection in order to maximize the video quality, and to minimize the fluctuations in quality and the number of stalling events \cite{csca1,csca2,csca3,csca4,csca5,csca6}.
The main purpose of all these client-centric algorithms is to adjust the video source rate of users according to available resource in order to prevent congestion and corresponding stalling events. However, in our work, we concentrate on the time period between the two bit-rate selection instants and show that it is possible to reduce the number of stalling events further via implementing an additional low-complexity scheduling algorithm at the server side.
\subsection{Client-Side Approaches}
It is reported that client-side approaches give rise to other problems due to over/under-estimation of the actual available bandwidth. In \cite{compete1} and \cite{compete2}, the authors have identified three main performance issues: (i) player instability, (ii) unfairness between the players and (iii) the under-utilization of the available bandwidth. It is pointed out that the main cause of these performance issues is the successive activity and inactivity periods, which leads to miscalculation of the available bandwidth. In line of this work, other studies have investigated the efficient and fair utilization of the available bandwidth \cite{cscax,compete3,compete4,compete5}.
In \cite{cscax}, two control mechanisms are implemented at the client side; one for controlling the playout buffer, and one for selecting the appropriate video quality level that matches the best-effort bandwidth. Furthermore, two actuators are implemented at the server; one changing the video quality, and the other throttling the video streaming rate. The fairness issues are addressed in \cite{compete3} and \cite{compete4}. In \cite{compete3}, the authors aim to adjust the rate of each segment to eliminate the off-periods (time period during which the client stops requesting a video segment). It was conjectured that the elimination of the off-periods improves the accuracy of TCP-based bandwidth estimation procedures.
In \cite{compete4},
a randomized segment scheduler is used to prevent the problem of overestimation/underestimation of the available bandwidth due to a biased view of the network state. Li {\em et al.} \cite{compete5} proposed {\em probe and adapt (PANDA)} method to improve bandwidth utilization. PANDA probes the available bandwidth via an {\em additive increase and multiplicative decrease} (AIMD) based method over a segment cycle in order to prevent misleading bandwidth estimation. Proposed algorithms in \cite{compete3,compete4,compete5} aim to increase accuracy of the resource estimation process, and unlike our method introduced in the paper, these approaches are client centric.
\subsection{Server-Side Approaches}
In \cite{proxy1}, a proxy based traffic and resource management framework for LTE networks is introduced. In this framework, a QoE optimizer is used to obtain the optimal transmission rates of the clients which maximize the aggregate video utility. These rates are utilized by the LTE scheduler as the target transmission rates. Further, the proxy may overwrite the clients' segment requests according to the feedback signals received from the QoE optimizer and the buffer levels of the clients in order to maximize the aggregate video utility. A similar framework is considered in \cite{proxy2}, but with the performance objective being a fair QoE maximization rather than being the aggregate video utility maximization as in \cite{proxy1}. Similarly, network assisted video quality assignment and bandwidth allocation schemes for improving QoE in HTTP video streaming systems are also studied in some recent papers such as \cite{channelbased, Colonnese15} and \cite{networkassist2}. The solutions in these papers require various side information to run properly such as network throughput rates, channel quality indicators, clients' instantaneous buffer states and clients' buffer occupancy trends. On the other hand, the BDRA scheme we derive in this paper operates without requiring any such information at the server-side and minimizes the system-wide stalling probability under such no-feedback conditions.
\begin{table*}[t]
\caption{Summary of related work and comparison with the derived BDRA algorithm.}
\label{Table: Literature Review}
\centering
\begin{tabular}{M{1.95cm}|M{1.95cm}|M{2.8cm}|M{1.95cm}|M{1.95cm}|M{1.95cm}|M{1.95cm}|}
\cline{2-7}
& \multicolumn{2}{c|}{\textbf{Algorithm Objective}} & \multicolumn{4}{c|}{\textbf{Required Side Information}} \\ \cline{1-7}
\multicolumn{1}{|c|}{Algorithm} & Resource Allocation & Bit-Rate Guidance & Client Buffer Status & Channel State Information & Rate-Based Utility Function & Blind (only HTTP-GET Requests) \\ \hline
\multicolumn{1}{|c|}{\cite{proxy1,proxy2,networkassist2,future}} & $\times$ & $\times$ & $\times$ & $\times$ & $\times$ & \\ \hline
\multicolumn{1}{|c|}{\cite{networkassist},\cite{sdn1}} & $\times$ & $\times$ & & & $\times$ & \\ \hline
\multicolumn{1}{|c|}{\cite{main2}} & $\times$ & & $\times$ & $\times$ & & \\ \hline
\multicolumn{1}{|c|}{\cite{bufferaware}} & $\times$ & $\times$ & $\times$ & & & \\ \hline
\multicolumn{1}{|c|}{\cite{sdn2}} & $\times$ & $\times$ & $\times$ & & $\times$ & \\ \hline
\multicolumn{1}{|c|}{\cite{channelbased}} & $\times$ & $\times$ & & $\times$ & & \\ \hline
\multicolumn{1}{|c|}{BDRA - Our Solution} & $\times$ & & & & & $\times$ \\ \hline
\end{tabular}
\end{table*}
Last but not least, there are also other recent papers that customize the software-defined-networking (SDN) architecture for the HTTP adaptive streaming scenarios in order to have a central controller helping users to select the optimal video quality level and helping the access point to share available resources among the users intelligently \cite{sdn1, sdn2, bufferaware, networkassist}. In \cite{bufferaware}, the authors propose a guidance mechanism for distributing HTTP-GET requests over time in order to improve QoE in HTTP video streaming systems. Their solution depends on the accurate estimation of data rate fluctuations over time, which can be an onerous task for when channel conditions change rapidly as in the wireless communication environments. Different from \cite{bufferaware}, our BDRA scheme is not restricted by the estimation process of randomly varying network conditions over time thanks to its blind operation. The papers \cite{sdn1,sdn2} aim to improve QoE of users via jointly optimizing the resource allocation and video quality levels on a segment scale. Different from them, we are mainly interested in a GoP level optimization that spans the time interval between two segment requests in this paper. To this end, we obtain a deadline-based rate allocation policy that can track the bit-rate fluctuations of the video files without requiring an additional feedback mechanism on top of the usual HTTP-GET requests available in a DASH based video streaming system. The solution in \cite{networkassist} is based on the existence of a network hypervisor that either determines an allocation of bandwidth slices to streaming clients or guide streaming clients in their video bit-rate selection process by using its network-wide knowledge on link capacities. Our BDRA scheme complements any bit-rate guidance mechanism in the SDN setup due to its operation blind to clients' quality adaptation mechanism. The bandwidth slicing approach in \cite{networkassist} divides the available bandwidth in proportion to the bit-rates of the requested video files, which is akin to the weighted rate-fair resource allocation (WRFRA) scheme implemented in our paper. And, we show that our BDRA scheme performs significantly better than the bandwidth slicing approach in terms of the frequency of stalling events.
\subsection{Summary of the key issues}
To conclude this part, we summarize the related work that is most relevant to our study in this paper along two main dimensions of algorithm objective and required side information in Table \ref{Table: Literature Review}. As this table makes it further clear, the BDRA algorithm allocates communication resources to multiple video streaming clients by the assistance of HTTP-GET requests only, whereas the required side information to run properly is much larger for other existing HTTP-based video streaming solutions. Hence, an emergent salient feature of our solution is its being blind to instantaneous channel conditions, clients' buffer sizes and clients' video adaptation mechanisms while scheduling the users in order to minimize the stalling probability in DASH based multiuser video streaming systems. The specific fundamental differences of the derived optimum BDRA algorithm in this paper, when compared to the previous work, can be listed as (i) its operation without requiring dedicated feedback communication, (ii) its operation that is blind to the operation of clients and their experienced channel conditions, and (iii) its operation that does not directly intervene with the quality selection process of the clients.
\section{Conclusions and Future Directions}
\label{sec:conclusion}
This work introduces a DASH compatible network assisted control mechanism to be implemented at the edge server. We have first presented the notion of optimal slot based resource allocation policy to minimize the segment stalling event probability. Then, we have analytically showed that the derived blind deadline-based rate allocation (BDRA) scheme minimizes the system-wide segment stalling probability when only average channel state information is available.
We have demonstrated the efficacy of the algorithm with a realistic NS-3 simulation depicting its performance over an ideal transport layer with perfect feedback, as well as over a more common TCP transport layer. The simulations also demonstrate that the BDRA scheme better utilizes the channel as compared to other rate-fair resource allocation schemes.\\
Note that the access point is oblivious to the instantaneous channel states in our model and the BDRA algorithm is proven to be optimal when instantaneous channel state information is not available. Although it is possible to achieve a higher network performance with an opportunistic rate allocation scheme using the instantaneous channel conditions, this will induce significant overhead and complexity in the system design. However, with the BDRA scheme, only certain features of the DASH protocol is utilized, e.g., GoP structure and HTTP-GET requests. Thus, the BDRA scheme is a blind algorithm in a sense that it is executed without provisioning the client side video bit-rate adaptations. In this network model, the BDRA scheme is totally excluded from the video quality selection procedure and only aims to minimize the number stalling events over all clients for chosen video qualities. A hybrid approach, where both client and server have control over the video quality selection, may improve the network performance.
In fact, there is a recent work that investigates a hybrid control mechanism for multi-user video streaming taking advantage of the computational efficiency of cloud computing \cite{future}. Although a joint control mechanism may improve the performance of the network, it requires a complete redesign of the DASH protocol. Furthermore, a joint control mechanism induces complexity at the server-side and requires additional feedback from the clients. In particular, server should also intervene in the video quality selection of client. An intelligent blind resource allocation algorithm that provisions the client side video bit-rate adaptations with the knowledge of DASH structure and allocates slots to users in a way that not only minimizes the segment stalling events but also forces some clients to decrease their video quality in favor of the overall network performance can be considered as an interesting extension of our work.
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The area of Macdonald polynomials and related combinatorics has been very active in the last $10$ years.
In 2006, James Haglund, Mark Haiman and Nick Loehr described a combinatorial formula
for the non-symmetric Macdonald polynomials, \cite{HaglundNonSymmetricMacdonald2008}.
This combinatorial model specializes to a model for Demazure characters,
or \emph{key polynomials} and \emph{Demazure atoms},
which is studied in \emph{e.g.} \cite{Pun2016Thesis, Mason2009}.
This model includes a \emph{basement}, a certain parameter $\sigma \in S_n$,
which can be modeled combinatorially, or via Demazure--Lusztig
operators, see \cite{Ferreira2011,Alexandersson15gbMacdonald}.
A generalization to other types using the Ram--Yip combinatorial
model \cite{RamYip2011} can be found in \cite{Feigin2016}.
The fillings we consider in this paper are the combinatorial objects that
generate the specialization $\mathrm{E}^\sigma_\alpha(\mathbf{x};q,0)$
of \emph{permuted-basement non-symmetric Macdonald polynomials} and the closely related ${\mathrm{\tilde{H}}}^\sigma_\alpha(\mathbf{x};q,0)$ which are a specialization of
\emph{permuted-basement modified Macdonald polynomials}.
Note that the non-symmetric Macdonald polynomial $\mathrm{E}_\alpha(\mathbf{x};q,t)$
specialize to the \emph{key polynomial} (or Demazure character) $\mathcal{K}_\alpha(\mathbf{x})$ at $q=t=0$,
so the specialization $\mathrm{E}_\alpha(\mathbf{x};q,0)$ can be considered as a
$q$-deformation of key polynomials.
These can be seen as a non-symmetric extension of
the modified Hall--Littlewood polynomials ${\mathrm{\tilde{H}}}_\lambda(\mathbf{x};q)$ in the following sense:
we have that $\omega \mathrm{E}_\lambda(\mathbf{x};q,0) = q^{\ast(\lambda)} {\mathrm{\tilde{H}}}_\lambda(x_1,\dotsc,x_n;q^{-1})$ whenever $\lambda$
is a partition of length $n$, and $\ast(\lambda)$ is an appropriate integer\footnote{The maximum number of inversion triples in the diagram of shape $\lambda$.}.
This identity follows from properties of LLT polynomials, see \emph{e.g.,} \cite{Haglund2005Macdonald}.
\subsection{Overview of results}
We construct maps between certain fillings that give the evaluation of non-symmetric Macdonald polynomials at
$t=0$. These maps preserve the major index statistic.
The existence of such maps are implied by relations given by Demazure operators,
but have not been constructed explicitly.
We construct such maps with particularly nice properties not implied by the operators themselves ---
in particular, we show that the maps can be made to preserve column sets.
These maps and their properties allow us to solve several problems in this area:
\begin{itemize}
\item In \cite{Mason2009}, two models for key polynomials are given, with a column-set preserving
bijection showing that these are equal.
We generalize this map to incorporate a $q$-parameter, corresponding to major index.
It is worth noting that even when $q=0$, the bijection between the sets we consider is nontrivial ---
this case was treated in \cite{Kurland2016} and the map given in the present paper specializes to the one in \cite{Kurland2016} when dealing with $q=0$.
\item We explicitly construct a biword for coinversion-free fillings
and extend the biword given for modified Macdonald polynomials in \cite{Haglund2005Macdonald}.
In the first case, the charge of the biword is shown to be equal to the major index of the filling,
while in the second case the cocharge of the biword is equal to the major index of the filling.
The extension of the second biword beyond the partition case
proves a generalization of a conjecture given in \cite{Nelsen2005}.
\item We demonstrate a bijective proof that
${\mathrm{\tilde{H}}}^\sigma_{\sigma \lambda}(\mathbf{x};q,0) = {\mathrm{\tilde{H}}}_{\lambda}(\mathbf{x};q,0)$,
for any fixed basement $\sigma \in S_n$.
In particular, we show that the bijection between the corresponding fillings can be taken to be \emph{column-set-preserving},
a property that uniquely defines this bijection.
\end{itemize}
Our proof method amounts to first constructing the maps for fillings with two rows, and then proving that
these maps are compatible with a larger filling.
The second part of this proof is partially done through computer verification,
due to the large (but finite) number of cases that needs to be considered.
\medskip
The paper is structured as follows:
In \cref{sec:preliminaries}, we introduce the necessary terminology regarding the
combinatorial model for non-symmetric Macdonald polynomials.
In \cref{sec:main} and \cref{sec:main2}, we define and prove properties of the maps.
Finally, in \cref{sec:applications}, we examine the various applications and consequences of the maps.
\section{Non-symmetric Macdonald polynomials and fillings}\label{sec:preliminaries}
In this section, we review the necessary terminology regarding the combinatorial model for non-symmetric Macdonald polynomials,
and modified Macdonald polynomials. We use the same notation as in
\cite{Alexandersson15gbMacdonald}, which differs slightly from the one used in \cite{qtCatalanBook,HaglundNonSymmetricMacdonald2008}.
In particular, we use \emph{English notation}, and not the ``skyline'' way of presenting fillings.
\medskip
An \emph{augmented diagram} of shape $\alpha$ is a Young diagram where the length of row $i$ from the top is given by $1+\alpha_i$.
The leftmost column is considered special and is referred to as the \emph{basement}.
An \emph{augmented filling} is an assignment of natural numbers to the boxes in the diagram.
We specify the entries in the basement by listing them from top to bottom --- in most places,
the basement is a permutation expressed in the one-line notation.
The \emph{weight} of a filling is the multiset of entries that are not part of the basement.
We let $\mathbf{x}^F \coloneqq \prod_{u \in F} x_u$, where $u$ ranges over all non-basement entries in $F$.
\begin{example}\label{ex:nonAttackingFilling}
Below is an augmented filling with shape $(2,4,0,3,2)$, basement given by $(4,5,1,3,2)$ and
$\mathbf{x}^F = x_1^2 x_2^2 x_3^3x_4 x_5^2$.
\[
\begin{ytableau}
\mathbf 4 & 1 & 2 \\
\mathbf 5 & 5 & 4 & 1 & 3\\
\mathbf 1 \\
\mathbf 3 & 3 & 3 & 5 \\
\mathbf 2 & 2
\end{ytableau}
\]
\end{example}
Let $F$ be an augmented filling. Two boxes $a$, $b$, are said to be \defin{attacking}
if $F(a)=F(b)$ and the boxes are either in the same column,
or they are in adjacent columns with the rightmost box in a row strictly below the other box.
A filling is \defin{non-attacking} if there are no attacking pairs of boxes.
The filling in \cref{ex:nonAttackingFilling} is non-attacking.
\subsection{Inversions, coinversions and descents}
A \defin{triple of type $A$} is an arrangement of boxes, $a$, $b$, $c$,
located such that $a$ is immediately to the left of $b$, and $c$ is somewhere below $b$,
and the row containing $a$ and $b$ is at least as long as the row containing $c$.
In a similar fashion, a \defin{triple of type $B$} is an arrangement of boxes, $a$, $b$, $c$,
located such that $a$ is immediately to the left of $b$, and $c$ is somewhere above $a$,
and the row containing $a$ and $b$ is \emph{strictly} longer than the row containing $c$.
A type $A$ triple is an \defin{inversion triple} if the entries ordered increasingly,
form a \emph{counter-clockwise} orientation. Similarly, a type $B$ triple is an inversion triple
if the entries ordered increasingly form a \emph{clockwise} orientation.
In the case of equal entries, the one with largest subscript in \cref{eq:invTriplets}
is considered to be largest.
\begin{equation}\label{eq:invTriplets}
\text{Type $A$:}\quad
\ytableausetup{centertableaux,boxsize=1.2em}
\begin{ytableau}
a_3 & b_1 \\
\none & \none[\scriptstyle\vdots] \\
\none & c_2 \\
\end{ytableau}
\qquad
\text{Type $B$:}\quad
\ytableausetup{centertableaux,boxsize=1.2em}
\begin{ytableau}
c_2 & \none \\
\none[\scriptstyle\vdots] & \none \\
a_3 & b_1 \\
\end{ytableau}
\end{equation}
A triple which is not an inversion triple is called a \emph{coinversion triple}.
\medskip
Let $F$ be an augmented filling and suppose $b$ is a non-basement box,
and $a$ is the box immediately to the left of $b$.
We say that $b$ is a \defin{descent} of $F$ if $F(a) < F(b)$.
The set of descents of $F$ is denoted $\Des(F)$.
The \defin{leg} of a box $u$, $\leg(u)$, in an augmented diagram is the number of boxes to the right of $u$.
The \defin{arm}, $\arm(u)$, of $u = (r,c)$ in an augmented diagram $\alpha$ is the total cardinality of
the sets
\begin{align*}
\{ (r', c) \in \alpha : r < r' \text{ and } \alpha_{r'} \leq \alpha_r \} \text{ and } \\
\{ (r', c-1) \in \alpha : r' < r \text{ and } \alpha_{r'} < \alpha_r \}.
\end{align*}
Given an augmented filling $F$, the \emph{major index},
$\maj(F)$, is defined as
\[
\maj(F) \coloneqq \sum_{u \in \Des(F)} \leg(u)+1,
\]
and the number of inversions, $\inv(F)$ is the number of inversion triples in $F$.
Similarly, $\coinv(F)$ is the number of coinversion triples in $F$.
\medskip
Let $\sigma \in S_n$ and let $\alpha$ be a composition with $n$ parts,
and let $\mathrm{NAF}(\alpha,\sigma)$ denote the set of non-attacking fillings of the augmented diagram of shape $\alpha$
and basement $\sigma$, with entries in $1\dotsc,n$.
The \defin{non-symmetric permuted basement Macdonald polynomial}
$\mathrm{E}^\sigma_\alpha(\mathbf{x};q,t)$ is defined as
\begin{equation}\label{eq:nonSymmetricMacdonaldBasement}
\mathrm{E}^\sigma_\alpha(\mathbf{x}; q,t) = \sum_{ F \in \mathrm{NAF}(\alpha,\sigma)} \mathbf{x}^F q^{\maj F} t^{\coinv F} \!\!\!
\prod_{ \substack{ u \in F \\ F(u_-)\neq F(u) }} \!\!\! \frac{1-t}{1-q^{1+\leg u} t^{1+\arm u}},
\end{equation}
where $u_-$ denotes the box to the left of $u$,
and we consider $F(u_-)$ not to be equal to $F(u)$ if $u$ is a box in the basement.
The ordinary non-symmetric Macdonald polynomial $\mathrm{E}_\alpha(\mathbf{x};q,t)$ considered in \cite{HaglundNonSymmetricMacdonald2008}
is recovered when taking $\sigma = w_0 = (n,\dotsc,2,1)$, that is, the unique longest permutation in $S_n$.
The basement $w_0$ is often referred to as \emph{the key basement} --- the reason will be evident further down.
\begin{example}
The set $\mathrm{NAF}(\alpha,\sigma)$ for $\alpha = (1,0,2,2)$, $\sigma=(2,1,3,4)$
consists of the following augmented fillings:
\begin{align*}
\substack{\young(21,1,332,444)\\ \coinv: 0\\ \maj: 0} \quad
\substack{\young(21,1,333,442)\\ \coinv: 0\\ \maj: 0} \quad
\substack{\young(21,1,333,444)\\ \coinv: 0\\ \maj: 0} \quad
\substack{\young(21,1,334,442)\\ \coinv: 2\\ \maj: 1}\\
\substack{\young(22,1,331,444)\\ \coinv: 1\\ \maj: 0} \quad
\substack{\young(22,1,333,441)\\ \coinv: 1\\ \maj: 0} \quad
\substack{\young(22,1,333,444)\\ \coinv: 0\\ \maj: 0} \quad
\substack{\young(22,1,334,441)\\ \coinv: 3\\ \maj: 1}
\end{align*}
\end{example}
\bigskip
Furthermore the \emph{modified Macdonald polynomials} are a class of symmetric functions,
defined\footnote{This is the same definition as in \cite{Haglund2005Macdonald},
by using the fact that ${\mathrm{\tilde{H}}}_\alpha(\mathbf{x}; q,t) = {\mathrm{\tilde{H}}}_{\alpha'}(\mathbf{x}; t,q)$.
} via
\begin{equation}\label{eq:symmetricMacdonaldDef}
{\mathrm{\tilde{H}}}_\alpha(\mathbf{x}; q,t) = \sum_{ F \in \mathrm{FIL}(\alpha,w_0)} \mathbf{x}^F q^{\maj F} t^{\inv F}
\end{equation}
where $\mathrm{FIL}(\alpha,w_0)$ is the set of fillings of \emph{partition shape} $\alpha$
with no restriction whatsoever.
The basement $w_0$ is now a \emph{big basement} --- a decreasing sequence of ``infinities'',
that is, $w_0 = (\infty_n,\infty_{n-1},\dotsc,\infty_1)$,
where we consider $\infty_i < \infty_j $ if $i<j$,
and $\infty_i>k$ for all natural numbers $i$, $k$ in the definition of major index
and inversion triples, see \cite{Haglund2005Macdonald} for details.
It is possible to generalize the modified Macdonald polynomials
to \emph{permuted basement modified Macdonald Polynomials} by considering a
big basement $\sigma$ as the sequence $(\infty_{\sigma_1},\infty_{\sigma_{2}},\dotsc,\infty_{\sigma_n})$. Then define
\begin{equation}\label{eq:symmetricMacdonaldDef2}
{\mathrm{\tilde{H}}}^\sigma_\alpha(\mathbf{x}; q,t) = \sum_{ F \in \mathrm{FIL}(\alpha,\sigma)} \mathbf{x}^F q^{\maj F} t^{\inv F}
\end{equation}
where $\mathrm{FIL}(\alpha,\sigma)$ is the set of fillings of \emph{composition shape} $\alpha$
with no restriction whatsoever.
However, one can show that these polynomials (up to a multiple of a power of $t$)
only depend on the parts of $\alpha$, see \emph{e.g.} \cref{eq:generalMacdonaldHIdentity} further down.
This fact is not easy to see from the definition here.
\subsection{Specializations of Macdonald polynomials}
The main topic of this paper is the
specialization of \eqref{eq:nonSymmetricMacdonaldBasement} and \eqref{eq:symmetricMacdonaldDef2}
at $t=0$. In the first case we have the simplified expression
\begin{equation}\label{eq:qAtoms}
\mathrm{E}^\sigma_\alpha(\mathbf{x}; q,0) = \sum_{ F \in \mathrm{CoInvFree}(\alpha,\sigma)} \mathbf{x}^F q^{\maj F}
\end{equation}
where $\mathrm{CoInvFree}(\alpha,\sigma)$ is the set of coinversion-free fillings of shape $\alpha$ and basement $\sigma$.
Note that a coinversion-free filling is also automatically non-attacking.
We define the \emph{key polynomial} $\mathcal{K}_\alpha(\mathbf{x})$ as the
specialization $\mathrm{E}^{w_0}_\alpha(\mathbf{x}; 0,0)$ --- note the use of the key basement $w_0$.
\emph{Caution:} our notation differs slightly from other literature: $\kappa_{\alpha} = \mathcal{K}_{\reverse(\alpha)}$,
where $\kappa_{\alpha}$ is the notation for key polynomials in \emph{e.g.} \cite{ReinerShimozono1995}.
Our notation fulfills the relation $\mathcal{K}_{\lambda}(x_1,\dotsc,x_n) = \mathrm{s}_{\lambda}(x_1,\dotsc,x_n)$
whenever $\lambda$ is a partition with $n$ parts.
\bigskip
Similarly, we have the simplified expression
\begin{equation}\label{eq:basementPermuted}
{\mathrm{\tilde{H}}}^\sigma_\alpha(\mathbf{x}; q,0) = \sum_{ F \in \mathrm{InvFree}(\alpha,\sigma)} \mathbf{x}^F q^{\maj F}
\end{equation}
where $\mathrm{InvFree}(\alpha,\sigma)$ is the set of inversion-free fillings of shape $\alpha$ and basement $\sigma$.
The standard \emph{modified Hall--Littlewood polynomials}, ${\mathrm{\tilde{H}}}_\alpha(\mathbf{x}; q,0)$ are recovered
by letting $\sigma=w_0$ and $\alpha$ be a partition.
\begin{figure}[!ht]
\includegraphics[width=0.9\textwidth]{polynomial-graph}
\caption{
Here is an overview of the polynomials.
Arrows down indicate the relation \emph{specializes to}, and plus-arrows down indicate \emph{expands positively into}.
Note that one relation is \cref{conj:positiveKeyExp}, which we discuss further down.
Here, $\lambda$ is a partition, and $\alpha$ is a composition.
}\label{fig:polynomialsOverview}
\end{figure}
\subsection{Brief background on Demazure operators}\label{subsec:operators}
We only use the following operators briefly in \cref{prop:bijExistence} below,
and this subsection is only to give some context.
\medskip
The non-symmetric Macdonald polynomials and the more general permuted basement Macdonald polynomials
can also be defined using Demazure--Lusztig operators,
see \emph{e.g.}, \cite{HaglundNonSymmetricMacdonald2008,Alexandersson15gbMacdonald} for details.
\medskip
Let $s_i$ act by simple transposition on the indices of the $x_i$, and define
\[
\partial_i = \frac{1-s_i}{x_i-x_{i+1}}, \quad \pi_i = \partial_i x_i, \quad \theta_i = \pi_i - 1.
\]
It is straightforward to see that $\partial_i$, $\pi_i$ and $\theta_i$ are operators on $\mathbb{C}[\mathbf{x}]$.
The operators $\pi_i$ and $\theta_i$ are used to define the key polynomials and Demazure atoms,
see \emph{e.g.} \cite{Lascoux1990Keys,Mason2009}.
We define the following $t$-deformations of the above operators.
\begin{align}
\tilde{\pi}_i(f) = (1-t)\pi_i(f) + t s_i(f) \qquad \tilde{\theta}_i(f) = (1-t)\theta_i(f) + t s_i(f).
\end{align}
The $\tilde{\theta}_i$ are the \emph{Demazure--Lusztig operators}, and generators of the affine Hecke algebra.
The $\tilde{\pi}_i$ and $\tilde{\theta}_i$ both satisfy the braid relations, and $\tilde{\pi}_i \tilde{\theta}_i = \tilde{\theta}_i\tilde{\pi}_i = t$.
These operators act on the basement of permuted basement Macdonald polynomials, as well as the indexing composition.
\section{A column-set preserving map}\label{sec:main}
By using the operators in \cref{subsec:operators} together with
the properties proved in \cite{Alexandersson15gbMacdonald},
it is possible to show the existence of a weight preserving and
major index preserving bijection between two sets of coinversion-free fillings:
\begin{proposition}\label{prop:bijExistence}
Let $\mathrm{CoInvFree}(\alpha,\sigma)$ be the set of coinversion-free fillings with shape $\alpha$ and basement $\sigma$.
Suppose $\sigma_i = \sigma_{i+1}+1$ and $\alpha_i < \alpha_{i+1}$. Then there exists a bijection
\begin{align}\label{eq:bijExistence}
\phi : \mathrm{CoInvFree}(\alpha,\sigma) \longleftrightarrow \mathrm{CoInvFree}(s_i \alpha,\sigma) \sqcup \mathrm{CoInvFree}(s_i \alpha, s_i\sigma)
\end{align}
with the property that $\phi$ preserves major index.
\end{proposition}
\begin{proof}
Using relations in \cite{Alexandersson15gbMacdonald}, we have
\begin{align}
\mathrm{E}^\sigma_{\alpha}(\mathbf{x};q,0) &= \pi_i \mathrm{E}^\sigma_{s_i \alpha}(\mathbf{x};q,0) \label{eq:shapeperm}\\
&= (1+\theta_i) \mathrm{E}^\sigma_{s_i \alpha}(\mathbf{x};q,0) \notag \\
&= \mathrm{E}^\sigma_{s_i \alpha}(\mathbf{x};q,0) + \mathrm{E}^{s_i\sigma}_{s_i \alpha}(\mathbf{x};q,0) \notag,
\end{align}
where $s_i$ acts via simple transpositions on the parts of $\alpha$.
Since $\mathrm{E}^\sigma_{\alpha}(\mathbf{x};q,0)$ is the weighted sum over the elements in
$\mathrm{CoInvFree}(\alpha,\sigma)$, and $\mathrm{E}^\sigma_{s_i \alpha}(\mathbf{x};q,0)$
and $\mathrm{E}^{s_i\sigma}_{s_i \alpha}(\mathbf{x};q,0)$
are the weighted sums over $\mathrm{CoInvFree}(s_1 \alpha,\sigma)$ and $\mathrm{CoInvFree}(s_1 \alpha,\sigma)$ respectively.
It follows that a weight-preserving bijection must exist.
\end{proof}
\bigskip
Given a filling, its \emph{column-sets} is simply the list of (multi)sets of entries in each column.
For example, the filling
\[
\young(22,1,331,444) \text{ has column sets } (\{1,2,3,4\}, \{2,3,4\}, \{1,4\}).
\]
The purpose of this paper is to explicitly construct a bijection $\phi$
with the additional property that it is \emph{column-set preserving},
that is, $F$ and $\phi(F)$ have the same column sets.
Note that it is not clear a priori that we can impose such a strong condition on $\phi$.
\medskip
In order construct such a $\phi$ --- which turns out to be unique ---
we first prove the following statement:
\begin{proposition}\label{prop:colSetPresInj}
Let $\mathrm{CoInvFree}(\alpha,\sigma)$ be the set of coinversion-free fillings with shape $\alpha$ and basement $\sigma$.
Suppose $\sigma_i >\sigma_{i+1}$ and $\alpha_i > \alpha_{i+1}$. Then there is an \emph{injection}
\[
\phi : \mathrm{CoInvFree}(\alpha,\sigma) \hookrightarrow \mathrm{CoInvFree}(s_i \alpha,\sigma) \sqcup \mathrm{CoInvFree}(s_i \alpha, s_i\sigma)
\]
with the property that $\phi$ preserves column sets and major index.
Furthermore, $\phi$ is a bijection whenever $\sigma_i=\sigma_{i+1}+1$.
\end{proposition}
The first part of this proposition together with \cref{prop:bijExistence} implies that $\phi$
is a bijection whenever $\sigma_i=\sigma_{i+1}+1$.
The proof of \cref{prop:colSetPresInj} is broken into two major parts:
we first construct $\phi$ for fillings with two rows in \cref{lem:tworowCase},
and then show that the result is compatible with the remainder of the filling
in \cref{lem:fillingcompatible}.
\begin{lemma}\label{lem:tworowCase}
Let $\mathrm{CoInvFree}(\alpha,\sigma)$ be the set of coinversion-free two-row
fillings with shape $\alpha=(\alpha_1,\alpha_2)$ and basement $\sigma=(\sigma_1, \sigma_2)$
with $\sigma_1 >\sigma_{2}$ and $\alpha_1 > \alpha_{2}$.
Then there is a column-set and major-index preserving injection $\phi$
\[
\phi : \mathrm{CoInvFree}(\alpha,\sigma) \hookrightarrow \mathrm{CoInvFree}(s_1 \alpha,\sigma) \sqcup \mathrm{CoInvFree}(s_1 \alpha, s_1\sigma).
\]
\end{lemma}
\begin{proof}
We define a \emph{filling rule} to transform a filling $F$
in $\mathrm{CoInvFree}(\alpha,\sigma)$ to some $F' \in \mathrm{CoInvFree}(s_1 \alpha,\sigma) \sqcup \mathrm{CoInvFree}(s_1 \alpha, s_1\sigma)$:
Start at the end of the first row of $F$ and map that entry to the bottom row (and corresponding column) in $F'$.
Label this entry $C$ and consider the adjacent column to the left.
If this adjacent column in $F$ has one entry, map that entry to the bottom row of $F'$ and let $C$ now denote this entry.
Otherwise, this column has entries $\{A,B\}$. If either entry is greater than $C$, choose the least
element greater than $C$ to be in the bottom row. If neither entry is greater than $C$,
choose the smallest entry to be in the bottom row. Repeat this procedure by letting $C$ be the bottom
entry in $F'$ in the column to the right of the column being processed.
It suffices to show that the resulting $F'$ is both coinversion-free and that it has the same major index as $F$.
\medskip
The first part is easy --- note that $A \neq B$ since otherwise, $F$ would contain a coinversion.
Then exactly one of the arrangements of $A$ and $B$ can produce a coinversion together with $C$,
and it is straightforward to verify that the filling rule gives the coinversion-free choice.
The main difficulty therefore is demonstrating that the filling rule is major index preserving.
We proceed by strong induction on the length of the shortest row $\alpha_2$
to demonstrate that the map is major index preserving.
For the remainder of the proof, let $<$ indicates the presence of a
descent between two entries while $\ge$ indicates the lack of such a descent.
For the base case, suppose that the shorter row has length $0$.
Suppose $F$ has the form
\begin{align*}
&a_1~c_1~c_2~\cdots~c_m\\
&b_1
\end{align*}
where $a_1$ and $b_1$ are the basement entries.
There are two separate cases to consider.
\noindent
\textbf{Base case a: $c_1>a_1$.} Since $a_1>b_1$ it follows that $F'$ has the form
\begin{align*}
&a_1\\
&b_1 < c_1~c_2~\cdots~c_m
\end{align*}
and therefore both the initial and final fillings have a descent between the
first and second columns and the major index is preserved.
\noindent
\textbf{Base case b: $c_1\le a_1$.} Then $F'$ is of the form
\begin{align*}
\begin{matrix}
a_1 \\
b_1 & \geq & c_1 & c_2 & \dotsc & c_m
\end{matrix}
\qquad
\text{ or }
\qquad
\begin{matrix}
b_1 \\
a_1 & \geq & c_1 & c_2 & \dotsc & c_m
\end{matrix}
\end{align*}
and in either case, there is no descent between the first and second column.
Hence, the major index is preserved and we have proved the base case.
\bigskip
For all the general cases below, we use filling
\begin{align*}
F \quad = \quad
\begin{matrix}
a_1 & a_2 & \dotsc & a_k & c_1 & \cdots & c_m \\
b_1 & b_2 & \dotsc & b_k
\end{matrix}
\end{align*}
to represent an $F\in \mathrm{CoInvFree}(\alpha,\sigma)$
where $a_1$, $b_1$ is the basement with $a_1>b_1$. As before, $F'$ is the filling obtained from $F$
via the filling rule. We proceed by casework based on how the filling rule was applied.
\medskip
\emph{The main outline for this casework is as follows:}
There are a few cases to consider depending on if there is some (smallest) index $\ell$, $1<\ell \leq k$
such that $a_\ell>b_\ell$, and if so, how this column appears in $F'$.
Note that we can use the induction hypothesis on all descents to the right of such a column,
and conclude that these contribute the same amount to the major index in the two fillings.
\medskip
\noindent
\textbf{Case 1:} There exist some $\ell \geq 3$, such that $a_\ell>b_\ell$,
$a_i<b_i$ for $2\le i\le \ell-1$ and $a_\ell$ appears on top of $b_\ell$ in $F'$.
Remember, $F$ is of the form
\begin{align*}
&a_1~\cdots~a_\ell~\cdots~a_k ~c_1~c_2~\cdots~c_m\\
&b_1~\cdots~b_\ell~\cdots~b_k
\end{align*}
By the inductive hypothesis, we have that the major index is preserved after the $\ell^{th}$ column.\footnote{Note that we are essentially treating the $\ell^{th}$ column as a basement in order to invoke the inductive hypothesis. This is done throughout the proof when invoking the inductive hypothesis.}
\noindent
\textbf{Subcase 1a:} Suppose $b_{\ell-1}<b_\ell$.
It follows that $a_{\ell-1}<b_{\ell-1}<b_{\ell}<a_{\ell}$.
Since $F$ has no coinversions it follows
that $b_{i+1}>a_i\ge a_{i+1}$ for $2\le i\le \ell-2$ and since $b_{i}>a_{i}$ over the same
indices it follows that $b_{i}>a_i\ge a_{i+1}$.
Therefore --- using the filling rule --- it follows that $F'$ is of the form
\begin{align*}
&\sigma(a_1,b_1)~b_2~\cdots~b_{\ell-1}< a_\ell~\cdots\\
&\sigma(a_1,b_1)~a_2~\cdots~a_{\ell-1}< b_\ell~\cdots~c_1~\cdots~c_m
\end{align*}
and the major index is clearly preserved from the second column onwards.
To prove that the entries in the first and second column preserves major index,
there are two subsubcases to consider. If $b_1\ge a_2$, then $b_2>a_1>b_1\ge a_2$ since $F$
has no coinversions. Thus,
\begin{align*}
F =
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & < & b_2 & \dotsc
\end{matrix}
\qquad
\text{and}
\qquad
F' =
\begin{matrix}
a_1 & < & b_2 & \dotsc \\
b_1 & \geq & a_2 & \dotsc
\end{matrix}
\end{align*}
Since the descent in both fillings is in the shortest row, it
is clear that in both cases it has the same contributes to the major index.
\medskip
In the second subsubcase, we have $a_2>b_1$, and since $F$ has no coinversions this implies that $b_2>a_1\ge a_2>b_1$.
Now we have
\begin{align*}
F =
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & < & b_2 & \dotsc
\end{matrix}
\qquad
\text{and}
\qquad
F' =
\begin{matrix}
b_1 & < & b_2 & \dotsc \\
a_1 & \geq & a_2 & \dotsc
\end{matrix}
\end{align*}
and major index is preserved again.
\noindent
\textbf{Subcase 1b:} Otherwise suppose $b_{\ell-1}\ge b_{\ell}$.
\noindent
\textbf{Subcase 1b.i:} If $a_{\ell-1}\ge b_{\ell}$ it follows
that $b_{\ell-1}>a_{\ell-1}\ge b_{\ell}$. Since $F$ is coinversion-free,
it follows that $b_{\ell}<a_{\ell}\le a_{\ell-1}<b_{\ell-1}$.
Since $F$ is coinversion-free it also implies that $b_{i+1}>a_i\ge a_{i+1}$ for $2\le i\le \ell-2$.
Furthermore, since $b_i>a_i$ over the same indices,
it follows that $b_i>a_i\ge a_{i+1}$ and $F'$ is of the form
\begin{align*}
&\sigma(a_1,b_1)~b_2~\cdots~b_{\ell-1}< a_\ell~\cdots\\
&\sigma(a_1,b_1)~a_2~\cdots~a_{\ell-1}< b_\ell~\cdots~c_1~c_2~\cdots~c_m.
\end{align*}
Using the same treatment as in Subcase 1a, it follows that $\maj F = \maj F'$.
\noindent
\textbf{Subcase 1b.ii:} Otherwise, $a_{\ell}>b_{\ell}>a_{\ell-1}$ and $b_{\ell-1}\ge b_{\ell}$.
\begin{itemize}
\item
If $b_{\ell-1}>b_{\ell-2}$, it follows that $b_{\ell-1}>b_{\ell-2}>a_{\ell-2}\ge a_{\ell-1}$
and $F$ has the form
\begin{align*}
&\cdots~a_{\ell-2}\ge a_{\ell-1}< a_\ell~\cdots~c_1~c_2~\cdots~c_m\\
&\cdots~b_{\ell-2}< b_{\ell-1}\ge b_\ell~\cdots
\end{align*}
while $F'$ is of the form
\begin{align*}
&\cdots~b_{\ell-2}\ge a_{\ell-1}< a_\ell~\cdots\\
&\cdots~a_{\ell-2}< b_{\ell-1}\ge b_\ell~\cdots~c_1~c_2~\cdots~c_m.
\end{align*}
The major index is preserved before the $(\ell-2)^{nd}$ column by the reasoning in Subcase 1a.
Hence, the major index is preserved on the entire filling and statement follows.
\item
Otherwise, $b_{\ell-1}\le b_{\ell-2}$.
Since $F$ is coinversion-free, it follows that $a_{\ell-1}\le a_{\ell-2}<b_{\ell-1}$.
In this case, both $F$ and $F'$ have two rows which are of the form
\begin{align*}
&\cdots~a_{\ell-2}\ge a_{\ell-1}< a_\ell~\cdots\\
&\cdots~b_{\ell-2}\ge b_{\ell-1}\ge b_\ell~\cdots
\end{align*}
Continuing left in such a manner, we either reach an index
such that $b_{j-1}<b_j$ for $3\le j\le {\ell-1}$ or there is no such index.
If $b_{j-1}<b_j$ is the greatest such $j$, it follows using logic identical
to above that $F$ has the form
\begin{align*}
&\cdots a_{j-1}\ge a_j\ge a_{j+1}~\cdots~\ge a_{\ell-1}< a_\ell~\cdots~c_1~c_2~\cdots~c_m\\
&\cdots b_{j-1} < b_j\ge b_{j+1}~\cdots~\ge b_{\ell-1}\ge b_\ell~\cdots
\end{align*}
while $F'$ has the form
\begin{align*}
&\cdots b_{j-1}\ge a_j\ge a_{j+1}~\cdots~\ge a_{\ell-1}< a_\ell~\cdots\\
&\cdots a_{j-1} < b_j\ge b_{j+1}~\cdots~\ge b_{\ell-1}\ge b_\ell~\cdots~c_1~c_2~\cdots~c_m
\end{align*}
and the major index is preserved from column $j$ through column $\ell$.
The remaining columns have major index preserved due to Subcase 1a, and thus major index is preserved overall.
If there is no such $j$, note that $b_2>a_1\ge a_2$ since $F$ is coinversion-free
with $a_1>b_1$. Therefore, $F$ is of the form
\begin{align*}
& a_1\ge a_{2}~\cdots~\ge a_{\ell-1}< a_\ell~\cdots~c_1~c_2~\cdots~c_m\\
&b_1< b_{2}~\cdots~\ge b_{\ell-1}\ge b_\ell~\cdots
\end{align*}
while $F'$ is of the form
\begin{align*}
& a_1\ge a_{2}~\cdots~\ge a_{\ell-1}< a_\ell~\cdots\\
&b_1< b_{2}~\cdots~\ge b_{\ell-1}\ge b_\ell~\cdots~c_1~c_2~\cdots~c_m
\end{align*}
and major index is preserved as the sum of leg lengths are the same.
\end{itemize}
\bigskip
\noindent
\textbf{Case 2:} There exist some $\ell \geq 3$, such that $a_\ell>b_\ell$,
$a_i<b_i$ for $2\le i\le \ell-1$ and $a_\ell$ appears below $b_\ell$ in $F'$.
\noindent
\textbf{Subcase 2a:} If $a_{\ell-1}\ge a_\ell$ then it follows that $a_{\ell-1}\ge a_{\ell}>b_\ell$.
Therefore, $F'$ is of the form
\begin{align*}
&\cdots b_{\ell-1}~b_\ell~\cdots~a_k\\
&\cdots a_{\ell-1}~a_\ell~\cdots~b_k~c_1~c_2~\cdots~c_m
\end{align*}
and the major index is preserved using Subcase 1a.
\noindent
\textbf{Subcase 2b:} Otherwise $a_{\ell}> a_{\ell-1}$.
\noindent
\textbf{Subcase 2b.i:} Suppose that $a_{\ell}>b_{\ell-1}$.
Then note that $a_{\ell}>b_{\ell-1}>a_{\ell-1}$ and $F$ of the form
\begin{align*}
&\cdots a_{\ell-1}~a_\ell~\cdots~a_k~c_1~c_2~\cdots~c_m\\
&\cdots b_{\ell-1}~b_\ell~\cdots~b_k
\end{align*}
while $F'$ is of the form
\begin{align*}
&\cdots b_{\ell-1}~b_\ell~\cdots~a_k\\
&\cdots a_{\ell-1}~a_\ell~\cdots~b_k~c_1~c_2~\cdots~c_m
\end{align*}
and thus the major index is preserved using the logic of Subcase 1a.
\noindent
\textbf{Subcase 2b.ii:} Otherwise, it follows that $b_{\ell-1}\ge a_{\ell}>b_{\ell}$.
Since $F$ is coinversion-free, it follows that $a_{\ell-1}<b_{\ell}<a_{\ell}$
and $F$ is of the form
\begin{align*}
&\cdots a_{\ell-1}< a_\ell~\cdots~a_k~c_1~c_2~\cdots~c_m\\
&\cdots b_{\ell-1}\ge b_\ell~\cdots~b_k
\end{align*}
which yields an $F'$ of the form
\begin{align*}
&\cdots a_{\ell-1}< b_\ell~\cdots~a_k\\
&\cdots b_{\ell-1}\ge a_\ell~\cdots~b_k~c_1~c_2~\cdots~c_m.
\end{align*}
The major index is preserved in this case by following Subcase 1b.ii.
\noindent
\textbf{Case 3:} Case $\ell=2$. This means $a_2>b_2$.
By inductive hypothesis, the major index is preserved from the second column onward.
We only need to verify that major index is preserved among descents between first and second column.
\noindent
\textbf{Subcase 3a:} Suppose that $b_2>a_1$ then it follows that $a_2>b_2>a_1>b_1$.
In this case, $F'$ is of the form
\begin{align*}
\begin{matrix}
a_1 & < & a_2 & \dotsc \\%& a_k \\
b_1 & < & b_2 & \dotsc
\end{matrix}
\qquad
\text{ or }
\qquad
\begin{matrix}
a_1 & < & b_2 & \dotsc \\%& a_k \\
b_1 & < & a_2 & \dotsc
\end{matrix}
\end{align*}
and
\begin{align*}
F\quad = \quad
\begin{matrix}
a_1 & < & a_2 & \dotsc \\%& a_k & c_1 & \cdots & c_m \\
b_1 & < & b_2 & \dotsc
\end{matrix}
\end{align*}
so the major index is preserved, since the descents between first and second columns appear in both fillings.
\noindent
\textbf{Subcase 3b:} Otherwise $a_1\ge b_2$ and then there are two possible cases.
\noindent
\textbf{Subcase 3b.i:} $a_2$ appear below $b_2$ in $F'$:
\begin{itemize}
\item Case $b_1\ge a_2$: It follows that we have the indicated non-descents:
\begin{align*}
F=
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F' =
\begin{matrix}
a_1 & \geq & b_2 & \dotsc \\
b_1 & \geq & a_2 & \dotsc
\end{matrix}
\end{align*}
and thus the major index is preserved.
\item Case $b_1<a_2$: With $F$ being coinversion-free, we have $a_1\ge a_2>b_2$, so
\begin{align*}
F=
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & \ast & b_2 & \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F' =
\begin{matrix}
b_1 & \ast & b_2 & \dotsc \\
a_1 & \geq & a_2 & \dotsc
\end{matrix}
\end{align*}
where $\ast$ contributes in the same way to both fillings if it is a descent.
\end{itemize}
\noindent
\textbf{Subcase 3b.ii:} $a_2$ appear above $b_2$ in $F'$:
There are again two possible cases.
\begin{itemize}
\item Case $b_1\ge b_2$: This condition together with $F$ being coinversion-free implies that
$b_2<a_2\le a_1$ and we must have
\begin{align*}
F=
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F' =
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\end{align*}
\item Case $b_1<b_2$: This gives $a_1\ge a_2>b_2>b_1$ and
it follows that
\begin{align*}
F=
\begin{matrix}
a_1 & \geq & a_2 & \dotsc \\
b_1 & < & b_2 & \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F' =
\begin{matrix}
a_1 & < & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\end{align*}
\end{itemize}
In both these subcases, major index is preserved.
\noindent
\textbf{Case 4:} The last case is occurs whenever $a_1>b_1$ and $a_i<b_i$ for $2\le i\le k$.
\noindent
\textbf{Subcase 4a:} If $c_1\le a_{k}<b_{k}$ or $a_{k}<b_{k}\le c_1$ then $F'$ is
\begin{align*}
&\cdots~b_2~\cdots b_{k}\\
&\cdots~a_2~\cdots a_{k}~c_1~c_2~\cdots~c_m\\
\end{align*}
and using the logic in Subcase 1a the result follows.
\noindent
\textbf{Subcase 4b:} Otherwise $a_{k}<c_1\le b_{k}$. Then we have
\begin{align*}
F\quad=\quad
\begin{matrix}
\dotsc & a_k & < & c_1 & \dotsc \\
\dotsc & b_k
\end{matrix}
\quad
\text{ and }
\quad
F'\quad=\quad
\begin{matrix}
\dotsc & a_k \\
\dotsc & b_k & \geq & c_1 & \dotsc
\end{matrix}
\end{align*}
Using the reasoning from Subcase 1b.ii,
there exists a $j \geq 1$ such that
\begin{align*}
F\quad=\quad
\begin{matrix}
\dotsc & a_j & \geq & a_{j+1}& \geq & \dotsc & a_k & < &c_1 & \dotsc \\
\dotsc & b_j & < & b_{j+1}& \geq & \dotsc & b_k
\end{matrix}
\end{align*}
and
\begin{align*}
F'\quad=\quad
\begin{matrix}
\dotsc & a_j & \geq & a_{j+1}& \geq & \dotsc & a_k \\
\dotsc & b_j & < & b_{j+1}& \geq & \dotsc & b_k & \geq &c_1 & \dotsc
\end{matrix}
\end{align*}
The two marked descents in $F$ has leg lengths that sum to
the leg length of the single marked descent in $F'$.
Furthermore the major index is preserved before the $j^{th}$ using the logic of Subcase 1a and the therefore the major index is preserved overall.
\medskip
All cases have now been covered, and this concludes the proof.
\end{proof}
The remainder of the proof is the verification that this two row filling rule is compatible
with the remainder of the filling ---
that is, applying the filling rule on two rows in a larger filling,
no coinversions are introduced.
Since the presence of a coinversion is a local condition,
we can reduce the proof to a finite list of possible configurations.
We verify these via computer verification and the procedure
is described in the detail in the following proof.
\begin{lemma}\label{lem:fillingcompatible}
The filling rule in \cref{lem:tworowCase} produces
fillings that are compatible with entries in a larger filling, \emph{i.e.},
no coinversions are introduced.
\end{lemma}
\begin{proof}
Since a coinversion only depends on elements in two separate rows,
it suffices to verify the algorithm for augmented fillings with
precisely three rows --- the two rows affected by the filling rule, and a fixed third row.
We may then assume that these three rows are adjacent.
\medskip
The general strategy is to consider a $2\times 3$ sub-grid of the diagram in $F$,
and then conclude that there are no coinversions in $F'$.
Furthermore, it suffices to restrict the entries between $1$ and $6$,
since only the relative order among the entries determines the presence of coinversions.
Let the two by three grid be as below
\begin{align*}
\begin{matrix}
a & d \\
b & e \\
c & f
\end{matrix}.
\end{align*}
Since inversions depend on the relative order of the row lengths,
we need to take that into consideration, by assigning each row a \emph{rank},
a number between $1$ and $3$, indicating the relative order of the row lengths.
In case two rows have the same length, the bottom row is assigned a lower rank.
The rank together with the up to six entries in the grid
allow us to determine which triples in the grid that are coinversions,
and we only consider such possible grids in $F$ which are
coinversion-free.
\medskip
Some of the entries in the grid might not be present in the case we are examining
the far right of a filling.
However, it is straightforward to see that there is no loss of generality
to assume that $a$, $b$ and $c$ are always present, as the other cases
can easily be verified by hand.
The filling rule processes the two rows being swapped from right to left,
and we need to verify that there is no coinversion in $F'$ being produced involving
the third fixed row.
Note that there might locally be \emph{two possible images}, $F'$ and $F''$
given a local grid $F$, depending on what the filling rule implied in the second column
--- in this example, the top two rows are swapped:
\begin{align}\label{eq:exampleSwapCases}
F =
\begin{matrix}
3: & a & d \\
1: & b & e \\
2: & c & f
\end{matrix}
\qquad
F' =
\begin{matrix}
1: & \ast & d \\
3: & \ast & e \\
2: & c & f
\end{matrix}
\qquad
F'' =
\begin{matrix}
1: & \ast & e \\
3: & \ast & d \\
2: & c & f
\end{matrix}
\end{align}
The entries marked $\ast$ are permutations of $a$ and $b$, and the positions
are determined by the filling rule. By construction, there are no coinversions involving only the rows being swapped,
so it suffices to check coinversions involving the third, fixed row.
\bigskip
There are exactly three things that can occur locally in the grid.
We verify this using the computer.
\noindent
\textbf{The diagram is degenerate.}
One of the rows being swapped has an element missing in the grid.
If this is the case, the position of the grid entries in $F'$ are uniquely determined
by the filling rule and there is only one case in \eqref{eq:exampleSwapCases}.
By checking all such local cases, we see that all corresponding grids are coinversion-free.
\noindent
\textbf{Both possibilities are valid.}
\item The non-fixed entries in the second column can either be swapped or not by the filling rule,
and \emph{both} these possibilities yield a coinversion-free grid using the filling rule.
This is verified by computer. As an example of this situation, we might have
\[
F =
\begin{matrix}
3: & 4 & 2 \\
1: & 1 & 1 \\
2: & 6 & 5
\end{matrix}
\qquad
F' =
\begin{matrix}
1: & 4 & 2 \\
3: & 1 & 1 \\
2: & 6 & 5
\end{matrix}
\qquad
F'' =
\begin{matrix}
1: & {1} & 1 \\
3: & {4} & {2} \\
2: & {6} & 5
\end{matrix}
\]
and in either case, the filling rule produces valid (coinversion-free) grid.
\noindent
\textbf{Only one of the grids is coinversion-free.}
This situation requires a more careful analysis,
and we need to do a non-local analysis to prove that $F'$ is indeed of the form
that produce a coinversion-free filling.
Computer check verifies that the event that only one of the two grids are valid
occurs only under the conditions in the following claim, which then determines that we are in the
case that produces a valid grid:
\begin{claim}\label{clm:badCaseFix}
Suppose we swap the longest and shortest row in the $2\times 3$ grid, as in
\begin{align}\label{eq:23diagramcases}
\begin{matrix}
{\text{swap}\left.\rule{0pt}{12pt}\right\{ \; }\\
\phantom{a}
\end{matrix}
\begin{matrix}
3: & a & d \\
1: & b & e \\
2: & c & f
\end{matrix}
\qquad
\text{ or }
\qquad
\begin{matrix}
\phantom{a} \\
{\text{swap}\left.\rule{0pt}{12pt}\right\{ \; }
\end{matrix}
\begin{matrix}
2: & a & d \\
3: & b & e \\
1: & c & f
\end{matrix}
\end{align}
and one of $e>d>f$ or $d>f>e$ or $f>e>d$ hold (down-increasing condition).
Then the corresponding grid in $F'$ \emph{must} be of the respective forms
\begin{align}\label{eq:23diagramcases2}
\begin{matrix}
1: & \ast & e \\
3: & \ast & d \\
2: & c & f
\end{matrix}
\qquad
\text{ and }
\qquad
\begin{matrix}
2: & a & d \\
1: & \ast & f \\
3: & \ast & e
\end{matrix}
\end{align}
\end{claim}
\begin{proof}
Suppose that the entries in the second column of $F'$ are not as in \cref{eq:23diagramcases2},
that is, we assume they did not ``flip''.
If the adjacent column to the right of the second column in $F$ also has all three entries present,
it follows (via computer verification) that these entries
also have the down-increasing condition.
This third column in $F$ then also appears identically in $F'$.
The down-increasing condition is therefore an invariant, present in all further columns to the right,
via induction.
Eventually, we reach the end one of the shortest row, where last complete column satisfies
the down-increasing condition and is identical in both $F$ and $F'$.
Finally, an exhaustive search on the computer shows that
it is impossible for $F$ and $F'$ to be of these specified forms and simultaneously be coinversion-free, a contradiction.
\medskip
Hence, the entries in $F'$ must be arranged as in \cref{eq:23diagramcases2}.
Computer verification on the local situation in \cref{eq:23diagramcases2} verifies that
the filling rule produces no coinversions.
\end{proof}
To conclude the proof, here is a summary of properties that needs to be verified via computer:
\begin{itemize}
\item Only the three local cases listed above appear among all possible $2\times 3$-grids;
degenerate, two valid possibilities, one valid possibility of specified form.
\item The filling rule in \cref{clm:badCaseFix} preserve the down-increasing condition.
\item Preserving the down-increasing condition eventually leads to a contradiction --- that is,
once we reach the end of the shortest row, $F$ and $F'$ cannot simultaneously be coinversion-free.
This forces $F$ to be of the expected form.
\end{itemize}
We have done this in Mathematica and Java with two different implementations.
\end{proof}
\section{Another column-set preserving map}\label{sec:main2}
It is possible to mimic the above proposition in the case
of modified Hall--Littlewood polynomials --- fillings without inversions.
\begin{proposition}
Let $\mathrm{InvFree}(\alpha,\sigma)$ be the set of inversion-free fillings with
shape $\alpha$ and big basement $\sigma$.
Suppose $\sigma_i >\sigma_{i+1}$ and $\alpha_i \ge \alpha_{i+1}$.
Then there is a bijection
\[
\varphi : \mathrm{InvFree}( \alpha,\sigma) \longleftrightarrow \mathrm{InvFree}(s_i \alpha, s_i\sigma)
\]
with the property that $\varphi$ preserves column sets and major index.
\end{proposition}
First note that the above proposition it trivial when $a_i=a_{i+1}$ as we can simply
interchange the two basement entries and preserve the remainder of the filling, since this action
does not introduce inversions and it clearly preserves the major index.
In the case when $a_i>a_{i+1}$,
we proceed as before and first establish a bijection in the two row case,
followed by proving that this is compatible with a larger filling.
\medskip
\begin{lemma}\label{lem:tworowCase2}
Let $\mathrm{InvFree}(\alpha,\sigma)$ be the set of inversion-free two-row fillings with
shape $\alpha=(\alpha_1,\alpha_2)$ and big basement $\sigma=(\sigma_1,\sigma_2)$,
with $\sigma_1 >\sigma_{2}$ and $\alpha_1 > \alpha_{2}$.
Then there is a bijection $\varphi$
\[
\varphi : \mathrm{InvFree}( \alpha,\sigma) \longleftrightarrow \mathrm{InvFree}(s_1 \alpha, s_1\sigma)
\]
with the property that $\varphi$ preserves column sets and major index.
\end{lemma}
\begin{proof}
We first consider the case when $\alpha_2=0$; in this case we can simply interchange
the rows and the result follows trivially. Otherwise for the remainder of the proof,
treat the column immediately after the basement to be the first column and note
the entries this column are in non-increasing order in order for the filling to have no inversion.
We now define a filling rule to transform a filling $F$ in $\mathrm{InvFree}(\alpha,\sigma)$
to a $F'\in \mathrm{InvFree}(s_1 \alpha,s_1\sigma)$: Start at the end of the first row
and map that entry into to the bottom row in $F'$.
Label this entry $C$ and consider the adjacent column to the left.
If this adjacent column in $F$ has one entry, map that entry to the bottom row
of $F'$ and let $C$ denote this entry.
Otherwise, the adjacent column in $F$ has entries $\{A,B\}$. If both entries are greater
than or equal to $C$ or less than $C$, map the larger entry to be in the bottom row
of $F'$. Otherwise map the lesser entry to be in the bottom row of $F'$.
Repeat this procedure with $C$ being the bottom entry in the adjacent column to
the right of the column being processed. It suffices to demonstrate that $F'$ is
inversion-free and that it has the same major index as $F$.
The first part is note that given a particular $\{A,B\}$ there is exactly one
arrangement of $A$ and $B$ that does not give a coinversion with $C$, and it
follows upon a straightforward verification that the
filling rule gives the inversion-free choice.
We now proceed by strong induction on the length of the shorter row $\alpha_2$ to
demonstrate that the filling rule is major index preserving. As in the previous
proof let $<$ indicate the presence of a descent between two entries while $\ge$
indicates the lack of a descent. For the base case, suppose that the shorter row
has length $1$. Suppose that $F$ has the form
\begin{align*}
&a_1~c_1~c_2~\dotsc~c_m\\
&b_1
\end{align*}
where $a_1$ and $b_1$ are the first entries after the basement.
Then there are two separate cases to consider.
\noindent
\textbf{Base case a:} $a_1\geq c_1$. Since $a_1\le b_1$ it follows that $F'$ has the form
\begin{align*}
&a_1\\
&b_1 \ge c_1~c_2~\dotsc~c_m\\
\end{align*}
and thus both $F$ and $F'$ lack a descent between the first and second columns and the major index is preserved.
\noindent
\textbf{Base case b:} $a_1< c_1$. The $F'$ is therefore
\begin{align*}
\begin{matrix}
a_1 \\
b_1 & < & c_1 & c_2 & \dotsc & c_m
\end{matrix}
\qquad
\text{ or }
\qquad
\begin{matrix}
b_1 \\
a_1 & < & c_1 & c_2 & \dotsc & c_m
\end{matrix}
\end{align*}
and in both cases there is a descent between the first and second column as in the original filling.
Hence the major index is preserved and the base case is proved.
\medskip
For all the cases below let original filling $F$ be represented by
\begin{align*}
F \quad = \quad
\begin{matrix}
a_1 & a_2 & \dotsc & a_k & c_1 & \dotsc & c_m \\
b_1 & b_2 & \dotsc & b_k
\end{matrix}
\end{align*}
were $a_1$ and $b_1$ are the first element \emph{after} the basement and $a_1\le b_1$.
As before, $F'$ is the filling obtained from the reverse filling rule.
We now proceed by casework on how the filling appears after the filling rule is used.
\noindent
\textbf{Case 1:} There exists $\ell\ge 3$ such that $a_\ell\le b_\ell$, $a_i>b_i$
for $2\le i\le \ell-1$ and $a_\ell$ appears on top on $b_\ell$ in $F'$.
By inductive hypothesis, the major index is preserved after the $\ell^{th}$ column.\footnote{This can be done as taking the original basement
and appending the filling as it appears at the $\ell^{th}$ column and after gives an inversion free filling. This logic is used whenever the inductive hypothesis is invoked.}
\noindent
\textbf{Subcase 1a:} Suppose that $b_{\ell-1}\ge b_\ell$.
It follows that $a_{\ell-1}>b_{\ell-1}\ge b_{\ell}\ge a_{\ell}$.
Since $F$ is inversion-free, it follows that $b_{i+1}\le a_i< a_{i+1}$ for $2\le i\le \ell-2$
and since $a_i>b_i$ over the same indices it follows that $b_i<a_i<a_{i+1}$.
Therefore, using the filling rule it follows that $F'$ is
\begin{align*}
&\sigma(a_1,b_1)~b_2~\dotsc~b_{\ell-1}\ge a_\ell~\dotsc\\
&\sigma(a_1,b_1)~a_2~\dotsc~a_{\ell-1}\ge b_\ell~\dotsc~c_1~c_2~\dotsc~c_m
\end{align*}
and the major index is clearly preserved from the second column onwards.
To prove that the major index is preserved between the first and second column,
there are two cases. If $b_1< a_2$, then $b_2\le a_1\le b_1< a_2$ as $F$ has no inversions.
Thus
\begin{align*}
F =
\begin{matrix}
a_1 & < & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\qquad
\text{and}
\qquad
F' =
\begin{matrix}
a_1 & \geq & b_2 & \dotsc \\
b_1 & < & a_2 & \dotsc
\end{matrix}
\end{align*}
Since the descent is in the longest row in both fillings, it is clear that in both cases it has the same contribution to major index.
In the second subcase, $a_2\le b_1$ and since $F$ has no inversions
it follows that $b_2\le a_1< a_2\le b_1$.
In this case it follows that we are in the situation
\begin{align*}
F =
\begin{matrix}
a_1 & < & a_2 & \dotsc \\
b_1 & \geq & b_2 & \dotsc
\end{matrix}
\qquad
\text{and}
\qquad
F' =
\begin{matrix}
b_1 & \geq & b_2 & \dotsc \\
a_1 & < & a_2 & \dotsc
\end{matrix}
\end{align*}
and major index is preserved in this case as well.
\noindent
\textbf{Subcase 1b:} Otherwise $b_{\ell-1}< b_{\ell}$.
\noindent
\textbf{Subcase 1b.i:} If $a_{\ell-1}< b_{\ell}$ then it follows $b_{\ell-1}<a_{\ell-1}<b_{\ell}$.
Furthermore since $F$ has no inversions it follows that $b_{\ell-1}<a_{\ell-1}<a_{\ell}\le b_{\ell}$.
This property also implies that $b_i<a_i<a_{i+1}$ for $2\le i\le \ell-2$ and thus $F'$ has the form
\begin{align*}
&\dotsc~b_{\ell-1}< a_\ell~\dotsc\\
&\dotsc~a_{\ell-1}< b_\ell~\dotsc~c_1~c_2~\dotsc~c_m.
\end{align*}
Using the same treatment in Subcase 1a, it follows that major index is preserved.
\noindent
\textbf{Subcase 1b.ii:} Otherwise $a_{\ell}\le b_{\ell}\le a_{\ell-1}$ and $b_{\ell-1}<b_{\ell}$.
\begin{itemize}
\item
If $b_{\ell-1}\le b_{\ell-2}$ then it follows that $b_{\ell-1}\le b_{\ell-2}<a_{\ell-2}< a_{\ell-1}$ as
there are no inversions in $F$. Then $F$ and $F'$ are of the forms
\begin{align*}
F =
\begin{matrix}
\dotsc a_{\ell-2} & < & a_{\ell-1} & \geq & a_{\ell} \dotsc \\
\dotsc b_{\ell-2} & \geq & b_{\ell-1} & < & b_{\ell} \dotsc
\end{matrix}
\end{align*}
and
\begin{align*}
F' =
\begin{matrix}
\dotsc b_{\ell-2} & < & a_{\ell-1} & \geq & a_{\ell} \dotsc \\
\dotsc a_{\ell-2} & \geq & b_{\ell-1} & < & b_{\ell} \dotsc
\end{matrix}
\end{align*}
and the major index is preserved up to the $(\ell-2)^{nd}$ column by the reasoning in Subcase 1a.
Therefore the major index is preserved on the entire filling and statement follows.
\item
Otherwise, $b_{\ell-1}> b_{\ell-2}$. Since $F$ is inversion-free it follows that $b_{\ell-1}\le a_{\ell-2}<a_{\ell-1}$. In this case, both $F$ and $F'$ have two rows which are of the form
\begin{align*}
\begin{matrix}
\dotsc a_{\ell-2} & < & a_{\ell-1} & \geq & a_{\ell} \dotsc \\
\dotsc b_{\ell-2} & < & b_{\ell-1} & < & b_{\ell} \dotsc
\end{matrix}.
\end{align*}
Continuing left, we either reach an index such that $b_{j+1}\le b_{j}$ for $2\le j\le {\ell-2}$ or there is no such index.
If $b_{j+1}\le b_j$ is the greatest such index it follows using identical reasoning as above that
\begin{align*}
\setcounter{MaxMatrixCols}{20}
F=
\begin{matrix}
\dotsc & a_j & < & a_{j+1} & \dotsc & < & a_{\ell-1} & \geq & a_{\ell} & \dotsc & c_1 & \dotsc & c_m\\
\dotsc & b_j & \geq & b_{j+1} & \dotsc & < & b_{\ell-1} & < & b_{\ell} & \dotsc
\end{matrix}
\end{align*}
while $F'$ has the form
\begin{align*}
F'=
\begin{matrix}
\dotsc & b_j & < & a_{j+1} & \dotsc & < & a_{\ell-1} & \geq & a_{\ell} & \dotsc \\
\dotsc & a_j & \geq & b_{j+1} & \dotsc & < & b_{\ell-1} & < & b_{\ell} & \dotsc & c_1 & \dotsc & c_m
\end{matrix}
\end{align*}
and the major index is preserved between from column $j$ and column $\ell$.
The remaining columns have major index preserved by the logic of Subcase 1a, and
thus the major index is preserved overall. If there is no such $j$,
note that $a_2>b_2$ and $b_1\ge a_1$ and since $F$ is inversion-free it
follows that $b_2\le a_1<a_2$. Therefore $F$ is of the form
\begin{align*}
\setcounter{MaxMatrixCols}{20}
F=
\begin{matrix}
a_1 & < & a_{2} & \dotsc & < & a_{\ell-1} & \geq & a_{\ell} & \dotsc & c_1 & \dotsc & c_m\\
b_1 & \geq & b_{2} & \dotsc & \geq & b_{\ell-1} & < & b_{\ell} & \dotsc
\end{matrix}
\end{align*}
while $F'$ has the form
\begin{align*}
F'=
\begin{matrix}
a_1 & < & a_{2} & \dotsc & < & a_{\ell-1} & \geq & a_{\ell} & \dotsc \\
b_1 & \geq & b_{2} & \dotsc & \geq & b_{\ell-1} & < & b_{\ell} & \dotsc & c_1 & \dotsc & c_m
\end{matrix}
\end{align*}
and major index is preserved as the sum of leg lengths are the same.
\end{itemize}
\noindent
\textbf{Case 2:} There exist $3\le \ell\le k$ such that $a_\ell\le b_\ell$, $a_i>b_i$ for $2\le i\le \ell-1$ and $a_{\ell}$ is below $b_{\ell}$ in $F'$.
\noindent
\textbf{Subcase 2a:} If $a_{\ell-1}< a_\ell$ then it follows that $b_{\ell-1}<a_{\ell-1}< a_{\ell}\le b_\ell$.
Thus $F'$ has the form
\begin{align*}
&\dotsc b_{\ell-1}~b_\ell~\dotsc~a_k\\
&\dotsc a_{\ell-1}~a_\ell~\dotsc~b_k~c_1~c_2~\dotsc~c_m
\end{align*}
and the major index is preserved using Subcase 1a.
\noindent
\textbf{Subcase 2b:} Otherwise $a_{\ell}\le a_{\ell-1}$.
\noindent
\textbf{Subcase 2b.i:} Suppose $a_{\ell}\le b_{\ell-1}$.
Since $a_{\ell-1}>b_{\ell-1}$ it follows that $a_{\ell-1}>b_{\ell-1}\ge a_{\ell}$.
Therefore, it follows that
\begin{align*}
F=
\begin{matrix}
\dotsc a_{\ell-1} & a_{\ell} & \dotsc & c_1 \dotsc \\
\dotsc b_{\ell-1} & b_{\ell} & \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
\dotsc b_{\ell-1} & b_{\ell} & \dotsc \\
\dotsc a_{\ell-1} & a_{\ell} & \dotsc & c_1 \dotsc
\end{matrix}
\end{align*}
and the major index is preserved using the reasoning in Subcase 1a.
\noindent
\textbf{Subcase 2b.ii:} Otherwise $b_{\ell-1}< a_{\ell}\le b_{\ell}$, $a_{\ell-1}\ge a_{\ell}$.
Furthermore $a_{\ell-1}\ge b_{\ell}\ge a_{\ell}$ as $F$ has no
inversions and it follows that the $F$ and $F'$ have the form
\begin{align*}
F=
\begin{matrix}
\dotsc a_{\ell-1}&\geq & a_{\ell} \dotsc a_k & c_1 \dotsc \\
\dotsc b_{\ell-1}& < & b_{\ell} \dotsc b_k
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
\dotsc a_{\ell-1}& \geq & b_{\ell} \dotsc a_k \\
\dotsc b_{\ell-1}& < & a_{\ell} \dotsc b_k & c_1 \dotsc
\end{matrix}
\end{align*}
The major index is preserved using the reasoning in Subcase 1b.ii
\noindent
\textbf{Case 3:} Case $\ell=2$. In this case $a_2\le b_2$ and by inductive hypothesis,
the major index is preserved from the second column onward.
We need therefore simply to verify that the major index is preserved among descent between the first and second column.
\noindent
\textbf{Subcase 3a:} Suppose that $b_2\le a_1$ then it follows that $a_2\le b_2\le a_1\le b_1$.
In this case, we have
\begin{align*}
\setcounter{MaxMatrixCols}{20}
F=
\begin{matrix}
a_1 & \geq & a_{2} \dotsc a_{k} & c_1 \dotsc c_m\\
b_1 & \geq & b_{2} \dotsc b_{k}
\end{matrix}
\end{align*}
while $F'$ is in one of the forms
\begin{align*}
\begin{matrix}
a_1 & \geq & a_{2} \dotsc a_{k} \\
b_1 & \geq & b_{2} \dotsc b_{k} & c_1 \dotsc c_m
\end{matrix}
\quad
\text{ or }
\quad
\begin{matrix}
a_1 & \geq & b_{2} \dotsc a_{k} \\
b_1 & \geq & a_{2} \dotsc b_{k} & c_1 \dotsc c_m\\
\end{matrix}
\end{align*}
The major index is preserved between the first and second columns since
descents are not present in either possible position.
\noindent
\textbf{Subcase 3b:} Otherwise $b_2>a_1$ and then there are two possible cases.
\noindent
\textbf{Subcase 3b.i:} $a_2$ appears below $b_2$ in $F'$:
\begin{itemize}
\item
Case $b_1< a_2$, where it follows that $a_1\le b_1<a_2\le b_2$ so
\begin{align*}
F=
\begin{matrix}
a_1 & < & a_{2} \dotsc & c_1 \dotsc c_m \\
b_1 & < & b_{2} \dotsc
\end{matrix}
\end{align*}
and
\begin{align*}
\quad
F'=
\begin{matrix}
a_1 & < & b_{2} \dotsc \\
b_1 & < & a_{2} \dotsc & c_1 \dotsc c_m \\
\end{matrix}
\end{align*}
and major index is preserved.
\item
Case $b_1\ge a_2$. With $F$ being inversion-free and $b_1\ge a_2$, $a_2\le b_2$, $a_1<b_2$
it follows that $a_1<a_2\le b_2$. Therefore,
\begin{align*}
F=
\begin{matrix}
a_1 & a_{2} \dotsc & c_1 \dotsc c_m \\
b_1 & b_{2} \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
b_1 & b_{2} \dotsc \\
a_1 & a_{2} \dotsc & c_1 \dotsc c_m
\end{matrix}
\end{align*}
and major index is preserved.
\end{itemize}
\noindent
\textbf{Subcase 3b.ii:} $a_2$ appears above $b_2$ in $F'$:
\begin{itemize}
\item
Case $b_1\ge b_2$. It follows that $b_1\ge b_2>a_1$ and $F$ has no
inversions it follows that $a_1<a_2\le b_2\le b_1$.
Thus, the situation is
\begin{align*}
F=
\begin{matrix}
a_1 & < & a_{2} \dotsc & c_1 \dotsc c_m \\
b_1 & \geq & b_{2} \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
b_1 & \geq & a_{2} \dotsc \\
a_1 & < & b_{2} \dotsc & c_1 \dotsc c_m
\end{matrix}
\end{align*}
and major index is preserved.
\item
Case $b_1< b_2$. Since $b_2>a_1$ and $F$ has no inversions it follows that $a_1<a_2\le b_2$.
Furthermore $a_1\le b_1$ and we have
\begin{align*}
F=
\begin{matrix}
a_1 & < & a_{2} \dotsc & c_1 \dotsc c_m \\
b_1 & < & b_{2} \dotsc
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
a_1 & < & a_{2} \dotsc \\
b_1 & < & b_{2} \dotsc & c_1 \dotsc c_m
\end{matrix}
\end{align*}
and major index is preserved.
\end{itemize}
\noindent
\textbf{Case 4:} The final case occurs if $a_1\leq b_1$ and $a_i>b_i$ for $2\leq i\leq k$.
\noindent
\textbf{Subcase 4a:} If $c_1> a_{k}>b_{k}$ or $a_{k}>b_{k}\ge c_1$ then $F'$ is
\begin{align*}
&\dotsc~b_2~\dotsc b_{k}\\
&\dotsc~a_2~\dotsc a_{k}~c_1~c_2~\dotsc~c_m
\end{align*}
using the logic of Subcase 1a and the major index is preserved similarly to Subcase 1a.
\noindent
\textbf{Subcase 4b:} Otherwise $a_{\ell}\ge c_1> b_{\ell}$.
We have
\begin{align*}
F=
\begin{matrix}
\dotsc & a_{\ell} & \geq & c_1 \dotsc c_m \\
\dotsc & b_{\ell}
\end{matrix}
\quad
\text{ and }
\quad
F'=
\begin{matrix}
\dotsc & a_{\ell} \\
\dotsc & b_{\ell} & > & c_1 \dotsc c_m
\end{matrix}
\end{align*}
Using the reasoning from Subcase 1b.ii it follows that there exists $j\ge 1$ such that
\begin{align*}
F=
\begin{matrix}
\dotsc a_j & < &a_{j+1}\dotsc< & a_{\ell}& \geq & c_1 \dotsc c_m \\
\dotsc b_j & \geq &b_{j+1}\dotsc< & b_{\ell}
\end{matrix}
\end{align*}
and
\begin{align*}
F'=
\begin{matrix}
\dotsc b_j & < &a_{j+1}\dotsc < & a_{\ell} \\
\dotsc a_j & \geq &b_{j+1}\dotsc < & b_{\ell}& < & c_1 \dotsc c_m
\end{matrix}
\end{align*}
The marked descents in $F$ and $F'$ have leg lengths that sum to the same value so
the major index is preserved after column $j$ while the major index is preserved
from column $1$ to column $j$ using Subcase 1a.
We have now verified all possible cases and the result that the filling rule is
major index preserving follows via strong induction on the length of the shorter row.
To demonstrate that is is a bijection,
we refer to Theorem\footnote{Theorem 5.1.1 only treats unaugmented fillings. However, since we fix a basement
on both sides that cannot introduce inversions, the result follows.} 5.1.1 in \cite{Haglund2005Macdonald}, with $t=0$.
It follows that the two sets of fillings $\varphi$ map between are equinumerous.
This together with the fact that $\varphi$ is injective implies that $\varphi$ is also bijective.
\end{proof}
Analogous to \cref{prop:colSetPresInj}, it remains to show that $\varphi$ is compatible
with a larger filling. We proceed as before, and reduce this to a finite set of
verifications. The approach is similar to the case when we treated coinversion-free fillings,
and the main difference is the details in \cref{clm:badCaseFix2}.
\begin{lemma}\label{lem:fillingcompatible2}
The map $\varphi$ applied to two adjacent rows in a larger filling
does not introduce any inversions.
\end{lemma}
\begin{proof}
As before, it suffices to consider the three-row case, where $\varphi$
is applied to two adjacent rows and the third row is fixed.
Again, we consider a $2\times 3$-grid with entries in $1,\dotsc,6$,
and all three entries in the first column are present.
There are again exactly three things that can occur locally in the grid.
We verify this using the computer.
\noindent
\textbf{The diagram is degenerate.}
One of the rows being swapped has an element missing in the grid
--- the entries in $F'$ are uniquely determined.
By checking all such local cases, we see that all corresponding grids are inversion-free.
\noindent
\textbf{Both possibilities are valid.}
\item The non-fixed entries in the second column can either be swapped or not by the filling rule,
and \emph{both} these possibilities yield an inversion-free grid using the filling rule.
This is verified by computer. As an example of this situation, we might have
\[
F =
\begin{matrix}
3: & 3 & 4 \\
1: & 5 & 6 \\
2: & 1 & 2
\end{matrix}
\qquad
F' =
\begin{matrix}
1: & 3 & 4 \\
3: & 5 & 6 \\
2: & 1 & 2
\end{matrix}
\qquad
F'' =
\begin{matrix}
1: & {5} & 6 \\
3: & {3} & {4} \\
2: & {1} & 2
\end{matrix}
\]
and in either case, the filling rule produces valid (inversion-free) grid.
\noindent
\textbf{Only one of the grids is inversion-free.}
This situation requires a more careful analysis,
and we need to do a non-local analysis to prove that $F'$ is indeed of the form
that produce a inversion-free filling.
Computer check verifies that the event that only one of the two grids are valid
occurs only under the conditions in the following claim, which then determines that we are in the
case that produces a valid grid:
\begin{claim}\label{clm:badCaseFix2}
Suppose we swap the longest and shortest row in the $2\times 3$ grid, as in
\begin{align
\begin{matrix}
{\text{swap}\left.\rule{0pt}{12pt}\right\{ \; }\\
\phantom{a}
\end{matrix}
\begin{matrix}
3: & a & d \\
1: & b & e \\
2: & c & f
\end{matrix}
\qquad
\text{ or }
\qquad
\begin{matrix}
\phantom{a} \\
{\text{swap}\left.\rule{0pt}{12pt}\right\{ \; }
\end{matrix}
\begin{matrix}
2: & a & d \\
3: & b & e \\
1: & c & f
\end{matrix}
\end{align}
Furthermore, suppose that one of ${e<d<f}$ or $d<f<e$
or $f<e<d$ or $d=f$ and ${e\neq f}$ hold ({up-increasing condition}).
Then the corresponding grid in $F'$ \emph{must} be of the respective forms
\begin{align}\label{eq:23diagramcases32}
\begin{matrix}
1: & \ast & e \\
3: & \ast & d \\
2: & c & f
\end{matrix}
\qquad
\text{ and }
\qquad
\begin{matrix}
2: & a & d \\
1: & \ast & f \\
3: & \ast & e
\end{matrix}
\end{align}
\end{claim}
\begin{proof}
Suppose that the entries in the second column of $F'$ are not as in \cref{eq:23diagramcases32},
that is, we assume they did not ``flip''.
If the adjacent column to the right of the second column in $F$ also has all three entries present,
it follows (via computer verification) that these entries
also have the up-increasing condition.
This third column in $F$ then also appears identically in $F'$.
The up-increasing condition is therefore an invariant, present in all further columns to the right,
via induction.
Eventually, we reach the end one of the shortest row, where last complete column satisfies
the up-increasing condition and is identical in both $F$ and $F'$.
Finally, an exhaustive search on the computer shows that
it is impossible for $F$ and $F'$ to be of these specified forms and simultaneously be inversion-free.
\medskip
Hence, the entries in $F'$ must be arranged as in \cref{eq:23diagramcases32}.
Computer verification on the local situation in \cref{eq:23diagramcases32} verifies that
the filling rule produces no inversions.
\end{proof}
Thus, $\varphi$ is an injection and together with Theorem 5.1.1 in \cite{Haglund2005Macdonald},
it is also a bijection. It is also possible to prove that this is a bijection by
a computer-style method similar to what is done above.
\end{proof}
As a final comment on this section, note that the set $\mathrm{InvFree}(\alpha,\sigma)$ (with big basement) is empty
if there are $i<j$ such that $\alpha_i < \alpha_j$ and $\sigma_i < \sigma_j$.
This is due to the fact that the first entry in row $j$,
and the two basement entries $\sigma_i$ and $\sigma_j$ form an inversion triple.
\section{Applications}\label{sec:applications}
Despite the relatively lengthy proofs given, the realization the fact that these
column set preserving maps exist allow for a multitude of short corollaries.
We begin with a technical result that has been partially proved in a variety of special cases.
\begin{proposition}\label{prop:atMostOneCoInvFree}
Given fixed column sets, there exists at most one coinversion-free
filling with shape $\alpha$ and decreasing basement.
\end{proposition}
\begin{proof}
To prove that there is at most one coinversion free filling
with shape $\alpha$ and decreasing basement, simply apply
the column set preserving operator $\phi$ until the shape becomes a partition.
In the partition case, any list of column sets with sizes compatible with $\alpha$
admits a unique coinversion-free filling. This unique filling can be constructed from the column sets
via the following iterative process going column by column, and in each column, top to bottom:
Given an entry $e$ in column $i$, the adjacent entry in column $i+1$
is found by taking the largest unused entry in column set $i+1$ less or equal to $e$,
and if there is no such element, simply take the largest entry.
It is easy to verify that this gives a coinversion-free filling and given $\mathrm{E}_{\lambda}(\mathbf{x};1,0)=e_{\lambda'}(\mathbf{x})$ the result follows.
\end{proof}
\medskip
Given a coinversion-free filling $F$, we construct the \emph{biword} of $F$ as follows:
Let the top row be the non-basement entries of $F$ listed in increasing order
and the bottom row be the corresponding columns the entries belong to,
listed in increasing order in case of a tie in the first row, see \cref{eq:biworexample} for an example.
\begin{align}\label{eq:biworexample}
\begin{ytableau}
\mathbf{4} &3&3&2&1&3\\
\mathbf{3} &2&2&1\\
\mathbf{2} &1&4\\
\mathbf{1}
\end{ytableau}
\qquad \longrightarrow \qquad
\begin{pmatrix}
1&1&1&2&2&2&3&3&3&4 \\
4&3&1&3&2&1&5&2&1&2
\end{pmatrix}
\end{align}
We let $\cw(F)$ denote the lower row in this biword, the \emph{charge word} of $F$.
Finally, define the charge of $F$ as $\charge(\cw(F))$,
where $\charge(\cdot)$ is defined as in \emph{e.g.}, \cite{qtCatalanBook},
by decomposing $\cw(F)$ into \emph{standard subwords},
computing $\charge(w)\coloneqq \maj(\reverse(w^{-1}))$ of each such subword $w$ (interpreted as a permutation), followed by adding the results.
The standard subwords are extracted iteratively by finding the rightmost occurrence of the
smallest element, then scanning right to left for the next smallest element,
looping around it necessary, and then repeating this process
until one has found an occurrence of largest element to the word.
These letters form the first subword
and this process is repeated on the remaining letters until there are no more letters left.
For example, the word $1322133241214$ has the subword decomposition
$w_1=3214$, $w_2=3241$, $w_3=321$, $w_4=12$, extracted as
\begin{align*}
132 213 \bar{3} 241 \bar{2} \bar{1}\bar{4} &\to 3214\quad (1)\\
132 21 \phantom{3}\bar{3} \bar{2}\bar{4}\bar{1}\phantom{2}\phantom{1}\phantom{4} &\to 3241 \quad (1) \\
1\bar{3}2 \bar{2}\bar{1} \phantom{3}\phantom{3} \phantom{2}\phantom{4}\phantom{1}\phantom{2}\phantom{1}\phantom{4} &\to 321 \quad (0) \\
\bar{1}\phantom{3}\bar{2} \phantom{2}\phantom{1} \phantom{3}\phantom{3} \phantom{2}\phantom{4}\phantom{1}\phantom{2}\phantom{1}\phantom{4} &\to 12 \quad (1)
\end{align*}
The charge for the permutations is displayed to the right, so the total charge of the word is $3$.
\medskip
\begin{theorem}\label{thm:composition-charge}
Let $F$ be a coinversion-free filling of shape $\alpha$ with the key basement $w_0$.
Then $\charge(\cw(F))=\maj(F)$.
\end{theorem}
\begin{proof}
Note that $\cw(F)$ is uniquely determined by the column sets of $F$ by construction.
Therefore, after applying $\phi$ repeatedly, it suffices to prove the
theorem when $F$ is of partition shape, and from hereon, $F$ is assumed to be of partition shape.
Under this assumption, $F$ can be constructed from the column sets
via the following iterative process given in \cref{prop:atMostOneCoInvFree}.
\medskip
\textbf{Claim:} The first column (the column adjacent to the basement) of $F$ is strictly decreasing.
Suppose the largest entry in the basement (and thus in the filling) is $n$.
Note that the $i^{th}$ largest entry in the first column is at most $n+1-i$,
since all entries in a fixed column are distinct.
Furthermore, since the $i^{th}$ largest element in the basement is exactly $n+1-i$,
the previous construction rule implies that the first column is strictly decreasing.
\medskip
Now note that the subwords obtained in the word decomposition of $\cw(F)$ naturally correspond to rows of $F$:
The claim ensures that the largest entry in the first column is in the topmost row.
It follows that entry is also the first entry of the first subword of $\cw(F)$.
Now, by the iterative process on how to recover $F$ from its column sets,
it is straightforward to see that the first subword of $\cw(F)$ corresponds to the first row of $F$.
By using the fact that the first column is decreasing,
this argument can be repeated for the remaining rows to
prove that subword $j$ of $\cw(F)$ corresponds to row $j$ of $F$.
Finally, since the inverse of the $j^{th}$ subword $w_j$ can be seen to have the same relative
order as the entries in row $j$, it follows that $\maj(\reverse(w_j^{-1}))$
is equal to the contribution of row $j$ to $\maj(F)$.
Hence, $\maj(F)$ is equal to $\charge(\cw(F))$.
\end{proof}
Note that \cref{thm:composition-charge} critically relies on the presence of a decreasing basement,
and the equality does not hold for non-decreasing basements.
\medskip
We are now prepared to give an analog of the famous cocharge formula
for partition case of the non-symmetric Macdonald polynomials.
Note that the following identity can also be proven using properties of
LLT polynomials proved in \cite{Haglund2005Macdonald} along with the symmetric cocharge formula,
however this proof is far more in the style of the proof of the original cocharge formula given
in \cite{Haglund2005Macdonald}. This particular version also has the advantage
that seems to it generalizes to a non-symmetric setting, see \cref{conj:positiveKeyExp} below.
\begin{theorem}
Let $\lambda$ be a partition.
Then
\[
\mathrm{E}_{\lambda}(\mathbf{x};q,0)=\sum_{\mu}\mathrm{s}_{\mu}(\mathbf{x})\sum_{P \in \mathrm{SSYT}(\mu',\lambda')}q^{\charge(P)}.
\]
\end{theorem}
\begin{proof}
The partition case of \cref{prop:atMostOneCoInvFree}
implies that every possible second row (with strictly decreasing blocks in the second row),
in a biword can be obtained from some coinversion-free filling of shape $\lambda$.
By performing RSK on such biwords, every possible pair $(P,Q)$ of SSYT's
with the insertion tableau $P$ having shape $\mu$
and the recording tableau $Q$ having shape $\mu'$ appears.
The proof of this is identical to the one given for the traditional RSK in \cite{StanleyEC2}.
Since charge is Knuth-invariant, two coinversion-free fillings $F_1$ and $F_2$ with identical
insertion tableau $P$ also have the same charge, and
\cref{thm:composition-charge} implies that $\maj(F_1)=\maj(F_2)=\charge(P)$.
It now follows that
\begin{align*}
\mathrm{E}_{\lambda}(\mathbf{x};q,0) &=\sum_{F\in \mathrm{CoInvFree}(\lambda,w_0)} q^{\maj(F)}\mathbf{x}^F
= \sum_{F\in \mathrm{CoInvFree}(\lambda,w_0)} q^{\charge(c(F))} \mathbf{x}^F \\
&\stackrel{RSK}{=} \sum_{P\in \mathrm{SSYT}(\mu',\lambda')} q^{\charge(P)} \sum_{Q\in \mathrm{SSYT}(\mu)} \mathbf{x}^{Q} \\
&= \sum_{\mu} \mathrm{s}_{\mu}(\mathbf{x}) \sum_{P\in \mathrm{SSYT}(\mu',\lambda')} q^{\charge(P)}
\end{align*}
as desired.
\end{proof}
We believe that the above theorem carries over in a more general setting:
\begin{conjecture}\label{conj:positiveKeyExp}
Let $\alpha$ be a composition. Then
\[
\mathrm{E}_\alpha(\mathbf{x};q,0) = \sum_{ F \in \mathrm{CoInvFree}(\alpha,w_0) } q^{\maj F}\mathbf{x}^F
= \sum_{P \in \mathrm{SSYT}(\mu',\lambda')} q^{\charge(P)} \mathcal{K}_{\gamma(P,\alpha)}(\mathbf{x})
\]
where $\lambda = \sort(\alpha)$ and $\gamma(P,\alpha)$ is a
composition\footnote{That rearranges to the conjugate shape of $P$.} determined by $\alpha$ and $P$.
In particular, if $S_Q$ is the subset of fillings in $\mathrm{CoInvFree}(\alpha,w_0)$
with recording tableau $Q \in \mathrm{SSYT}(\mu',\lambda')$,
then $\sum_{F \in S_Q} \mathbf{x}^F$ is a key polynomial.
\end{conjecture}
Note that a recent combinatorial proof of the key expansion of $\mathrm{E}_\alpha(\mathbf{x};q,0)$
using \emph{weak dual equivalence} was given in \cite{AssafKostka}.
\medskip
The following example illustrates \cref{conj:positiveKeyExp}.
\begin{example}
We have the following expansions in the key basis:
\begin{align*}
\mathrm{E}_{320}(\mathbf{x};q,0) &= \mathcal{K}_{320} + q \mathcal{K}_{311} + q \mathcal{K}_{221} + q^2 \mathcal{K}_{221} \\
\mathrm{E}_{203}(\mathbf{x};q,0) &= \mathcal{K}_{203} + q \mathcal{K}_{113} + q \mathcal{K}_{212} + q^2 \mathcal{K}_{221} \\
\mathrm{E}_{023}(\mathbf{x};q,0) &= \mathcal{K}_{023} + q \mathcal{K}_{113} + q \mathcal{K}_{122} + q^2 \mathcal{K}_{221} \\
\end{align*}
The corresponding recording tableaux are
\[
\young(11,22,3),\quad
\young(112,2,3),\quad
\young(113,22),\quad
\young(112,23).
\]
Observe that all coefficients are identical, and only the
composition indexing the key polynomials differ by some permutation.
Furthermore, several different permutations of the same partition might be present,
as with $\mathcal{K}_{212}$ and $\mathcal{K}_{221}$ above.
This refines the Kostka--Foulkes polynomials, as different terms
in the Kostka--Foulkes polynomials might be associated to different key polynomials
in the non-symmetric setting.
\end{example}
The question of finding an appropriate generalization of charge that explains this phenomena
was raised by Lascoux in \cite[p. 267--268]{LascouxPolynomialsBook},
and thus \cref{thm:composition-charge} answers this question.
\subsection{Inversion-free fillings and Hall--Littlewood polynomials}
Similar to \cref{prop:atMostOneCoInvFree}, we have the following result:
\begin{proposition}\label{prop:atMostOneInvFree}
Given fixed column sets there exists at most one inversion-free
filling with shape $\alpha$ and big basement $\sigma$.
\end{proposition}
\begin{proof}
Proof is similar to the one in \cref{prop:atMostOneCoInvFree}.
\end{proof}
A number of analogous results are also possible for the modified Hall--Littlewood polynomials.
The following identity is given in \cite[Thm. 5.1.1]{HaglundNonSymmetricMacdonald2008}:
Let $\tau$ be a permutation and $\lambda$ be a partition. Then
\begin{align}\label{eq:generalMacdonaldHIdentity}
{\mathrm{\tilde{H}}}^{\tau w_0}_{\tau\lambda}(\mathbf{x};q,t) = {\mathrm{\tilde{H}}}_\lambda(\mathbf{x};q,t).
\end{align}
The proof of this is rather indirect, and J.~Haglund asked for a bijective proof of this identity.
With our bijection $\varphi$, we can prove the $t=0$ case:
\begin{theorem}
For every $\tau\in S_n$ and partition $\lambda$,
there is a column-set preserving bijection that establish the identity
\begin{align}\label{eq:macdonaldHIdentity}
{\mathrm{\tilde{H}}}^{\tau w_0}_{\tau\lambda}(\mathbf{x};q,0) = {\mathrm{\tilde{H}}}_\lambda(\mathbf{x};q,0).
\end{align}
\end{theorem}
\begin{proof}
Consider the fillings contributing to ${\mathrm{\tilde{H}}}^{\tau w_0}_{\tau\lambda}(\mathbf{x};q,0)$.
We may repeatedly apply $\varphi$ until the resulting fillings are
the ones contributing to ${\mathrm{\tilde{H}}}_\lambda(\mathbf{x};q,0)$.
\end{proof}
Although it might be tedious to carry out the bijections, remember that the resulting
bijection is uniquely defined by the column-set preserving property.
We have not been able to prove the more general \cref{eq:generalMacdonaldHIdentity},
but computer experiments suggests the following refinement of the equality:
\begin{conjecture}
Let $\mathrm{FIL}(\lambda,\tau,C)$ be all fillings with shape $\lambda$, big basement $\tau$
and column sets $C$.
Then
\[
\sum_{F \in \mathrm{FIL}(\lambda,w_0,C)} q^{\maj(F)} t^{\inv(F)}
=
\sum_{F \in \mathrm{FIL}(\tau\lambda,\tau w_0,C)} q^{\maj(F)} t^{\inv(F)}.
\]
\end{conjecture}
This indicates that one should be able to find a \emph{column-set preserving bijection}
proving Haglund's identity.
\begin{remark}
One can modify \cref{lem:tworowCase} to show that for every $\tau$, we have
\[
[t^{top}]{\mathrm{\tilde{H}}}^{\tau w_0}_{\tau\lambda}(\mathbf{x};q,t) = [t^{top}]{\mathrm{\tilde{H}}}_\lambda(\mathbf{x};q,t),
\]
where $[t^{top}]$ indicate the coefficient of the maximal power of $q$,
and the bijection proving this is given by $\phi$.
To give an outline of the proof, note first that fillings contributing to
both sides of the above identity are exactly coinversion-free fillings.
Furthermore, notice in this case with the big basement, is that the first set in the right hand side of
\cref{prop:colSetPresInj} is empty.
\end{remark}
\bigskip
We conclude this paper by proving an analogue of \cref{thm:composition-charge},
in the case of inversion-free fillings.
Let $F$ be an inversion-free filling with a basement.
Define the following biword of $F$ (different from the above) as follows:
Let the top row be the non-basement entries of $F$ listed in decreasing order
and the bottom row be the corresponding columns the entries belong to,
listed in \emph{increasing} order in case of a tie in the first row.
Let the cocharge word of $F$, $\ccw(F)$, be the bottom row in this biword.
As an example,
\[
\young(71,91423,8213,6)
\qquad \longrightarrow \qquad
\begin{pmatrix}
9 & 8 & 7 & 6 & 4 & 3 & 3 & 2 & 2 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 3 & 4 & 5 & 2 & 4 & 2 & 2 & 3 \\
\end{pmatrix}
\]
has subword decomposition $13542$, $1423$, $12$, $1$,
and cocharge value $4+1+0+0=5$.
\begin{theorem}
\label{TheoremCocharge}
Let $F$ be an inversion-free filling of composition
shape $\tau \lambda$ where $\lambda$ is partition and basement $\tau w_0$.
Then
\[
\cocharge(\ccw(F))=\maj(F).
\]
where $\cocharge(\cdot)$ is defined as in \cite{Haglund2005Macdonald}.
\end{theorem}
\begin{proof}
The case where $\alpha$ is partition is given in \cite{Haglund2005Macdonald}.
\footnote{The proposition in \cite{Haglund2005Macdonald} does not include a basement
explicitly, but adding $\omega_0$ as basement in the partition case
leaves the analysis in \cite{Haglund2005Macdonald} unchanged.}
The result then follows by noting that the $\ccw(F)$ is only dependent on the columns
sets of $F$, so by applying the column-set and major-index preserving map $\varphi$
until the partition shape is reached, the statement follows.
\end{proof}
This answers a conjecture given in \cite{Nelsen2005} \footnote{In \cite{Nelsen2005}, only fillings of
shape $(v_1,v_2,\ldots,v_k)$ and an index $\ell$ such that $v_1\ge v_2\ge\ldots\ge v_{\ell}$
and $v_1<v_{\ell+1}<v_{\ell+2}\ldots<v_k$ are considered and the (implicit) basement is of the
same form as in \cref{TheoremCocharge}.}
\subsection*{Acknowledgement}
The authors would like to thank Jim Haglund for helpful discussions.
The second author would also like to thank the first author for his mentorship.
PA is funded by the \emph{Knut and Alice Wallenberg Foundation} (2013.03.07).
\bibliographystyle{amsalpha}
|
\section{\label{sec:majdm} Majoron dark matter}
As it was shown in the previous section, our dark matter candidate corresponds to the {\it massive} majoron state ($J_{\rm DM}$).
In our model, the $J_{\rm DM}$ is a decaying DM candidate~\cite{Aranda:2009yb,Akhmedov:1992hi,Lattanzi:2014mia} where its decay channels are mainly to neutrinos and {\it massless} majorons.
In this section, we focus on these modes and also on the majoron dark matter production in the Early Universe.\\
\subsection{\label{sec:decay} Dark matter decay}
In the case of decaying DM, the main phenomenological constraint comes from the DM lifetime.
We assume in our case that the majoron DM has a lifetime $\tau_{\rm DM} > 10^{27}$~s ($\Gamma_{\rm DM} < 10^{-52}$~GeV)~\cite{Cirelli:2012ut}.
Besides, in this model, we have two classes of decay modes: fermionic and scalar.
The first one comes from Eq.~\ref{eq:slangran} and corresponds to $J_{\rm DM} \rightarrow \nu \nu$, which is also the typical majoron signature~\cite{Aranda:2009yb,Akhmedov:1992hi,Lattanzi:2014mia,Schechter:1981cv}.
The second class corresponds to scalar modes coming from the potential (Eq.~\ref{eq:vscal}).
In this case, they are 2-body decays ($J_{\rm DM} \rightarrow \zeta_i \zeta_j$) and 3-body decays ($J_{\rm DM} \rightarrow \zeta_i \zeta_j \zeta_k$).
Since we assume a keV majoron, these modes are reduced to $J_{\rm DM} \rightarrow 2 \zeta_1$ and $J_{\rm DM} \rightarrow 3 \zeta_1$. \\
\subsubsection{Decay into neutrinos:}
The decay rate to neutrinos from $J_{\rm DM}$ in this model is
\begin{eq}
\label{eq:nudecay}
\Gamma_\nu \simeq \frac{M_J}{32\pi} \left( \left|\left| O_L \right|\right|^2\,+\, \left|\left| O_R \right|\right|^2\right) \, ,
\end{eq}
where $m_\nu \ll M_J$ is taken.
The terms $O_L$ and $O_R$ are the couplings to neutrinos which come from the $J_{\rm DM}$ projection on the scalar states $H,S,X$ and $\nu$ projection on the fermionic states $\nu_L, N_{1,2}$.
The decay rate can be written in our model as
\begin{eq}
\label{eq:J2nu}
\Gamma_\nu = \frac{M_J}{32\pi} f\left( m_\nu,m_D,M,v_S\right) \, ,
\end{eq}
where the function $f$ is described in~\ref{app:fercoup} and it contains the dependence on the parameters of the couplings between neutrinos and the majoron DM.
The decay rate can be expanded in powers of $\displaystyle \frac{\mu}{M} = \alpha \sim 10^{-7}$, then, the expansion up to order $\alpha$ is:
\begin{eq}
\label{eq:gammanu}
\Gamma_\nu = \Gamma_{0\nu}(\omega) \, \left\{(2-3\omega^2)\left(2-3\omega^2(1+2 \alpha)\right) + \mathcal{O}(\alpha^2) \right\}\, ,
\end{eq}
where the overall factor is:
\begin{eq}
\Gamma_{0\nu}(\omega) = \frac{M_J {m_{\nu}}^2}{256 \pi v_S^2} \, \frac{1}{\omega^2 (1 + 9 \omega^2)} \, .
\end{eq}
The decay rate vanishes for $\omega_0 = \sqrt{2/3}$ up to order $\alpha$ and the error carried by this choice produces a decay rate of $\Gamma_\nu = \Gamma_{0\nu}(\omega_0)\, 4 \alpha^2$.
This indicates that $v_X$ and $v_S$ should have similar values satisfying the previous ratio since further powers of $\alpha$ will act as perturbations around that value for $\omega_0$.
The overall factor can be evaluated at $\omega_0$ giving rise to:
\begin{eq}
\Gamma_{0\nu}(\omega_0) \simeq 10^{-40} \, {\rm GeV} \left(\frac{m_{\nu}}{0.1\,{\rm eV}} \right)^{2} \left(\frac{M_J}{1 \, {\rm keV}}\right) \left(\frac{v_S}{100\,{\rm TeV}}\right)^{-2} \, ,
\end{eq}
this indicates that we still require a factor $10^{-12}$ in order to satisfy the DM stability constraint.
However, the $\Gamma_{\nu}$ at $\omega_0$ gets a $10^{-14}$ suppression due to the $\alpha^2$ factor.
Hence it ensures that $\Gamma_{\nu} < \Gamma_{\rm DM}$.
Nonetheless, the solution at all orders in $\alpha$ for $\Gamma_{\nu} = 0$ is exact and it is given by:
\begin{eq}
\label{eq:omega23}
\omega = \frac{\sqrt{4 - 4 \alpha^2 + \alpha^3 - \alpha^4}}{\sqrt{3(2 + 2 \alpha + \alpha^4)}}
\end{eq}
It is quite remarkable that the $\mu/M$ ratio is at the same time relevant for the neutrino mass value in the inverse seesaw mechanism and for the DM lifetime.
\subsubsection{Decay into scalars:}
The scalar modes can occur only for $J_{\rm DM} \rightarrow 2 \zeta_1$ and $J_{\rm DM} \rightarrow 3 \zeta_1$ due to $M_J$ is $\mathcal{O}($keV) and the heavy scalars have masses larger than 100~GeV.
The couplings needed to calculate the decay rate come from the scalar potential after writing it in terms of the mass eigenstates and taking the corresponding derivatives:
\begin{eq}
\lambda_{ijk} = \frac{\partial^3 V_{\rm scalar}}{\partial \zeta_i \partial \zeta_j \partial \zeta_k} \quad {\rm and} \quad \lambda_{ijkl} = \frac{\partial^4 V_{\rm scalar}}{\partial \zeta_i \partial \zeta_j \partial \zeta_k \partial \zeta_l} \, .
\end{eq}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{lambdaeffv2.pdf}
\caption{\label{fig:lambeff} Diagrams involve in the calculation of $\lambda_{2111}$ for the process $J_{\rm DM} \rightarrow 3 \zeta_1$}
\end{figure}
Afterwards, the decay rate for the 3-body process can be calculated as
\begin{eq}
\Gamma_{3\zeta} &=& \frac{1}{(64\pi)^3} M_J \left|\left| \lambda^{\rm eff}_{2111} \right|\right|^2 \label{eq:3zeta}
\end{eq}
where $\lambda^{\rm eff}_{2111}$ includes the diagrams depicted in Fig.~\ref{fig:lambeff}.
Notice that the contributions for this coupling come directly from the scalar potential and from the combined contributions from the heavy scalars.
Given a $J_{\rm DM}$ with a mass in the keV range, we can safely integrate out the effect of the heavy mediators giving rise to:
\begin{eq}
\label{eq:scaleff}
\lambda^{\rm eff}_{2111} &=& \lambda_{2111}\,-\,\frac{\lambda_{213}\lambda_{113}}{m_3^2}\,-\,\frac{\lambda_{214}\lambda_{114}}{m_4^2}\,-\,\frac{\lambda_{215}\lambda_{115}}{m_5^2} \, ,
\end{eq}
where the relevant expressions for $\lambda_{2ij}$ and $\lambda_{2111}$ are shown in \ref{app:scacoup}.
Our aim is to obtain $\lambda^{\rm eff}_{2111} \simeq 0$ in order to stabilize the DM.
The first 3 terms come from Eq.~\ref{eq:potsx} and they are mainly unsuppressed quartic couplings.
On the other hand, we found that last term in Eq.~\ref{eq:scaleff} contains an overall factor of $\left(M_J/v_S \right)^4$ which comes from Eq.~\ref{eq:potvi}.
This last term, eventually, will provide the main contribution to the decay rate whether all the other terms are vanished,
so that we can estimate a typical value of the decay rate when only this term is present. In such conditions, the decay is given by:
\begin{eq}
\Gamma_{3\zeta} &=& \frac{1}{(64\pi)^3}\, M_J \left|\left| \frac{\lambda_{215}\lambda_{115}}{m_5^2} \right|\right|^2 \\
&=& \frac{1}{(64\pi)^3}\, M_J\, \left( \frac{M_J}{v_S} \right)^8\cdot \left( \frac{\psi^2 \omega^3}{(1 + 9 \omega^2)^3}\,\mathcal{F}\left(A,\psi,\omega,\lambda_h,\lambda_{HS},\lambda_{HX}\right) \right)^2 \, ,\nonumber
\end{eq}
where the function $\mathcal{F}\left(A,\psi,\omega,\lambda_h,\lambda_{HS},\lambda_{HX}\right)$ goes to zero when both couplings $\lambda_{HS}$ and $\lambda_{HX}$ go to zero.
For a wide range of the values of the couplings, we obtain that $\mathcal{F} \sim \mathcal{O}(1)$.
This implies that the decay rate is suppressed mainly by $\left( M_J/v_S \right)^8$.
Hence, an orders of magnitude evaluation of the decay rate is given by:
\begin{eq}
\Gamma_{3\zeta} \sim 10^{-12} \, \left(\frac{M_J}{1 \, {\rm keV}}\right) \left( \frac{M_J}{v_S} \right)^8\, {\rm GeV}\,\, \sim \,\, 10^{-100}\, {\rm GeV} \, .
\end{eq}
The last estimation indicates that $\Gamma_{3\zeta}$, when only the contribution of the higgs is considered, is larger than our benchmark for $\Gamma_{\rm DM}$ by 48 orders of magnitude.
This implies that the decay mediated does not spoil the DM lifetime whatsoever, and it can be safety neglected under our assumptions.
Now our concern is to vanish the remaining terms in Eq.~\ref{eq:scaleff}:
\begin{eq}
\label{eq:ampsimpeq0}
\lambda_{2111}\, -\, \lambda_{213}\lambda_{113}/m_3^2\, -\, \lambda_{214}\lambda_{114}/m_4^2 \, \simeq \, 0 \, .
\end{eq}
The full expressions involved in this equation in the limit $\lambda_{HX}$, $\lambda_{HS} \ll 1$ are given in \ref{app:scacoup}.
The first term comes directly from the scalar potential and it is not necessarily suppressed.
However it does contain terms proportional to $(M_J/v_S)^2$ that can be neglected.
The last 2 terms are always present and require that $0<\psi<1$ to be well defined (see Eqs.~\ref{eq:lam213} and \ref{eq:lam214}).
This condition forces to have a non-zero $\lambda_5$ via the value of $A$.
At this point and due to the complexity of the expressions involved at Eq.~\ref{eq:ampsimpeq0}, we will explore the parameter space with a numerical scan which we are going to discuss in the next section.\\
\subsection{Dark matter production}
In this section, we aim to describe a tentative framework for the DM production in the Early Universe.
Our DM candidate $J_{\rm DM}$ shares similar properties with a Feebly Interacting Massive Particle (FIMP)~\cite{Hall:2009bx} by means of suppressed coupling with the SM-like higgs and active neutrinos.
However, the couplings of $J_{\rm DM}$ to the heavy scalars $\zeta_3$ and $\zeta_4$ may not be necessarily suppressed, and in turn, the couplings of these particles with the SM-like higgs might take a wide range of values.
This makes the heavy scalars to be able to interact with the rest of the thermal bath, and subsequently decay to $J_{\rm DM}$.
Indeed, the same logic could be applied to the process involving heavy neutrinos.
Under these conditions, and since we assume a keV DM candidate, the production mechanism cannot be addressed with the typical freeze-out for $J_{\rm DM}$, which is used in WIMP-DM models to reproduce the relic abundance~\cite{Gondolo:1990dk}.
On top of having $J_{\rm DM}$ as a FIMP, the Lightest Observable Sector Particle (LOSP) could be either the lighest of the heavy neutrinos or the lightest between $\zeta_{3,4}$ because all of them have $U(1)_l$ charges.
Due to the interplay among all particles of the model, the relic abundance calculation has many edges which at first sight are unclear, however, we will sketch some relevant processes involved in the production.
Some of the prototype processes for the DM production can be summarized either by a quartic interaction like $\lambda \, \zeta_{3} \zeta_{4} J_{\rm DM}^2 $ and ${\displaystyle \frac{\lambda^{\prime}}{v_S}} \, \overline{\mathcal{N}} \mathcal{N}^{\prime} J_{\rm DM}^2$, or a triple interaction like $y\, \overline{\mathcal{N}} \mathcal{N}^{\prime} J_{\rm DM}$ and $\lambda^{''} v_S \, \zeta_{3} \zeta_{1} J_{\rm DM}$.
Focusing now only on the couplings of the scalar sector, we realize that the ones between $J_{\rm DM}$--$\zeta_{3,4}$, and $\zeta_{3,4}$--$\zeta_{5}$ are not necessarily suppressed, and they are controlled mainly by $\lambda_{X,S}$ and $\lambda_{HX,HS}$ respectively.
Oppositely, the coupling in $J_{\rm DM}$--$\zeta_{5}$ is suppressed by the ratio $\left(M_J/v_S\right)^2$.
The lagrangian $\mathcal{L}_{\rm int} = y_L \bar{L} H N_1$, where the yukawa value is proportional to $m_D$, gives rise to the interactions $\nu \zeta_5 \mathcal{N}^{(')}$.
Besides, the lagrangian $\mathcal{L}_{\rm int} = y_S S N_1 N_2$, where the coupling is proportional to $M$, gives rise to interactions $J_{\rm DM} \mathcal{N}^{(')} \mathcal{N}^{(')}$.
In both cases, the couplings are controlled by the inverse seesaw and, in our setup, they are $y_L \simeq 0.1$ and $y_S \simeq 1$.
All of this produces an interplay between neutrinos, higgs, and DM, similar to the scalar sector and the interplay among heavy scalars
and majorons.
In the Early Universe, the evolution of the DM yield depends directly on the interaction of $\zeta_{3,4}$ and/or $\mathcal{N}^{(')}$ with the SM.
In this way, the yields of LOSPs act as portals between SM and DM.
The combined processes are present meanwhile $T \gtrsim m_{\rm LOSP}$.
This means that LOSPs are likely in thermal equilibrium.
After that, they will decouple from the thermal bath in a similar way to the freeze-out, transferring subsequently their yields to $J_{\rm DM}$ and $\zeta_1$ via LOSP decays.
However, not all of the final DM yield comes necessarily from these decays.
If the DM-LOSP couplings are large enough, DM could reach thermal equilibrium assisted by LOSP's interactions and thus a fraction of the DM yield comes from the DM freeze-out.
Otherwise, i.e. for small couplings, the outcoming fraction of the DM could be explained via freeze-in.
A more complete calculation of the DM abundance will be given in a future work.
\section{\label{sec:intro}Introduction}
The success of the Standard Model of particle physics (SM) has been established thanks to accurate predictions of many experimental observations~\cite{Weinberg:1967tq,Higgs:1964pj,Arnison:1983rp,Bagnaia:1983zx,Aad:2012tfa,Chatrchyan:2012xdj}.
Nevertheless, the SM presents some theoretical and experimental issues that it cannot describe.
One of these is the Dark Matter (DM), which is the largest matter component ($\sim 85 \%$) present in the Universe~\cite{Oort:1932BAN,Rubin:1980zd,Ade:2015xua}.
Another one is the neutrino oscillations~\cite{Pontecorvo:1957cp,Maki:1962mu,Ahmad:2001an,Fukuda:1998tw}, which are a consequence of the still-not-measured neutrino masses.
Both issues provide a tantalizing connection between DM and neutrinos that can be realized in many ways (see, for instance,~\cite{Lattanzi:2014mia,Hirsch:2013ola}).
From the observations in neutrino oscillation experiments, we get two important features of neutrinos: $i)$~the leptonic mixing angles have very large values when compared with the ones in the hadronic sector, $ii)$~the neutrino mass scale is very small with respect to the masses of the rest of the SM fermions~\cite{Olive:2016xmw,Esteban:2016qun,Capozzi:2016rtj,Forero:2014bxa}.
These two features could be interpreted as an indication of new physics beyond the SM scale.
The simplest SM-like framework with massive neutrinos assumes that neutrinos are Majorana particles.
In this framework, Majorana masses arises via the dimension-5 Weinberg operator~\cite{Weinberg:1979sa}.
This operator respects main SM symmetries like Lorentz and the $SU(3)\times SU(2)\times U(1)$ gauge structure, and it is build exclusively with SM fields.
This operator can be generated at tree level only by minimally extending the SM in three ways.
These are renormalizable constructions that are known as type I~\cite{Minkowski:1977sc,Mohapatra:1979ia}, II~\cite{Schechter:1980gr,Mohapatra:1980yp}, and III~\cite{Foot:1988aq} seesaw mechanisms.
Each seesaw construction provides different predictions that can be tested in current and future experiments.
The inclusion of new particles is unavoidable in these constructions and the explanation for the neutrino mass scale is closely related to the mass scale for the new particles.
Although in the SM, baryon and lepton numbers are accidental global symmetries, the seesaw mechanism explicits the breaking of lepton number
through of the presence Majorana neutrino masses.
This issue, in turn, comes from the fact that lepton number is already not a symmetry at the new particle scale.
One way to alleviate this is assuming that lepton number is preserved at higher scale but it is spontaneously broken at some intermediate scale.
In this scheme, the goldstone boson of the lepton number breaking appears and it is historically know as \emph{The Majoron}~\cite{Riazuddin:1981hz}.
Although the Majoron appears as a massless particle, there are conjectures saying that global symmetries must be broken due to Planck scale effects.
This would explain how the Majoron get its small mass~\cite{Rothstein:1992rh} and why fully stable DM might not be possible~\cite{Boucenna:2012rc}.
When the Majoron becomes massive, the most important decay channel for this particle is through neutrinos~\cite{Schechter:1980gr,Berezinsky:1993fm,Aranda:2009yb,Lattanzi:2007ux,Garcia-Cely:2017oco}.
For sub-keV majorons, those particles have lifetimes larger than the age of the Universe~\cite{Berezinsky:1993fm,Lattanzi:2007ux}.
A massive Majoron is electrically neutral and might have weak interactions, something that turns it into a potentially suitable Dark matter candidate.
However, it is still not well understood how it acquires mass and how to produce a feasible relic DM abundance of massive Majorons in the early Universe.
At this point, we could argue that the \emph{goldstone} Majoron and the Majoron DM are different particles with a common origin.
On top of that, within the type I and III seesaw mechanisms and in order to give rise to small neutrino masses, the mass scale at which the new physics lives is around $10^{12}$~GeV.
However, the inclusion of extra fermions singlets, and interactions among themselves and the SM lepton doublet, can show variations of the seesaw mechanism in which the mass scales for the new particles is not necessarily large.
In the literature, those variations are called as {\it low scale seesaw}~\cite{Mohapatra:1986aw,Mohapatra:1986bd,Barr:2003nn}, and our focus
will be on the {\it inverse seesaw mechanism}~\cite{Mohapatra:1986bd,Abada:2014vea}.\\
In this work we propose a mechanism in which neutrino physics and a Majoron DM candidate are joined together.
We use this setting as a scheme based of the inverse seesaw mechanism which is in turn spontaneously generated.
The spontaneous symmetry breaking gives rise to a goldstone Majoron and a massive Majoron which is our DM candidate.
We describe the basics of the inverse sessaw and our model in Section~\ref{sec:model}.
The implications of our model for the Majoron DM are shown in Sections~\ref{sec:majdm} and \ref{sec:disc}.
Finally, the conclusions are in Section~\ref{sec:conc}.
\section{\label{sec:disc} Discussion}
Up to this moment, we have described the expression for the $J_{\rm DM}$ decay into neutrinos and three $\zeta_1$.
However, in this section, we aim to perform numerical analysis based on the stability of the DM candidate since this is the observable that constraints most the couplings.
Since we know the DM lifetime is extremely large and our model does not include a ad-hoc stabilizing symmetry, it is expected that the correlations among parameters must be strong.
In any case, we could assume that the correlations are the consequence of an unknown unified symmetry.
Moreover, in this part, we will not include constraints coming from the DM relic abundance.\\
\begin{table}[t]
\centering
\begin{tabular}{|c|c|}
\hline
\phantom{xx} Parameter \phantom{xx}&\phantom{xx} Value \phantom{xx}\\
\hline
\hline
$M$ & 100 TeV \\
$\mu$ & 10 MeV \\
$m_D$ & 10 GeV \\
\hline
$v_S$ & $10^8$ -- $10^{12}$ GeV\\
$\omega$ & 0.4 -- 1.6 \\
\hline
\end{tabular}
\caption{
\label{tab:fixedpar}
Benchmarks and scan range for parameters in the $J_{\rm DM} \rightarrow \nu \nu$ decay.
}
\end{table}
In the first place, the channel to neutrinos will be analyzed.
As we can see from \ref{app:fercoup}, the formuli regarding the couplings $O_L$ and $O_R$ shown at the Eq.~\ref{eq:nudecay} just depend on the parameters $m_D$, $M$, $\omega$, $v_S$, and $m_{\nu}$.
Although in this case, a dependence on the mass of the neutrino appears in order to make the parameter space compatible with $m_\nu\, \sim 0.1$~eV (See Tab.~\ref{tab:fixedpar}).
It is evident to realize that vanishing couplings imply a vanishing amplitude $\Gamma\left( J_{\rm DM} \rightarrow 2\nu\right)$.
Thus, given a set of $M$, $m_D$ and $v_S$, one can search for an $\omega$ that makes $J_{\rm DM}$ decay to neutrinos in cosmological times.
In Fig.~\ref{fig:omegavs}, we present the result of a scan on the $J_{\rm DM}$ decay width in the plane: $v_S$ versus $\omega$.
We showed in Eq.~\ref{eq:omega23} that $\omega \simeq \sqrt{2/3}$ makes the couplings $O_L$ and $O_R$ vanish.
In this plot, we present the decay width variability for a range of $\omega$ values.
We highlight, with the dashed line, the frontier of the DM lifetime.
For smaller values of $v_S$, DM lifetime requires that omega must be very close to $\sqrt{2/3}$, indicating a rather strong vev alignment among $S$ and $X$.
In opposition, larger values of $v_S$ weakens this alignment, since there is an overall factor $v_S^{-2}$ in the decay width (Eq.~\ref{eq:gammanu}).\\
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{omega-vs.pdf}
\caption{
\label{fig:omegavs}
Plot $v_S$ versus $\omega$. The color palette indicates the value of the $J_{\rm DM}$ decay width to neutrinos.
The dashed line shows the benchmark value for the DM lifetime $\tau_{DM} = 10^{27}$~s.
The value $\omega = \sqrt{2/3}$ makes the decay width vanish regardless of $v_S$ value.
For $v_S \gg 10^{11}$~GeV, $\omega$ starts to be irrelevant to satisfy the DM lifetime constraint.}
\end{figure}
In the analysis of the scalar decay, as it was described in the previous section, the approach is to vanish the non-Higgs part of the coupling $\lambda_{2111}^{{\rm eff}}$ (i.e. Eq.~\ref{eq:ampsimpeq0}).
The Higgs part is neglected because it is extremely suppressed.
Besides, the scalar sector parameter space is mostly independent of the fermion sector, although it is connected by $\omega$ and $v_S$.
Thus, we look for an interplay among the parameters $A$, $\psi$, $\lambda_X$,
and $\omega$ that satisfies Eq.~\ref{eq:ampsimpeq0}.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{psivsomegav2.pdf}
\includegraphics[width=0.75\columnwidth]{psivsomega.pdf}
\caption{
\label{fig:psivsomega}
Plot $\psi$ versus $\omega$ for $A=0.2$ (top) and $A=0.5$ (bottom).
The bluish lines correspond for the combination of $\psi$ and $\omega$ for a fixed value of $\lambda_X$ that makes the decay $J_{\rm DM}\rightarrow 3\zeta_1$ zero.
The vertical magenta dashed line correspond to $\omega = \sqrt{2/3}$.
The green area is the $\omega$ range that passes the DM lifetime constraint for the neutrino channel for $v_S \simeq 10^{11}$~GeV.
}
\end{figure}
In Fig.~\ref{fig:psivsomega}, we show the combinations of $\psi$ and $\omega$ that vanish the decay width for 2 values of $A = (0.2, 0.5)$ and 5 for $\lambda_X$ in different shades of blue.
This selection was made in order to show a general trend in the dependence between $\psi$ and $\omega$.
The blue curves range from the largest possible $\lambda_X$ (lightest blue line)
given by the perturbation limit, to a smaller value ($\lambda_X = 0.1$, darkest blue).
The left-most value of each curve indicates a solution when $\psi$ starts to become complex or stops to represent a cosine of an angle.
The light green zone corresponds to the range of $\omega$ compatible with the decay to neutrinos for $v_S \sim 10^{11}$~GeV.
The vertical dashed line is simply $\omega = \sqrt{2/3}$.
The scanned range of $\omega$ was up to $3$ in order to explore a ratio $v_X$ more or less in the same order of magnitude of $v_S$.
Besides, we include the constraint that the heaviest scalars are above the TeV scale.
By comparing both plots in Fig.~\ref{fig:psivsomega}, we observe that the perturbative limit for $\lambda_X$ sets a minimum $\psi(\omega)$-curve.
This minimum curve grows with larger values of $A$.
In opposition, the maximum $\psi(\omega)$-curves are related with the smallness of $\lambda_X$, however $\lambda_X \simeq 0$ puts problems with the vacuum stability.
Similar information is shown in Fig.~\ref{fig:Avsomega}, where we present the zero decay width solution for $A$ versus $\omega$.
Here we observe that the pertubative limit of $\lambda_X$ produces a maximum $A(\omega)$-curve for each choice of $\psi$.
For both Figs.~\ref{fig:psivsomega} and~\ref{fig:Avsomega}, we show that there is a smooth transition for different values of $\lambda_X$ and the combinations of $\psi$, $A$, and $\omega$ that make the decay width zero.
This implies that the solutions belong to a smooth volume in the parameter space, and therefore, one can always find one parameter when the other 3 have been given.
Besides, we find that extreme values of $A \, (\sim 0)$, and $\psi \, (\gtrsim 1.0 \, , \lesssim 0.0)$, are not favored by the DM stability condition and these values could lead to tachyonic states of $\zeta_3$ or $\zeta_4$.
Moreover, when we focus on the green region, we find that the most of the curves pass through it.
This is showing that there is a natural compatibility among the solutions for the neutrino and scalar decay modes independently.
\begin{figure}[t]
\includegraphics[width=0.75\columnwidth]{Avsomegav2.pdf}
\includegraphics[width=0.75\columnwidth]{Avsomega.pdf}
\centering
\caption{
\label{fig:Avsomega}
Plot $A$ versus $\omega$ for $\psi=0.1$ (top) and $\psi=0.4$ (bottom).
The bluish lines correspond for the combination of $A$ and $\omega$ for a fixed value of $\lambda_X$ that makes the decay $J_{\rm DM}\rightarrow 3\zeta_1$ zero.
The vertical magenta dashed line correspond to $\omega = \sqrt{2/3}$.
The green area is the $\omega$ range that passes the DM lifetime constraint for the neutrino channel for $v_S \simeq 10^{11}$~GeV.
}
\end{figure}
An interesting case regards the higgs physics, in our model the mixing among the CP-even scalars gives a SM-like higgs that is weakly mixed with the rest of the scalars by a factor $v_h/v_S$.
However, this mixing does not forbid a contribution to the invisible higgs decay, namely $H \rightarrow J_{\rm DM} J_{\rm DM},\, J_{\rm DM} \zeta_1,\, \zeta_1 \zeta_1$.
These processes come directly from the scalar potential via couplings $\lambda_{HS}$, $\lambda_{HX}$, and $\lambda_{\rm cp}$, which are translated into $\lambda_{215}$, $\lambda_{115}$, and $\lambda_{225}$ (See Eqs.~\ref{eq:l215}, \ref{eq:l115}, and \ref{eq:l225}, respectively).
We observe that all these couplings are suppressed by $(M_J/v_S)^2$, therefore the decay width is suppressed by $(M_J/v_S)^4$.
After evaluation, we obtain that the higgs decay width is $\mathcal{O}\left( 10^{-44} \right)$~GeV and thus these processes cannot be constrained using the measurement of the invisible higgs decay~\cite{Chatrchyan:2014tja,Baek:2014jga}.
The role of CP-phases in the decay width either in the scalar or neutrino modes is not an issue.
In the scalar sector, most of the effect is washed out by the tadpole equations that fix the relation among the 3 phases: $\theta$, $\tau$, and $\delta$.
In the case of neutrinos, in addition to our CP-phases, we could include extra phases in the yuwakas: $y_L$, $y_S$, and $y_X$.
However, we decided to keep the inverse seesaw mass terms real, and hence, the possible impact of CP-phases in the phenomenology is absorbed.
If we wanted to add effect of CP-phase, we should either relax the condition of real mass terms or add more families of neutrinos.
This addition of CP-phases effects in the DM decay adds an improvement on this setup.
A different improvement is to promote from a global $U(1)_l$ symmetry to a gauge one.
This would relax the correlations in the scalar sector, because the $\zeta_1$ would be eaten by the corresponding gauge boson after the SSB.
The latter feature is going to be worked out in a future work.
\section{Lepton charge assignments} \label{app:charge}
In this section, we examine the most general way in which the lepton charges are fixed for the fields $N_1$, $N_2$, $S$ and $X$.
As we advanced, the Yukawa couplings at Eq.~\ref{eq:slangran} will fix the value for the lepton numbers of the field $N_1$. However, this does not fix the charges for the new scalars $S$ and $X$, nor for $N_2$.
The final assignation can be obtained after considering the following general scalar potential:
\begin{eq}
V_{\rm I} = \lambda_{\rm cp} e^{i\delta} X^{m}S^{\dagger n} + {\rm h.c.} \, , \label{eq:deltagen}
\end{eq}
After demanding that $V_I$ is renormalizable, we can choose the values of $m$ and $n$ that subsequently will fix the values of lepton number for $S$ and $X$.
Thus, one has a collection of models formed by taking $m+n=2,3,4$.
Notice that we still have to choose one value for $m$ and $n$.
By now, we choose $m+n=4$ and, by following the assignation made at the Table~\ref{tab:charges}, we establish conditions in order
to make \ref{eq:deltagen} invariant under lepton number.
\begin{eq}
m\,+\,n\, =\, 4 \quad {\rm and}\quad m\left( 2x\right)\,-n\left( 1\,-\,x\right)\, =\, 0 \quad \Rightarrow \nonumber \\
x\, =\, \frac{n}{n+2m}\, =\, \frac{n}{8-n}
\end{eq}
Recall that $n,m$ are integers running from $1\dots 3$ ($0$ and $4$ will break lepton number explicitly).
Therefore, for $n=3$ and $m=1$, one has $x=3/5$, as it was stated in the section~\ref{sec:spinvseesaw}.
At the Table~\ref{tab:chgen}, we present the lepton number charges for different values of $n$ for $m+n=4$.
In order to show a case with a different choice of $m+n$, at the Table~\ref{tab:chgen2} we present the lepton numbers of the fields after considering $m+n=3$ at Eq.~\ref{eq:deltagen}.
\begin{table}[b]
\centering
\begin{tabular}{|c|ccccc|}
\hline
& \phantom{x} $L$ \phantom{x} & \phantom{x} $N_1$ \phantom{x} & \phantom{x} $N_2$ \phantom{x} & \phantom{x} $S$ \phantom{x} & \phantom{x} $X$ \phantom{x} \\
\hline\hline
\phantom{x} $n=1$ \phantom{x} & $1$ & $-1$ & $1/7$ & $6/7$ & $2/7$ \\
\phantom{x} $n=2$ \phantom{x} & $1$ & $-1$ & $1/3$ & $2/3$ & $2/3$ \\
\phantom{x} $n=3$ \phantom{x} & $1$ & $-1$ & $3/5$ & $2/5$ & $6/5$ \\
\hline
\end{tabular}
\caption{\label{tab:chgen} Charge assignment of different models for $m+n=4$.}
\end{table}
\begin{table}[b]
\centering
\begin{tabular}{|c|ccccc|}
\hline
& \phantom{x} $L$ \phantom{x} & \phantom{x} $N_1$ \phantom{x} & \phantom{x} $N_2$ \phantom{x} & \phantom{x} $S$ \phantom{x} & \phantom{x} $X$ \phantom{x} \\
\hline\hline
\phantom{x} $n=1$ \phantom{x} & $1$ & $-1$ & $1/5$ & $4/5$ & $2/5$ \\
\phantom{x} $n=2$ \phantom{x} & $1$ & $-1$ & $1/2$ & $1/2$ & $1$ \\
\hline
\end{tabular}
\caption{\label{tab:chgen2} Charge assignment of different models for $m+n=3$.}
\end{table}
\section{Couplings}
In this section we describe briefly the relevant couplings used in this work, with an special emphasis in the interactions participating in
the decays of the Majoron.
\subsection{Fermion Couplings} \label{app:fercoup}
The couplings shown below are related to the process $J_{DM} \longrightarrow \nu\nu$ that appears in the models involving majoron DM. Recall that in these couplings we got rid of the explicit dependence of the Yukawas and it was preferred to work with the mass parameters involved in neutrino mass generation via inverse seesaw, namely $\mu$, $m_D$ and $M$ (c.f. Eq.~\ref{eq:nudecay})
\begin{eq}
O_L &=& \frac{\mathcal{D}_1^{(L)}}{\mathcal{D}_2^{(L)}} \\
\mathcal{D}_1^{(L)} &=& im_\nu \left(4 m_D^6 + 4 M m_D^4 m_\nu +
M^3 m_\nu^3\right) \nonumber \\
& & \left[ \left( -4 m_D^8 + 4 M^2 m_D^4 m_\nu^2 -
M^3 m_D^2 m_\nu^3 + M^4 m_\nu^4\right) \right. \nonumber \\
&+& \left. 3 \left(2 m_D^8 + 2 M m_D^6 m_\nu +
M^4 m_\nu^4\right) \omega^2 \right] \\
\mathcal{D}_2^{(L)} &=& \left(2m_D^2 + 3 M m_\nu\right)^2 \left(m_D^2 -
M m_\nu\right) \nonumber \\
& & \left(2 m_D^4 - M m_D^2 m_\nu + M^2 m_\nu^2\right)^2 v_s \omega \sqrt{2 + 18 \omega^2} \\
O_R &=& \left( O_L \right)^{*}\, .
\end{eq}
By using the definitions from above, the function $f$ at Eq.~\ref{eq:J2nu} can be expressed as
\begin{eq}
f\left( m_\nu,m_D,M,v_S \right) = \left|\left| O_L \right|\right|^2\,+\, \left|\left| O_R \right|\right|^2
\end{eq}
\subsection{Scalar Couplings} \label{app:scacoup}
Since the relevant couplings for the DM decay in the scalar sector are cuartic, they have no mass dimensions.
This is respected by the effective coupling at Eq.~\ref{eq:scaleff}, which makes the entire coupling independent of mass scales, and thus, it just depends on the adimensional parameters we set (namely $A$, $\omega$, $\psi$, $\lambda_h$, $\lambda_X$ and $M_J/v_S$).
First, it is shown the formula for the direct contribution to this coupling:
\begin{eq}
\lambda_{2111} &=& \frac{\mathcal{D}_1^{(1)}}{\mathcal{D}_2^{(1)}} \\
\mathcal{D}_1^{(1)} &=& -3 \left[3 A \left(1 + 9 \omega^2\right) \left\{-2 \psi \omega +
\sqrt{1 - \psi^2} \left(-1 + 9 \omega^2\right)\right\} \right. \nonumber \\
&+& \left. 8 \psi \omega \left\{ \left(\frac{M_J^2}{v_s^2}\right) -
27 \left( \frac{M_J^2}{v_s^2} \right) \omega^2 +
6 \lambda_{X} \omega^2 \left(1 + 9 \omega^2\right)\right\}\right] \\
\mathcal{D}_2^{(1)} &=& 4 \psi (1 + 9 \omega^2)^3
\end{eq}
Now, we show the formuli for the contributions coming from the integrated effect of the heavy scalars.
On the one hand, we explicit the formuli for $\frac{\lambda_{213}\lambda_{113}}{m_3^2}$ and $\frac{\lambda_{214}\lambda_{114}}{m_4^2}$,
which share some similarities.
\begin{eq}
\label{eq:lam213}
\frac{\lambda_{213}\lambda_{113}}{m_3^2} &=& \frac{\mathcal{D}_1^{(3)}}{\mathcal{D}_2^{(3)}} \\
\mathcal{D}_1^{(3)} &=& 3 \left[A \left(-1 + \psi\right) \left(-1 + \psi -
9 \sqrt{1 - \psi^2} \omega\right) \left(1 + 9 \omega^2\right) \right. \nonumber \\
&-& \left. 2 \psi \omega \left(12 \left(\frac{M_J^2}{v_s^2} \right) \left(\sqrt{1 - \psi^2} +
5 \left(-1 + \psi\right) \omega\right) \right. \right. \nonumber \\
&+& \left. \left. \lambda_{X} \omega \left(1 - \psi + 9 \sqrt{1 - \psi^2} \omega\right) \left(1 +
9 \omega^2\right)\right)\right] \nonumber \\
& & \left[A \left(-1 + \psi\right) \left(\sqrt{1 - \psi^2} + \left(-1 + \psi\right) \omega\right) \left(1 +
9 \omega^2\right) \right. \nonumber \\
&+& \left. 2 \psi \omega \left(\lambda_{X} \omega \left(\sqrt{1 - \psi^2} + \left(-1 + \psi\right) \omega\right) \left(1 +
9 \omega^2\right) \nonumber \right. \right. \\
&-& \left. \left. 4 \left(\frac{M_J^2}{v_s^2} \right) \left(1 + 3 \sqrt{1 - \psi^2} \omega -
6 \omega^2 + \psi \left(-1 +
6 \omega^2\right)\right)\right)\right] \\
\mathcal{D}_2^{(3)} &=& 8 \left(-1 + \psi\right) \psi \left(1 + 9 \omega^2\right)^3 \nonumber \\
& & \left[-A \left(1 + \psi\right) \left(1 + 9 \omega^2\right) +
2 \psi \left(-\lambda_{X} \omega^2 \left(1 +
9 \omega^2\right) \right. \right. \nonumber \\
&+& \left. \left. \left( \frac{M_J^2}{v_s^2} \right) \left(-1 + \psi -
6 \sqrt{1 - \psi^2} \omega + 3 \omega^2 +
3 \psi \omega^2\right)\right)\right]\, ,
\end{eq}
\begin{eq}
\label{eq:lam214}
\frac{\lambda_{214}\lambda_{114}}{m_4^2} &=& \frac{\mathcal{D}_1^{(4)}}{\mathcal{D}_2^{(4)}} \\
\mathcal{D}_1^{(4)} &=& 3 \left[ A \left( 1 + \psi\right) \left( 1 + \psi -
9 \sqrt{1 - \psi^2} \omega\right) \left( 1 + 9 \omega^2\right) \nonumber \right. \nonumber \\
&-& \left. 2 \psi \omega \left( 12 \left( \frac{M_J^2}{v_s^2} \right) \left( \sqrt{
1 - \psi^2} +
5 \left( 1 + \psi\right) \omega\right) \right. \right. \nonumber \\
&-& \left. \left. \lambda_{X} \omega \left( 1 +
\psi - 9 \sqrt{1 - \psi^2} \omega\right) \left( 1 +
9 \omega^2\right)\right)\right] \nonumber \\
& & \left[ A \left( 1 + \psi\right) \left( \sqrt{
1 - \psi^2} + \omega + \psi \omega\right) \left( 1 +
9 \omega^2\right) \right. \nonumber \\
&+& \left. 2 \psi \omega \left( \lambda_{X} \omega \left( \sqrt{
1 - \psi^2} + \omega + \psi \omega\right) \left( 1 +
9 \omega^2\right) \right. \right. \nonumber \\
&-& \left. \left. 4 \left( \frac{M_J^2}{v_s^2} \right) \left( -1 +
3 \sqrt{1 - \psi^2} \omega +
6 \omega^2 + \psi \left( -1 +
6 \omega^2\right)\right)\right)\right] \\
\mathcal{D}_2^{(4)} &=& 8 \psi \left(1 + \psi\right) \left(1 + 9 \omega^2\right)^3 \nonumber \\
& & \left[A \left(-1 + \psi\right) \left(1 + 9 \omega^2\right) +
2 \psi \left(\lambda_{X} \omega^2 \left(1 + 9 \omega^2\right) \right. \right. \nonumber \\
&+& \left. \left. \left(\frac{M_J^2}{v_s^2} \right) \left(1 + \psi -
6 \sqrt{1 - \psi^2} \omega - 3 \omega^2 +
3 \psi \omega^2\right)\right)\right]
\end{eq}
The contributions of the Higgs field to the majoron decay (the fraction $\frac{\lambda_{215}\lambda_{115}}{m_5^2}$) are written below.
The expressions for $\lambda_{215}$, $\lambda_{225}$ and $\lambda_{115}$ are proportional to the ratio $\left( \frac{M_J}{v_s} \right)^4$, thus, having a keV majoron implies a natural supression for the contributions to the decay.
\begin{eq}
\frac{\lambda_{215}\lambda_{115}}{m_5^2} &=& \frac{\mathcal{D}_1^{(5)}}{\mathcal{D}_2^{(5)}} \\
\mathcal{D}_1^{(5)} &=& -1152 \left( \frac{M_J}{v_S} \right)^4 \psi^2 \omega^3 \left[2 \psi \left(5
\lambda_{HS} - \lambda_{HX}\right) \lambda_{X} \omega^3 \right. \nonumber \\
&-& \left. A \left \{\sqrt{1 - \psi^2} \lambda_{HS} + \lambda_{HX} \omega \left(2
\psi - 5 \sqrt{1 - \psi^2} \omega\right)\right\}\right] \nonumber \\
& & \left[-2 \psi \lambda_{X} \omega \left(\lambda_{HS} - 6 \lambda_{HS} \omega^2 +
3 \lambda_{HX} \omega^2\right) \right. \nonumber \\
&-& \left. A \left\{3 \sqrt{1 - \psi^2} \lambda_{HS} + \lambda_{HX} \left(6 \psi \omega + \sqrt{1 - \psi^2}
\left(1 - 6 \omega^2\right)\right)\right\}\right]\nonumber \\ \\
\mathcal{D}_2^{(5)} &=& \left(1 + 9 \omega^2\right)^3 \nonumber \\
& & \left[-A^4 \left(-1 + \psi^2\right)^2 \lambda_{h} \left(1 + 9 \omega^2\right) +
8 A^3 \psi \left(-1 + \psi^2\right) \omega \right. \nonumber \\
& & \left. \left(2 \sqrt{1 - \psi^2} \lambda_{HS} \lambda_{HX} + \psi \left(2
\lambda_{HX}^2 - \lambda_{h} \lambda_{X}\right) \omega\right) \left(1 + 9 \omega^2\right) \right. \nonumber \\
&+& \left. 16 \psi^4 \lambda_{X}^2 \omega^6 \left[8 \left(\frac{M_J}{v_S}\right)^2
\left(3 \lambda_{HS}^2 + 6 \lambda_{HS} \lambda_{HX} - \lambda_{HX}^2\right) \right. \right. \nonumber \\
&+& \left. \left. \lambda_{X}
\left(1 + 9 \omega^2\right) \left(4 \lambda_{HS}^2 + \left(4 \lambda_{HX}^2 - \lambda_{h}
\lambda_{X}\right) \omega^2\right)\right] \right. \nonumber \\
&+& \left. 32 A \psi^3 \lambda_{X} \omega^3 \left[\lambda_{X} \omega \left(1 +
9 \omega^2\right) \left\{2 \sqrt{1 - \psi^2} \lambda_{HS} \lambda_{HX} \omega \right. \right. \right. \nonumber \\
&+& \left. \left. \left. \psi \left(2 \lambda_{HS}^2 + \left(4 \lambda_{HX}^2 - \lambda_{h} \lambda_{X}\right)
\omega^2\right)\right\} \right. \right. \nonumber \\
&+& \left. \left. 4 \left(\frac{M_J}{v_S}\right)^2 \left\{ 3 \sqrt{1 - \psi^2} \lambda_{HS}^2 + \lambda_{HX}^2 \omega
\left(-2 \psi + 3 \sqrt{1 - \psi^2} \omega\right) \right. \right. \right. \nonumber \\
&+& \left. \left. \left. \lambda_{HS} \lambda_{HX} \left(6 \psi \omega + \sqrt{1 - \psi^2}
\left(-1 + 3 \omega^2\right)\right)\right\}\right] \right. \nonumber \\
&+& \left. 8 A^2 \psi^2 \left\{\lambda_{X} \omega^2 \left(1 + 9 \omega^2\right)
\left(2 \left(-1 + \psi^2\right) \lambda_{HS}^2 \right. \right. \right. \nonumber \\
&+& \left. \left. \left. 8 \psi \sqrt{1 - \psi^2} \lambda_{HS} \lambda_{HX} \omega + \left(2
\left(-1 + 5 \psi^2\right) \lambda_{HX}^2 + \left(1 - 3 \psi^2\right) \lambda_{h} \lambda_{X}\right) \omega^2\right) \right. \right. \nonumber \\
&+& \left. \left. 4 \left(\frac{M_J}{v_S}\right)^2 \left[\left(-1 + \psi^2\right) \lambda_{HS}^2 -
2 \lambda_{HS} \lambda_{HX} \omega \left(2 \psi \sqrt{1 - \psi^2} - 3 \omega +
3 \psi^2 \omega\right) \right. \right. \right. \nonumber \\
&+& \left. \left. \left. \lambda_{HX}^2 \omega^2 \left(12 \psi \sqrt{1 - \psi^2} \omega
+ 3 \omega^2 - \psi^2 \left(4 + 3 \omega^2\right)\right)\right]\right\}\right]
\end{eq}
Finally, we show the expressions for the couplings that lead to the Higgs invisible decays $H\longrightarrow 2 \zeta_1$, $H\longrightarrow 2 \zeta_2$
and $H\longrightarrow \zeta_1 \zeta_2$.
These formuli show that the contributions from the new fields $\zeta_{1,2}$ to the invisible Higgs decay are heavily supressed, since they are proportional to $\left( \frac{M_J}{v_s} \right)^2$.
On top of that, observe that these expressions also depend on $\lambda_{HX}$ and $\lambda_{SH}$, couplings that have been taken to be $\ll 1$.
\begin{eq}
\label{eq:l215}
\lambda_{215} &=& -24 \left(\frac{M_J}{v_s}\right)^2 \, \frac{ \, v_h\, \psi\, \omega^2 }{ \left(1 + 9 \omega^2\right)^2 \left\{A^2 \left(-1 + \psi^2\right) +
4 A \psi^2 \lambda_{X} \omega^2 +
4 \psi^2 \lambda_{X}^2 \omega^4\right\} } \nonumber \\
& &\left[\lambda_{HS} \left\{ 3 A \sqrt{1 - \psi^2} +
2 \psi \lambda_{X} \omega -
12 \psi \lambda_{X} \omega^3\right\} \right. \nonumber \\
&+& \left. \lambda_{HX} \left\{6 \psi \lambda_{X} \omega^3 +
A \left(6 \psi \omega +
\sqrt{1 - \psi^2} \left(1 - 6 \omega^2\right)\right)\right\} \right]
\end{eq}
\begin{eq}
\label{eq:l115}
\lambda_{115} &=& -24 \left(\frac{M_J}{v_s}\right)^2 \, \frac{ v_h\, \psi\, \omega}{ \left(1 + 9 \omega^2\right)^2 \left\{A^2 \left(-1 + \psi^2\right) +
4 A \psi^2 \lambda_{X} \omega^2 +
4 \psi^2 \lambda_{X}^2 \omega^4\right\} } \nonumber \\
& & \left[ \lambda_{HS} \left\{ A \sqrt{1 - \psi^2} -
10 \psi \lambda_{X} \omega^3\right\} \right. \nonumber \\
&+& \left. \lambda_{HX} \left\{2 \psi \
\lambda_{X} \omega^3 +
A \omega \left(2 \psi - 5 \sqrt{1 - \psi^2} \omega\right)\right\}\right]
\end{eq}
\begin{eq}
\label{eq:l225}
\lambda_{225} &=& - 72 \left( \frac{M_J}{v_S}\right)^2 \frac{ v_h\, \psi\, \omega^3 }{\left(1 +
9 \omega^2\right)^2 \left\{A^2 \left(-1 + \psi^2\right) +
4 A \psi^2 \lambda_{X} \omega^2 +
4 \psi^2 \lambda_{X}^2 \omega^4\right\}} \nonumber \\
& & \left[2 \psi \lambda_{X} \omega \left\{3 \lambda_{HX} \omega^2 + \lambda_{HS}
\left(2 + 3 \omega^2\right)\right\} \right. \nonumber \\
&+& \left. A \left\{3 \sqrt{1 - \psi^2} \lambda_{HS} + \lambda_{HX} \left(6 \psi
\omega + \sqrt{1 - \psi^2} \left(2 + 3 \omega^2\right)\right)\right\}\right]
\end{eq}
\section{\label{sec:model} The spontaneous inverse seesaw model}
As it was advanced, throughout the work we will focus on the inverse seesaw mechanism
for neutrino mass generation~\cite{Mohapatra:1986bd}. In particular, during this section, we will
embed this model in a scheme where it is spontaneously generated (although efforts in
this way have been presented before, see for instance~\cite{Humbert:2015epa}).
On top of that, throughout this section the particle content of the model and its
interactions will be described.
\subsection{\label{sec:invseesaw} The inverse Seesaw}
Among the different schemes for neutrino mass mechanism, the inverse seesaw
scenario is characterized by 2 mass scales, which are associated to 2 new
fermion singlets per active neutrino species added to the neutrino sector~\cite{Abada:2014vea}.
These new singlets give rise to heavy neutrinos with masses above the TeV scale.
This scenario is adequate to be tested with current or near-future planned experiments.\\
The mass lagrangian for the inverse seesaw can be written as~\cite{Abada:2014vea,Dias:2012xp}:
\begin{eq}
\label{eq:massnu}
\mathcal{L} = - \frac{1}{2} n_L^T C \mathcal{M} n_L \, + \, h.c. \, ,
\end{eq}
where $n_L^T = (\nu_L, N_1^c, N_2)$ is composed by the SM neutrino $\nu_L$ and
the new singlet fermions $N_{1,2}$, then the mass matrix $\mathcal{M}$ is
\begin{eq}
\mathcal{M} &=& \left(\begin{array}{ccc}
0 & m_D & 0 \\
m_D^T & 0 & M \\
0 & M^T & \mu
\end{array} \right) \, . \label{eq:massmat}
\end{eq}
The blocks $M$ and $\mu$ characterize the inverse seesaw and $m_D$ is the Dirac block.
This matrix can be perturbatively diagonalized in a similar way to the type-I seesaw when $\mu \ll m_D \ll M$.
Even though in the present work we are not going to explore neutrino physics,
without loosing generality, we will consider 1 active neutrino and 2 singlets.
This framework will provide just one massive light neutrino.
Under this setup, the masses of this sector at leading order are:
\begin{eq}
m_\nu &=& \left(\frac{m_D}{M}\right)^2 \mu ,\\
m_\mathcal{N} &=& M \,-\, \frac{m_D^2}{M} \,-\, \frac{\mu}{2} \, , \\
m_{\mathcal{N}^{'}} &=& M \,-\, \frac{m_D^2}{M} \,+\, \frac{\mu}{2} \, .
\end{eq}
For active neutrino mass of $m_\nu \sim$0.1~eV, heavy neutrinos with masses
of $M \sim$100~TeV, and a dirac term of $m_D \sim$10~GeV, we require a $\mu$
parameter to be around 10~MeV~\cite{Dias:2012xp}, which matches our requirement
$\mu \ll m_D \ll M$.
In this regime, the neutral fermions mixing matrix is
\begin{eq}
U &=& \left(\begin{array}{ccc}
1 & 0 & m_D/M \\
-m_D/\sqrt{2} M & 1/\sqrt{2} & 1/\sqrt{2} \\
-i m_D/\sqrt{2} M & -i/\sqrt{2} & i/\sqrt{2}
\end{array} \right) \, ,
\end{eq}
where the mass eigenstates are given by $ (\nu, \mathcal{N}, \mathcal{N}^{'}) = \left( U\, n_L \right)^T$ and the mass matrix $\mathcal{M}$ is diagonalized by $m_{\nu}^{diag} = U \mathcal{M} U^T$.
\subsection{The spontaneous inverse seesaw}\label{sec:spinvseesaw}
The mass parameters in the inverse seesaw can be generated by means of spontaneous symmetry breaking (SSB)
of a global $U_{l}(1)$ symmetry associated to the lepton number (see, for instance \cite{Humbert:2015epa,Humbert:2015yva}).
Our approach uses the following lagrangian:
\begin{eq}
\label{eq:slangran}
\mathcal{L} = - y_{L} \bar{L} H N_1^c - y_S S^\dagger \overline{N_2} N_1^c - \frac{y_X}{2} X^\dagger \overline{N_2^c} N_2 + h.c. \, ,
\end{eq}
where $y_i$ are yukawa couplings that after SSB give rise to Eq.~\ref{eq:massnu}.
The Higgs doublet is defined by $H^T = \left( \chi^+, \left(v_h + \sigma_h + i \chi_h\right)/\sqrt{2} \right)$ where
$\sigma_h (\chi_h)$ is the (pseudo)scalar component of the Higgs doublet whose vev is $v_h \simeq 246$~GeV, while $\chi^{+}$ is its charged component.\\
We have included 2 complex scalar $S$ and $X$ charged with lepton number, but both singlets under $SU(2)_L$ and with zero hypercharge.
After the SSB, these fields acquire non-zero vevs, and thus, the mass parameters of the inverse seesaw are defined by:
\begin{eq}
m_D = \frac{y_L v_h}{\sqrt{2}} \, , \, M = \frac{y_{S} v_S}{\sqrt{2}} \, {\rm , and } \, \, \mu = \frac{y_{X} v_X}{\sqrt{2}} \, .
\end{eq}
By fixing the values of $M$ and $\mu$ and since the yukawa couplings cannot exceed the perturbative limit of $\sqrt{4\pi}$, we obtain lower bounds for $v_S$ and $v_X$:
\begin{eq}
v_S > \frac{M}{\sqrt{2 \pi}} \, , \\
v_X > \frac{\mu}{\sqrt{2 \pi}} \, ,
\end{eq}
which can be translated to $v_S > 50$~TeV and $v_X > 5$~MeV for the values previously selected of $M$ and $\mu$.
On the contrary, the value of $m_D$ is completely fixed by $y_L$ since $v_h$ has a determined value.\\
\begin{table}[t]
\centering
\begin{tabular}{|c|ccccc|}
\hline
& \phantom{x} $L$ \phantom{x} & \phantom{x} $N_1$ \phantom{x} & \phantom{x} $N_2$ \phantom{x} & \phantom{x} $S$ \phantom{x} & \phantom{x} $X$ \phantom{x} \\
\hline\hline
\phantom{x} $SU(2)_{L}$ \phantom{x} & $2$ & $1$ & $1$ & $1$ & $1$ \\
\hline
$U(1)_{Y}$ & $1/2$ & $0$ & $0$ & $0$ & $0$ \\
\hline
$U(1)_{l}$ & $1$ & $-1$ & $x$ & $1-x$ & $2x$ \\
\hline
\end{tabular}
\caption{\label{tab:charges} Charge assignment of the model. }
\end{table}
The $U(1)_l$ charges have been assigned by requiring Eq.~\ref{eq:slangran}
to be lepton number invariant, and they are shown at Table~\ref{tab:charges}.
Note that not all the charges can be fixed by Eq.~\ref{eq:slangran}, which leaves
the assignations of the fields $N_2$, $S$ and $X$ completely free as a function
of the lepton number of $N_2$, which is called $x$.
This value can be restricted depending on the scalar potential, an issue depicted in the following subsection.
\subsection{\label{sec:scalpot} Scalar potential}
The scalar potential for the new singlets $S$ and $X$ is given by
\begin{eq}
\label{eq:potsx}
V_{SX} = -\mu_{S}^2 \left|S\right|^2 + \frac{\lambda_S}{4} \left|S\right|^4 - \mu_{X}^2 \left|X\right|^2 + \frac{\lambda_X}{4} \left|X\right|^4 + \lambda_5 \left|S\right|^2 \left|X\right|^2 + V_{\rm I} \, ,
\end{eq}
where $\mu_i^2$ are positive mass terms, $\lambda_i$ are adimensional couplings
allowed by perturbative limit, and $V_I$ is an $S\,-\,X$ interaction term given by:
\begin{eq}
\label{eq:potvi}
V_{\rm I} = \lambda_{\rm cp} e^{i\delta} X {S^{\dagger}}^3 + {\rm h.c.} \, \label{eq:vcp},
\end{eq}
where $\delta$ is a CP-phase for the coupling $\lambda_{\rm cp}$ which is positive\footnote{Different models
with a similar underlying idea can be found at~\cite{Berezinsky:1993fm,Chulia:2016giq,Gu:2010ys,Aranda:2009yb}}.\\
The addition of $V_I$ fixes the lepton number of the new fields as a function of $x$.
For this particular case, the charge assignations are $L_{N_1} = 1$,
$L_{N_2} = x = 3/5$, $L_X = 2x = 6/5$, and $L_S = 1-x = 2/5$, where we
have demanded that ${X^{m}} S^{\dagger n}$ is dimension 4.
We can also take $V_I$ with lower mass dimension in order to keep renormalizability, and thus producing different values of $x$ (see~\ref{app:charge}).\\
After SSB, the fields $S$ and $X$ can be written as:
\begin{eq}
S &=& \frac{1}{\sqrt{2}} \left(v_S e^{i\theta} + \sigma_S\, + i\chi_S \right) \label{eq:seps} \, \\
X &=& \frac{1}{\sqrt{2}} \left(v_X e^{i\tau} + \sigma_X + i\chi_X \right) \label{eq:sepx} \, ,
\end{eq}
where $\theta$ and $\tau$ are CP-phases. These complex phases in the vev are indicating spontaneous violation
of CP. Moreover, considering the role of $\delta$ at~\ref{eq:vcp}, we have two sources of CP-violation in this model,
whose consequence is allowing the mixing scalar/pseudoscalar~\cite{Geng:1988gr}.\\
The remaining terms that include the Higgs part are:
\begin{eq}
V_{\rm HSX} = -\mu_{H}^2 H^{\dagger}H + \frac{\lambda_{H}}{4}(H^{\dagger}H)^2 + \lambda_{HS} \left|S\right|^2 H^{\dagger}H + \lambda_{HX} \left|X\right|^2 H^{\dagger}H \, , \label{eq:higgsmix}
\end{eq}
where $\mu_H^2$ and $\lambda_H$ are the higgs mass parameter and its quartic self-interaction, while $\lambda_{HS}$ and $\lambda_{HX}$ are the couplings between $H$ and the new scalars.\\
The full scalar potential is the sum of Eqs.~\ref{eq:potsx} and \ref{eq:higgsmix},
\begin{eq}
\label{eq:vscal}
V_{\rm scalar} = V_{SX} + V_{HSX} \, .
\end{eq}
The physical fields are extracted after the minimization of the $V_{\rm scalar}$ and plugging back into the Lagrangian the
solution of the tadpole equations:
\begin{eq}
\left. \frac{\partial V_{\rm scalar}}{\partial s_0^i} \right|_{s_0^{j\neq i} =0} = 0 \, \label{eq:tadpoles},
\end{eq}
where ${s^0}^T = \left( \sigma_S,\sigma_X,\sigma_h,\chi_S,\chi_X \right)$, i.e. it represents all the neutral scalar and
pseudoscalar fields of the model but the field $\chi_h$. Afterwards, we still need to write and diagonalize the mass matrices
in order to get the physical fields.
Regarding the tadpole equations for $\chi^{+}$ and $\chi_h$, they are trivially satisfied and do not add relevant information in our context.
On the other hand, the minimization conditions~\ref{eq:tadpoles} give rise to the following relations among the parameters:
\begin{eq}
\tau &=& 3\theta - \delta - \pi \, ,\\
\mu_S^2 &=& \frac{v_S^2}{4} \left( 2 \epsilon^2 \lambda_{HS} + \lambda_S - 6 \lambda_{\rm cp} \omega + 2 \lambda_5 \omega^2 \right) \, , \\
\mu_X^2 &=& \frac{v_S^2}{4} \left( 2 \epsilon^2 \lambda_{HX} - 2 \lambda_{\rm cp}\, \omega^{-1} + 2 \lambda_5 + \lambda_X \omega^2\right) \, , \\
\mu_H^2 &=& \frac{v_S^2}{4} \left( \epsilon^2 \lambda_H + 2 \lambda_{HS} + 2 \lambda_{HX} \omega^2 \right) \, ,
\end{eq}
where $\omega = v_X/v_S$ and $\epsilon = v_h/v_S$ which can be as small as $5 \times 10^{-3}$.
The first condition implies that all CP-phases are aligned and therefore some level of CP violation must be present.
\subsection{\label{sec:spectrum} Mass spectrum}
The mass matrix for the scalar and pseudoscalar fields is obtained from
\begin{eq}
\frac{\partial^2 V_{\rm scalar}}{\partial s^0_i \partial s^0_j} \equiv \left\{M^2\right\}_{ij} \, .
\end{eq}
Recall that we ignore the $\chi^{+,h}$ fields since they are completely decoupled from the rest of scalars,
and thus they correspond to the electroweak goldstone bosons.\\
The CP-phases just mix up the scalar and pseudoscalar sectors inside $M^2$.
However, one can transform the mass matrix in such a way that this mixing is rotated away.
The corresponding rotation matrix depends exclusively on CP-phases and it takes the form,
\begin{eq}
R_{\rm cp} = \left( \begin{array}{ccccc}
c_\theta & 0 & 0 & s_\theta & 0 \\
0 & c_{\delta - 3\theta} & 0 & 0 & s_{\delta - 3\theta} \\
0 & 0 & 1 & 0 & 0 \\
-s_\theta & 0 & 0 & c_\theta & 0 \\
0 & -s_{\delta - 3\theta} & 0 & 0 & c_{\delta - 3\theta}
\end{array} \right) \, .
\end{eq}
As the result of this rotation, the mass matrix takes the following block diagonal form,
\begin{eq}
M_{s}^2 = R_{\rm cp} M^2 R_{\rm cp}^T =
\left( \begin{array}{cc}
M_{\rm scal}^2 & 0 \\
0 & M_{\rm pscal}^2
\end{array} \right) \, ,
\end{eq}
where
\begin{eq}
M_{\rm scal}^2 = \frac{v_S^2}{2}
\left( \begin{array}{ccc}
\lambda_S - 3\lambda_{\rm cp}\omega & 3\lambda_{\rm cp} - 2\lambda_5\omega & 2 \epsilon \lambda_{HS} \\
3\lambda_{\rm cp} - 2\lambda_5\omega & \displaystyle{\frac{(\lambda_{\rm cp} + \lambda_X\omega^3)}{\omega}} & -2 \epsilon \lambda_{HX} \omega \\
2 \epsilon \lambda_{HS} & -2 \epsilon \lambda_{HX} \omega & \epsilon^2 \lambda_H \\
\end{array} \right) \label{eq:matscal}
\end{eq}
corresponds to the scalar mass matrix in the limit $\delta, \theta \rightarrow 0$, and the same happens for the pseudoscalar mass matrix,
\begin{eq}
M_{\rm pscal}^2 = \frac{v_S^2}{2} \, \lambda_{\rm cp} \,
\left( \begin{array}{cc}
9 \omega & 3 \\
3 & \omega^{-1}
\end{array} \right) \, ,
\end{eq}
for vanishing CP-phases.\\
The model has 5 mass eigenstates labeled as $\zeta_i, i=1\dots5$.
The first 2 correspond to those from $M^2_{\rm pscal}$, the rest comes $M^2_{\rm scal}$ where $\zeta_5$ is reserved to the SM-like higgs with a mass of $m_{\zeta_5} = m_{h} = 125$~GeV.\\
The pseudoscalar mass matrix has two eigenstates which are obtained by $R_{\rm pscal} M_{\rm pscal} R_{\rm pscal}^T = diag(m_{\zeta_1}, m_{\zeta_2})$ where
\begin{eq}
R_{\rm pscal} = \frac{1}{\sqrt{1 + 9 \omega^2}}\left( \begin{array}{cc}
1 & -3 \omega \\
3 \omega & 1
\end{array} \right) \, .
\end{eq}
The $\zeta_1$ state corresponds to the goldstone boson of the $U(1)_l$ breaking and it is commonly known as {\it massless} majoron.
The second eigenstate has a mass of
\begin{eq}
m^2_{\zeta_2} = M_J^2 = \frac{v_S^2}{2 \omega} \lambda_{\rm cp} (1 + 9 \omega^2) \label{eq:majmass} \, ,
\end{eq}
and it can be thought as a {\it massive} majoron.
This state will be considered as the DM candidate of the model and we label it as $\zeta_2 = J_{\rm DM}$.
For the purpose of this work, we will keep its mass around the keV range.
Therefore it turns out to be $\lambda_{\rm cp} \simeq \frac{M_J^2}{v_S^2} < 10^{-22}$ for $\omega \simeq \mathcal{O}(1)$ and $v_S > 50$~TeV. \\
On the other hand, the scalar mass matrix provides 3 massive states which can be found by means of a perturbative diagonalization by using $\frac{M_J^2}{v_S^2}$ and $\epsilon$ as perturbative parameters.
In that expansion, the SM-like higgs has a mass of
\begin{eq}
\label{eq:hmass}
m_{h}^2 \simeq \frac{v_h^2}{2} \, \left\{ \frac{\lambda_H}{2} + 2 \left(\frac{ \lambda_{HX}^2 \lambda_S +
\lambda_{HS}^2 \lambda_X -4 \lambda_5 \lambda_{HS} \lambda_{HX}}{4 \lambda_5^2 - \lambda_S \lambda_X} \right) \right\} \, ,
\end{eq}
which is valid within this limit.
The constraint due to the higgs mass provides restrictions on the values of the couplings of the second term (Eq.~\ref{eq:hmass}).
A simplified version for the constraint arises when taking $\lambda_5 = 0$ i.e. $\sigma_S$ and $\sigma_X$ as quasi-decoupled states.
In this case, the higgs mass is
\begin{eq}
m_{h}^2 \simeq \frac{v_h^2}{2} \, \left\{ \frac{\lambda_H}{2} - 2 \left( \frac{\lambda_{HX}^2}{\lambda_X} + \frac{\lambda_{HS}^2}{\lambda_S} \right) \right\} \, ,
\end{eq}
and it gets a contribution from the couplings $\lambda_{HX}$, $\lambda_{HS}$, $\lambda_{X}$, and $\lambda_{S}$.
Regardless of the size of the couplings, the mixing between $\sigma_h$ and $\sigma_S$--$\sigma_{X}$ is suppressed by terms $\mathcal{O}(\epsilon)$,
this produces that $\zeta_5 = \sigma_h + \mathcal{O}(\epsilon) \cdot \left(\sigma_{S,X}\right)$ and thus the $\zeta_5$ is mostly SM-like higgs.
Nevertheless, throughout the paper we will stick to the limit $\lambda_{HX}$, $\lambda_{HS} \ll 1$, which guarrantees a higgs-like $\zeta_5$ as well
and a small mixing of $\sigma_h$ with $\sigma_S$--$\sigma_{X}$.\\
In the same perturbative scheme, the remaining 2 massive states ($\zeta_{3,4}$) have masses:
\begin{eq}
M_{\zeta_3}^2 &\simeq& \frac{v_S^2}{2} \left( \frac{-A\,+\,A\psi\, +\, 2\lambda_X\omega \psi}{2\psi} \right) \, , \\
M_{\zeta_4}^2 &\simeq& \frac{v_S^2}{2} \left( \frac{ A\,+\,A\psi\, +\, 2\lambda_X\omega \psi}{2\psi} \right) \, ,
\end{eq}
where $A$ and $\psi$ come from:
\begin{eq}
\lambda_S &=& A\, +\, \lambda_X \omega^2 \, ,\\
\lambda_5 &=& -A\, \left( \frac{\sqrt{1 - \psi^2}}{ 4\, \omega\, \psi} \right) \, .
\end{eq}
The $A$ parameter can be seen as an alignment between $\lambda_S$ and $\lambda_X$ and therefore it has a range value of a typical adimensional coupling.
The $\psi$ term is $\cos{\left(\phi/2 \right)}$ where $\phi$ is the mixing angle of the sector $\sigma_S$ and $\sigma_X$, and since it is a trigonometric function, $\lambda_5$ can take positive and negative values.
Without imposing any fine tunning of $A$ and $\psi$, the mass values for $\zeta_{3,4}$ are expected to be $\mathcal{O}\left(v_S\right)$.\\
The mass spectrum of the scalar sector has 3 well defined mass scales.
First, we have the light states ($< \mathcal{O}({\rm keV})$) corresponding to the massless and massive majorons.
The second scale is determined by the mass of SM-like higgs.
And the last one corresponds to the states $M_{\zeta_3}$ and $M_{\zeta_4}$, given by $v_S$ ($> 50$~TeV).
\section{\label{sec:conc}Conclusions}
In this work, we propose an extension of SM where neutrinos are Majorana particles and they become massive through an inverse seesaw mechanism which arises from the spontaneous symmetry breaking of the lepton number.
Our model allows us to have a massive Majoron as a DM candidate.
This latter particle has the following characteristics: \emph{i)} Its mass comes from the mixing in the pseudoscalar sector.
The mixing arises due to the lepton number charges needed by the neutrino and scalar sectors to make the lagrangian invariant under lepton number.
The mass range could go from the keV's up to TeV's, although we have explored just the keV region.
\emph{ii)} It is metastable and its main decay channels are to neutrinos and massless
Majorons.
These channels are similar to models with Majorons as \emph{pseudo-goldstone bosons}.
For simplicity, we test the model assuming one family of active neutrinos, although the extension to 3 families can be easily implemented.
The introduced scalars, that give rise to the inverse seesaw mechanism, also allow the spontaneous breaking of CP invariance.
Nevertheless the effect is not present in our model's phenomenology because we included just one family of active neutrinos.
The DM candidate stability is very fragile in this model because we did not include any \emph{ad-hoc} stabilizing symmetry.
However, we found that there is always a region in the parameter space where the massive Majoron has a lifetime longer than $10^{27}$~s and, therefore, it can be considered as a good DM candidate.
Moreover, we found that the ratio among vevs, $\omega$, has a very important role in the decay channel to neutrinos.
The value $\omega = \sqrt{2/3}$ can vanish the decay mode to neutrinos presenting a tantalizing vev alignment for model building.
The scalar decay modes are the most crucial because the drastical effect on the total DM lifetime and from the point of view of scan of the parameter space.
Nevertheless, we found that the decay width vanished in a region of the parameter space of the scalar sector.
We also discussed how to rid off the scalar decay modes by promoting the global lepton symmetry to a local one.
We discuss possible ways on how to estimate the DM relic abundance in terms of a freeze-in scenario.
In general words, the model presents an interesting relation between the neutrino mass mechanism and origin of the massive Majoron as a DM candidate.
|
\section{Introduction}
HR\,7428 (=~V~1817~Cygni) is a bright (V=6.3) long-period (108.578d) spectroscopic \mbox{RS\,CVn}\ binary composed by a K2 II-III star and a main sequence A2 star \citep{Parsons&Ake87}. The magnetic activity of the system is well known: \ion{Ca}{ii}\,H\,\&\,K\ emission was first reported by \cite{Gratton50}, by a detailed analysis of photometric observations \cite{Hall_etal90} were able to detect starspot signatures on the K2 primary star. It is now well established that the stellar atmosphere of cool stars is characterized by that temperature gradient inversion. An inversion explained in the framework of the magnetic activity theories, but not yet definitively understood.
\begin{figure*}
\begin{flushleft}
\includegraphics*[width=1.0\textwidth]{caos.eps}
\end{flushleft}
\caption{HR7428 normalized spectrum obtained using the new Catania Astrophysical Observatory Spectropolarimeter (CAOS) showing the line profiles of \mbox{H-$\beta$}, \ion{Na}{i}\,D, \mbox{H-$\alpha$}\ and \ion{Ca}{ii}\,IRT\ lines.}
\label{caos}
\end{figure*}
Late-type stars, with \mbox{H-$\alpha$}\ in emission, show also a fairly stable chromospheric emission outside flares, see. e.g., \cite{Byrne_etal98}. This reinforces the hypothesis that chromospheres are globally in a quasi-stationary state, modulated mainly by the stellar activity cycle, whose temperature-density structure results from the balance between global dissipation of
non-radiative energy and radiative cooling \citep{Kalkofen_etal99}.
Most of what we know about stars and systems of stars is derived from an analysis of their radiation and this knowledge will be secure only as long as the analytical technique is physically reliable. A well tested technique to get information on physical properties of chromospheric layers of active stars is the NLTE radiative transfer semi-empirical modeling: for different temperature vs height distributions, the NLTE populations for hydrogen are computed, by solving simultaneously the equations of hydrostatic equilibrium, radiative transfer, and statistical equilibrium. The emerging profiles for some chromospheric lines and continua are computed and compared to the observations. Then, the modeling is iterated until a satisfactory match is found. (see, e.g.,\cite{Vernazza_etal81}; \cite{Fontenla_etal93}). These models are built to match the observations in different spectral features, and make no assumption about the physical processes responsible for the heating of the chromosphere, but can be used as constraints for these processes. The obtained models describe the variations of the essential physical parameters, in particular the temperature, pressure and electron density across the outer atmosphere, and give information on its ''mean'' state, both temporally and spatially.
The most important problem of this approach lies in the uniqueness of the solution. In fact, knowing that a particular atmosphere
would emit a line profile like the one we observe for a given star does not imply that the star has indeed this atmospheric structure, since we do not know
whether some other atmosphere would produce the same profile. To solve, or
at least to reduce, this problem, the modeling has to be based on several spectral features, with different regions of formation. The amount and the kind of diagnostics used to build an atmospheric model is in fact, very important, by combining several spectral lines that are formed at different but overlapping depths in the atmosphere, we can obtain a more reliable model \citep{Mauas_etal06}.
Certainly the best known semiempirical model is the one for the average Quiet-Sun, Model C by \cite{Vernazza_etal81}.
Semiempirical modeling was successfully applied also to the atmospheres of cool stars. An extensive modeling of dM stars, has been done, starting with the work by \cite{Cram&Mullan79}, \cite{Short&Doyle98}, \cite{Mauas&Falchi94}, and \cite{Mauas_etal97}. Furthermore, a rich history of semiempirical chromospheric modeling has been also carried out for cool giant and supergiant stars (see, e.g.,\cite{Kelch_etal78}; \cite{Basri_etal81}; \cite{Luttermoser_etal94}). In cool stars the application of NLTE semi-empirical chromospheric modelling can be based on optical and ultraviolet (UV) observations. This is because lines such as the \mbox{H-$\alpha$}\ \ion{Na}{i}\,D\ \ion{Ca}{ii}\,IRT\ become dominated by electron-collision excitation processes, which make them effective chromospheric diagnostics \citep{Houdebine_96}. The possibility of using \mbox{H-$\alpha$}\ profile as a diagnostic of stellar chromospheres was discussed in detail by \cite{Cram&Mullan79} and \cite{Mullan&Cram82} in terms of control of the source function by photons or collisional processes. \mbox{H-$\alpha$}\ is observed in active stars in a wide variety of shapes and sizes; when the effective temperature is low enough, the collisional control of the \mbox{H-$\alpha$}\ source function becomes possible over a wide range of chromospheric pressures, and, under this conditions, \mbox{H-$\alpha$}\ can be a good chromospheric diagnostic. \ion{Mg}{ii}\,h\,\&\,k\ UV lines, due to their large opacity, provide excellent diagnostic over a wide range of heights of the outer chromospheric layers \citep{Uitenbroek92}, and \ion{Ca}{ii}\,IRT\ triplet is a constraint for the shape of the middle chromosphere from the temperature minimum up to the plateau \citep{Andretta_etal2005}.
Here we applied the NLTE semi-empirical chromospheric modelling to the K2 star of HR~7428 binary system basing the analysis on the \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Na}{i}\,D, \ion{Ca}{ii}\,IRT\ triplet, \ion{Mg}{ii}\,h\,\&\,k\ lines and UV continuum diagnostics.
\label{Sec:intro}
\section{Data acquisition and reduction}
\label{data}
\begin{figure}
\begin{flushleft}
\includegraphics*[width=0.5\textwidth]{mgii_iue.eps}
\caption{HR7428 high-resolution flux calibrated spectrum in the wavelength of the \ion{Mg}{ii}\ line profile obtained by International Ultraviolet Explorer (IUE).}
\label{mgii_iue}
\end{flushleft}
\end{figure}
\begin{figure}
\begin{flushleft}
\includegraphics*[width=0.5\textwidth]{mgvar.eps}
\caption{Comparison of two IUE spectra of the system, obtained in different dates: LWP26775HL.FITS obtained in November 1997 and LWR10313HL.FITS obtained in December 1995.}
\label{mgvar}
\end{flushleft}
\end{figure}
\mbox{H-$\alpha$}\ \ion{Na}{i}\,D\ \mbox{H-$\beta$}\ \ion{Ca}{ii}\,IRT\ spectroscopic observations of HR\,7428 were carried out at the 91-cm telescope of Catania Astrophysical Observatory, ''M. G. Fracastoro'' station (Serra La Nave, Mt. Etna, Italy), using the new Catania Astrophysical Observatory Spectropolarimeter (CAOS) which is a fiber fed, high-resolution, cross-dispersed echelle spectrograph (\cite{leone_etal16}; \cite{Spano_etal04}, \cite{Spano_etal06})
The spectra were obtained in September 2015. Exposure times have been tuned in order to obtain a signal-to-noise ratio of at least 200 in the continuum in the 390-900 nm spectral range, with a resolution of R=$\frac {\lambda}{\Delta(\lambda)}$~=~45,000, as measured from ThAr and telluric lines.
Echelle IRAF packages have been used for data reduction, following
the standard steps: bias subtraction, background subtraction, trimming, flat-fielding and scattered light subtraction, extraction for the orders, and
wavelength calibration. Several thorium lamp exposures were obtained
during each night and then used to provide a wavelength calibration of the observations.
Each spectral order was normalized by a polynomial
fit to the local continuum.
A reduced spectrum, in the wavelength ranges of the line of interest is shown in
Fig.~\ref{caos}. In the plotted spectrum the average S/N obtained at the continuum
close to \mbox{H-$\alpha$}\ is about 200.
\ion{Mg}{ii}\,h\,\&\,k\ spectroscopic observations have been obtained, in 1997, by the IUE satellite.
IUE spectra have been corrected for interstellar extinction.
According to the Hipparcos distance d~=~323$\pm$52~pc and to the typical
value of 1 Mag per kilo-parsec for interstellar extinction we adopted for the
computation A(V)~=~0.32. Assuming the standard reddening law $A(V)$~=~3.1 $\times$ E(B-V), a color excess E(B-V)~=~0.10 has been derived. IUE spectra have been de-reddened according to the selective extinction function of \cite{Cardelli_etal89}.
The spectral resolution is about 0.2 \AA\ for the \ion{Mg}{ii}\ region.
The IUE flux calibrated HR7428 spectrum, is shown in Fig.~\ref{mgii_iue}.
Unfortunately, we have no simultaneous UV and optical observations.
Therefore we have to take into account the activity-variability that could affect the HR7428 system, due to different levels of the stellar activity and/or different distribution of the active regions on the visible surface, in different times. In order to have an estimate of how much UV data are affected by activity-variability we have compared two IUE spectra of the system, obtained in different dates: LWP26775HL.FITS obtained in November 1997 and LWR10313HL.FITS obtained in December 1995 (see Fig.~\ref{mgvar} where the two \ion{Mg}{ii}\,h\,\&\,k\ IUE spectra are shown with a shift in wavelength in order to overlap.)
$1.060 \times 10^{-12} \pm 1 \times 10^{-15}$
We find that we can neglect the long term variability as far as the UV continuum is concerned, the two spectra show in fact the same continuum mean value (\ion{Mg}{ii}\ Continuum Flux(1995)~=~$1.06 \times 10^{-12} \pm 1 \times 10^{-14}$ erg cm$^{-2}$\,s$^{-1}$, \ion{Mg}{ii}\ Continuum Flux(1997)~=~$1.060 \times 10^{-12} \pm 7 \times 10^{-15}$ erg cm$^{-2}$\,s$^{-1}$). As far as the \ion{Mg}{ii}\,h\,\&\,k\ line profiles are concerned, we measure approximately equivalent observed fluxes at Earth ($Flux_{\small \ion{Mg}{ii}\,h\,\&\,k}$(1995)~=~$3.33 \times 10^{-11} \pm 3 \times 10^{-13}$ erg cm$^{-2}$\,s$^{-1}$, $Flux_{\small \ion{Mg}{ii}\,h\,\&\,k}$(1997)~=~$3.20 \times 10^{-11} \pm 3 \times 10^{-13}$ erg cm$^{-2}$\,s$^{-1}$), but the profile shape is quite different as shown in Fig.~\ref{mgvar}, most probably indicating a different distribution of active regions in different times, that we cannot take into account. Therefore an higher weight will be done to the best fit of the UV continuum with respect to the \ion{Mg}{ii}\,h\,\&\,k\ line profile.
\section{Computational method}
Historically, K2 giants were classified as "non-coronal" stars, however, in the late 1990s, evidence for transition region emission was detected for Arcturus (K2 III) and Aldebaran (K5 III) \citep{Ayres_etal03} and for other representative ''non-coronal'' red giants (see, e.g.,\cite{Ayres_etal97}; \cite{Robinson_etal98}) Therefore, here, the atmospheric model of the K2 primary magnetic active component has been built computing a photospheric model, a chromospheric model and a transition region model and joining the three together; we assume a plane-parallel geometry in our modeling efforts.
\subsection{Photospheric Model}
Taking into account the \cite{Marino_etal01} physical parameters (see Table~\ref{mar_par}) we selected from the Castelli LTE synthetic spectra database (http://www.oact.inaf.it/castelli/castelli/grids.html), a spectrum with parameters \ensuremath{\log g}\~=~2.0,
$T\rm_{eff}$\~=~4400~K and solar metalicity that describes the photospheric contribution of the K2 primary component, and a spectrum with parameters \ensuremath{\log g}\~=~4.0, $T\rm_{eff}$\~=~9000~K and solar metalicity that describes the flux contribution of the A2 secondary component of the binary system HR~7428.
\begin{table}
\caption{HR\,7428 K2II-III and A2 components \citep{Marino_etal01}.}
\label{mar_par}
\begin{center}
\begin{tabular}{ccc}
\hline
\\
Element & Primary (cooler K2II-III) & Secondary (hotter A2) \\
\hline
\\
R &$ 40.0 \pm\ 6.5 R_{\odot} $ & $ 2.25 \pm\ 0.5 R_{\odot} $ \\
$T\rm_{eff}$\ &$ 4400~K \pm\ 150~K $&$ 9000~K \pm\ 200~K $\\
\ensuremath{\log g}\ &$ 2.0 \pm\ 0.5 $ & $ 4.0 \pm\ 0.5 $ \\
\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\begin{flushleft}
\includegraphics*[width=0.5\textwidth]{continui_mg.eps}
\end{flushleft}
\caption{HR7428 Castelli \mbox{LTE}\ synthetic photospheric fluxes at Earth for the K star (dotted line) and A star (dashed line) where the conversion factor $R^2$/$d^2$ has been calculated for the K and A stars according to the parameters given in Tab.\ref{mar_par}. The plot zoom around 2000-5000~$ \AA\ $ shows that before $ \lambda\ $~=~3200~$ \AA\ $ the A star dominates the continuum emission, while for wavelength greater than $ \lambda\ $~=~3200~$ \AA\ $ we can neglect the A star contribution, the continuum is in fact dominated by the K star.}
\label{pesostellaa}
\end{figure}
In Fig.~\ref{pesostellaa} the fluxes at Earth of the two LTE models are shown together with their sum. The original fluxes have been converted to flux at Earth by the conversion factor $R^2$/$d^2$ that has been calculated for the K and A star according the parameters given in Tab.\ref{mar_par}. From the plot we conclude that, for \mbox{NLTE}\ radiative transfer calculations of the \mbox{H-$\beta$}, \ion{Na}{i}\,D, \mbox{H-$\alpha$}, \ion{Ca}{ii}\,IRT\ triplet profiles of the binary system, we can neglect the A2 star contribution. In fact, from Fig.~\ref{pesostellaa} we can see that for wavelength longer than 3200 \AA, the continuum flux of the binary system is dominated by the K star. This is in agreement with \cite{Marino_etal01} that find the A2 star contribution to the total flux in the \mbox{H-$\alpha$}\ region is only 4\%. On the other hand, in order to calculate the \mbox{NLTE}\ \ion{Mg}{ii}\,h\,\&\,k\ (wavelength lower than 3200 \AA) HR\,7428 line profile, we have to take into account the A2 star contribution. In the \ion{Mg}{ii}\,h\,\&\,k\ spectral range in fact (triangles show the observed IUE \ion{Mg}{ii}\,h\,\&\,k), the observed continuum is dominated by the A star contribution.
Therefore, we computed the A2 secondary component atmospheric model
by the ATLAS9 code \citep{Kurucz93} using the parameters \ensuremath{\log g}=4.0, $T\rm_{eff}$= 9000~K, \mbox{[A/H]}=0.0 .
The atmospheric model of the K2 primary magnetic active component has been built computing separately a photospheric model, a chromospheric model and a transition region model and joining the three together.
The K2 photospheric model was computed using the ATLAS9 code and the parameters \ensuremath{\log g}=2.0, $T\rm_{eff}$= 4400~K, \mbox{[A/H]}=0.0 for the K2 primary component.
\subsection{Transition Region Model}
A first estimation of a plane-parallel model for the lower transition region of the HR\,7428 K2 primary component, has been built using the method of the Volumetric Emission Measure. The use of Emission Measure techniques to construct transition region models is well established (see for example \cite{JordanBrown81}, \cite{Harper92}), the flux at the star, for lines forming at temperature $T_e$~$\approx$~$10^5$~K is in fact dominated by collisions. This results in emission lines that are optically thin and with a contribution function sharply picked in temperature, that is, typically formed over a temperature range of $\Delta\log(T_e)$~=~0.30 .
The above conditions allow to determine the temperature gradient as a function of the averaged Emission Measure over $\Delta\log(T_e)$~=~0.30 that we indicate as $EM_{0.3}$.
By imposing hydrostatic equilibrium, including turbulent pressure, the transition region model can be obtained combining the temperature gradient as a function of $EM_{0.3}$ and the pressure variation from the equation of hydrostatic equilibrium (see,e.g., \citet{Harper92}).
The expression for the temperature gradient combined with the equation of hydrostatic equilibrium gives the relationship:
\begin{equation}
\small
\begin{tabular}{l}
$P_{Tot}^2$($T_2$) - $P_{Tot}^2$($T_1$)~=~2 $\times$ $(1.4)^2 \, m_p \,g \,k \,\times$ \\
$\times \, \int_{T_1}^{T_2}{[EM_{0.3} \times \left(1 \dotplus 1.1 x \right) + \frac{1.4\,x\, m_p \, v_{turb}^2 EM_{0.3}}{2\, k\, T} ] dT}$\\
\end {tabular}
\label{eqintegr}
\end {equation}
\begin{figure}
\begin{flushleft}
\includegraphics*[width=0.5\textwidth]{modelgrid_upperhr7428.ps}
\end{flushleft}
\caption{Grid of transition region models used for the study of HR7428 atmosphere. The grid has been built as described in the text.}
\label{grigliatr}
\end{figure}
\noindent
where $m_p$ is the $H^{+}$ mass, $k$ is the Boltzmann constant, $v_{turb}$ is the turbulent velocity and x$\equiv$$N_H/N_e$.
The Eq. \ref{eqintegr} together with an estimate of electron density and turbulent velocity in a layer, allows to find the pressure as a function of temperature for the TR model. The total particle density ($N_{tot}$), gas pressure
($P_G$), turbulence pressure ($P_{turb}$) and electron density ($N_e
$) are then obtained according to the following relations
\begin{equation}
\begin{tabular}{l}
\label{tante}
$N_{tot}$~=~$ \sqrt{P_{Tot}}$ \,(k\, T + 0.5 $m_p$ \,$v_{turb}^2$)\\
$P_G$~=~$ N_{tot}$\,k\,T \\
$P_{turb}$~=~$ 0.5 N_{tot}$ \,$m_p$ \,$\mu$ \,$v_{turb}^2$\\
$N_e$~=~$ N_{tot}$ / (1.+1.1\,x) \\
\end {tabular}
\end {equation}
where $\mu \equiv \frac{1.4 N_H/N_e}{\left(1 \dotplus 1.1 N_H/N_e \right)}$ is the molecular weight.
We used the equations above in order to build plane-parallel models for the lower transition region of the HR\,7428 K2 primary component.
In order to have an estimate of Emission Measures for the HR\,7428 binary system, we used, the
values of Volumetric Emission Measure ($VEM$) vs $T\rm_{eff}$\ measured by \cite{GriffithsJordan98} for another \mbox{RS\,CVn}\ system, HR\,1099 primary component ($R_{HR1099 \,K1IV}$~=~3.9 $R{_\odot}$), opportunely scaling them, in order to take into account the bigger radius ($R_{K2}$~=~40 $R{_\odot}$) of HR\,7428 K2 star, according the formula $EM_{0.3}$~=~$VEM/4\pi R^2$ \citep{Brownetal91} and the parameters of Table, \ref{mar_par}.
From the estimated $EM_{0.3}$ values, a grid of 160 transition region models, shown in Fig.~\ref{grigliatr} has been built by means of equations \ref{eqintegr} and \ref{tante} using a grid of 23 values of electron density at the fixed temperature $T_e$~=~50000~K obtained scaling of a factor from 0.5 up to 30 the values of electron density at 50,000~K measured in HR\,1099 ($N_e$~=~5~$\times 10^{+11}$cm$^{-3}$). For each of these 23 electron density values, a grid of seven values of the turbulent velocity in the layer with $T_0$~=~$10^4$~K, $v_{turb}\left(T_0 \equiv ~10^4~ \mbox{K} \right)$~=~10, 20, 30, 40, 50, 60, 70 km~s$^{-1}$\ has been considered for the calculation of turbulent velocity distribution according the empirical law by \cite{GriffithsJordan98} $v_{turb}$(T)=$v_{turb}\left(T_0\right) \times \left(T/T_0\right)^{1/4}$ between $\log{(T)}$~=~4.0 and $\log{(T)}$~=~5.3
These transition region models provide the upper boundaries for the radiative transfer calculations of the chromospheric models, while the adopted photospheric model provides the lower boundaries of the chromospheric models.
\subsection{Chromospheric Model}
\begin{figure}
\begin{flushleft}
\includegraphics*[width=0.5\textwidth]{chromosperic_modelgrid_hr7428.eps}
\end{flushleft}
\caption{Grid of chromospheric structures obtained by spline interpolation between a fixed Transition Region model and a fixed Photospheric model, using as free parameters a grid of 25 knots point. The figure shows in dotted lines the models that do not satisfy the negative gradient condition $dT/dm \leq 0$.}
\label{25modello_chr}
\end{figure}
The term 'chromosphere' indicates the region above a stellar photosphere where non-radiative heating processes (either magnetic or acoustic) become important in the energy balance and where hydrogen is partially ionized. Since hydrogen is nearly completely ionized when temperatures reach 20,000-30,000~K, we consider this the top of chromosphere. Once hydrogen is nearly completely ionized, mechanical heating produces a steep thermal gradient because there is no cooling channel as effective as hydrogen, and \ion{Ca\,\textsc{II}}, \ion{Mg}{ii}\ have disappeared before hydrogen is ionized \citep{linsky_01}.
Chromospheres lies in the difficult regimes of non-LTE and non-ionization equilibrium. While photosphere can be calculated in detail when having \ensuremath{\log g}, \mbox{[A/H]}\ and $T\rm_{eff}$, and transition region can be constrained by observations because lines form in optically thin conditions as seen before, for the chromospheric layers we have no constraints, lines are optically thick, and the chromospheric model has to be based upon spectral diagnostic methods by means of semi-empirical modeling technique, that is changing an hypothetic model iteratively, in order to match as many as possible observations that form in the chromospheric layers. We can only suppose that the temperature increases with a smooth thermal gradient because, the non-radiative sources heats the plasma producing bright emission in the \mbox{H-$\alpha$}, \ion{Ca}{ii}\,IRT, \ion{Mg}{ii}\,h\,\&\,k\ which together are the dominant cooling channels of the chromosphere. In the Sun the chromosphere goes from $\approx$~600~km and $\approx$~2000~km above the photosphere and is characterized by temperature gradient sign change, where temperature smoothly grow from $\approx$~4500~K up to $\approx$~15000~K over which the gas density changes by several orders of magnitude due to a sharp decrease of turbulent pressure while the electron density only smoothly decrease.
The solar chromosphere has been described for the first time by \citet{Vernazza_etal73} by means of a one-component model of the solar atmosphere, including in that model the photosphere, chromosphere and transition zone, and then in detail, by \citet{Fontenla_etal99}, who using the PANDORA NLTE radiative transfer code determine semiempirical models for seven semiempirical models for sunspots, plages, network, and quiet atmosphere constructed to reproduce observed emergent intensities and profiles at wavelengths from the UV to radio wavelengths.
While \citet{Vernazza_etal73} determine the chromospheric model of the Sun by adjusting the temperature as a function of height to that distribution that gives best agreement between synthesized and observed line spectra, here we have built a wide grid of chromospheric models from which to look for the best model by means of a $\chi^2$ minimization selection method.
In particular, for each one of the 160 transition region model and for each of nine values of $T_{min}$ in the range between $\sim$2800~K and 4200~K that we have chosen with a step of less than 200~K as points where to cut the photospheric Kurucz model, a grid of 25 chromospheric models are generated by a smooth spline interpolation between the photosphere and the transition region using as free parameters a grid of $5 \times 5$ interpolation knots, (see Fig.~\ref{25modello_chr}, where the grid of 25 chromospheric models is shown for a fixed TR model and a fixed minimum of photospheric model $T_{min}$. We impose the chromospheric structures to have a monotonic temperature dependence on Column Mass ($dT/dm \leq 0$). Fig.~\ref{25modello_chr} shows how the temperature gradient constraint, strongly cuts the number of useful model of the grid, for example in the 25 model grid of the figure, only eight models satisfy the gradient constraint and can be used in the analysis.
The total grid of models includes 160*9*25=36225 models, and only 15691 satisfy the $dT/dm \leq 0$ constraint and have been considered in the \mbox{NLTE}\ radiative transfer.
\subsection{Computation: applying or not hydrostatic equilibrium equations}
\begin{figure*}
\includegraphics*[width=0.9\textwidth]{figura4hse_nohse_1.eps}
\caption{\mbox{NLTE}\ \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Ca}{ii}\,IRT, \ion{Na}{i}\,D\ normalized profiles computed for the best $HSE$ (dotted line) and the best $NOHSE$ (dashed line) models compared with observations.}
\label{righe}
\end{figure*}
The coupled equations of radiative transfer and statistical equilibrium were
solved using the version 2.2 of the code \mbox{\it Multi}\ \citep{Carlsson86}, for the \ion{H}{}, \ion{Ca}{}, \ion{Na}{}, \ion{Mg}{} atomic models.
The \ion{H}{} atomic model incorporates 16 states of \ion{H}{}, with 84 {\it b-b}
and 9 {\it b-f} transitions. The \ion{Ca}{} atomic model incorporates 8 states
of \ion{Ca}{i}, the lowest 5 states of \ion{Ca}{ii} and the ground state of
\ion{Ca}{iii}, 9 {\it b-b} and 13 {\it b-f} transitions are treated in detail.
The \ion{Mg}{ii}\ atomic
model is made of 3 states \ion{Mg}{i}, the lowest 6 states of \ion{Mg}{ii} and the ground state of \ion{Mg}{iii}, 9 {\it b-b} and 9 {\it b-f} transitions are treated in detail.
The \ion{Na}{} atomic model incorporates 12 levels: 11 levels of \ion{Na}{i} and the ground state of \ion{Na}{ii} and 29 {\it b-b} and 11 {\it b-f} transitions are treated in detail.
The opacity package included in the code takes into account free-free opacity, Rayleigh scattering, and bound-free transitions from hydrogen and metals, We included the line blanketing contribution to the opacity using the method described in \cite{Busa_etal01}.
\begin{figure}
\includegraphics*[width=0.5\textwidth]{figura4hse_nohse_2.eps}
\caption{\ion{Mg}{ii}\,h\,\&\,k\ absolute flux profiles computed from the best $HSE$ and the best $NOHSE$ atmospheric models compared with the IUE observation.}
\label{mg}
\end{figure}
As a first step we imposed hydrostatic equilibrium to the hydrogen, that is, we calculated the hydrogen radiative transfer and the equation of hydrostatic equilibrium consistently.
In detail, starting with the LTE hydrogen populations for the electron pressure and density calculation, hydrogen is iterated to convergence. Then hydrostatic equilibrium $dP_{Tot}$~=~$\rho g\, dh$ is solved and electron pressure and hydrogen populations are updated, the loop continues until we obtain a convergence.
The electron density obtained is then used to solve the \ion{Ca}{}, \ion{Mg}{}, and \ion{Na}{} radiative transfer and statistical equilibrium equations. The population densities obtained from the \ion{H}{} calculation are used to obtain the background NLTE source function in the \ion{Ca}{}, \ion{Mg}{}, and \ion{Na}{} calculations. This computation modifies the initial grid because a new column of electron density is obtained, we find that only 2052 models converge to a solution and we call them HSE models.
\begin{figure*}
\includegraphics*[width=0.9\textwidth]{figura4_4.eps}
\caption{Contribution functions of the \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Mg}{ii}\,h\,\&\,k, \ion{Ca}{ii}\,IRT, \ion{Na}{i}\,D\ lines for the best $NOHSE$ model. The maximum of the contribution function is set equal to one. The contour plot indicates
the fractions 0.9 (solid), 0.3 (dashed) 0.01 (dotted). The atmospheric model is plotted as a solid line in each $CF$ plot. The last plot, where all lines contribution functions are plotted together, indicates that our selected diagnostics mainly constraint the whole atmosphere from the photosphere (\mbox{H-$\alpha$}\ and \mbox{H-$\beta$}\ wings), and chromosphere (all the lines) up to the transition region (\ion{Mg}{ii}\,h\,\&\,k\ cores) giving strength to our semiempiral modeling.}
\label{cntrbfnohse}
\end{figure*}
We also considered not imposing hydrostatic equilibrium that means fixing the electron densities to the ones of the original grid.
In this case we consider all the 15691 models that only satisfy the $dT/dm \leq 0$ constraint and we call these models, $NOHSE$ models.
Even if an atmospheric model in non-hydrostatic equilibrium is not realistic and should be rejected, here we assume the $NOHSE$ models as good as the $HSE$ ones. Of course hydrostatic equilibrium has to happen in the star, but, because we are not taking into account all the pressure contributions to the total pressure, therefore we can assure that also the $HSE$ models should be rejected, or, at least can be, good or not, just like the $NOHSE$ models. In a sense, we can assume that, imposing $HSE$, the resulting model is not in hydrostatic equilibrium if a contribution to the pressure is missing in the $dP_{Tot}$~=~$\rho g\, dh$ equation.
\begin{figure*}
\includegraphics*[width=0.9\textwidth]{figura4hse_4.eps}
\caption{Contribution functions of the \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Mg}{ii}\,h\,\&\,k, \ion{Ca}{ii}\,IRT, \ion{Na}{i}\,D\ lines for the best $HSE$ model. Like for the $NOHSE$ model (see Fig.~\ref{cntrbfnohse} also for the $HSE$ model the outer atmosphere is enough constrained by our diagnostics.)}
\label{cntrbfhse}
\end{figure*}
\begin{figure}
\includegraphics*[width=0.5\textwidth]{figura4_6.eps}
\caption{Best atmospheric $HSE$ and $NOHSE$ models for the mean K2 primary component of the \mbox{RS\,CVn}\ system HR~7428 compared with the Solar Val~3c model.The plot describes the distribution of temperature versus height. Height is given in kilometers measured above a zero point where $\tau_{5000}$~=~1.}
\label{bestmodel}
\end{figure}
And this is the case of cool active stars.
We know in fact that the atmospheres of active stars are permeated by magnetic fields that emerge from deeper layers. With increasing height we might expect the structure to be more greatly influenced by magnetic fields, since the energy density of the magnetic fields should fall off more slowly than the energy density of the gas (this is the case of solar atmosphere where $\beta$~=~$8 \pi N K T_e / |(B)|^2$ is $\ge 1$ in photosphere and $<< 1$ in transition region layers). Therefore, magnetic pressure should be added in the computation of the total pressure and is not. This lack together with the lack of any other possible contribution not yet identified, let us to say that the electron densities and the hydrogen populations obtained from imposing $HSE$ differ from the ones we would have obtained introducing a magnetic filed contribution.
\begin{figure*}
\includegraphics*[width=0.9\textwidth]{figura4_5hse_BRAZIL.eps}
\caption{Temperature, electron density, density mass and pressure versus Column Mass for the best $HSE$ and $NOHSE$ models of the K2 primary component of the \mbox{RS\,CVn}\ system HR~7428. In the last plot, the two best models are shown for references.}
\label{best_comparison}
\end{figure*}
This approach takes into account the possibility that some pressure contributions are neglected and becomes a method to derive an estimate of the lacking pressure component. In such an approach we use the electron density as a free parameter, looking for the electron density distribution that best fit the data. We accept as best solution, also a distribution whose total pressure is not balancing the gravity, accepting the hypothesis that a new pressure component could be considered for achieving the equilibrium.
Furthermore, also different approaches in the treatment of line blanketing, produces different $N_e$ vs temperature. It is well known in fact, that the source function of a line can be strongly coupled to radiation fields even at very different wavelengths, via radiative rates determining the statistical equilibrium of the species producing the line.
If we would find as the best model reproducing our data, an $HSE$ model, this would mean that magnetic or other contributions to the total pressure are negligible in the whole upper atmosphere of the K2 HR\,7428 star.
In the case where a $NOHSE$ model should best reproduce the observations this would implies that other pressure contributions are important in the outer atmosphere of our star and we can then infer information on these additive contributions to the total pressure.
Therefore both the $HSE$ grid of 2052 models and the initial grid of 15691 model have been considered in the radiative transfer calculations for \ion{H}{}, \ion{Na}{}, \ion{Ca}{} and \ion{Mg}{}. For each grid, the models that converge to solution for all the atoms have been taken into account for the comparison with the observed spectrum.
The computed profiles of \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Ca}{ii}\,IRT, \ion{Na}{i}\,D, have been convolved with a rotational profile with $v \sin i$=17 km~s$^{-1}$ \citep{Marino_etal01} and an instrumental profile with R=$\frac {\lambda}{\Delta(\lambda)}$~=~$45,000$, normalized and then compared with observations.
The computed profiles of \ion{Mg}{ii}\,h\,\&\,k\ have been obtained by the weighted sum of the K2 star profile and the A2 star profile and weighted for the $d^2/R^2$ factor.
Wavelength shifts to account for orbital velocities, are applied to synthetic spectra.
\section{Comparison with observations: the best model}
We used a $\chi^2$ minimization procedure for the selection of the model that best describes the mean outer atmosphere of HR\,7428.
\begin{figure*}
\includegraphics*[width=0.9\textwidth]{magnetic_color.eps}
\caption{(Left panel:)The new pressure component vs Column Mass is shown (dashed line) in comparison to electron pressure (blue), gas pressure (red), turbulent pressure (green); the atmospheric model is overlapped (dot-dashed line) in order to identify the atmospheric layers where we are comparing the pressure components. (Right panel:) Magnetic field distribution vs Column Mass obtained considering the whole lacking pressure as magnetic pressure.}
\label{magneticfield}
\end{figure*}
For each line the $\chi^2$ between each observed and the synthetic line profile has been performed interpolating to the same wavelength grid the two profiles and choosing opportunely the wavelength range for the $\chi^2$ determination. Therefore, for each model, the whole set of observed lines has been compared with the corresponding synthetic lines and a global $\chi^2_{Tot}$ for each model has been calculated as the average of the $\chi^2$ obtained for the single line profiles, a 0.5 weight has been applied to the \ion{Mg}{ii}\,h\,\&\,k\ lines $\chi^2$ in order to take into account the activity-variability as seen in Section~\ref{data}. We defined a selection box formed of those models which give a $\chi^2$ less than a fixed value for all used lines and UV continuum, the best model is than selected as the one of the box with the lowest $\chi^2_{Tot}$.
We find that the best $NOHSE$ model has a $\chi^2_{Tot}$~=~1.22 while the best $HSE$ model has a $\chi^2_{Tot}$~=~2.60. This result lets conclude that the $NOHSE$ best model distribution of temperature, gas pressure, electron and population densities versus height is the best description of the mean outer atmosphere of the K2 star of the binary system HR7428.
We will describe here both the two $HSE$ and $NOHSE$ best models as possible representations of the mean outer atmosphere of the K2 star.
The best $\chi^2_{Tot}$ of the $NOHSE$ model with respect to the $HSE$ one is clearly understable from Fig.~\ref{righe} where the \mbox{H-$\alpha$}, \mbox{H-$\beta$}, \ion{Ca}{ii}\,IRT, \ion{Na}{i}\,D\ synthetic profiles computed for the best $HSE$ (dotted line) and the best $NOHSE$ (dashed line) models are shown and compared with the observed profiles.
The $NOHSE$ model gives \ion{Na}{i}\,D\ lines somewhat less deep than the observed profile and an \mbox{H-$\alpha$}\ profile a bit deeper than the observed one. Nevertheless, the $NOHSE$ model reproduce much better than the $HSE$ both the \ion{Ca}{ii}\,IRT\ line profiles as well as the \mbox{H-$\beta$}\ profile, furthermore, the also the $HSE$ model gives \ion{Na}{i}\,D\ lines less deep and \mbox{H-$\alpha$}\ deeper than the observed one.
The $NOHSE$ model gives also a better fit of the \ion{Mg}{ii}\,h\,\&\,k\ lines as can be seen from Fig.~\ref{mg} where the IUE observed flux at Earth of the \ion{Mg}{ii}\,h\,\&\,k\ is compared with the synthetic profiles obtained from the $HSE$ (dotted line) and the $NOHSE$ (dashed line) best models. We can see how the $NOHSE$ best model better reproduce both the continuum emission and the \ion{Mg}{ii}\,h\,\&\,k\ line profiles.
The discrepancy in the peak region of the h and k profiles is compatible with the activity-variability or also to the unresolved ISM absorption line that has not been taken into account in the synthetic lines computation.
Fig.~\ref{cntrbfnohse} and Fig.~\ref{cntrbfhse} show the line contribution functions ($CF$) to the emergent radiation ($CF$) as defined by \cite{Achmad_etal91}, of the synthesized lines in the case of $NOHSE$ and $HSE$ respectively. The plots mainly indicates the region of formation of the different part of each line, having on the $x$ axis the $\Delta \lambda$ from the center of the line and on the $y$ axis the Column Mass that corresponds to an the atmospheric layer. In order to make this more clear we overplotted the atmospheric model as a continuous line in each $CF$ plot.
In the case of \ion{Na}{i}\,D, \ion{Mg}{ii}\,h\,\&\,k\ and \ion{Ca}{ii}\,IRT, the $CFs$ of the multiple lines are overplotted together.
We can see that both in the $NOHSE$ and in the $HSE$ atmospheres, the line formation is very similar.
We find that the base of the transition region is quite well constrained by the \mbox{H-$\alpha$}\ and \ion{Mg}{ii}\,h\,\&\,k\ cores. Also the central part of the \ion{Na}{i}\,D, \mbox{H-$\beta$}\ and \ion{Ca}{ii}\,IRT\ cores are affected by the transition region conditions.
\ion{Mg}{ii}\,h\,\&\,k\ lines, except for the core, form entirely in the chromosphere which is also constrained by the \ion{Ca}{ii}\,IRT\ lines.
The line wings of all the used lines, except for \ion{Mg}{ii}\,h\,\&\,k\ wings, form in the atmospheric region that goes from the photosphere up to the temperature minimum region.
The last plot on the right, in Fig.~\ref{cntrbfnohse} and Fig.~\ref{cntrbfhse} shows the overplot of all the lines $CF$ and puts in evidence how the used diagnostics enough constrain the whole atmosphere, giving strength to the method and to the atmospheric model derived for the K2 star of HR\,7428.
The $HSE$ and $NOHSE$ best models are shown in Fig.~\ref{bestmodel} compared with the Solar Val~3c model. We can see that the shape of both the $HSE$ and $NOHSE$ best models is quite similar to the Solar one. It is worthwhile to notice that the two models describe strongly different geometries. The $HSE$ best model describes an atmosphere of the mean K2 primary component of the \mbox{RS\,CVn}\ system HR~7428 that extends up to $8 \times 10^5$~km, the temperature decreases and reaches the minimum temperature of about 3,000~K at about 500,000~Km above the photosphere, the chromosphere extends for 250,000~Km from the temperature minimum up to the base of the transition region that is located at about 800,000~Km. The best $NOHSE$ models is instead less extended of the $HSE$ one; the temperature from the photosphere decreases, reaching the minimum of $\approx$~3200~K at about 140,000~Km above the photosphere, the chromosphere extends for 60,000~Km from the temperature minimum up to the base of transition region that is located at about 200,000~Km.
The difference between the $HSE$ and the $NOHSE$ models can be better understood from the plot of Fig.~\ref{best_comparison} where the temperature, electron density and total density are plotted versus Column Mass density (CM) defined as $P/g$ for the best $HSE$ and $NOHSE$ models.
From Fig.~\ref{best_comparison} panel a), we can see that the $HSE$ best model has a minimum of temperature close to 3000~K when the $\log CM$~$\approx$~-1.04 and a chromospheric 'plateau' at a temperature of about 6500~K at Column Mass $\approx$~$\log CM$~=~-2 then temperature rise up to 10,000~K and at $\log CM$~=~-2.25 we find the base of transition region. In the transition region the temperature rises abruptly up to 200,000~K while the Column Mass remains approximately constant.
The $NOHSE$ best model has a minimum of temperature $\approx$~3250~K (250~K higher with respect to the $HSE$ model) at a $\log CM$ $\approx$~-0.48 a chromospheric plateau colder than the $HSE$ one at about 5500~K and $\log CM$~=~-2.1. The base of transition region is found here at $\log CM$~=~-3.14; then temperature rise abruptly up to 200,000~K while the Column Mass remains approximately constant.
It is worthwhile to notice (panel b) that in the chromospheric region the $HSE$ model presents an electron density rising continuously, while in the $NOHSE$ model we observe an abrupt rise of the electron density in a few kilometers while in the main part of the chromosphere the electron density is approximately constant. This result is the same as that found from semi-empirical models of other cool stars \citep{Harper92}. This abrupt enhancement of electron density corresponds to a density enhancement and to an enhancement of the total pressure. Both in the $HSE$ and in the $NOHSE$ best models, in the chromospheric layers the density decrease uniformly up to the transition region. Here both electron density and density decrease abruptly.
The most important difference between the two models is shown in Fig.~\ref{best_comparison} panel d), where the total pressure is plotted versus Column Mass. While, in the $HSE$ model, the total chromospheric pressure is imposed to be equal to the Column Mass multiplied for the gravity, that is $P_{Tot}$~=~CM~$\times g$, and we obtain a straight line, in the $NOHSE$ model, the total chromospheric pressure exceed the gravity pressure (dashed line is not a straight line but, in the chromospheric layers lays well above of the gravity pressure).
This exceeding pressure has to be balanced by an equal and opposite pressure otherwise we cannot have a stable star. Therefore, we assumed the difference $P_{Tot}$~-~CM~$\times ~g$ as equal, with opposite sign, to the lacking pressure in our calculations, that is: $P_{Totnew}$~=~$P_{turb}$~+~$P_{gas}$~+~$P_{new}$~=~CM~$\times ~g$ and therefore $P_{new}$~=~CM~$\times ~g$~-~$P_{Tot}$.
It is worthwhile to notice that not imposing hydrostatic equilibrium only refers to the chromospheric layers, both the used photospheric and transition regions models were calculated imposing hydrostatic equilibrium, therefore this new component pressure estimation refers only to this layer.
In Fig.~\ref{magneticfield} the new pressure component is shown in comparison to electron pressure, gas pressure, turbulent pressure. It is clear that this pressure component is not negligible being of the same order of strength of the gas pressure. In the hypothesis that the additive pressure, could be a magnetic pressure, We calculated a $\approx$~500~gauss magnetic field corresponding to a magnetic field that, from the base of the chromosphere decrease toward the outer layers.
\section{Conclusions}
We present here the semi-empirical modeling for the K2 star of the \mbox{RS\,CVn}\ binary system HR~7428. The model has been computed to match a wide set of observations from the UV continuum to a set of chromospheric lines. The fine coverage in the parameter space used in the modeling let us able to find a good agreement between the observed and computed spectral features. Furthermore a good agreement obtained in matching the UV continuum when use is made of the line-blanketing approximation method of \cite{Busa_etal01} reinforce the confidence on the method itself.
Although we have obtained an acceptable agreement between the calculations and the observations when the $HSE$ is imposed, as it is usual, we do not find an $HSE$ model that has a good match with all the used diagnostics. This could be due to many reasons, here we have explored the possibility that we are neglecting some components to the total pressure in the hydrostatic equilibrium equation. Therefore we considered also the radiative transfers without imposing $HSE$. The best model in $NOHSE$ gives a much better agreement with observations both in line profiles and in UV continuum.The stability of the best $NOHSE$ model, implies the presence of an additive (toward the center of the star) pressure, that decreases in strength from the base of the chromosphere toward the outer layers. Interpreting this additive pressure as a magnetic pressure we estimated a magnetic field intensity of about 500~gauss at the base of the chromosphere.
\section*{Acknowledgements}
We wish to thank the referee, Prof. Donald G. Luttermoser, for his careful reading of the manuscript and for his useful and kind comments and suggestions that we have greatly appreciated. IRAF is distributed by the NOAO which is operated by AURA under contract with NFS.
|
\section{Introduction}
Increased autonomy in applications ranging from transportation to energy systems necessitates the synthesis of controllers that perform safely in the presence of uncertainties. Often, such control laws need to satisfy complex specifications. To move beyond single objective and point-to-point motion planning towards complex specifications, formal methods, such as linear temporal logic (LTL) \cite{pnueli1977}, can be used. LTL combines propositional logic with temporal operators to enable the expression of logical statements in time.
Once a specification is formulated, standard model checking tools can be used to synthesize hybrid controllers \cite{fainekos2005b,tabuada2006,belta2007} based on a finite-state abstraction which bisimulates the original continuous system.
For discrete-time linear or mixed-logical dynamical systems \cite{bemporad1999}, the problem of finding trajectories that satisfy LTL specifications can be posed as a mixed-integer linear or quadratic program (MILP/MIQP) \cite{karaman2008,wolff2014} and can be applied in a receding horizon fashion \cite{gol2015}. This approach naturally allows for the consideration of continuous states and inputs. It benefits from the computational advances in mixed-integer optimization algorithms, rather than constructing a possibly large discrete abstraction.
Dealing with uncertainties is an area of active research in the context of control synthesis with formal specifications. The source of such uncertainties can be broadly classified into two categories: \emph{internal}, meaning uncertainty in the system dynamics or the system model, and \emph{external}, meaning uncertainty affecting the environment. Such external uncertainties can correspond to uncertainty in target locations or uncertain behavior of the environment. Furthermore, the objective when dealing with uncertainty is usually two-fold \begin{enumerate*}[label=(\roman*)]
\item to improve controller performance by incorporating knowledge about the uncertainty into the controller synthesis, e.g., taking into account known probability distributions of, or bounds on, the disturbance, and
\item to robustify the controller against the disturbances, avoiding violation of the specifications for any disturbance realization.
\end{enumerate*}
Stochastic models have been used to cope with probabilistic uncertainty affecting the dynamical system, whereas probabilistic specification languages have been introduced to address external uncertainty. In \cite{kamgarpour2013} stochastic hybrid systems and a subset of LTL specifications are considered. Maximizing the probability of satisfying the specification is cast as a stochastic reachability problem. This is generalized to include probabilistic uncertainties in the location of goal and obstacle sets \cite{kamgarpour2017}. In \cite{ding2011,wolff2012} general LTL specifications for a finite-state Markov decision process (MDP) are considered.
A dynamic programming approach is proposed to synthesize controllers that satisfy the specification.
Finally in \cite{lahijanian2012} MDPs are used with a probabilistic specification language.
These approaches aim to achieve satisfaction of the specification in probability and do not consider other performance criteria.
This is in contrast to \cite{aoude2013}, where a minimum-time objective is pursued, while ensuring that uncertain obstacles are avoided with a given probability. However, \cite{aoude2013} does not address general LTL specifications.
In a \emph{robust} setting, past work has addressed uncertain environments.
A powerful concept for temporal logic planning in uncertain environments is \emph{reactive planning/synthesis}. In this framework, generalized reactivity(1) specifications \cite{piterman2006} are used that capture both the task specifications and the allowed uncertain behavior of the environment. In \cite{kress-gazit2009} standard tools for LTL controller synthesis are used to generate controllers that are robust to uncertain environment behavior. In addition, \cite{wongpiromsarn2012} addresses the receding horizon case. However, these methods cannot easily capture dynamic disturbances.
Alternatively, in \cite{raman2015} signal temporal logic, a more advanced specification language that captures robustness, is used. It quantifies the degree of satisfaction of a specification. Trajectories that maximize robustness can be generated via repeated solution of MILPs.
\subsection{Contribution}
In this work, we focus on LTL specifications in uncertain or time-varying environments. Our motivation stems from the presence of uncertainties in the obstacles or goal sets. Hence, we introduce a framework in which the atomic propositions in LTL are themselves affected by uncertainty.
In contrast to methods presented in the literature, we synthesize a \emph{robust} control \emph{policy}.
We propose a novel approach based on the use of affine disturbance feedback policies to cope with uncertainties. The use of feedback improves performance compared to open-loop control policies by taking into account measurements of past disturbances in real time. While this approach is well-known in the model predictive control literature, its application to formal method control synthesis, to the best of our knowledge, has not been explored before. To deal with the computational complexity of the resulting robust MILP/MIQP, we propose an inner approximation and illustrate its performance via a case study.
\section{Problem Formulation} \label{sec:problem}
We consider discrete-time linear systems
\begin{equation} \label{eqn:sys}
x_{k+1} = Ax_{k} + B u_{k}\,,
\end{equation}
where $x_k \in \reals{n_x}$ is the system state at time $k$ and $u_k \in \reals{n_u}$ is the control input applied between time $k$ and $k+1$.
Note that results of this work extend to systems affected by additive disturbances and can further be extended to discrete-time hybrid dynamics described via mixed-integer constraints in the framework of mixed logical dynamical systems \cite{bemporad1999}.
\subsection{Uncertain Temporal Logic Specifications}
In safety critical planning problems it is often desirable to impose strict specifications for the allowed trajectories. Specifications can include statements such as \emph{reach-avoid}, reaching a goal set while avoiding obstacles, or \emph{coverage}, visiting a collection of regions.
Linear temporal logic allows the rigorous description of such specifications.
For system~\eqref{eqn:sys}, we define a finite trajectory, or \emph{run}, of length $L$ starting at $x_j$, as
\begin{equation*}
\mathbf{x}_j^L := \begin{bmatrix}x_j^\transp & x_{j+1}^\transp & \ldots & x_{j+L}^\transp\end{bmatrix}^\transp\,,
\end{equation*}
a sequence of states $x_k$ such that for each $k=j,\ldots,j+L-1$ there exists an input $u_k$ such that $x_{k+1} = Ax_{k} + Bu_{k}$.
We consider LTL specifications that are affected by an uncertain \emph{disturbance} vector $w \in \reals{n_w}$. Given a specification $\varphi$, length $L$ and index $j$, we want to find a control input sequence $\mathbf{u}_j^L := \begin{bmatrix}u_j^\transp & \cdots & u_{j+L-1}^\transp\end{bmatrix}^\transp$ such that the run $\mathbf{x}_j^L$ satisfies the specification $\varphi$ for all disturbance sequence realizations $\mathbf{w}_j^L := \begin{bmatrix}w_j^\transp & \ldots & w_{j+L}^\transp\end{bmatrix}^\transp$ contained in a bounded polyhedron $\setbf{W}_j^L \subseteq \reals{(L+1)n_w}$, i.e.,
\begin{equation*}
(\mathbf{x}_j^L,\mathbf{w}_j^L) \satisfies \varphi\quad \forall \mathbf{w}_j^L \in \setbf{W}_j^L\,.
\end{equation*}
For simplicity we consider LTL formulae in positive normal form \cite{baier2008}. To avoid issues with unbounded effects of the disturbances, we furthermore use bounded LTL formulae, without loops \cite[Definition~2.1]{biere2006}, a subset of the usual LTL semantics.
A formula in LTL is a combination of \emph{atomic propositions} $p$ taken from a finite set ${\rm AP} := \{p_1, \ldots , p_m\}$, propositional logic operators $\lnot$~(\emph{not}), $\land$~(\emph{and}) and $\lor$~(\emph{or}), and temporal operators $\lnext$~(\emph{next}), $\until$~(\emph{until}) and $\release$~(\emph{release}). More formally, we define LTL formulae via the grammar
\begin{equation*}
p \sep{} \lnot p \sep{} \phi \land \psi \sep{} \phi \lor \psi \sep{} \lnext \phi \sep{} \phi \until \psi \sep{} \phi \release \psi\,,
\end{equation*}
where $\phi, \psi$ are LTL formulae. Atomic propositions take values in $\{\true,\false\}$. In the context of this work, the disturbance $w$ enters the description of the atomic propositions $p_i \in {\rm AP}$, i.e., each $p_i$ is associated with a polyhedral set
\begin{equation*}
\set{P}_i := \{(x,w) \in \reals{n_x+n_w} \sep{} P_i^x x \leq P_i^w w + \rho_i \}\,,
\end{equation*}
defined over the state-disturbance space.
Given a sequence of disturbance realizations $\mathbf{w}_j^L$, a run $\mathbf{x}_j^L$ satisfies an atomic proposition $p_i$ if the \emph{augmented state} $\mathbf{z}_j^L:=(z_j,\ldots,z_{j+L})$, with $z_{{\hat{\jmath}}} := (x_{\hat{\jmath}},w_{\hat{\jmath}})$, satisfies $z_j \in \set{P}_i$. The satisfaction of the formula $p_i$ is denoted by $\mathbf{z}_j^L \satisfies p_i$. The propositional operators are defined as
\begin{subequations} \label{eqn:LTLSemantics}
\begin{align}
&\mathbf{z}_j^L \satisfies \lnot p_i &&\text{ iff } z_j \not\in \set{P}_i\,,\label{eqn:LTLSemanticsNot}\\
&\mathbf{z}_j^L \satisfies \phi \land \psi &&\text{ iff } \mathbf{z}_j^L \satisfies \phi \text{ and } \mathbf{z}_j^L \satisfies \psi\,,\label{eqn:LTLSemanticsAnd}\\
&\mathbf{z}_j^L \satisfies \phi \lor \psi &&\text{ iff } \mathbf{z}_j^L \satisfies \phi \text{ or } \mathbf{z}_j^L \satisfies \psi\,,\label{eqn:LTLSemanticsOr}\\
\intertext{and the temporal operators are defined as}
&\mathbf{z}_j^L \satisfies \lnext \phi &&\text{ iff } \mathbf{z}_{j+1}^{L-1} \satisfies \phi\,, \label{eqn:LTLSemanticsNext}\\
&\mathbf{z}_j^L \satisfies \phi \until \psi &&\text{ iff } \exists \hat{\jmath} \in \{0, \ldots, L-1\}\suchthat \mathbf{z}_{j+\hat{\jmath}}^{L-\hat{\jmath}} \satisfies \psi \text{ and } \forall i \in \{0,\ldots,\hat{\jmath}-1\}\,: \mathbf{z}_{j+i}^{L-i} \satisfies \phi\,,\label{eqn:LTLSemanticsUntil}\\
&\mathbf{z}_j^L \satisfies \phi \release \psi &&\text{ iff } \forall \hat{\jmath} \in \{0,\ldots,L-1\}:\mathbf{z}_{j+\hat{\jmath}}^{L-\hat{\jmath}} \satisfies \psi \text{ or } \exists i\in\{0,\ldots,\hat{\jmath}-1\} \suchthat \mathbf{z}_{j+i}^{L-i} \satisfies \phi\,.\label{eqn:LTLSemanticsRelease}
\end{align}
\end{subequations}
We introduce the additional temporal operators $\eventually \phi := \true \until \phi$ (\emph{eventually}) and $\always \phi := \false \release \phi$ (\emph{always}).
\subsection{Robust Policy Synthesis}
Given a fixed \emph{planning horizon} $N$ and \emph{initial state} $x_0$, we define the state trajectory $\mathbf{x} := \mathbf{x}_0^N$, disturbance sequence $\mathbf{w} := \mathbf{w}_0^N$ and corresponding input $\mathbf{u} := \mathbf{u}_0^N$, with
\begin{equation*}
\setbf{W} := \setbf{W}_0^N := \{\mathbf{w} \in \reals{(N+1)n_w} \sep{} \mathbf{W}\mathbf{w} \leq \mathbf{v} \}\,,
\end{equation*}
a closed and \emph{bounded} polyhedron, with $\mathbf{W} \in \reals{n_v \times (N+1)n_w}$ and $\mathbf{v} \in \reals{n_v}$.
Employing causal disturbance feedback policies allows us to synthesize a control law that can react to past disturbances in real-time based on measured data. Such robust feedback policies can be generated, even though only bounds on the disturbances are known during control synthesis.
\begin{problem}
Given an LTL specification $\varphi$, find a sequence of causal disturbance feedback policies
\begin{equation*}
u_0(w_0),\ldots,u_{N-1}(w_0,\ldots,w_{N-1})\,,
\end{equation*}
such that for all realizations of the uncertainty $\mathbf{w} \in \setbf{W}$:
we minimize an objective function and satisfy
\begin{enumerate*}[label=(\roman*)]
\item input constraints $u_k \in \set{U} \subseteq \reals{n_u}$,
\item state constraints $x_{k+1} \in \set{X} \subseteq \reals{n_x}$, and
\item the specification $(\mathbf{x},\mathbf{w}) \satisfies \varphi$
\end{enumerate*}.
\end{problem}
\section{Solution Approach} \label{sec:solution}
Searching for general feedback policies is intractable. Hence, we focus on linear feedback.
We define a causal, affine disturbance feedback policy with parameters $\mathbf{H} \in \reals{Nn_u\times Nn_{w}}$ and $\mathbf{h} \in \reals{Nn_u}$:
\begin{equation} \label{eqn:affpolicy}
\mathbf{u} =
\underbrace{\begin{bmatrix}H_{1,1} & & & 0\\
\vdots & \ddots & & \vdots\\
H_{N,1} & \cdots & H_{N,N} & 0 \end{bmatrix}}_{\begin{bsmallmatrix}\mathbf{H} & 0\end{bsmallmatrix}}\mathbf{w}
+ \underbrace{\begin{bmatrix}h_1 \\ \vdots \\ h_N\end{bmatrix}}_{\mathbf{h}}\,.
\end{equation}
Such affine policies are used in robust control \cite{bemporad1998, zhang2017} to improve performance compared to open-loop policies. By setting $\mathbf{H} = 0$ we have an open-loop policy, which is similar to the approach in \cite{raman2015}. There, worst-case disturbance sequences are computed via MILPs and then the trajectory is robustified against those sampled sequences. In our approach, using linear programming duality, only one mixed-integer program (MIP) needs to be solved to obtain guarantees for all disturbances.
Substituting policy \eqref{eqn:affpolicy} into the discrete-time system equations \eqref{eqn:sys}, we express the trajectory $\mathbf{x}$ as a function of the initial state $x_0$, the disturbance sequence $\mathbf{w}$, and the parameters $\mathbf{H}$ and $\mathbf{h}$ of our policy:
\begin{align*}
\mathbf{x} &= \underbrace{\begin{bmatrix}A^0 \\ A^1 \\ \vdots \\ A^N\end{bmatrix}}_{\mathbf{A}}x_0 + \underbrace{\begin{bmatrix}0 \\ B \\ AB & B \\ \vdots & & \ddots \\ A^{N-1}B & \cdots & AB & B \end{bmatrix}}_{\mathbf{B}}\mathbf{u}\\
&= \mathbf{A}x_0 + \begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix}\mathbf{w} + \mathbf{B}\mathbf{h}\,.
\end{align*}
We consider the following robust policy synthesis problem:
\begin{problem}[Robust policy synthesis] \label{prob:aff}
\setlength{\abovedisplayskip}{0.3em} \setlength{\abovedisplayshortskip}{0.3em}
\begin{align}
\min_{\mathbf{H},\mathbf{h}} \: & J(\mathbf{H},\mathbf{h})\nonumber\\
\suchthat & \begin{rcases}\begin{bmatrix}\mathbf{H} & 0\end{bmatrix}\mathbf{w} + \mathbf{h} \in \setbf{U}\,,\\
\mathbf{A}\theta + \begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix}\mathbf{w} + \mathbf{B}\mathbf{h} \in \setbf{X}\,,\\
(\mathbf{A}\theta + \begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix}\mathbf{w} + \mathbf{B}\mathbf{h},\mathbf{w}) \satisfies \varphi
\end{rcases} \forall \mathbf{w} \in \setbf{W}\,,\label{eqn:specconst}
\end{align}
where, $J : \reals{Nn_u\times N n_{w}} \times \reals{Nn_u} \rightarrow \reals{}$ is the convex quadratic objective and $\theta \in \reals{n_x}$ is the parametric initial state. Furthermore, $\setbf{U} := \set{U} \times \ldots \times \set{U} \subseteq \reals{Nn_u}$ and $\setbf{X} := \set{X} \times \ldots \times \set{X} \subseteq \reals{Nn_x}$ are the ``stacked'' input and state constraints, with $\set{U}$ and $\set{X}$ compact, convex polyhedra.
\end{problem}
\begin{remark}
The case where the disturbance has known linear dynamics $v_{k+1} = A^w v_k +B^w w_k$ with an uncertain input $w_k$, naturally fits into the presented framework. Disturbance state feedback $u_k = K_{k+1} v_k + \kappa_{k+1}$ can be equivalently posed as disturbance feedback by appropriate choice of $\setbf{W}$ and restrictions on the structure of $\mathbf{H}$.
\end{remark}
\subsection{Conversion to Robust Mixed-Integer Program} \label{sec:mipconversion}
It is known that specification constraints of the form $\mathbf{z} \satisfies \varphi$ can be transformed into linear mixed-integer inequalities \cite{wolff2014}. This is achieved by introducing auxiliary variables, some of which are restricted to be binary. We transform the constraint
\begin{equation*}
(\mathbf{A}\theta + \begin{bmatrix}\mathbf{B} \mathbf{H} & 0\end{bmatrix}\mathbf{w} + \mathbf{B}\mathbf{h},\mathbf{w}) \satisfies \varphi\,,
\end{equation*}
of \refprob{prob:aff} into a set of mixed-integer inequalities:
\begin{align} \label{eqn:miineq}
f^\varphi(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta) :=& \left(F^x\begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix} + F^w\right)\mathbf{w} + F^x\mathbf{B}\mathbf{h} + F^x\mathbf{A}\theta + F^\delta \delta + f \leq 0\,,
\end{align}
where the auxiliary vector $\delta \in \Delta := \reals{n_c} \times \binaries{n_b}$ consists of $n_c$ continuous and $n_b$ binary variables.
The matrices $F^x$, $F^w$, $F^\delta$ and vector $f$ are of appropriate dimensions.
Notice, that \eqref{eqn:miineq} is linear in $\theta$, $\mathbf{h}$ and $\delta$, and bilinear in $\mathbf{H}$ and $\mathbf{w}$.
We augment the specification constraint \eqref{eqn:miineq} with the input and state constraints of \refprob{prob:aff}. This yields the following set of mixed-integer inequalities:
\begin{align} \label{eqn:ltl+ineq}
g^\varphi(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta) :=& \left(G^H\begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix} + G^w\right)\mathbf{w} + G^H\mathbf{B}\mathbf{h} + G^\theta\mathbf{A}\theta + G^\delta \delta + g \leq 0\,,
\end{align}
where $g^\varphi(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta)$ consists of $m^\varphi$ constraints and the matrices $G^H$, $G^w$, $G^\theta$ and $G^\delta$, as well as the vector $g$ have appropriate dimensions.
The number $n_b$ of binary variables does not change.
Substituting the set of mixed-integer inequalities \eqref{eqn:ltl+ineq} into \refprob{prob:aff} yields an optimization problem with linear mixed-integer constraints that need to be satisfied robustly, i.e., for all $\mathbf{w} \in \setbf{W}$. To make this problem tractable, we will reduce the robust constraint to a set of mixed-integer constraints.
\subsection{Reduction to Mixed-Integer Program}
The set of robustly admissible feedback gains $\mathbf{H}$ and $\mathbf{h}$, parametrized by the initial state $\theta$, is given as follows:
\begin{align} \label{eqn:miineq+}
\set{C}^\varphi :=& \big\{ (\mathbf{H},\mathbf{h},\theta) \sep{\big} \eqref{eqn:specconst} \text{ holds for } \mathbf{H},\mathbf{h} \text{ given } \theta \, \big\}\nonumber\\
=& \big\{ (\mathbf{H},\mathbf{h},\theta) \sep{} \forall\mathbf{w} \in \setbf{W}\:\exists \delta \in \Delta \suchthat g^\varphi(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta) \leq 0 \, \big\}\nonumber\\
=& \big\{ (\mathbf{H},\mathbf{h},\theta) \sep{} \hspace*{-0.4em} \max_{\mathbf{w}\in\setbf{W}} \min_{\delta \in \Delta} \max_{i} g^\varphi_i(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta) \leq 0 \, \big\},
\end{align}
where $g^\varphi_i$ denotes the $i$-th component of $g^\varphi$. The solution of $\min_{\delta \in \Delta} \max_{i} g^\varphi_i(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta)$ is a piecewise affine function in $\mathbf{H}\mathbf{w}$, $\mathbf{h}$ and $\theta$ \cite{ramirez2001}. The challenge in solving \eqref{eqn:miineq+} is due to the max-min structure and the bilinear dependence on $\mathbf{H}$ and $\mathbf{w}$. In particular, the maximization over $\mathbf{w}$ leads to non-linear constraints \cite{khalilpour2014}. Using linear programming duality, $\set{C}^\varphi$ can be represented by constraints that are linear in $\mathbf{h}$ and $\theta$, but contain bilinear terms involving~$\mathbf{H}$. This results in a representation of \refprob{prob:aff} as a bilinear program, which can be solved via spatial branch-and-bound \cite{androulakis1995}. For the case of $\mathbf{H} = 0$, the set $\set{C}^\varphi$ can be represented by linear mixed-integer inequalities, leading to a mixed-integer formulation which can be solved using general-purpose MIQP solvers.
To avoid the difficulty of solving a bilinear optimization problem for the general case $\mathbf{H}\neq 0$, we propose a simple inner approximation. This approximation produces a mixed-integer program and usually helps to preserve sparsity in the optimization problem. We consider the set
\begin{align} \label{eqn:innerapprox}
\overline{\set{C}}^\varphi := \Big\{ (\mathbf{H},\mathbf{h},\theta) \sep{\Big} &\exists \delta \in \Delta\,,\:\greekbf{\lambda}_i \in \posreals{n_v} \text{ for } i=1,\ldots,m^\varphi \suchthat \text{for all } i = 1,\ldots,m^\varphi:\nonumber\\[-0.2em]
& \quad \mathbf{v}^\transp \greekbf{\lambda}_i + G^H_{i\bcdot}\mathbf{B}\mathbf{h} + G_{i\bcdot}^\theta\mathbf{A}\theta + G_{i\bcdot}^\delta \delta + g_i \leq 0\,,\nonumber\\[-0.2em]
& \quad \mathbf{W}^\transp \greekbf{\lambda}_i = \left(G^H_{i\bcdot}\begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix} + G_{i\bcdot}^w\right)^\transp \Big\}\,,
\end{align}
where $G_{i\bcdot}$ denotes the $i$-th row of matrix $G$.
\begin{lemma} \label{lem:inner}
The constraint set $\overline{\set{C}}^\varphi$ is an inner approximation of the robust specification constraint set $\set{C}^\varphi$, i.e., any feasible policy that satisfies \eqref{eqn:innerapprox} also satisfies \eqref{eqn:miineq+}.\\
\begin{proof}
First, we exchange the maximization over $\mathbf{w}$ with the minimization over $\delta$ in \eqref{eqn:miineq+}. This leads to the inner approximation
\begin{align*}
\overline{\set{C}}^\varphi := \big\{ (\mathbf{H},\mathbf{h},\theta) \sep{\big} &\exists \delta \in \Delta \suchthat \hspace*{-0.8em} \max_{\substack{\mathbf{w}\in\setbf{W}\\i=1,\ldots,m^\varphi}} \hspace*{-0.5em} g^\varphi_i(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta) \leq 0 \big\} \subseteq \set{C}^\varphi\nonumber\,.
\end{align*}
For a given $i \in 1,\ldots,m^\varphi$, we use a standard robust optimization technique~\cite[p.~472]{bertsimas2011}, applying linear programming duality to replace $\max_{\mathbf{w}\in\setbf{W}} g^\varphi_i(\mathbf{H},\mathbf{h},\theta,\mathbf{w},\delta)$ with its dual
\begin{align*}
\min_{\greekbf{\lambda}_i \in \posreals{n_v}} & \mathbf{v}^\transp \greekbf{\lambda}_i + G^H_{i\bcdot}\mathbf{B}\mathbf{h} + G_{i\bcdot}^\theta\mathbf{A}\theta + G_{i\bcdot}^\delta \delta + g_i\\[-0.8em]
\suchthat & \mathbf{W}^\transp \greekbf{\lambda}_i = \left(G^H_{i\bcdot}\begin{bmatrix}\mathbf{B}\mathbf{H} & 0\end{bmatrix} + G_{i\bcdot}^w\right)^\transp\,.
\end{align*}
Dropping the minimization over $\greekbf{\lambda}_i$ gives an upper bound. Collecting these upper bounds for $i = 1,\ldots,m^\varphi$ yields \eqref{eqn:innerapprox}.
\end{proof}
\end{lemma}
Using the inner approximation $\overbar{\set{C}}^\varphi$ in \refprob{prob:aff} results in \refprob{prob:affmip}, a mixed-integer program that can be solved using general-purpose MIQP solvers.
\begin{problem}[MIQP] \label{prob:affmip}
\setlength{\abovedisplayskip}{0.3em} \setlength{\abovedisplayshortskip}{0.3em}
\begin{align*}
\min_{\mathbf{H},\mathbf{h},\delta,\greekbf{\lambda}} \: & J(\mathbf{H},\mathbf{h})\\
\suchthat & \mathbf{\overbar{V}}^\transp \greekbf{\lambda} + G^H\mathbf{B}\mathbf{h} + G^\theta\mathbf{A}\theta + G^\delta \delta + g \leq 0\,,\\
& \overbar{\mathbf{W}}^\transp \greekbf{\lambda} = (I_{m^\varphi} \otimes \begin{bmatrix}\mathbf{H} & 0\end{bmatrix}^\transp) \mathbf{G}^\transp + {\mathbf{g}^\transp}\,,\\
& \delta \in \Delta\,, \greekbf{\lambda} \in \posreals{m^\varphi n_v}\,,
\end{align*}
where $\otimes$ denotes the Kronecker product and $\greekbf{\lambda} := \begin{bmatrix}\greekbf{\lambda}_1^\transp & \ldots & \greekbf{\lambda}_{m^\varphi}^\transp\end{bmatrix}^\transp$, $\mathbf{\overbar{V}} := \diag(\mathbf{v}, \ldots, \mathbf{v})$,
$\overbar{\mathbf{W}} := \diag(\mathbf{W},\ldots,\mathbf{W})$,
$\mathbf{G} := \begin{bmatrix}G^H_{1\bcdot} \mathbf{B} & \cdots & G^H_{m^\varphi\bcdot} \mathbf{B} \end{bmatrix}$ and
$\mathbf{g} := \begin{bmatrix}G^w_{1\bcdot} & \cdots & G^w_{m^\varphi\bcdot} \end{bmatrix}$.
\end{problem}
\begin{theorem} \label{thm:mip}
Any solution $\mathbf{H}^\star$, $\mathbf{h}^\star$ of \refprob{prob:affmip} is a feasible, possibly suboptimal, solution of \refprob{prob:aff}.\\
\begin{proof}
The proof follows directly from \reflem{lem:inner}.
\end{proof}
\end{theorem}
From \refthm{thm:mip} it follows that the policies obtained by solving \refprob{prob:affmip} robustly satisfy the specification for all disturbance realizations.
\begin{remark}
\refprob{prob:affmip} has $O(m^\varphi n_v + N^2 n_u n_w + n_c)$ continuous and $n_b$ binary decision variables, as well as $O(m^\varphi n_w)$ constraints. Mixed-integer solvers that can be used for instances of \refprob{prob:affmip} have a worst-case exponential complexity in the number of binary variables, $n_b$, whereas the complexity is polynomial, usually cubic, in the number of continuous variables and constraints. As remarked in \refsec{sec:mipconversion}, $n_b$ depends linearly on $N$. Solving \refprob{prob:affmip} therefore requires time that is in the worst-case exponential in the planning horizon $N$.
\end{remark}
\subsection{On the role of feedback and the approximation scheme}
The purpose of this section is to illustrate two points. First, in \refexa{exa:ex1}, we illustrate a simple case in which the inner approximation is not tight. In \refexa{exa:ex2}, we illustrate a case in which no feasible open-loop policy exists but a feasible disturbance feedback policy can be found.
\begin{figure*}
\centering
\begin{minipage}[b]{0.73\columnwidth}
\begin{subfigure}[b]{0.31\linewidth}
\centering
\begin{tikzpicture}[scale=1,cap=round]
\tikzstyle{dotted}=[dash pattern=on 0.5pt off 2pt]
\begin{axis}[
every outer x axis line/.append style={white!20!black},
every x tick label/.append style={font=\color{white!20!black}},
xmin=-1.5,
xmax=1.5,
xtick={-1,0,1},
ymin=-1,
ymax=1,
ytick={-2,-1,0,1},
axis lines = middle,
enlargelimits = true,
xlabel={$x{-}w$},
x label style={at={(axis description cs:1.2,0.25)}},
ylabel={$e^\varphi(x,w)$},
y label style={at={(axis description cs:0.5,1.2)}},
axis on top,
width = 1.8in, height = 1.4in]
\addplot[name path=mfl,domain=-1.2:-0.5,color=blue,line width=1.5pt] {-x-1}; \addplot[name path=mfl,domain=-1.5:-1.2,color=blue,line width=1.5pt, dotted] {-x-1};
\addplot[name path=mfl,domain=-0.5:0,color=blue,line width=1.5pt] {x};
\addplot[name path=mfl,domain=0:0.5,color=blue,line width=1.5pt] {-x};
\addplot[name path=mfl,domain=0.5:1.2,color=blue,line width=1.5pt] {x-1}; \addplot[name path=mfl,domain=1.2:1.5,color=blue,line width=1.5pt, dotted] {x-1};
\end{axis}
\end{tikzpicture}
\caption{Piecewise affine constraint function $e^\varphi(H,h,\theta,w)$ for $x = \theta + Hw + h$ and $w$.\\\\}\label{fig:ex1:exact:pwa}
\end{subfigure}
\hspace*{0.02\linewidth}
\begin{subfigure}[b]{0.31\linewidth}
\centering
\begin{tikzpicture}[scale=1,cap=round]
\tikzstyle{dotted}=[dash pattern=on 0.5pt off 2pt]
\begin{axis}[
every outer x axis line/.append style={white!20!black},
every x tick label/.append style={font=\color{white!20!black}},
xmin=-1.5,
xmax=1.5,
xtick={-1,0,1},
ymin=-1,
ymax=1,
ytick={-2,-1,0,1},
axis lines = middle,
enlargelimits = true,
xlabel={$h{+}\theta$},
x label style={at={(axis description cs:1.2,0.22)}},
ylabel={$\overbar{e}^\varphi(1,h,\theta,\delta)$},
y label style={at={(axis description cs:0.5,1.2)}},
axis on top,
width = 1.8in, height = 1.4in]
\addplot[name path=mfl,domain=-1.2:-0.5,color=blue,line width=1.5pt] {-x-1}; \addplot[name path=mfl,domain=-1.5:-1.2,color=blue,line width=1.5pt, dotted] {-x-1};
\addplot[name path=mfl,domain=-0.5:0.2,color=blue,line width=1.5pt] {x}; \addplot[name path=mfl,domain=0.2:0.5,color=blue,line width=1.5pt,dotted] {x};
\addplot[name path=mfl,domain=-0.2:0.5,color=lightred,line width=1.5pt, dashed] {-x}; \addplot[name path=mfl,domain=-0.5:-0.2,color=lightred,line width=1.5pt,dotted] {-x};
\addplot[name path=mfl,domain=0.5:1.2,color=lightred,line width=1.5pt, dashed] {x-1}; \addplot[name path=mfl,domain=1.2:1.5,color=lightred,line width=1.5pt, dotted] {x-1};
\end{axis}
\end{tikzpicture}
\caption{Mixed-integer constraint function $\overbar{e}^\varphi(H,h,\theta,\delta)$ for $H=1$. Cases $\delta = 1$ (\textbf{\color{blue}solid}), $\delta = 0$ (\textbf{\color{lightred}dashed}) illustrated for values of $h$ and $\theta$}\label{fig:ex1:inner:pwa}
\end{subfigure}
\hspace*{0.02\linewidth}
\begin{subfigure}[b]{0.31\linewidth}
\centering
\begin{tikzpicture}[scale=1,cap=round]
\begin{axis}[
every outer x axis line/.append style={white!20!black},
every x tick label/.append style={font=\color{white!20!black}},
xmin=-1,
xmax=1,
xtick={-1,0,1},
xticklabels={${-}1{-}\theta$,${-}\theta$,$1{-}\theta$},
ymin=0,
ymax=2.2,
ytick={1, 2},
axis lines = middle,
enlargelimits = true,
xlabel={$h$},
x label style={at={(axis description cs:1.1,0.1)}},
ylabel={$H$},
y label style={at={(axis description cs:0.5,1.1)}},
axis on top,
after end axis/.code={\path (axis cs:0,0) node [anchor=north,yshift=-0.2em] {{\color{white!20!black}${-}\theta$}};},
width = 1.8in]
\addplot[name path=mfl,domain=-1:0,color=blue,line width=1.5pt] {-x};
\addplot[name path=mfu,domain=-1:0,color=blue,line width=1.5pt] {x+2};
\addplot[name path=pfu,domain=0:1,color=blue,line width=1.5pt] {2-x};
\addplot[name path=pfl,domain=0:1,color=blue,line width=1.5pt] {x};
\addplot[fill=verylightblue] fill between [of=mfl and mfu,soft clip={domain=-1:0}];
\addplot[fill=verylightblue] fill between [of=pfl and pfu,soft clip={domain=0:1}];
\addplot[name path=mfl,domain=-1:-0.5,color=lightred,line width=1.5pt,dashed] {-x};
\addplot[name path=mfu,domain=-1:-0.5,color=lightred,line width=1.5pt,dashed] {x+2};
\addplot[name path=pfu,domain=-0.5:0,color=lightred,line width=1.5pt,dashed] {1-x};
\addplot[name path=pfl,domain=-0.5:0,color=lightred,line width=1.5pt,dashed] {1+x};
\addplot[fill=verylightred] fill between [of=mfl and mfu,soft clip={domain=-1:-0.5}];
\addplot[fill=verylightred] fill between [of=pfl and pfu,soft clip={domain=-0.5:0}];
\addplot[name path=mfl,domain=0:0.5,color=lightred,line width=1.5pt,dashed] {1-x};
\addplot[name path=mfu,domain=0:0.5,color=lightred,line width=1.5pt,dashed] {x+1};
\addplot[name path=pfu,domain=0.5:1,color=lightred,line width=1.5pt,dashed] {2-x};
\addplot[name path=pfl,domain=0.5:1,color=lightred,line width=1.5pt,dashed] {x};
\addplot[fill=verylightred] fill between [of=mfl and mfu,soft clip={domain=0:0.5}];
\addplot[fill=verylightred] fill between [of=pfl and pfu,soft clip={domain=0.5:1}];
\end{axis}
\end{tikzpicture}
\caption{The set $\set{C}^\varphi$ of all feasible affine disturbance feedback policies ({\color{blue}\textbf{solid}}) and its inner approximation $\overbar{\set{C}}^\varphi$ ({\color{lightred}\textbf{dashed}}). Both are parametrized by $\theta$.}\label{fig:ex1:feas}
\end{subfigure}
\caption{Illustrating the difference between the exact robust specification constraint formulation $\set{C}^\varphi$ and the mixed-integer inner approximation $\overbar{\set{C}}^\varphi$ for \refexa{exa:ex1}.} \label{fig:ex1}
\end{minipage}
\hspace*{0.02\columnwidth}
\begin{minipage}[b]{0.23\columnwidth}
\centering
\begin{tikzpicture}[scale=1,cap=round]
\begin{axis}[
every outer x axis line/.append style={white!20!black},
every x tick label/.append style={font=\color{white!20!black}},
xmin=-4,
xmax=4,
xtick={-4,-2,0,2,4},
xticklabels={$-4$,$ $,$0$,$ $,$4$},
ymin=-4,
ymax=4,
axis y line=none,
axis x line=middle,
enlargelimits = true,
axis on top,
width = 1.8in]
\filldraw[blue,fill=verylightblue,line width=1.5pt] (axis cs:-2,-2) rectangle (axis cs:2,2);
\addplot[color=red,mark=x,line width = 1.5pt,forget plot]
table[row sep=crcr]{0 0\\};
\node[right] at (axis cs:0.1,0.5) {{\color{red}$x_0$}};
\draw[->,someblue,line width=1.5pt] (axis cs:2,0) -- (axis cs:3,0) node[near end, above,someblue] {$w_0$};
\draw[->,someblue,line width=1.5pt] (axis cs:-2,0) -- (axis cs:-3,0);
\node at (axis cs:0,-3) {$k=0$};
\end{axis}
\end{tikzpicture}\\
\vspace*{-0.5cm}
\begin{tikzpicture}[scale=1,cap=round]
\begin{axis}[
every outer x axis line/.append style={white!20!black},
every x tick label/.append style={font=\color{white!20!black}},
xmin=-4,
xmax=4,
xtick={-4,-2,0,2,4},
xticklabels={$-4$,$ $,$0$,$ $,$4$},
ymin=-4,
ymax=4,
axis y line=none,
axis x line=middle,
enlargelimits = true,
x label style={at={(axis description cs:1.1,0.1)}},
y label style={at={(axis description cs:0.5,1.1)}},
axis on top,
width = 1.8in]
\filldraw[blue,fill=verylightblue,line width=1.5pt] (axis cs:-1,-2) rectangle (axis cs:3,2);
\addplot[color=red,mark=x,line width = 1.5pt,forget plot]
table[row sep=crcr]{1 0\\};
\draw[->,black,line width=1.5pt] (axis cs:0,0) -- (axis cs:1,0) node[near start, above,black] {$u_0$};
\node[right] at (axis cs:1.1,0.5) {{\color{red}$x_1$}};
\draw[->,someblue,line width=1.5pt] (axis cs:3,0) -- (axis cs:4,0) node[near end, above,someblue] {$w_1$};
\draw[->,someblue,line width=1.5pt] (axis cs:-1,0) -- (axis cs:-2,0);
\node at (axis cs:0,-3) {$k=1$};
\end{axis}
\end{tikzpicture}
\caption{Illustration of robustly feasible feedback policy for \refexa{exa:ex2}.} \label{fig:ex2}
\end{minipage}
\end{figure*}
\begin{example} \label{exa:ex1}
We consider a planning horizon of $N=1$ and a simple dynamical system $x = \theta + u$ with initial state $\theta$. A scalar disturbance $w \in \set{W}:= [-1,1]$ affects the specification $\varphi := p \lor q$, where the polyhedra associated with $p$ and $q$ are
\begin{align*}
\set{P} &:= \{x \in \reals{} \sep{} x \in [-1,0]+w \} \text{ and } \set{Q} := \{x \in \reals{} \sep{} x \in [0,1]+w \}\,.
\end{align*}
Given the affine disturbance feedback policy $u : = Hw + h$, we can construct the piecewise affine constraint function
\begin{equation*}
e^\varphi(H,h,\theta,w) := \min_{\delta \in \Delta} \max_{i} g^\varphi_i(H,h,\theta,w,\delta)\,,
\end{equation*}
illustrated in \reffig{fig:ex1:exact:pwa}. We see that $e^\varphi(H,h,\theta,w) \leq 0$, i.e., the constraint is feasible, for all $x$ and $w$ such that $-1 \leq x-w \leq 1$. This yields a description of the set $\set{C}^\varphi$ with $H$ and $h$, parametrized by $\theta$ and illustrated ({\color{blue}\textbf{solid}}) in \reffig{fig:ex1:feas}. In this example $\set{C}^\varphi$ is a convex polyhedron. However, in general it may be neither polyhedral nor convex.
We also consider the inner approximation $\overbar{\set{C}}^\varphi$ with the corresponding mixed-integer constraint function
\begin{equation*}
\overbar{e}^\varphi(H,h,\theta,\delta) := \max_{w \in \set{W}} \max_{i} g^\varphi_i(H,h,\theta,w,\delta)\,,
\end{equation*}
illustrated in \reffig{fig:ex1:inner:pwa} for $H=1$. For $H \neq 1$ the illustration in \reffig{fig:ex1:inner:pwa} needs to be shifted upwards by $|H-1|$, reducing the available choices for $h$. The set $\overbar{\set{C}}^\varphi$ of $H$ and $h$, parametrized by $\theta$, is depicted ({\color{red}\textbf{dashed}}) in \reffig{fig:ex1:feas}.
We see that $\overbar{\set{C}}^\varphi \subset \set{C}^\varphi$. Furthermore, when $H=0$, the inner approximation is empty, while the exact solution has a unique feasible policy: $h = -\theta$.
\end{example}
\begin{example} \label{exa:ex2}
We consider simple integrator dynamics $x_{k+1} = x_k + u_k$ with initial state $x_0 = 0$. We define a safe set $\set{P}_{\rm safe} := \{x \in \reals{2} \sep{} x_1 \in [-2,2]+w\,, x_2 \in [-2,2] \}$ with associated atomic proposition $p_{\rm safe}$. The disturbance $w_k \in \reals{}$ is in $[-1,1]$ for all time steps $k$. The initial state is illustrated in \reffig{fig:ex2}. The objective is to find an input sequence $u_0,u_1,\ldots$ that satisfies the specification $\varphi := \always p_{\rm safe}$ for all disturbance realizations and all time steps $k$. Clearly this is not possible, because the safe set can move both either left or right, i.e., no open-loop policy that robustly satisfies this specification exists. However, it is easy to see, that the affine disturbance feedback policy $u_k = w_k$ is robustly feasible.
\end{example}
\section{Case Study} \label{sec:study}
We consider a motion planning task on a two-lane highway illustrated in \reffig{fig:cs:track}. Two cars with known initial positions are cruising on the upper lane at uncertain velocities in the range $[\unitfrac[93.4]{km}{h},\unitfrac[106.6]{km}{h}]$. The controlled car is driving on the lower lane and has an initial velocity of $\unitfrac[110]{km}{h}$. It is modeled as a 2-dimensional double-integrator affected by a constant drag term:
\begin{equation} \label{eqn:doubleint}
\begin{bmatrix}\dot{x}_1 \\ \dot{x}_2\end{bmatrix} = \begin{bmatrix}x_3 \\ x_4\end{bmatrix}\,, \text{ and } \begin{bmatrix}\dot{x}_3 \\ \dot{x}_4\end{bmatrix} = \begin{bmatrix}u_1 \\ u_2\end{bmatrix} - c_d\,,
\end{equation}
where the state $x \in \reals{4}$ contains the position $(x_1,x_2)$ of the car and the longitudinal and lateral velocities $(x_3,x_4)$. The accelerations $(u_1,u_2)$ are the control inputs and are limited to the set $\set{U}$: $u_1 \in [\unitfrac[-4]{m}{s^2},\unitfrac[2]{m}{s^2}]$ and $u_2 \in [\unitfrac[-3]{m}{s^2},\unitfrac[3]{m}{s^2}]$. The drag term $c_d$ is assumed to be $\unitfrac[0.3]{m}{s^2}$. The forward velocity is limited according to driving regulations to $x_3 \in [\unitfrac[90]{km}{h},\unitfrac[120]{km}{h}]$ and the lateral velocity $x_4$ is limited to $\pm \unitfrac[20]{km}{h}$.
Additionally, we impose the lane constraint $x_2 \in [-\unit[3.5]{m},\unit[3.5]{m}]$.
We consider a reference frame moving in longitudinal direction with a constant velocity of $\unitfrac[100]{km}{h}$. All quantities are with respect to this frame. In \reffig{fig:cs:track} the origin of this frame is always $(0,0)$.
A discrete-time version of \eqref{eqn:doubleint} with sampling time $T_s = 0.2$~seconds is used.
A truck is approaching the controlled car from behind with an uncertain initial distance from the controlled car in the range $[\unit[7.5]{m},\unit[18]{m}]$ and an uncertain velocity between $\unitfrac[110]{km}{h}$ and $\unitfrac[120]{km}{h}$. The goal is to escape the approaching truck by performing a lane change. This needs to be accomplished without crashing for any realization of the stated position and velocity uncertainties of the other vehicles.
To model this scenario as an LTL specification, we introduce the three atomic propositions $p_{\rm car 1}$, $p_{\rm car 2}$ and $p_{\rm truck}$ for the obstacles and a proposition $p_{\rm goal}$ for the goal set that we want to reach before being hit by the truck. Hence, we want to satisfy the specification $\varphi := \always\big( \lnot p_{\rm car 1} \land \lnot p_{\rm car 2} \land \lnot p_{\rm truck}\big) \land \eventually \always p_{\rm goal}$ for all realizations $\mathbf{w} \in \setbf{W}$ of the uncertainty. The sets corresponding to the atomic propositions are
\begin{align*}
\set{P}_{\rm car 1} &:= \big\{ (x,w) \sep{\big} 0 \leq x_1 \leq 6.75 + w_1\,, 0 \leq x_2 \leq 3.5 \big\}\,,\\
\set{P}_{\rm car 2} &:= \big\{ (x,w) \sep{\big} 31.25 - w_2 \leq x_1 \leq 38\,, 0 \leq x_2 \leq 3.5 \big\}\,,\\
\set{P}_{\rm truck} &:= \big\{ (x,w) \sep{\big} 0 \leq x_1 \leq w_3\,, -3.5 \leq x_2 \leq 0 \big\}\,,\\
\set{P}_{\rm goal} &:= \big\{ (x,w) \sep{\big} 0 \leq x_2 \leq 3.5 \big\}\,,
\end{align*}
and the set of disturbances $\setbf{W}$ is defined as
\begin{align*}
\setbf{W} := \big\{ \mathbf{w} \in \reals{3(N+1)} \sep{\big} & w_{k,i} = \sum_{j=0}^k\nolimits \omega_{j,i} \text{ for } i=1,\ldots,3 \text{, with }\\
&\: \omega_{j,1} \in [-d_c,d_c]\,,\: \omega_{j,2} \in [-d_c,d_c] \text{ for } j=0,\ldots,N\\
&\: \omega_{0,3} \in [\unit[-13.5]{m},\unit[0]{m}]\,,\\
&\: \omega_{j,3} \in [T_s \cdot \unitfrac[10]{km}{h},T_s \cdot \unitfrac[20]{km}{h}] \text{ for } j=1,\ldots,N\: \big\}\,,
\end{align*}
with $d_c := T_s \cdot \unitfrac[6.6]{km}{h}$.
Because a point model is used, the position constraints of all vehicles have to be modified to take into account the car's shape which is $\unit[4.5]{m}$ long and $\unit[2]{m}$ wide. For simplicity, these margins are omitted in the illustration.
The planning horizon is $N=20$ and we use an objective function that promotes \emph{minimum time} solutions, additionally penalizing the control effort.
\begin{equation*}
J(\mathbf{H},\mathbf{h},\delta) := \underbrace{\sum_{k=0}^N k \delta_{{\rm goal},k}}_{\text{minimum time}} + \gamma\big( \underbrace{\|\mathbf{h}\|_2^2 + \|\mathbf{H}\|_F^2 \max_{\mathbf{w}\in\setbf{W}} \|\mathbf{w}\|_2^2}_{\text{control effort}} \big)\,,
\end{equation*}
with $\gamma = 0.001$. The binary variable $\delta_{{\rm goal},k}$ equals one if the goal is reached at time $k$. We denote by $\|\cdot\|_2$ the 2-norm and by $\|\cdot\|_F$ the Frobenius norm.
We generated both an open-loop and a disturbance feedback policy using YALMIP \cite{lofberg2004} and Gurobi \cite{gurobi} to solve the resulting MIQP on an Intel i7 CPU at 2.8GHz. The open-loop policy problem has $4198$ continuous and $273$ binary variables, and $7401$ constraints. An optimal solution was found in $\unit[0.4]{s}$. For the disturbance feedback policy, the proposed inner approximation was used leading to a problem with $35338$ continuous and $273$ binary variables, and $52401$ constraints. An optimal solution was found in $\unit[17.3]{s}$.
\begin{figure}
\centering
\begin{minipage}[b]{0.49\linewidth}
\centering
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth]{run_6_final.tikz}
\end{subfigure}\\
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth]{run_13_final.tikz}
\end{subfigure}\\
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth]{run_20_final.tikz}
\end{subfigure}
\caption{Trajectory for time steps $k=\{6,13,20\}$ with the feedback policy applied to the worst-case disturbance realization. The \emph{effective goal} is the feasible part of the goal, additionally taking into account the shape of the controlled car.} \label{fig:cs:track}
\end{minipage}
\hspace*{0.005\linewidth}
\begin{minipage}[b]{0.245\linewidth}
\centering
\includegraphics[width=\linewidth]{run_comparison.tikz}
\caption{Trajectories of controlled car corresponding to random disturbances ({\color{lightred}\textbf{solid}}), best-case ({\color{blue}\textbf{dotted}}) and worst-case ({\color{blue}\textbf{dashed}}) disturbance.} \label{fig:cs:traj}
\end{minipage}
\hspace*{0.005\linewidth}
\begin{minipage}[b]{0.23\linewidth}
\centering
\includegraphics[width=\linewidth]{controleffort.tikz}
\caption{Control effort $\|\mathbf{u}\|_2$ for different disturbance realizations ({\color{blue}\textbf{solid}}), best ({\color{red}\textbf{circle}}), worst ({\color{red}\textbf{square}}) and open-loop case ({\color{blue}\textbf{dashed}}).} \label{fig:cs:obj}
\end{minipage}
\end{figure}
In \reffig{fig:cs:track} the trajectory resulting from applying the synthesized feedback policy to the worst-case disturbance realization is illustrated for the time steps $k=6,13$ and $20$. The trajectory is feasible. Furthermore, the goal set is reached after $13$ time steps and the car remains there for the remainder of the planning horizon.
\reffig{fig:cs:traj} illustrates trajectories for different disturbances taken uniformly randomly from $\setbf{W}$ ({\color{lightred}\textbf{solid}}). Additionally, the trajectories for the best- and worst-case disturbance are illustrated ({\color{blue}\textbf{dotted}} and {\color{blue}\textbf{dashed}}, respectively). The open-loop trajectory does not differ substantially from the feedback trajectory resulting from the worst-case disturbance and is therefore omitted.
Finally in \reffig{fig:cs:obj} we give the overall control effort $\|\mathbf{u}\|_2$ (sorted) that was needed for $1000$ different disturbance realizations ({\color{blue}\textbf{solid line}}), as well as for the best-case ({\color{red}\textbf{circle}}), worst-case ({\color{red}\textbf{square}}) and the open-loop case ({\color{blue}\textbf{dashed line}}). This illustrates that the feedback policy allows the reduction of the control effort compared to the open-loop policy. Furthermore, for all disturbance samples, the feedback policy required $13$ time steps to reach the goal set whereas the open-loop policy required $14$ time steps.
Note that, in practice the controlled car would estimate the disturbances based on measured positions of the vehicles and apply disturbance feedback using these estimates.
\section{Conclusions}
We have addressed the problem of control synthesis for linear systems given bounded LTL specifications affected by uncertain disturbances. We have formulated a robust policy synthesis problem with affine disturbance feedback to synthesize policies that satisfy the specification for all considered disturbances.
We cast this problem as a robust mixed-integer program. Then we introduced a simple inner approximation of the constraint set resulting in an MIQP that can be solved using general-purpose mixed-integer solvers. The proposed method therefore enables the generation of control policies that satisfy the specification robustly and incorporates performance criteria such as minimum time or minimum control effort. The framework was applied to a numerical case study of guiding a car in a lane changing maneuver on a highway.
\bibliographystyle{ieeetr}
|
\section{Introduction}
An evanescent field of the light confined in a semiconductor optical microcavity extends out of the microcavity boundaries. This effect allows for a coupling between optical modes of adjacent microcavities provided that their sizes, thus resonant frequencies, are close enough.\cite{Stanley:APL1994} The coupling brings a rich variety of phenomena, like an ultralow threshold optical parametric oscillation \cite{Diederichs:Nature2006} or a spontaneous mirror-symmetry breaking.\cite{Hamel:NaturePhot2015} It is also predicted to enable a generation of single photons under cw excitation thanks to a polariton blockade effect.\cite{Liew:PRL2010} Possible applications of coupled microcavities include areas of photonics, quantum information science and laser technology.\cite{Borishkina:Chapter2010}
The research devoted to coupled semiconductor microcavities has been so far limited mostly to the systems based on III-V compounds. At the origin a dominant direction of the investigations was the light-matter coupling in III-V planar structures coupled through a partially transparent Bragg reflector\cite{Stanley:APL1994, Armitage:PRB1998, Panzarini:PRB1999}. Next, the studies turned to {\it horizontally} 3-D coupled photonic molecules providing a stronger light confinement than the planar structures. They were realized in geometries such as: side-coupled micropillars etched out of a planar microcavity\cite{Bayer:PRL1998, Karl:OptExpress2007, deVasconcellos:APL2011}, neighbouring photonic crystal defect cavities\cite{Sato:NaturePhot2012, Haddadi:OptExpress2014} or microdisks side-coupled through an air gap.\cite{Ishii:APL2005}
In turn, studies of photonic systems involving II-VI compounds involved mostly single microcavities either in the planar\cite{Kelkar:PRB1995, LeSiDang:PRL1998, Saba:Nature2001, Martin:PRL2002} or the micropillar geometry.\cite{Robin:APL2005, Lohmeyer:APL2006, Jakubczyk:ACSNano2014} The only work reporting on coupled photonic molecules in II-VI based structures, authored by Sebald {\it et al.} (Ref.~[\cite{Sebald:OptExpr2011}]), was realized in the geometry of side-coupled micropillars etched out of a single ZnSe microcavity. However, a technology for the production of coupled microcavities based on other II-VI materials, such as ZnTe, CdTe or ZnO has not been established so far.
The large exciton binding energy (in tens of meV range) in ZnTe and ZnO makes them a particularly good material for implementation in photonic devices operating at higher temperatures with respect to their III-V counterparts and should facilitate advanced experiments, like for example coupling of distant quantum emitters through a delocalized cavity mode.\cite{Benyoucef:PRB2008}. Several interesting phenomena resulting from the strong light-matter interactions in these materials have been already shown.\cite{Jakubczyk:JAP2013, Zamfirescu:PRB2002, Sturm:NJofPhys2011}
Here we present double, coupled, ZnTe based microcavities in planar (1-D) and single micropillar (3-D) geometry. We start with designing the structures using transfer matrix method (TMM) simulations and calculations in the frame of the tight-binding approximation\cite{Bayindir:PRL2000, Bayindir:JOA2001} applied to the modes of a 1-D photonic structure. In particular, we determine the coupling strength of the modes of neighbouring cavities as a function of the cavities separation, \textsl{i. e.}, the number of pairs forming the Distributed Bragg Reflector (DBR) inserted between them. Next, we epitaxially grow the structures, changing the cavities separation from sample to sample. The gradients of the cavities width are intentionally perpendicular to each other. Since the energy of the cavity resonant mode depends on the cavity width, this allows us to tune the relative detuning between the modes by selecting a place on the sample. In such a way, we are able to tune the cavities interaction strength, as predicted by our calculations and verified in photoluminescence and reflectivity measurements. In particular, we confirm the increase of the cavity modes coupling strength with decreasing the number of DBRs pairs separating the cavities predicted by our simulations. Etching micropillars out of the planar sample enables us to study the impact of an increasing photon confinement on the modes energy and the coupling strength. In contrast to the work of Sebald {\it et al.} (Ref.~[\cite{Sebald:OptExpr2011}]) in our case the microcavities are coupled \emph{vertically}, \emph{i.e.}, in the direction perpendicular to the microcavities plane. The cylindrical symmetry of a single micropillar and the resulting decreased sidewall losses should allow for a higher photon collection efficiency with respect to previous double micro-pillar geometries.\cite{Bayer:PRL1998, Karl:OptExpress2007, deVasconcellos:APL2011, Sebald:OptExpr2011}
As such, the present work provides a proof of concept for the \emph{vertically} coupled planar cavities and 3-D photonic molecules in single micropillar geometry based on II-VI semiconductor compounds, which are propitious for the realization of hybrid optoelectronic circuits involving photonic elements. Moreover, a large Huang-Rhys factor proper for the ZnTe\cite{Zhang:PRB2012} makes the studied system highly promising for implementation in an innovative laser cooling\cite{Pringsheim:ZPhysik1929} scheme involving the pumping through the optical mode of a double microcavity structure. The maximum efficiency of the scheme would be expected for the pump laser tuned to the lower energy mode in a region of the sample where the modes energy difference is one or two longitudinal optical (LO) phonon energy. In such a case an efficient anti-Stokes emission accompanied by an annihilation of the LO phonon would occur through the higher cavity mode, while the Stokes emission would be suppressed by the stopband.
\section{Sample design and epitaxial growth}
\subsection{Sample design}
The structures comprising two ZnTe $\lambda$-cavities embedded between DBR mirrors lattice matched to ZnTe are designed using TMM calculations. In the calculations we assume that the \emph{bottom} and \emph{top} mirrors are formed by 16 and 15 DBRs pairs, respectively. The role of high refractive index material in the DBRs is played by ZnTe layers, while a short period MgSe/ZnTe/MgTe/ZnTe superlattice serves as a low refractive index material.~\cite{Pacuski:APL2009} The wavelength dependent refractive indices of ZnTe, MgSe and MgTe are taken from Ref.~[\cite{Pacuski:APL2009}]. We expect that the use of only binary compounds should result in a more stable growth and a higher quality of the structure than it was in previous approaches to II-VI photonic structures, which involved ternary compounds\cite{LeSiDang:PRL1998, Peiris:JAP1999, Saba:Nature2001, Martin:PRL2002, Lohmeyer:APL2006, Sebald:OptExpr2011} or combined epitaxial growth with oxide deposition.\cite{Kelkar:PRB1995,Robin:APL2005}
We conduct two series of TMM simulations: (i) in the first of them, only the thickness of one of the cavities is varied, while all the other parameters, in particular the width of the second cavity, are fixed. This way we obtain the cavities coupling strength as a function of their relative detuning. We perform the calculations for a number $N_{DBR}$ of the DBR pairs separating the microcavities varied in the range from 3 to 12. (ii) In the second one, only $N_{DBR}$ is varied (in the range from 3 to 17), while all the other parameters are fixed. In particular, in this case we assume that both cavities have the same thickness, thus they are maximally coupled. In this way we obtain the cavities coupling strength as a function of their spatial separation.
The sample design assumes that the coupled optical modes are detuned from the absorption edge of ZnTe (2.26 eV at room temperature). However, in order to emulate a realistic spectral linewidth, which is usually increased due to non-zero absorption of the light within the sample, we introduce a net imaginary contribution to a dielectric function of the ZnTe layers forming the cavities and the DBRs. For the sake of simplicity we assume that its value ($0.01$) is independent of the light wavelength.
As a result of the TMM calculations we obtain reflectivity spectra of the structure. We determine the eigenmodes energy by fitting a sum of two Lorentzian curves to the two minima in the stopband region of the simulated spectra. The obtained dependence of the eigenmodes energy on the relative detuning between the uncoupled modes (simulation (i)) is shown in Fig.~\ref{fig:Simulation}a) for an exemplary structure, where the fixed uncoupled cavity mode is centred at 650~nm (1910~meV)
When the detuning attains zero, a clear modes anticrossing is evidenced. The energy separation of the eigenmodes at the zero detuning increases when the $N_{DBR}$ decreases, which reflects the increasing modes interaction strength with a decreasing inter-cavity distance.
As shown by Bayindir and co-workers, a single dimensionless constant $\kappa$ derived from the eigenmodes energy difference at resonance is enough to characterize the coupling strength of 1-D microcavity modes.\cite{Bayindir:PRL2000, Bayindir:JOA2001} In the frame of a tight binding approximation applied to the optical modes localized in neighbouring microcavities $\kappa = \beta - \alpha$, where $\alpha$ and $\beta$ are determined from the following equation:\cite{Bayindir:PRL2000, Bayindir:JOA2001}
\begin{equation}
\omega_{1,2}^2 = \Omega^2 \left( 1 \pm \beta \right) / \left(1 \pm \alpha \right),
\end{equation}
with $\omega_{1}$ and $\omega_{2}$ the eigenmodes energy of the coupled cavities at zero detuning, while $\Omega$ is the energy of an uncoupled individual cavity mode. A plot of |$\kappa$| as a function of $N_{DBR}$ (simulation (ii)) is presented in Fig.~\ref{fig:Simulation}b).\cite{comment_kappa} It shows clearly that the cavities interaction strength decreases with the cavities separation, with an asymptote to zero for an infinite $N_{DBR}$. We note that the order of magnitude of the simulated $\kappa$ is in agreement with the values reported previously for 1-D coupled photonic structures.\cite{Bayindir:PRL2000,Bayindir:JOA2001} Based on the results of the simulations we epitaxially grow two structures (see the next subsection) adjusting their parameters so that the mode interaction at resonance is characterized by two significantly different values of |$\kappa$| equal to 0.0261 or 0.0054.
\begin{figure}
\includegraphics[width=0.6\linewidth]{Fig1_simulation}
\caption{a) Energy of the eigenmodes of two coupled ZnTe planar microcavities calculated as a function of the cavities ratio (bottom X axis) and the detuning between the modes of uncoupled microcavities (top X axis), for a number $N_{DBR}$ of the DBR pairs separating the cavities in the range from 3 to 12.
b) Constant |$\kappa$| characterizing the strength of the microcavities mode coupling calculated as a function of $N_{DBR}$.}
\label{fig:Simulation}
\end{figure}
\subsection{Sample epitaxial growth}
The two structures are grown by molecular beam epitaxy on a 2-inch diameter GaAs substrate with a 1~$\mu$m thick ZnTe layer acting as a buffer. They are designed as described in the previous subsection (in consecutive order): 16 DBR pairs, the \emph{bottom} microcavity, a separating DBR, the \emph{top} microcavity, 15 DBRs pairs. The separation between the two ZnTe $\lambda$-microcavities is either 6 or 12 DBR pairs (|$\kappa$| of 0.0261 or 0.0054, respectively). The \emph{top} microcavity contains an active layer of CdTe quantum dots formed out of 3 monolayers of CdTe. The QDs are introduced to provide an emission source feeding the mode emission in photoluminescence studies.\cite{Suffczynski:PRL2009}
We deliberately do not spin the substrate during the epitaxial growth, despite the spinning being usually used in order to avoid any gradient of thickness of the deposited layers resulting from a different position of different effusion cells with respect to the substrate. In this way we achieve that the thickness of the layers changes across the lateral position on the sample. For the growth of the top microcavity we rotate the substrate by 90 degrees and fix its position. Afterwards, we rotate the substrate back to the original orientation and we continue the growth. As a result of the assumed procedure, we obtain a mutually perpendicular gradient of the cavities thicknesses (see a schematic in Fig.~\ref{fig:Structures}). It enables us to continuously change the detuning between the modes of the cavities, hence the strength of the coupling between the microcavities, by adjusting the position at the sample surface.
Note that within such a growth procedure the stopband spectral position also varies with the position at the sample surface, but for a given point it remains the same for the all three DBRs.
The continuous loss of material volume in effusion cells during $\sim$~12 hours of the sample growth results in a decrease of the growth rate. We achieve a correct layers thickness by \emph{live} monitoring of the sample thickness during the growth by \emph{in situ} reflectivity and by adjusting the deposition time and/or the temperature of the sources, which controls the elements flow rate.
The result of scanning transmission electron microscopy characterization of the structure is shown in Figure~\ref{fig:Structures}. One can see a very good structural quality of the sample, with an absence of any dislocations.
\begin{figure}
\includegraphics[width=1\linewidth]{sample_structure}
\caption{A scheme showing the sample design comprising two vertically coupled ZnTe microcavities separated by either 6 or 12 DBR pairs with mutually perpendicular gradient of cavities thicknesses (lower-left corner). Scanning transmission electron microscopy images of the cross-section of the epitaxially grown sample (12 DBR pairs) with consecutive magnifications, as indicated by the respective scale bars.}
\label{fig:Structures}
\end{figure}
\section{Experiment}
The samples are studied by a reflectivity spatial mapping at room temperature performed in a custom, home made setup assuring an automated 7.5 cm per 7.5 cm 2-D movement with a resolution of 0.6 $\mu$m. The signal from the entire 2 inch wafer is collected with a step size of 1~mm. A halogen lamp light is focused to the spot diameter of around 0.1~mm at the sample surface.
Next, selected regions of the wafer are cleaved out into around 7 mm per 17 mm pieces. The cleaved sample is put onto the cold finger of a helium flow cryostat at a temperature adjusted in the 7~K - 300~K range. Micro-photoluminescence measurement is performed at the temperature of 50~K, which assures that optical transitions of individual quantum dots broaden and efficiently feed the optical modes of the cavities.\cite{Suffczynski:PRL2009} The signal from a 1 $\mu$m diameter spot is excited and collected through a microscope objective of numerical aperture NA = 0.75. The CW laser emitting at E$_{exc}$ = 2.33 eV ($\lambda_{exc}$ = 532 nm) serves as an excitation source.
Angular resolution in microphotoluminescence and microreflectivity measurements is obtained by imaging the Fourier plane of the microscope objective on the slit of the spectrometer. To achieve this, an additional lens of focal length $f$ is inserted at a distance $f$ from the objective's Fourier plane. The lens imaging the signal on the spectrometer entrance is set at the position determined from the lens equation. The high numerical aperture of our objective enables a detection of photons emitted within a wide range of angles, from -50$^{\circ}$ to +50$^{\circ}$, without the need of any tilting of the sample.
Micropillars with diameters varying from 0.7 $\mu$m to 3 $\mu$m are etched out of the planar samples using a focused beam of Ga ions with an accelerating voltage up to 30 kV in a FEI Tecnai instrument. The procedure involves the etching of a large micropillar in the initial step and a subsequent decrease of its size to the desired diameter in subsequent steps. The ion current kept as low as 1.5~pA in the final polishing step assures a minimum roughness of the side surface of the micropillars.
\section{Results and discussion}
\subsection{Coupled ZnTe planar optical microcavities}
An exemplary reflectivity spectrum of the 12 DBR sample is shown in Fig.~\ref{fig:refl_mapping}a). As predicted by the TMM calculations (see Suppl. Movie), it contains a stopband with two optical modes in its centre. The stopband is relatively wide spectrally (around 200 meV) and its reflectivity is above 99\%. This confirms that the layers involving binary compounds provide a reasonable contrast of refractive indices between the DBR layers and an overall good quality of the DBRs.
\begin{figure}
\includegraphics[width=1\linewidth]{refl_mapping}
\caption{a) An exemplary reflectivity spectrum of the 12 DBR sample at T = 300 K. b) Colour map showing the energy difference between the eigenmodes and c) the ratio of eigenmodes intensity determined for the whole area of the 12 DBR sample. The dashed line in panel b) highlights the region of the sample shown in enlargement in Fig.~4.
}
\label{fig:refl_mapping}
\end{figure}
In analogy to the case of electronic states localized in potential wells, the coupling of identical cavities leads to the emergence of spatially delocalized symmetric and antisymmetric coupled optical modes. As a result of the coupling they are split in energy. the presence of two extrema corresponding to the two modes is expected, in principle, for the optical spectrum of any photonic structure that contains two cavities. Thus, in order to prove the modes interaction one should plot the modes energy as a function of the modes detuning. The anticrossing behaviour of the modes is an unequivocal sign of the coupling in this case.
In order to prove the cavities coupling we perform a reflectivity mapping to plot the modes behaviour as a function of the detuning. We determine the modes energy and intensity by fitting with a sum of Lorentzian curves to the two minima in the stopband region of the spectra. The energy of the stopband centre is also determined (see Suppl. Fig.~\ref{fig:SIstopband_center}a)).
The energy difference between the eigenmodes and the ratio of the eigenmodes intensity for the 12 DBRs sample determined for the whole area of 2-inch sample are presented in Figs.~\ref{fig:refl_mapping}b) and ~\ref{fig:refl_mapping}c), respectively. As it can be seen, the ratio of the eigenmodes intensity is equal to unity for the region of the sample, where the energy difference between the modes is minimum. Such behaviour strongly suggests a mixing of the individual cavity modes resulting from their coupling. In regions of the sample, where the modes are strongly detuned, the energy and intensity of both modes vary independently of each other.
The reflectivity spectra for consecutive points selected every 1~mm along the thickness gradient of the bottom microcavity in the 12 DBRs sample are shown in Figure~\ref{fig:mode_energy}a) in the spectral range limited to the modes vicinity. With the change of the position on the sample, the mode initially at a lower energy (peaked at around 1850~meV at the bottom spectrum in Fig.~\ref{fig:mode_energy}a)) significantly shifts spectrally towards higher energy. The second mode initially keeps its position, however, when it is approached by the lower energy mode, a mutual pushing of the modes is evidenced. This clear anticrossing behaviour of the modes confirms that we truly accomplish the intended coupling between the two microcavities.
As a side remark, the mode initially at around 1850 meV is much weaker than the mode at around 1900 meV in the bottom curve of Fig.~\ref{fig:mode_energy}a). Its lowered intensity indicates that it is indeed linked to the bottom cavity, which signal is filtered by the stopband of the top cavity. Naturally, such attribution is possible only when the modes are far from resonance, since when the detuning approaches zero, their wavefunctions become mixed and spatially delocalized between the two microcavities.
\begin{figure}
\includegraphics[width=1\linewidth]{modes_energy}
\caption{a) Reflectivity spectra for consecutive points on the 12 DBR sample along the gradient width of the bottom microcavity. The reflectivity spectrum when the modes enter into resonance is indicated in red. b) Surface plot of the relative energies of the modes determined by the reflectivity mapping \emph{vs} the position on the sample. The dots indicate the positions, at which the reflectivity spectra shown in panel a) are taken. The surfaces interpolate the discrete values of the modes determined with 1~mm step in each dimension.
}
\label{fig:mode_energy}
\end{figure}
In order to determine precisely the energy difference between the eigenmodes at resonance, we plot in Fig.~\ref{fig:mode_energy}b) the modes energy relatively to the stopband centre as a function of the position on the sample in the resonance region. The subtraction of the stopband energy allows us to eliminate a possible energy shift of the modes induced by the stopband shift resulting from the thickness gradient of the layers forming the DBRs. The eigenmodes splitting of 17 meV at resonance is found for the 12 DBRs sample. The splitting increases to 45 meV when the separation is decreased to 6 DBRs pairs (see Suppl. Figure~\ref{fig:17vs45}), in agreement with the expectation. The corresponding |$\kappa$| factors determined from the experiment amount to $0.0088 \pm 0.0025$ or $0.0220\pm 0.0025$, in the respective case of the 12 DBRs or 6 DBRs sample. As it is seen in Figure~\ref{fig:Simulation}, they remain in a good agreement with the predictions based on our calculations.
We perform angle resolved photoluminescence and reflectivity measurements, e.g., to determine the quality factor of the studied microcavities. The emission angle registered in the experiment encodes the in-plane component {\bf k$_{\parallel}$} of the momentum of the photon confined in the microcavity. The dependence between the photon energy and the {\bf k$_{\parallel}$} can be approximated by:\cite{AAmo:PhD2008} $E\left(k_{\parallel}\right) \sim \frac{c \, \hbar}{n_{\text{cav}}}\left(k_{\perp}+\frac{1}{2 k_{\perp}}k_{\parallel}^2\right)$, where $n_{\text{cav}}$ is the refractive index of the ZnTe microcavity. The $k_{\perp}$ being the photon momentum perpendicular to the microcavity plane is quantized due to the cavity resonant condition and takes a constant value for a given mode. Since the light rays leaving the microcavity at a given angle meet in a common point on the Fourier plane, a cross-section through the centre of the Fourier plane provides us with the dispersion relation E({\bf k$_{\parallel}$}).
The dispersion relation found for the 12 DBR sample is shown in Fig.~\ref{fig:kspace}. Let us note first that the E({\bf k$_{\parallel}$}) obtained from the reflectivity agrees well with the one coming from the photoluminescence measurement. It is also seen that the two extrema representing two eigenmodes of the coupled structure shift towards higher energy with an increasing photon in-plane wavevector. The E({\bf k$_{\parallel}$}) dependences of both eigenmodes take approximately a parabolic shape, similarly to what was previously observed in the case of microcavities confining a single mode.\cite{Kelkar:PRB1995, Armitage:PRB1998, Panzarini:PRB1999} The images shown in Figure~\ref{fig:kspace} are acquired for the point on the sample, where the modes are close to resonance. We have checked, however, that the observed E({\bf k$_{\parallel}$}) dependences are not sensitive to the modes relative detuning.
The linewidth $\Delta E$ of the cavity resonance, that is the linewidth of the dispersion relation at {\bf k$_{\parallel}$ = 0} provides a quality factor of the microcavity Q = $E / \Delta E$. The linewidths of the modes at {\bf k$_{\parallel}$} = 0 are determined from the reflectivity measurement to 1.2~meV and 1.5~meV for the lower and higher energy mode, respectively. This yields a quality factor Q of the cavities in the range of 1600-1250. We link the slightly larger width found for the higher energy mode with the fact that this mode represents an "antibonding" photon state. In that case, the broadening may result from a higher degree of the delocalized photon penetration in the barrier. A similar broadening was previously observed for excitons confined at excited levels of a semiconductor quantum well.\cite{Dingle:PRL1974}
\begin{figure}
\includegraphics[width=0.65\linewidth]{kspace}
\caption{Photoluminescence and reflectivity spectra of the 12 DBR sample resolved in photon energy and in plane momentum {\bf k$_{\parallel}$} for an exemplary point of the sample. T = 50 K.
}
\label{fig:kspace}
\end{figure}
The results presented in this Section show that the assumed sample design allows one to efficiently control the interaction between the modes of epitaxially grown, coupled ZnTe optical microcavities exhibiting a good crystal quality. The coarse adjustment is realized by the variation of the number of DBRs pairs separating the cavities. The fine control is obtained by selecting a position at the surface of the sample grown with a perpendicular gradient of the cavities thickness.
\subsection{Coupled 3-D photonic molecules in single micropillar geometry}
A microstructuration of a planar cavity sample provides a mean for an additional confinement of the light. In structures embedding quantum emitters it allows therefore enhancing of the light-matter interaction constants resulting from a decreased mode volume.\cite{Yoshie:Nature2004, Reithmaier:Nature2004} In the case of the studied structures one should expect that the microstructuration will result in an increased interaction strength between the coupled modes resulting from a stronger radial localization of the photon.
Micropillars are etched out of the coupled planar microcavities (see Fig.~\ref{fig:micropillars}a) and Suppl. Movie) in a region where the microcavities are maximally coupled. Their diameter ranges from 3 $\mu$m down to 0.7 $\mu$m. Since a discontinuity of the refractive index at the micropillar sidewalls provides a photon confinement in the radial direction, the obtained coupled photonic molecule has a three-dimensional character. In previous approaches involving DBRs based micropillar microcavities\cite{Bayer:PRL1998, Karl:OptExpress2007, deVasconcellos:APL2011, Sebald:OptExpr2011} the photonic molecules were realized in the side-coupled micropillar geometry, where the cavities were \emph{horizontally} coupled. In contrast, the presented 3-D photonic molecule is realized in a single pillar geometry and with a \emph{vertical} coupling of the cavities.
The study of the molecules involves reflectivity and photoluminescence measured at T = 50 K. A decrease of the light spot down to 1~$\mu$m diameter enables us to address and collect the signal from individual micropillars. The microstructuration results in a quantization of the cavity mode into a set of discrete submodes (see Fig.~\ref{fig:micropillars}b and inset to Fig.~\ref{fig:micropillars}c). The highest energy submodes originating from the lower energy eigenmode spectrally overlap with the lowest energy submodes originating from the higher energy eigenmode. The photoluminescence measurement performed with angular resolution enables us to properly determine the origin and energy of a given submode.
The energies of the first two submodes emerging from each eigenmode of the planar structure are determined by fitting with Lorentzian curve (see the inset to Fig.~\ref{fig:micropillars}c)). It is evidenced that with decreasing the pillar diameter, both the submodes energy and energy spacing between the consecutive submodes increase (see Fig.~\ref{fig:micropillars}c). This demonstrates an increased coupling strength of the modes resulting from the increasing photon confinement. The observed shape of the dependences of the submodes energy on the pillar diameter is such as observed previously for the modes of individual micropillars etched out of a single planar microcavity.\cite{Reithmaier:PRL1997, Lohmeyer:APL2006}
\begin{figure}
\includegraphics[width=1\linewidth]{micropillars}
\caption{a) Scanning electron microscopy image of an exemplary micropillar etched out of the 6 DBR sample. The 1~$\mu$m scal bar is indicated. b) Emission of a 3~$\mu$m micropillar from the 12 DBR sample as a function of the photon energy and the in-plane momentum, which reveals discrete submodes of the coupled 3-D photonic molecule. c) Energy dependence of the first two submodes on the micropillar diameter for the 12 DBR sample. Inset: exemplary reflectivity and photoluminescence spectra measured without angular resolution, along with fitted Lorentzian curves.}
\label{fig:micropillars}
\end{figure}
This Section indicates that the sample microstructuration is an effective tool for tuning of the coupled modes energy, as well as for a control of the modes coupling strength. Thanks to a a high symmetry and resulting low sidewall losses, the presented innovative design of coupled 3-D photonic molecules in single micropillar geometry is promising for practical applications.
\section{Conclusions}
In the present work, we provide a proof of concept for the double coupled optical microcavities and the coupled photonic molecules in a single micropillar geometry based on ZnTe, being a representative member of II-VI compounds family. The design, growth and results of spectroscopy measurements confirming the interaction and control of the microcavities optical modes are presented for two samples differing by the microcavities separation.
The ability for tuning of the energy between the eigenmodes to one or two LO phonon energy in ZnTe ($\sim$ 26 meV)\cite{Nahory:PR1967} in 12 DBR od 6 DBR structure, respectively, confirms the potential of the presented structures for the application in the laser cooling scheme mentioned in the introductory section. A high exciton oscillator strength makes the studied system suitable also for achieving the interaction between distant quantum emitters such as quantum dots or quantum wells mediated by the delocalized optical mode. Thanks to the photon contribution to the wavefunction of the exciton and resulting emergence of polaritons,\cite{Kelkar:PRB1995,LeSiDang:PRL1998,Armitage:PRB1998,Panzarini:PRB1999} the exciton tunneling between the quantum wells embedded in two microcavites should be made possible. The tunneling in such a case would take place for a distance of the order of a micrometer, which exceeds for more than an order of magnitude the distances reported on previously for quantum well excitons.\cite{Lawrence:PRL1994} Moreover, the performance of the studied system could be further enhanced through the use of semimagnetic quantum wells or quantum dots embedding individual magnetic ions.\cite{Pacuski:CGDesign2014}
\section{Acknowledgements}
An access to the FEI Tecnai Osiris TEM instrument located at the Facility for Electron Microscopy \& Sample Preparation of the University of Rzesz\'{o}w, as well as a financial support of Polish National Science Centre (NCN) under the projects no. UMO-2013/10/E/ST3/00215 and no. UMO-2015/16/T/ST3/00506 and of Polish Ministry of Science and Higher Education under the project no. IP2014040473 are acknowledged.
\begin{suppinfo}
\beginsupplement
\begin{figure}
\includegraphics[width=1\linewidth]{SI_stopband_center}
\caption{a) A real color photograph, b) a color map showing a central wavelength of the stopband determined from the reflectivity mapping performed on the 12 DBR sample.
}
\label{fig:SIstopband_center}
\end{figure}
The Suppl. Fig.~\ref{fig:SIstopband_center}a) shows the real photograph of the 12 DBR sample as well as a color map representing the spectral position of the stopband centre determined from the reflectivity mapping. It is seen that the both images are mutually consistent.
The difference of the eigenmodes energy as a function of the position on the sample, thus of the detuning, for 12 DBR and 6 DBR sample is shown in Fig.~\ref{fig:17vs45}a) and Fig.~\ref{fig:17vs45}b), respectively. In agreement with the expectation and the results of our simulations, the eigenmodes splitting at resonance is larger when the cavities separation is smaller.
\begin{figure}
\includegraphics[width=0.65\linewidth]{two_samples}
\caption{Difference of the eigenmodes energy as a function of the position on the sample for coupled ZnTe planar microcavities separated by a) 12 and b) 6 DBR pairs.}
\label{fig:17vs45}
\end{figure}
\end{suppinfo}
|
\section{Introduction}
Gravitational potential wells of galaxy clusters have been powerful laboratories to test the current model of dark matter ($\Lambda$CDM) and its alternatives.
While acknowledging many shortfalls of the $\Lambda$CDM model in galaxies \citep[e.g.][]{cuspcore1, cuspcore2, missing1, missing2, TBTF1, TBTF2, plane1,Pawlowski}, in the cluster arena few models can compete with the $\Lambda$CDM model, especially those alternatives which a priori assume that particle dark matter does not exist, but rather what we are seeing is a breakdown of Newtonian dynamics.
Observations of rotation curves in galaxies showed that dark matter effects were only required in low acceleration environments $\lessapprox 1.2\times10^{-10}$ ms$^{-2}$. This eventually led to the construction of the empirical gravitational paradigm known as Modified Newtonian Dynamics (MOND) \citep{milgrom19832,milgrom19833,milgrom19831,bekenstein1984}. The main function of MOND is to modify gravity in these low acceleration environments such that the gravitational acceleration falls proportional to $1/r$ in contrast to the Newtonian $1/r^{2}$. Newtonian dynamics are still preserved in the high acceleration environments. In order to achieve this, an acceleration scale was introduced to define what is meant by high and low acceleration environments, $a_{0} \approx 1.2\times10^{-10}$ ms$^{-2}$, such that Newtonian behaviour is recovered when $a>>a_{0}$ and the $1/r$ gravity law (Deep-MOND regime) occurs when $a<<a_{0}$, where $a$ is the total gravitational acceleration.
The MOND paradigm has had success on the galaxy scale, see \cite{famaeyreview} for an extensive review. One of the main problems in MOND is its inability to explain galaxy clusters, see for example \cite[][]{sanders1999,sanders2003}. Galaxy clusters tend to have an internal acceleration of the order $a_{0}$ and thus the MOND effect is weak. However, galaxy clusters show a large mass discrepancy from Newtonian predictions, much more than MOND is able to account for. This means that either { 1) There exists a $\Lambda$CDM dark matter halo, 2) There is missing matter which we are yet to detect in the form of non-luminous baryonic matter or some form of neutrinos or 3) MOND is not a complete gravity theory and needs to be generalised.} Work on { point 2} has achieved mixed results. \cite{angus20081} have shown that the 2 eV neutrino was insufficient to explain the galaxy cluster problem as their inclusion could not explain mass discrepancy in the centre of the clusters. The neutrino idea was then reinvestigated in \cite{angus2009} where 11 eV sterile neutrinos were tested. This work enjoyed more success in explaining the galaxy cluster problem in MOND and also had success in explaining the CMB anisotropies. However, cosmological simulations conducted by \cite{angusneutrino1} and \cite{angusneutrino2} showed that using neutrinos as hot dark matter in the MOND paradigm produces too many high mass galaxy clusters.
A recent addition to the galaxy cluster problem in MOND is the discovery of ultra-diffuse galaxies (UDGs) \citep[][]{UDGComa,UDGA168,UDGVirgo,UDGDF}. These galaxies have very little gas and are composed almost entirely of dark matter. Recent studies of a UDG in the Virgo cluster \citep{UDGVirgoObject} and the Coma cluster \citep{UDGComaObject} have shown that two UDGs, VC1287 and DF44 show a very high dark-to-stellar mass ratio.
In this work, we are interested in the nature of UDGs in the context of MOND. In a MONDian paradigm, it is possible to create a large dark matter-like effect if the gravitational acceleration across the system is very low. The MOND paradigm has an interesting feature called the external field effect (EFE) (see for example \cite{EFE1}, \cite{EFE2} and \cite{decliningcurves} for the interested reader). The EFE states that even a constant acceleration from an external source can affect the internal dynamics of a system. For example, a stellar cluster located close to the Milky Way disk should behave differently if it is moved further away from the disk as the gravitational acceleration across the cluster from the Milky Way would be less. In the context of the UDGs, if they were isolated objects, MOND would predict a large dark matter like effect, but as they are within the strong gravitational field of the galaxy cluster, MOND predicts they should behave closer to Newtonian.
Taking this into consideration, if MOND is to be generalised to try and explain the missing mass in galaxy clusters, it must also explain the nature of these UDGs. One modification to MOND which has been proposed is that of Extended MOND (EMOND) \citep{EMOND}. This extension of MOND changes the acceleration scale $a_{0}$ from being constant to being a function of gravitational potential, $A_{0}(\Phi)$, such that the effective acceleration scale in galaxy clusters is much larger than $a_{0}$. This allows deviations from Newtonian dynamics, and hence the inducing of dark matter-like effects, to occur at higher accelerations. We explored EMOND in \cite{HodsonEMOND} with a sample of 12 galaxy clusters. \cite{HodsonEMOND} showed that EMOND has some success with the basic formulation, but no attempt was made to explore the boundary conditions of the gravitational potential to try to get better fits. Also, the exact form of the baryonic mass profile is relevant when determining the EMOND prediction and thus different mass models should be tested in future. As a consequence of this, the paradigm requires more rigorous testing.
Recent work on the UDGs has allowed dynamical mass estimates to be made, that is the total mass of the UDGs including any dark component, for a sample of galaxies from the Coma and Virgo clusters using scaling relations \citep{UDGFM}. This method takes advantage of the fundamental manifold (FM) \citep[][]{FM1,FM2,FM3} to calculate velocity dispersions of the UDGs from their effective radius and surface brightness. The FM is an extension of the fundamental plane \citep[][]{FP1,FP2}. From the velocity dispersions, it is then possible to estimate a dynamical mass for the objects. It is also possible to estimate the stellar mass of the UDGs from their g-i colour. This technique was performed in \cite{UDGComaObject} for DF44 by using the colour-$M_{\star}/L$ correlation from \cite{GAMAStellar}. Therefore it is possible to get both dynamical and stellar mass estimates for a sample of UDGs.
By modelling the Coma cluster in the context of EMOND, we can find the value of $A_{0}(\Phi)$ in the cluster as a function of radius. Assuming this value is constant across any UDG, we can estimate the stellar mass of the UDGs from dynamical mass estimates using the EMOND recipe and compare the result to stellar mass predicted from the colour. By doing this, we can determine whether the EMOND recipe can predict both the mass profile of the Coma cluster and the dynamical to stellar mass fraction of the UDGs simultaneously.
This paper is organised as follows. Section \ref{EMOND} discusses the MOND and EMOND paradigm. Section \ref{Coma} discusses the Coma cluster model which we adopt. Section \ref{UDG} discusses the UDG dynamical and stellar mass estimates from the literature. The UDG modelling in the context of MOND and EMOND is discussed in Section \ref{MOND}. We show our results in Section \ref{Results}. In Section \ref{refine} we show how the constraints on the EMOND formalism from the UDG modelling affect the results of \cite{HodsonEMOND}. In Section \ref{contention}, we discuss possible contention with observational data. We then conclude in Section \ref{Conclusion}.
\section{Extended MOND}\label{EMOND}
We begin our discussion of EMOND by reviewing the standard MOND equations. In gravitational dynamics, the gravitational acceleration and matter density are linked via a Poisson equation. The MOND Poisson Equation is \citep{bekenstein1984},
\begin{equation}
4\pi G \rho = \nabla \cdot \left[ \mu\left(\frac{|\nabla \Phi|}{a_{0}} \right) \nabla \Phi \right]
\end{equation}
\noindent where $\rho$ is the matter density and $\Phi$ is the total gravitational potential. The function, $\mu(x)$ is called the interpolation function which models the transition between the Newtonian regime and the Deep-MOND regime. $\mu(x)$ must have limits such that when $x<<1$, $\mu(x) = x$ and when $x>>1$, $\mu(x)= 1$. The form for the interpolation function which we will use in this work is a modified simple interpolation (see \citep{simplemu,simple2} for simple interpolation function)
\begin{equation}
\mu(x)= \rm max \left[\frac{x}{1+x}, \frac{\epsilon}{1 + \epsilon} \right],
\end{equation}
\noindent where $\epsilon$ is a small number. The EMOND version of the MOND Poisson Equation is \citep{EMOND}\footnote{The additional $T_{2}$ term arises from the non-relativistic EMOND Lagrangian. Merely making the change $a_{0} \rightarrow A_{0}(\Phi)$ in the Poisson equation will not satisfy the Euler-Lagrange equation.}
\begin{equation}\label{EMONDPoiss}
4 \pi G \rho = \nabla \cdot \left[\mu\left( \frac{|\nabla \Phi|}{A_{0}(\Phi)} \right) \nabla \Phi \right] - T_{2},
\end{equation}
where
\begin{equation}
T_{2} = \frac{1}{8\pi G}\left| \frac{d (A_{0}(\Phi))^{2}}{d\Phi} \right|\left[ y F'(y) - F(y) \right].
\end{equation}
\noindent Also, $dF(y)/dy = \mu(\sqrt{y})$ and $y = |\nabla \Phi|^{2}/A_{0}(\Phi)^{2}$. It was shown explicitly in \cite{HodsonEMOND} that the $T_{2}$ term is negligible in clusters and thus the approximate spherical version of the EMOND Poisson Equation reduces to,
\begin{equation}\label{EMONDPoiss2}
\nabla \Phi_{N} \approx \mu\left( \frac{|\nabla \Phi|}{A_{0}(\Phi)} \right) \nabla \Phi
\end{equation}
\noindent where $\nabla \Phi_{N}$ is the Newtonian acceleration. The functional form of $A_{0}(\Phi)$ we use here is\footnote{In \cite{HodsonEMOND}, the function for $A_{0}(\Phi)$ was written as $A_{0}(\Phi) = a_{0} + (A_{0~max}-a_{0})\left[ \frac{1}{2}\tanh \left[ \log \left(\frac{\Phi}{\Phi_{0}} \right)^{q} \right] +\frac{1}{2}\right]$.}
\begin{equation}\label{A0tan}
A_{0}(\Phi) = \frac{a_{0}}{\epsilon} \mu\left[ \left( \frac{\Phi}{\Phi_{0}} \right)^{2q}\right]
\end{equation}
\noindent where $A_{0~max}$ is the maximum value which we allow $A_{0}$ to take $\approx 100 a_{0}$, $\Phi_{0}$ is a scale potential analogous to the MOND scale acceleration with units of m$^{2}$s$^{-2}$ and $q$ is a dimensionless parameter which controls the slope of $A_{0}(\Phi)$. We define $\epsilon$ to be $\epsilon = a_{0}/A_{\rm 0~max}$. Equation \ref{A0tan} says that when the potential is high ($\Phi >> \Phi_{0}$), $A_{0}(\Phi) = A_{0~max}$ and when the potential is low, ($\Phi << \Phi_{0}$), $A_{0}(\Phi) = a_{0}$. This is analogous to the MOND interpolation function, $\mu(x)$. In the work of \cite{HodsonEMOND}, the parameter choice of $q=2$ was used. In this work, we will also show results for $q=1$. The change in choice of $q$ warrants a change of scale potential, $\Phi_{0}$, as well. For $q=2$, the scale potential is unchanged from \cite{HodsonEMOND} with magnitude $\Phi_{0} \approx -2700000^{2}$ m$^{2}$s$^{-2}$ . For $q=1$, the scale potential is empirically chosen to be $\Phi_{0} \approx -3800000^{2}$ m$^{2}$s$^{-2}$. Therefore given a boundary potential, we can solve Eqn \ref{EMONDPoiss2} and determine the EMOND predicted acceleration profile and hence EMOND predicted dynamical mass.
\section{Modelling The Coma Cluster}\label{Coma}
The first step to modelling the Coma cluster UDGs is to build a model of the Coma cluster itself. We adopt the model of \cite{ComaModel} which has an intra-cluster gas component and cluster galaxy component. There is also a dark matter component in standard gravity, which we can compare with the effective phantom halo predicted by EMOND.
The gas is modelled via a $\beta$ density profile for which the expression for enclosed mass is,
\begin{equation}
M_{g}(r) = \frac{4}{3} \pi n_{0} (m_{e} + \gamma m_{p}) r^{3} F_{3/2,\beta}\left(\frac{r^{2}}{r_{c}^{2}}\right)
\end{equation}
\noindent where $n_{0}$ is the central electron number density of the emitting X-ray gas in the cluster, $\beta$ is a dimensionless parameter, $r_{c}$ is a scale length of the gas density, $\gamma$ is a parameter which converts the electron number density into a mass density and $F_{\alpha,\beta}(x) \equiv {}_{2}F_{1}\left( 3-\alpha, (3-\alpha)\beta);4-\alpha; -x\right),$ where ${}_{2}F_{1}$ is a hyper-geometric function.
The galaxies were modelled via,
\begin{equation}
M_{Gal}(r) = 4 \pi L_{\star} \Upsilon r_{s}^{3} \left[ \log\left( \frac{r+ r_{s}}{r_{s}} \right)-\frac{r}{r+r_{s}} \right]
\end{equation}
\noindent where $r_{s}$ is a scale radius, $L_{\star}$ is a luminosity normalisation constant and $\Upsilon$ is a mass-to-light ratio.
Finally, in the work of \cite{ComaModel} the dark matter mass was modelled via,
\begin{equation}
M_{DM}(r) = M_{v} \left( \frac{r}{r_{v}} \right)^{3-\alpha} \frac{F_{\alpha,1}(c r/r_{v})}{F_{\alpha,1}(c)}
\end{equation}
\noindent where $M_{v}$ is the virial mass, $r_{v}$ is the virial radius, $c$ is the concentration and $\alpha$ is a dimensionless parameter.
We plot the mass components of the Coma cluster as a function of radius in Figure \ref{ComaMass}, mimicking the top panel of Figure 8 in \cite{ComaModel}. We over-plot the EMOND predicted mass profile (blue solid line) determined by first solving Equation \ref{EMONDPoiss2}{ for EMOND gravity, which we will call $\nabla \Phi_{EMOND}$ }, and then calculating the effective EMOND mass via,
\begin{equation}\label{EMONDMass}
M_{EMOND}(r) = \frac{r^{2} \nabla \Phi_{EMOND}}{G}.
\end{equation}
{ where the Newtonian gravity used to determine $\nabla \Phi_{EMOND}$ is}
\begin{equation}
\nabla \Phi_{N} = \frac{G (M_{g}(r) + M_{Gal}(r))}{r^{2}}.
\end{equation}
To make the plot, we empirically take a value of the EMOND gravitational potential at the virial radius to be $\Phi(r_{v}) = -2.5\times10^{12}$ m$^{2}$s$^{-2}$\footnote{ The task of EMOND would be to determine the boundary potential from cosmological constraints. Due to the lack of a consistent cosmology we are, at this stage, limited to empirically fitting. Future work on EMOND can determine whether our empirical fit is acceptable.}. We can see that the dark matter dominates over the gas and galaxy contributions.
\begin{figure}
\includegraphics[scale=0.7]{ComaClusterMass.pdf}
\caption{Model of the Coma cluster which we adopt from \cite{ComaModel}. Green line shows the contribution from the intra-cluster gas, red thin line is the contribution from the { stars}. Using these, we can calculate the EMOND predicted dynamical mass from Equation \ref{EMONDMass} which is the solid blue line for $q=1$ model and solid magenta line for $q=2$ model (see Eqn \ref{A0tan}). We also plot the dark matter profile from \cite{ComaModel} (black dashed line) for comparison. We see that our EMOND mass matches the dark matter mass very well. For this, we assume an EMOND boundary potential at the virial radius $\Phi(r_{v}) = -2.5 \times 10^{12}$ m$^{2}$ s$^{-2}$. }
\label{ComaMass}
\end{figure}
The plot also shows that the EMOND predicted mass seems to match the dark matter profile to exceptional accuracy. This is a very good result for the EMOND paradigm. In the previous EMOND work, \cite{HodsonEMOND} found that EMOND had mixed success in describing the clusters. The work in question used a different baryonic mass profile for both the gas and the galaxies. This result suggests that the EMOND modelling of \cite{HodsonEMOND} might be improved by invoking a different functional form for the baryonic mass profile.
Now that we have derived the EMOND mass profile of the Coma cluster, we can make a plot of $A_{0}(\Phi)$ vs radius to show how the EMOND acceleration scale varies in the cluster environment. We show this in Figure \ref{ComaA0}.
\begin{figure}
\includegraphics[scale=0.6]{ComaClusterA0.pdf}
\caption{Profile of the EMOND calculated $A_{0}(\Phi)/a_{0}$ as a function of cluster radius. Blue dashed line is the $q=2$ model and red solid line is $q=1$ model. The $q=1$ model produces a shallower transition from high to low $A_{0}(\Phi)$ and a smaller in magnitude of $A_{0}(\Phi)$ than the $q=2$ model (see Eqn \ref{A0tan}). We only show radii $> 100$ kpc as this is the important range for the UDGs.}
\label{ComaA0}
\end{figure}
From Fig~\ref{ComaA0}, it is clear that the $q=1$ model creates a gentler transition of $A_{0}(\Phi)$ from the outside of the cluster to the centre than the $q=2$ model. It is also clear that the magnitude of $A_{0}(\Phi)$ which the $q=1$ model predicts is much lower than the $q=2$ model. This is due in part from the gentler transition, but mainly the choice of $\Phi_{0}$, which stops $A_{0}(\Phi)$ reaching $A_{0~max}$. We will see the effect of this in the following sections.
\section{UDG Properties}\label{UDG}
{ A note on conventions. Throughout the following sections we will be referring to several mass quantities which we should define clearly. The ``dynamical" mass of the UDGs is the inferred mass from dynamics, thus is the total mass of the system. In a $\Lambda$CDM context this would be the mass of the stars + the mass of the dark matter halo. In MOND/EMOND, this would be the baryons + phantom dark matter. The Newtonian mass is the total mass of the baryons, in this case stars. }
When we model the UDGs in EMOND, we will take the dynamical mass of the UDG at the effective radius and determine the Newtonian stellar mass at that radius using the EMOND recipe. We will then compare the Newtonian mass to the estimated stellar mass of these galaxies from their colour and mass-to-light ratio. Therefore, we need to determine both the dynamical mass and stellar mass for these systems. To do this, we follow the techniques used in \cite{UDGFM} and \cite{FM3} for the dynamical mass and \cite{UDGDF} for the stellar mass. We outline the techniques used in these works below.
\subsection{Dynamical Mass}\label{DynamicalMassSec}
The dynamical mass of the UDGs is determined from the velocity dispersion and effective radius via the formula \citep{wolf} (also see Eqn 1 in \cite{UDGComaObject}),
\begin{equation}\label{MWolf}
M_{\rm dyn}\left|\right._{r_{\rm s}={4 \over 3} r_e} \approx 3 \sigma^2 r_{\rm s}/G = 9.3 \times 10^{5} \sigma^{2} r_{\rm e}
\end{equation}
where $M_{\rm dyn}(<r_{\rm s}) $ is the total enclosed dynamical mass at the spherical half-mass radius $r_{\rm s} \approx {4 \over 3} r_e$, where the $r_{e}$ is the usual effective radius, i.e., the projected circularised half-light radius, $\sigma$ is the velocity dispersion in km/s and $r_{e}$ is the effective 2D radius in kpc. The effective radius was determined and corrected for ellipticity for 46 UDGs within the Coma cluster\footnote{The sample actually has 47 objects, but one object has incomplete data in the table and thus we disregard this entry.} and are given in \cite{UDGDF}. Currently, there is only one UDGs (Dragonfly 44 (DF44)) in the Coma cluster which has a measured value for the velocity dispersion. Therefore, to estimate velocity dispersions for all 46 galaxies in the Coma cluster sample, some assumptions have to be made.
We have taken a slightly different approach for our study instead of that of \cite{UDGFM}.
\footnote{In \cite{UDGFM}, the velocity dispersions are determined for the UDGs in the Coma cluster by making use of the fundamental manifold (FM). This relation links effective radius, mean surface brightness within the effective 2D radius and the internal kinematics of the system in question via a nearly power-law-like relation
$\log \Upsilon_{e} = 0.24 \left( \log V \right)^{2} + 0.12 \left( \log I_{e} \right)^{2} - 0.32 \log V - 0.83 \log I_{e} - 0.02 \log \left( V I_{e} \right)+ 1.49.$
where $\Upsilon_{e}$ is the mass-to-light ratio, $I_{e}$ is the mean surface brightness within $r_{e}$ and $V$ describes the kinematics of the system. This was then solved along with the known relation
$\log r_{e} = 2 \log V - \log I_{e} - \log \Upsilon_{e} - C,$
which is derived from Eqn \ref{MWolf}, to determine $V$ and $\Upsilon_{e}$. It was then assumed in \cite{UDGFM} that $V \approx \sigma$. This value of the velocity dispersion was then corrected via $\log \sigma_{corr} = (\log \sigma - 0.061)/0.833$ to account for a {\it `slight systematic deviation from the expectation'}. Therefore \cite{UDGFM} were able to obtain estimates for the velocity dispersions and thus dynamical masses for the UDGs.}
Using the FM relation from a previous study by \cite{FM3}, there exists a relationship between $I_{e}$, $r_{e}$ and $\sigma$, without having to solve the system equations in \cite{UDGFM},
\begin{equation}\label{zaritsky2}
\log r_{e} = -\alpha^{2} \log^{2}\sigma + (2 + 2 \alpha \beta) \log \sigma + B \log I_{e} + C_{2},
\end{equation}
In this equation, $\alpha$, $\beta$ and $C_{2}$ are constants which are empirically determined, taking values (Equation 8 and Figure 11 from \cite{FM1}) $\alpha^{2} \approx 0.63$, $2 + 2\alpha \beta \approx 3.7$, $B \approx -0.705$ and $C_{2} \approx -2.75$. We can use Eqn \ref{zaritsky2} to find the velocity dispersion analytically using the data given in \cite{UDGDF}. The only other difference between our method and \cite{UDGFM} is that we will not make the correction to the velocity dispersion\footnote{The FM line in Figure 11 of \cite{FM1} seems to align well with the data points, hence we do not make a correction.} and assume, for now, that all the UDGs lie on the FM. We will discuss the implications of this later.
The final discussion point is to convert the data table in \cite{UDGDF} to the correct units for the fundamental manifold equation. The fundamental manifold has a 2D effective radius in units of kpc and a mean surface brightness in units of $L_{\odot}/pc^{2}$. To determine the correct radius, we need to take the radii in column 5 (which is the major axis radius) of the table in \cite{UDGDF} and multiply in by the square root of the axis ratio, given in column 7 of the table. For the surface brightness, we need to use a standard conversion to change the central surface brightness, given in column 4 of the table in \cite{UDGDF} in mags/arcsec$^{2}$, into mean surface brightness within an effective radius in $L_{\odot}/pc^{2}$. This is done by
\begin{equation}\label{SBprofile1}
\log <I_{e}> = -\frac{I_{0} + 1.822 - 0.699 - M_{\odot} - 21.572}{2.5},
\end{equation}
where in this case, $M_{\odot}$ is the solar magnitude in the given band, $<I_{e}>$ is the mean surface brightness within an effective radius in $L_{\odot}/pc^{2}$ and $I_{0}$ is the central surface brightness in mags/arcsec$^{2}$. See Appendix \ref{Appendix1} for derivation of Eqn \ref{SBprofile1}.\footnote{The given formula for converting the surface brightness can be more general depending on the S$\acute{\rm e}$rsic index of the modelling. As \cite{UDGDF} used a S$\acute{\rm e}$rsic value of 1 for all UDGs, the above formula is valid for all the galaxies in our sample. }
Once we apply these conversions, we can use Eqn \ref{zaritsky2} to find the estimated velocity dispersion for each UDG and use Eqn \ref{MWolf} to determine the enclosed mass within the 3D radius.
\subsection{Estimating The Stellar Mass at the Effective Radius}
{ In the following sections we will be outlining how to infer the predicted stellar mass in the UDGs in the EMOND paradigm from the dynamical mass estimate described above. Therefore, test the validity of the EMOND formula, we require the approximate enclosed stellar mass at the effective radius for each UDG in the Coma cluster.} In order to do this we follow the technique used in \cite{UDGDF}. This work takes advantage of the relation between colour and mass-to-light ratio, used in \cite{GAMAStellar} which describes a link between the ({\it g}-{\it i}) colour and the stellar mass-to-light ratio in the {\it i}-band,
\begin{equation}\label{Mstellar}
\log_{10} \left[ M_{\star}{\cancel /M_{\odot}}\right] = 1.15 + 0.7(g-i) - 0.4 M_{i}
\end{equation}
\noindent where $M_{i}$ is the absolute magnitude in the {\it i}-band and $M_{\odot}$ is the solar mass, not to be confused with the solar magnitude used previously. From this, we can calculate the stellar mass using only colour and magnitude. The g-band magnitude is given for 46 UDG in the Coma cluster in \cite{UDGDF}. For the sample, the average {\it g}-{\it i} colour is $<g-i> \approx 0.8 \pm 0.1$. { This is the value we adopt for each UDG}. Therefore the {\it i}-band magnitude can be calculated from the quoted g-band magnitude via $M_{i} \approx M_{g}-0.8$. We therefore have all the necessary quantities to derive a stellar mass for the UDGs. Note, the mass calculated via Eqn \ref{Mstellar} is the total mass. The stellar mass within $r_s$, which is what we are interested in, is half of $M_{\star}$.
\subsection{Distance From Centre of Cluster}\label{UDGDistanceSec}
As we only have the 2D projected map of the Coma cluster and the UDGs, it is not possible to get their exact radii from the centre of the cluster. We can however calculate the minimum radius the UDGs should be from the right ascension and declination of the UDGs, as given in \cite{UDGDF}. If we assume that all the UDGs lie at the same distance as the Coma cluster itself, we can find their minimum distance from,
\begin{equation}
d_{\rm UDG-Coma} \approx d_{\rm Coma}\theta_{\rm UDG-Coma}
\end{equation}
where $d_{\rm Coma}$ is the distance to the Coma cluster and $\theta_{\rm UDG-Coma}$ is the angular separation in radians between the UDG and the Coma cluster centr
\begin{figure}
\includegraphics[scale=0.7]{D_UDG_Coma.pdf}
\caption{The minimum projected distance between the centre of the Coma cluster and the UDGs in kpc. The average distance is approximately 1300 kpc. Note that this is the projected distance and the true 3D distance will be higher than this.}
\label{D_UDG_Coma}
\end{figure}
We can see from Fig \ref{D_UDG_Coma} that the the average distance is approximately 1300 kpc, quite far from the cluster centre, with minimum and maximum values of 296 kpc and 2811 kpc respectively. However, as stated, the actual 3D radii will be on average higher than this.
\section{MOND and EMOND Modelling}\label{MOND}
In this section we describe how the UDGs were modelled in the regular MOND and EMOND paradigms. To do this we take the dynamical mass, derived from the predicted velocity dispersions (Section \ref{DynamicalMassSec}), and substitute the value into the MOND (and EMOND) formula. From this, we can then determine the Newtonian mass which is required to satisfy the MOND equations. Assuming that the galaxy is dominated by stellar mass, we can then compare this Newtonian mass to the stellar mass derived in Section \ref{UDG}. If the MOND paradigm is correct, these two methods should be consistent. All this modelling is conducted under the assumption that the UDGs are spherical\footnote{ The average b/a ratio for the sample is 0.74}.
\subsection{MOND}
To begin the MOND modelling, we start by assuming that the UDGs are isolated systems. If they were isolated, we can use the simple spherical MOND relation to model them,
\begin{equation}\label{MOND1}
\nabla \Phi_{Newt} = \mu\left( \frac{\nabla \Phi_{dyn}}{a_{0}} \right)\nabla \Phi_{dyn}
\end{equation}
where $\nabla \Phi_{Newt} = G M_{Newt}(r)/r^{2}$ is the Newtonian acceleration and $\nabla \Phi_{dyn} = G M_{dyn}(r)/r^{2}$ is the dynamical acceleration. As discussed, we can then find the MOND predicted Newtonian mass from the calculated dynamical mass of the UDGs.
However, this is not the correct picture as UDGs are not isolated, they are within the external field of the cluster. The MOND formula has to be modified to take into consideration the external field of the cluster,\footnote{ The results are found to be nearly the same when we assume $a$ and $g_{ext}$ are orthogonal.}
\begin{equation}\label{MONDPoissExt}
\begin{split}
&\sqrt{(\nabla \Phi_{N})^2 + (\nabla \Phi_{N~ext})^{2}} \approx \\ &\mu \left( \frac{\sqrt{(\nabla \Phi_{dyn})^{2} + (\nabla \Phi_{ext})^{2}}}{a_{0}} \right)\sqrt{(\nabla \Phi_{dyn})^{2} + (\nabla \Phi_{ext})^{2}}
\end{split}.
\end{equation}
Assuming that the external field is entirely dominated by the Coma cluster, we determine the magnitude of the external field from our model of the Coma cluster in Section \ref{Coma}. The external field used for each UDG is determined from the distance it is from the centre of the cluster, which we calculated in Section \ref{UDGDistanceSec}.
We expect that the external field increases the overall acceleration across the UDGs, pushing the internal dynamics closer to Newtonian as the MOND interpolation function argument is increased. This highlights the tension between the MOND paradigm and the UDG observations.
\subsection{EMOND}
As we have seen in our Coma cluster EMOND model, the effective value of $a_{0}$ is increased within the cluster. This could raise the dark matter-like effects within the UDGs even with the external field of the cluster dominating the dynamics. This is due to something called the external potential effect. As the UDGs are in the deep potential well of the Coma cluster, under the prediction of the EMOND paradigm, the internal dynamics of the UDGs are affected. The modified version of Equation \ref{MONDPoissExt} for EMOND is,
\begin{equation}\label{EMONDPoissExt}
\begin{split}
&\sqrt{(\nabla \Phi_{N})^2 + (\nabla \Phi_{N~ext})^{2}} \approx \\ &\mu \left( \frac{\sqrt{(\nabla \Phi_{dyn})^{2} + (\nabla \Phi_{ext})^{2}}}{A_{0}\left( \Phi_{dyn} + \Phi_{ext} \right)} \right)\sqrt{(\nabla \Phi_{dyn})^{2} + (\nabla \Phi_{ext})^{2}}
\end{split}.
\end{equation}
Making the assumption that $A_{0}(\Phi)$ is approximately constant across the UDGs as they are so small\footnote{Although the gravitational potential of the Coma clusters dominate the UDGs in our model, the gravitational accelerations of the UDGs are still relevant and thus we do not neglect them.}, we can rewrite Equation \ref{EMONDPoissExt} as
\begin{equation}\label{EMONDPoissExt2}
\begin{split}
(\nabla \Phi_{N})^{2} &= \mu \left( \frac{\sqrt{(\nabla \Phi)^{2} + (\nabla \Phi_{ext})^{2}}}{A_{0}(\Phi_{ext})} \right)^{2} (\nabla \Phi)^{2} + (\nabla \Phi_{ext})^{2} \\ &- \mu \left( \frac{\nabla \Phi_{ext}}{A_{0}(\Phi_{ext})} \right)^{2} \nabla \Phi_{ext}^{2}.
\end{split}
\end{equation}
\noindent where we have eliminated $\nabla \Phi_{N~ext}$ from Eqn \ref{EMONDPoissExt} via $\nabla \Phi_{N~ext} = \mu\left(\frac{\nabla \Phi_{ext}}{A_{0}(\Phi_{ext})}\right)\nabla \Phi_{ext}$.
Equations \ref{MOND1}, \ref{MONDPoissExt} and \ref{EMONDPoissExt2} can then be used to calculate the predicted Newtonian mass of the UDGs given the dynamical mass of the UDGs and the external field and potential, which is derived from the fundamental manifold and the Coma model respectively.
\section{Results}\label{Results}
For our results, we do not perform a rigorous error analysis as there are many sources of error from all the measurements and modelling of the UDGs as well as scatter from the FM and the model of the Coma cluster. We aim to determine simply whether EMOND is a possible explanation for the UDG over-massive dark haloes.
In the following plots we show the ratio of the Newtonian mass, predicted by the MOND and EMOND models, and the stellar mass calculated from the colour. Ideally, this ratio should be 1. If the ratio is less than 1, either the MOND paradigm predicts that there is less mass than is permitted by the stellar mass estimates or the stellar mass estimate is too high. If the ratio is more than 1, the MOND formulation predicts that there is more mass present than is permitted by the stellar mass estimates or the stellar mass estimates are too low.
We begin by showing the result for a MOND model with no effects from the Coma cluster (Figure \ref{MONDnoExt}). We see that for regular MOND the overall trend seems to be that the ratio is less than one by a factor of approximately 2. Therefore, perhaps within the errors, MOND with no external field might be sufficient in explaining the UDG masses.
\begin{figure}
\includegraphics[scale=0.7]{MOND_No_Ext_Ratio.pdf}
\caption{Figure showing the ratio of the MOND estimated Newtonian mass to the estimated stellar mass from colour { as a function of the distance to the cluster centre}. No effect from the Coma cluster considered. }
\label{MONDnoExt}
\end{figure}
We next show in Figure \ref{MONDExt} how the external field affects the result. As expected, the cluster boosts the acceleration across the UDGs, increasing the argument in the MOND interpolation function, and thus driving the systems closer to Newtonian. We therefore see that including the external field makes the MOND model worse, the ratio is larger than 1, therefore requiring much more stellar mass than is available according to the colour estimate.
\begin{figure}
\includegraphics[scale=0.7]{MOND_Ext_Ratio.pdf}
\caption{Same as Figure \ref{MONDnoExt} except with the inclusion of the external field from the Coma cluster. The MOND predicted mass is much larger than the colour predicted stellar mass.}
\label{MONDExt}
\end{figure}
We then show how the EMOND effect of increasing $a_{0}$ across the UDGs changes the result (Figure \ref{EMONDExt}). We find that the EMOND prediction improves the MOND fit substantially within the expected errors. We also note that in the $q=2$ model (top panel), there seems to be a trend such that the further out the UDG, the higher the predicted Newtonian mass from the EMOND formalism compared to the stellar mass. This is less of an issue with the $q=1$ model, demonstrating that the UDGs provide a stringent constraint in the allowed functional form of $A_{0}(\Phi)$. This might be an indication that rigorous numerical testing and a larger sample of UDGs might find that further refining the EMOND parameters and interpolation function might produce an even better fit. This is beyond the scope of this paper. Another point of note is the fact that the outer UDG values are similar in the MOND and EMOND case. This is due to the EMOND formalism asymptotically tending to MOND in the outer part of the cluster, as desired.
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=0.7]{EMOND_Ratio.pdf}\\
\includegraphics[scale=0.7]{EMOND_Ratio2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MONDExt} except with the EMOND correction to the MOND acceleration scale. Top panel is the $q=2$ model and bottom panel is the $q=1$ model (see Eqn \ref{A0tan}). The EMOND paradigm is predicting a reasonable Newtonian mass for the UDG sample in both models. The $q=2$ model shows that the required mass to light increases with distance, which is an undesirable feature. The $q=1$ model shows that a constant mass-to-light with distance is a good fit to the data, which seems more plausible.}
\label{EMONDExt}
\end{figure}
The above results seem to show that if we take at face value the dynamical mass of the UDGs, stellar mass of the UDGs, EMOND function and the model of the Coma cluster, that EMOND is able to explain the Coma cluster mass profile, and the UDGs within it. The $q=1$ model produces a better fit to the data than the $q=2$ model in terms of how the distance of the UDGs from the centre of the Coma cluster is affected by EMOND.
There will undoubtedly be sources of error within this calculations from spherical symmetry assumptions, scatter around the FM, the error in the Coma cluster mass model etc which will alter the result. The main source of error is most likely the uncertainty in the stellar mass-to-light ratio and the use of the M/L - (g-i) relation.
To get an idea of the error in the stellar mass-to-light ratio, we recreate Figure 13 from \cite{GAMAStellar} in Fig~\ref{GAMAPlot2} with the stellar mass-to-light used in this work (red band) and the \citep{BellStellar} function (blue band). There is quite a bit of contention between these two estimates of the M/L ratio. It is possible to reverse engineer the question by assuming that the EMOND formalism is correct and determining the required value of the stellar mass-to-light ratio of each object. For this, we assume that each UDG lies on the <{\it g}-{\it i}> = 0.8 line. We assume both functions have an approximate error of 0.1 dex (coloured band region for each function), which is reasonable according to the literature \citep{GAMAStellar}. We then determine the required value of the stellar mass to match the EMOND predicted mass and determine where on the mass-to-light plot each UDG lies. We show this in Fig \ref{GAMAPlot2}. We should remind ourselves at this point that the 0.8 value is an average with an error of $\pm 0.1$, therefore there is an extra source of uncertainty.
\begin{figure}
\begin{tabular}{c}
\includegraphics[scale=0.7]{MLratioGAMA2.pdf}\\
\includegraphics[scale=0.7]{MLratioGAMA1.pdf}
\end{tabular}
\caption{Plot showing the stellar mass-to-light functions from \cite{GAMAStellar} (red) and \cite{BellStellar} (blue) as a function of g-i colour. We show approximate error bars of 0.1 dex for each case. Top panel shows results for $q=2$ model and bottom panel shows the $q=1$ model (see Eqn \ref{A0tan}). The blue dots show where the UDGs must lie assuming that the EMOND formulation is correct. This shows that it may be possible for most of the UDGs to be explained by EMOND within the range of stellar mass-to-light allowed. The $q=1$ model again shows better results.}
\label{GAMAPlot2}
\end{figure}
Figure \ref{GAMAPlot2} shows that there seems to be a very large scope for error in the mass-to-light of the stars on the UDG, within which, most UDGs in our sample lie. We can therefore conclude that adjusting the stellar mass-to-light ratio can explain the UDGs mass, within the error bars, assuming that the EMOND modelling of the UDGs is valid.
\section{Adjusting the EMOND Formulation}\label{refine}
The above results show that the $q=1$ model fits the UDGs better than the $q=2$ model used in \cite{HodsonEMOND}. Therefore, for completeness, we should redo the analysis of \cite{HodsonEMOND} to check the $q=1$ model is consistent with the cluster sample of \cite{sample}. To do this, we will briefly review the \cite{HodsonEMOND} work and recreate their Figures 17-22 with the updated function for $A_{0}(\Phi)$.
One method of testing modified gravity theories is by comparing the estimated mass, derived from Newtonian dynamics and the mass calculated by assuming the inter-cluster gas is in hydrostatic equilibrium, which we call the dynamical mass. We have discussed how to find the Poisson predicted mass in the above sections. The dynamical mass is determined by solving the equation of hydrostatic equilibrium,
\begin{equation}\label{DynamicalMass}
M_{dyn}(r) = -\frac{kT(r)r}{G wm_{p}}\left[ \frac{d \ln \rho_{g}(r)}{d \ln r} + \frac{d \ln T(r)}{d \ln r} \right]
\end{equation}
where $\rho_{g}(r)$ is the density of the gas, $T(r)$ is the temperature of the gas, $k$ is the Boltzmann constant, $m_{p}$ is the proton mass and $w$ is the mean molecular weight. Therefore, for a given gas density and temperature, the dynamical mass can be calculated.
The last aspect to discuss is the determining of the boundary potential used to solve the Poisson Equations for each cluster. To get an estimate, \cite{HodsonEMOND} used the analytical best fit NFW profiles for each cluster and assumed that $\Phi(r_{\rm out}) \approx \Phi_{NFW}(r_{\rm out})$ where $r_{\rm out}$ was defines as some boundary outside the cluster. They then showed the range of solutions from $\Phi(r_{\rm out}) = (0.5 - 1.5) \times \Phi_{NFW}(r_{\rm out})$ to get an idea of how changing the boundary potential affects the result. { Here, to be consistent, we set each boundary potential to take the same value used for the Coma cluster. We also have modelled the galaxies for each cluster to have a similar mass profile as the Coma cluster in the central regions. { We note that each cluster will in practice have a different baryonic profile for the galaxies within the cluster.} We also show the boundary potential for $\Phi(r_{v}) = (0.9 - 1.1) \times \Phi(r_{v})$ in contrast to the previous work.} Better fits might be possible by numerically playing with this value, not addressed here\footnote{Better fits may also be found by using a different galaxy and/or gas model. Again, not addressed here.}.
Therefore, redoing the above steps for the new $A_{0}(\Phi)$ function, we show the updated mass plots for the cluster sample (Figs \ref{MassPlotEMOND} - \ref{MassPlotEMOND6}).
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{A133EMOND2.pdf} & \includegraphics[scale=0.5]{A262EMOND2.pdf}
\end{tabular}
\caption{Plot showing recreated Figs 17-22 from \cite{HodsonEMOND} with the modified $A_{0}(\Phi)$ function found under the UDG constraints. { Red dashed line is the best-fit $\Lambda$CDM model from \cite{sample}, black line is the dynamical mass derived from Eqn \ref{DynamicalMass} and the blue shaded region is the EMOND predicted mass. Here we show clusters A133 and A262.}}
\label{MassPlotEMOND}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{A478EMOND2.pdf} & \includegraphics[scale=0.5]{A1413EMOND2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MassPlotEMOND} for clusters A478 and A1413.}
\label{}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{A1795EMOND2.pdf} & \includegraphics[scale=0.5]{A1991EMOND2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MassPlotEMOND} for clusters A1795 and A1991.}
\label{}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{A2029EMOND2.pdf} & \includegraphics[scale=0.5]{RXJ1159EMOND2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MassPlotEMOND} for clusters A2029 and RXJ1159.}
\label{}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{MKW4EMOND2.pdf} & \includegraphics[scale=0.5]{A383EMOND2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MassPlotEMOND} for clusters MKW4 and A383.}
\label{}
\end{figure*}
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[scale=0.5]{A907EMOND2.pdf} & \includegraphics[scale=0.5]{A2390EMOND2.pdf}
\end{tabular}
\caption{Same as Figure \ref{MassPlotEMOND} for clusters A907 and A2390.}
\label{MassPlotEMOND6}
\end{figure*}
In these plots, the blue denotes the predicted mass by EMOND with the shaded region showing how the value is affected by different choices of the boundary potential, the red dashed curve is the NFW prediction from \cite{sample} and the black line is the dynamical mass, predicted from hydrostatic equilibrium. The EMOND blue curve should match the black curve, within errors related to modelling assumptions. We can see from Figs \ref{MassPlotEMOND}-\ref{MassPlotEMOND6} that the new form of $A_{0}(\Phi)$ gives a better fit than those found in \cite{HodsonEMOND}.{ It is also possible to see in cases such as RXJ1159 and MKW4 that even though the EMOND predicted dynamical mass is far from the dynamical mass from hydrostatic equilibrium, the EMOND curve is consistent with the best fit NFW curve (red dashed). Therefore we can conclude that changing the form of the interpolation function is still consistent with the previous EMOND work.}
\section{Contention With Observation}\label{contention}
Although our above analysis has shown a consistency between the Coma cluster and UDGs masses under the EMOND paradigm, we have made some rather large assumptions, the main assumption being that the FM can be used to determine the velocity dispersion of the UDGs. If we take the estimate for DF44, which is the only UDG in the sample which has been observed ($\approx$ $47^{+8}_{-6}$ km/s), the FM under-predicts the velocity dispersion by a factor of $\approx$ 2.7 (FM predicted velocity dispersion for DF44 is $\approx$ 17.4 km/s). If we then take the published observed data from \cite{UDGComaObject}, EMOND would predict a Newtonian mass of $\approx$ $7.7 \times 10^{8}$ M$_{\odot}$ and thus the ratio between this EMOND (q=1 model) predicted Newtonian mass and the stellar mass is $\approx$ 7.5 which is quite a substantial difference. This result is improved if the EMOND boundary potential chosen for the Coma cluster is increased. Choosing the boundary to be $3.5 \times 10^{12} \rm m^{2}s^{-2}$, the ratio is reduced to $\approx$ 6. If this potential was chosen, and we took the lowest bound for the velocity dispersion (41 km/s), the ratio is further decreased to $\approx$ 4.5. This could be further improved by choosing a higher stellar mass-to-light ratio than is used in \cite{UDGComaObject}. However, it must be checked what values for the boundary potential are allowed by the data for Coma. This would require further work, beyond the scope of this paper.
The reason our result differs from the work of \cite{UDGFM} is that the velocity dispersion is corrected due to there being a discrepancy between the observed velocity dispersion and the FM estimated value (see Figure 1 of \cite{UDGFM}). In our analysis, we used a different form of the FM. Further study as to the source of the discrepancy should be investigated in further work.
More detailed observations of more UDGs in the Coma cluster will be required to determine whether the over-massive dark halo of DF44 is a statistical outlier in the sample, or whether the interpretation of the FM used in our work is at a contention with the current observations.
\section{Conclusion}\label{Conclusion}
In this work, we modelled the Coma cluster in the EMOND paradigm and compared the predicted enclosed mass profile to that of a pure Newtonian model. We find that the EMOND result bears an extraordinary resemblance to the DM profile used in \cite{ComaModel}. This is quite a successful result for the EMOND paradigm. The success of this result should warrant further study into EMOND, taking into careful consideration the functional form of the baryonic mass profile and the boundary potential used to solve the Poisson equation.
We then moved on to make a model of UDGs in EMOND. We used this to determine the predicted Newtonian mass required to satisfy the EMOND formula. We then compared this to the stellar mass predicted by the UDG galaxy colour.
Our model seemed to give consistent values of the EMOND predicted Newtonian mass and the stellar mass derived from colour, within the error bars of the stellar mass-to-light. Further to this, the UDG sample gave a constraint on the exact function of $A_{0}(\Phi)$. Using a slightly different function to that of \cite{HodsonEMOND} yielded better results. This function was also checked against the earlier work of \cite{HodsonEMOND} yielding not only consistent, but better results. We can therefore conclude that the $q=1$ model is preferred by the EMOND paradigm.
However, the results of this work seem to be at contention with observations of DF44. A reanalysis of this calculation must be conducted when more UDG velocity dispersions are observed.
The UDGs serve as a very good test for MOND-like gravity theories and should be studied in more detail. The next step would be to conduct the same analysis for the Virgo cluster and its UDG population.
UDGs are still a relatively new discovery, with limited observations and a small sample size. We predict that more accurate measurements will be made of the velocity dispersions for the UDGs in the near future and with that comes more accurate dynamical mass estimates. It is hard to discuss possible formation scenarios in the context of EMOND as it is still a relatively new theory with limited research conducted. We hope that the take away message of this work is that a possible solution to the mass discrepancy in galaxy clusters in a MOND-like paradigm, EMOND, may also hold the answer to the nature of these UDGs. When two problems have one solution, it warrants further investigation and we hope that EMOND will be investigated further as a result of this.
\section*{ACKNOWLEDGEMENTS}
We would like to thank Anne-Marie Weijmans and Benoit Famaey for general comments on the draft. We would also like to thank Dennis Zaritsky for discussions on the fundamental manifold. AOH is supported by Science and Technologies Funding Council (STFC) studentship (Grant code: 1-APAA-STFC12).
\bibliographystyle{aa}
|
\section{Introduction}
Quantum information one may truly define as an ingenious investigation into the operational properties that differentiate classical systems from quantum systems. We believe it important from a fundamental perspective, to extend this definition so as to encompass all the properties discriminating classical from generically \emph{non-classical} systems~\cite{lamiatesi}. A useful framework for treating physical theories from a more general point of view, without losing the operational interpretation associated with them, is known as general probabilistic theories (GPTs, for short). Intuitively, one can think of GPTs as theories that are analogous to quantum mechanics in many ways, but with the fundamental difference that the cone of (unnormalised) states is no longer that of positive semidefinite matrices, but rather a generic convex cone in a finite-dimensional real vector space. In the rich landscape provided by GPTs, we investigate one particular phenomenon appearing in non-classical theories, \emph{data hiding}~\cite{dh original 1, dh original 2}.
In quantum mechanics, data hiding is usually intended as the existence of pairs of states of a bipartite system that are perfectly distinguishable with global measurements yet almost indistinguishable when only protocols involving local operations and classical communication (LOCC) are allowed. We extend the relevant definitions to encompass more general form of data hiding in arbitrary GPTs (Definition~\ref{dh}). In this context, the effectiveness of discriminating protocols is measured by the minimal probability of error $P_e^{\mathcal{M}}(\rho,\sigma; p)$ in the task of distinguishing the two states $\rho,\sigma$ with a priori probabilities $p,1-p$ respectively, when only operations from the class $\mathcal{M}$ are available. For instance, the archetypical example of a pair of states exhibiting data hiding is given by the normalised projectors onto the symmetric and antisymmetric subspace in $\mathds{C}^{n}\otimes\mathds{C}^{n}$, denoted by $\rho_{\mathcal{S}}$ and $\rho_{A}$, respectively. While $P_e^{\text{ALL}}\left(\rho_{S}, \sigma_{\mathcal{A}}, \frac12\right) = 0$, because the two states have orthogonal support, it can be shown that $P_e^{\text{LOCC}}\left(\rho_{S}, \sigma_{\mathcal{A}}, \frac12\right) = \frac{2}{n+1}$~\cite{dh original 1, dh original 2}.
For these discussions, a more convenient quantity is the \emph{distinguishability norm} associated with $\mathcal{M}$, denoted by $\|\cdot\|_{\mathcal{M}}$ and defined by $P_{e}^{\mathcal{M}}(\rho,\sigma; p)\, =\, \frac12 \left( 1 - \| p\rho-(1-p)\sigma\|_{\mathcal{M}} \right)$, as detailed in Definition~\ref{def d norm} and Lemma~\ref{discr GPT}; it quantifies the advantage of making an observation over pure guessing (i.e. the prior information). It is immediately obvious that higher values of $\| p\rho-(1-p)\sigma\|_{\mathcal{M}}$ correspond to an increased discriminating power of the set $\mathcal{M}$. Thus, the central object of our investigation is a quantity that we name \emph{data hiding ratio}, which depends on the GPT as well as on the restricted set of measurements we consider, and is given by
\begin{equation*}
R(\mathcal{M}) = \max \frac{\| p\rho-(1-p)\sigma\|_{\text{ALL}}}{\| p\rho-(1-p)\sigma\|_{\mathcal{M}}},
\end{equation*}
where $\|\cdot\|_{\text{ALL}}$ denotes the norm associated with the whole set of possible measurements and the maximisation ranges over all pairs of states $\rho,\sigma$ and a priori probabilities $p$ (Definition~\ref{dh} and Proposition~\ref{dh ratio}).
Our first interest is the determination of the value of $R(\mathcal{M})$ for relevant classes of models and physically or operationally significant restricted sets of protocols $\mathcal{M}$. Concretely, we focus on the case when the system under examination is bipartite, and $\mathcal{M}$ is the set of operations that are implementable if various kind of \emph{locality constraints} are imposed (Definition~\ref{locally constr}). We will look at local operations assisted by one-way classical communication ($\text{LOCC}_\rightarrow$) or the generally broader set of measurements with separable effects (called separable measurements and denoted by $\text{SEP}$). Throughout the rest of this section, $\mathcal{M}$ will always denote one of these locally restricted sets of measurements.
Our interest in this kind of questions has been spurred by a result in~\cite{VV dh}, stating that on a finite-dimensional quantum mechanical system $\mathds{C}^{n_A}\otimes\mathds{C}^{n_B}$ the data hiding ratio against LOCC protocols satisfies $\Omega\left(n\right) \leq R_{\text{QM}}(\text{LOCC}) \leq O \left( \sqrt{n_A n_B} \right)$, where $n = \min\{n_A, n_B\}$. In particular, when $n_A=n_B$ one obtains $R_{\text{QM}}(\text{LOCC})=\Theta(n)$. Our first contribution is an intuitive argument, using the quantum teleportation protocol, to show that $R_{\text{QM}}(\text{LOCC})=\Theta(n)$ holds in fact for all $n_A, n_B$ (Theorem~\ref{dh QM}). Note that the local real dimensions of the cones of states (i.e. the cones of positive semidefinite matrices) satisfy $d_A=n_A^2$ and $d_B=n_B^2$, so that one could write $R_{\text{QM}}(\text{LOCC})= \min\{ \sqrt{d_A}, \sqrt{d_B} \}$.
Along the same line of thought, we compute the data hiding ratios against locally constrained sets of measurements for another significant GPT called spherical model (Example~\ref{ex sph}). As the name suggests, the state space of such a model is a Euclidean ball. We prove that for fixed local real dimensions, the data hiding ratio displayed by the spherical model is quadratically larger than the quantum mechanical one, i.e. $R_{\text{Sph}}(\mathcal{M}) = \Theta \left( \min\{ d_A, d_B \} \right)$.
Our second line of investigation aims at establishing an \emph{optimal, universal upper bound} for $R(\mathcal{M})$ that depends only on the local real dimensions $d_A, d_B$, and where $\mathcal{M}$ is a locally constrained set of measurements, as usual. The answer to this question is the content of the main result of the present paper, Theorem~\ref{thm univ}, which states that $R(\mathcal{M})\leq \min\{d_A, d_B\}$ holds for all bipartite GPTs of local dimensions $d_A, d_B$. Since we have seen that such a scaling characterises spherical models, we deduce that $\min\{d_A, d_B\}$ is the optimal universal upper bound on the data hiding ratio against locally constrained sets of measurements. This answers our fundamental question on the ultimate effectiveness of data hiding when the local systems have bounded size.
The rest of the paper is organised as follows. In Section~\ref{sec GPT} we set the notation and provide a brief introduction to the main feature of the GPT framework. Section~\ref{sec dh GPTs} is devoted to giving precise definitions of what we mean by data hiding in the GPT setting. Then, throughout Section~\ref{sec ex} we show this machinery at work by examining some examples of significant physical theories, including quantum mechanics. In Section~\ref{sec univ} we present the main result of the paper, i.e. the determination of the optimal universal upper bound on data hiding ratios against locally constrained sets of measurements (Theorem~\ref{thm univ}). Finally, Section~\ref{sec special} presents a body of techniques to compute data hiding ratios in specific classes of GPT models satisfying some further assumptions.
\section{General probabilistic theories} \label{sec GPT}
That branch of the study of cones, which turns into the study of ordered Banach spaces, gives rise to an axiomatic framework to describe probabilistic physical models. Since it has been initially proposed as a way to generalise quantum probability rules~\cite{MACKEY, Ellis dual base, Davies 1970}, this framework, now widely known under the name of general probabilistic theories (GPTs), has received growing attention and many important refinements were put forward~\cite{Edwards 1, Edwards 2, LUDWIG, tensor rule 1, tensor rule 2}. The main advantage of GPTs is that within their realm a rich variety of physical theories can be encompassed. Prominent examples include classical probability theory and quantum mechanics as well as more abstract objects such as Popescu-Rohrlich boxes~\cite{PR boxes, Barrett original}. Nowadays, many intriguing features of quantum information processing have been investigated in the more general context provided by GPTs~\cite{Barnum no-broad, telep in GPT, nonloc polygon, Barnum steering, Pfister no disturbance, ent therm GPT}.
For the sake of completeness, throughout this section we provide an overview of the mathematical machinery of GPTs. For further details we refer the reader to more complete reviews that can be found in the literature~\cite{telep in GPT, Pfister Master, Barnum review, lamiatesi}. In conformity with these authors, we adopt the so-called {\it abstract state space} formalism.
\subsection{Single systems}
We shall describe a physical system by means of a {\it state space} $\Omega$, and measurement outcomes (called {\it effects}) by functions $e:\Omega\rightarrow[0,1]$, were $e(\omega)$ stands for the probability of getting the outcome $e$ on the state $\omega\in\Omega$. If we represent the probabilistic process of preparing the state $\omega_{1}\in\Omega$ with probability $p$ and the state $\omega_{2}\in\Omega$ with probability $1-p$ as $p\,\omega_{1}+(1-p)\omega_{2}$, and we require this to be a physically allowable state, we obtain that the state space must be convex. The compatibility with the rule determining how the probabilities of measurement outcomes are formed forces $e$ to be a convex-linear functional. A special effect is the {\it unit effect} $u$, defined by $u(\omega)\equiv 1$ for all $\omega\in\Omega$. Following a common convention, the set of all effects will be denoted with $[0,u]$. We shall see in a moment that this is really an interval with respect to a natural ordering on the dual space. In this context, a measurement (usually called positive operator-valued measurement or POVM in quantum mechanics) is a {\it finite} collection of effects $(e_{i})_{i\in I}\subset [0,u]$ such that $\sum_{i\in I} e_{i}=u$. We shall denote the set of all measurements in a GPT as $\mathbf{M}$.
A more convenient picture can be obtained by exploiting the natural embedding of $\Omega$ in the dual of the vector space $\mathcal{A}_{\Omega}$ of affine functionals $f:\Omega\rightarrow \mathds{R}$. Namely, to a state $\omega\in\Omega$ we can associate $\hat{\omega}\in \mathcal{A}_{\Omega}^{*}$ whose action is given by $\hat{\omega}(f)=f(\omega)$, for all $f\in \mathcal{A}_{\Omega}$. A common technical assumption we will make everywhere in the paper is that $\mathcal{A}_{\Omega}$ is {\it finite-dimensional}. In this way, $\Omega$ is endowed with a canonical topology we do not need to specify. As customary, we will consider only compact state spaces. In what follows, we will adopt the shorthand $\mathcal{A}_{\Omega}^{*}=V$ and hence $\mathcal{A}_{\Omega}=V^{*}$. We stress that while the assumptions of finite dimension and compactness can be regarded as purely technical, the convexity of the state space plays a decisive role in the physical interpretation.
We see that the vector space $V^{*}$ can be equipped with a natural ordering: if $f,g\in V^{*}$, we say that $f\leq g$ whenever $f(\omega)\leq g(\omega)$ for all $\omega\in \Omega$. Observe that the set $[0,u]$ of all effects coincides with the interval $\{f\in V^*: 0\leq f\leq u\}$, as our notation suggested. The ordering on $V^*$, in turn, induces a dual ordering on the space $V$: if $x,y\in V$, we declare $x\leq y$ whenever $f(x)\leq f(y)$ for all positive functionals $V^{*}\ni f\geq 0$. It is easy to see that the positive cone $C=V_{+}\subset V$ coincides with the set of non-negative multiples of states. Observe that $V,V^{*}$ equipped with these translationally invariant, positively homogeneous order relations become ordered vector spaces. The cones of positive elements $C,C^{*}$ are dual to each other and both closed, convex, salient and generating. Furthermore, we notice that the unit functional $u$ belongs to the interior of the cone $C^{*}$, which we signify by calling it it {\it strictly positive}. We remind the reader that given a cone $K\subset V$, its dual $K^{*}$ is defined as $K^{*}=\{f\in V^{*}:\ f(a)\geq 0\ \forall\ a\in K\}\subset V^{*}$. Moreover, $K$ is said to be salient if $K\cap (-K)=\{0\}$, and generating if $\text{span}(K)=V$. These two latter notions are elementarily seen to be dual to each other, meaning that a closed cone is salient iff the dual is generating, and vice versa. From the above discussion we abstract the following definition.
\vspace{2ex}
\begin{Def}[General probabilistic theories~\cite{Ellis dual base, Davies 1970}]
A \emph{general probabilistic theory} (GPT) is a triple $(V, C, u)$ consisting of a real, finite-dimensional vector space $V$ ordered by a closed, convex, salient and generating cone $C$ and of a strictly positive element $u\in \text{\emph{int}}(C^{*})$. We call $d=\dim V$ the dimension of the GPT.
\end{Def}
\vspace{0ex}
\begin{note}
All vector spaces in the present paper are from now on understood to be real and finite-dimensional.
\end{note}
Let us note in passing that in~\cite{Davies 1970} GPTs are called \emph{state spaces}. However, here we reserve this latter term for the set $\Omega$ of normalised states. As it turns out, in every GPT there is a natural notion of norm induced on the vector space by the unit in the dual.
\vspace{2ex}
\begin{Def}[Base norm] \label{def base}
Let $(V,C,u)$ be a GPT. The \emph{base norm} $\|\cdot\|$ on $V$ is defined as
\begin{equation}
\|x\|\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max_{f\in [-u,u]} f(x)\, =\, \max_{e\in [0,u]} \{|e(x)| + |(u-e)(x)|\}\, ,
\label{base norm}
\end{equation}
where $[-u,u]=\{f\in V^{*}:\ -u\leq f\leq u\}$. In other words, the dual base norm has $[-u,u]$ as the unit ball.
\end{Def}
\begin{note}
Whenever there is room for ambiguity, we will denote by a subscript the GPT to which the base norm refers.
\end{note}
Observe that if $a\geq 0$ then $\|a\|=u(a)$, i.e. its base norm coincides with the expectation value of the unit effect. For general $x\in V$, we will find useful a dual formula for the base norm as given in the following lemma.
\vspace{2ex}
\begin{lemma}\emph{\cite{Krein dual base, Ellis dual base}} \label{dual base}
Let $(V,C,u)$ be a GPT. Then its base norm $\|\cdot\|$ as defined in~\eqref{base norm} is also given by
\begin{equation}
\|x\|\, =\, \min\left\{ u(x_{+}) + u(x_{-})\ :\ x_{\pm}\geq 0,\ x=x_{+}-x_{-} \right\} .
\label{dual base eq}
\end{equation}
\end{lemma}
\begin{proof}
For the original proof we refer the reader to~\cite{Krein dual base, Ellis dual base}. Alternatively, in finite dimension one can observe that~\eqref{dual base eq} is exactly the dual of the convex program in~\eqref{base norm} (first equality). Since $f=0$ is a strictly feasible point of the primal problem, Slater's condition~\cite{B&V} ensures that the two values coincide, concluding the proof.
\end{proof}
\subsection{Bipartite systems}
A physical theory modelled by a GPT should encompass a way to build a bipartite system $AB$ out of two single systems $A$ and $B$. Throughout this section, we will review the basic physical requirements to be imposed on such a construction. First, performing a local measurement $(e_{A},e_{B})$ should be allowed for all effects $e_{A}\in[0,u_{A}]$ and $e_{B}\in [0,u_{B}]$, no matter what the joint state of the system is. Secondly, preparing a separate state $(\omega_{A},\omega_{B})$ should be possible for all local states $\omega_{A}\in \Omega_{A}$ and $\omega_{B}\in \Omega_{B}$. Finally, it is reasonable to postulate the {\it local tomography principle}: bipartite states are fully determined by the statistics resulting from local measurements. Under these assumptions, it can be shown~\cite{tensor rule 1, tensor rule 2} that the joint vector space $V_{AB}$ can be identified with the tensor product of the two local spaces, i.e.
\begin{equation}
V_{AB}\, \simeq\, V_{A}\otimes V_{B}\, .
\label{tensor spaces}
\end{equation}
Similarly, the joint unit effect can be taken as the product of the two local unit effects, $u_{AB}=u_{A}\otimes u_{B}$. Within this framework, the non-signalling principle is automatically implemented by the mathematically formalism: if $AB$ is in a state $\omega_{AB}$, whatever operation is carried out on the local party $B$, the reduced state of the system $A$ will be given by $u_B(\omega_{AB})$, implicitly defined by the equations $f\left( u_B(\omega_{AB})\right)\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = (f\otimes u_B)(\omega_{AB})$ for all $f\in V_A^*$.
Remarkably, the joint cone of positive elements $C_{AB}$ is not fully determined by the above axioms. Instead, it can be shown that
\begin{equation}
C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}\, \subseteq\, C_{AB}\, \subseteq\, C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! C_{B}\, .
\label{CAB bound}
\end{equation}
Here, the `lower bound' $C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}$, called {\it minimal tensor product}, is given by
\begin{equation}
C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \, \text{conv}\left( C_{A}\otimes C_{B} \right) ,
\label{minimal}
\end{equation}
where $C_{A}\otimes C_{B}=\left\{ a\otimes b:\ a\in C_{A},\ b\in C_{B} \right\}$, and $\text{conv}$ denotes the convex hull. States in the minimal tensor product are conventionally called {\it separable}. We should mention here that the above definition of minimal tensor product yields automatically a closed convex cone. To see that this, it suffices to show that the corresponding state space $\text{conv}(\Omega_{A}\otimes \Omega_{B})$ is closed. However, in the case considered in this work this is particularly easy, since the fact that $\Omega_{A}$ and $\Omega_{B}$ are compact sets immediately implies that the set $\text{conv}(\Omega_{A}\otimes \Omega_{B})$ is compact too~\cite[Theorem 3.25(b)]{RUDIN}. We refer to~\cite[Exercise 4.14]{AUBRUN} for a more general statement where compactness is not assumed.
The complementary `upper bound' in~\eqref{CAB bound}, called {\it maximal tensor product}, can be alternatively defined as
\begin{equation}
C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! C_{B}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \{ Z\in V_{A}\otimes V_{B}:\ (f\otimes g)(Z)\geq 0\ \forall\ f\in C_{A}^{*},\, g\in C_{B}^{*} \}\, =\, \Big( C_{A}^{*}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}^{*}\Big)^{*} .
\label{maximal}
\end{equation}
Observe that also~\eqref{maximal} identifies a closed convex cone. The original definitions as well as the (elementary) proof of the above equivalence were given by Namioka and Phelps in~\cite{NP}. For a more comprehensive introduction to the theory of tensor product of cones, we refer the reader to~\cite{Mulansky}.
In the following, we will see how the two-sided bound~\eqref{CAB bound} applies in many concrete cases. With a slight abuse of notation, given two GPTs $A=(V_{A},C_{A},u_{A}),\, B=(V_{B},C_{B},u_{B})$ we will refer to the composites
\begin{equation}
A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! B\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \Big( V_{A}\otimes V_{B},\, C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B},\, u_{A}\otimes u_{B}\Big) ,\qquad A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! B\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \Big( V_{A}\otimes V_{B},\, C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! C_{B},\, u_{A}\otimes u_{B}\Big)
\label{min max theories}
\end{equation}
as the minimal and maximal tensor product of the GPTs $A$ and $B$, respectively.
Now, let us devote some time to examine a particularly simple instance of a GPT, namely classical probability theory with a finite alphabet. Although it is trivial in some respects, its understanding is crucial in appreciating more complicated examples such as quantum theory (Example~\ref{ex QM}) or spherical models (Example~\ref{ex sph}), not to mention even more exotic GPTs to be treated later on (Example~\ref{cubic}).
\vspace{2ex}
\begin{ex}[Classical probability theory] \label{ex class}
The state space of classical probability theory is the set of probability distributions over a finite alphabet. The corresponding GPT can be defined as the triple
\begin{equation}
\text{Cl}_{d}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \left(\mathds{R}^{d},\, \mathds{R}^{d}_{+},\, u\right)\, ,
\label{classical}
\end{equation}
where $\mathds{R}^{d}_{+}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\{ x\in \mathds{R}^{d}:\, x_{i}\geq 0\ \forall\, i=1,\ldots, d\}$ is the positive orthant, and the unit effect acts as $u(y)=\sum_{i=1}^{d}y_{i}$ for all $y\in \mathds{R}^{d}$. The base norm associated with classical probability theory is easily seen to be the $l_{1}$-norm $|x|_{1}=\sum_{i=1}^{d}|x_{i}|$. Since classical cones are simplicial, it can be seen that the lower and upper bound in~\eqref{CAB bound} coincide and the composition rule is trivial: if $A=\text{Cl}_{d_{A}}$ or $B=\text{Cl}_{d_{B}}$, then necessarily $C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B=C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! C_B$.
\end{ex}
\section{Data hiding in GPTs} \label{sec dh GPTs}
\subsection{State discriminiation}
The binary distinguishability problem consists of choosing secretly one of the two states $\rho,\sigma\in \Omega$ (where $\Omega$ is a generic state space) with known a priori probabilities $p, 1-p$ and handing it over to an agent, whose task is to discriminate between the two alternatives. Naturally, the larger the set of measurements the agent has at his disposal, the lower the associated probability of error will be (for fixed states and a priori probabilities). In general, it will make sense to consider measurements that are at least {\it informationally complete}, meaning that from their complete statistics the full state can be reconstructed unambiguously. A formal definition is below.
\vspace{2ex}
\begin{Def} \label{info complete}
Let $(V,C,u)$ be a GPT. Then a measurement $(e_{i})_{i\in I}\in \mathbf{M}$ (i.e. finite family of effects $(e_{i})_{i\in I}\subset[0,u]$ such that $\sum_{i\in I} e_{i}=u$) is said to be \emph{informationally complete} if $\text{\emph{span}}\{e_{i}:\, i\in I\}=V^{*}$. A set $\{\mu_t\}_{t\in T}$ made of measurements $\mu_t=\big(e_i^{(t)}\big)_{i\in I_t}$ is deemed informationally complete if $\text{\emph{span}}\big\{e_{i}^{(t)}:\, t\in T, i\in I_t\big\}=V^{*}$.
\end{Def}
\vspace{2ex}
If $\mathcal{M}\subseteq \mathbf{M}$ is a set of measurements in an arbitrary GPT, we can define an associated norm by translating to GPTs the analogous definition in~\cite{VV dh}.
\vspace{2ex}
\begin{Def} \label{def d norm}
Let $\mathcal{M}\subseteq\mathbf{M}$ be an informationally complete set of measurements in a GPT $(V,C,u)$. The associated \emph{distinguishability norm} $\|\cdot\|_{\mathcal{M}}$ is a norm on $V$ given by
\begin{equation}
\|x\|_{\mathcal{M}} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \sup_{(e_{i})_{i\in I}\in \mathcal{M}} \sum_{i} |e_{i}(x)|
\label{d norm}
\end{equation}
for all $x\in V$.
\end{Def}
\vspace{2ex}
As is easy to see, the function defined in~\eqref{d norm} is truly a norm on $V$ thanks to the informational completeness of $\mathcal{M}$. Among its elementary properties, we note the following: (i) the identity $\|a\|_{\mathcal{M}}=u(a)$, valid on positive states $a\geq 0$ (independently of $\mathcal{M}$); (ii) the general bound $\|x\|_{\mathcal{M}} \geq |u(x)|$, an elementary consequence of the definition~\eqref{d norm}; (iii) the monotonicity of $\|\cdot\|_\mathcal{M}$ in $\mathcal{M}$, with the partial order defined by the inclusion; and (iv) the fact that $\|\cdot\|_{\mathbf{M}}$ coincides with the base norm, as seen from the second equality in~\eqref{base norm}. The above formal definition becomes relevant by virtue of its link to the operational task of state discrimination as given by the following lemma, totally analogous to~\cite[Theorem 5]{VV dh}.
\vspace{2ex}
\begin{lemma} \label{discr GPT}
Let $(V,C,u)$ be a GPT with state space $\Omega$, and consider an informationally complete set of measurements $\mathcal{M}\subseteq \mathbf{M}$. Then the lowest probability of error for discriminating between two states $\rho,\sigma\in \Omega$ with a priori probabilities $p,1-p$, respectively, is given by
\begin{equation}
P_{e}^{\mathcal{M}}(\rho,\sigma; p)\, =\, \frac12 \left( 1 - \| p\rho-(1-p)\sigma\|_{\mathcal{M}} \right) , \label{pr error}
\end{equation}
where $\|\cdot\|_{\mathcal{M}}$ is the distinguishability norm given by~\eqref{d norm}.
\end{lemma}
\begin{proof}
This goes in complete analogy with the corresponding argument for quantum mechanics~\cite{Helstrom original, VV dh, state discr GPT}, but we explain it here for the sake of completeness. Without loss of generality, we can assume that the protocol consists of measuring the state with a measurement $(e_{i})_{i\in I}$ and performing a (possibly probabilistic) post-processing of the classical outcome $i$. Assume that the outcome $i$ yields $\rho$ or $\sigma$ as final answers with probabilities $q_{i}$ and $1-q_{i}$, respectively. Then the probability of error is given by
\begin{equation*}
P_{e}\, =\, p \sum_{i\in I} e_{i}(\rho) (1-q_{i}) + (1-p) \sum_{i} e_{i}(\sigma) q_{i}\, =\, p - \sum_{i} q_{i} e_{i}\left(p\rho-(1-p)\sigma\right) ,
\end{equation*}
where we employed the normalisation relation $\sum_{i\in I} e_{i}=u$. Minimising over all probabilities $q_{i}$ one obtains
\begin{equation*}
P_{e}\, =\, \frac12 \left( 1 - \sum_{i\in I} |e_{i}\left(p\rho-(1-p)\sigma\right)| \right) ,
\end{equation*}
and finally~\eqref{pr error} after a minimisation over all measurements $(e_{i})_{i\in I}\in \mathcal{M}$.
\end{proof}
\vspace{2ex}
The analogy with quantum mechanics goes much beyond this. In fact, all the results of~\cite[Section 2]{VV dh} (with the exception of Proposition 8 there) carry over to GPTs. In translating the statements one has just to remember that the quantum mechanical trace norm $\|\cdot\|_{1}$ becomes the base norm in the GPT framework, and that similarly the identity becomes the unit effect. For details on the interpretation of quantum mechanics as a GPT, we refer the reader to Example~\ref{ex QM}. Here we limit ourselves to provide a formulation of~\cite[Theorem 4]{VV dh} for GPTs. In what follows, for a set of measurements $\mathcal{M}\subseteq\mathbf{M}$ we will denote by $\langle \mathcal{M} \rangle$ the set generated by $\mathcal{M}$ via {\it coarse graining}, i.e. by a posteriori declaring some of the outcomes of a measurement in $\mathcal{M}$ as the same. In formula,
\begin{equation}
\langle \mathcal{M} \rangle\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \bigg\{ (e_{j})_{j\in J}:\ \exists\ I\ \text{finite},\ \{I_{j}\}_{j\in J}\ \text{partition of}\ I,\ (e'_{i})_{i\in I}\in\mathcal{M}:\ e_{j}=\sum_{i\in I_{j}} e'_{i}\ \, \forall\, j\in J \bigg\} .
\label{coarse}
\end{equation}
\vspace{0ex}
\begin{lemma} \label{unit ball dual d norm}
The unit ball of the dual to the distinguishability norm~\eqref{d norm} is given by
\begin{equation}
B_{\|\cdot\|_{\mathcal{M},*}}\, =\, \overline{\text{\emph{conv}} \big\{ 2e-u:\ ( e, u-e )\in \langle \mathcal{M}\rangle \big\}}\, .
\label{unit ball dual d norm eq}
\end{equation}
Equivalently, $\|\cdot\|_{\mathcal{M}}$ can be computed as
\begin{equation}
\|x\|_{\mathcal{M}}\, =\, \sup\left\{ f(x):\ \left(\frac{u+f}{2},\, \frac{u-f}{2} \right) \in \langle \mathcal{M}\rangle \right\}\, .
\label{d norm altern}
\end{equation}
Consequently, there is a one-to-one correspondence between distinguishability norms~\eqref{d norm} and closed symmetric convex bodies $K$ such that $\pm u\in K\subseteq [-u,u]$.
\end{lemma}
\begin{proof}
See~\cite{VV dh}.
\end{proof}
\vspace{1ex}
\begin{rem}
From Lemma~\ref{unit ball dual d norm} it follows in particular that $\|\cdot\|_{\mathcal{M}}$ depends only on the set of measurements generated by $\mathcal{M}$ via coarse graining, and in fact only on the right-hand side of~\eqref{unit ball dual d norm eq}.
\end{rem}
At this point, the reader should be familiar enough with the body of techniques we have discussed so far, to be able to work out the translations of the other results in~\cite[Section 2]{VV dh} by herself. As for us, instead of reporting them, we believe it more appropriate to tell the antecedent story of data hiding in quantum mechanics, and to show in detail how the definitions can be generalised, as to encompass arbitrary GPTs.
\subsection{Data hiding and statement of the problem}
Throughout this section, a generalisation of the concept of data hiding against LOCC measurements in quantum mechanics as originally conceived in~\cite{dh original 1, dh original 2} is discussed. On the one hand, we will extend this notion to an arbitrary GPT, and on the other hand we will allow for data hiding against an arbitrary set of measurements, without any a priori assumption on its nature. We give the following definition.
\vspace{2ex}
\begin{Def} \label{dh}
Let $(V,C,u)$ be a GPT. For an informationally complete set of measurements $\mathcal{M} \subseteq \mathbf{M}$, we say that there is \emph{data hiding against $\mathcal{M}$ with efficiency $R\geq 1$} if there are two normalised states $\rho,\sigma\in\Omega$ and a real number $p\in [0,1]$ such that the probability of error defined in~\eqref{pr error} satisfies
\begin{equation}
P_{e}^{\mathbf{M}}(\rho,\sigma; p) = 0\, ,\qquad P_{e}^{\mathcal{M}}(\rho,\sigma; p) = \frac12 \left( 1-\frac1R\right) .
\label{dh eq}
\end{equation}
The highest data hiding efficiency against $\mathcal{M}$ is called \emph{data hiding ratio against $\mathcal{M}$} and will be denoted by $R(\mathcal{M})$.
\end{Def}
\vspace{1ex}
\begin{rem}
A priori, the only meaningful way to define $R(\mathcal{M})$ is as the supremum of all achievable data hiding efficiencies. However, we will see in a moment that this supremum is actually a maximum (Proposition~\ref{dh ratio}).
\end{rem}
In the above Definition~\ref{dh}, we chose not to restrict ourselves to equiprobable pairs of states. There are several reasons that justify this choice. On the one hand, it can be shown that any pair that exhibits data hiding with high efficiency is approximately equiprobable, the approximation becoming better and better for higher efficiencies. On the other hand, even considering only the case of exact equality $p=1/2$ from the start, the obtained data hiding ratio does not differ by more than a factor of two (additive constants apart) from that we have defined here. We devote Appendix~\ref{app equi} to exploring the consequences of restricting the definition of data hiding to the equiprobable case.
We now go back to the investigation of data hiding in the sense of Definition~\ref{dh}.
An elementary yet fruitful observation is that if a set of measurements exhibits data hiding with high efficiency, then the associated distinguishability norm -- given by~\eqref{d norm} -- has to be very different from the base norm. This is again a straightforward consequence of Lemma~\ref{discr GPT}. The following result establishes the converse, i.e. that if the two norms are very different on some vectors then there is highly efficient data hiding.
\vspace{2ex}
\begin{prop} \label{dh ratio}
For an informationally complete set of measurements $\mathcal{M}\subseteq\mathbf{M}$ in an arbitrary GPT, the data hiding ratio $R(\mathcal{M})$ is given by
\begin{equation}
R(\mathcal{M})\, =\, \max_{0\neq x\in V} \frac{\|x\|}{\|x\|_{\mathcal{M}}}\, .
\label{dh ratio eq}
\end{equation}
Adopting the terminology of~\cite{VV dh}, we can rephrase~\eqref{dh ratio eq} by saying that $R(\mathcal{M})$ is the constant of domination of $\|\cdot\|_{\mathcal{M}}$ on $\|\cdot\|$, i.e. the smallest $k\in \mathds{R}$ such that $\|\cdot\|\leq k\|\cdot\|_{\mathcal{M}}$.
\end{prop}
\begin{proof}
Let the GPT $(V,C,u)$ have state space $\Omega$. Then, from Definition~\ref{dh} and from~\eqref{pr error} we see that
\begin{equation*}
R(\mathcal{M})\, =\, \sup \left\{ \|p\rho-(1-p)\sigma\|_{\mathcal{M}}^{-1}:\ \rho,\sigma\in\Omega,\ 0\leq p\leq 1,\ \|p\rho-(1-p)\sigma\|=1 \right\} .
\end{equation*}
The crucial observation is that the set $K$ of vectors $x\in V$ that can be represented as $x=p\rho-(1-p)\sigma$ for appropriate $\rho,\sigma\in\Omega$ and $p\in[0,1]$ coincides with the unit ball of the base norm, i.e. $K=B_{\|\cdot\|}$. In order to prove this, start by observing that $\|p\rho-(1-p)\sigma\|\leq p\|\rho\|+(1-p)\|\sigma\|=p+(1-p)=1$, so $\|x\|\leq 1$ is a necessary condition for $x$ to belong to $K$. To see that it is also sufficient, notice that $K$ is convex by construction and $0\in K$, and therefore it suffices to consider the case $\|x\|=1$. With this hypothesis, Lemma~\ref{dual base} yields a decomposition $x=x_{+}-x_{-}$ such that $x_{\pm}\geq 0$ and $u(x_{+})+u(x_{-})=1$, from which it follows that $x_{+}=p\rho$ and $x_{-}=(1-p)\sigma$ for $p=u(x_{+})\in [0,1]$ and $\rho,\sigma\in \Omega$ normalised states. Thanks to this observation, we rewrite the above representation of $R(\mathcal{M})$ as
\begin{equation*}
R(\mathcal{M})\, =\, \sup\left\{\|x\|_{\mathcal{M}}^{-1}:\, \|x\|=1\right\}\, =\, \max_{x\neq 0}\, \frac{\|x\|}{\|x\|_{\mathcal{M}}}\, .
\end{equation*}
For the last step, we used the positive homogeneity of the norms, and we converted the supremum over the (compact) unit ball of the base norm into a maximum.
\end{proof}
\vspace{2ex}
Some elementary properties of the data hiding ratio are as follows.
\vspace{2ex}
\begin{lemma} \label{elem dh}
Let $\mathcal{M}$ be an informationally complete set of measurements in an arbitrary GPT. Then:
\begin{itemize}
\item the data hiding ratio given in Definition~\ref{dh} satisfies $R(\mathcal{M})\geq 1$, with equality iff the set on the left hand side of~\eqref{unit ball dual d norm eq} coincides with the full interval $[-u,u]$;
\item $R(\mathcal{M})$ is monotonically non-increasing as a function of $\mathcal{M}$, where the partial order on sets of measurements is the one given by inclusion.
\end{itemize}
\end{lemma}
\vspace{2ex}
As expected, not much can be said about data hiding ratios for a single system and when the sets of measurements are completely arbitrary. In fact, it is easy to show that already in a classical GPT (Example~\ref{ex class}) such as $(\mathds{R}^2, \mathds{R}^2_+, u)$ with $u=(1,1)$ such that $u(x,y)\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = x+y$, for the particular case when $\mathcal{M}$ is made of just one measurement $\big((\varepsilon,0),\, (1-\varepsilon, 1)\big)$ we have $R(\mathcal{M})=(2-\varepsilon)/\varepsilon$, so that the data hiding ratio can even be unbounded in a system of fixed dimension.
The situation changes dramatically when we consider bipartite systems, as originally done in the context of data hiding. Let us discuss first bipartite quantum systems. Here, some restricted sets of measurements come into play quite naturally as deriving from operational constraints. Examples of such sets include local operations (LO), local operations assisted by shared randomness (LOSR) or one-way classical communication ($\text{LOCC}_{\rightarrow}$), and general LOCC protocols when both communication directions are allowed. It is also convenient to introduce mathematical relaxations of these classes. For instance, consider the set of measurements $(e_{i})_{i\in I}$ such that $E_{i}$ is a separable positive operators for all $i\in I$ (i.e. it belongs to the quantum mechanical equivalent of~\eqref{minimal} as given by~\eqref{separable}). We call these measurements {\it separable}, and denote them collectively as SEP. It is easy to see that $\text{LOCC}\subseteq\text{SEP}$, and less trivially the inclusion can be shown to be strict.
Until now we have discussed only quantum theory. Perhaps surprisingly, it turns out that all these restricted classes of measurements can be defined in any bipartite arbitrary GPT $AB=(V_A\otimes V_B,\, C_{AB},\, u_A\otimes u_B)$ constructed out of two local theories $A$ and $B$ in such a way that the constraints~\eqref{CAB bound} are met. Before we provide general definitions below, let us briefly discuss how to add {\it dynamical prescriptions} to our structure. The purpose of these rules is to specify how states transform after a measurement, extending the picture we have been describing so far, mostly orientated towards the outcomes and their probabilities. In practice, we will not make use of these prescriptions, and in fact our results are totally independent of any assumption concerning them beyond the mere consistency with the operational interpretation of the theory. However, this apparatus is needed to define a generic LOCC protocol, which requires multiple, interactive rounds of operations on the same systems.
Following~\cite{Davies 1970}, we can define \emph{instruments} on one of the two system, say $A=(V_A, C_A, u_A)$, as collections $(\phi_i)_{i\in I}$ of linear maps $\phi_i: V_A\rightarrow V_A$ that are \emph{completely positive}, i.e. satisfy $\left((\phi_i)_A \otimes I_B\right)(C_{AB})\subseteq C_{AB}$, and sum up to a normalisation-preserving map, i.e. $\sum_{i\in I} \phi_{i}^*(u_A) = u_A$, with $\phi_i^*: V_A^*\rightarrow V_A^*$ being the dual (or transpose) of $\phi_i$. A totally analogous definition can be given for instruments on the $B$ system. In the operational interpretation of the theory, an instrument describes a non-destructive measurement, with $\phi_i(\rho)$ representing the unnormalised post-measurement state when the outcome $i$ has been recorded on the initial state $\rho$, and the normalisation coefficient $u\left(\phi_{i}(\rho) \right) = \left(\phi_{i}^*(u)\right)(\rho)$ being the probability that the process yields the outcome $i$ (accordingly, observe that $\left(\phi_i^*(u)\right)_{i\in I}$ is a valid measurement in the GPT sense).
With the concept of instrument at hand, in order to define LOCC protocols we can follow the steps described in~\cite[Section 2.2]{LOCC}. We will not repeat the construction here since it is totally analogous to the quantum mechanical one, once the concept of instrument in GPTs has been clarified.
\vspace{2ex}
\begin{Def} \label{locally constr}
Let $A=(V_{A}, C_{A}, u_{A})$ and $B=(V_{B}, C_{B}, u_{B})$ be two GPTs, and let the composite system \mbox{$AB=(V_A\otimes V_B,\, C_{AB},\, u_A\otimes u_B)$} satisfy~\eqref{CAB bound}. Then local operations (LO), local operations assisted by one-way classical communication ($\text{LOCC}_{\rightarrow}$) or two-way classical communication ($\text{LOCC}$), and separable measurements ($\text{SEP}$) are subsets of the set $\mathbf{M}_{AB}$ of all measurements on $AB$ given by:
\begin{align}
\text{\emph{LO}}\, &\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\langle \left\{ (e_{i}\otimes f_{j})_{(i,j)\in I\times J}:\ (e_{i})_{i\in I}\in\mathbf{M}_{A},\ (f_{j})_{j\in J}\in\mathbf{M}_{B} \right\}\right\rangle , \label{LO} \\[1ex]
\text{\emph{LOCC}}_{\rightarrow}\, &\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\langle\left\{ (e_{i}\otimes f_{j}^{(i)})_{(i,j)\in I\times J}:\ (e_{i})_{i\in I}\in\mathbf{M}_{A},\ (f_{j}^{(i)})_{j\in J}\in\mathbf{M}_{B}\ \forall\ i\in I \right\}\right\rangle , \label{1-way LOCC} \\[1ex]
\text{\emph{LOCC}}\, &\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\{ \left( \Phi_i^*(u_A\otimes u_B) \right)_{i\in I}:\ \text{$\left( \Phi_i \right)_{i\in I}$ LOCC instrument on $AB$} \right\} \\[1ex]
\text{\emph{SEP}}\, &\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\{ (E_i)_{i\in I} \in \mathbf{M}_{AB}:\ E_i\in C_A^* \ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! C_B^*\ \forall\ i \right\} . \label{SEP}
\end{align}
Here, $\langle\cdot\rangle$ denotes coarse graining as defined by~\eqref{coarse}. The above sets will be collectively called \emph{locally constrained sets of measurements}.
\end{Def}
\vspace{2ex}
It is easy to verify that
\begin{equation}
\text{LO}\, \subseteq\, \text{LOCC}_{\rightarrow}\, \subseteq\, \text{LOCC}\, \subseteq\, \text{SEP}\, .
\label{chain M}
\end{equation}
The last inclusion is slightly less trivial than the others, but its proof follows closely the quantum mechanical one. Namely, referring for details and nomenclature to~\cite{LOCC}, one can observe that: (i) one-way local instruments are separable, in the sense that each component is a positive sum of tensor products of completely positive maps; (ii) coarse-graining preserves separability; (iii) an instrument that is LOCC-linked to a separable one is again separable; (iv) separability is preserved under limits; and finally (v) if $(\Phi_i)_{i\in I}$ (acting on $AB$) is separable as an instrument, $\left( \Phi_i(u_A\otimes u_B) \right)_{i\in I}$ is separable as a measurement.
\vspace{1ex}
\begin{rem}
From an operational point of view, there is at least another notable set of locally constrained measurements that is worth mentioning. This is the set of local operations assisted by shared randomness, formally defined as
\begin{equation}
\text{LOSR}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\langle\left\{ \big(p_{i}\, e^{(i)}_{j}\otimes f^{(i)}_{k}\big)_{(i,j,k)\in I\times J\times K}:\ I\ \text{finite},\ p\in\mathcal{P}(I),\ \big(e^{(i)}_{j}\big)_{j\in J}\in\mathbf{M}_{A},\ \big(f_{k}^{(i)}\big)_{k\in K}\in\mathbf{M}_{B}\ \forall\ i\in I \right\}\right\rangle , \label{LOSR}
\end{equation}
where $\mathcal{P}(I)$ stands for the the set of probability distributions on a finite alphabet $I$. It is very easy to see that $\text{LOSR}$ lies between $\text{LO}$ and $\text{LOCC}_\rightarrow$, i.e.
\begin{equation}
\text{LO}\, \subseteq\, \text{LOSR}\, \subseteq\, \text{LOCC}_{\rightarrow}\, .
\end{equation}
The reason why we did not include this additional set in the above definition is that elements of $\text{LOSR}$ are convex combinations of local measurements, hence it is easy to check that
\begin{equation}
\|\cdot\|_{\text{LOSR}}\, =\, \|\cdot\|_{\text{LO}}\, .
\label{LOSR = LO norm}
\end{equation}
In other words, once they are restricted to local measurements, providing the parties with additional shared randomness does not enhance their distinguishability power.
\end{rem}
\vspace{1ex}
\begin{rem}
Let us stress here that $\text{LOCC}$ is the only locally constrained set of measurements that depends explicitly on the choice of the positive cone $C_{AB}$ of the bipartite system. In fact, it is easy to realise that only the structure of the local GPTs matters in~\eqref{LO},~\eqref{1-way LOCC}, and~\eqref{SEP} (as well as in~\eqref{LOSR}, incidentally). Following the discussion before Definition~\ref{locally constr}, we see that the dependence of $\text{LOCC}$ on $C_{AB}$ is hidden inside the concept of completely positive map, which is in turn necessary to define local instruments.
\end{rem}
An important feature of locally constrained sets of measurements is informational completeness. This is a consequence of the {\it local tomography principle}, valid in an arbitrary bipartite GPT and stating that the statistics of local measurements contain enough information to determine the global state completely. In order to prove this elementary fact, consider a local measurement $(e_{i}\otimes f_{j})_{(i,j)\in I\times J}\in \text{LO}$ such that $\text{span}\{e_{i}\}_{i\in I}=V_{A}^{*}$ and $\text{span}\{f_{j}\}_{j\in J}=V_{B}^{*}$. Then, obviously, $\text{span}\{e_{i}\otimes f_{j}\}_{(i,j)\in I\times J}=V_{A}^{*}\otimes V_{B}^{*}$, which yields the claim.
As we said, our primary interest lies in understanding data hiding against those restricted sets of measurements, whose corresponding constraints have an operational nature. In this context, the above locally constrained sets of measurements are thus excellent candidates, and in fact the rest of the present paper is devoted to the study of the data hiding ratios $R(\mathcal{M})$, with $\mathcal{M}=\text{LO},\, \text{LOCC}_\rightarrow,\, \text{LOCC},\, \text{SEP}$.
As an preliminary observation, note that~\eqref{chain M} and Lemma~\ref{elem dh} imply that
\begin{equation}
R(\text{SEP})\, \leq\, R(\text{LOCC})\, \leq\, R(\text{LOCC}_{\rightarrow})\, \leq\, R(\text{LO})
\label{chain dh}
\end{equation}
for all fixed GPTs.
Among the many questions one could ask at this point, one seems particularly relevant to us. Namely, we can wonder, how the best data hiding ratio {\it scales} with the dimensions of the local GPTs. To be more precise, we give the following definition.
\vspace{2ex}
\begin{Def}[Ultimate data hiding ratio] \label{univ dh ratio}
For a locally constrained set of measurements $\mathcal{M}$, the \emph{ultimate data hiding ratio against $\mathcal{M}$} for fixed local dimensions, denoted by $R_{\mathcal{M}}(d_{A},d_{B})$, is the supremum over all data hiding ratios $R(\mathcal{M})$ achieved by composite GPTs that satisfy~\eqref{CAB bound} and have local dimensions $d_{A},d_{B}$.
\end{Def}
\vspace{2ex}
With this concept in hand, we are ready to formulate the question lying at the heart of our investigation, namely, {\it what is the scaling of the ultimate ratio $R_\mathcal{M}(d_{A}, d_{B})$ with the local dimensions $d_{A},d_{B}$?} Clearly, thanks to the chain of inequalities~\eqref{chain dh}, we find
\begin{equation}
R_{\text{SEP}}(d_{A},d_{B})\, \leq\, R_{\text{LOCC}}(d_A,d_B)\, \leq\, R_{\text{LOCC}_{\rightarrow}}(d_{A},d_{B})\, \leq\, R_{\text{LO}}(d_{A},d_{B})\, .
\label{chain univ dh}
\end{equation}
We stress that the supremum in Definition~\ref{univ dh ratio} has to be taken over {\it all} local GPTs of the given dimensions, and over all {\it composition rules} to join the system (i.e. among all the global cones respecting the bounds~\eqref{CAB bound}). Now, we will show that at least this latter maximisation can be carried out explicitly when $\mathcal{M}$ is a locally constrained set of measurements different from $\text{LOCC}$, the optimal composite being always given by the {\it minimal} tensor product. To see why, notice that the exclusion of $\text{LOCC}$ implies that for a fixed $X\in V_A\otimes V_B$ only the global base norm $\|X\|$ depends on the composition rule we chose. On the contrary, the locally constrained norm $\|\cdot\|_{\mathcal{M}}$ ($\mathcal{M}\neq \text{LOCC}$) will depend on the local structure only, as already observed. Then, maximising the ratio between the former and the latter amounts to maximising the global base norm. In order to do so, a large set of global effects and thus a small set of states are required, and according to~\eqref{CAB bound} the smallest possible positive cone in a bipartite system is given by the minimal tensor product.
The above reasoning is perhaps not obvious from Definition~\ref{dh} alone, because restricting the set of available bipartite states gives less freedom in choosing the data hiding pair. However, this restriction plays no role once Proposition~\ref{dh ratio} is available. This way around the problem is made possible by the fact that any difference of two normalised states can be thought of as a positive multiple of the difference of two separable states, the multiplication coefficient being given by the base norm induced by the minimal tensor product. We summarise this whole discussion stating the following result, whose proof reproduces the informal argument presented above.
\vspace{2ex}
\begin{prop} \label{min is optimal}
Given two local GPTs $A=(V_A,C_A,u_A)$, $B=(V_B,C_B,u_B)$, and a locally constrained set of measurements $\mathcal{M}\neq \text{\emph{LOCC}}$, the maximal data hiding ratio against $\mathcal{M}$ is achieved when the bipartite GPT is constructed according to the minimal tensor product, i.e. $AB=A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! B$.
\end{prop}
\begin{proof}
Let us denote by $\|\cdot\|_{AB}$ the global base norm, whose dependence from the choice of the GPT $AB$ (i.e. of the bipartite cone $C_{AB}$) has been made explicit. It is always implicitly assumed that $C_{AB}$ obeys the two-sided bound in~\eqref{CAB bound}.
As is easy to verify, for all fixed $x\in V_{A}\otimes V_B$ the norm $\|x\|_{AB}$ is monotonically non-increasing as a function of $C_{AB}$, in the sense that
\begin{equation*}
C_{AB}\subseteq \widetilde{C}_{AB}\quad\Longrightarrow\quad \|x\|_{AB}\geq \|x\|_{\widetilde{AB}}\, .
\end{equation*}
To see why, it suffices to go back to the definition of base norm and to observe that: (i) since taking the dual reverses the inclusions, if $C_{AB}\subseteq \widetilde{C}_{AB}$ then $\widetilde{C}_{AB}^*\subseteq C_{AB}^*$; (ii) from this it follows that $[0,u_{AB}]_{\widetilde{C}_{AB}^*}\subseteq [0,u_{AB}]_{C^*_{AB}}$, where $[a,b]_{K} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \{c:\ b-c, c-a\in K \}$ denotes the interval according to the ordering determined by a cone $K$; and finally (iii) the base norm can be written as a maximisation over the interval $[0,u]_{C^*}$, as detailed in~\eqref{base norm}, hence one finds $\|x\|_{\widetilde{AB}}\leq \|x\|_{AB}$, as claimed.
An immediate corollary of this inequality is that
\begin{equation}
\|x\|_{AB} \leq\, \|x\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}
\label{Cmin largest norm}
\end{equation}
for all $x\in V_A \otimes V_B$ and for all admissible composites $AB$ whose corresponding cones $C_{AB}$ satisfy~\eqref{CAB bound}.
Now that dependence of the global base norm on the choice of the positive cone has been addressed, we can turn our attention to the other object appearing in the formula~\eqref{dh ratio eq} for computing data hiding ratios, i.e. the distinguishability norm $\|\cdot\|_{\mathcal{M}}$. We already observed that for the locally constrained sets of measurements $\mathcal{M}= \text{LO},\,\text{LOCC}_\rightarrow,\, \text{SEP}$, the distinguishability norm $\|\cdot\|_{\mathcal{M}}$ does not actually depend on $C_{AB}$. Therefore,
\begin{equation*}
R_{AB}(\mathcal{M})\, =\, \max_{0\neq x\in V_A\otimes V_B} \|x\|_{AB}\big/ \|x\|_{\mathcal{M}}\, \leq\, \max_{0\neq x\in V_A\otimes V_B} \|x\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B} \Big/ \|x\|_{\mathcal{M}}\, =\, R_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}(\mathcal{M})\, ,
\end{equation*}
where the dependence of the data hiding ratio on the GPT has been made explicit. The data hiding ratio $R(\mathcal{M})$ is thus maximised by the minimal tensor product $A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! B$, as claimed.
\end{proof}
\vspace{1ex}
\begin{rem}
Some intuitive understanding of Proposition~\ref{min is optimal} can be gained by looking at the opposite case, i.e. when the composite system is formed via the maximal tensor product. When $AB=A \ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! B$, in fact, the global base norm coincides with the separability norm. This is ensured by the fact that every allowed effect within this theory is automatically separable.
\end{rem}
\section{Examples: quantum mechanics and spherical model} \label{sec ex}
Throughout this section, we will investigate from the point of view of data hiding two well-known examples of GPTs, namely quantum mechanics and the so-called spherical model. Besides seeing all the mathematical machinery of GPTs in action in some concrete case, the purpose of doing so is twofold. First of all, the GPTs we chose to look into, especially quantum mechanics, are interesting on their own and deserve a complete solution. Second of all, computing the data hiding ratio against a locally constrained set of measurements for a specific case still yields a general lower bound on the maximal value achievable for fixed local dimensions.
\vspace{2ex}
\begin{ex}\label{ex QM}
Let us start by describing $n$-level quantum mechanics as a GPT. As is well-known, the cone of states is composed of the positive semidefinite $n\times n$ matrices (collectively denoted by $\text{PSD}_{n}$), embedded in the real space of hermitian matrices (called $\mathcal{H}_{n}$) whose real dimension is $d=n^2$. Since the density matrices are the positive matrices with trace one, the unit effect is easily seen to coincide with the trace. Therefore, we will write symbolically
\begin{equation}
\text{QM}_{n} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \left( \mathcal{H}_{n},\, \text{PSD}_{n},\, \text{Tr}\, \right) ,
\label{quantum}
\end{equation}
remembering that
\begin{equation}
\dim \text{QM}_{n}\, =\, n^{2}\, .
\label{dim quantum}
\end{equation}
Observe that the positive semidefinite cone is self-dual, i.e. $\text{PSD}_{n}^{*}=\text{PSD}_{n}$. The base norm in quantum mechanics can be proved to coincide with the trace norm $\|X\|_{1} = \text{Tr}\, |X| =\sum_{i=1}^{n}|\lambda_{i}(X)|$, where $\lambda_{i}(X)$ are the eigenvalues of $X\in\mathcal{H}_{n}$ and the last equality holds because $X$ is hermitian.
Now, let us discuss the composition rules for bipartite systems. Using the definitions~\eqref{minimal} and~\eqref{maximal}, we see that
\begin{align}
\text{PSD}_{n_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! \text{PSD}_{n_{B}}\, &=\, \Big\{ \sum\nolimits_{i\in I} P_{i}\otimes Q_{i} :\ I\ \text{finite},\ P_{i}\in\text{PSD}_{n_{A}},\, Q_{i}\in \text{PSD}_{n_{B}}\ \forall\ i\in I \Big\}\, , \label{separable} \\[1ex]
\text{PSD}_{n_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! \text{PSD}_{n_{B}}\, &=\, \left\{ W\in \mathcal{H}_{n_{A}n_{B}}:\ \text{Tr}\,[ P\otimes Q\, W ]\geq 0\quad \forall\ P\in\text{PSD}_{n_{A}},\, Q\in \text{PSD}_{n_{B}} \right\} . \label{witnesses} \end{align}
In quantum information, elements of~\eqref{witnesses} are variously called entanglement witnesses, separability witnesses or block-positive operators. This latter name comes from the fact that in~\eqref{witnesses} we can restrict $P$ and $Q$ to be pure states (i.e. rank-one projectors), and the defining condition for belonging to the set amounts to impose the positivity of the diagonal block(s) in all product bases. Interestingly enough, Nature has a preferred choice for the cone of bipartite states, which is neither the maximal nor the minimal tensor product. Instead, if $A=\text{QM}_{n_{A}}$ and $B=\text{QM}_{n_{B}}$ then $AB=\text{QM}_{n_{A}n_{B}}$, i.e.
\begin{equation}
\text{PSD}_{n_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! \text{PSD}_{n_{B}}\, \subsetneq\, C_{AB}\, =\, \text{PSD}_{n_{A}n_{B}}\, \subsetneq\, \text{PSD}_{n_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! \text{PSD}_{n_{B}}\, .
\label{cone bipartite quantum}
\end{equation}
While this is clearly not optimal in the sense of data hiding (because of Proposition~\ref{min is optimal}), it deserves special attention because of its prime importance in physics. Toward the end of this example, we will also look into a modified version of quantum mechanics designed to encompass the minimal tensor product rule suggested by Proposition~\ref{min is optimal}.
The computation of data hiding ratios in quantum mechanics has been the subject of many papers, whose main results we summarise briefly. The original example of a data hiding pair involves the normalised projectors onto the symmetric and antisymmetric subspace in $\mathds{C}^{n}\otimes\mathds{C}^{n}$, denoted by $\rho_{S}$ and $\rho_{A}$, respectively~\cite{dh original 1, dh original 2}. While $\|\rho_{S}-\rho_{A}\|_{1}=2$ because the two states have orthogonal support, it can be shown that $\|\rho_{S}-\rho_{A}\|_{\text{LOCC}}=2/(n+1)$~\cite{dh original 1, dh original 2}. The fact that the two states are mixed is crucial for this construction to work, as it can be shown that for pure states trace norm and LOCC norm always coincide~\cite{no dh pure 1, no dh pure 2}. In general, from~\cite{VV dh Chernoff, VV dh} it is known that
\begin{equation}
\frac{\min\{n_A, n_B\}+1}{2}\, \leq\, R_{QM}(\text{SEP})\, \leq\, R_{QM}(\text{LOCC})\, \leq\, R_{QM}(\text{LO})\, \leq\, \sqrt{153\, n_A n_B}\, , \label{VV bound 1}
\end{equation}
while for $R_{QM}(\text{SEP})$ the tighter bound
\begin{equation}
\frac{\min\{n_A, n_B\}+1}{2}\, \leq\, R_{QM}(\text{SEP})\, \leq\, \sqrt{n_A n_B} \label{VV bound 2}
\end{equation}
is available. As shown in~\cite[Corollary 17]{VV dh}, the above relations solve the problem of determining the optimal scaling in $n$ of all the data hiding ratios against locally constrained measurements when the subsystems have equal dimensions $n_A=n_B=n$. Instead, a problem arises when $n_A$ and $n_B$ are very different and thus $\min\{n_A,n_B\} \ll \sqrt{n_A n_B}$. In this case, the leftmost and rightmost side of~\eqref{VV bound 1} are no longer of the same order of magnitude, and an alternative argument has to be designed.
This scaling problem is somehow mitigated by~\cite[Lemma 20]{Brandao area law}, which implies that
\begin{equation}
R_{QM}(\text{LOCC})\, \leq\, R_{QM} (\text{LO})\, \leq\, \min\{n_A^2,\, n_B^2\}\, .
\end{equation}
Although this upper bound behaves better than that in~\eqref{VV bound 1} when $n_A$ and $n_B$ are very different from each other, its quadratic nature prevents us from determining -- for instance -- the exact scaling of the operationally relevant data hiding ratio against LOCC protocols.
Here we provide a simple reasoning that shows that in fact $O(\min\{n_A,n_B\})$ is still an upper bound for $R_{QM}(\text{LOCC})$ (and hence for $R_{QM}(\text{SEP})$, too). Furthermore, our reasoning yields much better constants for both the leftmost and the rightmost side of~\eqref{VV bound 1} (where $R_{QM}(\text{LO})$ is excluded, though). In fact, these constants are so close to being optimal that we are even able to compute $R_{QM}(\text{SEP})$ {\it exactly} when $n_A=n_B=n$.
\vspace{2ex}
\begin{thm}[Teleportation argument] \label{dh QM}
For a bipartite quantum mechanical system with Hilbert space $\mathds{C}^{n_A}\otimes \mathds{C}^{n_B}$, define $n\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\min\{n_A,n_B\}$. Then the data hiding ratios against separable and $\text{LOCC}_{(\rightarrow)}$ protocols satisfy
\begin{equation}
n\, \leq\, R_{QM}(\text{\emph{SEP}})\, \leq\, R_{QM}(\text{\emph{LOCC}})\, \leq\, R_{QM}(\text{\emph{LOCC}}_{\rightarrow})\, \leq\, 2n-1\, ,
\end{equation}
where the communication direction in $\text{LOCC}_{\rightarrow}$ is from the smaller to the larger subsystem. Moreover, if $n_A=n_B=n$ then $R_{\text{\emph{QM}}}(\text{\emph{SEP}})=n$.
\end{thm}
\begin{proof} Let us assume without loss of generality that $n=n_{A}\leq n_{B}$, and that classical communication goes from $A$ to $B$. We start by reminding the reader that the maximally entangled state $\ket{\Phi}=\frac{1}{\sqrt{n}}\,\sum_{i=1}^{n} \ket{ii}\in\mathds{C}^{n}\otimes \mathds{C}^{n}$, whose corresponding rank-one projector we denote by $\Phi$, has the property that there is a separable state $\sigma$ such that $\frac1n \Phi+\frac{n-1}{n}\, \sigma$ is again separable (in the language of~\cite{VidalTarrach}, $\Phi$ has entanglement robustness $r(\Phi)=n-1$). For instance, it is not difficult to see that $\sigma = \frac{\mathds{1} - \Phi}{n^{2}-1}$ satisfies all the requirements. This follows, for instance, from the characterisation of the separability region for isotropic states given in~\cite{Horodecki97}.
Now, since we can always produce any separable state with $\text{LOCC}_{\rightarrow}$ operations, we are free to evaluate the $\text{LOCC}_{\rightarrow}$ norm on $X_{AB} \otimes \left( \frac1n \Phi+\frac{n-1}{n}\, \sigma \right)_{A'B'}$ instead of $X_{AB}$. Here, the systems $A',B'$ have dimension $n_{A'}=n_{B'}=n_A=n$, and the operations are $\text{LOCC}_{\rightarrow}$ with respect to the splitting $AA'|BB'$.
Now, we are ready to apply the quantum teleportation protocol from $A$ to $B$~\cite{teleportation}. This is an $\text{LOCC}_{\rightarrow}$ operation $\tau$ mapping states of the system $AA'BB'$ to states of $B'B$, which can be defined as follows.
For $p,q=0,\ldots,n-1$, introduce the unitary matrices
\begin{equation}
\mathbf{X}(p)\,\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \sum_{k=1}^n \ket{k \oplus p}\!\!\bra{k}\, ,\qquad \mathbf{Z}(q)\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \sum_{k=1}^n e^{2qk\pi i/n} \ket{k}\!\!\bra{k}\, ,\qquad U(p,q)\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \mathbf{X}(p) \mathbf{Z}(q)\, ,
\label{HW}
\end{equation}
where $\oplus$ denotes sum modulo $n$.
Then the teleportation $\tau$ is given by
\begin{equation}
\tau(X_{AA'BB'})\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \sum_{p,q=0}^{n-1} U(p,q)_{B'}\ \text{Tr}_{AA'}\,\big[ X_{AA'BB'}\, U(p,q)_A \Phi_{AA'} U(p,q)_A^\dag \big]\ U(p,q)_{B'}^{\dag}\, .
\label{telep}
\end{equation}
Most notably, observe that $\tau\left(X_{AB} \otimes \Phi_{A'B'}\right)=X_{B'B}$ (meaning that the same operator $X$ is written in the registers $B'\simeq A$ and $B$). Now, on the one hand, after the protocol has been performed, the local constraint plays no role any more, and any measurement can be applied to $B'B$, showing that $\|X_{B'B}\|_{\text{LOCC}_{\rightarrow}} = \|X_{AB}\|_{1}$. On the other hand, $\tau(X_{AB}\otimes \sigma_{A'B'})$ is obtained from $X_{AB}$ via an $\text{LOCC}_{\rightarrow}$ protocol, hence $\|\tau (X_{AB} \otimes \sigma_{A'B'})\|_{\text{LOCC}_{\rightarrow}}\leq\|X_{AB}\|_{\text{LOCC}_{\rightarrow}}$.
Putting all together, we obtain the following chain of inequalities:
\begin{align}
\|X_{AB}\|_{\text{LOCC}_{\rightarrow}}\, &=\, \left\| X_{AB} \otimes \left( \frac1n \Phi+\frac{n-1}{n}\, \sigma \right)_{A'B'} \right\|_{\text{LOCC}_{\rightarrow}}
\label{telep 1} \\[0.5ex]
&\geq\, \left\| \tau\left(X_{AB} \otimes \left( \frac1n \Phi+\frac{n-1}{n}\, \sigma \right)_{A'B'}\right) \right\|_{\text{LOCC}_{\rightarrow}} \label{telep 2} \\[0.5ex]
&=\, \left\| \frac1n\, X_{B'B} +\frac{n-1}{n}\, \tau(X_{AB}\otimes \sigma_{A'B'}) \right\|_{\text{LOCC}_{\rightarrow}} \label{telep 3} \\
&\geq\, \frac1n\, \|X_{B'B}\|_{\text{LOCC}_{\rightarrow}} - \frac{n-1}{n}\, \|\tau(X_{AB}\otimes \sigma_{A'B'}) \|_{\text{LOCC}_{\rightarrow}} \label{telep 4} \\[0.5ex]
&\geq\, \frac1n\, \|X_{AB}\|_{1} - \frac{n-1}{n}\, \|X_{AB} \|_{\text{LOCC}_{\rightarrow}}\, . \label{telep 5}
\end{align}
We conclude that
\begin{equation}
\|X_{AB}\|_{\text{LOCC}_{\rightarrow}}\, \geq\, \frac{1}{2n-1}\, \left\| X_{AB} \right\|_1\, , \label{bound telep}
\end{equation}
enforcing $R_{QM}(\text{LOCC}_{\rightarrow})\leq 2n-1$ in view of Proposition~\ref{dh ratio}.
In order to deduce the lower bound $R_{QM}(\text{SEP})$, we appeal to Werner states~\cite{Werner, Werner symmetry}. These can be thought of as convex combinations of the normalised projectors onto the symmetric and antisymmetric subspace of $\mathds{C}^{n}\otimes \mathds{C}^{n}$, denoted by $\rho_{S}$ and $\rho_{A}$, respectively. In terms of the `flip operator' $F$ defined by $F\ket{\alpha\beta} = \ket{\beta\alpha}$ for all $\ket{\alpha},\ket{\beta}\in\mathds{C}^{n}$, we have
\begin{equation}
\rho_{S}\, =\, \frac{\mathds{1}+F}{n(n+1)}\, ,\qquad \rho_{A}\, =\, \frac{\mathds{1}-F}{n(n-1)}\, .
\label{symm antisymm proj}
\end{equation}
Since $n=\min\{n_{A},n_{B}\}$, we can safely imagine to give one share of this bipartite system to $A$ and the other to $B$. We already saw how the preparation with equal a priori probabilities of the two extremal states is well-known to produce data hiding, as shown by the fact that $\|\rho_{S}-\rho_{A}\|_{1}=2$ but $\|\rho_{S}-\rho_{A}\|_{\text{SEP}}=\|\rho_{S}-\rho_{A}\|_{\text{LOCC}}=2/(n+1)$~\cite{VV dh, VV dh Chernoff}. Curiously, there is an optimised version of this construction with different weights that does not seem to have been considered before. Namely, via the same techniques it can be shown that
\begin{equation}
\left\| \frac{n+1}{n}\, \rho_{S} - \frac{n-1}{n}\, \rho_{A} \right\|_{1}\, =\, 2\, ,\qquad \left\| \frac{n+1}{n}\, \rho_{S} - \frac{n-1}{n}\, \rho_{A} \right\|_{\text{SEP}}\, =\, \frac{2}{n}\, . \label{bound Werner}
\end{equation}
Since the proof of~\eqref{bound Werner} is just a variation of a standard calculation, we relegate it to Appendix~\ref{app Werner}. Thanks to Proposition~\ref{dh ratio}, this yields the lower bound in the claim. Finally, combining $R_{QM}(\text{SEP})\geq n$ with the upper bound in~\eqref{VV bound 2}, we see that when $n_{A}=n_{B}=n$ we must have $R_{QM}(\text{SEP})=n$.
\end{proof}
\vspace{1ex}
\begin{rem}
The fact that the upper bound for $R_{QM}(\text{LOCC})$ in Theorem~\ref{dh QM} scales only linearly in $\min\{n_{A},n_{B}\}$ is crucial in solving the data hiding problem in quantum mechanics (up to constants) for all pairs $(n_{A}, n_{B})$. To our knowledge, this complete solution was not known before. We find the simplicity of the above proof quite instructive on its own, but it does not seem like the teleportation argument can encompass the case of purely local measurements, by its very nature. Therefore, we must leave open the problem of finding the optimal scaling of $R_{QM}(\text{LO})$ in the general case.
\end{rem}
\vspace{1ex}
\begin{rem}
In terms of the real dimensions of the local spaces, given by $d_A=n_A^2,\, d_B=n_B^2$ according to~\eqref{dim quantum}, Theorem~\ref{dh QM} shows that the data hiding ratio against separable protocols scales as $\min\{\sqrt{d_{A}},\,\sqrt{d_{B}} \}$. Thus, we deduce a first estimate $R_{\text{SEP}}(d_{A},d_{B})\geq \min\{\sqrt{d_{A}},\,\sqrt{d_{B}}\}$ (valid when $\sqrt{d_{A}},\sqrt{d_{B}}$ are integers).
\end{rem}
\vspace{2ex}
In view of Proposition~\ref{min is optimal}, the reader might wonder, whether considering a modified version of quantum mechanics in which composite systems are obtained via the minimal tensor product~\eqref{separable} exhibits better data hiding properties. In what follows, we call such a theory {\it witness theory} (or {\it $W$-theory}, for short). In $W$-theory, the only allowed states of a multipartite system are fully separable, while the set of possible effects contains all entanglement witnesses (equivalently, all elements in the cone~\eqref{witnesses}).
Thus, the base norm of a bipartite operator $X_{AB}$ will be given by
\begin{equation}
\|X\|_W\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \|X\|_{\text{QM}_{n_A}\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! \text{QM}_{n_B}} =\, \max\big\{\text{Tr}\, XY:\ Y\in \mathcal{H}_{n_An_B},\ \big|\!\braket{\alpha\beta|Y|\alpha\beta}\!\big|\leq \braket{\alpha|\alpha}\! \braket{\beta|\beta}\ \, \forall\, \ket{\alpha}\in\mathds{C}^{n_A},\, \ket{\beta}\in\mathds{C}^{n_B} \big\}\, .
\label{W norm}
\end{equation}
Let us consider the class of protocols that can be realised in the $W$-theory framework when two agents $A$ and $B$ are allowed to use just local operations and $A\rightarrow B$ classical communication. We denote this class of protocols by $\text{{\bf LWCC }}_\rightarrow$. Despite the fact that $\text{{\bf LWCC }}_\rightarrow$ constitutes a more general class of {\it protocols} than $\text{LOCC}_\rightarrow$ in standard quantum mechanics, it is not difficult to see that they are not more powerful than the latter within the context of state discrimination. This is a consequence of the fact that the {\it measurements} one can implement with $\text{{\bf LWCC}}_\rightarrow$ operations are -- up to coarse graining -- necessarily of the form $\big(E_i\otimes F_j^{(i)}\big)_{(i,j)\in I\times J}$ for some local measurements $(E_i)_{i\in I}$ on $A$ and $\big( F^{(i)}_j\big)_{j\in J}$ on $B$. Since local measurements in $W$-theory are the same as in standard quantum mechanics, the same measurement is also obtainable via $\text{LOCC}_\rightarrow$ operations. More generally, we saw already that the locally constrained sets of measurements defined in~\eqref{LO},~\eqref{1-way LOCC},~\eqref{SEP} do not depend on the composition rule we choose for assembling multipartite systems.
The above discussion allows us to write the identity
\begin{equation}
\|X\|_{\text{LOCC}_{\rightarrow}} =\, \|X\|_{\text{{\bf LWCC}}_{\rightarrow}}\, ,
\label{confusing}
\end{equation}
where the right-hand side is defined through the usual formula~\eqref{d norm}, and it is understood that the corresponding set $\mathcal{M}$ includes in this case all those measurements that are implementable through an $\text{{\bf LWCC}}_{\rightarrow}$ protocol.
As it turns out, data hiding in $W$-theory is not much better than in quantum theory, in the sense that the scaling with the local dimensions is exactly the same.
The proof of this latter result constitutes another example of how the techniques used in~\cite{VV dh Chernoff, VV dh} seem not to be applicable in a more general scenario. In fact, the approach taken there relies on the inequality $\|\cdot\|_{\text{LOCC}_{\rightarrow}}\geq C \|\cdot\|_2$ (for $C$ universal constant), where $\|\cdot\|_2$ stands for the Hilbert-Schmidt norm. When the system under examination is $\mathds{C}^n\otimes \mathds{C}^n$, the elementary relation $\|\cdot\|_2\leq \frac{1}{n} \|\cdot\|_1$ yields $\|\cdot\|_{\text{LOCC}_{\rightarrow}}\geq \frac{C}{n} \|\cdot\|_1$, which is optimal up to a constant factor. However, it is not difficult to prove that the ratio between the norm $\|\cdot\|_W$ defined in~\eqref{W norm} and the Hilbert-Schmidt norm $\|\cdot\|_2$ can be asymptotically as large as $n^{3/2}$ (see Appendix~\ref{app W norm}). Therefore, the tighter inequality we can deduce by making use of the Hilbert-Schmidt norm in an intermediate step is $\|\cdot\|_{\text{LOCC}_\rightarrow}\geq C \|\cdot\|_2 \geq \frac{C}{n^{3/2}} \|\cdot\|_W$. As we will see in a moment, the scaling of the lower bound is not tight.
While a direct approach via the other techniques previously exploited in the literature does not lead to a satisfactory answer to the problem, the teleportation argument can be quickly adapted to compute {\it exactly} the data hiding ratios against $\text{SEP}$ or $\text{LOCC}_{\rightarrow}$ in $W$-theory.
\vspace{2ex}
\begin{prop} \label{dh W}
For a bipartite $W$-theory with local Hilbert spaces of dimensions $n_{A},n_{B}$, define $n\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\min\{n_A,n_B\}$. Then the data hiding ratios against separable and $\text{LOCC}_{\rightarrow}$ measurements are given by
\begin{equation}
R_{W}(\text{\emph{SEP}})\, =\, R_{W}(\text{\emph{LOCC}}_{\rightarrow})\, =\, 2n-1\, ,
\end{equation}
where the communication direction in $\text{LOCC}_{\rightarrow}$ is from the smaller to the larger subsystem.
\end{prop}
\begin{proof}
We assume without loss of generality that $n_B\geq n_A=n$. Let us start by showing that the above data hiding ratios can be upper bounded by $2n-1$. The idea is that the argument in~\eqref{telep 1}-\eqref{telep 5} can be adapted to encompass also the case of $W$-theory, by replacing everywhere $\|\cdot\|_{\text{LOCC}_{\rightarrow}}$ with $\|\cdot\|_{\text{LWCC}_{\rightarrow}}$ and the trace norm $\|\cdot\|_{1}$ with the minimal tensor product base norm~\eqref{W norm}.
The first step consists in acknowledging the fact that we can compute the $\text{LOCC}_{\rightarrow}$ distinguishability norm by making use of more general $\text{{\bf LWCC}}_{\rightarrow}$ protocols that are available in $W$-theory, as expressed in~\eqref{confusing}.
Now, we choose a particular $\text{{\bf LWCC}}_\rightarrow$ protocol in order to lower bound the norm $\|\cdot\|_{\text{\textbf{LWCC}}_{\rightarrow}}$. Such a protocol resembles the one we devised for the proof of Theorem~\ref{dh QM}, with one important difference.
Since separable states can be created with local operations and shared randomness, we can safely start by supplying $A$ and $B$ with a separable isotropic state $\left( \frac1n \Phi+\frac{n-1}{n}\, \sigma \right)_{A'B'}$, defined on an ancillary system $A'B'$ with local dimension $n_{A'}=n_{B'}=n$. As usual, $\sigma$ is an appropriate normalised, separable state. Then, we perform the teleportation $\tau$ defined in~\eqref{telep}, which is an $\text{{\bf LWCC}}_\rightarrow$ (even $\text{LOCC}_\rightarrow$) operation with respect to the splitting $AA'|BB'$, where classical communication goes from $AA'$ to $BB'$. After applying the triangle inequality, we are left with two terms, that is, $\|X_{B'B}\|_{\text{{\bf LWCC}}_\rightarrow}$ and $\|\tau(X_{AB}\otimes \sigma_{A'B'})\|_{\text{{\bf LWCC}}_\rightarrow}$. The first one can be computed exactly, since the operator $X_{B'B}$ obtained after teleportation belongs to the local subsystem $BB'$, and therefore measuring any witness $Y_{B'B}$ satisfying the constraints in~\eqref{W norm} is an allowed $\text{{\bf LWCC}}_\rightarrow$ operation, leading to the equality $\|X_{B'B}\|_{\text{{\bf LWCC}}_\rightarrow}=\|X\|_W$. As for the second term, we observe that $\|\tau(X_{AB}\otimes \sigma_{A'B'})\|_{\text{{\bf LWCC}}_\rightarrow}\leq \|X\|_{\text{{\bf LWCC}}_\rightarrow}$, since adding the ancillary system $A'B'$ in a separable state $\sigma_{A'B'}$ and subsequently applying $\tau$ is clearly an $\text{{\bf LWCC}}_\rightarrow$ protocol (which is why this latter inequality is in fact an equality). The above reasoning can be summarised in the following chain of inequalities, totally analogous to~\eqref{telep 1}-\eqref{telep 5}:
\begin{align*}
\|X_{AB}\|_{\text{LOCC}_\rightarrow}\, &=\, \|X_{AB}\|_{\text{\textbf{LWCC}}_\rightarrow} \\[0.5ex]
&=\, \left\| X_{AB} \otimes \left( \frac1n \Phi +\frac{n-1}{n}\, \sigma \right)_{A'B'} \right\|_{\text{\textbf{LWCC}}_\rightarrow} \\[0.5ex]
&\geq\, \left\| \tau\left(X_{AB} \otimes \left( \frac1n \Phi +\frac{n-1}{n}\, \sigma \right)_{A'B'}\right) \right\|_{\text{\textbf{LWCC}}_\rightarrow} \\[0.5ex]
&=\, \left\| \frac1n\, X_{B'B} +\frac{n-1}{n}\, \tau(X_{AB}\otimes \sigma_{A'B'}) \right\|_{\text{\textbf{LWCC}}_\rightarrow} \\
&\geq\, \frac1n\, \|X_{B'B}\|_{\text{\textbf{LWCC}}_\rightarrow} - \frac{n-1}{n}\, \|\tau(X_{AB}\otimes \sigma_{A'B'}) \|_{\text{\textbf{LWCC}}_\rightarrow} \\[0.5ex]
&\geq\, \frac1n\, \|X_{AB}\|_{W} - \frac{n-1}{n}\, \|X_{AB} \|_{\text{\textbf{LWCC}}_\rightarrow} \\[0.5ex]
&=\, \frac1n\, \|X_{AB}\|_{W} - \frac{n-1}{n}\, \|X_{AB} \|_{\text{LOCC}_\rightarrow}\, .
\end{align*}
In conclusion, we find
\begin{equation}
\|X_{AB}\|_{\text{LOCC}_{\rightarrow}}\, =\, \|X_{AB}\|_{\text{LWCC}_{\rightarrow}}\, \geq\, \frac{1}{2n-1}\, \left\| X_{AB} \right\|_W\, , \label{bound telep W}
\end{equation}
which implies $R_{W}(\text{LOCC}_{\rightarrow})\, \leq\, 2n-1$. Once more, to derive a lower bound on $R_{W}(\text{SEP})$ we use Werner states~\cite{Werner, Werner symmetry}. With the same notation as in the proof of Theorem~\ref{dh QM}, it can be shown that
\begin{equation}
\left\| \frac{n+1}{2n-1}\, \rho_{S}\, -\, \frac{n-1}{2n-1}\, \rho_{A} \right\|_{W}\, =\, 2\, ,\qquad \left\| \frac{n+1}{2n-1}\, \rho_{S}\, -\, \frac{n-1}{2n-1}\, \rho_{A} \right\|_{\text{SEP}}\, =\, \frac{2}{2n-1}\, , \label{bound Werner W}
\end{equation}
enforcing the complementary bound $R_{W}(\text{SEP})\, \geq\, 2n-1$. The proof of~\eqref{bound Werner W} is provided in Appendix~\ref{app Werner}.
\end{proof}
\end{ex}
\vspace{2ex}
\begin{ex} \label{ex sph}
The state space of $\text{QM}_{2}$ is well-known to be identifiable with a $3$-dimensional ball (Bloch ball). This observation is the starting point for defining a hypothetical class of physical models whose state space is a Euclidean ball of arbitrary dimension. These GPTs have been considered recently in connections to attempts of reconstructing quantum mechanics starting from few physically motivated axioms~\cite{quantum 5 axioms, quantum 4 1/2 axioms, quantum info unit}. This is to be expected in light of a famous classification theorem by Koecher and Vinberg~\cite{Koecher hom cones, Vinberg hom cones}, implying among other things that spherical models are one of the few classes of models enjoying a distinctive property of quantum mechanics called strong self-duality~\cite{telep in GPT}.
In the following, given $x\in \mathds{R}^{d}$ we call $x_{i}$ its $i$th entry ($i=0,\ldots,d-1$), while $\widebar{x}$ will denote the column vector obtained from $x$ by eliminating the zeroth component. The GPT corresponding to this {\it spherical model} has the form
\begin{equation}
\text{Sph}_{d} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \left( \mathds{R}^{d},\, C_d,\, u \right) .
\label{spherical}
\end{equation}
Here, $C_{d}$ is the `ice cream cone'
\begin{equation}
C_{d}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \left\{\, x \in \mathds{R}^{d}:\ |\widebar{x}|_{2}\, \leq\, x_{0}\, \right\} , \label{ice cream}
\end{equation}
where $|\cdot|_{2}$ is the standard Euclidean norm in $\mathds{R}^{d-1}$, and $u$ is defined as $u(y)\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = y_{0}$ for all $y\in \mathds{R}^{d}$. In light of the canonical identification $\left( \mathds{R}^{d} \right)^{*}\simeq \mathds{R}^{d}$, we find it convenient to adopt from now on a column notation for the dual as well as for the primal space. Within this convention, we shall write unambiguously $u=(1,0,\ldots,0)^{T}$.
This simplification of the notation is going to pay off because just like quantum mechanics, also the spherical model is self-dual, i.e. $C_{d}=C_{d}^{*}$.
As for the base norm in $\text{Sph}_{d}$, it can be shown that $\|x\|=\max\{|x_{0}|,|\widebar{x}|_{2}\}$.
Since our primary interest is in the exploration of the data hiding properties, according to Proposition~\ref{min is optimal} we construct a bipartite system $AB$ by joining two spherical models $A=\text{Sph}_{d_{A}}$ and $B=\text{Sph}_{d_{B}}$ via the minimal tensor product, i.e. taking $AB=\text{Sph}_{d_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! \text{Sph}_{d_{B}}$. This latter assumption will be made throughout the rest of this example. Tensors belonging to the bipartite vector space $\mathds{R}^{d_{A}}\otimes\mathds{R}^{d_{B}}$ (or to its dual) can be thought of as $d_{A}\times d_{B}$ matrices $X\in\mathds{R}^{d_{A}\times d_{B}}$. We shall find useful to denote by $\widebar{X}$ the $(d_{A}-1)\times (d_{B}-1)$ submatrix of $X\in\mathds{R}^{d_{A}\times d_{B}}$ which is obtained by cutting off the zeroth components $X_{i0},\, X_{0j}$ of the latter. Complementarily, given $M\in\mathds{R}^{(d_{A}-1)\times (d_{B}-1)}$ we call $\hat{M}$ the $d_{A}\times d_{B}$ `lifted' matrix whose entries are
\begin{equation*}
\hat{M}_{ij}\, =\, \left\{ \begin{array}{cl} 0 & \text{ if $i=0$ or $j=0$,} \\[0.5ex] M_{ij} & \text{ if $i,j\geq 1$.} \end{array} \right.
\end{equation*}
According to Proposition~\ref{dh ratio}, the data hiding ratio against separable measurements can be computed once we know the expressions for both the separability norm and the base norm induced by the minimal tensor product. Instead of treating the general case, we show how to compute these norms for a restricted yet large class of matrices, that is, those having zero entries in the zeroth row and column.
\vspace{2ex}
\begin{lemma} \label{norms sph}
Consider a bipartite system $AB=\text{\emph{Sph}}_{d_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! \text{\emph{Sph}}_{d_{B}}$. Then for all $M\in\mathds{R}^{(d_{A}-1)\times (d_{B}-1)}$ we have
\begin{equation}
\big\| \hat{M} \big\| = \|M\|_1\, ,\qquad \big\| \hat{M}\big\|_{\text{\emph{SEP}}} = \|M\|_\infty\, , \label{norms sph eq}
\end{equation}
where it is understood that the base norm $\| \cdot \|$ is induced by the minimal tensor product, and $\|M\|_{1},\, \|M\|_{\infty}$ denote the trace and operator norm of $M$, i.e. the sum and the largest of its singular values, respectively.
\end{lemma}
\begin{proof}
Consider an arbitrary dual tensor $E\in\mathds{R}^{d_{A}\times d_{B}}$. We claim that if $E$ is separable then necessarily $\big\|\widebar{E}\,\big\|_{1}\leq E_{00}$, and that this condition is also sufficient when all cross terms $E_{i0}, E_{0j}$ ($i,j\geq 1$) vanish. Let us start by proving the necessity of the above inequality. If $E=\sum_{k}x^{(k)} \big(y^{(k)}\big)^{T}$ with $\big|\widebar{x}^{(k)}\big|_{2}\leq x^{(k)}_{0},\, \big|\widebar{y}^{(k)}\big|_{2}\leq y^{(k)}_{0}$ for all $k$, then
\begin{equation}
\big\|\widebar{E}\,\big\|_{1}\, =\, \bigg\|\sum_{k} \widebar{x}^{(k)}\big(\widebar{y}^{(k)}\big)^{T}\bigg\|_1\, \leq\, \sum_{k} \Big\| \widebar{x}^{(k)} \big(\widebar{y}^{(k)}\big)^T\Big\|_{1}\, =\, \sum_{k} \big|\widebar{x}^{(k)}\big|_{2} \big|\widebar{y}^{(k)}\big|_{2}\, \leq\, \sum_{k} x^{(k)}_{0} y^{(k)}_{0}\, =\, E_{00}\, .
\label{norms sph proof 1}
\end{equation}
Now, let us turn to the sufficiency claim. Suppose that $E_{ij}=0$ whenever $i=0,\, j\geq 1$ or $i\geq 1,\, j=0$, and that $\big\|\widebar{E}\,\big\|_{1}\leq E_{00}$. Then by the singular value decomposition theorem there are vectors $v^{(k)}\in\mathds{R}^{d_{A}-1}$ and $w^{(k)}\in\mathds{R}^{d_{B}-1}$ such that $\widebar{E}=\sum_{k} v^{(k)} \big(w^{(k)}\big)^T$ and $\big\|\widebar{E}\,\big\|_{1}=\sum_{k} \big|v^{(k)}\big|_{2} \big|w^{(k)}\big|_{2}$. Define vectors $x^{(k)}_\pm\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \big|v^{(k)}\big|_{2} u_{A} \pm \hat{v}^{(k)}\in \mathds{R}^{d_{A}}$, i.e.
\begin{equation}
\left(x_{\pm}^{(k)}\right)_{i}\, =\, \left\{ \begin{array}{cl} \big|v^{(k)}\big|_{2} & \text{ if $i=0$,} \\[0.5ex] \pm\, v^{(k)}_i & \text{ if $i\geq 1$,} \end{array} \right.
\label{norms sph proof 2}
\end{equation}
and analogously for $y_{\pm}^{(k)}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \big|w^{(k)}\big|_{2} u_{B} \pm \hat{w}^{(k)} \in \mathds{R}^{d_{B}}$. Observe that the definition~\eqref{ice cream} tells us that $x_\pm^{(k)} \in C_{d_A}$ and $y_\pm^{(k)} \in C_{d_B}$. Defining $U\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = u_{AB}=u_A\otimes u_B=u_A u_B^T$, it is easy to see that
\begin{equation}
E\, =\, \left(E_{00} -\big\|\widebar{E}\,\big\|_1 \right) U\, +\, \frac12\, \sum_{k} \left( x^{(k)}_{+} \big(y^{(k)}_{+}\big)^{T} + x^{(k)}_{-} \big(y^{(k)}_{-}\big)^{T} \right) ,
\label{norms sph proof 3}
\end{equation}
which shows that $E$ is separable, as claimed.
Now, let us compute the separability norm. On the one hand, the above necessary condition for separability of effects shows that
\begin{align}
\big\|\hat{M}\big\|_\text{SEP}\, &=\, \max_{E,\,U-E\,\in\, C_{d_A}^*\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! C_{d_B}^*} \left( \big|E\big(\hat{M}\big)\big| + \big|(U-E)\big(\hat{M}\big)\big| \right) \, =\, \max_{U \pm F\,\in\, C_{d_A}^*\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! C_{d_B}^*} F\big(\hat{M}\big) \label{norms sph proof 4}\\
&\leq\, \max_{\left\|\bar{F}\right\|_1\, \leq\, 1\pm F_{00}} F\big(\hat{M}\big)\, =\, \max_{\left\|\bar{F}\right\|_1 +|F_{00}|\, \leq\, 1} \text{Tr}\, \Big[M \widebar{F}^{\,T}\Big]\, =\, \|M\|_\infty\, ,
\label{norms sph proof 5}
\end{align}
where we employed~\eqref{d norm altern} to find a more compact expression for the separability norm, and we exploited the fact that trace norm and operator norm are dual to each other. On the other hand, the fact that $\|N\|_1\leq 1$ is sufficient to guarantee the separability of $U \pm \hat{N}$ leads us, again via~\eqref{d norm altern}, to the complementary bound
\begin{equation*}
\big\| \hat{M} \big\|_\text{SEP}\, \geq\, \max_{\|N\|_1\leq 1} \hat{N}(\hat{M})\, =\, \max_{\|N\|_1\leq 1} \text{Tr}\,\left[ N^T M\right]\, =\, \|M\|_\infty\, .
\end{equation*}
Our final task is the calculation of the base norm induced by the minimal tensor product. Thanks to the formula~\eqref{dual base eq}, we can write
\begin{equation*}
\big\|\hat{M}\big\|\, =\, \min \Big\{ U(X_+)+U(X_-):\ X_\pm\in C_{d_A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{d_B},\ \hat{M}=X_+ - X_- \Big\}\, .
\end{equation*}
Since $X_\pm\in C_{d_A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{d_B}$ implies $U(X_\pm)\geq \big\|\widebar{X}_\pm\big\|_1$ and $\hat{M}=X_+ - X_-$ implies $M=\widebar{X}_+ - \widebar{X}_-$, we see that
\begin{equation*}
\big\|\hat{M}\big\|\, \geq\, \min \left\{ \big\|\widebar{X}_+\big\|_1 + \big\|\widebar{X}_-\big\|_1:\ M=\widebar{X}_+ - \widebar{X}_- \right\}\, \geq\, \|M\|_1\, ,
\end{equation*}
where the last lower bound follows from the triangle inequality. On the other hand, we construct an ansatz for $X_\pm$ achieving the above lower bound. From the singular value decomposition theorem, it is immediately seen that for all real matrices $M$ there exists a decomposition $M=M_+ - M_-$ such that $\|M_+\|_1=\|M_-\|_1=\frac{\|M\|_1}{2}$, and consequently $\|M\|_1=\|M_+\|_1+\|M_-\|_1$. Then, consider $X_\pm=\frac{\|M\|_1}{2}\, U + \hat{M}_\pm$, so that $X_+-X_-=\hat{M}$. Since there are no cross terms $(X_\pm)_{i0}$ or $(X_\pm)_{0j}$, the condition $(X_\pm)_{00}\geq \big\|\widebar{X}_\pm\big\|_1$ (satisfied by construction) is sufficient to ensure the separability of $X_\pm$, hence this is a valid ansatz. We find
\begin{equation*}
\big\|\hat{M}\big\|\, \leq\, U(X_+) + U(X_-)\, =\, \|M_+\|_1+\|M_-\|_1\, =\, \|M\|_1\, ,
\end{equation*}
and we are done.
\end{proof}
\vspace{1ex}
\begin{cor} \label{dh sph}
In the bipartite GPT $AB=\text{\emph{Sph}}_{d_{A}}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! \text{\emph{Sph}}_{d_{B}}$, the data hiding ratio against separable measurements can be lower bounded as
\begin{equation}
R_{\text{\emph{Sph}}}(\text{\emph{SEP}})\, \geq\, \min\{d_{A},d_{B}\} - 1\, .
\end{equation}
In particular, the ultimate data hiding ratios against locally constrained sets of measurements obey
\begin{equation}
R_{\text{\emph{LO}}}(d_{A},d_{B}) \geq R_{\text{\emph{LOCC}}_{\rightarrow}}(d_{A},d_{B}) \geq R_{\text{\emph{LOCC}}}(d_A,d_B) \geq R_{\text{\emph{SEP}}}(d_{A},d_{B}) \geq \min\{d_{A},d_{B}\} - 1\, . \label{univ dh ratio lower}
\end{equation}
\end{cor}
\begin{proof}
From Lemma~\ref{norms sph} we know that $R_\text{Sph}(\text{SEP})$ can be lower bounded by the maximal ratio between trace norm and operator norm of matrices in $\mathds{R}^{(d_{A}-1)\times (d_{B}-1)}$, which is well known to be $\min\{d_A-1,d_B-1\}=\min\{d_A,d_B\}-1$. Since we can provide an example of bipartite GPT with local dimensions $d_{A},d_{B}$ for which the data hiding ratio against separable measurements is no smaller than $\min\{d_A,d_B\}-1$, this constitutes a lower bound on the ultimate data hiding ratio $R_{\text{SEP}}(d_{A},d_{B})$.
\end{proof}
\vspace{2ex}
The problem of finding a complementary upper bound for the data hiding ratios in the spherical model is solved by the general result expressed in Theorem~\ref{thm univ} below, which implies that $R_{\text{Sph}}(\text{LO})\leq \min\{d_A,d_B\}$. This yields the almost tight, two-sided bound
\begin{equation}
\min\{d_{A},d_{B}\} - 1 \leq R_{\text{Sph}}(\text{SEP}) \leq R_{\text{Sph}}(\text{LOCC}) \leq R_{\text{Sph}}(\text{LOCC}_\rightarrow) \leq R_{\text{Sph}}(\text{LO}) \leq \min\{d_A,d_B\} ,
\label{double bound sph}
\end{equation}
which fully determines the scaling of all the data hiding ratios against locally constrained measurements up to an additive constant.
\vspace{1ex}
\begin{rem}
Corollary~\ref{dh sph} shows that quantum mechanics, even when it is modified to encompass the minimal tensor product composition rule according to Proposition~\ref{min is optimal}, is not optimal from the point of view of data hiding, in the sense that there exist GPTs with the same local dimensions but with a higher data hiding ratio against all locally constrained sets of measurements.
\end{rem}
\end{ex}
\section{Ultimate bound on data hiding effectiveness} \label{sec univ}
Throughout this section, we will prove our main result on ultimate data hiding ratios against locally constrained sets of measurement. Namely, we will show that the lower bound we found in Corollary~\ref{dh sph} by analysing the spherical model is optimal up to an additive constant. As discussed at the end of Section~\ref{sec dh GPTs}, this answers the central question of our investigation, that is, the determination of the general constraints that data hiding is forced to obey in any GPTs.
Before coming to our main theorem, we need some mathematical preliminary on tensor norms. We remind the reader~\cite{DEFANT, RYAN} that given two finite-dimensional Banach spaces $V_{A},V_{B}$ (whose norms will be equally denoted by $\|\cdot\|$ for simplicity), there are two notable ways in which one can construct a norm on the tensor product $V_{A}\otimes V_{B}$. The first of these construction yields the so-called {\it injective norm}, which can be expressed as
\begin{equation}
\|X\|_{\varepsilon}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{ (f\otimes g)(X):\ \, f\in V_{A}^{*},\ \|f\|_{*}\leq 1,\ g\in V_{B}^{*},\ \|g\|_{*}\leq 1 \right\} .
\label{inj}
\end{equation}
The second norm we are interested in goes under the name of {\it projective norm}, and is defined as
\begin{equation}
\|X\|_{\pi}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \inf\left\{ \sum_{i=1}^{n} \|x_{i}\|\,\|y_{i}\| :\ \, n\in\mathds{N},\ X=\sum_{i=1}^{n} x_{i}\otimes y_{i} \right\} .
\label{proj}
\end{equation}
These two norms are dual to each other in the following sense. Thinking of $V_{A}^{*}, V_{B}^{*}$ as Banach spaces equipped with the dual norms $\|\cdot\|_{*}$, we can construct the associated injective and projective tensor norms on $V_{A}^{*} \otimes V_{B}^{*}$, denoted by $\|\cdot\|_{*\varepsilon}$ and $\|\cdot\|_{*\pi}$, respectively. One could wonder, how these norms compare to the dual to~\eqref{inj} and~\eqref{proj}, denoted by $\|\cdot\|_{\varepsilon*}, \|\cdot\|_{\pi*}$, respectively. As it turns out, one has
\begin{equation}
\|\cdot\|_{*\varepsilon}\, =\, \|\cdot\|_{\pi*}\, ,\qquad \|\cdot\|_{*\pi}\, =\, \|\cdot\|_{\varepsilon*}\, .
\label{dual inj proj}
\end{equation}
\vspace{1ex}
\begin{note}
The notation is intended to help the reader via simple graphic rules. For instance, the symbol $\|\cdot\|_{\varepsilon*}$ stands for `first construct the injective norm, then take the dual', and conversely $\|\cdot\|_{*\pi}$ means `first take the dual norms, then construct the projective norm out of them'. Then, the above identities can be recovered easily by remembering that taking a $*$ from the external to the internal position (or vice versa) causes an exchange $\varepsilon\leftrightarrow \pi$.
\end{note}
\vspace{2ex}
Concerning the comparison between injective and projective norms, the inequality $\|X\|_{\varepsilon}\, \leq\, \|X\|_{\pi}$ is easily seen to hold for all $X\in V_{A}\otimes V_{B}$.
To see why this is the case, take a tensor $X\in V_A\otimes V_B$, and consider a decomposition $X=\sum_{i=1}^n x_i\otimes y_i$ as in~\eqref{proj}. For any two functionals $f\in V_A^*$ and $g\in V_B^*$ such that $\|f\|_*, \|g\|_*\leq 1$, one has $|f(x_i)|\leq \|x_i\|$ and $|g(y_i)|\leq \|y_i\|$, so that
\begin{equation*}
|(f\otimes g)(X)|\, =\, \bigg| \sum_{i=1}^n f(x_i) g(y_i) \bigg|\, \leq\, \sum_{i=1}^n |f(x_i)|\, |g(y_i) |\, \leq\, \sum_{i=1}^n \|x_i\|\, \|y_i\|\, \leq \, \|X\|_\pi\, ,
\end{equation*}
where in the last step we employed~\eqref{proj}. Maximising over $f,g$ as in~\eqref{inj} yields the claim.
We will write symbolically
\begin{equation}
\|\cdot\|_{\varepsilon}\, \leq\, \|\cdot\|_{\pi}\, .
\label{inj=<proj}
\end{equation}
As it turns out, for simple tensors $x\otimes y$ both norms coincide with the product of the local norms, i.e. $\|x\otimes y\|_{\varepsilon}=\|x\otimes y\|_{\pi}=\|x\|\, \|y\|$ (`reasonable cross norms').
In this work we are mainly interested in understanding the asymptotic behaviour of certain quantities as the dimension of the systems goes to infinity. This study fits in the so called local theory of Banach spaces, which is concerned in particular with the quantitative analysis of $d$-dimensional normed spaces (as $d\rightarrow \infty$). The investigation of tensor norms and their relations with operator ideals and factorizing operators is a crucial part of those studies~\cite{DEFANT}. Thus, it is not surprising that tensor norms play an important role in our approach. In this work, we will only care about the relation between projective and injective tensor norms when they are defined on finite dimensional Banach spaces. On the one hand, we are interested in the study of upper and lower bounds for the quotient $\|\cdot\|_{\pi}/ \|\cdot\|_{\varepsilon}$ for some particular spaces of relevance in our study. Let us mention in passing that the problem of comparing the projective and the injective tensor norms in the setting of infinite dimensional Banach leads to very deep results in the theory~\cite{Pisier, PisierII}. On the other hand, we ask ourselves what happens when we consider general finite dimensional Banach spaces with fixed dimensions. That is to say, we want to find the smallest $\mu(d_{A},d_{B})$ such that $\|\cdot\|_{\pi}\leq \mu\, \|\cdot\|_{\varepsilon}$ holds true for all Banach spaces $V_{A},V_{B}$ of dimensions $d_{A},d_{B}$, respectively. At this stage it is not even obvious that such a quantity will be finite. The answer to this question is provided by the following result, which could be known to experts in the topic, although we did not find any explicit reference.
\vspace{2ex}
\begin{prop} \label{prop proj=<ninj}
For all pairs of finite dimensional Banach spaces $V_{A},V_{B}$ with dimensions $d_{A}=\dim V_{A}$, $d_{B}=\dim V_{B}$, we have
\begin{equation}
\|\cdot\|_{\pi}\, \leq\, \min\{d_{A},d_{B}\}\, \|\cdot\|_{\varepsilon}\, .
\label{proj=<ninj}
\end{equation}
Furthermore, the constant on the right-hand side of the above inequality is the best possible.
\end{prop}
\begin{proof}
We start by recalling Auerbach's lemma~\cite[Vol I, Sec. 1.c.3]{LT}, which states that any finite-dimensional Banach space admits a basis $\{v_{i}\}_{i}$, whose associated dual basis we denote by $\{v_{i}^{*}\}_{i}$ (so that $v_{i}^{*}(v_{j})=\delta_{ij}$), such that
\begin{equation*}
\|v_{i}\| \leq 1\, ,\quad \|v_{i}^{*}\|_{*}\leq 1\qquad \forall\ i\, .
\end{equation*}
Suppose without loss of generality that $d_{A}\leq d_{B}$. Expand any tensor $X\in V_{A}\otimes V_{B}$ in the local Auerbach basis for $V_{A}$, that is, $X=\sum_{j=1}^{d_{A}} v_{j}\otimes y_{j}$. Choose $d_{A}$ functionals $g_{i}\in V_{B}^{*}$ such that $\|g_{i}\|_{*}\leq 1$ and $g_{i}(y_{i})=\|y_{i}\|$. Since also $\|v_{i}^{*}\|_{*}\leq 1$, using~\eqref{inj} we can lower bound the injective norm as
\begin{equation*}
\|X\|_{\varepsilon}\, \geq\, (v_{i}^{*}\otimes g_{i})(X)\, =\, \sum_{j=1}^{d_{A}} v_{i}^{*}(v_{j}) g_{i}(y_{j})\, =\, \sum_{j=1}^{d_{A}} \delta_{ij} g_{i}(y_{j})\, =\, g_{i}(y_{i})\, =\, \|y_{i}\|\, .
\end{equation*}
Then, the definition of projective norm as given in~\eqref{proj} tells us that
\begin{equation*}
\|X\|_{\pi}\, \leq\, \sum_{i=1}^{d_{A}} \|v_{i}\|\, \|y_{i}\|\, \leq\, \sum_{i=1}^{d_{A}} \|v_{i}\|\, \|X\|_{\varepsilon}\, =\, d_{A}\, \|X\|_{\varepsilon}\, ,
\end{equation*}
where we employed also the other defining property of the Auerbach basis, i.e. $\|v_{i}\|\leq 1$ for all $i$.
To see that the constant $\min\{d_{A}, d_{B}\}$ is optimal for all $d_{A},d_{B}$, we could appeal to Lemma~\ref{norms sph} together with the results of Subsection~\ref{subsec centr}. However, we resort here to an independent argument.
Consider two Euclidean spaces $V_{A},V_{B}$, whose norms we denote by $|\cdot|_{2}$. We can identify $V_{A}\otimes V_{B}$ with the set of $d_{A}\times d_{B}$ real matrices, denoted by $\mathds{R}^{d_{A}\times d_{B}}$. With this convention, for given $X\in V_{A}\otimes V_{B}$, $f\in V_{A}^{*}\simeq \mathds{R}^{d_{A}}$, and $g\in V_{B}^{*}\simeq \mathds{R}^{d_{B}}$, we can write $(f\otimes g)(X)=f^{T}Xg$. Then, remembering that Euclidean norms are self-dual, we obtain
\begin{equation*}
\|X\|_{\varepsilon}\, =\, \max \left\{ f^{T}Xg:\ |f|_{2*},|g|_{2*}\leq 1 \right\}\, =\, \max \left\{ f^{T}Xg:\ |f|_{2},|g|_{2}\leq 1 \right\}\, =\, \|X\|_{\infty}\, .
\end{equation*}
The fact that $|\cdot|_{2*}=|\cdot|_{2}$ also implies that $\|\cdot\|_{*\varepsilon}=\|\cdot\|_{\varepsilon}$. This together with~\eqref{dual inj proj} shows that
\begin{equation*}
\|\cdot\|_{\pi}\,=\,\|\cdot\|_{*\varepsilon*}\,=\,\|\cdot\|_{\varepsilon*}\,=\,\|\cdot\|_{\infty *}\,=\,\|\cdot\|_{1}\, .
\end{equation*}
Since $\|X\|_{1}=\min\{d_{A},d_{B}\}\, \|X\|_{\infty}$ whenever all the singular values of $X$ coincide, we can conclude.
\end{proof}
\vspace{2ex}
Note that the previous proof still applies if one of the spaces has infinite dimension. In fact, most of the results of this paper can be extended to that case straightforwardly. However, the case of two infinite-dimensional Banach spaces is completely different, and its study would require the use of more sophisticated Banach space techniques.
In order to apply Proposition~\ref{prop proj=<ninj} to our problem, we need to relate the distinguishability norms on a composite system to the injective and projective norm constructed out of the local base norms. This translation is the subject of our next result.
\vspace{2ex}
\begin{prop} \label{d norms inj proj}
Let $A=(V_{A},C_{A},u_{A})$ and $B=(V_{B},C_{B},u_{B})$ be two GPTs. The local base norms turn $V_{A},V_{B}$ into Banach spaces, and we can construct injective and projective tensor norms on $V_{A}\otimes V_{B}$, denoted simply by $\|\cdot\|_{\varepsilon}$ and $\|\cdot\|_{\pi}$. Considering the composite $A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! B$ given by the minimal tensor product as in~\eqref{min max theories}, we can also define on $V_{A}\otimes V_{B}$: (i) the base norm $\|\cdot\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! B}$ associated with $A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! B$; and (ii) the local distinguishability norm $\|\cdot\|_{\text{\emph{LO}}}$. Then the following holds true:
\begin{equation}
\|\cdot\|_{\pi}\, =\, \|\cdot\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! B}\, ,\qquad \|\cdot\|_{\varepsilon}\, \leq\, \|\cdot\|_{\text{\emph{LO}}}\, . \label{d norms inj proj eq}
\end{equation}
\end{prop}
\begin{proof}
We start by showing the first relation in~\eqref{d norms inj proj eq}. Consider $X\in V_{A}\otimes V_{B}$, and decompose it as $X=\sum_{i=1}^{n} x_i\otimes y_i$ in such a way that $\|X\|_{\pi} =\sum_{i=1}^{n} \|x_i\|\,\|y_i\|$ (see the definition~\eqref{proj}). According to Lemma~\ref{dual base}, we can construct $x_i^{\pm}\in C_{A}$ and $y_i^{\pm}\in C_{B}$ such that $x_i=x_i^+ - x_i^-$, $\|x_i\|=u_A\big(x_i^{+}\big)+u_A\big(x_i^{-}\big)$ and analogously for $y_i^\pm$. Then,
\begin{equation*}
X\, =\, \sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{+} + x_i^{-} \otimes y_i^{-}\right)\, -\, \sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{-} + x_i^{-}\otimes y_i^{+}\right)\, ,
\end{equation*}
and since
\begin{equation*}
\sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{+} + x_i^{-} \otimes y_i^{-}\right),\ \sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{-} + x_i^{-} \otimes y_i^{+}\right)\ \in\ C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}\, ,
\end{equation*}
the dual formula~\eqref{dual base eq} yields
\begin{align*}
\|X\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}\, &\leq\, (u_{A}\otimes u_{B})\bigg( \sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{+} + x_i^{-} \otimes y_i^{-}\right)\bigg)\, +\, (u_{A}\otimes u_{B})\bigg( \sum_{i=1}^{n} \left(x_i^{+}\otimes y_i^{-} + x_i^{-} \otimes y_i^{+}\right)\bigg) \\
&=\, \sum_{i=1}^{n} \left( u_A\big(x_i^{+}\big)+u_A\big(x_i^{-}\big) \right) \left( u_B\big(y_i^{+}\big)+u_B\big(y_i^{-}\big)\right)\, =\, \sum_{i=1}^{n} \big\|x_i\big\|\, \big\|y_i\big\|\, =\, \|X\|_{\pi}\, .
\end{align*}
To show the converse, notice first that $\|X\|_{\pi}=(u_{A}\otimes u_{B})(X)$ holds true for all separable $X\in C_{A} \ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}$. In fact, writing $X=\sum_{i} a_{i}\otimes b_{i}$ for $a_{i},b_{i}$ positive, we see that on the one hand $\|X\|_{\pi}\leq \sum_{i} \|a_{i}\|\, \|b_{i}\|= \sum_{i} u_{A}(a_{i}) u_{B}(b_{i}) = (u_{A}\otimes u_{B})(X)$, while on the other hand $(u_{A}\otimes u_{B})(X)\leq \|X\|_{\varepsilon}\leq\|X\|_{\pi}$ by inequality~\eqref{inj=<proj}. Now, for a generic $X\in V_{A}\otimes V_{B}$ apply once again Lemma~\ref{dual base} to construct $X_{\pm}\in C_{A}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_{B}$ such that $X=X_{+}-X_{-}$ and
\begin{equation*}
\|X\|_{A\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}=(u_{A}\otimes u_{B})(X_{+}) + (u_{A}\otimes u_{B})(X_{-})\, .
\end{equation*}
By using the above observation and the triangle inequality, we find
\begin{equation*}
\|X\|_{\pi} - (u_{A}\otimes u_{B})(X_{-})\, =\, \|X\|_{\pi} - \|X_{-}\|_{\pi}\, \leq\, \|X+X_{-}\|_{\pi}\, =\, \|X_{+}\|_{\pi}\, =\, (u_{A}\otimes u_{B})(X_{+})\, ,
\end{equation*}
from which
\begin{equation*}
\|X\|_{\pi}\, \leq\, (u_{A}\otimes u_{B})(X_{-}) + (u_{A}\otimes u_{B})(X_{+})\, =\, \|X\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}\, .
\end{equation*}
This completes the proof of the equality $\|\cdot\|_{\pi}\, =\, \|\cdot\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}$. Now, we move on to the second relation in~\eqref{d norms inj proj eq}. To find a lower bound on the separability norm, we will employ the expression~\eqref{d norm altern} with $\mathcal{M}=\text{LO}$. For arbitrary $X\in V_{A}\otimes V_{B}$ and $f\in V_{A}^{*},\, g\in V_{B}^{*}$ such that $\|f\|_{*}, \|g\|_{*}\leq 1$, we show that $f\otimes g$ is a valid test functional to be plugged into~\eqref{d norm altern} since $\mu\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \left( \frac12 (u_{A}\otimes u_{B} + f\otimes g),\, \frac12 (u_{A}\otimes u_{B} - f\otimes g)\right) \in \text{LO}$. In fact, notice that
\begin{align}
\frac12 \left( u_{A}\otimes u_{B} + f\otimes g \right)\, &=\, \frac{u_{A}+f}{2}\otimes \frac{u_{B}+g}{2}\, +\, \frac{u_{A}-f}{2}\otimes \frac{u_{B}-g}{2}\, , \label{fg local 1} \\
\frac12 \left( u_{A}\otimes u_{B} - f\otimes g \right)\, &=\, \frac{u_{A}+f}{2}\otimes \frac{u_{B}-g}{2}\, +\, \frac{u_{A}-f}{2}\otimes \frac{u_{B}+g}{2}\, . \label{fg local 2}
\end{align}
Thanks to the fact that the unit balls of the dual to the local base norms have the form $B_{\|\cdot\|_{*}}=[-u,u]$ (Definition~\ref{def base}), we know that $\left( \frac12 (u_{A}+ f),\, \frac12(u_{A}-f) \right)$ is a valid measurement on $A$, and analogously $\left( \frac12 (u_{B}+ g),\, \frac12(u_{B}-g) \right)$ is a measurement on $B$. Using~\eqref{LO}, it is easy to see that $\mu$ is indeed obtainable from a product measurement via a coarse graining procedure as defined in~\eqref{coarse}. Then, equation~\eqref{d norm altern} yields $\|X\|_{\text{LO}}\geq (f\otimes g)(X)$, which becomes in turn $\|X\|_{\text{LO}}\geq \|X\|_{\varepsilon}$ once we maximise over the functionals $f,g$ satisfying $\|f\|_{*},\|g\|_{*}\leq 1$.
\end{proof}
\vspace{1ex}
\begin{cor} \label{cor d norms}
Any global base norm $\|\cdot\|$ in a bipartite GPT constructed according to~\eqref{CAB bound} must obey $\|x\otimes y\|=\|x\otimes y\|_{\text{\emph{LO}}}=\|x\|\, \|y\|$ for all $x\in V_A$ and $y\in V_B$, where $\|\cdot\|$ stands for (both) the local base norms.
\end{cor}
\begin{proof}
On the one hand, since~\eqref{CAB bound} holds, the global base norms must be upper bounded by the one associated with the minimal tensor product, which coincides with $\|\cdot\|_\pi$ by Proposition~\ref{d norms inj proj}. On the other hand, the second identity in~\eqref{d norms inj proj eq} ensures that $\|\cdot\|_{\text{LO}}\geq \|\cdot\|_\varepsilon$. We already saw how injective and projective norm coincide on simple tensors. Then, putting all together we obtain
\begin{equation*}
\|x\|\, \|y\|\, =\, \| x \otimes y\|_\varepsilon\, \leq\, \|x\otimes y\|_{\text{LO}}\, \leq\, \|x\otimes y\|\, \leq\, \|x\otimes y\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}\, =\, \|x\otimes y\|_\pi\, =\, \|x\| \, \|y\|\, ,
\end{equation*}
concluding the proof.
\end{proof}
\vspace{2ex}
We are finally ready to prove one of our main results, that is, the quasi-optimality of the lower bound~\eqref{univ dh ratio lower}.
\vspace{2ex}
\begin{thm}[Upper bound on ultimate data hiding ratios] \label{thm univ}
Let $A,B$ be two GPTs, and let their composite $AB$ obey~\eqref{CAB bound}. Then the data hiding ratios against locally constrained sets of measurements satisfy the bound
\begin{equation}
R(\text{\emph{SEP}})\, \leq\, R(\text{\emph{LOCC}})\, \leq\, R(\text{\emph{LOCC}}_\rightarrow)\, \leq\, R(\text{\emph{LO}})\, \leq\, \max_{0\neq X\in V_A\otimes V_B} \frac{\| X \|_\pi}{\| X \|_\varepsilon}\, . \label{thm univ eq1}
\end{equation}
In particular, the corresponding ultimate data hiding ratios, as given in Definition~\ref{univ dh ratio}, are upper bounded as follows:
\begin{equation}
R_{\text{\emph{SEP}}}(d_{A},d_{B})\, \leq\, R_{\text{\emph{LOCC}}}(d_A,d_B)\, \leq\, R_{\text{\emph{LOCC}}_{\rightarrow}}(d_{A},d_{B})\, \leq\, R_{\text{\emph{LO}}}(d_{A},d_{B})\, \leq\, \min\{d_A,d_B\}\, . \label{thm univ eq2}
\end{equation}
\end{thm}
\begin{proof}
Since the inequalities~\eqref{chain dh} hold, we have to upper bound only the data hiding ratio against local operations. Using~\eqref{dh ratio} and Proposition~\ref{min is optimal}, we know that for a fixed pair of GPTs $A=(V_{A},C_{A},u_{A})$ and $B=(V_{B},C_{B},u_{B})$, this latter ratio satisfies $R(\text{LO})\leq \max_{X\neq 0} \|X\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B} \big/ \|X\|_{\text{LO}}$. Now, Proposition~\ref{d norms inj proj} states that $\|X\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont min} \end{array}}{\otimes}$}\! B}=\|X\|_{\pi}$ and $\|X\|_{\text{LO}}^{-1}\leq \|X\|_{\varepsilon}^{-1}$, from which we deduce~\eqref{thm univ eq1}. Using also Proposition~\ref{prop proj=<ninj}, we see that the right-hand side of~\eqref{thm univ eq1} is upper bounded by $\min\{d_A,d_B\}$ for all GPTs of fixed local dimensions $d_A,d_B$, proving also~\eqref{thm univ eq2}.
\end{proof}
\vspace{1ex}
\begin{rem}
Specialising Theorem~\ref{thm univ} to the modified version of quantum mechanics that we called $W$-theory in Example~\ref{ex QM} yields an improvement of~\cite[Lemma 20]{Brandao area law}, in the form of the inequality
\begin{equation}
\|X\|_W\, \leq\, n^2 \|X\|_\varepsilon\, ,
\label{improvement Brandao}
\end{equation}
valid for all Hermitian operators $X$ on $\mathds{C}^{n_A}\otimes\mathds{C}^{n_B}$, where $n=\min\{n_A, n_B\}$, as usual. Observe that the right-hand side of~\eqref{improvement Brandao} coincides with the right-hand side of~\cite[Eq. (C1)]{Brandao area law}, while the left-hand side of~\eqref{improvement Brandao} is larger than the corresponding left-hand side of~\cite[Eq. (C1)]{Brandao area law}, as follows from~\eqref{Cmin largest norm},~\eqref{W norm}, and from the fact that the trace norm $\|\cdot\|_1$ is the base norm corresponding to the standard quantum mechanical composition rule.
Let us stress that neither of the two bounds $R_{QM}(\text{LO})\leq c\, \sqrt{n_A n_B}$~\cite{VV dh} and $R_{QM}(\text{LO})\leq \min\{n_A^2, n_B^2\}$~\cite{Brandao area law} is tight. If $n_A$ and $n_B$ are very different from each other the latter bound will be more effective, while if they are of the same order the former will be preferable. In conclusion, as we already highlighted when discussing data hiding in quantum mechanics, determining the optimal scaling of $R_{QM}(\text{LO})$ remains an interesting open problem. At the same time, this example shows how consequences drawn from general results like Theorem~\ref{thm univ} can shed some light even on well-studied problems.
\end{rem}
\section{Data hiding in special classes of GPTs} \label{sec special}
Until now, we have been mainly interested in investigating the strongest cases of data hiding, in order to study the ultimate, intrinsic bounds characterising this non-classical phenomenon. For how well-motivated this inclination to universality can be, throughout this section we want to take a different approach and look into particular classes of models, chosen as suitable generalisations of Examples~\ref{ex QM} and~\ref{ex sph}.
Subsection~\ref{subsec centr} is concerned with all the GPTs whose state space, just like in the spherical model, is centrally symmetric. It turns out that in this case all the data hiding ratios against locally restricted sets of measurements are equal up to an additive constant, and moreover can be computed as a maximal ratio between a projective and an injective norm. On the one hand, this is reminiscent of the proof of Theorem~\ref{thm univ}, where we upper bounded $R(\text{LO})$ with the maximal projective-injective ratio induced by the local base norms. However, while in that case the only piece of information we could extract was the existence of the upper bound, here computing such a quantity yields the exact data hiding ratio up to additive constants (Theorem~\ref{centr ratio}). On the other hand, we see thanks to Example~\ref{ex sph} how this reduction can simplify the solution of the model. In that case, in fact, we were able to translate the data hiding problem into the computation of the maximal ratio between two matrix norms, and the answer to this latter question was a natural consequence of basic theorems in matrix analysis. We show how this approach can be pushed further, by using the machinery we develop to solve another `natural' model (Example~\ref{cubic}).
Throughout Subsection~\ref{subsec Werner}, we look into those GPTs whose vector space is endowed with a representation of a compact group that maps states to states but is otherwise irreducible on the section $\{x\in V:\, u(x)=0\}$. Although this assumption is quite strong, we show how it encompasses the main physically relevant examples of GPTs, like classical probability theory and quantum theory. With the group integral at hand, we are able to generalise the construction of Werner states and to use them for estimating data hiding in terms of simple geometrical parameters of the model (Theorem~\ref{dh Werner}).
\subsection{Centrally symmetric models} \label{subsec centr}
The solution of the spherical model we gave in Example~\ref{ex sph} through Lemma~\ref{norms sph} and Corollary~\ref{dh sph} was based on two remarkable facts: on the one hand, we could achieve the data hiding ratio (up to an additive constant) by employing only tensors of the form $\hat{M}$ for $M\in \mathds{R}^{(d_{A}-1)\times (d_{B}-1)}$, and on the other hand we were able to find simple expressions for base and separability norm of tensors of this simplified form. Here, we want to take the chance to generalise these intuitions a bit further, to encompass any model whose state space is centrally symmetric. Let us start with the following definition.
\vspace{2ex}
\begin{Def}[Centrally symmetric models] \label{centr}
A GPT of the form $(\mathds{R}^{d}, C, u)$ such that $u=(1,0,\ldots,0)^T$ is said to be \emph{centrally symmetric} if there exists a norm $|\cdot|$ on $\mathds{R}^{d-1}$ such that $C=\{(x_{0},\widebar{x})\in \mathds{R}\oplus\mathds{R}^{d-1}:\ x_{0}\geq |\widebar{x}|\}$.
\end{Def}
\vspace{2ex}
As can be easily verified, the dual cone to $C=\{(x_{0},\widebar{x})\in \mathds{R}\oplus\mathds{R}^{d-1}:\ x_{0}\geq |\widebar{x}|\}$ shares the same structure, being given by
\begin{equation*}
C ^*\, =\, \{(y_{0},\widebar{y})\in \mathds{R}\oplus\mathds{R}^{d-1}:\ y_{0}\geq |\widebar{y}|_*\}\, ,
\end{equation*}
where $|\cdot|_*$ is the dual to the norm $|\cdot|$. The base norm of a centrally symmetric GPT is given by $\left\|(x_0,\widebar{x})\right\|=\max\{|x_0|, |\widebar{x}|\}$. Another peculiarity of centrally symmetric models is the existence of a privileged state, denoted by $u_{*}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = (1,0,\ldots,0)^T$. The vector space can then be written as
\begin{equation}
\mathds{R}^{d}\, =\, \mathds{R} u_{*}\oplus \mathds{R}^{d-1}\, .
\label{V decomp}
\end{equation}
We will keep denoting by $x_0, \widebar{x}$ the two components of $x\in \mathds{R}^d$ according to the above decomposition. Along the same lines, for $v\in \mathds{R}^{d-1}$ we will write $\hat{v}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = 0\oplus v$. A remarkable feature of centrally symmetric models is the existence of a simple linear map $\Lambda:\mathds{R}^{d}\rightarrow \mathds{R}^{d}$, given by
\begin{equation}
\Lambda \,\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, 1 \oplus (-I_{d-1})
\label{Lambda}
\end{equation}
according to the decomposition~\eqref{V decomp} and with $I_{d-1}$ denoting the identity on $\mathds{R}^{d-1}$, that is also an order isomorphism (i.e. such that $\Lambda(C)=C$). Consequently, for all $x\in \mathds{R}^d$ one has $\|x\|=\|\Lambda(x)\|$ (also obvious from the explicit formula for the base norm given above). Now, let us turn to bipartite systems. From~\eqref{V decomp} we infer the natural decomposition
\begin{equation}
V_{AB}\, =\, \mathds{R}^{d_{A}}\otimes \mathds{R}^{d_{B}}\, =\, \left(\mathds{R}\, u_{A*}\otimes u_{B*}\right) \oplus \left(u_{A*}\otimes \mathds{R}^{d_{B}-1}\right) \oplus \left(\mathds{R}^{d_{A}-1}\otimes u_{B*}\right) \oplus \left(\mathds{R}^{d_{A}-1}\otimes \mathds{R}^{d_{B}-1}\right) . \label{VAB decomp}
\end{equation}
As a side remark, observe that the maps $\Lambda\otimes I,\, I\otimes \Lambda,\, \Lambda\otimes \Lambda$ preserve separability of states and (up to an irrelevant transposition) separability of effects. Therefore, the norm $\|\cdot\|_{\text{SEP}}$ is left invariant by any of these maps. As is easy to verify, the same is true for all the four locally constrained distinguishability norms, with the possible exception of $\|\cdot\|_{\text{LOCC}}$.
Concerning the analysis of the data hiding properties, what we did in the case of the spherical model was to consider only the latter of the above addends, and to a posteriori justify this restriction by employing Theorem~\ref{thm univ} to show that the obtained result is tight up to an additive constant. As it turns out, this procedure can be always followed for centrally symmetric GPTs without loss of generality. This is the content of our first result. As throughout Example~\ref{ex sph}, for a tensor $X\in V_{AB}$ we denote by $\widebar{X}$ its component pertaining to the fourth addend of~\eqref{VAB decomp}.
\vspace{2ex}
\begin{prop} \label{prop centr R r}
Let $AB$ be a bipartite system formed by two centrally symmetric GPTs joined with any rule that respects~\eqref{CAB bound}. Then for all $X\in V_{AB}$ we have
\begin{equation}
\frac{\|X\|}{\|X\|_{\text{\emph{LO}}}}\, \leq\, \frac{\big\|\widebar{X}\big\|}{\big\|\widebar{X}\big\|_{\text{\emph{LO}}}} + 2\, . \label{prop centr R r eq}
\end{equation}
\end{prop}
\begin{proof}
Consider $X=s\, u_*\otimes u_* + a\otimes u_* + u_*\otimes b + \widebar{X}\in V_{AB}$, where we omitted the subscripts $A,B$ for the sake of simplicity. Applying the triangle inequality to the global base norm, we see that $\|X\|\leq \| (s\,u_*+a)\otimes u_*\| + \|u_*\otimes b\| + \big\|\widebar{X}\big\|$. Now, Corollary~\ref{cor d norms} guarantees that $\| (s\,u_*+a)\otimes u_*\| = \| (s\,u_*+a)\otimes u_*\|_{\text{LO}}=\|s\,u_*+a\|$ and $\|u_*\otimes b\|=\|u_*\otimes b\|_{\text{LO}}=\|b\|$. Moreover, it is also clear by discarding the system $B$ and performing an arbitrary operation on $A$, that $\|X\|_{\text{LO}}\geq \|s\, u_*+a\|$. Proceeding in an analogous fashion with exchanged subsystems, remembering that $\Lambda$ defined in~\eqref{Lambda} leaves the base norm invariant, and exploiting the triangle inequality, yields
\begin{equation*}
\|X\|_{\text{LO}}\, \geq\, \|s\, u_* + b\|\, =\, \|\Lambda (s\, u_*+b)\|\, =\, \|s\,u_* - b\|\, \geq\, \left\|\, \frac12 (s\, u_* + b) - \frac12 (s\, u_* - b) \, \right\|\, =\, \|b\|\, .
\end{equation*}
Finally, using the readily verified invariance of $\|\cdot\|_{\text{LO}}$ under any of the maps $\Lambda\otimes I,\, I\otimes \Lambda,\, \Lambda\otimes \Lambda$, we obtain
\begin{equation*}
\|X\|_{\text{LO}}\, \geq\, \left\| \,\frac14 X - \frac14 (\Lambda\otimes I)(X) - \frac14 (I\otimes \Lambda)(X) + \frac14 (\Lambda\otimes\Lambda)(X)\, \right\|_{\text{LO}}\, =\, \big\|\widebar{X}\big\|_{\text{LO}}\, .
\end{equation*}
Putting all together, we find
\begin{align*}
\frac{\|X\|}{\|X\|_{\text{LO}}}\, &\leq\, \frac{\| s\,u_*+a\| + \|b\| + \big\|\widebar{X}\big\|}{\|X\|_{\text{LO}}}\, =\, \frac{\| s\,u_*+a\|}{\|X\|_{\text{LO}}}\, +\, \frac{\|b\|}{\|X\|_{\text{LO}}}\, +\, \frac{\big\|\widebar{X}\big\|}{\|X\|_{\text{LO}}}\, \leq\, 2 \, +\, \frac{\big\|\widebar{X}\big\|}{\big\|\widebar{X}\big\|_{\text{LO}}}
\end{align*}
\end{proof}
\vspace{2ex}
What Proposition~\ref{prop centr R r} is telling us is that up to an additive constant we can restrict the search for data hiding against local operations to projected tensors of the form $\widebar{X}$. From now on, we denote by $\widebar{R}(\mathcal{M})$ the {\it restricted data hiding ratio} against a set of measurements $\mathcal{M}$ that is obtained by considering only those tensors. With this notation, Proposition~\ref{prop centr R r} can be cast into the form of the inequality $R(\text{LO})\leq \widebar{R}(\text{LO})+2$. In order to carry out the analysis of restricted ratios, we need an analogue of Lemma~\ref{norms sph}.
\vspace{2ex}
\begin{prop} \label{norms centr}
Let $|\cdot|_\varepsilon,|\cdot|_\pi$ be the injective and projective norm constructed on $\mathds{R}^{(d_A-1)\times (d_B-1)}\simeq \mathds{R}^{d_A-1}\otimes \mathds{R}^{d_B-1}$ out of the local norms $|\cdot|$ on $\mathds{R}^{d_A-1}, \, \mathds{R}^{d_B-1}$. Then for $M\in\mathds{R}^{(d_A-1)\times (d_B-1)}$, whose embedding into $V_{AB}$ according to~\eqref{VAB decomp} we denote by $\hat{M}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = 0\oplus 0\oplus 0\oplus M$, the locally constrained distinguishability norms and the global base norm induced by the minimal tensor product are respectively given by
\begin{align}
&\big\|\hat{M}\big\|_{\text{\emph{LO}}}\, =\, \big\|\hat{M}\big\|_{\text{\emph{LOCC}}_\rightarrow}\, =\, \big\|\hat{M}\big\|_{\text{\emph{LOCC}}}\, =\, \big\|\hat{M}\big\|_{\text{\emph{SEP}}}\, =\, |M|_\varepsilon\, , \label{norms centr eq1} \\
&\big\|\hat{M}\big\|_{A \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! B}\, =\, |M|_\pi\, .
\end{align}
\end{prop}
\begin{proof}
The argument follows the guidelines of the proof of Proposition~\ref{norms sph}, so we omit some of the details. First of all, we notice that if a dual tensor $E\in\mathds{R}^{d_{A}\times d_{B}}$ is separable then necessarily $\big|\widebar{E}\big|_{*\pi} \leq E_{00}$, and this condition is also sufficient when all cross terms $E_{i0}, E_{0j}$ ($i,j\geq 1$) vanish. The proof of this claim follows exactly the steps of~\eqref{norms sph proof 1},~\eqref{norms sph proof 2} and~\eqref{norms sph proof 3}, with the dual norm $|\cdot|_*$ displacing the (self-dual) Euclidean norm, and the associated projective norm $|\cdot|_{*\pi}$ displacing the trace norm.
To compute the separability norm, on the one hand as in~\eqref{norms sph proof 4} and~\eqref{norms sph proof 5} we find
\begin{equation*}
\big\|\hat{M}\big\|_{\text{SEP}}\, \leq\, \max_{|\bar{F}|_{*\pi} + |F_{00}|\, \leq\, 1} \text{Tr}\, \Big[ \widebar{F}^{\,T} M\Big]\, =\, |M|_{*\pi*}=|M|_\varepsilon \, ,
\end{equation*}
where in the last step we used~\eqref{dual inj proj}. On the other hand, the complementary inequality $\big\|\hat{M}\big\|_{\text{LO}}\geq |M|_\varepsilon$ is a simple consequence of the second relation in~\eqref{d norms inj proj eq}. In fact, since for all $v\in\mathds{R}^{d_A-1}$ the identity $|v|_*=\|\hat{v}\|$ holds true, we obtain
\begin{equation*}
\big\|\hat{M}\big\|_{\text{LO}}\, \geq\, \big\|\hat{M}\big\|_\varepsilon\, =\, \max_{\|f\|_*,\|g\|_*\leq 1} (f\otimes g)\big(\hat{M}\big)\, \geq\, \max_{|v|_*,|w|_*\leq 1} v^T M w\, =\, |M|_\varepsilon\, .
\end{equation*}
Putting all together, we obtain
\begin{equation*}
|M|_\varepsilon\, \leq\, \big\|\hat{M}\big\|_{\text{LO}}\, \leq\, \big\|\hat{M}\big\|_{\text{LOCC}_\rightarrow}\, \leq\, \big\|\hat{M}\big\|_{\text{LOCC}} \leq\, \big\|\hat{M}\big\|_{\text{SEP}}\, \leq\, |M|_\varepsilon\, ,
\end{equation*}
from which~\eqref{norms centr eq1} follows.
\end{proof}
\vspace{1ex}
\begin{thm} \label{centr ratio}
For a fixed pair of centrally symmetric GPTs whose composite is formed with the minimal tensor product rule according to Proposition~\ref{min is optimal}, all the four restricted data hiding ratios against locally constrained measurements coincide, and can be computed as
\begin{equation}
\widebar{R}\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \widebar{R}(\text{\emph{LO}})\, =\, \widebar{R}(\text{\emph{LOCC}}_\rightarrow)\, =\, \widebar{R}(\text{\emph{LOCC}})\, =\, \widebar{R}(\text{\emph{SEP}})\, =\, \max_{0\,\neq\, M\,\in\, \mathds{R}^{(d_A-1)\times (d_B-1)}} \frac{\,|M|_\pi}{\,|M|_\varepsilon}\, . \label{centr ratio eq1}
\end{equation}
Moreover, the `true' data hiding ratios are equal to the restricted ones up to an additive constant. Namely,
\begin{equation}
\widebar{R}\, \leq\, R(\text{\emph{SEP}})\, \leq\, R(\text{\emph{LOCC}})\, \leq\, R(\text{\emph{LOCC}}_\rightarrow)\, \leq\, R(\text{\emph{LO}})\, \leq\, \widebar{R} + 2\, . \label{centr ratio eq2}
\end{equation}
\end{thm}
\begin{proof}
First of all,~\eqref{centr ratio eq1} follows directly from Proposition~\ref{norms centr}. To show~\eqref{centr ratio eq2}, note that, by definition, restricted data hiding ratios are smaller than the true ones. In particular, $R(\text{SEP})\geq \widebar{R}(\text{SEP})=\widebar{R}$. Using~\eqref{chain dh} and the result of Proposition~\ref{prop centr R r} in the form of the inequality $R(\text{LO})\leq \widebar{R}(\text{LO})+2$, we obtain immediately~\eqref{centr ratio eq2}.
\end{proof}
\vspace{1ex}
\begin{rem}
Sometimes the quantity $\widebar{R}$ appearing in~\eqref{centr ratio eq1} can be better computed with the help of a simple trick. Namely, remember that for all norms $|\cdot|^{(1)},|\cdot|^{(2)}$ on any space $V$ we have $\max_{0\neq x\in V} \frac{|x|^{(1)}}{|x|^{(2)}} = \max_{0\neq f\in V^*} \frac{|f|^{(2)}_*}{|f|^{(1)}_*}$. Thanks to this identity and to the duality relation~\eqref{dual inj proj}, we see that $\widebar{R}$ can alternatively be expressed as
\begin{equation}
\widebar{R}\, =\, \max_{0\,\neq\, M\,\in\, \mathds{R}^{(d_A-1)\times (d_B-1)}} \frac{\,|M|_{*\varepsilon}}{\,|M|_{*\pi}}\, . \label{Rbar dual}
\end{equation}
\end{rem}
\vspace{2ex}
Theorem~\ref{centr ratio} suggests a precise recipe for computing data hiding ratios of centrally symmetric models. It is in fact enough to analyse the behaviours of injective and projective norms, compute $\widebar{R}$ defined in~\eqref{centr ratio eq1}, and finally use~\eqref{centr ratio eq2} to determine all the ratios against locally constrained measurements up to a universal additive constant. In fact, this is basically what we did in Example~\ref{ex sph}, since trace norm and operator norm are exactly the injective and projective norms constructed out of local Euclidean spaces, as we saw in the proof of Proposition~\ref{prop proj=<ninj}. In the remaining part of this subsection, we are going to show how to apply all this machinery to solve another explicit example.
\vspace{2ex}
\begin{ex}[Cubic model] \label{cubic}
A generalised bit, or {\it gbit} for short~\cite{Barrett original}, is perhaps the simplest example of a non-classical system. As a GPT, we can define it as $G_{2} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \left( \mathds{R}^{3},\, C_{\square},\, u \right)$, where $C_{\square} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \{ (x,y,z)\in\mathds{R}^{3}:\ |x|+|y|\leq z \}$ and $u(x,y,z)\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = z$. As is easily seen, the state space of $G_{2}$ is a two-dimensional square. Thus, a straightforward generalisation of the above concept takes as state space an $n$-dimensional hypercube. For simplicity, in the following we refer to $G_n$ as a {\it cubic model}, a GPT with dimension $\dim G_n=n+1$.
The main reason for introducing these toy models is that they provide a useful tool to understand non-local correlations. In fact, the celebrated Popescu-Rohrlich (PR) box~\cite{PR boxes} can be seen as a maximal tensor product of the form $G_{2}\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize max} \end{array}}{\otimes}$}\! G_{2}$. Here, we are mainly interested in discussing the data hiding properties of GPTs of the form $G_n\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! G_m$, for all positive integers $n,m$. Since the $n$-dimensional hypercube is centrally symmetric, this is a perfect playground for testing the machinery we developed throughout this subsection. Notice that in the case of $G_n$ the norm $|\cdot|$ on $\mathds{R}^n$ appearing in Definition~\ref{centr} is given by the $\infty$-norm $|v|_\infty\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \max_{1\, \leq\, i\, \leq\, n} |v_i|$, with dual $|w|_{\infty*}=|w|_1=\sum_{i=1}^n |w_i|$. It is worth noticing that the extreme points of the unit ball $B_{|\cdot|_1}$ coincide with the elements of the standard basis of $\mathds{R}^n$ up to a sign.
According to Theorem~\ref{centr ratio}, the first step in solving the GPT $G_n\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! G_m$ is the determination of the restricted data hiding ratio $\widebar{R}=\max_{0\neq M\in \mathds{R}^{n\times m}} \frac{|M|_\pi}{|M|_\varepsilon}$, where $|\cdot|_\varepsilon,|\cdot|_\pi$ denote respectively the injective and projective norm induced on $\mathds{R}^{n\times m}$ by the local $\infty$-norms. We already observed how this can be dually rephrased as~\eqref{Rbar dual}. In our case, this is helpful because the injective and projective dual norms $|\cdot|_{*\varepsilon}, |\cdot|_{*\pi}$ can be easily computed as follows.
\begin{align}
|M|_{*\varepsilon}\, &=\, \max_{|v|_\infty,|w|_\infty\leq 1} v^T M w\, =\, \max_{s\in \{\pm 1\}^n,\, t\in \{\pm 1\}^m} s^T M t\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \|M\|_{\infty\rightarrow 1}, \\
|M|_{*\pi}\, &=\, |M|_{\varepsilon*}\, =\, |M|_{\infty *}\, =\, |M|_1\, ,
\end{align}
where in the second line we used
\begin{equation*}
|N|_{\varepsilon}\, =\, \max_{|v|_1, |w|_1\leq 1} v^T N w\, =\, \max_{i,j} |N_{ij}|\, =\, |N|_\infty\, ,
\end{equation*}
so that $|M|_{\varepsilon*}=|M|_1= \sum_{i,j} |M_{ij}|$, where we extended the definitions of $\infty$-norm and $1$-norm to matrices in an obvious way.
Now, we have to find the largest ratio $|M|_1/ \|M\|_{\infty\rightarrow 1}$ that is achievable by $M\in \mathds{R}^{n\times m}$. The argument comprises two parts: first, we have to exhibit an explicit example $M_0$ displaying a high ratio, and secondly, we have to show that this is optimal up to a constant factor.
Let us start with the first task. From now on, we will assume without loss of generality $n\leq m$. Suppose that there exists a matrix $H\in \{\pm 1\}^{n\times m}$ made of signs such that $HH^T=m\mathds{1}_n$ (such matrices are often called {\it partial Hadamard matrices}). Then simple considerations very close in spirit to Lindsey's lemma~\cite{Erdos74,cut norm H} (see also~\cite{A&S,Alon sparse}) show that for all $s\in \{\pm 1\}^n$ and $t\in \{\pm 1\}^m$ we have
\begin{equation*}
s^T H t\, \leq\, \sum_{j} \bigg| \sum_i s_i H_{ij}\bigg|\, \leq\, \sqrt{m}\, \sqrt{\sum_j \bigg| \sum_i s_i H_{ij}\bigg|^2}\, =\, \sqrt{m}\, \sqrt{\sum_{ijk} s_i s_k H_{ij} H_{kj}}\, =\, \sqrt{m}\, \sqrt{\sum_{ik} s_i s_k \, n\delta_{ik}}\, =\, n\sqrt{m}\, .
\end{equation*}
Maximising over sign vectors $s,t$ yields $\|H\|_{\infty\rightarrow 1}\leq n\sqrt{m}$. Since $|H|_1=nm$, we obtain the lower bound $|H|_1 \big/ \|H\|_{\infty\rightarrow 1}\geq \sqrt{n}$. Now, the existence of a matrix $H$ with the properties required for the above construction is not guaranteed for arbitrary integers $n,m$. In fact, this is never the case when $n=m>2$ and $n$ is not a multiple of $4$, while when $n=m=4k$ it is the content of the so-called {\it Hadamard conjecture}. However, observe that Hadamard matrices are elementarily guaranteed to exist for $n=m=2^k$, as the explicit example $\left( \begin{smallmatrix} 1 & 1 \\ -1 & 1 \end{smallmatrix} \right)^{\otimes k}$ shows. Since for all $n$ there is $n'$ power of $2$ such that $\frac{n}{2}\leq n'\leq n$, we get $|H|_1 \big/ \|H\|_{\infty\rightarrow 1}\geq \sqrt{n'}\geq \sqrt{\frac{n}{2}}$. This shows that
\begin{equation}
\max_{0\, \neq\, M\, \in\, \mathds{R}^{n\times m}} \frac{|M|_{*\pi}}{|M|_{*\varepsilon}}\, =\, \max_{0\, \neq\, M\, \in\, \mathds{R}^{n\times m}} \frac{|M|_1}{\|M\|_{\infty\rightarrow 1}}\, \geq\, \sqrt{\frac{n}{2}}\, .
\label{lower Hadamard}
\end{equation}
We note in passing that the crude estimate~\eqref{lower Hadamard} can be improved to $\sqrt{n}$ when $n\leq \frac{m}{2}$ by the same tricks. Further asymptotic refinements can be obtained with the help of sophisticated results such as~\cite{half Hadamard}, but this is beyond the scope of the present paper.
Now, we have to show that the above lower bound is optimal, i.e. that the scaling of the two sides of~\eqref{lower Hadamard} are the same. In order to do so, we resort to a celebrated inequality known as Khintchine inequality~\cite{bound B,best constant 1,best constant 2}, which states that for all $c\in \mathds{R}^n$, if $s_1,\ldots, s_n$ is an i.i.d. sequence with Rademacher distribution, then
\begin{equation}
\mathds{E}_s \bigg|\sum_i s_i c_i \bigg|\, \geq\, \frac{|c|_2}{\sqrt{2}}\, ,
\label{K ineq}
\end{equation}
where $|c|_2^2=\sum_i c_i^2$. Thanks to this result, we know that whenever $M\in\mathds{R}^{n\times m}$ we have~\cite[Corollary 2.4]{Alon sparse}
\begin{align*}
\|M\|_{\infty\rightarrow 1}\, &=\, \max_{s\in \{\pm 1\}^n,\, t\in \{\pm 1\}^m} s^T M t\, =\, \max_{s\in \{\pm 1\}^n} \sum_j \bigg| \sum_i s_i M_{ij} \bigg| \\[0.5ex]
&\geq\ \mathds{E}_s \sum_j \bigg|\sum_i s_i M_{ij}\bigg|\ =\ \sum_j \mathds{E}_s \bigg|\sum_i s_i M_{ij}\bigg|\ \geq\ \frac{1}{\sqrt{2}}\, \sum_j \sqrt{\sum_i M_{ij}^2} \\
&\geq\ \frac{1}{\sqrt{2n}}\,\sum_{ij} |M_{ij}|\ =\ \frac{1}{\sqrt{2n}}\,|M|_1\, .
\end{align*}
This is nothing but the complementary upper bound to~\eqref{lower Hadamard}, and reads
\begin{equation}
\max_{0\, \neq\, M\, \in\, \mathds{R}^{n\times m}} \frac{|M|_{*\pi}}{|M|_{*\varepsilon}}\, =\, \max_{0\, \neq\, M\, \in\, \mathds{R}^{n\times m}} \frac{|M|_1}{\|M\|_{\infty\rightarrow 1}}\, \leq\, \sqrt{2n}\, .
\label{upper Hadamard}
\end{equation}
The two inequalities~\eqref{lower Hadamard} and~\eqref{upper Hadamard} together solve the problem of data hiding in the cubic model. We summarise this solution as follows.
\vspace{2ex}
\begin{prop} \label{dh cubic}
Let the composite of two cubic models $A=G_n$, $B=G_m$ be given by the minimal tensor product, $AB=G_n\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize \emph{min}} \end{array}}{\otimes}$}\! G_m$. Then the restricted data hiding ratio $\widebar{R}_{G_n\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! G_m}$ defined via Proposition~\ref{norms centr} satisfies
\begin{equation}
\sqrt{\frac{\min\{n,m\}}{2}}\, \leq\, \widebar{R}_{G_n\ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! G_m}\, \leq\, \sqrt{2\, \min\{n,m\}}\, .
\end{equation}
Consequently, all the data hiding ratios against locally constrained measurements scale as
\begin{equation}
R_{G_n \ensuremath \! \raisebox{2.1pt}{$\scriptstyle\underset{\begin{array}{c} \vspace{-4.1ex} \\ \text{\fontsize{2}{4}\selectfont \emph{min}} \end{array}}{\otimes}$}\! G_m}(\mathcal{M})\, =\, \Theta\left(\sqrt{\min\{n,m\}}\right)\, =\, \Theta\left(\sqrt{\min\{d_A,d_B\}}\right) , \qquad \text{for $\mathcal{M}=\text{\emph{LO}},\, \text{\emph{LOCC}}_\rightarrow,\, \text{\emph{LOCC}},\, \text{\emph{SEP}}$.}
\end{equation}
\end{prop}
\end{ex}
\subsection{Werner construction for symmetric models} \label{subsec Werner}
A remarkable feature possessed by all the examples of GPTs we have seen so far, including the two that are undoubtedly the most physically relevant, i.e. classical probability theory (Example~\ref{ex class}) and quantum theory (Example~\ref{ex QM}), is the existence of a wide group of symmetries, i.e. linear transformations sending states to states. In the classical case, these are just permutations of the entries of the probability vector, while in the quantum case a symmetry is any conjugation by a unitary matrix. As for the spherical and cubic models (Examples~\ref{ex sph} and~\ref{cubic}), there are the natural actions of $SO(d-1)$ and of signed permutations on the last $d-1$ components of the state vector. We will show how these symmetries can be exploited in order to define a relevant class of bipartite states in a theory made of two copies of the same symmetric GPT. Let us start with the following definition.
\vspace{2ex}
\begin{Def}[Completely symmetric models] \label{compl symm}
A GPT $(V,C,u)$ is said to be \emph{completely symmetric} if there is a compact group $G$ and a representation $\zeta: G\rightarrow \mathcal{L}(V)$ that: (i) is such that $\zeta_g$ sends normalised states to normalised states, for all $g\in G$; and (ii) is irreducible and nontrivial on the preserved subspace $V_0\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \{x\in V:\, u(x)=0\}$.
\end{Def}
We will see at the end of this Subsection that all the examples of GPTs we have considered so far (namely, classical and quantum theory, and spherical and cubic model) are in fact completely symmetric. However, for now we are interested in keeping the reasoning as abstract and general as possible.
Thus, consider a completely symmetric GPT as in Definition~\ref{compl symm}. We start by listing some elementary consequences of the existence of a group symmetry. Let us remind the reader that if $\zeta:G\rightarrow \mathcal{L}(V)$ is a representation, then the dual space $V^*$ is naturally endowed with the dual representation $\zeta^*:G\rightarrow \mathcal{L}(V^*)$ given by $\zeta^*_g=\zeta_{g^{-1}}^T$. With this notation, the fact that $V_0$ is preserved by the action of $G$ can be written as $\zeta_g^*u=u$ for all $g\in G$. In other words, $\mathds{R}u$ is a trivial component of $V^*$ under the action of $G$ through $\zeta^*$. Since it is well-known that if $V$ is real and $G$ is compact then $\zeta$ and $\zeta^*$ are isomorphic, there must necessarily exist also a $G$-invariant vector $u_*\in V$. We deduce from the fact that $V_0$ is irreducible and nontrivial that $u_*\notin V_0$, so that we are free to rescale it in such a way that $u(u_*)=1$. Note that this fixes the decomposition of $V$ into $G$-irreducible representations as
\begin{equation}
V\, =\, \mathds{R}u_*\oplus V_0\, , \label{decomp V}
\end{equation}
where the two pieces are non-isomorphic. Now, we proceed to show that $u_*$ is in fact a state, as implied by property (i) in Definition~\ref{compl symm}. In order to do so, we use the Haar integral on $G$, denoted by $\int_G dg$ and whose existence is guaranteed by the compactness of $G$. For every state $\omega\in\Omega$, we have that $u_*=\int_G dg\, \zeta_g \omega$, because the right-hand side is $G$-invariant and normalised and therefore must coincide with $u_*$. This expression for $u_*$ as a positive combination of states $\zeta_g \omega$ reveals that $u_*$ is itself a state.
As is easy to see by referring to~\eqref{decomp V}, the dual vector space $V^*$ decomposes as
\begin{equation}
V^*\, =\, \mathds{R}u\oplus V_0^* \label{decomp V*}
\end{equation}
under the action of $G$, where the first addend is a trivial representation and the second one is $G$-isomorphic to $V_0$. Therefore, any $G$-isomorphism $\chi:V^*\rightarrow V$ must map $u$ into a multiple of $u_*$ and $V_0^*$ into $V_0$.
If we think of $\chi$ as a tensor in $V\otimes V$, we can write $\chi\,=\,\alpha\, u_*\!\otimes u_* + \beta\, \mathcal{E} = \alpha U_* +\beta \mathcal{E}$, where we used the shorthand $U_*\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = u_*\otimes u_*$, and $\mathcal{E} \in V_0\otimes V_0$ is (canonically identified with) a fixed $G$-isomorphism $V_0^*\rightarrow V_0$. Depending on the representation we choose, we can alternatively write $(\zeta_g\otimes \zeta_g)(\mathcal{E})=\mathcal{E}$ for all $g\in G$ or $\mathcal{E} \zeta_g^*=\zeta_g \mathcal{E} $ for all $g\in G$. From this latter expression we see that $\mathcal{E}_*\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = (\mathcal{E}^{-1})^T:V\rightarrow V^*$ is also a $G$-isomorphism. As for the tensor $\mathcal{E}_*\in V_0^*\otimes V_0^*$, we will rephrase this as $(\zeta_g^*\otimes \zeta_g^*)(\mathcal{E}_*)=\mathcal{E}_*$ for all $g\in G$.
It is perhaps convenient to think of $\mathcal{E}$ and $\mathcal{E}_*$ also as scalar products $\braket{\cdot,\cdot}_\mathcal{E}$, $\braket{\cdot,\cdot}_{\mathcal{E}_*}$ on $V_0^*$ and $V_0$, respectively. This can be done via the definitions $\braket{f,g}_\mathcal{E} \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = (f\otimes g)(\mathcal{E})$ and $\braket{v,w}_{\mathcal{E}_*}\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = \mathcal{E}_*(v\otimes w)$. It is well known that it is possible to choose $\mathcal{E}$ such that both these scalar product are positive definite. We will always make this assumption throughout the rest of this Section. If $\{f_i\}_{i=1}^{d-1}$ is an orthonormal basis for $\braket{\cdot,\cdot}_\mathcal{E}$, we will have
\begin{equation}
\mathcal{E}\, =\, \sum_{i=1}^{d-1} f_i^*\otimes f_i^*\, \in\, V_0\otimes V_0\, ,\qquad \mathcal{E}_*\,=\, \sum_{i=1}^{d-1} f_i\otimes f_i\,\in\, V_0^*\otimes V_0^*\, ,
\label{varepsilon expl}
\end{equation}
with $\{f_i^*\}_i$ being the dual basis to $\{f_i\}_i$. As a simple consequence, we see that $\mathcal{E}_*(\mathcal{E})=d-1$, and the two norms induced by the above scalar products are dual to each other. Moreover, observe that the completion $\{u,f_1,\ldots,f_d\}$ is an orthonormal basis for a global $G$-invariant scalar product. We are now ready to give the following definition.
\vspace{2ex}
\begin{Def}[Werner states] \label{Werner def}
For a composite $AB$ made of two copies $A,B$ of the same completely symmetric GPT $(V,C,u)$ and such that~\eqref{CAB bound} is obeyed, a \emph{Werner state} is a normalised state in $V\otimes V$ that corresponds to a $G$-isomorphism $V^*\rightarrow V$.
\end{Def}
\vspace{2ex}
With the language developed throughout the above discussion, we can express a generic Werner state as
\begin{equation}
\chi_\varphi\, =\, U_* + \varphi\, \mathcal{E}\, ,
\label{Werner}
\end{equation}
where $\varphi$ is a real parameter whose range depends on specific features of the model. For any bipartite cone $C_{AB}$ satisfying~\eqref{CAB bound}, let us define
\begin{equation}
k_\pm\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{ k:\, \chi_{\pm k}\in C_{AB} \right\}\, ,
\label{k+-}
\end{equation}
so that the allowed range of $\varphi$ in~\eqref{Werner} will be $[-k_-,k_+]$. Besides the obvious observation that $k_\pm\geq 0$, there seems to be nothing we can say a priori about these parameters. In order to proceed further, we need a little lemma.
\vspace{2ex}
\begin{lemma} \label{lemma G projector}
Let $(V,C,u)$ be a completely symmetric GPT with Werner states given by~\eqref{Werner}. If $\mathcal{E}, \mathcal{E}_*$ are given by~\eqref{varepsilon expl}, the identity
\begin{equation}
\int_G dg\ \zeta_g\otimes \zeta_g\, =\, U_* U\, +\, \frac{1}{d-1}\ \mathcal{E}\, \mathcal{E}_*
\label{G projector}
\end{equation}
between linear operators on $V\otimes V$ holds true. Here, $U\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = u\otimes u$, $U_*\mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} = u_*\otimes u_*$, and $u_*u:V\rightarrow V$ acts as $(u_* u)(x)=u(x)\, u_*$ for all $x\in V$.
\end{lemma}
\begin{proof}
For an arbitrary $X\in V\otimes V$, the properties of the Haar measure guarantee that $\int_G dg\, \zeta_g\otimes \zeta_g (X)$ is a $G$-invariant tensor, and thus can be expressed as $\alpha\, U_* + \beta\, \mathcal{E}$ for some $\alpha,\beta\in \mathds{R}$. Applying $U$ on both sides we see that
\begin{equation*}
\alpha\, =\, \int_G dg\, U\left(\zeta_g\otimes \zeta_g (X)\right)\, =\, \int_G dg \left(\zeta_g^* \otimes \zeta_g^*\, U\right)(X)\, =\, \int_G dg\, U(X)\, =\, U(X)\, .
\end{equation*}
Moreover, it is easy to verify that since $\mathcal{E}_*(\mathcal{E})=d-1$ and $\left(\zeta_g^*\otimes \zeta_g^*\right)(\mathcal{E}_*)=\mathcal{E}_*$ for all $g\in G$, the analogous equality $(d-1)\beta=\mathcal{E}_*(X)$ holds, thus completing the proof of~\eqref{G projector}.
\end{proof}
\vspace{2ex}
With Lemma~\ref{lemma G projector} in our hands, we can say a bit more about the constants $k_\pm$ introduced in~\eqref{k+-}. Namely, we can determine lower bounds that correspond to the {\it separability region} for the family of Werner states~\eqref{Werner}. As expected, these lower bounds will depend only on the local structure of the model, not on the particular choice of the composite cone, which will instead affect the values of $k_\pm$.
\vspace{2ex}
\begin{prop} \label{prop sep Werner}
For a completely symmetric GPT $(V,C,u)$ of dimension $d$ and a scalar product $\braket{\cdot,\cdot}_{\mathcal{E}_*}$ on the corresponding $V_0$, introduce the quantities
\begin{equation}
m_+ \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{\braket{v,v}_{\mathcal{E}_*}:\ v\in V_0,\, u_*+v\in C\right\}, \qquad m_- \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{ \braket{v,w}_{\mathcal{E}_*}:\ v,w\in V_0,\, u_*+v, u_*-w\in C\right\}\, .
\label{m+-}
\end{equation}
Then the constants $k_\pm$ defined through~\eqref{k+-} satisfy $k_\pm\geq \frac{m_\pm}{d-1}$, and a Werner state $\chi_\varphi$ as given in~\eqref{Werner} is separable if and only if $-\frac{m_-}{d-1}\leq \varphi\leq \frac{m_+}{d-1}$.
\end{prop}
\begin{proof}
Take $v\in V_0$ such that $u_*+v\in C$ and $\braket{v,v}_{\mathcal{E}_*}=m_+$. Compute
\begin{align*}
\int_G dg\ (\zeta_g\otimes \zeta_g)\left((u_*+v)\otimes (u_*+v)\right)\, &= \left( U_* U\, +\, \frac{1}{d-1}\ \mathcal{E}\, \mathcal{E}_* \right)\left((u_*+v)\otimes (u_*+v)\right) \\
&=\, U_*\, +\, \frac{\mathcal{E}_*(v\otimes v)}{d-1}\ \mathcal{E} =\, U_*\, +\, \frac{\braket{v,v}_{\mathcal{E}_*}}{d-1}\ \mathcal{E}\, =\, U_*\, +\, \frac{m_+}{d-1}\ \mathcal{E}\, .
\end{align*}
From the leftmost side we see that this is an allowed separable state of the bipartite system. Looking at the rightmost side and comparing it to~\eqref{Werner}, we see that $k_+\geq \frac{m_+}{d-1}$ and that $\chi_{m_+/(d-1)}\in C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B$. An analogous construction shows that we can find $v,w\in V_0$ such that $u_*+v, u_*-w\in C$ and
\begin{equation*}
\int_G dg\ \zeta_g\otimes \zeta_g \left((u_*+v)\otimes (u_*-w)\right)\, =\, U_*\, -\, \frac{m_-}{d-1}\ \mathcal{E}\, \in\, C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B\, ,
\end{equation*}
so that $k_-\geq \frac{m_-}{d-1}$ and $\chi_{-k_-}\in C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B$.
In order to show that $\chi_\varphi\in C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B$ if and only if $\varphi\in \left[- \frac{m_-}{d-1},\, \frac{m_+}{d-1}\right]$, we start by proving that $\omega\in C_A\ensuremath \! \raisebox{2pt}{$\underset{\begin{array}{c} \vspace{-3.7ex} \\ \text{\scriptsize min} \end{array}}{\otimes}$}\! C_B$ implies $-m_-\leq \mathcal{E}_*(\omega)\leq m_+$. In fact, if $\omega=\sum_i p_i\, (u_*+v_i)\otimes (u_*+w_i)$ we obtain $\mathcal{E}_*(\omega)\, =\, \sum_i p_i\, \braket{v_i,w_i}_{\mathcal{E}_*}$, and the claim follows from the inequalities
\begin{equation*}
\braket{v_i,w_i}_{\mathcal{E}_*}\, =\, - \braket{v_i,-w_i}_{\mathcal{E}_*}\, \geq\, -m_-\, ,
\end{equation*}
valid because $u_*+v_i,u_*-(-w_i)\in C$ and thus $\braket{v_i,-w_i}_{\mathcal{E}_*}\leq k_-$, and
\begin{equation*}
\braket{v_i,w_i}_{\mathcal{E}_*}\, \leq\, \sqrt{\braket{v_i,v_i}_{\mathcal{E}_*} \braket{w_i,w_i}_{\mathcal{E}_*}}\, \leq\, m_+\, .
\end{equation*}
Applying this to a Werner state $\chi_\varphi$ we obtain that a necessary condition for separability is $-m_-\leq \mathcal{E}_*(\chi_\varphi)=(d-1)\varphi\leq m_+$, concluding the proof.
\end{proof}
\vspace{2ex}
In order to study the data hiding properties of the family of Werner states, we need to understand the separability conditions at the dual level. Luckily enough, there is no need to repeat any calculation. This is because if $(V,C,u)$ is completely symmetric then $(V^*,C^*,u_*)$ is itself a completely symmetric GPT. Therefore, we start by giving the following `dual' definitions:
\begin{equation}
k_\pm^*\, \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{ k:\, U\pm k \mathcal{E}_*\in C^*_{AB} \right\}\, ,
\label{k+-*}
\end{equation}
\vspace{-4.5ex}
\begin{equation}
m_+^* \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{\braket{f,f}_{\mathcal{E}}:\ f\in V_0^*,\, u+f\in C^*\right\}, \qquad m_-^* \mathrel{\vcenter{\baselineskip0.5ex \lineskiplimit0pt \hbox{\scriptsize.}\hbox{\scriptsize.}}} =\, \max\left\{ \braket{f,g}_{\mathcal{E}}:\ f,g\in V_0^*,\, u+f, u-g\in C^*\right\}\, .
\label{m+-*}
\end{equation}
Exactly as in Proposition~\ref{prop sep Werner}, we obtain
\begin{equation}
k_\pm^*\, \geq\, \frac{m_\pm^*}{d-1}\, .
\label{k+-* lower bound}
\end{equation}
Moreover, one can prove the following.
\vspace{2ex}
\begin{lemma}
In a completely symmetric GPT $(V,C,u)$, the constants $k_\pm,k_\pm^*,m_\pm,m_\pm^*$ defined by~\eqref{k+-},~\eqref{k+-*},~\eqref{m+-},~\eqref{m+-*}, respectively, satisfy $m_-\leq m_+$, $m_-^*\leq m_+^*$, and moreover
\begin{equation}
k_\pm^*\,k_\mp\, =\, \frac{1}{d-1}\, ,\qquad 1\,\leq\, m_\pm m_\mp^*\, \leq\, d-1\, .
\label{k m dual}
\end{equation}
\end{lemma}
\begin{proof}
The inequalities $m_-\leq m_+$ and $m_-^*\leq m_+^*$ follow trivially from the definition by applying once Cauchy-Schwartz inequality. Thus, let us proceed to show the relation between $k_\pm$ and $k_\pm^*$. By definition of dual cone, $U\pm k \mathcal{E}_*\in C^*_{AB}$ if and only if $(U+ k \mathcal{E}_*)(\omega)\geq 0$ for all $\omega \in C_{AB}$, which can be assumed to be normalised without loss of generality. However, because of the group symmetry it is enough to test Werner states. In fact, since $(\zeta_g^*\otimes \zeta_g^*)(U+ k \mathcal{E}_*)=U+ k \mathcal{E}_*$ for all $g\in G$, we obtain
\begin{equation*}
(U+ k \mathcal{E}_*)(\omega)\, =\, \int_G dg\, \left((\zeta_g^*\otimes \zeta_g^*)(U+ k \mathcal{E}_*)\right) (\omega)\, =\, \int_G dg\ (U+ k \mathcal{E}_*) \left((\zeta_g\otimes \zeta_g) (\omega) \right)\, =\, (U+ k \mathcal{E}_*) (\chi_\varphi)\, =\, 1 + (d-1) k \varphi\, ,
\end{equation*}
where $(d-1)\varphi=\mathcal{E}_*(\omega)$. The rightmost side of the above equation is positive for all $\varphi\in [-k_-,k_+]$ if and only if $-\frac{1}{(d-1) k_+}\leq k\leq \frac{1}{(d-1) k_-}$.
Now, let us devote our attention to the bounds on the $m$ quantities in~\eqref{k m dual}. Clearly, it is enough to show that $1\leq m_+ m_-^*\leq d-1$, up to exchanging primal and dual GPT. Consider $f\in V_0^*$ such that $\braket{f,f}_{\mathcal{E}}=1$. Then it is easy to see that $u-\frac{f}{\sqrt{m_+}}\in C^*$, since for all states $\omega\in \Omega$, which can be conveniently parametrised as $\omega=u_*+v$ with $v\in V_0$ satisfying $\braket{v,v}_{\mathcal{E}_*}\leq m_+$, one has
\begin{equation*}
\Big(u \pm \frac{f}{\sqrt{m_+}}\Big)(\omega)\, =\, 1 \pm \frac{1}{\sqrt{m_+}}\, f(v)\, \geq\, 1 - \frac{1}{\sqrt{m_+}}\, \sqrt{\braket{f,f}_{\mathcal{E}} \braket{v,v}_{\mathcal{E}_*}}\, \geq\, 1 - \sqrt{\braket{f,f}_{\mathcal{E}}}\, =\, 0\, .
\end{equation*}
Since $u\, \pm \frac{f}{\sqrt{m_+}}\in C^*$, from the definition~\eqref{m+-*} we infer that $m_-^*\geq \braket{\frac{f}{\sqrt{m_+}},\frac{f}{\sqrt{m_+}}}_{\mathcal{E}}=\frac{1}{m_+}$. In order to prove the upper bound $m_+m_-^*\leq d-1$, let us write
\begin{equation*}
m_+\, \leq\, (d-1) k_+\, =\, \frac{1}{k_-^*}\, \leq\, \frac{d-1}{m_-^*}\, ,
\end{equation*}
where we employed in order: the results of Proposition~\ref{prop sep Werner}, the first relation in~\eqref{k m dual} and finally~\eqref{k+-* lower bound}.
\end{proof}
\vspace{2ex}
With the tools we have developed so far, we are ready to discuss quantitatively the existence of allowed and separable measurements displaying Werner symmetry.
\vspace{2ex}
\begin{prop} \label{prop dual Werner}
In a completely symmetric GPT $(V,C,u)$, for $\alpha,\beta\in\mathds{R}$ the functionals $\left( \alpha U + \beta \mathcal{E}_*,\, (1-\alpha) U - \beta \mathcal{E}_* \right)$ form an allowed measurement if and only if
\begin{equation}
- k_-^*\alpha\, \leq\, \beta\, \leq\, k_+\alpha\, ,\qquad -k_+^*(1-\alpha)\,\leq\, \beta\, \leq\, k_-^*(1-\alpha)\, .
\label{dual Werner}
\end{equation}
In the $(\alpha, \beta)$-plane, these two conditions identify a parallelogram with vertices
\begin{equation}
(0,0),\, (1,0),\ \left(\frac{k_-^*}{k_+^* + k_-^*},\ \frac{k_+^* k_-^*}{k_+^* + k_-^*} \right),\ \left(\frac{k_+^*}{k_+^* + k_-^*},\, -\frac{k_+^* k_-^*}{k_+^* + k_-^*} \right) .
\end{equation}
The separability conditions for the measurement under examination can be deduced from~\eqref{dual Werner} by making the substitutions $k_\pm^*\mapsto \frac{m_\pm^*}{d-1}$.
\end{prop}
\begin{proof}
We omit the details, since the proof consists in a systematic application of the definitions~\eqref{k+-*}, together with the dual conditions to those already given in Proposition~\ref{prop sep Werner}.
\end{proof}
\vspace{2ex}
Equipped with Proposition~\ref{prop dual Werner}, we are ready to explore the data hiding properties of Werner states. Since we did not make any general claim about anything but separability, we will compute the highest data hiding efficiency against separable measurements that is obtainable by using Werner states in Definition~\ref{dh}.
\vspace{2ex}
\begin{thm}[Data hiding with Werner states] \label{dh Werner}
For a composite system made of two copies of the same completely symmetric GPT, the highest data hiding efficiency against separable measurements that is achievable with only Werner states is given by
\begin{equation}
R_{\text{\emph{Werner}}}(\text{\emph{SEP}})\, =\, \max_{0\neq (a,b)\in\mathds{R}^2}\, \frac{\| a U_* + b \mathcal{E} \|}{\,\| a U_* + b \mathcal{E} \|_{\text{\emph{SEP}}}}\ =\ 1\, +\, \frac{2}{k_+ + k_-}\, \max\left\{\frac{1}{m_+^*} - k_-,\, \frac{1}{m_-^*} - k_+ \right\} .
\label{dh Werner eq}
\end{equation}
\end{thm}
\begin{proof}
Let us start by computing the base norm $\|a U_*+ b \mathcal{E}\|$, for $a,b\in \mathds{R}$. Thanks to Proposition~\ref{prop dual Werner}, we have to test just one nontrivial measurement, namely
\begin{equation*}
\left( \frac{k_-^*}{k_+^* + k_-^*}\, U + \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, \mathcal{E},\ \frac{k_+^*}{k_+^* + k_-^*}\, U - \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, \mathcal{E} \right) .
\end{equation*}
Thus, using~\eqref{d norm} we find
\begin{align}
\| a U_*+ b \mathcal{E} \|\, &=\, \left| \left( \frac{k_-^*}{k_+^* + k_-^*}\, U + \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, \mathcal{E}\right) \left( a U_* + b\mathcal{E}\right)\right| \, +\, \left|\left( \frac{k_+^*}{k_+^* + k_-^*}\, U - \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, \mathcal{E}\right) \left( a U_* + b \mathcal{E}\right)\right|\, = \label{base Werner 1} \\[1ex]
&=\, \left| \frac{k_-^*}{k_+^* + k_-^*}\, a + \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, (d-1) b\right|\, +\, \left| \frac{k_+^*}{k_+^* + k_-^*}\, a - \frac{k_+^* k_-^*}{k_+^* + k_-^*}\, (d-1) b\right|\, = \label{base Werner 2}\\[1ex]
&=\, \max\left\{ |a|,\ \frac{| 2(d-1)\,b\, k_+^* k_-^* + a (k_-^* - k_+^*) |}{k_+^* + k_-^*} \right\} , \label{base Werner 3}
\end{align}
where we used the elementary formula $|x+y|+|z-y|=\max\{ |x+z|,\, |2y+x-z| \}$ in the last step. As suggested by Proposition~\ref{prop dual Werner}, we can obtain the separability norm by replacing everywhere $k_\pm^*$ with $\frac{m_\pm^*}{d-1}$:
\begin{equation}
\| a U_*+ b \mathcal{E} \|_{\text{SEP}}\, =\, \max\left\{ |a|,\ \frac{| 2\,b\, m_+^* m_-^* + a (m_-^* - m_+^*) |}{m_+^* + m_-^*} \right\}\, . \label{sep Werner}
\end{equation}
Observe that plugging~\eqref{k m dual} into~\eqref{base Werner 3} yields the somehow more handy expression
\begin{equation}
\| a U_*+ b \mathcal{E} \|\, =\, \max\left\{ |a|,\ \frac{| 2\,b\, + a (k_- - k_+) |}{k_+ + k_-} \right\} .
\label{base Werner}
\end{equation}
Now, computing the ratio $\max_{0\neq (a,b)\in\mathds{R}^2}\, \frac{\| a U_* + b \mathcal{E} \|}{\,\| a U_* + b \mathcal{E} \|_{\text{SEP}}}$ amounts to minimising~\eqref{sep Werner} for a fixed value of~\eqref{base Werner}. It is very easy to see that such a minimum is always achieved for pairs $(a,b)$ such that the two expressions in the maximum appearing in~\eqref{sep Werner} coincide. Up to scalar multiples, there are exactly two such pairs, namely those satisfying $b=\pm \frac{a}{m_\mp^*}$. By substituting these values into~\eqref{base Werner} and dividing by~\eqref{sep Werner}, we find
\begin{equation}
R_{\text{Werner}}(\text{SEP})\, =\, \max_{0\neq (a,b)\in\mathds{R}^2}\, \frac{\| a U_* + b \mathcal{E} \|}{\,\| a U_* + b \mathcal{E} \|_{\text{SEP}}}\, =\, \max\left\{ 1,\ \frac{\big| \frac{2}{m_-^*}\, + k_- - k_+ \big|}{k_+ + k_-},\ \frac{\big| -\frac{2}{m_+^*}\, + k_- - k_+ \big|}{k_+ + k_-} \right\} . \label{dh Werner 1}
\end{equation}
Now, using~\eqref{k m dual} it is easy to see that
\begin{equation*}
\frac{2}{m_\pm^*} \pm (k_+-k_-)\, \geq\, \frac{2}{(d-1) k_\pm^*} \pm (k_+-k_-) \, =\, 2 k_\mp \pm (k_+-k_-)\, =\, k_+ + k_-\, .
\end{equation*}
Thanks to the above inequalities, we can further simplify~\eqref{dh Werner 1} to
\begin{align*}
R_{\text{Werner}}(\text{SEP})\, &=\, \max\left\{ \frac{\big| \frac{2}{m_-^*}\, + k_- - k_+ \big|}{k_+ + k_-},\ \frac{\big| -\frac{2}{m_+^*}\, + k_- - k_+ \big|}{k_+ + k_-} \right\}\, =\\
&=\, \frac{1}{k_+ + k_-}\, \max\left\{ \frac{2}{m_+^*} + (k_+-k_-),\ \frac{2}{m_-^*} - (k_+-k_-) \right\}\, =\\
&=\, 1\, +\, \frac{2}{k_+ + k_-}\, \max \left\{ \frac{1}{m_+^*} - k_-,\ \frac{1}{m_-^*} - k_+ \right\} ,
\end{align*}
finally proving~\eqref{dh Werner eq}.
\end{proof}
\vspace{2ex}
The above result shows how the maximal data hiding ratio obtainable by employing only Werner states depends on just few geometric parameters characterising the model under examination. The usefulness of this theorem rests on its immediate applicability to several natural classes of highly symmetric GPTs. Since computing the relevant parameters is often a simple and intuitive task, as we shall see in a moment,~\eqref{dh Werner eq} gives a quick lower bound on all the data hiding ratios against locally constrained sets of measurements.
To demonstrate the power of Theorem~\ref{dh Werner}, we apply it to the symmetric models that we have examined so far: classical probability theory, quantum mechanics, spherical model, and cubic model.
\begin{itemize}
\item {\it Classical probability theory} (Example~\ref{ex class}). By looking at the definition~\eqref{classical}, it is easy to see that the relevant group here is the symmetric group $S_d$. We have $u=(1,\ldots,1)^T=d u_*$, and the action on the subspace $V_0=\{x\in\mathds{R}^d:\ \sum_i x_i=0\}$ is well-known to be irreducible. This can be either proved with elementary tools or shown in one line via character theory and Burnside's formula. In fact, denote with $\eta:S_d\rightarrow \mathds{R}^{d\times d}$ the standard representation of $S_d$ on $\mathds{R}^d$. Since we can exhibit an explicit one-dimensional irrep (vectors of constant entries), it is enough to show that $\frac{1}{d!} \sum_{\pi\in S_d} \left( \text{Tr}\, \eta(\pi)\right)^2 = \langle \chi_\eta, \chi_\eta\rangle=2$. Observing that $\text{Tr}\, \eta(\pi)$ is the number of elements fixed by $\pi$ and applying Burnside's formula to the natural action of $S_d$ on $\{1,\ldots, d\}^{\times 2}$, which has exactly $2$ orbits, we find $\frac{1}{d!} \sum_{\pi\in S_d} \left( \text{Tr}\, \eta(\pi)\right)^2=2$.
Going back to the analysis of classical probability theory as a completely symmetric model, and exploiting the identification between elements in the tensor product and $d\times d$ real matrices we can also write $U=uu^T= d^2 U_*$ and $\mathcal{E}=\mathds{1}-\frac{1}{d}uu^T=\mathcal{E}_*$. Furthermore, observe that since the upper and lower bound in~\eqref{CAB bound} coincide, the bipartite cone is composed of all the entrywise positive matrices, collectively denoted by $\mathds{R}^{d\times d}_+$. Computing the relevant quantities is now elementary:
\begin{align*}
m_+^* &=\, \max\left\{v^Tv:\ u+v\in\mathds{R}^d_+,\ \sum\nolimits_i v_i=0 \right\}\, =\, d(d-1)\, =\, d^2 m_+\, ,\\
m_-^* &=\, \max\left\{v^Tw:\ u+v,\, u-w\in\mathds{R}^d_+,\ \sum\nolimits_i v_i=\sum\nolimits_j w_j = 0 \right\}\, =\, d\, =\, d^2 m_-\, ,\\
k_+ &=\, \max\left\{ k:\ u_*u_*^T+ k \left(\mathds{1} -d u_*u_*^T\right)\in \mathds{R}^{d\times d}_+\right\}\, =\, \frac{1}{d}\, ,\\
k_- &=\, \max\left\{ k:\ u_*u_*^T- k \left(\mathds{1} -d u_*u_*^T\right)\in \mathds{R}^{d\times d}_+\right\}\, =\, \frac{1}{d(d-1)}\, ,
\end{align*}
Applying~\eqref{dh Werner eq}, we find $R_{\text{Werner}}^{\text{Cl}}(\text{SEP})=1$, which is expected since the collapse of the hierarchy~\eqref{CAB bound} when either of the two cones is simplicial forbids the existence of data hiding altogether.
\item {\it Quantum mechanics} (Example~\ref{ex QM}). We refer to~\eqref{quantum} for the notation. The symmetry group in quantum mechanics is $U(n)$, the group of unitary $n\times n$ matrices. It is easy to see that we can choose $u=\mathds{1}=n\, u_*$ and that the orthogonal complement to this trivial representation (that is, the space of traceless hermitian matrices) is irreducible. In fact, this follows already from the same result for the classical case, since permutation matrices are also unitaries.
For a bipartite system we see that $U=\mathds{1}=n^2\, U_*$ and $\mathcal{E}=F-\frac{\mathds{1}}{n}=\mathcal{E}_*$, where $F\ket{\alpha\beta}=\ket{\beta\alpha}$ denotes the flip operator, as usual. If the composite is assembled according to the standard quantum mechanical rule~\eqref{cone bipartite quantum}, we find
\begin{align*}
m_+^* &=\, \max\left\{\text{Tr}\, X^2:\ \mathds{1}+X\geq 0,\ \text{Tr}\, X=0 \right\}\, =\, n(n-1)\, =\, n^2 m_+\, ,\\
m_-^* &=\, \max\left\{\text{Tr}\, XY:\ \mathds{1}+X,\, \mathds{1}-Y\geq 0,\ \text{Tr}\, X=\text{Tr}\, Y=0 \right\}\, =\, n\, =\, n^2 m_-\, ,\\
k_+ &=\, \max\left\{ k:\ \mathds{1}/ n^2+ k \left( F-\mathds{1}/n \right)\geq 0 \right\}\, =\, \frac{1}{n(n+1)}\, ,\\
k_- &=\, \max\left\{ k:\ \mathds{1}/ n^2 - k \left( F-\mathds{1}/n \right)\geq 0 \right\}\, =\, \frac{1}{n(n-1)}\, ,
\end{align*}
from which it follows easily $R_{\text{Werner}}^{\text{QM}}(\text{SEP})=n$. According to Theorem~\ref{dh QM}, this is even the optimal data hiding ratio against all separable measurements.
If the composite is formed with the minimal tensor product rule (what we called `$W$-theory' in Section~\ref{sec ex}), we write instead
\begin{equation*}
k_+ =\, \frac{m_+}{n^2-1}\, =\, \frac{1}{n(n+1)}\, ,\qquad k_- =\, \frac{m_-}{n^2-1}\, =\, \frac{1}{n(n^2-1)}\, ,
\end{equation*}
and thus $R_{\text{Werner}}^{\text{W}}(\text{SEP})=2n-1$, in accordance with Proposition~\ref{dh W}.
\item {\it Spherical model} (Example~\ref{ex sph}). Using the same notation as in~\eqref{spherical}, we see that in this case the relevant group is $O(d-1)$, the set of $(d-1)\times (d-1)$ orthogonal matrices acting on the last $d-1$ components of $\mathds{R}^d$. We can choose $u=(1,0,\ldots,0)^T=u_*$, and again the orthogonal complement is irreducible as follows from the same result for classical probability theory. For a bipartite system $U=uu^T=U_*$ and $\mathcal{E}=\hat{\mathds{1}}_{d-1}=\mathcal{E}_*$. If the composite is formed with the minimal tensor product rule, then it is very easy to verify that
\begin{equation*}
m_\pm^*=\, \max\{ v^Tv:\ u+v\in C_d\}\, =\, 1\, =\, m_\pm\, , \qquad k_\pm\, =\, \frac{m_\pm}{d-1}\, =\, \frac{1}{d-1}\, ,
\end{equation*}
so that $R_{\text{Werner}}^{\text{Sph}}(\text{SEP})=d-1$, which coincides with the lower bound for the data hiding ratio against separable measurements given in Corollary~\ref{dh sph}.
\item {\it Cubic model} (Example~\ref{cubic}). We follow the notation previously established. The symmetry group of the cubic model in $d$ dimensions is simply the symmetry group of the $(d-1)$-dimensional hypercube. As for the spherical model, we have $u=(1,0,\ldots,0)^T=u_*$, $U=uu^T=U_*$ and $u^\perp$ is irreducible. Furthermore, we can choose $\mathcal{E}=\hat{\mathds{1}}_{d-1}=\mathcal{E}_*$. However, even for a minimal tensor product composite we have
\begin{align*}
m_\pm^*&=\, \max\{ v^Tv:\ v\in\mathds{R}^{d-1},\ |v|_1\leq 1 \}\, =\, 1\, ,\\
m_\pm &=\, \max\{ v^Tv:\ v\in\mathds{R}^{d-1},\ |v|_\infty\leq 1 \}\, =\, d-1\, ,\\
k_\pm &=\, \frac{m_\pm}{d-1}\, =\, 1\, ,
\end{align*}
and therefore $R_{\text{Werner}}^{\text{G}}(\text{SEP})=1$. This example shows how it might be the case that Werner states do not display any data hiding property, despite the fact that there is global data hiding in the cubic model, as shown by Proposition~\ref{dh cubic}.
\end{itemize}
\section{Conclusions}
We have presented a general theory of data hiding in bipartite systems composed of GPTs (general probabilistic theories), a framework comprising quantum states and probability distributions, and generally operational theories with states and measurements. It significantly extends ideas and results from quantum mechanics; in particular, we were able to determine the maximum so-called data hiding ratio in terms of the state space dimension, by making a connection between this problem and Grothendieck's tensor norms. This maximum is essentially attained for GPTs over spherical cones.
Inspired by the prominent role played by Werner states in quantum mechanical data hiding, we investigated Werner-like states in classes of theories with symmetric state spaces. By using these states as ansatzes, we could find a general lower bound on the data hiding ratio that depends on few geometrically meaningful parameters.
For quantum mechanics on finite-dimensional spaces we proved a new upper bound on the data hiding ratio against LOCC operations by exploiting the celebrated quantum teleportation protocol. We saw how this improves previous results and definitively settles the problem of determining the optimal data hiding ratio against LOCC for fixed local dimensions. However, the same problem for the smaller set of local operations remains open, as we showed that none of the bounds that we provide or that are available in the literature is generally optimal.
Perhaps surprisingly, data hiding ratios in quantum mechanics are not as large as the maximal conceivable ones, being of the order of the square root of the latter, which are exhibited by spherical cones. Thus, if one were to summarise the results of our research in one single sentence, one could say that Nature is non-classical, but not as non-classical as it could have been.
Although this conclusion has been found by many other authors in relation to different investigations into non-locality phenomena, it appears to us so just, that we offer it here, as the essence, so to speak, of the whole story.
Lastly, while here we have focused on the maximum data hiding ratios among $R(\text{SEP})$ for various local GPTs, the opposite end is very interesting, too, from a foundational point of view. Namely, it is clear that to have $R(\text{SEP})>1$, it is necessary that both cones $C_A$ and $C_B$ are non-classical in the sense that they can not be simplicial, for the reason that otherwise min- and max-tensor product of the cones would coincide. It is however not known that this is necessary, and one might conjecture that $R(\text{SEP})=1$ if and only if one of the cones $C_A$ or $C_B$ is simplicial. A general result along these lines is due to Namioka and Phelps~\cite{NP}, which says that for $C_B$ the cone of the so-called gbit, $R(\text{SEP})=1$ iff $C_A$ is simplicial. If the more general conjecture were true, it would make a good case for $R(\text{SEP})$ as a measure of non-classicality of a GPT. For a thorough discussion of this and related problems, see \cite[Chapter~II]{lamiatesi}.
\vspace{2ex}
{\it Acknowledgements.} We thank Guillaume Aubrun for many insightful discussions on functional analysis, in particular for having suggested the use of Auerbach's lemma in the proof of Proposition~\ref{prop proj=<ninj}. LL is also grateful to Howard Barnum and Matthias Christandl for convincing him that defining LOCC in arbitrary GPTs is possible.
LL and AW are indebted to John Calsamiglia and C\'ecilia Lancien for enlightening observations on data hiding and certain aspects of convex analysis, respectively. Finally, we thank the anonymous referees for useful comments on the first version of this paper.
We acknowledge financial support from the European Research Council (AdG IRQUAT No. 267386), the Spanish MINECO (Project no. FIS2013-40627-P, no. FIS2016-86681-P, and no. MTM2014-54240-P), the Generalitat de Catalunya (CIRIT Project no. 2014 SGR 966), and the Comunidad de Madrid (QUITEMAD+ Project S2013/ICE-2801). CP was partially supported by the ``Ram\'on y Cajal program" (RYC-2012-10449).
|
\section{Introduction}
\label{sec:intro}
\POL{Encountering inappropriate or unwanted personalised online content can be awkward, depending on social context. What may appear humorous in one situation may be embarrassing, or worse, in another context. When presented with content regarded as inappropriate or discreditable, a user may wish to \emph{deny interest} in the content.}
\POL{The Oxford English Dictionary defines plausible deniability in terms of reasonable doubt as \emph{``the possibility of denying a fact (especially a discreditable action) without arousing suspicion''}, \cite{oxford}. Informally, user activity observed by the search engine exhibits plausible deniability when user activity is consistent with the user interest in \emph{any} one of several topics, at least one of which is not sensitive for the user, with sufficiently high probability. }
\POL{Accordingly, we assess threats to plausible deniability during web search by testing if content appearing on search result pages can be attributed to user interest in a specific sensitive topic, versus user interest in any other topic, on the balance of probabilities. We ask when can a user \emph{plausibly deny} interest in a range of sensitive topics during online web-search sessions?}
We provide guarantees on the best-possible level of plausible deniability a user can expect during web search in our model. We also introduce a new \emph{Plausible Deniability Estimator}, called \textbf{PDE}{}, that can be used to assess privacy threats. Outputs from \textbf{PDE}{} can be represented in terms of readily interpretable probabilities thereby providing an informative indication of risk to the user.
Our methods are chosen to be straightforward to implement using openly available technologies. We use our results to design and assess counter-measures against threats to plausible deniability during online web-search sessions, using the Google Search as a source of data. We are able to assess threats to plausible deniability from sensitive topic learning in a range of potentially sensitive topics, such as health, finance and sexual orientation.
Our experimental measurements indicate that, by observing as few as 3-5 revealing queries, a search engine can infer a user is interested in a sensitive topic on the balance of probabilities in $100\%$ of topics tested when no effective defence is provided. In the case of topics related to health and sexual preferences measurements from \textbf{PDE}{} suggest that the probability a user is interested in sensitive topics related to sexual preference is as high as $90\%$ greater than their probability of interest in any other topic.
We show that defence strategies based on random query injection of random noise queries and misleading click patterns may provide some protection for individual, isolated queries, but that search engines are able to learn quickly. Significant levels of threat to plausible deniability are detected even when very high levels of random noise are included in the query session or when misleading click patterns are used. These approaches seem to offer little or no improvement to user privacy when considering plausible deniability over the longer term.
In contrast, we find that a defence employing topics that are commercially relevant but uninteresting to the user as \emph{proxy topics} is effective in protecting plausible deniability in the case of $100\%$ of sensitive topics tested. The proxy topic defence differs from traditional obfuscation approaches in actively exploiting the observed ability of the search engine to learn topics quickly, deflecting the focus of interest toward the proxy topic and away from the true topic of interest to the user.
The proxy topic defence works in our experiments, and is simple to apply. However it is important to recognise that we are faced with commercially motivated and increasingly powerful systems with a history of adapting quickly. Our results suggest that search engine capability is continuously evolving so that we can reasonably expect search engines to respond to privacy defences with more sophisticated learning strategies. Our results also point towards the fact that the text in search queries plays a key role in search engine learning. While perhaps obvious, this observation reinforces the user's need to be circumspect about the queries they ask if they want to avoid search engine learning of their interests. Equally, our results suggest that simple countermeasures, such as proxy topics, that make accurate personalisation more expensive for the online system represent a promising approach in developing new techniques for practical user privacy.
\section{Related Work}
\label{sec:related:work}
We model a search engine as a black-box by making minimal assumptions about its internal workings. The technique of using predefined \emph{probe queries}, injected at intervals into a stream of true user queries as fixed sampling points, was used in \cite{mac2015don} \POL{where the focus of the paper was detection of possible privacy threats. Extending the idea of probe queries, discussed in \cite{mac2015don}, several new applications are presented in this current paper such as the model of plausible deniability and the associated \textbf{PDE}{} estimator, the proxy topic defence model and the evaluation of multiple noise and click models for each of these.}
The technique of using predefined probe queries is borrowed from black-box testing. Modelling an adversary as a black-box, where internal details of recommender systems algorithms and settings are unknown to users, is mentioned in \cite{datta2014privacy} and \cite{Hannak:2013:MPW:2488388.2488435}.
The importance of control over appropriate flow of information is discussed extensively in legal and social science fields. Individual control over personal information flow is discussed in a critique of the \emph{nothing to hide} defence for widespread surveillance in \cite{Solove:2007}. Individual privacy and its social consequences are discussed in \cite{Nissenbaum:2009:PCT:1822585,boyd2011talk}, where agency or control over appropriate disclosure is identified as a key concern.
Plausible deniability as a privacy defence for web search is addressed in the literature. In \cite{arampatzis2013versatile} alternative, less revealing queries are mixed with sensitive topic queries to obfuscate true user interest. In \cite{Arampatzis:2011} queries with generalised terms are used to approximate the search results of a true query, which is never revealed. Plausible deniability for database release has been studied in the context of user data anonymization. For example, in \cite{DBLP:journals/pvldb/BindschaedlerSG17} a definition of plausible deniability is applied to examine mechanisms for differentially private data set release. More generally, plausible deniability to counteract the impact of personalisation is examined in \cite{cummings2016empirical} for the case of a privacy aware user who knows they are being observed. The authors show that no matter what the behaviour of the user is, it is always compatible with some concern over privacy. In this way the user can offer their awareness of privacy concerns as a general alibi to justify any range of preferences. Plausible deniability for providers of online services is also discussed in the literature. For example, in \cite{vera2014doccloud} a distributed virtual machine infrastructure is used to provide deniability to online data providers by obfuscating the origin of index data used in recommendations.
The potential of online profiling and personalisation resulting in censorship and discrimination have received growing attention in the research literature. Personalisation as a form of censorship -- termed a filter bubble in \cite{Pariser:2011:FBI:2029079} -- is explored in \cite{Hannak:2013:MPW:2488388.2488435}. In a filter bubble, a user cannot access subsets of information because the recommender system algorithm has decided it is irrelevant for that user. In \cite{Hannak:2013:MPW:2488388.2488435} a filter bubble effect was detected in the case of Google Web Search in a test with 200 users. Discrimination associated with personalisation has been shown for topics generally regarded as sensitive. In \cite{Sweeney:2013:DOA:2460276.2460278} an extensive review of adverts from Google and \url{Reuters.com} showed a strong correlation between adverts suggestive of an arrest record, and, an individual's ethnicity. In \cite{Guha:2010:CMO:1879141.1879152}, the authors used online advertising targeted {exclusively} to gay men to demonstrate strong profiling in the case of sexual preference.
Several approaches exist for obfuscating user interactions with search engines with the aim of disrupting online profiling and personalisation. GooPIR, \cite{Domingo-Ferrer:2009h,SaNchez:2013:KSC:2383079.2383150}, attempts to disguise a user's ``true'' queries by adding masking keywords directly into a true query before submitting to a recommender system. Results are then filtered to extract items that are relevant to the user's original true query. PWS, \cite{balsa2012ob}, and TrackMeNot, \cite{howe2009trackmenot,peddinti2011limitations}, inject distinct noise queries into the stream of true user queries during a user query session, seeking to achieve an acceptable level of anonymity while not overly upsetting overall utility. Search engine algorithm evolution regarded as a continuous ``arms-race'', is evidenced in the case of Google, for example, by major algorithm changes such as \emph{Caffeine} and \emph{Search+ Your World} have included additional sources of background knowledge from Social Media, improved filtering of content such as \emph{Panda} to counter spam and content manipulation, most recently semantic search capability has been added through \emph{Knowledge Graph} and \emph{HummingBird}, \cite{search_timeline_1}, \cite{search_history}, \cite{search_history_2}.
\POL{Consent to share data for agreed purposes is critical to user trust in service providers and is a key feature of the EU General Data Privacy Regulation (GDPR), \cite{GDPR:2016}.
Several notable browser add-ons, such as Mozilla Lightbeam, \cite{lightbeam}, and PrivacyBadger, \cite{badger}, facilitate more active user awareness of possible consent issues by helping a user understand where their data is shared with third parties through the sites they visit. XRay, \cite{DBLP:journals/corr/LecuyerDLPPSCG14}, reports high accuracy in identifying which sources of user data such as email or web search history might have {triggered} particular results from online services such as adverts. Active consensual sharing of personal data is investigated in \cite{Fredrikson:2011} through an in-browser capability, called RePriv, allowing a user to select which portions of their personal data they wish to share with requesters. Both PrivAd, \cite{Guha:2011}, and Adnostic, \cite{Liu:2016} investigate safe profiling through generalisation of user interests in the browser. Both Adnostic and PrivAd seek to protect the true interests of the user by obfuscating and filtering personalised content through a published interface.}
Evaluation of the effectiveness of privacy defences in the wild was performed by \cite{peddinti2010privacy} in the case of \emph{TrackMeNot} where the authors demonstrate that by using only a short-term history of search queries it is possible to break the privacy guarantees of TrackMeNot using readily available machine-learning classifiers. The importance of background information in user profiling is explored in \cite{Petit2016}. Here a similarity metric measuring distance between \emph{known} background information about a user, given by query history, and subsequent queries is shown to identify 45.3\% of TrackMeNot and 51.6\% of GooPIR queries. \POL{Anti-tracking is an ongoing area of research and recently in \cite{DBLP:conf/ndss/PanCC15} an anti-tracking browser called TrackingFree was reported to be effective at disrupting all of the trackers in the Alexa top-500 list}. Self-regulation has also proven problematic, in \cite{balebako2012measuring}, six different privacy tools, intended to limit advertising due to behavioural profiling, are assessed. The tools assessed implement a variety of tactics including cookie blocking, site blacklisting and Do-Not-Track (DNT) headers. DNT headers were found to be ineffective in tests at protecting against adverts based on user profiling.
Examples of unsubstantiated and misleading claims by providers of technology to enhance individual privacy are common, \cite{zdnet:charlatans,zdnet:fake}. Concerns about objective evaluation of the claims by providers of such technologies have attracted the attention of Government, where the need for \emph{``Awareness and education of the users \ldots ''} is identified in \cite{enisa:guide} as a key step to building trust and acceptance of privacy technologies for individuals. Accountability and enforcement of accountability for privacy policy is also attracting attention. Regulatory requirements for data handling in industries such as Healthcare (HIPPA) and Finance (GLBA) are well established. The position with respect to handling of data collected by online recommender systems is less clear. In \cite{datta2014privacy}, the author reviews computational approaches to specification and enforcement of privacy policies at large scale.
\POL{Our contribution in this paper is orthogonal to the contributions in the works discussed here. We address the complimentary challenge of privacy monitoring by detecting possible inappropriate use of personal user data by observing personalised outputs. In this respect our approach can be deployed in conjunction with the technologies mentioned.}
\section{General Setup}
\label{sec:gen:setup}
\subsection{Threat Model}
\label{sec:notation}
The setup we consider is that of a general user of a commercial, for-profit online search engine. The relationship between the user, denoted \ensuremath{\mathcal{U}}{}, and the online system, denoted by \ensuremath{\mathcal{S}}{}, is based on mutual utility where both parties obtain something useful from the interaction -- \ensuremath{\mathcal{U}}{} gets useful information and recommendations -- while \ensuremath{\mathcal{S}}{} gets an opportunity to ``up-sell'' to \ensuremath{\mathcal{U}}{} through targeted content such as advertising. As a commercial business, \ensuremath{\mathcal{S}}{} recognises cost per user interaction and responsiveness of service are critical to competitiveness. Accordingly content based on user profiling is intended to adapt dynamically to the changing interests of \ensuremath{\mathcal{U}}{}. \ensuremath{\mathcal{U}}{} is generally informed regarding good personal privacy practice and is alert to unwanted or embarrassing personalisation. When \ensuremath{\mathcal{U}}{} detects threats to her privacy she wishes to assess her ability to plausibly deny her interest in compromising content to avoid awkward social implications. The relationship between \ensuremath{\mathcal{U}}{} and \ensuremath{\mathcal{S}}{} is generally described as ``honest but curious'' in the literature. Accordingly we will refer to \ensuremath{\mathcal{S}}{} as an \emph{observer} rather than the more traditional \emph{adversary}.
Let $\{c_1, \ldots, c_N\}$ denote a set of sensitive categories of interest to \ensuremath{\mathcal{U}}{}, \emph{e.g.} bankruptcy, cancer, addiction, \emph{etc}. Gather all other uninteresting categories into a catch-all category denoted $c_{0}$. The set $\SET{C} = \{c_0, c_1, \ldots, c_N\}$ is \emph{complete} in the sense that all user topic interests can be represented as subsets of $\SET{C}$ with the usual set operations. We are interested in threats from search engine learning that compromise the ability of \ensuremath{\mathcal{U}}{} to plausibly deny their interest in sensitive topics. We will assess threats to plausible deniability by testing if content appearing on search result pages can be attributed to interest in a specific topic $c_i \in \SET{C}$, versus interest in any other topic, on the balance of probabilities.
\begin{figure}
\centering
\includegraphics[scale=0.45]{overview}
\caption{Overview of user--search engine interaction\label{fig:overview}}
\end{figure}\\[-4mm]
We treat \ensuremath{\mathcal{S}}{} as a black-box with internal state unknown to \ensuremath{\mathcal{U}}{}. As a starting point, our initial assumption is that \ensuremath{\mathcal{S}}{} is motivated to use its internal state of knowledge of \ensuremath{\mathcal{U}}{} when producing personalised outputs for \ensuremath{\mathcal{U}}{}, thereby revealing something about its internal state.
\begin{assumption}[Revealing Observations]
A search engine selects personalised page content, such as adverts, it believes are aligned with our interests.
\label{assumption:1}
\end{assumption}
When a search engine infers that a particular advertising category is likely to be of interest to a user, and it is more likely to generate click through and sales, it is obliged to use this information when selecting which adverts to display. This suggests that, by examining \emph{advert content} recommended by the search engine, it is possible to detect evidence of sensitive topic profiling by the search engine. Assumption~\ref{assumption:1} is fundamental to the application of our approach in that, if \ensuremath{\mathcal{S}}{} does not produce content that reveals evidence of learning then, since \ensuremath{\mathcal{S}}{} is a black-box in our model, our approach has nothing to say about observer learning. In summary, we rely on the observer to show his hand through adverts -- our approach can only observe what is shown. In practice this does not appear to be a significant limitation with regard to many topics regarded as sensitive to users. In our experiments we observe an average of $2-3$ adverts per probe query with less than $10\%$ of probe queries resulting in no advert content. We also note that our scope is limited to examining advert content. We note that in addition to adverts commercial search engines also typically provide additioenal personalised content that could also be tested for evidence of learning, for example Google provides a variety of personalised content such as ``top stories'', related Tweets. However we leave consideration of these as future work.
A user interacts with a search engine by issuing a query, receiving a web page in response and then clicking on one or more items in the response. In the case of web-search, a single such interaction, labeled with index $j$, consists of a \emph{query, response page, item-click} triple, denoted
$\omega_j = \left(q_j, p_j, l_j \right)$.
We model construction of a query $q_j$ as selection of words from a generally available dictionary denoted $\SET{D}$. We assume that words in $\SET{D}$ are matched to topics in $\SET{C}$. The word--topic category matching is not unique and words may be matched to multiple topic categories.
A user session of length $k>0$ steps consists of a {sequence} of $k$ individual steps, and is denoted $\left\{\omega_k\right\}_{k\ge1}$. The sequence of interactions $\left\{\omega_k\right\}_{k\ge1}$ is jointly observed by the user and the search engine -- and perhaps several other third-party observers.
The relationship between prior and posterior background knowledge at each step $k$ is
\begin{align}
\{\omega_k, \EV{E}_{k-1}\} &= \EV{E}_{k}, \quad k=1,2,\ldots
\label{eqn:e:k:1}
\end{align}
\noindent where $\EV{E}_0$ denotes the initial background knowledge state of \ensuremath{\mathcal{S}}{} at the beginning of a session immediately before $\omega_1$ is observed. The detail of $\EV{E}_0$ is unknown to \ensuremath{\mathcal{U}}{} who treats \ensuremath{\mathcal{S}}{} as a black-box.
Figure~\ref{fig:overview} illustrates the interaction between user and search engine in our model.
Let the random variable $\VEC{X}$ with sample space $\{0,1\}^{N}$ represent user interest in categories in $\SET{C}$ during a session. A value of $1$ in element $i$ of $\VEC{X}$ indicates evidence is detected of user interest in topic $c_i$.
After each step $k$ of a query session, \ensuremath{\mathcal{S}}{} can construct a posterior probability distribution for $\VEC{X}$, namely, for $\VEC{x} \in \{0,1\}^{N}$
\begin{align}
\text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \Omega_k = \omega_k, \EV{E}_k)
= \text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \EV{E}_{k+1})
\label{eqn:actions:001}
\end{align}
\noindent We use $\Omega_k$ to represent the observation at step $k$, so that $\Omega_k = \omega_k$ indicates $\omega_k$ is observed at step $k$.
The individual interest vector for topic $i$ is denoted $\VEC{c}_i$, a vector with a single $1$ in the $i^{th}$ position and $0$ in all other positions. The probability of that a user is interested \emph{only} in topic $i$ at step $k$ of a session and the posterior probability of detecting evidence that a user is interested only in topic $i$ at query step $k$, conditioned on observing $\omega_k$ and background knowledge $\EV{E}_k$, are
\begin{align*}
\text{\rm P}(\VEC{X} = \VEC{c}_i \vert \, \EV{E}_k)
&\quad\text{and}\quad
\text{\rm P}(\VEC{X} = \VEC{c}_i \vert \, \Omega_k = \omega_k, \, \EV{E}_{k} )
\end{align*}
\noindent respectively.
Since $\mathcal{C}$ contains all possible topics:
\begin{align*}
\sum_{i=0}^{N} \text{\rm P}(\VEC{X} = \VEC{c}_i \vert \, \EV{E}_k) = 1, \quad k=1,2,\ldots
\end{align*}%
\subsection{Example: Single Sensitive Category}
\label{sec:example:bso}
\POL{To illustrate mathematical results as we go, we use a simple ideal model, consisting of a single sensitive category, as an illustrative example. We will refer to it as the Single Sensitive Category (SSC) model. The single sensitive topic is denoted $c^{1}$ and the catch-all, non-sensitive topic representing every other topic that is not part of the sensitive topic is denoted $c^{o}$. }
\POL{Suppose \ensuremath{\mathcal{U}}{} can issue queries related to either of two topics $\{c^{1}, c^{o} \}$ denoting sensitive and non-sensitive interests respectively. \ensuremath{\mathcal{S}}{} models the process by which \ensuremath{\mathcal{U}}{} draws queries according to an initial probability model
\begin{align}
\text{\rm P}(\VEC{X} = c^{1} \vert \EV{E}_0) := p_{0}^{s}, \quad
\text{\rm P}(\VEC{X} = c^{o} \vert \EV{E}_0) := p_{0}^{n}
\label{eqn:transition:bso:0}
\end{align}
\noindent with
$p^{1} + p^{0} = 1$.
}
\POL{On observing a query from \ensuremath{\mathcal{U}}{} at step $k$ the observer \ensuremath{\mathcal{S}}{} outputs one of $\{\omega^{s}_{k}, \omega^{n}_{k}\}$ according to the associated conditional probabilities at step $k$ given by
\small
\begin{align}
\text{\rm P}(\Omega_{k} = \omega_{k}^{s}\vert \VEC{X} = c^{1}, \EV{E}_{k}) = \text{\rm P}(\Omega_{k} = \omega_{k}^{n}\vert \VEC{X} = c^{o}, \EV{E}_{k}) &:= \pi^{k} \notag \\
\text{\rm P}(\Omega_{k} = \omega_{k}^{n}\vert \VEC{X} = c^{1}, \EV{E}_{k}) = \text{\rm P}(\Omega_{k} = \omega_{k}^{s}\vert \VEC{X} = c^{o}, \EV{E}_{k}) &:= 1-\pi^{k}
\label{eqn:transition:bso:1}
\end{align}
\normalsize}%
\POL{The SSC model is deliberately simple as the intention is to illustrate mathematical concepts. The model is generally useful for exploring black-box interactions and can be readily extended to include more sophisticated scenarios such as allowing \ensuremath{\mathcal{U}}{} to select from multiple topics, as would happen when \ensuremath{\mathcal{U}}{} attempts to obfuscate her interests by switching topics.}%
\subsection{Plausible Deniability}
\label{sec:by:design}
Our threat assessment model is based on \emph{plausible deniability}. Informally, the user activity observed by the search engine exhibits plausible deniability when, with high probability, is consistent with the user being interested in any one of several topics at least one of which is not sensitive for the user. That is, the user activity supports reasonable doubt about the user's actual interest in a given sensitive topic.
In our setup, the topics are $c_i\in\SET{C}$ while the observed activity is $\{\omega_j\}_{j=1}^{k}$ at step $k$ in a session (i.e. the queries, search result pages and associated user clicks). We formalise plausible deniability as follows.
\begin{definition}[{\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{}\label{def:plausible}]
\label{def:plausible:detect:001}
For privacy parameters $\epsilon >0$ and $m>0$ and a set of $N+1$ topics $\SET{C} = \{c_0, \ldots c_N \}$, a user with a true user interest vector $\VEC{x} \in \{0,1\}^{N+1}$ is said to have \emph{{\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{}} at step $k$ in the query session, if, for observations $\{\omega_{j}\}_{j=1}^{k}$, made at each step $j=1,\ldots k$ of a session by an observer possessing initial background knowledge $\EV{E}_0$ at the beginning of the session, there exist at least $m-1$ other $\VEC{x}_i \in \{0,1\}^{N+1}\setminus\{\VEC{x}\} $ such that
\begin{align}
e^{-\epsilon} < \ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k}) < e^{\epsilon}
\label{eqn:delta:deny}
\end{align}
\noindent where
\begin{align}
\ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k}) = \frac{\text{\rm P}(\Omega_{k}=\omega_{k}, \ldots \Omega_{1}=\omega_{1} \, \vert \, \VEC{X} = \VEC{x}, \EV{E}_0 )}
{\text{\rm P}(\Omega_{k}=\omega_{k}, \ldots \Omega_{1}=\omega_{1} \, \vert \, \VEC{X} = \VEC{x}_i, \EV{E}_0 )}
\label{eqn:delta:deny:1}
\end{align}
\label{def:delta:deny}
\end{definition}
\noindent For \eqref{eqn:delta:deny} to be well-defined, all probabilities are assumed to be non-zero. In practice, this is not a significant restriction since categories with zero probability are gathered into the catch-all topic $c_0$.
By applying the chain-rule for conditional probability, \eqref{eqn:delta:deny:1} can be rewritten as
{\small
\begin{align}
&\ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k}) = \frac{\text{\rm P}(\Omega_{k}=\omega_{k}, \ldots \Omega_{1}=\omega_{1} \, \vert \, \VEC{X} = \VEC{x}, \EV{E}_0 )}
{\text{\rm P}(\Omega_{k}=\omega_{k}, \ldots \Omega_{1}=\omega_{1} \, \vert \, \VEC{X} = \VEC{x}_i, \EV{E}_0 )} \notag\\
&= \prod_{j=0}^{k-1} \frac{\text{\rm P}(\Omega_{k-j}=\omega_{k-j} \vert \VEC{X} = \VEC{x}, \Omega_{k-j-1}=\omega_{k-j-1}, \ldots \Omega_{1}=\omega_{1}, \EV{E}_0 )}
{\text{\rm P}(\Omega_{k-j}=\omega_{k-j} \vert \VEC{X} = \VEC{x}_i, \Omega_{k-j-1}=\omega_{k-j-1}, \ldots \Omega_{1}=\omega_{1}, \EV{E}_0 )}
\label{eqn:delta:deny:2} \\
&= \prod_{j=0}^{k-1} \frac{\text{\rm P}(\Omega_{k-j}=\omega_{k-j}\, \vert \, \VEC{X} = \VEC{x}, \EV{E}_{k-j} )}
{\text{\rm P}(\Omega_{k-j}=\omega_{k-j} \, \vert \, \VEC{X} = \VEC{x}_i, \EV{E}_{k-j} )}
= \prod_{j=0}^{k-1} \WeePD[k-j](\VEC{x}, \VEC{x}_i, \omega_{k-j})
\label{eqn:delta:deny:3}
\end{align}
}%
\noindent where
\begin{align}
\WeePD(\VEC{x}, \VEC{x}_i, \omega_j) := \frac{\text{\rm P}(\Omega_{j}=\omega_{j}\, \vert \, \VEC{X} = \VEC{x}, \EV{E}_j )}
{\text{\rm P}(\Omega_{k}=\omega_{j} \, \vert \, \VEC{X} = \VEC{x}_i, \EV{E}_j )}
\label{eqn:delta:deny:4}
\end{align}
\noindent is the incremental change in {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} arising from the single observation $\omega_k$ at step $k$ of the session.
\POL{In the case of the SSC model, there are two topics -- sensitive and non-sensitive -- so that \ensuremath{\mathcal{U}}{} can at best hope for $(\epsilon, m=2)$--Plausible Deniability for the sensitive topic, in which case \eqref{eqn:delta:deny:4} can be written as
\begin{align}
\WeePD(c^{1}, c^{o}, \omega^{s}_j) &= \frac{\text{\rm P}(\Omega_{j}=\omega^{s}_{j}\, \vert \, \VEC{X} = c^{1}, \EV{E}_j )}
{\text{\rm P}(\Omega_{k}=\omega^{s}_{j} \, \vert \, \VEC{X} = c^{o}, \EV{E}_j )}
&=\frac{\pi^{j}}{1-\pi^{j}}
\label{eqn:deny:bso:1}
\end{align}
\noindent when \ensuremath{\mathcal{S}}{} emits a sensitive output $\omega^{s}_{j}$ at step $j$, and
\begin{align}
\WeePD(c^{1}, c^{o}, \omega^{s}_j)
&=\frac{1-\pi^{j}}{\pi^{j}}
\label{eqn:deny:bso:3}
\end{align}
\noindent when a non-sensitive output is emitted at step $j$. Substituting these values into \eqref{eqn:delta:deny:3} and assuming $a$ sensitive outputs and $k-a$ non-sensitive outputs in a session of length $k$, let $k_s$ denote the sub-sequence of steps where sensitive outputs are detected and $k_n = [k] \setminus k_s$ be the steps where non-sensitive outputs are detected so that
\begin{align}
\ensuremath{\mathbb{D}_{k}}(c^{1}, c^{o}, \{\omega_{j}\}_{j=1}^{k})
= \prod_{i\in k_s}\frac{\pi^{i}}{1-\pi^{i}}\prod_{j\in k_n}\frac{1-\pi^{j}}{\pi^{j}}
\end{align}
}
\subsection{Comparison with Other Anonymity Measures}
\label{sec:compare:k:dp}
\POL{Intuitively, Definition~\ref{def:plausible:detect:001} is similar to $k$--anonymity in that an observer can only explain observations $\{\omega_{j}\}_{j=1}^{k}$ to within a generalised set consisting of at least $k=m$ topic vectors with probability bounded by the choice of $\epsilon$. Definition~\ref{def:plausible:detect:001} differs from regular k--anonymity in requiring both upper and lower bounds in \eqref{eqn:delta:deny:1} since evidence of \emph{loss of interest} in a sensitive topic may be as revealing and potentially embarrassing as evidence of \emph{increase of interest}.}
\POL{Definition~\ref{def:plausible:detect:001} can also be compared with a slightly weaker form of Differential Privacy. Informally, making an observation should not make \ensuremath{\mathcal{S}}{} significantly more, or less, confident of user interest in a particular sensitive topic.}
\POL{From \eqref{eqn:delta:deny:4} the incremental change due to a single observation $\Omega_j = \omega_j$ is }
\begin{align}
\frac{\text{\rm P}(\Omega_{j}=\omega_{j}\, \vert \, \VEC{X} = \VEC{x}, \EV{E}_j )}
{\text{\rm P}(\Omega_{k}=\omega_{j} \, \vert \, \VEC{X} = \VEC{x}_i, \EV{E}_j )}
= \frac{\text{\rm P}( \VEC{X} = \VEC{x}\, \vert \, \Omega_{j}=\omega_{j}, \EV{E}_j )}
{\text{\rm P}(\VEC{X} = \VEC{x}_i \, \vert \, \Omega_{k}=\omega_{j}, \EV{E}_j )}
\label{eqn:compare:1}
\end{align}
\noindent\POL{by applying Bayes Theorem. Since \eqref{eqn:delta:deny:4} is bounded above and below for at least $m-1$ other $x_i$ when Definition~\ref{def:plausible:detect:001} holds, it follows that }
\begin{align}
e^{-\epsilon} < \frac{\text{\rm P}( \VEC{X} = \VEC{x}\, \vert \, \Omega_{j}=\omega_{j}, \EV{E}_j )}
{\text{\rm P}(\VEC{X} = \VEC{x}_i \, \vert \, \Omega_{k}=\omega_{j}, \EV{E}_j )} < e^{\epsilon}
\label{eqn:compare:2}
\end{align}
\noindent\POL{for at least $m-1$ other topic vectors $x_i$ -- but not necessarily for \emph{all} topic vectors. In which case we say that m--Differential Privacy holds for $\epsilon > 0$ whenever Definition~\ref{def:plausible:detect:001} holds, meaning that for any topic vector $x$ it is impossible to distinguish it from at least $m-1$ other topic vectors in $\{0,1\}^{N+1}$. This is a slightly weaker statement of Differential Privacy from the usual global definition.}
\subsection{Testing for Plausible Deniability}
\label{sec:p:d:i}
The following \emph{indistinguishability} definition of \emph{privacy risk} measures the change in belief by a search engine due to inference from observed user events relative to its prior belief conditioned on the background data available at the start of the query session. It is adapted from work begun in \cite{mac2015don} and using it allows us to adapt tools originally developed there.
\begin{definition}[\textit{\ensuremath{\epsilon}-Indistinguishability }]
\label{def:eps:001}
For a privacy parameter $\epsilon > 0$, a user with interest vector $\VEC{x} \in \{0,1\}^{N}$ is said to be s said to be $\epsilon$-Indistinguishable with respect to an observation of user actions $\omega_k$ at step $k$, if
\begin{align}
e^{-\epsilon} &\leq \FPRI (\VEC{x}, \omega_k) \leq e^{\epsilon}
\label{def:eps:indist:001}
\end{align}
\noindent where
\begin{align}
\FPRI (\VEC{x}, \omega_k) &= \frac{\text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \Omega_k = \omega_k, \EV{E}_{k} )}{\text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \EV{E}_{0})}
\label{eqn:m:u:1} \\
&= \frac{\text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \EV{E}_{k+1} )}{\text{\rm P}(\VEC{X} = \VEC{x} \, \vert \, \EV{E}_{0})} \quad\text{(Applying \eqref{eqn:e:k:1})}
\label{eqn:m:u:2}
\end{align}
\noindent is called the \emph{\textit{\ensuremath{\epsilon}-Indistinguishability } score} of the interest vector $\VEC{x}$ for observation $\omega_k$ and background knowledge $\EV{E}_{k}$ at step $k$.
\label{def:eps:indist1}
\end{definition}
In other words, for \textit{\ensuremath{\epsilon}-Indistinguishability } to hold at step $k$ of a query session, the conditional posterior distribution should be approximately equal to the prior distribution at the beginning of the query session for the true interests of the user. To ensure \eqref{eqn:m:u:1} is well defined we assume all probabilities in \eqref{eqn:m:u:1} are non-zero, so that $0 < \FPRI (\VEC{x}, \omega_k) < \infty $. Expression \eqref{eqn:m:u:1} implies that if \textit{\ensuremath{\epsilon}-Indistinguishability } holds at step $k$ for an interest vector $\VEC{x}$, then $e^{-\epsilon} < (\FPRI[k-1](\VEC{x}, \omega_{k}))^{-1} < e^{\epsilon}$.
The next result provides the necessary connection between \textit{\ensuremath{\epsilon}-Indistinguishability } and {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }} to apply tools, developed in \cite{mac2015don} for \textit{\ensuremath{\epsilon}-Indistinguishability }, to {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}.
\begin{proposition}\label{prop:4eps}
If \textit{\ensuremath{\epsilon}-Indistinguishability } holds on a subset $\SET{I} \subseteq \{0,1\}^{N+1}$ for $\epsilon>0$ for step $k$ and the initial step $1$, then {\text{(\ensuremath{4\epsilon, m})--Plausible Deniability }} holds on $\SET{I}$ for $m \leq \vert \SET{B} \vert$. Furthermore
\begin{align}
&\ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k})
= \frac{\FPRI[k](\VEC{x}_1, \omega_{k})}{\FPRI[k](\VEC{x}_2, \omega_{k})} \frac{\FPRI[1](\VEC{x}_2, \omega_{1})}
{\FPRI[1](\VEC{x}_1, \omega_{1})}
\label{eps:deny:proof:0}
\end{align}
\end{proposition}
\begin{proof}
Assume \textit{\ensuremath{\epsilon}-Indistinguishability } holds on $\SET{I} \subseteq \{0,1\}^{N+1}$ then for any $\VEC{x}_1, \VEC{x}_2 \in \SET{I}$. From \eqref{eqn:delta:deny:4}
\small
\begin{align}
&\WeePD(\VEC{x}_1, \VEC{x}_2, \omega_k) = \frac{\text{\rm P}(\Omega_k = \omega_k \vert \VEC{X} = \VEC{x}_1, \EV{E}_k)}{\text{\rm P}(\Omega_k = \omega_k \vert \VEC{X} = \VEC{x}_2, \EV{E}_k)} \notag \\
&=\frac{\text{\rm P}(\VEC{X} = \VEC{x}_1 \vert \Omega_k = \omega_k, \EV{E}_k)}{\text{\rm P}(\VEC{X} = \VEC{x}_1 \vert \EV{E}_k)}
\frac{\text{\rm P}(\VEC{X} = \VEC{x}_2 \vert \EV{E}_k)}{\text{\rm P}(\VEC{X} = \VEC{x}_2 \vert \Omega_k = \omega_k, \EV{E}_k)} \label{eps:deny:proof:1} \\
&\text{(Bayes Theorem)} \notag
\end{align}
\begin{align}
&=\underbrace{\frac{\text{\rm P}(\VEC{X} = \VEC{x}_1 \vert \Omega_k = \omega_k, \EV{E}_k)}{\text{\rm P}(\VEC{X}=\VEC{x}_1 \vert \EV{E}_0)}}_{(a)}
\underbrace{\frac{\text{\rm P}(\VEC{X}=\VEC{x}_1 \vert \EV{E}_0)}{\text{\rm P}(\VEC{X} = \VEC{x}_1 \vert \EV{E}_k)}}_{(b)} \times \ldots \notag \\
&\ldots\times \underbrace{\frac{\text{\rm P}(\VEC{X} = \VEC{x}_2 \vert \EV{E}_k)}{\text{\rm P}(\VEC{X}=\VEC{x}_2 \vert \EV{E}_0)}}_{(c)}
\underbrace{\frac{\text{\rm P}(\VEC{X}=\VEC{x}_2 \vert \EV{E}_0)}{\text{\rm P}(\VEC{X} = \VEC{x}_2 \vert \Omega_k = \omega_k, \EV{E}_k)}}_{(d)} \label{eps:deny:proof:2}
\end{align}
\begin{align}
= \frac{\FPRI[k](\VEC{x}_1, \omega_{k})\FPRI[k-1](\VEC{x}_2, \omega_{k-1})}
{\FPRI[k-1](\VEC{x}_1, \omega_{k-1})\FPRI[k](\VEC{x}_2, \omega_{k})}
\label{eps:deny:proof:3}
\end{align}%
\normalsize
\noindent Where expressions (b) and (c) in \eqref{eps:deny:proof:2} are $(\FPRI[k-1]({\VEC{x}_1, \omega_{k-1}}))^{-1}$ and $(\FPRI[k-1](\VEC{x}_2, \omega_{k-1}))^{-1}$ respectively from the definition in \eqref{eqn:m:u:2}.
Therefore from the definition of $\ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k})$ in \eqref{eqn:delta:deny:3}
\begin{align}
&\ensuremath{\mathbb{D}_{k}}(\VEC{x}, \VEC{x}_i, \{\omega_{j}\}_{j=1}^{k})
= \frac{\FPRI[k](\VEC{x}_1, \omega_{k})}{\FPRI[k](\VEC{x}_2, \omega_{k})} \frac{\FPRI[1](\VEC{x}_2, \omega_{1})}
{\FPRI[1](\VEC{x}_1, \omega_{1})}
\label{eps:deny:proof:4}
\end{align}
\noindent So that \eqref{eps:deny:proof:0} holds. Since individual elements in \eqref{eps:deny:proof:4} satisfy \textit{\ensuremath{\epsilon}-Indistinguishability } for $\epsilon>0$ it follows that {\text{(\ensuremath{4\epsilon, m})--Plausible Deniability }}{} holds as required.
\QED
\end{proof}%
Proposition~\ref{prop:4eps} provides a basic strategy for asserting when {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} holds. By establishing a value of $\epsilon$ for which a collection of topics $\SET{C} = \{c_0, \ldots c_N\}$ satisfies \textit{\ensuremath{\epsilon}-Indistinguishability }, {\text{(\ensuremath{4\epsilon, m})--Plausible Deniability }} follows with, at least, $m = \vert\SET{C}\vert = N+1$. This is a \emph{minimum} guarantee, as there may be topics for which \textit{\ensuremath{\epsilon}-Indistinguishability } fails but {\text{(\ensuremath{4\epsilon, m})--Plausible Deniability }} holds.
In our experiments we test whether \ensuremath{\mathcal{U}}{} can plausibly deny whether or not observed actions can be uniquely associated with interest in a given sensitive topic $c_i$ versus interest in ``any other'' topic in $\SET{C} \setminus \{c_i\}$, so that $m=2$.
For a topic $c_i\in\SET{C}$ the expression for {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} becomes
\small
\begin{align}
&\ensuremath{\mathbb{D}_{k}}(\VEC{c}_i, \VEC{c}_{-i} , \{\omega_j\}_{j=1}^k) :=
\frac{\text{\rm P}(\Omega_k = \omega_k, \ldots \Omega_1 = \omega_1 \vert \VEC{X} = \VEC{c}_i, \EV{E}_0)}{\text{\rm P}(\Omega_k = \omega_k, \ldots \Omega_1 = \omega_1 \vert \VEC{X} = \VEC{c}_{-i}, \EV{E}_0)}
\label{eqn:discuss:estim:2}
\end{align}
\normalsize
\noindent where $\VEC{c}_{-i}$ denotes the topic interest vector representing interest in the topics $\SET{C}\setminus \{c_i\}$.
The following result connects $\ensuremath{\mathbb{D}_{k}}(\VEC{c}_i, \VEC{c}_{-i}, \{\omega_j\}_{j=1}^k)$ to variation in probabilities
\begin{proposition}
\label{prop:estim}
If {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} holds for $\VEC{c}_{i}$ and $\VEC{c}_{-i}$ with $\epsilon>0$ and $m=2$ then
\begin{align}
\vert &\text{\rm P}(\Omega_k = \omega_{k}, \ldots \Omega_1=\omega_1 \, \vert \, \VEC{X} = \VEC{c}_i, \EV{E}_0 ) \notag\\
&- \text{\rm P}(\Omega_k = \omega_{k}, \ldots \Omega_1=\omega_1 \, \vert \, \VEC{X} = \VEC{c}_{-i}, \EV{E}_0 ) \vert \notag \\
&\geq\left\vert\log\left( \frac{\FPRI[k](\VEC{x}_1, \omega_{k})}{\FPRI[k](\VEC{x}_2, \omega_{k})} \frac{\FPRI[1](\VEC{x}_2, \Omega_{k-1})}
{\FPRI[1](\VEC{x}_1, \Omega_{k-1})} \right)\right\vert
\label{eqn:prop:estim:2}
\end{align}
And so
\begin{align}
\epsilon_{*} := \left\vert\log\left( \frac{\FPRI[k](\VEC{x}_1, \omega_{k})}{\FPRI[k](\VEC{x}_2, \omega_{k})} \frac{\FPRI[1](\VEC{x}_2, \omega_{k-1})}
{\FPRI[1](\VEC{x}_1, \omega_{k-1})}\right)\right\vert
\label{eqn:prop:estim:2b}
\end{align}
\noindent is a lower bound for the best possible achievable level of {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{}.
\end{proposition}
\begin{proof}
If {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} holds for $\epsilon>0$ then
\begin{align}
\left\vert\log( \ensuremath{\mathbb{D}_{k}}(\VEC{c}_i, \VEC{c}_{-i}, \{\omega_j\}_{j=1}^{k}) )\right\vert < \epsilon
\label{eqn:prop:estim:3}
\end{align}
\noindent so that $\vert\log(\ensuremath{\mathbb{D}_{k}}(\VEC{c}_i, \VEC{c}_{-i}, \{\omega_j\}_{j=1}^{k}))\vert$ is a lower bound for all $\epsilon >0$ for which {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} holds.
The result follows by substituting the expression in \eqref{eps:deny:proof:3} for $\ensuremath{\mathbb{D}_{k}}(\VEC{c}_i, \VEC{c}_{-i}, \{\omega_j\}_{j=1}^{k})$ in \eqref{eqn:prop:estim:3}.
\QED
\end{proof}%
Proposition~\ref{prop:estim} will be used later to create an estimator for $\epsilon_{*}$ that can be measured in experiments. From now on we simplify our discussion to the case $m=2$ and so experimental results are reported accordingly for the two-topic case, $\{c_i, c_{-i}\}$.
\section{Implementation}
\label{sec:experiment:setup}
\subsection{Preliminaries}
\label{sec:bayesian:formulation}
During testing we wish to use \eqref{eqn:prop:estim:2b} to create an estimator, we call \textbf{PDE}{}, to estimate the level of {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }} afforded. Since estimating the quantities in \eqref{eqn:prop:estim:2b} uses \textbf{PRI}, we recap the bare essentials of \textbf{PRI}{} and refer the reader to \cite{mac2015don} for more details.
To test for learning PRI injects a predefined \emph{probe query} into a stream of ``true'' queries during a query session. In this way, any differences detected in advert content in response to probe queries can be compared to identify evidence of learning. An ideal probe should not disrupt the learning process of \ensuremath{\mathcal{S}}{}. Denote the event a probe query is selected from $\SET{D}$ at step $k+1$ by $\Omega_{k+1} = \omega_{k+1}^{P}$. We formalise the notion of an ideal probe query by demanding that observing $\Omega_{k+1} = \omega_{k+1}^{P}$ should be conditionally independent of the user topic $\VEC{X}$ given the existing background knowledge of the observer
\begin{align}
&\text{\rm P}(\VEC{X} = \VEC{c}_i, \Omega_{k+1} = \omega_{k+1}^{P} \,\vert\, \EV{E}_k) \notag \\
&=
\text{\rm P}(\VEC{X} = \VEC{c}_i, \,\vert\, \EV{E}_k)\cdot \text{\rm P}(\Omega_{k+1} = \omega_{k+1}^{P} \,\vert\, \EV{E}_k)
\label{eqn:choose:probe:10}
\end{align}
\noindent and so observing the probe query and associated clicks does not provide any more information to the observer about the interests of \ensuremath{\mathcal{U}}{} than the current background knowledge already provides. From \eqref{eqn:choose:probe:10}
\begin{align}
&\text{\rm P}(\VEC{X} = \VEC{c}_i \,\vert\, \Omega_{k+1} = \omega_{k+1}^{P}, \EV{E}_k) \notag \\
&=
\text{\rm P}(\VEC{X} = \VEC{c}_i, \,\vert\, \EV{E}_k)
\label{eqn:choose:probe:2}
\end{align}
In practice, choosing an ideal probe query is achieved by selecting words from $\SET{D}$ that match words for several topics in $\SET{C}$ so that it is not possible to associate a single topic in $\SET{C}$ with the probe query.
Construction of \textbf{PRI}{} is based on several assumptions, the first of these assumptions is that the background knowledge at the first step of a query session, $\EV{E}_1$, provides sufficient description of background knowledge for all subsequent steps of that query session, $\EV{E}_k$.
\begin{assumption}[Sufficiently Informative Responses]\label{a:1}
Let $\SET{K}\subset\{1,2,\cdots\}$ label the sub-sequence of steps at which a probe query is issued.
At each step $k\in\SET{K}$ at which a probe query is issued,
\begin{align}
\frac{\text{\rm P}(\VEC{X} = \VEC{c}_i | \Omega_k = \omega_k, \EV{E}_{k} )}{\text{\rm P}(\VEC{X} = \VEC{c}_i | \EV{E}_{1} )}&=\frac{\text{\rm P}(\VEC{X} = \VEC{c}_i | \Omega_k = \omega_k,\EV{E}_{1} )}{\text{\rm P}(\VEC{X} = \VEC{c}_i| \EV{E}_{1})}
\label{eqn:assumption:informative}
\end{align}
\noindent for each topic ${c}_i \in \SET{C}$.
\end{assumption}
\noindent So that it is not necessary to explicitly use knowledge of the search history during the current session when estimating $\FPRI$ for a topic $c$ as this is already reflected in the search engine response, $\omega_k$, with the initial background knowledge $\EV{E}_1$ capturing background knowledge up to the start of the session, at step $k$. Assumption \ref{a:1} greatly simplifies estimation as it means we do not have to take account of the full search history, but requires that the response to a query reveals search engine learning of interest in sensitive category $c$ which has occurred. Assumption~\ref{a:1} was called the ``Informative Probe'' assumption in \cite{mac2015don}.
The next assumption is that adverts are selected to reflect search engine belief in user interests. In this way adverts are assumed to be the principal way in which search engine learning is revealed. Given this assumption, conditional dependence on $\omega_k$ can be replaced with dependence on the adverts appearing on the screen.
\begin{assumption}[Revealing Adverts]\label{a:2} In the search engine response to a query at step $k$ it is the adverts ${a}_k$ on a response page which primarily reveal learning of sensitive categories.
\begin{align}
\text{\rm P}(\VEC{X} = \VEC{c}_i | \Omega_k = \omega_k,\EV{E}_{1}) = \text{\rm P}(\VEC{X} = \VEC{c}_i | {a}_k,\EV{E}_{1} ), \ k\in \SET{K}\label{eqn:assumption:revealing}
\end{align}
\noindent for each topic ${c}_i \in \SET{C}$.
\end{assumption}
We estimate background knowledge $\EV{E}_1$ by selecting a training data-set, denoted $\SET{T}$, consisting of (label, advert) pairs; where the label is the category in $\SET{C}$ associated with the corresponding advert. For example, when testing for evidence of a single, sensitive topic, called ``Sensitive'', $\SET{T}$ contains items labeled ``Sensitive' or ``Other'', where ``Other'' is the label for the uninteresting, catch-all topic $c_0$. In this way $\SET{T}$ approximates the prior observation evidence available at the start of the query session so that $\SET{T}$ is an estimator for $\EV{E}_1$.
Text processing of $\SET{T}$ produces a dictionary $\SET{D}$ of \emph{keyword} features. This processing removes common English language high-frequency words and maps each of the remaining keywords to a stemmed form by removing standard prefixes and suffixes such as ``--ing'' and ``--ed''. The dictionary $\SET{D}$ represents an estimate of the known universe of keywords according to the background knowledge contained in the training data.
Text appearing in the adverts in a response page is preprocessed in the same way as \SET{T} to produce a sequence of keywords from $\SET{D}$ for each advert; denoted $W=\{w_1,w_2,\cdots,w_{|W|}\}$. Words not appearing in $\SET{D}$ are ignored in our experimental setup for simplicity since sessions are short. In an operational setting it is possible, for example, to update $\SET{D}$ when new keywords are encountered and refactor $\EV{E}_1$ accordingly.
Let $n_{\SET{D}}(w | W) := \vert \{i:i \in \{1,\cdots,|W|\},w_i = w\} \vert$, denote
the number of times an individual keyword $w \in \SET{D}$ occurs in a sequence $W=\{w_1,w_2,\cdots,w_{|W|}\}$. The relative frequency of an individual keyword $w\in\SET{W}$ is therefore,
\begin{align}
\phi_{\SET{D}}(w | W) &=
\frac{n_{\SET{D}}(w | W)}{\sum_{w\in\SET{D}} n_{\SET{D}}(w | W)}
\label{eqn:phi:define}
\end{align}
\noindent recalling that only keywords $w$ appearing $\SET{D}$ are admissible due to the text preprocessing in our setup.
Let $c_i \in \SET{C}$ be a sensitive topic of interest,
and let $\SET{T}(c_i)$ denote the subset of $\SET{T}$ where the labels corresponds to $c_i$. Let $T(\SET{C})$ denote the set of adverts labelled for any topic in $\SET{C}$. The \textbf{PRI}{} estimator for $\FPRI(\VEC{x}, \omega_k)$ given adverts $a_k$ appearing on the result page for query number $k$, is\footnote{Note that in \cite{mac2015don} the expression given for $\widehat{M}_k (\VEC{c}_i, \omega_k)$ is incorrect and is corrected here.}:
\begin{align}
\widehat{M}_k (\VEC{c}_i, \omega_k) &= \sum_{w\in{\SET{D}}}\left( \frac{\phi_{\SET{D}}(w | \SET{T}(c_i))}{ \phi_{\SET{D}}(w | \SET{T}(\SET{C})) } \cdot {\phi_{\SET{D}}(w|a_k)} \right) \label{eqn:mk001}
\end{align}
where we concatenate all of the advert text on page $k$ into a single sequence of keywords and $\psi_{\SET{D}}(w|a_k)$ is the relative frequency of $w$ within this sequence. Similarly, concatenating all of the keywords in the training set $ \SET{T}(c_i)$, respectively $ \SET{T}(\SET{C})$, into a single sequence then $\phi_{\SET{D}}(w | \SET{T}(c_i))$, respectively $\phi_{\SET{D}}(w | \SET{T}(\SET{C}))$, is the relative frequency of $w$ within that sequence.
\subsection{Tuning the \textbf{PRI}{} Estimator}
\label{sec:sparsity}
The quantity $\psi_{\SET{D}}(w|a_k)$ in the expression for the \textbf{PRI}{} estimator, \eqref{eqn:mk001}, is problematic when the adverts $a_k$ on page $k$ do not contain any of the topic keywords in dictionary $\SET{D}$ i.e. when $a_k=\emptyset$, indicating there is no detectable evidence of a particular topic. To be consistent with the definition of \textit{\ensuremath{\epsilon}-Indistinguishability } in Section \ref{def:eps:001}, should result in a \textbf{PRI}{} score of one for that topic. We therefore replace $\phi_{\SET{D}}(w|a_k)$ with
\small
\begin{align}
\psi_{0, \SET{D}}(w | a_k) &=
\begin{cases}
\phi_{\SET{D}}(w | {a_k})
& \mbox{if } a_k\ne \emptyset \\
1
& \mbox{otherwise}
\end{cases}
\end{align}
\normalsize
Training data is based on a sample of all possible adverts for a particular topic. We may be unlucky so that during the training phase we fail to observe adverts containing infrequently occurring keywords for a particular topic. In this case the relative frequency of such a keyword will be zero and it will not contribute when estimating \textbf{PRI}{} if encountered in an advert. To address this we introduce a Laplace smoothing parameter $\lambda$ as follows
\small
\begin{align}
n_{\lambda, \SET{D}}(w | W) &= \lambda + n_{\SET{D}}(w | W) \\
\phi_{\lambda,\SET{D}}(w | W)
&= \frac{n_{\lambda,\SET{D}}(w | W)}{\sum_{w\in\SET{D}} n_{\lambda, \SET{D}}(w | W)}
\\
\psi_{\lambda, \SET{D}}(w | a_k) &=
\begin{cases}
\phi_{\lambda,\SET{D}}(w | {a})
& \mbox{if } a_k\ne \emptyset \\
1
& \mbox{otherwise}
\end{cases}
\label{eqn:phi:define2}
\end{align}
\normalsize
The parameter $0\le \lambda < 1$ enforces a minimum frequency of $1/|\SET{D}|$ on every keyword.
The expression \eqref{eqn:mk001} is adjusted correspondingly to give a new estimator we call \textbf{PRI+}:
\begin{align}
\ensuremath{\widehat{\mathbb{M}}_{k}} (\VEC{c_i}, \omega_k) &= \sum_{w\in{\SET{D}}}\left( \frac{ \phi_{\lambda, \SET{D}}(w | \SET{T}(c_i))}{ \phi_{\lambda, \SET{D}}(w | \SET{T}(\SET{C})) } \cdot \psi_{\lambda, \SET{D}}(w | {a_k}) \right) \label{eqn:mk001:2}
\end{align}
We will use the \textbf{PRI+}{} estimator, given by \eqref{eqn:mk001:2}, from now on in this paper, unless stated otherwise. In our experiments we find empirically, through verification with the training data, that choosing the parameter $\lambda = 0.001$ worked well.
\subsection{The \textbf{PDE}{} Estimator}
\label{sec:dpe}
Substituting the \textbf{PRI+}{} estimator $\widehat{\FPRI{}}$ for $\FPRI{}$ in \eqref{eqn:prop:estim:2b} gives the \textbf{PDE}{} estimator
\small
\begin{align}
\ensuremath{\widehat{\epsilon}_{*,k}}{} &= \left\vert\log\left(\frac{\widehat{\FPRI[k]}(\VEC{c}_i, \omega_{k})}{\widehat{\FPRI[k]}(\VEC{c}_{-i}, \omega_{k})}\frac{\widehat{\FPRI[1]}(\VEC{c}_{-i}, \omega_{1})}
{\widehat{\FPRI[1]}(\VEC{c}_i, \omega_{1})}\right)\right\vert
\label{eqn:dpe:estim:1}
\end{align}
\normalsize
From Proposition~\ref{prop:estim}, the \textbf{PDE}{} estimator in \eqref{eqn:dpe:estim:1} can be interpreted directly as the best possible level of {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} a user can claim in the case $m=2$. We report the maximum value of \textbf{PDE}{} measured by probe step in our experiments to show the worst possible {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} scenario for \ensuremath{\mathcal{U}}{}. We also report the median value of \textbf{PDE}{} as a representative bound for approximately $50\%$ of the samples. An example of reporting is shown in Table~\ref{tbl:pde:example:measured} for the reference topic ``gay''.%
\begin{table}[h]
\captionsetup{position=bottom,skip=0pt,belowskip=0pt}
\caption{Measured \ensuremath{\widehat{\epsilon}_{*,k}} for Reference Topic versus Any Other Topic, reported as ``max (median)'', by Probe Query Sequence\label{tbl:pde:example:measured}}
\setlength{\tabcolsep}{0.35cm}
\begin{scriptsize}
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{gay} & 64 (33) & 47 ( 5) & 72 (25) & 48 (25) & 48 (19) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}\\[-4mm]
\end{scriptsize}
\end{table
For example, from Table~\ref{tbl:pde:example:measured}, a reported maximum value of \textbf{PDE}{} of $47\%$ in the second column indicates that the difference in probabilities that \ensuremath{\mathcal{U}}{} is uniquely interest in the reference topic versus being interested in any other topic is \emph{at least} $47\%$ in the worst case by probe step $5$. The median value of $25\%$ in parentheses in the Probe 3 and 4 columns indicates that the difference in probabilities can be expected to be at least $25\%$ in $50\%$ of cases by probes $3$ and $4$. Overall the results suggest that {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} is unlikely to constitute a reasonable defence in this case.
Reported values of \textbf{PDE}{} may increase, or decrease, during a session as individual queries are judged as more, or less, revealing by the \textbf{PDE}{} estimator. Inspection of the query scripts generated for the topic $c_i=\text{Gay}$, for example, shows that the queries associated with probe step $3$ are \emph{same sex relationships} and \emph{how do i know if I'm gay}, both of which appear revealing. The queries from the test script corresponding to probe steps $4$ and $5$ are \emph{HIV symptoms}, \emph{HIV treatment}, \emph{HIV men} and \emph{aids men} which may not point as distinctly to specific interest in the $c_i=\text{Gay}$ as they could reasonably be associated with health concerns.
The zeroth probe in a session is always run first, before any other query, to establish a baseline \textbf{PRI+}{} score for the session. As a result the measured \textbf{PDE}{} values for the zeroth probe is always $0$ for both maximum and median values and is not reported in our results.
One popular approach to designing defences of {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} is to attempt to \emph{hide in the crowd}. For example, by injecting varying degrees of noise in the stream of observations $\{\omega_j\}$ in the hope that \ensuremath{\mathcal{S}}{} will not detect the true sub-stream of sensitive events.
In \cite{mac2015don}, the authors observe that varying click patterns is seen to change the absolute volume of adverts appearing on a page. As both user clicks and queries are potential indicators of user interest for an observer we test injected noise from both queries and clicks as possible defence strategies.
An alternative tactic is to invert the previous approach by instead attempting to \emph{hide in plain sight}. By choosing a non-sensitive \emph{proxy topic}, chosen to attract personalised content \ensuremath{\mathcal{U}}{} can then carefully hide true, sensitive queries in a stream of proxy topic queries.
By demonstrating clear interest in a \emph{proxy} non-sensitive topic \ensuremath{\mathcal{U}}{} may tip the balance of probability toward the proxy topic by drawing the attention of the observer \ensuremath{\mathcal{S}}{}.
\section{Experimental Results}
\label{sec:experiment:results}
\subsection{Preliminaries}
\label{sec:topical:queries}
To facilitate easy comparison we use the same experimental data collection setup as \cite{mac2015don}. We summarise the key elements here with additional detail in the Appendix and refer the reader to \cite{mac2015don} for full details.
User interest topic categories taken from \cite{mac2015don}, are used in our experiments. Of the user interest topics, (i) ten are sensitive categories associated with subjects generally identified as causes of discrimination (medical condition, sexual orientation \emph{etc}) or sensitive personal conditions (gambling addiction, financial problems \emph{etc}), (ii) a further sensitive topic is related to London as a specific destination location, providing an obviously interesting yet potentially sensitive topic that a recommender system might track, (iii) the last topic is a catch-all category labeled ``Other''.
To construct sequences of queries for use in test sessions, we select a \emph{probe query}, providing a predefined sampling point for data collection. Numbering the probes in a session starting from $0$, the zeroth query issued in every session is a probe query. The zeroth probe is used to establish the baseline for calculations of the \textbf{PDE}{} estimator for subsequent probe queries. The \textbf{PDE}{} estimator, from \eqref{eqn:dpe:estim:1}, of the zeroth probe in a session is $0$ and so is not included in reports of experimental results. Measurements of \textbf{PDE}{} values are reported for each of the probe queries $1$--$5$ during experiments providing a consistent sample for analysis.
In our experiments, when implementing the ``Proxy Topic'' defence model, we choose three uninteresting, proxy topics likely to attract adverts, namely \emph{tickets for music concerts}, searching for \emph{bargain vacations} and \emph{buying a new car}.
All scripts were run for $3$ registered users and $1$ anonymous user on the Google search engine, yielding a data set consisting of $21,861$ probe queries in total across all of the test user interest topics. Test data was divided into individual test data sets based on different test configurations with each test data set consisting of approximately $1,000$ probe queries.
A separate hold-back was created for a common training data set of approximately $1,000$ queries. The \textbf{PDE}{} estimator in \eqref{eqn:dpe:estim:1} uses the training data-set to model the prior background knowledge $\EV{E}_0$. We do not re-train \textbf{PDE}{} during testing as new adverts are encountered. Experimental measurements of \textbf{PDE}{} are with respect to the common training set for consistent comparison.
All queries in a test session were automatically labelled with the intended topic of the test session as given by the query script used. For example, all queries from a session about ``prostate`` are labeled as ``prostate'' including probe queries. In this respect the labels capture intended behaviour of queries, rather than attempting an individual interpretation of specific query keywords during a user session. Test data is automatically divided into $7$ folds for processing so that, reported statistics are taken over $7$ distinct, randomised sub-samples of test data.
Before proceeding to testing with \textbf{PDE}, we verify \textbf{PRI+}{} by comparing its detection capability with previous results obtained in \cite{mac2015don} for the \textbf{PRI}{} estimator \POL{and compare the performance of \textbf{PRI+}{} with alternative implementations using Naive Bayes and Support Vector Machine as sensitive topic detectors}.
Comparison results between \textbf{PRI}{} and \textbf{PRI+}{} are shown in Table~\ref{tbl:old:new}. and were produced by processing data taken from \cite{mac2015don} but applying the \textbf{PRI+}{} estimator to decide which topic is detected. For comparison with \cite{mac2015don}, we declare a topic $c_i$ has been detected during a query session, consisting of $5$ probe queries, if \emph{at least one} of the $5$ probe queries is detected as topic $c_i$. For comparison, detection results for the \textbf{PRI}{} estimator from Table XIV(b) in \cite{mac2015don}, are reproduced as Table~\ref{tbl:old:new}(b). The True Detection rates using \textbf{PRI+}{} estimator are better or equal for each topic than the rates reported in \cite{mac2015don}. The False Detection rates are also better or equal in the case of all topics tested comparing favourably with the results obtained in \cite{mac2015don}.
\POL{Comparison of \textbf{PRI+}{} with alternative implementations was performed by taking results from Multinomial Naive Bayes (NB) and Linear SVM (SVM) classifiers to estimate the probabilities in the definition of $\FPRI$ in \eqref{eqn:m:u:2}. The intent of the comparison is to determine which of the NB, \textbf{PRI+}{} and SVM estimators detect privacy threats, using the definition of $\FPRI$ in \eqref{eqn:m:u:2}, for test items previously labeled as ``sensitive'' by examining the topic of the query used. To qualify as a privacy threat we choose a value of $e^{\epsilon} > 1.1$. We expect precision to be substantially less than 100\% for all estimators because the threshold will filter out weaker detections where $1.0 < e^{\epsilon} \leq 1.1$.}
\begin{figure}[h]
\centering
\includegraphics[scale=0.55]{compare_models}
\caption{Comparison of Naive Bayes, \textbf{PRI+}{} and Support Vector Machine estimators. (as Threat Detection Rate by Topic)\label{fig:compare:nb:pri:svm}}
\end{figure}%
\POL{Other than varying how $\FPRI$ was estimated, all other inputs and calculations were identical. A common test data set was constructed by selecting 5,500 result pages for each sensitive topic and then randomly selecting an additional 5,500 result pages labeled for the non-sensitive topic. In this way each sensitive topic had a balanced verification data set of 11,000 labeled items. Each verification data-set was divided randomly into $20\%-80\%$ test--training sets and calculations repeated 5 times for 5-fold verification of each of the NB, \textbf{PRI+}{} and SVM estimators. The Multinomial Naive Bayes and Linear SVC modules from the Python Sklearn package were used to construct the NB and SVM estimators, \cite{scikit-learn}.}
\POL{After common preprocessing each of the NB, \textbf{PRI+}{} and SVM classifiers were trained and probability estimates captured for the 5-fold test data-sets. A threat is declared ``detected'' if the calculated valued of $\FPRI$ for the sensitive topic exceeds $1.0$. Precision of sensitive topic threat detection is shown by topic in Figure~\ref{fig:compare:nb:pri:svm} for the NB, \textbf{PRI+}{} and SVM approaches.}
\POL{The results Figure~\ref{fig:compare:nb:pri:svm} indicate that that the \textbf{PRI+}{} estimator detects significantly more true-positive detection results than either of the NB or SVM estimators for all sensitive topics tested. The initial detection sensitivity of each of these estimators is influenced by the labelling assigned to examples in the training set. We adopt the perspective that privacy tools should err on the side of caution so that high detection sensitivity in the initial ``out of the box'' stage is a prudent approach. In a real-world application of \textbf{PRI+}{} the user would provide incremental training examples over time reflecting their tolerance of privacy risk and so tune \textbf{PRI+}{}. }
\subsection{Establishing a Baseline}
\label{sec:test:setup}
We begin with a sequences of queries, interleaved with probe queries, in what we term a ``no click, no noise'' model. Here there is no injected noise and no items are clicked on any of the search results pages. This model provides a baseline, where the queries alone are available to the recommender to learn about a user session as it progresses. Measurements of \textbf{PDE}{} for all topics using the ``no click, no noise'' model are shown in Table~\ref{tbl:base:noclick:nonoise}.
\begin{table}[h]
\captionsetup{position=bottom,skip=0pt,belowskip=0pt}
\caption{Measured \ensuremath{\widehat{\epsilon}_{*,k}} for Reference Topic versus Any Other Topic, reported as ``max (median)'', by Probe Query Sequence\label{tbl:base:noclick:nonoise}}
\setlength{\tabcolsep}{0.30cm}
\begin{scriptsize}
\subfloat[No Click, No Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 56 (52) & 56 (52) & 56 (52) & 56 (52) & 56 (52) \\
\textbf{bankrupt} & 1 ( 1) & 55 (43) & 55 (39) & 58 (48) & 56 (48) \\
\textbf{diabetes} & 40 (38) & 40 (38) & 40 (38) & 40 (38) & 40 (38) \\
\textbf{disabled} & 9 ( 9) & 9 ( 9) & 9 ( 9) & 40 (40) & 40 (33) \\
\textbf{divorce} & 41 (31) & 75 (65) & 56 (46) & 79 (68) & 79 (68) \\
\textbf{gambling} & 16 (12) & 18 (16) & 66 ( 4) & 57 (17) & 18 ( 3) \\
\textbf{gay} & 64 (33) & 47 ( 5) & 72 (25) & 48 (25) & 48 (19) \\
\textbf{location} & 10 ( 2) & 11 ( 3) & 11 (10) & 18 ( 7) & 18 ( 9) \\
\textbf{payday} & 2 ( 2) & 2 ( 2) & 21 ( 2) & 2 ( 2) & 2 ( 2) \\
\textbf{prostate} & 52 (17) & 52 (17) & 52 (17) & 52 (17) & 52 (17) \\
\textbf{unemployed} & 7 ( 5) & 7 ( 6) & 7 ( 6) & 13 ( 7) & 7 ( 7) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
}\\[-4mm]
\end{scriptsize}
\end{table
For the health-related topics $\text{Anorexia, Diabetes, Prostate, Bankrupt, Divorced, Gay}$ the reported results are high, indicating lack of plausible deniability for each of these topics.
It is concerning that personal circumstances, health status and sexual orientation appear to be the most revealing topics according to our experiments. In the case of the topic $\text{Disabled}$ there is more cause of concern about {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{} as the session progresses. On inspection of the associated query script this appears to be again related to the specificity of the queries at each probe step. At the beginning of this script the queries are related to availability of services -- for example, \emph{locations of disabled parking} -- while later queries are more specific to named conditions -- for example, \emph{treatment for spina bifida}.
The topics $\{Location, Payday, Unemployed\}$ appear among the topics of least concern from the perspective of {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{}. Both of the topics Payday and Unemployed asked queries about availability of social support services whereas queries for the topic Bankrupt asked about availability of paid professional services such as lawyers and accountants. It is perhaps an illustration of the motivations of a for-profit service where users seeking social supports are of less interest than users seeking expensive paid services.
Overall, measurements of \textbf{PDE}{} in experiments appear to agree with expectations from inspection of the underlying queries.
Our results suggest that queries are a strong signal to the observer of user interest, and that estimates from \textbf{PDE}{} appear to distinguish queries that are strongly revealing of specific topic interest from more generic queries where plausible deniability is clearer.
\subsection{The Effect of Random Noise Injection}
\label{sec:testing:noise}
Following from Section~\ref{sec:by:design}, we now consider the impact of injecting non-informative queries chosen at random from our popular query list into a user session. We simply refer to these as ``random noise'' queries. We consider three levels of random noise queries for testing purposes:
\begin{description}[style=unboxed,leftmargin=0cm]
\item [``Low Noise''] The automation scripts select uninteresting queries uniformly at random from the top-query list and inject a single random noise query after every topic-specific query so that the ``signal-to-noise ratio'' of sensitive to noise queries in this case is $1:1$.
\item [``Medium Noise''] Here the automation scripts inject two randomly selected queries after each topic-specific query for a signal to noise ration of $1:2$.
\item [``High Noise''] In this noise-model with the highest noise setting, three random noise queries are injected, resulting in a signal-to-noise ratio of $1:3$.
\end{description}
Note also that the automation scripts were configured to ensure the relevant number of noise queries was always injected \emph{immediately before} each probe query. Our intention was to construct a ``worst case'' for detection of learning, where probe queries are always separated from sensitive user queries by the specified number of noise queries.
\begin{table}[h]
\captionsetup{position=bottom,skip=0pt,belowskip=0pt}
\caption{Measured \ensuremath{\widehat{\epsilon}_{*,k}} for Reference Topic versus Any Other Topic, reported as ``max (median)'', by Probe Query Sequence\label{tbl:base:noclick:allnoise}}
\setlength{\tabcolsep}{0.30cm}
\begin{scriptsize}
\subfloat[No Click, Low Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 54 (45) & 54 (45) & 54 (45) & 54 (45) & 54 (45) \\
\textbf{bankrupt} & 16 ( 9) & 56 (50) & 52 (39) & 54 (45) & 56 (45) \\
\textbf{diabetes} & 46 (35) & 46 (35) & 46 (35) & 46 (35) & 46 (35) \\
\textbf{disabled} & 9 ( 3) & 9 ( 8) & 9 ( 7) & 33 ( 7) & 40 (32) \\
\textbf{divorce} & 13 ( 7) & 123 ( 8) & 54 ( 8) & 85 ( 6) & 85 ( 6) \\
\textbf{gambling} & 18 (16) & 18 (16) & 52 (18) & 18 (10) & 18 (18) \\
\textbf{gay} & 73 (61) & 73 (70) & 76 (46) & 79 (74) & 79 (70) \\
\textbf{location} & 18 (16) & 18 (10) & 18 (10) & 18 (10) & 18 (10) \\
\textbf{payday} & 3 ( 2) & 3 ( 2) & 4 ( 3) & 4 ( 3) & 4 ( 3) \\
\textbf{prostate} & 21 (16) & 21 (16) & 21 (16) & 21 (16) & 21 (16) \\
\textbf{unemployed} & 7 ( 3) & 7 ( 3) & 13 ( 9) & 13 ( 9) & 13 ( 9) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\subfloat[No Click, Med Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 55 (53) & 53 (53) & 53 (53) & 53 (53) & 53 (53) \\
\textbf{bankrupt} & 11 ( 8) & 48 (33) & 51 (43) & 52 (38) & 52 (38) \\
\textbf{diabetes} & 38 (38) & 38 (38) & 38 (38) & 38 (38) & 38 (38) \\
\textbf{disabled} & 4 ( 4) & 8 ( 7) & 1 ( 1) & 40 (36) & 40 (36) \\
\textbf{divorce} & 19 ( 9) & 65 (31) & 44 (31) & 72 (50) & 72 (50) \\
\textbf{gambling} & 18 (16) & 18 (17) & 18 (18) & 31 ( 3) & 18 (10) \\
\textbf{gay} & 89 (68) & 89 (69) & 88 (64) & 93 (73) & 93 (64) \\
\textbf{location} & 18 (10) & 18 (10) & 18 ( 7) & 18 ( 7) & 10 ( 7) \\
\textbf{payday} & 6 ( 3) & 6 ( 3) & 6 ( 3) & 6 ( 2) & 6 ( 1) \\
\textbf{prostate} & 32 (14) & 32 (14) & 18 (13) & 18 (13) & 18 (13) \\
\textbf{unemployed} & 13 ( 5) & 13 (10) & 13 ( 7) & 13 ( 9) & 7 ( 4) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\subfloat[No Click, High Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 48 (48) & 48 (48) & 48 (48) & 48 (48) & 48 (48) \\
\textbf{bankrupt} & 16 (10) & 65 (51) & 65 (48) & 65 (49) & 65 (49) \\
\textbf{diabetes} & 41 (38) & 41 (38) & 41 (38) & 41 (38) & 41 (38) \\
\textbf{disabled} & 9 ( 9) & 9 ( 9) & 9 ( 5) & 9 ( 7) & 9 ( 8) \\
\textbf{divorce} & 41 (27) & 75 (38) & 56 (22) & 75 (29) & 75 (29) \\
\textbf{gambling} & 21 (16) & 21 ( 3) & 21 ( 4) & 29 (16) & 18 ( 4) \\
\textbf{gay} & 86 (64) & 86 (64) & 80 (43) & 94 (59) & 94 (59) \\
\textbf{location} & 10 (10) & 8 ( 8) & 8 ( 8) & 18 (13) & 18 (13) \\
\textbf{payday} & 3 ( 2) & 4 ( 2) & 4 ( 2) & 4 ( 2) & 3 ( 1) \\
\textbf{prostate} & 17 (15) & 17 (15) & 17 (15) & 17 (15) & 17 (15) \\
\textbf{unemployed} & 10 ( 7) & 13 ( 7) & 13 ( 7) & 13 ( 7) & 13 ( 7) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-6mm]
\end{scriptsize}
\end{table
Table~\ref{tbl:base:noclick:allnoise}(a-c) shows the measured \textbf{PDE}{} values for Low, Medium and High levels of noise respectively for the ``no click'' model. The \textbf{PDE}{} values for all levels of noise are similar to the ``no click, no noise'' baseline values in Table~\ref{tbl:base:noclick:nonoise}.
Overall, there is no consistent reduction in values across all topics for all noise levels, indicating that injecting random noise queries does not have a consistent effect. In some cases, such as topic $\text{Gay}$, measured values of \textbf{PDE}{} increase for all noise levels indicating that noise injection \emph{worsens} the user's ability to assert {\text{(\ensuremath{\epsilon, m})--Plausible Deniability }}{}.
These results indicate that even the ``High Noise'' model fails to reduce the measured values of \textbf{PDE}{} in a coherent way, so that injecting random noise has not improved plausible deniability significantly with any consistency. We conclude that injection of random noise, even at substantial levels, is not observed to provide a useful defence for plausible deniability in our experiments.
\subsection{The Effect of Click Strategies}
\label{sec:testing:clicks}
We now consider whether it is possible to disrupt search engine learning by careful clicking of the links on response pages.
Intuitively, from the search engine's point of view, clicking on links is a form of active feedback by a user and so potentially informative of user interests. This is especially true when, for example, a user is carrying out exploratory search where their choice of keywords is not yet well-tuned to their topic of interest. Previous studies have also indicated that there is good reason to believe that user clicks on links are an important input into recommender system learning. In \cite{mac2015don} (Section $6.4$), user clicks emulated using the ``Click Relevant'' click-model were reported to result in increases of $60\%$ -- $450\%$ in the advert \emph{content}, depending on the ``Sensitive' topic tested.
We consider four different click strategies to emulate a range of user click behaviours:
\begin{description}[style=unboxed,leftmargin=0cm]
\item [``No Click''] No items are clicked on in the response page to a query. This user click-model does not provide additional user preference information to the recommender system due to click behaviour. This click model is used in the baseline measurements presented in Sections~\ref{sec:test:setup}.
\item [``Click Relevant''] Given the response page to a query, for each search result and advert we calculate the Term-Frequency (TF) of the visible text with respect to the keywords associated with the test session topic of interest. When $TF > 0.1$ for an item, the item is clicked, otherwise it is not clicked. This user click-model provides relevant feedback to the recommender system about the information goal of the user.
\item [``Click Non-relevant''] TF is calculated for each item with respect to the category of interest for the session in question as for the ``Click Relevant'' click-model, \emph{except} that items are clicked when the TF score is below the threshold and so they are deemed non-relevant to the topic, that is when $TF \leq 0.1$. This user click-model attempts to confuse the recommender system by providing feedback that is not relevant to the true topic of interest to the user.
\item [``Click All''] All items on the response page for a query are clicked. This user click-model gives the recommender system a ``noisy'' click signal, including clicks on items relevant and non-relevant to the user's information goal.
\item [``Click 2 Random Items''] Two items appearing on the response page for a query are selected uniformly at random with replacement and clicked.
\end{description}
In all cases, when uninteresting, noise queries are included in a query session, the relevant user click-strategy is also applied to the result pages of these queries. In this way we hope to avoid providing an obvious signal to the recommender system that might differentiate uninteresting queries from queries related to sensitive topics. Items on the result page in response to probe queries are \emph{not} clicked so that the probe query does not provide any additional information to the recommender system.
\begin{table}[h]
\captionsetup{position=bottom,skip=0pt,belowskip=0pt}
\caption{Measured Plausible Deniability versus any other tested topics as probability of interest, by Probe Query Sequence when the true topic of interest is ``Other'' with range $(\mu \pm 3\sigma)$\label{tbl:base:allclicks:nonoise}}
\setlength{\tabcolsep}{0.30cm}
\begin{scriptsize}
\subfloat[Click Relevant, No Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 59 (50) & 59 (50) & 59 (50) & 59 (50) & 59 (50) \\
\textbf{bankrupt} & 16 ( 8) & 65 (42) & 65 (36) & 59 (40) & 54 (38) \\
\textbf{diabetes} & 36 (36) & 36 (36) & 36 (36) & 36 (36) & 36 (36) \\
\textbf{disabled} & 7 ( 4) & 7 ( 4) & 9 ( 9) & 40 ( 4) & 40 ( 7) \\
\textbf{divorce} & 30 (24) & 30 ( 9) & 30 ( 9) & 30 ( 8) & 30 ( 8) \\
\textbf{gambling} & 6 ( 0) & 18 (16) & 32 (16) & 18 (16) & 18 ( 5) \\
\textbf{gay} & 92 (51) & 92 (77) & 78 (51) & 94 (72) & 94 (80) \\
\textbf{location} & 18 (18) & 10 (10) & 10 (10) & 18 (10) & 18 (10) \\
\textbf{payday} & 2 ( 2) & 2 ( 2) & 3 ( 2) & 3 ( 2) & 2 ( 2) \\
\textbf{prostate} & 17 (17) & 17 (17) & 17 (17) & 17 (17) & 17 (17) \\
\textbf{unemployed} & 13 ( 2) & 13 ( 4) & 13 ( 7) & 13 ( 7) & 7 ( 6) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\subfloat[Click Non-relevant, No Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 18 ( 5) & 22 (12) & 26 ( 5) & 31 (13) & 32 ( 6) \\
\textbf{bankrupt} & 57 ( 3) & 53 (36) & 50 (34) & 43 (33) & 48 (36) \\
\textbf{diabetes} & 4 ( 2) & 13 ( 8) & 11 ( 8) & 5 ( 3) & 11 ( 2) \\
\textbf{disabled} & 5 ( 2) & 6 ( 2) & 9 ( 3) & 29 (10) & 26 ( 8) \\
\textbf{divorce} & 49 (25) & 51 (33) & 49 (30) & 43 (29) & 43 (29) \\
\textbf{gambling} & 6 ( 2) & 18 ( 4) & 36 (24) & 35 (13) & 31 (13) \\
\textbf{gay} & 36 (33) & 75 (33) & 51 (32) & 39 (20) & 31 (27) \\
\textbf{location} & 9 ( 2) & 11 ( 1) & 7 ( 2) & 6 ( 2) & 9 ( 1) \\
\textbf{payday} & 3 ( 3) & 3 ( 1) & 4 ( 2) & 3 ( 2) & 4 ( 3) \\
\textbf{prostate} & 55 (38) & 68 (36) & 65 (48) & 61 (48) & 64 (42) \\
\textbf{unemployed} & 9 ( 1) & 6 ( 6) & 7 ( 1) & 9 ( 4) & 5 ( 2) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\subfloat[Click All, No Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 66 (57) & 66 (57) & 66 (57) & 66 (57) & 66 (57) \\
\textbf{bankrupt} & 51 (42) & 51 (42) & 51 (42) & 55 (46) & 56 (46) \\
\textbf{diabetes} & 35 (35) & 35 (35) & 35 (35) & 35 (35) & 35 (35) \\
\textbf{disabled} & 9 ( 9) & 9 ( 9) & 9 ( 9) & 31 (31) & 31 (31) \\
\textbf{divorce} & 30 ( 8) & 73 (54) & 54 (34) & 100 (49) & 100 (49) \\
\textbf{gambling} & 3 ( 1) & 16 (16) & 53 (11) & 16 ( 6) & 6 ( 2) \\
\textbf{gay} & 69 (65) & 77 (73) & 70 (60) & 82 (75) & 81 (71) \\
\textbf{location} & 18 (10) & 10 ( 6) & 10 ( 6) & 14 (10) & 18 ( 7) \\
\textbf{payday} & 2 ( 2) & 2 ( 2) & 2 ( 2) & 2 ( 2) & 2 ( 2) \\
\textbf{prostate} & 17 (17) & 17 (17) & 17 (17) & 17 (17) & 17 (17) \\
\textbf{unemployed} & 4 ( 4) & 7 ( 7) & 7 ( 7) & 7 ( 7) & 7 ( 6) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\subfloat[Click 2 Random Items, No Noise]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{anorexia} & 50 (12) & 27 ( 9) & 26 ( 9) & 36 (10) & 33 (11) \\
\textbf{bankrupt} & 5 ( 3) & 43 (33) & 39 (37) & 36 (35) & 38 (35) \\
\textbf{diabetes} & 38 ( 6) & 18 ( 7) & 17 ( 5) & 17 ( 7) & 11 ( 5) \\
\textbf{disabled} & 2 ( 1) & 4 ( 1) & 5 ( 3) & 39 (25) & 40 (25) \\
\textbf{divorce} & 24 (17) & 37 (31) & 37 (31) & 35 (25) & 35 (25) \\
\textbf{gambling} & 24 ( 0) & 7 ( 4) & 54 (23) & 33 (23) & 68 (20) \\
\textbf{gay} & 68 (68) & 68 (65) & 54 (52) & 46 (36) & 47 (42) \\
\textbf{location} & 8 ( 8) & 8 ( 8) & 8 ( 8) & 8 ( 8) & 8 ( 8) \\
\textbf{payday} & 4 ( 1) & 2 ( 2) & 4 ( 2) & 4 ( 3) & 4 ( 4) \\
\textbf{prostate} & 59 (57) & 67 (62) & 58 (56) & 60 (54) & 51 (44) \\
\textbf{unemployed} & 4 ( 3) & 8 ( 3) & 10 ( 4) & 3 ( 2) & 10 ( 1) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-6mm]
\end{scriptsize}
\end{table
Measure values of \textbf{PDE}{} are shown in Table~\ref{tbl:base:allclicks:nonoise}. As random noise injection had no observable effect on measurements of \textbf{PDE}{} for different click models in experiments, only the ``No Noise'' results are presented here for space reasons.
Taken overall, the results in Table~\ref{tbl:base:allclicks:nonoise}(a) for the ``non-relevant click, no noise'' model suggest clicking on non-relevant advert items is the best strategy of the click models tested. The only difference between the ``non-relevant click'' model and other click models is that non-relevant items \emph{only} are clicked, whereas in other click models it is possible that relevant items are clicked. It seems reasonable to postulate that clicking on relevant items provides ``fine-tuned'' feedback about user interests which is more informative for the observer. Clicking on non-relevant items may divert attention to a modest degree, but not to the extent of masking the sensitive topic revealed by the query.
Comparing the baseline ``No Click'' \textbf{PDE}{} observations in Table~\ref{tbl:base:noclick:nonoise} each of the subtables in Table~\ref{tbl:base:allclicks:nonoise} shows similar lack of consistency to the noise injection models. In out experiments there is no consistent change observed in \textbf{PDE}{} across topics due to variation in the click patterns tested. As with the noise injection case, there are sporadic increases and decreases in values of \textbf{PDE}{} but the lack of overall consistency makes using click models as a defence impractical.
It would appear in summary, that clicks transmit information to the observer, but not as consistently as does a revealing query. Consequently none of the user click-models tested appear to change the baseline level of plausible deniability associated with the query in a predictable way so that there is no globally discernible pattern with which to construct practical defence tools based on clicks.
\subsection{The Effect of Proxy Topics}
\label{sec:proxy:topic}
The next privacy protection strategy we consider is the introduction of proxy topics. In this case sequences of queries, with each sequence related to a single proxy topic which is not sensitive for the user but capable of attracting personalised advert content, are injected into a user session. The idea here is that each such sequence of queries emulates a user session where the proxy topic is the topic of interest. In this way we hope to misdirect learning by the search engine of user interests. The results in Section \ref{sec:testing:noise} are relevant here since they suggest that isolated, individual queries -- such as randomly selected noise queries -- tend not to provoke search engine learning. Our hope is that this can be exploited by inverting the notion of random noise injection so that individual \emph{sensitive} queries are injected as the noise in proxy topic sessions. Isolated sensitive queries will hopefully not provoke learning whereas the larger number of uninteresting proxy sessions will. In this way we can misdirect learning by the observer.
In out tests the following proxy topics are used:
\begin{description}[style=unboxed,leftmargin=0cm]
\item[\textbf{Tickets}] Searching for tickets for events in a well-known local stadium
\item[\textbf{Vacation}] Queries related to a vacation such as flights and accommodation.
\item[\textbf{Car}] Searches by a user seeking to trade in and change their car.
\end{description}
and related queries are constructed by selecting related keywords through the same process as was used for the sensitive topics.
\begin{table}[h]
\captionsetup{position=bottom,skip=0pt,belowskip=0pt}
\caption{Measured Plausible Deniability versus any other tested topics as probability of interest, by Probe Query Sequence when the true topic of interest is ``Other'' with range $(\mu \pm 3\sigma)$\label{tbl:base:proxy:topics}}
\setlength{\tabcolsep}{0.30cm}
\begin{scriptsize}
\subfloat[All Click and Noise Models]{
\begin{tabular}{@{}lccccc@{}}
\textbf{\shortstack[l]{Reference\\Topic\\{}}} & \textbf{Probe 1} & \textbf{Probe 2} & \textbf{Probe 3} & \textbf{Probe 4} & \textbf{Probe 5} \\
\cmidrule[0.5pt]{1-6}
\textbf{all topics} & 0 ( 0) & 0 ( 0) & 0 ( 0) & 0 ( 0) & 0 ( 0) \\
\cmidrule[0.5pt]{1-6}
\end{tabular}
} \\[-3mm]
\end{scriptsize}
\end{table
Proxy topic query scripts where constructed by selecting a sensitive topic, and then selecting an uninteresting proxy topic from the list of $3$ proxy topics. Having decided on a sensitive query we wish to issue, we select at least three and no more than four queries related to the proxy topic from a prepared list of proxy topic queries. We next randomly shuffle the order of the selected sensitive and proxy topic queries. In this way there is always a subgroup of at least two proxy topic queries next to each other in each query session. Finally, for testing purposes, we place a probe query before and after each block of 3-4 proxy + 1 sensitive queries to measure changes in \textbf{PRI+}{} score. We repeat this exercise using the same proxy topic until a typical query session consisting of $5$ probe queries is created.
Data was collected for $2,300$ such proxy topic sessions. This included each of the sensitive topics and each of the click models described in previous sections. The same \textbf{PRI+}{} and \textbf{PDE}{} setup, including the same training set, as before was used to process the search results.
Measured detection rates are shown in Table~\ref{tbl:base:proxy:topics}. The measured probability calculated from \textbf{PDE}{} is $0$ for all topics and for all click-models tested. That is, we find it is possible to claim full plausible deniability of interest in all of the topics tested. Since our detection approach is demonstrated to be notably sensitive to observer learning in earlier sections, we can reasonably infer that this result is not due to a defect in the detection methodology but rather genuinely reflects successful misdirection of the search engine away from sensitive topics.
This result is encouraging, especially in light of the negative results in previous sections for other obfuscation approaches. It suggests use of sequences of queries on uninteresting proxy topics may provide a defence of plausible deniability. The trade-offs for the user include the overhead of maintaining proxy topics and associated queries and the additional resources required to issue proxy topic queries in a consistent way. However since both of these tasks were readily automated during our testing it seems reasonable that these trade-offs could be readily managed by software in a way that is essentially transparent to the user.
\section{Conclusions and Discussion}
\label{sec:conclusions}
Our observations suggest that modern systems, such as Google, are able to identify user interests with high accuracy, exploit multiple signals, filter out uninteresting noise queries and adapt quickly when topics change. Furthermore learning appears to be sustained over the lifetime of query sessions. The power and sophistication of these systems make designing a robust defence of user privacy non-trivial.
The \textbf{PDE}{} estimator was tested via a comprehensive measurement program using online search engines to show that topic learning results in measurable impacts on the ability of a user to deny their interest in all sensitive topics tested. We find that revealing queries provide a significant signal for search engine adaptation. While user clicks provide additional feedback, we do not observe the same degree of associated learning with click behaviour as is observed with revealing queries. Overall, testing with \textbf{PDE}{} suggests that defences based on random noise injection and variable click models do not provide a reliable strategy for defence of plausible deniability.
By contrast, our experiments show that proxy topics that are uninteresting to the user but capable of generating commercial content provide observable privacy protection in our experiments. Wrapping sensitive queries in a stream of coherent proxy topic queries appears to distract the online system into adapting to the proxy topic while allowing the sensitive query noise to slip through. Our observation that proxy topics provide some relief indicates that defence of plausible deniability is not impossible, but indicates that increasingly sophisticated approaches are required in the face of ever improving search engine capability. In choosing proxy topics, for example, a user must be careful to not stimulate unintended learning of the proxy topics which may influence the utility of future search results.
Subtle tactics like proxy topics, that exploit the observer's strengths to tip the balance slightly in favour of the user, suggest an interesting avenue for future research. The simplicity of the approach means it should be possible to extend it in several ways, for example, by injecting a range of uninteresting single topic queries as additional noise in the proxy query stream it may be possible to provide additional guarantees of privacy such as \emph{k-anonymity} or \emph{differential privacy} for the sensitive topic. \POL{More investigation of proxy topics is an interesting line of future research. Experiments to compare the effectiveness of different proxy topics including, for example, inclusion of proxy topics that are more relevant to the user's known interests versus proxy topics that are less relevant to user topics. Similarly proxy topics with higher commercial value may have more potential to distract search engine learning than proxy topics with lower commercial value}
As discussed in Section~\ref{sec:related:work}, user click patterns may be used by recommender systems to rank page content, placing content likely to attract user clicks in more prominent positions on pages. In our experiments, we observed changes in volume of advert content on samples of probe query response pages. There are several plausible avenues of investigation that may help explain the mechanism behind this, such as user click patterns and the semantics of the true and noise queries chosen. The approach taken in this paper does not distinguish between items based on rank or order on the page. How the semantics of queries, the interaction between user click-models and the effect of content ranking may impact user privacy is beyond the scope of this current paper and an avenue for future research.
Overall our results point towards an arms race, where search engine capability is continuously evolving. In this setting, even if injection of proxy topic sessions were to become widely deployed then we can reasonably expect search engines to respond with more sophisticated learning strategies. Our results also point towards the fact that the text in search queries plays a key role in search engine learning. While perhaps obvious, this observation reinforces the user's need to be circumspect about the queries that they ask if they want to avoid search engine learning of their interests.
\bibliographystyle{unsrt}
\begin{scriptsize}
|
\section{Probabilistic proofs: a toy example}
There are many well known probabilistic proofs that objects with some properties exist. Such a proof estimates the probability for a random object to violate the requirements and shows that it is small (or at least strictly less than $1$). Let us look at a toy example.
Consider a $n\times n$ Boolean matrix and its $k\times k$ minor (the intersection of $k$ rows and $k$ columns chosen arbitrarily). We say that the minor is \emph{monochromatic} if all its elements are equal (either all zeros or all ones).
\begin{proposition}
For large enough $n$ and for $k=O(\log n)$, there exists a $(n\times n)$-matrix that does not contain a monochromatic $(k\times k)$-minor.
\end{proposition}
\begin{proof}
We repeat the same simple proof three times, in three different languages.
(Probabilistic language) Let us choose matrix elements using independent tosses of a fair coin. For a given $k$ colums and $k$ rows, the probability of getting a monochromatic minor at their intersection is $2^{-k^2+1}$. (Both zero-minor and one-minor have probability $2^{-k^2}$.) There are at most $n^k$ choices for columns and the same number for rows, so by the union bound the probability of getting at least one monochromatic minor is bounded by
$$
n^k \times n^k \times 2^{-k^2+1}= 2^{2k\log n - k^2 + 1}=2^{k(2\log n-k)+1}
$$
and the last expression is less then $1$ if, say, $k=3\log n$ and $n$ is suffuciently large.
(Combinatorial language) Let us count the number of bad matrices. For a given choice of columns and rows we have $2$ possibilities for the minor and $2^{n^2-k^2}$ possibilities for the rest, and there is at most $n^k$ choices for raws and columns, so the total number of matrices with monochromatic minor is
$$
n^k \times n^k \times 2\times 2^{n^2-k^2}=2^{n^2+2k\log n-k^2+1}=2^{n^2+k(2\log n - k)+1},
$$
and this is less than $2^{n^2}$, the total number of Boolean $(n\times n)$-matrices.
(Compression language) To specify the matrix that has a monochromatic minor, it is enough to specify $2k$ numbers between $1$ and $n$ (rows and column numbers), the color of the monochromatic minor ($0$ or $1$) and the remaining $n^2-k^2$ bits in the matrix (their positions are already known). So we save $k^2$ bits (compared to the straightforward list of all $n^2$ bits) using $2k\log n+1$ bits instead (each number in the range $1\ldots n$ requires $\log n$ bits; to be exact, we may use $\lceil \log n\rceil$), so we can compress the matrix with a monochromatic minor if $2k\log n+1 \ll k^2$, and not all matrices are compressible.
%
\end{proof}
Of course, these three arguments are the same: in the second one we multiply probabilities by $2^{n^2}$, and in the third one we take logarithms.
However, the compression language provides some new viewpoint that may help our intuition.
\section{A bit more interesting example}
In this example we want to put bits (zeros and ones) around the circle in a ``essentially asymmetric'' way: each rotation of the circle should change at least a fixed percentage of bits. More precisely, we are interested in the following statement:
\begin{proposition}
There exists $\varepsilon>0$ such for every suffuciently large $n$ there exists a sequence $x_0 x_1\ldots x_{n-1}$ of bits such that for every $k=1,2,\ldots,n-1$ the cyclic shift by $k$ positions produces a sequence
$$
y_0=x_k, y_1=x_{k+1},\ldots,y_{n-1}=x_{k-1},
$$
that differs from $x$ in at least $\varepsilon n$ positions \textup(the Hamming distance between $x$ and $y$ is at least $\varepsilon n$\textup).
\end{proposition}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1.0]{rotation.pdf}
\end{center}
\caption{A string $x_0\ldots x_{n-1}$ is bad if most of the dotted lines connect equal bits}\label{pic:rotation}
\end{figure}
\begin{proof}
Assume that some rotation (cyclic shift by $k$ positions) transforms $x$ into a string $y$ that coincides almost everywhere with $x$. We may assume that $k\le n/2$: the cyclic shift by $k$ positions changes as many bits as the cyclic shift by $n-k$ (the inverse one). Imagine that we dictate the string $x$ from left to right. First $k$ bits we dictate normally. But then the bits start to repeat (mostly) the previous ones ($k$ positions before), so we can just say ``the same'' or ``not the same'', and if $\varepsilon$ is small, we know that most of the time we say ``the same''. Technically, we have $\varepsilon n$ different bits, and at least $n-k\ge n/2$ bits to dictate after the first $k$, so the fraction of ``not the same'' signals is at most $2\varepsilon$. It is well known that strings of symbols where some symbols appear more often than others can be encoded efficiently. Shannon tells us that a string with two symbols with frequencies $p$ and $q$ (so $p+q=1$) can be encoded using
$$
H(p,q)=p \log\frac{1}{p} +q \log \frac{1}{q}
$$
bits per symbol and that $H(p,q)=1$ only when $p=q=1/2$. In our case, for small $\varepsilon$, one of the frequencies is close to $0$ (at most $2\varepsilon$), and the other one is close to $1$, so $H(p,q)$ is significantly less than $1$. So we get a significant compression for every string that is bad for the theorem, therefore most string are good (so good string do exist).
More precisely, every string $x_0\ldots x_{n-1}$ that does not satisfy the requirements, can be described by
\begin{itemize}
\item $k$\qquad [$\log n$ bits]
\item $x_0,\ldots,x_{k-1}$\qquad [$k$ bits]
\item $x_k\oplus x_0, x_{k+1}\oplus x_1, \ldots, x_{n-1}\oplus x_{n-k-1}$ \qquad [$n-k$ bits where the fraction of $1$s is at most $2\varepsilon$, compressed to $(n-k)H(2\varepsilon,1-2\varepsilon)$ bits]
\end{itemize}
For $\varepsilon<1/4$ and for large enough $n$ the economy in the third part (compared to $n-k$) is more important than $\log n$ in the first part.
\end{proof}
Of course, this is essentially a counting argument: the number of strings of length $(n-k)$ where the fraction of $1$s is at most $2\varepsilon$, is bounded by $2^{H(2\varepsilon,1-2\varepsilon)(n-k)}$ and we show that the bound for the number of bad strings,
$$
\sum_{k=1}^{n/2} 2^{k} 2^{H(2\varepsilon,1-2\varepsilon)(n-k)}
$$
is less than the total number of strings ($2^n$). Still the compression metaphor makes the proof more intuitive, at least for some readers.
\section{Lov\'asz local lemma and\\ Moser--Tardos algorithm}
In our examples of probabilistic proofs we proved the existence of objects that have some property by showing that \emph{most} objects have this property (in other words, that the probability of this property to be true is close to $1$ under some natural distribibution). Not all probabilistic proofs go like that. One of the exceptions is the famous Lov\'asz local lemma~(see, e.g.,~\cite{alon-spencer}). It can be used in the situations where the union bound does not work: we have too many bad events, and the sum of their probabilities exceeds $1$ even if probability of each one is very small. Still Lov\'asz local lemma shows that these bad events do not cover the probability space entirely, assuming that the bad events are ``mainly independent''. The probability of avoiding these bad events is exponentially small, still Lov\'asz local lemma provides a positive lower bound for it.
This means, in particular, that we cannot hope to construct an object satisfying the requirements by random trials, so the bound provided by Lov\'asz local lemma does not give us a randomized algorithm that constructs the object with required properties with probability close to $1$. Much later Moser and Tardos~\cite{moser,moser-tardos} suggested such an algorithm --- in fact a very simple one. In other terms, they suggested a different distribution under which good objects form a majority.
We do not discuss the statement of Lov\'asz local lemma and Moser--Tardos algorithm in general. Instead, we provide two examples when they can be used, and the compression-language proofs that can be considered as ad hoc versions of Moser--Tardos argument. These two examples are (1)~satisfiability of formulas in conjunctive normal form (CNF) and (2)~strings without forbidden factors.
\section{Satisfiable CNF}
A CNF (\emph{conjunctive normal form}) is a propositional formula that is a conjuction of \emph{clauses}. Each clause is a disjunction of \emph{literals}; a literal is a propositional variable or its negation. For example, CNF
$$
(\lnot p_1 \lor p_2 \lor p_4)\land (\lnot p_2 \lor p_3\lor \lnot p_4)
$$
consists of two clauses. First one prohibits the case when
$p_1 = \textsc{true}$, $p_2=\textsc{false}$, $p_4=\textsc{false}$; the second one prohibits the case when $p_2=\textsc{true}$, $p_3=\textsc{false}$, $p_4=\textsc{true}$. A CNF is \emph{satisfiable} if it has a \emph{satisfying assigment} (that makes all clauses true, avoiding the prohibited combinations). In our example there are many satisfying assigments. For example, if $p_1=\textsc{false}$ and $p_3=\textsc{true}$, all values of other variables are OK.
We will consider CNF where all clauses include $n$ literals with $n$ different variables (from some pool of variables that may contain much more than $n$ variables). For a random assignment (each variable is obtained by an independent tossing of a fair coin) the probability to violate a clause of this type is $2^{-n}$ (one of $2^n$ combinations of values for $n$ variables is forbidden). Therefore, \emph{if the number of clauses of this type is less than $2^n$, then the formula is satisfiable}. This is a tight bound: using $2^n$ clauses with the same variables, we can forbid all the combinations and get an unsatisfiable CNF.
The following result says that we can guarantee the satisfiability for formuli with much more clauses. In fact, the total number of clauses may be arbitrary (but still we consider finite formulas, of course). The only thing we need is the ``limited dependence'' of clauses. Let us say that two clauses are \emph{neighbors} if they have a common variable (or several common variables). The clauses that are not neighbors correspond to independent events (for a random assignment). The following statement says that if the number of neighbors of each clause is bounded, then CNF is guaranteed to be satisfisable.
\begin{proposition}
Assume that each clause in some CNF contains $n$ literals with different variables and has at most $2^{n-3}$ neighbor clauses. Then the CNF is satisfiable.
\end{proposition}
Note that $2^{n-3}$ is a rather tight bound: to forbid all the combinations for some $n$ variables, we need only $2^n$ clauses.
\begin{proof}
It is convenient to present a proof using the compression language, as suggested by Lance Fortnow. Consider the following procedure $\textsc{Fix}(C)$ whose argument is a clause (from our CNF).
%
\begin{flushleft}
\qquad \{ $C$ is false \}\\
\qquad $\textsc{Fix}(C)$:\\
\qquad\qquad $\textsc{Resample}(C)$\\
\qquad\qquad \textbf{for} all $C'$ that are neighbors of $C$:\\
\qquad\qquad\qquad \textbf{if} $C'$ is false \textbf{then} $\textsc{Fix}(C')$\\
\qquad\{ $C$ is true; other clauses that were true remain true \}
\end{flushleft}
Here $\textsc{Resample}(C)$ is the procedure that assigns fresh random values to all variables in $C$. The pre-condition (the first line) says that the procedure is called only in the situation where $C$ is false. The post-condition (the last line) says that \emph{if the procedure terminates}, then $C$ is true after termination, and, moreover, all other clauses of our CNF that were true before the call remain true. (The ones that were false may be true or false.)
Note that up to now we do not say anything about the termination: note that the procedure is randomized and it may happen that it does not terminate (for example, if all \textsc{Resample} calls are unlucky to choose the same old bad values).
\textbf{Simple observation}: if we have such a procedure, we may apply it to all clauses one by one and after all calls (assuming they terminate and the procedure works according to the specification) we get a satisfying assignment.
\textbf{Another simple observation}: it is easy to prove the ``conditional correctness'' of the procedure $\textsc{Fix}(C)$. In other words, it achieves its goal assuming that (1)~it terminates; (2)~all the recursive calls $\textsc{Fix}(C')$ achieve their goals. It is almost obvious: the $\textsc{Resample}(C)$ call may destroy (=make false) only clauses that are neighbors to $C$, and all these clauses are \textsc{Fix}-ed after that. Note that $C$ is its own neighbor, so the \textbf{for}-loop includes also a recursive call $\textsc{Fix}(C)$, so after all these calls (that terminate and satisfy the post-condition by assumption) the clause $C$ and all its neighbors are true and no other clause is damaged.
Note that the last argument remains valid even if we delete the only line that really changes something, i.e., the line $\textsc{Resample}(C)$. In this case the procedure never changes anything but still is conditionally correct; it just does not terminate if one of the clauses is false.
It remains to prove that the call $\textsc{Fix}(C)$ terminates with high probability. In fact, it terminates with probability $1$ if there are no time limits and with probability exponentially close to $1$ in polynomial time. To prove this, one may use a compression argument: we show that \emph{if the procedure works for a long time without terminating, then the sequence of random bits used for resampling is compressible}. We assume that each call of $\textsc{Resample}()$ uses $n$ fresh bits from the sequence. Finally, we note that this compressibility may happen only with exponentially small probability.
Imagine that $\textsc{Fix}(C)$ is called and during its recursive execution performs many calls
$$
\textsc{Resample}(C_1),\ldots,\textsc{Resample}(C_N)
$$
(in this order) but does not terminate (yet). We stop it at some moment and examine the values of all the variables.
\begin{lemma}
Knowing the values of the variables after these calls and the sequence $C_1,\ldots,C_N$, we can reconstruct all the $Nn$ random bits used for resampling.
\end{lemma}
\begin{proof}[Proof of the lemma] Let us go backwards. By assumption we know the values of all variables after the calls. The procedure $\textsc{Resample}(C_N)$ is called only when $C_N$ is false, and there is only one $n$-tuple of values that makes $C_N$ false. Therefore we know the values of all variables before the last call, and also know the random bits used for the last resampling (since we know the values of variables after resampling).
The same argument shows that we can reconstruct the values of variables before the preceding call $\textsc{Resample}(C_{N-1})$, and random bits used for the resampling in this call, etc.
\end{proof}
Now we need to show that the sequence of clauses $C_1,\ldots,C_N$ used for resampling can be described by less bits than $nN$ (the number of random bits used). Here we use the assumption saying each clause has at most $2^{n-3}$ neighbors and that the clauses $C'$ for which $\textsc{Fix}(C')$ is called from $\textsc{Fix}(C)$, are neighbors of $C$.
One could try to say that since $C_{i+1}$ is a neighbor of $C_i$, we need only $n-3$ bits to specify it (there are at most $2^{n-3}$ neighbors by assumption), so we save $3$ bits per clause (compared to $n$ random bits used by resampling). But this argument is wrong: $C_{i+1}$ is not always the neighbor of $C_i$, since we may return from a recursive call that causes resampling of $C_i$ and then make a new recursive call that resamples $C_{i+1}$.
To get a correct argument, we should look more closely at the tree of recursive calls generated by one call $\textsc{Fix}(C)$ (Fig.~\ref{pic:treecall}). In this tree the sons of each vertex correspond to neighbor clauses of the father-clause.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1.0]{treecall.pdf}
\end{center}
\caption{The tree of recursive calls for $\textsc{Fix}(C_1)$ (up to some moment)}\label{pic:treecall}
\end{figure}
The sequence of calls is determined by a walk in this tree, but we go up and down, not only up (as we assumed in the wrong argument). How many bits we need to encode this walk (and therefore the sequence of calls)? We use one bit to distinguish between steps up and down. If we are going down, no other information is needed. If we are going up (and resample a new clause), we need one bit to say that we are going up, and $n-3$ bits for the number of neighbor we are going to. For accounting purposes we combine these bits with a bit needed to encode the step back (this may happen later or not happen at all), and we see that in total we need at most $(n-3)+1+1=n-1$ bits per each resampling. This is still less than $n$, so we save one bit for each resampling. If $N$ is much bigger than the number of variables, we indeed compress the sequence of random bits used for resampling, and this happens with exponentially small probability.
This argument finishes the proof.
%
\end{proof}
\section{Tetris and forbidden factors}
The next example is taken from word combinatorics. Assume that a list of binary strings $F_1,\ldots,F_k$ is given. These $F_i$ are considered as ``forbidden factors'': this means that we want to construct a (long) string $X$ that does not have any of $F_i$ as a factor (i.e., none of $F_i$ is a substring of $X$). This may be possible or not depending on the list. For example, if we consider two strings $0, 11$ as forbidden factors, every string of length $2$ or more has a forbidden factor (we cannot use zeros at all, and two ones are forbidden).
The more forbidden factors we have, the more chances that they block the growth in the sense that every sufficiently long string has a forbidden factor. Of course, not only the number of factors matters: e.g., if we consider $0, 00$ as forbidden factors, then we have long strings of ones without forbidden factors. However, now we are interested in quantitative results of the following type: \emph{if the number of forbidden factors of length $j$ is $a_j$, and the numbers $a_j$ are ``not too big'', then there exists an arbitrarily long string without forbidden factors}.
This question can be analyzed with many different tools, including Lov\'asz local lemma (see~\cite{rumyantsev-ushakov}) and Kolmogorov complexity. Using a complexity argument, Levin proved that if $a_j=2^{\alpha j}$ for some constant $\alpha<1$, then there exists a constant $M$ and an infinite sequence that does not contain forbidden factors of length smaller than $M$. (See~\cite[Section 8.5]{usv} for Levin's argument and other related results.) A nice sufficient condition was suggested by Miller~\cite{miller}: we formulate the statement for the arbitrary alphabet size.
\begin{proposition}
Consider an alphabet with $m$ letters. Assume that for each $j\ge 2$ we have $a_j$ ``forbidden'' strings of length $j$. Assume that there exist some constant $x>0$ such that
$$
\sum_{j\ge 2} a_j x^j < mx -1
$$
Then there exist arbitrarily long strings that do not contain forbidden substrings.
\end{proposition}
\textbf{Remarks}. 1. We do not consider $j=1$, since this means that some letters are deleted from the alphabet.
2. By compactness the statement implies that there exists an infinite sequence with no forbidden factors.
3. The constant $x$ should be at least $1/m$, otherwise the right hand side is negative. This means that $a_j/m^j$ should be small, and this corresponds to our intution ($a_j$ should be significantly less than $m^j$, the total number of strings of length $j$).
The original proof from~\cite{miller} uses some ingenious potential function defined on strings: Miller shows that if its value is less than $1$, then one can add some letter preserving this property. It turned out (rather misteriously) that exactly the same condition can be obtained by a completely different argument (following~\cite{goncalves,ochem}) ---
so probably the inequality is more fundamental than it may seem! This argument is based on compression.
\begin{proof}
Here is the idea. We start with an empty string and add randomly chosen letters to its right end. If some forbidden string appears as a suffix, it is immediately deleted. So forbidden strings may appear only as suffixes, and only for a short time. After this ``backtracking'' we continue adding new letters. (This resembles the famous ``tetris game'' when blocks fall down and then disappear under some conditions.)
We want to show that if this process is unsuccessful in the sense that after many steps we still have a short string, then the sequence of added random letters is compressible, so this cannot happen always, and therefore a long string without forbidden factors exists. Let us consider a ``record'' (log file) for this process that is a sequence of symbols ``$+$'' and ``$+\langle\text{deleted string}\rangle$'' (for each forbidden string we have a symbol, plus one more symbol without a string). If a letter was added and no forbidden string appears, we just add `$+$' to the record. If we have to delete some forbidden string $s$ after a letter was added, we write this string in brackets after the $+$ sign. Note that we do \emph{not} record the added letters, only the deleted substrings. (It may happen that several forbidden suffixes appear; in this case we may choose any of them.)
\begin{lemma}
At every stage of the process the current string and the record uniquely determine the sequence of random letters used.
\end{lemma}
\begin{proof}[Proof of the lemma] Having this information, we can reconstruct the configuration going backwards. This reversed process has steps where a forbidden string is added (and we know which one, since it is written in brackets in the record), and also steps when a letter is deleted (and we know which letter is deleted, i.e., which random letter was added when moving forwards).
\end{proof}
If after many (say, $T$) steps we still have a short current string, then the sequence of random letters can be described by the record (due to the Lemma; we ignore the current string part since it is short). As we will see, the record can be encoded with less bits than it should have been (i.e., less than $T\log m$ bits). Let us describe this encoding and show that it is efficient (assuming the inequality $\sum a_j x^j< mx-1$).
We use arithmetic encoding for the lengths. Arithmetic encoding for $M$ symbols starts by choosing positive reals $q_1,\ldots,q_M$ such that $q_1+\ldots+q_M=1$. Then we split the interval $[0,1]$ into parts of length $q_1,\ldots,q_M$ that correspond to these $M$ symbols. Adding a new symbol corresponds to splitting the current interval in the same proportion and choosing the right subinterval. For example, the sequence $(a,b)$ corresponds to $b$th subinterval of $a$th interval; this interval has length $q_aq_b$. The sequence $(a,b,\ldots, c)$ corresponds to interval of length $q_aq_b\ldots q_c$ and can be reconstructed given any point of this interval (assuming $q_1,\ldots,q_M$ are fixed); to specify some binary fraction in this interval we need at most $-\log (q_aq_b\ldots q_c)+O(1)$ bits, i.e., $-\log q_a-\log q_b-\ldots-\log q_c+O(1)$ bits.
Now let us apply this technique to our situation. For $+$ without brackets we use $\log (1/p_0)$ bits, and for $+\langle s\rangle$ where $s$ is of length $j$, we use $\log (1/p_j)+\log a_j$ bits. Here $p_j$ are some positive reals to be chosen later; we need $p_0+\sum p_j =1$. Indeed, we may split $p_j$ into $a_j$ equal parts (of size $p_j/a_j$) and use these parts as $q_s$ in the description of arithmetical coding above; splitting adds $\log a_j$ to the code length for strings of length $j$.
To bound the total number of bits used for encoding the record, we perform amortised accounting and show that the average number of bits per letter is less than $\log m$. Note that the number of letters is equal to the number of $+$ signs in the record. Each $+$ without brackets increases the length of the string by one letter, and we want to use less that $\log m - c$ bits for its encoding, where $c>0$ is some constant saying how much is saved as a reserve for amortized analysis. And $+\langle s\rangle$ for a string $s$ of length $j$ decreases the length by $j-1$, so we want to use less than $\log m+c(j-1)$ bits (using the reserve).
So we need:
\begin{align*}
\log (1/p_0) &< \log m - c;\\
\log (1/p_j)+\log a_j &< \log m +c(j-1)
\end{align*}
together with
$$
p_0+\sum_{j\ge 2} p_j =1.
$$
Technically is it easier to use non-strict inequalities in the first two cases and a strict one in the last case (and then increase $p_i$ a bit):
$$
\log (1/p_0) \le \log m - c;\ \log (1/p_j)+\log a_j \le \log m +c(j-1);\ p_0+\sum_{j\ge 2} p_j <1.
$$
Then for a given $c$ we take minimal possible $p_i$:
\begin{align*}
p_0 &= \frac{1}{m 2^{-c}}\\
p_j& = \frac{a_j (2^{-c})^j}{m2^{-c}}
\end{align*}
and it remains to show that the sum is less than $1$ for a suitable choice of $c$. Let $x=2^{-c}$, then the inequality can be rewritten as
$$
\frac{1}{mx}+\sum_{j\ge 2} \frac{a_jx^j}{mx}< 1,
$$
or
$$
\sum_{j\ge 2} a_jx^j < mx-1,
$$
and this is our assumption.
Now we see the role of this mystical $x$ in the condition: it is just a parameter that determines the constant used for the amortised analysis.
%
\end{proof}
\textbf{Acknowledgement}.
Author thanks his LIRMM colleagues, in particular Pascal Ochem and Daniel Gon\c calves, as well as the participants of Kolmogorov seminar in Moscow.
|
\section{Introduction}
Interstellar media (ISM) of spiral galaxies is both magnetized and turbulent (see \citealt{Armstrong1995ElectronMedium, Chepurnov2010ExtendingData}) with turbulent magnetic fields playing a critical role for many key processes, including the process of star formation (see \citealt{MO07,MK04}), the propagation and the acceleration of cosmic rays (see \citealt{J66,YL08}) and the regulating heat and the mass transfer between different ISM phases (see \citealt{D09} for the list of the different ISM phases). In addition, galactic magnetic fields are responsible for the polarized radiation that presents a serious obstacle for the studies of the polarization of cosmological origin.
Therefore it is essential to have a reliable way to study the properties of magnetic fields in the ISM.
Magnetic fields make turbulence anisotropic, with turbulent eddies elongated along the magnetic field (see \citealt{2013SSRv..178..163B}, for a review). As a result, observed velocity correlations are expected to be elongated along the underlying magnetic field with this property, which was demonstrated with synthetic observations in previous studies (\citealt{2002ASPC..276..182L,2005ApJ...631..320E}, henceforth EL05). Later the study of anisotropies was performed using the Principal Component Analysis (PCA) and applied to observations \cite{2008ApJ...680..420H}.\footnote{The idea behind the studies, which is employing the PCA is the same and, in fact, our study in \cite{2016ApJ...818..118C} shows that there is no practical advantage of using the PCA compared to the velocity centroids. On the contrary, the anisotropies of centroids, unlike the PCA eigen-images, are analytically related to the properties of the underlying turbulence, i.e. to the properties of the Alfven, slow and fast modes that constitute the MHD cascade (see \citealt{2017MNRAS.464.3617K}). This opens prospects of separating the contribution of the compressible (slow and fast) and the in-compressible (Alfven) modes using velocity centroids.}
There is another way to employ properties of MHD turbulence in order to study magnetic fields. The aforementioned turbulent eddies are aligned with magnetic field which entails that the velocity gradients should have larger values for gradients calculated in the direction perpendicular to the magnetic field. This property of magnetic turbulence was employed in \citeauthor{GL17} (\citeyear{GL17}, henceforth GL17), in which the approach of using velocity centroids gradients (VCGs) to trace magnetic field orientations was proposed. The technique was further extended and elaborated in \citeauthor{YL17} (\citeyear{YL17}, henceforth YL17a) and the new way of magnetic field tracing was successfully compared with observations on polarization from {\it PLANCK}. The subsequent studies in \citeauthor{YL17b} (\citeyear{YL17b}, henceforth YL17b) revealed the synergy of simultaneous use of VCGs and intensity gradients (IGs). \footnote{By itself, the IGs were shown to be inferior to the VCGs in the ability of tracing magnetic field (GL17, YL17a,b), but synergistic in terms of studying shocks and regions dominated by self-gravity.} These papers introduced a new way of studying magnetic fields with spectroscopic data as well as studying other key processes taking place in the ISM.
In a separate development, the same idea of tracing magnetic fields with gradients was employed with synchrotron intensity maps in \citeauthor{LYLC17} (\citeyear{LYLC17}, henceforth LYLC). Using synthetic data which LYLC showed that synchrotron intensity gradients (SIGs) can reliably trace the magnetic field in the ISM, and the study also confirmed the conclusion by comparing {\it PLANCK} synchrotron polarization maps with the SIGs maps. The synergies of SIGs, IGs and VCGs in self-gravitating media were all demonstrated in \cite{YL17b}, showing that the relative rotations between different types of gradient vectors are informative of the stage of collapse of a piece of self-gravitating cloud.
Velocity centroid is a way of representing ISM velocities using observations, and this way has its limitations. For example, centroids reflect the contributions arising from both the velocity and the density fluctuations along the line of sight (see EL05). However, the density fluctuations are not as well aligned with magnetic fields as velocity (see \citealt{CL03,2005ApJ...624L..93B,2007ApJ...658..423K}, or the comparison of VCGs to IGs in YL17b ). At the same time, the analytical study in \cite{2000ApJ...537..720L,2004ApJ...616..943L} revealed that channel maps are sensitive only to velocity fluctuations if the corresponding turbulent density fluctuations are dominated by large scale contributions. In addition, studies of different velocity channels in some cases may allow separated contributions which are coming from spatially different regions along the line of sight, and therefore enabling one to study the 3D structure of magnetic field as well as other effects that the velocity gradients are sensitive to. This suggests that in a number of cases the VChGs may have advantages compared to the VCGs. This motivates our present study to explore the ability of velocity channel gradients (VChGs) in tracing the orientations of the magnetic field.
In some aspects the VChGs have similarities with the technique based on studies of filaments in the velocity HI channel maps in \cite{Clark15}, where these filaments were shown to be correlated with the magnetic field directions as revealed the {\it PLANCK} polarimetry. On the basis of \cite{2000ApJ...537..720L} one can conclude that the filaments observed by \cite{Clark15} in thin channel maps can be identified with caustics caused by velocity crowding. The relationship between the VChGs and the underlying velocity gradients is more straightforward than for the filaments, and therefore, we expect a better correlation between the VChGs and the magnetic field than with the filaments. Not to mention, the gradient technique shows its capability identifying shock and self-gravitating regions (YL17b), while no similar conclusion has ever been reported using filaments. Whether or not there can be any synergistic usage of the simultaneous use of the VChGs and tracing the channel filaments should be answered by further additional studies.
In what follows, we discuss in \S \ref{sec:theory} the theoretical motivation of this work. \S \ref{sec:numerics} discusses the numerical methods on simulations and the ways of doing analysis. We explore the gradients in the channel maps from turbulent velocities in \S \ref{sec:gradients-in-fsa}.
We examine the performance of channel map gradients and correlation anisotropies in \S \ref{sec:anisotropy}. In \S \ref{sec:fluctuations} we compare the performance of density fluctuations and velocities from channel maps. We explore the reduced centroid gradient in \S \ref{sec:reducedcentroid}. We testify our method with observational data in \S \ref{sec:obs}. We discuss our result in \S \ref{sec:discussion}, and make our conclusion in \S \ref{sec:conclusion}.
\section{Theoretical motivation and expectations}
\label{sec:theory}
\subsection{Anisotropy of MHD turbulence: illustration}
MHD turbulence theory is an old subject that has been boosted recently by the ability of performing high resolution 3D numerical simulations. Before that there was no way of testing theoretical constructions and many competing theories describing MHD turbulence were able to co-exist. \footnote{We are talking about realistic 3D MHD turbulence theories. MHD turbulence in 2D is very different from the one in 3D (see \citealt{2011ApJ...743...51E}).} For instance, the original studies of Alfvenic turbulence by \cite{I64} and \cite{K65} were based on a hypothetical model of isotropic MHD turbulence, while the later studies (see \citealt{1981PhFl...24..825M, 1983PhRvL..51.1484M, 1983JPlPh..29..525S,1984ApJ...285..109H}) demonstrated the anisotropic nature of the MHD cascade.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{Fig3_from_nsf.pdf}
\caption{\label{fig:illus0}A 3D snapshot of velocity iso-contours (red structures) from a sub-Alfvenic simulation showing they are all elongated with the magnetic field directions (Blue shows the mean directions, which is along the z-axis). One can imagine if gradients are calculated in this snapshot, they will be perpendicular to the field directions. }
\end{figure}
Figure \ref{fig:illus0} shows the visualization\footnote{Produced using Visit 2.8.1: https://wci.llnl.gov/simulation/computer-codes/visit} of our numerical simulations. In particular, the iso-contours of velocity in for the sub-Alfvenic turbulence are shown. It is obvious that the velocity gradients are directed perpendicular to the magnetic field and therefore by studying the direction of velocity gradients one can study the direction of magnetic field. In what follows, we explain that, in fact, one is expected not only to trace mean magnetic field in the sub-Alfvenic turbulence, but trace the magnetic field in its complexity both in sub-Alfvenic and super-Alfvenic turbulence.
\subsection{Anisotropy of MHD turbulence: local directions of magnetic field}
The modern theory of MHD turbulence originates from the prophetic work by (\citealt{GoldreichP.Sridhar1995GS95IITurbulence}, henceforth GS95). Given originally rather lukewarm acceptance by the MHD turbulence community, this theory nevertheless was supported by further theoretical and numerical studies (\citealt{Lazarian1999ReconnectionField}, henceforth LV99, \citealt{2000ApJ...539..273C,2001ApJ...554.1175M,Lithwick2001CompressiblePlasmas,Cho2001SimulationsMedium,Cho2002CompressiblePlasmasb,CL03,Kowal2010VelocityScalingsb}, see \citealt{2013SSRv..178..163B} for a a review) that extended the theory and provided their rigorous testing. Our present study is based on the modern understanding of the MHD turbulence cascade and the statistical properties of MHD turbulence that are confirmed numerically.
In the sub-Alfvenic regime, i.e. for the injection velocity $V_L$ being less than the Alfven velocity $V_A$, the Alfven modes initially evolve by increasing the perpendicular wavenumber while keeping the parallel wavenumber the same (see LV99, \citealt{Gal2005}). The increase of the perpendicular wave number makes the Alfvenic wave-vectors more and more perpendicular to magnetic field. Therefore the gradients of velocity structures tend to be perpendicular to the magnetic-field direction. Eventually at a scale $l_{trans}\approx L M_A^2$, where $L$ is the turbulence injection scale and $M_A=V_L/V_A$ is the Alfven Mach number (see LV99, \citealt{Lazarian2006}), the parallel scale of eddies starts changing, signifying the start of the GS95 regime. In the GS95 cascade both parallel and perpendicular wave-numbers of Alfvenic perturbations increase but the eddies get more and more elongated. To quantify this, one should adopt the system of reference aligned with the local magnetic field.\footnote{ In the works that followed groundbreaking GS95 paper (namely, LV99, \citealt{2000ApJ...539..273C,2001ApJ...554.1175M,Cho2002CompressiblePlasmasb}), it was shown that to describe MHD turbulence one should use the {\it local} magnetic field rather than the mean magnetic field.} In such a system of reference the fluid motions perpendicular to magnetic fields are not constrained by magnetic tension. This is the consequence of the fast turbulent reconnection that was shown to change the topology of the interacting magnetic flux tubes within the eddy turnover time (LV99). Therefore it is not surprising that the turbulent energy is channeled along this path of the least resistance. It is not surprising either that in the absence of magnetic resistance to the mixing motions, the perpendicular eddies evolve along the Kolmogorov cascade with the Kolmogorov scaling $v_l\sim l^{1/3}$.
However, the motions mixing magnetic field in the perpendicular direction also induce wave propagation along magnetic fields. The period of these waves $l_{\|}/V_A$ is equal of the period of the magnetized plasma mixing in the perpendicular direction $l_{\bot}/v_l$. The equality
of the two time scales is usually referred as the {\it critical balance} and the scales $l_{\|}$ and $l_{\bot}$ are associated with the parallel and perpendicular scales of the eddies. The relation between these scales for sub-Alfvenic turbulence is (LV99):
\begin{equation}
l_{\|}\approx L \left(\frac{l_{\bot}}{L}\right)^{2/3} M_A^{-4/3},
\label{lpar}
\end{equation}
which testifies that, for $l_{bot}$ is much smaller than the injection scale $L$, and therefore the eddies are strongly elongated. Note that these eddies are aligned with the local magnetic field. For trans-Alfvenic turbulence $M_A=1$ and Eq. (\ref{lpar}) provides the original GS95 scaling. Note that the alignment of the motions in respect to the local system of reference is an essential part of MHD turbulence as we understand it now. It is easy to see that due to this peculiar property of MHD turbulence, the velocity gradients are expected to be aligned with the {\it local} magnetic field. It is easy to show that the gradients of the smallest eddies produce the strongest signal. Indeed, the gradient of the eddy aligned mixing up local magnetic field lines is proportional to $v_l/l_{\bot}\sim l_{\bot}^{-2/3}$. The smallest eddies are the eddies at the resolution scale of the telescope. Thus we can claim that the velocity gradients trace magnetic field at the scale of the instrument resolution.
It was shown numerically (see Cho \& Lazarian 2002, 2003) that compressible MHD turbulence can be presented as a composition of three cascades that marginally exchange energy between them\footnote{This is not true for the relativistic MHD turbulence where the coupling between fast and Alfvenic fundamental modes can be significant \citep{2016ApJ...831L..11T}.} . These are cascades of Alfvenic, slow and fast modes. We use the word "modes" rather than "waves" as in strong MHD turbulence the properties of motions may not be wave-like. As we discussed earlier, Alfvenic modes are essentially eddies and they non-linearly decay within one period (see more in a review Brandenburg \& Lazarian 2013). Within the GS95 picture and its generalization to the compressible media (GS95, \citealt{Lithwick2001CompressiblePlasmas, Cho2002CompressiblePlasmasb, CL03, Kowal2010VelocityScalingsb}) the slow modes are slaved by Alfven modes. Indeed, Alfven modes shear and cascade slow mode perturbations, which is the process confirmed by numerical simulations. Thus, in agreement with numerical studies, the anisotropic scaling given by Eq. (\ref{lpar}) is valid for slow mode eddies. As a result, both the gradients of Alfven modes and of slow modes are expected to be {\it perpendicular} to the {\it local} directions of magnetic field. This is the key idea behind tracing magnetic-field directions with velocity gradients.
\subsection{Intensity fluctuations in thin channel maps: effect of velocity fluctuations}
While velocity gradients are directly not available from astrophysical observations of diffuse media, a number of measures can be constructed using observational data. In \cite{GL17} we explored Velocity Centroid Gradients (VCGs) of the first order and considered the higher order VCGs in \cite{YL17}. In this paper we prove another way to probe turbulent velocities, namely, to use velocity channel maps that can be constructed using spectroscopic observations of Doppler shifted lines. The statistics of these maps has been described in \citeauthor{2000ApJ...537..720L} (\citeyear{2000ApJ...537..720L}, henceforth LP00) for the optically thin data and in \cite{2004ApJ...616..943L} for the observations in the presence of absorption. In what follows we concentrate on the optically thin case and only mention some the possible effects of the optically thick data. Note that when we discuss thin and thick velocity slices, we mean not the effects of absorption but the thickness of channel maps. The minimal thickness of the latter is determined by the spectral resolution $\Delta v$ of the instrument, and it can be increased by integrating the spectroscopic data over larger $\Delta v$.
An important prediction in LP00 is that velocity caustics create fluctuations of intensity of the channel maps and the relative importance of the velocity and density fluctuations changes with the thickness of the channel maps. In particular, LP00 identified a regime of "thin velocity slices" and found that in this regime the intensity fluctuations in the slice are dominated by the velocity fluctuations, provided that the three dimensional density spectrum is steep, i.e. most of the energy is concentrated at the large scales. In what follows we use the terms "velocity slices" and "channel maps" interchangeably.
The aforementioned statement about the steep density spectrum can be expressed in terms of the 3D power spectrum $P(k)$. In terms of $P(k)$ the Kolmogorov cascade corresponds to $k^{-11/3}$ and it is steep. The borderline spectrum is $k^{-3}$, with turbulence having spectrum $k^{-\alpha}$, $\alpha<3$ containing more energy at the small scale and therefore being shallow. For subsonic flows the density spectra in MHD turbulence are steep (see \citealt{2007ApJ...658..423K}).\footnote{The traditionally used definition of the Kolmogorov spectrum is obtained via spatial integration of $P(k)$ in the k-space, which is equivalent to multiplying the spectrum by $k^2$. Thus the usually referred value of the Kolmogorov spectrum is $k^{-5/3}$ and the border line spectrum between the shallow and the steep is $k^{-1}$.} Thus for such flows the intensity fluctuations in thin velocity slices of Position-Position-Velocity (PPV) data are influenced only by the turbulent velocity statistics. Naturally, the thermal broadening interferes with the minimal slice of the PPV for which the fluctuations can be studied. This means that for studying thin slices of subsonic flows one should use heavier species that those of the main hydrogen astrophysical flow. On the other hand, if the density spectrum is shallow, i.e. $\alpha<3$, the contribution of density and velocities to the statistics of intensity fluctuations of the velocity slices was evaluated in LP00. For thin slices, both velocity and density are important, while the contribution of velocity decreases as the slice thickness increases. For instance, when the integration is performed over the entire line, only density fluctuations determine the fluctuations of the resulting intensity distribution. In fact, no matter how steep the spectrum is, velocity contribution can be enhanced by reducing the channel width, which significantly simplifies the interpretation of the VChGs' directions in terms of magnetic field tracing.
In view of the above, it is important to perform the study of the gradients of intensities within PPV slices. In terms of separating compressible and incompressible components in KLP16, such a study provides an additional test that the anisotropies are actually caused by turbulence. By varying the thickness of the slice one can study the variation of gradients as the relative contribution of density and velocity changes. This may be important as the density and velocity, in general, have different statistics and the fluctuations of these fields are aligned with magnetic field to a different degree. For instance, the strongly supersonic flows demonstrate an isotropic spectra of density. In addition, for studies of galactic flows the regular galactic shear opens a way to study different turbulent regions separately and therefore different channel maps can be associated with different locations within the galactic volume. In addition, velocity and density gradients can behave differently, e.g in self-gravitating regions and shocks (See YL17b). In such cases comparing alignments of VChGs from different channel widths opens a new way on locating self-gravitating media, regardless of the steepness of density spectrum.
In what follows the advancements of the understanding of the theory of PPV anisotropies in KLP16 guide us in studying gradients in the PPV velocity slices. Note that the criterion for the velocity slice of being thin or thick as it given in LP00 is as follows: the velocity slice is thin if the square root of turbulent velocity dispersion on the scales that the slice is being studied $\sqrt{\langle \delta v_R^2\rangle}$ is {\it greater} than the thickness of the slice $\Delta v$, i.e.
\begin{equation}
\sqrt{\langle \delta v_R^2\rangle}>\Delta v,
\label{criterion}
\end{equation}
where $R$ is the separation of the correlating points over plane of the sky, i.e. the PP separation. In the following, we distinguish thin and thick slices depending on whether the ratio $\Delta v/\sqrt{\langle \delta v_R^2\rangle}$ smaller or larger than unity. In observations the thickness $\Delta_v$ can be constrained either by the velocity resolution of the instrument or by the thermal line width $\sigma_{th}$that according to LP00 acts similarly. Therefore to make sure that the gradients are measured in the thin slice regime one should make sure that the scale over which the velocity gradients are calculated are sufficiently large.\footnote{LP00 study shows that velocity fluctuations are still important for the slice thicknesses larger than that given by the Eq. (\ref{criterion}). However, their relative contribution compared to density is gradually decreasing. Incidentally, this regime was termed in LP00 "thick slice regime". In the present paper we do not use this terminology and refer to "thick slice" only to the situations where the velocity information is integrated out.} For an individual turbulent volume $\delta v_R$ can be measured using the dispersion of the structure functions.
We note that if $\sqrt{\langle\delta v_R^2\rangle}<\sigma_{th}$ then it is advantageous to use velocity centroids that can represent velocity statistics for subsonic turbulence (see Esquivel \& Lazarian 2005,
Kandel et al. 2017b). The traditional centroids, however, have disadvantages to channels, e.g. one cannot isolate a particular region of the galaxy according to the rotation curve. To problem in the paper we discuss the use of a new construction that we term "reduced velocity centroids" that make use of the part of the line only. Gradients of the reduced velocity centroids and the VChGs can be used together, with the VChGs applied to scales $R$ for which the criterion given by Eq. (\ref{criterion}) is valid and with the reduced velocity centroid gradients for smaller scales.
This all suggestive that the scale of the data blocks for calculating the gradients is another important parameter for the calculating of gradients. In YL17a we dealt with the velocity centroids we showed that the uncertainty of the calculation of velocity centroid gradients (VCGs) decreases with the increase of the data block size. This was augmented in YL17b by a finding that there was an optimal size for the VCGs which provided the maximal resolution without compromising the accuracy. Below we prove that this property is also a part of the VChGs technique.
\section{Numerical Simulations}
\label{sec:numerics}
\subsection{MHD turbulence simulations and mode decomposition}
The numerical data is obtained by 3D MHD simulations using a single fluid, operator-split, staggered grid MHD Eulerian code ZEUS-MP/3D \citep{2006ApJS..165..188H} to set up a three-dimensional, uniform, isothermal turbulent medium. Periodic boundary conditions are applied to emulate a part of interstellar cloud. Solenoidal turbulence injections are employed. Our simulations employ various Alfvenic Mach numbers $M_A=V_L/V_A$ and sonic Mach numbers $M_s=V_L/V_s$, where $V_L$ is the injection velocity, while $V_A$ and $V_s$ are the Alfven and sonic velocities respectively, which they are listed in Table \ref{tab:simulationparameters}. The domain $M_A<M_s$ corresponds to the simulations of plasma with magnetic pressure larger than the thermal pressure, i.e. plasma with low $\beta/2=V_s^2/V_A^2<1$, while the domain $M_A>M_s$ corresponds to the pressure dominated plasma with $\beta/2>1$. Further we refer to the simulations in the table by their model name. For instance, our figures will have the model name indicating which data cube was used to plot the figure. The simulations are named with respect to a variation of $M_s$ \& $M_A$ in ascending values of $\beta$. The ranges of $M_s, M_A, \beta$ are selected so that they cover different possible scenarios of astrophysical turbulence from very subsonic to supersonic cases. In this study we devoted much of our analysis to sub- and trans- Alfvenic cases only, and postpone the discussion on super-Alfvenic simulations to our next paper (Yuen \& Lazarian). We expect the velocity gradients to successfully trace magnetic field in superAlfvenic turbulence after the appropriate filtering of low frequency spatial modes. The practical difficulty of such a study is that it is that the inertial range of the MHD turbulence is rapidly shrinking with the increase of $M_A$. If for $M_A>1$ the injection scale is $L$ the transition to the MHD regime happens at the scale $l_{A}\approx LM_A^{-3}$ (see Lazarian 2006), which requires very large cubes to study velocity gradients arising from MHD turbulence.
\begin{table}[h]
\centering
\label{tab:simulationparameters}
\caption {Simulations used in our current work. The magnetic criticality $\Phi = 2 \pi G^{1/2} \rho L/B$ is set to be 2 for all simulation data. Resolution of them are all $480^3$}
\begin{tabular}{c c c c}
Model & $M_s$ & $M_A$ & $\beta=2(\frac{M_A}{M_s})^2$\\ \hline \hline
Ms0.2Ma0.02 & 0.2 & 0.02 & 0.02 \\
Ms0.4Ma0.04 & 0.4 & 0.04 & 0.02 \\
Ms0.8Ma0.08 & 0.8 & 0.08 & 0.02 \\
Ms1.6Ma0.16 & 1.6 & 0.16 & 0.02 \\
Ms3.2Ma0.32 & 3.2 & 0.32 & 0.02 \\
Ms6.4Ma0.64 & 6.4 & 0.64 & 0.02 \\ \hline
Ms0.2Ma0.07 & 0.2 & 0.07 & 0.22 \\
Ms0.4Ma0.13 & 0.4 & 0.13 & 0.22\\
Ms0.8Ma0.26 & 0.8 & 0.26 & 0.22\\
Ms1.6Ma0.53 & 1.6 & 0.53 & 0.22\\\hline
Ms0.2Ma0.2 & 0.2 & 0.2 & 2 \\
Ms0.4Ma0.4 & 0.4 & 0.4 & 2 \\
Ms0.8Ma0.8 & 0.8 & 0.8 & 2 \\\hline
Ms0.13Ma0.4 & 0.13 & 0.4 & 18 \\
Ms0.20Ma0.66 & 0.20 & 0.66 & 18 \\
Ms0.26Ma0.8 & 0.26 & 0.8 & 18 \\\hline
Ms0.04Ma0.4 & 0.04 & 0.4 & 200 \\
Ms0.08Ma0.8 & 0.08 & 0.8 & 200 \\
Ms0.2Ma2.0 & 0.2 & 2.0 & 200\\\hline \hline
\label{tt1}
\end{tabular}
\end{table}
The numerical simulations that we employ for the study are listed in Table \ref{tt1}. The names of the simulations reflect both the sonic and Alfven Mach numbers. For instance, $Ms0.4Ma0.04$ corresponds to $M_s=0.4$
and $M_A=0.04$.
To investigate the detail structure of gradients from different wavemodes, we employ the wave mode decomposition method in \cite{Cho2002CompressiblePlasmasb,CL03} to extract Alfven, slow and fast modes from velocity data. The corresponding equations determining the basis for the decomposition into modes are:
\begin{subequations} \label{eq:fsa-decompositions}
\begin{align}
\hat{\zeta}_f &\propto (1+\frac{\beta}{2}+\sqrt{D}) k_\perp \hat{\bf{k}}_\perp +(-1+\frac{\beta}{2}+\sqrt{D}) k_{||}\hat{\bf{k}}_{||} \\
\hat{\zeta}_s &\propto (1+\frac{\beta}{2}-\sqrt{D}) k_\perp \hat{\bf{k}}_\perp +
(-1+\frac{\beta}{2}-\sqrt{D}) k_{||}\hat{\bf{k}}_{||}\\
\hat{\zeta}_a &\propto -\hat{\bf{k}}_\perp\times \hat{\bf{k}}_{||}
\end{align}\end{subequations}
where $D=(1+\beta/2)^2-2\beta \cos^2\theta$\footnote{The proportional constant for $\cos\theta$ is correct only in CL02, i.e, $\cos^2\theta$.} , $\beta=\frac{\bar{P}_g}{\bar{P}_B}=\frac{2M_A^2}{M_s^2}$, and $\cos\theta= \hat{k}_{||} \cdot \hat{B}$.
We would only use the LOS component of the decomposed velocities for velocity channels calculations.That is to say, the three velocity modes can then be acquired by
\begin{equation}
v_{(f,s,a),z}= [\mathscr{F}^{-1}(\mathscr{F}(\bf{v})\cdot\hat{\zeta}_{f,s,a})](\hat{\zeta}_{f,s,a} \cdot \hat{\zeta}_{LOS})
\end{equation}
where $\mathscr{F}$ is the Fourier transform operator.
\begin{figure*}[t]
\centering
\includegraphics[width=0.96\textwidth]{CL03-decompo.pdf}
\includegraphics[width=0.96\textwidth]{rhoconstant-Vf-Va-Vs-annotated.pdf}
\includegraphics[width=0.96\textwidth]{rhoconstant-Vf-Va-Vs-b42x.pdf}
\caption{\label{fig:fsa} (Top) Illustration of the decomposition method. From \cite{CL03}. (Middle \& Bottom) The resultant projected maps from CL03-decomposition from two sample cubes with low (middle panel) and high (low panel) $\beta$ }
\end{figure*}
The upper panel of Figure \ref{fig:fsa} illustrates the decomposition procedure that takes place in the Fourier space. The resulting 3 data cubes are dominated by the Alfven, slow and fast modes, respectively. In the middle and lower panels of Figure \ref{fig:fsa} we show the results of decomposed velocity cube which being projected along the x-axis. We illustrate the decomposition method using two cubes with low and high $\beta$. The properties of the fast and slow modes differ in low and high $\beta$ plasmas, therefore we study these two cases separately in the following sections.
\subsection{Calculations of gradients in channel maps}
Gradients were calculated following the procedures described in YL17a, including the gradient calculation method and the {\it sub-block averaging}.\footnote{This procedure is very different from employed for gradient calculation by other authors, e.g. \cite{Soler2013}). It is very advantageous to apply this procedure for calculating other types of gradients.} In short, the gist of the {\it sub-block averaging} method is to define the most probable direction within a block by finding a local Gaussian-fitting peak for the distribution of gradients. The fitting error provides a quantitative estimate whether the block size is large enough for reliably determining the direction.
The concept of thin channel maps was introduced in \cite{LP00}, where it was shown that the fluctuations within thin channel maps are mostly determined by velocity fluctuations. For determining whether the gradients are probing thin or thick channel map we use the following criterion: the channel is thin if the gradient is calculated over the scale of size $R$ for which the criterion given by Eq. (\ref{criterion}) is satisfied; otherwise, the channel is thick. To study velocity gradients over the scale larger than $R$ we construct the {\it velocity channel map}:
\begin{equation}
C(x,y) =\int dv \rho_{ppv} (x,y,v) e^{-\frac{|v-v_{0}|^2}{\delta v_R^2}}\\.
\end{equation}
As the thickness $v_R$ should not be less than the maximum of the thermal linewidth and spectrometer resolution, it means for images with high resolution the gradients at the smallest scales should be calculated using velocity centroids.
Readers should be reminded that channel map carries {\it intensity} information within a velocity channel. The reference velocity slice is convenient to define at the center of the spectral line. However, if the spectral line is broadened by the regular shear, as this is the case of Galactic atomic hydrogen, different velocity channels carry information about the 3D distribution of turbulence and therefore it is advantageous to study gradients in different bunches of channels in order
magnetic fields and the interstellar medium (ISM) physical processes.
Similarly, in the presence of absorption, the wings of the lines are advantageous to used rather than the entire absorption lines. The latter are saturated and thus not informative at center of the line.
\begin{comment}
{\it number density channel map} about some velocity slice $v_0$ by:
\begin{equation}
I(x,y) = \int dv n_{ppv} (x,y,v) e^{-\frac{|v-v_{0}|^2}{R^2}}\\
\end{equation}
The respective
\end{comment}
To produce channel maps that contain only spatial frequencies for which the slice is thin we provide filtering of the high spatial frequencies for which the criterion given by Eq. (\ref{criterion}) is not satisfied.
The filtered spacial frequencies can still be studied using the VCGs.
\subsection{Uncertainties of magnetic field tracing}
Unlike the traditional technique of calculating gradients (see \cite{Soler2013}) our approach allows us provides us with the uncertainty of the determination of the magnetic field direction. To test how well this works in
Figure \ref{fig:vchgerr} we show the error estimate for the VChG in the cube $M_s1.6M_A0.53$ that we calculate within our technique. As we fit the Gaussians into the distribution of the gradient directions we get an estimate of the fitting error, which is the lowest curve in Figure \ref{fig:vchgerr}. The half width of the Gaussians provides a significantly larger uncertainty shown by the black curve in Figure \ref{fig:vchgerr}. In numerical simulations we also can calculate the difference between the projected magnetic field and the measured gradient directions. This measurement is given by black curve in Figure \ref{fig:vchgerr}. These measures are changing are the function of the block size and our can observe our procedure of evaluating our error for magnetic field tracing is in reasonable correspondence with the actual measurements of the differences of the projected magnetic field direction and the direction given by the velocity gradients.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{stderror.pdf}
\caption{\label{fig:vchgerr} The average error estimate from the Gaussian fitting function (Stderr), the angular differences between magnetic field directions from polarization and VChG orientation ($\phi_{pol}-\phi_{VG}$) and the dispersion of VChGs ($\delta \phi_{VChG}$) for the velocity centroid map of cube $M_s1.6M_A0.53$ plotted against the change of block size. }
\end{figure}
\section{Alignment Measure and Density Effects for the VChGs}
\label{sec:gradients-in-fsa}
Before showing the gradient maps, we would like to illustrate the difference of supersonic and subsonic simulation in terms of density spectra. Figure \ref{fig:illus-0} shows the normalized density spectra from two simulation, one of it is supersonic and the other one is subsonic. The normalization is done by making the amplitudes of the spectra the same at the injection and it helps the reader to have easy comparison between the two spectra. The subsonic spectrum of density in strongly magnetized media scales as pressure and therefore follows not the Kolmogorov $k^{-5/3}$ law, but a steeper $k^{-7/3}$ law (see \cite{2007ApJ...658..423K}). For this definitions of spectra, the boderline between the shallow and the steep spectra corresponds to $k^{-1}$. It is very clear that the high $M_s$ spectra is shallow compared to the steep density spectra in subsonic systems. Our results on density studies agree well with those in \cite{2005ApJ...624L..93B} and \cite{2007ApJ...658..423K}.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{spectrum-dppv2.pdf}
\caption{\label{fig:illus-0} The 3D density spectrum from supersonic (blue,b15) and subsonic (red,b11) runs. While the subsonic spectrum has a steep slope, the supersonic one has a shallow one. The straight line shows the theoretically expected $-7/3$ density slope in subsonic strongly magnetized turbulence (see Kowal \& Lazarian 2010).}
\end{figure}
As the velocities, unlike densities, are directly related to MHD turbulence, we first consider only intensity fluctuations that arise only from turbulent velocity. For this purpose we create data cubes using the velocity field obtained from our 3D numerical simulations but substitute the actual densities by a constant density, which is just the {\it number density channel map} constructed above. Figure \ref{fig:illus-1} illustrates the relative orientation velocity channel gradients (VChGs) and the projected magnetic field on a {\it thin} slice setting. The gradients are all rotated for 90 degrees, which will be annotated as {\it rotated} gradients. The rotated gradients according to the theoretical considerations above correspond to the magnetic field directions.In the following sections, gradients are assumed to be rotated, unless emphasized specifically. One can see from Figure \ref{fig:illus-1} that both number density and velocity channel map behave very good in terms of gradient alignment. That gives us confidence on using the velocity channel maps in our later analysis.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{b23-hdfaa-008x}
\caption{\label{fig:illus-1} Intensity gradients (left panel, green color) and VChGs (right,red) from a thin slice map with respect to the projected magnetic field (blue color). The maps prior to the CL03 decomposition into modes are used.}
\end{figure}
The fact that the VChGs for thin maps are strongly influenced by the velocity field is suggesting that the tracing of magnetic fields using thin channel maps may be more accurate due to velocity being a better tracer of MHD turbulence. Below we consider the gradients arising from Alfven, slow and fast modes separately. The theoretical considerations giving the guidance for this study are provided in KLP16, where the anisotropies of the correlation of the intensities arising from velocities were studied.
The {\bf Alignment measure (AM)} is used by us to express quantitatively how well a certain type of gradient vectors trace magnetic field. AM is defined by the expression\footnote{This 2D measure is analogous to the 3D measure that is used in the theory of grain alignment (see Lazarian 2007 for a review).}:
\begin{equation}
AM=\langle2 cos^2\theta-1\rangle,
\end{equation}
and it ranges from -1 to 1. $\theta$ in the equation stands for the angle between the VChG and the projected magnetic field. The physical meaning of AM is as the following: if $AM \sim 1$, then the average angular difference in the directions of two vectors across the map is negligibly small; if $AM \sim 0$, then there is no alignment between two vectors. If the two directions tend to be perpendicular to each other, the $AM \rightarrow -1$.
\section{Applying VChGs to Basic MHD modes}
It is important to understand how what is the effect of different modes on the VChGs. The composition of MHD turbulence in terms of the basic modes is changing with $M_A$ and plasma $\beta$ (see Cho \& Lazarian 2003).
\subsection{Alfven modes}
Most important for the MHD turbulence are Alfven modes (see CL03). We will first demonstrate some properties of VChGs with these modes.
The dynamics of Alfvenic modes in the strong MHD turbulence is very different from the dynamics of freely propagating Alfven waves. The modes cascade on the scale of the order of one period with the wavevector of the Alfvenic perturbations in strong turbulence being nearly perpendicular to the local direction of the magnetic field. As a result, the anisotropy and the iso-contours of intensity correlation are both elongated {\it parallel} to magnetic field. The latter is essential for tracing of magnetic fields, as the gradients that these modes are inducing are also perpendicular to the local direction of magnetic field.
For Alfven modes the theoretical expectation is very natural: the anisotropy of the intensity correlations in a thin slice increases with the decrease of the Alfven Mach number, i.e. the weaker the velocity perturbations, the more turbulent motions are dominated by magnetic field which causes the eddies to be more anisotropic. Figure \ref{fig:alf-1} illustrates the channel maps for low and high $\beta$ and their gradients within a thin slice that are induced by Alfvenic modes. According to \cite{CL03} the change of Alfven modes with $M_s$ is marginal. The corresponding change in terms gradients is shown in the left panel of Figure \ref{fig:alf-2}.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{hdfaa-Vta-VG.pdf}
\caption{\label{fig:alf-1} The velocity channel map for low (left) and high (right) $\beta$ from Alfven mode in a thin slice map. The rotated 90 degrees gradients (red for VChGs, green for IGs) are over-plotted over the projected magnetic field (blue).}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{b3-51-Ms-AM.pdf}
\caption{\label{fig:alf-2} The change of $AM$ of Alfven mode VChGs with respect to $M_s$ }
\end{figure}
As we increase the thickness of channel maps, the $AM$ of Alfven modes is decreasing, which is shown in the blue curves in Figure \ref{fig:alf-3} for both low and high $\beta$ cases. This is because the contribution from velocities decreases as the channel the channel width increases. However, we observe that the change is gradual (see also theory in LP00) rather than a sharp jump, which indicates even thin slices satisfying Eq. (\ref{criterion}) may not be available for some data, selection of slices as thin as possible can already help enhancing the contribution from velocity information in observation data.
We can double check the above statement by testing the gradient variation dependence on the channel width. In this experiment, we use the density-weighted PPV cubes to illustrate the relative dominance of velocity over density fluctuations when one changes the channel width, as illustrated in Figure \ref{fig:alf-4}. When we increase the channel width, the VChGs variations, given by $\sqrt{\langle \delta I_{ch}^2\rangle}/ \langle I_{ch}\rangle$, are decreasing. Due to this decrease of gradient deviations, the role of velocity fluctuations in the channel map fluctuations is decreasing and the thick channels get dominated by density fluctuations. A more detailed discussion on the contribution of density effect will be in \S \ref{sec:fluctuations}.
One should remember, however, that whether the channels are thin or thick depends on the size of the eddies that are being studied (LP00). In our case, the block size is an additional factor limiting the size of the eddies that we can study, which usually have a large velocity dispersion than a small-block sampled channel maps. The verification of criterion of the thin slice (see Eq \ref{criterion}) has to be done every time in the case of small blocks.
\begin{figure*}
\centering
\includegraphics[width=0.48\textwidth]{AM-width-alfven-slow.pdf}
\includegraphics[width=0.48\textwidth]{AM-width-alfven-slow-b42.pdf}
\caption{\label{fig:alf-3} The change of $AM$ of slow (green) and Alfven (blue) modes for the VChGs as the channel width increases. (Left panel): The results for low-$\beta$ (left) simulations. (Right panel): The results for high-$\beta$ (right) systems simulations.}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{I-ch2.pdf}
\caption{\label{fig:alf-4} Variation of $\sqrt{\langle \delta I_{ch}^2\rangle}/\langle I_{ch}\rangle$ with respect to channel map thickness, for both density weighted (red) and constant density (blue) cases. The random walk fits are added accordingly. }
\end{figure}
According to LP00 the thermal line width acts similarly to the channel thickness. Here we used the data of the cube Ms1.6Ma0.528 by adding additional thermal dispersion of the velocities before constructing of the PPV cube. Figure ~\ref{fig:thermal_channel} shows the relationship between the AM versus the ratio thermal width over a channel width of $\delta v=0.2 \sqrt{\delta v_R^2}$ and the alignment measure. The decreasing trend resulted by an increase of thermal width is similar to the effect of increasing channel width illustrated in Figure \ref{fig:alf-3}. Moreover, when the ratio between thermal width to channel width is larger than 1, the thermal velocities dominates over the contribution from turbulent motions. Different from the effect from increasing channel width, the increase of thermal line width retain the velocity information yet washes away the fluctuations. This explains why the AM is oscillating around zero once the noise width is larger than the channel width. In spite of this, such an effect would not prevent the application of our technique in most of the region of the sky, which is mostly supersonic. This effect also sets a lower bound for observers to select the channel width when applying the method.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{Thermal-Channel.pdf}
\caption{\label{fig:thermal_channel} The thermal width relative to channel thickness versus the alignment measure value analyzed from data obtained in the cube Ms1.6Ma0.528. }
\end{figure}
\subsection{Slow modes}
Slow waves present perturbations that propagate along magnetic field lines. In the limit of incompressible media, slow waves are pure magnetic compressions that propagate along magnetic field lines. Formally the incompressible case corresponds to $\beta=\infty$ and in this limit the slow modes are frequently called preudo-Alfven modes. On the contrary, for $\beta\ll 1$ the slow waves are density perturbations propagating along magnetic field lines.
In the presence of Alfvenic turbulence, slow modes do not evolve on their own, but are sheared by Alfven modes. As a result, the features of Alfvenic turbulence, e.g. spectrum and anisotropies, are imprinted on the slow modes (see GS95, \citealt{Lithwick2001CompressiblePlasmas,Cho2002CompressiblePlasmasb,CL03}). This also means the perpendicular velocity gradients characteristic of Alfven modes are also inherited by slow modes.
Figure \ref{fig:slow-1} illustrates the channel maps for a thin slice for the case of low $\beta$ (left panel) and high $\beta$ (right panel). One can see that low $\beta$ VChGs performs better, which is also shown in Figure \ref{fig:alf-3}. In fact, when $\beta$ is small, the gradients of the corresponding channel maps are similar those of the Alfven mode channel maps. On the other hand, the slow mode channel maps from high $\beta$ systems are not as highly structured as the maps with low $\beta$. Moreover, the AM of slow modes from both cases decrease faster than that of Alfven modes as channel width increases. In reality both modes contribute to gradients of the channel maps. The result of Figure \ref{fig:alf-3} suggests that as the velocity channel thickness increases the AM decreases. The infinite channel thickness corresponds to using total intensities. It is obvious, that in terms of magnetic field tracing the intensities (thick channels) are inferior to the thin channels.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{hdfaa-008-Vts-VG.pdf}
\caption{\label{fig:slow-1} Slow modes:The {\it thin} slice velocity channel maps for low $\beta$ (left) and high $\beta$ (right) simulations.}
\end{figure}
\subsection{Fast modes}
\label{subsec:fastmodes}
Similar to the properties of the slow modes, the properties of fast modes are different in low and high $\beta$ plasmas. For high $\beta$ plasmas the fast wave are similar to sound waves that propagate with sound speed irrespectively of the magnetic field direction. Similar to acoustic turbulence, the corresponding turbulence is expected to be isotropic. In low $\beta$ the fast modes correspond to magnetic field compressions that propagate with the Alfven velocity. In terms of correlation function anisotropy of fast modes, the elongation of the iso-contours of correlation is {\it perpendicular} to magnetic field. Therefore for fast mode dominated environments, one should expect the gradients to be parallel instead of perpendicular to local magnetic field.We note that the alignment of velocity gradients from the actual MHD turbulence contain contributions from the three modes with fast modes decreasing the alignment induced by the Alfven and slow modes.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{hdfaa-008-Vtf-VG-unrotated.pdf}
\caption{\label{fig:fast-1}The {\it thin} slice velocity channel maps with {\it unrotated} gradients for low $\beta$ (left) and right $\beta$ (right) for fast modes. The red vectors are the {\it unrotated} VChGs, and the blue vectors are projected magnetic field direction. }
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.48\textwidth]{AM-width-fast-rotated.pdf}
\includegraphics[width=0.48\textwidth]{AM-width-fast-b42-rotated.pdf}
\caption{\label{fig:fast-2} The change of $AM$ of fast modes {\it rotated} VChGs with respect to a change of channel width for low (left) and high $\beta$ (right), respectively. }
\end{figure*}
Figure \ref{fig:fast-1} shows, respectively, the channel maps with low and high $\beta$ plasmas. The low $\beta$ map carries more structures than that of high $\beta$ map, but they are obviously carrying different anisotropy direction when examining the contours in the maps. In the language of fast mode gradients, the maximal gradient is {\it parallel} to the magnetic field direction. Figure \ref{fig:fast-2} illustrates the $AM$ of fast modes VChGs.
In contrast to the slow decreasing $AM$ for rotated Alfven and slow mode VChGs, the $AM$ of rotated fast mode VChGs decreases more significantly when the channel width passes through unity for both low and high $\beta$ case. This suggests a decrease of channel width suppresses the unwanted contribution of fast modes in velocity gradient calculations which makes thin channels preferable for tracing magnetic fields. The difference of the behavior of gradients from fast mode on one side and the Alfven and slow modes on the other side, makes it possible to separate the contributions from these modes. This issue will be discussed elsewhere (see also \S \ref{subsec:modesep})
Numerical simulations (e.g. \citealt{Cho2002CompressiblePlasmasb,CL03}) indicate that fast modes are subdominant at least for the cases of incompressible driving of turbulence. Our study indicates that in terms of VChGs one can expect a further suppression of the fast mode contribution in thin channels.
\section{Comparison with the correlation anisotropies within channel maps}
\label{sec:anisotropy}
It was demonstrated in \cite{2002ASPC..276..182L} that the correlation anisotropies in channel maps can trace magnetic fields. Further studies of this effect are provided in (Esquivel \& Lazarian 2005, Esquivel et al. 2015). We believe that correlation anisotropies can be very informative in terms of determining the relative contribution of compressible versus incompressible modes (see \citealt{2016MNRAS.461.1227K,2017MNRAS.464.3617K}). However, in terms of tracing of the detailed magnetic field structure they are inferior to the velocity gradients. Indeed, for correlation function to be informative, one requires a large number of sampling points to acquire enough statistics. As a result, one can expect that the correlation anisotropy requires a lot more data points for tracing magnetic field vector than the VChGs. Incidentally, a similar effect was demonstrated in LYLC for a test using the synchrotron intensity gradients and YL17b for a test using the VCGs. Here we test this statement for the VChGs.
In Figure \ref{fig:cfa-1} we show the VChGs alignment measure $AM$ versus the alignment measure of the directions of the maximal anisotropy (elongation) of the correlation functions of the channel map intensities. We clearly see that the channel map intensities are not good for detailed magnetic field tracing, which is consistent with our previous studies (see YL17, LYLC, YL17b).
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{VChG-anisotropy.pdf}
\caption{\label{fig:cfa-1} A comparison of $AM$ obtained with the VChGs and the correlation function anisotropy for the case of Alfven modes. }
\end{figure}
The relative efficiency of velocity gradients is easy to understand. The velocity gradients correspond to the angular velocities of turbulent eddies. These individual eddies are aligned with the local magnetic field. On the contrary, the correlation functions are the statistical measures that are well defined only after the averaging. Thus the averaging over small patches does not produce good statistics necessary for the correlation functions to be well defined. The shortcomings of the technique of magnetic field tracing based on the correlation function anisotropies is also seen when the corresponding modifications of technique are compared with the VCGs and SIGs (see LYLC, YL17b).
\section{Relative importance of velocity versus density fluctuations, and scale dependent gradients}
\label{sec:fluctuations}
Both velocities and densities contribute to the intensity fluctuations within channel maps, but their relative contribution is changing with the change of the channel map thickness (LP00).
To study the relative importance of density and velocity fluctuations, we use velocity and density data cubes obtained through our 3D MHD compressible simulations. The structures in these cubes are elongated in the same direction which we take to be x-direction. For our study we create the Position-Position-Velocity (PPV) cubes of synthetic spectral line by turning 90 decrees the density cube in respect to the velocity data cube, so the elongated structures in the density get perpendicular to the magnetic field. Mathematically,
\begin{equation}
\label{eq:tc}
C_t(y,z) = \int_{|v_x|<\Delta v} dx \rho(x,y,z) v_x(x,z,y)
\end{equation}
In other words, we are creating synthetic maps by transposing the velocity and density data by 90 degrees within the plane-of-sky plane, i.e. through the rotation along z-direction.
In the PPV cube created this way the directions of anisotropy for the velocity and density are orthogonal and the resulting anisotropy determines which contribution dominates. When examining the anisotropy using the correlation function $R_2({\bf r}) = \langle C_t({\bf x}) C_t({\bf x}+{\bf r}) \rangle_{x}$ in the centroid map, the ratio between the axis perpendicular (the elongation direction of $C_t$) to magnetic field direction over that of parallel one (the elongation direction of $I$) will reflect the relative importance of velocity to density fluctuations to the centroids measured this way.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{aniso.pdf}
\caption{\label{fig:fluc-1} The change of anisotropy (the axis ratio) as the velocity slice thickness changes for the case of centroid $C_t$ (see Eq. (\ref{eq:tc}) for the 90 degree rotated velocity and density data cubes. }
\end{figure}
Figure \ref{fig:fluc-1} shows the decreasing trend of the aforementioned axis ratio as velocity slice thickness increases. It is evident that even though the density fluctuations can contaminate the contribution of the velocity fluctuations within the channel maps, the overall anisotropy comes from the velocities provided that the channel map is sufficiently thin. Therefore using a thin slice one can reduce the contribution from density data (see Figure \ref{fig:fluc-1}.)
We use the same set up employing density and velocity cubes turned 90 degrees to each other to explore the effects of channel thickness for the VChGs. We study {\it transposed channel map} by picking the slices satisfying Eq. \ref{criterion} in a PPV cube constructed with transposed density. Figure \ref{fig:transC} illustrates a thick and a thin slice result from such construction. The thin slice use contribution of PPV cubes with channel width $\sim$ 0.8 of $\delta v$, while the thick slice is just the integration of the transposed density on the full spectral line. It is very obvious from visual inspection that, thin slice structures are parallel to magnetic field in the velocity cube, while that of thick slice reflects the direction of magnetic field in the density cube. The latter is reflected in the AM being negative. In agreement with Figure \ref{fig:fluc-1} we see that the velocity information dominates over density one in when the velocity channel is thin.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Transposed_slice_2.pdf}
\caption{\label{fig:transC} Illustration of gradients calculated for the thick (left panel) and thin (right panel) slices for the case when the density and velocity are turned 90 degrees in respect to each other. The rotated gradients are shown in green for the left panel and in red for the right panel. The magnetic field directions in both panels are shown in blue. }
\end{figure}
A more quantitative result can be seen in Figure \ref{fig:transCAM}, which is a plot of the alignment measure. When the channel width $\Delta v/\sqrt{\delta v_R^2}$ is smaller than unity, velocity contribution is dominant and therefore the AM is positive. On the other hand, the AM for thick channel width maps are negative, indicating contribution from density is dominant. The point where $AM=0$ correspond to the channel width being unity.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{transposed-centroid-width-to-AM-2.pdf}
\caption{\label{fig:transCAM} The alignment measure of the for VChGs as the channel map width increases. }
\end{figure}
\section{Reduced velocity centroids obtained from the data}
\label{sec:reducedcentroid}
Our study above suggests that the decrease of channel width improves the alignment of the gradient direction and the magnetic field. However, as we discussed earlier, thermal line width provides a lower limit on the effective thickness of the velocity channels. To study velocities on the scales less than the thermal velocity broadening one may invoke velocity centroids, which properties have been studied e.g. in Esquivel \& Lazarian (2005), Kandel et al. (2017b). The potential disadvantage of traditional centroids is that the entire spectral line is used while in some cases different parts of the line reflect magnetic fields in spatially different regions. This is the case, for instance, for the HI measurements where the galactic rotation curve provides rough information about the location of the emitting material. In addition, if the center of spectral line is saturated due to absorption effects, it is also good to use only the informative part of the line. Thus it is important to explore whether velocity centroids can be made a more flexible tool that can be used to study only a part of the spectral line. Below we experiment with such a measure.
Given the channel map data we introduce a new measure, i.e. Reduced Velocity Centroids that contain only part of the line:
\begin{equation}
C_{\Delta v, n} ({\bf x})=\int_{\Delta v} dv v^{n} \rho_{ppv} ,
\end{equation}
where the index $n$ determines the order of the centroid. Increasing $n$ enhances the effects of velocity, but, in practice, it also increases the noise. Reduced Velocity Centroids are useful both in the case of studying gradients from extended galactic disk data and also for the data from the wings of the absorption lines. In the study we explore how Reduced Velocity Centroid Gradients (RVCGs) trace magnetic field. To study the new measure we separate the line into three regions, namely the central part, the middle part and the wing. Figure \ref{fig:rc-1} illustrates the $n=1$ reduced centroid from a selected Alfven mode velocity data cube. The central portion is very much the same as the thin slice result, giving a good alignment in respect to the magnetic field. The middle portion of the reduced centroid also shows a fair alignment, but the number of data points get limited due to the limitations of the numerical resolution and the shot noise increase. Nevertheless, the gradient map after sub-block average still provides a nice fit to magnetic field direction. The wing part has also a good alignment to magnetic field. That actually tells us that one can use the part of the line to study magnetic fields. This can be portion of the line broadened by the galactic shear, e.g. the 21 cm line from the galactic plane hydrogen as we discuss below.\footnote{This can also be a part of the line corresponding to clouds at velocities formally forbidden by galactic rotation curve (see Stil et al. 2006), but arising from the turbulent stochastic nature of the velocity distribution in the galactic disk. High velocity clouds also present a case where RVCGs can be useful.} This provides a way to study the 3D structure of magnetic field and other phenomena (e.g. self-gravity and shocks) that are being traced by gradients.\footnote{Naturally, the mapping from the velocity space to the real space is distorted by turbulent velocities and therefore is accurate up to the turbulent velocity dispersion.} Using only wings potentially provides a way to study gradients using lines with high optical depth e.g. 12CO line. All of this should be discussed in detail in other publications, while here we provide a preliminary study of the new measure.
\begin{figure*}
\centering
\includegraphics[width=0.33\textwidth]{reduced-centroid-b23-alf-central.pdf}
\includegraphics[width=0.33\textwidth]{reduced-centroid-b23-alf-middle.pdf}
\includegraphics[width=0.33\textwidth]{reduced-centroid-b23-alf-wing-nanaddx.pdf}
\caption{\label{fig:rc-1} Maps of the reduced centroid for the central part of the line (left, $\delta v<0.2 \sqrt{\delta v_R^2}$), middle its middle part (middle,$0.2 \sqrt{\delta v_R^2}<\delta v< \sqrt{\delta v_R^2}$) and its wing (right,$\delta v>\sqrt{\delta v_R^2}$ ). Gradient directions are shown are in red, while the projected magnetic field directions are shown in blue.}
\end{figure*}
In fact, in MHD turbulence we expect that the overall line profile to be determined by the largest eddies that produce most of the dispersion. The gradients are not expected to depend on this. Therefore gradients in all three parts of the line reflect the alignment of the small scale eddies that are aligned in respect to the magnetic fields local to them.
\section{Comparison with the observational data}
\label{sec:obs}
\subsection{Application of VChGs and RVCGs}
For observations of galactic HI the use of different portions of the line is not any more just a test of our theoretical concepts. Due to the galactic rotation, different parts of the galactic HI have different velocities in respect to the observer. Therefore one can trace the magnetic field directions using the VChGs and the RVCGs. We illustrate our techniques using three regions consists of both diffuse HI regions and also 13CO data from a self-gravitating molecular cloud Vela C (Fissel et al. 2016). The HI region we compare with are from \cite{Clark15} for the studies of fliamentary structures in velocity channel spaces coming from the Galactic Arecibo L-band Feed Array HI (GALFA-HI) survey. GALFA-HI is an survey of the Galaxy in the $21 cm$ neutral hydrogen line. The data is obtained with the Arecibo Observatory 305 meter telescope. With the large aperture of the telescope, one get the angular resolution $\sim4'$, which is similar to Planck's best resolution ($\sim 5'$). For Vela C we compare our calculation with Balloon-borne Large Aperture Submillimeter Telescope for Polarimetr (BLASTPol, Galitzki 2014, Fissel et.al 2016), which is a 1.8-meter Cassegrain inferometric telescope detecting linearly submillimeter dust polarization.
\begin{figure*}[h]
\centering
\includegraphics[width=0.49\textwidth]{Fig20p.pdf}
\includegraphics[width=0.49\textwidth]{VelaCx.pdf}
\caption{\label{fig:AM_BS} Change of standard errors versus block size when applying different Gaussian Filters for GALFA-HI and Vela C. }
\end{figure*}
The regions we selected here are across $209.7^\circ - 196.6^\circ$, and declination (Dec.) across $22.5^\circ - 35.3^\circ$ for HI and $l\sim264.75-265.60$ and $b\sim 1.2-11.8$ for Vela C. .For the spectroscopic HI data, we stick with the channel with velocity ranges $0-2.944km/s$, which the velocity spectral line is peaked at. As a comparison, the dispersion of velocity across the whole map $10.3 km/s$, which means our selection of velocity channel is both thin (with a relative channel thickness of $\sim0.3$) and most representative compared to other channels. The analysis of Vela C is produce in Yuen et al. 2018. Here we reproduce one figure from that study to illustrate the ability of the VChGs to trace magnetic field in molecular clouds. Due to the thinner channels provided in the 13CO Vela C data (0.183 km/s), we use the relative channel thickness of $\sim 0.5$ for Vela C, which is about $1km/s$. We plot our VChGs for HI against Planck 353 Ghz polarization in the respective region by rotating both of them for $90\deg$. For Vela C, as it is a self-gravitating region, we keep the center part of VChGs unrotated while the boundary rotated for $90\deg$ (See YL17b for the criteria and the method of determining the gradient rotation density threshold) by estimating the volume density around the core region. We vary our block size and Gaussian kernel width (See Fig \ref{fig:AM_BS}) to acquire the optimal value ensuring the best combination of the alignment and the resolution. For the case of HI, the Gaussian filters have their respective gradient errors converge at block size of 100. For Vela C there is no such convergence plot, therefore we pick the lowest error point from the smallest Gaussian Filter to retain the data quality. The result is shown in Fig\ref{fig:GALFA} using the line integral convolution method (LIC, Cabral \& Leedom 1993). We find that the alignment measure is in general pretty well for the two regions.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Proposal1x.pdf}
\caption{\label{fig:GALFA} Magnetic field drapery patterns probed using VChGs (rotated $90\deg$ created by the line integral convolution method (LIC, Cabral \& Leedom 1993) for two observational examples GALFA-HI and Vela C (Fissel et al. 2016, reproduced from Yuen et al 2018). We here compare polarimetry data from different sources as the cloud density increases, and plot two of the magnetic field directions with LIC by rotating polarization for $90\deg$. The intensities are the {\it LIC-processed} texture that is proportional to the tracers intensity of the two maps.}
\end{figure*}
Note that taking the thickness of the channel map of the order of the velocity dispersion one produces intensity-dominated channel maps, which provides a rough structure of the column densities with the grid size of the order of the large scale turbulent eddies. The gradient study of such a map of which is equivalent to the study using the IGs. Therefore the corresponding channel map gradients can be associated with the IGs that trace both the combination of the effects arising from magnetic field and shocks, as we discussed in YL17b. Combining this with the magnetic field information e.g. obtained with the VChGs is very synergetic for studying interstellar shocks propagating through galactic disk (See Sec 11.3 for a complete discussion)
\section{Additional effects and prospects of the research}
\label{subsec:selfgrav}
The technique of studying magnetic fields with velocity gradients shows many promising directions for further studies. We outline a few directions
without getting into much detail within this publication.
\subsection{Effect of self-gravity}
In the presence of self-gravity we expect that the gradients to change their alignment in relation to magnetic field. Naturally, if this happens over the block over which the calculation of the gradients is performed, this will result in the higher than average uncertainty in determining the direction of gradient. This is what we observe in our simulations, as illustrated in Fig \ref{fig:grav_disp} shows how the {\it raw} gradients (i.e. gradients before block-averaging) react when they come close to the self-gravitating core. Instead, we calculate the dispersion of raw gradients within a block of selected size, and move the block from the left of the core to the right. From Fig \ref{fig:grav_disp} we clearly see, as we come closer to the gravitational center (core center), the dispersion of gradients significantly increases. We also observe that the increase of dispersion is smoothed out and the peak of dispersion is deviated from the core center as we are using bigger block size. However, the average level of dispersion around the core region (about $\sim 1.2 radian$is still significantly higher than that of the diffuse region, which is $\sim 0.8 radian$.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{grav_disp_1.pdf}
\caption{\label{fig:grav_disp} Dispersion of {\it raw} gradients with respect to block position (which we pick the center of the block to be the x-axis) for several choices of block size.}
\end{figure}
In addition, our earlier work, in particular YL17b, identified the synergistic way of using simultaneously the VCGs and the IGs in order to reveal both regions dominated by self-gravity and shocks and to trace magnetic in the presence of these phenomena. In particularly, in the presence of self-gravity both the VCGs and the IGs are changing their direction by 90 degrees, thus getting parallel to the magnetic field direction. However, the change direction of the IGs happens earlier than that of the VCGs and this provides the way of, first of all, identifying regions dominated by self-gravity, second, tracing magnetic fields within self-gravitating regions without any polarimetry data. Naturally, this effect is also present when we use the VChGs. The case thick velocity channels provides the information about the IGs, while the case of thin velocity channels provides the information about the velocities. Figure \ref{fig:grav} illustrates the difference in the response with time of the gradients measured within thin and thick channels. The time is measured in simulations from the moment of self-gravity being turned on. Due to this effect in observations we expect the region of where the gradients in the thick velocity channels are turning 90 degrees in respect to magnetic field to be more extended than the region over which the gradients in thin velocity channels are turned 90 degrees in respect to the magnetic field. Therefore while at the regions far from the self-gravity center, the gradients in thin and thick velocity channels are going to be aligned, they will turn 90 degrees in respect to each other at some distance from the self-gravity center and closer to the self-gravity center they are expected to get back being aligned. The detection of such a change is a signature of a self-gravity effect which can be used to indicate when it is rotated 90 degrees gradients that trace magnetic fields.
Turning of the gradients in thin velocity channels compared to thick velocity channels is also expected to happen within strong shocks. However, shocks can be distinguished from the self-gravity regions both through morphological differences and due to differences in the column densities. Thus in most cases, one can trace magnetic fields, shocks and identify regions of gravitational collapse through combining the gradients within thin and thick velocity channels.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{tAM-VChG-dn60.pdf}
\caption{\label{fig:grav} The variation of the AM of thick (red) and thin (blue) slices in time after the gravity is turned on.}
\end{figure}
\subsection{Use of interferometers and the effect of noise}
We would also like to examine the effect of noise for both RVCGs and VChGs. After obtaining channel and reduced centroid maps from cube Ms1,6Ma0.53 at relative channel width of $0.2$, we add white noises as we did in \cite{LYLC17} on top of the maps, i.e. add white noises according to the root-mean-squared value of map intensities. Figure ~\ref{fig:noise} shows that with noise level being $0.1$ of the map intensities. One may already observe from the map that there are already some dirt altering the alignment of gradients to projected magnetic field. We therefore add a Gaussian filter with width of 2 pixels according to \cite{LYLC17} to negate the effect of noise. Figure ~\ref{fig:noise-change} shows the alignment measure versus the amplitude of noise for VChGs and RVCGs respectively after the Gaussian filter. The alignment remains fairly good until the ratio between noise amplitude relative to the map intensities is close to 1, and significantly drops to zero as the ratio is around 3. This simple test illustrates the strength of VChGs and RVCGs in predicting alignments in strong noise environments with the aid of small-width Gaussian filters.
Our results also show that RVCGs and VChGs are comparable in their performance both when the turbulent broadening is larger and smaller than the turbulent line width. Revealing the relative advantages of these two techniques requires a more detailed study and is beyond the scope of the present paper.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Fig18x.pdf}
\caption{\label{fig:noise} The alignment between gradients and projected magnetic field with white noises added to the original maps. }
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Fig19pp.pdf}
\caption{\label{fig:noise-change} (Left) The alignment measure versus the amplitude of noise for VChGs and RVCGs respectively. A Gaussian Filter of width 2 pixels have been applied to all the map with noises. (Right) The same procedure as from the left but we keep density constant for low $M_s$ case.}
\end{figure*}
In Figure \ref{fig:practical} we show how the VChGs change in the presence of noise. We see that it is advantageous to remove spatial frequencies in order to decrease the noise aside from the Gaussian Filter method. Incidentally, similar results have been achieved in mimicking LYLC and YL17b with other types of gradients. We would like to note that the interferometric studies are frequently missing the low frequencies and therefore the filtering of the low frequencies is happening naturally in this case.
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{sigma2.pdf}
\includegraphics[width=0.49\textwidth]{wavenumber-02dv.pdf}
\caption{\label{fig:practical} (Left) Illustration of the noise suppression technique works for VChGs. (Right) Effect of the low wavenumber removal on the VChGs.}
\end{figure*}
\subsection{Pure velocity caustics}
Both velocity and density gradients in subsonic turbulence are directed perpendicular to magnetic field. Therefore to isolate the effect of velocity we create PPV cubes using constant density. For such cubes all the intensity variations arise from the velocity crowding. To mimic subsonic turbulence we introduce the thermal line width which produces broadening larger than the turbulent velocities, in our case the thermal line width is 3.78 times of that of turbulent velocities. Our studies of velocity channel intensity correlations in \cite{2000ApJ...537..720L,2004ApJ...616..943L} suggests a significant suppression of the amplitude of the correlations in such a case and, as a result, high influence of noise. Thus in the subsequent publications, e.g. \cite{2017MNRAS.464.3617K}, we claimed that velocity centroids provide a better way on studying velocity fluctuations when the thermal linewidth exceeds the turbulent line broadening. Nevertheless, for the gradients we do not see such a lose of information (See right of Figure \ref{fig:noise-change}), which we relate to the fact that for gradients the differences of the velocities are important and therefore the constant thermal width does not much affect them. This is an interesting effect that calls for further studies of the relative advantages of the channel map gradients and velocity centroid gradients. In particular, the studies of gradients of the velocity centroids may be advantageous in order to provide the complementary information.
\subsection{Physics of velocity gradients at different scales: ways to determine B-strength}
Real astrophysical flows include both regular and turbulent motions. However, an interesting property of turbulence that it increases the velocity gradients as the scale decreases. Therefore the velocity gradients induced by turbulence are expected to dominate. In addition, to increase the signal to noise, the contribution of the large scale motions can be removed using spatial filtering.
Properties of MHD turbulence are different at different scales. For instance, if turbulence is injected at scale $L$ with velocities $v_L$ larger than the Alfven velocity $V_A$, up to the scale $l_{A}=L M_A^{-3}$, where $M_A=v_L/V_A$ is the Alfven Mach number, the turbulent motions are marginally affected by the magnetic field presence. Therefore if the velocity gradients are measured at scales larger than $l_A$ they cannot trace magnetic field. As a result, for super-Alfvenic motions the influence of gradients from scales larger than $l_A$ is pernicious in terms of magnetic field tracing. Motions at these large scales should be filtered out in order to increase the VChGs accuracy as a magnetic field tracer.
At the same time the change of the properties gradients around the scale $l_A$ can be used to find the fluid magnetization. Indeed, by observing the change of the VChGs dispersion which changing the scale at which the gradients are calculated, one can determine $l_A$ and then calculate
$M_A=(L/l_A)^{1/3}$. The calculations of $L$ can be done spectroscopically $L$ as demonstrated by Chepurnov et al. (2010) using the Velocity Coordinate Spectrum (VCS) technique suggested in Lazarian \& Pogosyan (2006). If $M_A$ is determined, the Alfven velocity can be obtained by associating the $v_L$ with the velocity dispersion that is measured spectroscopically. Note, that this technique is applicable to finding the magnetic field in super-Alfenic turbulence, i.e. turbulence with $M_A>1$ where the traditional Chandrasekhar-Fermi technique fails (see Falceta-Golzalez et al. 2008)
In practical terms, we expect to observe that the magnetic field to be organized over the patches of the scale $l_A$ with the mean magnetic field in different patches to change randomly from one patch to another. As gradients reflect magnetic field direction we expect a similar coherent organization of velocity gradients over the patches of $l_A$.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{a5-islands.pdf}
\caption{\label{fig:islands} The distribution of alignment measure between {\it raw channel gradients} and projected magnetic field in our super-Alfvenic example Ms0.2Ma2.0. }
\end{figure}
The VChGs that we have discussed are tested with the simulations of MHD turbulence. The MHD approximation is applicable to partially ionized gas to the scales significantly larger than the scale of the neutral-ion decoupling (see Lithwick \& Goldreich 2001, Xu \& Lazarian 2017). At such scale the rate of turbulent variations is slow compared to the rate of neutral-ion collisions and therefore both species behave as one magnetized fluid. For Kolmogorov or GS95 scaling the rate of the eddy rotations changes as $v_l/l\sim l^{-2/3}$, i.e. increases with the decrease of the scale. Thus at a sufficiently small scale $l_{decoup}$ the ions do not collide with the ions frequently enough and get decoupled. For $l<l_{decoup}$ neutrals can start their own unmagnetized cascade and therefore their motion will stop reflecting magnetic field.
We have discussed that the VChGs are tracing the magnetic field at the smallest scale that is resolved by the telescope. If this scale is larger than $l_{decoup}$ then the velocity gradients represent magnetic field as we discussed in this paper. The change in the orientation of the velocity gradients in the vicinity of $l_{decoup}$ can be used to establish this important scale. In particular, we expect to see the change of the relative orientation of the gradients obtained with neutrals with maps smoothed over scales $l>l_{decoup}$ and the high resolution maps resolving $l<l_{decoup}$. We also expect to see the differences of the velocity gradient orientation of neutrals and ions at scales $l<l_{decoup}$. Both effects can be used to establish $l_{decoup}$. The latter scale is related to the magnetic field strength as discussed in Xu et al. (2016) which presents a new way of magnetic field strength measurements. In other words, the VChGs present a tool that can be used both to trace magnetic field in molecular clouds and, given enough resolution, to test the neutral-ion decoupling scale.
We believe that velocity gradients can trace magnetic field in most turbulent astrophysical environments. In particular, we expect the velocity gradients to trace magnetic fields in the accretion disks, where recent polarimetry showed that the grain alignment happens in respect to the radiation (see Rue et al. 2017, Lazarian \& Hoang 2007).
Naturally, the changes of velocity gradient properties that we described for the case of the VChGs can also be studied with the VCGs and the RVCGs.
We will present detailed testing of the ability of the velocity gradient technique to get the magnetic field strength elsewhere.
\subsection{Constraining VChGs directions using expectations from MHD turbulence theory}
It is important to make magnetic field tracing as precise as possible. In the sub-block averaging (YL17b), the blocks are created by evenly dividing the maps into localized regions and identifying the peaks of the Gaussian fits in these blocks as the most probable values within the block. Theory of MHD turbulence provides ways to improve the procedures for magnetic field tracing. Below we experiment with two ways of improving the alignment and we view this as just initial steps to the MHD turbulence theory based automated procedure of velocity gradient tracing that employs Machine Learning (see Le Cun 1990).
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{method.pdf}
\caption{\label{fig:newmethod} A pictorial illustration on how Moving Window (left) and Angle Constraining (right) should be pictorially in our gradient technique. }
\end{figure}
The Moving Window (MW) approach is an attempt to employ the sub-block averaging in a continues rather than a discrete manner. As magnetic field provides a continues representation it is advantageous to move the block and provide calculations as the block moves along the magnetic field lines that it traces. On the left of Fig \ref{fig:newmethod} shows pictorially how we improve the alignment: When there is an abnormal gradient vector compared to the neighboring vectors, we rotate the abnormal vector so that a smooth field line is formed. Mathematically the rotation can be handled by performing smoothing on both the cosines and sines of the {\it raw} gradient angle, which is a convolution of an averaging kernel with the raw cosine and sine data. We therefore pick the Gaussian kernel with different width to test how good the moving window is in our synthetic map. After the smoothing we apply the sub-block averaging to the processed gradient angle. Figure \ref{fig:MW} shows the strength of moving window with respect to the smoothing strength. One can see the alignment quickly rises until a window size of 0.75 pixel and being saturated after that. Similar to the sub-block averaging, one can estimate errors of the fitting in order to find the optimal moving window size.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{AM-MW.pdf}
\caption{\label{fig:MW} The effect of moving window to the alignment measure. }
\end{figure}
Another possible improvement is to use the expectations of the MHD theory in terms of applying the Angle Constraint (AC). Observations necessarily use the global system of reference related to the mean magnetic field. The probability distribution of magnetic field in terms of wavevectors parallel $k_{\|}$ and perpendicular $k_{\bot}$ to the mean magnetic field is provided in \cite{LP12}. This distribution predicts that the mean angle variations are the same at all scales. This variation can be determined with higher accuracy with the larger blocks and then can be used as a constraint when the direction of gradients is established within noisy small sub-blocks. Our application of the procedure of the AC provided a moderate improvement of the alignment measure of ~0.1 depending on the system's $M_A$. However, this procedure has minimal effect when $M_A$ is larger than unity. We believe that AC can be a useful part of the future algorithms of the VChGs calculation.
\subsection{Dependence on the Alfven Mach numbers}
For observational tracing of magnetic field it is important to know what is to expect in terms of $AM$ dependence on the Alfven Mach number $M_A$. The Figure \ref{fig:maam} provides these dependence for our simulations. We see that the alignment decreases as $M_A$ approaches to unity. We attribute this to large angle variations of magnetic field along the line of sight. As $M_A$ increases, especially when the turbulence gets superAlfvenic, it is important to remove the low spatial frequencies. The corresponding procedure was shown to work in LYLC17, but here we present the direct application of the VChGs to the data without any spatial filtering. The procedures to improve the $AM$ for the VChGs are elaborated elsewhere.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{maam.pdf}
\caption{\label{fig:maam} A plot showing the dependence of $M_A$ for alignment measure of channels (Selected relative channel width 0.2, block size 40, set B from YL17b.). The color of each point shows the logarithm of sonic mach number. }
\end{figure}
\section{Discussion}
\label{sec:discussion}
\subsection{Comparison with earlier studies}
The current study continues our series of research papers that are motivated by the modern understanding of the theory of MHD turbulence (see \citealt{2013SSRv..178..163B} for a review), in particular, the notion of the perpendicular cascade that is at the heart of this theory. Due to this cascade which can be visualized as hierarchy of turbulent eddies mixing magnetized plasmas perpendicular to the local direction of magnetic field (see \S \ref{sec:theory}), the gradients of velocity are expected to be maximal perpendicular to the local magnetic field, thus revealing its direction.
The first paper that employed this theoretical arguments to trace magnetic fields using velocity centroids was GL17. This study attempted a quantitative approach, e.g. it introduced the alignment measure $AM$ that is used in the subsequent publications, including this one. However, the actual start of the quantitative tracing of magnetic fields with gradients relies on the block-averaging procedure introduced in YL17b. This procedure allowed determining both the pointwise direction of magnetic field and estimate the uncertainty of this direction. As a result, this procedure was used in all the papers that followed. In this paper, the block averaging approach used to identify the regions where the direction of gradients changes 90 degrees as a result of the effect of self-gravity. We noticed that the blocks over which this transition happens exhibit a much worse single direction fitting, which corresponds to higher level of fitting errors. We showed that my measuring the level of the errors it is possible to identify the regions of gravitational infall.
A sister technique tracing magnetic fields using synchrotron intensity gradients (SIGs) was introduced in LYLC, where it was was shown that it is possible to establish the optimal size of the block for averaging providing the maximum resolution of the reliable magnetic field tracing. We employed this approach in the present paper.
The correlation between density gradients and magnetic fields was empirically revealed by \cite{Soler2013}. The technique did not attempt to trace the
magnetic field but was intended for establishing the statistical correlation between the observed structures and the magnetic field direction. In terms of observational studies the technique in \cite{Soler2013} relied on the comparison of the observed intensity gradients and polarization and the study was answering the question whether the preferential orientation of density structures at a given column density is parallel or perpendicular to the magnetic field. \cite{Soler2013} introduced Histograms of Relative Orientation (HRO) that provided an empirical insight of the probability of the intensity gradients to be aligned with magnetic field depending on the density of the media. Note, that if plotted in the pictorial plane, the orientation of magnetic field and intensity gradient present a rather chaotic patten, which prevents one from tracing magnetic field with the gradients. It is in the statistical sense that the HRO establishes the change of the alignment with the ambient density. A detailed comparison of the HRO (see also Soler et al 2013,2017, Soler \& Hennebelle 2017) with our approach is provided YL17b.
The effect of partial alignment of density gradients with magnetic field in diffuse medium is easy to understand from the point of view of the gradient theory in this paper. The velocity mixing motions corresponding to the MHD cascade in diffuse media project their properties on the density structures. Thus low sonic Mach numbers such structures are preferentially elongated along magnetic field and velocity and density gradients behave similarly. However, the difference between density and velocity structures becomes significant as the Mach number increases (see Kowal et al. 2007).
Thus the density gradients are inferior in tracing magnetic fields compared to velocity gradients.
Note, that the procedures of calculation of gradients in HRO do not included block averaging and therefore the maps of density gradients obtained in that study exhibit chaotic variations of magnetic fields, which makes it impossible to trace magnetic fields even in low Mach number turbulence. By applying our block averaging procedure to densities we can introduce the Intensity Gradient (IG) technique that is different from the HRO. Within
this technique one can determine both the direction of the density gradients and the uncertainty over the image plane. This information is complementary to that from HRO. The sensitivity of density gradients to shocks within the IG technique can be used to identify shocks in the interstellar medium, as it discussed in YL17b.
The preferential alignment of filaments within HI channel maps in the direction of magnetic field was established observationally in \cite{Clark15}. This work got significant resonance as it suggested the possibility of using this alignment to improve the separation of CMB polarization from the galactic microwave foreground. In the work by \cite{Clark15} the thickness of channels was empirically chosen to be 3 km/s and the authors identify the structures with actual density filaments that exist in HI. Indeed, we were told (Clark, private communication) of the correspondence of the HI filaments in channel maps and the filaments observed in dust which suggests that the filaments are actual real physical objects rather than caustics. In this situation, it is necessary to study what is happening with the filaments as the channel map thickness increases. The theory in Lazarian \& Pogosyan (2000) suggests that \cite{Clark15} studies "thin channel maps" where most of the structures should be due to velocity caustics. Indeed, assuming that the typical turbulent velocity dispersion is $\sim 10 km/s$ and that the turbulence is Kolmogorov-type, i.e. $v_l\sim l^{1/3}$, one can estimate that the turbulent motions at scales larger than $L/(10/3)^3\approx 0.03L$, where $L$ is the injection scale, are in the "thin slice regime" (Lazarian \& Pogosyan 2000). Using the injection scale of $\sim 100$~pc obtained in Chepurnov et al. (2011), one concludes that on scales larger than 3 pc the intensity of fluctuations in the channel maps is produced to a significant degree by velocity caustics rather than the real physical entities, i.e. filaments. The calculations in \cite{Clark15} are produced on the maps smoothed up to $FWHM=30'$. Based on the galactic rotation curve, the structures from the channel maps with velocity ranges $-56km/s \sim -10km/s$ or $10km/s \sim 56km/s$ are mostly consists of intensity structures. However, the "velocity crowding effect" is expected to be more severe in channels maps at $-10km/s \sim 10km/s$, in which the structures in the channel maps are coming from contributions from clouds in different physical locations along the line of sight.
What is peculiar about HI is still an issue of further investigations. Apparently, we do not observe any aligned features similar to filaments in our CO studies. At the same time velocity gradients trace successfully magnetic fields both in HI and denser regions like CO. We defer the issue of the nature of the filaments in \cite{Clark15} to future publications.
We feel that the advantage of the velocity gradient technique is that it is rooted both in the MHD turbulence theory (see Brandenburg \& Lazarian 2013 for a review) and the theory of PPV statistics (see Lazarian \& Pogosyan 2002, Kandel et al. 2016). It is important that the velocity gradients have a straightforward relation to the compressible MHD turbulence theory. For instance, using the VChGs, we explore the effect of three fundamental modes of MHD turbulence, i.e. Alfven, slow and fast, on the analysis. These modes affect the properties of the VChGs differently and our present opens a new avenue for the development of the gradient analysis. In particular, we feel that similar to the techniques of separating Alfven, slow and fast modes using correlation anisotropy as in the synchrotron study in \cite{LP12} and subsequent studies of velocity correlations in \cite{2016MNRAS.461.1227K,2017MNRAS.464.3617K}, the velocity gradients can be used to determine the regions with the prevalence of compressible or incompressible turbulence as well as determining the relative contribution of different modes. The corresponding procedures will be discussed elsewhere.
\subsection{Towards a unified gradient technique}
\label{subsec:upic}
\begin{table*}[t]
\centering
\label{tab:gradient_table}
\caption {A summary on the gradient alignment in different physical regime}
\begin{tabular}{c c c c c c c}
Papers & Gradients & Diffuse Alfven modes &Diffuse Slow modes &Diffuse Fast modes & Self-grav. & Self-grav. rotation speed\\ \hline \hline
GL17,YL17 & VCGs & $\bot$ & $\bot$ & $\parallel$ &$\parallel$ & slow\\
YL17,YL17b & IGs & $\bot$ & $\bot$ & $\parallel$ &$\parallel$ & fast\\
LYLC & SIGs & $\parallel$ & $\parallel$ & $\bot$ & $\parallel$ & not rotated\\
This work & thin VChGs & $\bot$ & $\bot$ & $\parallel$ &$\parallel$ & slow\\
This work & thick VChGs & $\bot$ & $\bot$ & $\parallel$ &$\parallel$ & fast\\ \hline \hline
\end{tabular}
\end{table*}
The properties of magnetic field in magnetized turbulence are well established by a series of groundbreaking papers starting with GS95. In fact, the application of the turbulence theory to observations has been explored in a number of studies. In particular, using the analytical treatment of the PPV developed in LP00 and \cite{2004ApJ...616..943L,2006ApJ...652.1348L,2017MNRAS.464.3617K}, we provided the treatment of the anisotropies of correlation functions of the channel map intensities and of the velocity centroids in \cite{2017MNRAS.464.3617K}. These studies shed light on the properties of velocity gradients that we study in the present paper.
Our series of papers starting with GL17 and YL17a provide a way on tracking magnetic field direction in realistically turbulent interstellar medium (see also LYLC,YL17b). We illustrated that the gradient technique is superior in terms of tracing magnetic field directions to the anisotropy tracing technique described in \S \ref{sec:anisotropy}. Nevertheless, these methods are complementary. Their synergy could provide both the detailed structure of magnetic field through using the velocity gradients as we discuss in the present paper and the information about the turbulence compressibility as it can be tested using the anisotropy of the correlation functions of the channel map intensities and velocity centroids (see Kandel et al. 2016, 2017).
In our earlier papers we were focused on velocity centroids as the measures of velocity that are available from observations, the present paper introduces new measures, namely, gradients within channel maps or the VChGs and reduced velocity centroids gradients (RVCGs). These measures enhance the abilities of the gradient technique. In particular, both measures can analyze a part of a spectral line which has significant advantages, e.g. in the case of using galactic rotation curve to get the magnetic field direction variation along the line of sight. Other ways of constructing measures of turbulent velocity from observed spectral lines are also possible. In YL17b we showed advantages of using higher moments of velocity centroids when the observational data has low noise. Similarly, higher moments of reduced centroids were suggested for the RVCGs within this paper. The study of these different measures is another important avenue for the gradient research. These measures are affected by different degree by the fluctuations of density and this opens a way to study better disentangling density and velocity effects in the gradient studies. As we discussed in detail in YL17b the two fields have different properties and the difference is important for e.g. identifying the regions of collapse induced by self-gravity.
The application the technique to different spectral lines corresponding allows one to study magnetic fields separately in different regions along the line of sight. This, in particular, is valuable for molecular clouds where by choosing the corresponding transitions one can study a cloud magnetic fields in depth layer by layer.
These and other ways of gaining new information from gradients will be demonstrated elsewhere.In addition, velocity and intensity gradients (IGs) can be used together with the synchrotron intensity gradients (SIGs) and this was demonstrated in LYLC.
The procedures that we developing for the VChGs are also applicable to the IGs, the VCGs and the SIGs. In fact, the block averaging makes our IG technique very different from that in Soler et al. (2013). The latter work s
Compared to the VCGs and the IGs, the VChGs include the advantages of both diagnostic tools, while being influenced by the interfering effects. In fact, as the thickness of the velocity channels increases the VChGs gradually transfer to the IGs. However, compared the IGs, the VChGs contain much more information as the transfer from the thin to thick gradients is taking place. At the same time, it is not right to think that VChGs completely overshadows the IGs. For instance, the IGs can be applied to dust emission, where the is no lines to study using the VChGs.
The VChGs and the RVCGs provide a useful tool for studying self-absorbing media (see \citealt{GLB17} for a discussion of the VCGs in self-absorbing media) and lines with complex structure, e.g. velocity lines have multi-peaks. Both the VChGs and the RVCGs allow using a part of a spectral line. For strongly self-adsorbing media the useful parts could be the wings. At the same time, for the lines with multi-peaks, the complexity of the line may present the complexity of the spatial structure of the emitting region. In the latter case studies of the separate peaks can provide the information about the magnetic field within a complex region. Detailed studies of the utility of the VChGs and the RVCGs as well as their complementary nature will be provided elsewhere.
\subsection{Prospects of the technique}
The galactic rotation opens new prospects for studying the detailed structure of magnetic field in the Milky Way. The position of the Solar System within the galactic disk makes it impossible for us to use far infrared polarimetry to study magnetic fields of most of the molecular clouds. Indeed, for most of such studies the line of sight inevitably crosses more than one cloud. This confusion, combined with the failure of grain alignment at large optical depths (see Lazarian 2007) makes a polarimetric study most of the star formation hotbeds impossible. In comparison, the approach based on using gradients can employ the galactic rotation curve to separate the contribution from different clouds. Moreover, the velocity gradients can probe the magnetic connection of the diffuse gas and molecular clouds. Such studies are impossible with far infrared polarimetry due to the signal from the diffuse media being too weak.
Velocity gradients add up differently compared to the Stokes parameters that are used in polarization. Therefore for exploring magnetic field in the galactic plane where we may expect significant direction variation along the line of sight, the VChGs and the RVCGs are advantageous compared to the VCGs. Using the former tools one can subdivide the line into velocity segments of the order of the turbulent velocity dispersion of the order of $\sim 7$ km/s and calculate the magnetic field direction for these segments. Within these segments the thin channel maps are used for calculating the VChGs and RVCGs. To compare with the polarization we then can use the mock Stock parameters similar as it is done in \cite{Clark15} for the filaments.
In terms of resolving the spacial structure of magnetic field using the galactic rotation curve, for the atomic hydrogen studies the spatial resolution depends both on the direction of study, the velocity range as well as on the turbulent velocity dispersion. The latter provides the grid size $\delta V$ that should be multiplied to the visual shear along the line of sight $(dV_{gal}/dz)^{-1}$ in the direction of the observations.
If the direction of magnetic field is changing along the line of sight the VChGs and the RVCGs can provide a coarse graded picture of the changes of the magnetic field direction perpendicular to the line of sight. In this way we can distinguish the variations of regular magnetic field over the cells that are larger than $V_L (dV_{gal}/dz)^{-1}$.
There is a particular domain where velocity gradients present the only way for magnetic fields tracing. This is the case of high velocity clouds. They are tenuous compared to the interstellar medium along the line of sight and therefore any polarization associated with them is not possible to detect. However, these clouds are vivid in the velocity space, which gives a way to use either the VChGs or the RVCGs to map their magnetic fields.
Another domain where dust polarimetry fails to trace magnetic field is related to circumstellar accretion disks. The dust there is aligned by radiative
torques, which according to the alignment theory in Lazarian \& Hoang (2007) align grains in the vicinity of stars in respect to the radiation direction
rather than the magnetic field direction (see also Tazaki et al. 2017). Therefore in spite of the ability of ALMA to resolve some of such disks, it
cannot really study their magnetic field structure. We expect that velocity gradient will be able to trace magnetic fields in circumstellar accretion disks.
The change of the channel thickness changes the relative contributions of the velocity and density fluctuations into the channel map. Taking the thickness of the channel map of the order of the velocity injection one produces intensity-dominated channel maps and therefore the gradient study of such maps is equivalent to the study using the IGs.
It is also important that the VChGs can be applied to interferometric data with in the situation when no single dish observations are available. This extends the application of the technique to distant and extragalactic objects in an important way.
We stated above that the observational information on turbulence is available only in the system of the mean field. There are exception to this, however. For instance, if only a thin slice of turbulence is seen due to the dust absorption effects (see \citealt{2017MNRAS.470.3103K}), the effects of the local system of reference may get important. Similarly, if the object under study and the turbulence scale are comparable, then the dispersion of magnetic field is going to decrease with the scale.
The synergetic use of the polarimetry and the velocity gradients can be very beneficial. For instance, velocity gradients can be employed to study magnetic fields in the disk of the galaxy, where the traditional far infrared polarimetry suffers from the effects of confusion as many clouds can be along the same line of sight. The galactic rotation curve can help isolate different clouds in the velocity space and allow studying their magnetic fields separately.
The available large telescopes and interferometers can provide much better resolution that the far infrared telescopes on the balloons. The comparison of the polarimetry and velocity gradients can provide the information about the regions of the gravitational collapse as the velocity gradients will change their direction for such regions. As this low resolution identification of the gravity-dominated regimes is done, the velocity gradients may be used to study the details of the magnetic field structure that are not available with the existing far infrared polarimetry.
As different molecules are produced at different depths inside clouds, the velocity gradients can study 3D magnetic field structure of the clouds. This provides a new dimension for the magnetic field studies.
It is important to note, that the far infrared polarimetry and velocity gradients do not provide identical information. First of all, the magnetic fields that are traced by the polarimetry are weighted by the dust density, the latter is being proportional to the gas density. Within the VChG approach the contribution of the density fluctuations are reduced and potentially the directions of magnetic field measured by the VChGs can be closer to the actual projected magnetic field.\footnote{Velocity Centroid gradients of different orders (see YL17a) should have different weighting of the fluctuations of density along the line of sight. As a result, potentially, combining those with the VChGs one can get the actual projected magnetic field.} Moreover, the addition of gradients along the line of sight is happening differently from the addition of the Stocks parameters Q and U in the case of polarimetry. This opens a way of probing the 3D magnetic structure combining polarimetry and velocity gradients. We discuss this possibility in Yuen et al. (2018).
While synergetic with polarimetry, velocity gradients present an independent way of magnetic field studies that does not require polarimetric information. Naturally, testing the velocity gradients with as much polarimetric data is advantageous to get more confidence in the new technique. Eventually, velocity gradients should be used on their own, however.
We have identified (see YL17b) the 90 degree change in the relative orientation of velocity gradients and magnetic field directions in the case of self-gravitational collapse. Other situations when regular flows dominate turbulence can be present. Therefore it is important to provide more numerical studies of velocity gradients in expanding HI regions, supernovae explosions etc. in order to see whether one should expect the change of the direction of velocity gradients in other astrophysical settings. If this happens, it is important to test our approach based on calculating of the uncertainties of the fitting of velocity gradients as the way of identifying the change.
Studies of the microwave foregrounds are extremely important for the attempts to detect and explore the polarization induced enigmatic cosmological B-modes. Separating this polarization from the polarization arising from galactic foreground requires a breakthrough in understanding of galactic magnetic fields. Velocity gradients, e.g. the VChGs and RVCGs, as well as SIGs provide an independent way of mapping magnetic fields. The corresponding information can be used both independently to predict the foreground polarization, or preferably, as a prior for the polarization studies.
\section{Summary}
\label{sec:conclusion}
In this paper we have shown that a new measure, i.e. gradients calculated within velocity channel maps (VChGs) can trace magnetic field both in diffuse media and in molecular clouds. The VChGs of the thin channel maps that carry the information about turbulent velocities, while VChGs of the thick maps carry the information about the turbulent densities. The essence of emerging technique is to vary the channel thickness to get complementary information on both magnetic fields and shocks. As velocity gradients are a more direct tracers of magnetic field most of the paper is devoted to the VChGs calculated for the thin velocity channels. We compared the abilities of the VChGs and another new measure Reduced Velocity Centroid Gradients (RVCGs) that also traces magnetic fields with spectroscopic data.
In particular,
\begin{enumerate}
\item We demonstrated the alignments of VChGs obtained with basic MHD modes, i.e.
\begin{enumerate}
\item the VChGs from Alfven modes are perpendicular to magnetic field, and demonstrate the highest alignment;
\item the VChGs from Slow modes are also perpendicular to magnetic field but show somewhat reduced alignment compared to that from Alfven modes; their alignment drops faster as channel width increases;
\item the VChGs from Fast modes are parallel to magnetic field.
\end{enumerate}
\item We showed that the VChGs are more powerful in tracing magnetic field directions than the channel map correlation functions that we proposed earlier, namely, the VChGs provide more detailed information about magnetic field and can trace magnetic field for both supersonic and subsonic turbulence.
\item We applied the VChGs to the observational HI data and compared the VChGs tracing of magnetic field with the {\it PLANCK} polarimetry data.
\item We found that the RVCGs are comparable in their performance to the VChGs and both techniques can be used to trace magnetic field in diffuse interstellar gas and its interface with molecular clouds, molecular clouds, high velocity clouds etc.
\item We believe that the VChGs are synergetic to other ways of magnetic field studies, in particular, to the far infrared polarimetry. Nevertheless, it is a independent way of studying magnetic field which can be used to trace magnetic fields on it own.
\item We claim that the VChGs and RVCGs can trace magnetic fields in situations when the traditional far-infrared polarimetry fails, e.g. due to the failure of dust to be aligned or due to the confusion effect typical for studying molecular clouds at low galactic latitudes.
\item We demonstrated the advantages of the synergistic use of the different types of gradients (e.g. synchrotron intensity, spectral line intensity) paving a way for a new Gradient Technique of studying magnetic field ecosystem, shocks and self-gravitational collapse. The technique can provide the magnetic field structure that is valuable for disentangling galactic foregrounds and CMB polarization.
\end{enumerate}
{\bf Acknowledgments.} Elucidating discussions with Chris McKee and Susan Clarke are acknowledged. We thank Victor Lazarian for a number of suggestions in improving our presentation. We also thank Susan Clark and Laura Fissel in providing easy access to their data for our analysis. We thank the technical support from Ka Wai Ho for the 3D visualization of our data. AL acknowledges the support the NSF grant DMS 1622353 and AST 1715754. The stay of KHY at UW-Madison is supported by the Fulbright-Lee Fellowship.
|
\section*{Introduction}
Microporous flexible metal-organic framework materials
are fascinating both from a fundamental point of
view and for their numerous potential applications such as
gas storage, gas separation, sensors, drug delivery, etc.\cite{Ferey09,
Alhamami14,Schneemann14,Coudert15,Ferey16}
A well-studied example is the MIL-53
family,\cite{Serre02} with formula
M(OH)(C$_8$H$_4$O$_4)$,
where is M is a trivalent species
such as Cr, Sc, Al, Ga or Fe.
These structures consist of zigzag M-OH-M-OH$\dots$ chains,
crosslinked by 1,4-benzodicarboxylate O$_2$C-C$_6$H$_4$-CO$_2$ (bdc)
units (\fig{mil53x}). Each M is coordinated by two oxygens of OH units
and four carboxylate oxygens yielding octahedral oxygen
coordination.
\begin{figure}
\includegraphics[width=85mm]{mil53x.pdf}
\caption{Structure of MIL-53(Cr). Cr atoms green, O red,
C gray, and H yellow. (a) bdc linkers joining zigzag Cr-OH-Cr-$\dots$ chains.
(b) Each zigzag chain is coordinated with four neighboring chains;
each Cr is octahedrally coordinated with six O. (c) Narrow pore ($np$) phase
showing bdc rotations. (d) Large pore ($lp$) phase. In (c) and (d), the
H are not shown.}
\label{fig:mil53x}
\end{figure}
These MIL-53 compounds exhibit a variety of topologically
equivalent structures with different volumes, but
generally include a narrow pore ($np$) structure
and a large pore ($lp$)
structure, both with formula M$_4$(OH)$_4$(bdc)$_4$ per conventional
unit cell, but with significantly different volumes. In MIL-53(Al),
the phase transition between $np$ and $lp$ forms can be reversibly achieved by
cycling the temperature;\cite{Liu08}
the cell parameter corresponding to the direction of the
short axis of the lozenge pores was found to increase by
87~\% in the $np$-$lp$ transformation.
By way of comparison, the strain variations achieved or predicted in functional
``hard" materials such as
(PbMg$_{1/3}$Nb$_{2/3}$O$_3$)$_{(1-x)}$-(PbTiO$_3$)$_{x}$\cite{Park97} or
BiFeO$_3$\cite{Dieguez11} are much smaller.
The large hysteresis\cite{Liu08} in the $np$-$lp$ phase transition of
MIL-53(Al) indicates that the transition is first-order.
Taking the transition temperature as the midrange of the hysteresis
loop, the transition temperature $T_c$ is approximately 260 K; an estimate based on
experimental sorption measurements places the transition at a somewhat lower
temperature of 203 K.\cite{Boutin10}
For empty MIL-53(Cr), the $lp$ structure is thermodynamically
preferred at all temperatures. In this system, a phase transition
to a $np$ structure has instead been observed in the case of
(1) sorption of a variety of sorbates; (2) pressure.
The hysteresis of the process in each case\cite{Serre07} indicates
again that there is a transition barrier. By fitting sorption isotherms, it
was determined that the free energy difference between the $lp$ and $np$ forms
of MIL-53(Cr) was only about 12 kJ mol$^{-1}$ of Cr$_4$(OH)$_4$(bdc)$_4$.\cite{Coudert08,
DevatourVinot09,Coombes09} An experiment that put the system under
hydrostatic pressure\cite{Beurroies10} came
up with a similar free energy difference.
The phase transition of MIL-53(Al) was explained
by Walker et al.\cite{Walker10} in 2010.
Van der Waals interactions stabilize the $np$
structure at low temperature, and vibrational
entropy drives the structural transition to the
$lp$ phase above $T_c$. Density functional theory (DFT)
phonon calculations were used to quantify the vibrational
entropy.
In that work, however,
the DFT energy and vibrational entropy were
determined for only the $np$ and $lp$ structures.
However, to build an accurate picture of the $np$-$lp$ phase
transition, including the hysteresis
and possible coexistence of $np$ and $lp$
phases,\cite{Triguero12} it is necessary to know
the quantitative free energy landscape over
the {\em full} volume range spanning the $np$ and $lp$ structures.
This free-energy landscape of MIL-53 systems has previously
been modeled in an {\it ad hoc} manner.\cite{Triguero11,Ghysels13}
This paper uses density functional total energy and
phonon linear response calculations to compute the
Helmholtz and Gibbs free energy in MIL-53(Cr)
as a function of temperature, pressure, and cell
volume, under the quasiharmonic approximation.
MIL-53(Cr) was chosen because of its relatively simple
phase transformation behavior and because it is
well-characterized experimentally.
The thermodynamic calculations are performed within the
quasiharmonic approximation. In the quasiharmonic approximation,
the anharmonic lattice dynamics that leads to thermal expansion,
etc., is approximated by harmonic lattice dynamics where the
phonon frequencies are volume-dependent. Suppose that one has a
crystal where the rank-ordered frequencies ${\nu_{\mu} (V)}$ can be
determined for an arbitrarily large supercell (equivalently at
arbitrary points in the Brillouin zone of the primitive cell).
The contribution of phonons to the thermodynamics is then given
well-known expressions.\cite{Maradudin71,vandeWalle02,Fultz10,Huang16}
Defining a dimensionless parameter $x_{\mu}(V,T) =
\frac{h \nu_{\mu}(V)}{k_B T}$, the molar internal energy as a function of
volume and temperature is given by
\begin{eqnarray}
\frac{U}{N}(V,T) = {\rm Lim}_{|a_{\rm min}|\rightarrow \infty}
\frac{1}{N} \bigl(U_0(V) + \nonumber \\
k_B T \sum_{\mu = 4}^{3 N_A}
[\frac{x_{\mu}(V,T)}{2} {\rm coth}(\frac {x_{\mu}(V,T)}{2})]\bigr),
\label{eq:inten}
\end{eqnarray}
the Helmholtz free energy by
\begin{eqnarray}
\frac{F}{N}(V,T) = {\rm Lim}_{|a_{\rm min}|\rightarrow \infty}
\frac{1}{N} \bigl(U_0(V) + \nonumber \\
k_B T \sum_{\mu = 4}^{3 N_A}
[\frac{x_{\mu}(V,T)}{2} + {\rm ln} (1 - e^{-x_{\mu}(V,T)})]\bigr),
\label{eq:helm}
\end{eqnarray}
and the Gibbs free energy is given by
$\frac{G}{N}(V,T) = \frac{F}{N}(V,T) + P V$.
$U_0(V)$ is the ground state energy neglecting zero-point vibrations, $N$
the number of moles and $N_A$ the number of atoms in the supercell,
and the summation begins at $\mu = 4$ to avoid the weak singularity
due to the zero-frequency translational modes.
First principles density functional theory calculations, as encoded in
the {\sc VASP} software~(\onlinecite{Kresse96,disclaim}), were used to
compute $U_0(V)$ and ${\nu_{\mu} (V)}$ for a 152-atom supercell
of MIL-53(Cr), doubled along $c$ so as to make $a$, $b$, and $c$ similar in
magnitude for the $lp$ phase. Two different sets of calculations were performed:
GGA calculations using the PBEsol functional\cite{Perdew08} and meta-GGA calculations
using the PBEsol+RTPSS\cite{Sun11} functionals. These functionals were chosen because
we have had success with them in past studies of microporous
materials.\cite{Cockayne12,Cockayne15} For each level of DFT, the nonlocal
van der Waals interactions were treated using three different approximations
of Grimme et al.: DFT-D2,\cite{Grimme06} DFT-D3,\cite{Grimme10}
and DFT-D3(BJ).\cite{Grimme11}
Anisotropic Hubbard parameters\cite{Liech95} were used for Cr and O atoms (GGA: U(Cr) = 4.0 eV,
J(Cr) = 0.5 eV; metaGGA: U(Cr) = 2.8 eV, J(Cr) = 0.5 eV; U(O) = 7.05 eV).
Spin polarized calculations were performed using the most-stable antiferromagnetic
arrangement of charges on the Cr$^{3+}$ ions.
Further details of the DFT calculations are given in the Supplementary
Information (SI).
Determination of $U_0(V)$ for each functional was done via straightforward
fixed-volume relaxation for (primitive cell) increasing in 50~\AA$^3$ steps
from 650~\AA$^3$ to 1700~\AA$^3$. The phonon frequencies for the
152-atom supercell were calculated using ab initio linear response. As
this method converges toward exact second derivatives of the energy, it is more
accurate than fitting frozen-phonon results. Due to the large
number of degrees of freedom, the phonon calculations are very expensive, and
eventually only three calculations were used for the thermodynamics:
V = 710~\AA$^3$, V = 1200~\AA$^3$, and V = 1506~\AA$^3$. Linear response was only done
using GGA and DFT-D2; the same phonon frequencies $\nu_{\mu}(V)$ were used for each
functional in \eq{helm}; only the $U_0$ changed.
Because the variation in volume between the $np$ and $lp$ phases is so large, one
does not expect the conventional linear Gr\"uneisen approximation for $\nu_{\mu} (V)$
to apply. Instead, we fit the phonon frequencies at intermediate volumes
by fitting to the following physically-motivated expression:
\begin{equation}
\nu_{\mu}^2 (V) = \nu_{\mu \infty}^2 + C_1/V + C_2/V^2 .
\label{eq:phofit}
\end{equation}
The coefficients in~\eq{phofit} were determined by fitting the results for the three
frequencies calculated. If $\nu_{\mu \infty}^2$ in the fit was less than
zero, it was set to zero and the fit recalculated.
Due to computational limitations, it is not possible to calculate larger
supercells for use in~\eq{helm}. Instead, the contribution of optical phonons to
the thermodynamics was approximated by the phonon spectra calculated for
the single 152-atom supercell. The contribution of acoustic phonons to the
thermodynamics was approximated by numerical integration of estimated acoustic
frequencies over the first Brillouin zone. Further details are given in the
Supplementary Information.
First, the phonons were calculated for the $np$ and $lp$
structures. All modes were stable for the $np$ structure.
For the $lp$ structure, instabilities were found. The
most unstable modes, for both the force-constant and dynamical
matrices, were hydrogen ``flopping" modes in which
the H in each hydroxyl group move in the $\pm x$ direction
so as to decrease the distance to a pair of carboxylate
oxygens (\fig{flop}).
Fully relaxing this mode maintains orthorhombic symmetry,
the 152-atom cell is now a primitive cell.
\begin{figure}
\includegraphics[width=42mm]{flop.pdf}
\caption{Local geometry of MIL-53(Cr) (lp) after DFT relaxation of
``H flopping" mode. Each H relaxes to sit approximately 2.4 \AA~from
each of a pair of oxygens (dashed lines); the O are superposed
from this vantage point.}
\label{fig:flop}
\end{figure}
The structure obtained upon relaxation of the flopping instability
was taken as the reference $lp$ structure. To obtain the initial
structure for the fixed volume relaxations used to determine $U_0(V)$,
the ionic coordinates were interpolated (or extrapolated) from the initial
$np$ and $lp$ structures.
\begin{figure}
\includegraphics[width=85mm]{uo.pdf}
\caption{Calculated DFT energy for MIL-53(Cr) at 0 K as a function of
volume for different density functionals, neglecting zero-point motion. Each curve is
scaled so that its minimum is zero.}
\label{fig:uo}
\end{figure}
The $U_0(V)$ determined for the various density functionals are shown in
\fig{uo}. The $F(V)$ for T = 293 K are shown in \fig{helm}.
For every plot in \fig{helm}, there are two minima in the free
energy, corresponding to $lp$ and $np$ structures. The effect of
phonon entropy is to reduce the free energy of the $lp$ structure with
respect to the $np$ structure, as expected. Calculations show that the
free energies for temperatures up to 500 K and pressures between
-30 MPa and 30 MPa maintain two minima for all density functionals tested.
\begin{figure}
\includegraphics[width=85mm]{helm.pdf}
\caption{Calculated Helmholtz free energy for MIL-53(Cr) at 293 K as a function of
volume for different density functionals. Each curve is scaled so that its minimum
is zero. The effect of atmospheric pressure of about 0.1 MPa is negligible on this
scale.}
\label{fig:helm}
\end{figure}
\tab{break} summarizes and compares the results for the
different functionals used. The volumes at which the minima for $U_0$ occur
are given by $V_{0np}$ and $V_{0lp}$. The locations of the minima in
$F$ at room temperature (RT; 293K) are given by $V_{np}(RT)$ and
$V_{lp} (RT)$. The calculated difference in $F$ between the $np$ and $lp$ minima
is $\Delta{F}(RT) = F_{lp}(RT) - F_{np}(RT)$.
The critical pressure $P_c$ is where the calculated Gibbs free energy of the $np$
and $lp$ phases becomes equal at T = 293 K. $G_{b}(RT;P_c)$ is the
calculated free energy barrier between the phases at this pressure.
\begin{table}
\caption{Calculated structural and thermodynamic results for MIL-53(Cr) for different choices of
the density functional. See text for explanation of the column headings.
$V$ are in~\AA$^3$; $F$ and $G$ in kJ mol$^{-1}$
(1 mole = 1 mole of Cr$_4$(OH)$_4$(bdc)$_4$); $P$ in MPa. RT is
room temperature, or 293 K.}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccccccccc}
\hline
xc & vdW & $V_{0np}$ & $V_{0lp}$ & $V_{np} (RT)$ & $V_{lp} (RT)$ & $\Delta{F}$ & $P_c$ & $G_{b}(RT;P_c$) \\
GGA & D2 & 720 & 1533 & 728 & 1534 & +62.3 & -127.0 & 45.6 \\
GGA & D3 & 811 & 1532 & 835 & 1534 & +13.5 & -32.0 & 14.1 \\
GGA & D3(BJ) & 806 & 1443 & 822 & 1483 & +32.1 & -76.8 & 18.7 \\
metaGGA & D2 & 778 & 1461 & 798 & 1466 & +10.4 & -25.6 & 16.0 \\
metaGGA & D3 & 892 & 1505 & 948 & 1512 & -4.9 & 14.6 & 3.2 \\
metaGGA & D3(BJ) & 875 & 1393 & 919 & 1493 & +10.6 & -30.7 & 6.0 \\
\hline
\label{tab:break}
\end{tabular}
}
\end{table}
Substantial differences are seen depending on what density functional is used.
The general trend is for the GGA functionals and the D2 vdW term to give lower
$V_{np}$ and higher $\Delta{F}$ than the metaGGA functionals and D3 or D3(BJ)
choices for the vdW interaction. Which functional gives the best agreement with
experiment? The experimental unit cell volume of the $lp$ phase of MIL-53(Cr) is
1486~\AA$^3$.(Ref.~\onlinecite{Llewellyn08}) The volume of the $np$ phase formed upon
sorption of H$_2$O is 1012 ~\AA$^3$,(Ref.~\onlinecite{Llewellyn08}) but this cannot
be directly compared with the calculation for the empty cell reported here.
As the $np$ phase of MIL-53(Cr) is thermodynamically unstable experimentally, we
take the experimental volume\cite{Liu08,Nanthamathee15} of MIL-53(Al) $np$, 864~\AA$^3$,
and estimate that the volume of MIL-53(Cr) should be about 900~\AA$^3$ due to the larger ionic
radius of Cr$^{3+}$. The best agreement with experiment for the lattice parameters is
for the metaGGA-D3(BJ) parameterization, while the second best is for metaGGA-D3.
On the other hand, the relative stability of the $lp$ phase found experimentally,
$\Delta{F} \approx$ -12.0 kJ mol$^{-1}$ is underestimated by {\em all} the functionals
chosen. The metaGGA-D3 calculation is best in this regard, as it is the only
calculation to yield a negative $\Delta{F}$. All of the metaGGA calculations perform
better than GGA in predicting the relative phase stability. As the metaGGA-D3 and
metaGGA-D3(BJ) have the best agreement with experiment, their low values
of the transition barrier $G_b$, 3.2 to 6.0 kJ mol$^{-1}$ should be considered most
reliable.
It is interesting to put the comparative results in context of previous studies.
In MIL-53, it has previously been found that the D2 vdW overbinds the
$np$ phase;\cite{Haigis14} this work confirms that result. Benchmarking
the performance of DFT calculations
is currently receiving a great deal of attention\cite{Kirklin15,Lejaeghere16,Tran16}.
In Ref.~\onlinecite{Tran16}, over sixty different density functionals are
compared. Although the RTPSS functional is not tested, the related metaGGA functional
TPSS-D3 gives good results for graphite, which suggests that these parameterizations
may work well for MIL-53, where the $np$ phase has benzyl rings of carbon approaching
each other. Further work is needed to make a full comparison among methods because the
current work: (1) includes Hubbard U and J parameters; (2) needs a vdW functional that
reproduces the vdW interactions correctly over a wide range of structural distortion,
not merely at one equilibrium point.
The metaGGA-D3 calculation predicts that the $lp$ phase of MIL-53(Cr) is stable
at room temperature, in agreement with experiment. Interestingly, it predicts
a transition to the $np$ phase below T = 160 K, similar to what actually occurs for MIL-53(Al).
The estimated change in $\Delta{F}$ with temperature is about
-0.036 kJ mol$^{-1}$ K$^{-1}$. Applying this to the experimental $\Delta{F} \approx$
-12.0 kJ mol$^{-1}$, the $lp$ phase is expected to remain stable down to T = 0 K,
albeit with a free energy advantage of less than 2 kJ mol$^{-1}$.
\begin{figure}
\includegraphics[width=85mm]{gibb.pdf}
\caption{Calculated Gibbs free energy for MIL-53(Cr) at 293 K as a function of
volume and pressure for the metaGGA-D3 density functional. Each curve is
scaled so that its minimum is zero.}
\label{fig:gibb}
\end{figure}
The shallowness of the free energy profile suggests that sufficiently large
positive or negative pressure would drive the Gibbs free energy
G(V, T = 293 K) into a regime where it has only one minimum corresponding to
either a $np$ or a $lp$ structure. In \fig{gibb}, we show G(V, T = 293 K) for
various pressures -80 MPa to 80 MPa, using the metaGGA-D3 results.
At pressures above about 60 MPa, there is a unique minimum at the $np$ phase;
below about -40 MPa, there is one minimum at the $lp$ phase. If the zero in
pressure is shifted to correct for the error in the metaGGA-D3 $\Delta{F}$ with
respect to experiment, the predicted pressures are shifted to about 80 MPa and -20 MPa,
respectively. Of course the prediction of the pressures at which the free energy converts to
a single minimum only sets an upper bound on the width of the pressure hysteresis loop;
in practice, fluctuations will cause the transitions to occur at less extreme pressures.
With this is mind, experimental transition pressures for the hysteresis loop
of roughly 50 MPa and 20 MPa for MIL-53(Cr)\cite{Rodriguez16} are
consistent with the DFT results. Note that negative pressures do have physical
relevance in microporous materials in the case of sorption- the effective solvation pressure
can be either positive or negative depending on the sorbate concentration.\cite{Ravikovitch06}
\begin{figure}
\includegraphics[width=85mm]{cell.pdf}
\caption{Calculated MIL-53(Cr) lattice parameters and cell angle $\beta$ versus volume.}
\label{fig:cell}
\end{figure}
In \fig{cell}, the crystallographic data for the DFT metaGGA-D2 structural
relaxations are shown. The lattice parameters are scaled to the volume of the
conventional unit cells. To make the orthorhombic-monoclinic transition
clear, the monoclinic cell parameters $a$ and $\beta$ are for an unconventional
body-center monoclinic setting.
The orthorhombic-monoclinic transition occurs at $V \sim 1500$~\AA$^3$,
intriguingly close to the experimental cell volume.
In addition to the structural transitions, there are three regimes in
the behavior of the lattice constants: (1) below about 850~\AA$^3$, $a$ $b$
and $c$ all increase with volume; (2) between about 850~\AA$^3$ and
1650 \AA$^3$, $a$ decreases with volume $b$ increases with volume, and
$c$ is nearly flat as the structure flexes; (3) above about 1650 \AA$^3$, all
lattice parameters increase again. The crossover between regimes
(2) and (3) does not occur at the same volume as the
monoclinic-orthorhombic transition. To a first approximation, the free energy
is nearly flat in regime (2) and increases rapidly above and below this range.
The three regimes agree qualitatively with those seen in a recent
experiment on the related material MIL-53(Al) under pressure.\cite{SerraCrespo15}
To summarize, we used density functional theory
total energy and linear response phonon calculations
to compute the free energy profile of MIL-53(Cr)
under the quasiharmonic approximation.
The density functionals that best match the experimental
results give remarkably flat free energy profiles,
with a transition barrier of only about a 3 to 6 kJ mol$^{-1}$
between the the narrow pore and large pore phases.
I thank Laura Espinal, Kevin F. Garrity,
and Winnie Wong-Ng for helpful discussions.
This paper was published as J. Phys. Chem. C 2017,
{\bf 121}, 4312-4317 (DOI:10.1021/acs.jpcc.6b11692).
The Supporting Information is available free of charge on the
ACS Publications website at the DOI given immediately
above.
|
\section*{Generation and distillation of remote entangled states}
To run entanglement distillation on a quantum network, several copies of a raw entangled state must first be shared between the nodes. This can be achieved using a network primitive of two nodes with two qubits each: a communication qubit with an optical interface for generating remote entanglement and a memory qubit for storage (\fref{Fig1}B). First the communication qubits run the entangling protocol, which due to photon loss is intrinsically probabilistic. After photon detection heralds the generation of a raw entangled state on the communication qubits, this state is swapped onto the memory qubits. The communication qubits are then used to generate a second raw entangled state. At this point, the network nodes share two nominally identical copies of the raw state, from which an entangled state of higher fidelity can be distilled. This protocol thus exploits the combination of heralded generation of remote entanglement with robust quantum state storage, high-fidelity quantum logic gates and non-demolition qubit readout within each node.
These demanding experimental requirements have so far limited the exploration of distillation on entangled qubits to four ions within a single node~\cite{reichle_experimental_2006} and to all-photonic protocols without memories in which the distilled state was unavoidably lost upon success~\cite{pan_experimental_2003,kwiat_experimental_2001,pan_multiphoton_2012}. As an important step towards the desired quantum network, heralded entanglement between distant stationary qubits has recently been achieved with ions, atoms, NVs, quantum dots and superconducting qubits~\cite{moehring_entanglement_2007,hofmann_heralded_2012,bernien_heralded_2013,narla_robust_2016,delteil_generation_2016,stockill_phase-tuned_2017}. However, the potential memory qubits investigated so far in conjunction with these protocols~\cite{pfaff_unconditional_2014,hucul_modular_2015} suffered from rapid dephasing during remote entangling attempts due to unwanted couplings, thus precluding the generation of multiple remote entangled states as required for distillation.
\begin{figure*}[hbtp!]
\centering
\includegraphics[width=\textwidth]{Fig1_arXiv-01.png}
\caption{\textbf{Entanglement distillation on a quantum network.} (\textbf{A}) Working principle: a remote entangled state of higher quality (right) is distilled via local operations and classical communication from several lower-quality states (left) that are shared between remote qubits (depicted as colored spins).
(\textbf{B}) Protocol overview. Each network node consists of a communication qubit (purple) and a memory qubit (yellow). First, the communication qubits are prepared in a remote entangled state by generating entanglement between a photon (red wave packet) and the spin, interfering the optical modes on a beam splitter (gray cube) and subsequently detecting a single photon \ding{172}. Next, the remote entangled state (purple waves represent entanglement) is swapped onto the memory qubits \ding{173}, followed by another round of entangled state generation \ding{174}. Finally, local operations (black circuit) distill a state of higher fidelity \ding{175}.
(\textbf{C}) Gate circuit implementing steps \ding{172}-\ding{175}. We include the photonic modes of each setup (red wave packets) in which a qubit is encoded such that vacuum and a single photon represent $\ket{0}$ and $\ket{1}$ respectively. Entanglement between the electron spin and photonic mode is experimentally realized by an optical $\pi$-pulse (depicted as CNOT symbol). The photonic Bell-state projection $\ket{\Psi^\pm}\bra{\Psi^\pm}$ is probabilistically realized by a beam splitter and subsequent detection of a single photon. Dashed-bordered gates indicate phase-shifts of the memory due to free evolution during entangling attempts. Colored boxes indicate logical blocks of the circuit and are used throughout the manuscript.}
\label{fig:Fig1}
\end{figure*}
We realize the distillation of entangled states on an elementary quantum network consisting of a pair of two-qubit solid-state nodes separated by two meters (\fref{Fig1}B). We achieve this result by implementing a single-photon-based entangling protocol using diamond electron-spin qubits (communication qubits) while capitalizing on recent progress on quantum control~\cite{cramer_repeated_2016} and robust state storage~\cite{reiserer_robust_2016} in nuclear-spin-based quantum memories. Real-time feedback is implemented to compensate memory qubit phase-shifts induced by the probabilistic nature of the remote entangling protocol. As an immediate advantage, the demonstrated protocol distinctly increases the efficiency of entanglement generation compared to the standard two-photon-coincidence protocols used in earlier works~\cite{hucul_modular_2015,hensen_loophole-free_2015}, while removing the optical path-length dependence of stand-alone probabilistic single-photon protocols~\cite{slodicka_atom-atom_2013,delteil_generation_2016}. More generally, by demonstrating the key capabilities for a quantum network in a single experiment, we realize a universal backbone that opens the door to extended quantum networks powered by high-quality remote quantum entanglement.
\section*{Quantum network nodes}
Our implementation of a quantum network node employs a nitrogen-vacancy (NV) electron spin in diamond as a communication qubit and a nearby carbon-13 nuclear spin as a memory qubit. The diamond chips holding these qubits reside in individual closed-cycle cryostats ($T=4\,\mathrm{K}$) that are separated by two meters~\cite{aa_suppmatt_2017}. The electron spin state is manipulated using amplitude-shaped microwave pulses. Electron spin decoherence occurs on timescales exceeding a millisecond and has negligible impact on the presented results.
Spin-selective resonant optical excitation enables high-fidelity initialization and single-shot non-demolition read-out of the electron spin~\cite{blok_manipulating_2014}, as well as generation of spin-photon entanglement for connecting distant nodes~\cite{bernien_heralded_2013}. We employ nuclear spins with intrinsic dephasing times $T_\mathrm{2}^*$ of $3.4(1)$ ms and $16.2(3)$ ms for node A and B, respectively~\cite{aa_suppmatt_2017}. We implement universal control on each of these nuclear spin qubits by exploiting its hyperfine coupling to the electron spin through recently developed dynamical-decoupling-based gate sequences~\cite{taminiau_universal_2014}. This complete quantum toolbox enables the implementation of all four steps in the distillation protocol.
\fref{Fig1}C shows the compilation of the full gate circuit into the quantum control operations of our platform. This compilation maximizes the repetition rate and minimizes the number of local quantum gates following the generation of the first remote state. In particular, by initializing the memory qubit at the start of the protocol we are able to implement the SWAP operation with just two conditional quantum gates instead of the three that would be required for arbitrary input states~\cite{aa_suppmatt_2017}. Note that our SWAP implementation maps the communication qubit energy eigenstates onto memory superposition states $\ket{\pm X} \equiv (\ket{0}\pm \ket{1})/\sqrt{2}$.
To benchmark the performance of the local quantum logic we execute a combination of the SWAP (yellow box in \fref{Fig1}C) and the gates of the distillation step (purple box in \fref{Fig1}C) to generate a maximally-entangled Bell-state between the communication and memory qubits (see \fref{Fig2}A). The full density matrix of the resulting two-qubit state is reconstructed via quantum state tomography (QST, see ref.~\cite{aa_suppmatt_2017} for further details). We find a fidelity with the ideal Bell state of $0.96(1)$ ($0.98(1))$ for node A (B) indicating high-quality operations in both nodes (\fref{Fig2}B).
\begin{figure}[ht!]
\centering
\includegraphics[width=\columnwidth]{Fig2_arXiv-01.png}
\caption{ \textbf{Benchmarking local control.} (\textbf{A}) Gate circuit for entanglement generation within one node. All local operations of the purification circuit are employed to generate a entangled state between communication and memory qubit. Color coding of local operations corresponds to \fref{Fig1}C. (\textbf{B}) Absolute value of the real part of the local density matrix obtained via sequential QST~\cite{aa_suppmatt_2017}. We find fidelities with the desired entangled state of $0.96(1)$ (node A) and $0.98(1)$ (node B). Transparent bars give the values of the ideal state.}
\label{fig:Fig2}
\end{figure}
\section*{Robust storage of quantum information}
A critical capability for the network nodes is the robust storage of quantum information in the memories while the communication qubits are used to generate remote entangled states. This requires the memory qubits to have long coherence times and be resilient to operations on the communication qubit. The generation of remote entanglement, in particular, poses two challenges as its probabilistic nature means that an \emph{a priori} unknown number of attempts is required.
First, each failed entangling attempt leaves the communication qubits in an unknown state which necessitates a reset by optical pumping. This reset is a stochastic process which, in combination with the always-on hyperfine interaction between communication and memory qubit, causes dephasing of stored memory states~\cite{reiserer_robust_2016,jiang_coherence_2008}. Here we employ memories with a small parallel hyperfine coupling so that the precession frequency of these memories exhibits only a weak dependency $\Delta \omega$ on the state of the communication qubit during the repumping process of a few hundred nanoseconds ($\Delta \omega_\mathrm{A} = 2\pi \cdot 22.4(1)\,\mathrm{kHz}$ and $\Delta \omega_\mathrm{B} = 2\pi \cdot 26.6(1)\,\mathrm{kHz}$). Decoherence via the perpendicular hyperfine component is suppressed by an applied magnetic field of about $40\,\mathrm{mT}$.
\begin{figure}[tb!]
\centering
\includegraphics[width=\columnwidth]{Fig3_arXiv-01.png}
\caption{ \textbf{Quantum state storage during entangling operations.} (\textbf{A}) Real-time feedback circuit for memory qubits. We initialize memory A/B, which then experiences a phase-shift of $\varphi_{A/B}$ per executed entangling attempt. After reinitialization of the communication qubit, the distillation step of the protocol is performed (\fref{Fig1}C). Blue-rimmed gate indicates the feedback. (\textbf{B}) Memory state as a function of the number of entangling attempts. The oscillation observed without feedback (orange) is successfully compensated (blue) by the feedback. Solid lines are fits to the data.
(\textbf{C}) Memory lifetime of node A (triangles) and node B (circles). We initialize the memory in one of the six cardinal states of the Bloch sphere (see right panel, where all non-relevant expectation values are assumed to be zero), sweep the number of entangling attempts, apply feedback, and read-out the relevant expectation value. The average state fidelities are separately plotted for phase-sensitive superposition states (blue) and phase-insensitive eigenstates (green). Blue solid lines depict a generalized exponential fit~\cite{aa_suppmatt_2017}. The decay is limited by the stochastic repumping process, microwave pulse errors and/or environmental dephasing. The color gradient of the left and right panel match to facilitate comparisons. Error bars represent one standard deviation.}
\label{fig:Fig3}
\end{figure}
Second, the interaction between communication and memory qubit leads to a deterministic phase-shift $\varphi_{A/B}$ on the memory per entangling attempt. Since it is unknown which entangling attempt will herald success, real-time feedback on the memory is required to compensate for these phase-shifts before the final two-qubit gate of \fref{Fig1}C is applied. In addition, the feedback must preserve the coherence of the communication qubit as it holds the second copy of the raw entangled state. We realize such real-time feedback through dynamical decoupling of the electron spin synced with the nuclear spin precession frequency that induces an electron-state-independent phase gate on the memory~\cite{taminiau_universal_2014}. At each node, the number of entangling attempts until success $N$ is tracked by a microprocessor that terminates the subsequent decoupling sequence when the desired rotation $\mathrm{R_z}(\varphi_{A/B})^N$ has been applied. Ideally, this leaves the memory with the desired phase relation regardless of the number of entangling attempts. We calibrate and verify this feedback at each node separately (see \fref{Fig3}A and \fref{Fig3}B) and measure a negligible effect on the memory state fidelity while the state of the communication qubit is preserved as desired.
With this feedback realized, we investigate the robustness of the memory as a function of the elapsed entangling attempts. We initialize the memory in one of the six cardinal states of the Bloch sphere ($\ket{0}$,$\ket{1}$,$\ket{\pm \mathrm{X}}$ and $\ket{\pm \mathrm{Y}} \equiv (\ket{0} \pm i\ket{1})/\sqrt{2}$), execute a number of entangling attempts followed by phase-feedback and measure the relevant memory expectation value (see \fref{Fig3}C). We observe that dephasing-sensitive states $\ket{\pm \mathrm{X}},\ket{\pm \mathrm{Y}}$ decay with $1/e$-values of $273(5)$ ($272(4)$) entangling attempts in node A (node B) whereas the energy eigenstates $\ket{0},\ket{1}$ are preserved with high fidelity as expected. The memories thus provide faithful storage during remote entangling attempts.
\section*{Experimental entanglement distillation}
With local control and storage in place, we now turn towards the execution of the full distillation protocol. Following Ref. \cite{campbell_measurement-based_2008}, we generate the remote states that provide the resources for distillation by first initializing both communication qubits in a superposition with variable angle $\theta$, $\ket{\theta} \equiv \sin\theta\ket{0}-i\cos\theta\ket{1}$. Subsequent optical excitation for state $\ket{0}$ and overlap of the emission of both communication qubits on a beam splitter (see steps $1$ and $3$ of \fref{Fig1}C) generates the raw remote state $\rho_\mathrm{raw}$ if a single photon is detected \cite{campbell_measurement-based_2008}. For equal and small detection probabilities for both nodes and negligible dark counts, $\rho_\mathrm{raw}$ reads:
\begin{equation}
\rho_\mathrm{raw} = \left(1-\sin^2 \theta \right)\ket{\Psi^\pm_\phi}\bra{\Psi^\pm_\phi}+
\sin^2 \theta \,\ket{0,0}\bra{0,0}.
\label{eq:rawstate}
\end{equation}
The states $\ket{\Psi^\pm_\phi}\equiv (|01\rangle \pm e^{i\phi} |10\rangle ) / \sqrt{2} $ are entangled states, with a relative phase depending on which detector clicked ($\pm$) and an additional internal phase $\phi$ due to the unknown path length between both emitters and the beam splitter. The fraction of the non-entangled admixture $\ket{0,0}\bra{0,0}$ can be directly controlled through the choice of the initial communication qubit state $\ket{\theta}$; note that the choice of $\ket{\theta}$ also affects the probability of successful entanglement generation (scaling as $\sin^2\theta$). We next swap the raw state onto the memories such that the communication qubit is free for another round of remote state generation (step $2$ in \fref{Fig1}C). Once a second state is successfully generated (step $3$), we apply a conditional quantum gate within each node and read out the communication qubits in a single shot. Owing to the quantum non-demolition nature of this readout the memory qubits do not experience additional dephasing during this step \cite{blok_manipulating_2014}. Readout of the communication qubit projects the memories into one of four states depending on the readout results~\cite{aa_suppmatt_2017}:
\[\def1.5{1.5}
\begin{array}{rl}
(0_A,0_B):& \frac{1}{2} \,\cos^4\theta \,U\ket{\Psi^\pm_0}\bra{\Psi^\pm_0}U^\dagger, \\
(0_A,1_B):& \frac{1}{2} \,\sin^2\theta \cos^2\theta \, U(\ket{0,1}\bra{0,1}+\ket{1,1}\bra{1,1})U^\dagger,\\
(1_A,0_B):& \frac{1}{2} \,\sin^2\theta \cos^2\theta \,U (\ket{1,0}\bra{1,0}+\ket{1,1}\bra{1,1})U^\dagger,\\
(1_A,1_B):&U\left(\sin^4\theta\ket{1,1}\bra{1,1} + \frac{1}{2} \cos^4\theta \ket{\Psi^\pm_{2\phi}}\bra{\Psi^\pm_{2\phi}}\right)U^\dagger.
\label{eq:PurifyOutcome}
\end{array}
\]
Here the states are left unnormalized; their traces indicate their probabilities of occurrence. The unitary $U$ corresponds to a Hadamard gate on each memory that arises from the swapping operation in each node. Observation of the readout combination ($0_A$,$0_B$) heralds successful distillation and leaves the system in the state
\begin{equation}
\ket{\psi}_\mathrm{c}\otimes \ket{\psi}_\mathrm{m} = e^{i\phi}\ket{0,0} \otimes U\ket{\Psi^\pm_0}
\end{equation}
with the relative phase of the final Bell state given by the photon detection signature, i.e. the photons in step 1 and 3 were detected in the same (+) or in different (-) output ports. Importantly, the protocol is agnostic to correlated dephasing of the raw states and is therefore only sensitive to optical path length drifts that occur within an individual run of the protocol~\cite{campbell_measurement-based_2008}. This is in stark contrast to probabilistic single-photon protocols~\cite{cabrillo_creation_1999,slodicka_atom-atom_2013,delteil_generation_2016,stockill_phase-tuned_2017} that require path length stabilization over the full course of data acquisition.
The experimental implementation of entanglement generation requires that the communication qubits' optical transitions are kept on resonance despite shot-to-shot fluctuations and long-term drifts of the respective local charge environments. We employ an automatic feedback loop and resonance search routine to compensate for charge jumps such that the experiment is push-button and runs without human intervention. To further optimize the data rate we bound the number of remote entangling attempts to 1000 for step 1 and up to 500 rounds for step 3, leading to event rates (i.e. two remote states were successfully generated) of around 10~Hz. These bounds are a compromise between maximizing the success probability (favoring more attempts) and minimizing effects of drifts and of memory decoherence (favoring fewer attempts).
\section*{Distillation results}
We start by running the complete protocol using $\theta=\pi/6$ and perform full quantum state tomography on the distilled state. This way, using the complete information obtained on the resulting output, we can verify whether the protocol works as desired. \fref{Fig4}A shows the resulting data for a maximum of $50$ entangling attempts in the second round of state generation. This truncation yields optimal state storage during each run of the protocol~\cite{aa_suppmatt_2017}. Quantitatively, the measured fidelity with the ideal Bell state of $0.65(3) > 0.5$ proves entanglement of the distilled state. Furthermore, the density matrix has high populations in the Bell-state-subspace only, showing that the distillation successfully diminishes the separable admixture.
To gain further insight into the performance of the protocol we measure the fidelity of the distilled state for different amounts of the separable admixture in the raw states; i.e. for different $\theta$ (see \fref{Fig4}B, blue dots). The results are again truncated after a maximum of $50$ entangling attempts in the second round. The state fidelities are averaged over both detection signatures.
\begin{figure} [htbp!]
\centering
\includegraphics[width=\columnwidth]{Fig4_arXiv-01.png}
\caption{\textbf{Experimental realization of entanglement purification.} (\textbf{A}) Two-qubit density matrices for $\theta = \pi/6$ and a maximum of 50 entangling attempts in the second round. Right panel: different detectors clicked. Left panel: the same detector clicked twice. We find a fidelity with the ideal state of $0.65(3)$ for both states. Transparent bars represent the ideal state. (\textbf{B}) Fidelity with the ideal state as a function of $\theta$ for a maximum of $50$ entangling attempts in the second round. Blue data is the two-memory state fidelity. Dashed lines are derived from our model~\cite{aa_suppmatt_2017}. Purple data is the measured raw state fidelity on the communication qubits. Solid orange (purple) line is the modeled fidelity of the raw state on the memories (communication qubits) that would be obtained if the initial internal phase was known. The memory state is calculated for the average number of entangling attempts until success ($25$). The orange shaded region is the modeled memory fidelity for minimal ($0$ attempts) and maximal ($50$ attempts) dephasing. Fidelities were obtained by measuring the expectation values $\langle \hat{X}\hat{X} \rangle$, $\langle \hat{Y}\hat{Y} \rangle$ and $\langle \hat{Z}\hat{Z} \rangle$. We denote the Pauli operators as $ \hat{X}$, $\hat{Y}$ and $\hat{Z}$. (\textbf{C}) State decay for $\theta = \pi/6$. Data are binned according to the number of second entanglement generation attempts until success. Shown are the state fidelity (blue) and the absolute value of the relevant expectation values. The dashed lines are derived from our theoretical model~\cite{aa_suppmatt_2017}. Error bars represent one standard deviation.}
\label{fig:Fig4}
\end{figure}
The hallmark of successful distillation is an increase in fidelity of the distilled state compared to that of the raw states. Whereas in the textbook description both raw states are assumed to be equal, in our experiment they are different due to imperfections in the swap operation and memory storage that only affect the raw state held by the memories, and path length variations on short timescales that only affect the raw state held by the communication qubits. To make a meaningful comparison we therefore consider the state fidelities of each of these raw states separately.
Because of the unreferenced internal phase of the raw states, all coherences are washed out due to optical path length variations. Direct tomography will therefore yield state fidelities that cannot surpass $0.5$. As a result, the measured fidelities of the distilled states far exceed the electron state fidelities measured after step 1 (\fref{Fig4}B, orange dots). Although these numbers reflect the current experiment, we now turn to a more strict comparison by taking into account that the internal phase may become accessible in future experiments through optical path stabilization.
We model the raw state fidelities at the start of the distillation step (step 4) using independently determined parameters under the assumption of a perfectly known initial path length difference (\fref{Fig4}B, solid purple line for raw state on the electrons and solid orange line for raw state on the nuclei) \cite{aa_suppmatt_2017}. For small values of $\theta$ (small separable admixtures) the fidelity increase due to distillation is offset by the errors introduced with the additional quantum operations of the distillation step. However, we find that for larger values of $\theta$ the distilled state fidelity significantly surpasses both of the raw state fidelities (see~\cite{aa_suppmatt_2017} for hypothesis test). This result demonstrates the realization of entanglement distillation on our elementary quantum network.
For a more detailed understanding of the different error sources contributing to the measured fidelity, we develop an extensive model of the full protocol using independently measured quantities and two free parameters: one factor accounting for additional memory control errors and the second for phase fluctuations of the raw states~\cite{aa_suppmatt_2017}. We find good agreement between the modeled fidelity and the data for each of the different separable admixtures (see \fref{Fig4}B blue dashed line and~\cite{aa_suppmatt_2017}) and for the evolution of the correlations with number of entangling attempts (\fref{Fig4}C). The model indicates that the state fidelities are mainly limited by memory qubit dephasing and control errors as well as non-zero two-photon distinguishability. The latter effect, quantified by a measured two-photon interference visibility of 0.73(3)~\cite{aa_suppmatt_2017}, is especially harmful in the above comparison with the raw state, as this occurs twice for the distillation protocol but only once for the raw state generation. A visibility of 0.9 as observed on different NV center pairs~\cite{hensen_loophole-free_2015} would thus yield an even stronger entanglement enhancement.
\section*{Ebit rate}
Previously demonstrated entangling protocols based on two-photon coincidences~\cite{hucul_modular_2015,hensen_loophole-free_2015} require steps 1 and 3 to succeed in subsequent attempts leading to a success probability scaling with the square of the photon detection probability $p_{\mathrm{det}}$. In contrast, the distillation protocol allows step 3 to succeed in one of many attempts following success in step 1, leading to a success probability scaling linearly with $p_{\mathrm{det}}$ in the ideal case. Given that in a typical quantum network setting $p_{\mathrm{det}}$ will be small (in our case $p_{\mathrm{det}} \approx 10^{-3}$), the distillation protocol can provide a distinct rate advantage despite the overhead of the additional local quantum logic.
\begin{figure}[tbp!]
\centering
\includegraphics[width=\columnwidth]{Fig5_arXiv-01.png}
\caption{\textbf{Ebit rate comparison.} Ebit rate as a function of excitation angle $\theta$. The blue data is derived from the measured success rates and state fidelity over the entire data set~\cite{aa_suppmatt_2017}. The blue dashed line is the estimated ebit rate for distillation including the overhead of local operations. The solid (dashed) orange line gives the estimated ebit rate of a standard two-photon protocol including (excluding) imperfections. All data are averaged over both detection signatures. Error bars represent one standard deviation.}
\label{fig:Fig5}
\end{figure}
To quantitatively compare our results with two-photon-coincidence protocols we upper bound the rate $r$ of entangled bit (ebit) generation for each protocol using $r = \nu E_N$ with the logarithmic negativity $E_N$ and the rate of success $\nu$. \fref{Fig5} compares the ebit rate of the presented distillation protocol to the modeled rate of the Barrett-Kok two-photon-coincidence protocol~\cite{barrett_efficient_2005} used in earlier experiments on NV centers~\cite{bernien_heralded_2013,hensen_loophole-free_2015}. We find that the distillation protocol (blue dots) outperforms the two-photon-coincidence protocol for identical experimental conditions, not only when assuming the measured two-photon indistinguishability (orange solid line) but even for the case that the two-photon-coincidence protocol would be able to access perfect two-photon indistinguishability (orange dashed line).
\section*{Conclusion and outlook}
The combination of generating, storing and processing remote entangled qubits as demonstrated in the current distillation experiment provides a universal primitive for realizing extended quantum networks. The distillation itself is a powerful method to counteract unavoidable decoherence as entanglement is distributed throughout the network. Also, the protocol enables a speedup of entanglement generation that can be harnessed in related platforms such as other solid-state defect centers~\cite{christle_isolated_2015} and trapped ions~\cite{monroe_scaling_2013}. Future improvements can be achieved by encoding qubits into decoherence-protected subspaces~\cite{reiserer_robust_2016}, by using isotopically purified materials with longer qubit dephasing times~\cite{balasubramanian_ultralong_2009,maurer_room-temperature_2012,tyryshkin_electron_2012}, by implementing a faster reset or a measurement-based reset of the communication qubit and by increasing the entangling rates through photonic cavities~\cite{faraon_resonant_2011,sipahigil_integrated_2016}. Furthermore, the techniques employed in recent demonstrations of multi-qubit control and quantum error correction on a 4-qubit node~\cite{cramer_repeated_2016,kalb_experimental_2016} are fully compatible with the current experiment, thus highlighting the potential for scaling to more qubits and extending network functionality in the near future. Finally, the methods developed here open the door to exploration and utilization of many-particle entanglement on a multi-node quantum network.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.